www.nature.com/nature Vol 464 | Issue no.

7286 | 11 March 2010

Climate of fear
The integrity of climate research has taken a very public battering in recent months. Scientists must now
emphasize the science, while acknowledging that they are in a street fight.

limate scientists are on the defensive, knocked off balance Senator James Inhofe (Republican, Oklahoma), the leading climate

C by a re-energized community of global-warming deniers who, by sceptic in the US Congress, who labelled several respected climate
dominating the media agenda, are sowing doubts about the fun- scientists as potential criminals — nonsense that was hardly a surprise
damental science. Most researchers find themselves completely out of considering the source. Some scientists have responded by calling for
their league in this kind of battle because it’s only superficially about the a unified public rebuttal to Inhofe, and they have a point. As a mem-
science. The real goal is to stoke the angry fires of talk radio, cable news, ber of the minority party, Inhofe is powerless for now, but that may
the blogosphere and the like, all of which feed off of contrarian story one day change. In the meantime, Inhofe’s report is only as effective
lines and seldom make the time to assess facts and weigh evidence. as the attention it receives, which is why scientists need to be careful
Civility, honesty, fact and perspective are irrelevant. about how they engage such critics.
Worse, the onslaught seems to be working: some “Scientists must The core science supporting anthropogenic global
polls in the United States and abroad suggest that it is not be so naive as to warming has not changed. This needs to be stated again
eroding public confidence in climate science at a time assume that the data and again, in as many contexts as possible. Scientists
when the fundamental understanding of the climate speak for themselves.” must not be so naive as to assume that the data speak
system, although far from complete, is stronger than for themselves. Nor should governments. Scientific
ever. Ecologist Paul Ehrlich at Stanford University in California says agencies in the United States, Europe and beyond have been oddly
that his climate colleagues are at a loss about how to counter the attacks. silent over the recent controversies. In testimony on Capitol Hill last
“Everyone is scared shitless, but they don’t know what to do,” he says. month, the head of the US Environmental Protection Agency, Lisa
Researchers should not despair. For all the public’s confusion about Jackson, offered at best a weak defence of the science while seeming to
climate science, polls consistently show that people trust scientists distance her agency’s deliberations from a tarnished IPCC. Officials
more than almost anybody else to give honest advice. Yes, scientists’ of her stature should be ready to defend scientists where necessary,
reputations have taken a hit thanks to headlines about the leaked and at all times give a credible explanation of the science.
climate e-mails at the University of East Anglia (UEA), UK, and an These challenges are not new, and they won’t go away any time soon.
acknowledged mistake about the retreat of Himalayan glaciers in a Even before the present controversies, climate legislation had hit a
recent report from the Intergovernmental Panel on Climate Change wall in the US Senate, where the poorly informed public debate often
(IPCC). But these wounds are not necessarily fatal. leaves one wondering whether science has any role at all. The IPCC’s
To make sure they are not, scientists must acknowledge that they fourth assessment report had huge influence leading up to the climate
are in a street fight, and that their relationship with the media really conference in Copenhagen last year, but it was always clear that policy-
matters. Anything strategic that can be done on that front would be makers were reluctant to commit to serious reductions in greenhouse-
useful, be it media training for scientists or building links with cred- gas emissions. Scientists can’t do much about that, but they can and
ible public-relations firms. In this light, there are lessons to be learned must continue to inform policy-makers about the underlying science
from the current spate of controversies. For example, the IPCC error and the potential consequences of policy decisions — while making
was originally caught by scientists, not sceptics. Had it been promptly sure they are not bested in the court of public opinion. ■
corrected and openly explained to the media, in full context with the
underlying science, the story would have lasted days, not weeks. The
IPCC must establish a formal process for rapidly investigating and,
when necessary, correcting such errors.
The unguarded exchanges in the UEA e-mails speak for them-
Scientific glasnost
selves. Although the scientific process seems to have worked as it Russia’s scientific reputation will continue to
should have in the end, the e-mails do raise concerns about scientific dwindle unless it embraces international research.
behaviour and must be fully investigated. Public trust in scientists
is based not just on their competence, but also on their perceived ver since the Soviet Union fell apart in 1991, Russian leaders
objectivity and openness. Researchers would be wise to remember
this at all times, even when casually e-mailing colleagues. E have been vowing to transform their old-line, industrial society
into a modern, knowledge-based economy driven by innova-
US scientists recently learned this lesson yet again when a private tive science and technology. The current Russian president, Dmitry
e-mail discussion between leading climate researchers on how to deal Medvedev, has repeated that ambition frequently — not least as a
with sceptics went live on conservative websites, leading to charges way to overcome Russia’s dependence on oil and gas exports. Unfor-
that the scientific elite was conspiring to silence climate sceptics (see tunately, that transformation continues to be hobbled by outdated
page 149). The discussion was spurred by a report last month from attitudes at the top of Russia’s academic hierarchy.
141
© 2010 Macmillan Publishers Limited. All rights reserved
EDITORIALS
A small, but telling example came to light last month when the
popular online newspaper gazeta.ru published an interview with Yuri
Osipov, president of the Russian Academy of Sciences in Moscow.
Pressed by the reporter about the very low citation rate for articles
published in Russian-language science journals, Osipov dismissed
the relevance of citation indices, questioned the need for Russian
scientists to publish in foreign journals and said that any top-level
specialist “will also study Russian and read papers in Russian”.
From anyone else, such a response might be dismissed as an off-
hand comment, perhaps reflecting a bit of stung national pride. But
Osipov is head of the largest and most powerful research organiza-
tion in Russia, the employer of around 50,000 scientists in more
than 400 research institutes, and the publisher of some 150 Russian-
language research journals. What he says and thinks has a big effect
on Russian science. Moreover, the undercurrent of scientific nation-
alism in his remarks is widely shared by other senior members of the
academic establishment — many of whom are products of Soviet
times, when Russian science was pretty much an all-Russian affair
(see Nature 449, 524–527, 528–529; 2007).
Such parochialism is hopelessly at odds with any dreams of a
knowledge-based economy. The knowledge in question flows from
basic research and technological innovation, which have long since
moved beyond being just national endeavours. If nothing else, inter-
national scrutiny and feedback are essential for winnowing the good
ideas from the dead ends. And, as Osipov himself acknowledged
Europe’s research future
The region’s member states must follow through on
their political and scientific commitments.
t is hard to believe on a first reading of the European Commission’s
I 3 March proposal for ‘Europe 2020’ — the new ten-year economic
strategy for the European Union (EU) — that it is important for
scientists. In bland-yet-grandiose prose that invites the eye to slide
right off the page, it defines a strategy for Europe to survive the
current financial crisis and “emerge stronger”. It pompously pledges
to reform research and development and innovation systems “to
foster excellence and smart specialisation”.
But this draft is indeed important for scientists. It endorses
research as the basis of an economically and socially strong
Europe. It maintains the EU’s goal of raising research expenditure
to 3% of its gross domestic product. It recommits EU member
states to the concept of the European Research Area, which seeks
to remove legal hindrances to the free movement of researchers
across the region. It endorses the concept of a single, European-
wide patent system, which is badly needed as an alternative to
the current costly system that requires patents to be registered in
individual countries. And it explicitly reiterates the EU commit-
ment to achieve a significant reduction in greenhouse-gas emis-
sions, while promoting the development of clean and efficient
energy sources.
142
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
in the interview, English, not Russian, is the international language
of science.
Russian science is already lagging behind that of other nations.
According to an analysis published in January by Thomson Reuters,
Russia produced just 2.6% of the research papers published between
2004 and 2008 and indexed by the firm — fewer than China (8.4%)
and India (2.9%) and only slightly more than the Netherlands
(2.5%). Moreover, Russia’s publication output has remained almost
flat since 1981, even as the output of nations such as India, Brazil
and China was exploding. The situation is so bleak that in October
last year, 185 Russian expatriate scientists signed an open letter
to Medvedev and Prime Minister Vladimir Putin warning of an
imminent collapse of Russian science unless something was done
to improve the inadequate funding, strategic planning and teach-
ing of science.
Self-imposed scientific isolationism can only make matters worse
— and accelerate the already large emigration of Russian scientists
seeking better opportunities in the West. And those who remain
in Russia are also starting to recognize the danger. Many young
researchers now eagerly collaborate with Western groups. And many
older Russian professors continue to produce excellent science under
often difficult conditions. They know very well what a grave dis-
service they would do to their students by asking them to publish
in low-profile journals for the supposed sake of national pride. The
answer isn’t to close Russia in, but to open it up. ■
These targets are worthy indeed. But then, they were worthy when
they were outlined in the Lisbon Strategy, the ten-year economic
plan that was adopted by the EU in 2000. Today that strategy is
widely regarded as a failure, largely because the EU member states
seemed content to bask in the warmth of their good ideas rather than
implementing them.
The European Council, which comprises the 27 EU mem-
ber states, will meet on 25–26 March to start turning the draft
proposal into detailed policy. Only at this point will it become
clear exactly to what extent the council
members — whose loyalty to the inter- “Politics is a dicey
ests of their own country can sometimes business and EU
conflict with their allegiances to the EU research is highly
— will allow themselves to be pinned political.”
down.
The commission will then have to write a second draft proposal
that reflects those details. This final document must be approved by
the European Council in June. The consequences will then begin
to unfold for researchers on the ground — for example, through
the design of the Eighth Framework Programme for Research and
Technological Development, a multibillion-euro grant scheme that
is due to launch in 2014.
Politics is a dicey business and EU research is highly political.
There are still many ways in which things could go wrong for
research as a consequence of Europe 2020 strategy. But this time,
at least, this draft makes it possible to imagine that things could go
right. ■
 Vol 464|11 March 2010

RESEARCH HIGHLIGHTS
The long and the short

M. STEIN
Zool. J. Linn. Soc. 158, 477–500 (2010)
The fossils of many arthropods — invertebrates with hard, external skeletons
and segmented bodies — from 540 million to 250 million years ago have a large
pair of ‘great appendages’ that they probably used to grasp and handle food.
Euarthropods had shorter versions of the appendage, whereas many of the
now-extinct anomalocaridids had longer, leg-like structures.
The position of these animals in the arthropod family tree is
controversial. Martin Stein, now at the University of
Kansas in Lawrence, reports a new arthropod
species that bridges a gap between the
two groups.
The new species, Kiisortoqia
soperi (reconstruction pictured),
discovered in Greenland, was clearly
a euarthropod but had the longer type of appendage.
Stein says this supports a shared ancestry of the long and short
‘great appendages’ and suggests that both evolved from the limbs
of a particular segment of the arthropod body.

METABOLISM
Warm milk
Cell Metab. 11, 206–212 (2010)
Mother’s milk fires up a heat-generating
metabolic pathway in newborn mice.
Newborns have to rapidly adjust to the
comparatively chilly environment outside the
womb. Francesc Villarroya of the University
of Barcelona in Spain and his colleagues
found that expression of a metabolic regulator
gene called Fgf21 increases immediately after
birth. However, this increase was seen only
when pups were allowed to suckle or were
given a lipid-rich emulsion to drink. Pups fed
a glucose solution — a diet similar to what
they would have received in utero — did not
activate Fgf21.
When the FGF21 protein was injected
into newborns that were not allowed to
nurse, genes associated with heat generation
in brown fat, a tissue specialized in heat
production, were activated in the pups.
CHEMISTRY
Cellulose busters
adding water to a mixture of acid catalyst and
ionic liquid as it attacked untreated cellulosic
biomass. The process avoids the hazards of
working with concentrated acids, and in a
few hours produces sugar yields of 70–90%,
which enzymes take days to achieve.
However, scaling this up for commercial
use could be problematic because of the need
to recycle ionic liquids, which are expensive
relative to other solvents.
NEUROSCIENCE
Nerve cell talk
Science 327, 1250–1254 (2010)
Neuroscientists had long believed that neural
cells called astrocytes (pictured below)
provide structural support and nutrients to
the neurons they surround. But a debate has
erupted over whether these cells also release
signalling molecules that affect neuronal
communication — and, if they do, how.
It is thought that an increase in astrocyte
calcium-ion concentration might trigger the
release of these signalling molecules. Cendra
Agulhon at the University of North Carolina at
Chapel Hill and her team engineered mice in
which they could either stimulate or suppress
calcium signalling. They showed that neither
activating nor deactivating the calcium rise in
astrocytes affected neurotransmission in the
brain’s hippocampus, suggesting that other
mechanisms underlie astrocyte signalling.
BIOMATERIALS
Squishy particles
Angew. Chem. Int. Edn doi:10.1002/anie.200906606
(2010)
Researchers have created tiny microgel
particles that can squeeze through pores just
one-tenth of their size. This makes them
potentially useful as a biomaterial for tasks
such as drug delivery — squishy nanoparticles
can be easily filtered out by the kidneys, so
don’t need to be degradable by the body.
Grant Hendrickson and Andrew Lyon
at the Georgia Institute of Technology in
Atlanta found that 116-nanometre-wide
microgel particles, when subject to pressure,
can pass through a material with pores of
10 nanometres — similar in diameter to those
of the kidneys.
A. TOUSSON/PHOTOLIBRARY

Proc. Natl Acad. Sci. USA doi:10.1073/

pnas.0912073107 (2010) CANCER BIOLOGY
Finding a safe, low-cost and high-yield
way to break down cellulose — a major Arsenic activation
component of plant-fuel sources such as Cancer Res. 70, 1981–1988 (2010)
maize stalks — into fermentable sugars Arsenic, a carcinogen found at unsafe
is a challenge in biofuel development. A levels in drinking water in many parts
mild mixture of acid and ionic solvent of the world, may cause cancer by
may prove quicker and cheaper than increasing the activity of the Hedgehog
the enzymes commonly used. signalling pathway, which is known to
Ronald Raines and Joseph Binder at promote cell proliferation.
the University of Wisconsin, Madison, David Robbins, now at the
improved an existing recipe by slowly University of Miami, Florida, and his
144
© 2010 Macmillan Publishers Limited. All rights reserved
 Vol 464|11 March
NATURE|Vol 464|11
2010
March 2010
team have shown that arsenic destabilizes
a protein that regulates Hedgehog activity,
which may allow the pathway to boost cell
growth. Mice exposed to arsenic-laced water
had higher Hedgehog activity.
The researchers also analysed 265 human
bladder cancer samples along with the
arsenic levels in the patients’ tap water.
Higher arsenic levels correlated with
increased tumour expression of a key player
in the Hedgehog pathway.
PHYSICS
Photon storage for telecoms
Phys. Rev. Lett. 104, 080502 (2010)
Photons have a quantum mechanical spin,
which can be ‘up’, ‘down’ or both. Storing a
photon using conventional methods alters this
state, destroying its quantum information.
Björn Lauritzen and his colleagues at
the University of Geneva in Switzerland
have found a way to store infrared photons
without changing them. The team shone
AM. CHEM. SOC.
a weak infrared laser pulse at a crystal
containing erbium atoms, which had
previously been excited with a different light
pulse. As an infrared photon was absorbed
by the crystal, its quantum state spread
across many erbium atoms.
Using an electric-field gradient, the group
triggered the crystal’s re-emission of a photon
encoding the same information as the
incident photon a few hundred nanoseconds
later. The efficiency was well below 1%, but
the technique could prove useful in quantum-
communication devices, the authors say.
CANCER GENOMICS
Melanoma’s mutations
Genome Res. doi:10.1101/gr.103697.109 (2010)
Some cancer-associated genetic changes are
not easily detected with standard technologies.
Researchers have now found mutations linked
to melanoma using RNA sequencing.
Levi Garraway of the Dana-Farber Cancer
Institute in Boston, Massachusetts, and his
colleagues used a high-speed sequencing
technology to sequence RNA from ten
patients’ melanoma samples. They identified
11 abnormal RNAs resulting from genes that
had fused in the genome — the first reported
gene fusions for melanoma. They also found
12 instances in which two separate genes were
transcribed, or ‘read’, together to produce
a mutated RNA, seven of which occurred
in more than one sample. In addition, the
researchers confirmed previous findings that
melanoma has a higher mutation rate than
other cancers, reflecting DNA damage caused
by exposure to ultraviolet light.
© 2010 Macmillan Publishers Limited. All rights reserved
NANOTECHNOLOGY
Harvesting heat
Nano Lett. doi:10.1021/nl903267n (2010)
Waste heat from vehicle exhaust pipes
and industrial waste streams could offer
a sustainable energy source, but current
technologies for harvesting thermal energy
are costly and inefficient.
Ray Baughman at the University of
Texas at Dallas and his colleagues have
created a ‘thermocell’ that can be wrapped
around pipes (pictured below). Made of
carbon nanotube electrodes, the device
is three times as efficient as conventional
platinum-based thermocells. The difference
in temperature between the two electrodes
creates an electrochemical potential
difference, which the thermocell uses to
generate electricity.
One of the team’s prototypes can be
attached to hot nuclear-reactor pipes.
EVOLUTION
Creating cooperation
Evolution doi:10.1111/j.1558-5646.2010.00959.x (2010)
How cooperation evolves between species is
much debated. William Harcombe, currently
at Harvard University in Cambridge,
Massachusetts, used bacteria to observe this
evolution in the lab.
He plated out Petri dishes with an
Escherichia coli mutant unable to produce an
essential amino acid, and a Salmonella species
that consumes waste from E. coli and excretes
small amounts of the amino acid. In two out
of ten dishes, Salmonella mutants arose that
made large amounts of the amino acid.
When grown with E. coli and normal
Salmonella, the cooperative mutants rapidly
increased from 1% of the Salmonella
population to more than 80%. When the
bacteria were grown on different media such
that the Salmonella no longer relied on the
E. coli for food, cooperative mutant numbers
crashed. This also happened when the bacteria
were grown in flasks, suggesting that spatial
structure of the bacterial colonies is needed for
cooperation to evolve.
RESEARCH HIGHLIGHTS
JOURNAL CLUB
Markus Reichstein
Max Planck Institute for
Biogeochemistry, Jena, Germany
A biogeochemist looks at where
all the emitted carbon dioxide is
going.
Humanity is currently performing
a huge global experiment, emitting
increasing amounts of CO2 into the
atmosphere by burning fossil fuels.
I find it astonishing that although
we scientifically explore other
planets, we still don’t understand
Earth’s important carbon cycle.
Corinne Le Quéré at the
University of East Anglia in Norwich,
UK, and her team have put together
the pieces of the contemporary
global carbon cycle. They analysed
observations and modelling results
on fossil-fuel emissions and the
terrestrial and ocean carbon cycle,
which are the major contributors
to the atmospheric carbon budget
(C. Le Quéré et al. Nature Geosci. 2,
831–836; 2009).
The bottom line is that humans
are emitting more CO2 than
projected in the pessimistic
scenarios outlined by the United
Nations Intergovernmental
Panel on Climate Change.
The researchers find that only
40–45% of this CO2 remains in the
atmosphere; the rest is ‘cleaned
up’ by the ocean and land — the
‘carbon sink’. It would be interesting
to know whether the fraction taken
up by oceans and land remains
constant, because any alterations
will change the global climate–
carbon-cycle feedback.
The study also indicates that
we are moving towards saturation
of the carbon sink, but the
uncertainties are large. Many
carbon pools and processes,
particularly those below ground in
the soil, are not well understood
and are hardly accounted for in
carbon-cycle models (P. Ciais
Nature 462, 393; 2009).
The message from Le Quéré
et al. is that more observations
are needed, that data should be
fully integrated with models, and
that these efforts must be more
targeted and coordinated if we are
to understand what is going on
with the Earth system in our huge
experiment.
Discuss this paper at http://blogs.
nature.com/nature/journalclub
145
 Vol 464|11 March 2010

NEWS BRIEFING
● POLICY

ESO/S.BRUNIER
European vision: In a new
decadal economic agenda,
the European Union (EU) has
said it hopes to raise spending
on research and development
(R&D) to 3% of the region’s gross
domestic product by 2020. That
measure was supposed to have
been hit this year, but spending
never rose above 2%. The Europe
2020 strategy, released on
3 March, suggests streamlining
bureaucracy to boost private
R&D spending, a single EU-wide
patent system and completion
of the European Research
Area, which would enable the
free movement of researchers
and funding across national
borders. Europe also needs
to refocus its research agenda
onto societal challenges such as
climate change, the plan says. See
Editorial, page 142.
GIANT EYE IN CHILE
Polio drive: More than 85 million The €800-million (US$1.1-billion) European Extremely Large Telescope will probably be built at Cerro
children under the age of five Armazones, in Chile’s Atacama Desert. The European Southern Observatory (ESO) announced on 4 March
will be immunized against polio that its advisory committee had recommended the site (pictured) above Spain’s La Palma and three other
in 19 countries across Africa. Chilean competitors. The Atacama Desert has the best overall sky quality and would allow the 42-metre-
The effort, part of the Global diameter telescope to operate in sync with ESO’s existing Paranal Observatory, the committee said. A final
Polio Eradication Initiative, was decision will be made no later than June.
announced by the World Health
Organization on 4 March. It is
an attempt to stop an ongoing Canada budget hope: Canada’s OCI, which investigates crimes
epidemic of the disease. The 2010–11 federal budget features such as the sale of counterfeit
initiative tried to prevent polio modest increases for major drugs, increased by 73% (after
spreading in 2009, but not enough science funding agencies, but adjustment for inflation) to
children received vaccines. also includes plans to review all $41 million in the decade that
federal support of research and ended in 2008. But the 230-plus-
Swiss animal rights: In development. See page 153 for person office runs with little
a 7 March referendum, more. accountability to senior agency
Switzerland’s voters rejected managers, the report says. The
a proposal for animals to be UK facilities funding: A FDA largely agreed with the
represented by state-funded
lawyers in court proceedings
structural shake-up may aid
Britain’s troubled Science and
SOUND findings. Its commissioner,
Margaret Hamburg, says it is
involving animal rights. The
Zurich canton already provides
Technology Facilities Council,
which distributes most of the
BITES taking action.

lawyers for animals in cases
of alleged animal cruelty, but
campaigners hoped to extend
the provision across the country,
which already has some of the
world’s toughest animal-rights
laws. The referendum did pass a
separate proposal to amend the
constitution in order to preserve
human dignity in biomedical
research; more detailed legislation
is expected to follow on stem-cell
research and gene therapy.
146
country’s physics and astronomy
grants but has suffered huge
budget holes since it was created
in 2007. See page 155 for more.
Lax drug oversight: The US
Food and Drug Administration
(FDA) has been lax in
overseeing its Office of Criminal
Investigations (OCI), a
report from the Government
Accountability Office concluded
last week. The budget for the
© 2010 Macmillan Publishers Limited. All rights reserved
“My feeling is that
it is going to be very
difficult to get a
treaty.”
Connie Hedegaard, European
commissioner for climate
change, is pessimistic over
whether the world will draw up
a new treaty on climate change
in 2010.
Source: Financial Times
● BUSINESS
Pharma cuts: AstraZeneca has
announced details of how it
plans to reshape its research and
development programme. The
London-based pharmaceutical
company will slash several
early-stage projects in topics
ranging from neuroscience
(schizophrenia, depression,
bipolar disorder and anxiety) to
infectious disease (hepatitis C
 Vol 464|11 March
NATURE|Vol 464|11
2010
March 2010 NEWS BRIEFING

and most vaccine projects).

Research sites in Lund, Sweden,
and in Charnwood and
Cambridge, UK, will close. The
company is holding to its earlier
estimate of a net reduction of
1,800 people affiliated with
research and development.
Solar support: Germany’s
government is to reduce state
incentives for solar power from
1 July. Solar feed-in tariffs — the
price an electricity utility must
pay to generators of solar energy
— have been dropping by 8–10%
a year since they were introduced
in 2000; now they will be cut
by a further 16% for roof-top
panels and abandoned for panels
installed in converted farmland.
Those installing and generating
US ARMY/AP
NUMBER of confidence” in protection
CRUNCH
of workers and the public at
biocontainment laboratories in
$1bn–6bn
Fort Detrick, Maryland, despite
finding numerous deficiencies in
the way environmental hazards
The liability that at the site were modelled. The lab
(where anthrax researcher David
experts polled by
Norwood is pictured, below) was
financial analysts at the source of the anthrax spores
Swiss bank UBS think that killed five people in 2001.
GlaxoSmithKline faces The US Army Medical Research
as a result of lawsuits Institute of Infectious Diseases
filed over its diabetes is expanding its facilities there
drug Avandia. under a congressional mandate,
Source: The Guardian
but citizens living nearby had
complained of risks including
an inadequate environmental
impact statement. The council’s
report concluded that it “would
not be useful” to redo the analysis
because construction has already
THE WEEK
AHEAD
13–25 MARCH
Conservation deals, including
a ban on international trade
in bluefin tuna and greater
protection for the African
elephant, are on the table at a
meeting of the Convention on
International Trade in Endangered
Species in Doha, Qatar.
➧ www.cites.org
15–19 MARCH
The American Physical Society
hosts its spring meeting in
Portland, Oregon.
➧ go.nature.com/Y8YXiL

solar power will see lower begun on the project.

returns — meaning that solar
panel manufacturers will have
to drop their prices. But as last
year’s plummeting price for solar
modules has not yet been fully
passed on to installers, Germany
should end up installing about
the same solar capacity as in
2009, according to analysts
Bloomberg New Energy Finance
in London.
Alzheimer’s disappointment:
A closely watched Alzheimer’s
treatment has failed to show
an effect on the disease in two
late-stage clinical trials, the
pharmaceutical company Pfizer,
based in New York, announced
on 3 March. Latrepirdine
(dimebon) — already used
as an antihistamine in Russia
— had seemed promising
enough for Pfizer to enter into a
$225-million deal in 2008 with
the drug’s developer, Medivation,
a biopharmaceutical company
in San Francisco, California. But
in a trial involving 598 patients
it performed no better than
placebo. Medivation’s share price
plunged after the announcement.
● RESEARCH
Pathogen lab confidence: A
US National Research Council
report published on 4 March
pronounced “a high degree
Palace in space: China plans to
launch an unmanned orbiting
space module in 2011, the
first component of a proposed
permanent space station.
Tiangong-1, or ‘Heavenly Palace’,
will first be used for orbital
docking practice with three
spacecraft; after that, crewed
ships will visit the module.
Aerospace official Qi Faren told
state media on 3 March that
“technical reasons” had pushed
back an original launch target of
late 2010.
16–19 MARCH
Miami, Florida, hosts a ‘State of
the Arctic’ conference, which will
discuss current understanding
of the rapidly changing Arctic —
and what can be done about it.
➧ http://soa.arcus.org
Energy call: The US
Department of Energy last
week announced $100 million
for a third round of research
programmes at its fledgling
Advanced Research Projects
Agency-Energy (ARPA-E).
The agency is geared towards
energy research that might not
be funded through traditional
science grants. Research
proposals are requested in
energy storage, electrical
power delivery and energy-
efficient cooling technology.
The announcement came at the
inaugural ARPA-E summit.

MARKET WATCH
SOURCE: PNAS

CONSUMER-BASED
More than one-fifth of the carbon dioxide California, published their findings this week CARBON ACCOUNTING
produced by China in 2004 was emitted to (S. J. Davis and K. Caldeira Proc. Natl Acad. Sci Because of exports, China produces
more CO2 than expected from
provide goods and services for non-Chinese USA doi:10.1073/pnas.0906974107; 2010).
domestic consumption.
consumers, mainly in western Europe, the The most comprehensive of its kind, the 5.8
United
6.5
United States and Japan (see chart). The analysis takes trade figures from 113 countries States
CO2 emissions
statistic comes from the latest study to look at across 57 industry sectors and finds that 23% produced
an alternative style of carbon accounting — one of global CO2 emissions (or 6.2 billion tonnes
CO2 emissions
that assigns CO2 emissions to the consumers of CO2) was traded internationally in 2004.
consumed 4.1 Europe 5.1
responsible for them. By contrast, inventories In some wealthy countries, including Britain
such as those reported under the United Nations and France, more than 30% of consumption-
Framework Convention on Climate Change based emissions are imported; in the United 5 billion
tonnes
(UNFCCC) simply tally the amount of gas each States, the figure is 11%. Caldeira thinks that the
5.1 China 4.0
country produces. UNFCCC’s production-based inventories should
Ken Caldeira and Steve Davis, at the Carnegie be supplemented with figures that track where Figures for 2004
Institution for Science’s lab in Stanford, the carbon is ultimately being consumed.
147
© 2010 Macmillan Publishers Limited. All rights reserved
 Vol 464|11 March 2010

NEWS

Outcry over scientists’ dismissal

Following years of acrimony, two high-profile researchers in Mexico have been expelled from their institute.

R. ROMERO/NEWSCOM
To the scores of scientific luminaries winners of the Nobel Prize in
who support them, the Terrones Chemistry. They unsuccessfully
brothers are two of the brightest stars lobbied for a compromise with
of Mexican science and have raised Mexican President Felipe Calderón
the nation’s profile in nanotechnology. Hinojosa and other top officials.
Yet the federal Institute for Scien- Kroto and some 70 other scientists
tific and Technological Research are again petitioning in support of
of San Luis Potosí (IPICYT) fired the brothers (see Correspondence
Humberto and Mauricio Terrones from H. W. Kroto et al., page 160).
Maldonado in December, and their The appeals by researchers out-
future in their home country is now side Mexico have not calmed the
looking dim. situation. In a written response to
International science leaders say enquiries from Nature, Rosaura Ruiz
that the case serves as an example of Gutiérrez, president of the Mexican
how entrenched scientific bureauc- National Academy of Science, said
racies in developing nations can that the letter sent to President Cal-
drive away promising researchers, derón in 2008 “created a great deal
especially those who have been of unease in the Mexican scientific
trained abroad. community, since its arguments lack
“This is a major loss for Mexican sufficient knowledge of the regula-
science,” says Mildred Dresselhaus, tions in Mexican institutions”. Ruiz
a nanotechnology researcher at the said the case has divided the Mexi-
Massachusetts Institute of Tech- can scientific community and that
nology (MIT) in Cambridge and a “proper channels exist for resolving
former president of the American this disagreement”.
Association for the Advancement Some are pessimistic about the
of Science, who has advocated on chances for resolving the dispute.
behalf of the Terrones. Ljubisa Radovic is a Spanish-speak-
David Ríos Jara, IPICYT’s direc- ing materials scientist at Pennsyl-
tor, rejects these arguments. “The vania State University in University
foreign scientists’ perception that Park who visited San Luis Potosí in
Mexican science is in imminent an unsuccessful attempt to broker a
peril is a flat misconception,” he says, compromise. He blames the impasse
adding that “there are about 20,000 Mauricio Terrones is an established nanotech researcher. on the IPICYT administration. “If
researchers in the country carrying you have stars like the Terrones,” he
out high-quality research in many areas”. the brothers were expelled from their lab in says, “you take care of them.”
Ríos says that he was forced to terminate December. Dresselhaus got a personal view of the
the employment of the brothers because they “I won’t work in a developing nation again,” escalating conflict in mid-December, when she
had violated institute rules and Mexican laws. says Mauricio. “Other Mexican universities are visited IPICYT to participate in the doctoral
Scientists outside Mexico have heard only one afraid to hire us,” adds Humberto. defence of Jessica Campos-Delgado, one of the
side of the story, he says. Terrones’ students. An unusual number of uni-
After receiving their doctorates in the United Centre of attention formed guards were present, apparently to head
Kingdom and doing postdoctoral work abroad, In an e-mail to Nature on 25 February, Ríos off a disturbance by students supporting the Ter-
the Terrones returned to Mexico and estab- accuses the brothers of several “irregularities”: rones, according to the brothers and Campos.
lished the country’s first nanotech lab, located failing to include IPICYT in four technol- The Terrones were intermittently called away
at IPICYT, about a decade ago. They secured ogy patent applications, not securing proper for discussions with their lawyers and adminis-
several large grants, collaborated with top authority to travel extensively last year and trators. Around that time, their e-mail accounts
researchers abroad and published well-cited improperly working for a private university. were shut down, and by the end of December,
papers on carbon nanotubes and buckyballs The Terrones deny any impropriety and blame their pay cheques stopped coming. “The whole
in high-impact journals. professional jealousy for their firing, a charge event was bizarre,” says Dresselhaus.
But several years ago, disagreements arose that Ríos rejects. Kroto, who oversaw Mauricio’s doctorate at
between the Terrones and the administration In 2008, the Terrones brothers’ situation the University of Sussex in Brighton, UK, says
at IPICYT over the operation of their lab (see attracted the attention of prominent scien- that he worries about the Terrones’ remaining
Nature 454, 143; 2008). Tensions between the tists, including British researcher Harold students. In an e-mail addressed to the “Inter-
two sides came to a head late last year, and Kroto and Mexico’s Mario Molina, both national Community” in late January, several
148
© 2010 Macmillan Publishers Limited. All rights reserved
Vol 464|11 March
NATURE|Vol 464|11
2010
March 2010
students and staff in the Terrones’ nanotech
group said “we remain extremely worried that
we might be fired on a whim” and that they had
been pressured to retract their criticisms of the
administration.
Ríos says that the IPICYT administration
has offered its full support to the remaining
students. The current students declined inter-
view requests from Nature.
Campos, who received her degree, says that
she remains shaken by the Terrones’ firing. She
will soon leave Mexico and head for a postdoc-
toral fellowship in Brazil. “I don’t know if I can
or will return, if I have to deal with the people
who did this,” she says.
That will be a loss for Mexico. During her
doctoral training, Campos spent six months
in Dresselhaus’s lab at MIT, providing a key
contribution to a method to enhance graphene
nanoribbons for mass production as semicon-
ductors (X. Jia et al. Science 323, 1701–1705;
2009). MIT, in conjunction with IPICYT,
applied for a US patent application on the
technology, and MIT now is seeking licensing
agreements.
Helpful discussions
Two of the patents that Ríos says the Terrones
failed to disclose involve Mexico’s largest juice
producer, Grupo Jumex in Tulpetlac. The
other two are owned by institutions in Japan:
the National Institute for Materials Science
(NIMS) in Tsukuba and Shinshu University
in Wakasato.
The science for the patents was “partly or
fully developed” at IPICYT, Ríos wrote, and
“it goes without saying how serious stealing
intellectual property is”.
Gerd Reiband, a head engineer at Jumex,
says that the company filed the two patent
applications on its own because the work was
conducted there, not at IPICYT. There was no
agreement to financially reward the Terrones
and their names were included as a courtesy for
their helpful discussions. Scientists at the two
Japanese institutions say similar conditions
applied to their patents.
Ajayan Vinu, a materials scientist at NIMS,
worked with Mauricio when the Mexican sci-
entist was on a fellowship in Japan. Vinu says
that he was astounded to hear that advice that
Mauricio had offered had played a part in his
dismissal.
“Tell them they are crazy,” says Vinu, who
was not contacted for details by IPICYT offi-
cials. “This is dangerous for science.”
The brothers remain in San Luis Potosí and
are considering legal action to fight the termi-
nation of their employment. ■ Stanford University in California who
Rex Dalton
See Correspondence, page 160.
© 2010 Macmillan Publishers Limited. All rights reserved
NEWS
A HIDING PLACE FOR HIV
S. KAULITZKI/ISTOCKPHOTO
AIDS virus escapes
treatment inside progenitor
blood cells.
go.nature.com/ZGu9wW
Climate e-mail rerun
Arriving at work on 5 March, Stanford urging colleagues to calm down and stick
University ecologist Paul Ehrlich found to the science. And he hopes that the Inhofe
a rambling and highly profane voice report — which says the scientists “violated
message from someone identifying himself fundamental ethical principles governing
as John Q Public. In one of his more lucid taxpayer-funded research and, in some
moments, the caller labelled Ehrlich cases, may have violated federal laws” —
and his colleagues in the climate-science will spark a backlash. “If we don’t have civil
community as “progressive communists discourse, where reality rather than spin is
attempting to destroy America”. the basis of decisions, how are we going to
Once again, a private e-mail function as a society?” Schneider asks.
conversation about global warming had Yet Schneider himself compares Inhofe
gone public, inflaming critics and giving to the infamous Senator Joe McCarthy,
climate scientists yet another lesson in the who led the discredited campaign
challenge of responding effectively. against communists during the 1950s.
Several days earlier, Ehrlich had taken Inhofe spokesman Matt Dempsey calls
part in an informal discussion with leading the comparison “flat out false” and says
climate researchers on a National Academy the report merely lists a number of laws
of Sciences e-mail list. They — including the Freedom of
were exploring how climate “Progressive Information Act and the False
scientists should respond communists Claims Act — that the scientists
to a recent US Senate report might have violated. Inhofe has
on the ‘Climategate’ e-mails attempting to asked the inspectors general
leaked late last year from the destroy America.” of a number of federal science
University of East Anglia in agencies to determine whether
the United Kingdom. In the 23 February violations actually occurred.
report, Senator James Inhofe (Republican, As a member of the minority party, Inhofe
Oklahoma), who has long been the leading can’t do much more than issue statements
sceptic in Congress, accused at least and reports. But Schneider says that if the
17 climate scientists of unethical, even Republicans later regain a majority in the
potentially criminal, behaviour in their Senate, Inhofe could take more concrete
handling of climate data. steps, such as forcing climate scientists to
The researchers’ private — and quite testify before Congress and pursuing his
spirited — discussion was soon disclosed claims in congressional hearings.
in conservative media outlets. The initial Scientific societies, including the
story, which ran in The Washington Times, American Association for the Advancement
quoted Ehrlich as saying that climate of Science (AAAS) and the American
scientists must recognize they are in a Geophysical Union, should respond
“street fight against well-funded, merciless to Inhofe’s report with a declaration of
enemies who play by entirely different support, says Michael Oppenheimer, a
rules”. Other participants went so far as climate scientist at Princeton University
to suggest taking out advertisements to in New Jersey who is also on Inhofe’s list.
counter the sceptics, perhaps through a “Defending the scientific community and
new non-profit organization. scientists from attacks ought to be a central
Ehrlich stands by his statement, but says part of their mission.”
the affair has made it “crystal clear” that The AAAS and other scientific
there is no such thing as a private e-mail. organizations have repeatedly affirmed the
“I’m not putting anything in an e-mail that fundamental science of global warming,
I don’t want to appear on the front page of but some officials are wary of responding
The Washington Times or Fox News,” he directly to Inhofe. Alan Leshner, AAAS
says, adding that the sceptics “are trying chief executive, says he is focused on trying
to keep the scientists busy and to keep the to ensure that policy-makers distinguish
scientists from doing their job, and they between the controversies and the science.
are doing extremely well”. “What I don’t want to see is that this set of
Stephen Schneider, a climatologist at incidents is used as an excuse to deny the
scientific findings.” ■
is on Inhofe’s list and participated in the Jeff Tollefson
National Academy discussion, says he is See Editorial, page 141.
149
 Vol 464|11 March
NATURE|Vol 464|11
2010
March 2010
300 μm 30 μm
Apollo 14
Apatite grains in a basalt
sample contain about as much
water as has been found in
some volcanic rocks on Earth.
for its formation, and that the heavy comet
MOON: LICK OBSERVATORY
water mixed into the Moon’s magma ocean.
The comets would have struck Earth too, but
because the young planet had a significantly
larger supply of water, the heavy water depos-
ited on Earth was greatly diluted.
Alternatively, perhaps the heat of the
impact evaporated the lighter water from the
Moon, leaving it enriched in heavier water.
Or maybe the impactor itself contained large
amounts of heavy water.
According to Elkins-Tanton, “this may all
lead to a reconsideration of the giant-impac-
tor theory,” especially in how it relates to the
origin of lunar water.
But not yet. The heavy-water results could
reflect nothing more than a specific spot of
enrichment — the impact site of an ancient
comet, for example. And some researchers
aren’t convinced that the Moon is a wet as the
new results suggest.
Chip Shearer at the University of New
Mexico in Albuquerque has been analysing
Apollo 15
A grain of the
mineral apatite
shows off
rainbow hues
under a
microscope.
Locked inside are
hydroxyl ions
(orange, inset)
revealing the
presence of
water.
chlorine in volcanic rocks, which can pro-
vide some information about ancient water
because of the way chlorine would have
bound to it to form other compounds. He
says that the estimated water concentrations
reported last week are at the upper limit of the
range allowed by his chlorine results.
Sorting out the debate will require new
samples. To Clive Neal at the University of
Notre Dame in Indiana and chair of a NASA
lunar advisory group, that means returning
to the Moon. But before then, researchers will
be revisiting the rocks collected by Apollo
astronauts.
At the meeting last week, Taylor drew an
index card out of his wallet that contained
the five-digit codes for three more samples
he was going to retrieve from the vaults at the
Johnson Space Center, on the other side of
Houston from the conference. He has made
the trip many times before. “Being an old
fart,” he says, “I know the ropes.” ■ funding so complicated.”
Eric Hand
© 2010 Macmillan Publishers Limited. All rights reserved
NEWS
JUSTUS SUSTERMANS (1636)
GALILEO BACKED
COPERNICUS DESPITE
DATA
Early telescopes suggested
that Earth stood still.
go.nature.com/OV49Sv
Biology thinks
big to stay cuts
Biologists with grandiose research
F. MCCUBBIN, CARNEGIE INST. WASHINGTON
proposals often turn to the risk-tolerant
Human Frontier Science Program (HFSP)
to get themselves started. However, the
world’s only intercontinental research
funding agency is now working hard to
convince its members — 13 individual
nations plus representation from the
European Union — that frontier research is
a wise investment in tough economic times.
The 20-year-old HFSP distributes about
US$60 million a year to support research
into complex biological systems. Last
week, the programme’s secretary-general,
geneticist Ernst-Ludwig Winnacker, met
with an elite group of about two dozen
scientists to discuss where the frontiers of
biology might lie.
Participants agreed that genetic
technologies will continue to be central
to frontier biology; that fields such as
microscopy will continue to drive it; and that
new developments in chemistry will feed it.
NASA
Tom Henzinger of the Institute of Science
and Technology in Vienna also convinced
participants that pure mathematics could
become an increasingly useful tool in
understanding biological processes.
“We remain open to any approach — the
three sole criteria remain excellence, risk
and biology,” says Winnacker, who has
previously served as secretary-general of the
European Research Council and president
of the German research agency the DFG.
The general messages from the
brainstorming session will feed into the
organization’s first formal strategy report,
which will be crucial when the HFSP’s
members meet in May. Winnacker is already
bracing himself for trouble: Japan, which
normally contributes around half of the
budget, has this year cut its contribution
by 5%, and the global financial crisis is
squeezing all members. Winnacker hopes
that the report will help to secure continued
funding for this unique programme, whose
unbureaucratic support of big ideas is
popular with scientists. “It is a paradox
that science is international, but funding
is regional or national,” Winnacker notes.
“The HFSP is the only agency with none
of the political borders that make joint
■
Alison Abbott
151
NEWS Vol 464|11 March 2010
NATURE|Vol

Plant biologists fear for cress project

Is enthusiasm withering for funding studies into Arabidopsis thaliana?
The brilliant career of a diminutive weed may INSIDE ARABIDOPSIS was this idea that if we march through every

CUSTOM LIFE SCIENCE IMAGES/ALAMY

have hit a snag. Arabidopsis thaliana has been The Arabidopsis 2010 project has linked gene and look at its phenotype with a knockout
the darling of plant biologists for some 30 years thousands of the plant's genes and proteins mutant, that would give us great insight into the
because of its small genome and rapid growth, to their biological function. functional identity of every gene,” says Philip
and in 2000 it became the first plant to have its Benfey, a plant biologist at Duke University in
genome sequenced. To capitalize on this, the Durham, North Carolina. “The reality was
US National Science Foundation (NSF) Identified more than
much more complicated.”
soon afterwards dedicated US$200 mil- Studied microRNAs
350 genes with an
and the regulatory
lion towards determining the function of networks that control
essential role in seed Systems approach
every Arabidopsis gene by 2010. development In 2008, the project team drafted a proposed
flower development
Now time is nearly up, although the Arabidopsis 2020 programme that would
work is far from done. And with the NSF focus on systems biology, an approach that
unlikely to extend the project, Arabidopsis Analysed the would use the large data sets generated in the
researchers fear that the plant’s popularity with expression of 2010 programme to develop models of plant
27,103 specific
funders is on the wane. sections of DNA
function.
The appeal of Arabidopsis is as a stand-in That programme, however, is unlikely to be
for unwieldy food crops that grow funded, according to Parag Chitnis, deputy
slowly, take up lots of space and have Identified more than 4,000 gene director of the NSF’s Division of Molecular
large genomes. By studying thousands of mutations that cause defects in Studied transcription and Cellular Biosciences. “Now we are
photosynthesis, chloroplast factors involved in
plants in a single greenhouse, scientists responses to infection
thinking beyond just one model organism,”
can conduct experiments in a fraction of metabolism and development he says.
the time and space required for crop spe- The NSF’s Plant Genome Research Pro-
cies. But sequencing technology is improv- gram and the Department of Agriculture’s com-
ing so quickly that genome size is no longer petitive funding programme have increasingly
the barrier it once was. As more crop genomes favoured grants for work on agricultural species
are sequenced, including maize (corn) and rice, in recent years. The NSF has also begun phas-
researchers can increasingly study crop species ing out funding for TAIR, which is partially
directly, rather than relying on a model plant. funded by 2010 project funds (see Nature 462,
“There’s obviously a drive back to increased 258–259; 2009).
funding of crop plant research,” says Mark Stitt “People seem to feel, ‘Oh, Arabidopsis is
at the Max Planck Institute of Molecular Plant Mapped gene done. Let’s just move to the real stuff,’” says
Physiology in Potsdam, Germany. “This is expression in Natasha Raikhel at the University of California,
fair and good, but there is a quite serious risk developing roots Riverside. “Well, it’s not done.” Raikhel adds
that some of the advances made in Arabidop- that researchers have only just begun to tackle
sis research in the past ten years may not be some key biological questions in Arabidopsis,
sustained.” such as how cell walls form. Researchers devel-
Data from the NSF’s 2010 project (see oping ways to turn cellulose into ethanol need
‘Inside Arabidopsis’ for a few highlights) have to know how plants build their cell walls, but it
been used to tackle fundamental questions Washington in Stanford, California. However, is often more difficult to characterize the func-
about development, pathogen resistance and Huala estimates that direct experimental data tion of genes in biofuel crops.
hormone signalling. Annotation of the Arabi- about function are available for only a third of And although Benfey welcomes additional
dopsis genome — the linking of biological data all Arabidopsis genes. funding for crop research, he notes that the
to sequence information — is Money was one hurdle. extensive data already available for Arabidopsis
now considered among the “People seem to feel, Although the project was make it the best species to advance plant systems
highest quality of all sequenced ‘Oh, Arabidopsis is designed for an annual budget of biology. It would be difficult to build general
genomes. $100 million, it actually received models of flowering, for example, using data
But the project has fallen done. Let’s just move about a fifth of that sum. Then collected from many different plants, he says.
short of its original goal of to the real stuff.’ Well, there were biological obstacles. Even crop researchers have a good word
determining the function of it’s not done.” Arabidopsis has many large for the little weed. “It’s very important for
every gene. Most Arabidopsis gene families whose constitu- the entire plant and agricultural community
genes have been characterized in some fashion, ent genes have overlapping functions — a com- that Arabidopsis research should continue
for instance by observing their expression in a mon phenomenon in plants. Knocking out one to be funded.” says Edward Buckler, a maize
high-throughput assay, says Eva Huala, direc- gene in the family often did not affect the plant, geneticist at Cornell University in Ithaca,
tor of the Arabidopsis Information Resource because other genes were able to compensate. New York. ■
(TAIR), a database at the Carnegie Institution of “When the programme was designed, there Heidi Ledford
154
© 2010 Macmillan Publishers Limited. All rights reserved
Vol
NATURE|Vol
464|11 March
464|11
2010
March 2010 NEWS

WOODY SHRUBS DON’T

C. TAYLOR
SUCK UP WATER
Clearing savannahs might
make drought worse.
go.nature.com/VlScpq

GRAPHIC DETAIL

*SCIENCE & ENGINEERING INDICATORS: 2010 (NATL SCI. BD);

Securing
UK science
British politicians have given no
assurances that science will be
protected from the deep cuts in
public spending that are predicted
to follow this year’s general
election (see Nature 463, 410–411;
2010). But Britain’s Royal Society
is trying to ward off the expected
blow. In a report entitled The
Scientific Century: Securing Our
Future Prosperity, released on
9 March, the venerable society
argues that slashing support
for research would be a false
economy, and would harm the
nation’s long-term economic
prospects.
INTERNATIONAL COMPARATIVE PERFORMANCE OF THE UK RESEARCH BASE (EVIDENCE LTD, 2009)
60 an influential 2005 report by
the US National Academies
50
Share of global funding
for research (2007 data)*
that advocated increased funding
to maintain America’s pre-eminent
40
Share of world's most
highly cited papers
position in science. The
(1998–2008)†
Royal Society argues along
Percentage similar lines: “Our scientific
30 leadership, which has taken
decades to build, can quickly
20 be lost,” it warns.
The report asks the UK
10 government to outline plans
for increased research funding
over a 15-year period (2011–26).
0
United Japan China Germany France United It also recommends expanding
States Kingdom a tax credit for research
and development to stimulate
bigger proportion of the most rising research funding in India, investment by businesses, and
highly cited research papers than China and Brazil, along with outlines the need for policies
its share of global funding would significant investment for science that encourage researchers
suggest (see graphic). But seen in the stimulus packages from overseas to work in the
this impressive track record is in the United States, France and country. ■

†
The United Kingdom has “fragile” and brings a “risk of Germany. Richard Van Noorden
traditionally punched above its complacency”, the report The dossier models itself on ➧ http://royalsociety.org/
weight in science, garnering a says, noting competition from Rising Above the Gathering Storm, the-scientific-century

A rescue plan for UK physics funding

Britain’s most troubled research council
is about to undergo radical surgery. On
4 March, UK science minister Paul Drayson
unveiled his plan to reform the Science
and Technology Facilities Council (STFC),
the main research council that provides
funding for particle physics, space science,
nuclear physics and astronomy.
The plan, which promises to end years
of financial turbulence for the council, has
gained tentative support from the physics
community, which had previously been
incensed over repeated budget shortfalls
that forced the STFC to slash grants and cut
operations (see Nature 462, 396; 2009).
“We’re very pleased,” says Robert Kirby-
drastic cuts. But this year it has again
come up short, by £40 million. “That was
leading to damage on the grants end of its
responsibilities,” says Drayson.
Drayson set up a ministerial commission
to look into what he describes as the
“tough problem” of balancing the STFC’s
commitments and grants, resulting in a
proposed three-part reform. First, the
council will keep the running costs for
its large facilities separate from grant
funding and budget them over a six-year
period. Second, it will enlist the Bank
of England to help it manage the cost of
overseas subscriptions to institutions such
as the European Southern Observatory,
headquartered in Garching, Germany, and
CERN, Europe’s particle-physics laboratory
near Geneva, Switzerland, in order to
minimize the impact of the pound’s poor
exchange rate. Third, the UK subscription
to the European Space Agency (ESA),
the STFC’s most costly international
commitment, will be moved to a new
British space agency that is scheduled to be
launched later this month.
The plan “is just good business sense”,
says Kirby-Harris. By making sure that
the STFC’s fixed costs are appropriately
budgeted, the council should be able
to better serve the physicists and
DIAMOND LIGHT SOURCE

Harris, chief executive of the London-based
Institute of Physics, which represents the
country’s physicists. But missing from the
plan is more money for research funding.
Grants will still be cut back, and Britain still
plans to withdraw from the multinational
Gemini telescope project in 2012.
The STFC was formed in 2007 by
merging two councils: one that managed
large facilities and another that disbursed
physics and astronomy grants. Almost
immediately it became clear that the
organization faced an £80-million
(US$121-million) shortfall, leading to
© 2010 Macmillan Publishers Limited. All rights reserved
New proposals could safeguard funding for big
UK facilities such as the Diamond Light Source.
astronomers who depend on it for grants,
he says. Moving the ESA subscription
to the new space agency makes sense,
adds Andrew Fabian, president of the
Royal Astronomical Society in London,
although the implications of this move for
astronomers are not yet clear.
The proposed changes should happen
after the government’s next spending
review, which is currently scheduled to
follow a national election that is expected
to be called in May. Any change of
government could yet upend the plan. ■
Geoff Brumfiel
155
 NEWS FEATURE Vol 464|11 March 2010
NATURE|Vol

Welcome to the Atomic

Weapons Establishment
With the launch of a powerful laser facility, Britain’s most secretive
lab is opening up to academics. Geoff Brumfiel secures a preview.

A
ldermaston is a picturesque English target chamber — and the fact that the AWE is have started to change over the past decade,
village with red-brick houses, tidy sharing it at all. says Steven Rose, a physicist at Imperial Col-
gardens and an inviting local pub. The motivations for collaboration are not lege London who headed the AWE’s plasma-
But just to the south lies a much entirely selfless. The defence establishment physics division from 2001 to 2004. “The old
more forbidding set of buildings. Behind dou- wants to provide scientists inside the security argument that secrecy was paramount I think
ble-fencing and guarded gates is the sprawling cordon with the sort of intellectual stimulation perhaps holds less sway these days,” he says.
campus of the Atomic Weapons Establishment needed to keep them on their toes; it also has “They’ve realized that you can do some things
(AWE), Britain’s most hush-hush research lab- an incentive to nurture the wider community in collaboration with the outside world.”
oratory. The 304-hectare facility, a converted of physicists from which it draws its staff. “We Enter Orion, the £183-million (US$274-
Second World War airfield, is home to the want to demonstrate that we maintain high million) laser on which construction began in
roughly 4,000 scientists, engineers and tech- standards for our science,” says 2005. The machine has ten con-
nicians who are responsible for maintaining Daryl Landeg, the AWE’s chief “The defence ventional lasers (see graphic),
Britain’s nuclear warheads. scientist. And academics far ministry is each of which can deliver 500
The Ministry of Defence, which oversees the beyond Aldermaston are keen joules of energy in about a nano-
AWE, is notoriously tight-lipped about the lab’s to cross the fence. At present, notoriously tight- second (a billionth of a second).
activities, and scientists who work there gen- Europe has only a handful of lipped about the It also has two short-pulse lasers,
erally try to keep a low profile. But within the comparable laser facilities, says lab’s activities.” which deliver the same amount
next few months, the AWE’s most ambitious François Amiranoff, director of of energy in just half a picosec-
and expensive scientific project is due to be the Laboratory for the Use of Intense Lasers ond (a trillionth of a second).
completed, and it is prompting the lab to open at the École Polytechnique near Paris. “A facil- Inside Orion’s main building, workers in
up its doors — at least a chink. ity like Orion is very, very interesting for the Teflon suits and hairnets are busily scrambling
The new Orion facility will be home to scientific community,” he says. around gigantic white scaffolding. The struc-
12 high-powered laser beams capable of heat- ture will soon hold the mirrors, amplifiers and
ing and compressing material to millions of Secrecy rules lenses needed to boost and focus the 12 beams
degrees Celsius in less than a nanosecond. Britain established its atomic-weapons pro- onto their target, which lies in a separate cham-
Housed in a gleaming building the size of gramme at the current AWE site in 1950, after ber behind 1.5 metres of solid concrete, a shield
a soccer pitch, the laser system will pro- close involvement in the Manhattan Project. The necessary to contain the radiation generated
vide physicists at Aldermaston with crucial campus, which also houses facilities for manu- when the laser beams hit. The exceptional
data about how components of their ageing facturing and storing sensitive nuclear materi- cleanliness in the laser halls and target area (and
nuclear weapons behave. Under current plans, als, is shielded by the nation’s strict secrecy laws, expected of visitors) is about more than aesthet-
around 15% of Orion’s time will be offered to and has historically shunned visitors. Things ics: a stray hair in the path of the intense beam
academics wanting to study con- could cause irregularities and crack
CROWN COPYRIGHT

ditions on stars or inside giant ORION LASER an expensive mirror or grating.

planets. And in that open spirit,
researchers there invited a Nature
reporter in for a look around.
In most respects, Orion is
the smaller cousin of the US
National Ignition Facility (NIF)
in Livermore, California, which is
already running academic experi-
ments. When operating at full steam,
the NIF will use 192 lasers to create
around 4 million joules of energy,
some 100 times more powerful than
Orion. What makes the AWE’s laser
notable is the exquisite precision
that it will give researchers in con-
trolling the heat and compression
exerted on the materials placed in its
156
Ten long-pulse and two short-pulse lasers allow precise control of temperature
and pressure in target materials.
Orion’s main mission, like that of
4-metre-diameter target chamber the NIF, is to explore how nuclear
Radiation shielding weapons work, particularly as they
get older. In 1998, Britain ratified
the Comprehensive Nuclear-Test-
Ban Treaty, an international agree-
ment prohibiting tests of nuclear
weapons. Scientists in Britain and
worldwide have therefore been
busy developing computer models
to simulate nuclear warheads and
work out whether the weapons
Two short-pulse will still detonate after decades
beamlines
in storage, and what type of deto-
Laser-pulse nation will result. What is missing,
Ten long-pulse compressors
however, are actual data.
beamlines
US scientists hope that the more
© 2010 Macmillan Publishers Limited. All rights reserved
 Vol 464|11 March
NATURE|Vol 464|11
2010
March 2010
Target chamber: Orion’s twelve lasers will explore conditions inside ageing thermonuclear weapons.
powerful NIF will contribute some of those
A. COX/AWE CROWN COPYRIAGHT 2010
data, by generating temperatures and pressures
so high that they will spark nuclear fusion in
small quantities of two hydrogen isotopes, deu-
terium and tritium. This fusion process would
resemble conditions inside the most powerful
stage of a modern thermonuclear weapon.
If the NIF is a thermonuclear hammer, then
Orion is a scalpel. The smaller facility will never
achieve full-scale fusion, but it will be able to
carefully control conditions inside test materials
such as uranium. Pressure and temperature usu-
ally go hand-in-hand, explains Peter Roberts,
head of the AWE’s plasma-physics department.
“You pump up a tyre with a bicycle pump and it
gets hot,” he says. But Orion can get around this.
It can compress a material with its long, nano-
second pulses then suddenly heat it with its very
short, half-picosecond pulses. The result is ‘iso-
choric heating’, an unusual condition in which a
material is heated so quickly that it doesn’t have
time to expand. This capability allows Orion to
probe materials at wide-ranging combinations
of temperatures and pressures.
In particular, researchers will use Orion
to explore two key parameters for materials
used in nuclear weapons: their opacity and
their equation of state. The first describes how
radiation travels through a material — in this
© 2010 Macmillan Publishers Limited. All rights reserved
case, the two stages that make up a weapon.
The first stage, or primary, is a few kilograms of
plutonium that are compressed by conventional
explosives until they begin a runaway nuclear
reaction. The radiation from that reaction is
then focused onto the ‘secondary’, the stage in
which hydrogen isotopes create a much larger
blast using nuclear fusion. Researchers want to
know what the opacity is and how it changes
with age so that they can model radiation’s flow
from the primary to the secondary and verify
whether the warheads will still work. The other
parameter — the equation of state — describes
how a material behaves at enormous pressures
and temperatures. By generating data on these
and other crucial parameters, Orion will give
nuclear-weapons scientists the information they
need to ensure that their models are correct.
“You can’t look this stuff up,” Rose says.
The researchers running the NIF often
emphasize the giant laser’s applications in
energy production and fundamental science
over its military role; it could, for example, lead
to new reactors that produce electricity using
tiny fusion implosions. Orion’s scientists are
much less circumspect. “We’re working on
weapons physics fundamentally,” Roberts says.
Nevertheless, he and others at the lab are eager to
give civilian scientists an opportunity to use the
NEWS FEATURE
laser — and the academics are eager to try out its
capabilities. “It will have significant characteris-
tics that no other laser in the United Kingdom or
indeed Europe has,” says Mike Dunne, director
of the Central Laser Facility at the Rutherford
Appleton Laboratory near Didcot, UK. Dunne
says that Orion should relieve the strain on the
facility’s Vulcan laser, Orion’s massively oversub-
scribed civilian counterpart.
Extreme conditions
Physicists studying astronomical objects find
themselves in a situation not dissimilar to
that of nuclear-weapons scientists: unable to
recreate the extreme conditions inside a star,
for example, and largely dependent on com-
plex computer models to show how they work.
Orion will not reach stellar temperatures and
pressures, but it will be able to create flows of
hot ionized gas with the sorts of high magnetic
fields and temperatures that mimic parts of
stars and other astronomical objects. By scaling
up the data from Orion, astrophysicists should
be able to improve their models, Dunne says.
The laser facility can also inform more
abstruse, fundamental work in atomic physics.
Current theory does well at describing normal
matter and extremely hot matter that has been
stripped of electrons. But it cannot describe
situations in which atoms are subjected to high
levels of radiation without losing all their elec-
trons. By heating a test material, then probing it
with its short-pulse beams, Orion can provide
data that can extend existing theories into this
middle region, says Amiranoff. This warm,
dense matter may exist at the heart of gas giants
such as Jupiter, or even within Earth’s core.
Scientific collaborations are not slated to
begin on Orion until the second half of 2012,
and proposals to use the facility will be sub-
mitted through the system being used for
Vulcan and other UK lasers, rather than being
determined by weapons scientists. This open
peer-review of proposals “is absolutely critical
to gaining the confidence of the community”,
Dunne says. “If some committee inside of
the wire assessed relative merits, there would
always be the suspicion that they were picking
topics that helped their programme rather than
just because they were good science.”
Back on the AWE’s campus, the weapons
scientists are eagerly preparing for their new
guests. Tom Bett, who helps to manage con-
struction, shows off a data-analysis room pur-
pose-built for unclassified visitors and lined
with sleek computer terminals. (The weap-
ons work will be done in a separate room, he
explains, and the data will be stored on servers
locked inside vaults.)
On the ground level, floor-to-ceiling
windows illuminate a bright reception area —
the first thing that visiting researchers will see
as they are welcomed to the new laser facility.
“Those windows,” Bett adds proudly, “are all
bullet-proof.” ■
Geoff Brumfiel is a senior reporter for Nature
based in London.
157
 NEWS FEATURE
What to make with DNA origami
Chemists looking to create complex self-assembling nanostructures are turning
to DNA. Katharine Sanderson looks at the science beneath the fold.
D
NA is the kind of polymer that chem-
ists dream about. Because its comple-
mentary sequences can bind to one
another, individual molecules of the
right sequence will assemble all by themselves
into intricate shapes and structures at the nano-
scale. DNA can weave together and bind other
molecules, allowing it to serve as a scaffold for
complex nanomachinery.
DNA nanoengineering is dreamy, but dif-
ficult. Researchers have been putting together
carefully chosen segments of DNA to form
P. W. K. ROTHEMUND
sheets, tubes, even simple machines such as
tweezers since the early 1980s. But back then,
designing these structures could take months
to years. And because researchers were focused
on designing them from scratch, they could
use only the short segments, no more than
150-base-pairs long, that DNA synthesizers
could manufacture. This in turn constrained
the size and complexity of the designs. “The
problem is that we don’t just want to make
small stuff, we want to make complicated small
stuff, cheaply and easily,” says Paul Rothemund,
a computational bioengineer at the California
Institute of Technology in Pasadena.
Rothemund wondered whether he could
create the complicated stuff using a longer,
naturally occurring piece of DNA, such as
the genome of a virus, and folding it over on
158
itself. So in 2004 and 2005 he spent months, he
says, programming in his underpants, trying
to work out a way to bend a 7,000-base-pair
viral genome to his will. In his design he visu-
alized how the genome could be folded into
a predetermined, two-dimensional shape.
Knowing the sequence of the virus at every
twist and turn, he was able to write comple-
mentary DNA sequences, about 16-base-pairs
long, that would essentially staple
the folds in place. He ordered
the ‘staples’ from a DNA-syn-
thesis company, mixed them
with his virus in a buffer
that stabilized the DNA and
then heated and cooled the
mixture, allowing the single
stranded viral DNA to bind
with the staples (see graphic,
opposite). The result, viewed using
atomic-force microscopy, was the smiley
face and several other shapes, created by
what he called DNA origami1.
The ease of DNA origami was a break-
through, dispensing with the intricacies of pre-
cise DNA engineering and other metamaterials
development. “It’s like being able to bake a cake
and not pay attention to the ingredient ratios,”
says Rothemund. But with the right ingredients
complex structures can be built with the kind of
© 2010 Macmillan Publishers Limited. All rights reserved
Vol 464|11 March 2010
NATURE|Vol
precision that many people have been looking
H. DIETZ/TECHNICAL UNIV. MUNICH
for. Origami scaffolds, sheets or bricks of folded
DNA, are packed with known sequences that
could be used to position DNA-binding mol-
ecules just a few nanometres apart. And the
new, larger structures can contain upwards of
200 sites for affixing such molecules, compared
with only a handful on pre-origami structures.
This type of precision engineering could be a
boon to nanoengineers wanting to position
components on nanoelectronic circuits or for
bioengineers looking to place proteins in
close, accurate proximity to one another.
Now the challenge is to go beyond the
novelty of Rothemund’s smileys and
a dozen or so other demonstration
patterns and build structures
with a practical purpose. Here’s
what several researchers are
dreaming of doing.
Make a ruler
Rothemund’s technique was a door
opener for Friedrich Simmel, a biophys-
icist at the Technical University of Munich in
Germany. Suddenly, Simmel says, he was able
to have even “rather sloppy” physics students
making DNA structures with ease. Simmel has
used DNA origami to make a ruler to meas-
ure distances between single molecules and
NATURE|Vol
Vol 464|11 March
464|11
2010
March 2010
STAPLING A SMILEY
With about 250 specific ‘staple’ strands of DNA,
Paul Rothemund folded a 7,000-base-pair viral genome
into a 10-nanometre ‘disk with three holes’.
Viral genome
Staple strands
calibrate super-high resolution microscopes.
Simmel designed a DNA origami rectangle
measuring 100 nm by 70 nm and included
some staple strands labelled with fluorescent
dye molecules. When the DNA folds, two
labelled staples sit at opposite ends of the rec-
tangle in precise locations. This ruler can be
used to calibrate high-tech microscopes that
can resolve objects smaller than the diffraction
limit of light — roughly 200 nm (ref. 2). The
kinds of molecules generally used for calibrat-
ing such scopes, such as loose pieces of DNA or
filamentous proteins, are not ideal because they
are flexible and their dimensions can change.
Simmel wants to use the ruler and related
structures to help track the movement of
molecular motors. Because fluorescent tags
allow them to use light microscopy, as opposed
to atomic-force microscopy or electron micros-
copy, researchers will be able to view molecular
processes as they happen. That’s important,
Rothemund says, once scientists begin cou-
pling proteins to DNA origami. Rulers like
Simmel’s will be useful to watch how those
proteins behave.
Build an artificial leaf
DNA origami could allow researchers to put
biomolecules together according to speci-
fication. A grail of sorts for many engineers
working at the nanoscale is photosystem II,
a complex of more than 20 protein subu-
nits and accessories that helps to split water
into hydrogen ions and oxygen during
photosynthesis.
Attempts to recreate photosystem II, or even
some of the catalysts involved in electron trans-
port, have been disheartening. DNA origami
could provide the scaffold to hold proteins in
place, says Hao Yan, a biochemist at Arizona
State University in Tempe. As part of a collabo-
ration funded by the US Department of Energy
at the university’s Center for Bio-Inspired Solar
Fuel Production, he has been looking to use
DNA origami as the basis for an artificial leaf
that makes hydrogen fuel from water.
Yan plans to use a cage-like DNA structure to
position the proteins and to bind manganese,
a crucial component of the water-oxidation
process. The goal is to better direct electron
flow in an artificial system that requires pre-
cise placement of components. “If we can really
© 2010 Macmillan Publishers Limited. All rights reserved
NEWS FEATURE
Stapled DNA
control all the electron-transfer sites then we that were complementary to loose ends on the
can improve the efficiency,” says Yan. DNA triangles5. The work is still quite crude,
says Rothemund. They haven’t got an active
Put a drug in a box device yet. “Much of the next five years will
Last year, Kurt Gothelf, director of the Centre be spent perfecting techniques to place DNA
for DNA Nanotechnology at Aarhus Univer- origami on surfaces where we want them and
sity in Denmark, and his colleagues, reported making this technique widely available to other
that they had made a box from DNA origami3. scientists,” he says. Importantly, they’ve tacked
One strand of DNA holds the lid shut; a sepa- down the shapes rather than having them
rate DNA ‘key’ springs it open. The invention “bobbing around like jellyfish in solution”, says
prompted many to wonder what could be put Rothemund. “I didn’t think anyone would solve
into the box and subsequently released. that problem in ten years. Now we’ve got DNA
William Shih, a DNA nanotechnologist at origami lined up like little ducks in a row.”
Harvard Medical School in Boston, Massachu-
setts, is keen to exploit this kind of vessel for Pushing past smileys
drug delivery, but the challenges are significant. The grand schemes go on. Researchers imagine
Getting the box to pass through a cell membrane an artificial ribosome capable of building cus-
is going to be difficult, he says. He proposes cov- tom enzymes, a matrix for supporting artificial
ering the box with a membrane similar to those organs, or a DNA origami network designed to
sported by some enveloped viruses. support a neuronal network connected to elec-
These viruses also have special proteins to trical circuitry. Researchers are attempting all of
facilitate entry. Shih says that he can take viral these, but they will face some serious hurdles.
proteins or related proteins and fix them to Buffer solutions must be finely tuned, otherwise
the outside of his DNA cages, but structures fall apart. And large,
he faces a long list of challenges.
“Now we’ve got complicated structures can take
“There’s a lot of stuff,” Shih says. DNA origami up to a week to fold completely.
“At this point it’s just a concept.” lined up like little Change is needed to the
basic, almost slapdash synthesis
Go for gold
ducks in a row.” approach of using unpurified
Many have talked about using DNA origami as DNA that isn’t adequately sequenced, says Shih,
a substrate for nanoelectronic circuitry, such “just because we got away with it so far doesn’t
as in plasmonic devices. mean we want to get away with it forever”.
Plasmonic devices couple light waves with But as long as this growing group of research-
charges on a metal surface and offer the speed of ers keeps trying to make different shapes, other
information transfer that light provides, but at applications will appear. Rothemund can’t pre-
sizes smaller than those to which technologies dict quite how, though. “The problem is that
such as fibre optics are limited. Current litho- you can do all these little cool things but they
graphic techniques run into trouble when trying in no way form a whole system. It’s not exactly
to arrange metallic materials such as gold into clear what kinds of systems we’re going to be
patterns with features smaller than about 100 nm able to build.”
in size. Rothemund says that gold spheres might He’s working hard to find out. “In the past six
be positioned using DNA origami to make struc- months, I’ve mostly been back at home, coding
tures with better optical qualities. in my underwear again, which hopefully means
He has already taken steps, in concert with good things are happening.” ■
IBM Almaden in San Jose, California, towards Katharine Sanderson is a reporter for Nature
arranging metal nanoparticles. Last year they in London.
managed to arrange DNA triangles on a litho-
graphically patterned surface4. Jennifer Cha 1. Rothemund, P. W. K. Nature 440, 297–302 (2006).
from IBM and her colleagues subsequently 2. Steinhauer, C., Jungmann, R., Sobey, T. L., Simmel, F. C. &
showed that they could place gold nanopar- Tinnefeld, P. Angew. Chem. Int. Edn 48, 8870–8873 (2009).
3. Andersen, E. S. et al. Nature 459, 73–76 (2009).
ticles smaller than 10 nm onto each origami 4. Kershner, R. J. et al. Nature Nanotech. 4, 557–561 (2009).
structure by affixing strands of DNA to the gold 5. Hung, A. M. et al. Nature Nanotech. 5, 121–126 (2009).
159
OPINION
CORRESPONDENCE
Science and Mexico
are the losers in
institute politics
Events at Mexico’s Instituto
Potosino de Investigación
Científica y Tecnológica (IPICYT)
have escalated to crisis point (see
page 148). We the undersigned
call on the world’s academic
community to help reverse the
damage currently being done
in this research institution,
once a shining example for all
developing nations.
After battling for two years,
Humberto and Mauricio Terrones
— acclaimed leaders of IPICYT’s
prestigious nanoscience and
nanotechnology group — have
been removed from office. This
flies in the face of the presumed
commitment of the Consejo
Nacional de Ciencia y Tecnología
(CONACYT), Mexico’s highest
research authority, to seek a
“solution that will be a product
of open negotiations carried out
with tolerance, good will, great
objectivity and agreements that
avoid personal aggression and
radicalization of positions”. This is
the wording of a recommendation
solicited by CONACYT from a
prestigious group of scientists
formed to help prevent this
outcome.
An international group of some
75 scientists has been working
hard with us for two years to broker
a solution to this sorry affair, to no
avail. Our hope is that President
Felipe Calderón will step in and
avert further damage. Otherwise,
the prestige of Mexico’s science
and the prospects for its
technological development will
suffer, as young Mexican scientists
won’t return after being trained in
research abroad.
The academic community
should join forces to reverse
this situation.
Harold W. Kroto Florida State
University, USA
e-mail: kroto@chem.fsu.edu
Pulikel M. Ajayan Rice University,
Texas, USA
Anthony K. Cheetham University of
Cambridge, UK
160
Mildred S. Dresselhaus Massachusetts
Institute of Technology, USA
Morinobu Endo Shinshu University,
Japan
Alan L. Mackay Birkbeck College,
University of London, UK
Ljubisa R. Radovic Pennsylvania State
University, USA
Colour-coded targets
would help clarify
biodiversity priorities
In this International Year of
Biodiversity, we should be setting
ambitious but realistic targets
for biodiversity policy over the
next ten years. Those shaped
at last month’s sixth Trondheim
Conference on Biodiversity in
Norway will be refined by the
Subsidiary Body on Scientific,
Technical and Technological
Advice in May and at the
Conference of the Parties of the
Convention on Biological Diversity
(CBD) in October.
As participating scientists in
the international biodiversity
programme DIVERSITAS, we
welcome the draft set of 2020
targets proposed by the CBD. But
the targets continue to mix the
biodiversity we value highly (that
is, the conservation agenda) and
the biodiversity we urgently need
to secure the benefits people
derive from fully functioning
ecosystems.
To resolve competing demands,
these different priorities should be
made explicit by categorizing the
targets according to their primary
motivation. We suggest the use
of red targets to stem urgent
deleterious biodiversity loss,
green targets for conservation
priorities and blue targets to
secure the long-term benefits
from functioning ecosystems.
The CBD should work closely
with the science community
to develop these targets for
changing environments and in the
light of new scientific discoveries.
The proposed Intergovernmental
Platform for Biodiversity and
Ecosystem Services (H. Mooney
and G. Mace Science 325,
© 2010 Macmillan Publishers Limited. All rights reserved
1474; 2009) and the global
biodiversity observation system
GEO BON (R. J. Scholes et al.
Science 321, 1044–1045; 2008)
will be valuable tools in this
collaboration.
Anne Larigauderie DIVERSITAS, 57
Rue Cuvier–CP 41, 75231 Paris Cedex
05, France
e-mail: anne@diversitas-international.org
Georgina M. Mace Imperial College
London, Centre for Population Biology,
Silwood Park, Ascot SL5 7PY, UK
Harold A. Mooney Department of
Biology, Stanford University, Stanford,
California 94306, USA
Barriers to carbon
capture and storage
may not be obvious
There’s more to ‘Buried trouble’
than whether carbon dioxide
should be injected under urban
areas or offshore (Nature 463,
871–873; 2010). Some barriers to
carbon-sequestration measures
are less immediately noticeable
than public opinion.
For example, the technology
should be incorporated into
developing energy systems. In
most scenarios produced by
the Intergovernmental Panel on
Climate Change (IPCC), much
larger volumes of CO2 will be
captured in China and India by
2050 than in developed countries.
But the scale and pace of energy-
systems development, and
the necessary carbon capture
and storage (CCS) technology
transfers, are daunting.
People optimistic about CCS
technology transfer to developing
countries should remember that
transferring even cost-saving
technologies (transgenic seeds,
for example) has been difficult.
We should develop transfer
incentives by recognizing CCS
investments within the Clean
Development Mechanism of the
Kyoto Protocol. We must also
limit uncertainties surrounding
CCS investments in developing
countries, particularly in
protecting intellectual property in
capture technologies.
NATURE|Vol 464|11 March 2010
Another barrier to CCS is
that countries with a weak
manufacturing capability are
not in a strong position to
develop lucrative carbon capture
technologies. The Australian
government, for example, has
committed Aus$2.4 billion
(US$2.2 billion) to its CCS
Flagships Programme. But two
of the projects rely on capture
technologies from Japan
(Mitsubishi in the ZeroGen
Project) and the United States
(GE in the Wandoan Power
Project). Governments need to
balance their desire to support
emerging domestic CCS
technologies against importing
potentially better technologies
from abroad.
CCS solutions are also
subject to the vested interests
of national politics. In the coal-
heavy economies of Canada,
the United States and Australia,
for example, governments
promote CCS in their emissions-
reduction promises, but they have
been reluctant to mandate the
technology.
The real barrier to CCS
is that, even in enthusiastic
countries, the focus is on selling
CCS solutions rather than on
mandatory CCS deployment.
Advocates should commit to a
firm timeline for mandatory CCS
on all new and retrofitted large
emitters.
Frances Bowen International Resource
Industries and Sustainability Centre,
Haskayne School of Business,
University of Calgary, 2500 University
Drive NW, Calgary, Alberta T2N 1N4,
Canada, and Smith School of Enterprise
and the Environment, University of
Oxford, UK
e-mail: fbowen@ucalgary.ca
Contributions to Correspondence
may be submitted to
correspondence@nature.com.
Please see the Guide to Authors at
go.nature.com/cMCHno. Science
publishing issues that may be of
interest to authors are regularly
featured at our blog Nautilus
(http://blogs.nature.com/
nautilus), where we welcome
comments and debate.
Vol 464|11 March 2010
Accelerating HIV vaccine development
Translational-research programmes supported by flexible, long-term, large-scale grants are needed to turn
advances in basic science into successful vaccines to halt the AIDS epidemic, says Wayne C. Koff.
he world needs a vaccine as soon as
T possible to help halt the inexorable spread
of HIV. But vaccine development has
not kept pace with breakthroughs in basic sci-
ence. However, during the past year, advances
in vaccine research and development (R&D)
have generated great excitement. These include
the first, partial protection in humans by an
HIV vaccine1; identification of a new target
for broadly neutralizing antibodies on HIV’s
surface2; and evidence for significant vaccine-
induced control of simian immunodeficiency
virus (SIV) in non-human primates3.
New leads for vaccine discovery should come
from linking HIV vaccine research with the
broader field of viral immunology, and under-
standing the early events in HIV pathogenesis
— as recommended elsewhere in this issue of
Nature4,5. Translational-research programmes
need to be expanded to connect this basic sci-
ence with existing vaccine-development tools,
including hypothesis-driven clinical trials to
assess novel immunogen designs. This will
probably require solutions to two problems
that have remained unsolved for more than
20 years: the design of immunogens to elicit
broadly neutralizing antibodies to prevent
HIV infection, and the design of immunogens
to elicit robust cellular immune responses to
control HIV infection.
Most importantly, the field should launch
flexible, large-scale, long-term funding mech-
anisms ideally by the beginning of 2011 that
invest in multidisciplinary teams rather than
in projects. Such mechanisms would foster
greater innovation and shorten the timeline to
a safe and effective HIV vaccine.
The neutralizing-antibody problem
Licensed viral vaccines protect against dis-
ease by priming the immune system before
pathogen exposure. This generates antibody
responses that can prevent infection, and
cellular responses that target and eliminate
virus-infected cells. Virus-specific neutraliz-
ing antibodies bind to proteins on the surface
of viral particles and stop them from infecting
host cells. Neutralizing antibodies can also bring
about the destruction of virus-infected cells, via
cellular effector mechanisms. Where natural
virus infection elicits robust neutralizing-anti-
body responses, vaccines have been developed
by using attenuated versions of the live virus
© 2010 Macmillan Publishers Limited. All rights reserved
IN NEED OF A VACCINE
The number of people living with HIV* continues to grow, and AIDS-
related illnesses remain one of the leading causes of death globally.
Estimated HIV
prevalence (millions)
0 0.02 0.15 1.0 2.5 6.0
No data
(such as for measles, mumps and rubella);
inactivation (for polio); and virus surface protein
subunits or virus-like particles (for hepatitis B
and human papilloma virus).
There are, however, three formidable
obstacles for vaccine developers attempting
to raise HIV-specific neutralizing antibodies.
First, the virus is hypervariable so a vaccine
must elicit broadly neutralizing antibodies
to counter myriad circulating isolates of HIV.
Second, the target for broadly neutralizing anti-
bodies, the envelope spike, is very unstable and
so is difficult to recreate in a form that can act
as a vaccine6. Lastly, most highly conserved tar-
gets for HIV-neutralizing antibodies are hard
to access on the virus spikes7. However, broad
and potent neutralizing antibodies against
HIV have been identified from HIV-infected
subjects8, demonstrating the feasibility of
inducing such antibodies. Unfortunately,
experimental attempts to elicit such antibodies
have to date universally failed.
Recent technological advances — in struc-
tural biology, cryoelectron tomography, com-
putational biology, B-cell immunobiology, and
high-throughput robotic micro-neutralization
screening assays — have been brought together
to try to ‘reverse engineer’ a solution to this
problem9. These studies may inform the rational
design of vaccines against other highly vari-
able viruses such as influenza and hepatitis C.
OPINION
2008 REPORT ON THE GLOBAL AIDS EPIDEMIC (UNAIDS, 2008)
*2007 estimates
This approach identifies infected subjects
with broadly neutralizing serum antibody
responses, isolates the corresponding broadly
neutralizing monoclonal antibodies (bnMAbs),
characterizes the interaction of these bnMAbs
with the virus envelope, and then engineers
immunogens based on this information.
However, in an era of global economic
uncertainty, the multidisciplinary centres and
consortia likely to be required to adequately
address this problem are currently too few and
in jeopardy of not achieving critical mass or
long-term commitment. This is due in large
part to increasingly restrictive funding mech-
anisms such as milestone based, short-term,
project-specific funding.
The cellular immune-response problem
Several lines of evidence, from human HIV
natural history and simian immunodeficiency
virus (SIV) studies, suggest that cell-mediated
immune responses are required to control
HIV by keeping the viral load very low or
undetectable. Because control of infection is
required to prevent disease, and as the best
licensed vaccines against other pathogens
do not necessarily completely prevent infec-
tion, a successful HIV vaccine will probably
also need to elicit cell-mediated immune
(CMI) responses capable of controlling HIV
infection. This is difficult because the virus
161
OPINION
becomes persistently established in host-cell
reservoirs within days of exposure, allowing
only a small window of opportunity for CMI-
based control.
Three HIV vaccine concepts have now
advanced through efficacy trials and none
controlled HIV infection as measured by viral
load in the blood. In contrast, a small subset of
HIV-infected people can control HIV infection
to nearly undetectable levels without antiret-
roviral therapy10, and some SIV vaccines can
control pathogenic SIV infections to similarly
low viral loads11. To solve this problem, three
key issues need to be addressed. First, elucidat-
ing the mechanisms of CMI-based control of
HIV infection in humans and of SIV in SIV-
immunized monkeys would provide important
leads. Second, improving functional assays of
CMI responses would enhance preclinical and
clinical prioritization of candidate vaccines.
Third, researchers need to design and screen
immunogens capable of outwitting HIV’s ability
to mutate rapidly and evade CMI responses.
Again, such complex challenges require
multidisciplinary teams, and a different
approach to funding translational research.
Enhancing translational research
Total global investment for HIV vaccine R&D
for 2008 was US$868 million (see chart), a
10% decrease from 2007 (ref. 12). However,
only 5–10% was focused directly on trying to
solve the two major vaccine-design problems
detailed above. Moreover, funding for trans-
lational research has generally been limited
to 3–5 year programmes that primarily sup-
port small academic consortia, and lack many
of the industrial-style disciplines needed for
vaccine discovery and development. Collec-
tively, such projects have a very low probabil-
ity of success. In addition, academic scientists
must split their time between research, getting
grants, teaching and fulfilling administrative
and managerial responsibilities, limiting their
time focused on the scientific problems.
The funding mechanisms from national
public-sector research agencies have been
effective in fostering basic research, primarily
through investigator-initiated programmes.
These should continue. However, these mecha-
nisms are not optimal for attracting new talent
to HIV vaccine research, fostering innovation in
vaccine discovery, establishing links with indus-
try or providing long-term multidisciplinary
commitment. It is unfortunate then that phil-
anthropic foundations and other donors aiming
to fill the translational science gap have increas-
ingly tilted towards the public-sector model.
There are examples of successful integration
of basic and translational research practices
that should be learned from and expanded
162
HIV VACCINE INVESTMENT
In 2008, global research-and-development
spend totalled US$868 million.
Public sector
$731 million
Philanthropy
$104 million
Industry
$33 million
on. These include the Neutralizing Antibody
Consortium (NAC) set up by the Interna-
tional AIDS Vaccine Initiative (IAVI); the
US National Institute of Allergy and Infec-
tious Diseases (NIAID) intramural Vaccine
Research Center; and the recently established
Ragon Institute in Boston, Massachusetts.
Teams associated with these three institutions
have been responsible for many of the most
promising recent advances2,7,10,13–15.
These institutions have established prac-
tices that emphasize long-term (eight years or
more) institutional commitment. They reward
proven scientific leadership and track records
rather than project-specific proposals; they
try to link researchers with those from out-
side HIV vaccine R&D; and they limit major
project reporting and reviews of laboratories
to multi-year cycles, increasing time for HIV
vaccine design.
Building on these successes, three new trans-
lational-research funding mechanisms should
be instituted.
Young investigator awards should be set up
by major stakeholders of the Global HIV Vac-
cine Enterprise (www.hivvaccineenterprise.org)
for scientists under age 35. These awards would
attract the best new scientists to the challenges
of HIV vaccine R&D by offering 5–7 years of
salary and flexible funding. Investigators would
be free, within broadly defined research pro-
grammes, to redirect their funds as new data
emerge, in exchange for expending at least 75%
of their effort on the two major scientific prob-
lems noted above. Developing such grants will
require stakeholders, including public-sector
research agencies and non-profit foundations
to invest in young scientists as components of
multidisciplinary teams, rather than relying on
traditional investments in specific projects.
Incentives for expanded biopharmaceutical
investment need to be set up. Vaccines are
made in industry, yet industrial involvement
in HIV vaccine R&D is minimal because of
the scientific challenges, market disincentives
and opportunity costs. Support for industrial
investment — from the Bill & Melinda Gates
Foundation and IAVI — to Theraclone Sci-
ences in Seattle, Washington, and Monogram
Biosciences in San Francisco, California, led
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
to the application of B-cell screening and
ADAPTED FROM REF. 12
micro-neutralization technologies central to
the recent identification of a new target on
HIV for broadly neutralizing antibodies2.
Additional mechanisms to enhance bio-
pharmaceutical investment, particularly
in process development for vaccines, could
include advanced market commitments,
intellectual-property incentives, and direct
public-sector support for vaccine R&D.
Translational-research teams must be estab-
lished. The Howard Hughes Medical Institute
(HHMI) successfully supports internation-
ally recognized basic-research scientists with
a salary plus a portion of laboratory costs for a
minimum of five years, without restriction on
their research direction. This attracts talented
scientists and fosters innovation. Few HHMI-
like mechanisms exist for teams focused on
HIV vaccine translational research. Ten per
cent of the annual investment in HIV vaccine
R&D should be refocused to provide such a
mechanism for HIV vaccine development.
In sum, major funders will bring a successful
HIV vaccine closer by establishing long-term
translational-research programmes, attracting
scientific talent and technologies, providing
greater incentives for industry participation
and focusing on rational vaccine design. These
initiatives should be formulated over the com-
ing months, debated at the international AIDS
Vaccine 2010 conference starting on 28 Sep-
tember in Atlanta, Georgia, and implemented
by the beginning of next year. This effort will
also provide a framework for the rational
design of vaccines against other global infec-
tious diseases. ■
Wayne C. Koff is senior vice-president of
research and development, and chief scientific
officer at the International AIDS Vaccine
Initiative, 110 William Street, 27th Floor, New
York, New York 10038, USA.
e-mail: wkoff@iavi.org
1. Rerks-Ngarm, S. et al. N. Engl. J. Med. 361, 2209–2220
(2009).
2. Walker, L. M. et al. Science 326, 285–289 (2009).
3. Hansen, S. G. et al. Nature Med. 15, 293–299 (2009).
4. Virgin, H. W. & Walker, B. D. Nature 464, 224–231 (2010).
5. Haase, A. T. Nature 464, 217–223 (2010).
6. Pancera, M. et al. Proc. Natl Acad. Sci. USA 107, 1166-1171
(2010).
7. Chen, L. et al. Science 326, 1123–1127 (2009).
8. Stamatatos, L., Morris, L., Burton, D. R. & Mascola, J. R.
Nature Med. 15, 866–870 (2009).
9. Burton, D. R. et al. Nature Immunol. 5, 233–236 (2004).
10. Miura, T. et al. J. Virol. 83, 3407–3412 (2009).
11. Koff, W. C. et al. Nature Immunol. 7, 19–23 (2006).
12. HIV Vaccines and Microbicides Resource Tracking
Working Group Adapting to Realities: Trends in HIV
Prevention Research Funding (July 2009); available at
http://www.hivresourcetracking.org.
13. Simek, M. D. et al. J Virol. 83, 7337–7348 (2009).
14. Hessell, A. J. et al. Nature Med. 15, 951–954 (2009).
15. Košmrlj, A. et al. Nature (in the press).
See Reviews, page 217, and Perspectives, page 224.
Vol 464|11 March 2010
Evolution of the motor car
A proposed reinvention for urban motoring based on ultra-small electric vehicles does not
address the bigger environmental or social challenges, finds Daniel Sperling.
Reinventing the Automobile: Personal
Urban Mobility for the 21st Century
by William J. Mitchell, Christopher E.
Borroni-Bird and Lawrence D. Burns
MIT Press: 2010. 240 pp. $21.95, £16.95
How do we reconcile people’s desire to own
cars with the need to reduce pollution and
dependency on oil? This dilemma has inspired
much technological and policy innovation,
from hybrid electric vehicles and smart traffic
information to congestion charging and car-
sharing schemes. Yet these developments have
had little impact on the function and perform-
ance of our urban transport systems, which
have barely changed in the past 50 years.
As remarkable consumer products, conven-
tional cars can last for more than 10 years with
little maintenance and withstand huge forces
if they crash. But they still require someone to
grasp the steering wheel and press the pedals.
Because they run using combustion engines and
petroleum fuels, they have a huge carbon foot-
print. And they take up valuable space. They
are also an extravagance, sitting unused for 90%
of the time on average and costing more than
US$8,000 a year to own and operate.
The solution, according to authors
William Mitchell, Christopher Borroni-
Bird and Lawrence Burns, is to “reinvent the
automobile” to accommodate the growing
demand for cars in the megacities of India,
China and the rest of the world. What’s
needed, they propose, is an ultra-small vehi-
cle powered by electricity or hydrogen that
will not cause the environmental and social
© 2010 Macmillan Publishers Limited. All rights reserved
BOOKS & ARTS
disruption of today’s models, particularly be integrated with a smart grid and how the
SMART CITIES
in cities. Such a proposal might be expected adoption of road pricing might encourage the
from Mitchell, an architect who directs the transition to small, energy-efficient vehicles.
smart cities research group at the Media Lab Focusing on large cities, they do not address
of the Massachusetts Institute of Technology inter-city or even suburban travel.
in Cambridge, Massachusetts. Yet the promo- It is easy to be sceptical: small, electric,
tion of small vehicles seems surprising com- connected cars in big cities will not solve our
ing from Borroni-Bird, a General Motors transportation and energy challenges. Large,
executive, and Burns, the firm’s former head dense cities house only about 20% of the world’s
of research and development and strategic cars and light trucks. And there are many chal-
planning. General Motors is a company with lenges that the authors hardly mention. Will
a long commitment to large cars and trucks ultra-small vehicles be economically attractive?
that, in the past, has vociferously opposed Is there much demand for small cars? Will the
regulations on fuel and greenhouse gases pricing of road space be widely accepted? Is
aimed at reducing vehicle size. it feasible to build specialized roads for small
Reinventing the Automobile is vehicles in cities?
provocative, aiming to overturn “Reinventing the Yet they introduce some big
more than a century of vehicle ideas that could play an impor-
and urban design. It proposes car will be harder tant part in building livable cit-
four main ideas: transform the than putting a ies and a low-carbon world. More
basic principles of car design; man on the Moon connected and automated vehi-
connect vehicles wirelessly to cles — not just cars but trucks
share information on roads and — and much more operating on dedicated high-
traffic; set up ‘smart’ electric- important.” speed roads — are desirable and
ity grids that manage energy inevitable.
supply and demand; and impose real-time Reinventing the Automobile is well written
pricing and travel-on-demand services that and its ideas are illustrated by many draw-
give people greater flexibility and reduce the ings, graphs and tables. Above all the authors
cost of transportation systems. deserve credit for taking on a challenging issue.
The authors focus mainly on upgrading car As they put it, reinventing the car will be harder
design. They also address the practicalities of than putting a man on the Moon — and much
shifting to their proposed system, including more important. ■
how to convert combustion-engine vehicles Daniel Sperling is professor of engineering
to electric propulsion, how information and and environmental science and policy at the
communication technologies might be used University of California, Davis, California 95616,
to increase vehicle flows without increas- USA, and co-author of Two Billion Cars.
ing road space, how electric vehicles might e-mail: dsperling@ucdavis.edu
163
OPINION
Space to contemplate
The Edge of Physics: A Journey to Earth’s
Extremes to Unlock the Secrets of the
Universe/Dispatches from the Frontiers of
Cosmology
by Anil Ananthaswamy
Houghton Mifflin Harcourt/Gerald
Duckworth: 2010. 336 pp. $25/£16.99
Amazing discoveries tend to follow a satellite
launch or telescope’s first light. Physicists have
traced galaxies across most of the visible Uni-
verse and have seen the top quark. Yet they lie
awake at night worrying that they don’t know
what makes up 96% of the Universe, what dark
energy is or why anything has mass.
This gap between theory and observation
concerns science writer Anil Ananthaswamy.
In The Edge of Physics, he travels the planet to
see for himself the experiments that are chal-
lenging our view of nature. Along the way he
meets the researchers who suffer for their sci-
ence in bleak locations: from macho Siberians
who sleep on the heaving ice of Lake Baikal
to catch neutrinos, to technicians who work
every day deep in a Minnesotan mine to detect
dark matter. Ananthaswamy seeks to capture
the peculiar motivations that draw people to
these remote sites of discovery.
Confessing that his journey is a pilgrim-
age, Ananthaswamy begins and ends his tale
at observatories that share peaks with mon-
asteries. The first site is Mount Wilson in
164
California, where Edwin Hubble measured
the expansion of the Universe in the 1920s.
Ananthaswamy explains how a reverential atti-
tude was expected from astronomers working
there, and how for many years women were
banned from the site because they were con-
sidered a distraction. Reflecting this austere
atmosphere, the astronomers’ basic sleeping
accommodation is still called The Monastery.
Ananthaswamy stays on for a few nights with
the neighbouring community of Camaldolese
monks to contemplate the site in solitude and
silence. One of the monks, a distant relative of
George Ellery Hale who founded the observa-
tory in 1904, explains how he too seeks “a deep
experience that one can’t express”.
Ananthaswamy takes a lift some 700 metres
below ground into the vast caverns of Soudan
mine in Minnesota to interview scientists
working on the Cryogenic Dark Matter Search
experiment installed there. He flies to Siberia
to visit the neutrino-detection experiment in
Lake Baikal’s crystal-clear waters. Scientists
here must endure sub-zero winter temperatures
to check the detectors that hang on strings like
jellyfish tentacles deep below the frozen
surface. To bring good luck, the Russian physi-
cists toss some of their vodka onto the ice.
Chile’s Paranal observatory and Hawaii’s
Mauna Kea telescopes highlight clashes
between science and locale. Mauna Kea’s
volcanic peak is considered sacred by many
Hawaiians, who leave offerings on its summit
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
and have tried to limit development of the
ESO/H.H. HEYER
site. Astronomers have generally responded
with respect, by negotiating agreements to
protect the mountain’s unique environment.
By contrast, Paranal’s only neighbours in the
high Atacama desert are vast copper mines.
Chileans travel far to work in this arid land-
scape, where nothing grows except the lush
vegetation cocooned in the geodesic dome
of the observatory’s space-age hotel. The
gardener who tends this isolated oasis empa-
thizes with the astronomers: “We are devoted
to the things we know, and we are a very good
complement.”
The political agendas behind these experi-
ments are largely skated over. An exception is
the negotiations that have led to South Africa
— together with Australia — becoming a pro-
posed site for the Square Kilometre Array, a
giant radio telescope due to be built over the
next decade to scan for pulsars and the first
stars. This ambitious plan for South African
astronomy hinges on the enthusiasm and
savvy of Bernie Fanaroff, radio-astronomer-
turned-political adviser to the South African
government.
The final chapters explore the most
extreme experiments. Ananthaswamy heads
to Antarc tica to witness balloon launches
delayed by bad weather and the struggles to
bury the IceCube neutrino detectors beneath
the ice. Struck by the complexity and scale of
the Large Hadron Collider at CERN, Europe’s
particle-physics lab near Geneva, Switzer-
land, he pauses to note the tendency of the
researchers to describe the ephemeral par-
ticles they seek in certain terms. Although
NATURE|Vol 464|11 March 2010
Clear views from Chile’s Paranal Observatory
are protected through agreements to limit
light pollution by neighbouring mines.
he does not travel into space to visit the
Planck spacecraft in person, he writes of
how its forthcoming results on the cos-
mic microwave background will soon
challenge cosmologists’ knowledge of the
early Universe.
The Edge of Physics is an accomplished
and timely overview of modern cosmology
and particle astrophysics. Ananthaswamy’s
characterizations of the many physicists
he meets are on the mark; yet the actors’
number and brevity on the stage may
baffle the uninitiated. It is especially gratify-
ing that he spends time with the dedicated
technicians and support staff, not just the
big scientific names.
Ananthaswamy conveys that cutting-
edge science is a human endeavour.
Ending his journey, and his story, at India’s
Hanle observatory in the Himalayas, he
again notes the confluence of an observa-
tory and a monastery at a remote location.
He urges that these sites must be protected
from environmental threats such as climate
change and oil pipelines. A more press-
ing brake to grand experimentation is the
global recession. Yet his interviews show
that understanding ultimately comes from
individual curiosity and motivation, not
expensive equipment. Perhaps in the quieter
economic times ahead, physicists will have
more liberty to contemplate the Universe
around them.
Joanne Baker is Nature’s Books & Arts editor,
and an ex-observational cosmologist. She is
author of 50 Physics Ideas You Really Need to
Know.
Genius who shuns the limelight
Perfect Rigor: A Genius and the
Mathematical Breakthrough of the Century
by Masha Gessen
Houghton Mifflin Harcourt: 2009.
256 pp. $26
Mathematics rarely makes the news. One
exception was the proof of the century-old
Poincaré conjecture in 2003. When Russian
mathematician Grigory ‘Grisha’ Perelman
posted his solution on the Internet, people
took notice. Experts found that it was correct
and Perelman was awarded the highest
honour in maths, the Fields Medal, by the
International Mathematical Union in 2006.
Yet the laureate refused to attend the award
ceremony in Madrid, leaving King Juan Car-
los, who was to present the medal, waiting in
vain. Media interest again soared: here was a
long-haired weirdo who had flouted the King,
thus confirming every prejudice about math-
ematicians. Tales of priority disputes and
plagiarism quickly followed.
Given his supreme problem-solving ability
and eccentric behaviour, Perelman’s personal-
ity begs analysis. Perfect Rigor fills that need.
Science writer Masha Gessen has researched
the reclusive mathematician thoroughly,
interviewing his teachers, maths coaches and
colleagues in the United States, Russia and
Israel. She also consulted psychologists about
his behaviour, which, she suggests, has much
in common with Asperger’s syndrome.
Born in St Petersburg (at that time Lenin-
grad), Perelman became a maths prodigy. At
16 he won a gold medal at the 1982 Interna-
tional Mathematical Olympiad with a perfect
score. After studies at Leningrad State Univer-
sity, the Steklov Institute of Mathematics in St
Petersburg and a postdoc at the University of
■ Perfect Rigor reminds us that it is journalists
Grigory Perelman disappeared after proving the
Poincaré conjecture about the topology of a sphere.
© 2010 Macmillan Publishers Limited. All rights reserved
OPINION
California, Berkeley, he vanished for several
years. None of his colleagues knew it, but he
was working in solitude on a proof of the
Poincaré conjecture — a famous problem in
topology about three-dimensional boundaries
of four-dimensional spheres. After reappear-
ing briefly to explain his proof in a whirlwind
US tour, he vanished again. His whereabouts
today are unknown.
Gessen describes the best and the worst that
Soviet maths had to offer; the best being the
support given to gifted youths such as Perel-
man through maths clubs and competitions,
the worst being the rampant anti-Semitism
that pervaded maths departments until
perestroika in the late 1980s. Had it not been
for the strenuous efforts of some professors to
support Perelman, even a prodigy of his stature
might have been prevented from pursuing
graduate studies because of his Jewish back-
ground. Gessen, a self-described ‘maths junkie’
brought up in Moscow, is ideally placed to tell
of discriminatory practices in the Soviet
academic world — her parents are Jewish
engineers who suffered through such ordeals.
The book falls short in one important
respect: like others who have tried, Gessen
was unable to contact Perelman in person.
Putting on a brave face, she writes that she was
not burdened by any allegiance to Perelman’s
narrative. Unfortunately, lacking tangible
facts, she nevertheless speculates on why he
shunned both the medal and the public. With-
out any evidence from supporting interviews,
Gessen suggests that his great achievement
made him condescending towards the work
of others. This guesswork is unwarranted.
One person who knows the Russian mathe-
matician’s true motives is John Ball, the then-
president of the International Mathematical
Union who travelled to St Petersburg in an
attempt to persuade Perelman to accept the
F. ROBERTS/NEWSCOM
award. Ball reveals only that Perelman was
allegedly disappointed by the dishonourable
behaviour of some unnamed mathematicians.
Gessen, by proffering gratuitous speculations,
both misleads the reader and does Perelman
grave injustice.
Until 2006, the Poincaré conjecture was one
of the most famous open problems in maths;
now it is one more theorem. For Perelman,
proving the conjecture was sufficient reward in
itself — no prize or recognition was needed.
and the public who want more. ■
George Szpiro is a writer based in Jerusalem,
Israel. His forthcoming book is Numbers Rule.
e-mail: g.szpiro@nzz.ch
165
OPINION
Q&A: Peter Hessler on urbanization in China
In Country Driving, the final book in his China trilogy, Peter Hessler recounts his 11,000-kilometre drive across China to
see at first hand the effects of rapid industrialization. The New Yorker journalist explains how mass migration to cities
brings out people’s resourcefulness, but also how the speed of social and environmental change leads them to seek
meaning in their lives.
What did you hope to reveal about China
in your trilogy?
Each book has a different focus: River Town
(2001) looks at geography and place, Oracle
Bones (2006) looks at history and time, and
Country Driving looks at development and
economics. I examine those issues through
the experiences of the average Chinese
person, in particular people who are making
the transition from the countryside to the
city. They are the driving force of China’s
industrial revolution. A main point is
to look at how individuals, families and
communities respond to those changes and
new challenges.
How have improved roads and the
automobile boom affected the nation’s
rural regions?
Along my journey I was amazed by the pace
of transition. I saw how quickly villages in
northern China are emptying out because
young people are migrating to cities for
jobs, leaving young children behind to be
raised by their grandparents. Many rural
regions are far from urban centres so there
are few opportunities for villagers to improve
their livelihood. Roads and infrastructure
allow them to get out — the young people
hit the road for the east and the south.
Building highways is normally the first step
to urbanization. But in China, people can’t
buy and sell rural land. To develop it, local
governments often move farmers off their
land, turn it into urban land by building
highways and infrastructures, and then sell it
on at market rates.
You followed the progress of a newly
established factory town. What did you
notice?
I was struck by the adaptability of
individuals. Spending time in a factory town,
you realize just how quickly things are being
built and how rapidly urbanization is taking
place. One thing that impressed me was the
functionality of the development system —
it is flawed, but the flaws are not fatal. It is
very chaotic, the rules may be unfair, there is
corruption, people tend to have a short-term
view, but it works — for now at least. People
know the rules and know how to advance
to improve their situation. Those who move
166
Peter Hessler: struck by Chinese adaptability.
into factory towns become resourceful
and self-reliant. It is a tough world: when
negotiating jobs, they have to learn how to
manipulate people. There are positive aspects
to this, such as increased individualism and
capability, but the climate can be a bit brutal
and lawless.
How is China’s industrial revolution
different from those in Europe and the
United States?
One of the main differences is motivation.
A shortage of labour in Europe and
the United States inspired people to be
innovative and efficient. China has so many
workers that there is no need to save labour.
This makes it a good place for low-margin
production, but not so good for innovation
or for making high-value-added products.
China’s political system, which lacks party
competition, and its educational system,
which is very controlled and traditional,
also limit the possibility of innovation.
Has the country’s rapid development
affected people’s sense of wellbeing?
The pace of change can be overwhelming.
There is a breakdown in values and sense
of community. Families become more
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
separated, geographically and emotionally.
M. LEONG/REDUX/EYEVINE
There is a sense of emptiness, a sense of
something lacking. I met many ‘searchers’ in
China, people who are concerned about what
else there is in life beyond material success.
They want a community and people they
feel connected to. Many have developed a
religious instinct — becoming followers of all
sorts of faiths — but the religious institutions
in China are poorly developed.
How do people reconcile their sense of
individual identity with the imposed
collectivism?
Individualism is becoming more
characteristic in China. It arises outside
the education system. The people who
become migrants or entrepreneurs
educate themselves, change their lives
and adjust their outlook. And people have
learned to take different roles in different
circumstances: a farmer may act sometimes
as a villager, as a city person, an entrepreneur
or an obedient Party member. This is part
of the general flexibility and adaptability of
this generation: they are not stuck in a single
framework.
Does China’s political system help
or hinder its long-term economic
development?
In some ways it has been an asset in the
past few decades. This is a society that has
pulled more people out of poverty than any
other in history; it has a high literacy rate,
life expectancy has gone up, the per capita
income has increased. It deserves credit
for that. The question is how it gets to the
next stage of development. It may reach a
point when people become more resistant
and demanding. When that happens, the
system may either collapse or adapt to new
challenges. But we haven’t yet reached that
point. ■
Interview by Jane Qiu, Nature’s retained
correspondent in Beijing.
e-mail: jane@janeqiu.com
Country Driving: A Journey Through China
from Farm to Factory
by Peter Hessler
HarperCollins: 2010. 448 pp. $27.99
Vol 464|11 March 2010
STRUCTURAL BIOLOGY
When four become one
Robert Craigie
Every machine is made of parts. But, as the new structure of the HIV integrase enzyme in complex with
viral DNA shows, one could not have predicted from the individual parts just how this machine works.
Retroviruses such as human immunodeficiency
virus type 1 (HIV-1) integrate a DNA copy of
their genome into their host genome as an
obligatory step in their replication cycle. The
DNA cutting-and-pasting reactions that lead to
integration are mediated by the virus’s integrase
enzyme. The structures of this enzyme’s three
domains — the core catalytic domain, the
amino-terminal domain and the carboxy-
terminal domain — have been solved previously.
But the spatial organization of the domains with
viral DNA when in the active nucleoprotein
complex (known as the intasome), which car-
ries out integration, has remained unknown. In
this issue (page 232), Hare et al.1 describe the
structure of a retroviral intasome. The struc-
ture clarifies many poorly understood aspects
of retroviral-DNA integration. What’s more,
it provides a solid platform for understanding
the mechanism by which the inhibitors of this
essential viral enzyme act (Fig. 1).
Integration of retroviral DNA has been
extensively studied at the biochemical level.
Such analyses showed that the DNA cutting-
and-pasting steps are mediated by a mecha-
nism similar to that used by the transposase
enzymes found in bacteria and more complex
organisms. The structures of the catalytic
domain of retroviral integrases revealed2,3 an
active site that is common not only to these
transposase enzymes, but also to members of
a larger family of polynucleotidyl transferase
enzymes4. Beyond confirming the mechanism
of catalysis, however, these structures were
disappointingly uninformative. For instance,
they indicated that the catalytic domain is
dimeric, with the two active sites on opposite
sides of the nearly spherical dimer. This spacing
is incompatible with the spacing of the catalysis
sites on the two strands of host DNA. Clearly,
the functional complex would need at least a
pair of such dimers for two active sites to be
close to each other.
In addition to the structures of the individual
amino- and carboxy-terminal domains of retro-
viral integrases, the two-domain structures of
the catalytic domain with either the amino- or
the carboxy-terminal domain have been deter-
mined5. On the basis of these ‘partial’ struc-
tures, many models have been proposed for the
© 2010 Macmillan Publishers Limited. All rights reserved
NEWS & VIEWS
Amino-terminal domain Core catalytic domain Carboxy-terminal domain
Figure 1 | The pieces of the intasome puzzle. An integrase monomer consists of three structural domains:
the amino-terminal helical bundle, the core catalytic domain and the carboxy-terminal domain. Viral
DNA integration into the host genome is, however, catalysed by a highly stable nucleoprotein complex
called the intasome, in which four integrase monomers bridge a pair of viral DNA ends — a structure
that Hare et al.1 have now solved, putting the pieces of the puzzle together. Inhibitors of HIV-1 integrase
bind to the intasome rather than to free integrase, making the new structure particularly valuable.
structure of the intasome complex. Hare et al.1 tetramer and DNA, however, is unlike any of
show that all of the earlier models are wrong. the postulated models.
But why has the path to this level of under- Extensive surface contacts between integrase
standing been so tortuous? Retroviral inte- and viral DNA, as well as protein–protein
grases are notoriously difficult proteins to contacts, are the glue holding the structure
work with: even using mutations that improve together. The major contribution of DNA–
their solubility, the integrase and its DNA protein contacts to the stability of the com-
substrate aggregate under the conditions plex is reminiscent of the complex formation
required for assembly of the active complex. between the Tn5 transposase and DNA7. The
A more fundamental problem is that, on dimer interface of the catalytic domains,
initial binding to viral DNA, the integrase which is present in all crystal structures of this
does not show strong sequence specificity, domain, is preserved as expected. Only two of
and stable complexes are formed only after the the four active sites in the tetramer, one from
initial binding6. Mixing integrase with DNA each catalytic-domain dimer, associate with
therefore traps nonspecific aggregates that are viral DNA, whereas the amino- and carboxy-
unsuitable for structural studies. terminal domains of these subunits contribute
Hare et al. took advantage of their finding to stabilizing the tetramer. The other catalytic
that the integrase enzyme of the prototype domains in each dimer lie far from the DNA
foamy virus (PFV) — a close relative of HIV-1 substrate, and their corresponding amino- and
— has much more favourable biophysical carboxy-terminal domains are disordered.
properties than its HIV-1 counterpart. Using The past few years have seen great progress
PFV integrase, they thus successfully crystal- in the development of therapeutic antiviral
lized an integrase in complex with viral DNA, drugs that target integrase. These drugs have
a feat that has defied the best efforts of many been developed through large-scale screening
groups for the past two decades. efforts. The inhibitor raltegravir — like other
As predicted from the biochemical studies, promising integrase inhibitors — binds inte-
the authors’ intasome structure consists of a grase in complex with DNA rather than bind-
tetramer of integrase bridging a pair of viral ing the free enzyme in solution. Knowing little
DNA ends. The organization of the integrase about the integrase structures when in complex
167
NEWS & VIEWS
50 YEARS AGO
Strange World of the Moon. An
Enquiry into Lunar Physics. By
V. A. Firsoff — Persistent observers
of the Moon, many of them
amateurs with small instruments
and little scientific training, have
often reported changes of various
kinds in the appearance of the
lunar surface; these have tended
to be met with scepticism by
most professional astronomers
… The present book, which
attempts to interpret changes
seen by the author and other
observers in terms of known
physical and chemical laws, is
timely … in view of the increased
interest in the Moon resulting
from the new facilities offered
by space rockets … Mr. Firsoff
attributes ray systems to erosion
by running water, fed from large
deposits of ice accumulated
below the spongy surface rocks,
and he suggests that the maria,
also, are filled by impure ice …
Basically, however, the author’s
arguments only show that his
ideas are barely possible; he has
certainly not shown that they are
at all likely. In particular, the idea
that water is present on the lunar
surface is highly implausible.
From Nature 12 March 1960.
50 & 100 YEARS AGO
100 YEARS AGO
Survival of Man. A Study in
Unrecognised Human Faculty. By
Sir Oliver Lodge — To most people
the question of the survival of
human personality is the greatest
problem of life … At present
the chief interest of the subject
centres round the theory of cross
correspondences, emanating
chiefly from the so-called
controls that are manifested in
certain well-known and much
discussed automatic scripts, and
claim to represent the surviving
personalities of Myers, Hodgson,
and others … Similarly, indications
of known personalities, examples
of typical intellectual activities
continuing after earthly existence
has ceased, may accumulate to
such intensity that no hypothesis
is so simple or so effective as that
which involves the acceptance of
the belief in their manifest survival.
From Nature 10 March 1910.
168
with DNA, therefore, it has not been possible to
visualize the details of how these drugs interact
with the enzymes’ active site, much less to seri-
ously consider applying rational design to their
development.
Hare et al.1 now change that situation. Their
structure succinctly shows how integrase
inhibitors displace the 3ʹ end of the viral DNA
from the enzyme’s active site, thus blocking its
integration activity. Furthermore, they find
that the active site, which seemed to lie in a
featureless landscape in the earlier structures of
the catalytic domain alone, is partially buried
in a rugged surface — features that may afford
potential sites for anchoring inhibitors.
Despite the giant leap forward that this
study1 takes our understanding of retroviral
integrases, questions remain. First, four of the
domains in the integrase tetramer are disor-
dered. Are they involved in the interaction
with the host DNA, as Hare et al. suggest?
Do they mediate another function, or are
they simply redundant? Also, the target host
DNA is not present in the structure, although
there seems to be only one place for it to
ATMOSPHERIC CHEMISTRY
Wider role for airborne chlorine
Roland von Glasow
Unexpected chlorine chemistry in the lowest part of the atmosphere can
affect the cycling of nitrogen oxides and the production of ozone, and
reduce the lifetime of the greenhouse gas methane.
When people hear about chlorine in the atmos-
phere, many probably think of the chlorine
that comes from man-made chlorofluoro-
carbons (CFCs) and harms the ozone layer
in the stratosphere, some 25 kilometres above
Earth’s surface. However, on page 271 of this
issue, Thornton et al.1 report their observa-
tions of large amounts of reactive chlorine
compounds at Earth’s surface, in the middle
of the continental United States.
The authors’ measurement site is far from
known natural chlorine sources, such as the
ocean, or strong anthropogenic sources (for
example, coal combustion or steel production).
These findings are important because they
reveal a previously unconsidered chemical
pathway that generates a large amount of
chlorine radicals in the troposphere (the lowest
10 kilometres or so of the atmosphere). Such
radicals will react with many volatile organic
compounds (VOCs), including the potent green-
house gas methane (CH4), speeding up their
removal from the atmosphere. Furthermore,
this pathway will extend the lifetime of nitro-
gen oxides, the main precursors for ozone pro-
duction in the troposphere. Ozone is central to
atmospheric chemistry, but can also be a pollut-
ant — in high concentrations it is toxic to animals
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
go without drastic structural rearrangement.
The stark contrast between reality and the
now-redundant previous models cautions
against embarking on exuberant model build-
ing with few constraints imposed by experi-
mentation. That said, the integrase enzymes
of HIV-1 and PFV are sufficiently similar to
justify the use of PFV integrase in complex
with both viral DNA and antiviral inhibitors
for reliably modelling the binding of these
inhibitors to the HIV-1 enzyme. ■
Robert Craigie is in the Laboratory of Molecular
Biology, National Institute of Diabetes and
Digestive and Kidney Diseases, National Institutes
of Health, Bethesda, Maryland 20892, USA.
e-mail: bobc@helix.nih.gov
1. Hare, S., Gupta, S. S., Volkov, E., Engelman, A. &
Cherepanov, P. Nature 464, 232–236 (2010).
2. Dyda, F. et al. Science 266, 1981–1986 (1994).
3. Bujacz, G. et al. J. Mol. Biol. 253, 333–346 (1995).
4. Rice, P., Craigie, R. & Davies, D. R. Curr. Opin. Struct. Biol. 6,
76–83 (1996).
5. Chiu, T. K. & Davies, D. R. Curr. Top. Med. Chem. 4, 965–977
(2004).
6. Li, M. et al. EMBO J. 25, 1295–1304 (2006).
7. Davies, D. R. et al. Science 289, 77–85 (2000).
and plants, and it acts as a greenhouse gas.
To understand the relevance of Thornton
and colleagues’ discovery, it is necessary to
briefly explain the general role of nitrogen
oxides in the atmosphere. Nitrogen oxides
are released from anthropogenic combus-
tion processes (for example, in cars and power
plants) or the burning of biomass, but are also
produced naturally by lightning. In the tropo-
sphere, the light-induced reaction of nitrogen
dioxide (NO2) is the only way to form ozone.
Nitrogen dioxide can also react with ozone to
produce nitrate radicals (NO3•), which can then
react with more NO2 to form dinitrogen pent-
oxide (N2O5). In other words, two molecules of
NO2 are ‘stored’ in one molecule of N2O5.
Dinitrogen pentoxide is stable only at low
temperatures in the absence of light. If exposed
to sunlight, the two molecules of NO2 stored
within it are released and can continue to pro-
duce ozone. It has been shown2,3 that a strong
loss process for N2O5 is its reaction on the
surface of aerosol (Fig. 1) — tiny omnipres-
ent particles in the atmosphere that originate
from sources such as combustion processes,
dust and salt from the oceans.
Thornton and colleagues’ atmospheric meas-
urements1, taken at night just outside Boulder,
NATURE|Vol 464|11 March 2010 NEWS & VIEWS

States alone is much larger than the first global

a Night-time b Night-time Daytime estimate of the compound’s production6, and
ClNO2 Sunlight is similar to recent estimates of global ClNO2
production in coastal regions5.
HCl N2O5 HCl N2O5
Clƭ + NO2 If other continental regions also act as
CH4 and sources of ClNO2, then the cycling of nitrogen
Many volatile oxides and chlorine through this compound is
steps organic
compounds
of global relevance. A more thorough investiga-
2NO3– NO3– tion of this reaction pathway is therefore war-
Cl– Cl– ranted, involving many more measurements of
ClNO2 and of background chlorine compounds
Aerosol in general. As a large part of the uncertainties
particle
in Thornton and colleagues’ model relate to the
O3
actual reaction of N2O5 on aerosol particles, this
Deposition Deposition reaction should also be studied in more detail,
in the hope of finding generally applicable
Ground
parameters for the process.
Numerical models of atmospheric halogen
Conventional model New model
chemistry have to include many compounds to
Figure 1 | Schematic of the uptake of N2O5 on aerosol particles. a, Conventional picture of aerosol fully describe the various reaction cycles. Yet
uptake of hydrochloric acid (HCl) and dinitrogen pentoxide (N2O5), leading to chloride (Cl−) and many of these compounds, including certain
nitrate (NO3−) ions, respectively. In the absence of further reactions both are eventually deposited on key intermediates such as hypobromous acid
the ground with the aerosol particle. b, Mechanism supported by the measurements of Thornton (HOBr) and hypochlorous acid (HOCl), have
et al.1: nitryl chloride (ClNO2) is formed at night in the reaction of N2O5 on particles containing so far not been identified in the atmosphere,
chloride. The next day, sunlight breaks ClNO2 into chlorine radicals (Cl•) and nitrogen dioxide
despite the fact that new instruments have pro-
(NO2). The chlorine radicals can then attack the greenhouse gas methane (CH4) and volatile organic
compounds, and the resulting peroxy radicals and NO2 lead to the formation of ozone (O3).
vided much insight into halogen reaction cycles
in recent years1,5. To keep making progress,
scientists working in the laboratory, the field
Colorado, indicate an additional major reaction understanding of where this kind of chemistry and with numerical models must collaborate
pathway for N2O5 (Fig. 1). If it reacts at night on might be most prominent. at the early stages of projects, to find ways of
aerosol particles that contain chloride, then one On this basis, Thornton et al.1 assumed that detecting and quantifying ‘new’ compounds7.
of the two NO2 molecules stored within it can the ingredients for ClNO2 formation are ubiq- Such an approach would be invaluable in
form nitryl chloride (ClNO2). This is released uitous in the atmosphere of the continental answering the questions raised by Thornton and
to the atmosphere, whereas the other NO2 mol- United States and used a computer model of colleagues’ study1. ■
ecule forms a nitrate ion (NO3−). Under most the atmosphere to estimate the continent-wide Roland von Glasow is at the School of
conditions, nitrate ions remain in the particles impact of this reaction. A notable uncertainty Environmental Sciences, University of East
and are eventually deposited on the ground. The in the model is the efficiency of ClNO2 forma- Anglia, Norwich NR4 7TJ, UK.
following morning, ClNO2 is broken down by tion, which depends on the available surface e-mail: r.von-glasow@uea.ac.uk
sunlight into a chlorine radical and NO2. This area of aerosol particles, the particles’ composi-
1. Thornton, J. A. et al. Nature 464, 271–274 (2010).
effectively extends the lifetime of NO2, allowing tion (especially their chloride content) and the 2. Dentener, F. J. & Crutzen, P. J. J. Geophys. Res. 98,
it to produce more ozone and to be transported relative humidity. The authors’ simulations sug- 7149–7163 (1993).
by air currents away from its source. The chlor- gest that 8–22% of all nitrogen oxides emitted 3. Evans, M. J. & Jacob, D. J. Geophys. Res. Lett.
doi:10.1029/2005GL022469 (2005).
ine radical, meanwhile, can react with VOCs, in the United States cycle through the ClNO2 4. Finlayson-Pitts, B. J., Ezell, M. J. & Pitts, J. N. Nature 337,
including CH4. The chain of reactions that fol- pathway, which will therefore affect the lifetime 241–244 (1989).
lows eventually leads to the production of per- and transport of those compounds as described 5. Osthoff, H. D. et al. Nature Geosci. 1, 324–328 (2008).
6. Erickson, D. J., Seuzaret, C., Keene, W. C. & Gong, S. L.
oxy radicals that (along with nitrogen oxides) earlier. The researchers’ estimate of the amount J. Geophys. Res. 104, 8347–8372 (1999).
are key ingredients of ozone formation. of ClNO2 produced over the continental United 7. www.hitt-task.net
This reaction pathway for N2O5 had already
been observed4 in the laboratory in 1989, but
confirmation of its role in the atmosphere
didn’t occur until 2008, when Osthoff et al.5 SUPRAMOLECULAR CHEMISTRY
observed high ClNO2 concentrations in coastal
regions and ship plumes — places where chlor-
ide from sea salt meets pollution. Crucially,
Sticking to sugars
Thornton et al.1 have now observed the same
pathway well away from likely local sources of
Anthony P. Davis
chloride, and provide compelling reasons to If evolution has had trouble making effective carbohydrate receptors,
believe that this chemistry is more widespread
than previously thought. The authors suggest
what hope do humans have of creating synthetic versions? A method for
that the chloride required for the production preparing libraries of such receptors boosts the chances of success.
of ClNO2 could be supplied from a plethora
of man-made and natural sources. Long-term Carbohydrate recognition — the process in receptors for carbohydrates, and are beginning
atmospheric measurements from deposition which receptors bind to specific carbohydrate to succeed. Reporting in Angewandte Chemie,
networks at many locations in the United States molecules — is a difficult task, best left, one Pal et al.1 describe a family of peptide-based
suggest that this might indeed be the case, but might think, to biological molecules that have receptors that incorporates synthetic, carbo-
the sources and amounts of ‘background’ chlor- evolved to do the job. Chemists have neverthe- hydrate-binding amino acids. By screening
ide need to be better quantified to improve our less taken up the challenge of finding synthetic 400 variants of these receptors, the authors
169
© 2010 Macmillan Publishers Limited. All rights reserved
NEWS & VIEWS
discovered one that selectively binds to a
tumour-associated carbohydrate that might
be used as a diagnostic marker for cancer.
The field of supramolecular chemistry
concerns systems that assemble from mol-
ecules (and/or ions), and therefore involves
the design of molecules that interact con-
trollably with other chemical species. Of the
various goals in the area, carbohydrate recog-
nition in water is among the most challeng-
ing. A typical carbohydrate, such as glucose, is
festooned with hydroxyl (OH) groups, which
can make strong hydrogen bonds to recep-
tors in non-polar solvents. But in water, which
has hydroxyl groups of its own, carbohydrate
hydroxyl groups merely serve as camouflage:
although carbohydrate hydroxyls can form
hydrogen bonds to receptors in aqueous solu-
tion, they must displace water molecules in the
process. This creates an extra energetic hurdle
to binding compared with the same process in
non-polar solvents.
To create a driving force for carbohydrate
recognition in water, substrates must there-
fore fit exceptionally well into their receptors.
Evidence suggests that even biological recep-
tors find this requirement difficult to achieve,
as the binding of carbohydrates to lectins (the
main class of carbohydrate-binding proteins)
is often quite weak2.
The complexity of carbohydrate structures
increases the problem even further. The build-
ing blocks of carbohydrates are sugars known
as monosaccharides. Unlike proteins or nucleic
acids, which form only one kind of bond
between their building blocks, there are sev-
eral different ways in which monosaccharides
can link together. A small set of these mono-
mers can therefore generate a huge range of
oligosaccharide structures3. Such complexity
is clearly a headache for chemists trying to
make synthetic carbohydrate receptors, but it
is useful in nature — where oligosaccharides
provide efficient systems for information stor-
age, and are used extensively as labels for cells
and proteins4. Carbohydrate recognition thus
has roles in such diverse processes as fertili-
zation, infection, cell–cell recognition and
protein transport.
But synthetic carbohydrate receptors are
worth the effort involved in making them, as
they hold great potential. For example, they
could be used in diagnostic methods to detect
carbohydrate biomarkers of disease, comple-
menting the limited supply of natural lectins.
Perhaps more importantly, if receptors can be
made that inhibit naturally occurring carbo-
hydrate–protein interactions, they would open
up new avenues for drug discovery.
Supramolecular chemists have therefore
worked on carbohydrate recognition for about
20 years. Broadly speaking, they have split into
two groups. The first group has devised recep-
tors that rely solely on non-covalent inter-
actions with their substrates5 — an approach
that, until recently6,7, has had limited success
in water. The second group has exploited a
170
a OH HO O
R1 B + R1 B + 2H2O
OH HO O a short peptide in which two other structural
Boronic Diol
acid
b OH OH OH OH
OO O the receptors to see which, if any, bound to
HO OR2 TF antigen
HO HN
O Although most of the library showed no sign
c O O structure (Fig. 1c) bound to TF antigen rela-
B OH HO B
NH HN O Moreover, Pal et al. found that carbohydrates
O
O H O H O
R3 N N O other than TF antigen did not bind to this
N N N receptor, unless they were closely related in
H O H O H
MeO
Figure 1 | Synthetic carbohydrate receptors.  that the affinity of TF antigen for this com-
a, Boronic acids react reversibly with diol
groups in molecules. B is boron, R1 can be any
organic group. b, Diol groups are ubiquitous in
carbohydrates, such as TF antigen, a tumour-
associated disaccharide. One of the diol groups is
highlighted in blue. R2 can be a lipid or a protein.
c, Pal et al.1 have made a library of 400 synthetic
carbohydrate receptors. The example shown
is the receptor that binds most strongly to TF
antigen. The boroxole groups (red) react with
diols in carbohydrates in a variant of the reaction
shown in a. The green parts of the molecule are
structural components that are varied to produce
other library members. R3 is a polymeric group
that is used to help dissolve the receptor, as
required for carbohydrate-binding assays.
reversible reaction between boronic acids
(a class of boron-containing compounds) and
diol groups (pairs of OH groups that are ubiqui-
tous in carbohydrates), thus making receptors
that form covalent bonds to their substrates8,9.
This reaction (Fig. 1a) proceeds well in water,
but it often requires high pH to work, a con-
dition that isn’t compatible with biological
processes. What’s more, it has proved difficult
to develop highly selective receptors based on
this process.
In spite of the problems, recent results6 sug-
gest that particular carbohydrates can indeed
be targeted by carefully designed synthetic
receptors. There remains, however, the issue of
generality — with so many targets of interest,
we really need a modular receptor system that
can adapt to bind to any given carbohydrate.
Pal et al.1 have succeeded in making just such a
system by merging the two main approaches of
carbohydrate-receptor design, and by using the
technique of combinatorial chemistry. Their
target substrate is TF antigen — a disaccharide
composed of galactose and N-acetylgalactos-
amine sugars that is present in 90% of human
cancers (Fig. 1b)10.
The authors’ general design for receptors1
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
involved two boron-based molecular units11,
selected for their ability to bind diols at neu-
tral pH. Pal et al. incorporated these units into
units could be varied (Fig. 1c). By installing
20 different variants at each of the two adjust-
able positions, the authors synthesized a library
of 400 different receptors. They then screened
TF antigen.
of binding, a few members were active. The best
tively weakly when compared with biological
receptors, but its performance compared well
with other state-of-the-art synthetic receptors.
structure to TF antigen.
When the authors made a compound that
mimicked the structure of their best receptor,
but which lacked boron atoms, they observed
pound was reduced — as would be expected
if the boron atoms are involved in substrate
binding. Intriguingly, however, the reduction
in affinity was not as much as would have been
anticipated. This implies that substrate binding
involves not only the boron-containing groups,
but also non-covalent interactions with other
parts of the receptor.
Screening of Pal and colleagues’ synthetic
receptor1 against a wide range of disaccharide
structures is required to show that it is genu-
inely selective for TF antigen. It may also be
necessary to increase the affinity of this class
of receptor for the target antigens before they
can be used in practical applications. Never-
theless, the success of Pal and colleagues’ strat-
egy for discovering carbohydrate receptors is
encouraging — there is no obvious reason why
receptors for other oligosaccharides could not
be found in this way, either by screening the
existing library of receptors or by synthesizing
a different set of compounds based on a modi-
fied general design. If this new method lives
up to its promise, carbohydrate recognition ‘on
demand’ may be just around the corner. ■
Anthony P. Davis is in the School of Chemistry,
University of Bristol, Bristol BS8 1TS, UK.
e-mail: anthony.davis@bristol.ac.uk
1. Pal, A., Bérubé, M. & Hall, D. G. Angew. Chem. Int. Edn 49,
1492–1495 (2010).
2. Toone, E. J. Curr. Opin. Struct. Biol. 4, 719–728 (1994).
3. Laine, R. A. Glycobiology 4, 759–767 (1994).
4. Gabius, H.-J. (ed.) The Sugar Code: Fundamentals of
Glycosciences (Wiley, 2009).
5. Davis, A. P. & Wareham, R. S. Angew. Chem. Int. Edn 38,
2978–2996 (1999).
6. Kubik, S. Angew. Chem. Int. Edn 48, 1722–1725 (2009).
7. Ferrand, Y., Crump, M. P. & Davis, A. P. Science 318,
619–622 (2007).
8. James, T. D., Sandanayake, K. R. A. S. & Shinkai, S. Angew.
Chem. Int. Edn Engl. 35, 1910–1922 (1996).
9. James, T. D., Phillips, M. D. & Shinkai, S. Boronic Acids in
Saccharide Recognition (R. Soc. Chem., 2006).
10. Yu, L.-G. Glycoconjugate J. 24, 411–420 (2007).
11. Bérubé, M., Dowlut, M. & Hall, D. G. J. Org. Chem. 73,
6471–6479 (2008).
NATURE|Vol 464|11 March 2010
SEX DETERMINATION
An avian sexual revolution
Lindsey A. Barske and Blanche Capel
Hormones are not all-powerful in determining whether birds develop with
male or female features. Chickens that are genetic sexual mosaics reveal
that individual cells also have a say in the matter.
Male and female vertebrates have long been
thought to owe their distinct sexual charac-
teristics — their sexual phenotypes — to the
action of hormones produced by the gonads.
This basic paradigm maintains that sex deter-
mination occurs during embryonic develop-
ment when genetic or environmental factors
instruct the bipotential gonads to differentiate
as ovaries or testes1,2. These organs, in turn,
secrete hormones that secondarily instil a
female or male phenotype in the reproductive
tract and brain — and, in species with external
sexual dimorphism, in tissues such as muscle
and feathers. According to this view, sex-
chromosome genotype is relevant only in the
gonadal cells where primary sex determination
occurs, because non-gonadal cells adopt a sex-
ual phenotype according to the hormones they
sense rather than their own complement of
sex chromosomes1. On page 237 of this issue,
however, Zhao et al.3 argue that, in birds, sex
determination occurs in cells across the entire
body, not just in the gonads.
Mammals use an XX/XY system to deter-
mine sex, whereby the Y-chromosome-linked
Sry gene activates testis differentiation. If an Sry
transgene is expressed in gonadal cells of an XX
embryo, the embryo will form testes and dif-
ferentiate as a phenotypic male, even though all
of its cells are XX. There is one clear exception:
marsupial mammals, in which differentiation
of a pouch or scrotum occurs before gonad dif-
ferentiation and depends on X-chromosome
dosage4,5. So animals with two X chromosomes
develop a pouch, and those with one develop
a scrotum.
Zhao and colleagues’ challenge3 to the classic
model is based on an analysis of avian gynan-
dromorphs. Gynandromorphs are naturally
occurring organisms that have both male and
female structures, with examples cropping up
among crustaceans, birds and insects. Depend-
ing on the stage of development at which
gynandromorphy is established, the division
between male and female tissues may fall along
the midline, producing a bilateral gynandro-
morph, or occur in a mosaic pattern across the
body. The condition can result from sex-chro-
mosome aneuploidy (that is, aberrant loss or
gain of a sex chromosome) or from abnormal
fertilization events.
Zhao et al. investigated the cellular basis
of this phenomenon in three chicken gynan-
dromorphs with bilateral asymmetry (Fig. 1).
Birds have ZZ/ZW sex chromosomes, where
males are homogametic (ZZ) and females are
© 2010 Macmillan Publishers Limited. All rights reserved
heterogametic (ZW). The authors analysed the
sex-chromosome genotype of numerous tissues
from each side of the body, expecting to find
evidence of aneuploidy. Instead, all assayed tis-
sues were mixtures of ZW and ZZ cells. Tissues
from the ‘male’ side of the body were biased
in favour of ZZ cells, whereas ‘female’ tissues
had higher proportions of ZW cells. Because all
tissues in the body are exposed to an identical
hormonal milieu, the authors argue that cells
must respond to hormones differently depend-
ing on their sex-chromosome constitution — in
other words, they have cell-autonomous sexual
identity. These results build on an analysis of
a zebra finch gynandromorph that was also
composed of Z- and W-bearing cells6. Like the
chicken gynandromorphs, the zebra finch dis-
played bilaterally asymmetric male and female
phenotypes, which, in this case, extended to
neural and behavioural traits.
To test their sexual-identity hypothesis,
Zhao and colleagues transplanted donor
tissue labelled with green fluorescent protein
into unlabelled hosts to construct same-sex or
mixed-sex chimaeric embryos. As expected,
donor cells integrated into host gonads, but
ZW donor cells did not adopt Sertoli (male)
cell fate and were not incorporated into testis
cords of ZZ hosts. Similarly, ZZ donors never
expressed the female marker aromatase in a
host ovary, suggesting that their sexual iden-
tity was already fixed by the time they arrived
in the host tissue.
These results contrast with mouse chimae-
ras established at the eight-cell stage, in which
XX (female) cells can adopt Sertoli cell fate and
XY (male) cells are incorporated into ovarian
follicles7,8. Likewise, human patients with a
mixed cellular complement of male and female
genotypes (46,XX/46,XY) display intersex
rather than mosaic phenotypes9, suggesting
that the dominant influence on mammalian
cells is the hormonal environment (although
there may be exceptions in the brain10).
Interestingly, another study11 involving arti-
ficial ZZ/ZW chicken chimaeras found not
only that these chimaeras were not gynandro-
morphic at sexual maturity, but that, in some
cases, ZZ donor cells actually induced ZW
hosts to develop as males with functional testes.
This discrepancy might be explained by differ-
ences in the developmental stage at which the
transplant was performed, the degree of mix-
ing between donor and host cells, or the pro-
portional representation of the two cell types.
Future work should determine when during
NEWS & VIEWS
Figure 1 | Funky chicken. Bilateral chicken
gynandromorphs, such as those studied by
Zhao et al.3, are composed of cells of both sexual
genotypes — ZZ (male, blue) and ZW (female,
pink). These mosaics resolve into male and
female patterns of coloration, wattle length and
comb height, despite exposure to an identical
hormonal environment. This suggests that avian
cells adopt a sexual identity on the basis of their
sex-chromosome genotype, not the gonadal
hormones they sense. (Concept by L. A. Barske.)
development sexual identity is established;
whether the proportions or level of mixing
between ZZ and ZW cells determines whether
a chimaera develops as a gynandromorph or a
simple phenotypic male or female; and whether
ZW-biased gynandromorph tissues that retain
a large ZZ component (or vice versa) retain a
subtle mosaic character or experience a com-
munity effect that permits resolution into a
pure female (or male) phenotype.
What factors could establish the sexual
identity of cells across the body? Genes that
are sex-linked and differentially expressed
early in development are good candidates.
Smith et al.12 recently offered compelling evi-
dence that the Z-chromosome gene DMRT1
acts as a dose-dependent master regulator of
male sex-determination in the chicken. But
DMRT1 is expressed at detectable levels only
in the gonads13 and is therefore unlikely to
determine sexual identity in other tissues. In
species such as the fruitfly Drosophila and
the nematode Caenorhabditis elegans, sex-
chromosome dosage compensation is closely
linked to the molecular pathway that controls
171
NEWS & VIEWS NATURE|Vol 464|11 March 2010

sex determination, such that each cell estab-
lishes its sexual identity at the same time that
X-chromosome dosage is regulated14. However,
even in Drosophila, a cell’s sexual identity can
be overridden by secreted signals15.
How should Zhao and colleagues’ results3 be
reconciled with the substantial body of work
involving hormone-induced sex reversal in
chickens? Seventy-five years of research have
definitively demonstrated that oestrogen is
necessary and sufficient for female develop-
ment: ZZ (male) embryos exposed to oestro-
gen develop as females, whereas ZW (female)
embryos depleted of oestrogen develop as
males16. ZZ ‘females’ revert to a male pheno-
type at puberty if the hormone treatment is dis-
continued16, possibly because ZZ cells cannot
produce enough oestrogen to maintain ovarian
structures. By contrast, adult ZW ‘males’ have
fairly normal-looking testes, indicating that,
when oestrogen is absent, ZW cells can become
Sertoli cells and organize into male patterns.
Do these hormone treatments simply override
cell-autonomous sexual identity? Investigation
into the adult phenotypes of castrated chick
embryos should show whether an underlying
sexual identity does indeed exist in the absence
of hormones. Although this procedure was pre-
viously attempted using irradiation, the animals
did not survive to hatching17.
A final question is whether cell-autonomous
sexual identity will turn out to be a common
element in the arsenal of sex-determining
systems in vertebrates, with variable influ-
ence on the outcome of sexual differentiation.
Perhaps it will. These funky chickens, oddities
of nature that they are, may well provide new
perspectives on questions of sexual identity
long thought to have been resolved. ■ been proposed to explain the observed recent
Lindsey A. Barske and Blanche Capel are in the
Department of Cell Biology, Duke University
Medical Center, Durham, North Carolina 27710,
USA.
e-mail: b.capel@cellbio.duke.edu
1. Jost, A. Phil. Trans. R. Soc. Lond. B 259, 119–130 (1970).
2. Lillie, F. R. Proc. Natl Acad. Sci. USA 3, 464–470 (1917).
3. Zhao, D. et al. Nature 464, 237–242 (2010).
4. O, W.-S., Short, R. V., Renfree, M. B. & Shaw, G. Nature 331,
716–717 (1988).
5. Renfree, M. B. et al. Phil. Trans. R. Soc. Lond. B 322, 41–53 (1988).
6. Agate, R. J. et al. Proc. Natl Acad. Sci. USA 100, 4873–4878
(2003).
7. Palmer, S. J. & Burgoyne, P. S. Development 112, 265–268
(1991).
COSMOLOGY
Gravity tested on cosmic scales
J. Anthony Tyson
Einstein’s theory of general relativity has been tested — and confirmed
— on scales far beyond those of our Solar System. But the results don’t
exclude all alternative theories of gravity.
Our understanding of the physics that under-
lies the dynamical evolution of the Universe
and the development of cosmic structure is
driven by astronomical observations. His-
torically, measurements on galaxy and larger
cosmological scales conflicted with predic-
tions based on a cosmological model that
combined Albert Einstein’s theory of gravity
(general relativity) and the standard model of
particle physics. Modifications to this model
were later introduced, involving dark matter
and dark energy (a dilute component that has
acceleration of the Universe’s expansion), that
8. Burgoyne, P. S., Buehr, M. & McLaren, A. Development 104,
683–688 (1988).
9. Krob, G., Braun, A. & Kuhnle, U. Eur. J. Pediatr. 153, 2–10
(1994).
10. Arnold, A. P. J. Neuroendocrinol. 21, 377–386 (2009).
11. Kagami, H., Clark, M. E., Verrinder Gibbins, A. M. & Etches,
R. J. Mol. Reprod. Dev. 42, 379–387 (1995).
12. Smith, C. A. et al. Nature 461, 267–271 (2009).
13. Raymond, C. S., Kettlewell, J. R., Hirsch, B., Bardwell, V. J. &
Zarkower, D. Dev. Biol. 215, 208–220 (1999).
14. Cline, T. W. & Meyer, B. J. Annu. Rev. Genet. 30, 637–702
(1996).
15. DeFalco, T. et al. Dev. Cell 14, 275–286 (2008).
16. Smith, C. A. & Sinclair, A. H. BioEssays 26, 120–132 (2004).
17. Wolff, E. & Wolff, E. J. Exp. Zool. 116, 59–97 (1951).
accounted for a wide range of observations1.
But the physics of dark matter and dark energy
is not understood. Although there are sugges-
tions from particle physics about the nature
of dark matter, that of dark energy remains a
mystery.
Frustrated by the lack of theoretical candi-
dates for dark energy, some researchers have
instead considered models in which the Uni-
verse’s dynamics, and so gravity itself, devi-
ates from that predicted by general relativity
on ‘cosmological’ scales — those larger than
galaxies and clusters of galaxies. But how can
one tell the difference between models that

APPLIED ECOLOGY

Grass and the X factor
Symbiosis between plants and
microorganisms is widespread,
but research into its effects on the
composition of a plant community
is in its infancy. Such studies
require patience. Jennifer Rudgers
and colleagues have spent six
years investigating the relative
performance of cultivars of tall
fescue grass with different forms
of a symbiont, and now report the
results and their recommendations
for the use of this versatile yet
vexatious grass (J. A. Rudgers et al.
J. Appl. Ecol. doi:10.1111/j.1365-
2664.2010.01788.x).
Tall fescue could well be the plant
that makes a lawn near you. It is
also grown for forage (pictured).
It is highly invasive, however, and
counts as a weed in situations in
which plant diversity is desirable.
Like many plants, it does not live
alone, and has what might be called
the X factor — symbionts that in the
case of tall fescue often include the
fungus Neotyphodium coenophialum.
Like the grass, the fungus comes
in different genotypic forms. For
example, one common form (KY-31)
produces alkaloids that are toxic to
certain herbivores; another form
(AR-542) does not.
By planting experimental plots,
Rudgers et al. set out to test, among
other things, how two cultivars of
the tall fescue Lolium arundinaceum
inoculated with KY-31, AR-542 or
neither, affect the plant species
composition of the plots. The two
cultivars chosen were Georgia-5
and Jesup.
The authors’ salient finding is
that Georgia-5 plus AR-542 allows
greatest growth of other grasses and
herbaceous flowering plants, and
produces fewest inflorescences (and
so fewer seeds). Those characteristics
would be appropriate for applications
in which tall fescue has a job to do but
where diversity of vegetation is the
aim, for example in preventing soil
erosion. There would also be reduced
fescue spread from the planting
site. Jesup with either symbiont is
preferable for monoculture.
CAPMC
The significance of this line of
research extends well beyond the
specific performance and application
of tall fescue. Plant breeders are
becoming ever-more imaginative in
exploring the intricate options offered
by different genotypic mixes of plants
and symbionts. Ecologists have a big
task on their hands in checking the
ecosystem consequences — and in
taking into account the many other
conditions that affect vegetation
growth and invasiveness.
Tim Lincoln

172
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010 NEWS & VIEWS

include dark energy and modified dynamics, they used Tegmark

M. BLANTON & SDSS

models of gravity? In 2007, Zhang
et al. 2 developed a method for
distinguishing between these two
cases that is minimally affected by
our lack of knowledge of some key
parameters. In this issue (page 256),
Reyes et al.3 apply Zhang and col-
leagues’ method to data from the
Sloan Digital Sky Survey (SDSS;
12 h
and colleagues’ measurement6
h of the distribution of galaxy
16 velocities (the redshift distortion
effect, which is visible in Figure
1 as radial (redshift) smear-
ing of many of the (red) galax-
ies). They took care to address
remaining systematic errors.
And after applying correction

8h
Fig. 1) and find consistency with factors derived from computer
general relativity and one modi- Redshift (z) simulations, they measured this
fied form of it. diagnostic of departure from

0.02

0.04

0.06

0.08

0.10

0.12

0.14
Reyes and colleagues’ measure- general relativity on scales of
20 h

ments are significant not just about 14–70 Mpc to a precision

because they are consistent within of 16%. Although the results are
error with general relativity, but consistent with general relativity
also because they point the way to and with one class of large-scale
future high-precision tests that will deviation from it, a particular
better distinguish between general example of a model of modified
relativity and some variant models. gravity, termed tensor–vector–
This is important because gravity h scalar, which aims to explain
4
is the least tested of the forces in both dark matter and the recent
nature. The predictions of grav- 0h cosmic acceleration, seems less
ity theory have undergone high- Figure 1 | Slices through the SDSS three-dimensional map of galaxies. Earth likely.
precision tests that have yielded is at the centre, and each point represents a galaxy. The radial coordinate is the Zhang et al.2 pointed out that
remarkable agreement4, but these redshift, and the angular coordinate, in units of hours, is the right ascension the next generation of surveys
were performed only on scales of (the celestial equivalent of longitude). The outer circle is at a distance of about will yield per-cent-level preci-
our Solar System or smaller. The 700 megaparsecs. Redder points denote galaxies composed of older stars. sion on this diagnostic. But these
effects of modified general rela- The regions between the slices are not mapped because dust in the Milky Way future surveys will do even more
tivity on scales a hundred billion obscures the view in these directions. The clustering of galaxies (caused by than that. With billions of gal-
times larger may leak down to the gravitational pull of huge unseen masses of dark matter) can be seen. Also axies charted across 80% of the
small-scale effects, and this could evident is the redshift-space distortion, a radial smearing of3 galaxies due to age of the Universe (100 times
their infall velocities near high-density regions. Reyes et al. combine these
be tested by high-precision meas- measures of the effects of gravity in a sample of galaxies from the Sloan Digital the number of galaxies in Reyes
urements of the Earth–Moon dis- Sky Survey (SDSS), which extends to somewhat higher redshift than those seen and colleagues’ study3), other
tance using laser ranging5. here, with another measure of the dark-matter mass in these cosmic structures: probes of dark energy7,8, which
Departures from general rela- the ‘gravitational lens distortion’ of the shapes of yet more distant galaxy images. could more sensitively test the
tivity on cosmological scales will modified general-relativity alter-
cause the growth of dark-matter native, will become possible. By
structure and the cosmic geometry to differ dark-matter fluctuation will cancel out in the combining observations of the cosmic micro-
from that predicted by Einstein’s theory, in ratio of two different methods for measuring wave background (relic radiation from the Big
ways that are in principle distinguishable. How- the dynamics of a large sample of galaxies. The Bang) with galaxy data, unprecedented studies
ever, a full investigation of any such effects, by two methods involve imaging and spectros- will become possible, such as directly observ-
tracing the dynamical history of the Universe copy. For a population of galaxies, the masses ing the homogeneity of the Universe on large
over cosmic time, is currently fraught with of the huge dark-matter haloes in which they scales. The recent acceleration of the Universe,
systematic errors. Perhaps if general relativ- are embedded can be determined in two ways: whether caused by dark energy or a manifes-
ity breaks down on scales of clusters of galax- from the distribution of galaxy velocities (via tation of a modification of gravity on scales
ies, there is sufficient information encoded in spectroscopy) and from the correlation of the a hundred billion times larger than the Solar
the dynamics and the mass of a large sample galaxies with the weak distortion of the shapes System, presages new physics. Experiments
of galaxies to detect this. Under the opposing of background galaxies by the effect of gravita- in the next decade promise even greater
influences of gravity due to dark matter and tional lensing (via imaging). Modified theories insight into the fundamental physics of the
the cosmic expansion, there is a natural rela- of gravity affect these two quantities differently Universe. ■
tionship between the clustering of galaxies, from Einstein’s gravity theory, producing an J. Anthony Tyson is in the Department of Physics,
the mass of the underlying dark matter and the observable change in their ratio compared with University of California, Davis, One Shields
velocity distribution of the galaxies that could the general-relativistic prediction. Avenue, Davis, California 95616, USA.
be used to test general relativity. But there are In a first measurement of this diagnostic, e-mail: tyson@physics.ucdavis.edu
challenges in going from observations of these Reyes et al.3 used imaging and spectroscopy of
properties to a test of any departure from grav- more than 70,000 luminous red galaxies in the 1. Spergel, D. N. et al. Astrophys. J. Suppl. Ser. 170, 377–408
(2007).
ity theory on large scales. SDSS at a mean distance of about 1,700 mega-
2. Zhang, P., Liguori, M., Bean, R. & Dodelson, S. Phys. Rev.
One problem is that galaxies are a biased parsecs. Clustering of these galaxies was meas- Lett. 99, 141302 (2007).
tracer of dark matter, and this bias varies with ured by correlating their relative positions in 3. Reyes, R. et al. Nature 464, 256–258 (2010).
scale. Another is that we do not know to suf- space. To calculate one measure of dynamics, 4. Will, C. M. Space Sci. Rev. 148, 3–13 (2009).
ficient precision how the density of dark matter the authors determined the mass versus scale 5. Murphy, T. W. Jr Space Sci. Rev. 148, 217–223
(2009).
fluctuates across space in the Universe. Zhang for the underlying dark-matter assembly by
6. Tegmark, M. et al. Phys. Rev. D 74, 123507 (2006).
et al.2 recognized that, if a combination of means of weak gravitational lensing, using the 7. Jain, B. & Zhang, P. Phys. Rev. D 78, 063503 (2008).
several types of measurement is used, the bias shapes of 30 million background galaxies, also 8. Song, Y.-S. & Koyama, K. J. Cosmol. Astroparticle Phys.
and other poorly known quantities such as the in the SDSS. For the second measure of the doi:10.1088/1475-7516/2009/01/048 (2009).

173
© 2010 Macmillan Publishers Limited. All rights reserved
www.nature.com/nature Vol 464 | Issue no. 7286 | 11 March 2010

EXOTIC MATTER

G
Cover illustration ases, liquids and solids are the three most REVIEW
Artwork by N. Spencer common types, or phases, of matter that
176 The enigma of supersolidity
we might think about day to day. But
these are not the only phases into which S. Balibar
matter can be classified. Consider, for instance, metals,
antiferromagnets, ferroelectrics, liquid crystals and PERSPECTIVE
Bose–Einstein condensates or the many solid phases
183 Superconductivity gets an
into which liquids can crystallize.
iron boost
There are numerous ways in which a large number
Editor, Nature
of particles can be ordered, whether the particles are I. I. Mazin
Philip Campbell
Publishing
fundamental (such as electrons) or composite (such
Nick Campbell as atoms and molecules). This ordering is determined REVIEW
Claudia Deasy by the quantum nature, mutual interactions and 187 Non-Abelian states of matter
Insights Editor symmetries that govern the particles. Crucially,
Karl Ziemelis A. Stern
the macroscopic properties of matter usually differ
Production Editor substantially from those of its microscopic constituents.
Davina Dadley-Moore PERSPECTIVE
In fact, numerous complex many-body systems show
Senior Art Editor
Martin Harrison
emergent phenomena that are associated with the 194 The birth of topological
Art Editor
‘whole’ but that cannot be understood solely in terms insulators
Nik Spencer of the fundamental laws that govern their microscopic
J. E. Moore
Sponsorship ‘parts’. This complexity is perhaps best encapsulated
Amélie Pequignot in Philip Anderson’s ‘more is different’ idea, from his
Reya Silao often-quoted 1972 Science article, and is part of the REVIEWS
Production challenge and the beauty of condensed-matter physics. 199 Spin liquids in frustrated
Jocelyn Hilton
In this Insight, we expose examples of matter magnets
Marketing
Elena Woodstock
for which ‘more is different’ holds true in striking L. Balents
Emily Elkins ways. We venture into the enigmatic world of
Editorial Assistant supersolids and superconductors, as well as into
Emma Gibson the frustrated landscapes inhabited by spin liquids 209 Electron liquids and solids in
and spin ice. We take a look at the one-dimensional one dimension
confines of nanowires and carbon nanotubes, where V. V. Deshpande, M. Bockrath,
correlated electrons form liquids and solids with L. I. Glazman & A. Yacoby
unique properties that are determined by their low
dimensionality. And we dive deep into the exotic
realms of topological insulators and non-Abelian states
of matter, where quantum effects come with a twist.
We hope that these articles provide a window onto
the world of many-body physics, which controls
myriad aspects of the world around us, and we
thank the authors and reviewers for their invaluable
contributions.
Dan Csontos, Associate Editor

175
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW NATURE|Vol 464|11 March 2010|doi:10.1038/nature08913

The enigma of supersolidity

Sebastien Balibar1

A ‘supersolid’ is a quantum solid in which a fraction of the mass is superfluid. As a remarkable consequence, it is
rigid, but part of its mass is able to flow owing to quantum physical processes. This paradoxical state of matter
was considered as a theoretical possibility as early as 1969, but its existence was discovered only in 2004, in
4
He. Since then, intense experimental and theoretical efforts have been made to explain the origins of this exotic
state of matter. It now seems that its physical interpretation is more complicated than originally thought.

Quantum properties of matter are paradoxical, especially when they
show up at the macroscopic level. This is true for superfluidity (Box 1),
and even more so for the newly found supersolidity of solid 4He. A solid
is rigid because its atoms (or molecules) occupy particular positions in
space — they are ‘localized’ — and are ‘distinguishable’, which means
that an atom at one position can be distinguished from another at a dif-
ferent position. By contrast, a superfluid moves without friction because
it is a collective wave of matter in which atoms are indistinguishable and
delocalized. Given this, it is not clear how a system can be ‘supersolid’,
which means being solid and superfluid at the same time. In a super-
solid, part of the mass flows without friction through the rest, which
remains solid, but there is no consensus on how this occurs. As will
become evident, the enigma of supersolidity “continues to defy agreed
theoretical explanation”, as A. J. Leggett said recently1.
In 2004, Kim and Chan2,3 found some experimental evidence for this
phenomenon in so-called torsional oscillator experiments on solid 4He
samples. Since then, several groups have confirmed that anomalies exist
not only in the rotation properties4–9 of solid 4He but also in its elastic
properties10–13 and in its specific heat14,15. The whole set of experimental
results strongly supports the hypothesis that solid 4He is indeed a super-
solid, but the interpretation of the data is probably not as simple as was
proposed 40 years ago.
There is some consensus among theorists that supersolidity does
not exist in perfect crystals16,17 and that defects such as dislocations in
single crystals, grain boundaries in polycrystals or glassy regions are
necessary. However, this consensus is not universal, as Anderson18–20
believes that supersolidity can be an intrinsic property of ideal crystals
that is only enhanced by disorder. The goal of this Review is to sum-
marize the present status of this intriguing field, to offer some possible
interpretations of the observed phenomena and to raise some issues
whose investigation could help solve the enigma of supersolidity. Three
extensive review articles have already been published16,21,22, but this sub-
ject evolves quickly.
From superfluids to supersolids
A classical liquid is made of atoms (or molecules) that move in a random
manner such as to make the liquid viscous (Box 1). In a superfluid,
however, atoms move coherently because together they form a macro-
scopic wave of matter. An important consequence of this coherence is
that superfluids do not rotate like classical liquids: in a slowly rotating
bucket, a superfluid stays at rest; it is not set in rotation by the moving
walls. If the rotation speed is increased beyond a critical value, then
some rotation appears inside the superfluid in the form of quantized
vortices. The same is true for superconductors subjected to magnetic
fields: a small field is totally excluded from the superconductor but
1
Laboratoire de Physique Statistique, Ecole Normale Supérieure and CNRS, 24 rue Lhomond, 75231 Paris, France.
beyond some threshold it may enter as so-called magnetic flux quanta,
which are quantized vortices of electrical charges.
In a classical solid, each individual particle resides on one particular
site — a lattice site if the solid is a crystal — and is localized such that the
crystal is rigid and responds elastically to shear stress. A classical solid
inside a box is forced to rotate with the box walls when the box rotates.
However, in a quantum solid, particles fluctuate a lot around their aver-
age positions, with the result that atoms may exchange places with their
neighbours. If this exchange is easy enough, some of the atoms may flow
through the otherwise rigid network and some of the mass may stay at
rest while the rest rotates. Eventually, if this flow becomes superfluid,
the solid is said to be supersolid. Some of the mass is delocalized and
the remainder is localized. The question to be answered is whether this
paradoxical state of matter really exists.
Let us start with some early ideas. In 1969, Thouless23 and, independ-
ently, Andreev and Lifshitz24 considered a quantum crystal in which, at
low temperatures, there could be ‘vacancies’, that is, empty sites where
atoms are missing (Fig. 1). These authors further conjectured that such
vacancies could behave like mobile quantum particles because they
could easily exchange their positions with neighbouring atoms through
quantum tunnelling. Anyone who has tried to solve a 16-tile magic-
square puzzle knows that exchange is much easier in a lattice if there are
empty sites. If the atoms are bosons, mobile vacancies should also obey
Bose–Einstein statistics and could therefore undergo Bose–Einstein
condensation, thus giving rise to superfluid-like behaviour below a cer-
tain temperature (Box 1). Crucially, a flow of vacancies being an inverse
flow of mass, it should be possible to observe the coexistence of solid
properties (an elastic response to shear stress) and superfluid proper-
ties (a fraction of the mass flowing without friction through the lattice).
This was the original proposal for the emergence of a new ‘superstate of
matter’ — a supersolid. Shortly after this suggestion, Leggett explained
that the rotation of this solid should be anomalous, like that of a super-
fluid liquid25. The problem with this explanation is that it relies on the
assumption that there exist vacancies in the zero-temperature limit, and
that is far from obvious.
First sighting of supersolidity
The suggestion of a new state of matter was attractive enough that many
research groups started looking for it26, but the first indication of its
existence was found in only 2004, by Kim and Chan2,3. They used a well-
documented method for the detection of anomalous rotation, which had
been used with superfluids. The experimental set-up consists of a tor-
sional oscillator containing a cylindrical cell with an annular space filled
with a sample material, such as liquid helium, suspended from a torsion
rod (Fig. 2a). At resonance, the period, τ, of the oscillator is related to

176
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010 REVIEW INSIGHT

Box 1 | Superstates of matter

Superfluidity, superconductivity and now
supersolidity — the list of superstates of
matter is growing. These quantum states
of matter differ significantly from classical
ones. In a classical fluid, particles (atoms or
molecules) move in a random manner and
are distinguishable from each other. Owing
to interparticle interactions, a classical fluid
has a non-zero viscosity. However, in 1937
in Cambridge, UK, Allen and Misener49
discovered that below a temperature of 2K,
liquid 4He enters a new state, which moves
without friction — a state named ‘superfluid’
by Kapitza50, who made similar observations
at the same time in Moscow51. A few months
later, Allen and Jones52 made spectacular
images of fountains of superfluid 4He. The
image (courtesy of J. F. Allen, University of St
Andrews) shows the famous ‘fountain effect’
(photographed in 1972). The inviscid flow of
superfluid 4He is produced by a small heater
below the neck of the glass bottle. This motion
could not be explained using the classical
laws of thermodynamics and hydrodynamics,
and triggered the proposal by London53 and
Tisza54 that superfluid 4He had to be a quantum
fluid in which atoms are indistinguishable
and accumulate in the same state in such a
way that they all move together in a coherent
manner — a phenomenon called Bose–Einstein
condensation (BEC) that pertains to bosons.
Helium-4 atoms are bosons, and it is now
generally accepted that superfluidity is
associated with a generalization of BEC that
was explained in 1951by Penrose55.
There is a close analogy with electrical
conductors. In an ordinary metal, electrical
resistance is a consequence of the incoherent
motion of electrons. However, Kamerlingh
Onnes56 discovered in 1911 (that is, more
than two decades before the discovery
of superfluidity) that at a sufficiently
low temperature, some metals enter a
‘superconducting’ state in which the electrical
resistance vanishes. In 1957, Bardeen, Cooper
and Schrieffer explained this phenomenon
as being due to the formation of a BEC of
electron pairs57,58 (the so-called BCS theory
of superconductivity) — single electrons
are fermions obeying Fermi statistics but
become bosons when paired up. Similarly,
the superfluidity of liquid 3He was discovered
near 2mK by Osheroff, Gully, Richardson and
Lee59,60 in 1972, and it was shown that because
3
He atoms are fermions, they also need to form
pairs to achieve a superfluid state. During the
past two decades, it has also been shown that
a similar BEC/BCS type of superfluidity can
be realized in bosonic61,62 and fermionic atom
gases. In fermionic atom gases, it has also been
shown how the superfluid transition can be
tuned continuously from BCS pairing to a BEC
of molecules as a function of the intensity of the
interactions between the atoms.
As discussed in this Review, the great
surprise today has been to find that a quantum
solid may also be superfluid — a new state of
matter called a ‘supersolid’.

the rotational inertia, I, through the relationship τ = 2π(I/K)1/2, where Κ
is an elastic constant determined mainly by the rigidity of the rod. The
tell-tale sign of the superfluid transition is a shift in the resonance period
of the oscillator because the inertia decreases when the material becomes
superfluid — the superfluid liquid stays at rest while the walls of the box
oscillate. The amplitude of the period shift thus varies with temperature,
T, as the fraction of the liquid that is superfluid increases from zero to
one as T decreases from the superfluid transition temperature to zero.
Furthermore, this period shift is observed only for oscillation velocities
smaller than a certain critical velocity (typically of the order of 1 cm s−1),
beyond which dissipation starts and superfluidity is destroyed.
The remarkable discovery by Kim and Chan in 2004 was that they
observed a similar shift in the oscillator period in solid, not liquid, 4He.
At a temperature of the order of 100 mK, the oscillator period decreased
(Fig. 2b), a behaviour that was absent in control experiments with either
an empty cell or a cell filled with 3He (recalling that 3He atoms are fermi-
ons, unlike 4He atoms, which are bosons (Box 1)). This was interpreted
as a strong indication of supersolidity. The main difference from super-
fluid liquid 4He was that, in Kim and Chan’s sample, the magnitude of
the shift indicated a superfluid fraction of the order of 1% of the total
4
He mass, but it seemed that at a sufficiently low temperature, and in
agreement with Leggett’s prediction, solid 4He did not rotate like a clas-
sical solid.
The nature of supersolidity
Kim and Chan’s discovery2,3 triggered a period of intense activity, with
several experimental groups trying to reproduce and better understand
the data and many theorists trying to explain the observed phenomenon
(see, for example, refs 16, 21 and 22). One of the critical issues early on
was the question of whether supersolidity could be an intrinsic property
of 4He crystals, as originally considered by Thouless23 and by Andreev
© 2010 Macmillan Publishers Limited. All rights reserved
and Lifshitz24, or a phenomenon solely dependent on the presence of
defects in the studied samples. The crucial point thus became determin-
ing whether there were vacancies in the ground state of a perfect 4He
crystal, that is, a crystal in equilibrium near T = 0 K, and whether the
presence of such vacancies could explain the observed anomalies.
Several theorists looked at the energy cost and gain associated with
creating vacancies in a crystal. On the one hand, removing one atom
from a crystal lattice costs potential energy because bonds between
atoms must be broken. On the other hand, if the vacancies are delocal-
ized, they are shared by all atoms instead of being confined inside a ‘cage’
made by nearest neighbours (Fig. 1). Heisenberg’s uncertainty principle
stipulates that a confinement in position space results in a large uncer-
tainty in momentum space and, consequently, in a large kinetic energy.
Thus, a delocalization of the vacancies would entail fewer fluctuations
of their velocity, meaning less kinetic energy; in short, the vacancies
would gain energy by being delocalized. If the energy gain from the
delocalization is larger than the potential energy cost of removing one
atom from the crystal lattice, the balance is negative and the ground
state must contain vacancies. In addition, if vacancies repel each other,
they form a dilute gas, which can be superfluid. In such circumstances,
the crystal would still be periodic in position space, but the number of
atoms would be less than the number of lattice sites; the crystal would
be ‘incommensurate’, and some of its mass, in the form of atoms jump-
ing between the lattice sites owing to the presence of vacancies, could
flow without friction. Below some superfluid transition temperature, it
would thus become supersolid.
For a long time, the simple vacancy-based model of supersolidity
looked possible. However, post-2004 Monte Carlo simulations showed
that the above energy balance is positive and large16,17,27; it was found
that the creation of vacancies costs more than 10 K (at the atomic level,
energies are often measured in kelvin, a temperature unit that translates
177
INSIGHT REVIEW
into energy according to E = kBT, where kB is the Boltzmann constant).
Because 10 K is much larger than 0.1 K, the typical temperature at which
supersolidity was observed, the probability of vacancies existing under
experimental conditions should be negligible, and they could not be
responsible for the observed phenomena.
Importance of disorder
Despite some disagreement18–20,22 about the validity of the Monte Carlo
simulations mentioned above, it was then proposed that supersolidity
might instead be caused by the presence of defects in the solid samples
under study. Two possible types of defect were suggested, dislocation
cores and grain boundaries (Fig. 3). One reason for this conjecture was
that these are regions of the crystal where local stresses naturally give
rise to the formation of vacancies, thus allowing the exchange of atoms
and mass flow. Indeed, theoretical studies predicted28,29 that part of the
mass inside defects could, under the right circumstances, become super-
fluid and propagate along dislocations and grain boundaries connected
in a three-dimensional network, thereby giving rise to the observed
supersolid behaviour at the macroscopic level.
At this stage, two types of experiment seemed to confirm that disorder
does have a crucial role in the supersolid-like behaviour. One line of
experiments continued the exploration of alternating (a.c.) mass flow
using torsional oscillators. Although Rittner and Reppy4,5 reproduced
Kim and Chan’s findings2,3, they also found that in annealed samples
(heat treatment typically reducing the defect density in crystals) the rota-
tion anomaly called ‘non-classical rotational inertia’ (NCRI) decreased
to below the noise level in the measurements. By contrast, when per-
forming the same type of measurement on quench-frozen samples, that
is, in the presence of large disorder, the NCRI was found to be as large as
20% of the total helium inertia. Subsequent experiments by the group of
Chan30 later confirmed that annealing reduces the NCRI. However, this
group found that even in their best samples, a very small fraction of the
mass, of the order of a few parts in 104, still decoupled from the oscillat-
ing walls. Annealing did not completely suppress the rotation anomaly.
Figure 1 | Quantum tunnelling of vacancies. In the original model of
supersolidity, the number of atoms is less than the number of lattice sites.
The lattice thus contains vacancies (red circles), which are able to exchange
their positions with neighbouring atoms (blue dots) so easily that they
are delocalized throughout the whole system. These vacancies are bosons,
which at low temperatures may form a Bose–Einstein condensate, that is,
a macroscopic wave that extends throughout the system. As a result, the
atoms themselves are not localized at particular sites as in a classical lattice.
Some of the mass can flow without friction through the rest, which remains
a rigid solid.
178
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
In any event, it became clear that the magnitude of the supersolid frac-
tion strongly depends on the amount of disorder.
The other type of experiment was pursued by Sasaki et al.31,32, who,
instead of investigating a.c. mass flow in torsional oscillator experiments,
decided to look for d.c. mass flow through a static, solid sample. To this
end, they built a two-part sample cell containing various types of 4He
crystals in equilibrium with liquid 4He. Using an optical cryostat, they
monitored the respective levels of the liquid–solid interface in the two
parts of the cell and measured their relaxation to the same height in the
cell. As any level change would necessitate mass flow through the solid
because of the different densities of liquid and solid 4He, such an equilibra-
tion would signal supersolid behaviour. The experiments indeed provided
evidence for superfluid d.c. mass flow, but only in polycrystals containing
grain boundaries. No such flow was observed in single crystals. This then
seemed to imply that grain boundaries are necessary for supersolidity.
However, as indicated below, the lowest temperature that the experi-
ments by Sasaki et al.31,32 reached, 50 mK, today seems not to have been
low enough to rule out the possibility of mass flow in single crystals.
Following the experiments by Sasaki et al., more evidence was found
that the few atomic layers forming grain boundaries might be super-
solid33; however, evidence was also found that some mass could flow
along liquid channels forming at the interface between grain boundaries
and the cell walls or at points where grain boundaries intersect with each
other34,35. In my opinion, the existence of d.c. mass flow inside single
crystals of 4He is not yet firmly established.
As the growth of solid samples at constant volume can lead to the
formation of polycrystals with a high density of grain boundaries, in light
of the above discussion the question became that of whether the observed
supersolid behaviour in torsional oscillator experiments could also be
attributed to mass flow along grain boundaries. To answer this question,
Chan’s group made new measurements30 using torsional oscillators in
which the cell was filled with single crystals of 4He. They found that the
period shift is smaller than in the corresponding experiments on poly-
crystals but is non-zero, indicating the presence of a supersolid fraction
of 0.04–0.4%. It is therefore possible that although grain boundaries could
be responsible for the supersolidity observed in polycrystals, supersolid-
ity could also be exhibited by single crystals, in which the origin of the
vacancy-induced mass flow more probably could be due to dislocations.
However, the interpretation of the data was to receive yet another twist.
Pinning at 3He impurities
The first experiments by Kim and Chan2 had already shown that the
phenomenon of supersolidity was very sensitive to the presence of tiny
amounts of 3He impurities. Helium-4 is extracted from oil wells where
it accumulates owing to the radioactive decay of uranium in the Earth.
In natural 4He, there are about 0.3 parts per million (3 × 10−7) of the
light isotope 3He. This may not seem much, but at low temperatures
the quantum statistical nature of particles (4He atoms being bosons
and 3He atoms being fermions) becomes important. This fundamental
difference in quantum behaviour makes 3He behave like a significant
impurity in a 4He system, even at such low concentrations.
The concentration of 3He in 4He can be controlled. In fact, it is possible
to prepare ultrapure 4He gas in which the concentration of 3He is only
one part per billion (10−9). Experiments have found that the supersolid
transition temperature depends on the presence of 3He, even at such
extremely low concentrations10,11,36. Because previous experiments had
shown that 3He impurities adsorb on dislocations at low temperatures37,38,
it was then suggested that the whole phenomenon could be attributed to
how the properties of defects such as grain boundaries and dislocations,
discussed above, are modified by the presence of 3He atoms.
Unexpectedly, measurements in 2007 by Day and Beamish10 and Day,
Syshchenko and Beamish11 found that the shear modulus, μ, increased
at the onset of supersolidity (Fig. 4) — when the mass started to flow,
they found that the solid 4He became stiffer. This was surprising, as the
naive expectation is that if part of the mass of a crystal starts flowing, the
quantum solid should somehow more closely resemble a liquid, such
that its shear modulus should decrease, contrary to what was observed.
NATURE|Vol 464|11 March 2010 REVIEW INSIGHT

a Figure 2 | First experimental evidence of

Base plate: thermal contact
to the refrigerator supersolidity in 4He. a, Schematic view of the
Magnesium barrier experimental set-up in a typical torsional oscillator
(control experiment)
experiment by Kim and Chan2,3. The oscillating cell
Torsion rod Filling line contains helium in an annular channel where the
(and filling line)
circulation can be blocked by inserting a vertical
magnesium barrier. The helium is introduced in the
Magnesium cell through a filling line that runs vertically inside
disk the torsion rod. Electrodes are used to drive and
Oscillating cell
Aluminium
detect the oscillation of the cell, and consequently to
shell measure its resonant period. This period is expected
Electrodes to decrease when solid helium becomes supersolid
because a fraction of the helium mass stops moving
with the oscillating walls of the box. This fraction
Solid helium is found to be 0.01–20%, depending on sample
Detection
in annular channel quality. When the magnesium barrier is inserted,
Drive the period change disappears. b, Left: one of the first
measurements by Kim and Chan2, showing that the
b period, τ, of their torsional oscillator decreases at
Empty-cell background temperatures, T, below about 0.1 K when filled with
920 Solid helium (31 Nm s–1) solid 4He, relative to when the empty cell is used.
0.015
4 Nm s–1 The reference period, τ*, is chosen as 971,000 ns
6 Nm s–1 for convenience. The maximum velocity of the
910 0.012 14 Nm s–1 cell rim in the experiment with 4He is as indicated.
33 Nm s–1
t – t* (ns)

Right: measurements, made by the same authors3,

0.009 117 Nm s–1 where the period change is converted into the non-
NCRIF

900 420 Nm s–1 classical rotational inertia fraction (NCRIF) of the

4
0.006 He sample. At low temperatures, the NCRIF is of
the order of 1%. Its value depends on the maximum
890 velocity of the cell rim (as indicated) when this
0.003
velocity is greater than 6 μm s−1. (Panel courtesy
of M. H. W. Chan, Pennsylvania State University,
880 0 University Park, Philadelphia.)
0.02 0.04 0.1 0.2 0.4 1 2 0.02 0.04 0.1 0.2 0.4
T (K) T (K)

These experiments further found that μ increased at temperatures below
the same critical temperature at which the rotational inertia began to
decrease in Kim and Chan’s2,3 experiments. In fact, the shape of the tem-
perature variation of μ looks exactly the same as that of the NCRI, and
even the dependence of the transition temperature on 3He impurity
content was found to be the same (Fig. 4).
Thus, these experiments10,11 showed that the anomalies observed in
both the rotational and elastic properties of solid 4He have a common
origin, which prompted Day and Beamish10 to interpret the stiffen-
ing as follows. In a quantum crystal, dislocations can be highly mobile
because they may have mobile kinks (Fig. 5). Kinks are regions where
the dislocation lines shift from one atomic row to the next. Quantum
fluctuations may allow kinks to tunnel from one lattice site to the next,
and so induce motion of the whole line. Consequently, in measure-
ments of the stiffness the response of the system to an applied stress is
the sum of the usual lattice deformation plus a contribution due to the
displacement of the dislocations. However, the mobility of dislocations
should depend on temperature and on the 3He concentration because
3
He impurities should bind to dislocations and pin them locally below a
certain temperature. Below this temperature, which could be of the order
of 0.1 K, the crystal should therefore be more rigid. This interpretation
looked likely to be correct, but it left open an important question, that
of the mechanism by which supersolidity is associated to the pinning
of dislocations.
Torsional oscillator experiments revisited
In view of the above developments, the very existence of supersolidity
was once again questioned. One proposal39,40 suggested that the anoma-
lies seen in torsional oscillator experiments could in fact be an arte-
fact — it was suggested that the period shift perhaps was not due to a
reduction of the inertia, I, in the relationship τ = 2π(I/K)1/2, but to an
increase in the elastic constant, K. Naturally, the contribution to K of
the helium stiffness would have been small in most experiments, but
the period shift itself was small. The respective groups of Chan and
© 2010 Macmillan Publishers Limited. All rights reserved
Beamish collaborated to check this possibility41. They built a very strong
cell in which the helium contribution to K was negligible. They found
a definite period shift that had to be the consequence of a decrease
in the rotational inertia41. Furthermore, they proved that the rotation
anomaly is indeed associated with the quantum statistics of the 4He;
they found a similar elastic anomaly in solid 3He, a fermionic system,
but did not observe a rotation anomaly there (see ref. 41 for details on
the importance of the crystal structure). Supersolidity thus seemed to
have been confirmed.
The existence of supersolidity had in fact already been strongly
supported by two other types of measurement. One type was the
so-called blocked annulus experiment. Kim and Chan3 and later Rit-
tner and Reppy42, with more accuracy, compared a torsional oscillator
with a free annular space containing solid helium with a similar one in
which the annulus was blocked along a diameter so as to interrupt the
macroscopic circulation of mass (Fig. 2). In the experiments with a free
annulus, a rotation anomaly signalling loss of inertia was found. In the
blocked case, however, such an anomaly was not observed. Even though
Reppy recently presented some puzzling results from measurements in
a more complicated cell geometry43, I believe that the rotation anomaly
observed in torsional oscillator experiments in general is due to a mac-
roscopic superflow. The other supporting experiments were done by
Lin and colleagues14,15. They found a peak in the specific heat of solid
helium at the temperature at which the rotation anomaly occurs. This
is a strong indication that the phenomenon is a true phase transition,
despite there being discussions about the magnitude of the peak relative
to that of the inertia change44.
At this point, I would like to mention a few interesting results related
to the response time and frequency dependence of the torsional oscil-
lator experiments. It has been found that as the frequency of the
oscillations is increased, the supersolid transition temperature increases
as well7. It also seems that the rotation velocity has a role similar to the
thermal agitation9. Furthermore, Aoki et al.7 and Penzev et al.8 have
shown that their oscillators are hysteretic at low temperatures: the
179
INSIGHT REVIEW
a period decreases when cooling below 50 mK, but a subsequent warm-
Edge dislocation
b vortices. Because such vortices should interact with the crystal lattice,
c To provide a better understanding of this set of observations, some
d
Figure 3 | Various defects in solids. a, In a single crystal, an edge dislocation is
the end line of an interrupted lattice plane. b, Owing to their mutual attraction,
dislocation lines usually organize themselves as three-dimensional networks.
For a detailed description, see the ParaDiS project (http://paradis.stanford.
edu). (Image courtesy of the ParaDiS project code team: A. Arsenlis, V. V.
Bulatov, W. Cai, R. D. Cook, M. Hiritani, M. Rhee and M. Tang, Lawrence
Livermore National Laboratory, Livermore, California.) c, In polycrystals, the
boundaries (green) between crystal grains of different orientations are regions
of atomic thickness where atoms are not ordered, owing to the competing
influences of the adjacent lattices. d, Photograph of a vertical grain boundary
between two helium crystals in equilibrium with liquid helium above. The cell
height and width are 11 mm. It is enclosed by two glass windows. Two crosses
are carved on the bottom right of the front window to help focus the camera.
The photograph shows that a grain boundary is not a liquid region32,34. Indeed,
as explained by Sasaki et al.32,34, the groove it makes where it meets the liquid
phase has a non-zero opening angle (of about 30°). As a consequence, grain
boundaries may be supersolid (that is, solid and partially superfluid) but not
entirely superfluid. (Image reproduced, with permission, from ref. 32.)
180
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
ing does not bring the period to its first measured values. This obser-
vation leads to many questions, especially about the true value of the
critical velocity measured by various groups (which ranges from a few
micrometres per second to at least 100 times more). Eventually, Hunt
et al.9 showed that their oscillator has a response time that diverges
near the transition temperature in a way that, as far as I know, has not
been understood either.
It is not known what mechanisms are behind this set of observations.
It could be the network of dislocations (or grain boundaries), which
relaxes with a T-dependent time, or it could be the diffusion of 3He
impurities, which becomes slow at low T. A totally different picture was
proposed by Anderson18–20. If mass really flows it may form vortices,
and Anderson invoked a ‘vortex liquid’, that is, a tangle of interacting
their proliferation near the supersolid transition temperature might lead
to an increase in the rigidity. Furthermore, the properties of a vortex
tangle should depend on frequency and temperature, but all these issues
are open problems and are difficult to solve, especially if the sample
disorder is unknown.
Towards a solution
Given the observations that I briefly describe above, I believe that even if
supersolidity remains mysterious, it has really been discovered. It appears
to be strongly dependent on disorder, and it seems that the supersolid
transition is accompanied by a surprising stiffening of the samples.
groups have tried to maximize the NCRI by preparing highly disordered
samples that in effect might be glassy. In my group, we have chosen an
opposite approach that seems promising. Having optical control of the
crystal growth34,45, we know how to prepare polycrystals, disordered
single crystals or very high-quality single crystals. The high-quality sin-
gle crystals have been shown to be free of screw dislocations if grown at
20 mK (ref. 46). Day and Beamish’s observations10,11 of stiffening were
done on polycrystals but not interpreted in terms of grain boundaries;
to clarify this, we have made new measurements of the elasticity of 4He,
this time in single crystals with variable concentration of 3He impuri-
ties13. Our most interesting result is on ultrapure single crystals (we have
discovered a simple method to eliminate all impurities47), in which we
also observed a stiffening13, this time below 40 mK. This temperature
is lower than has been observed in any other experiment. It is also the
temperature below which the search for d.c. mass flow through solid
4
He should be repeated. In view of these observations, the pinning of
dislocations by 3He impurities cannot be the only mechanism respon-
sible for the stiffening.
Interestingly, a new possible explanation for the observed stiffening in
the absence of 3He impurities was proposed by Kuklov and colleagues48
in 2009. They theoretically predicted that dislocations could change state
as a function of temperature. They calculated that at T = 0 K, a disloca-
tion is a straight line that is anchored by the lattice and does not fluctuate.
However, at a certain finite temperature, a transition occurs to a rough
fluctuating state in which the dislocation line becomes mobile because
it is invaded by kinks (Fig. 5). This ‘roughening transition’ could exist
in the absence of 3He and thus provide an explanation for the change in
stiffness observed in experiments. The sensitivity to the presence of 3He
impurities is a consequence of the fact that kinks become dressed with
3
He atoms that attach to them (Fig. 5), such that a higher temperature is
needed for their creation.
To verify these new ideas experimentally, it is necessary to see whether
an ideal crystal free of impurities and dislocations shows stiffening at low
temperatures. This experiment is in progress in our group, but it is not
easy to observe a single dislocation. In addition, these are very delicate
experiments because helium crystals are extremely fragile, probably
owing to their quantum nature. If we succeed in showing that perfect,
ultrapure crystals show no stiffening, the theory of Kuklov and col-
leagues48 will be strongly supported. If the crystals do show stiffening,
it will be necessary to search for another origin of the stiffening, which
NATURE|Vol 464|11 March 2010 REVIEW INSIGHT

a Figure 4 | Stiffness measurements and the role of 3He

1 p.p.b. 3He
1 85 p.p.b. 3He impurities. a, The original measurements by Day and Beamish10
0.3 p.p.m. 3He of the variation of the stiffness of 4He crystals as a function of 3He
0.8 impurities. For comparison, measurements by Kim and Chan3 of
0.3 p.p.m. 3He
% m/% m(18 mK)

(NCRI) the NCRI are shown on the same graph (see key). Both the stiffness
0.6 1 p.p.b. 3He and the NCRI are expressed as variation in shear modulus, μ, at
(NCRI)
0.4 temperature T relative to the variation at 18 mK. p.p.b., parts per 109;

p.p.m., parts per million. (Panel reproduced, with permission, from

0.2 ref. 10.) b, Further measurements have shown an even more striking
similarity between the temperature dependences of the shear
0 modulus and the NCRI: after adjustment of the vertical scales, the
0.02 0.03 0.04 0.06 0.1 0.2 0.3 0.4 two sets of data shown coincide (μo is the low-temperature value of
T (K) the stiffness). This is very strong evidence that the rotation anomaly
and the elastic anomaly have a common origin. (Panel courtesy of
b J. D. Beamish, University of Alberta, Edmonton, Canada.)
1.00 Shear modulus at 200 Hz 1.5
NCRI at 910 Hz

0.98
1.0

NCRI (%)
m/mo

0.96

0.5

0.94

0
0.92

0 0.1 0.2 0.3

T (K)

could perhaps be explained by Anderson’s vortex liquid picture18–20.
Naturally, we will also have to see whether the same ideal crystals show
a rotation anomaly. For this, we need a transparent torsional oscillator
to be able to monitor the growth and measure the crystal orientation,
which is an important parameter that should be explored. This experi-
ment is also in progress in our group, now in collaboration with J. West
and M. H. W. Chan. If the rotation anomaly disappears in ideal crystals,
then the dislocation network theory will be strongly supported.
We also have a suggestion for the connection between rigidity and
supersolidity: transverse fluctuations of dislocation lines should induce
mass currents, which should manifest as phase fluctuations because,
in a superfluid, a mass current is the gradient of the phase of the wave
function. When dislocations start fluctuating, phase fluctuations could
destroy the supersolidity. Conversely, if the dislocation is supersolid,
the phase coherence should force the dislocation to be a smooth, non-
fluctuating, straight line. If the rotation anomaly disappears in perfect
crystals, this theory will receive strong support. Otherwise, the favoured
theory will be Anderson’s, where supersolidity is an intrinsic crystal
property that is only enhanced by disorder, and the elastic anomaly is
due to the proliferation of vortices close to the supersolid transition.
In conclusion, I believe that supersolidity is a real phenomenon but
that it is not yet understood. The true origin of mass flow inside 4He
crystals is unclear, as are the origin of their stiffening at low temperatures
and the connection between the mechanical response of these quantum

Crystal lattice
Crysta

3He impurity

Kinks

Dislocation lines

Figure 5 | Dislocation lines. At the atomic scale and at a sufficiently high
temperature, dislocation lines show kinks where they shift from one atomic
row to the next. In a quantum crystal such as 4He, these kinks may move
very easily by quantum tunnelling from one site to the neighbouring one. In
this case, the dislocation line itself is a mobile line whose presence reduces
the crystal rigidity. However, Aleinikava et al.48 have proposed that below
a certain roughening transition temperature, kinks disappear and the
dislocation line is straight and immobile, such that the crystal rigidity is
greater than it is above the roughening temperature. This roughening process
may be at the origin of the stiffness anomaly of 4He crystals. The roughening
temperature should depend on the presence of 3He impurities, which are
known to bind to the dislocation lines, where they probably dress kinks.
Even if this set of hypotheses is correct, the coupling between stiffness and
supersolidity remains to be clarified.

181
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
crystals and the onset of supersolidity. However, I expect experimental
results in the coming year that should help to discriminate between the
two main types of interpretation that have been proposed. The most fun-
damental question remaining to be answered by experiments is whether
disorder is necessary for supersolidity to appear in a quantum crystal
such as solid 4He. ■ Rev. Lett. 99, 205302 (2007).
1. Leggett, A. J. A ‘superglass’ state in solid 4-He. J. Club Condens. Matter Phys. 〈http://
www.condmatjournalclub.org/wp-content/uploads/2009/04/jccm_april09_032.pdf〉
(2009).
2. Kim, E. & Chan, M. H. W. Probable observation of a supersolid helium phase. Nature 427,
225–227 (2004).
This paper reported the first observation of anomalous rotation properties in solid 4He.
3. Kim, E. & Chan, M. H. W. Observation of superflow in solid helium. Science 305, 1941–
1944 (2004).
4. Rittner, A. S. & Reppy, J. Observation of classical rotational inertia and nonclassical
supersolid signals in solid 4He below 250mK. Phys. Rev. Lett. 97, 165301 (2006).
5. Rittner, A. S. & Reppy, J. Disorder and the supersolid state of solid 4He. Phys. Rev. Lett. 98,
175302 (2007).
6. Kondo, M., Takada, S., Shibayama, Y. & Shirahama, K. Observation of non-classical
rotational inertia in bulk solid 4He. J. Low Temp. Phys. 148, 695–699 (2007).
7. Aoki, Y., Graves, J. C. & Kojima, H. Oscillation frequency dependence of non-classical
rotation inertia of solid 4He. Phys. Rev. Lett. 99, 015301 (2007).
8. Penzev, A. Yasuta, Y. & Kubota, M. AC vortex-dependent torsional oscillation response
and onset temperature T0 in solid 4He. Phys. Rev. Lett. 101, 065301 (2008).
9. Hunt, B. et al. Evidence for a superglass state in solid 4He. Science 324, 632–636 (2009).
10. Day, J. & Beamish, J. Low-temperature shear modulus changes in solid 4He and
connection to supersolidity. Nature 450, 853–856 (2007).
The paper reported the first observation of a stiffening linked to the appearance of
supersolidity.
11. Day, J., Syshchenko, O. & Beamish, J. D. Intrinsic and dislocation-induced elastic behavior
of solid helium. Phys. Rev. B 79, 214524 (2009).
12. Mukharsky, Yu., Penzev, A. & Varoquaux, E. Low-frequency acoustics in solid 4He at low
temperature. Phys. Rev. B 80, 140504 (2009).
13. Rojas, X., Pantalei, C., Maris, H. J. & Balibar, S. Acoustic properties of solid 4He in the limit
of zero impurity. J. Low Temp. Phys. 158, 478–484 (2009).
14. Lin, X., Clark, A. C. & Chan, M. H. W. Heat capacity signature of the supersolid transition.
Nature 449, 1025–1028 (2007).
15. Lin, X., Clark, A. C., Cheng, Z. G. & Chan, M. H. W. Heat capacity peak in solid 4He: effects
of disorder and 3He impurities. Phys. Rev. Lett. 102, 125302 (2009).
This paper presents the most recent measurements of a heat capacity peak signalling
that the supersolid transition is a true phase transition.
16. Prokof’ev, N. What makes a crystal supersolid? Adv. Phys. 56, 381–402 (2007).
This is a recent review on the theory of supersolidity, including results from numerical
simulations.
17. Ceperley, D. M. & Bernu, B. Ring exchanges and the supersolid phase of 4He. Phys. Rev.
Lett. 93, 155303 (2004).
18. Anderson, P. W. Two new vortex liquids. Nature Phys. 3, 160–162 (2007).
19. Anderson, P. W. Bose fluids above Tc: incompressible vortex fluids and ‘supersolidity’.
Phys. Rev. Lett. 100, 215301 (2008).
In this paper, Anderson argues that supersolidity is an intrinsic property of quantum
crystals that exists even in the absence of disorder.
20. Anderson, P. W. A. Gross–Pitaevskii treatment for supersolid helium. Science 324, 631–
632 (2009).
21. Balibar, S. & Caupin, F. Supersolidity and disorder. J. Phys. Condens. Matter 20, 173201
(2008).
This is a recent review on the whole field of supersolidity, with emphasis on experimental
results.
22. Galli, D. E. & Reatto, L. Solid 4He and the supersolid phase: from theoretical speculation
to the discovery of a new state of matter? A review of the past and present status of
research. J. Phys. Soc. Jpn 77, 111010 (2008).
23. Thouless, D. J. The flow of a dense superfluid. Ann. Phys. 52, 403–427 (1969).
24. Andreev, A. F. & Lifshitz, I. M. Quantum theory of defects in crystals. Sov. Phys. JETP 29,
1107–1113 (1969).
25. Leggett, A. J. Can a solid be superfluid? Phys. Rev. Lett. 25, 1543–1546 (1970).
26. Meisel, F. Supersolid 4He: an overview of past searches and future possibilities. Physica B
178, 121–128 (1992).
27. Boninsegni, M. et al. Fate of vacancy-induced supersolidity in 4He. Phys. Rev. Lett. 97,
080401 (2006).
28. Boninsegni, M. et al. Luttinger liquid in the core of screw dislocation in helium-4.Phys. Rev.
Lett. 99, 035301 (2007).
182
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
29. Pollet, L. et al. Superfluidity of grain boundaries in solid 4He. Phys. Rev. Lett. 98, 135301
(2007).
30. Clark, A. C., West, J. T. & Chan, M. H. W. Nonclassical rotational inertia in helium crystals.
Phys. Rev. Lett. 99, 135302 (2007).
31. Sasaki, S., Ishiguro, R., Caupin, F., Maris, H. J. & Balibar, S. Superfluidity of grain boundaries
and supersolid behaviour. Science 313, 1098–1100 (2006).
32. Sasaki, S., Caupin, F. & Balibar, S. Wetting properties of grain boundaries in solid 4He. Phys.
33. Ray, M. W. & Hallock, R. B. Observation of unusual mass transport in solid hcp 4He. Phys.
Rev. Lett. 100, 235301 (2008).
34. Sasaki, S., Caupin, F. & Balibar, S. Optical observations of disorder in solid helium-4.J. Low
Temp. Phys. 153, 43–76 (2008).
35. Balibar, S. & Caupin, F. Comment on “Observation of unusual mass transport in solid hcp
4
He”. Phys. Rev. Lett. 101, 189601 (2008).
36. Kim, E. et al. Effect of 3He impurities on the nonclassical response to oscillation of solid 4He.
Phys. Rev. Lett. 100, 065301 (2008).
37. Iwasa, I. & Suzuki, H. Sound velocity and attenuation in hcp 4He crystals containing 3He
impurities. J. Phys. Soc. Jpn 49, 1722–1730 (1980).
38. Paalanen, M. A., Bishop, D. J. & Dail, H. W. Dislocation motion in hcp 4He. Phys. Rev. Lett.
46, 664–667 (1981).
39. Nussinov, Z., Balatsky, A. V., Graf, M. J. & Trugman, S. A. Origin of the decrease in the
torsional-oscillator period of solid 4He. Phys. Rev. B 76, 014530 (2007).
40. Yoo, C. D. & Dorsey, A. T. Theory of viscoelastic behavior of solid 4He. Phys. Rev. B 79,
100504 (2009).
41. West, J. T., Syshchenko, O., Beamish, J. D. & Chan, M. H. W. Role of shear modulus and
statistics in the supersolidity of helium. Nature Phys. 5, 598–601 (2009).
42. Rittner, A. S. C. & Reppy, J. D. Probing the upper limit of nonclassical rotational inertia in
solid helium-4.Phys. Rev. Lett. 101, 155301 (2008).
43. Reppy, J. D. Is supersolid superfluid? Supersolids Banff 2009 〈http://www.phys.ualberta.ca/
supersolids/talks/〉 (2009).
44. Balatsky, A. V., Graf, M. J., Nussinov, Z. & Trugman, S. A. Entropy of solid 4He: the possible
role of a dislocation-induced glass. Phys. Rev. B 75, 094201 (2007).
45. Balibar, S., Alles, H. & Parshin, A. Y. The surface of helium crystals. Rev. Mod. Phys. 77,
317–370 (2005).
46. Ruutu, J. P. et al. Facet growth of helium-4 crystals at mK temperatures. Phys. Rev. Lett. 76,
4187–4190 (1996).
47. Pantalei, C., Rojas, X., Edwards, D. O., Maris, H. J. & Balibar, S. How to prepare an ideal
helium-4 crystal. J. Low Temp. Phys. doi: 10.1007/s10909-010-0159-6 (in the press).
48. Aleinikava, D., Dedits, E., Kuklov, A. B. & Schmeltzer, D. Mechanical and superfluid
properties of dislocations in solid 4He. Phys. Rev. B (submitted); preprint at
〈http://arxiv.org/abs/0812.0983〉 (2008).
49. Allen, J. F. & Misener, A. D. Flow of liquid helium II. Nature 141, 75 (1938).
50. Kapitza, P. Viscosity of liquid helium below the λ point. Nature 141, 74 (1938).
51. Balibar, S. The discovery of superfluidity. J. Low Temp. Phys. 146, 441–470 (2007).
52. Allen, J. F. & Jones, H. New phenomena connected with heat flow in helium II. Nature 141,
243–244 (1938).
53. London, F. The λ phenomenon of liquid helium and the Bose–Einstein degeneracy. Nature
141, 643–644 (1938).
54. Tisza, L. Transport phenomena in helium II. Nature 141, 913 (1938).
55. Penrose, O. On the quantum mechanics of helium II. Phil. Mag. 42, 1373–1377 (1951).
56. Kamerlingh Onnes, H. The resistance of pure mercury at helium temperatures. Commun.
Phys. Lab. Univ. Leiden 120b (1911); The disappearance of the resistivity of mercury.
Commun. Phys. Lab. Univ. Leiden 122b (1911); On the sudden change in the rate at which the
resistance of mercury disappears. Commun. Phys. Lab. Univ. Leiden 124c (1911).
57. Bardeen, J., Cooper, L. N. & Schrieffer, J. R. Theory of superconductivity. Phys. Rev. 106,
162–164 (1957).
58. Bardeen, J., Cooper, L. N. & Schrieffer, J. R. Microscopic theory of superconductivity. Phys.
Rev. 108, 1175–1204 (1957).
59. Osheroff, D. D., Richardson, R. C. & Lee, D. M. Evidence for a new phase in solid 3He. Phys.
Rev. Lett. 28, 885–888 (1972).
60. Osheroff, D. D., Gully, W. J. & Richardson, R. C. New magnetic phenomena in liquid 3He
below 3mK. Phys. Rev. Lett. 29, 920–923 (1972).
61. Anderson, M. H., Ensher, J. R., Matthews, M. R., Wieman, C. E. & Cornell, E. Observation of
Bose–Einstein condensation in a dilute atomic vapour. Science 269, 198–201 (1995).
62. Davis, K. B. et al. Bose–Einstein condensation in a gas of sodium atoms. Phys. Rev. Lett. 75,
3969–3973 (1995).
Acknowledgements I acknowledge support from the Agence Nationale de la
Recherche (grant BLAN07-1-215296).
Author Information Reprints and permissions information is available at www.
nature.com/reprints. The author declares no competing financial interests.
Correspondence should be addressed to the author (balibar@lps.ens.fr).
NATURE|Vol 464|11 March 2010|doi:10.1038/nature08914 PERSPECTIVE INSIGHT

Superconductivity gets an iron boost

Igor I. Mazin1

Superconductivity, the resistance-free flow of electrical charges, is one of the most exotic phenomena in
solid-state physics. Even though it was discovered almost a century ago, many questions remain unanswered,
in particular those concerning the physics of high-temperature superconductivity. The recent discovery of
iron-based superconductors was arguably the most important breakthrough in this field for more than two
decades and may provide new avenues for understanding this high-temperature phenomenon. Here I present
my view of the recent developments in this field that have led to the current understanding of this important
new class of superconductor.

A once popular joke asked how a physicist would interpret experimental
data on odd numbers. As the first experiments reveal that 1, 3, 5 and
7 are all prime numbers, the physicist becomes convinced that all odd
numbers are prime and that a correct theory of ‘primeness’ should be
able to explain this experimental fact. Further studies, however, show
that 9 is not a prime number. The community initially disregards this as
an experimental error; however, after more experiments, the research-
ers are forced to admit that 9 is indeed not prime, making it a unique
case. This view is reinforced when further experiments show that the
next odd numbers in the series, 11 and 13, are both prime. Only after
it is found that, in violation of ‘conventional wisdom’, 15 is not a prime
number, does the idea that there are infinitely many odd numbers, but
not prime numbers, take root in researchers’ minds.
This joke can be seen as an allegory of the modern history of
superconductivity. During the 1960s and 1970s, the recorded transi-
tion temperature, Tc, for superconductors very slowly inched its way
up, culminating in the 1976 finding of Tc = 23.2 K in Nb3Ge, largely
thanks to the efforts of Berndt Matthias. At about that time, according
to physicists’ folklore, Matthias formulated his famous six rules for a
successful search for new superconductors1. One, a high symmetry is
good; cubic symmetry is the best. Two, a high density of electronic states
is good. Three, stay away from oxygen. Four, stay away from magnet-
ism. Five, stay away from insulators. Six, stay away from theorists. All of
these rules, with the possible exception of rule six (I’ll leave that to the
judgement of my experimental colleagues), not only have been proved
incorrect but also their exact opposite seems to be true.
Following this stagnation in the search for superconductors with a
higher Tc than 23–24 K, between 1976 and 1986 an increasingly large
number of physicists became convinced, and openly sounded their
conviction (a notable exception being Vitaly Ginzburg2), that a quantitative
theory of superconductivity ultimately would provide proof that Tc has a
fundamental limit of about 25–30 K. It is ironic that in the chemical cabi-
nets of many of these distinguished scientists was magnesium diboride
(MgB2), which, as has been known since 2001, has a Tc of 40 K.
This conviction was radically challenged when copper-oxide-based
superconductors with a high Tc were discovered in 1986. The rapid climb
of the maximum known Tc up to 140 K created the first paradigm shift;
it became clear that high-Tc superconductivity (with critical tempera-
tures larger than the anticipated 25–30 K) was possible, even though the
underlying mechanism remained unknown. The following two decades,
however, brought little progress towards a further increase of Tc. No
new high-Tc materials were found, although some superconductors that
broke the old record of 23.2 K (but not by much) were discovered, for
example (Ba,K)BiO3, doped fullerenes and MgB2. Inevitably, a grow-
ing number of physicists started to look for the unique combination of
physical properties that makes copper oxides such a spectacular excep-
tion in the materials universe.
The new, iron-based, superconductors (initially called iron-pnictide
superconductors, before another, chalcogen-based subfamily was found)
were discovered in 2008 (ref. 3). Their epistemological value is that they
did not fit the same mould as copper oxides. Their discovery demonstrated
that unconventional (that is, not phonon-mediated) high-Tc superconduc-
tivity, such as that found in copper oxides, is not unique and is probably
just as ubiquitous as the conventional, ‘low-Tc’, kind, if researchers would
look in the right places. What is special about iron-based superconduc-
tors is that the more they are studied, the less they look like copper oxides,
suggesting that high-Tc superconductivity is not limited to copper oxides
and might not even be limited to any particular class of materials. (This

Table 1 | Properties of different classes of superconductor

Property Conventional superconductors Copper oxides MgB2 Iron-based superconductors
Tc (maximum) <30K 134K 39K 56K
Correlation effects None (nearly-free electrons) Strong local electronic interaction None (nearly-free electrons) Long-range (non-local)
magnetic correlations
Relationship to magnetism No magnetism Parent compounds are magnetic No magnetism Parent compounds are
insulators magnetic metals
Order parameter One band, same-sign s wave One band, sign-changing d wave Two band, same-sign s wave Two band, presumably sign-
changing s wave
Pairing interaction Electron–phonon Probably magnetic (no consensus) Electron–phonon Presumably magnetic
Dimensionality Three dimensional Two dimensional Three dimensional Variable

1
Naval Research Laboratory, code 6390, 4555 Overlook Avenue Southwest, Washington DC 20375, USA.

183
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT PERSPECTIVE
may not be entirely true; as discussed below, there might be some common
conditions for high-Tc superconductivity, such as proximity to magnetism.
But if these conditions exist, they are of a general character.)
In this Perspective article, I discuss the commonalities and differences
between the two previously known classes of high-Tc superconductors,
copper oxides and MgB2, and what can be learned from this comparison.
I then take a look at how far the understanding of iron-based supercon-
ductors has advanced and what the remaining challenges and prospects
are in this field.
True siblings or lookalikes
Eighteen months after the discovery of superconductivity in copper
oxides, there was almost no understanding of the underlying physics.
Eighteen months after superconductivity was found in MgB2, there was
almost a complete theory for it4. Some twenty months into the ‘iron
age’, and the physics community’s understanding lies between these
limits: much is known, and there is a plausible theoretical hypothesis
that the majority of researchers in the field seem to agree on (curiously,
proposed within days of the original discovery5,6). However, there is not
a complete theory, and there are several facts that are difficult to explain
and to reconcile with one another7.
In many respects, iron-based superconductors sit between copper
oxides and MgB2 (Table 1). Undoped (‘parent’) copper oxides are strong
magnets and insulators. This is because two electrons located on the
same copper ion are subject to strong Coulomb repulsion, leading to
strong correlations and hence electron localization8. The undoped
compounds have exactly one valence electron for each copper atom,
which makes them strong magnets — each electron has a spin of ½ and
therefore a magnetic moment. When extra charge carriers are intro-
duced by way of doping, these carriers are free to move about and screen
the Coulomb repulsion. Given enough doping, the static magnetism
disappears, and the electron states form a single band, as is the case in
simple metals. Crucially, there is a critical range of doping in which a
superconducting phase appears within a characteristic ‘superconduct-
ing dome’ (Fig. 1b). Many, but not all, researchers think that the ‘glue’
that binds electrons into Cooper pairs, which give rise to superconduc-
tivity, is provided by the exchange of magnetic fluctuations. Another
important characteristic of copper oxides is that the wavefunction of the
Cooper pairs — also known as the superconducting order parameter
— has d-wave symmetry (Fig. 2b). This means that the paired electrons
a b
Paramagnetic
Super-
Su
conducting
cond
Pseudo-
gap
Hole
Figure 1 | Iron-based superconductors and copper-oxide-based
superconductors. a, Generic crystal structure of iron-based
superconductors. Fe atoms are shown in red, and pnictogens (As or P) or
chalcogens (Se or Te) are indicated in blue. The filler layer, shown without
atomic detail, can contain any of the following constituents (for each two
iron atoms): one alkaline earth atom; two alkali atoms; two rare-earth
atoms and two oxygen atoms; or more complex layers with various atoms.
It is also possible for there to be no filler layer. b, Schematic phase diagram
of copper-oxide-based superconductors. Note how both electron doping
and hole doping suppress the magnetism of the parent compounds and
184
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
orbit each other with a particular angular momentum, avoiding close
contact with each other, and thereby reducing the effect of their mutual
Coulomb repulsion. Many think that such ‘Coulomb avoidance’ is an
important component of high-Tc superconductivity8.
In contrast to copper oxides, MgB2 has no trace of magnetism, has
delocalized electrons (which are not subject to strong on-site correla-
tions), and has a complex electronic structure, featuring two distinc-
tively different groups of electrons. These two groups form two kinds
of electronic band (hence the term two-band superconductors), which
give rise to two separate sets of Fermi surfaces4.
Iron-based superconductors, on the one hand, resemble copper oxides
in that they are strong magnets in parts of their phase diagram (Fig. 1c)
and in that superconductivity develops when magnetism is destroyed
by doping. On the other hand, unlike copper oxides, the parent (mag-
netic) iron-based compounds are metallic, so the increase in the charge
carrier concentration as a result of doping is not a decisive factor in the
development of superconductivity. Indeed, magnetism can be suppressed
and superconductivity induced in these materials without charge dop-
ing. This can be achieved by applying pressure, by partially substituting
arsenic with phosphorus, or by simply diluting the iron layer with a non-
magnetic species, such as ruthenium. With few exceptions, one of the
main characteristic features of iron-based superconductors is that when
magnetism is destroyed (by any means), superconductivity emerges9.
Another property that sets iron-based superconductors apart from cop-
per oxides is the strength of the Coulomb correlations. The latest calcula-
tions, as well as estimates from spectroscopic experiments, seem to be in
agreement that the correlation strength is weak to moderate10,11, although
not as weak as in MgB2, in which Coulomb correlations are almost absent.
Another important property, which iron-based superconductors have in
common with MgB2, is that electrons form a multi-sheet Fermi surface
that can be separated into two distinct sets of surfaces (Fig. 3). But, as dis-
cussed below, there seem to be significant differences between these two
classes of materials regarding the nature of the superconducting states.
Ironing out the last wrinkles
So what is known about these materials, and how has this information
been uncovered? It is known that iron-based superconductors
comprise a broad variety of materials9 and that they all have similar
phase diagrams (Fig. 1c) and presumably have the same mechanism
of superconductivity. All iron-based superconductors have the same
Temperature Temperature
c
Paramagnetic
Phase separation
Coexistence
Super-
Antiferro- conducting Antiferro-
magnetic magnetic
Doping Electron Hole Doping Electron
fraction fraction
induce superconductivity under the characteristic superconducting dome.
The green lines between the paramagnetic and pseuodogap phases (green
shading) represent crossover transitions, black lines between paramagnetic
and antiferromagnetic phases are well-defined transitions. c, Schematic
phase diagram for the 122 family of iron-based superconductors. An
example is BaFe2As2, in which Ba atoms can be substituted with K atoms,
thereby doping the parent compound with holes; Fe atoms can be substituted
with Co atoms, doping the parent compound with electrons; or Fe atoms
can be substituted with Ru atoms (or As atoms by P atoms), to suppress
magnetism without changing the carrier concentration.
NATURE|Vol 464|11 March 2010
a b
c d
Figure 2 | Superconducting order parameter. A schematic representation
of the superconducting order parameter in different cases: a conventional,
uniform, s wave, such as in an ‘old-fashioned’ superconductor (for example
aluminium) (a); a d wave, as is the case in copper oxides (b); a two-band
s wave with the same sign, as in MgB2 (c); an s± wave, as is thought to be the
case in iron-based superconductors (d). In a and b, the two-dimensional
Fermi surface is approximated by one circle. In c and d, the Fermi surface
is approximated by a small circle in the centre (the first band) surrounded
by four larger circles (to comply with the tetragonal symmetry; the second
band). In all cases, the height of the ‘rubber sheet’ is proportional to the
magnitude of the order parameter (including its sign).
crystallographic motif (Fig. 1a), with the main component being a
square lattice of iron atoms sandwiched between two square lattices
of pnictogen (arsenic or phosphorus) — hence the initial name — or
chalcogen (selenium or tellurium) atoms. Between these crucial trilay-
ers, various ‘filler slabs’ can be placed (although this is not essential): for
example, a single crystallographic layer of sodium, barium, strontium
or calcium; a trilayer consisting of a layer of oxygen between layers of
a rare-earth element (the highest Tc so far, 56 K, has been observed in
this family); or an even more complex filler slab.
Superconductivity can be induced in all materials by chemical doping
or pressure, or a combination of both, as long as this results in the sup-
pression of magnetism. In contrast to copper oxides, which have very
low electrical conductivity along the direction perpendicular to the
copper-oxide layers, none of these systems is truly two dimensional.
This is beneficial in terms of practical applications, because in polycrys-
talline two-dimensional materials superconductivity can be destroyed
by a relatively small current.
What can be inferred from the large number of iron-based
superconductors and the known properties of these materials? So far,
no simple correlation has been noticed between anisotropy, the distance
between planes in the crystal lattice, the temperature at which antifer-
romagnetic ordering occurs in the parent phase (or even the pattern of
the magnetic order of this phase) and the superconducting Tc. Initially, it
seemed as though Tc is optimized when the four anions around an iron ion
form an ideal tetrahedron12. However, after more iron-based supercon-
ductors had been uncovered, this finding seemed not to be universal13.
By contrast, phase diagrams of essentially all iron-based
superconductors have both superconducting phases and a strongly anti-
ferromagnetic phase (see Fig. 1, which is a schematic view of the generic
features of the phase diagram). All iron-based superconductors contain
iron in a valence state that is close or equal to Fe2+. All materials that have
been studied so far show a peak in inelastic neutron-scattering spectra
that corresponds to magnetic excitations at a particular wavevector, Qm
(even though for at least one family, FeTexSe1−x, the static magnetic order
occurs at a different wavevector). In cases in which the Fermi surface has
been mapped by angle-resolved photoemission spectroscopy (ARPES),
two sets of Fermi surfaces, roughly separated by the same wavevector,
Qm, have been revealed (Fig. 3).
© 2010 Macmillan Publishers Limited. All rights reserved
PERSPECTIVE INSIGHT
It is tempting to assume (and in fact almost the entire community has
succumbed to this temptation) that the features that such disparate iron-
based superconductors have in common reflect a common origin for the
observed superconductivity. Adopting this path, it can be concluded that
proximity to a magnetic quantum critical point (as is seen in the phase
diagram) signals that magnetic (spin) fluctuations play an important
role. The fact that neutron-scattering measurements always uncover
magnetic excitations with a particular wavevector, Qm, also suggests
that these excitations are instrumental for mediating the pairing of elec-
trons. Note that the two electrons in a singlet Cooper pair have the same
charge but opposite spins. A corollary of this is that magnetic excitations
lead to pairing only if the corresponding wavevector spans parts of the
Fermi surface with order parameters (that is, the pair wavefunction) of
opposite sign (see ref. 14 for further explanation). Now, noting that there
are two sets of Fermi surfaces that are roughly separated by the same
wavevector Qm (Fig. 3), the so-called s± superconductivity is derived, in
which the sign of the order parameter is switched between the two sets
of Fermi surfaces (Fig. 2).
In the previous paragraph, I describe how the s± superconductivity
model could have been arrived at, by using data from ARPES and neutron-
scattering experiments that became available roughly one year after the
initial discovery. It is gratifying that theorists were able to come up with
this model within a few weeks of the initial discovery, solely on the basis
of electronic structure calculations and theoretical models5,6.
Farther down the road
It is still not clear beyond a reasonable doubt that the superconducting
symmetry realized in iron-based superconductors is s± symmetry and
that pairing is due to spin fluctuations. The jury is still out. This is
a jury that is deeply convinced by the prosecution but is reluctant to
base its verdict solely on circumstantial evidence. However, circum-
stantial evidence is plenty in this case, and physicists might be en route
to uncovering a direct proof.
It is known (from nuclear-magnetic-resonance spectroscopy data)
that the Cooper pairs in iron-based superconductors are spin singlets
(formed by electrons with antiparallel spins). In this class, three possible
symmetries of the order parameter are compatible with the tetragonal
Electron momentum (ky)
Electron momentum (kx)
Figure 3 | A typical calculated Fermi surface of an iron-pnictide
superconductor. The Fermi surface (projected onto the kx–ky plane, where
k is electron momentum) shown is calculated for 10% electron-doped
LaFeAsO. Experimentally observed Fermi surfaces show similar geometries.
The momentum connecting the two sets of Fermi surfaces, Qm, is shown by
the arrow. Spin fluctuations with this moment were predicted theoretically
and found experimentally, and they are now thought to be instrumental for
creating high-Tc superconductivity in iron-based superconductors.
185
INSIGHT PERSPECTIVE
symmetry of iron-based superconductor crystals: a d-wave symmetry;
an s-wave symmetry without sign change; or the s±-symmetry model,
which combines s-wave symmetry with a sign change of the order
parameter. All of these models are illustrated in Fig. 2.
As can be seen in Fig. 2, an s-wave order parameter (of either flavour)
and a d-wave order parameter differ by their rotational symmetry: the
s wave does not change under a 90° rotation, whereas the d wave changes
sign. This difference allows the design of ‘phase-sensitive’ experiments15
that can distinguish between the two. The preliminary indications are
strongly against iron-based superconductors having d-wave symmetry.
However, arguably the most convincing evidence against the d-wave
model (or any finite angular momentum model) is that such a state,
according to its symmetry, has nodes in the order parameter, which
implies that there is a zero minimal excitation gap in the superconduct-
ing state. In other words, excitations with arbitrarily small energy exist.
Because this is mandated by the symmetry, such zero minimal excitation
gaps should be present in all iron-based superconductors, but numerous
experiments have established that some iron-pnictide superconductors
have a finite minimal gap. By contrast, an s state (whether s or s±) may
or may not have nodes (depending on the particular combination of
material-dependent parameters). Indeed, some of iron-pnictide super-
conductors (KFe2As2, LaFePO and BaFe2As2−xPx) show clear indications
of nodes, whereas others are fully gapped.
Taken together, the above considerations lead to the idea that the
overall symmetry of the superconducting state in iron-based super-
conductors is s-wave symmetry. It still remains to be experimentally
established whether this is a conventional s-wave state (such as is in
MgB2), which in some cases becomes disrupted (by magnetic or Cou-
lomb interactions) and develops nodes, or an s± state, which also may
exist with or without nodes, depending on the material. At present, the
preponderance of experimental evidence7,9 favours an s± state, which is
also favoured theoretically7.
Reproducibly confirming the proposed s± symmetry remains the main
experimental challenge. Phase-sensitive experiments15 have provided
such confirmation for the superconductivity in copper oxides, which
has a d-wave symmetry. Given that the s and the s± state belong to the
same rotational symmetry group, it is far more challenging to devise
a phase-sensitive experiment to distinguish between them. But, hav-
ing excluded d-wave symmetry, researchers need to prove only that the
order parameter in iron-based superconductors changes sign (while not
becoming zero anywhere on the Fermi surface). Several experimental
designs along these lines have been proposed, and several groups are
working to implement these. Within a year or two, the final experimen-
tal answer is likely to have been revealed.
So what comes next? Can we invent a new set of rules, Matthias’s rules
for the twenty-first century? It might be preposterous to think about a
general strategy for finding new high-Tc superconductors, given that
so far only two such families have been found. But this might be the
time to start thinking about it. There are good reasons to believe that
proximity to a magnetic transition is essential: magnetism is basically
the only quantum coherence phenomenon that commonly occurs at
room temperature. The magnetic exchange interaction is basically the
only interaction that is strong (on the order of the Fermi energy) and
yet does not conflict with electronic itineracy. Proximity to a magnetic
186
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
quantum critical point ensures the abundance of spin fluctuations that
may provide the necessary glue. It also seems important to have a Fermi
surface topology that matches the structure of magnetic excitations.
A layered crystal structure may be another not-so-accidental com-
ponent. One great advantage of a quasi-two-dimensional system is that
the density of electronic states that are available for superconductiv-
ity depends very weakly on the carrier concentration. Although it is
unclear why a low carrier concentration is beneficial, the possibility
of changing this parameter provides additional flexibility that is not
present in three-dimensional systems. In addition, low-dimensional
systems generally fluctuate more, which may also be helpful for super-
conducting pairing.
The following is my attempt to concoct a set of rules that could replace
Matthias’s rules. One, layered structures are good. Two, the carrier den-
sity should not be too high (compared with, say, conventional metals).
Three, transition metals of the fourth period (vanadium, chromium,
manganese, iron, cobalt, nickel and copper) are good. Four, magnetism
is essential. Five, proper Fermi surface geometry is essential (it must
match the structure of the spin excitations). Six, enlist theorists, at least
to compute the Fermi surfaces (I hope that theorists are more useful
than this but do not dare insist). Note that there is one corollary to these
rules. Materials of interest are likely to be complex chemical compounds
— work closely with solid-state chemists. ■
1. Pickett, W. E. The other high-temperature superconductors. Physica B 296, 112–119 (2001).
2. Mazin, I. I. Vitaly Ginzburg and high temperature superconductivity: personal
reminiscences. Physica C 468, 105–110 (2008).
3. Kamihara, Y., Watanabe, T., Hirano, M. & Hosono, H. Iron-based layered superconductor
La[O1−xFx]FeAs (x=0.05−0.12) with Tc=26K. J. Am. Chem. Soc. 130, 3296–3297 (2008).
4. Mazin, I. I. & Antropov, V. P. Electronic structure, electron–phonon coupling, and multiband
effects in MgB2. Physica C 385, 49–65 (2003).
5. Mazin, I. I., Singh, D. J., Johannes, M. D. & Du, M. H. Unconventional sign-reversing
superconductivity in LaFeAsO1−xFx. Phys. Rev. Lett. 101, 057003 (2008).
6. Kuroki, K. et al. Unconventional superconductivity originating from disconnected Fermi
surfaces in LaO1−xFxFeAs. Phys. Rev. Lett. 101, 087004 (2008).
7. Mazin, I. I. & Schmalian, J. Pairing symmetry and pairing state in ferropnictides: theoretical
overview. Physica C 469, 614–627 (2009).
8. Lee, P. A., Nagaosa, N. & Wen, X. G. Doping a Mott insulator: physics of high-temperature
superconductivity. Rev. Mod. Phys. 78, 17–85 (2006).
9. Chu, P. C. W. et al. (eds) Superconductivity in iron-pnictides. Physica C 469 (special issue),
313–674 (2009).
10. Anisimov, V. I., Kurmaev, E. Z., Moewes, A. & Izyumov, I. A. Strength of correlations in
pnictides and its assessment by theoretical calculations and spectroscopy experiments.
Physica C 469, 442–447 (2009).
11. Yang, W. L. et al. Evidence for weak electronic correlations in iron pnictides. Phys. Rev. B 80,
014508 (2009).
12. Lee, C.-H. et al. Effect of structural parameters on superconductivity in fluorine-free
LnFeAsO1−y (Ln=La, Nd). J. Phys. Soc. Jpn 77, 083704 (2008).
13. Hosono, H., Matsuishi, S., Nomura, T. & Hiramatsu, H. Iron-based superconducting
materials. Bull. Phys. Soc. Jpn 64, 807 (2009).
14. Scalapino, D. J. Superconductivity and spin fluctuations. J. Low Temp. Phys. 117, 179–188
(1999).
15. Van Harlingen, D. J. Phase-sensitive tests of the symmetry of the pairing state in the high-
temperature superconductors — evidence for dx2−y2 symmetry. Rev. Mod. Phys. 67, 515–535
(1995).
Acknowledgements I dedicate this article to the memory of Vitaly Ginzburg, a
relentless enthusiast of high-temperature superconductivity and my former teacher,
who passed away while this article was being written.
Author Information Reprints and permissions information is available at www.
nature.com/reprints. The author declares no competing financial interests.
Correspondence should be addressed to the author (mazin@dave.nrl.navy.mil).
NATURE|Vol 464|11 March 2010|doi:10.1038/nature08915 REVIEW INSIGHT

Non-Abelian states of matter

Ady Stern1

Quantum mechanics classifies all elementary particles as either fermions or bosons, and this classification
is crucial to the understanding of a variety of physical systems, such as lasers, metals and superconductors.
In certain two-dimensional systems, interactions between electrons or atoms lead to the formation of
quasiparticles that break the fermion–boson dichotomy. A particularly interesting alternative is offered by
‘non-Abelian’ states of matter, in which the presence of quasiparticles makes the ground state degenerate,
and interchanges of identical quasiparticles shift the system between different ground states. Present
experimental studies attempt to identify non-Abelian states in systems that manifest the fractional quantum
Hall effect. If such states can be identified, they may become useful for quantum computation.

Electrons are fermions; photons are bosons. Although seemingly a
mundane statement about the symmetries of a quantum mechani-
cal wavefunction when two identical particles are interchanged (for
instance with respect to degrees of freedom such as position or momen-
tum), this statement is in fact a pillar of the understanding of nature,
and a basis of the understanding of the periodic table and the existence
of metals, to list just two examples. Such quantum statistical consid-
erations also affect the behaviour of composites of quantum particles:
helium-4 atoms are bosons and form superfluids at low temperatures;
and pairs of electrons are effectively bosons, which is how Cooper pairs
of electrons make the phenomenon of superconductivity possible.
Non-Abelian systems1,2 contain composite particles that are neither
fermions nor bosons and have a quantum statistics that is far richer than
that offered by the fermion–boson dichotomy. The presence of such
quasiparticles manifests itself in two remarkable ways. First, it leads to
a degeneracy of the ground state that is not based on simple symmetry
considerations and is robust against perturbations and interactions with
the environment. Second, an interchange of two quasiparticles does not
merely multiply the wavefunction by a sign, as is the case for fermi-
ons and bosons. Rather, it takes the system from one ground state to
another. If a series of interchanges is made, the final state of the system
will depend on the order in which these interchanges are being carried
out, in sharp contrast to what happens when similar operations are per-
formed on identical fermions or bosons. It is this ‘ordering’ dependence
that justifies the name ‘non-Abelian’ (‘non-commutative’ in mathemati-
cal terms). Just as the minus sign that accompanies the interchange of
two fermions is independent of details such as their interaction or envi-
ronment, the effect of the interchange of two non-Abelian quasiparticles
is insensitive to noise from the environment around them.
Non-Abelian states have generated great interest recently1–5, for three
main reasons. The first is the theory behind them. Understanding of
their origin and properties is in its infancy. The second is the challenge to
observe them in an experiment. There are some systems in which strong
theoretical arguments suggest the existence of non-Abelian quasiparti-
cles, and that motivates a search for their experimental discovery. Third, if
non-Abelian states were shown to exist in realizable systems, they would
be ideal candidates for constructing topological quantum computers6–8.
A quantum computer needs a set of quantum states that is well separated
from the rest of the world. The degenerate ground states of a non-Abelian
system, separated by an energy gap from the rest of the spectrum, deliver
that. Furthermore, a quantum computer needs a minimum sensitivity
1
Department of Condensed Matter Physics, Weizmann Institute of Science, Rehovot 76100, Israel.
to noise and decoherence, and this requirement is amply satisfied by the
insensitivity of the effect of interchange of non-Abelian quasiparticles to
noise and perturbations. The topological aspects of these interchanges
motivate the name ‘topological quantum computation’.
There are real-life settings in which existing theory predicts
non-Abelian states of matter to exist, but they are not numerous2. Sys-
tems that manifest the fractional quantum Hall effect are believed to
have a series of states that are non-Abelian9–27. Closely related states may
be realized in cold atoms28,29, superconductors of p-wave pairing symme-
try2 and hybrid systems of superconductors with so-called topological
insulators30–32 (see page 194) and/or semiconductors33. Non-Abelian
lattice spin models have been proposed19 but are far from experimental
realization. Among all of these, the most prominently studied candidate
for an experimentally accessible non-Abelian state is the ν = ⁄ quantum
Hall state9–27.
In this Review, I discuss the properties of non-Abelian states, the systems
in which they are expected to emerge, the possible ways of identifying them
and the status of the experimental attempts to find them. I focus mainly
on the ν = ⁄ quantum Hall state but also touch on hybrid systems that
combine the effects of spin–orbit interaction with superconductivity.
Non-Abelian quasiparticles
Non-Abelian quasiparticles appeared on the quantum mechanical stage
after several decades of engagement with the quantum mechanics of
identical particles. The originators of quantum mechanics distinguished
fermions from bosons through the symmetry restrictions imposed
on their many-body wavefunctions. A wavefunction, Ψ(r1, …, rN), of
N identical spin-polarized particles whose coordinates are r1, …, rN must
be odd under the interchange of the positions of two of the particles if
they are fermions, and even if they are bosons.
In the past three decades, it has been realized that in two spatial dimen-
sions the statistics of quasiparticles formed as composites of the elemen-
tary particles of a system is not limited to the fermion–boson dichotomy.
As a first break from that dichotomy, the interchange of two quasiparticles
may multiply the wavefunction by a phase. That phase may take any value
and, as a consequence, the quasiparticles are known as ‘anyons’.
The second break from the fermion–boson dichotomy, embodied
by so-called non-Abelian quasiparticles, is much more surprising: an
interchange of two quasiparticles does not merely multiply the ground-
state wavefunction by a phase factor; rather, it shifts the system to a
different ground state.

187
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
Defining characteristics of non-Abelian states
To introduce the four defining characteristics of non-Abelian states of
matter, it is useful to think of a two-dimensional (2D) fluid in which
several vortices are localized at large distances from one another. Four
conditions need to be satisfied for these vortices to become non-Abelian
quasiparticles. First, there must be an energy gap separating the fluid’s
ground state from its excitations (Fig. 1a). Second, the ground state of
the system must be degenerate, with the degeneracy being exponentially
large in the number of vortices. Third, the degeneracy should be robust,
and insensitive to weak noise and perturbations. Fourth, when vortices
are made to ‘braid’, that is, to encircle one another or interchange their
positions (Fig. 1b), the system should transform from one ground state
to another, with the transformation being determined only by the topol-
ogy of the quasiparticles’ trajectory: two braiding trajectories that may
be deformed into one another without the quasiparticles ever getting
close should apply the same transformation to the system, up to a phase
factor. It is the combination of an energy gap, ground-state degeneracy,
robustness and topology — all of which are elaborated on below — that
defines non-Abelian states of matter.
Although energy gaps are common in many-body systems, the
combination of an energy gap with the three other characteristics is unique
to non-Abelian states. To elaborate on these characteristics, it is necessary
to be slightly more formal. The fluid in the picture above is made up of a
large number of microscopically identical particles (usually electrons, but
they could also be atoms, for instance) whose set of degrees of freedom is
{ri}. The vortices in turn are collective degrees of freedom, created by the
collective behaviour of the microscopic particles. Their number is much
smaller (Fig. 1b) than the total number of particles forming the back-
ground fluid. Assume that they are localized at the coordinates {Rj}. The
wavefunctions of the set of degenerate ground states, |Ψα(Rj)〉, enumerated
by the index α, are thus functions of the particles’ coordinates, {ri}, and
depend on the vortices’ coordinates, {Rj}, as parameters.
A braiding — a quasiparticle interchange operation — is carried out
when the vortices slowly move along a trajectory, {Rj(t)}, that starts and
ends in the same configuration, {Rj} (Fig. 1b). If there is just a single, non-
degenerate, ground state, the slowness (‘adiabaticity’) of the motion and
the existence of an energy gap guarantee that the final state of the system
is identical to the initial state, up to a possible phase factor. A series of
such braidings results in a product of phase factors. Such a product is
commutative and, hence, the quasiparticles are Abelian anyons.
By contrast, if the ground state is degenerate, a braiding may transform
one ground state, |Ψα(Rj)〉, into a different one, |Ψβ(Rj)〉. The transforma-
tion, Uαβ, which is a unitary matrix, depends only on the topology of the
a condensate is a Bardeen–Cooper–Schrieffer superconductor (see page 176
Energy
|Ya(Rj)°
b R1
Fluid
R2
Figure 1 | Characteristics of non-Abelian systems. In the presence of
vortices — non-Abelian quasiparticles — the spectrum includes a set
of degenerate ground states, |Ψα(Rj)〉, separated by an energy gap from a
continuous spectrum of excitations. a, Spectrum. b, Vortices and possible
braidings of their positions. The two winding trajectories on the right are
topologically equivalent.
188
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
quasiparticles’ trajectory, that is, on their windings around one another and
their mutual interchanges of position. It does not depend on the detailed
geometry of the trajectory. The transformation that corresponds to a set of
consecutive braidings is a product of such unitary matrices. Matrix product
is non-commutative, hence the term ‘non-Abelian’ and the naming of the
quasiparticles as non-Abelian anyons. The degeneracy of the ground state
and the non-Abelian statistics of the quasiparticles therefore go together.
Degeneracies in quantum spectra usually originate from a symmetry
of the underlying Hamiltonian, and are removed when the symmetry is
violated, for instance by the application of perturbations such as impuri-
ties or applied electric or magnetic fields. The degeneracy of the ground
state in a non-Abelian state is different. It does not originate from a simple
symmetry of the system, and as such it is robust. Small changes in the
details of the system — varied interactions or varied disorder — do not
lift the degeneracy. The degeneracy is only removed if a change to the
system is large enough to close the energy gap between the ground state
and the excitations (Fig. 1a), which would take the system to another
state of matter, or if the quasiparticles are brought close to one another. In
the second case, the degeneracy is lifted by terms that are exponentially
small in the ratio of the spatial separation between the quasiparticles and
a microscopic length, the coherence length of the underlying fluid.
The inner makings of a non-Abelian state
The question of what it is in the microscopic interactions between the
constituents of a system that gives rise to a non-Abelian state of matter
is a question of great interest and scarce understanding. Consider, for
example, the fractional quantum Hall state (Box 1), which so far is the
main arena for the study of non-Abelian states. A partially filled Landau
level is characterized by its ‘filling fraction’, ν, and by the type of interac-
tion between its electrons. For a wide range of interaction parameters,
the degeneracy of the ground state is insensitive to the details of the
interaction but is exponentially sensitive to ν. For example, the ground-
state degeneracy is exponential in (small) |ε| for ν = ⁄ + ε, thus making
the ν = ⁄ case non-Abelian (at least for a certain range of interactions),
whereas no such degeneracy appears for ν = ⁄ + ε.
A hint to a possible origin of non-Abelian systems may lie in the
observation that most theoretical constructs of these systems are fluids
in which microscopic particles of arbitrary quantum statistics group to
form clusters whose statistics is effectively bosonic1,18. These clusters
then form a Bose–Einstein condensate. Generally, breaking of a clus-
ter in that condensate costs energy, at least as much as the energy gap.
Crudely speaking, vortices formed in these Bose–Einstein condensates
become non-Abelian quasiparticles when their presence allows clusters
to be broken at no energy cost, leading to ground-state degeneracy.
In the simplest example, the cluster is a Cooper pair of fermions and the
for a discussion on the connections between superfluidity, superconduc-
tivity and Bose–Einstein condensation in systems of quantum particles).
Generally, the spectrum of a superconductor has excited states in which
some Cooper pairs are broken, and these states are separated by an energy
gap from the ground state. In the language of ‘second quantization’, these
excitations are created and annihilated by a set of creation and annihila-
tion operators. For some superconductors, in which the two fermions that
constitute the Cooper pair have a relative angular momentum equal to an
odd multiple of " (Planck’s constant divided by 2π), the presence of vortices
makes the energy cost for some of the excitations vanish, thus leading to
a degeneracy of the ground state. The simplest of these superconductors,
which has an angular momentum of " per pair, is known as a p-wave super-
conductor. The operators that create and annihilate the cost-free excitations
in this type of superconductor are known as Majorana operators. They are
Hermitian operators formed by equal-weight superposition of fermionic
creation and annihilation operators, and as such are electrically neutral
(see Box 2 and a recent review of Majorana fermions34).
Candidate systems
Given the well-developed understanding of Cooper pairs in supercon-
ductors, it is not surprising that the best-understood candidate systems
NATURE|Vol 464|11 March 2010
Box 1 | Integer and fractional quantum Hall systems
Ohmic
contacts
2D electron fluid
Ix
Edge
states
Magnetic field
out of page
Two-dimensional electronic systems subject to a perpendicular magnetic
field show the Hall effect, in which a current flow, Ix, is accompanied by a
voltage, Vy, perpendicular both to the current and to the magnetic field (see
figure, left, which shows a schematic of an experimental quantum Hall
system). The ratio of this voltage to the current, the Hall resistance, Rxy, may
be written as h/e2ν, where h is Planck’s constant, e is the electron charge
and ν is a dimensionless number. Classical considerations predict that
ν = nΦ0/B, where n is the electron density, Φ0 = hc/e is the flux quantum, c is
the speed of light and B is the magnetic field. At low enough temperature
and for clean enough samples, quantum Hall states are formed. These
states are characterized by quantized ν values, leading to plateaux in
the Hall resistance as a function of magnetic field (see figure, right).
of non-Abelian states are based on the model of Majorana operators in
a superconductor. Realizations of this model have been proposed in
fractional quantum Hall systems12,17, in p-wave superconductors and
in hybrid systems of superconductors with materials in which spin–
orbit coupling has a dominant role30–33. Non-Abelian states that are
not described by this model are believed to exist in certain fractional
quantum Hall states18,28, where they are based on clusters larger than
two fermions, and in certain spin models19.
Fractional quantum Hall systems
The first system in which non-Abelian quasiparticles were predicted to
exist12, the quantum Hall system (Box 1), is still the richest arena for the
search for non-Abelian states of matter. The fractional quantum Hall
effect occurs in a 2D electronic system subject to a strong magnetic field,
and is a consequence of the combination of the magnetic field and the
interaction between the electrons. It is characterized by an energy gap in
the bulk of the system, a gapless mode at its edge, a dissipationless flow of
current along the edges and a quantization of the Hall resistivity.
There are good reasons to believe that the ν = ⁄ fractional quantum
Hall state is non-Abelian9–27 and that the non-Abelian quasiparticles
carry one-quarter of an electron charge. Originally, deep considerations
of conformal field theories led Moore and Read12 to propose a trial wave-
function (known as the ‘Pfaffian’ wavefunction) for the ν = ⁄ state and
to conclude that the state may be non-Abelian.
A more intuitive picture that leads to the same understanding is based
on so-called composite fermion theory17,35,36. In this theoretical descrip-
tion, the problem of interacting electrons in a magnetic field B is mapped
onto a problem of composite fermions in a magnetic field Bʹ = B − 2Φ0n,
where Φ0 = hc/e is the flux quantum and n is the electron density. A com-
posite fermion is an electron to which two quanta of magnetic flux are
attached. This mapping modifies the interaction between the particles:
composite fermions interact both through their charges and through
the magnetic flux they carry.
The ν = ⁄ state is composed of two filled and inert Landau levels
and a half-filled level. Composite fermion theory maps the half-filled
Landau level (ν = ½) of interacting electrons onto spin-polarized com-
posite fermions at zero magnetic field (Bʹ = 0). These composite fer-
mions may then form a gapped state by pairing into Cooper pairs. For
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
0.50 ½(h/e2)
0.45
Rxy (h/e2)
Vy (h/e2)
0.40
0.35
⅓(h/e2)
4.5 5.0 5.5
Magnetic field (T)
Furthermore, these states are characterized by an energy gap in the bulk
and a chiral flow of the current along the edges.
The integer quantum Hall states, for which ν is an integer, may be
understood from a simple picture of non-interacting electrons under the
effect of a magnetic field and a disordered potential. For clean systems,
ν is roughly the number of filled Landau levels, that is, the number of highly
degenerate energy states of non-interacting electrons in a magnetic field.
The interaction between electrons is essential to the understanding of the
fractional quantum Hall states, for which ν is a fraction (see figure, right).
A useful tool for analysing how this interaction leads to the formation of
fractional quantum Hall states is composite fermion theory, as discussed
in the main text. (Right panel reproduced, with permission, from ref. 11.)
spin-polarized composite fermions, pairing is possible only if the relative
angular momentum per pair is odd, with the simplest example being
p-wave pairing. Vortices may be induced into the composite fermion
superconductor by variation of the magnetic field B such that the filling
fraction is shifted slightly away from ν = ⁄. These vortices carry Majo-
rana operators and are thus non-Abelian quasiparticles.
Numerical works20–27 support these theories in several ways, of which
I describe two. First, the ground state has been calculated numerically
for a small number of particles, and the Pfaffian wavefunction has been
found to approximate it remarkably well for a wide range of microscopic
interactions. Second, the spectrum has been obtained for a series of
Hamiltonians that slightly differ from one another in the microscopic
interactions between the electrons. At one end of this series, the Ham-
iltonian was artificially constructed such that the interaction between
its electrons makes the Pfaffian wavefunction its exact ground state. At
the other end, the Hamiltonian had the realistic Coulomb interaction
between the electrons built in. The intermediate Hamiltonians adiabati-
cally connect between these two extremes. The energy gap was found
not to vanish in all these cases, at least for the small systems that are
computationally tractable. Because topological properties are robust
as long as the energy gap does not close, this conclusion, if correct also
for an infinite system, implies that the exact ground state has the same
topological properties as the Pfaffian state.
If the ν = ⁄ is indeed a non-Abelian state, the first topological qubit
may be realized8, by means of an electronic analogue to the Fabry–Pérot
interferometer, as described below. It should be noted, however, that
a Majorana-based qubit has a limited potential (it is ‘non-universal’),
as the unitary transformations that can be applied to this type of qubit
through braiding of quasiparticles do not allow the system to reach
to all possible ground states7. It may be possible to realize a universal
topological qubit using other non-Abelian quantum Hall states, in
which the clusters consist of more than two electrons.
p-wave superconductors
If a p-wave superconductor of composite fermions such as the ν = ⁄
state is a good candidate for a non-Abelian state of matter, a p-wave
superconductor of less exotic particles, such as electrons or atoms, might
also be expected to be one. Although relatively rare, superconductors of
189
INSIGHT REVIEW
Box 2 | Majorana operators
Superconductivity results from the pairing of fermions into Cooper pairs.
When described in the formalism of second quantization, a fermion at a
point r in space is created by a creation operator ψ̂+(r) and is annihilated
by an annihilation operator ψ̂(r). The spectrum of a superconductor has
excitations in which some Cooper pairs are broken. These excitations,
which are separated by an energy gap from the ground state, are created
by a set of creation operators of the form Γ+ = ∫[g1(r)ψ̂(r) + g2(r)ψ̂+(r)] dr,
where g1(r) and g2(r) are functions determined by the details of the
system. The excitations are annihilated by a corresponding set of
operators, Γ, that are the Hermitian conjugates of Γ +. These excitations
are unique to superconductors: to create an excitation one needs a
superposition of operators creating a fermion and annihilating one.
For some superconductors, the presence of vortices allows Cooper
pairs to be broken at no energy cost. When that happens, the operators
that create and annihilate these zero-energy excitations have some
unique properties. First, they are Majorana operators, meaning that
Γ + = Γ or, equivalently, g1(r) = g2*(r). Creating such an excitation is the
same as annihilating it. Consequently, the zero-energy excitations
that are described by these operators cannot be counted in the way
particles are. Second, the functions g1(r) and g2(r) are localized within
the cores of the vortices. Third, the phases of the functions g1(r) and g2(r)
localized within the core of one vortex depend on the positions of all
other distant vortices in a multiply valued way. As a consequence, when
one vortex encircles another, the Majorana operators associated with
both are modified. These zero-energy excitations are the reason for the
degeneracy of the ground state. Their dependence on the position of the
vortices is what makes them sensitive to braiding.
this symmetry are formed in helium-3 under certain conditions, and
possibly also in electronic materials, such as SrRuO4. They have also
been conjectured to form from certain cold fermionic atoms29. Vortices
in these superconductors may then be good candidates in which to
observe non-Abelian statistics.
Unfortunately, none of these superconductors has so far been realized
in two dimensions. The vortices that carry Majorana modes in these
superconductors are generally not the lowest-energy vortices. Finally,
the energy difference between the zero-energy Majorana states and other
excitations within the cores of the vortices is likely to be very small, being
of the order of Δ2/EF, where Δ is the superconductor’s energy gap and
EF is its Fermi energy.
Hybrid systems
Another set of candidate systems for which Majorana modes are
proposed as a means of observing non-Abelian statistics is based on a
2D conductor in which spin–orbit coupling — a term in the Hamiltonian
that couples the orbital and spin degrees of freedom of electrons moving
in the crystal field potential of a solid — has a crucial role30–33. This 2D
conductor is to be positioned in proximity to a three-dimensional (3D)
superconductor in such a way that the latter induces superconductivity
in the former through the so-called proximity effect. The p-wave Cooper
pairing originates in this case from the spin–orbit coupling in the 2D
conductor. Conveniently, the 3D superconductor may be of the very
common s-wave or d-wave pairing, rather than the rare p-wave type.
Perhaps the most intriguing example for this idea is one in which
the normal conductor is the 2D metal formed on the surface of a 3D
topological insulator (see page 194 for a discussion of these newly dis-
covered materials). In this metal, spin–orbit coupling makes the elec-
tron spin lie in the plane of the metal and fixes its angle relative to the
electron’s linear momentum. Because Cooper pairs in a superconduc-
tor always comprise pairs of electrons with opposite linear momenta, a
superconductor formed at the surface of a topological insulator would
have Cooper pairs with electrons having a relative angular momentum
of ". The superconductor would therefore be of p-wave symmetry and
vortices formed therein would carry Majorana operators, as discussed
above. A similar situation should occur in semiconductors with strong
enough spin–orbit coupling27.
190
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
Detecting non-Abelian states
The difficulty of demonstrating that a many-body state is non-Abelian
is almost inherent in its nature. In experiments, the system under study
is usually perturbed, and then how it responds to perturbation is exam-
ined. However, the local perturbations used in experiments, such as
electric fields, do not couple different ground states of a non-Abelian
system, and therefore do not allow the multitude of ground states to be
probed. Nevertheless, several experiments have been proposed as ways
to probe non-Abelian states. Of these, I focus on those probing the ν = ⁄
fractional quantum Hall state.
Predicted signatures of the ν = 5⁄2 quantum Hall state
Most of the experimental effort directed at studying non-Abelian states
has so far been focused on the ν = ⁄ state. The motivation behind this
effort is twofold. The first is the interest in the nature of the ν = ⁄ state;
the non-Abelian Moore–Read phase is but one of several candidate
descriptions for this state, with Abelian and non-Abelian contenders37.
The second is the desire to observe non-Abelian quasiparticles and see
topological qubits realized.
The first of the four defining properties of a non-Abelian state, an
energy gap between the ground state and the excitations, characterizes all
quantum Hall states, including that with ν = ⁄. The second characteristic
to probe is the degeneracy of the ground state38,39. Being a property of
the spectrum, it may in principle be observed in thermodynamic meas-
urements, such as that of the heat capacity. In practice, the electronic
contribution to the heat capacity is overwhelmed by non-electronic,
low-energy degrees of freedom, such as phonons, nuclear spins and
so on. The electronic entropy, which is related to the degeneracy of the
ground state, would be better accessed through its effect on the tempera-
ture dependence of the electronic magnetization and chemical potential,
as dictated by Maxwell relations. Measurements of these two thermo-
dynamic quantities in 2D electronic systems have already been carried
out40,41, albeit not for the ν = ⁄ state.
Braiding and interferometry
The most intriguing — and challenging — characteristic to probe is
the unitary transformation applied by the braiding of quasiparticles.
Quasiparticles winding around one another may be observed most
directly through interferometry. The ‘holy grail’ of this line of study
is the realization of a qubit8, but an essential step along the way is a
definitive experimental demonstration of the quasiparticles being non-
Abelian. Generally, an interferometer interferes pairs of wave trajecto-
ries that enclose an interference loop. When the interfering particles are
non-interacting electrons in a strong magnetic field, the relative phase
between the paths is 2π times the number of flux quanta enclosed by the
interference loop, as dictated by the Aharonov–Bohm effect.
In interferometers in the fractional quantum Hall regime (Fig. 2), the
relative phase between the trajectories is affected also by the presence
of localized quasiparticles trapped within the loop42,43. In particular, for
the ν = ⁄ Moore–Read state44–47, no interference takes place when a non-
Abelian quarter-charged quasiparticle interferes around an interference
loop that encloses an odd number of similar quasiparticles. Interfer-
ence does take place when that number is even, in which case one of
two interference patterns will be seen, mutually shifted by a phase of π.
The choice between these two patterns is determined by the particular
ground state that the localized quasiparticles are in.
These general statements become concrete when two types of
interferometer are considered. A Fabry–Pérot interferometer42–47 is
shown schematically in Fig. 2a. In the limit where the two constrictions
introduce only a small amplitude for tunnelling between the edges, the
probability of the incoming current being backscattered involves an
interference of two trajectories only, one corresponding to each constric-
tion. Ideally, the number of localized quasiparticles in the interference
loop is controlled by the magnetic field, and the interference pattern is
probed as a function of the loop’s area, which is controlled by the appli-
cation of voltage to a side gate. Therefore, if the ν = ⁄ state is indeed the
non-Abelian Moore–Read state, interference should be turned on and
NATURE|Vol 464|11 March 2010
off periodically as the magnetic field is varied and the number of trapped
quasiparticles alternates between even and odd. For the even case, the
phase of the interference pattern may assume one of two possible values,
depending on the ground state occupied by the trapped quasiparticles.
When the inter-edge coupling is made stronger, the interferometer
becomes a quantum dot and the sinusoidal interference pattern becomes
a series of so-called Coulomb blockade peaks. These are resonances in
the transmission of the current through the dot that take place when
the dot can accommodate another electron at no electrostatic energy
cost. The signature of the Moore–Read state is then in two patterns of
peaks: equally spaced peaks when an odd number of quasiparticles is
trapped and peaks that are bunched into pairs when an even number
of quasiparticles is trapped. Again, for the even case there are two pos-
sible patterns of peaks, depending on the ground state of the trapped
quasiparticles44,47.
A Fabry–Pérot interferometer enclosing two quasiparticles is the
proposed ν = ⁄ qubit8 and a measurement of its interference pattern or
the positions of its Coulomb blockade peaks as a function of its area serves
as a read-out of its state. The strength of this qubit is in its insensitivity to
noise, such as that from phonons or photons. As long as the noise is too
weak to expel one of the localized quasiparticles from the interferometer,
it will have no effect on the ground state occupied by these quasiparticles.
To change this state, a winding is needed, this time a winding of an edge
quasiparticle around an odd number of the localized ones.
A Mach–Zehnder interferometer48,49 is shown schematically in
Fig. 2b. In contrast to the Fabry–Pérot interferometer, its interference
loop encloses one of the edges that separate the ν = ⁄ state from the
vacuum. The current enters the interferometer from the source located
at the exterior edge, and ends up in one of the two drains. One drain is
located at the interior edge and the other at the exterior edge. Whenever
a quasiparticle tunnels from the exterior edge to the interior edge, it
changes the parity of the number of quasiparticles within the interfer-
ence loop, and affects the tunnelling probability for the quasiparticles
that follow. The measured current is then a proper average over tunnel-
ling rates that are history dependent. As a striking consequence, in the
limit of weak tunnelling through the constrictions the current flows in
batches of quasiparticles. The average charge in each batch oscillates
with the magnetic flux through the interferometer, and varies between
one-quarter and about three electron charges. This average, which may
be observed in noise measurements, is limited to be less than or equal
to one electron charge for single-edge quantum Hall states, such as
the ν = 1 and ν = ⅓ states. Because the number of quasiparticles in the
loop inherently varies in time, its thermal fluctuations do not affect the
noise. Experiments carried out using a Mach–Zehnder interferom-
eter in the regime of the integer quantum Hall effect conform closely
to the ideal model, at least in the limit of small currents through the
interferometer (the ‘linear response limit’)50,51. Technically, however,
the Mach–Zehnder interferometer is complicated to realize, and has
so far been used only for integer filling fractions, which are easier to
study (see Fig. 3).
The non-Abelian nature of the ν = ⁄ state may also be probed by a
direct measurement of the edge, independent of the bulk and the local-
ized quasiparticles that it contains. Possible probes include tunnelling of
electrons into the edge from the outside, tunnelling between two edges
and thermal transport17,37,52–55. The current–voltage characteristics of
such transport, the associated current noise and the thermal Hall con-
ductance all carry signatures of the nature of the ν = ⁄ state.
Present experimental status of the ν = 5⁄2 quantum Hall state
A realization of the various experimental tests described here is being
vigorously attempted at present by many experimenters. Preliminary
results support the identification of the ν = ⁄ quasiparticles as carriers
of one-quarter of an electron charge, a necessary condition for non-
Abelian phenomena to be observed56,57. These results are, however,
non-definitive, and are not fully understood.
A Fabry–Pérot interferometer has already been thoroughly studied for
quantum Hall states with integer ν values, ν = ⅓ and ν = ⁄ (refs 58–61). The
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
a
D2
S D1
b
S D2 D1
Figure 2 | Interferometric measurement setups in quantum Hall
systems. a, In the context of the quantum Hall effect, the Fabry–Pérot
interferometer is a Hall bar perturbed by two constrictions (‘quantum point
contacts’, each represented by a pair of triangles). Current (black) flows
chirally along the edges. The edge along which it moves to the right is put at
a higher chemical potential than the edge along which it moves to the left.
The point contacts introduce amplitudes t1 and t2 for inter-edge tunnelling.
The presence of two point contacts makes the probability of backscattering
the quantum Hall analogue of the probability of reaching the interference
screen in a double-slit interference experiment. The backscattered current
is, to lowest order, proportional to |t1 + t2|2. The relative phase between the
two amplitudes may be varied by varying the magnetic field and/or the area
of the interference loop. The latter is implemented using a side gate that, to
the crudest approximation, ‘moves’ the walls of the cell without introducing
quasiparticles into the bulk. By contrast, the variation of the magnetic
field leads to the introduction of localized quasiholes or quasiparticles.
As explained in the text, if the ν = ⁄ state is non-Abelian, the interference
will be turned off when the number of quasiparticles localized in the bulk
is odd, and turned on when it is even. S, source; D1, D2, drains. b, The
Mach–Zehnder interferometer is another device that allows interference of
trajectories of particles that are backscattered through two point contacts.
In the Fabry–Pérot interferometer, the interference loop encloses only the
cell between the two point contacts, whereas here the entire internal edge is
part of the interference loop. The image shows the two trajectories leading
from the source to one of the drains.
constraints on the sample quality and temperature needed to observe these
states are less stringent than those needed for the ν = ⁄ state, primarily as a
result of their larger energy gaps. These studies showed that the magnetic
field and the voltage on the side gate affect the interference in a way that is
more complicated than that predicted by the idealized model described
above (Fig. 3a). For most of the interferometers studied, the magnetic
field and the gate voltage affect both the area of the interference loop and
the number of localized quasiparticles62. Although these difficulties have
to be understood in detail, they do not obscure the distinction, described
above, between the types of interference corresponding to even and odd
numbers of localized quasiparticles.
A Fabry–Pérot interferometer has also been realized for the ν = ⁄ state.
Recent studies yield encouraging, yet not definitive, results63,64 (Fig. 3). The
backscattering probability of the current, reflected in the interferometer’s
resistance, was observed to oscillate as a function of the voltage on a side
gate, and the periodicity of the oscillations was observed to switch between
two values. The oscillations were attributed to the effect of the side gate
on the interferometer’s area, and the switching between periodicities was
attributed to its effect on the number of localized quasiparticles. In this
interpretation, one periodicity corresponds to the interference of non-
Abelian quarter-charged edge quasiparticles, and dominates when the
number of localized bulk quasiparticles is even. The other periodicity cor-
responds to the interference of Abelian half-charged edge quasiparticles,
and dominates when the number of localized bulk quasiparticles is odd
and quarter-charged quasiparticles do not interfere65.
Although the observed switching of periodicities is suggestive, it is not
yet unambiguous. To make a conclusive identification, a two-pronged
effort is likely to be required. On the experimental side, a mapping of
191
INSIGHT REVIEW NATURE|Vol 464|11 March 2010

the backscattered current in the two-parameter plane of gate voltage and
magnetic field is probably the best way to establish the interpretation of the
oscillations. On the theoretical side, a connection should be made between
the ideal picture and the real world. In the ideal picture, quasiparticles are
kept very far away from one another, their number is controlled and their
positions may be braided at will. In the real world none of these fully holds.
Moreover, energy splittings that are exponentially small cannot always
be neglected, temperature may not be assumed to be zero and localized
bulk quasiparticles may have tunnel couplings to the edge. Some of these
issues have already been addressed in theoretical works66–71, and others
still need to be understood.
An example of the latter is the issue of spin polarization (refs 22, 72, 73
and T. D. Rhone and A. Pinczuk, personal communication). The original
Moore–Read state is fully spin polarized, and numerical works support
the assumption of full polarization of the ground state. Full polarization is,
however, not a condition for non-Abelian quasiparticles to exist. Experi-
ments indicate that in real samples spin polarization may not be complete,
either as a property of the ground state or as a finite-temperature conse-
quence of low-energy spin-reversed excitations. The effect of incomplete
spin polarization on the attempts to study non-Abelian statistics in the
ν = ⁄ state and to use it for topological quantum computation is not fully
understood yet.
Outlook
The experiments reviewed in the previous sections, those proposed and
those attempted, are aimed at the observation of a non-Abelian state of
matter. While this observation is awaited, researchers in the field actively
study its theory.
In the context of known candidates for non-Abelian states of matter,
much is still to be learned about probing states whose main defining prop-
erty is their insensitivity to perturbations. The list of possible definitive
tests for the identification of non-Abelian states is likely to be extended by
future proposals, and the subtleties associated with interpreting experi-
mental results will be better understood.
In the general context of non-Abelian states of matter, the theoretical
challenge is wide open — it is the search for physical systems that may
form non-Abelian states. Among the plethora of interacting systems that
condensed-matter physics makes it possible to realize, it is desirable to
identify those in which topology dominates the physics. It is to be hoped
that, armed with such an understanding, the fine control that experimen-
talists have over the synthesis of new materials, the fabrication of elec-
tronic systems and the interaction between cold atoms in condensates
can be used to design non-Abelian states of matter.
In the field of topological quantum computation, the potential of the
special properties of non-Abelian states to help quantum computers over-

a 2.0-Nm2 device b d
1.0
6 0.8
5
4 0.7
0.8
0
2 0.6

%B (G)
0.6 –5 0 0.5
V (mV)

–2 0.4
%R (10–3h/e2)

0.4 0.3
–4
0.2
–6
0.2 0.1
6 8 10 12 14 16 18
V (mV)
0.0
1.763 1.764 1.765 1.766 c
B (T)
Near filling fraction 5/2
e/2 e/2
0.01
18-Nm2 device
4

2 0.00
2
%R t200 (8)

0
0 –2 –0.01
0.3
V (mV)

–2 e/4 e/4 e/4

Amplitude

0.2
–0.02
%R (10–3h/e2)

–4
0.1 Period, 15.5 mV
e/2
–6 0
–0.03 0 0.1 0.2
Frequency (mV–1)
–8
0.544 0.545 0.546 0 40 80 120 160 200 240
B (T) Change in side-gate voltage V (mV)

Figure 3 | Experimental results from interferometers. a, Fabry–Pérot
interferometers were realized first in the integer59,61 and Abelian fractional61
quantum Hall states. The plots show the backscattered current (expressed
in terms of the change in the sample’s resistance, ΔR) as a function of
magnetic field, B, and voltage on a side gate, V, for small and large areas
of interference loops, for an integer quantum Hall state. The different
directions of the constant phase lines are due to the variation of the
interferometer area with magnetic field, which is crucial for small devices
and negligible for large ones. (Image reproduced, with permission, from
ref. 59.) b, Image of a ν = ⁄ Fabry–Pérot interferometer. Scale, 1 μm.
(Image reproduced from ref. 63.) c, Data from the ν = ⁄ Fabry–Pérot
192
© 2010 Macmillan Publishers Limited. All rights reserved
interferometer. The change in the resistance is shown as a function of
the change in side-gate voltage. The switching of periodicities may be an
observation of effects associated with non-Abelian statistics. Inset, Fourier
transform of the data in the main panel. (Panel reproduced from ref. 63.)
d, Mach–Zehnder interferometers have already been realized, but only for
states of the integer quantum Hall effect. The plot shows the probability
of the current reaching from the source to the drain at the exterior as a
function both of the variation of magnetic field, ΔB, around a value of 9 T,
and of voltage applied to a side gate. (Data are unpublished data associated
with ref. 50; panel courtesy of N. Ofek and M. Heiblum, Weizmann Institute
of Science, Rehovot, Israel.)
NATURE|Vol 464|11 March 2010
come decoherence is all but exhausted, even theoretically.
In conclusion, the past few decades have highlighted the surprisingly
constructive interference of the subtlety of topology, the richness of quan-
tum mechanics and the sophistication of materials science. In the next
decade, it is to be hoped that this combination will express itself in making
non-Abelian states a reality. ■ statistics in fractional quantum Hall liquids. Phys. Rev. B 79, 115317 (2009).
1. Stern, A. Anyons and the quantum Hall effect — a pedagogical review. Ann. Phys. 323,
204–249 (2008).
This paper is a pedagogical introduction to the concept of Abelian and non-Abelian
anyons.
2. Nayak, C., Simon, S. H., Stern, A., Freedman, M. & Das Sarma, S. Non-Abelian anyons and
topological quantum computation. Rev. Mod. Phys. 80, 1083–1159 (2008).
This is a comprehensive review of non-Abelian states of matter and their relevance to
quantum computation, and contains an extensive list of references.
3. Das Sarma, S., Freedman, M. & Nayak, C. Topological quantum computation. Phys. Today 59,
32–38 (2006).
4. Collins, G. P. Computing with quantum knots. Sci. Am. 4, 57 (2006).
5. Day, C. Devices based on the fractional quantum Hall effect may fulfill the promise of
quantum computing. Phys. Today 58, 21–23(2005).
6. Kitaev, A. Y. Fault-tolerant quantum computation by anyons. Ann. Phys. 303, 2–30 (2003).
This paper discusses the use of anyons for quantum computation.
7. Freedman, M. H. et al. Topological quantum computation. Bull. Am. Math. Soc. 40, 31–38
(2003).
8. Das Sarma, S., Freedman, M. & Nayak, C. Topologically protected qubits from a possible
non-Abelian fractional quantum Hall state. Phys. Rev. Lett. 94, 166802 (2005).
This paper proposes a qubit based on the ν=5⁄2 state.
9. Willett, R. et al. Observation of an even-denominator quantum number in the fractional
quantum Hall effect. Phys. Rev. Lett. 59, 1776–1779 (1987).
10. Xia, J. S. et al. Electron correlation in the second Landau level: a competition between many
nearly degenerate quantum phases. Phys. Rev. Lett. 93, 176809 (2004).
11. Eisenstein, J. P., Cooper, K. B., Pfeiffer, L. N. & West, K. W. Insulating and fractional quantum
Hall states in the first excited Landau level. Phys. Rev. Lett. 88, 076801 (2002).
12. Moore, G. & Read, N. Non-Abelions in the fractional quantum Hall effect. Nucl. Phys. B 360,
362–396 (1991).
This paper introduced the concept of non-Abelian states of matter.
13. Read, N. Non-Abelian adiabatic statistics and Hall viscosity in quantum Hall states and
px+ipy paired superfluids. Phys. Rev. B 79, 045308 (2009).
14. Nayak, C. & Wilczek, F. 2n-quasihole states realize 2n−1-dimensional spinor braiding
statistics in paired quantum Hall states. Nucl. Phys. B. 479, 529–553 (1996).
15. Ivanov, D. A. Non-Abelian statistics of half-quantum vortices in p-wave superconductors.
Phys. Rev. Lett. 86, 268–271 (2001).
16. Stern, A., von Oppen, F. & Mariani, E. Geometric phases and quantum entanglement as
building blocks for non-Abelian quasiparticle statistics. Phys. Rev. B 70, 205338 (2004).
17. Read, N. & Green, D. Paired states of fermions in two dimensions with breaking of parity and
time-reversal symmetries and the fractional quantum Hall effect. Phys. Rev. B 61, 10267–
10297 (2000).
This paper puts forward a composite fermion theory of the ν=5⁄2 state.
18. Read, N. & Rezayi, E. Beyond paired quantum Hall states: parafermions and incompressible
states in the first excited Landau level. Phys. Rev. B 59, 8084–8092 (1999).
This paper describes a proposed series of non-Abelian quantum Hall states of matter.
19. Levin, M. A. & Wen, X. G. String-net condensation: a physical mechanism for topological
phases. Phys. Rev. B 71, 045110 (2005).
20. Wen, X. G. Topological order and edge structure of ν=½ quantum Hall state. Phys. Rev. Lett.
70, 355–358 (1993).
21. Greiter, M., Wen, X.-G. & Wilczek, F. Paired Hall state at half filling. Phys. Rev. Lett. 66,
3205–3208 (1991).
22. Morf, R. H. Transition from quantum Hall to compressible states in the second Landau level:
new light on the ν=5⁄2 enigma. Phys. Rev. Lett. 80, 1505–1508 (1998).
23. Storni, M., Morf, R. H. & Das Sarma, S. The fractional quantum Hall state at ν=5⁄2 and the
Moore–Read Pfaffian. Preprint at 〈http://arxiv.org/abs/0812.2691〉 (2008).
24. Rezayi, E. H. & Haldane, F. D. M. Incompressible paired Hall state, stripe order and the
composite fermion liquid phase in half-filled Landau levels. Phys. Rev. Lett. 84, 4685–4688
(2000).
25. Tőke, C., Regnault, N. & Jain, J. K. Nature of excitations of the 5⁄2 fractional quantum Hall
effect. Phys. Rev. Lett. 98, 036806 (2007).
26. Wójs, W. & Quinn, J. J. Landau level mixing in the ν=5⁄2 fractional quantum Hall state. Phys.
Rev. B 74, 235319 (2006).
27. Feiguin, A. E., Rezayi, E., Nayak, C. & Das Sarma, S. Density matrix renormalization group
study of incompressible fractional quantum Hall states. Phys. Rev. Lett. 100, 166803 (2008).
28. Cooper, N. R., Wilkin, N. K. & Gunn, J. M. F. Quantum phases of vortices in rotating Bose–
Einstein condensates. Phys. Rev. Lett. 87, 120405 (2001).
29. Gurarie, V. & Radzihovsky, L. Resonantly paired fermionic superfluids. Ann. Phys. 322, 2–119
(2007).
30. Fu, L. & Kane, C. L. Superconducting proximity effect and Majorana fermions at the surface
of a topological insulator. Phys. Rev. Lett. 100, 096407 (2008).
This paper proposes a setting in which Majorana fermions are created in a hybrid system
of a topological insulator and a superconductor.
31. Fu, L. & Kane, C. L. Probing neutral Majorana fermion edge modes with charge transport.
Phys. Rev. Lett. 102, 216403 (2009).
32. Nilsson, J., Akhmerov, A. R. & Beenakker, C. W. Splitting of a Cooper pair by a pair of
Majorana bound states. Phys. Rev. Lett. 101, 120403 (2008).
33. Sau, J. D., Lutchyn, R. M., Tewari, S. & Das Sarma, S. A generic new platform for topological
quantum computation using semiconductor heterostructures. Preprint at 〈http://arxiv.org/
abs/0907.2239〉 (2009).
34. Wilczek, F. Majorana returns. Nature Phys. 5, 614–618 (2009).
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
35. Jain, J. Composite Fermions (Cambridge Univ. Press, 2007).
36. Heinonen, O. (ed.) Composite Fermions (World Scientific, 1998).
37. Boyarsky, A., Cheianov, V. & Fröhlich, J. Effective field theories for the ν=5⁄2 edge. Phys. Rev. B
80, 233302 (2009).
38. Cooper, N. R. & Stern, A. Observable bulk signatures of non-Abelian quantum Hall states.
Phys. Rev. Lett. 102, 176807 (2009).
39. Yang, K. & Halperin, B. I. Thermopower as a possible probe of non-Abelian quasiparticle
40. Eisenstein, J. P., Pfeiffer, L. N. & West, K. W. Compressibility of the two-dimensional electron
gas: measurements of the zero-field exchange energy and fractional quantum Hall gap.
Phys. Rev. B 50, 1760–1778 (1994).
41. Wiegers, S. A. J. et al. Magnetization and energy gaps of a high-mobility 2D electron gas in
the quantum limit. Phys. Rev. Lett. 79, 3238–3241 (1997).
42. Chamon, C. de C., Freed, D. E., Kivelson, S. A., Sondhi, S. L. & Wen, X. G. Two point-contact
interferometer for quantum Hall systems. Phys. Rev. B 55, 2331–2343 (1997).
43. Fradkin, E. et al. A Chern–Simons effective field theory for the Pfaffian quantum Hall state.
Nucl. Phys. B 516, 704–718 (1998).
44. Stern, A. & Halperin, B. I. Proposed experiments to probe the non-Abelian ν=5⁄2 quantum
Hall state. Phys. Rev. Lett. 96, 016802 (2006).
45. Bonderson, P., Kitaev, A. & Shtengel, K. Detecting non-Abelian statistics in the ν=5⁄2
fractional quantum Hall state. Phys. Rev. Lett. 96, 016803 (2006).
46. Bonderson, P., Shtengel, K. & Slingerland, J. K. Probing non-Abelian statistics with
quasiparticle interferometry. Phys. Rev. Lett. 97, 016401 (2006).
47. Ilan, R., Grosfeld, E., Schoutens, K. & Stern, A. Experimental signatures of non-Abelian
statistics in clustered quantum Hall states. Phys. Rev. B 79, 245305 (2009).
48. Feldman, D. E. & Kitaev, A. Detecting non-Abelian statistics with an electronic
Mach–Zehnder interferometer. Phys. Rev. Lett. 97, 186803 (2006).
49. Feldman, D. E., Gefen, Y., Kitaev, A., Law, K. T. & Stern, A. Shot noise in an anyonic
Mach–Zehnder interferometer. Phys. Rev. B 76, 085333 (2007).
50. Neder, I. et al. Interference between two indistinguishable electrons from independent
sources. Nature 448, 333–337 (2007).
51. Roulleau, P. et al. Direct measurement of the coherence length of edge states in the integer
quantum Hall regime. Phys. Rev. Lett. 100, 126802 (2008).
52. Milovanović, M. & Read, N. Edge excitations of paired fractional quantum Hall states. Phys.
Rev. B 53, 13559–13582 (1996).
53. Fendley, P., Fisher, M. P. A. & Nayak, C. Edge states and tunneling of non-Abelian
quasiparticles in the ν=5⁄2 quantum Hall state and p+ip superconductors. Phys. Rev. B. 75,
045317 (2007).
54. Fiete, G. A., Bishara, W. & Nayak, C. Multi-channel Kondo models in non-Abelian quantum
Hall droplets. Phys. Rev. Lett. 101, 176801 (2008).
55. Feldman, D. E. & Li, F. Charge-statistics separation and probing non-Abelian states. Phys.
Rev. B 78, 161304 (2008).
56. Dolev, M. et al. Observation of a quarter of an electron charge at the ν=5⁄2 quantum Hall
state. Nature 452, 829–834 (2008).
57. Radu, I. P. et al. Quasi-particle properties from tunneling in the ν=5⁄2 fractional quantum Hall
state. Science 320, 899–902 (2008).
58. Camino, F. E., Zhou, W. & Goldman, V. J. Quantum transport in electron Fabry–Perot
interferometers. Phys. Rev. B 76, 155305 (2007).
59. Zhang, Y. et al. Distinct signatures for Coulomb blockade and Aharonov–Bohm interference
in electronic Fabry–Perot interferometers. Phys. Rev. B 79, 241304 (2009).
60. Goldman, V. J. Superperiods and quantum statistics of Laughlin quasiparticles. Phys. Rev. B
75, 045334 (2007).
61. Ofek, N. et al. The role of interactions in an electronic Fabry–Perot interferometer operating
in the quantum Hall effect regime. Preprint at 〈http://arxiv.org/abs/0911.0794〉 (2009).
62. Rosenow, B. & Halperin, B. I. Influence of interactions on flux and back-gate period of
quantum Hall interferometers. Phys. Rev. Lett. 98, 106801 (2007).
63. Willett, R. L., Pfeiffer, L. N. & West, K. W. Measurement of filling factor 5⁄2 quasiparticle
interference with observation of charge e/4 and e/2 period oscillations. Proc. Natl Acad. Sci.
USA 106, 8853–8858 (2009).
64. Willett, R. L., Pfeiffer, L. N. & West, K. W. Alternating e/4 and e/2 period interference
oscillations consistent with filling factor 5⁄2 non-Abelian quasiparticles. Preprint at
〈http://arxiv.org/abs/0911.0345〉 (2009).
65. Bishara, W., Bonderson, P., Nayak, C., Shtengel, K. & Slingerland, J. K. Interferometric
signature of non-Abelian anyons. Phys. Rev. B 80, 155303 (2009).
66. Rosenow, B., Halperin, B., Simon, S. & Stern, A. Bulk-edge coupling in the non-Abelian ν=5⁄2
quantum Hall interferometer. Phys. Rev. Lett. 100, 226803 (2008).
67. Overbosch, B. J. & Wen, X. G. Dynamical and scaling properties of ν=5⁄2 interferometer.
Preprint at 〈http://arxiv.org/abs/0706.4339〉 (2007).
68. Rosenow, B., Halperin, B., Simon, S. & Stern, A. Exact solution for bulk-edge coupling in the
non-Abelian ν=5⁄2 quantum Hall interferometer. Phys. Rev. B 80, 155305 (2009).
69. Bishara, W. & Nayak, C. Odd–even crossover in a non-Abelian ν=5⁄2 interferometer. Phys.
Rev. B 80, 155304 (2009).
70. Saarikoski, H., Tölö, E., Harju, A. & Räsänen, E. Pfaffian and fragmented states at ν=5⁄2 in
quantum Hall droplets. Phys. Rev. B 78, 195321 (2008).
71. Overbosch, B. J. & Bais, F. A. Inequivalent classes of interference experiments with non-
Abelian anyons. Phys. Rev. A 64, 062107 (2001).
72. Feiguin, A. E., Rezayi, E., Yang, K., Nayak, C. & Das Sarma, S. Spin polarization of the ν=5⁄2
quantum Hall state. Phys. Rev. B 79, 115322 (2009).
73. Dean, C. R. et al. Contrasting behavior of the ν=5⁄2 and 7⁄3 fractional quantum Hall effect in a
tilted field. Phys. Rev. Lett. 101, 186806 (2008).
Acknowledgements This work was supported by the US–Israel Binational Science
Foundation, the Minerva foundation and Microsoft’s Station Q.
Author Information Reprints and permissions information is available at www.nature.
com/reprints. The author declares no competing financial interests. Correspondence
should be addressed to the author (adiel.stern@weizmann.ac.il).
193
INSIGHT PERSPECTIVE
The birth of topological insulators
Joel E. Moore1,2
Certain insulators have exotic metallic states on their surfaces. These states are formed by topological
effects that also render the electrons travelling on such surfaces insensitive to scattering by impurities. Such
topological insulators may provide new routes to generating novel phases and particles, possibly finding uses
in technological applications in spintronics and quantum computing.
Many aspects of condensed-matter physics are concerned with
understanding how order emerges when a very large number of simple
constituents, such as ions, magnetic moments or electrons, interact with
each other. In ordered phases such as crystals and magnets, the order is
described through symmetry breaking: in a crystal, ions are arranged
periodically owing to their electrostatic interactions, thereby breaking
the continuous symmetry of space under rotations and translations;
in typical magnets, some of the rotational symmetry of spin space is
broken, together with the time-reversal symmetry.
A major discovery in the 1980s was that electrons that are confined
to two dimensions and subject to a strong magnetic field show a com-
pletely different, topological, type of order, which underlies the quantum
Hall effect. Consequences of this order include dissipationless transport
and emergent particles with fractional charge and statistics. One of the
important discoveries of the past few years is that topological order also
occurs in some three-dimensional (3D) materials; in these materials, the
role of the magnetic field is assumed by the mechanism of spin–orbit
coupling, an intrinsic property of all solids. These materials have been
named topological insulators because they are insulators in the ‘bulk’
but have exotic metallic states present at their surfaces owing to the
topological order.
In this Perspective article, I provide an overview of the basic concepts
underlying topological insulators and recent studies of these remark-
able new materials. After an explanation of what makes some insulators
‘topological’ and a brief history of this rapidly developing field, I discuss
recent advances made in experiments on topological insulators — for
both bulk and nanostructured materials — and in the theoretical under-
standing of these materials. I conclude by explaining why many research
groups are seeking to use topological insulators to generate new particles
and phases, which has possible applications to quantum computing.
A primer on topological insulators
The easiest way to describe a topological insulator is as an insulator
that always has a metallic boundary when placed next to a vacuum
or an ‘ordinary’ insulator. These metallic boundaries originate from
topological invariants, which cannot change as long as a material
remains insulating. An intuitive illustration showing why the metallic
surfaces exist is presented in Fig. 1a: a trefoil knot is used to represent
a topological insulator and a closed loop to represent an ordinary one.
Topology is a branch of mathematics that studies the properties of
objects that are invariant under smooth deformations, a classic exam-
ple being a doughnut transforming into a coffee cup. In contrast to the
doughnut/coffee cup pair, the trefoil knot and the closed loop have
different topological invariants and thus neither can be deformed to
become the other, no matter how the string (or wire) is stretched or
1
Department of Physics, 366 Le Conte #7300, University of California, Berkeley, California 94720, USA. 2Materials Sciences Division, Lawrence Berkeley National Laboratory, Berkeley,
California 94720, USA.
194
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010|doi:10.1038/nature08916
twisted, without being cut. Nevertheless, these invariants do change
in crossing the interface between topological and ordinary insulators,
so by contradiction the surface cannot remain insulating, which is
analogous to cutting the knot.
In a topological insulator, the component that is ‘knotted’ is the
electron’s wavefunction as it moves through momentum space. Associ-
ated with this knotting are topological invariants (usually expressed as
integrals involving the wavefunction) that cannot change as long as the
material remains insulating. Even when the boundary between ordi-
nary insulators and topological insulators is atomically sharp and the
continuous description using topology might not seem to be applica-
ble, there are nevertheless metallic delocalized wavefunctions at this
boundary. In Fig. 1b, the simplest case of knotting in a 3D electronic
structure is illustrated: for one occupied electronic band and one empty
one, each point in 3D momentum space is associated with a unit vector
representing the occupied state, and the Hopf map shown in Fig. 1b is a
topologically non-trivial example.
Lessons from the past
The appearance of the unusual metal when the topology changes at a
surface is the main experimental signature that an insulator is indeed
topological. To explain the properties of the insulator, I will begin with
a short review of the historical developments that led to the theoretical
predictions that topological insulators exist. A simpler version of this
metal occurs at the edge of a quantum Hall droplet, the first reported
example of two-dimensional (2D) topological order. Quantum Hall
edges are perfect quantum wires wrapping around the insulating droplet
(Fig. 2a), and they result from the topological properties of the elec-
tronic wavefunctions when the constituent electrons are confined to
two dimensions and subject to a strong magnetic field perpendicular
to the plane to which the electrons are confined (see page 187 for more
details on the quantum Hall effect).
Work on topological insulators grew out of the idea that the quantum
Hall effect that arises in such 2D systems in the presence of a magnetic
field could occur even for electrons moving on a lattice in the absence
of a macroscopic magnetic field1. Instead of being driven by such a mag-
netic field, it was predicted in the late 1980s that electrons could, in
principle, form a quantum Hall state driven by forces that result from
their motion through the crystal lattice. Recent developments are based
on spin–orbit coupling, a relativistic effect in which the spin and orbital
angular momentum degrees of freedom of electrons are coupled; this
coupling causes electrons that are moving through a crystal to feel a
spin-dependent force, even in non-magnetic materials.
Although spin–orbit coupling does not have the symmetry required
to induce the quantum Hall effect (that is, it does not break time-reversal
NATURE|Vol 464|11 March 2010
a b
Figure 1 | Metallic states are born when a surface unties ‘knotted’ electron
wavefunctions. a, An illustration of topological change and the resultant
surface state. The trefoil knot (left) and the simple loop (right) represent
different insulating materials: the knot is a topological insulator, and the
loop is an ordinary insulator. Because there is no continuous deformation
by which one can be converted into the other, there must be a surface where
the string is cut, shown as a string with open ends (centre), to pass between
the two knots; more formally, the topological invariants cannot remain
symmetry as an applied magnetic field would), in simplified models
introduced in around 2003 it can lead to a quantum spin Hall effect,
in which electrons with opposite spin angular momentum (commonly
called spin up and spin down) move in opposite directions around the
edge of the droplet in the absence of an external magnetic field2 (Fig. 2b).
These simplified models were the first steps towards understanding
topological insulators. But it was unclear how realistic the models were:
in real materials, there is mixing of spin-up and spin-down electrons,
and there is no conserved spin current. It was also unclear whether the
edge state of the droplet in Fig. 2b would survive the addition of even
a few impurities.
In 2005, a key theoretical advance was made by Kane and Mele3.
They used more realistic models, without a conserved spin current,
and showed how some of the physics of the quantum spin Hall effect
can survive. They found a new type of topological invariant that could
be computed for any 2D material and would allow the prediction of
whether the material had a stable edge state. This allowed them to show
that, despite the edge not being stable in many previous models, there are
realistic 2D materials that would have a stable edge state in the absence of
a magnetic field; the resultant 2D state was the first topological insulator
to be understood. This non-magnetic insulator has edges that act like
perfectly conducting one-dimensional electronic wires at low tempera-
tures, similar to those in the quantum Hall effect.
Subsequently, Bernevig, Hughes and Zhang made a theoretical
prediction that a 2D topological insulator with quantized charge con-
ductance along the edges would be realized in (Hg,Cd)Te quantum
wells4. The quantized charge conductance was indeed observed in this
system, as a quantum-Hall-like plateau in zero magnetic field, in 2007
(ref. 5). These experiments are similar to those on the quantum Hall
effect in that they require, at least so far, low temperature and artificial
2D materials (quantum wells), but they differ in that no magnetic field
is needed.
Going 3D
The next important theoretical development, in 2006, was the
realization6–8 that even though the quantum Hall effect does not general-
ize to a genuinely 3D state, the topological insulator does, in a subtle way.
Although a 3D ‘weak’ topological insulator can be formed by layering
2D versions, similar to layered quantum Hall states, the resultant state
is not stable to disorder, and its physics is generally similar to that of the
2D state. In weak topological insulators, a dislocation (a line-like defect
© 2010 Macmillan Publishers Limited. All rights reserved
PERSPECTIVE INSIGHT
defined. If the topological invariants are always defined for an insulator,
then the surface must be metallic. b, The simplest example of a knotted 3D
electronic band structure (with two bands)35, known to mathematicians as
the Hopf map. The full topological structure would also have linked fibres
on each ring, in addition to the linking of rings shown here. The knotting
in real topological insulators is more complex as these require a minimum
of four electronic bands, but the surface structure that appears is relatively
simple (Fig. 3).
in the crystal) will always contain a quantum wire like that at the edge
of the quantum spin Hall effect (discussed earlier), which may allow 2D
topological insulator physics to be observed in a 3D material9.
There is also, however, a ‘strong’ topological insulator, which has a
more subtle relationship to the 2D case; the relationship is that in two
dimensions it is possible to connect ordinary insulators and topologi-
cal insulators smoothly by breaking time-reversal symmetry7. Such a
continuous interpolation can be used to build a 3D band structure that
respects time-reversal symmetry, is not layered and is topologically non-
trivial. It is this strong topological insulator that has protected metallic
surfaces and has been the focus of experimental activity.
Spin–orbit coupling is again required and must mix all components of
the spin. In other words, there is no way to obtain the 3D strong topologi-
cal insulator from separate spin-up and spin-down electrons, unlike in
the 2D case. Although this makes it difficult to picture the bulk physics of
the 3D topological insulator (only the strong topological insulator will be
discussed from this point), it is simple to picture its metallic surface6.
The unusual planar metal that forms at the surface of topological
insulators ‘inherits’ topological properties from the bulk insulator.
The simplest manifestation of this bulk–surface connection occurs at
a smooth surface, where momentum along the surface remains well
defined: each momentum along the surface has only a single spin state
at the Fermi level, and the spin direction rotates as the momentum
moves around the Fermi surface (Fig. 3). When disorder or impurities
are added at the surface, there will be scattering between these surface
states but, crucially, the topological properties of the bulk insulator do
not allow the metallic surface state to vanish — it cannot become local-
ized or gapped. These two theoretical predictions, about the electronic
structure of the surface state and the robustness to disorder of its metallic
behaviour, have led to a flood of experimental work on 3D topological
insulators in the past two years.
Experimental realizations
The first topological insulator to be discovered was the alloy BixSb1−x,
the unusual surface bands of which were mapped in an angle-resolved
photoemission spectroscopy (ARPES) experiment10,11. In ARPES exper-
iments, a high-energy photon is used to eject an electron from a crystal,
and then the surface or bulk electronic structure is determined from an
analysis of the momentum of the emitted electron. Although the surface
structure of this alloy was found to be complex, this work launched a
search for other topological insulators.
195
INSIGHT PERSPECTIVE
a Figure 3 shows the measured surface state of Bi2Se3 and a theoretical
e–
Integer quantum Hall state
(n = 1)
Ordinary insulator
b degeneracies was detected by scanning tunnelling microscopy experi-
Idealized quantum spin Hall state
Ordinary insulator
Figure 2 | Topological order in two dimensions. a, Edge of an integer
quantum Hall state. The electrons (e–) are confined to a 2D insulating
droplet with a metallic edge. Along the edge, electrons propagate only in
one direction, which is determined by the sign of the applied magnetic
field perpendicular to the droplet. One integer, n, the topological invariant,
determines the Hall conductance and the number of propagating edge
modes. b, Edge of an idealized quantum spin Hall state (that is, a 2D
topological insulator). Along the edge, spin-up electrons move clockwise,
whereas spin-down electrons move anticlockwise. Spin-up and spin-down
electrons are independent and are in oppositely directed quantum Hall
states. An applied electrical field generates a spin current but no charge
current. Each droplet is surrounded by an ordinary insulator.
For a topological insulator to form, spin–orbit coupling must be strong
enough to modify the electronic structure significantly, which suggests
that heavy-element, small-bandgap semiconductors are the most promis-
ing candidates. This suggestion stems from two points. First, spin–orbit
coupling is a relativistic effect and is only strong for heavy elements. Sec-
ond, if the bandgap is much larger than the energy scale of spin–orbit
coupling, then spin–orbit coupling will not be able to change the phase.
The search for topological insulators culminated in the recent discovery
of topological insulator behaviour in Bi2Se3 and Bi2Te3 (refs 12–14). These
‘next-generation’ materials both show topological insulator behaviour up
to higher temperatures than does the original material (BixSb1−x), with
bulk bandgaps of more than 0.1 eV, and have the simplest surface state
that is allowed. Beyond providing further confirmation of the theory
of topological insulators, the simplicity of the surface state in these
materials opens up the possibility of many experiments, some of which
are described below. Furthermore, the large bandgap means that these
experiments do not need to be carried out at extremely low temperatures.
The main remaining complication about these materials, especially when
using experimental techniques that do not distinguish directly between
bulk states and surface states (unlike ARPES), is that in the bulk state they
have residual conductivity arising from impurities.
196
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
A graphene lookalike
idealization of its state, including the electron spin. The surface state of
the next-generation topological insulators is closely related to the Dirac
electronic structure of graphene, which has a linear energy–momentum
relationship like that of a relativistic particle (and is known as a Dirac
cone). Graphene, which consists of a single layer of carbon atoms, has
been an extremely active subject of research in recent years15: it is inter-
esting both structurally, because it is the most 2D material possible, and
electronically, because of its linear energy–momentum relationship. The
main difference between the surface of a topological insulator and that
of graphene is that the topological insulator has only one Dirac point (or
valley) and no spin degeneracy, whereas graphene has two Dirac points
and is spin degenerate. This difference has far-reaching consequences,
including the possibility of generating new particles that have applica-
tions in quantum computing.
Another remarkable consequence of the absence of the extra
ments16–18: the interference patterns near defects or steps on the surface
show that electrons are never completely reflected when scattered. Even
if the disorder becomes much stronger and a description in terms of
well-separated scattering events is invalid, the surface remains metallic19.
This protection of the surface metal from Anderson localization (that
is, formation of an insulating state as a result of strong disorder20) is one
of the key differences between the surface of the topological insulator
and the ‘accidental’ surface states present in other materials, such as the
noble metals. Beyond just being stable to disorder, it is now understood
that the topological insulator, at least in some 2D models, forms as a
result of disorder21,22. In graphene, there is an approximate version of
this protection if the scattering has a very smooth potential, but real
graphene is likely to become localized with strong disorder.
There was, however, initially a disadvantage in using a topological
insulator for some purposes as opposed to graphene. In graphene, the
chemistry of the carbon atoms naturally locates the Fermi level at the
Dirac point (that is, the point at which the two cones intersect), where
the density of states vanishes. This means that the density of carriers in
graphene is highly tunable using an applied electrical field and allows
applications of graphene in both basic science and microelectronics. The
surface Fermi level of a topological insulator does not have any particu-
lar reason to sit at the Dirac point; however, through a combination of
surface and bulk chemical modification, tuning to the Dirac point in
Bi2Se3 was recently demonstrated23. This control of chemical potential is
important for applications, as well as for a proposal to create a topologi-
cal exciton condensate by biasing a thin film of topological insulator24.
This condensate is a superfluid state of electron–hole pairs that would
have a bound electronic state around superfluid vortices.
Materials challenges
Two recent advances in the nanostructuring of topological insulators
also deserve a mention. Nanoribbons of Bi2Se3 have been synthesized,
allowing observation of the Aharonov–Bohm effect in the metallic
surface state when a magnetic field is applied along the long direction
of the nanoribbon25, and molecular beam epitaxy has been used to
grow thin films of Bi2Se3 with controlled thickness down to a single
unit cell26. Such nanostructures are essential for many of the proposed
applications of topological insulators to spintronics and other fields.
For example, the spin structure in Fig. 3 means that a charge current
along the surface of the topological insulator automatically yields a
non-zero spin density; a heterostructure that combines a topologi-
cal insulator with a ferromagnet could allow the ferromagnet to be
switched by passing a current through the topological insulator’s sur-
face, and this would be a new type of spin torque device for magnetic
memory applications27. The next important steps are to carry out
direct transport and optical experiments on the metallic surface state
to measure its conductivity and spin properties, and for these experi-
ments improved materials with reduced residual bulk conductivity
might be required.
NATURE|Vol 464|11 March 2010
A nascent field
Several experiments that are under way are based on a different picture
of the essential properties of a topological insulator, a picture that is con-
nected to particle physics research from the 1980s. It has been known
for many years that the magnetic field and electric field are coupled in
the interior of some insulators. Topological insulators contain a special
type of such coupling that is quantized and was originally discussed by
particle physicists who were analysing the coupling of an axion particle
to ordinary electric and magnetic fields.
Particle physics and superfluidity
Condensed-matter physicists often seek to define a phase of matter
in terms of its response to some influence. For example, a solid is a
material with a non-zero stiffness in response to applied shear force,
and a superconductor is defined in terms of the Meissner effect (that
is, by its expulsion of applied magnetic fields). Such definitions are
connected directly to experimentation and are independent of details
on the microscopic level. It turns out that the defining response of a
topological insulator was almost worked out in the 1980s in an effort to
understand the properties of axion electrodynamics, which differs from
ordinary electrodynamics by the addition to the Lagrangian of a term
that is proportional to the scalar product of the electric field and the
magnetic field, E•B (ref. 28). There are only two values of the coefficient
of this term that are compatible with time-reversal symmetry, and these
correspond to ordinary insulators and topological insulators. Hence,
topological insulators are materials whose internal structure generates
the non-zero value of the axion-like coupling, in the same way that
insulators modify the dielectric constant, which is the coefficient of E2
in the Lagrangian29.
There are many predicted consequences of this term30,31, including
monopole-like behaviour and surface states with continuously vari-
able Hall conductivity. The most important for applications may be the
simple magnetoelectric effect in which an applied electrical field gen-
erates a magnetic dipole and vice versa. The potential advantage of the
magnetoelectric effect in topological insulators over that in multiferroic
materials32 is an increase not in strength but instead in speed and repro-
ducibility without fatigue, because this effect results purely from orbital
motion of the electrons. Whether a given material is a topological insu-
lator can be theoretically defined simply by asking how much its bulk
polarization changes in an applied magnetic field, and measurements
of this magnetoelectric polarizability in other materials suggest that it
a b
0.1 High
M ( M
0.0 SS
–0.1
EB (eV)
Low
–0.2
Surface
band
–0.3
Dirac
Bulk point
–0.4 bands
–0.15 0 0.15
ky (Å–1)
Figure 3 | Signatures of the exotic metallic surface states in topological
insulators. a, The electronic structure of Bi2Se3, as measured by
ARPES. Measured energy electron energy, EB, is plotted against electron
momentum, ky. High intensity (red and yellow areas) indicates a non-zero
electronic density of states. The surface bands crossing the bulk bandgap
–
enclose a single Dirac point at the Brillouin-zone centre (Γ), which is the
signature that this material is a topological insulator. —
M indicates the
centre of an edge of the Brillouin zone, and the path in the Brillouin zone
© 2010 Macmillan Publishers Limited. All rights reserved
PERSPECTIVE INSIGHT
should be possible to observe the effect in topological insulators.
Another recent theoretical advance was understanding how phases
similar to the 3D topological insulator might appear in systems with
different symmetries than those of the non-magnetic solids discussed
above33–35. Finding experimental examples of these other types of topo-
logical insulator and superconductor would be a major accomplish-
ment. Work in this area also led to a reconsideration of the properties of
superfluid 3He, in which neutral fermionic atoms form Cooper pairs: the
quasiparticles in this superfluid have some remarkable properties that
are closely connected to those of electrons in the topological insulator,
and finding direct experimental signatures of these quasiparticles is an
important challenge.
Emergent particles and quantum computing
The interface between a topological insulator and a superconductor
may allow the creation of an ‘emergent’ particle that neither material
supports by itself. Like any other metal, the metallic surface of a topo-
logical insulator becomes superconducting, by way of the proximity
effect, when placed next to an ordinary superconductor. However, the
superconducting surface of a topological insulator is novel: if a vortex
line runs from the superconductor into the topological insulator, then
a zero-energy Majorana fermion is trapped in the vicinity of the vor-
tex core36,37. Unlike vortex core states in ordinary superconductors, the
Majorana fermion has quantum numbers that differ from those of an
ordinary electron: it is its own antiparticle, is electrically neutral and
is, in most respects, ‘half ’ of an ordinary electron38. There are several
reasons for the excitement about this proposal and others of a simi-
lar spirit39. First, these proposals may allow a Majorana fermion to be
observed directly, a long-sought goal in particle physics and condensed-
matter physics. Second, Majorana fermions are one step towards a topo-
logical quantum computer, which would be exceptionally well protected
from errors40 because the quasiparticles obey a special kind of quantum
statistics that is non-Abelian (for a more in-depth discussion of non-
Abelian states of matter, see page 187). More complex materials might
even allow Majorana fermions as point excitations in the bulk of the
3D material41.
More generally, there are several aspects of the topological order in
quantum Hall states that have not yet been realized in topological insu-
lators, and the most important is the emergence of quasiparticles with
modified charge and statistics. The traditional wellspring of topological
order, the 2D electron gas in a strong magnetic field, has not dried up;
E
ky
kx
EF
ky
kx
Bulk
bandgap
is indicated by white arrows. The direction of electron spin is indicated by
blue arrows. (Panel modified, with permission, from ref. 12; data taken
from refs 12 and 23.) b, Theoretical idealization of the electronic structure
of Bi2Se3, showing the rotation of the spin degree of freedom (red arrows)
as an electron (with energy E) moves around the Fermi surface (with
Fermi energy EF). Scattering of the surface electrons by non-magnetic
disorder will modify the details of the electronic wavefunctions but will not
eliminate the metallic surface.
197
INSIGHT PERSPECTIVE
recent results from studies of this gas include the first observation of
the fractional quantum Hall effect in graphene and suggest the possible
observation of non-Abelian statistics in interferometry experiments.
Topological states have also been proposed to occur in frustrated spin
systems (such as those discussed on page 201). The variety of topological
phenomena observed in semiconductor heterostructures over the past
three decades suggests that the topological insulator materials that have
been uncovered so far are just the first examples and that new types of
topological order still await discovery.
Looking forward
The recent discovery of topological insulators, like other advances in
basic condensed-matter physics, allows new applications that build on
our new understanding. The unusual metallic surfaces of these insula-
tors may result in new spintronic or magnetoelectric devices. Further-
more, in combination with superconductors, topological insulators
could lead to a new architecture for topological quantum bits. These
insulators have already had a considerable impact on ‘pure’ condensed-
matter physics, making it clear that topological effects — which were
long thought to be restricted to low temperatures, reduced dimensional-
ity and high magnetic field — can determine the physics of seemingly
ordinary bulk materials under ordinary conditions. ■ 28. Wilczek, F. Two applications of axion electrodynamics. Phys. Rev. Lett. 58, 1799–1802
1. Haldane, F. D. M. Model for a quantum Hall effect without Landau levels: condensed-
matter realization of the ‘parity anomaly’. Phys. Rev. Lett. 61, 2015–2018 (1988).
2. Murakami, S., Nagaosa, N. & Zhang, S.-C. Spin-Hall insulator. Phys. Rev. Lett. 93, 156804
(2004).
3. Kane, C. L. & Mele, E. J. Z2 topological order and the quantum spin Hall effect. Phys. Rev.
Lett. 95, 146802 (2005).
This paper explains the theoretical requirements for a non-magnetic material to be a
2D topological insulator, with a quantum spin Hall effect.
4. Bernevig, B. A., Hughes, T. L. & Zhang, S.-C. Quantum spin Hall effect and topological
phase transition in HgTe quantum wells. Science 314, 1757–1761 (2006).
5. König, M. et al. Quantum spin Hall insulator state in HgTe quantum wells. Science 318,
766–770 (2007).
This paper reports the first experimental observation of a 2D topological insulator that
has a quantum spin Hall effect.
6. Fu, L., Kane, C. L. & Mele, E. J. Topological insulators in three dimensions. Phys. Rev. Lett. 98,
106803 (2007).
7. Moore, J. E. & Balents, L. Topological invariants of time-reversal-invariant band structures.
Phys. Rev. B 75, 121306(R) (2007).
8. Roy, R. Topological phases and the quantum spin Hall effect in three dimensions.
Phys. Rev. B 79, 195322 (2009).
9. Ran, Y., Zhang, Y. & Vishwanath, A. One-dimensional topologically protected modes in
topological insulators with lattice dislocations. Nature Phys. 5, 298–303 (2009).
10. Hsieh, D. et al. A topological Dirac insulator in a quantum spin Hall phase. Nature 452,
970–974 (2008).
By using ARPES experiments, this study observed a 3D topological insulator, the
theoretical predictions for which were made in refs 6–8.
11. Hsieh, D. et al. Observation of unconventional quantum spin textures in topological
insulators. Science 323, 919–922 (2009).
12. Xia, Y. et al. Observation of a large-gap topological-insulator class with a single Dirac cone
on the surface. Nature Phys. 5, 398–402 (2009).
13. Zhang, H. et al. Topological insulators in Bi2Se3, Bi2Te3 and Sb2Te3 with a single Dirac cone
on the surface. Nature Phys. 5, 438–442 (2009).
References 12 and 13 report experiments and theory on next-generation topological
insulator materials, which have a large bandgap and a single surface Dirac cone; these are
the most promising materials for future experiments.
198
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
14. Chen, Y. L. et al. Experimental realization of a three-dimensional topological insulator,
Bi2Te3. Science 325, 178–181 (2009).
15. Castro Neto, A. H., Guinea, F., Peres, N. M., Novoselov, K. S. & Geim, A. K. The electronic
properties of graphene. Rev. Mod. Phys. 81, 109–163 (2009).
16. Roushan, P. et al. Topological surface states protected from backscattering by chiral spin
texture. Nature 460, 1106–1109 (2009).
17. Alpichshev, Z. et al. STM imaging of electronic waves on the surface of Bi2Te3:
topologically protected surface states and hexagonal warping effects. Preprint at 〈http://
arxiv.org/abs/0908.0371〉 (2009).
18. Zhang, T. et al. Experimental demonstration of the topological surface states protected by
the time-reversal symmetry. Preprint at 〈http://arxiv.org/abs/0908.4136〉 (2009).
19. Nomura, K., Koshino, M. & Ryu, S. Topological delocalization of two-dimensional
massless Dirac fermions. Phys. Rev. Lett. 99, 146806 (2007).
20. Anderson, P. W. Absence of diffusion in certain random lattices. Phys. Rev. 109, 1492–1505
(1958).
21. Li, J. et al. Topological Anderson insulator. Phys. Rev. Lett. 102, 136806 (2009).
22. Groth, C. W. et al. Theory of the topological Anderson insulator. Phys. Rev. Lett. 103,
196805 (2009).
23. Hsieh, D. et al. A tunable topological insulator in the spin helical Dirac transport regime.
Nature 460, 1101–1105 (2009).
24. Seradjeh, B., Moore, J. E. & Franz, M. Exciton condensation and charge fractionalization in
a topological insulator film. Phys. Rev. Lett. 103, 066402 (2009).
25. Peng, H. et al. Aharonov–Bohm interference in topological insulator nanoribbons. Preprint
at 〈http://arxiv.org/abs/0908.3314〉 (2009).
26. Zhang, Y. et al. Crossover of three-dimensional topological insulator of Bi2Se3 to the two-
dimensional limit. Preprint at 〈http://arxiv.org/abs/0911.3706〉 (2009).
27. Garate, I. & Franz, M. Inverse spin-galvanic effect in a topological-insulator/ferromagnet
interface. Preprint at 〈http://arxiv.org/abs/0911.0106〉 (2009).
(1987).
29. Qi, X.-L., Hughes, T. L. & Zhang, S.-C. Topological field theory of time-reversal invariant
insulators. Phys. Rev. B 78, 195424 (2008).
30. Qi, X.-L. et al. Inducing a magnetic monopole with topological surface states. Science 323,
1184–1187 (2009).
31. Essin, A. M., Moore, J. E. & Vanderbilt, D. Magnetoelectric polarizability and axion
electrodynamics in crystalline insulators. Phys. Rev. Lett. 102, 146805 (2009).
32. Ramesh, R. & Spaldin, N. A. Multiferroics: progress and prospects. Nature Mater. 6, 21–27
(2007).
33. Schnyder, A. P. et al. Classification of topological insulators and superconductors in three
spatial dimensions. Phys. Rev. B 78, 195125 (2008).
34. Kitaev, A. Periodic table for topological insulators and superconductors. Preprint at
〈http://arxiv.org/abs/0901.2686〉 (2009).
35. Moore, J. E., Ran, Y. & Wen, X.-G. Topological surface states in three-dimensional
magnetic insulators. Phys. Rev. Lett. 101, 186805 (2008).
36. Fu, L. & Kane, C. L. Superconducting proximity effect and Majorana fermions at the
surface of a topological insulator. Phys. Rev. Lett. 100, 096407 (2008).
37. Jackiw, R. & Rossi, P. Zero modes of the vortex-fermion system. Nucl. Phys. B 190, 681–691
(1980).
38. Wilczek, F. Majorana returns. Nature Physics 5, 614–618 (2009).
39. Nilsson, J., Akhmerov, A. R. & Beenakker, C. W. Splitting of a Cooper pair by a pair of
Majorana bound states. Phys. Rev. Lett. 101, 120403 (2008).
40. Collins, G. P. Computing with quantum knots. Sci. Am. 294, 57–63 (2006).
41. Teo, J. & Kane, C. L. Majorana fermions and non-Abelian statistics in three dimensions.
Phys. Rev. Lett. 104, 046401 (2009).
Acknowledgements I have benefited from conversations about topological
insulators with L. Balents, B. A. Bernevig, A. Essin, M. Franz, D. Haldane, Z. Hasan,
C. Kane, D.-H. Lee, A. Ludwig, L. Molenkamp, S. Ryu, D. Vanderbilt, A. Vishwanath,
X.-G. Wen, C. Xu and S.-C. Zhang. My work on topological insulators is supported
by the US National Science Foundation.
Author Information Reprints and permissions information is available at www.
nature.com/reprints. The author declares no competing financial interests.
Correspondence should be addressed to the author (jemoore@berkeley.edu).
NATURE|Vol 464|11 March 2010|doi:10.1038/nature08917 REVIEW INSIGHT

Spin liquids in frustrated magnets

Leon Balents1

Frustrated magnets are materials in which localized magnetic moments, or spins, interact through competing
exchange interactions that cannot be simultaneously satisfied, giving rise to a large degeneracy of the
system ground state. Under certain conditions, this can lead to the formation of fluid-like states of matter,
so-called spin liquids, in which the constituent spins are highly correlated but still fluctuate strongly down to
a temperature of absolute zero. The fluctuations of the spins in a spin liquid can be classical or quantum and
show remarkable collective phenomena such as emergent gauge fields and fractional particle excitations. This
exotic behaviour is now being uncovered in the laboratory, providing insight into the properties of spin liquids
and challenges to the theoretical description of these materials.

Sometimes, a little frustration can make life interesting. This is true in
physics, where frustration often leads to exotic properties of materi-
als. In physics, frustration refers to the presence of competing forces
that cannot be simultaneously satisfied. The concept has been applied
broadly, from magnetism, which I discuss here, to negative thermal
expansion of solids1 and to soft materials2. Studies of frustration began
with antiferromagnets, in which frustration usually has a simple geo-
metric origin. Such geometric frustration occurs in systems of spins
on lattices that involve triangular motifs (Fig. 1), in which the nearest-
neighbour interactions favour anti-aligned spins (Box 1). On a triangle,
the three spins cannot all be antiparallel. Instead, depending on the
circumstances, the spins fluctuate, or order, in a less obvious manner.
An antiferromagnet consisting of a two-dimensional (2D) triangular
lattice of Ising spins (which point either upward or downward along
a single axis) provided one of the prototypes of frustration. Wannier
showed in 1950 that this model has a very large degeneracy of ground
states3. Now, such degeneracy is considered to be a key, or even defining,
characteristic of frustration. In the triangular Ising antiferromagnet, the
ground-state entropy is extensive and equals 0.323kBN, where kB is the
Boltzmann constant and N is the number of spins. At low temperatures,
the spins continue to fluctuate thermally, although in a correlated man-
ner because they are restricted to the ground states of the Ising anti-
ferromagnet. By analogy to an ordinary liquid, in which the molecules
form a dense, highly correlated state that has no static order, the spins in
the triangular Ising antiferromagnet form a ‘spin liquid’, or cooperative
paramagnet. The ‘frustration’ (or ‘fluctuation’) parameter, f (Box 1), pro-
vides a quantitative measure of the depth of the spin-liquid regime; f = ∞
if the spins remain liquid down to a temperature, T, of absolute zero.
Fluctuations of the spins in a spin liquid can be classical or quantum.
In quantum mechanics, the magnitude of a spin is quantized in half-
integer units of " (where " denotes Planck’s constant divided by 2π),
the quantum of angular momentum. Classical fluctuations dominate
for large spins (those with a size, S, much larger than the minimum
size of ½) and are driven by thermal energy. Spins can be thought of as
reorienting randomly with time, cycling through different microstates.
When the energy kBT becomes too small, classical fluctuations cease and
the spins either freeze or order. For small spins, with S comparable to
½, the quantum mechanical uncertainty principle produces zero-point
motions comparable to the size of the spin itself, which persist down to
T = 0 K. Although they are similar to thermal fluctuations in some ways,
quantum fluctuations can be phase coherent. If they are strong enough,
1
Kavli Institute for Theoretical Physics, University of California, Santa Barbara, Santa Barbara, California 93106, USA.
the result is a quantum spin liquid (QSL), which is a superposition state
in which the spins simultaneously point in many different directions. In
a QSL, the spins are highly entangled with one another in a subtle way
that involves entanglement between widely separated spins. QSLs are
more elusive than their classical counterparts but offer a much greater
conceptual pay-off. From theory, it is predicted that they should show
extremely exotic phenomena such as hosting exotic excitations with
fractional quantum numbers and artificial gauge fields. They may also
be associated with exotic forms of superconductivity4. Researchers are
seeking QSL states in solid-state materials known as Mott insulators (in
which electrons are localized to individual atomic or molecular orbit-
als but maintain their spin degree of freedom) and recently in artificial
frustrated lattices that have been created optically for ultracold atoms.
In this Review, I survey how the discovery of new materials and
improved experimental probes, together with complementary advances
in theory, have reinvigorated the study of spin liquids and frustrated mag-
netism in general. I begin with the best-understood example of spin ice,
a classical spin liquid, whose properties are far better understood after
progress made in the past year. I then turn to quantum frustrated systems.
After describing their basic physics, I discuss the experimental situation
and its challenges. I uncover some unexpected physics in ‘failures’ to find
QSLs. Finally, I discuss promising directions for QSL research.
Spin ice
To begin the survey of these developments, I start with a simple exam-
ple, in which the spins can be regarded as classical: spin-ice materials5,6,
which were first studied by Harris and colleagues in 1997 (ref. 7). In spin
ices — for example Dy2Ti2O7, Ho2Ti2O7 and Ho2Sn2O7 — only the ions
of the rare-earth atoms (Dy and Ho) are magnetic, and these reside on
a pyrochlore lattice, which is a network of corner-sharing tetrahedra
(Fig. 1c). Their f-electron spins are large and classical, and they behave
as Ising doublets aligned with the local 〈111〉 axis, which connects the
centres of the two tetrahedra shared by that spin. Their interactions are
predominantly long range and dipolar, but much of their physics can be
understood solely from an effective nearest-neighbour exchange energy,
Jeff, which is ferromagnetic8,9 (Box 2). This ferromagnetic interaction
is frustrated, owing to the different Ising axes of the spins. Indeed,
the states that minimize the energy of a single tetrahedron are highly
degenerate, consisting of all six configurations in which two spins point
inward and two spins point outward from the centre of the tetrahedron
(Fig. 2a). The name spin ice originates from a direct analogy between

199
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
these configurations and the positions of protons in the tetrahedrally
coordinated O2− framework of water ice10.
When kBT << Jeff, the system fluctuates almost entirely within the
two-in and two-out manifold of states. It turns out that the number of
such states is exponentially large, so a low-temperature entropy remains
even within this limit. This entropy, first estimated by Pauling in 1935
(ref. 11), has been measured in spin ice10. Although the spins remain
paramagnetic in this regime, the ‘ice rules’ imply strong correlations:
for instance, if it is known that two spins on a tetrahedron are pointing
out, then the other two spins must point in. The correlated paramagnet
is a simple example of a classical spin liquid. A key question is whether
the local constraints have long-range consequences: is the spin liquid
qualitatively distinguishable from an ordinary paramagnet? Interest-
ingly, the answer for spin ice is ‘yes’.
Analogies to electromagnetism
To understand how long-range effects arise from a local constraint,
it is helpful to use an analogy to electromagnetism. Each spin can be
thought of as an arrow pointing between the centres of two tetrahedra
(Fig. 2b). This defines a vector field of flux lines on the lattice, which
a At the pinch points, if the ice rules are obeyed perfectly, a sharp singular-
J
J’ J’
b
J
J’ J’
c must be flipped, starting from the tetrahedron in question (Fig. 2c).
Figure 1 | Frustrated magnetism on 2D and 3D lattices. Two types of 2D
lattice are depicted: a triangular lattice (a) and a kagomé lattice (b). The 3D
lattice depicted is a pyrochlore lattice (c). In experimental materials, the
three-fold rotational symmetry of the triangular and kagomé lattices may
not be perfect, allowing different exchange interactions, J and Jʹ, on the
horizontal and diagonal bonds, as shown. Blue circles denote magnetic ions,
arrows indicate the direction of spin and black lines indicate the shape of the
lattice. In b, ions and spins are depicted on only part of the illustrated lattice.
200
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
because of the two-in, two-out rule is divergence free. In this sense, the
vectors define an ‘artificial’ magnetic field, b, on the lattice (the field can
be taken to have unit magnitude on each link, with the sign determined
by the arrows). Because the spins are not ordered but fluctuating, the
magnetic field also fluctuates. However, because the magnetic field lines
do not start or end, these fluctuations include long loops of flux (Fig. 2b)
that communicate spin correlations over long distances.
The nature of the long-distance spin correlations was derived by Young-
blood and colleagues in a mathematically analogous model of a fluctuating
ferroelectric, in which the electric polarization is similarly divergence-
less12; this was subsequently rederived for spin ice13. The result is that the
artificial magnetic field, at long wavelengths, fluctuates in equilibrium just
as a real magnetic field would in a vacuum, albeit with an effective mag-
netic permeability. For the spins in spin ice, this implies power-law ‘dipo-
lar’ correlations that are anisotropic in spin space and decay as a power
law (~1/r3, where r is the distance between the spins) in real space. It is
remarkable to have power-law correlations without any broken symmetry
and away from a critical point. After Fourier transformation of these cor-
relations on the lattice, a static spin structure factor with ‘pinch points’ at
reciprocal lattice vectors in momentum space is obtained12–14.
Such dipolar correlations have recently been observed in high
resolution neutron-scattering experiments by Fennell and colleagues15.
ity is expected, as well as a precise vanishing of the scattering intensity
along lines passing through the reciprocal lattice vectors. The rounding
of this singularity gives a measure of the ‘spin-ice correlation length’,
which is estimated to grow to 2–300 Å (a large number) at a temperature
of 1.3 K. In the future, it may be interesting to see how this structure
changes at even lower temperatures, at which spin ices are known to
freeze and fall out of equilibrium. Although the argument of Young-
blood and colleagues12 and the model outlined above rely on equilib-
rium, arguments by Henley suggest that the pinch points could persist
even in a randomly frozen glassy state14.
Magnetic monopoles
Interestingly, the magnetostatic analogy goes beyond the equilibrium
spin correlations. One of the most exciting recent developments in this
area has been the discovery of magnetic monopoles in spin ice16. These
arise for simple microscopic reasons. Even when kBT << Jeff, violations of
the two-in, two-out rule occur, although they are costly in energy and,
hence, rare. The simplest such defect consists of a single tetrahedron
with three spins pointing in and one pointing out, or vice versa (Fig. 2c).
This requires an energy of 2Jeff relative to the ground states. From a
magnetic viewpoint, the centre of this tetrahedron becomes a source or
sink for flux, that is, a magnetic monopole. A monopole is a somewhat
non-local object: to create a monopole, a semi-infinite ‘string’ of spins
Nevertheless, when it has been created, the monopole can move by
single spin flips without energy cost, at least when only the dominant
nearest-neighbour exchange, Jeff, is considered.
Remarkably, the name monopole is physically apt: this defect carries
a real ‘magnetic charge’16. This is readily seen because the physical mag-
netic moment of the rare-earth atom is proportional to the pseudo-
magnetic field, M = gμBb, where g is the Landé g factor and μB is the Bohr
magneton. Thus, a monopole with the non-zero divergence =•b also
has a non-zero =•M. The actual magnetic charge (which measures the
strength of the Coulomb interaction between two monopoles) is, how-
ever, small: at the same distance, the magnetostatic force between two
monopoles is approximately 14,000 times weaker than the electrostatic
force between two electrons. Nevertheless, at low temperatures, this is
still a measurable effect.
A flood of recent papers have identified clear signatures of magnetic
monopoles in new experiments and in previously published data. Jaubert
and Holdsworth showed that the energy of a monopole can be extracted
from the Arrhenius behaviour of the magnetic relaxation rate17. They
found that in Dy2Ti2O7 the energy of a monopole is half that of a single
spin flip, reflecting the fractional character of the magnetic monopoles.
NATURE|Vol 464|11 March 2010
They also identified deviations from the Arrhenius form at lower tem-
peratures as a result of Coulomb interactions between monopoles. In
an intriguing paper, Bramwell and colleagues applied an old theory put
forward by Onsager18 for the electric-field dependence of thermal charge
dissociation in electrolytes — the Wien effect — to the magnetic analogue
in spin ice, that is, to the disassociation of monopole–antimonopole pairs
with a magnetic field19. The theory allows an extraction of the absolute
value of the magnetic charge of a monopole from the dependence of the
magnetic relaxation rate on the magnetic field. The authors measured
muon spin relaxation in Dy2Ti2O7 to obtain the rate, and from this they
extracted a magnetic charge in near perfect agreement with theoretical
expectations. In addition to these quantitative measures of the energy
and charge of a monopole, two recent papers presented neutron-scat-
tering measurements in magnetic fields, interpreting them as evidence
of monopoles and the ‘strings’ emanating from monopoles20,21. Many
more experiments in which the monopoles in spin ice are studied and
manipulated are likely to emerge soon.
Quantum spin liquids
In spin ice, as the temperature is lowered, the spins themselves fluctuate
ever more slowly, eventually falling out of equilibrium and freezing
below about 0.5 K (from theory, it is predicted that, in equilibrium,
the spins should order at T = 0.1−0.2 K (ref. 6)). This is a consequence
of the large energy barriers between different ice-rule configurations,
which require the flipping of at least six spins, and the weak quantum
amplitude for such large spins to cooperatively tunnel through these
barriers. By contrast, for materials with spins of S = ½ and approximate
Heisenberg symmetry, quantum effects are strong, and there are no
obvious energy barriers. I now turn to such materials in the search for
a QSL ground state, in which spins continue to fluctuate and evade
order even at T = 0 K. Such a QSL is a strange beast: it has a non-mag-
netic ground state that is built from well-formed local moments. The
theoretical possibility of QSLs has been hotly debated since Anderson
proposed them in 1973 (ref. 22).
Wavefunctions and exotic excitations
A natural building block for non-magnetic states is the valence bond,
a pair of spins that, owing to an antiferromagnetic interaction, forms a
spin-0 singlet state (Fig. 3a). A valence bond is a highly quantum object,
the two spins being maximally entangled and non-classical. If all of the
spins in a system are part of valence bonds, then the full ground state has
spin 0 and is non-magnetic. One way in which this can occur is by the
partitioning of all of the spins into specific valence bonds, which are static
and localized. Mathematically, such a ground state is well approximated as
a product of the valence bonds, so that each spin is highly entangled with
only one other, its valence-bond partner. This is known as a valence-bond
solid (VBS) state (Fig. 3a) and occurs in several materials23–25. VBS states
are interesting because, for instance, they provide an experimental way
of studying Bose–Einstein condensation of magnons (which are triplet
excitations of the singlet valence bonds) in the solid state26.
A VBS state is not, however, a true QSL, because it typically breaks
lattice symmetries (because the arrangement of valence bonds is not
unique) and lacks long-range entanglement. To build a QSL, the valence
bonds must be allowed to undergo quantum mechanical fluctuations.
The ground state is then a superposition of different partitionings of
spins into valence bonds (Fig. 3b, c). If the distribution of these partition-
ings is broad, then there is no preference for any specific valence bond
and the state can be regarded as a valence-bond ‘liquid’ rather than a
solid. This type of wavefunction is generally called a resonating valence-
bond (RVB) state22. RVB states became subjects of intense theoretical
interest when, in 1987, Anderson proposed that they might underlie the
physics of high-temperature superconductivity4. Only relatively recently,
however, have RVB wavefunctions been shown to be ground states of
many specific model Hamiltonians27–32.
It is now understood that not all QSLs are alike. Generally, different
states have different weights of each valence-bond partition in their
wavefunctions. The valence bonds need not be formed only from nearby
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
Box 1 | Elements of frustration
A triangle of antiferromagnetically interacting Ising spins, which must
point upward or downward, is the simplest example of frustration. All
three spins cannot be antiparallel. As a result, instead of the two ground
states mandated by the Ising symmetry (up and down), there are six
ground states (see figure; blue circles denote magnetic ions, arrows
indicate the direction of spin, and black and red lines indicate the shape
of the triangular lattice, with red lines denoting the axis on which the
spins are parallel). On 2D and 3D lattices, such degeneracies can persist.
When they do, fluctuations are enhanced and ordering is suppressed.
On the basis of this fact, Ramirez introduced a simple empirical measure
of frustration that has become widely used50. At high temperatures, the
spin (or magnetic) susceptibility of a local-moment magnet generally
has a Curie–Weiss form, χ≈ C/(T − ΘCW), where T is temperature and
C is the Curie constant. This allows extraction of the Curie–Weiss
temperature, ΘCW, from a plot of 1/χ versus T. |ΘCW| provides a natural
estimate for the strength of magnetic interactions (ΘCW<0 for an
antiferromagnet) and sets the scale for magnetic ordering in an
unfrustrated material. By comparing the Curie–Weiss temperature with
the temperature at which order freezes, Tc, the frustration parameter, f, is
obtained: f = |ΘCW|/Tc. Typically, f>5–10 indicates a strong suppression
of ordering, as a result of frustration. For such values of f, the temperature
range Tc<T<|ΘCW| defines the spin-liquid regime.
spins33. If a valence bond is formed from spins that are far apart, then
those spins are only weakly bound into a singlet and the valence bond can
be ‘broken’ to form free spins with relatively little energy. So states that
have a significant weight from long-range valence bonds have more low-
energy spin excitations than states in which the valence bonds are mainly
short range (see, for example, ref. 34). There are also other excitations that
do not break the valence bonds but simply rearrange them. Such excita-
tions can have low energy even in short-range RVB states.
Given the possibility of different QSL states, it is interesting to attempt
to classify these states. This problem is only partly solved, but it is clear
that the number of possible states is huge, if not infinite. For instance,
Wen has classified hundreds of different ‘symmetrical spin liquids’
(QSLs with the full space-group symmetry) for S = ½ antiferromagnets
on the square lattice35. Finding the correct QSL ground state among all
of the many possible RVB phases, many of which have similar energies,
is a challenge to theory, reminiscent of the landscape problem in string
theory, in high-energy physics.
Even with this diversity of possible states, one feature that theorists
expect QSLs to have in common is that they support exotic excitations.
What is meant by exotic? In most phases of matter, all of the excitations
can be constructed from elementary excitations that are either electron-
like (spin S = ½ and charge ±e) or magnon-like (spin S = 1 and charge
neutral). Only in rare examples, such as the fractional quantum Hall
states, do the elementary excitations differ from these, in this case by
carrying fractional quantum numbers. The magnetic monopoles in spin
201
INSIGHT REVIEW
Box 2 | Theoretical frameworks
Theoretical models of magnets with localized electrons typically start
with the Heisenberg Hamiltonian, for example:
1
H=− —
2
∑J S •S
ij
ij i j (1)
where Jij are exchange constants, and Si and Sj are quantum spin-S
operators. They obey Si2 = S(S + 1) and canonical commutation relations
[Siμ, Sjν] = iєμνλSiλδij, where єμνλ is the Levi-Civita symbol. In spin ice, owing
to the large S (S=15/2 for Dy and 8 for Ho), the spins can be regarded
simply as classical vectors (conventionally they are taken to be of unit
length, with factors of S absorbed into the exchange constants). For
spin ice, equation (1) above mimics the short-range effect of dipolar
exchange, by taking ferromagnetic Jij = Jeff > 0 on nearest neighbours,
as is discussed in the main text. In addition, in spin ice, strong Ising
anisotropy must be included by way of the crystal field term
Hcf = −D ∑ (S • n̂ )
i
i i
2
where n̂i is a unit vector along the Ising 〈111〉 axis of the ith spin and D is
the Ising anisotropy. The positive D >> J requires the spins to point along
this axis. In low-spin systems of transition metal spins, the exchange is
antiferromagnetic, Jij < 0, crystal field effects are much less significant
and S is much smaller. For the QSL candidates discussed in the main
text, Stakes its minimum value,½. Thus, when the assumption of
localized electrons is valid, equation (1) itself is a good starting point. If
this assumption is not valid, a better description, which includes charge
fluctuations, is the Hubbard model:
HHubbard = − ∑t c c
ij,α
†
ij iα jα ∑
+ U ni(ni − 1)
i
where c†iα and ciα are creation and annihilation operators, respectively,
for electrons with spin α = ±½ on site i, and ni = ∑α c†iαciα is the number
of electrons on the site. The electron operators obey canonical anti-
commutation relations {c†iα, cjβ} = δαβδij. The coefficient tij gives the
quantum amplitude for an electron to hop from site j to site i, and U is
the Coulomb energy cost for two electrons to occupy the same site.
When U>> t, and there is on average one electron per site, charge
fluctuations become negligible. Then the Hubbard model reduces to the
Heisenberg one, with Jij≈ −4tij2/U. However, when U/t is not very large,
the two models are inequivalent. For a small enough U/t, the Hubbard
model usually describes a metal, in which electrons are completely
delocalized. As U/t is increased from zero, there is a quantum phase
transition from a metal to an insulator; this is known as the Mott
transition. For U/t larger than this critical value, the system is known as
a Mott insulator. ‘Weak’ Mott insulators are materials with U/t close to
the Mott transition.
ice are examples of this: the elementary magnetic dipole ‘splits’ into a
monopole–antimonopole pair. However, these are classical excitations
and not true coherent quasiparticles. In QSLs, the most prominent exotic
excitations are in the form of spinons, which are neutral and carry spin
S = ½ (Fig. 4). Spinons are well established in one-dimensional (1D) sys-
tems, in which they occur as domain walls (Fig. 4a). A spinon can thus
be created similarly to a monopole in spin ice, by flipping a semi-infinite
string of spins. A key difference, however, is that in one dimension the
only boundary of such a string is its end point, so the string is guar-
anteed to cost only a finite energy from this boundary. By contrast, in
two or three dimensions, the boundary of a string extends along its full
length. A string would naturally be expected to have a tension (that
is, there is an energy cost proportional to its length). String tension
represents confinement of the exotic particle, as occurs for quarks in
quantum chromodynamics. This is avoided in spin ice by the special
form of the nearest-neighbour Hamiltonian. However, when spin ice
202
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
is in equilibrium, corrections to this form would be expected to lead to
monopole confinement at low temperatures.
In a true 2D or 3D QSL, the string associated with a spinon remains
robustly tensionless even at T = 0 K, owing to strong quantum fluctua-
tions (Fig. 4c). This can be understood from the quantum superposition
principle: rearranging the spins along the string simply reshuffles the
various spin or valence-bond configurations that are already super-
posed in the ground state. Detailed studies of QSL states have shown
that higher-dimensional spinons can have varied character. They may
obey Fermi–Dirac36, Bose–Einstein37 or even anyonic statistics (see page
187). They may be gapped (that is, require a non-zero energy to excite)
or gapless, or they may even be so strongly interacting that there are no
sharp excitations of any kind38.
Experimental search for QSLs
In contrast to the apparent ubiquity and variety of QSLs in theory, the
experimental search for QSLs has proved challenging. Most Mott insu-
lators — materials with localized spins — are ordered magnetically or
freeze into spin-glass states. A much smaller set of Mott insulators form
VBS states23–25. A clear identification of RVB-like QSL states has proved
more elusive.
How is a QSL identified in experiments? One good indication is a
large frustration parameter, f > 100–1,000 (f may be limited by material
complications, such as defects or weak symmetry-breaking interac-
tions, or the ability to cool). A more stringent test is the absence of static
moments, even disordered ones, at low temperatures, a feature that can
be probed by nuclear magnetic resonance (NMR) and muon spin reso-
nance experiments. Specific-heat measurements give information on the
low-energy density of states of a putative QSL, which can be compared
with those of theoretical models. Thermal transport can determine
whether these excitations are localized or itinerant. Elastic and inelastic
neutron-scattering measurements, especially on single crystals, provide
crucial information on the nature of correlations and excitations, and
these could perhaps uncover spinons. All told, this is a powerful arsenal
of experimental tools, but the task is extremely challenging. At the heart
of the problem is that there is no single experimental feature that identi-
fies a spin-liquid state. As long as a spin liquid is characterized by what it
is not — a symmetry-broken state with conventional order — it will be
much more difficult to identify conclusively in experiments.
Nevertheless, in recent years, increased activity in the field of highly
frustrated magnetism has led to a marked increase in the number of
candidate QSL materials, giving reason for optimism39,40. A partial list
of the materials that have been studied is presented in Table 1.
QSL materials
The last two entries in Table 1 (Cs2CuCl4 and FeSc2S4) have already
been ruled out as true QSLs, and I return to these later (in the section
‘Unexpected findings’). All of the remaining materials are promising
candidates, with S = ½ quantum spins on frustrated lattices. They show
persistent spin dynamics down to the lowest measured temperatures,
with every indication that the dynamics persist down to T = 0 K. With
the exception of Cu3V2O7(OH)2•2H2O, no signs of phase transitions
are seen on cooling from the high-temperature paramagnet to a low
temperature. The characteristics of these materials are now beginning
to be understood.
Surveying them, it is clear that there is a variety of structures and
properties. There are organic compounds — κ-(BEDT-TTF)2Cu2(CN)3
and EtMe3Sb[Pd(dmit)2]2 — and a variety of inorganic compounds. The
spins form several different structures: 2D triangular lattices; 2D kagomé
lattices; and a hyperkagomé lattice, which is 3D. In some cases, these
are ideal, isotropic forms. In other cases, there are asymmetries and,
consequently, spatial anisotropy. Existing samples are compromised
by varying degrees of disorder, from as much as 5–10% free defect
spins and a similar concentration of spin vacancies in ZnCu3(OH)6Cl2
(ref. 41) to as little as 0.07% impurity in recently improved samples of
Cu3V2O7(OH)2•2H2O (ref. 42). In several materials, the nature and
degree of disorder have not been well characterized.
NATURE|Vol 464|11 March 2010 REVIEW INSIGHT

A key physical distinction among the candidates is their degree of a

charge localization — this is important for understanding the mech-
anism of the spin-liquid behaviour. The kagomé materials can be
regarded as strong Mott insulators, in which the electrons constituting
the spins are approximately localized to a single atomic orbital. This can
be seen from the small exchange couplings (proportional to the Curie–
Weiss temperature, ΘCW) relative to the Mott charge gap, which is of the
order of electronvolts (eV) in such copper oxides. For these materials, a
microscopic description purely in terms of Heisenberg spins is justified.
By contrast, for the organic compounds and Na4Ir3O8, the exchange
energies are larger and the charge gaps are smaller. These materials are
therefore weak Mott insulators, in which the electronic spins are much b
less well localized, so charge fluctuations may have an important role.
In the organic compounds, the Hubbard repulsion, U (Box 2), which
sets the Mott gap, is much smaller than in inorganic compounds, owing
to the extended nature of the molecular orbital constituting each spin.
In Na4Ir3O8, U is likewise reduced because 5d orbitals are significantly
larger than 3d ones. The fact that metal–insulator transitions can be
induced by physical pressure or chemical pressure in all three cases
(for the two organic compounds and Na4Ir3O8)43,44, and the fact that
in κ-(BEDT-TTF)2Cu2(CN)3 no clear Mott gap is seen spectroscopi-
cally45, is direct experimental evidence that these materials are weak
Mott insulators. The applicability of a Heisenberg spin description to
these materials is thus open to question.
Thermodynamic studies of these materials are also important because
c
they reveal the spectrum of low-energy states, an abundance of which
is expected in frustrated systems. At very low temperatures, the func-
tional form of the specific heat and sometimes the spin (or magnetic)
susceptibility can distinguish between QSLs. The general thermody-
namic features of the materials are remarkably similar. In all cases in
which it has been measured, the magnetic specific heat shows a peak
well below the Curie–Weiss temperature46–49, indicating that significant
spin entropy is maintained above this peak temperature. This specific
heat then gradually decreases on further lowering of the temperature, in
a manner that can roughly be fitted to a quadratic behaviour, Cv ≈ AT2,
where Cv is molar specific heat at constant volume and A is a constant.
This non-exponential form clearly indicates the absence of an energy
gap. However, some caution is in order before interpreting this as char-
acteristic of QSLs: such approximately quadratic behaviour of Cv is com-
mon in frustrated magnets, even ones that are classical and for which
Figure 2 | Spins, artificial magnetic fields and monopoles in spin
QSL physics is clearly irrelevant50. At very low temperatures, the can- ice. a, A ground-state configuration of spins is shown in a pyrochlore
didate QSL materials generally show a linear behaviour, Cv ≈ γT, where lattice. The spins obey the constraint of the ice rules that mandates two
γ is the Sommerfeld coefficient (refs 46–49, 51), which varies widely inward-pointing spins and two outward-pointing ones on each tetrahedron.
(between 1 and 250 mJ K−2 mol−1). This is evidence of a large density of b, For the same lattice type, some of the loops of ‘magnetic flux’ are shown
low-energy states. (red lines), as defined by the mapping of spins to an artificial magnetic field.
Ideally, the spin susceptibility, χ, is a useful measure of the available c, For the same lattice type, a monopole (green) and antimonopole (red)
low-energy spin excitations. All of the candidate QSL materials also show are shown. These are created by flipping the ‘string’ of spins connected by
a large Pauli-like paramagnetic susceptibility, χ0, in the zero-temperature the yellow line (compare with the spins in b). For simplicity, the particles
limit. This susceptibility is smallest for the organic compounds: for these, themselves are not depicted. Note that the spins on the tetrahedron
containing a monopole (antimonopole) orient three-in, one-out (three-out,
χ0 ≈ 3 × 10−4 e.m.u. mol−1 (refs 52, 53), whereas it is 3–10 times larger
one-in), violating the ice rules.
for the inorganic compounds41,42,46,47,49. Before these values of χ0 can be
interpreted as additional evidence for gapless excitations in QSLs, it is
important to consider the influence of spin–orbit coupling. Spin–orbit larger than 1, with the exception of ZnCu3(OH)6Cl2 (for which an
coupling generally leads to a non-zero χ0 value irrespective of the pres- estimate of the intrinsic γ value, uncontaminated by impurities, is
ence or absence of an energy gap. An empirical indicator of the impor- not available) and κ-(BEDT-TTF)2Cu2(CN)3. The value of the Wilson
tance of spin–orbit coupling is the Wilson ratio, which is defined as ratio for ZnCu3(OH)6Cl2 is probably suppressed by the high degree
of disorder. The extremely large value for Na4Ir3O8 (R = 70) indicates
4π2kB2χ0 very strong spin–orbit effects, which are not surprising given the large
R = _________ (1) atomic number of iridium.
3(gμB)2γ The most direct evidence for a lack of magnetic ordering comes from
local probe measurements, by using techniques such as NMR and muon
and should be, at most, of order 1 in the absence of strong ferromagnetic spin resonance, in which local fields resulting from static moments affect
tendencies or spin–orbit coupling. Because all of the candidate QSL the nuclear or muon spins and can be readily detected by their influence
materials are strongly antiferromagnetic, the ferromagnetic interpreta- on the muon spins. Such measurements have confirmed the absence of
tion is untenable. In fact, most (but not all) QSL theories, in the absence local static moments down to T = 32 mK in κ-(BEDT-TTF)2Cu2(CN)3
of spin–orbit coupling, predict that R << 1. By contrast, the measured (ref. 52), T = 1.37 K in EtMe3Sb[Pd(dmit)2]2 (ref. 54) and T = 50 mK in
Wilson ratios for all of the QSL candidates in Table 1 are significantly ZnCu3(OH)6Cl2 (ref. 55). However, in Cu3V2O7(OH)2•2H2O, magnetic
203
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
order and/or freezing is observed, by using NMR spectroscopy, at T < 1 K
(ref. 56). Moreover, recent experiments show that this compound has a
complex series of low-temperature phases in an applied magnetic field56.
Given the exceptionally high purity of Cu3V2O7(OH)2•2H2O, an expla-
nation of its phase diagram should be a clear theoretical goal.
Theoretical interpretations
I now turn to the theoretical evidence for QSLs in these systems and
how the experiments can be reconciled with theory. Theorists have
attempted to construct microscopic models for these materials (Box 2)
and to determine whether they support QSL ground states. In the case
of the organic compounds, these are Hubbard models, which account
for significant charge fluctuations. For the kagomé materials, a Heisen-
berg model description is probably appropriate. There is general theo-
retical agreement that the Hubbard model for a triangular lattice has
a QSL ground state for intermediate-strength Hubbard repulsion near
the Mott transition57–59. On the kagomé lattice, the Heisenberg model
is expected to have a non-magnetic ground state as a result of frus-
tration60. Recently, there has been growing theoretical support for the
conjecture that the ground state is, however, not a QSL but a VBS with
a large, 36-site, unit cell61,62. However, all approaches indicate that many
competing states exist, and these states have extremely small energy dif-
ferences from this VBS state. Thus, the ‘real’ ground state in the kagomé
materials is probably strongly perturbed by spin–orbit coupling, dis-
order, further-neighbour interactions and so on63. A similar situation
applies to the hyperkagomé lattice of Na4Ir3O8 (ref. 64).
These models are difficult to connect directly, and in detail, to
experiments, which mainly measure low-energy properties at low tem-
peratures. Instead, attempts to reconcile theory and experiment in detail
have relied on more phenomenological low-energy effective theories
of QSLs. Such effective theories are similar in spirit to the Fermi liquid
theory of interacting metals: they propose that the ground state has a
certain structure and a set of elementary excitations that are consistent
with this structure. In contrast to the Fermi liquid case, however, the
elementary excitations consist of spinons and other exotic particles,
which are coupled by gauge fields. A theory of this type — that is, pro-
posing a ‘spinon Fermi surface’ coupled to a U(1) gauge field — has
had some success in explaining data from experiments on κ-(BEDT-
TTF)2Cu2(CN)3 (refs 65, 66). Related theories have been proposed for
ZnCu3(OH)6Cl2 (ref. 67) and Na4Ir3O8 (ref. 68). However, comparisons
a
+ +…
+…
Figure 3 | Valence-bond states of frustrated antiferromagnets. In a VBS
state (a), a specific pattern of entangled pairs of spins — the valence bonds
— is formed. Entangled pairs are indicated by ovals that cover two points
on the triangular lattice. By contrast, in a RVB state, the wavefunction is a
204
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
for these materials are much more limited. In all cases, the comparison
of theory with experiment has, so far, been indirect. I return to this
problem in the subsection ‘The smoking gun for QSLs’.
Unexpected findings
In the course of a search as difficult as the one for QSLs, it is natural for
there to be false starts. In several cases, researchers uncovered other
interesting physical phenomena in quantum magnetism.
Dimensional reduction in Cs2CuCl4
Cs2CuCl4 is a spin-½ antiferromagnet on a moderately anisotropic
triangular lattice69,70. It shows only intermediate frustration, with f ≈ 8,
ordering into a spiral Néel state at TN = 0.6 K. However, neutron-scat-
tering results for this compound reported by Coldea and colleagues
suggested that exotic physical phenomena were occurring69,70. These
experiments measure the type of excitation that is created when a neu-
tron interacts with a solid and flips an electron spin. In normal mag-
nets, this creates a magnon and, correspondingly, a sharp resonance is
observed when the energy and momentum transfer of the neutron equal
that of the magnon. In Cs2CuCl4, this resonance is extremely small.
Instead, a broad scattering feature is mostly observed. The interpreta-
tion of this result is that the neutron’s spin flip creates a pair of spinons,
which divide the neutron’s energy and momentum between them. The
spinons were suggested to arise from an underlying 2D QSL state.
A nagging doubt with respect to this picture was the striking similar-
ity between some of the spectra in the experiment and those of a 1D
spin chain, in which 1D spinons indeed exist71. In fact, in Cs2CuCl4 the
exchange energy along one ‘chain’ direction is three times greater than
along the diagonal bonds between chains (that is, Jʹ ≈ J/3 in Fig. 1a).
Experimentally, however, the presence of substantial transverse disper-
sion (that is, dependence of the neutron peak on momentum perpendic-
ular to the chain axis in Cs2CuCl4), and the strong influence of interchain
coupling on the magnetization curve, M(H), seemed to rule out a 1D
origin, despite an early theoretical suggestion72.
In the past few years, it has become clear that discarding the idea of
1D physics was premature73,74. It turns out that although the interchain
coupling is substantial, and thus affects the M(H) curve significantly,
the frustration markedly reduces interchain correlations in the ground
state. As a result, the elementary excitations of the system are simi-
lar to those of 1D chains, with one important exception. Because the
1 ( + )
–2
superposition of many different pairings of spins. The valence bonds may
be short range (b) or long range (c). Spins in longer-range valence bonds
(the longer, the lighter the colour) are less tightly bound and are therefore
more easily excited into a state with non-zero spin.
NATURE|Vol 464|11 March 2010
Figure 4 | Excitations of quantum antiferromagnets. a, In a quasi-1D
system (such as the triangular lattice depicted), 1D spinons are formed as
a domain wall between the two antiferromagnetic ground states. Creating
a spinon (yellow arrow) thus requires the flipping of a semi-infinite line
of spins along a chain, shown in red. The spinon cannot hop between
chains, because to do so would require the coherent flipping of an infinite
spinons are inherently 1D, they are confined to the chains, and to take
advantage of the transverse exchange, they must be bound into S = 1
‘triplon’ bound states (Fig. 3b). These triplons move readily between
chains and are responsible for the transverse dispersion observed in
the experiment. Thus, the observation of triplons provides a means to
distinguish 1D spinons from their higher-dimensional counterparts.
A quantitative theory of this physics agrees well with the data, with no
adjustable parameters. It is therefore understood that Cs2CuCl4 is an
example of ‘dimensional reduction’ induced by frustration and quan-
tum fluctuations. This phenomenon was unexpected and might have a
role in other correlated materials. Perhaps it is related to the cascade of
phases that is observed in the isostructural material Cs2CuBr4 in applied
magnetic fields75.
Spin–orbital quantum criticality in FeSc2S4
Among the entries in Table 1, FeSc2S4 stands out as a material that has
not only spin degeneracy but also orbital degeneracy. This is common
in transition-metal-containing compounds76,77. It is possible to imagine
a quantum orbital liquid78–80, analogous to a QSL. Like the more familiar
(theoretically) QSL, the quantum orbital liquid is experimentally elu-
sive. Nevertheless, experimentalists have observed that FeSc2S4, which
has a twofold orbital degeneracy, evades order down to T = 50 mK, and
on this basis it was characterized as a spin–orbital liquid81–83.
Recently, it was suggested that this liquid behaviour is due not to frus-
Table 1 | Some experimental materials studied in the search for QSLs
Material Lattice
κ-(BEDT-TTF)2Cu2(CN)3 Triangular†
EtMe3Sb[Pd(dmit)2]2 Triangular†
Cu3V2O7(OH)2•2H2O (volborthite) Kagomé†
ZnCu3(OH)6Cl2 (herbertsmithite) Kagomé
BaCu3V2O8(OH)2 (vesignieite) Kagomé†
Na4Ir3O8 Hyperkagomé
Cs2CuCl4 Triangular†
FeSc2S4 Diamond
BEDT-TTF, bis(ethylenedithio)-tetrathiafulvalene; dmit, 1,3-dithiole-2-thione-4,5-ditholate; Et, ethyl; Me, methyl. *R is the Wilson ratio, which is defined in equation (1) in the main text. For EtMe3Sb[Pd(dmit)2]2
and ZnCu3(OH)6Cl2, experimental data for the intrinsic low-temperature specific heat are not available, hence R is not determined. †Some degree of spatial anisotropy is present, implying that Jʹ≠J in Fig. 1a. ‡A
theoretical Curie–Weiss temperature (ΘCW) calculated from the high-temperature expansion for an S=½ triangular lattice; ΘCW=3J/2kB, using the J fitted to experiment.
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
number of spins, in this case all of the red spins and their counterparts on
the next chain. b, A bound pair of 1D spinons forms a triplon. Because a
finite number of spins are flipped between the two domain walls (shown
in red), the triplon can coherently move between chains, by the flipping of
spins along the green bonds. c, In a 2D QSL, a spinon is created simply as an
unpaired spin, which can then move by locally adjusting the valence bonds.
tration but to a competition between spin–orbit coupling and magnetic
exchange84. Microscopic estimates of the spin–orbit interaction, λ,
indeed show that its strength, λ/kB = 25–50 K, is comparable to the Curie–
Weiss temperature, 45 K. As a result, the material is serendipitously close
to a quantum critical point between a magnetically ordered state and
a ‘spin–orbital singlet’, induced by spin–orbit coupling84 (Fig. 5). This
picture seems to explain data from a variety of experiments, including
NMR81, neutron-scattering82, spin susceptibility83 and specific-heat83
measurements. Most notably, the anomalously small excitation gap of
2 K that was measured in neutron-scattering82 and NMR81 experiments
is understandable — this gap vanishes on approaching the quantum
critical point. If the theory is correct, FeSc2S4 can be viewed as a kind of
spin–orbital liquid with significant long-distance entanglement between
spins and orbitals. Because the material is not precisely at the quantum
critical point, however, there is a finite correlation length; therefore, this
entanglement does not persist to arbitrarily long distances, as would be
expected in a true RVB state.
Future directions
I have only touched the surface of the deep well of phenomena to
be explored, experimentally and theoretically, in frustrated magnets
and spin liquids. In spin ice, there are subtle correlations, collective
excitations and emergent magnetic monopoles, all of which are highly
amenable to laboratory studies. In several frustrated magnets with spin
S ΘCW (K) R* Status or explanation
½ −375‡ 1.8 Possible QSL
½ −(375–325)‡ ? Possible QSL
½ −115 6 Magnetic
½ −241 ? Possible QSL
½ −77 4 Possible QSL
½ −650 70 Possible QSL
½ −4 0 Dimensional reduction
2 −45 230 Quantum criticality
205
INSIGHT REVIEW
FeSc2S4
Temperature
Quantum
critical
region
Spin–orbital
Antiferromagnet
singlet
Quantum critical point
Competition between magnetic exchange and spin–orbit interaction, J/l
Figure 5 | Spin–orbital quantum criticality. Schematic phase diagram of
FeSc2S4, representing the competition between spin–orbit interaction,
λ, and magnetic exchange energy, J. For large values of J/λ, the spins
magnetically order into an antiferromagnetic state. For small J/λ, the
spins and orbitals combine into a non-magnetic spin–orbital singlet
state. The two ground states are separated by a quantum critical point,
which governs the quantum critical region (in the ‘fan’ above it) at T > 0.
The dashed line indicates the value of J/λ for FeSc2S4, which has a large
correlation length and small gap as a consequence of its proximity to the
quantum critical point.
S = ½, researchers seem finally to have uncovered QSL ground states,
which have long eluded them, but the detailed structures of these
have not yet been elucidated. Explaining them will be a theoretical
and experimental challenge. In addition, in the process of searching
for QSLs, examples of dimensional reduction and quantum criticality
were unexpectedly found.
Beyond spin ice
The recent observation of magnetic monopoles, in agreement with
theory, suggests that there is a good understanding of the interactions
and dynamics of the canonical spin-ice materials. The challenge is now
to improve this understanding by making a more detailed and accurate
comparison of theory and experiment and perhaps by ‘engineering’ new
phenomena through the control of magnetic monopoles. For instance, it
might be possible to make spectral measurements of magnetic monopo-
les by exciting them with neutrons. Some theoreticians have suggested
another consequence of emergent electromagnetism in spin ice. If the
material can be driven through certain types of phase transition, then
exotic critical behaviour in violation of the standard Landau–Ginzburg–
Wilson paradigm might be observed85–87.
Spin ice certainly proves that there is more to be learned about clas-
sical spin liquids, and that these materials are much simpler to under-
stand than their quantum counterparts. In fact, many other materials
have classical spin-liquid regimes, and perhaps the ideas that have
arisen from studies of spin ice will inspire an understanding of their
dynamics. Well-known classical spin liquids include gadolinium gal-
lium garnet and SrCr9–xGa3+xO19, both of which have large, predomi-
nantly classical spins with very large frustration parameters50. Another
set of examples is the spinel chromites, ACr2O4 (where A = Zn, Cd or
Hg), which form S = 3⁄2 Heisenberg pyrochlore antiferromagnets and
show a spin-liquid regime with unusual ‘loop’ structures seen in dif-
fuse neutron scattering88. More recently, the A-site spinels MnSc2S4
(ref. 83) and CoAl2O4 (ref. 89) have been interpreted in terms of a
‘spiral spin liquid’90.
Another way to move ‘beyond’ spin ice is to introduce quantum
mechanics. Quantum effects are seen in Tb2Ti2O7, which in a magnetic
field shows many of the anisotropies that are characteristic of spin ice. It
avoids freezing or ordering down to T = 50 mK and remains an intrigu-
ing theoretical puzzle6,91. The material Pr2Ir2O7 combines features of spin
ice with metallic conduction92 and has a zero-field anomalous Hall effect
that is indicative of a ‘chiral spin liquid’ phase.
206
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
The smoking gun for QSLs
The diversity of QSL candidates and the progress being made towards
their theoretical description are encouraging, but the research com-
munity is still in the unsatisfactory situation in which the evidence for
QSL states is circumstantial and heavily depends on theoretical inter-
pretation. For spin ice, by contrast, there is increasingly direct experi-
mental evidence that magnetic monopoles exist. For QSLs, so far, such
a ‘smoking gun’ is lacking. A conventional experimental technique that
could be a candidate for identifying QSLs is inelastic neutron scattering,
which the experiments on Cs2CuCl4 clearly show can reveal spinons
under appropriate conditions69. This is an important avenue for future
experiments, but it would be advantageous to design new experiments
specifically to identify QSLs. Some proposals have been made. It has
been suggested that a particular class of QSL could display a vortex
memory effect, which would probe the existence of the ‘vison’ excitation
of such a spin liquid93, but it is extremely difficult to measure this effect
in most materials94. For QSLs with a spinon Fermi surface or Fermi
points, it might be possible to measure ‘2kF’ oscillations, which would
be considerable if observed in an insulator. Such observations could be
made in neutron-scattering measurements, or they could be made in
real space by way of analogues of Friedel oscillation measurements, by
using local probes or composite structures95.
Connections
The physics of frustrated magnetism and spin liquids has intriguing
connections to diverse exotic phenomena, many of which are discussed
in other articles in this Insight. For example, the modern interest in spin
liquids grew largely from Anderson’s proposed connection between
RVB states and high-temperature copper-oxide-based superconduc-
tivity4. The idea is that, in an RVB state, electrons bound into valence
bonds are paired even though the state is non-superconducting. Thus,
part of the physics of superconductivity has already been achieved in
this state, and if the material can be made conducting and phase coher-
ent, for example by doping with mobile charges, it will naturally become
superconducting. Despite the elegance of this suggestion and decades of
theoretical work, RVB superconductivity remains unproven in experi-
ments, although the idea is still actively discussed96.
More generally, the problem of frustration and spin-liquid physics in
systems with mobile charges is relatively unstudied. There are several
interesting materials in which this type of physics can be explored exper-
imentally. I have discussed materials in which such physics might be
involved, including the triangular-lattice organic compounds, Na4Ir3O8,
and Pr2Ir2O7 (a type of metallic spin ice). Another interesting material in
this context is 2H-AgNiO2, a semi-metallic material that shows charge
ordering on a triangular lattice97. It has been suggested that a novel,
continuous Mott transition — rather than the usual discontinuous one
— might be observed in such a situation98,99.
From the above discussion, it is clear that spin–orbit coupling seems
to have an important role in frustrated systems and QSLs. Even in the
organic compounds, in which spin–orbit effects are expected to be mini-
mized by the extended molecular orbital states, spin–orbit effects are not
obviously negligible. Experiments on κ-(BEDT-TTF)2Cu[N(CN)2]Cl,
which is a magnetically ordered organic compound, demonstrated a
Dzyaloshinskii–Moriya (DM) coupling, D, of order D/kB ≈ 5 K (ref. 100).
If κ-(BEDT-TTF)2Cu2(CN)3 had a similar magnitude of DM coupling
it would probably affect many of the low-temperature measurements,
such as specific heat and thermal conductivity. In the inorganic materi-
als, DM interactions and other spin–orbit effects are clearly significant
and are even dominant in the case of Na4Ir3O8.
For such materials with strong spin–orbit coupling, it will be interest-
ing to explore the connection to topological insulators (see page 194),
which are known to occur in situations with strong spin–orbit coupling
and weak correlation. How does the topological insulator evolve with
increasing electron localization by Coulomb interactions? Ultimately,
with very large values of U, the strong Mott insulating limit is reached
and the system is described by a Heisenberg model, as I have discussed
here. It is also interesting to ask how far the physics of topological insula-
NATURE|Vol 464|11 March 2010
tors survives in this direction and whether anything qualitatively new
emerges in the presence of both spin–orbit and electronic correlations.
The combination of proximity to the Mott transition and strong spin–
orbit effects may be common in the late 5d transition metal oxides. A
preliminary theoretical study suggests that the physics in this regime
could be particularly interesting101.
I have only hinted at the abundance of exotic properties of frustrated
magnets. Nature has many surprises in store for researchers of these
complex materials. It is safe to assume that the exploration of spin liquids
will reveal more unexpected phenomena, which might be as interesting
as the original objects of the search. ■ normal state of underdoped high Tc superconductors. Phys. Rev. Lett. 86, 3871–3874
1. Ramirez, A. P., Broholm, C. L., Cava, R. J. & Kowach, G. R. Geometrical frustration, spin ice
and negative thermal expansion — the physics of underconstraint. Physica B 280, 290–295
(2000).
2. Kléman, M., Lavrentovich, O. D. & Friedel, J. Soft Matter Physics: An Introduction (Springer,
2003).
3. Wannier, G. H. Antiferromagnetism. The triangular Ising net. Phys. Rev. 79, 357–364
(1950).
4. Anderson, P. W. The resonating valence bond state in La2CuO4 and superconductivity.
Science 235, 1196–1198 (1987).
5. Bramwell, S. T. & Gingras, M. J. P. Spin ice state in frustrated magnetic pyrochlore materials.
Science 294, 1495–1501 (2001).
This is an excellent review of the physics of spin ice that was published before the recent
work on Coulomb correlations and monopoles.
6. Gingras, M. J. P. in Highly Frustrated Magnetism (eds Lacroix, C., Mendels, P. & Mila, F.)
(Springer, in the press); preprint at 〈http://arXiv.org/abs/0903.2772〉 (2009).
7. Harris, M. J., Bramwell, S. T., McMorrow, D. F., Zeiske, T. & Godfrey, K. W. Geometrical
frustration in the ferromagnetic pyrochlore Ho2Ti2O7. Phys. Rev. Lett. 79, 2554–2557
(1997).
8. Isakov, S. V., Moessner, R. & Sondhi, S. L. Why spin ice obeys the ice rules. Phys. Rev. Lett.
95, 217201 (2005).
9. den Hertog, B. C. & Gingras, M. J. P. Dipolar interactions and origin of spin ice in Ising
pyrochlore magnets. Phys. Rev. Lett. 84, 3430–3433 (2000).
10. Ramirez, A. P., Hayashi, A., Cava, R. J., Siddharthan, R. & Shastry, B. S. Zero-point entropy in
’spin ice’. Nature 399, 333–334 (1999).
11. Pauling, L. The structure and entropy of ice and of other crystals with some randomness of
atomic arrangement. J. Am. Chem. Soc. 57, 2680–2684 (1935).
12. Youngblood, R., Axe, J. D. & McCoy, B. M. Correlations in ice-rule ferroelectrics. Phys. Rev. B
21, 5212–5220 (1980).
13. Isakov, S. V., Gregor, K., Moessner, R. & Sondhi, S. L. Dipolar spin correlations in classical
pyrochlore magnets. Phys. Rev. Lett. 93, 167204 (2004).
14. Henley, C. L. Power-law spin correlations in pyrochlore antiferromagnets. Phys. Rev. B 71,
014424 (2005).
15. Fennell, T. et al. Experimental proof of a magnetic Coulomb phase. Preprint at
〈http://arXiv.org/abs/0907.0954〉 (2009).
16. Castelnovo, C., Moessner, R. & Sondhi, S. L. Magnetic monopoles in spin ice. Nature 451,
42–45 (2008).
This theoretical paper proposed that magnetic monopoles are present in spin ice.
17. Jaubert, L. D. C. & Holdsworth, P. C. W. Signature of magnetic monopole and Dirac string
dynamics in spin ice. Nature Phys. 5, 258–261 (2009).
This paper shows how the density and dynamics of magnetic monopoles can be inferred
from the experimental magnetization relaxation rate in spin ice.
18. Onsager, L. Deviations from Ohm’s law in weak electrolytes. J. Chem. Phys. 2, 599–615
(1934).
19. Bramwell, S. T. et al. Measurement of the charge and current of magnetic monopoles in
spin ice. Nature 461, 956–959 (2009).
This paper extracts an experimental value for the charge of a magnetic monopole in spin
ice, on the basis of an analogy of the monopole plasma to a classical electrolyte.
20. Morris, D. J. P. et al. Dirac strings and magnetic monopoles in the spin ice Dy2Ti2O7. Science
326, 411–414 (2009).
21. Kadowaki, H. et al. Observation of magnetic monopoles in spin ice. J. Phys. Soc. Jpn 78,
103706 (2009).
22. Anderson, P. W. Resonating valence bonds: a new kind of insulator. Mater. Res. Bull. 8,
153–160 (1973).
This paper proposed the RVB wavefunction for a QSL state.
23. Iwase, H., Isobe, M., Ueda, Y. & Yasuoka, H. Observation of spin gap in CaV2O5 by NMR.
J. Phys. Soc. Jpn 65, 2397–2400 (1996).
24. Azuma, M., Hiroi, Z., Takano, M., Ishida, K. & Kitaoka, Y. Observation of a spin gap
in SrCu2O3 comprising spin-½ quasi-1D two-leg ladders. Phys. Rev. Lett. 73, 3463–3466
(1994).
25. Kageyama, H. et al. Exact dimer ground state and quantized magnetization plateaus in the
two-dimensional spin system SrCu2(BO3)2.Phys. Rev. Lett. 82, 3168–3171 (1999).
26. Nikuni, T., Oshikawa, M., Oosawa, A. & Tanaka, H. Bose–Einstein condensation of dilute
magnons in TlCuCl3. Phys. Rev. Lett. 84, 5868–5871 (2000).
27. Moessner, R. & Sondhi, S. L. Resonating valence bond phase in the triangular lattice
quantum dimer model. Phys. Rev. Lett. 86, 1881–1884 (2001).
28. Kitaev, A. Anyons in an exactly solved model and beyond. Ann. Phys. (Leipz.) 321, 2–111
(2006).
29. Motrunich, O. I. & Senthil, T. Exotic order in simple models of bosonic systems. Phys. Rev.
Lett. 89, 277004 (2002).
30. Balents, L., Fisher, M. P. A. & Girvin, S. M. Fractionalization in an easy-axis kagomé
antiferromagnet. Phys. Rev. B 65, 224412 (2002).
31. Hermele, M., Fisher, M. P. A. & Balents, L. Pyrochlore photons: the U(1) spin liquid in a S=½
three-dimensional frustrated magnet. Phys. Rev. B 69, 064404 (2004).
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
32. Levin, M. A. & Wen, X. G. String-net condensation: a physical mechanism for topological
phases. Phys. Rev. B 71, 045110 (2005).
33. Liang, S., Doucot, B. & Anderson, P. W. Some new variational resonating-valence-bond-
type wave functions for the spin-½ antiferromagnetic Heisenberg model on a square
lattice. Phys. Rev. Lett. 61, 365–368 (1988).
34. Alet, F., Walczak, A. M. & Fisher, M. P. A. Exotic quantum phases and phase transitions in
correlated matter. Physica A 369, 122–142 (2006).
35. Wen, X. G. Quantum orders and symmetric spin liquids. Phys. Rev. B 65, 165113 (2002).
36. Read, N. & Sachdev, S. Large-N expansion for frustrated quantum antiferromagnets. Phys.
Rev. Lett. 66, 1773–1776 (1991).
37. Affleck, I. & Marston, J. B. Large-n limit of the Heisenberg–Hubbard model: implications for
high-Tc superconductors. Phys. Rev. B 37, 3774–3777 (1988).
38. Rantner, W. & Wen, X.-G. Electron spectral function and algebraic spin liquid for the
(2001).
39. Lee, P. A. An end to the drought of quantum spin liquids. Science 321, 1306–1307 (2008).
40. Ramirez, A. P. Quantum spin liquids: a flood or a trickle? Nature Phys. 4, 442–443 (2008).
41. Olariu, A. et al. 17O NMR study of the intrinsic magnetic susceptibility and spin dynamics
of the quantum kagomé antiferromagnet ZnCu3(OH)6Cl2. Phys. Rev. Lett. 100, 087202
(2008).
42. Yoshida, H. et al. Magnetization ‘steps’ on a kagomé lattice in volborthite. J. Phys. Soc. Jpn
78, 043704 (2009).
43. Kurosaki, Y., Shimizu, Y., Miyagawa, K., Kanoda, K. & Saito, G. Mott transition from a spin
liquid to a Fermi liquid in the spin-frustrated organic conductor κ-(ET)2Cu2(CN)3. Phys. Rev.
Lett. 95, 177001 (2005).
44. Shimizu, Y., Akimoto, H., Tsujii, H., Tajima, A. & Kato, R. Mott transition in a valence-bond
solid insulator with a triangular lattice. Phys. Rev. Lett. 99, 256403 (2007).
45. Kézsmárki, I. et al. Depressed charge gap in the triangular-lattice Mott insulator
κ-(ET)2Cu2(CN)3. Phys. Rev. B 74, 201101 (2006).
46. Okamoto, Y., Nohara, M., Aruga-Katori, H. & Takagi, H. Spin-liquid state in the S=½
hyperkagomé antiferromagnet Na4Ir3O8. Phys. Rev. Lett. 99, 137207 (2007).
47. Okamoto, Y., Yoshida, H. & Hiroi, Z. Vesignieite BaCu3V2O8(OH)2 as a candidate spin-½
kagomé antiferromagnet. J. Phys. Soc. Jpn 78, 033701 (2009).
48. Hiroi, Z. et al. Spin-½ kagomé-like lattice in volborthite, Cu3V2O7(OH)2·2H2O. J. Phys. Soc.
Jpn 70, 3377–3384 (2001).
49. Helton, J. S. et al. Spin dynamics of the spin-½ kagomé lattice antiferromagnet
ZnCu3(OH)6Cl2. Phys. Rev. Lett. 98, 107204 (2007).
50. Ramirez, A. P. Strongly geometrically frustrated magnets. Annu. Rev. Mater. Sci. 24,
453–480 (1994).
This review helped to define the field of highly frustrated magnets and is an excellent
discussion of the state of the science at the time.
51. Yamashita, S. et al. Thermodynamic properties of a spin-½ spin-liquid state in a κ-type
organic salt. Nature Phys. 4, 459–462 (2008).
52. Shimizu, Y., Miyagawa, K., Kanoda, K., Maesato, M. & Saito, G. Spin liquid state in an
organic Mott insulator with a triangular lattice. Phys. Rev. Lett. 91, 107001 (2003).
53. Itou, T., Oyamada, A., Maegawa, S., Tamura, M. & Kato, R. Quantum spin liquid in the
spin-½ triangular antiferromagnet EtMe3Sb[Pd(dmit)2]2. Phys. Rev. B 77, 104413 (2008).
54. Itou, T., Oyamada, A., Maegawa, S., Tamura, M. & Kato, R. Spin-liquid state in an organic
spin-½ system on a triangular lattice, EtMe3Sb[Pd(dmit)2]2. J. Phys. Condens. Matter 19,
145247 (2007).
55. Mendels, P. et al. Quantum magnetism in the paratacamite family: towards an ideal
kagomé lattice. Phys. Rev. Lett. 98, 077204 (2007).
56. Yoshida, M., Takigawa, M., Yoshida, H., Okamoto, Y. & Hiroi, Z. Phase diagram and spin
dynamics in volborthite with a distorted kagomé lattice. Phys. Rev. Lett. 103, 077207
(2009).
57. Koretsune, T., Motome, Y. & Furusaki, A. Exact diagonalization study of Mott transition in
the Hubbard model on an anisotropic triangular lattice. J. Phys. Soc. Jpn 76, 074719 (2007).
58. Kyung, B. & Tremblay, A. M. S. Mott transition, antiferromagnetism, and d-wave
superconductivity in two-dimensional organic conductors. Phys. Rev. Lett. 97, 046402
(2006).
59. Morita, H., Watanabe, S. & Imada, M. Nonmagnetic insulating states near the Mott
transitions on lattices with geometrical frustration and implications for κ-(ET)2Cu2(CN)3.
J. Phys. Soc. Jpn 71, 2109–2112 (2002).
60. Lecheminant, P., Bernu, B., Lhuillier, C., Pierre, L. & Sindzingre, P. Order versus disorder
in the quantum Heisenberg antiferromagnet on the kagomé lattice using exact spectra
analysis. Phys. Rev. B 56, 2521–2529 (1997).
61. Marston, J. B. & Zeng, C. Spin-Peierls and spin-liquid phases of kagomé quantum
antiferromagnets. J. Appl. Phys. 69, 5962 (1991).
62. Singh, R. R. P. & Huse, D. A. Ground state of the spin-½ kagomé lattice Heisenberg
antiferromagnet. Phys. Rev. B 76, 180407 (2007).
63. Cépas, O., Fong, C. M., Leung, P. W. & Lhuillier, C. Quantum phase transition induced
by Dzyaloshinskii–Moriya interactions in the kagomé antiferromagnet. Phys. Rev. B 78,
140405 (2008).
64. Chen, G. & Balents, L. Spin–orbit effects in Na4Ir3O8: a hyperkagomé lattice
antiferromagnet. Phys. Rev. B 78, 094403 (2008).
65. Motrunich, O. I. Variational study of triangular lattice spin-½ model with ring exchanges
and spin liquid state in κ-(ET)2Cu2(CN)3. Phys. Rev. B 72, 045105 (2005).
66. Lee, S.-S. & Lee, P. A. U(1) gauge theory of the Hubbard model: spin liquid states and
possible application to κ-(BEDT-TTF)2Cu2(CN)3.Phys. Rev. Lett. 95, 036403 (2005).
67. Ran, Y., Hermele, M., Lee, P. A. & Wen, X. G. Projected-wave-function study of the spin-½
Heisenberg model on the kagomé lattice. Phys. Rev. Lett. 98, 117205 (2007).
68. Lawler, M. J., Paramekanti, A., Kim, Y. B. & Balents, L. Gapless spin liquids on the three-
dimensional hyperkagomé lattice of Na4Ir3O8. Phys. Rev. Lett. 101, 197202 (2008).
69. Coldea, R., Tennant, D. A. & Tylczynski, Z. Extended scattering continua characteristic
of spin fractionalization in the two-dimensional frustrated quantum magnet Cs2CuCl4
observed by neutron scattering. Phys. Rev. B 68, 134424 (2003).
70. Coldea, R., Tennant, D. A., Tsvelik, A. M. & Tylczynski, Z. Experimental realization of a 2D
fractional quantum spin liquid. Phys. Rev. Lett. 86, 1335–1338 (2001).
207
INSIGHT REVIEW
71. Lake, B., Tennant, D. A., Frost, C. D. & Nagler, S. E. Quantum criticality and universal scaling
of a quantum antiferromagnet. Nature Mater. 4, 329–334 (2005).
72. Bocquet, M., Essler, F. H. L., Tsvelik, A. M. & Gogolin, A. O. Finite-temperature dynamical
magnetic susceptibility of quasi-one-dimensional frustrated spin-½ Heisenberg
antiferromagnets. Phys. Rev. B 64, 094425 (2001).
73. Starykh, O. A. & Balents, L. Ordering in spatially anisotropic triangular antiferromagnets.
Phys. Rev. Lett. 98, 077205 (2007).
74. Kohno, M., Starykh, O. A. & Balents, L. Spinons and triplons in spatially anisotropic
frustrated antiferromagnets. Nature Phys. 3, 790–795 (2007).
This theoretical paper shows how a model of interacting 1D spinons can quantitatively
explain inelastic neutron scattering in Cs2CuCl4.
75. Fortune, N. A. et al. Cascade of magnetic-field-induced quantum phase transitions in a
spin-½ triangular-lattice antiferromagnet. Phys. Rev. Lett. 102, 257201 (2009).
76. Khomskii, D. I. Role of orbitals in the physics of correlated electron systems. Phys. Scr. 72,
CC8–CC14 (2005).
77. Tokura, Y. & Nagaosa, N. Orbital physics in transition-metal oxides. Science 288, 462–468
(2000).
78. Ishihara, S., Yamanaka, M. & Nagaosa, N. Orbital liquid in perovskite transition-metal
oxides. Phys. Rev. B 56, 686–692 (1997).
79. Feiner, L. F., Olés, A. M. & Zaanen, J. Quantum melting of magnetic order due to orbital
fluctuations. Phys. Rev. Lett. 78, 2799–2802 (1997).
80. Khaliullin, G. & Maekawa, S. Orbital liquid in three-dimensional Mott insulator: LaTiO3.
Phys. Rev. Lett. 85, 3950–3953 (2000).
81. Büttgen, N., Zymara, A., Kegler, C., Tsurkan, V. & Loidl, A. Spin and orbital frustration in
FeSc2S4 probed by 45Sc NMR. Phys. Rev. B 73, 132409 (2006).
82. Krimmel, A. et al. Vibronic and magnetic excitations in the spin–orbital liquid state of
FeSc2S4. Phys. Rev. Lett. 94, 237402 (2005).
83. Fritsch, V. et al. Spin and orbital frustration in MnSc2S4 and FeSc2S4. Phys. Rev. Lett. 92,
116401 (2004).
84. Chen, G., Balents, L. & Schnyder, A. P. A. Spin–orbital singlet and quantum critical point on
the diamond lattice: FeSc2S4. Phys. Rev. Lett. 102, 096406 (2009).
85. Powell, S. & Chalker, J. T. Classical to quantum mappings for geometrically frustrated
systems: spin-ice in a [100] field. Phys. Rev. B 78, 024422 (2008).
86. Saunders, T. E. & Chalker, J. T. Structural phase transitions in geometrically frustrated
antiferromagnets. Phys. Rev. B 77, 214438 (2008).
87. Bergman, D. L., Fiete, G. A. & Balents, L. Ordering in a frustrated pyrochlore
antiferromagnet proximate to a spin liquid. Phys. Rev. B 73, 134402 (2006).
88. Lee, S. H. et al. Emergent excitations in a geometrically frustrated magnet. Nature 418,
856–858 (2002).
208
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
89. Tristan, N. et al. Geometric frustration in the cubic spinels MAl2O4 (M=Co, Fe, and Mn).
Phys. Rev. B 72, 174404 (2005).
90. Bergman, D., Alicea, J., Gull, E., Trebst, S. & Balents, L. Order by disorder and spiral spin
liquid in frustrated diamond lattice antiferromagnets. Nature Phys. 3, 487–491 (2007).
91. Gardner, J. S. et al. Cooperative paramagnetism in the geometrically frustrated pyrochlore
antiferromagnet Tb2Ti2O7. Phys. Rev. Lett. 82, 1012–1015 (1999).
92. Nakatsuji, S. et al. Metallic spin-liquid behavior of the geometrically frustrated Kondo
lattice. Phys. Rev. Lett. 96, 087204 (2006).
93. Senthil, T. & Fisher, M. P. A. Fractionalization in the cuprates: detecting the topological
order. Phys. Rev. Lett. 86, 292–295 (2001).
94. Wynn, J. C. et al. Limits on spin–charge separation from h/2e fluxoids in very underdoped
YBa2Cu3O6+x. Phys. Rev. Lett. 87, 197002 (2001).
95. Norman, M. R. & Micklitz, T. How to measure a spinon Fermi surface. Phys. Rev. Lett. 102,
067204 (2009).
96. Lee, P. A., Nagaosa, N. & Wen, X. G. Doping a Mott insulator: physics of high-temperature
superconductivity. Rev. Mod. Phys. 78, 17–85 (2006).
The main subject of this paper is superconductivity, but it provides a good summary of the
status of the theory of QSL states.
97. Wawrzyńska, E. et al. Orbital degeneracy removed by charge order in triangular anti-
ferromagnet AgNiO2. Phys. Rev. Lett. 99, 157204 (2007).
98. Podolsky, D., Paramekanti, A., Kim, Y. B. & Senthil, T. Mott transition between a spin-liquid
insulator and a metal in three dimensions. Phys. Rev. Lett. 102, 186401 (2009).
99. Senthil, T. Theory of a continuous Mott transition in two dimensions. Phys. Rev. B 78,
045109 (2008).
100. Smith, D. F. et al. Dzialoshinskii–Moriya interaction in the organic superconductor
κ-(BEDT-TTF)2Cu[N(CN)2]Cl. Phys. Rev. B 68, 024512 (2003).
101. Pesin, D. A. & Balents, L. Mott physics and band topology in materials with strong spin–
orbit interaction. Preprint at 〈http://arxiv.org/abs/0907.2962〉 (2009).
Acknowledgements I am grateful to countless friends and collaborators, especially
O. Starykh and M. Fisher for many stimulating interactions. I would also like to
thank R. Coldea, H. Takagi, A. Loidl and P. Mendels for sharing their insights into
experiments on frustrated magnets. My research is supported by the US National
Science Foundation, the US Department of Energy, the David and Lucile Packard
Foundation and the US Army Research Office.
Author Information Reprints and permissions information is available at www.
nature.com/reprints. The author declares no competing financial interests.
Correspondence should be addressed to the author (balents@kitp.ucsb.edu).
NATURE|Vol 464|11 March 2010|doi:10.1038/nature08918 REVIEW INSIGHT

Electron liquids and solids in one dimension

Vikram V. Deshpande1, Marc Bockrath2, Leonid I. Glazman3 & Amir Yacoby4

Even though bulk metallic systems contain a very large number of strongly interacting electrons, their properties
are well described within Landau’s Fermi liquid theory of non-interacting quasiparticles. Although many higher-
dimensional systems can be successfully understood on the basis of such non-interacting theories, this is not
possible for one-dimensional systems. When confined to narrow channels, electron interaction gives rise to such
exotic phenomena as spin–charge separation and the emergence of correlated-electron insulators. Such strongly
correlated electronic behaviour has recently been seen in experiments on one-dimensional carbon nanotubes
and nanowires, and this behaviour challenges the theoretical description of such systems.

According to Landau’s Fermi liquid theory1, a bulk metallic system of
interacting electrons has low-energy excitations called quasiparticles
that behave much like weakly interacting fermions. Therefore, many
experiments on such systems can be understood on the basis of theories
of non-interacting quasiparticles. The theoretical picture is qualitatively
different in materials in which the electrons are confined to one-dimen-
sional (1D) channels, or wires. In such wires, interactions cause electrons
to behave in a highly cooperative way, and the Fermi liquid theory no
longer applies. Progress in obtaining 1D materials realized as individual
nanostructures has provided highly tunable systems that allow the study
of the effects of interactions in many-electron systems. Other 1D tunable
systems, such as those based on cold atoms2,3, are also being developed
and keep producing interesting results (see, for example, refs 4 and 5).
Here, we discuss the many ways in which interacting electrons in one
dimension often produce surprising results that illuminate the role of the
dimension of an interacting many-body system. After discussing the basic
manifestations of 1D cooperative electron behaviour in simple tunnelling
experiments, we focus on a number of new developments in 1D physics
both in theory and in experiments on nanowires and carbon nanotubes.
We first discuss the 1D phenomenon of spin–charge separation and exper-
iments to observe it using the technique of momentum-conserved tun-
nelling between parallel nanowires6,7. The same kind of experiment raises
the question of how a charge with a definite momentum is accommodated
in a correlated 1D system8; this question is discussed in a section devoted
to the apparent charge ‘break-up’. The above physics can be understood
using a theory valid at sufficiently low energies; however, the question of
how 1D systems behave at higher energies is interesting as well. We discuss
the recently developed ‘nonlinear Luttinger liquid’ theory addressing this
question9–13 and possible experiments aimed at its verification.
The availability of clean, freely suspended carbon nanotubes has ena-
bled experiments to be performed in the low-electron-density regime.
We discuss this limit, in which electrons organize themselves in a state of
electron matter resembling a solid rather than a liquid14. Finally, we discuss
recent observations of electrons in carbon nanotubes showing behaviour
similar to that of a Mott insulator15, which is a special kind of insulating
state induced by electron–electron interactions.
From Fermi liquids to Luttinger liquids
Coulomb forces acting between electrons and ions hold in place the
atoms forming a metal and ultimately determine the low-energy
spectra of its elementary excitations, electrons in conduction bands and
phonons. Despite the strength of the Coulomb forces (the corresponding
energy is in the range of 10 eV), the classification of the low-energy
excitations, suggested by Landau’s Fermi liquid theory1, is remarkably
universal. In the language of this theory, the electric charge carriers in
a conventional metal can be viewed as almost-free quasiparticles with
Fermi statistics, the Coulomb interaction ultimately determining the
crystalline structure of a material and the specific form of the quasi-
particles’ dispersion relationship (their energy–momentum relation-
ship). The residual interaction between quasiparticles, stemming from
the Coulomb repulsion, leads to the quasiparticle states having a finite
energy uncertainty; the uncertainty for a given quasiparticle, however,
is small in comparison with its energy.
Box 1 | Luttinger liquid and bosonization in one dimension
In our discussion about the inapplicability of the Fermi liquid theory to
describe 1D interacting electron systems, shifts of the electron fluid
have a crucial role. The creation of a particle at some point, x, along the
wire may be viewed as a result of a shift of all the fluid to the left, say,
of x by the interparticle distance, while the fluid to the right of x stays
put; the ‘bump’ of density created at point x in this way will contain just
one extra particle. The displacements of the fluid may be viewed as a
superposition of harmonic waves. The quantized displacement waves
are bosons that can be thought of as quanta of sound waves. At low
energies, the interaction between bosons can be disregarded, which
allows a low-energy excitation of a 1D fluid to be considered a collection
of free bosons.
In one dimension, it is the above-mentioned bosons, rather than the
Fermi quasiparticles of the Fermi liquid theory, that are the well-defined
elementary excitations. This simple representation of the low-energy
excitations of a 1D many-body quantum system is known as a Luttinger
liquid. The lower the energy scale, the better this representation is
supposed to work. It becomes exact for the so-called Tomonaga–
Luttinger model38,75,76, which is an idealized system of interacting
fermions with a linear dispersion relation (this may be pictured as a
model of electrons with some fixed Fermi velocity but with an infinitely
deep conduction band). The operator that creates an electron at point x
in a Luttinger liquid is, in the ‘bosonized’ representation, proportional to
an operator that causes the finite displacement of a huge portion (one-
half) of the liquid. Thus, in terms of the bosonic variables, the electron
operator is a very non-local one. That inconvenience is a small price to
pay to be able to recast a Hamiltonian of interacting electrons in the
form of non-interacting particles38,75,76 (bosons).

1
Department of Physics, Columbia University, New York, New York 10027, USA. 2Department of Physics and Astronomy, University of California, Riverside, California 92521, USA. 3Departments
of Physics and Applied Physics, Yale University, New Haven, Connecticut 06520, USA. 4Department of Physics, Harvard University, Cambridge, Massachusetts 02138, USA.

209
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
a b
10–6
Current (A)
10–8
10–10
1/T 1+α (a.u.)
10–12
0.01 0.1 1 105 K
Bias (V)
0.001 0.01 0.1 1 10
γeV/kBT
Figure 1 | Power-law behaviour of electron tunnelling with energy for
molybdenum selenide wires of various diameters. a, Current–voltage data
at different temperatures, T, for a molybdenum selenide wire bundle, scaled
to a universal curve. Inset, the same data, but unscaled29. The parameter γ
is a constant that indicates how the voltage drop across the device is shared
Although Landau’s Fermi liquid theory provides a satisfactory account
of bulk three-dimensional (3D) metals, the situation is qualitatively
different in materials in which the electrons are confined to channels
that are narrow relative to the electron wavelength and as a result effec-
tively act as 1D wires for the electrons. Because the electrons are con-
strained to move in only two directions, right or left along the wire, any
perturbations will affect a large number of electrons collectively. For
instance, a charged particle trying to tunnel into the wire may create a
short-wavelength density perturbation, forcing the liquid of interacting
electrons to shift quickly away from the incoming particle. To accom-
modate the extra particle, however, ultimately all electrons to the left of it
must shift by one-half of the average interparticle distance along the wire
— in analogy to displacing equidistant beads on a string to make room
for an additional bead. Likewise, all liquid to the right of the tunnelling
particle must shift to the right by the same amount. This is in contrast to
the 3D case, in which the charge of an incoming particle is compensated
by a swift redistribution of the itinerant electrons, which are free to move
away from the incoming electron in all directions, thus causing less of a
collective disturbance of the so-called Fermi sea of electrons.
An immediate consequence of this intuitive implication of the role of
collective behaviour when confining charged particles to one dimension
is that the tunnelling rate for particles entering a 1D wire is suppressed
around zero bias — the so-called zero-bias anomaly. By contrast, the
tunnelling rate into a clean, conventional 3D metal is featureless (nei-
ther suppressed nor enhanced) at zero bias, confirming the unimpeded
creation of particles or holes close to the Fermi level. This qualitative
picture of tunnelling into a 1D electron system can be quantified within
the so-called Luttinger liquid theory, ‘bosonization’ being a practical
B nanowires that allow variation of the number of channels available for
d I
Figure 2 | Momentum-resolved tunnelling spectroscopy using two parallel
wires. Electrons or holes tunnelling (arrows) between two parallel wires a
distance d apart. The momentum boost in the tunnelling event, Δp = edB,
is achieved by application of a magnetic field, B, perpendicular to the
wires (into the page in this example). The energy resolution is achieved by
controlling the interwire bias, V.
210
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
10
5
35 K
55 K
70 K
95 K
α
125 K
140 K
160 K
180 K 1
200 K
Fit
0.5
100 0 5 10 15 20
Diameter (nm)
between wire segments in series between the electrodes. a.u., arbitrary
units; e, elementary charge; kB, Boltzmann constant. b, Power-law exponent,
α, as a function of wire diameter; 1/α scales as the square root of the number
of channels, in agreement with theory19. Error bars, diameter uncertainty.
(Figure reproduced, with permission, from ref. 29.)
tool that makes the theory remarkably simple (Box 1).
In experiments, the zero-bias anomaly can be observed in the
differential conductance, dI/dV, where I is the current and V the volt-
age, of a tunnel junction attached to a 1D wire; a signature of Luttinger
liquid behaviour would be a deep minimum at zero bias (V = 0). This
discussion also applies to tunnelling into (or from) an end of a wire.
The only difference is that now the electron liquid must shift away from
the end by the entire interparticle distance; tunnelling is therefore sup-
pressed more strongly.
A specific signature of Luttinger liquid behaviour predicted by the-
ory16–19 is that the zero-bias anomalies of dI/dV for tunnelling into a
point at an end or into a point within the ‘bulk’ of a wire take the form
of a power law in V (ref. 16): dI/dV ≈ V α. Similarly, the temperature
dependence of the tunnelling conductance at small, fixed bias also
must follow a power law. The exponent α depends on the inter-electron
interaction strength and the number of channels19. The quantization
of electron motion across the wire’s width results in the wire having a
finite number of channels. The wider the wire, the larger the number of
channels and the smaller the electron shift necessary to accommodate
an extra electron, leading to the decrease of α towards zero (the conven-
tional case of tunnelling into a clean conductor in higher dimensions).
These power-law dependences were specifically predicted for17,18,
and measured in, carbon nanotubes, for tunnelling both into an end
and into the bulk20–22. The experimental results20–22, notwithstanding
the difficulties in interpretation associated with the finite-size effects
and zero-bias anomalies in the resistive leads23–25, are considered to be a
clear manifestation of Luttinger liquid physics. Further confirmation of
the Luttinger liquid behaviour came from resonant tunnelling experi-
ments with gallium-arsenide-based quantum wire dots26 and nanotube
dots27, nanotube capacitance measurements28 and molybdenum selenide
V electron motion29 (Fig. 1).
Despite the agreement between the tunnelling anomaly measurements
and theory, sceptics may consider it to be circumstantial evidence of the
unusual behaviour of electrons associated with their confinement to one
dimension. Indeed, the presence of disorder typical for highly resistive,
conventional (non-1D) conductors also impedes the spreading of charge
of tunnelling electrons. Because both Luttinger liquids and such higher-
dimensional systems display a zero-bias anomaly23–25, perhaps the pro-
posed effect could be confused with the zero-bias anomaly in resistive
leads. It would help to know the range of biases over which the power
law for tunnelling conductance is expected to hold. An experiment on
multiwalled carbon nanotubes produced results hardly distinguishable
from the hallmark experiments20–22, with a power law observed over a
NATURE|Vol 464|11 March 2010
Box 2 | Spin density waves in a Luttinger liquid
Electron fluids support waves of spin density, as well as waves of charge
density. To see this, first consider a free-fermion gas, neglecting the
interparticle interaction. Imagine that in some portion of the 1D system,
an excess of spin-up particles is created and is compensated, so as
to maintain charge neutrality, by the deficit of spin-down ones (the
up–down axis in spin space may be picked arbitrarily). Even if the total
density of particles remains unchanged, the energy of the system will
increase, as locally the chemical potentials of the spin-up and spin-
down particles deviate from the common equilibrium value. This energy
increase provides the stiffness that defines the speed of the spin mode.
In the free-fermion gas, the speeds of mass density and spin density
modes both equal the Fermi velocity. The phases of oscillations of the
local Fermi levels of the spin-up and spin-down electrons are opposite
to each other and thus have no effect on the charge density (see figure,
where EF↑ is the Fermi energy for spin-up electrons, EF↓ is the Fermi
energy for spin-down electrons and x denotes position along the wire).
Therefore, a spin wave does not create a perturbation of charge density,
suspiciously wide range of biases30. Another uncertainty comes from
the limitations caused by finite-size effects in the first experiments20–22,
which, for example, made it impossible to distinguish between the
conducting and insulating states of a nanotube in the limit of very low
temperatures. The possibility of ‘dielectrization’, that is, formation of an
insulating state with a gap in the electron spectrum at low energies, was
considered theoretically18,31–34. The single-point tunnelling phenomena
described above gave only the first glimpse into the behaviour of elec-
trons confined to one dimension.
Spin–charge separation
Another distinct manifestation of the correlated behaviour of electrons
when confined to one dimension is the phenomenon of spin–charge
separation. The electron liquid, along with the waves of electron charge
density, supports waves of electron spin density. These are collective
excitations, propagating with two different velocities (Box 2).
The successful observation and direct visualization of spin–charge
separation has been achieved in experiments on so-called momentum-
resolved tunnelling6. In such experiments, the aim is to extract an elec-
tron from the wire with a given momentum, p. Such an electron would
not be associated with a unique energy, ε(p), owing to the interactions
in the liquid, and would therefore leave behind multiple excitations (as
above, the same considerations apply for holes and particles and, thus,
for tunnelling from and into the wire). The probability of forming a
state with energy ε if the added/extracted electron has momentum p is
quantified by the spectral function, A(p, ε). In 1D wires, this information
can be obtained from momentum-conserving tunnelling between two
a c 12
10
Magnetic field (injected momentum) (T)
8 Char
b 6
IL IR
VSD
6 4
d subtracted from the data. The colour scale indicates the magnitude of
4 0
1 at zero energy at a magnetic field of around 7 T. d, Measured spin velocities
vF /v
2
0 0
0 120
Carrier density (μm–1)
REVIEW INSIGHT
EF↑, EF↓
x
and a weak but long-range charge–charge interaction hardly affects
the spin velocity, vs, but increases the velocity of the charge mode, vρ.
Note that the presence of interactions makes the waves of charge and
spin truly collective excitations: they propagate with different velocities
and therefore cannot be associated with a motion of a single particle
endowed with the two qualities. The accommodation of a tunnelled
electron requires charge and spin displacements, which propagate along
the electron fluid with unequal velocities.
parallel 1D electron channels formed in a semiconductor field-effect
device6–8,35. In such a device, the momentum of the tunnelling electron
is tuned using a magnetic field, B, applied perpendicular to the wires,
and the energy resolution is achieved by controlling the interwire bias,
V (Figs 2 and 3a), yielding information about the electron spectral func-
tion, A(p, ε), through the interwire current, I(B, V). A complication of
this technique in comparison with, for example, angle-resolved photo-
emission spectroscopy36,37 is the unavoidable electrostatic interaction
between the wires. Because of this interaction, some effort is necessary
to extract A(p, ε) from the measurement of I(B, V) (refs 6, 7).
Ideally, spin–charge separation would manifest itself as two peaks in the
energy dependence of A(p, ε) at a fixed value of p. For definiteness, we con-
centrate here on the case of p values slightly less than the Fermi momen-
tum, pF, and ε values less than zero, that is, on the probability of extracting
a particle from a Luttinger liquid. Then, the two peaks signify the two
extremes in sharing of the momentum by the elementary charge and spin
excitations (holons and spinons, respectively38). The high-energy peak
(with the largest value of |ε|) is located at ε = vρ(p − pF) and corresponds
to the entire momentum being given to a holon. The low-energy peak,
at ε = vs(p − pF), corresponds to the entire momentum being carried away
by a spinon (Fig. 3). This peak actually corresponds to the lowest-energy
state of the Luttinger liquid, which may be created by removing from it a
particle with momentum p; A(p, ε) must equal zero for any smaller value of
|ε|. For |ε| values slightly above the threshold, |vs(p − pF)|, a small part of the
hole’s energy is spent on the creation of charge waves with small wavevec-
tors, |q| << |vs(p − pF)|/vρ, and the greater part is carried by a spinon.
Luttinger liquid theory predicts that the two described peaks in the
Figure 3 | Probing spin–charge separation and charge fractionalization
14 in interacting 1D wires using momentum-resolved tunnelling
spectroscopy. a, Schematic structure of the double-wire geometry, made
12
using the cleaved-edge overgrowth technique7. The double-wire system
10 resides at the edge of a gallium arsenide/aluminium gallium arsenide
Sp
double-well structure. Dark green, gates; blue, AlGaAs; white, GaAs; light
in
8 green, 1D wire segments; transparent box, epitaxial confining layer on the
ge
cleaved edge. b, Schematic of circuit used to measure the spectral function
and charge fractionalization. IL, left-moving current; IR, right-moving
current; VSD, bias voltage. c, Measured spectral function, ∂I(V, B)/∂V,
indicating the spin and charge excitation branches (plotted versus injection
2 energy, eV, on the horizontal axis). A smooth background has been
∂I(V, B)/∂V (arbitrary units). The features labelled spin and charge cross
–2
(open symbols) and charge velocities (filled symbols) for different carrier
–4 densities (colours), plotted as the Fermi velocity, vF, divided by the measured
velocity, v. A lower carrier density corresponds to stronger repulsion
–6 between electrons and hence to a larger charge excitation velocity. The
–15 –10 –5 0 5 spin modes are nearly independent of the carrier density. Solid curves are
Injection energy (meV) theoretical fits. (Panel reproduced from ref. 6.)
211
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
a Q
p → pF
Q/2 Q/2
υpl → ∞
b Q
p → pF
(1 – f)Q fQ Q
L momentum as a good quantum number7. Electrons with given momenta
υρ = υF υρ > υF υ ρ = υF
Figure 4 | Spreading of charge on tunnelling into an interacting wire. a, An
electron liquid in a homogeneous infinite wire in the absence of shielding
supports plasmons that transfer the tunnelled charge to the outer parts
of the system; in the limit vpl/vF → ∞, the initial momentum of the charge
does not affect the pattern of transient charges. b, Transient charges at a
finite charge mode velocity, vρ, in the presence of free-fermion leads. The
charge pulse splits into transmitted and reflected components (dashed) at
the boundary of two media with different ‘refraction indices’ for the charge
waves. In the set-up shown, multiple reflections at the two boundaries result
in the entire tunnelled charge, Q, being transferred into the right-hand lead
and zero charge exiting into the left-hand lead.
electron spectral function are power-law singularities, with exponents
defined by the strength of the inter-electron interaction39. Singular points
in the energy–momentum plane form one pair of lines, ε = vρ(p − pF) and
ε = vs(p − pF), that intersect at (0, pF), and a similar pair that intersect at
(0, −pF). In the experiment (Fig. 3), the momentum of a tunnelling elec-
tron is tuned using B and its maximal energy is determined using V. The
intersecting straight lines tracing the singularities of A(p, ε) correspond
to the intersecting straight lines tracing the singularity in the differential
conductance in the V–B plane.
Measurements6,7 reveal clear maxima of ∂I(V, B)/∂V along a system of
lines in the V–B plane, consistent with theory40. Two of the lines indeed
form a crossing (see Fig. 3c, where the injection energy, eV, is plotted
against B), at a point that can be unambiguously identified with the tun-
nelling of electrons with momenta approaching pF. The velocities of the
corresponding excitations can be extracted from the slopes of the two
lines at the crossing point. These velocities are in reasonable agreement
with theoretical estimates for the velocities of charge and spin modes
(Fig. 3d). Current measurements are not accurate enough for to tell
whether ∂I(V, B)/∂V is actually divergent along the lines corresponding
to the spectra of the two modes.
Further recent work exploiting a similar idea, based on tunnelling
between a wire array and a two-dimensional electron gas, also found
multiple excitation velocities consistent with theory41. Taken together,
experiments6,7,41 provide direct confirmation of the phenomenon of
spin–charge mode separation in 1D electron systems at low energies.
Although there is a qualitative agreement with the Luttinger liquid
theory for momenta approaching pF , the experiments on momentum-
conserving tunnelling have sharpened a question that is not addressed
within the Luttinger liquid theory, that of what happens to spin–charge
separation at higher energies. The two experiments6,7 do demonstrate
that a remnant of the charge mode spectrum survives at higher bias, in
the form of a parabolic dispersion relationship similar to a free-particle
excitation. However, traces of the spin mode fade away at higher ener-
gies. Finding a theoretical answer to the question requires accounting
for the nonlinearity of the generic electron spectrum, which necessitates
going beyond the Tomonaga–Luttinger model, in which the linearity of
the spectrum is a central assumption. Some suitable tools for develop-
ing a theory of a liquid of quantum particles with a generic dispersion
relationship (a nonlinear Luttinger liquid) were found only recently9,42,43
212
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
and several experimentally testable predictions have already been made,
among them the survival of the power-law singularity of A(p, ε) to higher
energies for certain modes43 and particle–hole asymmetry in the relaxa-
tion rates10. The second of these predictions can be tested using exist-
ing6 experimental techniques. We return to the question of spin–charge
separation at higher energies below.
Charge polarization and break-up
In the discussion of momentum-resolved tunnelling, we considered an
idealization of an infinitely long junction between two parallel wires. In
practice, the junction typically has a length of few micrometres, which
is long on the scale of the Fermi wavelength and does not limit the use of
are injected near the middle of a long quantum wire (Fig. 3b). This poses
the question of how the injected charge spreads once it is in the wire. Pre-
dictably, the answer depends on the circumstances, leading to interesting
considerations related to the break-up, or ‘fractionalization’, of charges.
First, look at the ideal case of a quantum wire in the absence of any
other conductors that may shield the electric field produced by charges
in the wire. In this case, the dispersion relationship for the charge mode
is ω(q) ≈ q√[ln(1/wq)], where w is the wire’s width and q is the wavevector
of the charge excitation, or plasmon. If the tunnelling electrons have
low energies (such that their momenta are close to ±pF), then only low-
energy plasmons can be excited. Such excitations have small wavevectors
(q → 0) and, thus, diverging velocities, vpl ≈ √[ln(1/wq)] → ∞. Relative
to the plasmons, the velocity of a tunnelling electron, v → vF (where vF
is the Fermi velocity), is so small that its direction is inconsequential;
tunnelling of a charge Q with a well-defined momentum thus creates
two pulses, each of charge Q/2, running in opposite directions (Fig. 4a).
Although this is somewhat counterintuitive, it is no different from the
dynamics of charge screening in higher dimensions: in a conducting liq-
uid, a polarization charge develops that locally compensates the charge
of the tunnelling electron and spreads it towards the edges of the system.
An electron tunnelling into a two-dimensional electron liquid, for exam-
ple, would create a circularly symmetric ‘splash’ of polarization charge.
Shielding of the long-range Coulomb forces reduces the infinite (for
q → 0) plasmon velocity, vpl, to a finite velocity, vρ, of charge wave propaga-
tion. For a finite ratio vρ/vF, the waves of polarization moving in opposite
directions away from the tunnelling point differ from each other (Fig. 4b).
The fraction, f, of the tunnelling charge, Q, that is carried by the wave
propagating in the direction of the incident electron motion can be found
from the momentum conservation law (for pedagogical reasons, it is bet-
ter to consider the case in which Q >> e). The momentum of the tunnelling
charge, (Q/e)mvF, is carried away by particles of the same mass, m, moving
in opposite directions with velocity vρ. The momentum conservation law
takes the form fvρ − (1 − f)vρ = vF, yielding f = ½(1 + vF/vρ); in the absence of
interactions (vρ = vF), the expected result, f = 1, is obtained.
We should emphasize the distinction between the notions of charge
quantization and fractionalization. Nowhere in the above discussion
was a particular value of the tunnelling charge, Q, important, consistent
with the idea of a liquid as a description of a continuous medium. The
quantized value of the charge of the particles forming the liquid must
come from a separate, microscopic consideration. (For example, the
quantized fractional charge of quasiparticles under the conditions of
the fractional quantum Hall effect is defined by the special properties
of the corresponding microscopic wavefunction44.)
This discussion of charge fractionalization is applicable as long as the
polarization charges do not hit the ends or some other inhomogeneity of
the wire. In principle, a direct determination of f can be achieved from a
time-resolved measurement of the transient currents in the wire or from
a measurement of the current noise spectrum45. However, d.c. measure-
ments cannot determine f. This limitation is illustrated here in the simplest
geometry of free-fermion leads attached to a Luttinger liquid of length L.
The contact with the leads may be modelled as a smooth spatial variation
of the charge wave’s speed from vρ to vF. Being smooth on the length scale
set by the Fermi wavelength, such an inhomogeneity does not backscatter
individual electrons; by analogy with optics, it rather acts as an interface of
NATURE|Vol 464|11 March 2010
two media with different refraction indices for the charge waves (Fig. 4b).
To analyse the d.c. measurement, consider the propagation of charges
over times much greater than that, L/vρ, needed to traverse the Luttinger
liquid. The propagation occurs in the free-fermion part of the system. As
the electrons do not backscatter, the momentum conservation law can be
applied as was done above, but with the replacement vρ → vF. This leads to
the conclusion that the entire tunnelled charge eventually ends up moving
in the direction of the initial electron’s momentum, regardless of the value
of f within the Luttinger liquid (Fig. 4b).
The concept of charge fractionalization provides a convenient basis for
calculations within Luttinger liquid theory. Experiments performed at low
injection energy, using the more sophisticated configuration of a three-
terminal d.c. measurement (Fig. 3b) with momentum-selective source and
drains attached to the wire, are in excellent agreement with theory8. Such
d.c. measurements can also be analysed within the framework of charge
fractionalization, but they do not allow determination of f.
A natural question to ask is how the asymmetry between currents
in the two drains (Fig. 3b) behaves at higher injection energies, once
the conventional Luttinger liquid theory breaks down. Answering this
question may provide information about the differences between the
kinetics of injected high-energy particles and holes.
Nonlinear Luttinger liquid
To analyse the above-mentioned experiments, aimed at measuring
the spectral functions of electrons and holes at higher energies and
at elucidating how the injected charge propagates along the wire at
these higher energies, it is necessary to go beyond conventional Lut-
tinger liquid theory. There is also an impetus to do so from within the
theory. The linearization of the particles’ spectrum that is assumed in
the Tomonaga–Luttinger model brings an immense degeneracy into
the energy spectrum of the excited many-body states: the energy of a
state composed of an arbitrary number of particle–hole pairs residing
on one branch of the single-particle spectrum depends only on the total
momentum of the pairs and is insensitive, for example, to the number of
pairs. Reintroduction of the generic nonlinear dispersion relationship of
the particles into the theory removes this degeneracy in a subtle way.
Recently, the nonlinear theory was developed for spinless quantum
particles9–11, and since then a first step has been made in generalizing the
dynamic response theory to spin-½ fermions with a generic dispersion
relationship12,13. The new methods make it possible to move beyond
the linear Luttinger liquid description in understanding the dynamic
response functions (such as the spectral function, A(p, ε)). They also
provide a basis for understanding the kinetics of interacting quantum
particles confined to one dimension, although the full kinetic theory
has not yet been constructed. The new theory predicts the persistence
of the power-law singularities in A(p, ε) even for momenta far from the
Fermi points and allows the evaluation of the corresponding exponents.
In striking contrast to the conventional Luttinger liquid theory, it also
predicts asymmetry in the relaxation rates for particles and holes.
Here we demonstrate the main idea of the new theory using the
example of weakly interacting fermions without spin. Free particles have
a parabolic dispersion relationship, ε(p) = p2/2m, and here it is assumed
that an equilibrium population of fermions occupies states with |p| ≤ pF.
The focus is on evaluating A(p, ε) for the hole excitations (ε < 0). In the
absence of interactions, only one state of the system is created by extract-
ing an electron with some momentum, p0 (−pF ≤ p0 ≤ pF), in the Fermi
distribution; this state consists of a single hole with momentum p0. The
velocity of this hole, |p0|/m, is less than vF, which is the velocity of the
lowest-energy particle–hole excitations. Therefore, interaction of this
hole with the particles of the Fermi sea does not lead to creation of excita-
tions, in much the same way that a particle moving at less than the speed
of light in a medium does not emit Cherenkov radiation. This means that
the existence of the energy threshold in A(p, ε) along some line, ε(p), at
a momentum |p| ≤ pF is robust with respect to the interaction (Fig. 5). At
the same time, the functional form of A(p, ε) is modified by interaction
away from its free-fermion limit, A(p, ε) ≈ δ(ε − ε(p)) (where δ denotes the
delta function). Remarkably, this modification is universal: the threshold
© 2010 Macmillan Publishers Limited. All rights reserved
REVIEW INSIGHT
behaviour of A(p, ε) is described by a power-law function of ε(p) − ε, as
in the Tomonaga–Luttinger model. Unlike in that model, however, the
power-law behaviour at the threshold is valid at any momentum, and the
exponent of the power law depends on the momentum.
To explain the power-law form of A(p, ε), here we allude46 to the effect
known as the Anderson orthogonality catastrophe47. For definiteness,
consider the part of A(p, ε) associated with the creation of a hole in the
system. The introduced hole produces a potential acting on other par-
ticles. The threshold in A(p0, ε) corresponds to the ground-state energy
of the many-body system adjusted to that potential. The ‘old’ (pre-hole)
and ‘new’ many-body ground states have zero overlap47. The overlap
with excited states is finite and scales as some power of the excitation
energy (measured relative to the new ground state). This non-zero over-
lap with excited states of the Fermi sea results in the redistribution of the
spectral weight in A(p0, ε) from a single point (ε = ε(p0)) to all energies
above the threshold, according to a power law in ε(p0) − ε.
This picture, which was initially developed for weakly interacting
particles, has been extended into a fully consistent phenomenological
theory9,43 valid at arbitrary interaction strength. The phenomenology
relates the threshold exponents of a dynamic correlation function to the
properties of the corresponding threshold energy spectrum; these two
types of characteristic can be measured independently of each other.
The limit of weakly interacting fermions also demonstrates the dif-
ference between the particles and holes. Particles move at velocities
greater than vF (Fig. 5), allowing the ‘Cherenkov radiation’ of particle–
hole pairs; unlike holes, energetic particles do relax, producing multiple
particle–hole pairs. Although particle–hole asymmetry has not yet been
observed, experiments to explore such deviations from Luttinger liquid
theory are underway.
Developing the nonlinear Luttinger liquid theory of spin-½ fermions is
of special interest. Such a theory should cover the case of strong interac-
tions keeping fermions ‘in place’ and thus suppressing their spin-exchange
interaction and creating an almost flat band of spinon excitations.
Correlated-electron solid at low density
In the previous sections, we discussed theory and experiment in the
regime where the charge density is large enough to ignore the strong
correlations between the electron positions that develop when the elec-
tron density, n, becomes sufficiently low. These correlations emerge
because although the long-range part of the electron–electron interac-
tion chiefly affects the value of vρ, as this depends on the electrostatic
energy required to create long-wavelength charge fluctuations, the
increase of its short-range part (~e2n) relative to the single-electron
ε
pF
p
Figure 5 | Effects of the finite mass of particles forming the nonlinear
Luttinger liquid. At positive curvature of the energy–momentum function,
ε(p) (solid black line), the velocity of a hole is less than vF and therefore a
hole state is stable, unlike a particle state (the particle spectrum is shown
as the dark-grey broadened line). The spectrum of holes (dashed black
lines) forms the threshold for the spectral function, A(p, ε), which is finite
everywhere in the light-grey shaded area. Creation of a ‘deep’ hole (lower
open red circle) perturbs the Fermi sea, creating low-energy particle–hole
pairs (open and filled red circles near the Fermi point); this process results
in the power-law threshold singularities in A(p, ε) discussed in the text.
213
INSIGHT REVIEW
a (ref. 57). The periodically ordered electrons forming the Wigner crystal
B
b Ground-state configuration
Region III Region II Region I
c G (μS)
0 >10.0
Hole number
40 30 20 10
B (T)
5
III II
0
–4.0 –3.0 –2.0 –1.0 0 1.0
Vg (V)
Figure 6 | Charges at low densities in a carbon nanotube form a 1D Wigner
crystal. a, Although the sense of motion of electrons around the nanotube
(red and blue arrows) depends on the direction of the magnetic field, B, in
an armchair tube, in a tube of general chirality electrons with both senses
of motion can be present. The sense of motion for each charge can be
labelled using an isospin quantum number. b, Ground-state configurations
of charges in the regions (I–III) described in the text. c, Differential
conductance, G, of a nanotube as a function of gate voltage, Vg, and
magnetic field. The vertical stripes are the traces of the conductance peaks.
(Panel reproduced, with permission, from ref. 14.)
kinetic energy (~"n2/m, where " denotes Planck’s constant divided by
2π and m denotes the electron mass) tends to suppress the electrons’
ability to pass by each other. This produces a 1D Wigner-crystal-like
structure with periodically ordered charges48–52. (Note that although
perfect charge ordering in one dimension is not expected, the correla-
tion length can readily exceed the sample length in realistic situations,
yielding a state that is much more like an electron solid than an electron
liquid.) For e2n >> "n2/m, the positional exchange between neighbour-
ing electrons becomes exponentially rare and the fluctuations of their
relative positions become small (the condition can be recast in the form
naB << 1, where aB is the effective Bohr radius for a conduction electron
in the material). The suppressed exchange decreases the spin mode
stiffness, lowering the energy cost of spin excitations. If exchange is
ignored entirely, the spin system becomes disordered for any positive
temperature. This so-called spin-incoherent regime has been a topic of
much recent theoretical interest53.
Studying the properties of such a system requires low-disorder
samples to avoid an insulating state where disorder pins the Wigner
crystal. Potential clean systems in which to study such behaviour include
wires made using cleaved-edge overgrowth35 and semiconducting car-
bon nanotubes54. In particular, advances in sample preparation for car-
bon nanotubes have allowed the measurement of as-grown, suspended
nanotubes, minimizing the tube–substrate interaction and producing
very clean and low-disorder 1D systems that may reveal Wigner crystal
physics. As discussed above, one of the hallmarks of a Wigner crystal
state is that the antiferromagnetic exchange energy, J, is exponentially
suppressed at low densities. This has an interesting consequence for
the spin polarization in an external magnetic field. In contrast to free
electrons, for which full polarization occurs only if the Zeeman energy,
gμBB (where B is the magnetic field, μB is the Bohr magneton and g is
the g factor), of an electron spin exceeds the Fermi energy, in a Wigner
crystal polarization occurs at exponentially smaller fields, as only the
condition gμBB > 2J is required.
In addition to their spin, electrons in nanotubes have an extra isospin
degree of freedom, which is an additional angular momentum acquired
by the electrons because of their motion around the axis of the nanotube
(either clockwise or anticlockwise; Fig. 6a). In an axial magnetic field,
the Zeeman effect splits the spin states55,56 and their different orbital
magnetic moments, ±μorb, split the isospin states57. The isospin splitting,
2μorbB, is typically larger than the Zeeman splitting by a factor of ~5–10
214
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
act as a spin and isospin chain. For B > 0, the ground state reflects a com-
petition between the magnetic energy, which favours parallel spins and
parallel isospins, and the exchange energy, which favours neighbouring
electrons having different spin and/or isospin directions14,58.
Because J increases rapidly with n, in a magnetic field it is expected
that there are three different regimes for a Wigner crystal state in a car-
bon nanotube as n and B are varied. In one (region I), the magnetic
energy is dominant and all the electrons are fully spin and isospin polar-
ized. In another (region II), the exchange dominates the Zeeman split-
ting but not the orbital splitting, which is larger, and only the isospin is
1 polarized. In the last (region III), the exchange energy is dominant and
the polarization is minimal for the given number of electrons. Figure 6b
I shows the ground-state configurations for these states.
Coulomb blockade spectroscopy provides a very useful tool to study
this experimentally. In Coulomb blockade spectroscopy, the device acts
as a single-electron transistor, producing peaks in the dependence of
conductance on gate voltage, Vg, whenever an electron is added to the
dot, thereby changing the electron density. The peak spacing reflects the
energy required to add each electron. Because the magnetic energy contri-
bution to this energy is different for each combination of quantum num-
bers, measurement of each peak’s energy shift as a function of B yields the
charges’ quantum numbers. Figure 6c shows the evolution of the Coulomb
peaks in a magnetic field for a semiconducting carbon nanotube, which
displays all three regimes: in region I, the parallel traces that conductance
peaks produce in the B–Vg plane indicate that each added electron has
the same spin and isospin; in region II, the alternating slopes indicate that
isospin is polarized and that spins alternate; and in region III, the zigzag
pattern indicates the minimum polarization state14.
The yellow line in Fig. 6c shows a single-parameter fit to theory58, with
the fitting parameter within the theoretically predicted range58,59. The
red line is a plot of the boundary between regions I and II made using
the same parameter value. The quantitative agreement with theory is
satisfactory. We note that a model of non-interacting electrons predicts
very few or no spin-polarized electrons under the same conditions14.
The success of the Wigner crystal picture in accounting for the rela-
tive ease of polarizing electron spins in nanotubes in a magnetic field
suggests that testing a theory49 predicting one-half the usual conduct-
ance of a ballistic channel in the spin-incoherent regime would be very
interesting. However, this would require progress in reliably obtaining
nearly perfect contacts to ultraclean semiconducting nanotubes.
Mott insulator
The presence of enhanced spatial correlations between the electrons
due to the interactions (that is, Wigner crystallization) increases the
tendency towards pinning by inhomogeneities. When pinned, an elec-
tron system becomes insulating at sufficiently low temperatures. Sup-
pression of conductance observed in gallium arsenide wires made using
cleaved-edge overgrowth has been tentatively explained35 in terms of
such a pinning phenomenon.
The atomic lattice can be another source of pinning in an interacting
electron system. Because of the lattice periodicity, pinning opens an
energy gap in the electron spectrum. Commensurability between the
mean electron spacing and the lattice period aids the opening of the gap.
When electrons are on a lattice at half-filling (one conduction electron
per atom), such a situation occurs. An example illustrating this is a 1D
lattice-site chain in which a large energy cost, U, exists for double occu-
pancy of a site and where the hopping matrix element, t, between sites
can be considered small. At half-filling, every site is occupied, an elec-
tron–hole pair costs an energy U to create and the system is an insulator.
This is essentially the physics of Mott insulators60. However, in contrast
to 3D systems, where the formation of a Mott gap requires U to exceed
some threshold, in one dimension a gap exists for any ratio U/t > 0.
Although such physics has been studied in bulk materials61, carbon
nanotubes afford the opportunity of potentially studying 1D Mott
physics in an individual nanostructure with readily tunable param-
eters. Indeed, there is a puzzle to be solved — nanotubes predicted to
NATURE|Vol 464|11 March 2010 REVIEW INSIGHT

be metallic in the single-particle description are found experimentally a

dI/dV (μS)
to be gapped. Data taken from a so-called armchair nanotube (Fig. 6a) 25 40

V (mV)
show such a gap (Fig. 7a) at zero doping. The simplest single-particle 0
20
perturbations, such as curvature or strain of an armchair nanotube, are –25 (18, 18)
0
prohibited from opening a gap by symmetry constraints62.
–5 0 5 10
Such a gap may possibly be due to the many-body effect in Mott Vg (V)
insulators. A number of works have included electron–electron inter-
actions in theories of carbon nanotubes and found that the interac-
tions should open a gap at half-filling18,31–34. The structure of a general b
U
nanotube is determined by a pair of integers (N, M), which specifies the U/N
tube’s chirality and radius, r. For N = M, an ‘armchair’ nanotube (Fig. 6a), t t
the honeycomb lattice can be mapped onto an equivalent 1D problem t⊥ t t
known as the two-leg ladder31 (Fig. 7b). When long-ranged interactions t⊥
are taken into account, theories predict an energy gap of Δ ≈ 1/r β for
metallic nanotubes, with β ranging from β = 1/(1 − g) to β = 2 (ref. 63),
depending on detailed assumptions. Here g = vF/vρ ≈ 0.2–0.3 for nano-
tubes18,33 and is a measure of the strength of the interactions therein.
Experiments show an approximate 1/r1.3 dependence15, in agreement c Gap
with the theoretical range of predicted values of the exponent. Holes Electrons
Nevertheless, even in the most robust case of an armchair nanotube it 5 30
remains possible that the observed gap originates from another single-

G (μs)
20

particle perturbation, a twist. It is therefore beneficial to develop an
approach that experimentally rules out the single-particle gap-opening
mechanisms for nanotubes of arbitrary structure. Theory predicts that
such gaps close at some value of an applied axial magnetic field and that
a metallic state is recovered62,64. Interestingly, experiments performed on
pristine, suspended carbon nanotubes show that the gap reaches a mini-
mum, non-zero value at a critical magnetic field (~2.5 T for the device
in Fig. 7c), before widening again15. The observation of a minimum
gap rules out all of the above non-interacting theories of gap creation
in these nanotubes.
Another indication of strongly interacting electrons in nanotubes
comes from the presence of low-energy neutral electronic excitations
with energies less than the gap energy15. (By contrast, if the electron
system were non-interacting, the lowest-energy excitation would be
that which moves an electron from the valence band to the conduction
band, costing an energy equal to the gap energy.) Although the origin of
these low-energy excitations is not yet fully understood, their presence
is consistent with the Mott theory31–34, which predicts neutral excita-
tions of the Mott insulator with energies less than the gap energy, such
as spin excitations, and an energy scale similar to that observed. This
provides further evidence for a many-body origin of the observed gap
in carbon nanotubes.
In future, signatures of a 1D Mott insulator around its metal–insulator
transition point65 may be investigated. Furthermore, the exotic spin liq-
uid excitations predicted for carbon nanotubes34 could be studied in
detail. Finally, gaps could be opened even in the absence of pinning. It
has been theorized for Wigner crystals48 that in wider channels the 1D
Wigner crystal will make a transition to a zigzag phase with a gapped
mode, in addition to a gapless one, as its hallmark. Recent conductance
measurements of wide quantum wires66 are consistent with this picture;
however, the gapped mode has not yet been directly observed.
Future directions
Further experiments, including those on systems and using techniques
not discussed here, are expected to provide more insight into the nature
of interacting particles confined to one dimension. Examples include
experiments on cold-atom systems (see, for example, ref. 4) and time-
resolved measurements of charge and spin dynamics (a recent experi-
ment in carbon nanotubes67 has yielded promising results; however, the
fast charge mode has not yet been observed, possibly owing to insuffi-
cient time resolution). In addition to contributing to the understanding
of the fundamental physics of excitons in one dimension68–70, optics
measurements may give new access to a broad range of quantum impu-
rity problems (see, for example, ref. 71). One-dimensional systems may
be controllably integrated with external structures such as an external
periodic potential to create an artificial Mott insulator that can provide
B (T)
2.5
10
0
0.5 1 1.5 2
Vg (V)
Figure 7 | Electron correlations in 1D carbon nanotubes give rise to
insulating gaps. a, Colour plot of differential conductance, dI/dV, as
a function of source–drain bias, V, and gate voltage, Vg, for an (18, 18)
nanotube. The large region of dI/dV suppression near the centre indicates
the presence of a gap. (Data courtesy of V.V.D., M. Takekoshi, P. Kim,
J. Hone and T. Heinz.) b, Mapping of the honeycomb lattice onto an
equivalent two-leg ladder system. Brown and green distinguish which
atoms in the nanotube are mapped onto which ladder leg. The parameter
U is the electronic on-site interaction discussed in the main text, and t
and t⊥ are the electron hopping matrix elements. The effective on-site
interaction in the equivalent ladder is reduced by a factor N, to U/N,
for an (N, N) nanotube. The black arrow indicates the tube’s cylindrical
axis. c, Conductance, G, as a function of magnetic field, B, and gate
voltage, showing a gap that reaches a minimum value for B ≈ 2.5 T. (Panel
reproduced from ref. 15.)
a highly tunable means of harnessing many-body physics in the solid
state and developing new applications such as a pump for quantized
fractional charge59. A number of interesting theories42,53 concerning
spin-incoherent and nonlinear Luttinger liquids may be investigated.
There is also significant theoretical work ahead aimed at filling the
gaps between the limits of the Tomonaga–Luttinger model and incoher-
ent Luttinger liquids and between the limits of ideally ordered 1D inter-
acting systems and general disordered ones. In the case of disordered
Luttinger liquids, the recognized72 need for further theory is driven by
experiments with carbon nanotubes out of equilibrium73 and by cur-
rent theoretical predictions such as many-body localization74. Given
the rapid evolution of its experimental and theoretical techniques, the
field of correlated many-body 1D systems looks set for more exciting
developments in the future. ■
1. Landau, L. D. & Lifshitz, E. M. Course of Theoretical Physics Vol. 9, 1–27 (Pergamon, 1980).
2. Kinoshita, T., Wenger, T. & Weiss, D. S. Observation of a one-dimensional Tonks–Girardeau
gas. Science 305, 1125–1128 (2004).
3. Paredes, B. et al. Tonks–Girardeau gas of ultracold atoms in an optical lattice. Nature 429,
277–281 (2004).
4. Kinoshita, T., Wenger, T. & Weiss, D. S. A quantum Newton’s cradle. Nature 440, 900–903
(2006).
5. Clément, D., Fabbri, N., Fallani, L., Fort, C. & Inguscio, M. Exploring correlated 1D Bose gases
from the superfluid to the Mott-insulator state by inelastic light scattering. Phys. Rev. Lett.
102, 155301 (2009).
6. Auslaender, O. M. et al. Spin–charge separation and localization in one dimension. Science
308, 88–92 (2005).
This paper reports the observation of the spin–charge separation using momentum-
resolved electron tunnelling.

215
© 2010 Macmillan Publishers Limited. All rights reserved
INSIGHT REVIEW
7. Auslaender, O. M. et al. Tunneling spectroscopy of the elementary excitations in a one-
dimensional wire. Science 295, 825–828 (2002).
8. Steinberg, H. et al. Charge fractionalization in quantum wires. Nature Phys. 4, 116–119
(2008).
9. Imambekov, A. & Glazman, L. I. Universal theory of nonlinear Luttinger liquids. Science 323,
228–231 (2009).
This paper describes the theory of momentum-resolved electron tunnelling into a liquid
of particles with a generic (nonlinear) dispersion relationship.
10. Khodas, M., Pustilnik, M., Kamenev, A. & Glazman, L. I. Fermi–Luttinger liquid: spectral
function of interacting one-dimensional fermions. Phys. Rev. B 76, 155402 (2007).
11. Pereira, R. G., White, S. R. & Affleck, I. Spectral function of spinless fermions on a one-
dimensional lattice. Phys. Rev. B 79, 165113 (2009).
12. Pereira, R. G. & Sela, E. Coulomb drag from spin–charge coupling at zero magnetic field.
Preprint at 〈http://arxiv.org/abs/0911.1391〉 (2009).
13. Schmidt, T. L., Imambekov, A. & Glazman, L. I. The fate of 1D spin–charge separation away
from Fermi points. Preprint at 〈http://arxiv.org/abs/0912.0326〉 (2009).
14. Deshpande, V. V. & Bockrath, M. The one-dimensional Wigner crystal in carbon
nanotubes. Nature Phys. 4, 314–318 (2008).
15. Deshpande, V. V. et al. Mott insulating state in ultraclean carbon nanotubes. Science 323,
106–110 (2009).
This paper reports the observation of energy gaps in nominally metallic carbon
nanotubes, as well as the presence of low-energy neutral excitations within the gap that
are interpreted using the theory of 1D Mott insulators.
16. Kane, C. L. & Fisher, M. P. A. Transport in a one-channel Luttinger liquid. Phys. Rev. Lett. 68,
1220–1223 (1992).
This paper predicted the zero-bias anomalies in tunnelling into a Luttinger liquid.
17. Egger, R. & Gogolin, A. O. Effective low-energy theory for correlated carbon nanotubes.
Phys. Rev. Lett. 79, 5082–5085 (1997).
18. Kane, C., Balents, L. & Fisher, M. P. A. Coulomb interactions and mesoscopic effects in
carbon nanotubes. Phys. Rev. Lett. 79, 5086–5089 (1997).
19. Matveev, K. A. & Glazman, L. I. Coulomb blockade of tunnelling into a quasi-one-
dimensional wire. Phys. Rev. Lett. 70, 990–993 (1993).
20. Yao, Z., Postma, H. W. Ch., Balents, L. & Dekker, C. Carbon nanotube intramolecular
junctions. Nature 402, 273–276 (1999).
This paper reports the power-law behaviour of the end-to-end tunnelling conductance
between carbon nanotubes, consistent with Luttinger liquid theory.
21. Postma, H. W. C., Jonge, M., Yao, Z. & Dekker, C. Electrical transport through carbon
nanotube junctions created by mechanical manipulation. Phys. Rev. B 62, R10653−R10656
(2000).
22. Bockrath, M. et al. Luttinger-liquid behaviour in carbon nanotubes. Nature 397, 598–601
(1999).
This paper studies two geometries for electron tunnelling into carbon nanotubes and
shows that the energy dependence of the tunnelling rate is consistent with the theoretical
predictions.
23. Nazarov, Y. V. Coulomb blockade of tunneling in isolated junctions. JETP Lett. 49, 126–128
(1989).
24. Devoret, M. H. et al. Effect of the electromagnetic environment on the Coulomb blockade
in ultrasmall tunnel-junctions. Phys. Rev. Lett. 64, 1824–1827 (1990).
25. Girvin, S. M., Glazman, L. I., Jonson, M., Penn, D. R. & Stiles, M. D. Quantum fluctuations
and the single-junction Coulomb blockade. Phys. Rev. Lett. 64, 3183–3186 (1990).
26. Auslaender, O. M. et al. Experimental evidence for resonant tunneling in a Luttinger liquid.
Phys. Rev. Lett. 84, 1764–1767 (2000).
27. Postma, H. W. C., Teepen, T., Yao, Z., Grifoni, M. & Dekker, C. Carbon nanotube single-
electron transistors at room temperature. Science 293, 76–79 (2001).
28. Ilani, S., Donev, L. A. K., Kindermann, M. & McEuen, P. L. Measurement of the quantum
capacitance of interacting electrons in carbon nanotubes. Nature Phys. 2, 687–691 (2006).
29. Venkataraman, L., Hong, Y. S. & Kim, P. Electron transport in a multichannel one-
dimensional conductor: molybdenum selenide nanowires. Phys. Rev. Lett. 96, 076601
(2006).
30. Bachtold, A. et al. Suppression of tunneling into multiwall carbon nanotubes. Phys. Rev. Lett.
87, 166801 (2001).
31. Balents, L. & Fisher, M. P. A. Correlation effects in carbon nanotubes. Phys. Rev. B 55,
R11973−R11976 (1997).
32. Krotov, Y. A., Lee, D. H. & Louie, S. G. Low energy properties of (n, n) carbon nanotubes.
Phys. Rev. Lett. 78, 4245–4248 (1997).
33. Odintsov, A. A. & Yoshioka, H. Universality of electron correlations in conducting carbon
nanotubes. Phys. Rev. B 59, R10457−R10460 (1999).
34. Nersesyan, A. A. & Tsvelik, A. M. Coulomb blockade regime of a single-wall carbon
nanotube. Phys. Rev. B 68, 235419 (2003).
35. Steinberg, H. et al. Localization transition in a ballistic quantum wire. Phys. Rev. B 73, 113307
(2006).
36. Kim, B. J. et al. Distinct spinon and holon dispersions in photoemission spectral functions
from one-dimensional SrCuO2. Nature Phys. 2, 397–401 (2006).
37. Damascelli, A., Hussain, Z. & Shen, Z.-X. Angle-resolved photoemission studies of the
cuprate superconductors. Rev. Mod. Phys. 75, 473–541 (2003).
38. Giamarchi, T. Quantum Physics in One Dimension 81–86 (Oxford Univ. Press, 2004).
39. Voit, J. One-dimensional Fermi liquids. Rep. Prog. Phys. 58, 977–1116 (1995).
40. Carpentier, D., Peça, C. & Balents, L. Momentum-resolved tunneling between Luttinger
liquids. Phys. Rev. B 66, 153304 (2002).
41. Jompol, Y. et al. Probing spin–charge separation in a Tomonaga–Luttinger liquid. Science
325, 597–601 (2009).
216
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
42. Pustilnik, M., Khodas, M., Kamenev, A. & Glazman, L. I. Dynamic response of one-
dimensional interacting fermions. Phys. Rev. Lett. 96, 196405 (2006).
43. Imambekov, A. & Glazman, L. I. Phenomenology of one-dimensional quantum liquids
beyond the low-energy limit. Phys. Rev. Lett. 102, 126405 (2009).
44. Laughlin, R. B. Anomalous quantum Hall effect: an incompressible quantum fluid with
fractionally charged excitations. Phys. Rev. Lett. 50, 1395–1398 (1983).
45. Berg, E., Oreg, Y., Kim, E. A. & von Oppen, F. Fractional charges on an integer quantum Hall
edge. Phys. Rev. Lett. 102, 236402 (2009).
46. Pustilnik, M. Dynamic structure factor of the Calogero–Sutherland model. Phys. Rev. Lett.
97, 036404 (2006).
47. Anderson, P. W. Infrared catastrophe in Fermi gases with local scattering potentials. Phys.
Rev. Lett. 18, 1049–1051 (1967).
48. Meyer, J. S. & Matveev, K. A. Wigner crystal physics in quantum wires. J. Phys. Condens.
Matter 21, 023203 (2009).
49. Matveev, K. A. Conductance of a quantum wire in the Wigner-crystal regime. Phys. Rev.
Lett. 92, 106801 (2004).
50. Schulz, H. J. Wigner crystal in one dimension. Phys. Rev. Lett. 71, 1864–1867 (1993).
51. Averin, D. V. & Nazarov, Y. V. Tunneling to and from the one-dimensional Wigner lattice.
Phys. Rev. B 47, 9944–9947 (1993).
52. Glazman, L. I., Ruzin, I. M. & Shklovskii, B. I. Quantum transport and pinning of a one-
dimensional Wigner crystal. Phys. Rev. B 45, 8454–8463 (1992).
53. Fiete, G. A. The spin-incoherent Luttinger liquid. Rev. Mod. Phys. 79, 801–820 (2007).
54. Cao, J., Wang, Q. & Dai, H. Electron transport in very clean, as-grown suspended carbon
nanotubes. Nature Mater. 4, 745–749 (2005).
55. Cobden, D. H., Bockrath, M., McEuen, P. L., Rinzler, A. G. & Smalley, R. E. Spin splitting and
even-odd effects in carbon nanotubes. Phys. Rev. Lett. 81, 681–684 (1998).
56. Tans, S. J., Devoret, M. H., Groeneveld, R. J. A. & Dekker, C. Electron–electron correlations in
carbon nanotubes. Nature 394, 761–764 (1998).
57. Minot, E. D., Yaish, Y., Sazonova, V. & McEuen, P. L. Determination of electron orbital
magnetic moments in carbon nanotubes. Nature 428, 536–539 (2004).
58. Levitov, L. S. & Tsvelik, A. M. Narrow-gap Luttinger liquid in carbon nanotubes. Phys. Rev.
Lett. 90, 016401 (2003).
59. Novikov, D. S. Electron properties of carbon nanotubes in a periodic potential. Phys. Rev. B
72, 235428 (2005).
60. Mott, N. F. Metal-insulator transition. Rev. Mod. Phys. 40, 677–683 (1968).
61. Dagotto, E. & Rice, T. M. Surprises on the way from one- to two-dimensional quantum
magnets: the ladder materials. Science 271, 618–623 (1996).
62. Charlier, J., Blase, X. & Roche, S. Electronic and transport properties of nanotubes. Rev.
Mod. Phys. 79, 677–732 (2007).
63. Chen, W., Andreev, A. V., Tsvelik, A. M. & Dror, O. Twist instability in strongly correlated
carbon nanotubes. Phys. Rev. Lett. 101, 246802 (2008).
64. Yang, L., Anantram, M. P., Han, J. & Lu, J. P. Band-gap change of carbon nanotubes: effect of
small uniaxial and torsional strain. Phys. Rev. B 60, 13874–13878 (1999).
65. Garst, M., Novikov, D. S., Stern, A. & Glazman, L. I. Critical conductance of a one-
dimensional doped Mott insulator. Phys. Rev. B 77, 035128 (2008).
66. Hew, W. K. et al. Incipient formation of an electron lattice in a weakly confined quantum
wire. Phys. Rev. Lett. 102, 056804 (2009).
67. Zhong, Z., Gabor, N. M., Sharping, J. E., Gaeta, A. L. & McEuen, P. L. Terahertz time-domain
measurement of ballistic electron resonance in a single-walled carbon nanotube. Nature
Nanotechnol. 3, 201–205 (2008).
68. Lefebvre, J. & Finnie, P. Polarized photoluminescence excitation spectroscopy of single-
walled carbon nanotubes. Phys. Rev. Lett. 98, 167406 (2007).
69. Wang, F., Dukovic, G., Brus, L. E. & Heinz, T. F. The optical resonances in carbon nanotubes
arise from excitons. Science 308, 838–841 (2005).
70. Maultzsch, J. et al. Exciton binding energies in carbon nanotubes from two-photon
photoluminescence. Phys. Rev. B 72, 241402 (2005).
71. Balents, L. X-ray-edge singularities in nanotubes and quantum wires with multiple
subbands. Phys. Rev. B 61, 4429–4432 (2000).
72. Gutman, D. B., Gefen, Y. & Mirlin, A. D. Tunneling spectroscopy of Luttinger-liquid
structures far from equilibrium. Phys. Rev. B 80, 045106 (2009).
73. Chen, Y.-F., Dirks, T., Al-Zoubi, G., Birge, N. O. & Mason, N. Nonequilibrium tunneling
spectroscopy in carbon nanotubes. Phys. Rev. Lett. 102, 036804 (2009).
74. Basko, D. M., Aleiner, I. L. & Altshuler, B. L. Possible experimental manifestations of the
many-body localization. Phys. Rev. B 76, 052203 (2007).
75. Tomonaga, S. Remarks on Bloch’s method of sound waves applied to many-fermion
problems. Prog. Theor. Phys. 5, 544–569 (1950).
76. Luttinger, J. M. An exactly soluble model of a many-fermion system. J. Math. Phys. 4,
1154–1162 (1963).
Acknowledgements A.Y. and L.I.G. acknowledge a discussion with B. I. Halperin
of the difference between charge fractionalization and quantization. L.I.G.
acknowledges support from the US National Science Foundation (NSF) Division of
Materials Research (grant no. DMR-0906498) and the Nanosciences Foundation
at Grenoble, France. M.B. acknowledges the US Office of Naval Research. A.Y. is
supported by the NSF under contract DMR-0707484.
Author Information Reprints and permissions information is available at www.
nature.com/reprints. The authors declare no competing financial interests.
Correspondence should be addressed to M.B. (marc.bockrath@ucr.edu).
Vol 464j11 March 2010jdoi:10.1038/nature08757
Targeting early infection to prevent HIV-1
mucosal transmission
Ashley T. Haase1
Measures to prevent sexual mucosal transmission of human immunodeficiency virus (HIV)-1 are urgently needed to curb the
growth of the acquired immunodeficiency syndrome (AIDS) pandemic and ultimately bring it to an end. Studies in animal
models and acute HIV-1 infection reviewed here reveal potential viral vulnerabilities at the mucosal portal of entry in the
earliest stages of infection that might be most effectively targeted by vaccines and microbicides, thereby preventing
acquisition and averting systemic infection, CD4 T-cell depletion and pathologies that otherwise rapidly ensue.
uring more than one-quarter of a century of the HIV-1
D pandemic, millions have died from AIDS, millions have
been orphaned, and more than 30 million people world-
wide are living with HIV-1 infection1. Although advances
in antiretroviral therapies and access to treatment are countering this
great and continued toll in the developed world and sub-Saharan
Africa, the epicentre of the pandemic2, prevention of sexual mucosal
transmission, the principal route of acquisition1, is clearly needed to
contain the continuing growth of the pandemic and ultimately eradi-
cate AIDS3.
The prevention record is mixed. Circumcision has proven effective
in preventing 50% to 60% of female to male transmission4–6, and
there are the first hints recently of vaccine and microbicide candi-
dates that could prevent sexual transmission to men and women,
albeit at low efficacy of about 30%7,8. However, vaccine and micro-
bicide candidates have not proved efficacious or have even enhanced
acquisition in previous trials9,10. These disappointing failures and lim-
ited successes reinforce the view of leading investigators11 that it is now
especially timely to reinvigorate and broaden research on the virus, the
immune response to it, and the pathogenesis of transmission and
infection to identify the correlates of prevention that will enable the
design of fully protective measures against acquisition of HIV-1.
In this review, I provide a personal perspective and synthesis of
studies on sexual mucosal transmission relevant to this goal. I largely
focus on the results of in vivo analyses of relevant tissues, and on
transmission to women, because of the increasing feminization of
the pandemic in sub-Saharan Africa12, and because of insights into
HIV-1 transmission to women derived from extensive studies of the
simian immunodeficiency virus (SIV) rhesus macaque nonhuman
primate model. I conclude from these studies that prevention strat-
egies should target the earliest stage of infection because of: (1) viral
vulnerabilities at this stage at the portal of entry in the small, infected
founder populations; (2) local expansion necessary to establish sys-
temic infection; and (3) the ill effects that otherwise rapidly ensue
from systemic infection.
Fast stage of slow lentivirus infections
The retrovirus genus to which HIV-1 and SIV belong was long ago
named Lentivirinae13 based on the long incubation period and slow
clinical course of disease that typify infection. However, it is now
clear from the SIV rhesus macaque model that local events critical
to establishing systemic infection take place very quickly in the early
1
Department of Microbiology, University of Minnesota, Minnesota 55455, USA.
©2010 Macmillan Publishers Limited. All rights reserved
REVIEWS
stages of lentivirus infections. CD4 T-cell depleting and pathological
processes are also very quickly set in motion in the early stages of
systemic infection that then play out over the months and years in the
slow phase of infection.
SIV rhesus macaque nonhuman primate model. These conclusions
are drawn from studies in the SIV rhesus macaque nonhuman primate
(NHP) model of transmission of HIV-1 to women14, which provides a
window through which mucosal transmission and the first 2 weeks of
infection can be viewed in vivo in ways that, for practical and ethical
reasons, are not possible in HIV-1 infection, which is only clinically
manifest after this time15. From NHP studies of the earliest stages of
infection16, we have the following picture (Fig. 1) of the critical events
following mucosal exposure to high doses of SIV: virus can cross the
mucosal epithelial barrier within hours17 to establish what has turned
out to be a small founder population of infected cells16,18. This founder
population then undergoes a necessary local expansion during the first
week of infection to generate sufficient virus and infected cells to
disseminate and establish a self-propagating systemic infection
throughout the secondary lymphoid organs16.
Beginning in the second week of infection, replication explodes in
the lymphatic tissues where virus has access to many more susceptible
target cells in close spatial proximity, compared to the dispersed low-
density populations of susceptible target cells at the portal of entry.
Virus levels in blood and tissues peak near the end of this second week
of infection, before declining towards relatively stable lower levels by
4 weeks after exposure. At this time, these infected lymphatic tissues
comprise a reservoir where virus is produced and stored and where
proviruses are harbored in latently infected cells in SIV-infected rhesus
macaques19, just as in HIV-1 infected humans20. Furthermore, this
reservoir is already the site of CD4 T-cell depletion and other patho-
logical processes that will eventually lead to progression to disease
(Fig. 2).
This description of transmission, early infection and ill effects of
systemic infection (Figs 1 and 2) is derived from studies in the SIV
high-dose vaginal challenge NHP model in which rhesus macaques
are atraumatically inoculated twice in the same day with two doses of
105 TCID50 (50% tissue culture infectious dose) and billions of viral
particles16. The strengths of this model are: (1) the window (Fig. 1) it
provides on critical events that precede the earliest time clinical signs
and symptoms of HIV-1 infection disease are manifest15; (2) infec-
tion of a high proportion of animals with a known time of exposure;
(3) access to relevant tissues in a relevant time frame, which increases
chances to observe directly virus–host cell interactions and critical
217
REVIEWS NATUREjVol 464j11 March 2010

Time frame Establishment

Vagina Cervix Lymphatic Activated CD4+
tissue T-cell target
Window
Crossing the reservoir
barrier
Hours
Weeks Thoracic duct Early
Immune
Local Infected 2-4 Lymph activation TReg cells
GALT
propagation founder node Suppress
Viral population
Small
vulnerabilities R0<1
Opportunities Spleen Lung
for prevention Brain Liver

Days Local Infected Massive depletion of gut lamina propria Prevent Too little Virus-specific
expansion cells CD4+ T cells too late CD8+ T cells
Virus

Sufficient Battlefield map

production Week 2
Draining
Dissemination lymph node Large numbers of
Self- SIV-specific CD8+ T cells
propagating +
Week 1 systemic
infection
Large numbers of
Establishment SIV-infected cells
Lymphatic
tissue E:T ratio <2
reservoir
Thoracic duct Lymph node
GALT
Weeks

Spleen Lung
Brain Liver

Figure 1 | Time frame, sites and major events in vaginal transmission and SIV-specific
effector CD8+ T cell
the fast phase of lentivirus infection. The SIV rhesus macaque animal
model provides a window through which to view early infection. Within SIV-infected
hours, virus in the inoculum may gain access through breaks in the mucosal target cell

epithelial barrier to susceptible target cells. The small focal infected founder Effector–target
population is initially composed mainly of infected resting CD4 T cells conjugate

lacking conventional markers of activation. The founder population

expands locally in these ‘resting’ and in activated CD4 T cells. Local
expansion is necessary to disseminate infection to the draining lymph node,
and subsequently through the bloodstream to establish a self-propagating
infection in secondary lymphoid organs. Crossing the barrier, small founder
populations (with the associated risk that the basic reproductive rate, R0, will
fall below one), and local expansion are vulnerabilities for the virus in week 1
of infection. These vulnerabilities create opportunities for prevention
targeting this stage. The male (blue) and female (pink) figures at the lower
left are positioned in the time frame marked weeks, after systemic infection
has been established. It is at this time that HIV-1 infection is first clinically
manifest, and, hence, the need for the animal model to view sexual mucosal
transmission and the earlier stages of infection. The larger female figure
symbolizes the disproportionate acquisition of HIV-1 infection and focus of
the review on transmission to women.
events; and (4) similarities in anatomy, physiology and immunology
of the female reproductive tract of rhesus macaques to humans, and
the general similarities of pathogenic SIV infection to HIV-1 infec-
tion in CD4 T-cell depletion, pathology and AIDS14,21.
This animal model also has weaknesses. First, the SIV challenge
dose exceeds the highest dose of HIV-1 to which humans are exposed
in semen, even in the acute stage of HIV-1 infection where virus
concentrations and transmission rates are the highest22. We thus have
to assume that virus–host interactions documented in tissue analyses
in the high dose challenge also occur with lower dose challenges at a
local level, but at such low frequencies that they would escape detec-
tion in tissue samples. In further defence of the high dose challenge,
single genome analysis (see below) in multiple low-dose intra-rectal
challenges that more closely approximate typical doses of HIV-1 in
semen provides evidence consistent with one of the key findings to
emerge from the high dose model: that infected founder populations
are small as a consequence frequently of infection with a single virus
genotype23.
In addition, this cell-free virus model of course cannot provide
insight into cell-associated virus transmission. An animal model of
the latter potential mode of HIV-1 transmission would certainly be
valuable, but would have to be developed, as infected cell inocula
218
©2010 Macmillan Publishers Limited. All rights reserved
Figure 2 | Ill effects of systemic infection and too little too late immune
response. After dissemination and establishment of systemic infection in
weeks 2–4 (upper left panel), peak viral replication in gut-associated and
other lymphatic tissues results in massive depletion of CD4 T cells in gut
lamina propria. Immune activation (upper right panel) has drawbacks in
supplying activated CD4 T-cell targets for viral replication and inducing a
Treg response that suppresses the immune response. Immune activation
does elicit a virus-specific CD8 T-cell response that is too little and too late to
prevent gut CD4 T-cell depletion or clear infection. Although there are large
numbers of SIV-specific CD8 T cells in, for example, lymph nodes (lower
panel), there are also large numbers of infected cell targets so that at the
effector to target (E:T) ratio achieved (,2), infection is only partially
controlled.
have so far proved an inefficient way to infect animals reliably24,
except when facilitated by genital ulceration25. I later discuss the
implications and relevance of rapid dissemination from the portal
of entry under these conditions.
Opportunities and lost opportunities. Studies in this high-dose vaginal
challenge model support the conclusion that prevention strategies
should target the first week of infection both to take advantage of viral
vulnerabilities and two opportunities for prevention (Fig. 1) and to avert
the ill effects of systemic infection in the second week and beyond
(Fig. 2). The first viral vulnerability and opportunity for prevention is
related to the establishment of the small founder population, which
must be sustained at a basic reproductive rate, R0, $ 1, otherwise infec-
tion will be aborted. Thus, early intervention measures that reduce
the growth rate to ,1 will have that desired outcome. The second
opportunity is preventing local expansion so that insufficient virus
and infected cells are produced to disseminate and establish systemic
infection. This second opportunity for prevention may not hold for
rectal and oral transmission, given the evidence that dissemination
beyond the portal of entry occurs within a few days of rectal or oral
exposure, without apparent local expansion26–28.
The ill effects depicted in Fig. 2 of losing those opportunities to
prevent systemic infection are similar in SIV and acute HIV infec-
tions, and include: massive depletion of memory CD4 T cells in the
NATUREjVol 464j11 March 2010
lamina propria of the gut mediated by direct effects of infection and
apoptosis of bystander cells29–35; acute enteropathy induced by pro-
inflammatory cytokines and virotoxic effects of the virus that induce
intestinal epithelial apoptosis36–38; and immune activation, the down-
side of which includes a new supply of activated CD4 T cells to
support virus replication, and a T regulatory (Treg) response that
has been called premature, because the immunosuppressive effects
that measurably dampen the SIV-specific cellular immune response
precede clearance or control of viral replication39. Although immune
activation does elicit a virus-specific CD8 T-cell response, it is ‘too
little and too late’ to prevent gut CD4 T-cell depletion or do more
than partially control infection40.
These serious ill effects on the host’s immune system in early
infection are only the beginning of immunopathological processes
set in motion at this stage that will have their greatest impact in the
later chronic stages of SIV and HIV-1 infection: microbial transloca-
tion associated with continued damage to the gut lining that con-
tributes to chronic immune activation and the continued loss of CD4
T cells41; and TGF-b1 Treg cells42 that initiate a fibrotic process, which
disrupts the lymphatic tissue architecture that supports the migra-
tion, growth and survival of T cells, thereby contributing both to CD4
T-cell depletion and limiting immune reconstitution with antiretro-
viral therapies43.
Small founder population and local expansion. The evidence for a
small, focal infected founder population of cells depicted in Fig. 1,
with local expansion as an antecedent and prerequisite for dissemi-
nation and systemic infection, is based on in vivo tissue analyses of
vaginal transmission. These analyses revealed only small foci of pro-
ductively infected cells, defined as cells with detectable SIV RNA, at 3
to 4 days after inoculation16,18. Infection then expands locally before
detection of infection in the draining lymph nodes and systemically
throughout the secondary lymphoid organs, consistent with the local
expansion and staged dissemination model shown in Fig. 1.
There is evidence that dendritic cells with detectable viral antigens
reach the draining lymph nodes much earlier—18 to 24 h after expo-
sure17—but apparently the threshold for a self-propagating infection
is not crossed until virus and infected cells are produced locally in
sufficient quantities to establish infection distal to the portal of entry,
because the draining lymph nodes are not the site where productive
infection is first detected18. Thus, dilution, trapping in cervical
mucus, the physical mucosal epithelial barrier and other mechanisms
transform exposure to a large quantity of virus in the inoculum to a
small founder population of infected cells that must then expand
locally for a few days before systemic infection is established.
Genetic bottleneck in HIV-1 transmission. The earliest stages of
mucosal transmission of HIV-1 have not been directly observed, but
we can infer parallels to SIV’s small, infected founder populations
from the genetic bottleneck at HIV-1 transmission. It has been known
for some time that in mother-to-child and heterosexual transmission,
virus isolates from the newly infected individual are genetically much
less diverse than viruses from the transmitter44,45. More recently,
sequencing viruses in heterosexual transmission pairs and in acute
HIV-1 infection provided evidence that a single virus (or infected cell)
initiated productive infection in close to 80% of the individuals tested,
and two to five viruses in the other 20%46,47. The implied small,
infected founder populations in HIV-1 mucosal transmission again
indicate that the greatest opportunities for prevention are strategies
that target these initially small and genetically homogeneous foci of
infection in the first week of infection.
However, the small size of the infected founder population does
not imply that the transmitted virus itself is inherently any easier to
contain. To the contrary, phenotypic analysis of transmitted HIV-1
revealed masking of CCR5 co-receptor binding regions, and equiva-
lent or enhanced resistance to fusion inhibitors and neutralizing
antibodies compared to virus isolated from chronically infected
individuals47,48.
©2010 Macmillan Publishers Limited. All rights reserved
REVIEWS
Mucosal front line
Crossing the barrier. I now return to a fuller description of the
earliest steps in transmission and local infection based on direct
observations on cervical vaginal tissues. Virus depicted in Fig. 1
enters at anatomical sites where the mucosal barrier is most easily
breached, by mechanisms such as the microtrauma associated with
sexual intercourse49 that provide immediate access to target cells in
the submucosa (Fig. 1). Only a single layer of columnar epithelium
guards the endocervix, and the transformation zone between ecto-
and endocervix is also a single layer of epithelium and a site of high cell
turnover that could facilitate entry50,51. These anatomical differences
and recent mapping studies of the sites where infected founder popu-
lations were identified in vaginal transmission of SIV are consistent
with the transformation zone and endocervix as preferential sites of
entry52. This hypothesis could also explain the increased acquisition of
HIV-1 associated with cervical ectopy—extension of the simple
columnar epithelium to cover a portion of the ectocervix—observed
in some studies53.
The transformation zone and endocervix, however, are not exclusive
sites of entry, as SIV has been transmitted to hysterectomized
animals54, and HIV-1 in the congenital absence of a cervix55, establish-
ing that there are target cells and independent entry mechanisms for
vaginal transmission. These observations, however, do not address
preferred sites of entry and the role of inflammation and trauma in
the individual described in ref. 55; nor does the failure of diaphragms to
decrease HIV-1 transmission exclude cervix as a preferred site of entry,
because of the confounding effects of poor adherence in that study56.
Initial target cells. In the SIV rhesus macaque model, the earliest
focal collections of infected cells are CD4 T cells with a surprising
phenotype. The initial expectation from growing HIV isolates and
SIV in tissue culture was that they would be macrophage-tropic and/
or replicate in activated CD4 T cells, using the CCR5 co-receptor
(reviewed in ref. 57). However, in vivo, it turned out that the initially
infected cells were CD4 T cells with none of the expected markers of
activation18. They were therefore called31 ‘resting’, with the implica-
tion that they probably had some residue of co-receptor expression
and prior activated state to enable them to support productive infec-
tion, in contrast to truly resting CD4 T cells. In the gut, a4b7 may
facilitate entry and preferential infection of T-helper 17 CD4 T cells
with this minimally activated phenotype58,59.
Infected ‘resting’ CD4 T cells comprise about 90% of the detectably
productively infected cells in the initial foci and subsequent local
expansion at the portal of entry, and the population of infected cells
that produces peak levels of virus in SIV infections18,31,60 in the
lymphatic tissues in the second week of infection. Over 60% of
CD4 T cells in gut with detectable SIV DNA in early infection have
also been shown to be memory CD4 T cells with relatively low levels
of CCR530; in early HIV-1 infection close to 90% of infected cells were
typed as CD4 T cells and the majority had a ‘resting’ phenotype18,61.
Moreover, replication-competent clones derived from single genome
amplification of isolates in acute HIV-1 infection primarily infected
CD4 T cells rather than macrophages in culture48. Thus, CD4 T cells
are probably the principal cell type infected at the portal of entry and
throughout the lymphatic tissue reservoir from the earliest stages of
infection.
Target cell availability. One reason that this may be the case is target
cell availability. CD4 T-cell susceptible target cell populations in the
vagina, ecto- and endocervix in rhesus macaques60,62 and humans50,63
are largely spatially dispersed populations lying just beneath the epi-
thelium, and, to a lesser extent, deeper submucosa, with some small
focal aggregates and intraepithelial lymphocytes. There are also
macrophages and dendritic cells in the submucosa, and dendritic
cells within the epithelium in both species, but ‘resting’ CD4 T cells
outnumber macrophages and dendritic cells by 4 to 5:1 and two to
three or more infected ‘resting’ CD4 T cells are often found clustered
in the small focal clustered infected founder populations60.
219
REVIEWS
These census analyses are consistent with the hypothesis that
numbers and spatial proximity contribute to preferential propaga-
tion initially in this cell type60. However, target cell availability is
unlikely to be the sole determinant of the cell tropisms in early infec-
tion, because activated CD4 T cells are the other cell type most fre-
quently infected at this stage, even though there are 70 times as many
‘resting’ CD4 T cells and relatively more abundant CD41CCR51
macrophages and dendritic cells in the vicinity.
Mucosal epithelium: immunity’s front line. Mucosal epithelial
defences against viral entry are conventionally portrayed as a passive
physical barrier protecting against invasion, but, in actuality, the
epithelium lining endocervix and upper female reproductive tract
is the active front line of the host immune system. This mucosal
epithelial front line mediates innate defences against microbes under
the hormonal control of oestradiol and progesterone64–66, and func-
tions as a sentinel and signalling system with Toll-like receptors that
recognize and respond to pathogen-associated molecular patterns by
secreting: (1) microbicidal defensins and other antimicrobial pep-
tides; (2) secretory leukocyte protease inhibitor; (3) the microbicidal
enzymes lactoferrin and lysozyme; (4) surfactant protein A; and (5)
complement. Mucosal epithelial signalling through chemokines and
cytokines also recruits plasmacytoid dendritic cells (pDCs)67,68 and
other cells that mediate innate defences and inflammation, and ini-
tiate and link innate and adaptive immunity.
The balancing act in innate defences. The physical barrier and active
mucosal front-line defences described above would be expected to
inhibit SIV and HIV entry and to provide an array of inhibitory
activities against viral replication: (1) SDF-1, MIP-1a/b and
RANTES to block viral entry mediated by the co-receptors CXCR4
(blocked by SDF-1) and CCR5 (blocked by MIP-1a/b and
RANTES)57; (2) MIP-3a expression in response to vaginal inocu-
lation of SIV to immediately recruit interferon (IFN)-a/b producing
pDCs to the endocervix52; and (3) large increases in IFN-c expression
in cervical vaginal tissues by 6 days after exposure69. These increases
in inhibitory chemokines immediately after exposure and the first
few days of infection and unknown contributions of innate defences
cited in the preceding section presumably provide a cumulative level
of protection that accounts in part for the small infected founder
populations in SIV infection despite exposure to a large dose of virus.
This innate immune response might also explain restricted cellular
tropisms in vivo, for example protection of IFN-producing pDCs
from infection. In HIV-1 transmission, similar defence mechanisms
probably also contribute to the generally low probability of vaginal
transmission of HIV-1 of ,1 in 100 to 1,000, depending on the viral
load of the transmitting partner70,71.
Paradoxically, these innate antiviral and inflammatory defence
mechanisms in SIV and HIV-1 infections at the same time may
facilitate transmission, by increasing target cell availability, and by
creating conditions for highly efficient cell-to-cell spread of infection
where HIV-1 replication has been shown to become less sensitive to
inhibition by interferon72. Mapping local expansion52 revealed
growth by accretion of newly infected cells around foci of infected
founder cells, and spread of infection along tracts of infiltrating
inflammatory cells. These findings led to the model shown in Fig. 3:
SIV overcomes the problem of replication in the low density and
dispersed populations of CD4 T cells at the portal of entry through
exploiting the innate immune and inflammatory response that brings
in large numbers of target cells to create a generally favourable environ-
ment to fuel expansion wherever the initial focus is established, and to
generate conditions where SIV has been shown to propagate most
efficiently by direct cell-to-cell spread or at close range73.
The mechanism inducing the influx of target cells fits nicely with the
concept of mucosal epithelium functioning as a sentinel and signalling
system65. In this case, exposure to virus or other components of the
inoculum stimulates increased expression of MIP-3a in the endocer-
vical epithelium, attracting pDCs beneath the epithelium, which in
turn recruit CD4 T cells through MIP-1b and other chemokines52. The
220
©2010 Macmillan Publishers Limited. All rights reserved
NATUREjVol 464j11 March 2010
Vagina Cervix
MIP-3α
pDCs
Thinned/disrupted
MIP-1β
mucosal barrier
Target
cells
Pre-existing
inflammation
Interferons
Anti-viral
New target
cells
chemokines
Fuel Restrict
infection infection
Local
expansion
Figure 3 | Inflammation, innate immunity, mucosal epithelial signalling
and target cell availability at mucosal front lines. Exposure of endocervical
epithelium to virus, a component of virus or inoculum initiates signalling
(jagged arrow) that increases expression of MIP-3a in the epithelium that
recruits pDCs. They in turn recruit, through chemokines such as MIP-1b,
the CD4 T cells that fuel local expansion. Interferons from the pDCs and
chemokines also suppress viral replication but the balance is tipped in favour
of the virus by the cells that fuel the local expansion necessary for
dissemination and establishment of systemic infection. Pre-existing vaginal
inflammation also facilitates infection by thinning and disrupting the
multilayered lining, and providing a pool of target cells for local expansion.
scales in Fig. 3 are meant to convey the counterbalancing effects of the
innate and inflammatory response in restricting viral infection, and
facilitating it, through increasing target cell availability, with the
greater target cell availability tipping the scales in favour of the virus.
The influx of new target cells to fuel local expansion is also a
vulnerability that can potentially be exploited in preventing trans-
mission. Glycerol monolaurate inhibits female reproductive tract
epithelial signalling and may have prevented local expansion to
thereby protect animals exposed to high doses of SIV52.
Inflammation and mucosal transmission. Innate defences and assoc-
iated inflammation generally facilitate transmission of SIV and HIV-1
by compromising the integrity of the mucosal barrier and increasing
target cell availability. In the model illustrated in Fig. 3, pre-existing
inflammation is shown in association with thinned and disrupted
vaginal epithelium. This provides immediate access to large numbers
of target cells in the underlying submucosa, and thus the association of
initial vaginal SIV infections at such a site52. The seemingly paradoxical
enhancing effects of imiquimod and CpG ODN74 applied vaginally on
SIV transmission may also be explained by the effects on mucosal
integrity and target cell availability of inflammation induced by these
Toll-like receptor agonists.
Pre-existing genital infections and ulcers may facilitate systemic
infection not only by effects on the mucosal barrier and target cell
availability, but also by providing a haematogeneous route for imme-
diate systemic spread. For example, SIV-infected and other cells
inoculated vaginally were quickly disseminated throughout the
lymphatic tissues in the setting of experimentally induced genital
ulceration25.
Co-existing genital infections also increase HIV-1 acquisition
through macro- or micro-ulcerations compromising the integrity of
the mucosal barrier and the increased availability of target cells asso-
ciated with these infections75,76. These conditions relieve the genetic
bottleneck described above, such that multiple genetic variants rather
than one or two are transmitted to individuals with underlying genital
infections77.
Target cell availability may also explain the failure of acyclovir treat-
ment of herpes simplex virus 2 (HSV-2) infection to reduce the assoc-
iated acquisition of HIV-1, despite decreased shedding and healing of
HSV-2 ulcers78,79. This result, in line with the generally disappointing
results in reducing HIV-1 acquisition by treating sexually transmitted
NATUREjVol 464j11 March 2010
infections75,76, can now be attributed to persistence of enriched target
cell populations after ulcer healing80.
Seminal contributions to transmission. Even without pre-existing
inflammation, exposure of the mucosal epithelium to semen initiates
signalling to orchestrate the necessary changes in the female re-
productive tract for tolerance of allogeneic sperm, implantation
and development of the fetus81,82 that may also facilitate transmis-
sion. Exposure of cervical vaginal epithelium to semen elicits
increases in chemokines such as MIP-3a83 (similar to the effect of
exposure to the viral inoculum in experimental SIV infection52) and
pro-inflammatory cytokines (GM-CSF, IL-1, IL-6 and IL-8) that
recruit neutrophils, dendritic cells, macrophages and lymphocytes,
which accumulate beneath the cervical and uterine epithelium. Thus,
a microenvironment conducive to transmission is created by the
recruited cells, the immunosuppressive effects of TGF-b and pros-
taglandin E in semen, other transmission-enhancing factors in semen
such as the amyloid fibres derived from prostatic acid phosphatase84,
alterations in mucosal integrity effected by neutrophils migrating
through the epithelium, and increased target cell availability
mediated by mucosal epithelial signalling.
Mucosal cellular immune response
Too little and too late. I now turn from early infection and virus
interactions with innate defences to the adaptive immune response.
Even though this response is not detectable until the end of the
second week of infection40, its timing, location and magnitude are
still instructive for designing effective preventive measures.
The cellular immune response that has so far been characterized
follows antigen expansion and peak replication, and therefore has
been described40 as ‘‘too late and too little’’, because it is too late to
clear virus locally or prevent systemic spread, and insufficient in
magnitude to prevent the massive CD4 T-cell depletion in gut and
other pathological consequences of early infection. Nonetheless, the
robustness of the immune response in specific tissue compartments
is a determinant of the extent of control of viral replication and thus
rate of disease progression.
Tissue-compartment-specific partial control. The changing rela-
tionships over time in numbers and locations of immune effectors—
virus-specific CD8 T cells—and their infected targets in infected tissues
can be visualized and quantified by combining in situ tetramer staining
of the effectors with in situ hybridization to detect viral RNA1 cells85.
This analysis, illustrated in Fig. 2, provides images of the spatial pro-
ximity and conjugates of effectors and targets, and a direct way to
determine the in vivo effector/target ratio. It is immediately apparent
in this infected lymph node that there are not only large numbers of
effectors but also large numbers of targets, so the effector/target ratio is
,2. Such a low effector/target ratio correlates with poor control of viral
replication—measured as reduction from peak viral load. By contrast,
at high effector/target ratios $50–100:1 achieved at 3 weeks after
vaginal exposure in cervical vaginal tissues, there was a commensurate
,100-fold reduction in viral load.
This analysis tells us that the extent of control by SIV-specific CD8
T cells is tissue-compartment-specific and related to the relative
numbers of both effectors and targets. The opportunities for preven-
tion in the earliest stages of infection tell us that both timing and
location may be critical determinants of outcome. When infected
founder populations are small and focal, and have not as yet
expanded sufficiently to have disseminated and established systemic
infection, a relatively small number of effectors at the mucosal site of
entry might be at the right place at the right time to be ‘enough and
soon enough’ to clear infection85.
Targeting early infection. The ‘soon enough’ part of this concept
implies that a very rapid local recall response and local expansion of
virus-specific cytotoxic T lymphocytes (CTLs) might eradicate founder
populations before local expansion spreads infection. Whether the
few days available in vaginal transmission would be sufficient or not
is unclear from experience with conventional vaccines and ideas
©2010 Macmillan Publishers Limited. All rights reserved
REVIEWS
about immunological memory. It may be that a vaccine would have
to induce and maintain a population of effectors at the portal of entry
to be up to the challenge. This is more likely to be the case for rectal
and oral transmission where infection is also even more quickly
disseminated26–28.
Consistent with the concept of a protective mucosal population of
immune effectors, vaccine-induced high-avidity CTLs have been
shown to delay simian-human immunodeficiency virus (SHIV) dis-
semination from mucosa following rectal challenge86, and prior
infection of rhesus macaques with an attenuated SHIV has been
shown to provide protection against vaginal challenge associated
with SIV-specific CTLs in cervical vaginal tissues87. Furthermore, a
persistent prior infection with rhesus cytomegalovirus expressing
SIV antigens that elicited and maintained virus-specific effector
memory CD4 and CD8 T cells at extra-lymphoid sites has recently
been shown to provide partial protection against repeated intra-
rectal challenge88.
More generally, measures that target the establishment and expan-
sion of small founder populations in the earliest stages of infection
are most likely to be effective in prevention (Fig. 1). Thus, pharmaco-
logical interventions with antiretroviral therapies within the window
in which the infected founder population has just been established
can prevent or increase control and moderate the pathological con-
sequences of infection89. Microbicides that block binding and co-
receptor-mediated entry and reverse transcription have been shown
to protect against SHIV and SIV vaginal and rectal challenges in the
rhesus macaque model90–92. Similarly, moderating the growth in
target cells available to fuel the local expansion on which the virus
depends can also prevent vaginal transmission52.
Challenges and gaps in knowledge
Studies described herein reveal both challenges in preventing mucosal
HIV-1 transmission and opportunities to do so, particularly with
strategies targeting early infection. Designing and assessing the effec-
tiveness of countermeasures would be informed and enabled by
addressing the following six issues.
First, discovering ways to generate a resident or rapidly responding
mucosal population to eliminate the small founder population at the
portal of entry that avoids the immunosuppressive and target cell
availability problems inherent in the immune activation necessary for
that response.
Second, identifying correlates of prevention in the recent Thai
trial7 of a two-component vaccine in which neither component alone
had previously shown efficacy.
Third, monitoring occult infections in trials of vaccines and micro-
bicides. Occult infections are defined as HIV-1-exposed individuals
who remain seronegative but who have detectable cellular immune
responses to HIV-1 and DNA in CD4 T cells93; or rhesus macaques
repeatedly challenged vaginally or rectally with low or high doses of
SIV94–96 that develop transient viraemia with low to undetectable
antibodies or cellular immune response, and have detectable provirus
in vaginal and lymphoid tissues94,96. They may later be manifest as
typical pathogenic systemic infections, have been documented after
apparent protection with a microbicide52, and thus represent a poten-
tial public health issue and point to the need for long-term follow up in
testing vaccine and microbicide candidates.
Fourth, defining mechanism(s) by which circumcision does or
does not protect4–6, and the lack of circumcision facilitates male
transmission.
Fifth, developing and investigating the following in NHP models:
(1) cell-free virus transmission under conditions that more closely
approximate human exposures; (2) cell-associated viral transmis-
sion; and (3) each route of mucosal transmission—vaginal, rectal,
oral and penile—which will probably differ, based on their distinct
anatomy and functions, both in local propagation and routes and
speed of systemic spread.
221
REVIEWS
Sixth, advancing fundamental knowledge of mucosal immunology:
(1) understanding why the interferon system and other innate
immune mechanisms are not more effective against SIV and HIV-1
despite their massive upregulation in early infection52,69,97; (2) the role
of natural killer cells and antibody-dependent cellular cytotoxicity in
host defences at mucosal sites; (3) CD8 T-cell mucosal immune res-
ponses and the trafficking, turnover and fate of these cells at mucosal
sites; (4) the role of CD4 T cells in host defences at mucosal sites; and
(5) discovering new mechanisms of host defences at mucosal sites by
genomic analyses and systems biology.
Looking to the future, a better understanding of mucosal-site-
specific infection and immunity, answering the fundamental
immunological questions posed in ref. 98, and engaging a broader
scientific community in the effort constitute the most promising path
to an effective vaccine and other interventions to prevent sexual
mucosal transmission of HIV-1.
1. UNAIDS. 2008 Report on the global AIDS epidemic Æhttp://www.unaids.org/en/
KnowledgeCentre/HIVData/GlobalReport/2008/2008_Global_report.aspæ
(2008).
2. Fauci, A. S. 25 years of HIV. Nature 453, 289–290 (2008).
3. Cohen, J. Treatment and prevention exchange vows at international conference.
Science 321, 902–903 (2008).
4. Bailey, R. C. et al. Male circumcision for HIV prevention in young men in Kisumu,
Kenya: a randomised controlled trial. Lancet 369, 643–656 (2007).
5. Auvert, B. et al. Randomized, controlled intervention trial of male circumcision for
reduction of HIV infection risk: the ANRS 1265 Trial. PLoS Med. 2, e298 (2005).
6. Gray, R. H. et al. Male circumcision for HIV prevention in men in Rakai, Uganda: a
randomised trial. Lancet 369, 657–666 (2007).
7. Rerks-Ngarm, S. et al. Vaccination with ALVAC and AIDSVAX to prevent HIV-1
infection in Thailand. N. Engl. J. Med. 361, 2209–2220 (2009).
8. Karim, S. A. et al. Safety and effectiveness of vaginal microbicides BufferGel and
0.5% PRO 2000/5 gel for the prevention of HIV infection in women: Results of
the HPTN 035 trial. Abstract 48LB (16th Conference on Retroviruses and
Opportunistic Infections, 2009).
9. Buchbinder, S. P. et al. Efficacy assessment of a cell-mediated immunity HIV-1
vaccine (the Step Study): a double-blind, randomized, placebo-controlled, test-
of-concept trial. Lancet 372, 1881–1893 (2008).
10. Check, E. Scientists rethink approach to HIV gels. Nature 446, 12 (2007).
11. Fauci, A. et al. HIV vaccine research: the way forward. Science 321, 530–532
(2008).
12. Quinn, T. C. & Overbaugh, J. HIV/AIDS in women: an expanding epidemic. Science
308, 1582–1583 (2005).
13. Haase, A. The slow infection caused by visna virus. Curr. Top. Microbiol. Immunol.
72, 101–156 (1975).
14. Miller, C. J. et al. Genital mucosal transmission of simian immunodeficiency virus:
animal model for heterosexual transmission of human immunodeficiency virus.
J. Virol. 63, 4277–4284 (1989).
15. Fiebig, E. W. et al. Dynamics of HIV viremia and seroconversion in plasma donors:
implications for diagnosis and staging of primary HIV infection. AIDS 17,
1871–1879 (2003).
16. Miller, C. J. et al. Propagation and dissemination of infection after vaginal
transmission of simian immunodeficiency virus. J. Virol. 79, 9217–9227 (2005).
Comprehensive tissue analysis of the pathogenesis of transmission and early
infection in the nonhuman primate model of vaginal transmission of HIV-1.
17. Hu, J., Gardner, M. B. & Miller, C. J. Simian immunodeficiency virus rapidly
penetrates the cervicovaginal mucosa after intravaginal inoculation and infects
intraepithelial dendritic cells. J. Virol. 74, 6087–6095 (2000).
18. Zhang, Z.-Q. et al. Sexual transmission and propagation of simian and human
immunodeficiency viruses in two distinguishable populations of CD41 T cells.
Science 286, 1353–1357 (1999).
First description of the CD4 T cell as the principal target in acute SIV and HIV-1
infections, and the surprising ostensibly resting phenotype of a majority of the
CD4 T cells initially infected.
19. Reinhart, T. et al. Simian immunodeficiency virus burden in tissues and cellular
compartments during clinical latency and AIDS. J. Infect. Dis. 176, 1198–1208
(1997).
20. Haase, A. T. Population biology of HIV-1 infection: viral and CD41 T cell
demographics and dynamics in lymphatic tissues. Annu. Rev. Immunol. 17,
625–656 (1999).
21. Miller, C. M. & Shattock, R. J. Target cells in vaginal HIV transmission. Microbes
Infect. 5, 59–67 (2003).
22. Christopher, D. P. et al. Brief but efficient: acute HIV infection and the sexual
transmission of HIV. J. Infect. Dis. 189, 1785–1792 (2004).
23. Keele, B. F. et al. Low-dose rectal inoculation of rhesus macaques by SIV smE660
or SIVmac251 recapitulates human mucosal infection by HIV-1. J. Exp. Med. 206,
1117–1134 (2009).
24. Sodora, D. L., Gettie, A., Miller, C. J. & Marx, P. A. Vaginal transmission of SIV:
assessing infectivity and hormonal influences in macaques inoculated with
222
©2010 Macmillan Publishers Limited. All rights reserved
NATUREjVol 464j11 March 2010
cell-free and cell-associated viral stocks. AIDS Res. Hum. Retroviruses 14 (Suppl. 1),
S119–S123 (1998).
25. Weiler, A. M. et al. Genital ulcers facilitate rapid viral entry and dissemination
following intravaginal inoculation with cell-associated simian immunodeficiency
virus SIVmac239. J. Virol. 82, 4154–4158 (2008).
26. Miyake, A. et al. Rapid dissemination of a pathogenic simian/human
immunodeficiency virus to systemic organs and active replication in lymphoid
tissues following intrarectal infection. J. Gen. Virol. 87, 1311–1320 (2006).
27. Stahl-Hennig, C. et al. Rapid infection of oral mucosal-associated lymphoid tissue
with simian immunodeficiency virus. Science 285, 1261–1265 (1999).
28. Milush, J. M. et al. Rapid dissemination of SIV following oral inoculation. AIDS 18,
1–10 (2004).
29. Veazey, R. S. et al. Gastrointestinal tract as a major site of CD41 T cell depletion
and viral replication in SIV infection. Science 280, 427–431 (1998).
30. Mattapallil, J. et al. Massive infection and loss of memory CD41 T cells in multiple
tissues during acute SIV infection. Nature 434, 1093–1097 (2005).
31. Li, Q. et al. Peak SIV replication in resting memory CD41 T cells depletes gut
lamina propria CD41 T cells. Nature 434, 1148–1152 (2005).
Tissue analysis that revealed the rapid kinetics of acute SIV infection after
vaginal exposure, predominance of productive infection in resting memory CD4
T cells, and the loss of gut lamina propria CD4 T cells mediated mainly by
apoptosis.
32. Clayton, F., Snow, G., Reka, S. & Kotler, D. P. Selective depletion of rectal lamina
propria rather than lymphoid aggregate CD4 lymphocytes in HIV infection. Clin.
Exp. Immunol. 107, 288–292 (1997).
33. Mehandru, S. et al. Primary HIV-1 infection is associated with preferential
depletion of CD41 T lymphocytes from effector sites in the gastrointestinal tract.
J. Exp. Med. 200, 761–770 (2004).
34. Brenchley, J. M. et al. CD41 T cell depletion during all stages of HIV disease occurs
predominantly in the gastrointestinal tract. J. Exp. Med. 200, 749–759 (2004).
35. Guadalupe, M. et al. Severe CD41 T-cell depletion in gut lymphoid tissue during
primary human immunodeficiency virus type 1 infection and substantial delay in
restoration following highly active antiretroviral therapy. J. Virol. 77, 11708–11717
(2003).
36. Kotler, D. P., Gaetz, H. P., Lange, M., Klein, E. B. & Holt, P. R. Enteropathy
associated with the acquired immunodeficiency syndrome. Ann. Intern. Med. 101,
421–428 (1984).
37. Heise, C., Miller, C. J., Lackner, A. & Dandekar, S. Primary acute simian
immunodeficiency virus infection of intestinal lymphoid tissue is associated with
gastrointestinal dysfunction. J. Infect. Dis. 169, 1116–1120 (1994).
38. Li, Q. et al. Simian immunodeficiency virus-induced intestinal cell apoptosis is the
underlying mechanism of the regenerative enteropathy of early infection. J. Infect.
Dis. 197, 420–429 (2008).
39. Estes, J. D. et al. Premature induction of an immunosuppressive T regulatory
response in acute SIV infection. J. Infect. Dis. 193, 703–712 (2006).
40. Reynolds, M. R. et al. The CD81 T-lymphocyte response to major
immunodominant epitopes after vaginal exposure to SIV: too late and too little.
J. Virol. 79, 9228–9235 (2005).
Comprehensive tissue analysis after vaginal exposure that revealed the ‘too late
and too little’ CD8 T-cell response to prevent massive gut CD4 T-cell depletion
in acute SIV infection but a robust response in cervical vaginal tissues that could
potentially prevent acquisition if elicited earlier by a vaccine.
41. Brenchley, J. M. et al. Microbial translocation is a cause of systemic immune
activation in chronic HIV infection. Nature Med. 12, 1365–1371 (2006).
42. Estes, J. D. et al. Simian immunodeficiency virus-induced lymphatic tissue fibrosis
is mediated by transforming growth factor b1-positive regulatory T cells and
begins in early infection. J. Infect. Dis. 195, 551–561 (2007).
43. Estes, J. D., Haase, A. T. & Schacker, T. W. The role of collagen deposition in
depleting CD41 T cells and limiting reconstitution of HIV-1 and SIV infections
through damage to the secondary lymphoid organ niche. Semin. Immunol. 20,
181–186 (2008).
44. Wolinsky, S. et al. Selective transmission of human deficiency virus type 1 variants
from mothers to infants. Science 255, 1134–1137 (1992).
45. Zhu, T. et al. Genetic characterization of human immunodeficiency virus type 1 in
blood and genital secretions: evidence for viral compartmentalization and
selection during sexual transmission. J. Virol. 70, 3098–3107 (1996).
46. Derdeyn, C. A. et al. Envelope-constrained neutralization-sensitive HIV-1 after
heterosexual transmission. Science 303, 2019–2022 (2004).
First description of the monophyletic nature of viruses transmitted
heterosexually.
47. Keele, B. F. et al. Identification and characterization of transmitted and early
founder virus envelopes in primary HIV-1 infection. Proc. Natl Acad. Sci. USA 105,
7552–7557 (2008).
Comprehensive sequence analysis showing that a single virus genotype initiates
the vast majority of HIV-1 infections.
48. Salazar-Gonzalez, J. F. et al. Genetic identity, biological phenotype, and
evolutionary pathways of transmitted/founder viruses in acute and early HIV-1
infection. J. Exp. Med. 206, 1273–1289 (2009).
49. Norvell, M. K., Benrubi, G. I. & Thompson, R. J. Investigation of microtrauma after
sexual intercourse. J. Reprod. Med. 269, 269–271 (1984).
50. Pudney, J., Quayle, A. J. & Anderson, D. J. Immunological microenvironments in
the human vagina and cervix: mediators of cellular immunity are concentrated in
the cervical transformation zone. Biol. Reprod. 73, 1253–1263 (2005).
NATUREjVol 464j11 March 2010
51. O’Connor, D. M. A tissue basis for colposcopic findings. Obstet. Gynecol. Clin.
North Am. 35, 565–582 (2008).
52. Li, Q. et al. Glycerol monolaurate prevents mucosal SIV transmission. Nature 458,
1034–1038 (2009).
Tissue analysis of vaginal transmission revealing local expansion of small,
infected founder populations, in vivo evidence of mucosal epithelial signalling
that elicited an inflammatory response to fuel local expansion, and the ability of
glycerol monolaurate to prevent transmission potentially by blocking signalling
at mucosal front lines.
53. Myer, L., Wright, T. C. Jr, Denny, L. & Kuhn, L. Nested case-control study of
cervical mucosal lesions ectopy, and incident HIV infection among women in
Cape Town, South Africa. Sex. Transm. Dis. 33, 683–687 (2006).
54. Miller, C. J., Alexander, N. J., Vogel, P., Anderson, J. & Marx, P. A. Mechanism of
genital transmission of SIV: a hypothesis based on transmission studies and
location of SIV in the genital tract of chronically infected female rhesus macaques.
J. Med. Primatol. 21, 64–68 (1992).
55. Kell, P. D., Barton, S. E., Edmonds, D. K. & Boag, F. C. HIV infection in a patient with
Meyer-Rokitansky-Kuster-Hauser syndrome. J. R. Soc. Med. 85, 706–707 (1992).
56. Padian, N. S. et al. Diaphragm and lubricant gel for prevention of HIV acquisition in
southern African women: a randomised controlled trial. Lancet 370, 251–261 (2007).
57. Berger, E. A., Murphy, P. M. & Farber, J. M. Chemokine receptors as HIV-1
coreceptors: Roles in viral entry, tropism, and disease. Annu. Rev. Immunol. 17,
657–700 (1999).
58. Arthos, J. et al. HIV-1 envelope binds to and signals through integrin a4b7, the gut
mucosal homing receptor for peripheral T cells. Nature Immunol. 9, 301–309 (2008).
59. Kader, M. et al. a4b7hiCD41 memory T cells harbor most TH -17 cells and are
preferentially infected during acute SIV infection. Mucosal Immunol. 2, 439–449
(2009).
60. Zhang, Z.-Q. et al. Roles of substrate availability and infection of resting and
activated CD41 T cells in transmission and acute simian immunodeficiency virus
infection. Proc. Natl Acad. Sci. USA 101, 5640–5645 (2004).
61. Schacker, T. S. et al. Productive infection of T cells in lymphoid tissues during
primary and early human immunodeficiency virus infection. J. Infect. Dis. 183,
555–562 (2001).
62. Ma, Z., Lü, F. X., Torten, M. & Miller, C. J. The number and distribution of immune
cells in the cervicovaginal mucosa remain constant throughout the menstrual
cycle of rhesus macaques. Clin. Immunol. 100, 240–249 (2001).
63. Edwards, J. N. T. & Morris, H. B. Langerhans’ cells and lymphocyte subsets in the
female genital tract. Br. J. Obstet. Gynaecol. 92, 974–982 (1985).
64. Wira, C. R., Fahey, J. V., Sentman, C. L., Pioli, P. A. & Shen, L. Innate and adaptive
immunity in female genital tract: cellular responses and interactions. Immunol.
Rev. 206, 306–335 (2005).
65. Wira, C. R., Grant-Tschudy, K. S. & Crane-Godreau, M. A. Epithelial cells in the
female reproductive tract: a central role as sentinels of immune protection. Am. J.
Reprod. Immunol. 53, 65–76 (2005).
66. Fahey, J. V. et al. Estradiol selectively regulates innate immune function by
polarized human uterine epithelial cells in culture. Mucosal Immunol. 1, 317–325
(2008).
67. Dieu-nosjean, M. et al. Macrophage inflammatory protein 3a is expressed at
inflamed epithelial surfaces and is the most potent chemokine known in attracting
Langerhans cell precursors. J. Exp. Med. 192, 705–717 (2000).
68. Cremel, M. et al. Characterization of CCL20 secretion by human epithelial vaginal
cells: involvement in Langerhans cell precursor attraction. J. Leukoc. Biol. 78,
158–166 (2005).
69. Abel, K., Rocke, D. M., Chohan, B., Fritts, L. & Miller, C. J. Temporal and anatomic
relationship between virus replication and cytokine gene expression after vaginal
simian immunodeficiency virus infection. J. Virol. 79, 12164–12172 (2005).
70. Gray, R. H. et al. Probability of HIV-1 transmission per coital act in monogamous,
heterosexual, HIV-1-discordant couples in Rakai, Uganda. Lancet 357, 1149–1153
(2001).
71. Wawer, M. J. et al. Rates of HIV-1 transmission per coital act, by stage of HIV-1
infection, in Rakai, Uganda. J. Infect. Dis. 191, 1403–1409 (2005).
72. Vendrame, D., Sourisseau, M., Perrin, V., Schwartz, O. & Mamano, F. Partial
inhibition of human immunodeficiency virus replication by type I interferons:
Impact of cell-to-cell viral transfer. J. Virol. 83, 10527–10537 (2009).
73. Rudnicka, D. et al. Simultaneous cell-to-cell transmission of human
immunodeficiency virus to multiple targets through polysynapses. J. Virol. 83,
6234–6246 (2009).
74. Wang, Y. et al. The toll-like receptor 7 (TLR7) agonist, imiquimod and the TLR9
agonist, CpG ODN, induce antiviral cytokines and chemokines but do not prevent
vaginal transmission of simian immunodeficiency virus when applied
intravaginally to rhesus macaques. J. Virol. 79, 14355–14370 (2005).
75. Glavin, S. R. & Cohen, M. S. The role of sexually transmitted diseases in HIV
transmission. Nature Rev. Microbiol. 2, 33–42 (2004).
76. Kaul, R. et al. The genital tract immune milieu: an important determinant of HIV
susceptibility and secondary transmission. J. Reprod. Immunol. 77, 32–40 (2008).
77. Haaland, R. E. et al. Inflammatory genital infections mitigate a severe genetic
bottleneck in heterosexual transmission of subtype A and C HIV-1. PLoS Pathog. 5,
1–13 (2009).
©2010 Macmillan Publishers Limited. All rights reserved
REVIEWS
78. Celum, C. et al. Effect of acyclovir on HIV-1 acquisition in herpes simplex virus 2
seropositive women and men who have sex with men: a randomized, double-
blind, placebo-controlled trial. Lancet 371, 2109–2119 (2008).
79. Watson-Jones, D. et al. Effect of herpes simplex suppression on incidence of HIV
among women in Tanzania. N. Engl. J. Med. 358, 1560–1571 (2008).
80. Zhu, J. et al. Persistence of HIV-1 receptor-positive cells after HSV-2 reactivation
is a potential mechanism for increased HIV-1 acquisition. Nature Med. 15,
886–893 (2009).
Report of persistent target cells in healed HSV-2 ulcers after acyclovir treatment
as the potential explanation for the failure of treatment to reduce HIV-1
acquisition.
81. Robertson, S. A. Seminal plasma and male factor signalling in the female
reproductive tract. Cell Tissue Res. 322, 43–52 (2005).
82. Robertson, S. A., Ingman, W. V., O’Leary, S., Sharkey, D. J. & Tremellen, K. P.
Transforming growth factor b—a mediator of immune deviation in seminal
plasma. J. Reprod. Immunol. 57, 109–128 (2002).
83. Berlier, W. et al. Seminal plasma promotes the attraction of Langerhans cells via
the secretion of CCL20 by vaginal epithelial cells: involvement in the sexual
transmission of HIV. Hum. Reprod. 21, 1135–1142 (2006).
84. Münch, J. et al. Semen-derived amyloid fibrils drastically enhance HIV infection.
Cell 131, 1059–1071 (2007).
85. Li, Q. et al. Visualizing antigen-specific and infected cells in situ predicts outcomes
in early viral infection. Science 323, 1726–1729 (2009).
Report of the tissue-compartment-specific nature of the cellular immune
response that reveals the importance of the relative numbers of both immune
effectors and infected targets in determining whether viral infection is cleared,
and the extent of control if infection is not cleared.
86. Belyakov, I. M. et al. Impact of vaccine-induced mucosal high-avidity CD81 CTLs
in delay of AIDS viral dissemination from mucosa. Blood 107, 3258–3264 (2006).
87. Genescà, M., Skinner, P. J., Bost, K. M., Lu, D. & Wang, Y. Protective attenuated
lentivirus immunization induces SIV-specific T cells in the genital tract of rhesus
monkeys. Mucosal Immunol. 1, 219–228 (2008).
88. Hansen, S. G. et al. Effector memory T cell responses are associated with
protection of rhesus monkeys from mucosal simian immunodeficiency virus
challenge. Nature Med. 15, 293–299 (2009).
Report of impressive protection against rectal transmission associated with
effector memory T-cell responses elicited by a rhesus CMV vaccine.
89. Lifson, J. D. et al. Containment of simian immunodeficiency virus infection:
Cellular immune responses and protection from rechallenge following transient
postinoculation antiretroviral treatment. J. Virol. 74, 2584–2593 (2000).
90. Veazey, R. S. et al. Protection of macaques from vaginal SHIV challenge by
vaginally delivered inhibitors of virus-cell fusion. Nature 438, 99–102 (2005).
91. Lederman, M. M. et al. Prevention of vaginal SHIV transmission in rhesus
macaques through inhibition of CCR5. Science 306, 485–487 (2004).
92. Cranage, M. et al. Prevention of SIV rectal transmission and priming of T cell
responses in macaques after local pre-exposure application of tenofovir gel. PLoS
Med. 5, 1238–1250 (2008).
93. Zhu, T. et al. Persistence of extraordinarily low levels of genetically homogeneous
human immunodeficiency virus type 1 in exposed seronegative individuals. J. Virol.
77, 6108–6116 (2003).
94. McChesney, M. B. et al. Occult systemic infection and persistent simian
immunodeficiency virus (SIV)-specific CD41-T-cell proliferative responses in
rhesus macaques that were transiently viremic after intravaginal inoculation of
SIV. J. Virol. 72, 10029–10035 (1998).
95. Ma, Z.-M., Abel, K., Rourke, T., Wang, Y. & Miller, C. J. A period of transient
viremia and occult infection precedes persistent viremia and antiviral immune
responses during multiple low-dose intravaginal simian immunodeficiency virus
inoculations. J. Virol. 78, 14048–14052 (2004).
96. Trivedi, P. et al. Intrarectal transmission of simian immunodeficiency virus in
rhesus macaques: selective amplification and host responses to transient or
persistent viremia. J. Virol. 70, 6876–6883 (1996).
97. Li, Q. et al. Microarray analysis of lymphatic tissue reveals stage-specific, gene
expression signatures in HIV-1 infection. J. Immunol. 183, 1975–1982 (2009).
98. Virgin, H. W. & Walker, B. D. Immunology and the elusive AIDS vaccine. Nature.
(in the press).
Acknowledgements I thank J. V. Carlis, R. P. Johnson, Q. Li, J. D. Lifson,
D. Masopust, S. Pambuccian, P. J. Southern, J. Estes, D. Douek, H. W. Virgin and
B. D. Walker for discussions. Errors of commission are mine as are errors of
omission, with apologies to the authors of work that I could not cite because of
space limitations and exclusive focus on tissue analyses. I thank C. O’Neill and
T. Leonard for help with the manuscript and figures. Work from my laboratory cited
in the review was supported by grants from the National Institutes of Health (AI
38565, AI 48484, AI 71976) and the International AIDS Vaccine Initiative.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The author declares no competing financial interests.
Correspondence should be addressed to the author (haase001@umn.edu).
223
 Vol 464j11 March 2010jdoi:10.1038/nature08898

PERSPECTIVES
Immunology and the elusive AIDS vaccine
Herbert W. Virgin1 & Bruce D. Walker2
Developing a human immunodeficiency virus (HIV) vaccine is critical to end the global acquired immunodeficiency syndrome
(AIDS) epidemic, but many question whether this goal is achievable. Natural immunity is not protective, and despite
immunogenicity of HIV vaccine candidates, human trials have exclusively yielded disappointing results. Nevertheless, there
is an indication that success may be possible, but this will be dependent on understanding the antiviral immune response in
unprecedented depth to identify and engineer the types of immunity required. Here we outline fundamental immunological
questions that need to be answered to develop a protective HIV vaccine, and the immediate need to harness a much broader
scientific community to achieve this goal.

nti-HIV drugs have extended the lives of millions of Gaps in knowledge

A infected people, but the epidemic continues its course,

with high rates of new infections in regions least able to
cope medically, socially and financially. For example, in
some particularly hard hit areas infection rates in young women soar
from less than 1% at age 15 to in excess of 50% by their mid-twenties.
A vaccine will be required to ultimately conquer AIDS, but many
scientists believe that this is probably decades away, if even possible.
Thus far, HIV vaccine trials in humans have resulted in either no or
relatively unimpressive protection despite measureable immuno-
genicity of administered HIV antigens1–4. Indeed, natural immunity
itself is far from protective, inducing both humoral and cellular
immune responses but not leading to spontaneous clearance or pro-
tecting against HIV superinfection5. Therefore, simply generating an
immune response similar to what is generated in natural infection is
unlikely to immunize effectively against HIV. This conclusion sup-
ports the view that our fundamental approach to HIV vaccination
needs to be re-examined.
We believe that the development of an effective HIV vaccine is an
achievable goal, and that there are many ‘knowable unknowns’ that
can be addressed immediately to advance this effort. The goal of this
Perspective is to outline progress from studies of HIV and the closely
related AIDS-inducing monkey retrovirus simian immunodeficiency
virus (SIV), together with insights derived from studies of other
viruses, showing that some level of control of lentiviruses can be
achieved. This suggests that there are immune mechanisms that, if
induced by a vaccine, may prevent HIV infection or at least limit the
level of viraemia in infected people. In addition to the critical value in
continuing and expanding current efforts to test vaccine candidates
in humans, we emphasize the importance of a long-term plan to
enhance understanding of basic immunological mechanisms that
may be able to prevent amplification and systemic spread from the
limited initial nidus of HIV infection (see the accompanying paper6).
We need better integration of novel concepts and information on
new immune system genes and mechanisms derived from work in
model systems into vaccine science. These advances need to be
rapidly translated into the HIV field to optimize chances for effective
vaccination against AIDS. Achieving this goal will be greatly facili-
tated by engagement of scientists from diverse but currently isolated
disciplines, who, by providing new insights, can accelerate progress
towards an effective vaccine.
1
Washington University School of Medicine and Midwest Regional Center of Excellence for Biodefense and Emerging Infectious Disease Research, Campus Box 8118, 660 South Euclid
Avenue, Saint Louis, Missouri 63110, USA. 2Ragon Institute of Massachusetts General Hospital, Massachusetts Institute of Technology and Harvard University, Howard Hughes
Medical Institute, 149 13th Street, Charlestown, Massachusetts 02129, USA.
Given the failure of empirical HIV vaccine approaches, understanding
the nature of the immune response needed for protection is the essential
missing ingredient. We still lack fundamental knowledge regarding the
nature, quality and quantity of immune responses that should be
induced, the ideal antigens to include, how to overcome the tremendous
sequence variability engendered by the error-prone HIV reverse tran-
scriptase, or even whether preventive vaccine strategies should focus on
protection from infection or protection from disease progression.
Classically, immunization has depended on programming memory
T and B cells to remember encounter with antigen. The current
vaccine failures indicate that this is not sufficient, and that immuni-
zation strategies need now to be engineered with close attention to the
type and differentiation state of the immune response generated,
the efficacy of vaccine-induced effector mechanisms against specific
vulnerable steps in the pathogenesis of HIV infection, the capacity of
the induced response to maintain activity at relevant body surfaces, the
influence of the virome and microbiome on vaccine responses, the
fitness cost of mutations forced on HIV by specific immune responses,
the capacity of vaccine-generated immune responses to avoid exhaus-
tion, and the capacity of the different parts of the immune response
generated by vaccine to synergize with one another at the relevant sites.
Because HIV initially establishes infection in a very small number of
cells6, appropriate pre-existing immunity has the potential to prevent
infection, clear infection, or limit replication so that the set point of
viraemia is low enough during chronic infection to prevent progres-
sion to AIDS and inhibit spread from one person to another. A better
understanding of the molecular mechanisms responsible for immune
control of acute and chronic viral infections, and immune interactions
with HIV at various stages of disease (Fig. 1 and ref. 6) will be needed to
direct effective vaccine strategies to meet this desperate global need.
Signals that vaccination may work
There are, finally, weak signals from a human vaccine trial, and more
robust findings from efforts to vaccinate against SIV, indicating that a
measure of protective immunity to HIV or SIV can be induced. A recent
phase III trial combining an HIV envelope protein immunogen with a
recombinant canarypox-HIV vector suggested an encouraging (albeit
insufficient) 30% reduction in acquisition1. In monkeys, induction of
virus-specific CD8 T-cell responses does not prevent infection, but if
strong and broad enough, can curtail SIV replication after homologous

224
©2010 Macmillan Publishers Limited. All rights reserved
NATUREjVol 464j11 March 2010 PERSPECTIVES

Figure 1 | HIV pathogenesis and outstanding

What form(s) of HIV spreads?
Free HIV? Cell-associated HIV?
HIV virus
Is HIV pathogenesis or immunity
different at different mucosal sites?
questions related to vaccination. Shown are the
steps in HIV pathogenesis and ‘knowable
unknowns’ that, when defined, may alter how we
approach vaccination against HIV. APC, antigen-
presenting cell.

Does compression of the quasispecies

associated with transmission have
CD4 T cell Dendritic cell
implications for vaccines?
Does HIV productively infect and/or
impair the function of APCs?
Can HIV latency be
Do innate iimmune mechanisms How are APCs optimally licensed prevented or eliminated?
c acquisition?
affect for induction of immunity?

Replicate in Viraemia
de
How does HIV evade HI be cleared by
Can HIV
mucosa nity once
immunity o infection
?
vaccine immunity?
i established?
is estab
t b

Blood

What determines HIV fitness?

ess?
For transmission?
Secondary infection ty?
For evasion of immunity?
lymphoid organs For establishment of set point?
Are specific viral genotypes
associated with elite controllers?
CD4 T-cell depletion
Lymph
node

AIDS

or heterologous virus challenge7,8. Use of simian cytomegalovirus
(CMV) as a vector to continuously deliver SIV antigens limits systemic
infection in some animals after intrarectal challenge with SIV9. These
and similar studies show that exposure to lentivirus antigens can lead to
resistance to or control of infection and provide controlled models to
address substantial gaps in our knowledge of what immune mechan-
isms are responsible for the observed effects.
Encouragement afforded by these vaccine results is enhanced by
studies of natural transmission that indicate that induction of even
modest protective immunity at mucosal surfaces may tip the balance
towards less efficient spread of HIV. Only one-third of infants born to
infected mothers acquire infection; sexual transmission occurs after
between 1 out of 100 and 1 out of 1,000 exposures (although the risk
may be far higher when the transmitting partner is acutely infected)10;
exposure to infected blood products does not consistently lead to
infection11; commercial sex workers may remain uninfected despite
extensive exposure12; and most infections are initiated by a single
virus13. These data indicate that the induction of even modest pro-
tective immunity at mucosal surfaces before infection may be effec-
tive at inhibiting spread of HIV. Indeed, this may be the mechanism
of apparent protection in the recent Thai vaccine trial, which limited
HIV acquisition but had no effect on set-point viral load once
systemic infection was established. Detailed studies of the earliest
events in transmission in animal models and humans are needed to
define the mechanisms that explain resistance to infection in the vast
majority of HIV exposures.
Détente teaches us how to go to war
Not all HIV infections lead to AIDS, indicating that it is possible for
immunity to establish a very long-term détente with lentiviruses;
the mechanisms involved might provide insight into the goals for
©2010 Macmillan Publishers Limited. All rights reserved
vaccination (Fig. 2). For example, about one in 300 people maintain
extremely low HIV loads without treatment, some for more than 30
years14, and HIV-2 pathogenesis is similar to HIV-1 pathogenesis and
yet causes disease far less frequently15. Similarly, SIV can cause high-
level viraemia, infect and deplete CD4 cells, and yet not cause AIDS16.
Together these data indicate that there are as-yet-unidentified viral or
host factors that attenuate lentivirus pathogenesis. Identification of
these may define the type of immune responses needed to reach the
goal of vaccinating to prevent disease progression and spread within
the population (Fig. 2). A key step in this direction would be iden-
tification of novel molecules involved in control of chronic virus
infection in model systems, and rapid translation of such new findings
for detailed analysis during HIV and SIV infection.
Fight using all the tools immunity offers
Current evidence overwhelmingly argues for vaccines that simulta-
neously elicit effective CD4, CD8 and B-cell responses; however, an
integrated understanding of the relationships between innate and
adaptive immune responses to HIV antigens, and how this might
lead to enhanced immunization, is yet to be achieved (Fig. 3).
Furthermore, although it is clear that innate immune mechanisms
contribute to control of HIV17, it is not clear whether it is possible to
harness this response via vaccination. Indeed, the HIV vaccine field
has been largely divided into B-cell approaches and CD8 T-cell
approaches. There is little evidence that any vaccine achieves protec-
tion through antibody alone, or ever prevents initial infection
entirely; the same holds for T-cell responses. Antibody functions
optimally in the setting of simultaneous CD4 T-cell responses that
provide help; CD4 and CD8 responses are probably needed to limit
or eliminate HIV replication at the initial site of infection through
innate and adaptive mechanisms. Similarly, T-cell vaccines largely
225
PERSPECTIVES
HIV: Immune people unavailable,
Define correlates elite controllers limit progression
of immunity SIV: Protection from acquisition and
lowering of set point demonstrated,
mechanisms unknown
HIV: Uncertain
Choose
se desired
SIV: Neutralizing antibody,
immune responses
es
CD8 T cells, CD4 T cells
Vaccine goal
Prevent infection? Vaccine strategy Vaccine
Limit set point? development
Prevent progression?
Vector?
Adjuvant? and trial
Prevent spread? Dose?
Frequency?
Choose antigen HIV: Highly variable
targets rapid mutation
Define genome HIV: Well described,
moving target as
and proteome epidemic expands
Figure 2 | Approach to vaccination and special challenges of HIV
vaccination. Shown are the steps in development of a vaccine strategy for
HIV. In blue are the key issues corresponding to each step that influence our
capacity to immunize against HIV.
focused on CD8 T-cell responses7,8 are unlikely to succeed without
antibody and HIV-specific CD4 T-cell responses9,18. Even if we knew
what we wanted to induce in a vaccine, the science of engineering
specific immune responses to exceed the capacity of the natural
immune response is in its infancy. A key step in this important
direction will be to define relationships and potential synergies
between CD4 T-cell, CD8 T-cell, antibody and B-cell responses
during vaccination and in situations in which long-term viral control
is achieved in infected people.
The case for re-embracing CD4 T cells
There is much controversy concerning the role of HIV-specific CD4
T cells in control versus promotion of HIV replication. HIV repli-
cates best in vitro in activated CD4 T cells, there is evidence for
selective infection of HIV-specific CD4 T cells by HIV19, and a cor-
relation between SIV recrudescence and CD4 T-cell activation after
depletion of CD8 T cells has been reported20. Such observations
support concerns that induction of a large population of HIV-
immune CD4 T cells might enhance rather than control HIV infec-
tion. This view is countered by overwhelming evidence that CD4 T
cells are essential for antiviral responses in vivo including CD8 T-cell
responses and effective B-cell responses in every system in which it
has been studied (for example, see refs 21–23 and references therein).
For example, CD8 T cells do not efficiently access mucosal sites to
combat viral infection without CD4 T-cell help18, and the outcome of
retroviral infection depends critically on speed and extent of virus-
specific CD4 T-cell responses24. The basic rules of immunology there-
fore indicate that it will be impossible to generate long-lived, stable,
high-level vaccine-induced immunity to HIV without inducing a
protective CD4 T-cell response, but most HIV vaccine strategies have
not specifically focused on inducing these responses.
Current data strongly suggest that HIV-specific CD4 T cells are
beneficial rather than detrimental. The initial burst of HIV and SIV
replication during acute infection does not occur in HIV- or SIV-
specific CD4 T cells6, and whereas infection of these cells could con-
ceivably add fuel to this early fire, a strong effector-memory CD4 T-cell
response is associated with diminished, not enhanced, SIV replication
after intrarectal challenge9. Moreover, although HIV preferentially
infects HIV-specific CD4 cells, the overall magnitude of this increase
is small, and the great majority of HIV-specific CD4 T cells are not
infected in vivo, even in the presence of high-level viraemia19. Blockade
of CD4 T-cell activation after CD8 T-cell depletion using anti-
interleukin (IL)-15 antibody in SIV-infected macaques has no effect
226
©2010 Macmillan Publishers Limited. All rights reserved
NATUREjVol 464j11 March 2010
on SIV replication25, indicating that previous associations between
CD4 T-cell activation and increased viraemia are correlative only20.
Furthermore, CD4 T cells can provide direct antiviral effects24, and
the presence of strong virus-specific CD4 T-cell responses distinguishes
non-pathogenic HIV-2 infection from HIV-1 infection15. Possible sup-
port for a role for immune CD4 T cells in vaccination may be found in
the Thai HIV vaccine trial results, in which decreased HIV acquisition
was observed with a vaccine that induces HIV-specific CD4 T cells1.
New data, A much greater emphasis on understanding the antiviral effector
assess correlates,
revise approach
properties of CD4 T cells is needed, and the role of these cells in control
of HIV at mucosal sites and vaccination must be one of the highest
priorities for future studies.
B cells and antibody come back into focus
Animal challenge studies indicate that a pre-existing neutralizing
antibody response can prevent AIDS virus infection. Notably, a signifi-
cant minority of HIV infections generate heterogeneous mixtures of
polyclonal high-affinity broadly neutralizing antibodies to multiple
different envelope epitopes26, giving hope that properly formulated
vaccines might generate protective responses. However, it is unknown
whether vaccine-induced elicitation of such antibodies can counter
enormous and growing HIV quasispecies diversity and prevent initial
infection on exposure. So far, vaccine-elicited antibody responses have
not been protective and the form of the virus (free versus intracellular)
that spreads is unknown (Fig. 1) so that epitopes exposed during
natural transmission are undefined. Moreover, the envelope glyco-
protein is very difficult to produce in its folded trimeric form; this
may be required to provide an effective immunogen. Furthermore,
because HIV spread between cells may occur by means of multiple
mechanisms, including virological synapse-mediated cell adhesion
coupled to viral endocytosis, it is not clear that antibody will be able
to inhibit all forms of viral spread in vivo27.
As only a tiny subset of an infected person’s HIV quasispecies
transmits6, the relationship between the diversity of sequences of
the surface glycoprotein gp120 worldwide and appropriate vaccine
epitopes is unknown, but it is possible that much of this diversity is
irrelevant to immunogen design28,29. Recent results from large popu-
lation screening studies have revealed the presence of novel broadly
neutralizing antibodies in some infected people, and provide important
new avenues for antigen design through targeting these epitopes that
are subdominant or entirely unrecognized in the majority of infected
persons (for example, see refs 28, 30). Reasonable levels of one of the
few available broadly neutralizing monoclonal antibodies blocks low-
dose intravaginal challenge with SIV using the HIV gp120 (SHIV)31,32,
and encouraging recent studies show that complete protection from
SIV challenge has been achieved by adeno-associated virus (AAV)-
mediated continuous endogenous expression of immunoadhesins
derived from neutralizing antibodies33.
Even if protective B-cell epitopes are found, the mechanisms
responsible for antibody action in vivo need to be much better under-
stood to foster induction of a polyfunctional humoral response
(Fig. 3)30,34,35,. Meeting this need is complicated by the lack of sur-
rogate in vitro assays, other than virus neutralization, for protection.
We may be missing epitopes that generate antibodies that protect
through antibody-dependent cell-mediated cytotoxicity, complement
fixation, or the capacity to block infection without neutralizing, as
observed for other viruses (see refs 36–39 and references therein).
Antibodies may also be important through interaction with Fc recep-
tors on granulocytes, macrophages, dendritic cells, or B cells to alter
antigen presentation, tropism, or replication in vivo36. Furthermore,
B cells are more than simply factories for making antibody, as B cells
unable to make antiviral antibody exert T-cell-dependent control of
chronic virus infection in mice40. Efficient antibody-mediated control
of Friend leukaemia retrovirus requires CD4 and CD8 T-cell responses
and is regulated by specific major histocompatibility complex (MHC)
alleles37,41. This may be particularly relevant to AIDS as specific MHC
alleles are associated with more effective immune control of HIV
NATUREjVol 464j11 March 2010 PERSPECTIVES

Can anybody prevent infection? Figure 3 | Immune responses to HIV and

Can the envelope be expressed as Is neutralization the only relevant
outstanding questions related to vaccination.
an antigen in authentic form? assay for protective antibody Shown are the steps in the immune response to
responses to HIV? HIV. For each component of the immune
How do we generate broadly response, colour-matched ‘knowable unknowns’
Can the initial conditions of vaccination: cross-protective antibodies? are specified which, when defined, may alter how
Enhance mucosal response? Antibody we approach vaccination against HIV.
Enhance memory? Would a strong antibody response
Deal with HIV diversity? improve the efficacy of T cells?
Generate effector cells present constantly? B cell How do we optimize antibody
Prevent T-cell or B-cell exhaustion?
responses at mucosal surfaces
Generate relevant effector mechansims?
where they are needed?
Essential help Can the immune system clear
established infection by targeting
replication and latently infected cells?
APC CD4 Target
T cell cell Are there undiscovered immune
effector functions relevant to HIV?
CD4?
C CD8? Antibody? Innate immunity?
Essential help
Is it possible to target T cell and antibody
Can APCs be altered epitopes that extract a high cost in viral
or targeted to optimize fitness when they mutate?
effective T-cell responses? CD8
T cell Can the most potent antiviral CD8 T-cell
responses be defined?
Are T-cell mechanisms operative
to control set point relevant correlates of Would pre-exisiting immune CD4 T cells
protection for vaccine development? help or hurt immune control of HIV infection?

How can we generate the optimal CD4 helper

response for protective antibody and CD8 responses?
How do we get immune T cells to the Are there undiscovered mechanisms
mucosal surfaces where they are needed? of T-cell antigen recognition?

infection. Unexplored interactions between B cells, antibody res-
ponses and T cells are knowable unknowns that are probably critical
to HIV vaccine strategies (Fig. 3).
A major future challenge will be converting information on pro-
tective epitopes and antibody-mediated mechanisms of protection
into protective vaccine antigens. Antibodies to HIV often target
accessible gp120 epitopes that can mutate without fitness cost rather
than subdominant epitopes that may be more conserved and thus
more targetable. A fundamental question is how to generate CD4
T-helper cell activity for vaccine antibody responses to potentially
protective subdominant epitopes. One approach to generating
protective responses to subdominant epitopes is inoculation
with viral vectors that express blocking antibody-like molecules33.
Alternatively, the HIV glycoprotein may have to be engineered to
optimize responses to cross-protective subdominant epitopes.
CD8 T cells are important yet insufficient
The failure of the first T-cell HIV vaccine in humans2,3 indicates that
more needs to be understood about the breadth, specificity and quality
of protective CD8 responses to HIV to optimize vaccine design
(Fig. 3). Targeting of Gag epitopes is associated with lower viral loads
than targeting Env, a finding that needs to be evaluated in SIV models
to assess whether this should inform vaccine design42. Some SIV- and
HIV-specific T cells are surprisingly ineffective at killing virus-infected
targets43,44, and the relevance of killing to the protective activity of CD8
T cells in vivo remains unknown. Immunoregulatory networks that
exhaust adaptive CD4 and CD8 responses during chronic infection
probably contribute to CD8 ineffectiveness45–48. Studies using assays
that more closely replicate the in vivo situation in which viral antigens
are processed and presented by primary cells rather than being pro-
vided as synthetic peptides at supra-physiological doses are urgently
needed, along with high-throughput single-cell analysis of T-cell phe-
notypes and gene expression, to define differences between effective
and ineffective T-cell responses49.
Tuning the immune system to fight AIDS
To vaccinate against HIV we need to induce B and T cells that have
and retain very specific differentiated properties allowing them to
control replication of established infection and to function during
©2010 Macmillan Publishers Limited. All rights reserved
the first critical hours of infection in the intestine or genitourinary
tract (Fig. 1 and ref. 6). A major opportunity to engineer vaccine
responses comes from studies in mice and non-human primates
showing that the immune system is dampened during persistent
virus infection, generating ‘exhausted’ T cells to prevent tissue
damage (reviewed in ref. 22). The key observation for vaccine design
is that manipulation of the mechanisms responsible for exhaustion at
the time of first antigen exposure can fundamentally alter vaccine
responses.
Exhausted cells express downregulatory receptors, lack multiple
effector functions, and show weak proliferative responses. Exhaus-
tion is best studied in T cells, but may also affect B cells50,51. For T cells,
exhaustion is mediated by factors including FOXP3-positive regula-
tory T (Treg) cells, inhibitory receptors such as PD1 and CTLA4 and
others22,48, and cytokines such as IL-10 (refs 52, 53) or transforming
growth factor-b54. In mice infected with lymphocytic choriomenin-
gitis virus (LCMV) recent data show that IL-21, perhaps made by
CD4 T cells, prevents exhaustion55–57. Blockade of IL-10 or PD1
increases control of chronic LCMV and or SIV infection52,58,59. This
is relevant to vaccine strategies as control of chronic LCMV infection
is enhanced by either blockade of IL-10 during therapeutic DNA
vaccination52 or blockade of PD1 during vaccinia vector vaccina-
tion60. These data support linking studies of immunoregulatory
mechanisms operating during chronic infection to vaccination
efforts. IL-10 and PD1 inhibit CD8 responses by independent
mechanisms; simultaneous blockade of both may be advantageous53.
New data show that administration of the ‘immunosuppressive’ drug
rapamycin counterintuitively enhances memory CD8 T-cell differ-
entiation61, indicating that pharmacological manipulations at the time
of vaccination may enhance differentiation of induced T or B cells.
Administration of antigen in specific contexts such as via dendritic
cells may induce immune responses that can enhance chronic
immune control of HIV62. This striking finding indicates that similar
approaches to vaccination might induce immune responses that
either eliminate the initial nidus of HIV infection or limit replication
and thus control set point viraemia. Similarly, one might argue to
block induction of Treg cells during vaccination. However, counter-
intuitively, elimination of Treg cells can blunt the mucosal immune
response and control of vaginal infection with herpes simplex virus, a
227
PERSPECTIVES
finding potentially related to altering the balance between mucosal
and lymphatic responses63. This provides a cautionary note as one
may not be able to translate all mechanisms of chronic control of HIV
to the vaccination effort. Another major issue is how to use adjuvants,
which may, in addition to enhancing the overall immunogenicity of
an antigen, be used to direct and optimize the differentiation of cells
responding to vaccine antigens. Here again are knowable unknowns:
it is important to determine how to engineer augmented vaccine
responses via manipulation of immunoregulatory pathways during
vaccination (Fig. 3).
Battling HIV in mucosal tissues
HIV delivers a near-lethal hit to the host immune system by killing a
huge proportion of T cells in the gastrointestinal tract (ref. 6 and
Fig. 1), and persistent viraemia may feed off of systemic immune
activation associated with mucosal damage during HIV infection.
Mucosal events may contribute to HIV and SIV progression,
variation in vaccine responses, and variations in measurements of
surrogate markers of protection. Continuous activation of mucosal
HIV-immune cells may be required for protective vaccine responses
given the speed of mucosal HIV pathogenesis6. The immune effector
mechanisms that target HIV in the mucosa are unknown, and rela-
tionships between mucosal and systemic immunity are poorly under-
stood, leading to uncertainty over the benefits of mucosal versus
systemic vaccination. Tissue-specific immune responses have a key
role in control of viral infections22,64,65, and thus differences among
the vagina, penis, intestine and oropharynx might determine vaccine
efficacy. Furthermore, emerging evidence from humans and mice
suggests that specific mechanisms maintain local immunity in skin
and genital tissues64,65. As studies of HIV or SIV immunity typically
analyse peripheral blood cells, we may be working from an inaccurate
view of immunity to HIV and SIV. An increased focus on the role of
mucosal events in HIV vaccination is therefore needed.
Although appropriate emphasis has been placed on local res-
ponses, there is nevertheless convincing evidence of effective induc-
tion of mucosal responses to enteric murine norovirus infection and
human papilloma virus after subcutaneous vaccination66,67, for
mucosal protection against SIV after subcutaneous vaccination with
a CMV-based vector9, and for generation of substantial mucosal
responses after systemic immunization in mice and macaques68.
Thus, systemic vaccination can elicit effective mucosal immunity,
and systemic immunoglobulin-G (IgG) responses, rather than
mucosal IgA responses, may be a reasonable vaccine goal. If lympho-
cytes with specific differentiated properties are essential for mucosal
immunity to HIV they must be better defined to optimize systemic
induction of mucosal immunity. The mucosal immune responses
associated with continuous antigen and potential adjuvant effects69
of live vaccine vectors such as cytomegalovirus9 or live attenuated
SIV70, and with the immunization scheme used in the Thai vaccine
trial, need to be defined.
One of the key sources of data for immune correlates of HIV
immunity that might be converted into vaccine strategies is the nature
of immunity to HIV during chronic infection (Fig. 2). Mucosal-
damage-associated translocation of intestinal contents, or immune-
suppression-associated intestinal infection, may lead to systemic
immune activation that fosters HIV disease progression16,71,. Further-
more, SIV-infected non-progressors show fewer signs of intestinal
barrier dysfunction than progressors16. Studies in mice deficient for
innate immune genes also show that alterations in mucosal barrier
function result in systemic activation of adaptive immunity72.
Mucosal damage is therefore a new factor to be considered with
regards to immune effects of HIV infection, and might well alter
measurements of HIV-immune cells or antibody and thus contribute
to false conclusions about the effects of HIV on immunity, and vice
versa, during chronic infection.
The antigens, pathogens, or microbial products that induce sys-
temic immune activation, and may therefore complicate assessment
228
©2010 Macmillan Publishers Limited. All rights reserved
NATUREjVol 464j11 March 2010
of correlates of HIV immunity, are yet to be defined. Translocated
intestinal products may not be the entire cause of systemic immune
activation. It has been presumed that standard diagnostic tests rule
out a role for pathogens in HIV- and SIV-related enteropathy, but
this is not tenable as metagenomic analyses reveal many novel enteric
viruses in humans and primates (for example, see refs 73, 74) whose
role in HIV/AIDS has not been assessed. Furthermore, monkeys
harbour widely variant populations of intestinal bacteria that may
influence progression to AIDS and systemic vaccine responses75. The
virome and microbiome of the host have profound effects on the
nature of the mucosal and systemic immune response to new patho-
gens, including the basal level of activation of innate immune cells and
the development of T-cell subtypes in the intestinal mucosa22,69,76.
Future studies of immune cell function in HIV-infected individuals
and of HIV vaccine responses need to be designed while bearing in
mind the potential role of the microbiome and virome in determining
vaccine responses.
Exploiting evolution in a vaccine trap
HIV sequences have tremendous diversity; circulating HIV isolates
exhibit up to 35% sequence variation in gp120 (ref. 77) and tremendous
variability is generated in an individual after HIV passes through the
transmission bottleneck6,13,78. The HIV quasispecies probably has
pathogenetic properties independent of the capacity to mutate to evade
immunity. For example, poliovirus studies show that a quasispecies can
confer pathogenetic properties such as neurotropism and virulence
during viral pathogenesis79.
The tremendous variability of HIV also exposes some weaknesses.
In simultaneously infected adult identical twins, the earliest CD8
epitope escape mutations arose at the same nucleotide in both indi-
viduals, indicating that there are constraints on HIV evolution, and
that some escape pathways are detrimental to viral fitness80. Popula-
tion studies show predictable pathways to escape within epitopes
presented by specific HLA alleles81,82. These mutations, particularly
in Gag, can impair viral fitness83,84. It is perhaps not surprising that
CD8 T cells targeting epitopes in Gag are associated with lower viral
load than CD8 T cells targeting the more mutable gp120 (ref. 42).
Part of the role of HLA B*57 in people able to control HIV replication
to undetectable levels without medication (‘elite controllers’) seems
to be generation of fitness-impairing epitope mutations that cannot
easily be compensated83,85. Such CD8 T-cell-induced mutations
associated with lower viral load can revert after transmission, pro-
viding in vivo evidence of impaired fitness86,87. The ability to define
the fitness landscape of HIV now exists, and this information could
have a significant impact on vaccine development.
Careful observation of HIV evolution under selective pressure in
vivo also has the potential to provide key new insights into the nature
of potent immune responses. Clear signatures of selection have been
identified that are not accounted for by known CD4 or CD8 T-cell, or
B-cell, epitopes82; known epitopes do not explain all CD4 and CD8
T-cell responses observed88. CD8 T-cell control of chronic infection
with herpesviruses and polyomaviruses can involve non-classical
MHC molecules, and similar responses have been reported in
humans22. Thus, there are additional epitopes, and also potentially
other mechanisms of innate or adaptive immune control, that await
discovery.
These findings define a weakness in HIV’s most potent defence, the
capacity to mutate, and indicate the potential for effective immuni-
zation via an ‘evolutionary trap’, driving HIV into a fitness dead end
by targeting epitopes in fitness-critical regions of viral proteins.
Experiments in both HIV and in animal models in which immune
pressure on viruses can be easily controlled will be needed to develop
the methods for evolutionary trap vaccination. This will require
integration of computational biology, evolutionary biology and
high-throughput sequencing to determine the sequence space within
which a ‘fit’ virus must reside. Full-length genomes will have to be
analysed to define escape mutations and linkages between these
NATUREjVol 464j11 March 2010
mutations and intragenic and extragenic suppressor mutations that
restore fitness. Studies have already shown that immune-induced
mutations in gp120 are more frequent yet less likely to confer a fitness
cost to HIV84 than those in Gag that impair fitness83,84. However, even
then compensatory mutations can sometimes rapidly restore viral
fitness89. Combining an evolutionary trap with mosaic vaccines based
on computationally derived sequences representing maximal coverage
of HIV diversity is an attractive approach90.
Weapons deployed by vaccination against HIV
In the end, inducing immune responses to HIV will be effective only if
potent effector mechanisms are deployed against the virus. Because
viral integration and establishment of a pool of latently infected cells
occurs very rapidly after transmission (Fig. 1 and ref. 6), these res-
ponses will have to be present continually or arise extremely rapidly to
prevent establishment of chronic infection. Unfortunately, we still do
not know what an effective immune response to HIV includes, and
therefore are compromised in our quest for a vaccine (Figs 1–3).
Recent focus has been on inducing polyfunctional T-cell responses
as a surrogate for effective responses, but which functions are truly
relevant remains unknown. Significant effort needs to be directed to
defining vaccine-relevant effector mechanisms that can eliminate the
initial nidus of HIV infection, prevent systemic spread from an initial
localized mucosal infection, or block the establishment of latency. The
prevention of latency might enhance vaccination by preventing the
virus from accessing a quiescent state of infection invisible to the
immune system (see below). In this area, particular emphasis needs
to be placed on defining novel mechanisms as currently known
effector responses do not consistently correlate with control of HIV.
Clues to potential novel immune effector mechanisms come from
analysing the complex cellular factors that promote or restrict HIV
replication91,92. HIV has evolved mechanisms to thwart many host
proteins that limit viral replication. The importance of defining the
relationship between immune evasion strategies, host restriction and
immunity is shown by the recent demonstration that Rfv3, which
determines the effectiveness of antibody against Friend retrovirus
in mice, is encoded by Apobec3, which encodes a protein that is a
homologue of human APOBEC3G and APOBEC3F proteins targeted
by HIV vif 93–95. The task we face is defining mechanisms that are not
effectively blocked by HIV and which can be deployed by vaccine-
generated immune cells. Cytotoxic killing of an infected cell requires
recognition that occurs optimally only after HIV is integrated into
the genome and proteins are expressed. Breaking the cycle of replica-
tion and latency may therefore require different or additional
mechanisms, in particular those that block early events in HIV patho-
genesis (Fig. 1 and ref. 6).
One potential weapon that might be used against HIV would be
prevention or regulation of latency. It is possible that the capacity of
HIV to enter an immunologically silent latent state in which the
provirus is present but antigenic proteins are not expressed early after
infection of the mucosal surface could be responsible for vaccine
failures or inefficient vaccination. Thus, immune responses that pre-
vent the establishment of latency might enhance vaccine effective-
ness. Latent reservoirs of HIV and SIV are incompletely defined96,
and this knowledge is critical to the vaccine effort. Reactivation of
HIV from latency is regulated by multiple overlapping factors includ-
ing chromatinization of the provirus, methylation, T-cell activation,
NFkB and cytokines including tumour-necrosis factor-a97–99.
Notably, the immune system can regulate the nature of latency in cells
infected with herpesviruses100, and thus the nature of latency is not
necessarily cell intrinsic—a potentially important clue that needs to be
exploited. Latency in macrophages or other cells that may have a role
in persistent infection is less well understood than in T cells. Therefore,
it is not known whether vaccine-induced effector mechanisms should
target HIV in macrophages or dendritic cells, in addition to T cells. A
further profound advantage of defining immune and cell-intrinsic
control of latency in more detail could be the development of
©2010 Macmillan Publishers Limited. All rights reserved
PERSPECTIVES
additional animal models in which mice are adapted to HIV98, or
HIV is adapted to additional primates99.
Tough choices ahead
There is hope for HIV vaccination. There are many knowable
unknowns (Figs 1 and 3) that can guide this effort, and there is much
scientific knowledge yet to be applied. However, there are also major
scientific challenges that must be met to optimize chances of generating
an effective vaccine. Although a successful vaccine might be around the
corner in the next human trial, instead the path to an effective vaccine
might be many years long and require novel discoveries and insights
into the functioning of the immune system. Thus, the scientific issues
addressed in this Perspective are best considered in the broader context
of how to structure the research effort to ensure success. Because many
of the basic immunological issues that need to be solved to conquer
AIDS are similar to those for hepatitis C, hepatitis B, malaria, tuber-
culosis and cancer, enhancing basic understanding of HIV immunity
will have broad impacts. Calls for a greater focus on basic science to
facilitate HIV vaccine research were expressed 16 years ago101. We, and
many others102, make the same call today, as the kind of effort com-
mensurate with the problem is far from being realized.
So what are the barriers, in addition to the complexity of the
immunology, to rapid progress in basic science relevant to HIV vac-
cination? Most important among these barriers, and most controversial
because it has important practical, philosophical, political, scientific
and financial implications, is the separation of most HIV-related
research from research in other areas of immunology and virology.
Considering the many issues that need to be solved to give us the best
long-term chance of effective HIV vaccination (for example, Figs 1 and
3), it becomes clear that not all of the answers can efficiently come from
studies of HIV. The history of research on HIV immunology is filled
with paradigm shifts based on studies of viruses and species commonly
defined as not AIDS related. Examples include T-cell exhaustion, Treg
cells, PD1, CTLA4, inhibitory cytokines, and many others. Many were
first defined in mice by studying LCMV, a virus unrelated to HIV, and
many have taken years to translate from murine to human studies.
Studies of other human pathogens are also highly relevant. For
example, systems-based studies of human vaccines that do work, such
as the yellow fever virus vaccine, provide new genes and pathways,
some entirely unexpected, that may contribute to highly functional
CD8 T-cell and B-cell vaccine responses49. Studies of cytomegalovirus
led to the development of an HIV vaccine approach effective against
mucosal challenge9. There is no debate that HIV vaccine research has
benefitted significantly from research in other areas, and vice versa. The
question is how best to move forward to realize the great potential
offered by integrating currently isolated efforts.
We need, as a scientific community, to break down barriers
between the AIDS effort and other relevant areas of science to access
previously untapped human and technological resources for HIV
vaccination. But this needs to be done in context, such that those
with expertise from other fields are able to effectively tap into the
wealth of current knowledge on HIV pathogenesis. It is not enough to
simply funnel more funding to current AIDS research efforts. The key
point here is that there is no lack of enthusiasm among scientists
outside the HIV field to get involved in HIV/AIDS related research;
rather there is merely a lack of a way in. The study section system and
foundation efforts should be reformed with this principle in mind.
Studies in multiple systems, especially those that define novel
mechanisms and pathways related to chronic viral infection, should
be linked to ongoing HIV efforts and pursued with urgency propor-
tionate to the severity of the global AIDS crisis and under the AIDS
umbrella.
We propose that AIDS-related research be redefined to promote
the integrated study of all mechanisms that may facilitate under-
standing HIV pathogenesis and immunity. The HIV/AIDS problem
is too severe to allow the relevant science to be restricted by policies
that separate disciplines that could have a real impact on this deadly
229
PERSPECTIVES
human pathogen. Although we must, of course, study vaccine can-
didates in humans, we must also effectively engage scientists who can
bring new perspectives to combating HIV/AIDS. A long-term plan to
enhance the basic science relevant to HIV pathogenesis and immuno-
logy, broadly defined, is as essential today as it was 16 years ago101.
Targeted funding, even in small amounts, specifically designed to
integrate currently isolated but potentially transforming efforts,
may be the best possible investment that can be made at this juncture.
We must finally apply the full power of modern science to the AIDS
vaccine efforts, and by defining the knowable unknowns translate this
knowledge to vaccine-mediated protection against a pathogen that
has already caused over 30 million deaths and shows no sign of
relenting.
1. Rerks-Ngarm, S. et al. Vaccination with ALVAC and AIDSVAX to prevent HIV-1
infection in Thailand. N. Engl. J. Med. 361, 2209–2220 (2009).
The first human HIV vaccine trial to show a decrease in HIV acquisition after
immunization, which consisted of a combination of four priming doses of a
recombinant canarypox vector plus two booster injections with recombinant
HIV gp120.
2. Buchbinder, S. P. et al. Efficacy assessment of a cell-mediated immunity HIV-1
vaccine (the Step Study): a double-blind, randomised, placebo-controlled, test-
of-concept trial. Lancet 372, 1881–1893 (2008).
3. McElrath, M. J. et al. HIV-1 vaccine-induced immunity in the test-of-concept Step
Study: a case-cohort analysis. Lancet 372, 1894–1905 (2008).
4. Gilbert, P. B. et al. Correlation between immunologic responses to a recombinant
glycoprotein 120 vaccine and incidence of HIV-1 infection in a phase 3 HIV-1
preventive vaccine trial. J. Infect. Dis. 191, 666–677 (2005).
5. Altfeld, M. et al. HIV-1 superinfection despite broad CD81 T-cell responses
containing replication of the primary virus. Nature 420, 434–439 (2002).
6. Haase, A. T. Targeting early infection to prevent HIV-1 mucosal transmission.
Nature doi:10.1038/nature08757 (this issue).
7. Liu, J. et al. Immune control of an SIV challenge by a T-cell-based vaccine in rhesus
monkeys. Nature 457, 87–91 (2009).
8. Wilson, N. A. et al. Vaccine-induced cellular responses control simian
immunodeficiency virus replication after heterologous challenge. J. Virol. 83,
6508–6521 (2009).
9. Hansen, S. G. et al. Effector memory T cell responses are associated with
protection of rhesus monkeys from mucosal simian immunodeficiency virus
challenge. Nature Med. 15, 293–299 (2009).
Persistent effector memory CD4 and CD8 T-cell responses were induced in
monkeys by CMV vectors expressing SIV Gag, Rev-Tat-Nef and Env, and after
repeated low-dose intrarectal challenge with SIV, 4 out of 12 immunized
monkeys were protected from systemic spread of infection.
10. Cohen, M. S., Hellmann, N., Levy, J. A., DeCock, K. & Lange, J. The spread,
treatment, and prevention of HIV-1: evolution of a global pandemic. J. Clin. Invest.
118, 1244–1254 (2008).
11. Kroner, B. L., Rosenberg, P. S., Aledort, L. M., Alvord, W. G. & Goedert, J. J. HIV-1
infection incidence among persons with hemophilia in the United States and
western Europe, 1978–1990. Multicenter Hemophilia Cohort Study. J. Acquir.
Immune Defic. Syndr. 7, 279–286 (1994).
12. Fowke, K. R. et al. Resistance to HIV-1 infection among persistently seronegative
prostitutes in Nairobi, Kenya. Lancet 348, 1347–1351 (1996).
13. Salazar-Gonzalez, J. F. et al. Genetic identity, biological phenotype, and
evolutionary pathways of transmitted/founder viruses in acute and early HIV-1
infection. J. Exp. Med. 206, 1273–1289 (2009).
14. Deeks, S. G. & Walker, B. D. Human immunodeficiency virus controllers:
mechanisms of durable virus control in the absence of antiretroviral therapy.
Immunity 27, 406–416 (2007).
15. Duvall, M. G. et al. Maintenance of HIV-specific CD41 T cell help distinguishes
HIV-2 from HIV-1 infection. J. Immunol. 176, 6973–6981 (2006).
16. Sodora, D. L. et al. Toward an AIDS vaccine: lessons from natural simian
immunodeficiency virus infections of African nonhuman primate hosts. Nature
Med. 15, 861–865 (2009).
17. Martin, M. P. et al. Innate partnership of HLA-B and KIR3DL1 subtypes against
HIV-1. Nature Genet. 39, 733–740 (2007).
18. Nakanishi, Y., Lu, B., Gerard, C. & Iwasaki, A. CD81 T lymphocyte mobilization to
virus-infected tissue requires CD41 T-cell help. Nature 462, 510–513 (2009).
In a murine HSV-2 infection model, it is shown that CD4 T-helper cells control
migration of CD8 T cells to the mucosa through secretion of IFN-c and induction
of local secretion of chemokines.
19. Douek, D. C. et al. HIV preferentially infects HIV-specific CD41 T cells. Nature
417, 95–98 (2002).
20. Klatt, N. R. et al. Availability of activated CD41 T cells dictates the level of viremia
in naturally SIV-infected sooty mangabeys. J. Clin. Invest. 118, 2039–2049
(2008).
21. Hasenkrug, K. J., Brooks, D. M. & Dittmer, U. Critical role for CD41 T cells in
controlling retrovirus replication and spread in persistently infected mice. J. Virol.
72, 6559–6564 (1998).
230
©2010 Macmillan Publishers Limited. All rights reserved
NATUREjVol 464j11 March 2010
22. Virgin, H. W., Wherry, E. J. & Ahmed, R. Redefining chronic viral infection. Cell 138,
30–50 (2009).
23. Matloubian, M., Concepcion, R. J. & Ahmed, R. Cd41 T-cells are required to
sustain Cd81 cytotoxic T-cell responses during chronic viral-infection. J. Virol. 68,
8056–8063 (1994).
24. Pike, R. et al. Race between retroviral spread and CD41 T-cell response determines
the outcome of acute friend virus infection. J. Virol. 83, 11211–11222 (2009).
25. Okoye, A. et al. Profound CD41/CCR51 T cell expansion is induced by CD81
lymphocyte depletion but does not account for accelerated SIV pathogenesis. J.
Exp. Med. 206, 1575–1588 (2009).
26. Scheid, J. F. et al. Broad diversity of neutralizing antibodies isolated from memory
B cells in HIV-infected individuals. Nature 458, 636–640 (2009).
27. Hübner, W. et al. Quantitative 3D video microscopy of HIV transfer across T cell
virological synapses. Science 323, 1743–1747 (2009).
28. Walker, L. M. et al. Broad and potent neutralizing antibodies from an African
donor reveal a new HIV-1 vaccine target. Science 326, 285–289 (2009).
Using a high-throughput screening method, two potent and broadly cross-
reactive neutralizing antibodies were isolated that target a novel broadly
neutralizing epitope expressed on trimeric envelope protein and spanning
conserved regions of variable loops of the gp120 subunit.
29. Derdeyn, C. A. et al. Envelope-constrained neutralization-sensitive HIV-1 after
heterosexual transmission. Science 303, 2019–2022 (2004).
30. Chen, L. et al. Structural basis of immune evasion at the site of CD4 attachment on
HIV-1 gp120. Science 326, 1123–1127 (2009).
31. Hessell, A. J. et al. Effective, low-titer antibody protection against low-dose
repeated mucosal SHIV challenge in macaques. Nature Med. 15, 951–954 (2009).
This study shows that systemic antibody can protect against mucosal challenge
with laboratory-derived viral stocks, indicating that induction of systemic
humoral immunity might inhibit HIV spread.
32. Hessell, A. J. et al. Broadly neutralizing human anti-HIV antibody 2G12 is effective
in protection against mucosal SHIV challenge even at low serum neutralizing
titers. PLoS Pathog. 5, e1000433 (2009).
33. Johnson, P. R. et al. Vector-mediated gene transfer engenders long-lived
neutralizing activity and protection against SIV infection in monkeys. Nature Med.
15, 901–906 (2009).
34. Nelson, S. et al. Maturation of West Nile virus modulates sensitivity to antibody-
mediated neutralization. PLoS Pathog. 4, e1000060 (2008).
This study shows that differences in the form of viral glycoproteins on the
surface of an enveloped virus that occur during viral morphogenesis can be a
critical determinant of the extent of virus neutralization.
35. Kwong, P. D. & Wilson, I. A. HIV-1 and influenza antibodies: seeing antigens in new
ways. Nature Immunol. 10, 573–578 (2009).
36. Hessell, A. J. et al. Fc receptor but not complement binding is important in
antibody protection against HIV. Nature 449, 101–104 (2007).
This study calls attention to the fact that in vivo effects of vaccine-generated
antibody may depend on interactions between antibody molecules and Fc receptors.
37. Messer, R. J., Dittmer, U., Peterson, K. E. & Hasenkrug, K. J. Essential role for virus-
neutralizing antibodies in sterilizing immunity against Friend retrovirus infection.
Proc. Natl Acad. Sci. USA 101, 12260–12265 (2004).
38. Tyler, K. L., Mann, M. A., Fields, B. N. & Virgin, H. W. Protective anti-reovirus
monoclonal antibodies and their effects on viral pathogenesis. J. Virol. 67,
3446–3453 (1993).
39. Mehlhop, E. et al. Complement activation is required for induction of a protective
antibody response against West Nile virus infection. J. Virol. 79, 7466–7477 (2005).
40. McClellan, K. B., Gangappa, S., Speck, S. H. & Virgin, H. W. Antibody-independent
control of gamma-herpesvirus latency via B cell induction of anti-viral T cell
responses. PLoS Pathog. 2, e58 (2006).
41. Hasenkrug, K. J., Brooks, D. M. & Chesebro, B. Passive immunotherapy for
retroviral disease: influence of major histocompatibility complex type and T-cell
responsiveness. Proc. Natl Acad. Sci. USA 92, 10492–10495 (1995).
42. Kiepiela, P. et al. CD81 T-cell responses to different HIV proteins have discordant
associations with viral load. Nature Med. 13, 46–53 (2007).
43. Chung, C. et al. Not all cytokine-producing CD81 T cells suppress simian
immunodeficiency virus replication. J. Virol. 81, 1517–1523 (2007).
44. Chen, H. et al. Differential neutralization of human immunodeficiency virus (HIV)
replication in autologous CD4 T cells by HIV-specific cytotoxic T lymphocytes. J.
Virol. 83, 3138–3149 (2009).
This study shows that CD8 T cells vary widely in their ability to inhibit HIV
replication in an in vitro assay, indicating that much of the CD8 T-cell response in
chronic infection may be irrelevant.
45. Day, C. L. et al. PD-1 expression on HIV-specific T cells is associated with T-cell
exhaustion and disease progression. Nature 443, 350–354 (2006).
46. Petrovas, C. et al. PD-1 is a regulator of virus-specific CD81 T cell survival in HIV
infection. J. Exp. Med. 203, 2281–2292 (2006).
47. Trautmann, L. et al. Upregulation of PD-1 expression on HIV-specific CD81 T cells
leads to reversible immune dysfunction. Nature Med. 12, 1198–1202 (2006).
48. Kaufmann, D. E. et al. Upregulation of CTLA-4 by HIV-specific CD41 T cells
correlates with disease progression and defines a reversible immune dysfunction.
Nature Immunol. 8, 1246–1254 (2007).
49. Querec, T. D. et al. Systems biology approach predicts immunogenicity of the
yellow fever vaccine in humans. Nature Immunol. 10, 116–125 (2009).
NATUREjVol 464j11 March 2010
This study shows that a number of previously unknown pathways are associated with
effective immune responses induced by the Yellow Fever virus vaccine, potentially
altering how we define effective responses and measure correlates of immunity.
50. Moir, S. et al. Evidence for HIV-associated B cell exhaustion in a dysfunctional
memory B cell compartment in HIV-infected viremic individuals. J. Exp. Med. 205,
1797–1805 (2008).
51. Moir, S. & Fauci, A. S. B cells in HIV infection and disease. Nature Rev. Immunol. 9,
235–245 (2009).
52. Brooks, D. G., Lee, A. M., Elsaesser, H., McGavern, D. B. & Oldstone, M. B. IL-10
blockade facilitates DNA vaccine-induced T cell responses and enhances
clearance of persistent virus infection. J. Exp. Med. 205, 533–541 (2008).
53. Brooks, D. G. et al. IL-10 and PD-L1 operate through distinct pathways to suppress
T-cell activity during persistent viral infection. Proc. Natl Acad. Sci. USA 105,
20428–20433 (2008).
54. Tinoco, R., Alcalde, V., Yang, Y., Sauer, K. & Zuniga, E. I. Cell-intrinsic transforming
growth factor-b signaling mediates virus-specific CD81 T cell deletion and viral
persistence in vivo. Immunity 31, 145–157 (2009).
55. Fröhlich, A. et al. IL-21R on T cells is critical for sustained functionality and control
of chronic viral infection. Science 324, 1576–1580 (2009).
56. Elsaesser, H., Sauer, K. & Brooks, D. G. IL-21 is required to control chronic viral
infection. Science 324, 1569–1572 (2009).
57. Yi, J. S., Du, M. & Zajac, A. J. A vital role for interleukin-21 in the control of a chronic
viral infection. Science 324, 1572–1576 (2009).
58. Barber, D. L. et al. Restoring function in exhausted CD8 T cells during chronic viral
infection. Nature 439, 682–687 (2006).
59. Velu, V. et al. Enhancing SIV-specific immunity in vivo by PD-1 blockade. Nature
458, 206–210 (2008).
60. Ha, S. J. et al. Enhancing therapeutic vaccination by blocking PD-1-mediated
inhibitory signals during chronic infection. J. Exp. Med. 205, 543–555 (2008).
61. Araki, K. et al. mTOR regulates memory CD8 T-cell differentiation. Nature 460,
108–112 (2009).
Administration of the ‘immunosuppressive’ drug rapamycin is shown to enhance
memory T-cell responses, raising the possibility that combining pharmacological
and vaccine approaches may enhance immunity.
62. Lu, W., Arraes, L. C., Ferreira, W. T. & Andrieu, J. M. Therapeutic dendritic-cell
vaccine for chronic HIV-1 infection. Nature Med. 10, 1359–1365 (2004).
63. Lund, J. M., Hsing, L., Pham, T. T. & Rudensky, A. Y. Coordination of early
protective immunity to viral infection by regulatory T cells. Science 320,
1220–1224 (2008).
64. Zhu, J. et al. Virus-specific CD81 T cells accumulate near sensory nerve endings in
genital skin during subclinical HSV-2 reactivation. J. Exp. Med. 204, 595–603 (2007).
65. Gebhardt, T. et al. Memory T cells in nonlymphoid tissue that provide enhanced
local immunity during infection with herpes simplex virus. Nature Immunol. 10,
524–530 (2009).
This study suggests that there may be tissue-specific events that control the
long-term presence of immune lymphocytes in a given location.
66. Chachu, K. A., LoBue, A. D., Strong, D. W., Baric, R. S. & Virgin, H. W. Immune
mechanisms responsible for vaccination against and clearance of mucosal and
lymphatic norovirus infection. PLoS Pathog. 4, e1000236 (2008).
67. Paavonen, J. et al. Efficacy of human papillomavirus (HPV)-16/18 AS04-
adjuvanted vaccine against cervical infection and precancer caused by oncogenic
HPV types (PATRICIA): final analysis of a double-blind, randomised study in
young women. Lancet 374, 301–314 (2009).
68. Kaufman, D. R. et al. Trafficking of antigen-specific CD81 T lymphocytes to
mucosal surfaces following intramuscular vaccination. J. Immunol. 181,
4188–4198 (2008).
69. Barton, E. S. et al. Herpesvirus latency confers symbiotic protection from bacterial
infection. Nature 447, 326–329 (2007).
70. Wyand, M. S. et al. Protection by live, attenuated simian immunodeficiency virus
against heterologous challenge. J. Virol. 73, 8356–8363 (1999).
71. Brenchley, J. M. et al. Microbial translocation is a cause of systemic immune
activation in chronic HIV infection. Nature Med. 12, 1365–1371 (2006).
72. Slack, E. et al. Innate and adaptive immunity cooperate flexibly to maintain host-
microbiota mutualism. Science 325, 617–620 (2009).
73. Nix, W. A., Jiang, B., Maher, K., Strobert, E. & Oberste, M. S. Identification of
enteroviruses in naturally infected captive primates. J. Clin. Microbiol. 46,
2874–2878 (2008).
74. Finkbeiner, S. R. et al. Metagenomic analysis of human diarrhea: viral detection
and discovery. PLoS Pathog. 4, e1000011 (2008).
75. McKenna, P. et al. The macaque gut microbiome in health, lentiviral infection, and
chronic enterocolitis. PLoS Pathog. 4, e20 (2008).
76. Ivanov, I. I. et al. Induction of intestinal Th17 cells by segmented filamentous
bacteria. Cell 139, 485–498 (2009).
77. Korber, B. T., Letvin, N. L. & Haynes, B. F. T-cell vaccine strategies for human
immunodeficiency virus, the virus with a thousand faces. J. Virol. 83, 8300–8314
(2009).
78. Keele, B. F. et al. Identification and characterization of transmitted and early
founder virus envelopes in primary HIV-1 infection. Proc. Natl Acad. Sci. USA 105,
7552–7557 (2008).
©2010 Macmillan Publishers Limited. All rights reserved
PERSPECTIVES
79. Vignuzzi, M., Stone, J. K., Arnold, J. J., Cameron, C. E. & Andino, R. Quasispecies
diversity determines pathogenesis through cooperative interactions in a viral
population. Nature 439, 344–348 (2006).
This study argues that a viral quasispecies acts in vivo to confer pathogenetic
properties on a virus, indicating that viral quasispecies have roles in
pathogenesis above and beyond immune escape.
80. Draenert, R. et al. Constraints on HIV-1 evolution and immunodominance revealed
in monozygotic adult twins infected with the same virus. J. Exp. Med. 203,
529–539 (2006).
81. Moore, C. B. et al. Evidence of HIV-1 adaptation to HLA-restricted immune
responses at a population level. Science 296, 1439–1443 (2002).
82. Brumme, Z. L. et al. HLA-associated immune escape pathways in HIV-1 subtype B
Gag, Pol and Nef proteins. PLoS One 4, e6687 (2009).
83. Miura, T. et al. HLA-B57/B*5801 human immunodeficiency virus type 1 elite
controllers select for rare gag variants associated with reduced viral replication
capacity and strong cytotoxic T-lymphocyte recognition. J. Virol. 83, 2743–2755
(2009).
84. Troyer, R. M. et al. Variable fitness impact of HIV-1 escape mutations to cytotoxic
T lymphocyte (CTL) response. PLoS Pathog. 5, e1000365 (2009).
85. Schneidewind, A. et al. Structural and functional constraints limit options for cytotoxic
T-lymphocyte escape in the immunodominant HLA-B27-restricted epitope in human
immunodeficiency virus type 1 capsid. J. Virol. 82, 5594–5605 (2008).
86. Goepfert, P. A. et al. Transmission of HIV-1 Gag immune escape mutations is
associated with reduced viral load in linked recipients. J. Exp. Med. 205,
1009–1017 (2008).
87. Allen, T. M. et al. Selection, transmission, and reversion of an antigen-processing
cytotoxic T-lymphocyte escape mutation in human immunodeficiency virus type
1 infection. J. Virol. 78, 7069–7078 (2004).
88. Goonetilleke, N. et al. The first T cell response to transmitted/founder virus
contributes to the control of acute viremia in HIV-1 infection. J. Exp. Med. 206,
1253–1272 (2009).
89. Schneidewind, A. et al. Transmission and long-term stability of compensated CD8
escape mutations. J. Virol. 83, 3993–3997 (2009).
90. Fischer, W. et al. Polyvalent vaccines for optimal coverage of potential T-cell
epitopes in global HIV-1 variants. Nature Med. 13, 100–106 (2007).
91. Neil, S. & Bieniasz, P. HIV, restriction factors, and interferon. J Interferon. Cytokine.
Res. 29, 569–80 (2009).
92. Bushman, F. D. et al. Host cell factors in HIV replication: meta-analysis of genome-
wide studies. PLoS Pathog. 5, e1000437 (2009).
93. Santiago, M. L. et al. Apobec3 encodes Rfv3, a gene influencing neutralizing
antibody control of retrovirus infection. Science 321, 1343–1346 (2008).
94. Sheehy, A. M., Gaddis, N. C., Choi, J. D. & Malim, M. H. Isolation of a human gene
that inhibits HIV-1 infection and is suppressed by the viral Vif protein. Nature 418,
646–650 (2002).
95. Sheehy, A. M., Gaddis, N. C. & Malim, M. H. The antiretroviral enzyme
APOBEC3G is degraded by the proteasome in response to HIV-1 Vif. Nature Med.
9, 1404–1407 (2003).
96. Dinoso, J. B. et al. A simian immunodeficiency virus-infected macaque model to
study viral reservoirs that persist during highly active antiretroviral therapy. J.
Virol. 83, 9247–9257 (2009).
97. Blazkova, J. et al. CpG methylation controls reactivation of HIV from latency. PLoS
Pathog. 5, e1000554 (2009).
98. Zhang, J. X., Diehl, G. E. & Littman, D. R. Relief of preintegration inhibition and
characterization of additional blocks for HIV replication in primary mouse T cells.
PLoS One 3, e2035 (2008).
99. Hatziioannou, T. et al. A macaque model of HIV-1 infection. Proc. Natl Acad. Sci.
USA 106, 4425–4429 (2009).
100. Tibbetts, S. A., Van Dyk, L., Speck, S. H. & Virgin, H. W. Immune control of the
number and reactivation phenotype of cells latently infected with a gamma-
herpesvirus. J. Virol. 76, 7125–7132 (2002).
101. Fields, B. N. AIDS: time to turn to basic science. Nature 369, 95–96 (1994).
102. Fauci, A. S. et al. HIV vaccine research: the way forward. Science 321, 530–532 (2008).
Acknowledgements We would like to acknowledge, while accepting full
responsibility for the content and opinions expressed, open discussions and
correspondence with: M. Altfeld, D. Brooks, D. Burton, M. Colonna, L. Corey,
B. Beutler, D. Barouch, A. Chakraborty, F. Chisari, R. Desrosiers, M. Diamond,
A. Garcia-Sastre, L. Glimcher, P. Goulder, H. Greenberg, A. Haase, K. Hasenkrug,
B. Haynes, M. Hirsch, S. Abdool-Karim, W. Koff, B. Lamb, J. Lifson, J. Loh,
R. Medzhitov, J. Moore, M. Nussenzweig, M. Oldstone, L. Picker, B. Roizman,
R. Siliciano, G. Silvestri, S. Wain-Hobson, D. Watkins, R. Weiss, J. Wherry and
O. Yang. This work was supported by grants from the National Institutes of Health
(H.W.V. and B.D.W.), the Bill and Melinda Gates Foundation (B.D.W.), and a gift
from the Mark and Lisa Schwartz Foundation (B.D.W.).
Author Contributions H.W.V. and B.D.W. contributed equally to all aspects of the
planning and writing of this review.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence should be addressed to H.W.V. (virgin@wustl.edu) or B.D.W.
(bwalker@partners.org).
231
 Vol 464 | 11 March 2010 | doi:10.1038/nature08784

ARTICLES
Retroviral intasome assembly and
inhibition of DNA strand transfer
Stephen Hare1*, Saumya Shree Gupta1*{, Eugene Valkov1{, Alan Engelman2 & Peter Cherepanov1

Integrase is an essential retroviral enzyme that binds both termini of linear viral DNA and inserts them into a host cell
chromosome. The structure of full-length retroviral integrase, either separately or in complex with DNA, has been lacking.
Furthermore, although clinically useful inhibitors of HIV integrase have been developed, their mechanism of action remains
speculative. Here we present a crystal structure of full-length integrase from the prototype foamy virus in complex with its
cognate DNA. The structure shows the organization of the retroviral intasome comprising an integrase tetramer tightly
associated with a pair of viral DNA ends. All three canonical integrase structural domains are involved in extensive
protein–DNA and protein–protein interactions. The binding of strand-transfer inhibitors displaces the reactive viral DNA
end from the active site, disarming the viral nucleoprotein complex. Our findings define the structural basis of retroviral DNA
integration, and will allow modelling of the HIV-1 intasome to aid in the development of antiretroviral drugs.

Retroviral integrase (IN) recognizes and acts on the termini of the
linear double-stranded DNA molecule that is produced by reverse
transcription1,2. Initially, in a reaction termed 39-processing, IN
removes two or three nucleotides from one or both viral DNA ends
to expose the 39 hydroxyl groups of the invariant CA dinucleotides.
Next, after import of the viral DNA into the nucleus, IN inserts both
39 ends of the viral DNA into opposing strands of cellular genomic
DNA. Mechanistically and structurally, IN belongs to a diverse family
of polynucleotidyl transferases3, which notably includes RNaseH4
and the transposases from Tn5 (ref. 5) and eukaryotic mobile ele-
ment Mos1 (ref. 6) (reviewed in refs 7, 8). The reactions catalysed by
these enzymes proceed by SN2-type nucleophilic substitution,
assisted by divalent metal cofactors4,9. In retroviral IN, a pair of
divalent metal cations (Mg21 or Mn21) are thought to be coordi-
nated by three carboxylates of the invariant D,D-35-E motif within
the catalytic core domain (CCD). To function, IN also requires its
amino-terminal domain (NTD), a three-helical bundle stabilized by
binding a Zn atom, and a carboxy-terminal domain (CTD) that
adopts an SH3-like fold10,11. In vivo, IN acts within a large nucleo-
protein complex that contains viral DNA and several virus- and host-
cell-derived components. The minimal functional complex involving
viral DNA and IN, referred to here as the intasome, can be assembled
in vitro from purified components12.
Despite its acute importance for antiretroviral drug discovery and
decades of rigorous research7,13, the complete structure of IN, either as a
separate protein or in the context of the functional intasome, is lacking.
Accordingly, the structural organization of the enzyme active site, which
is believed to adopt its functional state only after viral DNA binding, is
unknown. Because clinically useful HIV-1 IN strand-transfer inhibi-
tors14,15 (InSTIs) preferentially bind to and inhibit the intasome com-
plex as compared to free IN16, the mechanism of drug action is poorly
understood. We have now obtained diffracting crystals of the full-length
IN from the prototype foamy virus (PFV) in complex with its cognate
viral DNA. The availability of these crystals enabled us to determine
the long-sought structure of the retroviral intasome and explain the
mechanism of strand-transfer inhibitor action.
1
Division of Medicine, Imperial College London, St-Mary’s Campus, Norfolk Place, London W2 1PG, UK. 2Department of Cancer Immunology and AIDS, Dana-Farber Cancer Institute,
44 Binney Street, Boston, Massachusetts 02115, USA. {Present addresses: Hannover Biomedical Research School, D-30625 Hannover, Germany (S.S.G.); School of Chemistry and
Molecular Biosciences, University of Queensland, Brisbane, Queensland 4072, Australia (E.V.).
*These authors contributed equally to this work.
Crystallization of the PFV intasome
Most characterized INs predominantly promote the insertion of one
viral DNA end into one strand of a target DNA duplex in vitro. In
contrast, we recently reported that recombinant PFV IN catalyses
efficient concerted insertion of two PFV DNA ends into target
DNA17. Here, we obtained soluble and fully functional PFV intasome
preparations using recombinant PFV IN and double-stranded oligo-
nucleotides that mimic the viral DNA ends (Supplementary Fig. 1).
To bypass the initial catalytic step, 39-processing, the IN–DNA com-
plexes were assembled using ‘pre-processed’ oligonucleotides with
recessed termini, which model the viral DNA ends before their inser-
tion into host chromosomal DNA. The IN–DNA complexes were
remarkably stable and did not dissociate or lose activity, even after
prolonged incubations under high ionic strength conditions (Sup-
plementary Fig. 1 and data not shown).
After extensive crystallization trials we identified a crystal form of
full-length wild-type PFV IN in complex with a 19-base-pair (bp)
donor DNA that diffracted X-rays to 2.9 Å resolution, enabling us to
determine the three-dimensional structure of the intasome (Sup-
plementary Table 1). The asymmetric unit contained a single IN
dimer with a tightly associated viral DNA molecule, and a pair of
symmetry-related IN dimers formed an oblong tetramer (130 3 65 3
70 Å; Fig. 1a and Supplementary Movie 1). The IN dimer–dimer
interface is stabilized by intermolecular NTD–CCD interactions, as
previously observed in partial structures of lentiviral INs18,19. The
overall shape of the tetramer is reminiscent of the low-resolution
envelope obtained by a recent negative-stain electron microscopy
study of HIV-1 IN complexes20. Nonetheless, its architecture is markedly
different from all previously reported models.
Overall architecture of the PFV intasome
The inner subunits of the tetramer (coloured green and blue in Fig. 1)
are responsible for all contacts involved in tetramerization and viral
DNA binding. The CCDs of the outer subunits (gold in Fig. 1) seem
to provide supporting function, their NTDs and CTDs not resolved

232
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
90º
a b
CCD
CCD NTD
CTD
CTD
NTD CCD
CCD
Figure 1 | Architecture of the PFV intasome. a, Views along (left) and
perpendicular to (right) the crystallographic two-fold axis. The inner
subunits of the IN tetramer, engaged with viral DNA, are blue and green;
outer IN chains are yellow. The reactive and non-transferred DNA strands
are magenta and orange, respectively. Side chains of Asp 128, Asp 185 and
in the electron density maps. Previous retroviral IN CCD structures
showed a conserved dimeric interface7, and this interface is retained
between the inner and outer IN subunits of the intasome. Partial
structures of INs from HIV-1 (ref. 21), simian immunodeficiency22
and Rous sarcoma23 viruses showed considerable variability of the
CCD–CTD linker region. Within the PFV intasome, the CCD–CTD
linker adopts an extended conformation for most of its length, track-
ing parallel to the NTD–CCD linker from the same subunit (Fig. 1b).
The interdomain linkers truss both halves of the intasome together,
and the structure is further stabilized by a pair of CTDs interacting
with both inner CCDs (Fig. 1a), in addition to an extensive network
of protein–DNA interactions (see later). Thus, in contrast to previous
IN tetramer models based on two domain structures in which the
dimer–dimer interface appeared highly flexible18,19, the overall con-
formation of the assembled intasome is well constrained. Homology
modelling suggests that the notably shorter interdomain linkers in
HIV-1 IN (Supplementary Fig. 2) can extend sufficiently to allow a
similar overall architecture and topology in the HIV-1 intasome
(data not shown). A further small domain, which we refer to as the
NTD extension domain (NED), precedes the PFV IN NTD (Fig. 1b).
On the basis of amino acid sequence comparisons and secondary
structure predictions, NEDs are present in other spumaviral and
possibly gammaretroviral INs (data not shown).
IN–DNA interactions within the PFV intasome
In total, almost 10,000 Å2 of molecular surface is buried within IN–
DNA interfaces of the intasome. The protein–DNA interactions
involve amino acid residues from each domain of the inner IN sub-
units, their interdomain linkers, and 17 nucleotides from each viral
DNA end (Supplementary Figs 2 and 3). Thus, as observed in the case
of Tn5 transposase5, the canonical retroviral IN domains (NTD, CCD
and CTD) do not have discrete functions; each contributes to extensive
protein–protein and protein–DNA contacts within the functional
complex. The most intimate protein–DNA interactions are found
within the terminal six nucleotides, where the viral DNA deviates
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
CCD
Linkers:
CCD–CTD
NTD–CCD
CTD
NTD
NED
Glu 221 active-site residues are red sticks; Zn atoms are grey spheres.
Locations of the canonical IN domains (NTD, CCD and CTD) are indicated.
b, Inner (green) IN chain with domains and linkers indicated. The
orientation is the same as in the right panel of a.
considerably from the ideal B form. Each CTD makes contact with
the phosphodiester backbone of both viral DNA molecules, essentially
crosslinking the structure. Notably, the NED and NTD of each cata-
lytic subunit interact with the viral DNA molecule engaged at the active
site of the opposing CCD (Fig. 1a). The NED interacts with the phos-
phodiester backbone, whereas the other elements (NTD, CTD, CCD,
NTD–CCD and CCD–CTD linkers) further contribute to interactions
with DNA bases. These sequence-specific interactions include the
main-chain carbonyl group of Gly 218, which forms a hydrogen bond
with guanine 4 of the non-transferred strand (Fig. 2a). The protein–
DNA interactions extend into the beginning of the CCD a4 helix,
which packs into the minor groove at the end of the viral DNA duplex
(Fig. 2a and Supplementary Fig. 3). The side chain of Arg 222 is
involved in hydrogen bonding with T5 and C6 bases of the non-
transferred strand (Fig. 2a). Another set of sequence-specific contacts
involves residues from the NTD–CCD and CCD–CTD linkers extend-
ing from the opposing IN dimer. The side chain of Arg 313 intercalates
the minor groove of viral DNA, stacking its guanidinium group against
adenosine 12 of the reactive strand and forming a hydrogen bond with
cytosine 11 (Fig. 2b). Nearby, the side chain of Asn 106 interacts with
thymine 8 of the non-transferred strand. These interactions cause
notable widening of the minor groove of the viral DNA (Fig. 2b and
Supplementary Fig. 3b). The overall extent of the protein–DNA inter-
action agrees well with the ,16 bp IN footprint in functional HIV-1
intasomes12.
At the reactive viral DNA terminus, the base pair involving the
cytosine of the invariant CA dinucleotide (C16 and A17 of the reactive
strand; Supplementary Fig. 3a) is distorted by a buckle of ,30u,
whereas the terminal adenine is completely isolated from its com-
plement on the non-transferred strand (Fig. 2a). The active-site loop
(PFV IN residues 211–220) is directly involved in separating the viral
DNA strands, acting as a plough with highly conserved residues
Pro 214 and Gln 215 (corresponding to HIV-1 Pro 145 and Gln 146,
see Supplementary Fig. 2 for PFV/HIV-1 structure-based alignment)
forming its share. In particular, Gln 215 displaces thymine 3 of the
233
ARTICLES
a T2
D128
D185
P214
A17
T3 E221
Q215
G218 C16
G4
A15
T5
α4
R222
C6 G14
b the viral DNA, are located deep within the dimer–dimer interface.
N106
C11 R313 G9
C10
G10
R69
T9
A11
A12
T8
NTD
Figure 2 | Sequence-specific protein–DNA interactions. a, View of the IN
active site with bound DNA. Amino-acid side chains discussed in text, DNA
bases and the main-chain carbonyl of Gly 218 are shown as sticks.
b, Interactions of the same DNA molecule with the NTD and NTD–CCD and
CCD–CTD linkers from the opposing (blue) IN subunit. Note the widening
of the minor groove owing to insertion of Arg 313 and Asn 106 side chains.
The Arg 69 side chain packs into the major groove, forming hydrogen bonds
with guanine 10 of the non-transferred strand.
non-transferred strand, which turns away from the interior of the
DNA duplex (Fig. 2a and Supplementary Fig. 3b). The 59-overhang
of the non-transferred strand is threaded between the CCD and inter-
domain linkers, where it forms extensive contacts with the active-site
loop and the CTD from the same IN chain (Supplementary Fig. 3c).
The involvement of the CCD a4 helix and the active-site loop in
intimate interactions with the viral DNA end agrees well with results
of chemical and photo crosslinking of functional HIV-1 IN–DNA
complexes24–26.
The active site of the functional intasome
The reactive 39 termini of the donor DNA molecules are positioned
within close proximity of the Asp 128, Asp 185 and Glu 221 active-
site carboxylates (Fig. 2a). Although the crystals could be grown in
the presence of MgCl2, which considerably improved their diffrac-
tion limit (Supplementary Table 1), data resolution did not allow
unambiguous visualization of Mg21 cations in the active site.
Fortuitously, similar to other retroviral INs1, PFV IN can efficiently
use Mn21, a more electron-rich element, as a metal ion cofactor
(Supplementary Fig. 1d). A difference electron density map calculated
using diffraction data collected on crystals soaked in the presence of
MnCl2 revealed two strong positive peaks (9.4 and 12.4s) within the
active sites of the inner IN subunits. This result confirmed the
expected two-metal binding mode of retroviral INs and revealed the
positions of metal ion cofactors within the assembled active site, which
could be refined at full occupancy (Fig. 3a, see also Supplementary
Fig. 4a for metal atom omit and final electron density maps). On the
basis of the current model for two-metal active-site catalysis4, metal
atom B, coordinated by the carboxylates of Asp 128 and Glu 221, is in
place to activate the 39 hydroxyl group of the pre-processed viral DNA
for strand transfer, whereas metal A, bound by Asp 128 and Asp 185,
234
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
would be expected to destabilize the scissile phosphodiester group in
target DNA. Superposition of the Ca atoms of the active-site Asp and
Glu residues showed notable conservation between the metal and
DNA substrate binding modes of the Tn5 synaptic complex27 and
PFV intasome (Supplementary Fig. 5). Furthermore, the positions
of the metal ions are nearly identical to those of Cd21 and Zn21
cations observed in structures of the avian sarcoma virus IN CCD28.
Of note, soaking crystals in MgCl2 or MnCl2 did not change the
organization of the intasome active site (Supplementary Fig. 6).
Hence, the positioning of the 39 end of viral DNA is independent of
bound divalent metal ions.
Model for target DNA binding
The active sites of the inner IN subunits, engaged with the 39 termini of
Therefore, the only mode of host chromosomal DNA (target DNA)
binding that would not require marked rearrangement of the inta-
some or severe DNA bending is along the cleft between IN dimers
(Fig. 4). This target DNA binding mode could not have been predicted
based on previous partial IN structures, and starkly differs from what
we recently proposed18. Modelling B-form DNA within the cleft
results in near perfect alignment of the active sites with opposing
target DNA phosphodiester bonds separated by 4 bp, the known
spacing of concerted PFV integration17. It is easy to see how muta-
tions within the a2 helix of the CCD, described in ref. 29, would
prevent target DNA binding (Fig. 4). We tentatively speculate that
the NTDs and/or the CTDs of the outer IN subunits, disordered in our
structures, could be involved in target DNA capture30. However, this
target-binding model requires verification using mutagenesis or crys-
tallographic approaches.
Mechanism for strand-transfer inhibitor action
We have previously shown that PFV IN is sensitive to HIV-1 InSTIs17.
These compounds are thought to engage metal ion cofactors in the
IN active site by interactions with uniquely positioned oxygen atoms
of the pharmacophore31. The role of the remaining common InSTI
feature, a fluorobenzyl group, is enigmatic. Intasome structure
refinement using diffraction data collected on crystals soaked in
the presence of Mg21 and the clinical InSTI MK0518 (also known
as raltegravir)14 or GS9137 (elvitegravir)15 showed strong further
electron density within the active sites of the inner IN subunits.
Structures of MK0518 or GS9137 with pairs of Mg atoms could be
easily fitted into the maps and refined to 2.85 and 3.15 Å resolution,
respectively (Fig. 3b, c, Supplementary Fig. 4d, f and Supplementary
Table 1). Moreover, soaking crystals in the presence of the drugs and
Mn21 produced similar results, with manganese atoms and drug
molecules refining at almost precisely the same positions
(Supplementary Fig. 4b, c, e, g). Concordantly, InSTIs inhibited both
the Mg21- and Mn21-dependent activities of the PFV intasome
(Supplementary Fig. 1d).
On the basis of the structures, the two InSTIs seem to have very
similar modes of binding and action, involving an induced fit mech-
anism. Their metal chelating oxygen atoms orient towards the metal
cofactors of the active site, whereas their halobenzyl groups fit within
a tight pocket created by displacement of the 39 adenosine (A17).
Within it, the drugs make intimate Van der Waals interactions with
the bases of the invariant CA dinucleotide, guanine 4 from the non-
transferred strand, and conserved residues Pro 214 and Gln 215
(Fig. 3b, c). In addition, the isopropyl and methyl-oxadiazole groups
of MK0518 are involved in hydrophobic and stacking interactions
with the side chains of Pro 214 and Tyr 212, respectively (Fig. 3b),
further stabilizing this drug in the active site. Through its quinolone
base and isopropyl group, GS9137 interacts with Pro 214 (Fig. 3c).
Crucially, this mode of drug binding results in displacement of the
reactive 39 viral DNA end from the active site (Fig. 3b, c), which can
only result in deactivation of the intasome. Thus, after binding of
MK0518, the reactive 39 hydroxyl group moves away from the active
NATURE | Vol 464 | 11 March 2010 ARTICLES

a b c

Q215 Q215 Q215

P214 Y212 A17 P214 Y212 P214 Y212
A17 G4 A17 G4
G4

3′OH
S217
S217 3′OH S217 3′OH
E221 E221 E221
B A D185 D185 D185
D128 D128 D128
N224 β1 N224 β1 N224 β1
α4 α4 α4

Y212 Y212 Y212

P214 P214 P214

3′OH MK0518 GS9137

3′OH
3′OH
B
A

Figure 3 | PFV IN active site in committed and drug-bound states. yellow, C; blue, N; red, O; grey, F; green, Cl. The complex is shown as a
a–c, Views without drug (a) and with MK0518 (b) or GS9137 (c) bound. solvent accessible surface in bottom panels, coloured by atoms (light grey, C;
Protein and DNA in top panels are cartoons, with A17, DNA bases and the red, O; blue, N). Grey spheres are Mn21 (a, labelled A and B) or Mg21
side chains of indicated amino acids as sticks. Drug atoms are coloured: (b, c) ions.

site by more than 6 Å, compared to its positions in the Mg21- or
Mn21-containing, or apo crystals.
Because the core contact points consisting of invariant nucleotide
bases and amino acid residues are conserved in HIV-1, the mode of
InSTI binding and action are unlikely to significantly differ. The
extensive contacts with the viral DNA end observed in our structures
determine why the InSTIs preferentially interact with and inhibit the
DNA-bound form of HIV-1 IN16. Moreover, the induced fit caused
by displacement of the 39 adenosine by the halobenzyl groups of these
compounds explains why the deletion of this base markedly increased
InSTI on- and off-rates for binding to HIV-1 IN–DNA complexes32.
α2
α2
90º
Furthermore, mutations of HIV-1 IN residue Tyr 143, which, on the
basis of our structure, is expected to interact with the methyl-
oxadiazole group of MK0518 (Fig. 3b and Supplementary Fig. 2),
are known to confer resistance to this drug33. Common InSTI resist-
ance pathways involve mutations of HIV-1 IN Gln 148 or Asn 155
(ref. 33), which correspond to PFV IN residues Ser 217 and Asn 224,
respectively (Supplementary Fig. 2). Mutations at these positions are
likely to interfere with coordination of metal cofactors by the active-
site carboxylates, as proposed recently34. Conceivably, a slight shift in
metal ion cofactor positions might suffice to abrogate drug binding,
which relies on its spatially constrained metal chelating groups, albeit
at a steep price of impaired viral replication fitness due to detuning of
the IN active site structure.
Our findings will allow the generation of reliable HIV-1 IN and
InSTI pharmacophore models, which will be invaluable for the
development of next-generation strand-transfer inhibitors. Their
design should take advantage of the most conserved elements of
the IN active site determined here, such as the bases of the invariant
CA dinucleotide, positions of the metal co-factors and the main-
chain atoms of the protein.

α2
Figure 4 | Predicted target DNA binding orientation. Cartoon
representation of the intasome as in Fig. 1a with active-site side chains shown
as red sticks and a2 helices, known to contribute to target DNA binding29, in
cyan. The DNA molecule modelled in black and grey shows the most likely
orientation for target DNA binding. Note that the CTD, juxtaposed to the
target DNA in this model, is known to possess sequence non-specific DNA
binding activity35.
©2010 Macmillan Publishers Limited. All rights reserved
METHODS SUMMARY
Protein–DNA complexes were assembled using full-length wild-type PFV IN17
and synthetic double-stranded DNA that modelled the viral U5 end
(59-TACAAAATTCCATGACA-39/59-ATTGTCATGGAATTTTGTA-39) by dia-
lysis, reducing the NaCl concentration from 500 to 200 mM. A typical result of
complex assembly is shown in Supplementary Fig. 1a. The intasome was crystal-
lized by vapour diffusion in hanging drops with a reservoir solution containing
1.35 M ammonium sulphate, 25% (v/v) glycerol, 4.8% (v/v) 1,6-hexanediol and
50 mM 2-(N-morpholino) ethanesulphonic acid (MES), pH 6.5. The crystals
were soaked in the presence of MK0518, GS9137, Mg21 and/or Mn21, as
required. Diffraction data were obtained at the beamlines I02 and I04 of the
Diamond Light Source (Oxford, UK), and the structure was solved by molecular
replacement. Crystallographic and refinement statistics for the seven resulting
models are summarized in Supplementary Table 1. The structures, refined to
235
ARTICLES
2.85–3.25 Å resolution, included PFV IN residues 10–374 and 116–278 in chains
A and B, respectively, without gaps, and both complete strands of the pre-
processed donor DNA molecule (17 and 19 nucleotides in chains C and D,
respectively). The models had good geometry with 92.2–96.4 and not more than
0.4% of amino acid residues in most preferred and disallowed regions of the
Ramachandran plot, respectively.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 10 November 2009; accepted 6 January 2010.
Published online 31 January 2010.
1. Craigie, R. in Mobile DNA II (eds Craig, N. L., Craigie, R., Gellert, M. & Lambowitz,
A. M.) 613–630 (ASM Press, 2002).
2. Lewinski, M. K. & Bushman, F. D. Retroviral DNA integration - mechanism and
consequences. Adv. Genet. 55, 147–181 (2005).
3. Dyda, F. et al. Crystal structure of the catalytic domain of HIV-1 integrase:
similarity to other polynucleotidyl transferases. Science 266, 1981–1986 (1994).
4. Nowotny, M., Gaidamakov, S. A., Crouch, R. J. & Yang, W. Crystal structures of
RNase H bound to an RNA/DNA hybrid: substrate specificity and metal-
dependent catalysis. Cell 121, 1005–1016 (2005).
5. Davies, D. R., Goryshin, I. Y., Reznikoff, W. S. & Rayment, I. Three-dimensional
structure of the Tn5 synaptic complex transposition intermediate. Science 289,
77–85 (2000).
6. Richardson, J. M., Colloms, S. D., Finnegan, D. J. & Walkinshaw, M. D. Molecular
architecture of the Mos1 paired-end complex: the structural basis of DNA
transposition in a eukaryote. Cell 138, 1096–1108 (2009).
7. Jaskolski, M., Alexandratos, J. N., Bujacz, G. & Wlodawer, A. Piecing together the
structure of retroviral integrase, an important target in AIDS therapy. FEBS J. 276,
2926–2946 (2009).
8. Nowotny, M. Retroviral integrase superfamily: the structural perspective. EMBO
Rep. 10, 144–151 (2009).
9. Engelman, A., Mizuuchi, K. & Craigie, R. HIV-1 DNA integration: mechanism of
viral DNA cleavage and DNA strand transfer. Cell 67, 1211–1221 (1991).
10. Cai, M. et al. Solution structure of the N-terminal zinc binding domain of HIV-1
integrase. Nature Struct. Biol. 4, 567–577 (1997).
11. Eijkelenboom, A. P. et al. The DNA-binding domain of HIV-1 integrase has an SH3-
like fold. Nature Struct. Biol. 2, 807–810 (1995).
12. Li, M., Mizuuchi, M., Burke, T. R. Jr & Craigie, R. Retroviral DNA integration:
reaction pathway and critical intermediates. EMBO J. 25, 1295–1304 (2006).
13. Marchand, C., Maddali, K., Métifiot, M. & Pommier, Y. HIV-1 IN inhibitors: 2010
update and perspectives. Curr. Top. Med. Chem. 9, 1016–1037 (2009).
14. Summa, V. et al. Discovery of raltegravir, a potent, selective orally bioavailable
HIV-integrase inhibitor for the treatment of HIV-AIDS infection. J. Med. Chem. 51,
5843–5855 (2008).
15. Sato, M. et al. Novel HIV-1 integrase inhibitors derived from quinolone antibiotics.
J. Med. Chem. 49, 1506–1508 (2006).
16. Espeseth, A. S. et al. HIV-1 integrase inhibitors that compete with the target DNA
substrate define a unique strand transfer conformation for integrase. Proc. Natl
Acad. Sci. USA 97, 11244–11249 (2000).
17. Valkov, E. et al. Functional and structural characterization of the integrase from
the prototype foamy virus. Nucleic Acids Res. 37, 243–255 (2009).
18. Hare, S. et al. Structural basis for functional tetramerization of lentiviral integrase.
PLoS Pathog. 5, e1000515 (2009).
19. Wang, J. Y., Ling, H., Yang, W. & Craigie, R. Structure of a two-domain fragment of
HIV-1 integrase: implications for domain organization in the intact protein. EMBO
J. 20, 7333–7343 (2001).
20. Michel, F. et al. Structural basis for HIV-1 DNA integration in the human genome,
role of the LEDGF/P75 cofactor. EMBO J. 28, 980–991 (2009).
21. Chen, J. C. et al. Crystal structure of the HIV-1 integrase catalytic core and
C-terminal domains: a model for viral DNA binding. Proc. Natl Acad. Sci. USA 97,
8233–8238 (2000).
236
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
22. Chen, Z. et al. X-ray structure of simian immunodeficiency virus integrase
containing the core and C-terminal domain (residues 50–293)—an initial glance
of the viral DNA binding platform. J. Mol. Biol. 296, 521–533 (2000).
23. Yang, Z. N., Mueser, T. C., Bushman, F. D. & Hyde, C. C. Crystal structure of an
active two-domain derivative of Rous sarcoma virus integrase. J. Mol. Biol. 296,
535–548 (2000).
24. Alian, A. et al. Catalytically-active complex of HIV-1 integrase with a viral DNA
substrate binds anti-integrase drugs. Proc. Natl Acad. Sci. USA 106, 8192–8197
(2009).
25. Esposito, D. & Craigie, R. Sequence specificity of viral end DNA binding by HIV-1
integrase reveals critical regions for protein-DNA interaction. EMBO J. 17,
5832–5843 (1998).
26. Jenkins, T. M., Esposito, D., Engelman, A. & Craigie, R. Critical contacts between
HIV-1 integrase and viral DNA identified by structure-based analysis and photo-
crosslinking. EMBO J. 16, 6849–6859 (1997).
27. Steiniger-White, M., Rayment, I. & Reznikoff, W. S. Structure/function insights
into Tn5 transposition. Curr. Opin. Struct. Biol. 14, 50–57 (2004).
28. Bujacz, G. et al. Binding of different divalent cations to the active site of avian
sarcoma virus integrase and their effects on enzymatic activity. J. Biol. Chem. 272,
18161–18168 (1997).
29. Nowak, M. G., Sudol, M., Lee, N. E., Konsavage, W. M. Jr & Katzman, M. Identifying
amino acid residues that contribute to the cellular-DNA binding site on retroviral
integrase. Virology 389, 141–148 (2009).
30. Miller, M. D., Bor, Y. C. & Bushman, F. Target DNA capture by HIV-1 integration
complexes. Curr. Biol. 5, 1047–1056 (1995).
31. Grobler, J. A. et al. Diketo acid inhibitor mechanism and HIV-1 integrase:
implications for metal binding in the active site of phosphotransferase enzymes.
Proc. Natl Acad. Sci. USA 99, 6661–6666 (2002).
32. Langley, D. R. et al. The terminal (catalytic) adenosine of the HIV LTR controls the
kinetics of binding and dissociation of HIV integrase strand transfer inhibitors.
Biochemistry 47, 13481–13488 (2008).
33. Buzón, M. J. et al. The HIV-1 integrase genotype strongly predicts raltegravir
susceptibility but not viral fitness of primary virus isolates. AIDS 24, 17–25 (2010).
34. Grobler, J. A., Stillmock, K., Miller, M. D. & Hazuda, D. J. Mechanism by which the
HIV integrase active-site mutation N155H confers resistance to raltegravir.
Antivir. Ther. 13 (suppl. 3), A41 (2008).
35. Engelman, A., Hickman, A. B. & Craigie, R. The core and carboxyl-terminal
domains of the integrase protein of human immunodeficiency virus type 1 each
contribute to nonspecific DNA binding. J. Virol. 68, 5911–5917 (1994).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank F. Dyda for critical reading of the manuscript,
R. Clayton and M. Cummings for the generous gift of InSTIs and helpful
discussions, T. Sorensen and the staff of the I02 and I04 beamlines of the Diamond
Light Source for assistance with X-ray data collection. P.C. and co-workers are
funded by the UK Medical Research Council, and A.E. by the US National Institutes
of Health.
Author Contributions E.V. and P.C. carried out initial trials with truncated PFV IN
constructs; S.S.G. and P.C. obtained full-length PFV IN–DNA complexes, carried
out crystallization screening and optimization; S.H. soaked and prepared crystals
for data collection; S.H. and P.C. collected diffraction data and solved the
structures; S.H. refined the final models; S.H. and S.S.G. carried out gel-filtration
and activity assays; P.C., S.H. and A.E. wrote the paper.
Author Information Atomic coordinates and structure factors have been
deposited with the Protein Data Bank under accession codes 3L2Q (Apo), 3L2R
(Mg), 3L2S (Mn), 3L2T (Mg/MK0518), 3L2U (Mg/GS9137), 3L2V (Mn/MK0518)
and 3L2W (Mn/GS9137) structures. Raw diffraction images are available on
request. Reprints and permissions information is available at www.nature.com/
reprints. The authors declare no competing financial interests. Correspondence
and requests for materials should be addressed to P.C.
(p.cherepanov@imperial.ac.uk).
doi:10.1038/nature08784
METHODS
Preparation and analyses of the PFV intasome. Full-length PFV IN (corres-
ponding to residues 752–1143 of PFV POL) was produced in Escherichia coli
strain PC2 (ref. 36), transformed with pSSH6P-PFV-INFL17 and purified as
previously described17. The protein was stored in aliquots at 280 uC in 0.5 M
NaCl, 5 mM dithiothreitol (DTT), 10% glycerol, 50 mM Tris-HCl, pH 7.4. Ion-
exchange HPLC-purified oligonucleotides were purchased from Eurogentec.
Protein–DNA complexes were prepared by dialyses of mixtures containing
120 mM PFV IN, 50 mM synthetic DNA duplex, 500 mM NaCl, and 50 mM
BisTris propane-HCl, pH 7.45, against excess 200 mM NaCl, 2 mM DTT,
25 mM ZnCl2, 20 mM BisTris propane-HCl, pH 7.45, for 18–24 h at 18 uC.
Dialysed material was supplemented with an extra 120 or 800 mM NaCl (0.32
or 1 M NaCl final), incubated for 1 h on ice and analysed by size-exclusion
chromatography using a Superdex 200 HR 10/30 column, attached to an
ÄKTA Purifier system (GE Healthcare). The column was operated in 0.32 or
1 M NaCl supplemented with 20 mM BisTris propane-HCl, pH 7.45, at
1 ml min21, 20 uC. Strand-transfer assays with size-exclusion chromatography-
purified intasome were carried out using established buffer conditions17. A
typical reaction contained 300 ng supercoiled pGEM9 target DNA, 12 mA280 nm
units (,30 nM) intasome, 125 mM NaCl, 5 mM MgCl2 (or MnCl2), 10 mM
DTT, 4 mM ZnCl2, 25 mM BisTris propane-HCl, pH 7.45, in a final volume of
40 ml. The reaction conditions were modified as required. After incubation at
37 uC for 30–60 min, the products were deproteinized, separated in 1.5% (w/v)
agarose gels and visualized by staining with ethidium bromide, as previously
described17.
Crystallization and structure determination. More than 30 DNA constructs
were tested in initial crystallization trials with full-length wild-type and several
mutant PFV IN proteins in ,40,000 initial sparse matrix conditions. Although
several crystal forms could be identified and optimized, only one, obtained using
a 19-bp mimic of pre-processed U5 end of PFV DNA (59-TACAAAA
TTCCATGACA-39/59-ATTGTCATGGAATTTTGTA-39) and wild-type full-
length IN, diffracted X-rays to a resolution better than 6 Å. For crystallization,
the protein–DNA complex assembled by dialysis and supplemented with a
further 120 mM NaCl (320 mM NaCl final) was concentrated to 10–14 mg ml21
using Amicon Ultra-4 centrifugal devices (Millipore). Intasome preparations
were stable for extended periods of time when stored on ice. Crystals were grown
by vapour diffusion in hanging drops at 18 uC by mixing 1 ml IN–DNA complex
and 1 ml of reservoir solution containing 1.35 M ammonium sulphate, 25 mM
MgCl2, 25% (v/v) glycerol, 4.8% (v/v) 1,6-hexanediol, 50 mM MES-NaOH,
pH 6.5. MgCl2 was omitted to obtain apo crystals used in Mn21 and InSTI/
Mn21 soaking experiments. Crystals typically appeared within 24–48 h and
grew to a size of ,100 3 100 3 75 mm within 7–10 days. To obtain drug-bound
forms of the complex, intasome crystals were incubated in stabilizing solution
(reservoir plus 200 mM NaCl and 7 mM DTT), supplemented with 1 mM
MK0518 or GS9137 for 60 h at 18 uC. To visualize metal atoms in the active sites,
crystals grown without MgCl2 were soaked in the presence of 25 mM MnCl2 in
the absence or presence of 1 mM MK0518 or GS9137.
The crystals were frozen by rapid immersion in liquid nitrogen, and diffrac-
tion data, acquired on I02 and I04 beamlines of the Diamond Light Source
(Oxford, UK), were processed using Mosflm37 and SCALA38. Availability of
the PFV IN CCD structure17 and the high solvent content of the crystal form
(68.5%) made it possible to determine the intasome structure by molecular
©2010 Macmillan Publishers Limited. All rights reserved
replacement. The solution was found in Phaser39 in space group P41212 using
PFV IN CCD (Protein Data Bank (PDB) accession 3dlr)17, HIV-1 IN CTD
(residues 223–266 of chain A from PDB accession 1ex4)21 and a generic 7-bp
B-form DNA duplex as search models (Supplementary Fig. 7a). The initial phases,
modified with prime-and-switch mode of RESOLVE40 (Supplementary Fig. 7b)
were used as input for Buccaneer41, which built most of the protein part of the
structure, including the NTD, NTD–CCD and CCD–CTD linkers. The models
were improved by iterative manual building in Coot42, simulated annealing in
Phenix43 and maximum likelihood restrained positional refinement in Refmac44.
Non-crystallographic symmetry restraints for CCD regions similar in chains A
and B (residues 121–200, 241–257 and 263–273) were applied throughout.
Translation, liberation and screw (TLS) refinement45 was included in the final
stages. Geometry of the final structures was analysed using Molprobity46.
Examples of electron density and omit maps are given in Supplementary Figs 4
and 7. Supplementary Table 1 lists unit cell parameters and crystallographic
statistics. Coordinates and restraints files for the drug molecules were created
using the PRODRG2 server47. Figures were prepared using PyMOL (DeLano
Scientific) and ESPript48, molecular surface areas were calculated using
Areaimol38,49 and protein–DNA contacts were analysed with help of NUCPLOT50.
36. Cherepanov, P. LEDGF/p75 interacts with divergent lentiviral integrases and
modulates their enzymatic activity in vitro. Nucleic Acids Res. 35, 113–124 (2007).
37. Leslie, A. G. W. Recent changes to the MOSFLM package for processing film and
image plate data. Joint CCP4 and ESF-EAMCB Newsletter on Protein Crystallography
no. 26 (1992).
38. Collaborative Computational Project, Number 4. The CCP4 suite: programs for
protein crystallography. Acta Crystallogr. D 50, 760–763 (1994).
39. McCoy, A. J. et al. Phaser crystallographic software. J. Appl. Crystallogr. 40,
658–674 (2007).
40. Terwilliger, T. C. SOLVE and RESOLVE: automated structure solution and density
modification. Methods Enzymol. 374, 22–37 (2003).
41. Cowtan, K. The Buccaneer software for automated model building. 1. Tracing
protein chains. Acta Crystallogr. D 62, 1002–1011 (2006).
42. Emsley, P. & Cowtan, K. Coot: model-building tools for molecular graphics. Acta
Crystallogr. D 60, 2126–2132 (2004).
43. Zwart, P. H. et al. Automated structure solution with the PHENIX suite. Methods
Mol. Biol. 426, 419–435 (2008).
44. Murshudov, G. N., Vagin, A. A. & Dodson, E. J. Refinement of macromolecular
structures by the maximum-likelihood method. Acta Crystallogr. D 53, 240–255
(1997).
45. Winn, M. D., Isupov, M. N. & Murshudov, G. N. Use of TLS parameters to model
anisotropic displacements in macromolecular refinement. Acta Crystallogr. D 57,
122–133 (2001).
46. Davis, I. W. et al. MolProbity: all-atom contacts and structure validation for
proteins and nucleic acids. Nucleic Acids Res. 35, W375–W383 (2007).
47. Schüttelkopf, A. W. & van Aalten, D. M. PRODRG: a tool for high-throughput
crystallography of protein-ligand complexes. Acta Crystallogr. D 60, 1355–1363
(2004).
48. Gouet, P., Robert, X. & Courcelle, E. ESPript/ENDscript: Extracting and rendering
sequence and 3D information from atomic structures of proteins. Nucleic Acids
Res. 31, 3320–3323 (2003).
49. Lee, B. & Richards, F. M. The interpretation of protein structures: estimation of
static accessibility. J. Mol. Biol. 55, 379–400 (1971).
50. Luscombe, N. M., Laskowski, R. A. & Thornton, J. M. NUCPLOT: a program to
generate schematic diagrams of protein-nucleic acid interactions. Nucleic Acids
Res. 25, 4940–4945 (1997).
Vol 464 | 11 March 2010 | doi:10.1038/nature08852

ARTICLES
Somatic sex identity is cell autonomous in
the chicken
D. Zhao1*, D. McBride1*, S. Nandi1, H. A. McQueen3, M. J. McGrew1, P. M. Hocking2, P. D. Lewis4, H. M. Sang1
& M. Clinton1
In the mammalian model of sex determination, embryos are considered to be sexually indifferent until the transient action of
a sex-determining gene initiates gonadal differentiation. Although this model is thought to apply to all vertebrates, this has
yet to be established. Here we have examined three lateral gynandromorph chickens (a rare, naturally occurring
phenomenon in which one side of the animal appears male and the other female) to investigate the sex-determining
mechanism in birds. These studies demonstrated that gynandromorph birds are genuine male:female chimaeras, and
indicated that male and female avian somatic cells may have an inherent sex identity. To test this hypothesis, we
transplanted presumptive mesoderm between embryos of reciprocal sexes to generate embryos containing male:female
chimaeric gonads. In contrast to the outcome for mammalian mixed-sex chimaeras, in chicken mixed-sex chimaeras the
donor cells were excluded from the functional structures of the host gonad. In an example where female tissue was
transplanted into a male host, donor cells contributing to the developing testis retained a female identity and expressed a
marker of female function. Our study demonstrates that avian somatic cells possess an inherent sex identity and that, in
birds, sexual differentiation is substantively cell autonomous.

Sexual development in vertebrates is thought to be governed by
general principles defined in the early to mid-twentieth century1,2.
These principles state that the sexual phenotype of individuals is
dependent on the gonad: male and female somatic cells and tissues
are initially sexually indifferent and sexual dimorphism is imposed by
the type of gonad that develops. Although these principles have been
challenged, most notably by work on songbird neural development3–6
and marsupial development7,8, these observations are generally con-
sidered as exceptions. In the currently accepted model, gonadal differ-
entiation is triggered in sexually indifferent embryos by the transient
action of a sex-determining gene. In mammals, the sex-determining
gene is known to be the testis-determining Sry gene carried by the
male-specific Y chromosome9. Although all vertebrates are thought
to conform to this general model, with the exception of Sry in mam-
mals and Dmy in medaka10,11, no other vertebrate sex-determining
genes have been confirmed.
In terms of morphology, birds seem to conform to the mammalian
pattern: male and female embryos are sexually indistinguishable until
around days 5–6 of incubation (Hamburger and Hamilton12 (H&H)
stage 28/29) when the action of a sex-determining gene is thought to
initiate testis or ovary development13. However, in birds, not only is
the identity of the putative sex-determining gene unknown, the
nature of the sex-determining mechanism has not been established.
Current theories of sex determination in birds include the presence of
an ovary-determining gene on the female-specific W chromosome,
and a dosage mechanism based on the number of Z chromosomes
(females have one Z and one W sex chromosome whereas males have
two Z sex chromosomes)14. Currently, the best candidate for a testis-
determining gene in birds is DMRT1 (doublesex and mab-3-related
transcription factor 1). Expression of DMRT1 is restricted to the
gonads and it has recently been shown that repressing levels of
1
Division of Developmental Biology and 2Division of Genetics and Genomics, The Roslin Institute and Royal (Dick) School of Veterinary Studies, University of Edinburgh, Roslin,
Midlothian EH25 9PS, UK. 3Institute of Cell Biology, University of Edinburgh, West Mains Road, Edinburgh EH9 3JR, UK. 4Animal and Poultry Science Department, University of
KwaZulu-Natal, Pietermaritzburg, South Africa.
*These authors contributed equally to this work.
©2010 Macmillan Publishers Limited. All rights reserved
DMRT1 in male embryos has a ‘feminizing’ effect on the developing
testis15.
In an attempt to clarify the nature of the sex-determining mech-
anism in birds, we have investigated the composition of a number of
gynandromorph chickens. These birds are rare, naturally occurring
phenomena in which one side of the animal appears male and the
other female16. We investigated these birds with the expectation that
this condition resulted from a sex-chromosome aneuploidy on one
side of the bird, and that our analysis would provide evidence regard-
ing the nature of the avian sex-determining mechanism. Contrary to
expectations, our analysis established that the gynandromorphs were
in fact male:female chimaeras, and that the gynandromorphic pheno-
type was due to ZZ (male) and ZW (female) somatic cells responding
in different ways to the same profile of gonadal hormones. These
observations led to a series of transcriptome screens and embryonic
transplantation studies showing that male and female avian cells
possessed an inherent sex identity. Our studies demonstrate that in
chickens, gonadal development and the sexual phenotype are largely
cell autonomous and not principally dependent on sex hormones.
Gynandromorph birds are mixed-sex chimaeras
We obtained three adult lateral gynandromorph birds (designated
G1, G2 and G3) which we maintained and observed over a period of
24 months. These birds occur naturally and it has been suggested that
this condition results from the loss of a single sex chromosome at the
two-cell stage17. All three birds were ISA brown commercial hybrids
with sex-linked plumage colour. ISA brown males are heterozygous
for the dominant silver and recessive gold genes (Ss) and so have
white plumage; females possess only the gold gene (s-) and have
brown plumage. The birds displayed a marked bilateral asymmetry,
where one side of the animal appeared phenotypically female and the
237
ARTICLES
other side phenotypically male (Supplementary Fig. 1). Figure 1
shows a picture of G1 where the right side of the bird is female in
coloration (brown) and has a small wattle and small leg spur. In
contrast, the left side is male coloured (predominantly white), has
a large wattle and a large leg spur, a heavier leg structure and an
obviously greater mass of breast muscle, typical of a cockerel. Post
mortem, whole tissues from both sides were weighed and measured
and samples of all tissues were taken for later analysis. The measure-
ments performed on individual tissues from both sides of all three
animals supported the observation that these animals were, at least
phenotypically, half male and half female. On the side that appeared
male, tissues were larger and heavier and bones were longer and
denser than corresponding tissues and bones from the side with a
female appearance (Supplementary Table 1). Fluorescent in situ
hybridization (FISH) analysis using Z and W chromosome probes
was performed on preparations of blood cells from all three animals
and on multiple preparations of cultured skin cells from both sides
of birds G2 and G3. Whereas an autosomal probe demonstrated a
diploid chromosome constitution for G1 blood cells, the sex chro-
mosome probes demonstrated that approximately half of gynand-
romorph G1 cells were female (ZW) and half were male (ZZ)
(Fig. 2a). Similar FISH analyses of blood and primary fibroblast
cultures from birds G2 and G3 demonstrated that all three animals
were composed of a mixture of normal diploid male and female cells
(Supplementary Fig. 2 and Supplementary Table 2). Although a
recent analysis of a gynandromorph zebra finch3 demonstrated that
both Z-chromosome and W-chromosome containing cells were pre-
sent, the possibility remained that such animals were composed of a
mixture of ZW and Z0 cells. Here we show that gynandromorph
birds are genuine male:female chimaeras and provide an explanation
for a phenomenon that has been debated for centuries18.
We next investigated whether the apparent bilateral asymmetry
reflected the distribution of ZZ and ZW cells by examining the cellular
Right Left
Figure 1 | Image of gynandromorph bird (G1). ISA brown bird where the
right side has female characteristics and left side has male characteristics
(white colour and larger wattle, breast musculature and spur).
238
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
a b 100
Mean percentage
80
of cell type
60 ZZ cells
ZW cells
40
20
ZZ/ZW 0
Sk. Wa. BM Sk. Wa. BM
‘Female side’ ‘Male side’
Figure 2 | Male and female cells in gynandromorph birds. a, FISH analysis
of sex chromosomes in gynandromorph blood cells. Shown are interphase
nuclei prepared from cultured blood cells from gynandromorph G1
hybridized according to standard FISH procedures with probes specific to
both the W and Z chromosome (XhoI repeat on W chromosome, and Z
chromosome bacterial artificial chromosome (BAC) containing the VLDL
receptor gene identified by screening the HGMP chicken BAC library).
Erythrocytes were hybridized with probes for Z chromosome (green) and W
chromosome (red). Cells contain either two Z chromosomes or one Z and
one W chromosome. b, Mean relative proportions of ZZ and ZW cells in
tissues from male and female sides of gynandromorph birds. The average
percentage of ZW and ZZ cells (Supplementary Table 2) in three tissues from
the phenotypically female side and from the phenotypically male side of
three gynandromorph birds is shown. Tissues from the sides that appear
female contain more ZW (female) than ZZ (male) cells, whereas tissues from
the sides that appear male are composed predominantly of ZZ cells. BM,
breast muscle; Sk., skin; Wa., wattle.
composition of a variety of tissues from both sides of the individual
birds. Southern analysis using sex chromosome probes on genomic
DNA extracted from multiple tissues revealed that none of the tissues
from either side was composed exclusively of either ZZ- or ZW-
containing cells; that is, all tissues examined comprised a mixture of
both female and male cells (examples shown in Supplementary Fig. 3).
Multiple Southern analyses were performed on separate DNA samples
extracted from different regions of skin, from wattle and from breast
muscle from both sides of all three birds, to quantify the relative
proportions of male and female cells. Phosphorimager analyses com-
paring the hybridization signal obtained from DNA from gynandro-
morphic tissues with that obtained from known amounts of male and
female DNA produced a measure of the relative proportion of male
and female cells in each tissue. Figure 2b shows the mean proportion
of female and male cells in skin, wattle and breast muscle from the
‘male’ side and ‘female’ side of all three birds. It is clear that tissues
from the side that appeared female contained more ZW (female) than
ZZ (male) cells, whereas tissues from the side that appeared male were
composed predominantly of ZZ cells (Supplementary Table 2). Our
data establishing the presence of both ZZ- and ZW-containing cells
indicate that it is highly unlikely that these birds arise as a consequence
of mutation at the two-cell stage of development, and would support
the hypothesis that gynandromorphs arise as a result of failure of
extrusion of a polar body during meiosis and subsequent fertilization
of both a Z- and W-bearing female pronucleus19.
The development of gonads in the gynandromorph birds was of
obvious interest (Supplementary Fig. 4). The type of gonad present
did not correspond to the external appearance but rather reflected the
cellular composition of the individual organs. The gonads differed
for each gynandromorph: G1 contained a testis-like gonad on the left
side, G2 contained an ovary-like gonad on the left side, and G3
contained a swollen testis-like structure on the left side (in contrast
to G1 and G2, G3 appeared female on the left side and male on the
right). The G1 testis-like gonad was composed primarily of sperm-
containing seminiferous tubules, whereas the G2 ovary-like gonad
was composed predominantly of large and small follicles. The gonad
from G3 comprised a mixture of empty tubules and small follicular-
like structures (ovo-testis). Southern analyses demonstrated that the
morphological appearance of the gonads conformed to the cellular
composition in that the structures that appeared to be testis and
ovary were composed principally of ZZ- and ZW-containing cells,
NATURE | Vol 464 | 11 March 2010
respectively, whereas the ‘ovo-testis’ comprised a mixture of ZZ- and
ZW-containing cells.
Although the findings from our gynandromorph analyses are
uninformative in terms of elucidating the avian sex-determining
mechanism, they do lead to the conclusion that the classical dogma
of sex differentiation, where the phenotype is mainly determined by
gonadal hormonal secretions, does not apply to birds. These results
strongly indicate that the avian phenotype is dependent on the nature
of the cells comprising the individual tissue rather than being
imposed by the type of gonad formed: both sides of these animals
are exposed to an identical profile of gonadal products yet each side
responds differently to these stimuli. For example, although it is well
established that growth of the wattle is sensitive to testosterone20, it is
clear from Fig. 1 that a major determinant in wattle size is the cellular
composition of the tissue, and therefore cellular identity and gonadal
hormones both have a significant role in establishing the sexual
phenotype of this tissue. Our analyses led us to conclude that male
and female chicken somatic cells may have a cell-autonomous sex
identity.
Sex differences precede gonadal hormone influences
To investigate whether differences exist between male and female
cells independently of any possible gonadal influences we compared
the transcriptomes of male and female embryos at developmental
stages before the formation of the gonads (data not shown). These
analyses identified both messenger RNAs and microRNAs (miRNAs)
that were expressed in a sexually dimorphic fashion throughout the
embryos at stages before the formation of the gonadal precursor (the
genital ridge; H&H stage 21) and well before the generally accepted
point of sex determination in the chicken (that is, around day 5/6 of
incubation). Screening for mRNAs expressed exclusively in male or
in female embryos led to the identification of an mRNA encoded by a
W chromosome gene that was expressed ubiquitously in females.
This gene was designated FAF for female-associated factor and
sequences were deposited in the EMBL/GenBank databases (acces-
sion numbers AJ606294–AJ606297). Whole-mount in situ hybridiza-
tion analysis of embryos at stages before genital ridge formation
showed that FAF mRNA is expressed throughout the female embryo
as early as 18 h of incubation (H&H stage 4) (Fig. 3a). We also
identified a ubiquitously expressed miRNA that is present at levels
approximately tenfold higher in males than in females throughout
development, including at stages before the expected point of sex
determination (Fig. 3b and Supplementary Fig. 5). This miRNA is
encoded on the Z chromosome and the sequence has not previously
been reported in any other species (Gallus gallus mir-2954, accession
number AM691163). These observations are in agreement with other
studies that have identified sexually dimorphic gene expression in the
brain preceding morphological differentiation of the gonads, in both
chicken and mouse21,22. Although the functions of these particular
a H&H stage 4 H&H stage 14
m
f m f Stage 14 Stage 20
Figure 3 | Sexually dimorphic expression in early chick embryos.
a, Expression of FAF in male and female embryos before development of
genital ridge/gonads. Whole-mount ISH showing expression of FAF
(purple) in embryos at 18 h, 48 h and 72 h of development (H&H stages 4
(original magnification, 340), 14 (320) and 20 (310), respectively). FAF is
clearly expressed throughout the female embryos at all developmental stages
and is not expressed in male embryos. FAF is not expressed in extra-
embryonic tissues of the female. The FAF transcript is encoded by the
genomic DNA complementary to the intergenic regions between copies of
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
transcripts are unknown, these findings not only supported the
concept that the tissue phenotype was not dependent on gonadal
products, but also reinforced the suggestion that the phenotype
was defined by inherent differences between the male and female cells.
Chimaeras confirm cell-autonomous sexual differentiation
To test the hypothesis that the male and female cellular composition
defines phenotype, we generated embryos containing chimaeric
gonads comprised of a mixture of male and female cells. Gonadal
chimaeras were generated by transplantation of sections of presump-
tive mesoderm from green fluorescent protein (GFP)-expressing
embryos23 at developmental stage 12 (day 2) to replace the equivalent
tissue of non-GFP embryos at the same stage of development (Fig. 4a).
Donor tissue was transplanted only to the left side of recipient
embryos as only the left ovary develops fully in the chick. Trans-
planted embryos were returned to the incubator and allowed to
develop until stage 35 (day 9). By stage 35, donor cells were incor-
porated into tissues on the left side of the embryo in the region
between the fore and hind limbs (Supplementary Fig. 6). In addition
to the gonads, donor cells were observed in a variety of tissues includ-
ing skin, muscle, mesonephros, Wolffian duct and Müllerian duct. A
minimum of four donor:host chimaeric gonads were generated for
each of the four possible combinations: male:male, female:female,
male:female and female:male. For each of these donor:host com-
binations, chimaeras were generated with different levels of donor
contribution—ranging from examples where the contribution of
donor cells was limited to isolated individual cells dispersed through-
out the host gonad, to instances where areas of the host gonad were
almost exclusively composed of donor cells. Gonad:mesonephros
pairs were collected at stage 35 and longitudinal frozen sections
were prepared for confocal microscopy. GFP expression was used to
estimate the extent of donor contribution to the individual chimaeric
gonads and immunohistochemistry (IHC) analysis was performed
with antibodies for both anti-Müllerian hormone (AMH) and aro-
matase. AMH is a marker of functionally ‘male’ cells24 (expressed by
precursor Sertoli cells of the sex cords) whereas aromatase is a marker
of functionally ‘female’ cells25 (expressed by cells in the female medul-
lary region). At stage 35 of development, the male gonad is composed
of a thin layer of cortex tissue surrounding a medullary region which
contains the developing sex cords (expressing AMH) separated by
interstitial connective tissue. In contrast, the female left ovary com-
prises a greatly thickened cortex surrounding a smaller less-structured
medullary region (expressing aromatase). Figure 4b shows the normal
expression of AMH and aromatase in stage 35 male and female
gonads, respectively. It is clear that the testis is composed almost
exclusively of medullary tissue and that AMH is expressed in distinct
cord-like structures within this tissue. In contrast, the developing
ovary comprises a definitive cortex enclosing a reduced medulla and
aromatase is expressed in cells throughout the medulla. Examples of
H&H stage 20 b m f m f
f
miRNA
m
U6
the W-chromosome repeat gene Wpkci (also called HINTW)34,35 and
transcribed in the opposite orientation. f, female; m, male. b, Expression of
novel chicken miRNA (Gallus gallus mir-2954). Expression in whole
embryos at 48 h (H&H 14) and 72 h (H&H 20) of development is shown.
This miRNA is clearly expressed in a sexually dimorphic fashion at stages
before the sexual differentiation of the gonads. This miRNA matches
sequence present in chicken Z-chromosome BAC clones AC192757 and
AC187119. U6 RNA was used as a loading control.
239
ARTICLES
a c m GFP
g g
Transplant m m
t
Donor Host md
b m
d GFP AMH GFP/AMH m GFP
GFP AROM GFP/AROM m GFP
t
Figure 4 | Expression of male and female markers in chimaeric gonads.
a, Generation of chimaeras. Left: schematic illustrating transplantation of
presumptive mesoderm from GFP-expressing embryo to non-GFP embryo
at day 2. Right: image of mesonephros and gonads from chimaeric embryo at
day 9 showing donor contribution to left gonad (g), mesonephros (m) and
Müllerian duct (md). Original magnification, 320. b, Expression of female
and male markers in embryonic gonads. Expression of aromatase (AROM)
in ovary and anti-Müllerian hormone (AMH) in testis at day 9 is shown by
IHC. Original magnification, 3400. c, Integration of GFP-expressing donor
cells into host gonads. Panels in the first column show a low-magnification
view of sections through host gonads and illustrate the extent of donor cell
contribution. Panels to the right show higher-magnification views of
highlighted areas. Using IHC, donor cells are marked by GFP (green)
whereas expression of AMH and aromatase are shown in red. The fourth
column is a merged image of the images from the second and third columns.
In same-sex chimaeras, GFP-expressing donor cells co-localize with AMH-
expressing and aromatase-expressing cells in host testis and ovary,
respectively (yellow/orange in the fourth column). In mixed-sex chimaeras,
GFP-expressing donor cells do not co-localize with AMH or aromatase. m,
all four donor:host combinations of chimaeric gonads are shown in
Fig. 4c and Supplementary Fig. 7. It is clear that GFP expression did
not affect the ability of donor cells to contribute to host tissues and to
function normally: in each case of same-sex chimaeras, either male or
female donor cells were integrated into all somatic compartments of
the respective host testis and ovary (cortex, sex cords and interstitial
tissue). Moreover, when integrated into the appropriate ‘functional’
compartment, donor male cells expressed AMH and donor female
cells expressed aromatase. In contrast, in mixed-sex chimaeras the
donor cells did not integrate into the ‘functional’ structures of the
host gonad: female donor cells in host testis medulla were not
recruited into the AMH-expressing sex cords and were restricted to
the interstitial tissue, whereas male donor cells in host ovary were
excluded from the aromatase-expressing structures. In mixed-sex chi-
maeras the inability of donor cells to form functional host structures
was evident regardless of the relative contribution of male and female
cells (Supplementary Fig. 7). The fact that female chicken cells in an
environment and location that induces testicular development cannot
be recruited into the functionally ‘male’ Sertoli cell compartment, and
male cells in an ovary-inducing environment are excluded from a
functionally ‘female’ compartment, strongly supports the suggestion
that chicken somatic cells possess a cell-autonomous sexual identity.
240
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
AMH GFP/AMH
GFP AROM GFP/AROM
AROM GFP/AROM
AMH GFP/AMH
mesonephros; o, ovary; t, testis. d, Retention of female donor phenotype in
mixed-sex chimaeras. IHC showing expression of AMH (red in top row) and
aromatase (red in bottom row) in neighbouring sections from the gonad of a
female:male (donor:host) chimaera. Donor contribution is illustrated by
GFP (green) expression. The right column shows a merged image of the
images in left and middle columns. Regions containing a significant host
contribution (defined by the bottom bracket) formed sex-cord-like
structures and expressed AMH. Female donor cells were not incorporated
into AMH-expressing sex cords, as shown by the lack of GFP and AMH co-
localization. Regions composed primarily of female donor cells (defined by
top bracket) behaved as ovarian-like tissue and expressed aromatase, as
shown by co-localization of GFP and aromatase (yellow/orange). Scale bars
in c and d indicate 100 mm. IHC was performed following standard
procedures. Primary antibodies were (1:100) goat anti-human AMH (Santa
Cruz Biotechnology), (1:200) mouse anti-human cytochrome P450
aromatase (AbD Serotec) and (1:250) rabbit anti-GFP conjugated to Alexa
Fluor 488 (Invitrogen). Secondary antibodies were conjugated to Alexa
Fluor 594 (Invitrogen).
This is further supported by a striking example where the degree of
the female donor contribution was sufficient to effectively generate
an ‘ovo-testis’ in the host embryo (Fig. 4d). This mixed-sex chimaera
contained a gonad with an anterior portion composed almost exclu-
sively of female cells. Whereas the posterior portion contained testis-
like medulla with AMH-expressing sex cords, the region composed of
female cells did not form sex cords and did not express AMH. Most
surprisingly, the female cells in this region expressed aromatase. This
demonstrates that although female cells in a male embryo can correctly
interpret gonadal location and differentiation signals, they respond in
a cell-autonomous manner characteristic of a female genotype (and
express aromatase). Our findings are in contrast with those from
mammalian mixed-sex chimaeras, where XX cells can become func-
tional Sertoli cells and XY cells can become functional granulosa
cells26,27.
These studies demonstrate that avian somatic cells possess a cell-
autonomous sex identity. Our results support and extend previous
findings3 that showed that differences between male and female zebra
finch brains were a result of endogenous genetic differences in the
brain cells themselves. Our analysis of lateral gynandromorph birds,
showing that they are male:female chimaeras, and our experimental
generation of embryos with mixed-sex chimaeric gonads, together
NATURE | Vol 464 | 11 March 2010
Mammalian model
Sry Testis
Male
phenotype
Hormones
Genital Indifferent Hormones
ridge gonad Female
Ovary phenotype
Chicken model
Male Male
soma phenotype
Testis
es
Genital Hormon
ridge DMRT1(?) Hormo
nes
Ovary
Female Female
soma phenotype
Figure 5 | A novel mechanism of sex determination in the chicken. A sexual
identity is genetically imposed on the male and female chicken soma at
fertilization and is the major factor in determining the adult sexual
phenotype. At the appropriate stage in development, the sexually-dimorphic
transcripts underlying the male/female identity trigger expression in the
genital ridge of the gene cascade responsible for testis/ovary development.
The gonads have limited effects on the sexual phenotype. In contrast, in
mammals, gonadal fate is dependent on transient expression of the testis-
determining Sry gene in the indifferent early gonad. The mammalian gonads
have a major influence on the sexual phenotype.
indicate that male and female somatic cells possess a sex identity.
These observations indicate that there is a molecular mechanism
functioning in every cell that confers a sex-specific identity that
influences how individual cells respond to developmental and hor-
monal signals. We propose that cell-autonomous sex identity is
dependent on sexually dimorphic gene expression resulting from
the ‘dosage compensation’ system that operates to equalize the
phenotypic effects of characteristics determined by genes on the Z
chromosome. Recent evidence has shown that this system in birds is
not chromosome-wide and results in a large number of gene expres-
sion differences between male and female cells28–31. We have esti-
mated that this system of dosage compensation would result in at
least 300 non-compensated Z-chromosome genes31. Our identifica-
tion of sexually dimorphic transcripts that are expressed ubiquitously
from very early in development adds to these observations. On the
basis of our findings, and from evidence of the dosage compensation
system in birds, we propose that the overall mechanism of sex deter-
mination in birds differs significantly from the mammalian model
(Fig. 5). Although sexually dimorphic differentiation of the gonads
may be regulated independently from other somatic tissues, we pro-
pose that a male or female sex identity is imposed on the chicken
soma early in development by sex chromosome transcription and it is
this inherent molecular identity that triggers the appropriate testis or
ovary gene cascade in the developing genital ridge (for example, via
DMRT1 (ref. 15)). Although the gonads clearly have a significant
influence on the adult phenotype they do not dictate somatic differ-
ences to the same extent as in mammals. It is also possible that
elements of such a system are retained in certain mammals: previous
studies have shown that, in a marsupial mammal, the wallaby, forma-
tion of the mammary gland and scrotum is independent of gonadal
hormones32, and rather than exhibiting transient localized expres-
sion, Sry shows widespread expression in multiple tissues well before
the point of gonadal differentiation33. As Sry-type sex-determining
mechanisms have not yet been established for all vertebrate species, it
is possible that the model we propose where the phenotype of indi-
vidual tissues is largely defined by an inherent sex identity of the
somatic cells is not restricted to birds.
METHODS SUMMARY
Generation of chimaeric embryos:GFP embryos23 and ISA brown embryos at
H&H stage 11/12 (13–15 somites) were used as donor and host, respectively. The
blunt end of donor eggs was pierced to create an air hole and a ‘window’ cut on
the midline. The embryos were removed and pinned on a 3% agarose surface
containing 0.5% India ink. Embryos were kept moist by the addition of PBS
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
containing 100 units ml21 penicillin and 100 mg ml21 streptomycin (PBS-Pen/
Strep). A strip of presumptive mesoderm flanking presumptive somites 21–23
was removed and stored in CO2-independent medium (Invitrogen) containing
10% FBS and Pen/Strep. Host eggs were windowed as above and kept moist by
the addition of PBS-Pen/Strep. To help visualize somites, sterile India ink (20%
in PBS-Pen/Strep) was injected under the host embryos. Using a microneedle,
the vitelline membrane and a flap of ectoderm were folded back from the under-
lying mesoderm. A strip of presumptive mesoderm was then removed from the
host embryo taking care to leave the endoderm intact. The GFP-donor tissue was
then inserted into the host site and the ectodermal flap replaced. Two millilitres
of albumen was then withdrawn from the host eggs using a hypodermic syringe.
Transplanted eggs were tightly sealed with tape and incubated at 37 uC in a
humidified incubator. All other methods are standard.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 10 August 2009; accepted 18 January 2010.
1. Lillie, F. R. Sex-determination and sex-differentiation in mammals. Proc. Natl Acad.
Sci. USA 3, 464–470 (1917).
2. Jost, A. Hormonal factors in the sex differentiation of the mammalian foetus. Phil.
Trans. R. Soc. Lond. B 259, 119–131 (1970).
3. Agate, R. J. et al. Neural, not gonadal, origin of brain sex differences in a
gynandromorph finch. Proc. Natl Acad. Sci. USA 100, 4873–4878 (2003).
4. Wade, J. & Arnold, A. P. Functional testicular tissue does not masculinize
development of the zebra finch song system. Proc. Natl Acad. Sci. USA 93,
5264–5268 (1996).
5. Arnold, A. P. Sexual differentiation of the zebra finch song system: positive
evidence, negative evidence, null hypothesis, and a paradigm shift. J. Neurobiol. 33,
572–584 (1997).
6. Wade, J. & Arnold, A. P. Sexual differentiation of the zebra finch song system. Ann.
NY Acad. Sci. 1016, 540–559 (2004).
7. Renfree, M. B. & Short, R. V. Sex determination in marsupials: evidence for a
marsupial-eutherian dichotomy. Phil. Trans. R. Soc. Lond. B 322, 41–53 (1988).
8. Glickman, S. E., Short, R. V. & Renfree, M. B. Sexual differentiation in three
unconventional mammals: spotted hyenas, elephants and tammar wallabies.
Horm. Behav. 48, 403–417 (2005).
9. Sekido, R. & Lovell-Badge, R. Sex determination and SRY: down to a wink and a
nudge? Trends Genet. 25, 19–29 (2009).
10. Matsuda, M. et al. DMY gene induces male development in genetically female
(XX) fish. Proc. Natl Acad. Sci. USA 104, 3865–3870 (2007).
11. Volff, J.-N., Kondo, M. & Schartl, M. Medaka dmY/dmrt1Y is not the universal
primary sex-determining gene in fish. Trends Genet. 19, 196–199 (2003).
12. Hamburger, V. & Hamilton, H. L. A series of normal stages in the development of
the chick embryo. J. Morphol. 88, 49–92 (1951).
13. Smith, C. A. & Sinclair, A. H. Sex determination in the chicken embryo. J. Exp. Zool.
290, 691–699 (2001).
14. Clinton, M. Sex determination and gonadal development: a bird’s eye view. J. Exp.
Zool. 281, 457–465 (1998).
15. Smith, C. A. et al. The avian Z-linked gene DMRT1 is required for male sex
determination in the chicken. Nature 461, 267–271 (2009).
16. Hutt, F. B. Genetics of the Fowl (McGraw-Hill, 1949).
17. Cock, A. G. Half-and-half mosaics in the fowl. J. Genet. 53, 49–80 (1955).
18. Birkhead, T. The Wisdom of Birds. An Illustrated History of Ornithology (Bloomsbury,
2008).
19. Hollander, W. F. Sectorial mosaics in the domestic pigeon: 25 more years. J. Hered.
66, 197–202 (1975).
20. Owens, I. P. F. & Short, R. V. Hormonal basis of sexual dimorphism in birds:
implications for new theories of sexual selection. Trends Ecol. Evol. 10, 44–47
(1995).
21. Scholz, B. et al. Sex-dependent gene expression in early brain development of
chicken embryos. BMC Neurosci. 7, 12 (2006).
22. Dewing, P., Shi, T., Horvath, S. & Vilain, E. Sexually dimorphic gene expression in
mouse brain precedes gonadal differentiation. Brain Res. Mol. Brain Res. 118,
82–90 (2003).
23. McGrew, M. et al. Localised axial progenitor cell populations in the avian tail bud
are not committed to a posterior Hox identity. Development 135, 2289–2299
(2008).
24. Nishikimi, H. et al. Sex differentiation and mRNA expression of P450c17,
P450arom and AMH in gonads of the chicken. Mol. Reprod. Dev. 55, 20–30
(2000).
25. Nomura, O., Nakabayashi, O., Nishimori, K., Yasue, H. & Mizuno, S. Expression of
five steroidogenic genes including aromatase gene at early developmental stages
of chicken male and female embryos. J. Steroid Biochem. Mol. Biol. 71, 103–109
(1999).
26. Patek, C. E. et al. Sex chimaerism, fertility and sex determination in the mouse.
Development 113, 311–325 (1991).
27. Burgoyne, P. S., Buehr, M. & McLaren, A. XY follicle cells in ovaries of XX–XY
female mouse chimaeras. Development 104, 683–688 (1988).
241
ARTICLES
28. McQueen, H. A. et al. Dosage compensation in birds. Curr. Biol. 11, 253–257
(2001).
29. Melamed, E. & Arnold, A. P. Regional differences in dosage compensation on the
chicken Z-chromosome. Genome Biol. 8, R202 (2007).
30. Ellegren, H. et al. Faced with inequality: chicken do not have a general dosage
compensation of sex-linked genes. BMC Biol. 5, 40 (2007).
31. McQueen, H. A. & Clinton, M. Avian sex chromosomes: dosage compensation
matters. Chrom. Res. 17, 687–697 (2009).
32. O, W.-S., Short, R. V., Renfree, M. B. & Shaw, G. Primary genetic control of somatic
sexual differentiation in a mammal. Nature 331, 716–717 (1988).
33. Harry, J. L., Koopman, P., Brennan, F. E., Graves, J. A. & Renfree, M. B. Widespread
expression of the testis-determining gene SRY in a marsupial. Nature Genet. 11,
347–349 (1995).
34. Hori, T., Asakawa, S., Itoh, Y., Shimizu, N. & Mizuno, S. Wpkci, encoding an altered
form of PKCI, is conserved widely on the avian W chromosome and expressed in
early female embryos: implications of its role in female sex determination. Mol.
Biol. Cell 11, 3645–3660 (2000).
35. Smith, C. A., Roeszler, K. N. & Sinclair, A. H. Genetic evidence against a role for
W-linked histidine triad nucleotide binding protein (HINTW) in avian sex
determination. Int. J. Dev. Biol. 53, 59–67 (2009).
242
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements This work was supported by DEFRA and BBSRC (BB/
E015425/1). We thank G. Miele, G. Robertson, S. Wilson, A. Sherman,
M. Hutchison, F. Thomson and R. Mitchell for technical support and for provision of
fertilized eggs and embryos. We also thank T. Cannon for donation of
gynandromorph bird G1, and R. Field and N. Russell for photography.
Author Contributions D.Z. and D.M. performed transplantation studies,
transcriptome screens, Southern analyses and general molecular biology. S.N.
performed immunostaining, H.A.M. performed FISH analyses and P.M.H.
performed dissections and post-mortem measurements. M.J.M. performed ISH
and suggested transplantation strategy and P.D.L. obtained gynandromorph birds.
Overall project was conceived by M.C. and H.M.S. M.C. carried out day-to-day
supervision and wrote the manuscript. All authors edited the manuscript.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to M.C.
(Michael.clinton@roslin.ed.ac.uk).
doi:10.1038/nature08852
METHODS
Embryo transplantation. Generation of chimaeric embryos:GFP embryos23 and
ISA brown embryos at H&H stage 11/12 (13–15 somites) were used as donor and
host, respectively. The blunt end of donor eggs was pierced to create an air hole
and a ‘window’ cut on the midline. The embryos were removed and pinned on a
3% agarose surface containing 0.5% India ink. Embryos were kept moist by the
addition of phosphate buffered saline (PBS) containing 100 units ml21 penicillin
and 100 mg ml21 streptomycin (PBS-Pen/Strep). A strip of lateral plate meso-
derm flanking presumptive somites 21–23 was removed and stored in CO2-
independent medium (Invitrogen) containing 10% FBS and Pen/Strep. Host
eggs were windowed as above and kept moist by the addition of PBS-Pen/
Strep. To help visualize somites, sterile India ink (20% in PBS-Pen/Strep) was
injected under the host embryos. Using a microneedle, the vitelline membrane
and a flap of ectoderm were folded back from the underlying mesoderm. A strip
of lateral plate mesoderm was then removed from the host embryo, taking care to
leave the endoderm intact. The GFP-donor tissue was then inserted into the host
site and the ectodermal flap replaced. Two millilitres of albumen was then
withdrawn from the host eggs using a hypodermic syringe. Transplanted eggs
were tightly sealed with tape and incubated at 37 uC in a humidified incubator.
Immunostaining. Immunohistochemistry was carried out as described previ-
ously36. Briefly, tissues were fixed in 4% paraformaldehyde for 2 h at 4 uC,
equilibrated in 15% sucrose then embedded in 15% sucrose plus 7.5% gelatin
in PBS, pH 7.2. Sections, 15 mm thick, mounted on Superfrost slides (Menzel)
were washed for 30 min in PBS at 37 uC and blocked in PBS containing 10%
donkey serum, 1% BSA, 0.3% Triton X-100 and 0.05% Tween 20 for 2 h at 22–24
uC. Incubation with primary antibodies was carried out overnight at 4 uC, fol-
lowed by washing in PBS containing 0.3% Triton X-100 and 0.05% Tween 20,
and then incubation with secondary antibodies for 2 h at room temperature.
After washing, the sections were treated with Hoechst solution (10 mg ml21) for
5 min to stain nuclei.
Fluorescent in situ hybridization (FISH). FISH analysis of metaphase or inter-
phase preparations of chicken cells was performed by standard procedures37.
BAC clones containing the VLDL receptor, aldolase B, CHRN or SCII genes were
identified by screening the HGMP chicken BAC library and used to identify Z
chromosomes. A probe for the W chromosome was prepared by polymerase
chain reaction (PCR) amplification of a portion of the XhoI repeat region from
the W chromosome38. After gel purification, the probe was labelled by incor-
poration of either biotin-16-dUTP (Roche) or digoxigenin-11-dUTP (Roche)
during a further round of PCR. BAC DNA was prepared using Qiagen plasmid
columns following recommendations for low-copy plasmid purification. Biotin-
16-dUTP and digoxigenin-11-dUTP were incorporated into BAC DNA by nick
translation and labelled probes were concentrated by precipitation in the pres-
ence of 5 mg of salmon sperm DNA as a carrier and 2 mg of sonicated chicken
genomic DNA as competitor. The pellet was resuspended in 15 ml of hybridiza-
tion mix, denatured and pre-annealed for 15 min at 37 uC to block repetitive
sequences.
Whole-mount in situ hybridization. Chicken embryos and isolated gonads were
fixed in 4% paraformaldehyde for 1 h and whole-mount in situ hybridization was
©2010 Macmillan Publishers Limited. All rights reserved
carried out as described previously39. Digoxigenin-labelled probes were prepared
from linearized plasmid clones using a Roche DIG RNA labelling kit to incor-
porate digoxigenin-11-UTP by in vitro transcription with SP6 and T7 RNA
polymerases.
RNA preparation. Total RNA was extracted from pools of male and female chick
embryos and tissues using RNA-Bee (AMS Biotechnology) according to the
manufacturer’s instructions.
Differential display. RNA expression profiles in male and female embryos were
compared by differential display reverse transcription PCR (DDRT–PCR).
Embryos were sexed38 and pools of RNA from male and female embryos generated.
DDRT–PCR was performed as described previously40.
miRNA library construction. Low-molecular-mass RNAs (,40 nucleotides
long) were isolated from total RNA by the use of a flashPAGE fractionator
(Ambion). MicroRNA libraries were constructed essentially as described previ-
ously41,42.
MicroRNA northern analysis. Five micrograms of total RNA was separated by
electrophoresis through a 15% TBE/urea polyacrylamide gel (Bio-Rad) before
transfer to Hybond-N1 membrane (GE Healthcare). Locked nucleic acid (LNA)
oligonucleotides antisense to the mature miRNA were end-labelled (mirVana
Probe and Marker kit, Ambion) with 32P-dATP (Perkin-Elmer) and hybridized
to membranes containing miRNAs. Hybridization was carried out overnight in
ULTRAhyb-oligo (Invitrogen) at 42 uC and membranes washed at 63 uC in 0.13
SSC/0.1% SDS43 (22 uC below the estimated melting temperature of the LNA,
85 uC).
Southern analysis. High-molecular-mass genomic DNA was extracted from
tissues of embryonic and adult male and female chickens by standard phenol-
chloroform procedures43. DNA was digested with restriction endonucleases,
subjected to electrophoresis on a 1% TBE gel and transferred to Hybond-N
membrane. Probes labelled with 32P-dCTP were hybridized by standard proce-
dures and signal was recorded on high-sensitivity film (Kodak) and by phos-
phorimager analysis.
36. Stern, C. D. In Essential Developmental Biology: A Practical Approach (eds Stern, C.
D. & Holland, P. W. H.) 193–212 (IRL, 1993).
37. McQueen, H. A. et al. CpG islands of chicken are concentrated on
microchromosomes. Nature Genet. 12, 321–324 (1996).
38. Clinton, M., Haines, L., Belloir, B. & Mcbride, D. Sexing chick embryos: a rapid and
simple protocol. Br. Poult. Sci. 42, 134–138 (2001).
39. Henrique, D. et al. Expression of a Delta homologue in prospective neurons in the
chick. Nature 375, 787–790 (1995).
40. Clinton, M., Miele, G., Nandi, S. & McBride, D. In Differential Display Methods and
Protocols 2nd edn (eds Liang, P., Meade, J. D. & Pardee, A. B.) 157–178 (Humana,
2006).
41. Lau, N. C. et al. An abundant class of tiny RNAs with probable regulatory roles in
Caenorhabditis elegans. Science 294, 858–862 (2001).
42. Lee, R. C. & Ambros, V. An extensive class of small RNAs in Caenorhabditis elegans.
Science 294, 862–864 (2001).
43. Sambrook, J. & Russell, D. Molecular Cloning: A Laboratory Manual (Cold Spring
Harbor Laboratory Press, 2001).
Vol 464 | 11 March 2010 | doi:10.1038/nature08779

ARTICLES
Systems survey of endocytosis by
multiparametric image analysis
Claudio Collinet1, Martin Stöter2, Charles R. Bradshaw1, Nikolay Samusik1, Jochen C. Rink{, Denise Kenski{,
Bianca Habermann1, Frank Buchholz1, Robert Henschel3, Matthias S. Mueller3, Wolfgang E. Nagel3, Eugenio Fava2,
Yannis Kalaidzidis1,4 & Marino Zerial1

Endocytosis is a complex process fulfilling many cellular and developmental functions. Understanding how it is regulated and
integrated with other cellular processes requires a comprehensive analysis of its molecular constituents and general design
principles. Here, we developed a new strategy to phenotypically profile the human genome with respect to transferrin (TF)
and epidermal growth factor (EGF) endocytosis by combining RNA interference, automated high-resolution confocal
microscopy, quantitative multiparametric image analysis and high-performance computing. We identified several novel
components of endocytic trafficking, including genes implicated in human diseases. We found that signalling pathways such
as Wnt, integrin/cell adhesion, transforming growth factor (TGF)-b and Notch regulate the endocytic system, and identified
new genes involved in cargo sorting to a subset of signalling endosomes. A systems analysis by Bayesian networks further
showed that the number, size, concentration of cargo and intracellular position of endosomes are not determined randomly
but are subject to specific regulation, thus uncovering novel properties of the endocytic system.

Endocytosis is a fundamental process supporting many functions platform, on the basis of the custom-designed image analysis software
such as nutrient uptake, intracellular signalling1, morphogenesis MotionTracking12. First, images were segmented to determine mor-
during development2 and defence against pathogens3. Dysfunctions phological and positional descriptors for each endosome (Fig. 1b).
of the endocytic system lead to severe metabolic4, infectious5 and Second, a total set of 62 parameters were extracted, 4 serving as quality
neurodegenerative diseases6. Receptors are internalized by both control to reject sub-optimal images and 58 describing defined bio-
Clathrin-dependent and -independent mechanisms into early endo- logical properties of the endocytic system, such as total internal cargo,
somes and, from here, follow different intracellular routes. For number, size (mean apparent area) and position of endosomes
example, transferrin (TF) is recycled back to the plasma membrane, (Fig. 1c, Supplementary Table 1 and Supplementary Information
whereas epidermal growth factor (EGF) is routed to late endosomes/ for details). The latter were processed to suppress plate-to-plate
lysosomes for degradation. Despite a great deal of molecular random variations, normalized (see Supplementary Information)
insights7–9 our understanding of the endocytic machinery remains and the 46 most robust parameters (see Supplementary Information
fragmentary. Furthermore, the design principles of the endocytic and Supplementary Table 1) were combined into multiparametric
pathway, its integration in the overall cellular system and high-order profiles (Fig. 1d) describing quantitatively the endocytic phenotypes.
control are still largely unknown. Addressing these problems requires In preparation for the genomic screen, we validated the quality and
methods capable of seizing the complexity at the system level (that is, reproducibility of assay and QMPIA. First, we verified that the deple-
measuring many key parameters). tion of established components of the endocytic machinery, such as
Here, we designed a new analytical approach aimed at phenotypi- Clathrin (CLTC), Dynamin-2 (DNM2), EGF receptor (EGFR),
cally profiling genes with respect to endocytosis by genome-wide Transferrin receptor (TFRC), early endosome antigen 1 (EEA1)
RNAi screening through the quantitative assessment of many system and others by RNAi (Supplementary Figs 1 and 2), yielded the
parameters. expected alterations in phenotypic profiles. For example, silencing
of TFRC reduced the values of number of endosomes, total internal
Multiparametric profiling of the endocytic system cargo and mean cargo load per endosome for TF but not EGF (Fig. 1d
We visualized endocytosis of fluorescent TF and EGF in HeLa cells and Supplementary Fig. 1e). Conversely, EGFR downregulation per-
before we could apply our platform to more disease-relevant systems. turbed the parameters of EGF but not of TF (Supplementary Fig. 1e).
After small interfering RNA (siRNA) transfection, cells internalized Second, we validated the use of a single confocal section as representa-
the two cargo markers simultaneously for 10 min, were fixed, stained tive of the three-dimensional endosomal population of the entire cell
with 49,6-diamidino-2-phenylindole (DAPI)/SYTO42 blue for nuclei (Supplementary Fig. 3). Third, we validated the reproducibility of the
and cytoplasm detection, and imaged (12 images per well) by triple- observed phenotypic profiles using the same and different siRNAs
colour high-resolution automated confocal microscopy (Fig. 1a). targeting the same gene for a set of 78 genes and 468 siRNA (6
Instead of performing a phenotypic evaluation on the basis of single siRNAs per gene). We found that profiles of the same siRNAs were
or few parameters as commonly done in previous screens10,11, we highly reproducible (Fig. 2a) (mean Pearson correlation coefficient
developed a quantitative multiparametric image analysis (QMPIA) 0.6 6 0.02 s.e.m.), compared with other monoparametric screens13
1
Max Planck Institute for Molecular Cell Biology and Genetics, 2High-Throughput Technology Development Studio, MPI-CBG, Pfotenhauerstrasse 108, 01307 Dresden, Germany.
3
Center for Information Services and High Performance Computing (ZIH), Dresden University of Technology, D-01062 Dresden, Germany. 4Belozersky Institute of Physico-Chemical
Biology, Moscow State University, 119899, Moscow, Russia. {Present addresses: University of Utah School of Medicine, 401 MREB, 20 North 1900 East, Salt Lake City, Utah 84132-
3401, USA (J.C.R.); Sirna Therapeutics Inc., San Francisco, California 94158, USA (D.K.).

243
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES NATURE | Vol 464 | 11 March 2010

a a C12orf4 runs
EGF Synthetic EGF TF Other
TF 6.0
DAPI
4.0

Normalized z-values
2.0

0.0

–2.0
20 μm 20 μm
–4.0
b
–6.0

P1
P2

P8
P3
P4
P5
P6
P7

P9
P1

P 11

P2 8

P37
P10

P1 2
P13
P14
P1 5
P16
P17

P20
P23

P3 0
P31

P38
P2 6
P28
P39

P32
P3 3
P34
P35
P3 6

P4 9
P40
P 41

P43
P4 2

P 44

P5 7
P 50
P5 2
P55
P5 6
P57
P4
5

8
{

{
{
{
{
{
{
{
{
{
{
{

{
{
{
{
G

G
G

G
G
G

G
G

10
2

8
3
1

1
6
Parameter

c Multi-parametric analysis b C12orf4 all siRNAs

EGF TF Other
Endosome parameter groups/cargo Other endocytic parameters 6.0
G1. Number of endosomes (1). G9. Background intensity (2).
G2. Cargo uptake (2). G10. Colocalization (2).

Normalized z-values
4.0
G3. Endosome area (3).
G4. Vesicle elongation (3). Quality control parameters
G5. Endosomal cargo concentration (3). - Number of nuclei. 2.0
G6. Endosomal cargo content (3). - Nuclear size.
G7. Endosome distance from nucleus (3). - Nucleus intensity. 0.0
G8. Endosome clustering (9). - Focus score.

d EGF TF Other
–2.0

30 Control –4.0
Normalized z-value

20 Internalized TFRC
cargo Cargo load –6.0
Number Cargo density
10

P1
P2

P4

P8
P3

P5
P6
P7

P9
P1
P10
P 11
P1 2
P13
P14
P1 5
P16

P20

P39
P17

P3 6

P 38
P4 9
P2 8

P23
P2 6
P28

P3 0
P31
P32
P3 3
P34
P35

P37

P40
P41

P43

P4 5

P 50
P4 2

P44

P57

P5 2
P55
P5 6
P57
8
{

{
{
{
{
{
{
{
{
{
{
{

{
{
{
{
0

G
G

G
G
G

10
2

8
3
1

1
–10 Parameter
–20 Figure 2 | Reproducibility of multiparametric profiles by the same siRNA
–30 and different siRNAs/gene. a, Reproducibility of multiparametric profiles
–40 of a single siRNA silencing the C12orf4 gene. The same siRNA (ID: 28470)
was transfected in seven independent experiments (runs) and
P3
P1
P2

P5

P7

P9

P10
P11

P14

P16

P28

P23
P4

P6

P 36
P37
P38
P 49

P41
P 42

P57

P 56
P57
P8

P1

P12
P13

P 15

P17

P 28

P32

P35
P20

P 26

P39
P 30
P31

P 33
P34

P40

P43
P44

P50

P55
P 45

P 52

8
{
{
{
{
{

{
{
{
{
{
{
{
{
{
{
{

multiparametric profiles (differently coloured for each experiment) were

G
G

G
G
G

G
G

G
1
2

10
3

9
4

8
8

Parameter calculated and plotted as described in Fig. 1. b, Reproducibility of

Figure 1 | Multiparametric image analysis. a, Example of three-colour high-
resolution images collected with the automated spinning disk confocal
OPERA microscope (left). Nuclei and cytoplasm are pseudo-coloured in blue,
Alexa-488-labelled EGF is pseudo-coloured in red and Alexa-647-labelled TF
is pseudo-coloured in green. The synthetic image (right) was obtained after
background intensity subtraction and modelling of endosomal structures
(see Supplementary Information). b, Close-ups show the model structure of
individual endosomes. c, List of quantitative parameters divided into groups
(G1–G10). Ten groups of parameters (number of parameters in each group
indicated in brackets) were used to measure endosomal features and four
parameters as image quality control. See Supplementary Table 1 and
Supplementary Information for further description on the QMPIA
parameters. d, Example of QMPIA profile (TFRC; see pictures in
Supplementary Fig. 1c). The 46 most specific parameters (listed in
Supplementary Table 1 and Supplementary Information) are aligned on the
x-axis and normalized z-values are plotted on the y-axis. Parameter groups
(G1–10) and parameter numbers (P1–P58) are indicated corresponding to
the nomenclature in c and to the enumeration in Supplementary Table 1,
respectively. Black continuous lines separate the parameter groups and grey
dashed lines separate the parameters. The horizontal bars indicate the
endocytic markers which the parameters refer to. Key parameters affected in
the TFRC profile are indicated by arrows.
(see Supplementary Information), indicating that the experimental
noise is not increased by the higher complexity of the assay. In contrast,
different siRNAs targeting the same gene frequently yielded different
profiles (Fig. 2b) (mean Pearson correlation coefficient 0.047 6 0.01
s.e.m.), consistent with high incidence of RNAi off-target13–17. As we
excluded variations in protein depletion and silencing of alternatively
spliced transcripts (data not shown), the major cause of incoherence
between profiles of different siRNAs is most probably off-target effects.
multiparametric profiles of different siRNAs targeting the same gene
(C12orf4). C12orf4 was silenced with six different siRNAs (IDs: 133085,
133086, 133087, 28285, 28378, 28470) in the same experiment and
corresponding multiparametric profiles were calculated and plotted as
described in Fig. 1.
probabilistic approach to infer gene-specific phenotypes by ‘averaging’
the profiles of different siRNAs targeting the same gene. We defined the
‘gene profile’ as the most probable profile calculated as mode of the a
posteriori joint probability distribution considering all individual
siRNA profiles for a given gene. Notably, here every siRNA targeting
a given gene contributes, to a different extent, to the determination of
gene profiles (see Supplementary Fig. 4), circumventing the need of
similarity thresholds to define supporting siRNAs (2 of 3, 3 of 3, and
so on)18. Gene profiles instead of individual siRNA profiles were sub-
sequently used for phenotypic clustering and gene classification.
As test, we screened 1,000 genes including 35 known endocytic
regulators (Supplementary Table 2) and 92 kinases identified in a
previous screen19 (Supplementary Table 3) using seven siRNAs per
gene. On the basis of the QMPIA, 31 of the 35 endocytic regulator
genes (,90%) and 62 out of 92 kinase genes presented a x2 probability
(1.0 2 P) $ 0.95 corresponding to a statistically significant enrich-
ment (P 5 1.1 3 10211 and P 5 5 3 10212, respectively) in the list of
scores. Interestingly, only 7 out of the 35 endocytic regulators would
have scored considering only two parameters (EGF and TF total
internal cargo). Altogether, these data indicate that the QMPIA has
the advantage of higher comprehensiveness and specificity over
monoparametric read-outs in the identification of genes regulating
endocytic trafficking.

Gene phenotypic profiling Genome-wide survey of EGF and TF endocytosis

Assuming that each siRNA profile has both ‘on-target’ and ‘off- To improve accuracy and coverage in the survey, we screened directly
target’ phenotypic components, we exploited the QMPIA to suppress in the primary assay three genomic libraries, two siRNA and one
as much as possible the latter contaminant. For this, we developed a endoribonuclease-prepared siRNA (esiRNA)15, corresponding to
244
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 ARTICLES

7–8 si/esiRNAs per gene for a total of 161,492 si/esiRNAs (see providing both strong and mild, but specific, phenotypes. A list of
Methods). This approach presents two major advantages. First, the 4,609 genes, 3,804 with a ‘strong’ and 805 with a ‘mild’ phenotype
high number of si/esiRNAs analysed by QMPIA reduces the impact was compiled (Supplementary Table 4; images and profiles available
of false negatives owing to disagreement between siRNAs. Second, at http://gwsdisplayer.mpi-cbg.de). Considering both the design of the
reproducibility of phenotypes is assessed already in the primary assay and the integration of endocytosis with many cellular processes,
screen for all genes (not just the hits). We acquired 12 images per such a large number was expected because, besides the ‘core’ endocytic
sample to obtain statistically significant data, accounting for a total of machinery, many genes have a direct or indirect role in endocytosis. For
,2.5 3 106 images. Quality controls for transfection efficiency, comparison with the common practices in the field, we estimated that
QMPIA and toxicity (see Methods) were included in each plate (see screening a single library (2–3 siRNAs per gene) with a typical
Supplementary Fig. 5). Image analysis and gene profile determination two-parameter assay and validating the hits with an independent set
required ,4.5 3 106 CPU hours of calculation on a 2,584-cores com- of si/esiRNAs (4–5) would yield only 336 genes (data not shown),
puter cluster. indicating that the large number of genes in the list results from (1) a
broader window of detection (many parameters) and (2) high number
Phenotypic clustering and gene scoring of si/esiRNAs screened rather than permissive criteria for selection.
As our aim was not to identify a few hit genes but rather to conduct a To estimate the reproducibility of the hits, we re-screened twice the
systems analysis, we exploited the multidimensionality of gene profiles. kinome-phosphatome (9,330 siRNAs, 1,459 genes), including a large
First, we clustered all genes of the human genome using mean-shift fraction of the gene hits (235). We observed 84% reproducibility for
clustering20 (see Supplementary Information). We identified 14 pheno- ‘strong’ and 72.5% for ‘mild’ phenotypes. Nevertheless, to assign a
typic cluster groups (Fig. 3) stable over a wide range of the algorithm better probabilistic value to the scores, the genes in the list will be re-
parameters, indicating that the endocytic system is under constraints screened (the web database will be updated). The silencing efficacy of
and changes between a discrete set of possible states in response to the libraries was assessed by the manufacturers by quantitative RT–PCR
perturbations. Second, genes were scored on the basis of two criteria, (qRT–PCR) on a set of siRNAs and we independently confirmed it by
phenotype amplitude (probability of x2) and presence of specific western blotting (see Methods and Supplementary Fig. 7).
phenotypic traits (defined as the cluster group profiles; ‘pheno-score’; The largest portion of scored genes (,34%, 1,593; P-value of
Supplementary Information), as shown in Supplementary Fig. 6, enrichment P 5 7 3 10234) encoded for components of metabolic

a 0.4
EGF TF Other b 0.4 EGF TF Other i 0.4
EGF TF Other j Parameters list
Cluster 1 Cluster 3
Normalized z-value

Normalized z-value
Normalized z-value

0.3 0.3 0.3 Cluster 14

Cluster 2 Cluster 4 Endosome parameter groups/cargo
0.2 0.2 0.2
0.1 0.1 0.1
G1. Number of endosomes (1).
0.0
G2. Cargo uptake (2).
0.0 0.0
G3. Endosome area (3).
–0.1 –0.1 –0.1
G4. Vesicle elongation (3).
–0.2 –0.2 –0.2
G5. Endosomal cargo concentration (3).
–0.3 –0.3 –0.3
G6. Endosomal cargo content (3).
–0.4 –0.4 –0.4
G7. Endosome distance from nucleus (3).
G1
G2
G3
G4
G5

G7
G1
G1

G6

G2

G4

G9
G2
G3
G4
G5

G7
G1

G10

G3

G5

G7
G1

G6

G2

G4

G9

G6
G2
G3
G4
G5

G7
G1

G3

G5

G7

G10
G6

G2

G4

G9

G6
G3

G5

G7
G6

G10

Parameter group Parameter group Parameter group G8. Endosome clustering (9).
c EGF TF Other d EGF TF Other Other endocytic parameters
0.4 0.4 G9. Background intensity (2).
Normalized z-value

Cluster 5 Cluster 7
Normalized z-value

0.3 Cluster 6 0.3 G10. Colocalization (2).

Cluster 8
k
0.2 0.2
0.1 0.1
0.0 0.0
–0.1 –0.1 Cluster Description Hits Total
–0.2 –0.2 group size
–0.3 –0.3
–0.4 –0.4 1 Selective up-regulation of TF endocytosis 932 5,258
G1
G1

G2
G3
G4
G5

G7
G1
G2
G3
G4
G5

G7
G1

G6

G2

G4

G9
G6

G2

G4

G9

G3

G5

G7
G3

G5

G7

G6

G10
G6

G10

Parameter group Parameter group 2 Selective down-regulation of TF endocytosis 88 586

e EGF TF Other f EGF TF Other Specific effect on subcellular localization: endosomes appear clustered in
0.4 0.4 3 801 2,990
Cluster 9 the cell centre
Normalized z-value

Normalized z-value

0.3 0.3 Cluster 11

0.2
Cluster 10
0.2 Specific effect on subcellular localization: endosomes appear dispersed in
0.1 0.1 4 260 1,505
the cell periphery
0.0 0.0
–0.1 –0.1 Opposite effects on EGF and TF endocytosis: EGF endocytosis is increased
5 224 1,158
–0.2 –0.2 and TF endocytosis is decreased
–0.3 –0.3 Opposite effects on EGF and TF endocytosis: EGF endocytosis is
–0.4 –0.4 6 143 670
decreased and TF endocytosis is increased
G1
G2
G3
G4
G5

G7
G1

G10
G6

G2

G4

G9
G3

G5

G7
G6

G1
G2
G3
G4
G5

G7
G1

G10
G6

G2

G4
G3

G5

G7
G9
G6

Parameter group Parameter group 7 Effects on endocytosis of both markers: increased EGF and TF endocytosis 137 631
g 0.4 EGF TF Other h 0.4 EGF TF Other 8 Effects on endocytosis of both markers: decreased EGF and TF endocytosis 799 2,256
Normalized z-value

Normalized z-value

0.3 Cluster 12 0.3 Cluster 13 9 Selective up-regulation of EGF endocytosis 178 695
0.2 0.2
0.1 0.1 10 Selective down-regulation of EGF endocytosis 324 1,291
0.0 0.0 Selective up-regulation of EGF endocytosis with accumulation of
–0.1 –0.1 11 271 1,012
endosomes in cell centre
–0.2 –0.2
–0.3 –0.3 12 Reduced TF endocytosis with endosomes accumulated in the cell centre 204 955
–0.4 –0.4 13 Selective increase in EGF endosomes number and elongation 38 140
G1

G10
G2
G3
G4
G5

G7
G1
G6

G2

G4

G9

G1
G3

G5

G7

G10
G2
G3
G4
G5

G7
G1
G6

G6

G2

G4

G9
G3

G5

G7
G6

Parameter group Parameter group 14 Increase in elongation of TF endosomes with mild increase of TF endocytosis 37 176

Figure 3 | Phenotypic cluster groups profiles. a–i, Mean-shift genomic and cluster group 8 (blue) present opposite effects on endocytosis of both
clustering of the gene profiles (see text and Supplementary Information) EGF and TF (both increased for cluster group 7, both decreased for cluster
yielded 14 cluster groups. Multiparametric profiles of the cluster groups are group 8). e, Cluster group 9 (black) and cluster group 10 (blue) present
plotted as in Fig. 1. The amplitude of the phenotype vectors was normalized selective and opposite effects on EGF endocytosis: increased and decreased,
to 1. a, Cluster group 1 (black) and cluster group 2 (blue) present selective respectively. f, Cluster group 11 presents increased EGF endocytosis and
and opposite effects, that is, increased vs decreased internal TF, respectively. endosome accumulation in the perinuclear region. g, Cluster group 12
b, Cluster group 3 (black) and cluster group 4 (blue) present specific and presents reduced TF endocytosis and endosomes accumulated in the cell
opposite effects on the subcellular localization of endosomes (‘distance from centre. h, Cluster group 13 presents a selective increase in EGF endosomes
nucleus’); that is, endosomes are either clustered in the perinuclear region number and their elongation. i, Cluster group 14 presents mainly an increase
(cluster group 3) or dispersed in the periphery (cluster group 4). c, Cluster in TF endosomes elongation. j, List of parameter-groups compiled as in
group 5 (black) and cluster group 6 (blue) is a pair of opposite clusters Fig. 1c. k, Table providing a summary of the most representative phenotypes
presenting each opposite effects with respect to EGF and TF endocytosis: in each group (‘Description’). The number of hits and the total size of each
increased EGF endocytosis with concomitantly decreased TF endocytosis group are also indicated.
(cluster group 5) and vice-versa (cluster group 6). d, Cluster group 7 (black)
245
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
and signalling pathways; 10% (468, P 5 1 3 102214) could bioinfor-
matically be assigned to the ‘core’ endocytic machinery, embracing
established (CLTC, DNM2, EEA1) and predicted genes on the basis of
trafficking-relevant functional domains, such as VPS9 (8/10;
P 5 6 3 1024), TBC (13/38; P 5 0.18), FYVE-finger (9/27;
P 5 0.27)21, PX (14/41; P 5 0.17) or BAR (4/15; P 5 0.59)22 domains.
Some genes (17%; 798; P 5 1 3 10236) were ‘unknown’ owing to
absence of curated annotations. We found several novel potential
endocytic regulators such as four TBC domain-containing proteins
(MGC16169 (also known as TBCK), TBC1D9, TBC1D9B,
TBC1D21), two FYVE-finger proteins (MTMR4 and WDFY1), and
a protein with a VPS9 domain (RINL). Notably, the list was enriched
in genes implicated in different human diseases (Kyoto Encyclopedia
of Genes and Genomes database; see Supplementary Information)
such as different cancers (for example, glioma (P 5 1.26 3 1028),
colorectal (P 5 5.53 3 1026), pancreatic cancer (P 5 8.92 3 1026)
etc.), neurodegenerative and metabolic disorders (for example,
dentatorubral-pallidoluysian atrophy; P 5 7.8 3 1023), Huntington
disease (P 5 0.026) and type II diabetes (P 5 3.7 3 1023).
Regulation by metabolic and signalling pathways
The cluster analysis provided insights into how endocytosis is inte-
grated with other cellular functions and made predictions concerning
the function of uncharacterized genes. Many cluster groups were
anti-correlated (Fig. 3a–e and Supplementary Fig. 8a, b), reflecting
positive and negative regulation on the endocytic system. For
example, ablation of genes in cluster groups 1 and 2 increased and
decreased the amount of internalized TF, respectively (Fig. 3a and
Supplementary Fig. 8a). Interestingly, genes of metabolic pathways
were highly enriched in cluster group 1, most probably reflecting a
compensatory response to a metabolic deficit by generally increasing
endocytosis of nutrients.
Another example of anti-correlated phenotypes is provided by
cluster groups 3 and 4 (Fig. 3b and Supplementary Fig. 8b), where
the endosomes were localized either closer to or more distant from
the nuclei, respectively. We found enrichment (P 5 0.022 in cluster
group 3 and P 5 2.3 3 1023 in cluster group 4) of components of the
actin and tubulin cytoskeleton (for example, RAC1, ROCK2,
CAPZB, microtubule and actin motors KIF20A and MYH9).
Interestingly, in cluster group 4 we found genes involved in cytokin-
esis such as CEP55 (ref. 23) and KIF20A (ref. 24), strengthening the
link with endocytosis25.
Cluster groups 8 and 10 contained several known endocytic com-
ponents. Cluster group 8 (decreased internal TF and EGF; Fig. 3d)
contained many genes encoding components required for endocytic
uptake, such as CLTC, DNM2 and phosphatidylinositol kinases/
phosphatases (INPP4A, INPP5B, PIB5PA, INPPL1; P 5 1.48 3 1027).
Cluster group 10 contained several endosomal regulators such as
Rabankyrin-5 (ANKFY1)26, Rabenosyn-5 (ZFYVE20)27 and Hrs
(HGS), indicating that EGF endocytosis is particularly susceptible to
depletion of PI(3)P effectors. On the basis of these results, we predict
that many new components of the endocytic core machinery are in
these cluster groups.
Signalling pathways exert distinct effects on the endocytic system.
Here we confirmed the activity of mitogen-activated protein kinase
(MAPK), Ca21, integrin/cell adhesion and mTOR signalling path-
ways19,28 and uncovered the activity of several new ones, TGF-b/
activin, Wnt and Notch. Some (MAPK, Ca21, mTOR, Wnt) had a
wide range of effects on the endocytic system (enriched in several
cluster groups; not shown) but others produced very specific effects
(enriched in single or opposite groups). For example, in cluster
groups 3 and 4 we found enrichment in the integrin/cell adhesion
pathway (P 5 0.012 in cluster group 3 and P 5 6.6 3 1024 in cluster
group 4; ITGA5, ITGB1, ITGA9, PAK1, MAPK9, MAPK1).
Interestingly, the Notch and TGF-b pathways were enriched
(P 5 0.026 and P 5 5.6 3 1023, respectively) in cluster group 8
(reduced total internal cargo for both EGF and TF). As we excluded
246
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
the possibility that the reduction in total internal cargo was due to
downregulation of surface receptors (Supplementary Fig. 9), we sug-
gest a new role of these pathways in the regulation of endocytic
uptake. We found for the Notch pathway NOTCH4, PSENEN and
PSEN2, MAML1 and RBPJ and for the TGF-b pathway several genes
encoding for activin receptors (ACVR1B, ACVR2A, ACVR2B and
ACVRL1).
In conclusion, the cluster analysis assigns new functions to meta-
bolic and signalling pathways and further predicts functions for new
genes in endocytosis.
Differential regulation of TF and EGF trafficking
To reveal general design principles of the endocytic system we con-
sidered all ,160,000 knockdown conditions as independent per-
turbed states of the system and determined correlations between
parameters. A first interesting aspect emerging from such analysis
was a high divergence in EGF and TF endocytosis. Despite the good
correlation between the size of EGF- and TF-positive endosomes
(Fig. 4a) and their intracellular localization (Fig. 4b), we found an
unexpected low correlation between the levels of total internal cargo
(Fig. 4c) of the two markers. These results strongly indicate that
although the molecular regulators of endosome size and distance
from nucleus are essentially shared by EGF and TF, the machineries
controlling internalization and sorting are largely distinct. We
exclude a major contribution of alternative non-clathrin-dependent
entry pathways29 under our experimental conditions, because deple-
tion of CLTC almost completely blocked the uptake of both cargo
markers (Supplementary Fig. 1a). One possible interpretation is that
EGF and TF may preferentially be internalized into endocytic vesicles
subjected to different regulation. For example, depletion of AP-2
complex specifically reduces TF but not EGF internalization
(Supplementary Fig. 10)30, whereas knockdown of other genes such
as PLEKHB2, CCDC33 and RCHY1 specifically downregulated EGF
internalization.
Interestingly, silencing of certain genes resulted in the dispersal of
EGF into small, peripheral endosomes (Fig. 4e), a pattern reminiscent
of APPL1-endosomes31. These endosomes are largely distinct from
canonical EEA1-positive early endosomes31 and act as signalling plat-
forms32. Knockdown of ZFYVE20, ANKFY1, PIK3R1, PIK3R2 along
with novel genes such as CCDC128 (also known as KLRAQ1), C12orf4,
MGC16169 (also known as TBCK) specifically affected EEA1-positive
endosomes (Fig. 4d–e) and increased the colocalization of EGF with
APPL1 (Fig. 4h–i). Conversely, knockdown of INPP5B and RABEP1
specifically affected APPL1-positive endosomes (Fig. 4d–f) and con-
comitantly decreased their colocalization with EGF (Fig. 4h–i).
Altogether, we identified new components that regulate the sorting
of EGF to APPL1-positive endosomes.
New design principles of the endocytic system
Our analysis also showed that various endosomal parameters are
significantly correlated and, therefore, functionally linked. For both
EGF (Fig. 5a–c) and TF, we observed a negative correlation between
mean size and number of endosomes, a positive correlation between
the number of endosomes and their mean distance from nucleus and
a negative correlation between endosome size and distance from
nucleus. Total internal cargo and cargo concentration per endosome
(Fig. 5d) as well as number of endosomes (Fig. 5e) were also strongly
correlated positively. In contrast, there was no correlation between
total internal cargo and endosome size (Fig. 5f) and their distance
from nucleus (not shown).
The analysis of pair correlations indicate that the endocytic para-
meters are functionally linked, but does not infer cause–effect rela-
tionships. To infer directionality in these relationships and learn
general rules underlying the organization of the endocytic system,
we performed a Bayesian network analysis. The structure of Bayesian
networks is a directed acyclic graph where nodes denote the variables
of interest and edges correspond to logical dependencies between
NATURE | Vol 464 | 11 March 2010 ARTICLES

a c d e f
RABEP1/EGF

Total internal cargo TF

Correlation coefficient = 0.51 40 Correlation coefficient = 0.15
30
30
Mean size TF

20 20
10 10
0 0
–10
–10
–20 20 μm Control/EGF 20 μm CCDC128/EGF 20 μm
–20 –15 –10 –5 0 5 10 15 20 25 –20 –10 0 10 20 30 40 RABEP1/EEA1
Mean size EGF Total internal cargo EGF

b g h
30 Correlation coefficient = 0.86 0.6 Control 0.6 CCDC128

colocalization
colocalization
0.5

Percentage
Percentage 0.5
Distance from

20 0.4 0.4
nucleus TF

10 0.3 0.3
0.2 0.2
0 20 μm 20 μm
0.1 0.1 Control/EEA1 CCDC128/EEA1 20 μm
–10 0.0 0.0
–20 EGF>EEA1 EGF>APPL EGF>EEA1 EGF>APPL RABEP1/APPL
–30
i
–20 –10 0 10 20 30 0.6 RABEP1
colocalization

0.5
Percentage

Distance from nucleus EGF 0.4

0.3
0.2
0.1
0.0
EGF>EEA1 EGF>APPL Control/APPL1 CCDC128/APPL1

Figure 4 | Regulators of transport of cargo to EEA1 vs APPL endosomes. reduced the endosomal staining of APPL1 determining an increase of the
a–c, Correlation between endosome size (a), endosome distance from nuclei cytoplasmic staining. Insets show higher magnifications to visualize better
(b) and total internal cargo (c) for both EGF and TF. The data are the endosomal (co)-localization. Merge insets show always EGF overlaid on
represented as scatter plots of z-score values. The density of the data point is the indicated endosomal marker. Scale bars, 10 mm. White arrows indicate
indicated in blue for low- and in red for high-density values. Pearson the presence of EGF in the APPL1 and EEA1-positive endosomes in the
correlation values (correlation coefficient) are indicated for each scatter plot. merge images. Green arrows indicate the presence of EEA1 on the indicated
d–f, Examples of knockdown of genes affecting the accumulation of EGF in endosomes in the EGF and EEA1 insets. Light blue arrows indicate the
APPL1-positive endosomes. HeLa cells silenced for the indicated genes were presence of APPL1 on the indicated endosomes in the EGF and APPL1
fixed and stained for the indicated markers. d, Mock-transfected control insets. g–i, Quantification of EGF colocalization with the markers indicated.
cells. e, Silencing of CCDC128 selectively affected EEA1 endosomes EGF-to-APPL colocalization is increased upon silencing of CCDC128 (h) and
(enlarged endosomes in the perinuclear region), determining the decreased upon silencing of RABEP1 (i) compared to control cells (g). Error
accumulation of EGF in APPL1-positive endosomes. f, Silencing of RABEP1 bars indicate standard error of the mean (s.e.m.).

them33. We constructed two Bayesian networks for EGF and TF the endocytic system. The dependencies between the triplet endo-
(Fig. 5g–h), where the nodes correspond to the ‘key’ phenotypic some distance from nucleus–number–size reflect a progression from
parameters and the edges to their pair correlation. The two networks many small peripheral to few large perinuclear endosomes. Genes
shared a similar structure. Surprisingly, for both EGF and TF the involved in protein ubiquitination were found enriched among the
parameter measuring endosome distance from nucleus was found regulators of this process (cluster group 4 P 5 0.032). We found
to influence other parameters, indicating a key regulatory role in RBX1 and DDB1, encoding components of a multiprotein complex

a d
30 Correlation coefficient = –0.61 30 Correlation coefficient = 0.78
Total internal

20 20
g h
Mean size

0.86
10
cargo

10
0 0
–10 EGF TF
–10
–20 Total Total
–20 Distance Distance
internal internal
–20 –10 0 10 20 –20 –10 0 10 20 30 from nucleus from nucleus
cargo cargo
Number of endosomes Cargo concentration
b e
30 Correlation coefficient = 0.42
20 Correlation coefficient = 0.61 0.16 0.02
Distance from

Total internal

20 0.61 –0.33 0.27

0.79 0.89 0.38
10
nuclei

10
cargo

0 0
–10 –10 Cargo 0.79 Cargo Endosome Cargo 0.88 Cargo Endosome
–20 –20 concentration content number concentration content number
–30
–20 –10 0 10 20 –20 –10 0 10 20 30
Number of endosomes Number of endosomes –0.05
c f 30
–0.43 –0.61 0.03 0.3 –0.68
30 Correlation coefficient = –0.36 Correlation coefficient = –0.27
Distance from

Total internal

20 20
10 10
nuclei

Endosome Endosome
cargo

0 0 size size
–10 –10
–20 –20
–30 0.51
–20 –10 0 10 20 30 –20 –10 0 10 20 30
Mean size Endosome size

Figure 5 | Design principles of the endocytic system. a–f, Analysis of endosome number (e) and weak correlation between total internal cargo and
correlations between different endocytic parameters. The data are endosome size (f). g–h, Bayesian networks of the endocytic parameters for
represented as in Fig. 4 and only correlations between EGF parameters are the genome-wide analysis of EGF and TF endocytosis. The nodes represent
shown: negative correlation between endosome size and number (a), positive individual endocytic parameters measuring EGF or TF endocytosis and the
correlation between number and distance from nuclei (b), negative numbers indicate the Pearson correlation values. In black the direct
correlation between endosome size and endosomes distance from nuclei correlations shared by the two networks, in red (present) and grey (absent)
(c), positive correlation between total internal cargo and mean endosome correlations that present differences between the two networks.
concentration (d), positive correlation between total internal cargo and
247
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
including cullins catalysing ubiquitin ligation34 and other genes such
as FBXO22, FBXO30, FBXO46, FBXO6, USP31 and HECTD3, indi-
cating a new role of protein ubiquitination in the endosomal spatio-
temporal progression.
However, the two networks also presented some interesting differ-
ences. The negative dependency of endosome cargo content on dis-
tance from nucleus reflects a progressive concentration of EGF in
endosomes, in contrast to TF that is continuously removed and
recycled to the surface. A comparison of the triplet concentration–
size–content revealed some unexpected properties of the endocytic
system. The TF network has a structure typical of passively sorted
membrane proteins, where cargo content depends on size and con-
centration of cargo per endosome. In contrast, the mean concentra-
tion of EGF in the endosomes is adjusted so that the mean total
content of EGF per endosome remains constant, independently of
endosome size (negative dependence of the size on the concentration
and no correlation between size and content). Such a mechanism
cannot be explained only by sorting of EGF into intra-luminal vesicles9,
because (1) this is inconsistent with the high correlation between size of
EGF and TF endosomes and (2) the knockdown of HGS (Hrs) and
TSG101 (ref. 9) does not alter the size–concentration–content relation-
ship. These data rather indicate the existence of a quasi-deterministic
mechanism tightly coupling fusion and fission events to regulate the
concentration of EGF in each endosome.
Finally, the cell exerts a tighter control on the number of EGF- than
TF-positive endosomes, as indicated by the stronger dependence on
total internal cargo (corrEGF 5 0.61 vs corrTF 5 0.16). In other words,
to increase cargo capacity the system in addition to concentrating
EGF in endosomes (as for TF) also generates new endosomes.
Altogether, these findings indicate that the cell regulates the number,
size, loading and, therefore, the intracellular density of EGF-bearing
endosomes. These findings have profound implications for the role of
endosomes in signalling31,35,36.
Discussion
As a first step towards a systems analysis of endocytosis, we profiled
the human genome relative to a set of 58 parameters that quantita-
tively describe endocytic features. From this survey we could attri-
bute new functions to known and unknown (798) genes with respect
to EGF and TF endocytosis. Furthermore, we revealed unexpected
properties of the endocytic system. First, we uncovered the effect of
various signalling pathways, notably integrin/cell adhesion, TGF-b,
Notch and Wnt, on both TF and EGF endocytosis. The activity of
these pathways depends on endocytosis37–39, but our data further
indicate that they exert a feedback mechanism on the endocytic
machinery. Interestingly, TGF-b, Notch and Wnt pathways have
been implicated in the regulation of epithelial–mesenchymal trans-
ition (EMT)40, cell-fate determination41 and tissue morphogenesis42
and given the established role of endocytosis in these events2,43, we
propose that part of their morphogenetic activity depends on their
effects on the endocytic pathway. Second, a surprising result was the
low correlation between genes regulating TF and EGF endocytosis
(Figs 5g–h and 4a–c). Differences in the molecular machinery regu-
lating cargo uptake and recycling have been reported30, but such a
degree of diversity is unprecedented. Third, our analysis provided
insights into the design principles underlying the endocytic system.
The fact that the endosome distance from the nucleus is a key
regulatory parameter demonstrates directly that the spatio-temporal
progression of endosomes observed in living cells12 is not an epiphe-
nomenon but has a fundamental role in the function of the endocytic
system. The many components of the actin and tubulin cytoskeleton
identified in the screen (cluster groups 3 and 4) provide insights into
the molecular regulation underlying this mechanism.
An interesting aspect emerging from the comparison of the EGF
and TF Bayesian networks is that the cell specifically regulates the
number of EGF-positive endosomes, and strictly couples size and
concentration of cargo in endosomes. Given that endosomes can
248
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
act as intracellular signalling platforms32,35, the system might regulate
the concentration of EGF and the number of EGF-positive endo-
somes to quantitatively control signalling responses, for example,
for different cell-fate decisions, as shown for the MAPK network44.
Regulating the density of receptor tyrosine kinases, and number and
distribution of endosomes may be necessary to calibrate the concen-
tration of signalling molecules and their signalling output. In iden-
tifying genes that selectively affect cargo transport to APPL1-positive
endosomes, we unveiled a sorting mechanism to distribute cargo to
this compartment. This is the beginning of a more detailed molecular
characterization of APPL1 endosomes, their regulation of transport
and their role in signal propagation32,35. However, a limitation of our
approach remains the poor identification of highly homologous and
functionally redundant genes (for example, RAB5A, B, C and so on).
This could be overcome by combinations of several siRNAs, as for
CBL, CBLB and CBLC (Supplementary Table 4).
This exploratory endeavour required uncommon resources for a
primary screen. Nevertheless, it was necessary to establish the condi-
tions for streamlining the technology and rendering such genomic
profiling easier and more affordable in the future. Several genes iden-
tified in the screen are associated with human diseases, confirming the
fundamental role of endocytosis in the pathogenesis of many human
diseases and providing novel potential drug targets. It will be important
to determine which parameters of the QMPIA might act as physio-
pathological indicators reflecting disease-relevant alterations. When
applied to primary cells and disease model systems, the screening plat-
form described here will reveal its true potential for drug discovery,
through the identification of novel therapeutic targets, new mechan-
isms of action, and for discerning therapeutic effects of small molecules
from toxic ones.
METHODS SUMMARY
RNAi screen for endocytosis. HeLa cells were reverse transfected with three
genomic libraries (Ambion, Qiagen and a custom-made esiRNA library15) using
20 nM siRNAs or 25 ng per well esiRNAs. Seventy-two hours after transfection,
cells were incubated with 100 ng ml21 EGF-Alexa 488 and 5 mg ml21 Transferrin–
Alexa 647 (Molecular Probes) in serum-free medium for 10 min at 37 uC before
fixation with formaldehyde. Thereafter, nuclei and cytoplasm were stained with
0.4 mg ml21 DAPI and 0.2 mM SYTO42 blue (Molecular Probes).
Automated image acquisition. Triple-colour images were acquired in a fully
automated and unbiased manner using a spinning disk confocal microscope
(OPERA, Evotec Technologies-PerkinElmer) and a 340 water immersion
objective (NA 5 0.9). Twelve images per well were collected to obtain a sufficient
number of cells for reliable statistical analysis. Image correction and image
analysis were performed using custom designed image analysis software
(MotionTracking; see Supplementary Information).
EEA1-APPL1 endosome dissection of EGF trafficking. Genes of interest (78)
were screened to detect alterations in the EEA1- or APPL1-positive endosomes
with respect to EGF trafficking. Following siRNA transfection and EGF–Alexa
488 internalization, cells were stained with mouse monoclonal anti-EEA1 (BD
Biosciences; Pharmingen) and rabbit polyclonal anti-APPL1 (ref. 31) antibodies
and quadruple-colour images were acquired as described (Methods).
Detection of surface receptors. To exclude changes in the level of surface recep-
tors upon downregulation of Notch and TGF-b signalling pathways, cells were
transfected with si/esiRNAs targeting these genes and 1,034 randomly selected
si/esiRNAs, incubated with 100 ng ml21 EGF–Alexa 488 and 5 mg ml21 TF–Alexa
647 (Molecular Probes) in serum-free medium for 30 min on ice, fixed and
stained with DAPI-SYTO42 as described earlier. Signals were enhanced by stain-
ing with a rabbit anti-Alexa 488 (Molecular Probes) and mouse monoclonal
antibodies against the ectodomain of human TFRC (BD Biosciences;
Pharmingen) without permeabilization. Triple-colour images (ten per well) were
collected with a 320 water immersion objective (NA 5 0.7) as described earlier.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 1 December 2008; accepted 17 December 2009.
Published online 28 February; corrected 11 March 2010 (see full-text HTML version
for details).
1. Polo, S. & Di Fiore, P. P. Endocytosis conducts the cell signaling orchestra. Cell 124,
897–900 (2006).
NATURE | Vol 464 | 11 March 2010
2. Dudu, V., Pantazis, P. & González-Gaitán, M. Membrane traffic during embryonic
development: epithelial formation, cell fate decisions and differentiation. Curr.
Opin. Cell Biol. 16, 407–414 (2004).
3. Burgdorf, S. & Kurts, C. Endocytosis mechanisms and the cell biology of antigen
presentation. Curr. Opin. Immunol. 20, 89–95 (2008).
4. Pagano, R. E., Puri, V., Dominguez, M. & Marks, D. L. Membrane traffic in
sphingolipid storage diseases. Traffic 1, 807–815 (2000).
5. Gruenberg, J. & van der Goot, F. G. Mechanisms of pathogen entry through the
endosomal compartments. Nature Rev. Mol. Cell Biol. 7, 495–504 (2006).
6. Baskys, A., Bayazitov, I., Zhu, E., Fang, L. & Wang, R. Rab-mediated endocytosis:
linking neurodegeneration, neuroprotection, and synaptic plasticity? Ann. NY
Acad. Sci. 1122, 313–329 (2007).
7. Marsh, M. & McMahon, H. T. The structural era of endocytosis. Science 285,
215–220 (1999).
8. Stenmark, H. Rab GTPases as coordinators of vesicle traffic. Nature Rev. Mol. Cell
Biol. 10, 513–525 (2009).
9. Piper, R. C. & Katzmann, D. J. Biogenesis and function of multivesicular bodies.
Annu. Rev. Cell Dev. Biol. 23, 519–547 (2007).
10. Perrimon, N. & Mathey-Prevot, B. Applications of high-throughput RNA
interference screens to problems in cell and developmental biology. Genetics 175,
7–16 (2007).
11. Friedman, A. & Perrimon, N. A functional RNAi screen for regulators of receptor
tyrosine kinase and ERK signalling. Nature 444, 230–234 (2006).
12. Rink, J., Ghigo, E., Kalaidzidis, Y. & Zerial, M. Rab conversion as a mechanism of
progression from early to late endosomes. Cell 122, 735–749 (2005).
13. König, R. et al. A probability-based approach for the analysis of large-scale RNAi
screens. Nature Methods 4, 847–849 (2007).
14. Kulkarni, M. M. et al. Evidence of off-target effects associated with long dsRNAs in
Drosophila melanogaster cell-based assays. Nature Methods 3, 833–838 (2006).
15. Kittler, R. et al. Genome-wide resources of endoribonuclease-prepared short
interfering RNAs for specific loss-of-function studies. Nature Methods 4, 337–344
(2007).
16. Birmingham, A. et al. 39 UTR seed matches, but not overall identity, are associated
with RNAi off-targets. Nature Methods 3, 199–204 (2006).
17. Cullen, B. R. Enhancing and confirming the specificity of RNAi experiments. Nature
Methods 3, 677–681 (2006).
18. Echeverri, C. J. et al. Minimizing the risk of reporting false positives in large-scale
RNAi screens. Nature Methods 3, 777–779 (2006).
19. Pelkmans, L. et al. Genome-wide analysis of human kinases in clathrin- and
caveolae/raft-mediated endocytosis. Nature 436, 78–86 (2005).
20. Fukunaga, K. & Hostetler, L. The estimation of the gradient of a density function,
with applications in pattern recognition. IEEE Trans. Inf. Theory 21, 32–40 (1975).
21. Hayakawa, A., Hayes, S., Leonard, D., Lambright, D. & Corvera, S. Evolutionarily
conserved structural and functional roles of the FYVE domain. Biochem. Soc. Symp.
74, 95–105 (2007).
22. Habermann, B. The BAR-domain family of proteins: a case of bending and
binding? EMBO Rep. 5, 250–255 (2004).
23. Carlton, J. G. & Martin-Serrano, J. Parallels between cytokinesis and retroviral
budding: a role for the ESCRT machinery. Science 316, 1908–1912 (2007).
24. Hill, E., Clarke, M. & Barr, F. A. The Rab6-binding kinesin, Rab6-KIFL, is required
for cytokinesis. EMBO J. 19, 5711–5719 (2000).
25. Baluska, F., Menzel, D. & Barlow, P. W. Cytokinesis in plant and animal cells:
endosomes ’shut the door’. Dev. Biol. 294, 1–10 (2006).
26. Schnatwinkel, C. et al. The Rab5 effector Rabankyrin-5 regulates and coordinates
different endocytic mechanisms. PLoS Biol. 2, e261 (2004).
27. Nielsen, E. et al. Rabenosyn-5, a novel Rab5 effector, is complexed with hVPS45
and recruited to endosomes through a FYVE finger domain. J. Cell Biol. 151,
601–612 (2000).
28. Galvez, T. et al. siRNA screen of the human signaling proteome identifies the
PtdIns(3,4,5)P3-mTOR signaling pathway as a primary regulator of transferrin
uptake. Genome Biol. 8, R142 (2007).
29. Sigismund, S. et al. Clathrin-independent endocytosis of ubiquitinated cargos.
Proc. Natl Acad. Sci. USA 102, 2760–2765 (2005).
30. Motley, A., Bright, N. A., Seaman, M. N. & Robinson, M. S. Clathrin-mediated
endocytosis in AP-2-depleted cells. J. Cell Biol. 162, 909–918 (2003).
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
31. Miaczynska, M. et al. APPL proteins link Rab5 to nuclear signal transduction via an
endosomal compartment. Cell 116, 445–456 (2004).
32. Schenck, A. et al. The endosomal protein Appl1 mediates Akt substrate specificity
and cell survival in vertebrate development. Cell 133, 486–497 (2008).
33. Pearl, J. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
(Morgan Kaufmann, 1988).
34. Petroski, M. D. & Deshaies, R. J. Function and regulation of cullin-RING ubiquitin
ligases. Nature Rev. Mol. Cell Biol. 6, 9–20 (2005).
35. Miaczynska, M., Pelkmans, L. & Zerial, M. Not just a sink: endosomes in control of
signal transduction. Curr. Opin. Cell Biol. 16, 400–406 (2004).
36. Kermorgant, S. & Parker, P. J. Receptor trafficking controls weak signal delivery: a
strategy used by c-Met for STAT3 nuclear accumulation. J. Cell Biol. 182, 855–863
(2008).
37. Huang, S. S. & Huang, J. S. TGF-b control of cell proliferation. J. Cell. Biochem. 96,
447–462 (2005).
38. Le Borgne, R. Regulation of Notch signalling by endocytosis and endosomal
sorting. Curr. Opin. Cell Biol. 18, 213–222 (2006).
39. Gagliardi, M., Piddini, E. & Vincent, J. P. Endocytosis: a positive or a negative
influence on Wnt signalling? Traffic 9, 1–9 (2008).
40. Heldin, C. H., Landström, M. & Moustakas, A. Mechanism of TGF-b signaling to
growth arrest, apoptosis, and epithelial-mesenchymal transition. Curr. Opin. Cell
Biol. 21, 166–176 (2009).
41. Blank, U., Karlsson, G. & Karlsson, S. Signaling pathways governing stem-cell fate.
Blood 111, 492–503 (2008).
42. Weinmaster, G. Notch signal transduction: a real rip and more. Curr. Opin. Genet.
Dev. 10, 363–369 (2000).
43. Entchev, E. V. & González-Gaitán, M. A. Morphogen gradient formation and
vesicular trafficking. Traffic 3, 98–109 (2002).
44. Santos, S. D., Verveer, P. J. & Bastiaens, P. I. Growth factor-induced MAPK
network topology shapes Erk response determining PC-12 cell fate. Nature Cell
Biol. 9, 324–330 (2007).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We acknowledge T. Galvez and G. Marsico, members of the
Zerial group for discussions and scientific support. We are particularly indebted to
K. Korn who developed transfection protocols and organized various preparatory
steps and logistics before the screen and E. Krausz for the management of the
HT-TDS, the MPI-CBG screening facility. We thank J. Schmitt, A. Lohmann,
S. Christ, N. Tomschke, A. Niederlein, J. Wagner, M. Gierth, E. Krausz for technical,
robotics and computational assistance, and M. Boes and J. Oegema for IT support
that made the large-scale computational analysis possible. We acknowledge
I. C. Baines, T. Galvez, J. Howard, T. Hyman, M. McShane, G. O’Sullivan and
P. Tomancak for comments on the manuscript. This work was financially supported
by the Max Planck Society (MPG) and by the systems biology network HepatoSys
of the German Ministry for Education and Research BMBF, the DFG and the
Gottlieb Daimler und Karl Benz Stiftung. This work is also part of the project
‘Endotrack’, which received research funding from the European Community’s
Sixth Framework Programme.
Author Contributions C.C., Y.K. and M.Z. conceived the project and M.Z. directed
it. C.C. and D.K. developed the endocytosis assay under the guidance of M.Z. and
E.F.; M.S. performed the screen and image acquisition. Y.K. developed the QMPIA
with the help of C.C. and J.C.R. and performed all computational analysis of the
screening data. N.S. developed the clustering and pheno-score algorithms and
together with C.C. performed the analysis of the clusters. C.R.B., under the
supervision of B.H., performed the si/esiRNAs remapping and bioinformatics
analysis. CC. and Y.K performed the Bayesian Network analysis. F.B. provided the
esiRNA library. R.H., under the supervision of M.S.M. and W.E.N., provided IT
support for the use of the computer cluster located at the TUD. C.C.,Y.K. and M.Z.
wrote the manuscript.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to M.Z.
(zerial@mpi-cbg.de).
249
doi:10.1038/nature08779
METHODS
siRNA and esiRNA libraries and remapping to RefSeq 28. Three genomic
RNAi libraries, two commercial siRNA libraries (Ambion Silencer Genome wide
siRNA Library V.3 and Qiagen Human Whole Genome siRNA Library V.1) and
a custom-made esiRNA library15 were screened. The Ambion library included
51,880 siRNAs targeting 22,527 genes in the human genome and was designed
partly on the basis of the RefSeq database Release 15 and partly on the ENSEMBL
database. The Qiagen library, which included 91,956 siRNAs targeting 22,732
genes, was entirely designed on the basis of the RefSeq database (kinase/
phosphatase and gpcr subsets: Refseq Release 18, Druggable Genome Set:
Refseq Release 17, the rest of the Whole Genome Set and the Predicted Genes
Set RefSeq Release 12). The esiRNA library was prepared as described previ-
ously13 and included 17,188 esiRNAs targeting 16,722 human genes, designed on
the basis of the ENSEMBL (Release 36) database. Ambion and Qiagen validated
768 and 3,073 siRNAs by qRT–PCR, respectively. Out of these, 740 (Ambion)
and 3,072 (Qiagen) were shown to be efficient (.70% reduction) at silencing
their corresponding target mRNAs, providing an overall estimated efficiency of
.95% for the Ambion and .99% for the Qiagen libraries, respectively. The
sequences of all 3 libraries were remapped on the RefSeq Release 28 database by
BLAST. The sequence of an siRNA or esiRNA was mapped to a given gene on the
basis of the following three conditions: (1) 19/19 bases match on a region on the
mRNA, (2) matching with all transcripts (including differentially spliced
mRNAs) of a given gene, (3) absence of matching with any other transcripts
from other genes. All siRNAs/esiRNAs that did not meet these conditions were
discarded from further analysis. After re-mapping to RefSeq Release 28, the 3
libraries covered 33,223 genes of which 28,279 encode for proteins or contain an
open reading frame, corresponding to 62% coverage of the transcriptome and
73% of the proteome.
HT RNAi transfection and endocytic assay. HeLa cells were reverse transfected
with 20 nM siRNAs or 25 ng per well esiRNAs (final concentration) with 0.2 ml per
well Oligofectamine (Invitrogen) in a total volume of 50 ml. For HT-transfection a
‘Freedom Evo’ pipetting robot (Tecan) equipped with a 384-fixed-needle-head,
an 8-needle-pipettor and a gripper was used under a customized biological safety
cabinet. A fully automated workflow prepared transfection complexes in 384 wells
at a throughput of 5 plates/50 min (manuscript submitted). For cell seeding and
dispensing the assay solutions the MTP dispensers WellMate and Multidrop 384
(Thermo Scientific) were used; wash procedures were performed using the Power
Washer 384 (Tecan).
Briefly, 10 ml of 200 nM siRNAs or 5 ng ml21 esiRNAs were mixed with 10 ml of
pre-diluted Oligofectamine (40 ml ml21 OptiMem) for 15 min to allow complex
formation. Subsequently, 10 ml of this mixture were placed in 384-well
CellCarrier plates (Evotec Technologies). HeLa cells (1000 cells per well, counted
with CASY Model TT, Schärfe System) were seeded in 40 ml per well. Plates were
©2010 Macmillan Publishers Limited. All rights reserved
incubated for 72 h at 37 uC and 5% CO2. Thereafter, the cells were incubated with
100 ng ml21 EGF–Alexa 488 and 5 mg ml21 transferrin–Alexa 647 (Molecular
Probes) in serum-free medium for 10 min at 37 uC. Cells were subsequently
washed with PBS and fixed with 3.7% formaldehyde. Nuclei and cytoplasm were
stained with 0.4 mg ml21 DAPI and 0.2 mM SYTO42 blue (Molecular Probes).
For the analysis of protein downregulation, cells were grown in 96-well plates
and transfected with 20 nM siRNAs (final concentration) with 0.3 ml per well
Oligofectamine (Invitrogen) in a total volume of 100 ml. After 72 h total protein
extracts were prepared and analysed by quantitative western blotting using
antibodies against EEA1, ZFYVE20 and ANKFY1.
Automated spinning disk confocal imaging. Twelve images per well were col-
lected with a 340 water immersion objective (NA 5 0.9) in a fully automated and
unbiased manner with a spinning disk confocal microscope (OPERA, Evotec
Technologies-Perkin Elmer). Triple-colour images were acquired in two sequen-
tial exposures by three different CCD cameras: the Alexa 488 signal was collected
in a first exposure using 488 nm laser light and a 540/75 band-pass filter; Alexa 647
and DAPI-SYTO 42 blue signals were collected in a second exposure using 635 nm
and 405 nm laser light and with 700/90 and 450/50 band-pass filters, respectively.
Images were subsequently corrected (MotionTracking; see Supplementary
Information) for uneven field illumination and chromatic aberrations according
to reference images. For reference, images of 2.5 mm multicolour beads, and
images of reference dyes and dark field images were collected using the Opera
Adjustment Plate (Evotec Technologies-Perkin Elmer).
Three-dimensional imaging of the endosome distribution of the cell was per-
formed on a small set of 27 siRNAs (targeting 10 genes of the human genome).
Using the same microscope and automated imaging protocols as above 10 focal
planes with 0.5 mm steps were acquired.
EEA1-APPL1 endosome dissection of EGF trafficking. A smaller set of 78 genes
was further screened to detect alterations in the EEA1- or APPL1-positive endo-
somes with respect to EGF trafficking. For this, cells were transfected, EGF–Alexa
488 was internalized and cells were fixed as above. Cells were then permeabilized
with 0.1% saponin and 0.2% fish gelatin (as blocking agent) and subsequently
stained with mouse monoclonal anti-EEA1 (BD Biosciences Pharmingen) and
rabbit polyclonal anti-APPL1 (ref. 31) antibodies. Fluorescently conjugated goat
anti-rabbit–Alexa 568 and goat anti-mouse–Alexa 647 secondary antibodies
(Molecular Probes) revealed the antigen signals and DAPI-SYTO 42 blue stained
the nuclei and cytoplasm. Quadruple-colour confocal images were collected
(OPERA, Evotec Technologies-PerkinElmer) in two sequential exposures:
Alexa 488 and Alexa 647 signals were collected in a first exposure using 488 nm
and 635 nm laser light and with 540/75 and 690/50 band pass filters; Alexa 568 and
DAPI-SYTO 42 blue signals were collected in a second exposure using 405 nm and
561 nm laser light and with 450/50 and 605/40 band-pass filters, respectively.
Reference images were acquired and image correction performed as before.
 Vol 464 | 11 March 2010 | doi:10.1038/nature08756

ARTICLES
The primary transcriptome of the major
human pathogen Helicobacter pylori
Cynthia M. Sharma1, Steve Hoffmann2, Fabien Darfeuille3,4, Jérémy Reignier3,4, Sven Findeiß2, Alexandra Sittka1,
Sandrine Chabas3,4, Kristin Reiche5, Jörg Hackermüller5, Richard Reinhardt6, Peter F. Stadler2,5,7,8,9 & Jörg Vogel1,10
Genome sequencing of Helicobacter pylori has revealed the potential proteins and genetic diversity of this prevalent human
pathogen, yet little is known about its transcriptional organization and noncoding RNA output. Massively parallel cDNA
sequencing (RNA-seq) has been revolutionizing global transcriptomic analysis. Here, using a novel differential approach
(dRNA-seq) selective for the 59 end of primary transcripts, we present a genome-wide map of H. pylori transcriptional start
sites and operons. We discovered hundreds of transcriptional start sites within operons, and opposite to annotated genes,
indicating that complexity of gene expression from the small H. pylori genome is increased by uncoupling of polycistrons and
by genome-wide antisense transcription. We also discovered an unexpected number of ,60 small RNAs including the
e-subdivision counterpart of the regulatory 6S RNA and associated RNA products, and potential regulators of cis- and
trans-encoded target messenger RNAs. Our approach establishes a paradigm for mapping and annotating the primary
transcriptomes of many living species.

About 50% of all humans are infected with Helicobacter pylori, a micro- Fig. 1). We primarily analysed H. pylori strain 26695 grown to mid-
aerophilic, Gram-negative e-proteobacterium that thrives in the acidic logarithmic phase (ML2/1 libraries), or under acid stress at pH 5.2
environment of the stomach, and is associated with severe inflam- (AS2/1) resembling the host environment. To increase data depths,
mation, peptic ulcer disease and gastric cancer1,2. The 1.67 megabase bacteria were also grown in contact with responsive gastric epithelial
pairs genome of H. pylori strain 26695 carries 1,576 open reading cells (AG2/1) or non-responsive liver cells (HU2/1), or in cell
frames (ORFs), but surprisingly few genes of transcriptional regula- culture medium alone (PL2/1) (Supplementary Fig. 2). Following
tors3,4. Moreover, only ,55 transcriptional start sites (TSS; compiled in 454 pyrosequencing, ,217 million bases of cDNA (Supplementary
Supplementary Table 1) were known in this important human patho- Table 2) were mapped to the H. pylori chromosome (Fig. 1a).
gen, and the principal organization of its transcriptome remained to be dRNA-seq confirmed the known acid induction10–12 of major
elucidated. Small noncoding RNAs (sRNAs), an otherwise abundant H. pylori virulence loci such as the urease (ure) operon or the cag
class of post-transcriptional regulators in bacteria5,6, also seemed to be pathogenicity island (Fig. 1b), as evident from a three- to fourfold
lacking in H. pylori, perhaps reflecting the absence of the common higher cDNA coverage of the ureA, ureI or cag16 transcripts in the
sRNA chaperone, Hfq, in all e-proteobacteria7. AS2 vs ML2 libraries (Supplementary Table 3). Crucially, the
Global analyses using tiling arrays and RNA-seq have provided exonuclease treatment revealed TSS by means of a characteristic redis-
invaluable insights into the gene expression patterns and sRNA output tribution of a gene’s cDNAs towards a sawtooth-like profile with an
of diverse bacteria, including several pathogens8,9. Yet, these studies elevated sharp 59 flank. For example, the AS1 cDNAs clustered
did not directly detect TSS owing to a lack of discrimination between towards the nuclease-protected primary 59 end of ureA mRNA, match-
primary and processed transcripts. Specifically, primary transcripts ing the known TSS13,14, whereas the AS2 cDNAs were uniformly dis-
including most mRNAs and sRNAs carry a 59 tri-phosphate tributed over ureA (Fig. 1b). Thus, the selective destruction of
(59PPP) group, whereas processed transcripts such as mature ribo- processed 59P transcripts enriches TSS-specific cDNAs (Supplemen-
somal and transfer RNAs (rRNA, tRNA) have a 59 mono-phosphate tary Fig. 2). Similar patterns mark the cagA TSS15 and reveal primary
(59P). Using a novel dRNA-seq approach to selectively identify native and processed 59 ends of tRNA-Phe (Fig. 1c). Altogether, 69 H. pylori
59PPP ends we present a single-nucleotide resolution map of the TSS determined by independent methods were matched by dRNA-seq
primary transcriptome of H. pylori and unravel the unexpectedly with high accuracy (Fig. 1d).
complex RNA output from this small and compact genome.
Genome-wide TSS and operon maps
Differential RNA-seq Annotation of 59 ends enriched in (1) vs (2) libraries and satisfying
Our dRNA-seq approach discriminates primary from processed 59 other plausible criteria identified a total of 1,907 TSS (Supplementary
ends by sequencing differential cDNA library pairs: one library Table 4). These were grouped into five categories (Fig. 2a): primary
denoted (2) from untreated total bacterial RNA, and the other TSS having the most cDNAs within #500 bp upstream of annotated
(1) enriched for primary transcripts by terminator exonuclease mRNA start codons or processed sRNAs; secondary TSS associated
treatment that degrades 59P but not 59PPP RNA (Supplementary with the same gene but with fewer cDNAs; internal TSS within an
1
Max Planck Institute for Infection Biology, RNA Biology Group, D-10117 Berlin, Germany. 2University of Leipzig, Department of Computer Science & Interdisciplinary Centre for
Bioinformatics, D-04107 Leipzig, Germany. 3INSERM U869 and 4Université de Bordeaux, F-33076 Bordeaux Cedex, France. 5Fraunhofer Institute for Cell Therapy and Immunology,
RNomics Group, D-04103 Leipzig, Germany. 6Max Planck Institute for Molecular Genetics, D-14195 Berlin, Germany. 7Max Planck Institute for the Mathematics in Sciences, D-04103
Leipzig, Germany. 8University of Vienna, Institute for Theoretical Chemistry, A-1090 Vienna, Austria. 9The Santa Fe Institute, Santa Fe, 87501 New Mexico, USA. 10University of
Würzburg, Institute for Molecular Infection Biology, D-97080 Würzburg, Germany.

250
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 ARTICLES

a log2 c
mapped reads (–) cDNA, no treatment

Leading strand
18 (+) cDNA, 5′PPP enriched
(–)
0
(+)
tRNA
rRNA AUG
mRNA
cagA

Lagging strand
18
(–)
0
(+)

0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.67

b tRNA-Phe
Urease operon cag pathogenicity island
RNase P cleavage
ML –

Leading strand
ML +
d
50
Relative score

AS – 87%

Number of TSS
0.4 0.1 40
0.2 AS + 0.05
0 0 30
cag11 cag12 cag16 cag17
cagA 20
cag10 cag23
ureH ureF ureE ureI ureB ureA 10
ca 33
ca g13

8
9

ca 20
ca 21
2

cag24
5
4

HP0068
05

g1
g1

g2

g2
g1

g1
Relative score

g
g
HP

tRNA-Val

Lagging strand
ca
ca

ca

ca
ca

0 0 0
<–100 <–2 –2 –1 0 1 2 >2 >100
–0.2 ML – –0.1
–0.4 –0.2 Difference (nt)
ML +

AS –

AS +
cag22–cag18 suboperon
cag25–cag18 operon
00

00

00

00

0
00

00

00

00

00

00

00

00

00

00

00
,0

,0

,0

,0

4,

6,

8,

0,

2,

4,

6,

8,

0,

2,

4,
72

74

76

78

56

56

56

57

57

57

57

57

58

58

Figure 1 | H. pylori TSS revealed by dRNA-seq. a, Combined cDNA reads
without (blue, (2) libraries) or with (orange, (1) libraries) terminator
exonuclease treatment mapped to and plotted as log2 values over the H.
pylori chromosome. All libraries were adjusted to same scale. Vertical bars
indicate tRNA (green), rRNA (blue) or mRNA (grey) gene clusters. b, cDNAs
of mid-log growth (ML2/1) and acid stress (AS2/1) libraries mapped onto
urease operon (left) or cag pathogenicity island (right) annotations in
forward and reverse direction (genome position at the bottom). Black and
grey arrows denote published (ureA13,14, ureI49,50, cagA/B15) and newly
identified TSS, respectively. Annotation of cagB (open symbol) according to
H. pylori strain G27. Dotted lines exemplify the primary cag25–cag18 operon
and its associated cag22–cag18 suboperon transcribed by the internal TSS in
cag23. Scales give a relative score (% mapped reads per genome position)
58
and are consistent for all libraries of the same strand (maximum of 0.1 for
leading strand; minimum of 20.2 for lagging strand). c, Schematic drawing
of cDNA enrichment patterns (here in AS2/1 libraries) at primary 59 ends
of cagA mRNA (top) or tRNA-Phe precursor (bottom). Exonuclease
treatment (AS1 library; red curve) shifts the cagA cDNAs towards the
nuclease-protected 59-end, yielding a sawtooth-like profile with an elevated
sharp 59 flank that matches the published TSS15. In contrast, the mature
(RNase P-cleaved) 59 end of tRNA-Phe is predominant in the AS (2) library
(black curve). d, Histogram indicating observed distances between 69 TSS
identified by dRNA-seq vs independent experimental approaches
(Supplementary Table 1). Of these TSS 87% matched within 62 nt tolerance
(all 22 TSS mapped by 59 RACE in Supplementary Fig. 3 match with 61 nt).

annotated gene on the same strand; antisense TSS situated inside or revealed an extended Pribnow box (tgnTAtaAT) as the 210 motif of
within #100 bp of an oppositely encoded gene; and orphan TSS the housekeeping s80 in H. pylori (Fig. 2b; for s28 and s54, see Sup-
without annotated genes in proximity. Consequently, multiple asso- plementary Tables 6–8 and Supplementary Fig. 5). Intriguingly, the
ciations of TSS are not uncommon (Fig. 2a). 235 motif is replaced by a periodic AT-rich signal upstream of
H. pylori was thought to lack the extensive operon structure3,16 typical position 214. Thus, consistent with earlier analysis of much fewer
of other bacterial genomes. We complemented our TSS map with promoters in H. pylori20 or related Campylobacter jejuni21, our global
DOOR17 and conventional RNA-seq analyses to assign 87.5% of all H. TSS map shows that transcription in e-proteobacteria initiates at
pylori genes to 337 primary operons (Supplementary Table 5), present- extended Pribnow boxes downstream of periodic AT-rich stretches.
ing the first operon map based on knowledge of transcriptional ini-
tiation. dRNA-seq readily detected internal TSS of downstream Massive antisense transcription
cistrons within longer primary operons, predicting 126 additional sub- The H. pylori transcriptome is highly complex (Fig. 2a), and 27%
operons and 66 monocistrons overlapping the 39 part of polycistrons in of the primary TSS are also antisense TSS, indicating that—similar to
H. pylori. Independent validation of such suboperonic signals (Sup- E. coli22—antisense transcription occurs across the entire H. pylori
plementary Fig. 4) confirmed an acid-induced internal TSS in cag23 genome (Fig. 3a, Supplementary Fig. 6 and Supplementary Table 9).
(Fig. 1b), supporting the known upregulation of cag22–cag18 from Because the antisense TSS are uncorrelated with local GC content
the primary cag25–18 operon under acid stress10. Likewise, internal (Fig. 3a), they are unlikely to reflect promiscuous transcription ini-
TSS reveal potential uncoupling of suboperons starting with ftsH (cell tiation at AT-rich hexamers. Furthermore, analysis of biological repli-
division) or copP (copper transport) from the physiologically unrelated cates (ML2/1 libraries), control experiments using actinomycin D to
HP1067 (cheY) chemotaxis gene in the HP1067–HP1074 operon18. suppress unspecific cDNA synthesis (Supplementary Information),
Bacterial promoters usually contain specific sequences for binding and independent northern blot probing (Supplementary Fig. 7),
of RNA polymerase (RNAP)-associated s factors, for example, the consistently show that the many detected antisense TSS are not
235 (TTGACA) and 210 (TATAAT) boxes of the housekeeping s70 artefacts of library construction. Rather, their high number might
in Escherichia coli19. Correspondingly, motif searches at the 1,907 TSS indicate that transcriptional initiation is an important cause of global
251
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES NATURE | Vol 464 | 11 March 2010

a b 1,000
P I P P,I O
S 800
Gene 1 Gene 2 Gene 3 Gene 4

Counts
600
<500 <100 <100
400
A A
200
Secondary Internal
0
119 439 –20 –15 –10 –5
Primary Antisense Extended –10 box position relative to TSS
812 31 244 969
2
0 36 12
451 686
0 2

Bits
1
0 Orphan
142 50
3 0
38 0

–40
–39
–38
–37
–36
–35
–34
–33
–32
–31
–30
–29
–28
–27
–26
–25
–24
–23
–22
–21
–20
–19
–18
–17
–16
–15
–14
–13
–12
–11
–10
–9
–8
–7
–6
–5
–4
–3
–2
–1
+1
216
5′ 3′

c d Reannotation
30 Old AUG TGNTATAAT SD AUG
26695
Leaderless mRNA HPAG1
SD sequence 25 J99
2 Hac
Bits

1 Consensus

Number of mRNAs
0 20

15

8
01
P1
H
Relative score
10 pgsA rocE htrA
0.2
5 0.1 AS –
0
0 AS +
–400 –200 –100 –80 –60 –40 –20
5′ UTR length (nt) 1,079,000 1,079,500 1,080,000 1,080,500 1,081,000 1,081,500

Figure 2 | TSS annotation and 59 mRNA structure. a, Top: representation of which 1,645/1,906 (,86%) lie 211, 210 and 29 bp relative to TSS (top,
categories for TSS annotation based on expression strength and genomic histogram of distribution; bottom, logo of upstream TSS sequences). The
context: primary (P), secondary (S), internal (I), antisense (A), or orphan dotted line indicates the upstream periodic AT-stretch signal. See
(O). Bottom: Venn diagram showing overlap among TSS categories. Many of Supplementary Table 7 for statistical analysis. c, Frequency of individual 59
the 1,907 H. pylori TSS associate with multiple categories (see TSS in gene 2 UTR length based on 825 TSS (primary and secondary) of mRNAs. 59 UTR
above), yielding 2,496 TSS associations in total (Supplementary Table 4). lengths ,10 nt (red bars) reveal 34 leaderless mRNAs. The inset shows the
Antisense TSS were detected for 721/1,576 (,46%) of annotated ORFs; SD sequence motif of H. pylori. d, Re-annotation of rocE gene prompted by a
primary TSS for 717 ORFs; 106 secondary TSS for 98 ORFs; 428 internal TSS newly discovered TSS downstream of its originally annotated start codon
for 363 ORFs. Primary TSS commonly reside in upstream ORFs: 142/812 (grey box). Reads of AS2/1 libraries mapping to the pgsA-rocE-HP1018-
(,17%) primary TSS are also internal; 142/439 (,32%) internal TSS are htrA region are shown below an alignment of rocE sequences (59 region) of
also primary. 216/812 primary TSS (,27%) are antisense TSS, revealing diverse Helicobacter strains. The highly conserved re-annotated AUG, its SD
profound global antisense transcription. b, Motif searches upstream of and the 210 box of the new TSS are boxed.
H. pylori TSS reveal extended Pribnow boxes (210 signal: tgnTAtaAT) of

antisense transcription in H. pylori, in addition to possible imperfect ,6 nt (median distance) upstream of start codons as the consensus
termination22. Shine–Dalgarno sequence in H. pylori (Fig. 2c). We found correlations
At least one antisense TSS is associated with ,46% of all ORFs, of 59UTR length with cellular function. For example, nucleotide- and
including many housekeeping genes such as valS (valyl-tRNA synthe- nucleoside-related genes almost invariably have ,30-nt-long 59UTRs,
tase) or rpl21 (ribosomal protein L21), although there is no general as if optimized for translation. In contrast, regulatory genes of cellular
bias towards core or variable H. pylori genes23 (data not shown). processes such as cell division, pathogenesis or transformation possess
Moreover, ,28% of the tRNAs, and the 59 leaders of 23S and 16S significantly longer 59UTRs (Supplementary Fig. 8). Structural align-
rRNA precursors have antisense TSS. Notably, as H. pylori lacks ment of the H. pylori 59UTRs with Rfam database families detected a
endoRNases E/G and other common processing factors of stable thiamine pyrophosphate riboswitch upstream of pnuC. In support of
RNAs3, antisense-guided cleavage involving the dsRNA-specific the previously predicted transcriptional attenuation mechanism30, both
RNase III (ref. 3) might compensate for this paucity. RNA-seq and northern blot probing observed a short (,100 nt) tran-
We identified antisense TSS for 22 of 34 putative phase-variable script from this candidate riboswitch (Supplementary Fig. 9). Although
genes24 featuring homopolymeric tracts and dinucleotide repeats, no other known riboswitches were detected, conforming to their general
and with functions in lipopolysaccharide biosynthesis, surface struc- paucity in Gram-negative species, there are 337 UTRs long enough
ture and DNA restriction/modification (Supplementary Table 10). (.60 nt) to harbour novel cis-regulatory RNA structures31.
Of these, the two fucT copies (HP0651/0379) encode the fucosyl- Although leaderless mRNAs are considered rare and primarily
transferases that modify the major Lewis antigen of H. pylori. Our phage-associated in Gram-negative species32, we found that ,2.2% of
discovery of antisense TSS adjacent to the 59 poly(C) tracts involved all H. pylori mRNAs have a 59UTR ,10 nt (Supplementary Table 11).
in switching on/off the fucT genes by slipped-strand mispairing25,26 At 26 genes, including the dnaA, recR and hemH housekeeping genes,
raises the possibility of antisense regulation of surface structure transcription initiates exactly at an AUG start codon essential for stable
variations and host interactions. Moreover, acid-stress-induced ribosome binding of leaderless mRNAs32.
antisense RNAs opposite to known acid-stress-repressed genes (for Some primary TSS lay downstream of previously annotated start
example, rpl21, HP1186, HP0637; refs 10, 11) would indicate similar codons, as exemplified by rocE (Fig. 2d). Yet, sequence conservation
control at low pH (Supplementary Fig. 7). among Helicobacter strains strongly supported the rocE TSS mapped
by dRNA-seq, as well as a new start codon ,30 bp downstream.
59 UTRs and leaderless mRNAs Similarly, we propose corrections for 18 more genes (Supplementary
The 59UTR (untranslated region) ranging from TSS to start codon Table 12), complementing re-annotations by genome comparison33.
determines the translational efficiency of messengers. In striking accor-
dance with the structurally observed ribosome contacts27, our TSS An unexpected wealth of Helicobacter sRNAs
annotation reveals that most (,50%) of the 59UTRs are 20–40 nucleo- Previous annotations in H. pylori predicted stable tRNAs, rRNAs,
tides (nt) in length (Fig. 2c), and support the AAGGag motif28,29 located transfer-messenger RNA (tmRNA), RNase P and signal recognition
252
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 ARTICLES

a 20
experiments (Supplementary Figs 10–14 and Supplementary
log-normalized

Antisense sRNA Small ORF Sense in ORF 6S RNA

Table 14), among these are all five members of a ,200 nt sRNA
expression

10

0
family whose multiplicity is reminiscent of the Qrr sRNAs that con-
10
trol quorum sensing and virulence in Vibrio bacteria34.
20
Plasticity zone
6S We have also identified 6S RNA (,180 nt), an abundant and
GC content

0.45 ubiquitous riboregulator of RNAP. 6S RNA notoriously failed to be

0.35 discovered in the e-subdivision31,35,36, perhaps because it is expressed
0.25 Plasticity zone cag antisense to HP1219 (Fig. 3b,c), a poorly conserved hypothetical ORF
0 0.5 1.0 1.5 1.67 (Supplementary Fig. 15). Structural probing experiments in vitro
b pRNA* pRNA c (Fig. 3d) and conservation analysis (Supplementary Fig. 16) indicate
0.3 0.6 0.8 1.2 1.7 2.0 A600
AAUAAGGCUAGGAGGAU AACGAACUUGCC
HP1219 (hypothetical) 6S that H. pylori 6S RNA adopts the characteristic structure of a long
ML– hairpin with a central asymmetric bulge by which E. coli 6S RNA
ML+
5S
sequesters RNAP35–37. To disentangle itself, RNAP uses 6S RNA as
0.1
AS– template for transcription of 14–20 nt RNA products (pRNAs)37.
We detected two classes of pRNAs in H. pylori (Fig. 3b), one starting
Relative score

0.05 AS+
0

1,295,300 1,295,400 1,205,500 1,295,600 1,295,700

with the corresponding bulge-internal adenosine of E. coli pRNAs37,
0
–6 ML– and the other (pRNA*) originating from the opposite strand as previ-
–12
ML+ ously observed with certain 6S RNA mutants in vitro (K. Wassarman,
AS– personal communication). Our in vivo detection of pRNAs in a
AS+ remote relative of E. coli shows that 6S RNA regulation of RNAP
purD yhcG
6S RNA
activity is a widely conserved mechanism.
Several of the new H. pylori sRNAs seem as abundant as 6S RNA
d (Fig. 3a), and most of them are conserved at the sequence level among
Le as T1
Lead T2
++
ad +
e
e
RNas

90
Helicobacter species but not outside the e-subdivision (data not
H
RN
T1
O
C

shown). In global structural clustering analysis assaying conservation

G151
G143 of functional secondary structure rather than primary sequence, we
G130
G121 110 observed that whereas the c-proteobacterial sRNAs of E. coli and
G108
G100
70 Salmonella form large groups of similar structures, the H. pylori
G85 Lead(II) sRNAs fall in small groups exhibiting specific structural motifs
G77
G71
RNase T2
RNase T1
(Supplementary Fig. 17). Thus, except for 6S RNA and housekeeping
G64 RNAs, e-proteobacteria including H. pylori might have evolved a
G59
G54
unique sRNA repertoire.
pRNA* Many sRNAs in other bacteria repress trans-encoded mRNAs by
G44
G41 pRNA short base-pairing in the 59UTR5,6. Here, the TargetRNA program38
G37 predicted an interaction of the abundant HPnc5490 sRNA with the
G32
59UTR of tlpB encoding a canonical chemotaxis receptor of H. pylori
G29
(Fig. 4a). The involvement of accessible loop residues of HPnc5490,
and the calculated RNA duplex strength (DG 5 234.3 kcal mol21)
G19 added confidence to this prediction. An HPnc5490 deletion mutant
G17
was generated, and observed to have increased levels of the ,60 kDa
G12 170 TlpB protein (Fig. 4b) and the tlpB-HP0102 operon mRNA, yet
G9
10 normal expression of the remaining canonical chemotaxis receptor
genes, tlpA and tlpC (Fig. 4c). Thus, HPnc5490 probably regulates
5′ 3′
tlpB as a trans-antisense RNA.
Figure 3 | Discovery of H. pylori sRNAs including 6S RNA. a, Relative We identified a family of six structurally related ,80 nt sRNAs
expression of candidate and validated sRNAs according to log2 values of read expressed antisense to small ORFs of homologous 22–30 amino acid
numbers (first 50 nt downstream of TSS) of all sequenced libraries (upper peptides (Fig. 4d, e and Supplementary Fig. 18), henceforth referred
part) compared to local GC content (lower part) plotted across the H. pylori
to as IsoA1-6 (RNA inhibitor of small ORF family A) and aapA1-6
genome. b, Reads in the ML and AS libraries show a highly expressed 180 nt
sRNA candidate, now H. pylori 6S RNA, in the yhcG–purD intergenic region (antisense RNA-associated peptide family A), respectively. Five of the
and opposite to annotated (yet questionable) ORF HP1219. Determined aapA ORFs produced stable mRNAs in vivo (Fig. 4e). In vitro trans-
sequences of detected 6S RNA-specific pRNA/pRNA* including their 6S lation assays yielded the expected small peptides, except for aapA2
RNA-borne TSS are shown at the top. c, Northern blot detection of 6S RNA mRNA whose Shine–Dalgarno sequence is mutated in strain 26695
(and 5S rRNA as loading control) from exponential growth to late stationary (not shown). Furthermore, translation of aapA1 or aapA3 was
phase in BHI medium. d, In vitro structure probing (left) of 59end-labelled strongly and specifically inhibited in the presence of the cognate
6S RNA of H. pylori. C, T1 and OH identify no reaction, digestion under IsoA1 or IsoA3 sRNAs (Fig. 4f), thus revealing candidates of cis-
denaturing conditions with nuclease T1 (cleaved G residues given to the antisense regulation in H. pylori.
left), or alkali, respectively. The right four lanes reveal cleavages induced by
The AapA peptides are conserved in other H. pylori strains
RNase T1, T2, or lead(II), whose positions are indicated in the derived
secondary structure of 6S RNA (right panel) by black, red, and grey circles, (Supplementary Fig. 19) and might interact with membranes, as
respectively. The template nucleotides of the detected pRNA and pRNA* suggested by their predicted high hydrophobicity and a-helical struc-
sequences are framed in black or blue, respectively. ture (Fig. 4d and Supplementary Fig. 20), as well as similarities to
Antimicrobial Peptide Database39 entries such as human defensin
particle RNA (SRP RNA), all of which were here confirmed to be LL-37. Therefore, the aapA–isoA loci might be toxin-antitoxin
expressed. Moreover, the dRNA-seq data predicted hundreds of addi- systems that slow down growth of H. pylori or other organisms in
tional sRNA candidates (denoted HPnc) from intergenic regions, the gastric mucosa, protect against phages, or facilitate the proposed
antisense to ORFs, and also sense within ORFs (Fig. 3a; Supplemen- (though controversial) altruistic autolysis of H. pylori40 which was
tary Fig. 10 and Supplementary Table 13). We have so far validated postulated to involve as yet unknown ,3.5 kDa hydrophobic pep-
the expression of ,60 new sRNAs by independent northern blot tides41. Whereas the AapA peptides remain to be detected in vivo, our
253
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES NATURE | Vol 464 | 11 March 2010

interaction map43. The data provide new insights into the organiza-

0
a b c

49
c5
tion of the H. pylori transcriptome, and provide a framework for

Pn
tlpB
WT ΔHPnc5490

T
ΔH
HP0101 HP0102 HP0104

W
10
9
better analysis of individual genes. Knowledge of the vast majority

Relative mRNA expression

5′–CG CA–3′
GGGGGGGGGGGUG
||:|||||:||||
175 8 of H. pylori TSS and operons will help evaluate unclear phenotypes of
7
3′–UU
CCUCCCCCUCCAC
UG–5′ 80 6
transposon insertions44,45, and eliminate unwanted antisense tran-
HP1043 HP1044 5 scripts in heterologous expression constructs of H. pylori vaccine
HPnc5490
58 TlpB
4
3
candidates. Altogether, ,100 sRNAs are known in E. coli5, the model
2 organism of bacterial RNA research. Corrected by genome size, H.
46
5′
3′ 1 pylori rivals E. coli despite the lack of a conserved Hfq protein.
0
Combined with the success of artificial antisense RNAs in H. pylori46,

C
10

tlp
tlp

tlp
P0
our results show that many RNA-mediated regulations are yet to be

H
discovered in e-proteobacterial pathogens.
d aapA1 e
Other RNA-seq studies detected termini of bacterial transcripts8
omp27 deoD

7
but could not unequivocally assign TSS due to lack of 59 group
H S

H S
uh

uh
L

L
AG

AG
AS

AS
PL

PL
M

M
isoA1 331 242
aapA2 242 aapA1
190
aapA5 discrimination. As 59PPP ends mark native transcripts in all eubac-
HP1175 omp27
190
147
110 teria, dRNA-seq should help improve the genome annotations of
isoA2
HP1432 aapA3
110 IsoA5 other organisms, alone or through metatranscriptomics47. It could
67
HP1433 67
IsoA1
also complement the popular CAGE48 technique for eukaryotic
isoA3
aapA4
mRNAs because their cap structure blocks terminator exonuclease.
34
HP1533 242 Moreover, an improved dRNA-seq protocol might permit detection
tnpB aapA6
isoA5 190
IsoA2 147
of processed 59 hydroxyl ends that unlike the more prominent 59P
aapA5 67
110 ends are not presently captured in the (2) library. Semi-quantitative
omp2 IsoA6
HP0022
isoA5
331 analysis of TSS coverage revealed considerable gene expression
242 aapA3 67
aapA6 190 changes in H. pylori grown along with eukaryotic cells (Supplemen-
nusA
HP1517 67
IsoA3 tary Table 16). Thus, dRNA-seq has the potential to unravel gene
isoA6 HP1516
331
expression in pathogens and perhaps also their hosts during
aapA4
4,000 bp
242
f infections.
Is 1
3
oA
oA

1 30 190 –
M

Is

AapA1 110
AapA3
AapA5
AapA6
AapA4
67
Figure 4 | Trans and cis regulatory sRNAs. a, Genomic context of the operon
(dotted line) encoding chemotaxis receptor, TlpB, and of HPnc5490 sRNA,
together with the predicted 13 bp HPnc5490–tlpB RNA duplex (grey
background). A consensus structure of HPnc5490 homologues shown below
reveals presentation of the pairing residues by an accessible loop region.
b, SDS–PAGE showing accumulation of TlpB protein (identified by mass
spectrometry) in wild-type vs HPnc5490-deficient H. pylori bacteria in mid-
log growth. Marker bands (kDa) to the left. c, Quantitative real-time PCR
analysis of HPnc5490-dependent expression of four H. pylori mRNAs. Wild-
type mRNA levels (grey bars) are set to 1, and fold-regulation upon HPnc5490
deletion is indicated by black bars. Error bars indicate standard deviations
among three technical replicates. d, Genomic locations of IsoA1–6 RNA
(black boxes) and associated peptide-encoding aapA mRNAs (grey). IsoA2
overlaps with hypothetical ORF HP1176. In strain 26695, aapA2 lacks a
conserved peptide ORF. IsoA5 and IsoA6 lie opposite to hypothetical ORFs,
HP0024 (aapA5) and HP1515 (aapA5). In the alignment of AapA1 and
AapA3–6 peptides of strain 26695 (bottom), grey shades denote predicted
transmembrane domains, and ‘1’ a positively charged amino acid (H, K, or
R) conserved in at least one sequence. e, Northern blots of IsoA and aapA
transcripts of bacteria grown to mid-log phase (ML), under acid stress (AS),
or in cell culture flasks (PL) along with gastric (AGS) or liver (Huh7) cells. An
aapA2 mRNA failed to be detected. Triangles indicate expected transcripts
according to RNA-seq. Size marker positions are shown to the left.
f, Autoradiogram of gel-separated in vitro translation products of the aapA1
and aapA3 mRNAs alone, or in presence of fourfold excess of IsoA1 or IsoA3
antisense RNAs. Lane M contains a co-migrating 3.5 kDa marker peptide.
prediction of additional conserved 10–60 amino acid peptides
(Supplementary Table 15 and Supplementary Figs 20–21) shows that
small ORFs are more common in Helicobacter species than appre-
ciated. Moreover, the intriguing similarities of three antisense RNA-
associated peptides (aapC1/2, aapD) of H. pylori with the hydrophobic
Ibs family42 of E. coli indicate that such loci might be spread via hori-
zontal gene transfer.
Concluding remarks
Our single nucleotide resolution TSS map constitutes the third global
reference data set for the model organism H. pylori strain 26695,
complementing its genome sequence3,4 and global protein–protein
254
©2010 Macmillan Publishers Limited. All rights reserved
3.5 aapA1
IsoA4
METHODS SUMMARY
3.5 aapA3
Details of bacterial growth conditions, cDNA library construction, TSS annota-
tion, sRNA validation, and biocomputational as well as statistical analyses, as
well as other methods are provided as Supplementary Information. For dRNA-
seq, total RNA was extracted from H. pylori grown under various growth con-
ditions, and treated with Terminator-59-phosphate-dependent exonuclease
(Epicentre Biotechnologies) to deplete processed RNAs. The cDNA libraries
(vertis Biotechnologie AG) were constructed for RNA 2/1 prior exonuclease
treatment for each growth condition without size fractionation of RNA,
before cDNA synthesis. 225,000–540,000 cDNAs per library were sequenced
on a Roche FLX machine, and mapped to the H. pylori 26695 genome. For each
library, graphs representing the number of mapped reads per nucleotides were
calculated and visualized using the Affymetrix Integrated Genome Browser. To
corroborate operon predictions, we also analysed a total of ,43,000,000 strand-
specific cDNAs obtained from randomly fragmented RNA by Solexa (Illumina)
sequencing.
Received 6 August; accepted 14 December 2009.
Published online 17 February 2010.
1. Cover, T. L. & Blaser, M. J. Helicobacter pylori in health and disease.
Gastroenterology 136, 1863–1873 (2009).
2. Suerbaum, S. & Michetti, P. Helicobacter pylori infection. N. Engl. J. Med. 347,
1175–1186 (2002).
3. Tomb, J. F. et al. The complete genome sequence of the gastric pathogen
Helicobacter pylori. Nature 388, 539–547 (1997).
4. Alm, R. A. et al. Genomic-sequence comparison of two unrelated isolates of the
human gastric pathogen Helicobacter pylori. Nature 397, 176–180 (1999).
5. Waters, L. S. & Storz, G. Regulatory RNAs in bacteria. Cell 136, 615–628 (2009).
6. Majdalani, N., Vanderpool, C. K. & Gottesman, S. Bacterial small RNA regulators.
Crit. Rev. Biochem. Mol. Biol. 40, 93–113 (2005).
7. Valentin-Hansen, P., Eriksen, M. & Udesen, C. The bacterial Sm-like protein Hfq: a
key player in RNA transactions. Mol. Microbiol. 51, 1525–1533 (2004).
8. Sorek, R. & Cossart, P. Prokaryotic transcriptomics: a new view on regulation,
physiology and pathogenicity. Nature Rev. Genet. 11, 9–16 (2010).
9. Sharma, C. M. & Vogel, J. Experimental approaches for the discovery and
characterization of regulatory small RNA. Curr. Opin. Microbiol. 12, 536–546
(2009).
10. Wen, Y. et al. Acid-adaptive genes of Helicobacter pylori. Infect. Immun. 71,
5921–5939 (2003).
11. Merrell, D. S., Goodrich, M. L., Otto, G., Tompkins, L. S. & Falkow, S. pH-regulated
gene expression of the gastric pathogen Helicobacter pylori. Infect. Immun. 71,
3529–3539 (2003).
12. Bury-Moné, S. et al. Responsiveness to acidity via metal ion regulators mediates
virulence in the gastric pathogen Helicobacter pylori. Mol. Microbiol. 53, 623–638
(2004).
NATURE | Vol 464 | 11 March 2010
13. Shirai, M., Fujinaga, R., Akada, J. K. & Nakazawa, T. Activation of Helicobacter pylori
ureA promoter by a hybrid Escherichia coli–H. pylori rpoD gene in E. coli. Gene 239,
351–359 (1999).
14. Spohn, G. & Scarlato, V. Motility of Helicobacter pylori is coordinately regulated by
the transcriptional activator FlgR, an NtrC homolog. J. Bacteriol. 181, 593–599
(1999).
15. Spohn, G., Beier, D., Rappuoli, R. & Scarlato, V. Transcriptional analysis of the
divergent cagAB genes encoded by the pathogenicity island of Helicobacter pylori.
Mol. Microbiol. 26, 361–372 (1997).
16. Thompson, L. J. et al. Gene expression profiling of Helicobacter pylori reveals a
growth-phase-dependent switch in virulence gene expression. Infect. Immun. 71,
2643–2655 (2003).
17. Mao, F., Dam, P., Chou, J., Olman, V. & Xu, Y. DOOR: a database for prokaryotic
operons. Nucleic Acids Res. 37, D459–D463 (2009).
18. Beier, D., Spohn, G., Rappuoli, R. & Scarlato, V. Identification and characterization
of an operon of Helicobacter pylori that is involved in motility and stress
adaptation. J. Bacteriol. 179, 4676–4683 (1997).
19. Burgess, R. R. & Anthony, L. How sigma docks to RNA polymerase and what sigma
does. Curr. Opin. Microbiol. 4, 126–131 (2001).
20. Forsyth, M. H. & Cover, T. L. Mutational analysis of the vacA promoter provides
insight into gene transcription in Helicobacter pylori. J. Bacteriol. 181, 2261–2266
(1999).
21. Petersen, L., Larsen, T. S., Ussery, D. W., On, S. L. & Krogh, A. RpoD promoters in
Campylobacter jejuni exhibit a strong periodic signal instead of a 235 box. J. Mol.
Biol. 326, 1361–1372 (2003).
22. Selinger, D. W. et al. RNA expression analysis using a 30 base pair resolution
Escherichia coli genome array. Nature Biotechnol. 18, 1262–1268 (2000).
23. Gressmann, H. et al. Gain and loss of multiple genes during the evolution of
Helicobacter pylori. PLoS Genet. 1, e43 (2005).
24. Saunders, N. J., Peden, J. F., Hood, D. W. & Moxon, E. R. Simple sequence repeats
in the Helicobacter pylori genome. Mol. Microbiol. 27, 1091–1098 (1998).
25. Appelmelk, B. J. et al. Phase variation in Helicobacter pylori lipopolysaccharide.
Infect. Immun. 66, 70–76 (1998).
26. Wang, G., Rasko, D. A., Sherburne, R. & Taylor, D. E. Molecular genetic basis for
the variable expression of Lewis Y antigen in Helicobacter pylori: analysis of the
a(1,2) fucosyltransferase gene. Mol. Microbiol. 31, 1265–1274 (1999).
27. Ramakrishnan, V. Ribosome structure and the mechanism of translation. Cell 108,
557–572 (2002).
28. Vanet, A., Marsan, L., Labigne, A. & Sagot, M. F. Inferring regulatory elements
from a whole genome. An analysis of Helicobacter pylori s80 family of promoter
signals. J. Mol. Biol. 297, 335–353 (2000).
29. Yada, T., Totoki, Y., Takagi, T. & Nakai, K. A novel bacterial gene-finding system
with improved accuracy in locating start codons. DNA Res. 8, 97–106 (2001).
30. Rodionov, D. A., Vitreschak, A. G., Mironov, A. A. & Gelfand, M. S. Comparative
genomics of thiamin biosynthesis in procaryotes. New genes and regulatory
mechanisms. J. Biol. Chem. 277, 48949–48959 (2002).
31. Weinberg, Z. et al. Identification of 22 candidate structured RNAs in bacteria
using the CMfinder comparative genomics pipeline. Nucleic Acids Res. 35,
4809–4819 (2007).
32. Brock, J. E., Pourshahian, S., Giliberti, J., Limbach, P. A. & Janssen, G. R. Ribosomes
bind leaderless mRNA in Escherichia coli through recognition of their 59-terminal
AUG. RNA 14, 2159–2169 (2008).
33. Boneca, I. G. et al. A revised annotation and comparative analysis of Helicobacter
pylori genomes. Nucleic Acids Res. 31, 1704–1714 (2003).
34. Lenz, D. H. et al. The small RNA chaperone Hfq and multiple small RNAs control
quorum sensing in Vibrio harveyi and Vibrio cholerae. Cell 118, 69–82 (2004).
35. Barrick, J. E., Sudarsan, N., Weinberg, Z., Ruzzo, W. L. & Breaker, R. R. 6S RNA is a
widespread regulator of eubacterial RNA polymerase that resembles an open
promoter. RNA 11, 774–784 (2005).
36. Wassarman, K. M. & Storz, G. 6S RNA regulates E. coli RNA polymerase activity.
Cell 101, 613–623 (2000).
©2010 Macmillan Publishers Limited. All rights reserved
ARTICLES
37. Wassarman, K. M. & Saecker, R. M. Synthesis-mediated release of a small RNA
inhibitor of RNA polymerase. Science 314, 1601–1603 (2006).
38. Tjaden, B. et al. Target prediction for small, noncoding RNAs in bacteria. Nucleic
Acids Res. 34, 2791–2802 (2006).
39. Wang, Z. & Wang, G. APD: the Antimicrobial Peptide Database. Nucleic Acids Res.
32, D590–D592 (2004).
40. Phadnis, S. H. et al. Surface localization of Helicobacter pylori urease and a heat
shock protein homolog requires bacterial autolysis. Infect. Immun. 64, 905–912
(1996).
41. Fujita, Y. et al. A novel mechanism of autolysis in Helicobacter pylori: possible
involvement of peptidergic substances. Helicobacter 10, 567–576 (2005).
42. Fozo, E. M., Hemm, M. R. & Storz, G. Small toxic proteins and the antisense RNAs
that repress them. Microbiol. Mol. Biol. Rev. 72, 579–589 (2008).
43. Rain, J. C. et al. The protein–protein interaction map of Helicobacter pylori. Nature
409, 211–215 (2001).
44. Salama, N. R., Shepherd, B. & Falkow, S. Global transposon mutagenesis and
essential gene analysis of Helicobacter pylori. J. Bacteriol. 186, 7926–7935 (2004).
45. Jenks, P. J., Chevalier, C., Ecobichon, C. & Labigne, A. Identification of nonessential
Helicobacter pylori genes using random mutagenesis and loop amplification. Res.
Microbiol. 152, 725–734 (2001).
46. Croxen, M. A., Ernst, P. B. & Hoffman, P. S. Antisense RNA modulation of alkyl
hydroperoxide reductase levels in Helicobacter pylori correlates with organic
peroxide toxicity but not infectivity. J. Bacteriol. 189, 3359–3368 (2007).
47. Shi, Y., Tyson, G. W. & DeLong, E. F. Metatranscriptomics reveals unique
microbial small RNAs in the ocean’s water column. Nature 459, 266–269 (2009).
48. Kodzius, R. et al. CAGE: cap analysis of gene expression. Nature Methods 3,
211–222 (2006).
49. Akada, J. K., Shirai, M., Takeuchi, H., Tsuda, M. & Nakazawa, T. Identification of
the urease operon in Helicobacter pylori and its control by mRNA decay in
response to pH. Mol. Microbiol. 36, 1071–1084 (2000).
50. Pflock, M., Kennard, S., Delany, I., Scarlato, V. & Beier, D. Acid-induced activation
of the urease promoters is mediated directly by the ArsRS two-component
system of Helicobacter pylori. Infect. Immun. 73, 6437–6445 (2005).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank F. Seifert; H. Hamoutene and B. Timmermann for
technical support; M. Schmid for mass spectrometry analysis; H. De Reuse, A. van
Vliet and M. K. Waldor for discussions; F. Thümmler for library preparation;
M. Droege for pyrosequencing support. J.V. and R.R. are supported by NGFN1 grants
(BMBF, Germany), and J.V. and P.F.S. by DFG Priority Program SPP1258 Sensory and
Regulatory RNAs in Prokaryotes (Grants VO8751/2, VO8751/4; STA850/7-1). S.H.
was supported by a formel.1 grant of the University of Leipzig, the Freistaat Sachsen
(LIFE project), the German Research Foundation IZBI (BIZ6/1-4) and Volkswagen
Stiftung (I/82 720). F.D. is supported by the French Agence Nationale de la
Recherche (ANR-07-JCJC-0104-01), the French Association de la Recherche contre
le Cancer (ARC) and La Ligue Nationale contre le Cancer (LNCC). We thank D. Rose
for his supporting work and S. Washietl for a pre-release of the RNAcode software.
Author Contributions C.M.S., F.D., P.F.S. and J.V. designed the research; C.M.S.,
F.D., A.S., J.R., J.V. and S.C. performed all wet lab work. C.M.S., S.H., S.F., K.R., J.H.
and P.F.S. conducted biocomputational analyses; R.R. carried out sequencing. J.V.
wrote the manuscript, which all authors commented on, and supervised the
project. Author information and raw data are available from C.M.S, P.F.S. and J.V.
Author Information Raw data are available from the NCBI Short Read Archive
(http://www.ncbi.nlm.nih.gov/Traces/sra) under accession number SRA010186.
Reprints and permissions information is available at www.nature.com/reprints.
The authors declare no competing financial interests. Correspondence and
requests for materials should be addressed to J.V.
(joerg.vogel@uni-wuerzburg.de).
255

LETTERS
Confirmation of general relativity on large scales from
weak lensing and galaxy velocities
Reinabelle Reyes1, Rachel Mandelbaum1, Uros Seljak2,3,4, Tobias Baldauf2, James E. Gunn1, Lucas Lombriser2
& Robert E. Smith2
Although general relativity underlies modern cosmology, its
applicability on cosmological length scales has yet to be stringently
tested. Such a test has recently been proposed1, using a quantity,
EG, that combines measures of large-scale gravitational lensing,
galaxy clustering and structure growth rate. The combination is
insensitive to ‘galaxy bias’ (the difference between the clustering
of visible galaxies and invisible dark matter) and is thus robust to
the uncertainty in this parameter. Modified theories of gravity
generally predict values of EG different from the general rela-
tivistic prediction because, in these theories, the ‘gravitational slip’
(the difference between the two potentials that describe perturba-
tions in the gravitational metric) is non-zero, which leads to
changes in the growth of structure2 and the strength of the gravi-
tational lensing effect3. Here we report that EG 5 0.39 6 0.06 on
length scales of tens of megaparsecs, in agreement with the general
relativistic prediction of EG < 0.4. The measured value excludes a
model1 within the tensor–vector–scalar gravity theory4,5, which
modifies both Newtonian and Einstein gravity. However, the rela-
tively large uncertainty still permits models within f( ) theory6,
which is an extension of general relativity. A fivefold decrease in
uncertainty is needed to rule out these models.
Operationally, we define the probe of gravity EG as a combination
of three observable quantities:
1 U gm (R)
EG (R)~
b U gg (R)
Here b is the redshift distortion parameter7, which is a measure of the
radial anisotropy in the clustering pattern reconstructed by galaxy
redshift surveys. Growth of large-scale structure over cosmic time
induces coherent galaxy motions that give rise to the observed aniso-
tropy. In the regime where perturbations are linear, b is equal to the
ratio of the dimensionless linear growth rate of structure, f, and the
galaxy bias parameter, b. The galaxy–matter and galaxy–galaxy annular
differential surface densities8 (ADSDs), Ugm(R) and Ugg(R), are derived
from observations of large-scale gravitational lensing and galaxy
clustering, respectively. They have been constructed to suppress the
small-scale contributions to these observations. In particular, the
galaxy–matter ADSD is defined as

2
R0
U gm (R):DSgm (R){ DSgm (R0 )
R universe was about 77% of its current size, and is already well into the
ð
2 ð2Þ
2 R R0
~ 2 dR’ R’Sgm (R’){Sgm (R)z Sgm (R0 )
R R0 R
where DSgm(R) ; S  gm(,R) 2 Sgm(R) is a measure of the cross-
correlation between the distribution of galaxies and dark matter. The
1
Princeton University Observatory, Peyton Hall, Princeton, New Jersey 08544, USA. 2Institute for Theoretical Physics, University of Zurich, Zurich 8057, Switzerland. 3Physics and
Astronomy Department and Lawrence Berkeley National Laboratory, University of California, Berkeley, California 94720, USA. 4Institute for Early Universe, Ewha University, Seoul
120-750, South Korea.
256
©2010 Macmillan Publishers Limited. All rights reserved
Vol 464 | 11 March 2010 | doi:10.1038/nature08857
quantity DSgm(R) is directly proportional to the observable quantity
for galaxy–galaxy lensing9, the tangential shear, which is a statistical
measure of the weak distortion of shapes of background galaxies due to
the gravitational deflection of light by matter along the line of sight.
Here Sgm(R) is the projected surface mass density at a transverse radius
R and S  gm(,R) is its mean value within the region bounded by R;
DSgm(R) therefore measures an excess surface mass density. The scale
R0 sets the minimum length scale that contributes to the ADSD.
In analogy with the second line of equation (2), the galaxy–galaxy
ADSD is defined as
" ð
2 #
2 R R0
U gg (R):rc 2 dR’ R’wgg (R’){wgg (R)z wgg (R0 )
R R0 R
where rc is the critical matter density of the Universe and wgg is the
projected two-point galaxy correlation function, which is a measure
of how strongly galaxies are clustered together. Because the ADSDs
are insensitive to length scales less than R0, they have the advantage of
being robust to nonlinear, small-scale effects such as the stochasticity
and scale dependence of galaxy bias. In practice, we choose the mini-
mum scale R0 to be large enough to suppress these systematic effects
but small enough to preserve a high signal-to-noise ratio in the mea-
surement of EG.
Another distinct advantage of using EG to test gravity is its insensi-
tivity to both the galaxy bias parameter, b, and the amplitude of
ð1Þ matter perturbations1. This is true to first order in general relativity,
as well as in modified theories of gravity. In particular, in general
relativity, equation (1) simplifies to EG 5 Vm,0/f, where Vm,0 is the
present-day matter density parameter. Thus, unlike in previous
attempts to constrain the theory of gravity on cosmological length
scales10, we are not required to use additional observations and
assumptions to estimate galaxy bias, and are able to obtain results
that are more robust.
We measure Ugm(R) and Ugg(R) from a sample of 70,205 luminous
red galaxies11 (LRGs), from the Sloan Digital Sky Survey12, that satisfy
selection criteria13 used for large-scale structure studies. We use the
value b 5 0.309 6 0.035 measured14 from a galaxy sample selected in
the same way. Our sample covers 5,215 square degrees of sky and the
redshift range z 5 0.16–0.47 (the redshift of the radiation emitted by a
distant galaxy is a measure of the time of emission). The mean redshift,
z 5 0.32, corresponds to a look-back time of 3.5 3 109 yr, when the
accelerated phase of the cosmic expansion. The sample spans a large co-
moving volume, 1.02h23 Gpc3, where h ; H0/(100 km s21 Mpc21) is a
dimensionless form of the current Hubble parameter, H0.
Figure 1 shows the average galaxy–galaxy lensing profile, DSgm(R)
(Fig. 1a), and the projected two-point galaxy correlation function,
NATURE | Vol 464 | 11 March 2010
10
ΔΣ gm (h Mpc–2)
1 f( )
0.1
1 2 4 6 8 10 20 40
b
100
wgg (h–1 Mpc)
10
1 of the growth factor (EG 5 0.22, shown with a nominal error bar of 10% for
1 2 4 6 8 10 20 40
R (h–1 Mpc)
Figure 1 | Probes of large-scale structure measured from 70,000 LRGs.
a, b, Observed radial profiles for two complementary probes, galaxy–galaxy
lensing (a) and galaxy clustering (b), are shown for scales
R 5 1.2h21–47h21 Mpc (open circles). The error bars (1s) are estimated
from jackknife resampling of 34 equal-area regions in the sky. Profiles
measured from mock galaxy catalogues are also shown (solid curves). To
generate the mock galaxy catalogues, we use a standard five-parameter halo
occupation distribution (HOD) model with two parameters related to the
assignment of central galaxies and three parameters related to the
distribution of satellite galaxies (see Supplementary Information for more
details). To fix the HOD model parameters, we require the galaxy number
density to match the observed value and find the best joint fit to the observed
galaxy–galaxy lensing and galaxy clustering profiles. Despite this tuning, it is
remarkable that this simple model is able to reproduce both the overall shape
and particular features of the observed profiles.
wgg(R) (Fig. 1b), measured from the LRG sample for scales
R 5 1.2h21–47h21 Mpc. To achieve a high signal-to-noise ratio in
the lensing profile, we stack together shape measurements15 of more
than 3 3 107 source galaxies (see Supplementary Information for
details). To calculate wgg(R), we use a standard method of counting
galaxy pairs and comparing the result with that for a randomly dis-
tributed sample16.
Figure 2 shows our estimate of EG(R), with 1s error bars that
include the error in the measurement of b. We choose the minimum
scale, R0 5 1.5h21 Mpc, to be close to the typical virial radius of the
haloes of the most massive LRGs, above which we expect the distri-
bution of galaxies to trace that of the dark matter, but our results are
not very sensitive to this particular choice of R0. To estimate errors in
EG(R) while accounting for any correlations between radial bins, we
use jackknife resampling of 34 galaxy subsamples from equal-area
regions in the sky. To obtain numerical corrections accounting for
the effect of scale-dependent galaxy bias and other systematic effects,
we use a suite of dark-matter simulations17 that have been populated
with galaxies using the HOD model18 that best reproduces the obser-
vations (Fig. 1 and Supplementary Information). The correction
factors that we obtain are well below the statistical uncertainty in EG.
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
0.6
GR + ΛCDM
EG
0.4
TeVeS
0.2
2 4 6 8 10 20 40
R (h–1 Mpc)
Figure 2 | Comparison of observational constraints with predictions from
general relativity and viable modified theories of gravity. Estimates of
EG(R) with error bars (1s) including the statistical error in the measurement
of b (ref. 14). The grey shaded region is the 1s envelope of the mean EG on
scales R 5 10h21–50h21 Mpc, where the systematic effects are least
important (Supplementary Information). The horizontal line shows the
mean prediction of general relativity, EG 5 Vm,0/f(z), at the effective redshift
of the measurement, z 5 0.32. On the right-hand side of the panel, labelled
vertical bars show the predicted ranges from three different gravity theories:
general relativity (GR) plus L cold dark matter (LCDM) model
(EG 5 0.408 6 0.029 (1s)); a class of cosmologically interesting models in
f( ) theory with Compton-wavelength parameters21 B0 5 0.001–0.1
(EG 5 0.328–0.365); and a tensor–vector–scalar (TeVeS) model1 designed to
match existing cosmological data and to produce a significant enhancement
clarity).
We take the average of EG(R) over scales R 5 10h21–50h21 Mpc,
accounting for correlations in the data, and find it to be
ÆEGæ 5 0.392 6 0.065 (1s) (grey shaded region in Fig. 2). The 16%
error in EG is dominated by the 11% statistical error in b and the 12%
statistical error in the galaxy–galaxy lensing signal. In addition, there
is a 5% lensing calibration uncertainty15. As detailed in the Sup-
plementary Information, systematic effects on EG are least important
on length scales R . 10h21 Mpc, so the results are most robust there.
We note that the average for R 5 2h21–50h21 Mpc yields a result,
ÆEGæ 5 0.40 6 0.07, consistent with that above.
The general relativistic prediction is EG 5 Vm,0/f(z) 5 0.408 6 0.029
at redshift z 5 0.32, where f(z) < Vm(z)0.55 < 0.629 and Vm(z) is the
matter density parameter at redshift z. The allowed range is determined
by the size of current uncertainties on Vm,0 5 0.2565 6 0.018 (ref. 19).
The data are consistent with this prediction over the range of scales we
consider (Fig. 2, solid line and GR 1 LCDM bar). Unfortunately, pro-
viding model-independent constraints on the gravitational slip is com-
plicated, because changes in the gravitational slip will in turn affect the
rate of growth of structure. What is clear is that there is no evidence for
a non-zero gravitational slip from our data. Thus, we find no deviation
from general relativity on length scales 1011 times greater than those for
which classical tests20 have been performed.
We also compare our constraint on EG with predictions from two
viable modified theories of gravity: tensor–vector–scalar theory5 and
f( ) theory6 (Fig. 2, TeVeS and f( ) bars). Models of f( ) theory21 that
are designed to reproduce the observed cosmic expansion
history with a specific model for the gravitational slip predict that
EG 5 0.328–0.365 (Supplementary Information). The data favour
slightly higher values, but are consistent with this predicted range.
These models can be tested in the near future; limits on EG will
improve as a result of the larger data sets and better control of sys-
tematic errors allowed by the next generation of galaxy surveys.
Nevertheless, even with the current limits, we can tentatively rule
out particular models. For example, a particular tensor–vector–scalar
model1 predicts that EG 5 0.22, which is lower than the observed
value by more than 2.5s. Whether this result rules out the entire class
257
LETTERS
of tensor–vector–scalar models is an open issue22, but it in any case
serves as a concrete demonstration that our measurement of EG
presents a new and non-trivial challenge to both existing and future
proposals of modifications to general relativity.
Received 3 August 2009; accepted 19 January 2010.
1. Zhang, P., Liguori, M., Bean, R. & Dodelson, S. Probing gravity at cosmological
scales by measurements which test the relationship between gravitational lensing
and matter overdensity. Phys. Rev. Lett. 99, 141302 (2007).
2. Daniel, S. F., Caldwell, R. R., Cooray, A., Serra, P. & Melchiorri, A. Multiparameter
investigation of gravitational slip. Phys. Rev. D 80, 023532 (2009).
3. Uzan, J. P. & Bernardeau, F. Lensing at cosmological scales: a test of higher
dimensional gravity. Phys. Rev. D 64, 083004 (2001).
4. Milgrom, M. A modification of the Newtonian dynamics—implications for
galaxies. Astrophys. J. 270, 371–389 (1983).
5. Bekenstein, J. D. Relativistic gravitation theory for the modified Newtonian
dynamics paradigm. Phys. Rev. D 70, 083509 (2004).
6. Carroll, S. et al. Is cosmic speed-up due to new gravitational physics? Phys. Rev. D
70, 083509 (2004).
7. Hamilton, A. J. S. in The Evolving Universe (ed. Hamilton, D.) 185–275 (Astrophys.
Space Sci. Library 231, Kluwer, 1998).
8. Baldauf, T., Smith, R. E., Seljak, U. & Mandelbaum, R. An algorithm for the direct
reconstruction of the dark matter correlation function from weak lensing and
galaxy clustering. Phys. Rev. D (submitted).
9. Bartelmann, M. & Schneider, P. Weak gravitational lensing. Phys. Rep. 340,
291–472 (2001).
10. Guzzo, L. et al. A test of the nature of cosmic acceleration using galaxy redshift
distortions. Nature 451, 541–544 (2008).
11. Eisenstein, D. et al. Spectroscopic target selection for the Sloan Digital Sky Survey:
the luminous red galaxy sample. Astron. J. 122, 2267–2280 (2001).
12. York, D. G. et al. The Sloan Digital Sky Survey: technical summary. Astron. J. 120,
1579–1587 (2000).
13. Eisenstein, D. et al. Detection of the baryon acoustic peak in the large-scale
correlation function of SDSS luminous red galaxies. Astrophys. J. 633, 560–574
(2005).
258
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
14. Tegmark, M. et al. Cosmological constraints from the SDSS luminous red galaxies.
Phys. Rev. D 74, 123507 (2006).
15. Mandelbaum, R. et al. Systematic errors in weak lensing: application to SDSS
galaxy-galaxy weak lensing. Mon. Not. R. Astron. Soc. 361, 1287–1322 (2005).
16. Landy, S. D. & Szalay, A. S. Bias and variance of angular correlation functions.
Astrophys. J. 412, 64–71 (1993).
17. Smith, R. E. Covariance of cross-correlations: towards efficient measures for
large-scale structure. Mon. Not. R. Astron. Soc. 400, 851–865 (2009).
18. Seljak, U. Analytic model for galaxy and dark matter clustering. Mon. Not. R.
Astron. Soc. 318, 203–213 (2000).
19. Dunkley, J. et al. Five-year Wilkinson Microwave Anisotropy Probe observations:
likelihoods and parameters from the WMAP data. Astrophys. J. Suppl. Ser. 180,
306–329 (2009).
20. Will, C. M. The confrontation between general relativity and experiment. Space
Sci. Rev. 148, 60–71 (2009).
21. Song, Y.-S., Hu, W. & Sawicki, I. The large scale structure of f(R) gravity. Phys.
Rev. D 75, 044004 (2007).
22. Ferreira, P. G. & Starkman, G. D. Einstein’s theory of gravity and the problem of
missing mass. Science 326, 812–815 (2009).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements R.M. was supported for the duration of this work by NASA
through a Hubble Fellowship grant awarded by the Space Telescope Science
Institute, which is operated by the Association of Universities for Research in
Astronomy, Inc., for NASA. U.S. acknowledges the Swiss National Foundation. T.B.
acknowledges support by a grant from the German National Academic Foundation
during the initial phase of this project.
Author Contributions R.R., R.M., U.S. and J.E.G. worked on the observational
analysis, with R.R. doing most of the computations. T.B. and R.E.S. worked on the
numerical simulations, with T.B. calculating the correction factors. L.L. worked on
the theoretical predictions for comparison with the observations.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to R.R.
(rreyes@astro.princeton.edu).
Vol 464 | 11 March 2010 | doi:10.1038/nature08864

LETTERS
Deviations from a uniform period spacing of gravity
modes in a massive star
Pieter Degroote1, Conny Aerts1,2, Annie Baglin3, Andrea Miglio4, Maryline Briquet1, Arlette Noels4, Ewa Niemczura5,
Josefina Montalban4, Steven Bloemen1, Raquel Oreiro1, Maja Vučković1, Kristof Smolders1, Michel Auvergne3,
Frederic Baudin6, Claude Catala3 & Eric Michel3

The life of a star is dominantly determined by the physical pro-
cesses in the stellar interior. Unfortunately, we still have a poor
understanding of how the stellar gas mixes near the stellar core,
preventing precise predictions of stellar evolution1. The unknown
nature of the mixing processes as well as the extent of the central
mixed region is particularly problematic for massive stars2.
Oscillations in stars with masses a few times that of the Sun offer
a unique opportunity to disentangle the nature of various mixing
processes, through the distinct signature they leave on period spa-
cings in the gravity mode spectrum3. Here we report the detection
of numerous gravity modes in a young star with a mass of about
seven solar masses. The mean period spacing allows us to estimate
the extent of the convective core, and the clear periodic deviation
from the mean constrains the location of the chemical transition
zone to be at about 10 per cent of the radius and rules out a clear-
cut profile.
The young massive star HD 50230 is poorly studied, but it is
known to have spectral type B3V and a visual magnitude of 8.95.
HD 50230 is in its central nuclear-burning phase, just like the Sun,
transforming hydrogen into helium in its core (the so-called main
sequence phase). This important evolutionary phase covers about
90% of the life of the star, and the detailed internal structure of the
star during this phase determines its ultimate fate1,2. This star is at the
limiting mass of ending either as core-collapse supernova or as a
white dwarf, enriching the interstellar medium with helium and
heavy metals in the former case and carbon in the latter. A high-
quality continuous photometric light curve with 32-s sampling and
with a span of 137 days has been obtained for the star by the
Convection Rotation and Planetary Transits (CoRoT4) satellite. A
linear downward trend was removed from the data, after outliers
were excluded from the light curve. The duty cycle of the final set
of measurements analysed here is 88.6% and the noise has an ampli-
tude of 1 mmag (Fig. 1). The light curve reveals the presence of
numerous oscillation modes.
Stellar oscillations have been found to occur at different stages of
stellar life, for a range of stellar masses5. The best known case is that of
the solar oscillations, which are acoustic modes with periods near
5 min (refs 6–9). The seismic interpretation of the detected solar
oscillations led to a drastic improvement in the knowledge of the
internal structure of the Sun10,11. Meanwhile, similar acoustic modes
have been detected in various types of distant stars12–14. Gravity
modes, on the other hand, probe much deeper layers inside stars
and in principle allow the study of the core properties of stars far
better than acoustic modes. Although such modes have been detected
in massive stars similar to HD 5023015,16, where they have typical
periods of half a day to a few days, their seismic potential could
not be exploited because the number of detected modes in one star
was far too low.
Frequency analysis of HD 50230’s CoRoT light curve resulted in a
rich frequency spectrum (see Supplementary Information for an
outline of the analysis). We detected hundreds of gravity modes with
periods roughly in the range of half a day to five days and amplitudes
of millimagnitude level, as well as tens of high-frequency acoustic
modes with periods in the range of one to ten hours and with

–10 1.0
Δ Brightness (mmag)

–5 0.5
Δ Flux (%)

0 0

5 –0.5

10 –1.0
2,860 2,880 2,900 2,920 2,940 2,960 2,980
Time (days)

Figure 1 | The light curve of HD 50230 measured by the CoRoT satellite. Main panel, the light curve. Inset, magnified view of the boxed part of the light
curve. The residuals resulting from the frequency analysis are plotted in grey.
1
Instituut voor Sterrenkunde, KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium. 2IMAPP, Department of Astrophysics, Radboud University Nijmegen, PO Box 9010, 6500 GL
Nijmegen, The Netherlands. 3LESIA, UMR8109, Université Pierre et Marie Curie, Université Denis Diderot, Observatoire de Paris, 92195 Meudon, France. 4Institut d’Astrophysique et
de Géophysique de l’Université de Liege, Allée du 6 Aout 17, 4000 Liege, Belgium. 5Astronomical Institute, Wroclaw University, Kopernika 11, 51-622 Wroclaw, Poland. 6Institut
d’Astrophysique Spatiale (IAS), Batiment 121, F-91405, Orsay Cedex, France.

259
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
0.20
0.15
Amplitude (%)
0.10
0.05
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6
Period (days)
Figure 2 | Extracted frequency spectrum in the g mode region. The vertical
dashed lines denote the members of the detected period spacing.
considerably lower amplitudes of several micromagnitudes. This
high number of gravity mode frequencies is compatible with our
non-adiabatic excitation computations for an appropriate stellar
model based on an increase of the metal opacity at temperatures
around 200,000 K (ref. 17). The oscillation period content in the data
was extracted using classical and short-time Fourier analysis tech-
niques, combined with nonlinear fitting of the light curve5, and a
search for a recurrent period spacing was conducted18. A group of
eight consecutive periods with an almost constant spacing of 9,418 s
was detected in the data (Fig. 2). Monte Carlo simulations showed
that the probability of a coincidental occurrence of such spacing
among eight consecutive oscillation periods within the list of
detected frequencies is below 0.1%. Small deviations of up to 200 s
from the constant period spacing occur (Fig. 3). We interpret these as
being due to the presence of a region of varying chemical composi-
tion outside the convective core3. This region is progressively built up
during the main sequence by the boundary of the convective core
receding towards the centre.
The interior structure of massive stars basically consists of a con-
vective core surrounded by a radiative envelope. With this simplified
assumption, theory19 predicts the occurrence of a constant period
spacing among high-order gravity modes of the same degree. Such a
spacing has been found and interpreted previously in white dwarfs,
which are stellar remnants at the end of the life of stars like the
Sun20–22. They had hitherto not been detected in main-sequence stars.
The deviations from the equidistant period spacing carry informa-
tion on the details of the physical conditions in the stellar interior,
which are otherwise inaccessible for observational testing. Obser-
vations of binaries and stellar clusters show that a mixed zone larger
9,800
9,700
9,600
ΔPeriod (s)
9,500
9,400
9,300
9,200
9,100
1.0 1.1 1.2 1.3 1.4
Period (days)
Figure 3 | The components and values of the g mode period spacing. Filled
squares, the values of the period components (error bars, s.e.m.). Grey line, a
sinusoidal fit with decreasing amplitude, with a period of 2,450 s and
maximum amplitude of 240 s.
260
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
than what is imposed by the convection theory is required23. Such an
additional mixing could be the result of convective overshooting,
that is, the extension of convective motions above the boundary given
by the local treatment of convection used in the stellar modelling (the
mixing length theory). Other mixing processes include turbulent
diffusion or mixing induced by rotation. Theory predicts that the
relative importance of these various mixing processes can be separated,
as they leave a distinct marker on the spacing structure of gravity
modes3.
The deviations from the constant period spacing we determined
for HD 50230 allow us to evaluate the characteristics of the mixing.
To achieve this, the fundamental parameters of the star need to be
estimated. For this purpose, several ground-based high-resolution
1.8 2.0 spectra of HD 50230 were recorded with the CORALIE24 spectro-
graph attached to the 1.2-m Euler telescope at the La Silla Obser-
vatory in Chile, and with the High Efficiency and Resolution
Mercator Echelle Spectrograph (HERMES25) attached to the 1.2-m
Mercator telescope at La Palma. These follow-up spectra revealed a
decimal logarithmic gravity of 3.8 6 0.2 (c.g.s. units) and a projected
equatorial surface rotational velocity below 5 km s21 from the syn-
thesis of line profiles based on atmosphere models. The effective
temperature of the star was estimated to be 17,500 6 1,500 K from
its spectral energy distribution, based on archival photometry and
space spectroscopy measured with the International Ultraviolet
Explorer (see Supplementary Information for a description of the
fitting procedure). These results from classical spectroscopy are in
full agreement with the spectral type of B3V.
Stellar structure models in accordance with the observed fun-
damental parameters of HD 50230 were computed with the Code
Liégeois d’Évolution Stellaire (CLÉS26). We considered models with-
out extra mixing, as well as with instantaneous mixing due to core
overshooting over a distance parametrized as 0.1–0.3 local pressure
scale heights, or with parametrized turbulent mixing3. Under the
assumption of solar metallicity and dipolar oscillation modes, the
mean spacing of 9,418 s led us to conclude that the overshoot para-
meter needs to be at least 0.2 in the radiative regions adjacent to the
core to be compatible with the observed effective temperature (Fig. 4).
Lowering the metallicity or increasing the degree of the modes
increases the amount of extra mixing needed to fit the observed effec-
tive temperature.
The observed deviations from constant period spacing (Fig. 3) can
be characterized by a periodic behaviour3. Its periodicity gives an
indication of the evolutionary status in the main-sequence phase.
The current observations suggest that the period of the oscillatory
part is equal to 2,450 s, indicating that about 60% of the initial
12,000
11,000
Mean perios spacing (s)
10,000
9,000
8,000
7,000
6,000
5,000
4,000
4.40 4.35 4.30 4.25 4.20 4.15 4.10 4.05 4.00
Log10 [effective temperature (K)]
Figure 4 | The influence of the size of the convective mixing region on the
1.5 1.6
mean value of the period spacing. Dashed line, the location of the observed
period spacing and the estimate of the effective temperature. Solid lines, the
mean period spacing value as a function of the effective temperature, for
overshoot parameters of 0 (black), 0.2 (dark grey) and 0.3 (light grey), and
different masses (8–4 solar masses from left to right). The metallicity was
fixed at Z 5 0.02.
NATURE | Vol 464 | 11 March 2010
hydrogen in the centre of the star has already been consumed. The
amplitude of the periodic component is determined by the detailed
shape of the chemical composition gradient27. The observed small
amplitude (240 seconds), decreasing with increasing period, is a clear
signature of a smooth gradient of chemical composition. Such a
gradient is not compatible with an instantaneous mixing, as the latter
would leave such a sharp chemical composition gradient that the
amplitude of the periodic components of the deviations from con-
stant period spacing would have a much larger amplitude than that
observed—and show no variation with the period.
As a next step, the pressure modes of HD 50230 could in principle
be exploited, in addition to the gravity modes discussed here.
Without the establishment of a pressure mode frequency spacing,
such seismic exploitation of the pressure mode frequencies requires
an observational identification of their degree and azimuthal order.
This can only be achieved using high-precision multicolour pho-
tometric28 and high-resolution spectroscopic time series29,30, cover-
ing a time base similar to that of the CoRoT space photometry.
Received 9 November 2009; accepted 19 January 2010.
1. Schaller, G., Schaerer, D., Meynet, G. & Maeder, A. New grids of stellar models
from 0.8 to 120 solar masses at Z50.020 and Z50.001. Astron. Astrophys. Suppl.
Ser. 96, 269–331 (1992).
2. Maeder, A. Physics, Formation and Evolution of Rotating Stars (Springer, 2009).
3. Miglio, A., Montalban, J., Noels, A. & Eggenberger, P. Probing the properties of
convective cores through g modes: high-order g modes in SPB and Gamma
Doradus stars. Mon. Not. R. Astron. Soc. 386, 1487–1502 (2008).
4. Auvergne, M. et al. The CoRoT satellite in flight: description and performance.
Astron. Astrophys. 506, 411–424 (2009).
5. Aerts, C., Christensen-Dalsgaard, J. & Kurtz, D. W. Asteroseismology (Springer,
2009).
6. Leighton, R. B., Noyes, R. W. & Simon, G. W. Velocity fields in the solar
atmosphere. Astrophys. J. 135, 474–499 (1962).
7. Ulrich, R. K. The five-minute oscillations on the solar surface. Astrophys. J. 162,
993–1002 (1970).
8. Leibacher, J. W. & Stein, R. F. A new description of the solar five-minute
oscillation. Astrophys. Lett. 7, 191–192 (1971).
9. Deubner, F. L. Observations of low wavenumber nonradial eigenmodes of the sun.
Astron. Astrophys. 44, 371–375 (1975).
10. Christensen-Dalsgaard, J. Helioseismology. Rev. Mod. Phys. 74, 1073–1129 (2002).
11. Chaplin, W. J. & Basu, S. Perspectives in global helioseismology and the road
ahead. Sol. Phys. 1–2, 53–75 (2008).
12. Bedding, T. R. & Kjeldsen, H. Solar-like oscillations. Publ. Astron. Soc. Aust. 2,
203–212 (2003).
13. De Ridder, J. et al. Non-radial oscillation modes with long lifetimes in giant stars.
Nature 459, 398–400 (2009).
14. Michel, E. et al. CoRoT measures solar-like oscillations and granulation in stars
hotter than the sun. Science 322, 558–560 (2008).
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
15. Waelkens, C. Slowly pulsating B stars. Astron. Astrophys. 246, 453–468 (1991).
16. Krisciunas, K. A new class of pulsating stars. Bull. Am. Astron. Soc. 25, 1442–1443
(1993).
17. Dupret, M. A. Nonradial nonadiabatic stellar pulsations. Astron. Astrophys. 366,
166–173 (2001).
18. Degroote, P. et al. CoRoT’s view of newly discovered B-star pulsators: results for
358 candidate B pulsators from the initial run’s exoplanet field data. Astron.
Astrophys. 506, 471–489 (2009).
19. Tassoul, M. Asymptotic approximations for stellar nonradial pulsations.
Astrophys. J. Suppl. Ser. 43, 469–490 (1980).
20. Winget, D. E. et al. Asteroseismology of the DOV star PG 1159–035 with the
Whole Earth Telescope. Astrophys. J. 378, 326–346 (1991).
21. Winget, D. E. et al. Whole earth telescope observations of the DBV white dwarf
GD358. Astrophys. J. 430, 839–849 (1994).
22. Brassard, P., Fontaine, G., Wesemael, F. & Hansen, C. J. Adiabatic properties of
pulsating DA white dwarfs. II — Mode trapping in compositionally stratified
models. Astrophys. J. Suppl. Ser. 80, 369–401 (1992).
23. Andersen, J., Clausen, J. V. & Nordstrom, B. New strong evidence for the
importance of convective overshooting in intermediate-mass stars. Astrophys. J.
363, L33–L36 (1990).
24. Baranne, A. et al. ELODIE: a spectrograph for accurate radial velocity
measurements. Astron. Astrophys. Suppl. Ser. 119, 373–390 (1996).
25. Raskin, G. & Van Winckel, H. Hermes, a high-resolution fiber-fed spectrograph for
the Mercator Telescope. Proc. SPIE 7014, 178–189 (2008).
26. Scuflaire, R. et al. CLÉS, Code Liégeois d’Évolution Stellaire. Astrophys. Space Sci.
316, 83–91 (2008).
27. Noels, A., Montalban, J., Miglio, A., Godart, M. & Ventura, P. Overshooting and
semiconvection. Astrophys. Space Sci. 326, 233–243 (2009).
28. Aerts, C. et al. Asteroseismology of HD 129929: core overshooting and nonrigid
rotation. Science 300, 1926–1928 (2003).
29. Aerts, C. Core overshooting and nonrigid internal rotation of massive stars. Proc.
IAU 250, 237–244 (2008).
30. Aerts, C. & De Cat, P. Beta Cep stars from a spectroscopic point of view. Space Sci.
Rev. 105, 453–492 (2003).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements The research leading to these results received funding from
the European Research Council under the European Community’s Seventh
Framework Programme (PROSPERITY), from the Research Council of KU Leuven
and from the Belgian Federal Science Policy. The CoRoT space mission has been
developed, and is operated by, CNES, with contributions from Austria, Belgium,
Brazil, ESA, Germany and Spain.
Author Contributions P.D. and C.A. analysed and interpreted the light curve; M.B.,
A.M., A.N. and J.M. computed theoretical models; P.D., M.B. and E.N. analysed
spectroscopic data; S.B., R.O., M.V. and K.S. made spectroscopic observations; and
A.B., M.A., F.B., C.C. and E.M. are the CoRoT instrument builders.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to P.D.
(pieter.degroote@ster.kuleuven.ac.be).
261

LETTERS
Transmission of electrical signals by spin-wave
interconversion in a magnetic insulator
Y. Kajiwara1,2, K. Harii1, S. Takahashi1,3, J. Ohe1,3, K. Uchida1, M. Mizuguchi1, H. Umezawa5, H. Kawai5, K. Ando1,2,
K. Takanashi1, S. Maekawa1,3 & E. Saitoh1,2,4
The energy bandgap of an insulator is large enough to prevent
electron excitation and electrical conduction1. But in addition to
charge, an electron also has spin2, and the collective motion of spin
can propagate—and so transfer a signal—in some insulators3. This
motion is called a spin wave and is usually excited using magnetic
fields. Here we show that a spin wave in an insulator can be gen-
erated and detected using spin-Hall effects, which enable the direct
conversion of an electric signal into a spin wave, and its sub-
sequent transmission through (and recovery from) an insulator
over macroscopic distances. First, we show evidence for the trans-
fer of spin angular momentum between an insulator magnet
Y3Fe5O12 and a platinum film. This transfer allows direct conver-
sion of an electric current in the platinum film to a spin wave in the
Y3Fe5O12 via spin-Hall effects4–11. Second, making use of the trans-
fer in a Pt/Y3Fe5O12/Pt system, we demonstrate that an electric
current in one metal film induces voltage in the other, far distant,
metal film. Specifically, the applied electric current is converted
into spin angular momentum owing to the spin-Hall effect7,8,10,11 in
the first platinum film; the angular momentum is then carried by a
spin wave in the insulating Y3Fe5O12 layer; at the distant platinum
film, the spin angular momentum of the spin wave is converted
back to an electric voltage. This effect can be switched on and off
using a magnetic field. Weak spin damping3 in Y3Fe5O12 is
responsible for its transparency for the transmission of spin angu-
lar momentum. This hybrid electrical transmission method poten-
tially offers a means of innovative signal delivery in electrical
circuits and devices.
Figure 1 | Two types of non-equilibrium spin currents in solids. a, A
schematic illustration of a conduction-electron spin current: spin angular
momentum JS carried by electron diffusion. b, A schematic illustration of a
spin-wave spin current: spin angular momentum carried by collective
magnetic-moment precession. c, A schematic illustration of the spin-
1
Institute for Materials Research, Tohoku University, Sendai 980-8577, Japan. 2Department of Applied Physics and Physico-Informatics, Keio University, Yokohama 223-8522,
Japan. 3CREST, 4PRESTO, Japan Science and Technology Agency, Sanbancho, Tokyo 102-0075, Japan. 5FDK Corporation, Shizuoka 431-0495, Japan.
262
©2010 Macmillan Publishers Limited. All rights reserved
Vol 464 | 11 March 2010 | doi:10.1038/nature08876
A flow of spin angular momentum is called a spin current2. In
solids, there are two types of carriers for non-equilibrium spin cur-
rents. One is a conduction electron2,12,13 (Fig. 1a). The other is col-
lective motion of magnetic moments (Fig. 1b)—spin waves14,15,
comprising magnetostatic and exchange spin-wave modes14,15.
Here we call a spin current carried by spin waves a ‘spin-wave spin
current’ (see Supplementary Information section A for details).
Extensive studies of conduction-electron spin currents in metals
and semiconductors have clarified that the currents have a critical
problem; they disappear within a very short distance, typically hun-
dreds of nanometres16. In contrast, it has been shown that a spin-
wave spin current may persist for much greater distances because it is
carried by the collective motion of spins coupled by exchange inter-
action14,15. Significantly, a spin-wave spin current exists even in mag-
netic insulators, in which its decay is typically suppressed. This is
because the decay is caused mainly by conduction electrons, which
are absent in insulators. For instance, in the magnetic insulator
Y3Fe5O12, the spin-wave decay length can be several centimetres3
and thus the waves are propagated over a relatively long distance;
Y3Fe5O12 is an ideal conductor for spin-wave spin currents even
though it is an insulator for electric currents.
To make use of the spin-wave spin currents in insulators, it is
necessary to find methods for getting a d.c. spin current into and
out of the insulators. We show that this can be done by using spin
pumping and spin-transfer torque (STT). Here, spin pumping refers
to the transfer of spin angular momentum from magnetization-
precession motion to conduction-electron spin9,17–19, a phenomenon
pumping detection mechanism in the present system. If the magnetization
(M) dynamics in the Y3Fe5O12 layer (on the top face of the block) pumps a
spin current JS into the Pt layer, the current generates electromotive force
ESHE via ISHE.
NATURE | Vol 464 | 11 March 2010
which allows generation of a spin current from magnetization
motion. STT is, in contrast, the reverse process of this spin pumping,
that is, the transfer of angular momentum from conduction-electron
spin to magnetization20–22: the magnetization receives torque by
absorbing a spin current20–22. These two phenomena enable the
mutual conversion of angular momentum between conduction-elec-
tron spin and magnetization. However, up to now, experiments on
these phenomena have been limited to electric conductors. In this
Letter we show that by using Pt/Y3Fe5O12 films, both phenomena23
occur even at insulator/metal interfaces, and the phenomena allow
transmission of a d.c. electric signal through the insulator for a long
distance in a controllable manner at room temperature.
First, we show evidence for spin pumping across a Pt/garnet-type
Y3Fe5O12 interface. Y3Fe5O12 is a ferrimagnetic insulator whose
charge gap is 2.7 eV. Owing to this huge gap, Y3Fe5O12 exhibits very
high resistivity (,1012 V cm at room temperature, greater than that
of air). Also, the magnetization damping is very small:
a < 6.7 3 1025, where a is the Gilbert damping coefficient19,24 for
the sample used in the present study (Fig. 2b). Figure 2a is a schematic
illustration of the experimental set-up. The sample is a bilayer film
composed of a single-crystal Y3Fe5O12 layer and a Pt layer. Here, the
a c
Microwaves
dP/dH (a.u.)
V H
H 0
θ
Y3Fe5O12
Pt
b d
N=1
dV/dH (μV Oe–1)
SWR 3
7
5
dP/dH (a.u.)
0
V H
–0.5
4.73 4.75
H (kOe)
e f
Vmax(θ)/Vmax(θ = 90°)
ISHE 1 Exp.
90º Sinθ
V (μV)
0 0
0 observed voltage generation; electromagnetic artefacts are irrelevant.
2
θ = –90º
–1
2.5 2.6 2.7 –90º
H (kOe)
Figure 2 | Spin pumping in Pt/Y3Fe5O12. a, Schematic illustration of the
experimental set-up. The sample is a Pt/Y3Fe5O12 bilayer film, 3 mm
3 1 mm, composed of a 1.3-mm-thick garnet-type Y3Fe5O12 layer and a 10-
nm-thick Pt layer. The voltage (V) electrodes attached to the Pt film are
3 mm apart. H denotes the in-plane external magnetic field. b, SWR
spectrum measured when a magnetic field is perpendicular to the film.
Arrows, fitting results for the exchange spin-wave resonance fields29. N, spin-
wave mode number along the direction perpendicular to the film. Spectral
width and a for the N 5 1 mode was estimated via a fitting procedure using
Lorentz-type dispersion functions (Supplementary Information section E).
P and H, the microwave absorption intensity and the strength of
H, respectively. c, Ferromagnetic SWR spectrum for the Pt/Y3Fe5O12 film at
h 5 90u (h is defined in a). d, H dependence of dV/dH for the Pt/Y3Fe5O12
film measured by applying 1 mW microwaves at h 5 90u. Inset, H
dependence of V for the Pt/Y3Fe5O12 film at h 5 90u. Galvanomagnetic
effects24 in magnets are irrelevant to V, as Y3Fe5O12 is an insulator. e, H
dependence of V measured by applying the microwaves at various values of
h. f, h dependence of the maximum peak values Vmax in the V(H) curves
measured with application of microwaves. g, Microwave-power dependence
of Vmax. The experimental data (filled circles) are well reproduced by a curve
(solid line) calculated from a longitudinal spin-pumping model.
LETTERS
Pt layer is used as a spin-current detector, in which the inverse spin-
Hall effect9–11 (ISHE) converts a spin current into electromotive force
ESHE via the spin–orbit interaction. In ISHE, when a spin current
carries the spin polarization s along the spatial direction JS, ESHE is
given by9,11 (Fig. 1c):
ESHE // JS 3 s (1)
ISHE is known to be enhanced in heavy noble metals such as Pt (refs
9, 11).
During the measurements, a static in-plane magnetic field H is
applied and the sample is placed at the centre of a 9.44 GHz micro-
wave cavity. When H fulfils the ferromagnetic spin-wave resonance
(SWR) conditions14,15, precession of magnetization is induced. If this
precession pumps spin currents into the Pt layer, then voltage is
generated in the Pt layer via ISHE; spin pumping in this system is
sensitively detected by measuring the voltage difference V between
the ends of the Pt layer9,11.
Figures 2c and 2d are the SWR spectrum and the H dependence of
dV/dH, respectively, both measured with microwaves applied and
with the external magnetic field perpendicular (h 5 90u) to the dir-
ection across the electrodes. In the SWR spectrum, many resonance
signals appear, each dispersion corresponding to a SWR mode14,15 in
SWR V
the Y3Fe5O12 layer. These resonance fields are much greater than the
θ = 90º in-plane magnetization saturation field (HC 5 20 Oe) of this
Y3Fe5O12 film. In the dV/dH spectrum, as shown in Fig. 2d, multiple
peaks appear at these SWR fields, indicating that electromotive force
is induced in the Pt layer concomitant with SWR in the Y3Fe5O12
layer. Figure 2g shows the microwave-power dependence of the max-
imum peak values Vmax in the V(H) curves. The experimental data
(filled circles) are well reproduced by a curve (solid line) calculated
4
0.5 ISHE from a longitudinal spin-pumping model in which the spin accu-
V (µV)
θ = 90º
0
mulation in the Pt layer is taken into consideration (see
2.5 2.7
H (kOe) Supplementary Information section B for details). This voltage signal
0 was found to disappear in a Cu/Y3Fe5O12 system, where the Pt layer is
replaced by a Cu layer in which the spin–orbit interaction is very
weak22, indicating the important role of spin–orbit interaction, or
ISHE, in the voltage generation. The voltage signal also disappears in
2.5 2.7
H (kOe) a Pt/SiO2/Y3Fe5O12 system, where the Y3Fe5O12 and the Pt layers are
g separated by a thin (10 nm) film of insulating SiO2, and also in a Pt/
4 Gd3Ga5O12 system, where the Y3Fe5O12 layer is replaced by a para-
Exp.
Cal. magnetic Gd3Ga5O12 layer. These last two results indicate that direct
contact between the magnet Y3Fe5O12 and Pt is necessary for the
V (μV)
2
By changing the field direction h, the V signal for Pt/Y3Fe5O12 dis-
appears at h 5 0 (Fig. 2e, f) and then changes its sign at 290u , h , 0.
0 90º 0 0.2 0.4 0.6 This behaviour is consistent with equation (1), and is direct evidence
θ P (mW)
for ISHE induced by spin pumping from the insulator Y3Fe5O12.
The spin pumping in Pt/Y3Fe5O12 can be attributed to the small
but finite spin-exchange interaction17,19 between a conduction elec-
tron in Pt and a localized moment in Y3Fe5O12, or to the mixing
conductance of the conduction electrons23 at the interface. By taking
this interaction into account in coupled equations for the magnet-
ization and spin accumulation (the Landau–Lifshitz–Gilbert equa-
tion24 and the Bloch equation with spin diffusion17,19), we obtain the
pumped spin current, which when combined with ISHE9–11 allows
calculation of V as a function of the spin-exchange interaction at the
interface. From the experimental values of V together with values of
microwave field strength and the spin-Hall angle for Pt (ref. 11), the
magnitude of the spin-exchange energy density at the interface is
estimated to be ,16 erg cm22; alternatively, the magnitude of the
mixing conductance23 is estimated as 3 3 1012 cm22 (see
Supplementary Information section B for details).
The above observation of the spin pumping in Pt/Y3Fe5O12 sug-
gests the possibility of the reverse process: STT20–22 acting on the
insulator Y3Fe5O12. Second, we demonstrate STT across the Pt/
Y3Fe5O12 interface using the same system as follows. We applied an
263
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
electric current J to the Pt layer (Fig. 3a). In the Pt layer, J is converted
into a spin current via the spin-Hall effect7,8,11,22. This spin current
exerts STT on the magnetization in the Y3Fe5O12 layer. This is the
reverse process of the aforementioned spin pumping effect. When
J . 0 (J , 0) at h 5 90u (h 5 290u), STT is antiparallel to the mag-
netization damping torque in the Y3Fe5O12 layer22 (Fig. 3b). Here, h
represents the angle between J and the magnetization direction of the
Y3Fe5O12 layer. In this case, if the magnitude of STT is greater than
that of the damping torque, the damping is cancelled out by STT.
This means that when a sufficient amount of current is applied, the
a b
H DT
θ STT
Y3Fe5O12
J σ
M = Pt
M/Y3Fe5O12
d e
M = Pt
S(j) – S(–j)
20 θ = 0
20
4.5
f 5.5
H = 1.2 kOe
θ = 90º
S(j) – S(–j) (fW MHz–1)
0 0
4.5 5 5.5 4.5
f (GHz)
g h
5.2
W(j) – W(–j) (pW)
f0 (GHz)
5
0 10
1 j (108 A m–2) 1
M = Pt
0 5 10 0 5
j (108 A m–2)
Figure 3 | Magnetization oscillation induced by spin-transfer torque in Pt/
Y3Fe5O12. a, Schematic illustration of the experimental set-up; the sample
film is that shown in Fig. 2a. h denotes the angle between the external in-
plane magnetic field H and the direction along the electric current J in the Pt
layer. b, c, Schematic illustrations of the directions of spin-transfer torque
(STT) acting on the magnetization (M) and the magnetization-damping
torque (DT) of the Y3Fe5O12 layer at h 5 90u when J . 0 (b) and J , 0 (c). J,
amplitude of J; s, spin polarization of the spin current (SC) induced from
J by the spin-Hall effect in the Pt layer. d, e, DS(j) ; S(j) 2 S(2j) spectra for
various values of the current density j in the Pt layer at h 5 90u for Pt/
Y3Fe5O12 (d) and Cu/Y3Fe5O12 (e) films when H 5 1.2 kOe. S(j), the
frequency f power spectrum of the microwaves (see Methods). In DS(j),
background noise is eliminated as the antisymmetric component with
respect to j is extracted. Inset to d, DS(j) spectrum at h 5 0 for the Pt/
Y3Fe5O12 film measured with application of j 5 11.1 3 108 A m22 when
H 5 1.2 kOe. f, DS(j) spectra for various magnetic field strengths H for the
Pt/Y3Fe5O12 film measured when j 5 13.3 3 108 A m22 at h 5 90u.
g, h, Values of the frequency integral from 4.5 GHz up to 5.5 GHz of the DS(j)
spectra for the Pt/Y3Fe5O12 film (shown in d), and for the Cu/Y3Fe5O12 film
(shown in e), respectively, as functions of j. i, H dependence of the centroid
frequencies f0 of the DS(j) spectra shown in f.
264
NATURE | Vol 464 | 11 March 2010
magnetization spontaneously oscillates20–22 with the eigenfrequen-
cies14 and emits electromagnetic waves. As the magnetization damp-
ing in Y3Fe5O12 is very small3, this cancellation is achieved by a small
amount of STT. In contrast, when J , 0 (J . 0) at h 5 90u
(h 5 290u), STT and the damping torque are parallel and cannot
be cancelled out22 (Fig. 3c).
We measured the power spectra S(j), where j is electric current
density, of microwaves emitted from the Pt/Y3Fe5O12 film using an
antenna while applying an electric current to the Pt layer. In Fig. 3d,
we show DS(j) ; S(j) 2 S(2j) spectra at various values of j when
H 5 1.2 kOe (.HC). When the applied magnetic field is in the dir-
ection along the electric current, h 5 0, there are no emission signals
in DS(j). In contrast, when the field is perpendicular to the current,
Microwaves c h 5 90u, multiple peak signals appear in the spectra when j is greater
M M than 4.4 3 108 A m22 (;jc), as shown in Fig. 3d. The peak frequen-
H H cies in the spectra vary with the magnetic field (Fig. 3f and i). The
solid curve in Fig. 3i represents the ferromagnetic resonance frequen-
DT STT cies calculated from the Kittel equation (see Supplementary
Information section C for details); these peaks evidently originate
σ from the spontaneous magnetization oscillations induced when
SC
SC j . jc. The multiple peaks (see Fig. 3d) may be attributed to magneto-
static spin-wave modes which are excited simultaneously in the
f
M = Cu M = Pt θ = 90º
Y3Fe5O12 layer (see Supplementary Information section D for
details), but the strong change in the whole spectral shape with cur-
j = 13.3 108 A m–2
20
20 rent and field strengths suggests the appearance of chaos25 or the
H = 1.15 kOe spin-diffusion-induced instability26 in magnetization dynamics. In
H = 1.2 kOe
θ = 90º the j dependence of the frequency-integrated intensities for the
S(j) – S(–j) (fW MHz–1)
j = 11.1 108 A m–2 1.10
DS(j) spectra, a clear threshold is observed at jc (Fig. 3g). This implies
10.0
1.05
that STT compensates the magnetization damping torque at j 5 jc.
8.9
The centroid frequency of the emission spectra decreases slightly with
1.00 the current, a tendency which follows that in current-injection-type
6.7
magnetization oscillations27,28. We checked that these peak signals
0.95
4.4 disappear both in Pt/Gd3Ga5O12 and Cu/Y3Fe5O12 (Fig. 3e and h)
3.3 0.90
films.
2.2
Finally, we show electric-signal transmission in the insulator
0.85 Y3Fe5O12 film by making use of these phenomena together.
0
0 Figure 4a is a schematic illustration of the experimental set-up.
5 5.5 4 4.5 5 Two Pt films (i and o) are sputtered on a single-crystal Y3Fe5O12 film
f (GHz) f (GHz) and an electric current is applied to the Pt film i. The distance
i between the films i and o is 1 mm, shorter than the spin-wave decay
length in Y3Fe5O12 crystals3. In this set-up, the electric current
Exp.
8
applied to the Pt film i induces magnetization oscillation in the
(GHz)
Cal.
Y3Fe5O12 layer due to the STT across the Pt/Y3Fe5O12 interface, as
f0
4 demonstrated in Fig. 3. This magnetization oscillation then propa-
M = Cu M = Pt gates in the Y3Fe5O12 layer via a spin-wave spin current. Figure 4b
shows a numerical calculation of this propagation. When the oscil-
10 0 1 2
H (kOe) lation reaches the second interface, it generates electric voltage in the
j (108 A m–2)
Pt film o via spin pumping and ISHE, as demonstrated in Fig. 2. As
shown in Figs 1c and 3b, these effects are activated when j . 0 (j , 0)
at h 5 90u (h 5 290u)9,22. We measured the voltage generated in the
Pt film o while applying electric currents to the Pt film i.
Figures 4c and d show the voltage difference V between the ends of
the Pt film o as a function of the electric current density j in the Pt film
i. When the magnetization of the Y3Fe5O12 layer is along the electric
current, h 5 0 and 180u, no signal appears in V (Fig. 4c). At h 5 0 and
180u, the above spin-transfer and voltage-generation mechanisms are
inactive, and this result also confirms that the two Pt films are
well isolated electrically. When h 5 90u (h 5 290u), in contrast, the
voltage V signal appears with application of a current
j . 6.0 3 108 A m22 (j , 26.0 3 108 A m22), as shown in Fig. 4d.
We found that there is a clear threshold at 6.0 3 108 A m22 in the j
dependence of V (Fig. 4d). This threshold current density is compar-
able to that for the current-induced magnetization oscillation shown
in Fig. 3g. All these results are consistent with the prediction that
electric-signal transmission in the Y3Fe5O12 is activated when j . jjcj
(j , 2jjcj) at h 5 90u (h 5 290u), where jjcj is a threshold current
density. If the distance between the Pt films were less than
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 LETTERS

a b c d 1.2
Pt/Y3Fe5O12

V (nV)
Pt film i
V 1 H = 2.3 kOe 1 θ = 90º
Y3Fe5O12 Pt (o) ESHE
1
0 H (kOe) 3
Pt (i) H
θ θ = 0º

V (nV)

V (nV)
J 0 0
1 mm
θ = 180º
y
z
y x x –1 –1 θ = –90º Pt/Y3Fe5O12
JS H = 2.3 kOe
0 0.5 1.0
|my,z| (a.u.) –10 –5 0 5 10 –10 –5 0 5 10
j (108 A m−2) j (108 A m−2)

Figure 4 | Electric-signal transmission via spin-wave spin currents. a, A
schematic illustration of the experimental set-up. The sample is a 1.3-mm-
thick single-crystal Y3Fe5O12 (111) film on which two separate 15-nm-thick
Pt films (i and o) are sputtered. The distance between the Pt films is 1 mm.
The surfaces of the Y3Fe5O12 layer, Pt film i and Pt film o are rectangular
shapes of area (mm2) 35, 27.5 and 0.5, respectively. The distance between the
voltage electrodes (V) attached to the Pt film o is 5 mm. b, In-plane spatial
distribution of the time average of the magnetization-precession amplitude
| my,z | in the Y3Fe5O12 layer numerically calculated using a stochastic
Landau–Lifshitz–Gilbert equation30 at room temperature. In the calculation,
STT22 that compensates the magnetization-damping torque at the Pt film
i/Y3Fe5O12 interface (dashed rectangle) is taken into consideration. Time
average is taken for 1 ms (Supplementary Information section F). c, d, V as a
function of j in the Pt film i at h 5 0u (red curve in c), h 5 180u (blue curve in
c), h 5 90u (red curve in d) and h 5 290u(blue curve in d). h is defined in
a. An in-plane magnetic field of 2.3 kOe is applied. Inset to d, V at
j 5 16.6 3 108 A m22 as a function of H (0.2 kOe , H , 3 kOe) when
h 5 90u.

nanometres, electron tunnelling could convey spin angular Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
momentum; in the present macroscopic-sized system, in contrast,
the tunnelling is clearly irrelevant. We confirmed that the electric- Received 28 September 2009; accepted 1 February 2010.
signal transmission disappears again both in Pt/Gd3Ga5O12/Pt and
1.
Cu/Y3Fe5O12/Cu systems. The results also indicate that the electric
polarization in the insulator is irrelevant. The inset to Fig. 4d shows V 2.
at j 5 16.6 3 108 A m22 as a function of magnetic field strength H
(0.2 kOe , H , 3 kOe) when h 5 90u. In this field range, V is min- 3.
imally affected by the field-strength change and the role of the field
4.
seems to be no more than that of aligning the magnetization dir-
ection. This j dependence of V at h 5 90u above j 5 6.0 3 108 A m22 5.
deviates from the linear dependence observed in Fig. 3g. This might 6.
be because not all the modes contribute equally to this transmission
7.
and the population of each mode may depend on the excitation
strength (because of the intermode coupling or the spin-wave non- 8.
linearity), but this discrepancy needs to be quantitatively elucidated.
The observed voltage transmission in an insulator provides a new
9.
method of signal transfer, and opens the door to insulator-based
spintronics. The observed magnetization oscillation induced by the
spin-Hall effect could also be applied to the construction of a micro- 10.
wave generator. We note that spin pumping from the insulator
11.
enables spin injection free from the conventional impedance-match-
ing condition12. Finally, we anticipate that use of this spin transfer in 12.
insulators will lead to substantial advances in spintronics and elec-
tronics. 13.
14.
METHODS SUMMARY 15.
A single-crystal Y3Fe5O12 (111) film was grown on a Gd3Ga5O12 (111) single-
crystal substrate by liquid phase epitaxy. For the film growth, we used PbO-B2O3 16.
flux around 1,200 K. Then, a 10-nm-thick Pt layer was sputtered on the Y3Fe5O12
layer. Immediately before the sputtering, the surface was cleaned through the
metal mask by Ar-ion bombardment in a vacuum. For the spin pumping mea- 17.
surements shown in Fig. 2, the Pt/Y3Fe5O12 sample system was placed near the
centre of a TE011 microwave cavity; at this position, the magnetic-field compon-
18.
ent of the microwave mode is maximized while the electric-field component is
minimized. The microwave power was less than 10 mW, a value lower than the 19.
saturation of the ferromagnetic resonance absorption for the present sample. For
measuring voltage induced by the spin pumping, a twisted pair of thin coated Cu
wires (0.08 mm in diameter) are connected to the ends of the Pt layer. Microwave 20.
emission spectra were measured by attaching a gold coplanar-waveguide
antenna to the Pt surface of the Pt/Y3Fe5O12 sample film. The microwave signal 21.
received by the antenna was led to an amplifier via a microwave probe. The
22.
amplified microwave signal was analysed and recorded by a spectrum analyser.
Micromagnetic simulation was performed by solving numerically the Landau– 23.
Lifshitz–Gilbert equation in which the spin torque at the interface cells are taken
into consideration (for details, see Methods). 24.
©2010 Macmillan Publishers Limited. All rights reserved
Ashcroft, N. W. & Mermin, N. D. Solid State Physics Ch. 9 (Saunders College,
1976).
Maekawa, S. (ed.) Concepts in Spin Electronics Ch. 7 and 8 (Oxford Univ. Press,
2006).
Schneider, T. et al. Realization of spin-wave logic gates. Appl. Phys. Lett. 92,
022505 (2008).
Dyakonov, M. I. & Perel, V. I. Current-induced spin orientation of electrons in
semiconductors. Phys. Lett. A 35, 459–460 (1971).
Hirsch, J. E. Spin Hall effect. Phys. Rev. Lett. 83, 1834–1837 (1999).
Takahashi, S. & Maekawa, S. Hall effect induced by a spin-polarized current in
superconductors. Phys. Rev. Lett. 88, 116601 (2002).
Kato, Y. K., Myers, R. C., Gossard, A. C. & Awschalom, D. D. Observation of the
spin Hall effect in semiconductors. Science 306, 1910–1913 (2004).
Wunderlich, J., Kaestner, B., Sinova, J. & Jungwirth, T. Experimental observation of
the spin-Hall effect in a two-dimensional spin-orbit coupled semiconductor
system. Phys. Rev. Lett. 94, 047204 (2005).
Saitoh, E., Ueda, M., Miyajima, H. & Tatara, G. Conversion of spin current into
charge current at room temperature: inverse spin-Hall effect. Appl. Phys. Lett. 88,
182509 (2006).
Valenzuela, S. O. & Tinkham, M. Direct electronic measurement of the spin Hall
effect. Nature 442, 176–179 (2006).
Kimura, T., Otani, Y., Sato, T., Takahashi, S. & Maekawa, S. Room-temperature
reversible spin Hall effect. Phys. Rev. Lett. 98, 156601 (2007).
Valet, T. & Fert, A. Theory of the perpendicular magnetoresistance in magnetic
multilayers. Phys. Rev. B 48, 7099–7113 (1993).
Takahashi, S. & Maekawa, S. Spin current in metals and superconductors. J. Phys.
Soc. Jpn 77, 031009 (2008).
Kittel, C. Introduction to Solid State Physics 8th edn, Ch. 12 and 13 (Wiley, 2005).
Demokritov, S. O., Hillebrands, B. & Slavin, A. N. Brillouin light scattering studies of
confined spin waves: linear and nonlinear confinement. Phys. Rep. 348, 441–489
(2001).
Bass, J. & Pratt, W. P. Jr. Spin-diffusion lengths in metals and alloys, and spin-
flipping at metal/metal interfaces: an experimentalist’s critical review. J. Phys.
Condens. Matter 19, 183201 (2007).
Silsbee, R. H., Janossy, A. & Monod, P. Coupling between ferromagnetic and
conduction-spin-resonance modes at a ferromagnetic-normal-metal interface.
Phys. Rev. B 19, 4382–4399 (1979).
Tserkovnyak, Y., Brataas, A. & Bauer, G. E. W. Enhanced Gilbert damping in thin
ferromagnetic films. Phys. Rev. Lett. 88, 117601 (2002).
Mizukami, S., Ando, Y. & Miyazaki, T. Effect of spin diffusion on Gilbert damping
for a very thin permalloy layer in Cu/permalloy/Cu/Pt films. Phys. Rev. B 66,
104413 (2002).
Kiselev, S. I. et al. Microwave oscillations of a nanomagnet driven by a spin-
polarized current. Nature 425, 380–383 (2003).
Ji, Y., Chien, C. L. & Stiles, M. D. Current-induced spin-wave excitations in a single
ferromagnetic layer. Phys. Rev. Lett. 90, 106601 (2003).
Ando, K. et al. Electric manipulation of spin relaxation using the spin Hall effect.
Phys. Rev. Lett. 101, 036601 (2008).
Brataas, A., Bauer, G. E. W. & Kelly, P. J. Non-collinear magnetoelectronics. Phys.
Rep. 427, 157–255 (2006).
Chikazumi, S. Physics of Ferromagnetism Ch. 20 and 21 (Oxford Univ. Press, 1997).
265
LETTERS
25. Stiles, M. D., Xiao, J. & Zangwill, A. Phenomenological theory of current-induced
magnetization precession. Phys. Rev. B 69, 054408 (2004).
26. Wigen, P. E., Doetsch, H., Ming, Y., Baselgia, L. & Waldner, F. Chaos in magnetic
garnet thin films. J. Appl. Phys. 63, 4157–4159 (1988).
27. Rippard, W. H., Pufall, M. R., Kaka, S., Russek, S. E. & Silva, T. J. Direct-current
induced dynamics in Co90Fe10/Ni80Fe20 point contacts. Phys. Rev. Lett. 92,
027201 (2004).
28. Krivorotov, I. N. et al. Large-amplitude coherent spin waves excited by spin-
polarized current in nanoscale spin valves. Phys. Rev. B 76, 024418 (2007).
29. Laulicht, I., Suss, J. T. & Barak, J. The temperature dependence of the
ferromagnetic and paramagnetic resonance spectra in thin yttrium-iron-garnet
films. J. Appl. Phys. 70, 2251–2258 (1991).
30. Brown, W. F. Jr. Thermal fluctuations of a single-domain particle. Phys. Rev. 130,
1677–1686 (1963).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank K. Sato, Y. Suzuki, Y. Tserkovnyak, G. Tatara,
T. Ishibashi and K. M. Itoh for discussions. This work was supported by a
266
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
Grant-in-Aid for Scientific Research in Priority Area ‘Creation and control of spin
current’ (19048028) from MEXT, Japan, a Grant-in-Aid for Scientific Research (A)
from MEXT, Japan, the global COE for the ‘Materials integration international
centre of education and research’ and ‘High-level global cooperation for
leading-edge platform on access spaces (C12)’ from MEXT, Japan, a Grant for
Industrial Technology Research from NEDO, Japan, and Fundamental Research
Grant from TRF, Japan.
Author Contributions Y.K., K.H., K.U. and K.A. performed the measurements and
analysed the data; J.O. carried out the numerical analysis; S.T., S.M. and E.S.
provided the theoretical analysis; H.U. and H.K. contributed to the sample
fabrication; Y.K., K.H, K.U., M.M. and K.T. contributed to the experimental set-up;
Y.K., S.T., J.O., K.U., M.M., H.U., K.T., S.M. and E.S. wrote the manuscript; all authors
discussed the results and commented on the manuscript; and E.S. planned and
supervised the project.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to E.S.
(saitoheiji@imr.tohoku.ac.jp).
doi:10.1038/nature08876
METHODS
Film preparation. A single-crystal Y3Fe5O12 (111) film was grown on a 0.5-mm-
thick Gd3Ga5O12 (111) single-crystal substrate by liquid phase epitaxy. For the
film growth, we used PbO-B2O3 flux around 1,200 K. Then, a 10-nm-thick Pt
layer was sputtered on the Y3Fe5O12 layer through a metal mask using a sputter-
ing and etching system (K-Science KE604TT1-S2B). Immediately before the
sputtering, the surface was cleaned through the metal mask by Ar-ion bombard-
ment in a vacuum. The chemical composition and crystalline structure of the
film were confirmed by X-ray fluorescence spectroscopy and X-ray diffractome-
try, respectively. The magnetization-saturation field for the Y3Fe5O12 layer was
estimated by measuring the magnetic-field dependence of a magneto-optical
Kerr effect signal induced by He-Ne laser light (632.8 nm) in a cross-Nicol
configuration.
Spin pumping measurements shown in Fig. 2. The Pt/Y3Fe5O12 sample system
was placed near the centre of a TE011 microwave cavity at which the magnetic-
field component of the microwave mode is maximized while the electric-field
component is minimized. During the measurements, the microwave mode with
a frequency of 9.44 GHz was exited in the cavity, and an external static magnetic
field H was applied along the film. The microwave power was less than 10 mW, a
value lower than the saturation of the ferromagnetic resonance absorption for
the present sample. The microwave absorption was detected using a spectro-
meter (JEOL JES-FA200) connected to the cavity. For voltage measurements,
two 100-nm-thick 0.5-mm-wide Pt electrodes were sputtered on the ends of the
Pt film. A twisted pair of thin coated Cu wires (0.08 mm in diameter) is con-
nected to the electrodes with silver paste. The voltage difference between these
two wires was measured using a lock-in amplifier (NF LI5640) and a volt meter
(Keithley 2000) via an amplifier (DL-instruments 1201).
Microwave emission measurements shown in Fig. 3. A gold coplanar-wave-
guide antenna, whose centre-conductor width is 100 mm, was attached to the Pt
surface of the Pt/Y3Fe5O12 sample film. The microwave signal received by the
antenna was led to an amplifier (Miteq AFS44-00102000-30-10P-44) via a micro-
wave probe (Picoprobe 40A-GSG-100-VP-NM). The amplified microwave signal
was analysed and recorded by a spectrum analyser (Anritsu MS2687B).
Non-local I–V measurements shown in Fig. 4. The sample film was fixed on a
sample stage of a micro-prober system (Nagase GRAIL-103054LVMG) and
©2010 Macmillan Publishers Limited. All rights reserved
electric connections were made with the microprobes. An electric current was
applied to the Pt film i with a voltage-function generator (NF WF1946B) and a
hand-made current amplifier. The voltage signal in the Pt film o was led to a
storage scope (Tektronix TDS1001) and to a lock-in amplifier (NF LI5640) via a
preamplifier (DL-instruments 1201). Precise I–V measurements were performed
with the lock-in amplifier by modulating sinusoidally the input d.c. current and
then by integrating the lock-in signal with respect to the current. Before the
precise measurement, we performed rough measurements to see the overall
behaviour of V; we measured d.c. voltage V using the storage scope with chan-
ging input current without sinusoidal modulations.
Micromagnetic simulation. We numerically solved the Landau–Lifshitz–
Gilbert equation:
dM a dM
~{c(Heff |M)z M| ztstt
dt Ms dt
by using the object-oriented micromagnetic framework31 extended to the fourth
order Runge-Kutta method. Here, c 5 1.76 3 107 Oe21 s21 is the gyromagnetic
ratio, 4pMs 5 1,956 Oe is the saturation magnetization and a 5 6.7 3 1025 is
the Gilbert damping coefficient. The field Heff is an effective magnetic field
which includes an applied magnetic field Hdc 5 2.3 kOe, an exchange field
Hex 5 (2A/Ms2)=2M due to the ferromagnetic coupling where the exchange stiff-
ness A 5 4.73 3 10212 J m21, the demagnetization field Hdemag and the random
magnetic field due to the thermal effect30. tstt 5 cM 3 (s 3 M) represents the spin
torque term due to the transmission of angular momentum from Pt film i to
Y3Fe5O12 via an exchange interaction (see Supplementary Information section B
for details). When the static field Hdc is applied, the spin torque term tstt points
to the direction opposite to the damping vector M 3 dM/dt. For realizing the
experimental condition where the input is the threshold current, the parameter c
was chosen so that the magnitude is that of the damping term. We calculated the
time average of the out-of-plane component of the magnetization. The gradation
of the colour indicates that the spin-wave spin current propagates in the x
direction.
31. Donahue, M. J. & Porter, D. G. OOMMF v1.2a3 Object Oriented MicroMagnetic
Framework Software (NIST, 2004).
Vol 464 | 11 March 2010 | doi:10.1038/nature08863
Tunable polymer multi-shape memory effect
Tao Xie1
Shape memory polymers are materials that can memorize temporary
shapes and revert to their permanent shape upon exposure to an
external stimulus such as heat1, light2,3, moisture4 or magnetic field5.
Such properties have enabled a variety of applications including
deployable space structures6, biomedical devices7,8, adaptive optical
devices9, smart dry adhesives10 and fasteners11. The ultimate poten-
tial for a shape memory polymer, however, is limited by the number
of temporary shapes it can memorize in each shape memory cycle
and the ability to tune the shape memory transition temperature(s)
for the targeted applications. Currently known shape memory
polymers are capable of memorizing one or two temporary shapes,
corresponding to dual- and triple-shape memory effects (also count-
ing the permanent shape), respectively11–13. At the molecular level,
the maximum number of temporary shapes a shape memory poly-
mer can memorize correlates directly to the number of discrete
reversible phase transitions (shape memory transitions) in the poly-
mer11–13. Intuitively, one might deduce that multi-shape memory
effects are achievable simply by introducing additional reversible
phase transitions. The task of synthesizing a polymer with more than
two distinctive and strongly bonded13 reversible phases, however, is
extremely challenging. Tuning shape memory effects, on the other
hand, is often achieved through tailoring the shape memory transi-
tion temperatures, which requires alteration in the material com-
position14–16. Here I show that the perfluorosulphonic acid ionomer
(PFSA), which has only one broad reversible phase transition,
exhibits dual-, triple-, and at least quadruple-shape memory effects,
all highly tunable without any change to the material composition.
PFSA is a commercial thermoplastic polymer with a polytetrafluor-
oethylene backbone and perfluoroether sulphonic acid side chains
(Fig. 1a). Owing to its proton conducting capability, PFSA has been
extensively studied as proton exchange membranes for fuel cells17.
Besides fuel cells, PFSA has also been used in a number of other appli-
cations including chlor-alkali cells, sensors and actuators. Generally,
PFSA possesses an ionic cluster phase and a crystalline phase17. Above
room temperature (25 uC), PFSA typically shows two thermally rever-
sible transitions17. These two transitions (below 150 uC and above
240 uC) correspond respectively to the short-range segmental motions
(glass transition) within a static electrostatic network and the onset of
the long-range molecular mobility due to destabilization of the elec-
trostatic network18. Although a glass transition is usually used for a
shape memory transition1, the temperature persistence of an electro-
static network has also proved beneficial in the design of shape memory
polymers19. For PFSA, its combination of a glass transition and a mech-
anism for setting a permanent shape (the temperature-persistent elec-
trostatic network) motivated us to explore its shape memory
properties.
The PFSA used in this study is Nafion, which has an equivalent
weight (average molecular mass per mole of ionic groups in g mol21)
of 1,000. It possesses a broad glass transition from ,55 uC to ,130 uC
(Fig. 1b). Before evaluating its shape memory performance, the as-
received PFSA film was first annealed at 140 uC, upon which the
1
Mail Code: 480-106-710, Chemical Sciences and Materials Systems Laboratory, General Motors Research and Development Center, 30500 Mound Road, Warren, Michigan 48090-
9055, USA.
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
polymer shrank by about 26% and reached an equilibrium length
(Supplementary Fig. 1). The shrinkage was not related to water loss
(see Supplementary Information), but was due to the removal of
residual stress/strain from the polymer processing step. X-ray scatter-
ing analysis further suggests that the stress relaxation occurred in the
ionic phase of the polymer (Supplementary Fig. 2). The annealing also
led to polymer darkening, yet the infrared spectra of PFSA before and
after annealing (Supplementary Fig. 3) appear nearly identical. This
suggests that the primary polymer structure remained intact during
annealing, as supported by a previous study showing20 that the ther-
mal degradation of PFSA starts at 300 uC. In addition, more recent
work on the chemical durability of PFSA suggested that its end-groups
are the most unstable moieties on the polymer21, the thermal decom-
position of which is the probable source for the darkening observed
here. Given the high molecular weight of PFSA17 and the correspond-
ing low end-group concentration, minor degradation in polymer end-
groups would not be detectable by infrared21 or thermo-gravimetric
analysis20, nor is it expected to affect the polymer’s thermomechanical
properties.
The equilibrium length after annealing defines the permanent
shape for PFSA and its shape memory performance was evaluated
using thermal mechanical analysis. In a typical thermally activated
dual-shape memory cycle, a shape memory polymer is deformed at
an elevated temperature (deformation temperature, Td) and the
deformed temporary shape is fixed upon cooling. When heated to
a recovery temperature (Tr), the permanent shape is recovered. The
dual-shape memory effect can thus be quantified on the basis of the
percentage of shape fixation (shape fixity, Rf) and shape recovery
(shape recovery, Rr), calculated using equations (1) and (2) shown
in the Methods Summary.
Accordingly, PFSA was found to exhibit excellent dual-shape
memory performance when deformed and recovered at 140 uC (that
is, Td and Tr both above the upper end of the glass transition), with
both Rf and Rr approaching 100% (Fig. 1c). Shape fixing and recovery
can also be carried out near the peak (Supplementary Fig. 4,
Td 5 Tr 5 100 uC) or the onset (Fig. 1d, Td 5 Tr 5 60 uC) of the glass
transition, with Rf and Rr above 97% in both cases. Consecutive shape
memory cycling experiments (Td 5 Tr 5 80 uC) show that shape
memory performance is highly reproducible (Supplementary Fig. 5).
Additionally, upon consecutive cycling at the annealing temperature of
140 uC, PFSA experienced almost no deterioration in either Rf or Rr
(Supplementary Fig. 6), confirming that indeed the darkening due to
annealing did not affect the thermomechanical response of PFSA.
A prototype dual-shape memory cycle occurs with both shape fixa-
tion and recovery above the shape memory transition to maintain
optimal shape fixing and recovery properties. Although recent studies
have demonstrated the feasibility and benefits (for example, thermal
healing) of shape fixation near22,23 or even below24 the onset of the
shape memory transition, Rf under such conditions is not typically
reported. Presumably, undesirable instantaneous recovery (spring
back) may have occurred, leading to a compromised Rf, as confirmed
267
LETTERS NATURE | Vol 464 | 11 March 2010

by a recent example for which an Rf of 83% was reported24. So, that PFSA is tunable simply by deforming the same polymer at different
PFSA maintained near-perfect shape fixing and recovery performance temperatures, again, without compromising Rf.
at any temperature above the onset of its glass transition is rather The multi-stage recovery (Fig. 1e) further indicates that the polymer
unexpected. can memorize multiple temporary shapes in a single shape memory
Although the strain imparted at a temperature as low as 60 uC can cycle, that is, the multi-shape memory effect. Here, it is important to
be fully recovered at the same temperature (Fig. 1d), the deformation note that currently known triple-shape memory polymer systems
strain introduced at 140 uC was unable to recover fully at lower tem- possess two well separated phase transitions and that the two tem-
peratures. As shown in Fig. 1e, for PFSA deformed at 140 uC, increas- porary shapes are introduced above and between both transition tem-
ing the recovery temperature in a staged manner led to staged peratures, respectively11–13. Tuning the triple-shape memory effect for
recovery behaviour. Previous work has shown that shape recovery such systems would require varying of the ratio between the two
at a certain temperature may depend on the deformation temper- reversible phases or changing the reversible phase transition tempera-
ature at which the shape fixing occurred22,23, but a well-defined multi- tures11–13, which cannot be realized without changing the material
stage (beyond two-stage) recovery such as that shown in Fig. 1e has composition. PFSA, in contrast, has only one broad phase transition
not previously been realized. In contrast to the staged recovery, a and its triple-shape memory effect, therefore, should not be limited to
continuous shape recovery behaviour was observed when the two discrete temperature regions.
recovery temperature was increased in a continuous fashion to The triple-shape memory effect for PFSA is demonstrated in
140 uC for three samples deformed at 60, 80 and 100 uC, respectively Fig. 2a. The permanent shape S0 was deformed at 140 uC and fixed
(Supplementary Fig. 7). In those cases, the shape recovery at 68 uC to yield the first temporary shape S1, which was further
curve shifted to the right (higher temperatures) for PFSA with a deformed at 68 uC and fixed at 20 uC to yield the second temporary
higher Td. shape S2. Upon reheating to 68 uC, the recovered first temporary
The above experimental results clearly suggest that the polymer shape S1rec was obtained. Further heating to 140 uC yielded the
memorizes not just the strain, but also the thermomechanical history. recovered permanent shape S0rec. Quantitatively, two triple-shape
For instance, the polymer memorizes the deformation temperature, memory cycles are illustrated in Fig. 2b (with two deformation tem-
showing the so-called temperature memory effect25. As illustrated in peratures of 140 uC and 53 uC) and Supplementary Fig. 9 (with two
Supplementary Fig. 8a–c, when the deformation occurs at a certain Td deformation temperatures of 90 uC and 53 uC).
within the glass transition, the subsequent iso-strain stress recovery Equations (3) and (4) (see Methods Summary) were used to cal-
experiment shows that the peak of recovery stress appears at a tem- culate the Rf and Rr values for the triple-shape memory cycles. As
perature roughly the same as the corresponding Td. Compared to the such, a notable difference between Fig. 2b and Supplementary Fig. 9
earlier example of the polymer temperature memory effect25, this lies in the first-shape fixity (Rf(S0RS1) being 83.5% and 74.5%,
effect for PFSA occurs in a temperature range with both Rf and Rr respectively), suggesting that the first-shape fixity is closely related
approaching 100%. In addition, whereas adjusting the recovery stress to the difference between the two deformation temperatures in the
for known shape memory polymer systems is typically accomplished corresponding triple-shape memory cycle. Shape recoveries, on the
by changing the rubbery modulus through the adjustment of other hand, were above 95% in all cases. Together, Fig. 2a and b and
the crosslinking density8,14–16 or inclusion of reinforcing fillers3,26, Supplementary Fig. 9 confirm that the triple-shape memory effect for
Supplementary Figs 8a–c suggest that the peak recovery stress of PFSA can be realized at any two arbitrary temperatures above the

a b 1.0 c 140
0.9
3 40
(CF2CF2)m CF2CF 0.8 110

Temperature (ºC)
n
30

Stress (MPa)
O
log[(E′) (MPa)]

0.6
Strain (%)

2 0.6 80
tanδ

CF2
CF CF3 20
0.4
O 50
0.3
1 0.2 10
CF2
CF2 20
0.0 0
SO3H 0.0
0 –10
20 40 60 80 100 120 140 0 20 40 60 80
Temperature (ºC) Time (min)

d e 120
140
80 9
60 100 1.0
110
Temperature (ºC)

Temperature (ºC)

80
Strain (%)

Stress (MPa)
Stress (MPa)

6
40 60
Strain (%)

60 80
3 40 0.5
20 40 50
20
0
0 20
0 20 0.0
0 10 20 30 40 50 0 50 100 150 200 250 300
Time (min) Time (min)

Figure 1 | Structure, dynamic mechanical, and dual-shape memory modulus E9. c, Dual-shape memory cycle at Td 5 Tr 5 140 uC. Rf: 100%, Rr:
properties of PFSA. a, Structure of PFSA (m 5 5.56 for an equivalent weight 96.6%. d, Dual-shape memory cycle at Td 5 Tr 5 60 uC. Rf: 98.6%, Rr: 97.7%.
of 1,000; n is the number of repeating units). b, Dynamic mechanical analysis e, Multi-staged shape recovery (Td 5 140 uC).
curve of PFSA. tand is the ratio between the loss modulus E99 and the storage
268
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 LETTERS

a a

S0 S1
S0 S1 S2 S3 S2rec S1rec S0rec

b Fixing Recovery
S2 15
140
εS3, load 140
εS3
120 12
120

Temperature (°C)
100 εS2, load 9

Stress (MPa)
S1rec S0rec 100

Strain (%)
80 εS2 εS2, rec
80 6
b Fixing Recovery 60
εS1, load
60
120 40 εS1, rec 3
εS2, load εS1
εS2 40
140 20
100 10 εS0 εS0, rec 0
0 20
80 Temperature (ºC)

Stress (MPa)
30 60 90 120 150 180
Strain (%)

100 Time (min)

60 εS1, load
εS1, rec
εS1
40
20
εS0
0 20
40 60 80 100 120
Time (min)
Figure 2 | Triple-shape memory properties of PFSA. a, Visual
demonstration. b, Quantitative thermal mechanical cycle
(Td1 5 Tr2 5 140 uC, Td2 5 Tr1 5 53 uC). Rf(S0RS1): 83.5%,
Rf(S1RS2): 96.7%, Rr(S2RS1): 97.4%, Rr(S1RS0): 94.6%.
onset of the glass transition, provided that the two temperatures are
not too close.
The versatile shape memory properties for PFSA are also reflected in
a quadruple-shape memory effect. As demonstrated in Fig. 3a, starting
as a permanent shape S0, PFSA can memorize three temporary shapes
(S1, S2 and S3) in the same shape memory cycle. Subsequent heating
to the relevant temperatures led to the recovered shapes (S2rec, S1rec
and S0rec). Quantitatively (Fig. 3b), whereas the Rr values at all three
recovery stages were above 93%, the first and second Rf values were
around 60%. Theoretically, multi-shape memory effects beyond
quadruple- are feasible. In fact, the multi-stage recovery (Fig. 1e) itself
can be interpreted as memorizing five temporary shapes through a
one-step programming process, with each recovery step representing
the recovery of one temporary shape12. If more and more temporary
shapes were to be fixed individually into the PFSA instead of the one-
step programming process, the Rf for each individual fixing step would
be further compromised because the fixation of more temporary
shapes demands shape fixation at a temperature closer to the corres-
ponding deformation temperature.
The versatile shape memory properties demonstrated above for
PFSA stem from its broad glass transition. This broad transition
can be viewed as the collective contribution of an infinite number
of transitions, each corresponding to infinitely sharp transition tem-
peratures continuously distributed across the broad transition. In the
context of shape memory effects, these sharp transitions can be
understood as individual memory elements. Depending on the
deformation temperature(s) during the shape memory cycles, a vari-
able number of memory elements are activated for the memory func-
tion, leading to the tunable shape memory effects. Therefore, other
ionomers with broad glass transitions are likely to show similar shape
memory effects. Amorphous shape memory polymers possessing a
broad thermally reversible transition are another class of candidates.
In such cases, the broad thermal reversible transition may arise from
©2010 Macmillan Publishers Limited. All rights reserved
8
Figure 3 | Quadruple-shape memory properties of PFSA. a, Visual
demonstration. S0: permanent shape; S1: first temporary shape (Td1:
60 4 140 uC); S2: second temporary shape (Td2: 107 uC); S3: third temporary
shape (Td3: 68 uC); S2rec: recovered second temporary shape (Tr1: 68 uC);
εS0, rec S1rec: recovered first temporary shape (Tr2: 107 uC); S0rec: recovered
0 permanent shape (Tr3: 140 uC). b, Quantitative thermal mechanical cycle
(Td1 5 Tr3 5 140 uC, Td2 5 Tr2 5 90 uC, Td3 5 Tr1 5 53 uC).
140 160
Rf(S0RS1): 58.7%, Rf(S1RS2): 57.1%, Rf(S2RS3): 96.1%,
Rr(S3RS2): 100.0%, Rr(S2RS1): 99.6%, Rr(S1RS0): 93.0%.
strong and temperature-dependent molecular interactions (for
example, self-complementary hydrogen bonding27,28) or composi-
tional heterogeneity (for example, compositional gradient) in the
polymer structure. Some aspects of tunable multi-shape memory
effects may even be applicable to semi-crystalline shape memory
polymers. Crystallization of such polymers can be induced by either
cooling or strain3,24,29 and the structural perfection of crystallites (and
hence their corresponding melting transitions) depends on the ther-
momechanical history29. Using a combination of temperature and
strain control during the shape fixing step, it is potentially possible to
manipulate the melting transition, which would lead to tunable
shape recovery behaviour. The above concepts, althought they are
yet to be verified experimentally, suggest strategies for expanding
tunable multi-shape memory effects to a broad range of polymers,
the availability of which could significantly expand the technical
potential for shape memory polymers.
METHODS SUMMARY
All quantitative shape memory properties were evaluated in a tensile and force
controlled mode in a typical thermal mechanical analysis set-up under ambient
humidity conditions (relative humidity of 22–34%). The heating and cooling
rates were both 5 uC per minute. Rf and Rr for the dual-shape memory effect were
calculated using equations (1) and (2) below:
Rf ~100%|e=eload ð1Þ
Rr ~100%|(e{erec )=e ð2Þ
where eload represents the maximum strain under load, e is the fixed strain after
cooling and load removal, and erec is the strain after recovery.
Equations (1) and (2) are expanded to equations (3) and (4) to calculate Rf and
Rr for the triple-shape and quadruple-shape memory effects.
Rf (X?Y )~100%|(ey {ex )=(ey, load {ex ) ð3Þ
Rr (Y ?X)~100%|(ey {ex, rec )=(ey {ex ) ð4Þ
where X and Y denote two different shapes, respectively, ey, load represents the
maximum strain under load, ey and ex are fixed strains after cooling and load
removal, and ex,rec is the strain after recovery.
269
LETTERS
The visual demonstrations of the triple- and quadruple-shape memory effects
were carried out using oven heating. An equilibrium time of thirty minutes was
used for any temperature changes that occurred during the shape memory cycles.
In stress recovery experiments, the sample was first heated to a targeted tem-
perature, deformed at a strain rate of 10% per minute to a strain of 40%, and
cooled down to 20 uC. While maintaining such a strain (iso-strain), the recovery
stress was monitored while the sample was heated to 140 uC at 5 uC per minute.
Received 30 September 2009; accepted 25 January 2010.
1. Liu, C., Qin, H. & Mather, P. T. Review of progress in shape-memory polymers. J.
Mater. Chem. 17, 1543–1558 (2007).
2. Lendlein, A., Jiang, H., Junger, O. & Langer, R. Light-induced shape-memory
polymers. Nature 434, 879–882 (2005).
3. Koerner, H. et al. Remotely actuated polymer nanocomposites—stress-recovery of
carbon-nanotube-filled thermoplastic elastomers. Nature Mater. 3, 115–120 (2004).
4. Huang, W. et al. Water-driven programmable polyurethane shape memory
polymer: demonstration and mechanism. Appl. Phys. Lett. 86, 114105 (2005).
5. Mohr, R. et al. Initiation of shape-memory effect by inductive heating of magnetic
nanoparticles in thermoplastic polymers. Proc. Natl Acad. Sci. USA 103,
3540–3545 (2006).
6. NASA. Elastic memory composite hinges tested on shuttle. NASA Tech. Briefs 31,
20 Æhttp://www.techbriefs.com/component/content/article/1344æ (2007).
7. Lendlein, A. & Langer, R. Biodegradable, elastic shape-memory polymers for
potential biomedical applications. Science 296, 1673–1676 (2002).
8. Yakacki, C. M. et al. Strong, tailored, biocompatible shape-memory polymer
networks. Adv. Funct. Mater. 18, 2428–2435 (2008).
9. Kunzelman, J., Chung, T., Mather, P. T. & Weder, C. Shape memory polymers with
built-in threshold temperature sensors. J. Mater. Chem. 18, 1082–1086 (2008).
10. Xie, T. & Xiao, X. Self-peeling reversible dry adhesive. Chem. Mater. 20,
2866–2868 (2008).
11. Bellin, I., Kelch, S., Langer, R. & Lendlein, A. Polymeric triple-shape materials. Proc.
Natl Acad. Sci. USA 103, 18043–18047 (2006).
12. Behl, M. et al. One-step process for creating triple-shape capability of AB polymer
networks. Adv. Funct. Mater. 19, 102–108 (2009).
13. Xie, T., Xiao, X. & Cheng, Y. T. Revealing triple-shape memory effect by polymer
bilayers. Macromol. Rapid Commun. 30, 1823–1827 (2009).
14. Xie, T. & Rousseau, I. A. Facile tailoring of thermal transition temperatures of
epoxy shape memory polymers. Polymer 50, 1852–1856 (2009).
15. Liu, C. & Mather, P. T. Thermomechanical characterization of a tailored series of
shape memory polymers. J. Appl. Med. Polym. 6, 47–52 (2002).
16. Liu, C. et al. Chemically cross-linked polycyclooctene: synthesis, characterization,
and shape memory behavior. Macromolecules 35, 9868–9874 (2002).
17. Mauritz, K. A. & Moore, R. B. State of understanding of Nafion. Chem. Rev. 104,
4535–4585 (2004).
270
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
18. Page, K. A., Cable, K. M. & Moore, R. B. Molecular origins of the thermal transitions
and dynamic mechanical relaxations in perfluorosulfonate ionomers.
Macromolecules 38, 6472–6484 (2005).
19. Weiss, R., Izzo, E. & Mandelbaum, S. New design of shape memory polymers:
mixtures of an elastomeric ionomer and low molar mass fatty acids and their salts.
Macromolecules 41, 2978–2980 (2008).
20. Deng, Q., Wilkie, C. A., Moore, R. B. & Mauritz, K. A. TGA-FTi.r. investigation of
the thermal degradation of Nafion and Nafion/silicon oxide-based
nanocomposites. Polymer 39, 5961–5972 (1998).
21. Xie, T. & Hayden, C. A. A kinetic model for the chemical degradation of
perfluorinated sulfonic acid ionomers: weak end groups versus side chain
cleavage. Polymer 48, 5497–5506 (2007).
22. Yakacki, C. M. et al. Unconstrained recovery characterization of shape-memory
polymer networks for cardiovascular applications. Biomaterials 28, 2255–2263
(2007).
23. Gall, K. et al. Thermomechanics of the shape memory effect in polymers for
biomedical applications. J. Biomed. Mater. Res. A 73A, 339–348 (2005).
24. Rodriguez, E., Luo, X. & Mather, P. T. Miscible blends of linear and crosslinked
poly(e-caprolactone) for shape memory assisted self-healing. Mater. Res. Soc. Fall
Meet. Symp. BB (Polymer-Based Smart Materials–Process, Properties, and
Application) abstr. BB5.19 (2008).
25. Miaudet, P. et al. Shape and temperature memory of nanocomposites with
broadened glass transition. Science 318, 1294–1296 (2007).
26. Gall, K. et al. Internal stress storage in shape memory polymer nanocomposites.
Appl. Phys. Lett. 85, 290–292 (2004).
27. Li, J. Viveros, J. A., Wrue, M. H. & Anthamatten, M. Shape-memory effects in
polymer networks containing reversibly associating side-groups. Adv. Mater. 19,
2851–2855 (2007).
28. Kushner, A. M., Vossler, J. D., Williams, G. A. & Guan, Z. A biomimetic modular
polymer with tough and adaptive properties. J. Am. Chem. Soc. 131, 8766–8768
(2009).
29. Ping, P., Wang, W., Chen, X. & Jing, X. Poly(e-caprolactone) polyurethane and its
shape memory property. Biomacromolecules 6, 587–592 (2005).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements I thank I. Rousseau, H. Kia, P. Krajewski, X. Huang and
M. Verbrugge for comments, D. Eckel for Infrared analysis, and R. L. Speer Jr for
X-ray analysis.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The author declares no competing financial interests.
Correspondence and requests for materials should be addressed to the author
(tao.xie@gm.com).
Vol 464 | 11 March 2010 | doi:10.1038/nature08905
A large atomic chlorine source inferred from
mid-continental reactive nitrogen chemistry
Joel A. Thornton1, James P. Kercher1{, Theran P. Riedel1,2, Nicholas L. Wagner3, Julie Cozic3,4, John S. Holloway3,4,
William P. Dubé3,4, Glenn M. Wolfe1,2, Patricia K. Quinn5, Ann M. Middlebrook3, Becky Alexander1
& Steven S. Brown3
Halogen atoms and oxides are highly reactive and can profoundly
affect atmospheric composition. Chlorine atoms can decrease the
lifetimes of gaseous elemental mercury1 and hydrocarbons such as
the greenhouse gas methane2. Chlorine atoms also influence cycles
that catalytically destroy or produce tropospheric ozone3, a green-
house gas potentially toxic to plant and animal life. Conversion of
inorganic chloride into gaseous chlorine atom precursors within
the troposphere is generally considered a coastal or marine air phe-
nomenon4. Here we report mid-continental observations of the
chlorine atom precursor nitryl chloride at a distance of 1,400 km
from the nearest coastline. We observe persistent and significant
nitryl chloride production relative to the consumption of its nitro-
gen oxide precursors. Comparison of these findings to model pre-
dictions based on aerosol and precipitation composition data from
long-term monitoring networks suggests nitryl chloride production
in the contiguous USA alone is at a level similar to previous global
estimates for coastal and marine regions5. We also suggest that a
significant fraction of tropospheric chlorine atoms6 may arise
directly from anthropogenic pollutants.
Night-time reactions of nitrogen oxides are known to convert
inorganic chloride into chlorine atom (Cl .) precursors7 (for example,
Fig. 1). N2O5, a nocturnal NOx reservoir (NOx 5 NO 1 NO2), can
react on airborne particles to produce only HNO3 (reaction (1a)), or
both nitryl chloride (ClNO2) and nitrate (NO32) (reaction (1b)).
N2O5(g) 1 H2O(aq) R 2HNO3(aq) (1a)
2 2
N2O5(g) 1 Cl (aq) R ClNO2(g) 1 NO3 (aq) (1b)
The ClNO2 yield depends on water and chloride concentrations
within particles8. The latter is often sufficient for efficient ClNO2
production5. However, the moles of particulate chloride per volume
of air (pCl2) is typically small, and would limit reaction (1b) except
that gaseous HCl can almost always provide a larger reservoir
through equilibrium repartitioning to particles9. Depending on the
environment, soluble chloride or NOx availability and reactivity may
limit ClNO2 production. The few prior observations of ClNO2 within
polluted marine air have shown it is produced in high yields during
spring and summer, and exceeds previously predicted values by
factors of 2–30 (refs 5, 10, 11).
ClNO2 production has an impact on NOx and Cl budgets, both of
which affect the troposphere’s oxidizing capacity12. Current atmo-
spheric chemistry models predict reaction (1a) accounts for 30–50%
of total NOx removal in polluted regions13, but they typically neglect
1
Department of Atmospheric Sciences, 2Department of Chemistry, University of Washington, Seattle, Washington 98195, USA. 3Earth Systems Research Laboratory, National
Oceanic and Atmospheric Administration, Boulder, Colorado 80305, USA. 4Cooperative Institute for Research in Environmental Studies, Boulder, Colorado 80309, USA. 5Pacific
Marine Environmental Laboratory, National Oceanic and Atmospheric Administration, Seattle, Washington 98115, USA. {Present address: Department of Chemistry, Hiram College,
Hiram, Ohio 44234, USA.
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
reaction (1b). Owing to efficient deposition of HNO3, reaction (1a)
represents a terminal NOx sink in the lower troposphere, whereas
reaction (1b) recycles NOx through ClNO2 photolysis (reaction (2))
within a few hours after sunrise14.
ClNO2 1 hn R Cl. 1 NO2 (2)
Models incorporating reactions (1b) and (2) differ in their predicted
impacts, but generally show summer ozone concentrations in coastal
urban areas to be enhanced by faster cycling of oxidants produced
from Cl . attacking hydrocarbons3,5.
These previous studies have been restricted to coastal or marine
regions. However, our observations made near Boulder, Colorado,
an urban location in the middle of North America, demonstrate that
this chemistry extends well inland, and are consistent with calcula-
tions based on network observations of aerosol and precipitation
composition.
We made in situ measurements on 11–25 February 2009 at the
National Oceanic and Atmospheric Administration’s Kohler Mesa
facility, just west and 150 m above Boulder, Colorado. The site, subject
to large variability in pollutant levels, receives either the urban plume
from nearby cities, or much cleaner air from the Rocky Mountain
HCl Acid HNO3 HCl
HNO3 Cl
displacement H2SO4
N2O5
ClNO2
Acid precursor and
Inland NOx and NOx
chlorine emissions
NOx chlorine emissions
Aeolian sources
Sea
spray
Figure 1 | Schematic of chlorine activation by night-time NOx chemistry.
The emphasis here is on inland ClNO2 production, although production also
occurs in coastal and marine regions. Globally, sea spray is the dominant
source of tropospheric inorganic chloride9, which can be transported as fine-
mode sea spray particles or gaseous HCl following acid displacement by
HNO3 or H2SO4. Locally, other sources of chloride from industrial
activities9, biomass burning9, or transport of wind-blown soil dust15 may be
important. Reactions of N2O5 on chloride-containing particles produce
ClNO2, which photochemically converts to chlorine atoms and NO2 in the
morning. The chlorine atoms react with hydrocarbons, returning to HCl or
forming an organo-chlorine compound.
271
LETTERS
region. Typically residing above the urban nocturnal boundary layer,
the site is often minimally affected by direct, local night-time emissions.
It therefore allows overnight observation of the chemical evolution of
air masses characteristic of this region. More information on the site
and measurement methods is provided in Supplementary Information.
There are three salient features of the February 2009 time series
(Fig. 2). First, ClNO2 production was routinely observed in a region
of North America far removed from sea spray, but possibly affected by
chloride transport from coastal areas9 or inland salt beds15, and by
anthropogenic sources including combustion and transportation9.
Second, ClNO2 mixing ratios consistently reached 100–450 parts per
trillion by volume (p.p.t.v.) whenever the urban plume was sampled.
These levels are unexpectedly large, reaching a third to a half of the
maximum values observed in polluted coastal areas5,10. Third, consis-
tent with the above mechanism, ClNO2 was observed only at night or
in the early morning, and often correlated with N2O5 (Fig. 2, top).
However, their relationship was variable, probably due to changes in:
(1) the N2O5 heterogeneous loss rate, which depends on humidity,
particle composition and phase16, and surface area density (Sa); (2) air
mass age and non-aerosol losses of N2O5 (Supplementary Informa-
tion); and (3) chloride availability. We observed similar levels of N2O5
and ClNO2 in February 2008 at a nearby location, suggesting year-to-
year consistency (Supplementary Information).
As expected, the 2009 observations also indicate total available
chloride for ClNO2 production is greater than pCl2 alone5 (Fig. 3a).
The inferred minimum abundance of total available chloride
(100–500 p.p.t.v.) is consistent with admittedly uncertain HCl obser-
vations in other urban areas9, and with the pCl2 « HClg equilibrium
predicted using thermodynamic models constrained by our measure-
ments17 (Supplementary Information).
We performed time-dependent chemical box modelling to develop
a quantitative relationship between measured pCl2 and the ClNO2
yield from reaction (1) (wClNO2 ) needed to explain the observations.
The box model integrates explicit N2O5, ClNO2 and pCl2 mass balance
500
ClNO2 (p.p.t.v.)
400
ClNO2 (p.p.t.v.)
300
200
100
0 0
0 400 800 1,200
N2O5 (p.p.t.v.)
ClNO2 (p.p.t.v.)
400
200
0 the urban surface layer containing fresh NOx
1,500
N2O5 (p.p.t.v.)
1,000
500
0
pCl– (p.p.t.v.)
100
50
80
RH (%)
40
0
12 14
272
NATURE | Vol 464 | 11 March 2010
equations initialized with observations of NO2, O3, Sa, pCl2. The N2O5
reaction probability and wClNO2 are adjusted between 0.005–0.03 and
0.07–0.36, respectively, so the model matches the observed NO2, O3,
N2O5 and ClNO2 at a specific time. Model details and additional out-
put are given in Supplementary Information.
The night of 15–16 February illustrates the time evolution of
ClNO2 under low wind, providing a model-to-measurement com-
parison least affected by air mass variations due to transport. A con-
stant model wClNO2 of 0.14 adequately reproduces the observed
ClNO2 and pCl2 evolution (Fig. 3b), as well as NO2, O3 and N2O5
at specific times after sunset. These species are more sensitive to fresh
emissions or land-surface interactions caused by slight changes in
transport, and therefore exhibit larger deviations from the model
curve (Supplementary Information). This night did not show
evidence of a chloride limitation, and had the highest 24-h average
pCl2 and ClNO2 of the campaign. Most other nights did not allow
comparison to the box model with a single set of initial conditions
mainly because of transport effects (Supplementary Information).
The box-model-derived wClNO2 are generally lower than calculated
from laboratory parameterizations8,18 using measured pCl2 and
estimated particle water17 (Supplementary Information). Several
possible reasons for the apparent discrepancy exist. First, although
sufficient total chloride mass may exist, it probably partitions to
a fraction of the particle surface area19, reducing the population
average efficiency of reaction (1b) which our model-derived wClNO2
values represent. A more accurate calculation would require size-
resolved particle composition and pH measurements, which are chal-
lenging20. Additionally, the wClNO2 required in the model is different
for different nights, and even decreases within some nights, suggest-
ing significant variability in the factors controlling ClNO2 produc-
tion as well as a possible limitation by available chloride.
The existence of a mid-continental ClNO2 source was unknown
before the observations we report here. Yet, the ingredients for
this source, NOx and particulate chloride, are known from various
500 500 Figure 2 | Time series of key quantities
observed in Boulder, Colorado, from 11 to 25
400 400
ClNO2 (p.p.t.v.)
February 2009. Main panels: from top to
300 300
bottom, ClNO2, N2O5, pCl2 shown in mixing
200 200 ratio units (p.p.t.v.) for comparison with gases,
100 100 and relative humidity (RH). Insets (top) are
0 point-to-point comparisons of ClNO2 to N2O5
0 400 800 1,200 0 400 800 1,200 on 13–14 February (left), 15–16 February
N2O5 (p.p.t.v.) N2O5 (p.p.t.v.)
(centre) and 21–22 February (right). These nights
are characterized by high (85%), moderate (50%)
and low (35%) RH; and high, moderate and low
ratios of ClNO2:N2O5, respectively. The
ClNO2:N2O5 ratio is not solely a function of RH.
For example, on 14 February we sampled air from
emissions which can titrate N2O5. See
Supplementary Information for additional
details.
16 18 20 22 24
Days in February 2009 (local time)
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 LETTERS

a b 600

200 ClNO2 ClNO2

500
4× pCl– 4 × pCl–

Mixing ratio (p.p.t.v.)

Mixing ratio (p.p.t.v.)

150 400

300
100

200

50
100

0 0
0 3 6 9 12 15 18 21 24 0 4 8 12 16
Hour of day (local time) Hours since sunset

Figure 3 | Observed and modelled relationships of ClNO2 and particulate
chloride. a, 3-h averages of coincident ClNO2 and 4 3 pCl2 versus time of
day. Shading indicates the standard error of means; each bin contains more
than 15 points. b, Observed and modelled evolution of ClNO2 and pCl2 over
the 12-h period from sunset on 15 February. Symbols as in a, stars denote
pCl2 below the detection threshold, and solid lines represent model
predictions. The model reasonably reproduces the night-time evolution of
ClNO2 with a constant ClNO2 yield of 0.14, and matches the ozone, NO2 and
N2O5 observations 12 h after sunset. See Supplementary Information for a
more complete description of the model.

long-term databases to be ubiquitous across the North American
continent (and elsewhere). With these data sets, we now estimate
the magnitude of this halogen source across wider scales and com-
pare these estimates to multi-location observations of ClNO2.
Assuming lower tropospheric NOx is in steady-state, we calculate a
total annual ClNO2 production rate as a sum over seasonal averages
(s) via the following equation:
X XX
PClNO2 ~ s
PClNO 2
~ s
ENO x
(r)fNs 2 O5 (r)wsClNO2 (r) ð3Þ
s s
where r denotes a 1 3 1u grid cell in the contiguous US or its coastal
s
regions, ENOx
(r) is the anthropogenic NOx emission rate from the
EDGAR database21, fNs 2 O5 ðr Þ is the fraction of NOx removed by N2O5
heterogeneous chemistry predicted by the GEOS-Chem global
model13 and wsClNO2 ðr Þ is the ClNO2 yield. We constrained
wsClNO2 ðr Þ with a combination of precipitation and aerosol composi-
tion measurements from the National Atmospheric Deposition
Program (NADP)22 and the Interagency Monitoring of Protected
Visual Environments (IMPROVE) network23, respectively. This
approach provides wClNO2 consistent with all existing ClNO2 obser-
vations in Boulder and coastal regions5,10, while reproducing the
likely spatial distribution in total available chloride.
Figure 4 shows the annual average fields of ENOx (Fig. 4a), fN2 O5
(Fig. 4b), wClNO2 (Fig. 4c), and the resulting PClNO2 (Fig. 4d). Based on
r
the seasonal values of these components, the variability in the Boulder
and network data, and alternative approaches to equation (3), all of
which are presented in the Supplementary Information, we estimate
that PClNO2 for the contiguous US lies in the range 3.2–8.2 Tg yr21,
providing a photolytic Cl. source of 1.4–3.6 Tg Cl yr21. This US
ClNO2 source is far larger than the first global estimate of

a NOx emission rate b Fraction of NOx lost via N2O5

8

45 0.6
7.5

40 7
0.4

35 6.5

6 0.2
30
Latitude (°N)

5.5
c Yield of ClNO2 d Production rate of ClNO2
1
10
45

9
40
0.5
8
35

30 7

0
120 100 80 120 100 80
Longitude (°W) Longitude (°W)

Figure 4 | Annual average components of PClNO2 over the US. a, US NOx IMPROVE and NADP network observations; and d, the ClNO2 production
emissions (kg yr21) given by the EDGAR database with a logarithmic colour rate, PClNO2 , in g Cl yr21, obtained from the product of quantities shown in
scale starting at 105.5 kg NO2 yr21; b, annual average fraction of total nitrate panels a, b and c with a logarithmic colour scale starting at 106.5 g Cl yr21.
(0–2 km) formed by N2O5 reactions on particles predicted by the GEOS- Boulder, Colorado, is at approximately 40u N, 105u W.
Chem global model; c, annual average yield of ClNO2 calculated from
273
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
0.06 Tg Cl yr21 (ref. 24), and is similar to the recent 3.2 Tg Cl yr21
estimated for global coastal and marine regions5. More than half of
the predicted ClNO2 production occurs over land, and 40% occurs
during winter (December–February). It also suggests that, in the US,
an amount of NOx equivalent to 8–22% of that emitted cycles through
ClNO2, potentially forming a non-negligible fraction of reactive nitro-
gen at night’s end, primarily during winter above the nocturnal sur-
face layer (see Supplementary Information).
If the global distribution of pCl2 and NOx sources are similar to
those in the US, as independent measurements suggest25, we estimate
the global Cl. source from ClNO2 to be 8–22 Tg Cl. yr21, which is of
the same order as that inferred from methane isotopes in remote
regions (25–35 Tg Cl . yr21)2,6. ClNO2 production will mostly occur
in polluted regions; as such, Cl . from ClNO2 will react with other
hydrocarbons as well as methane. Thus, the Cl . source from ClNO2
may be in addition to that inferred from methane isotopes2,6.
Some fraction of the Cl. from ClNO2 may also convert to temporary
reservoirs, such as HOCl and ClONO2, possibly enhancing Cl. pro-
duction not captured by our estimates11.
Although likely to be the most rigorous observationally con-
strained estimate of continental-scale ClNO2 production to date,
the accuracy of the above approach (and probably any other) is
limited by significant uncertainty in fN2 O5 and wClNO2 that arise from
an incomplete understanding of N2O5 reactivity26,27 and of chloride
partitioning across the particle distribution19. More studies of the
NOx–ClNO2–pCl2–HCl system around the globe are critical to
refine these predictions.
Our results imply that a significant fraction of the tropospheric Cl .
source is anthropogenic, distributed over a relatively small area of the
Earth’s surface—polluted continental and coastal regions—and con-
centrated in a fraction of each day (morning). Long-range transport
of NOx reservoirs, such as peroxy nitrates28, or NOx emissions from
irradiated snow packs29, may also affect halogen activation far from
anthropogenic emissions, albeit on a smaller concentration scale.
Therefore, past and future trends in continental NOx emissions
may represent an important global influence on tropospheric
halogen sources that has largely gone unrecognized.
NOx influences climate by directly and indirectly regulating oxid-
ant budgets that determine the methane lifetime and affect aerosol
formation30. Nocturnal processing of NOx is typically considered a
reduction in the troposphere’s oxidizing capacity due to the conver-
sion of NOx and ozone into soluble, largely non-reactive species (for
example, HNO3). Widespread ClNO2 production instead renders
nocturnal NOx chemistry a potential source of oxidants and ozone
in polluted regions, and may have as yet unrecognized influences in
remote regions. Thus, anthropogenic NOx may have an even larger
effect on the oxidizing power of the lower troposphere than current
models estimate.
Received 2 October 2009; accepted 2 February 2010.
1. Donohoue, D. L., Bauer, D. & Hynes, A. J. Temperature and pressure dependent
rate coefficients for the reaction of Hg with Cl and the reaction of Cl with Cl: a
pulsed laser photolysis-pulsed laser induced fluorescence study. J. Phys. Chem. A
109, 7732–7741 (2005).
2. Platt, U., Allan, W. & Lowe, D. Hemispheric average Cl atom concentration from
13
C/12C ratios in atmospheric methane. Atmos. Chem. Phys. 4, 2393–2399 (2004).
3. Knipping, E. M. & Dabdub, D. Impact of chlorine emissions from sea-salt aerosol
on coastal urban ozone. Environ. Sci. Technol. 37, 275–284 (2003).
4. Von Glasow, R. & Crutzen, P. J. in The Atmosphere (ed. Keeling, R. F.) 21–64
(Oxford Univ. Press, 2007).
5. Osthoff, H. D. et al. High levels of nitryl chloride in the polluted subtropical marine
boundary layer. Nature Geosci. 1, 324–328 (2008).
6. Allan, W., Struthers, H. & Lowe, D. C. Methane carbon isotope effects caused by
atomic chlorine in the marine boundary layer: global model results compared with
Southern Hemisphere measurements. J. Geophys. Res. D 112, D04306,
doi:10.1029/2006JD007369 (2007).
7. Finlayson-Pitts, B. J., Ezell, M. J. & Pitts, J. N. Formation of chemically active
chlorine compounds by reactions of atmospheric NaCl particles with gaseous
N2O5 and ClONO2. Nature 337, 241–244 (1989).
274
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
8. Behnke, W., George, C., Scheer, V. & Zetzsch, C. Production and decay of ClNO2
from the reaction of gaseous N2O5 with NaCl solution: bulk and aerosol
experiments. J. Geophys. Res. D 102, 3795–3804 (1997).
9. Graedel, T. E. & Keene, W. C. Tropospheric budget of reactive chlorine. Glob.
Biogeochem. Cycles 9, 47–77 (1995).
10. Kercher, J., Riedel, T. & Thornton, J. Chlorine activation by N2O5: simultaneous, in
situ detection of ClNO2 and N2O5 by chemical ionization mass spectrometry.
Atmos. Meas. Techn. 2, 193–204 (2009).
11. Pechtl, S. & von Glasow, R. Reactive chlorine in the marine boundary layer in the
outflow of polluted continental air: a model study. Geophys. Res. Lett. 34, L11813,
doi:10.1029/2007GL02976 (2007).
12. Logan, J. A., Prather, M. J., Wofsy, S. C. & McElroy, M. B. Tropospheric chemistry:
a global perspective. J. Geophys. Res. 86, 7210–7254 (1981).
13. Alexander, B. et al. Quantifying atmospheric nitrate formation pathways based on
a global model of the oxygen isotopic composition (D17O) of atmospheric nitrate.
Atmos. Chem. Phys. 9, 5043–5056 (2009).
14. Nelson, H. H. & Johnston, H. S. Kinetics of the reaction of Cl with ClNO and ClNO2
and the photochemistry of ClNO2. J. Phys. Chem. 85, 3891–3896 (1981).
15. Abuduwailli, J., Gabchenko, M. V. & Xu, J. R. Eolian transport of salts — a case
study in the area of Lake Ebinur (Xinjiang, Northwest China). J. Arid Environ. 72,
1843–1852 (2008).
16. Thornton, J. A., Braban, C. F. & Abbatt, J. P. D. N2O5 hydrolysis on sub-micron
organic aerosols: the effect of relative humidity, particle phase, and particle size.
Phys. Chem. Chem. Phys. 5, 4593–4603 (2003).
17. Wexler, A. S. & Clegg, S. L. Atmospheric aerosol models for systems including the
ions H1, NH41, Na1, SO422, NO32,Cl2, Br2, and H2O. J. Geophys. Res. D 107,
4207, doi:10.1029/2001JD000451 (2002).
18. Roberts, J. M., Osthoff, H. D., Brown, S. S. & Ravishankara, A. R. Laboratory
measurements of ClNO2 yields from N2O5 reactions as a function of chloride
molarity in various inorganic salt solutions. Geophys. Res. Lett. 36, L20808,
doi:10.1029/2009GL040448 (2009).
19. Moffet, R. C., de Foy, B., Molina, L. T., Molina, M. J. & Prather, K. A. Measurement
of ambient aerosols in northern Mexico City by single particle mass spectrometry.
Atmos. Chem. Phys. 8, 4499–4516 (2008).
20. Keene, W. C. & Savoie, D. L. The pH of deliquesced sea-salt aerosol in polluted
marine air. Geophys. Res. Lett. 25, 2181–2184 (1998).
21. Olivier, J. G. J. & Berdowski, J. J. M. in The Climate System (eds Berdowski, J.,
Guicherit, R. & Heij, B. J.) 33–78 (Balkema, and Swets and Zeitlinger, Lisse, The
Netherlands, 2001).
22. National Atmospheric Deposition Program Office. National Atmospheric Deposition
Program (NRSP-3) (Illinois State Water Survey, Champaign, Illinois, 2010).
23. Malm, W. C., Sisler, J. F., Huffman, D., Eldred, R. A. & Cahill, T. A. Spatial and
seasonal trends in particle concentration and optical extinction in the United
States. J. Geophys. Res. D 99, 1347–1370 (1994).
24. Erickson, D. J., Seuzaret, C., Keene, W. C. & Gong, S. L. A general circulation model
based calculation of HCl and ClNO2 production from sea salt dechlorination:
reactive chlorine emissions inventory. J. Geophys. Res. D 104, 8347–8372 (1999).
25. Zhang, Q. et al. Ubiquity and dominance of oxygenated species in organic aerosols
in anthropogenically-influenced Northern Hemisphere midlatitudes. Geophys.
Res. Lett. 34, L13801, doi:10.1029/2007GL029979 (2007).
26. Bertram, T. H. et al. Direct measurements of N2O5 reactivity on ambient aerosol
particles. Geophys. Res. Lett. 36, L19803, doi:10.1029/2009GL040248 (2009).
27. Brown, S. S. et al. Variability in nocturnal nitrogen oxide processing and its role in
regional air quality. Science 311, 67–70 (2006).
28. Singh, H. B. & Hanst, P. L. Peroxyacetyl nitrate (PAN) in the unpolluted atmosphere
— an important reservoir for nitrogen oxides. Geophys. Res. Lett. 8, 941–944 (1981).
29. Grannas, A. M. et al. An overview of snow photochemistry: evidence, mechanisms
and impacts. Atmos. Chem. Phys. 7, 4329–4373 (2007).
30. Shindell, D. T. et al. Improved attribution of climate forcing to emissions. Science
326, 716–718 (2009).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements Funding for the ClNO2 observations and analysis was
provided by NSF grants ATM-0633897 and ATM-0846183 to J.A.T. Boulder field
measurements were supported in part by the NOAA Atmospheric Chemistry and
Climate Program. J.P.K. thanks the Camille and Henry Dreyfus Foundation for a
postdoctoral fellowship in environmental chemistry. N.L.W. thanks the National
Research Council for a postdoctoral fellowship.
Author Contributions This was largely a collaborative effort. All authors
contributed to the collection of observations during the 2009 intensive campaign
described here. J.P.K., W.P.D., J.A.T. and S.S.B conceived of the need and initial
designs for the measurement campaign; J.A.T. performed much of the final
analysis and modelling; T.P.R. performed the predictions of continental-scale
ClNO2 production; and J.A.T. wrote the paper with significant input from all
co-authors, especially S.S.B.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to J.A.T.
(thornton@atmos.washington.edu).
Vol 464 | 11 March 2010 | doi:10.1038/nature08798
Antagonistic coevolution accelerates molecular
evolution
Steve Paterson1*, Tom Vogwill1*, Angus Buckling2, Rebecca Benmayor2, Andrew J. Spiers3, Nicholas R. Thomson4,
Mike Quail4, Frances Smith4, Danielle Walker4, Ben Libberton1, Andrew Fenton1, Neil Hall1
& Michael A. Brockhurst1*
The Red Queen hypothesis proposes that coevolution of interact-
ing species (such as hosts and parasites) should drive molecular
evolution through continual natural selection for adaptation and
counter-adaptation1–3. Although the divergence observed at some
host-resistance4–6 and parasite-infectivity7–9 genes is consistent
with this, the long time periods typically required to study coevolu-
tion have so far prevented any direct empirical test. Here we show,
using experimental populations of the bacterium Pseudomonas
fluorescens SBW25 and its viral parasite, phage W2 (refs 10, 11),
that the rate of molecular evolution in the phage was far higher
when both bacterium and phage coevolved with each other than
when phage evolved against a constant host genotype. Coevolution
also resulted in far greater genetic divergence between replicate
populations, which was correlated with the range of hosts that
coevolved phage were able to infect. Consistent with this, the most
rapidly evolving phage genes under coevolution were those
involved in host infection. These results demonstrate, at both the
genomic and phenotypic level, that antagonistic coevolution is a
cause of rapid and divergent evolution, and is likely to be a major
driver of evolutionary change within species.
According to the Red Queen hypothesis, biotic interactions are a
fundamental driver of molecular evolution2. The Red Queen hypo-
thesis posits that for a given species, its effective environment is likely
to be comprised of the other species in the ecosystem, such that an
adaptation increasing the fitness of one species necessarily causes a
decline in fitness of those species with which it interacts1,3. Such
coevolutionary interactions give rise to continual natural selection
for adaptation and counter-adaptation by ecologically interacting
species1,3, thereby driving molecular evolution2. Nowhere are such
evolutionary dynamics thought to be so prevalent as in interactions
between hosts and virulent parasites, in which selection is strongly
antagonistic yet closely coupled12. Comparative studies have found
particularly high rates of molecular evolution in genes associated
with infection7–9 or resistance to infection4–6. However, there have
been no direct empirical tests of whether antagonistic host–parasite
coevolution accelerates molecular evolution in parasite genomes,
and whether such evolution is particularly rapid at genes determining
infectivity.
Here we use experimental evolution of populations of the bac-
terium Pseudomonas fluorescens SBW25 and its viral parasite, phage
W2. We have previously demonstrated that P. fluorescens and W2
undergo a persistent coevolutionary ‘arms race’ with reciprocal selec-
tion for the evolution of new resistance and infectivity phenotypes
through time in bacteria and phage, respectively10,13, but the link
1
School of Biological Sciences, Biosciences Building, University of Liverpool, Crown Street, Liverpool L69 7ZB, UK. 2Zoology Department, University of Oxford, South Parks Road,
Oxford OX1 3PS, UK. 3SIMBIOS Centre, Level 5 Kydd Building, University of Abertay Dundee, Bell Street, Dundee DD1 1HG, UK. 4Pathogen Genomics, The Wellcome Trust Sanger
Institute, Wellcome Trust Genome Campus, Hinxton, Cambridge CB10 1SA, UK.
*These authors contributed equally to this work.
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
between this rapid phenotypic evolution and the underlying pattern
of molecular evolution has not been resolved. Crucially, it is possible
to separate bacteria and phage when transferring populations to fresh
media14, which allows one partner to be held evolutionarily constant
while the other partner is allowed to evolve15–17. Initially isogenic,
replicate populations of P. fluorescens and W2 were propagated by
serial transfer under two conditions: (1) evolution, in which the
bacterial genotype was held constant and only the phage was allowed
to adapt, and (2) coevolution, in which both the bacterium and the
phage were allowed to evolve adaptations and counter-adaptations.
At the end of the selection experiment we obtained whole-genome
sequences of phage populations by high coverage second-generation
sequencing to determine the identity and frequency of mutations in
each population. Mutations were partitioned into synonymous and
non-synonymous changes; very few synonymous mutations were
observed and only non-synonymous mutations were used in analyses
(see Supplementary Information; note that each indel (that is, inser-
tions or deletions) was counted as one mutation regardless of its
length). From these data we calculated the number of sites that had
acquired mutations in each population relative to the ancestral ref-
erence W2 sequence (obtained as part of this study; see Supplementary
Information), and from allele frequencies, the genetic distance of each
population from the ancestral W2 sequence, the genetic divergence
among populations and the genetic diversity within each population.
Coevolved phage populations showed twice the genetic distance
from the ancestor as that of the evolved populations (average pairwise
genetic distances: coevolved, 22.7 6 1.9 standard error (s.e.); evolved,
11.1 6 0.4 s.e.; t 5 6.64, d.f. 5 4.46, P , 0.01), shown also by the
increased branch lengths for coevolved populations in the phylogen-
etic tree in Fig. 1a. Similarly, coevolved populations had a greater
number of sites exhibiting mutations than evolved populations (co-
evolved, mean 52.8, range 46–60; evolved, mean 37.5, range 29–42;
likelihood-ratio test (LRT) 5 14.3, P , 0.001). Furthermore, far
greater genetic divergence was observed among replicate coevolved
populations than was observed among replicate evolved populations
(WST 5 0.45 for coevolved populations versus WST 5 0.06 for evolved
populations (Supplementary Table 1), in which WST is a measure of
the proportion of the total molecular variation attributable to differ-
ences among populations18). The tree in Fig. 1a also shows that rep-
licate populations from the same treatment grouped together
genotypically. This topology is not due to co-ancestry as all popula-
tions were split at the start of the experiment. Instead, the topology
reflects parallel evolution: selection acting independently at the same
sites among replicate populations. Thus, the tree reflects three key
275
LETTERS NATURE | Vol 464 | 11 March 2010

a b Phage infectivity traits. To address this prediction, we phenotypically char-

acterized the infectivity profile of each phage population. Specifically,
C4 we used cross-infection experiments to test whether phage from each
Scale population were able to infect hosts from all coevolved populations.
C1
We found that coevolved phage populations varied in terms of both

C4
C1
C3
C5

C3
E3 E2
E1
E6
C6
E5
E4
C5 ref
Figure 1 | Genetic and phenotypic responses to selection. a, Phylogenetic
tree for evolved (E1–6) and coevolved (C1, C3–6) phage populations and
ancestral reference genotype (ref) based on Euclidean distances calculated
from the frequency and identity of mutations in each population. Scale bar
indicates a Euclidean distance of one. b, The phage-infectivity range based
on the ability of each coevolved population to infect 20 bacterial clones from
each host population. Infection by phage is shown in red, and resistance by
hosts is shown in grey. The dendrogram indicates phenotypic similarity
between phage populations.
evolutionary patterns. Specifically, the extent to which replicates: (1)
followed a similar trajectory away from the ancestral sequence,
presumably as they adapted to laboratory conditions; (2) evolved
similarly among replicates within a treatment but differently in res-
ponse to consistent differences between treatments; and (3) showed
independent evolution within each replicate, and at a far higher rate in
the coevolved than the evolved treatment.
Increased genetic divergence between parasite populations due to
coevolution is likely to be driven primarily by divergent selection on
a b
C6
the range and identity of host genotypes that they were able to infect
C1 hosts (Fig. 1b), but that phage from evolved populations failed to infect any
coevolved hosts (data not shown). Indeed, phenotypic divergence of
the infectivity profile of coevolved phage populations closely
C3 hosts
matched the genetic divergence of the phage genomes as demon-
strated by the similar topologies of trees constructed using genetic
C4 hosts or phenotypic traits (Fig. 1b).
The increased rate of molecular evolution observed in the co-
C5 hosts
evolved populations was not distributed uniformly across the phage
genome (Fig. 2a). Four genes showed significantly increased molecular
evolution in coevolved versus evolved phage genomes. SBWP25_0036
C6 hosts (EMBL accession FN594518), which encodes a tail-fibre protein (gp49),
had a greater number of sites with mutations in the coevolved versus
evolved treatment, and, based on the allele frequencies of mutations at
these sites, a substantially higher divergence from the ancestral genotype
(number of mutational sites: coevolved, mean 17.6, range 15–20;
evolved, mean 12.7, range 11–14; LRT 5 4.43, P , 0.05; pairwise
genetic distance: coevolved, 10.86 6 0.90 s.e.; evolved, 4.51 6 0.19 s.e.;
t 5 7.69, d.f. 5 4.38, P , 0.01). SBWP25_0027, which encodes a struc-
tural protein (gp40), also had a greater number of sites with mutations
and a higher divergence from the ancestor in the coevolved than the
evolved populations (number of mutational sites: coevolved, mean
5.8, range 4–7; evolved mean 1.0, range 0–2; LRT 5 20.9, P , 0.001;
pairwise genetic distance: coevolved, 1.79 6 0.23 s.e.; evolved,
0.17 6 0.05 s.e.; t 5 7.69, d.f. 5 4.38, P , 0.01). SBWP25_0034 and
SBWP25_0035, which encode internal virion structural proteins
gp47 and gp48, respectively, also showed higher rates of molecular
evolution in coevolved populations, although to a lesser extent
than SBWP25_0027 and SBWP25_0036 (pairwise genetic distance:
SBWP25_0034 coevolved, 1.21 6 0.13 s.e.; evolved, 0.34 6 0.06 s.e.;
t 5 6.79, P , 0.01; SBWP25_0035 coevolved, 1.23 6 0.22 s.e.; evolved,
Pairwise diversity

12 8
Pairwise distance per gene

per gene

10 6
SBWP25_0027

SBWP25_0032

SBWP25_0034

SBWP25_0035
SBWP25_0036

8 4
6 2
4
2
0
0 10 20 30 40
Position on phage genome (kb)

0 10 20 30 40
0.0 0.2 0.4 0.6 0.8 1.0
Position on phage genome (kb)
Frequency
c
SBWP25_0027 SBWP25_0036
C1
C3
C4
C5
C6
E1
E2
E3
E4
E5
E6
0 500 0 500 1,000 1,500
Length (bp)

Figure 2 | Patterns of molecular evolution in the W2 genome. a, b, Pairwise within each population are shown as bars underneath each coding sequence,
genetic distance between each phage population and the ancestral genotype with the colour of each bar indicating the frequency of each mutation within
(a), and genetic diversity within each phage population (b). Symbols denote each population (white, rare; red, common). c, Magnified view of identity
means 6 s.e.m. of replicate populations within the coevolved (magenta; and frequency of mutations in each population for SBWP25_0027 (gp40)
n 5 5) and the evolved (blue; n 5 6) treatments. The locations of mutations and SBWP25_0036 (gp49). bp, base pairs.
276
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
0.48 6 0.06 s.e.; t 5 3.61, P , 0.05). SBWP25_0027 and SBWP25_0036
had a similar density of mutations (Fig. 2c) as each other, but, because
of its smaller size, SBWP25_0027 contributed less to the overall diver-
gence of the coevolved genomes from the ancestor than SBWP25_0036.
Consistent with the observed evolution of the phage-infectivity range
(Fig. 1b), all four of these proteins are predicted to be involved in host
attachment19. In tailed bacteriophages, attachment is a two-step process
consisting of an initial reversible adsorption by tail fibres, followed by
irreversible adhesion by structural proteins20. The average size of dele-
tions in the tail fibre gene (SBWP25_0036) was positively correlated to
the number of bacterial genotypes the phage populations could infect
(infectivity range) (Supplementary Fig. 1), suggesting that tail fibres are
under strong directional selection for reduced protein length during
coevolution but that the precise genetic changes varied between popu-
lations. Shortened tail fibres also evolved in the evolution treatment,
although to a lesser degree than under coevolution and without a
concomitant increase in the infectivity range (Supplementary Fig. 1),
suggesting that to some extent shorter tail fibres may also be a general
adaptation to laboratory conditions, perhaps through increasing
adsorption efficiency. SBWP25_0032, encoding a tail tubular protein
(gp45), showed divergence from the ancestor in both evolved and
coevolved treatments, but at different sites in each treatment.
SBWP25_0027 (gp40) and SBWP25_0036 (gp49) accounted for most
of the divergence between replicate, coevolved phage populations
(Supplementary Fig. 2).
Coevolved populations also showed higher genetic diversity within
populations than did evolved populations (Supplementary Table 1).
This was predominantly due to variation in SBWP25_0027 (gp40), as
SBWP25_0036 (gp49) displayed within-population diversity in both
treatments (Fig. 2b). This high genetic diversity at SBWP25_0036 is
surprising given the apparent directional selection for reduced tail-fibre
protein length during coevolution. This suggests that SBWP25_0036
polymorphisms may be transient and the result of recurrent con-
tinuing selective sweeps, and/or clonal interference. Alternatively,
both SBWP25_0036 and SBWP25_0027 may be subject to diversifying
or fluctuating selection within populations. Together these genes
(SBWP25_0036 and SBWP25_0027) are believed to control host
adhesion19; thus, it is possible that diversity at these genes may deter-
mine fine-scale host-specificity differences between phage genotypes.
Such phenotypic differences between individual phage clones from
the same population have been observed in a previous study in this
system16.
Overall, our results are consistent with accelerated evolution in the
coevolution treatment that is driven by selective effects, rather than
purely demographic differences between treatments. Demographic
effects, such as reduced generation time or population size, or
reduced fidelity of DNA replication, would have led to a genome-
wide increase in divergence and diversity, which was not observed. By
contrast, genetic divergence and diversity for most phage genes were
roughly similar in the two treatments (Fig. 2), indicating selection
under coevolution on specific infectivity genes/traits, such as that for
tail-fibre protein length or infectivity range (Supplementary Fig. 1).
Furthermore, whereas greater genetic divergence among coevolved
populations (Fig. 1a) could be explicable simply if coevolved popula-
tions are also smaller, and hence more susceptible to genetic drift, this
explanation is incompatible with the higher genetic diversity
observed in SBWP25_0027 in coevolved populations (Fig. 2b). In
line with this, there was no significant difference in phage population
size between treatments over the course of the experiment (log10
(plaque-forming units (p.f.u.) ml21) averaged through time: coe-
volved, 7.39 6 0.14 s.e.; evolved, 7.51 6 0.08 s.e.; t 5 0.71, d.f. 5 9,
P 5 0.5).
Our results highlight coevolution as a fundamental driver of mole-
cular evolution, and emphasize the utility of genome re-sequencing
for quantifying evolutionary dynamics in experimental evolution21.
We directly demonstrate that antagonistic coevolution accelerates
molecular evolution and can generate genetic divergence both between
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
and within populations. By contrast, populations adapting to a fixed
host genotype showed a remarkable degree of parallel evolution, indi-
cating genetic constraints on the evolutionary trajectories taken by
replicate populations22,23. Coevolutionary interactions between species
are likely therefore to be responsible for rapid evolutionary change
within species, potentially causing sufficient between-population
genetic divergence to drive speciation itself24.
METHODS SUMMARY
Twelve replicate microcosms (30 ml glass universals containing 6 ml of King’s B
(KB) broth) were inoculated with 107 isogenic cells of P. fluorescens SBW25 and
104 isogenic particles of phage W2. Cultures were propagated by 12 serial trans-
fers in a static incubator at 28 uC. Transfers for the six coevolving populations
involved transferring 60 ml (1%) of culture to a fresh KB microcosm every 48 h.
Transfers for the six evolving populations involved isolating phage populations
using 0.1 vol. chloroform and centrifuging at 14,000g for 2 min, and then inocu-
lating fresh microcosms with 60 ml (1%) of the phage population plus 107 ances-
tral SBW25 cells every 48 h. Every two transfers we estimated phage population
density by plating dilutions of each phage population onto KB agar plates with a
semi-solid overlay bacterial lawn. At the end of the experiment phage DNA was
isolated from each population25 and sequenced on a Roche 454 Titanium pyro-
sequencer. Reads were mapped to the W2 reference sequence and mutations were
identified and their frequencies calculated using the Roche Newbler mapping
tool. We used all non-synonymous changes to construct a phylogenetic tree
using Euclidean genetic distance (the square root of pairwise differences), which
is suggested as an appropriate metric for molecular variation data18. To deter-
mine phage population infectivity profiles, 20 independent bacterial colonies
were isolated from each of the five coevolved populations by plating on KB agar,
and these were streaked against a perpendicular line of each phage population
that had been previously applied to a KB agar plate10. A bacterial colony was
deemed susceptible if it showed any inhibition of growth after encountering the
line of phage.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 8 September; accepted 23 December 2009.
Published online 24 February 2010.
1. Van Valen, L. A new evolutionary law. Evol. Theory 1, 1–30 (1973).
2. Van Valen, L. Molecular evolution as predicted by natural selection. J. Mol. Evol. 3,
89–101 (1974).
3. Stenseth, N. C. & Smith, J. M. Coevolution in ecosystems: Red Queen evolution or
stasis? Evolution 38, 870–880 (1984).
4. Clark, A. G. et al. Evolution of genes and genomes on the Drosophila phylogeny.
Nature 450, 203–218 (2007).
5. Hedrick, P. W. Evolutionary genetics of the major histocompatibility complex. Am.
Nat. 143, 945–964 (1994).
6. Obbard, D. J., Jiggins, F. M., Halligan, D. L. & Little, T. J. Natural selection
drives extremely rapid evolution in antiviral RNAi genes. Curr. Biol. 16, 580–585
(2006).
7. Blanc, G. et al. Molecular evolution of Rickettsia surface antigens: evidence of
positive selection. Mol. Biol. Evol. 22, 2073–2083 (2005).
8. Mu, J. et al. Genome-wide variation and identification of vaccine targets in the
Plasmodium falciparum genome. Nature Genet. 39, 126–130 (2006).
9. Barrett, L. G. et al. Diversity and evolution of effector loci in natural populations of
the plant pathogen Melampsora lini. Mol. Biol. Evol. 26, 2499–2513 (2009).
10. Buckling, A. & Rainey, P. B. Antagonistic coevolution between a bacterium and a
bacteriophage. Proc. R. Soc. Lond. B 269, 931–936 (2002).
11. Brockhurst, M. A., Morgan, A. D., Fenton, A. & Buckling, A. Experimental
coevolution with bacteria and phage: the Pseudomonas fluorescens–W2 model
system. Infect. Genet. Evol. 7, 547–552 (2007).
12. Woolhouse, M. E., Webster, J. P., Domingo, E., Charlesworth, B. & Levin, B. R.
Biological and biomedical implications of the co-evolution of pathogens and their
hosts. Nature Genet. 32, 569–577 (2002).
13. Brockhurst, M. A., Morgan, A. D., Rainey, P. B. & Buckling, A. Population mixing
accelerates coevolution. Ecol. Lett. 6, 975–979 (2003).
14. Morgan, A., Gandon, S. & Buckling, A. The effect of migration on local adaptation
in a coevolving host–parasite system. Nature 437, 253–256 (2005).
15. Morgan, A. & Buckling, A. Relative number of generations of hosts and parasites
does not influence parasite local adaptation in coevolving populations of bacteria
and phages. J. Evol. Biol. 19, 1956–1963 (2006).
16. Poullain, V., Gandon, S., Brockhurst, M. A., Buckling, A. & Hochberg, M. E. The
evolution of specificity in evolving and coevolving antagonistic interactions
between a bacteria and its phage. Evolution 62, 1–11 (2008).
17. Benmayor, R., Hodgson, D. J., Perron, G. G. & Buckling, A. Host mixing and disease
emergence. Curr. Biol. 19, 764–767 (2009).
277
LETTERS
18. Excoffier, L., Smouse, P. E. & Quattro, J. M. Analysis of molecular variance inferred
from metric distances among DNA haplotypes: application to human
mitocondrial DNA restriction data. Genetics 131, 479–491 (1992).
19. Ceyssens, P. J. et al. Genomic analysis of Pseudomonas aeruginosa phages LKD16
and LKA1: establishment of the WKMV subgroup within the T7 supergroup.
J. Bacteriol. 188, 6924–6931 (2006).
20. Calendar, R. L. The Bacteriophages (Oxford Univ. Press, 2005).
21. Barrick, J. E. et al. Genome evolution and adaptation in a long-term experiment
with Escherichia coli. Nature 461, 1243–1247 (2009).
22. Bull, J. J. et al. Exceptional convergent evolution in a virus. Genetics 147, 1497–1507
(1997).
23. Wichman, H. A., Badgett, M. R., Scott, L. A., Boulianne, C. M. & Bull, J. J. Different
trajectories of parallel evolution during viral adaptation. Science 285, 422–424
(1999).
24. Thompson, J. N. The Geographic Mosaic of Coevolution (Univ. of Chicago Press,
2005).
25. Santos, M. A. An improved method for the small scale preparation of
bacteriophage DNA based on phage precipitation by zinc chloride. Nucleic Acids
Res. 19, 5442 (1991).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
278
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
Acknowledgements We are grateful to M. Begon and G. Hurst for comments on
previous drafts of this manuscript. We acknowledge funding from the Natural
Environment Research Council, the Wellcome Trust, the Leverhulme Trust and the
European Research Council. We are grateful for technical assistance from staff at
the Centre for Genomic Research, University of Liverpool.
Author Contributions M.A.B. and S.P. conceived and designed the study; T.V.
performed selection experiments, infectivity assays and prepared samples for
population genomic sequencing; B.L. and M.A.B. performed further assays; R.B.
prepared samples for ancestral genome sequencing; N.R.T., M.Q., F.S. and D.W.
performed ancestral genome sequencing and assembly; A.J.S. and N.R.T.
performed genome annotation; S.P., T.V. and M.A.B. analysed the data; M.A.B.,
S.P., A.B., A.J.S. and N.R.T. wrote the manuscript; T.V. and N.H. commented on the
manuscript; M.A.B., A.F. and S.P. supervised the research; S.P., N.H., A.B. and
M.A.B. obtained funding.
Author Information The ancestral genome sequence of phage W2 has been
submitted to the EMBL Nucleotide Sequence Database under accession number
FN594518. Reprints and permissions information is available at www.nature.com/
reprints. The authors declare no competing financial interests. Correspondence
and requests for materials should be addressed to M.A.B.
(michael.brockhurst@liverpool.ac.uk).
doi:10.1038/nature08798
METHODS
Selection experiment. A single bacterial colony of P. fluorescens SBW25 was
isolated from laboratory stocks by plating on KB agar, and a single plaque of
W2 was isolated from an agar overlay of P. fluorescens infected with a sample of
W2 from laboratory stocks. Each was grown up overnight and used to found 12
populations with 107 SBW25 cells and 104 phage particles. Cultures were pro-
pagated in 30-ml glass universals containing 6 ml KB media by serial transfer for
24 days in a static incubator at 28 uC. Transfers for the coevolved treatment
involved transferring 60 ml (1%) of culture to a fresh KB microcosm every
48 h. The evolved treatment involved isolating phage populations using
0.1 vol. chloroform and centrifuging at 14,000g for 2 min, and then inoculating
fresh microcosms with 60 ml (1%) of the phage population plus 107 ancestral
SBW25 cells every 48 h. Every two transfers, phage densities were estimated by
plating dilutions of each population onto KB agar plates with a semi-solid
overlay bacterial lawn.
Molecular methods. The complete sequence of the ancestral W2 clone was
obtained through a whole-genome shotgun library approach. W2 DNA was
isolated from an infected SBW25 culture after lysis with 0.1 vol. chloroform,
centrifugation at 14,000g for 2 min, and extraction from the supernatant as
previously described25. The purified phage DNA was used to produce 2–4 kb
and 4–6 kb size-fractionated subclone libraries in pCR4-TOPO (Invitrogen).
Sequencing was performed using ABI BigDye Terminator cycle sequencing kits
with sample analysis performed on ABI 3730xl instruments. The finished
sequence was generated from a total of 1,104 good quality reads (93% pass rate;
867,704 bp; theoretical coverage of 320) including 363 finishing reads used to
close gaps and span low coverage regions. The genome was assembled using
Phrap and was tested by BLASTN against the P. fluorescens SBW25 genome
sequence to ensure that no chromosomal sequences had been incorporated into
the phage alignment. The sequence was annotated in Artemis26 (Wellcome Trust
Sanger Institute) where putative coding sequences (CDS) were investigated by
BLASTP against the non-redundant (nr) NCBI protein database. Highest homo-
logies were found to LKA1 bacteriophage, and CDS named accordingly. The
genome has been submitted to EMBL under accession number FN594518.
Phage DNA was extracted from selection experiment populations as earlier
(sufficient quality and quantity of DNA for subsequent pyrosequencing was
obtained for five of six coevolved populations and for all evolved populations).
Shotgun single-stranded template DNA (sstDNA) libraries were prepared for
five coevolved populations (C1, C3–6) and six evolved populations (E1–6) by
nebulising DNA and ligating multiplex identifier tags according to manufac-
turer’s instructions. Libraries were then sequenced on a Roche 454 Titanium
pyrosequencer. The average sequence coverage for each library ranged from
374–354. Reads were mapped to the W2 reference sequence and mutations were
identified and their frequencies calculated using the Roche Newbler mapping
tool. Although some mutations were fixed in individual populations, no single
mutation was fixed across all populations, indicating that the reference sequence
used was that of the ancestral genotype.
©2010 Macmillan Publishers Limited. All rights reserved
Infectivity profiles. Twenty independent bacterial colonies were isolated from
each of the five coevolved populations by plating on KB agar. Samples from each
coevolved and evolved phage population were isolated using 0.1 vol. chloroform
and centrifugation at 14,000g for 2 min to produce phage supernatants. An
infectivity profile for a phage population was determined by drying lines of
phage on KB agar plates, and perpendicularly streaking all isolated bacterial
colonies across it. A bacterial colony was deemed susceptible if it showed any
inhibition of growth after encountering the line of phage10.
Population genetic methods. The average number of pairwise genetic differences
between each population and the ancestral genotype was calculated from the
frequency of each mutation identified by the Newbler mapping tool. To minimise
potential bias due to variation in sequence coverage between libraries, all non-
synonyomous mutations identified were used, including potential sequencing
errors present at low frequency; although the results were essentially identical if
analyses were limited to only high-confidence mutations or to non-synonymous
and synonymous mutations combined. (Because the frequency of sequencing
errors and synonymous mutations were low they contributed little to genetic
distance measures.) Indels, regardless of size, were counted as a single mutation.
The number of sites with mutations was calculated based on sites with a mutation
frequency of .5%. This threshold was used to avoid including sequencing errors,
which were likely to have occurred, on average, at comparable magnitudes in both
treatments (for further discussion of handling sequencing errors in heterogeneous
population samples see ref. 27). The significance of different genetic distances
between treatments was determined using Welch two sample t-tests, and for
numbers of sites containing mutations using generalized linear models with a
Poisson error distribution. A phylogenetic tree was constructed using Euclidean
genetic distance (the square root of pairwise differences), which is suggested as an
appropriate metric for molecular variation data18. We note that it is possible that a
small number of the mutations that we observed could have arisen in the original
outgrowth from which all the populations were initiated; thus, some allelic vari-
ation could have been shared by all the populations at the beginning of the
selection experiment. However, any such effect would not be sufficient to explain
our results, as replicate populations from the same treatment would only group
together within the tree (as observed in Fig. 1a) where selection increases the
frequency of the same allele independently in replicate populations in one treat-
ment but not in the other treatment. Analysis of molecular variation (AMOVA)
was conducted to apportion molecular variation into components due to vari-
ation within populations, among populations and between treatments18. To
achieve this, AMOVA was conducted on populations in each treatment separately
and on all populations combined. All analyses were conducted in R v2.9 and
Perl v5.8.
26. Rutherford, K. et al. Artemis: sequence visualization and annotation. Bioinformatics
16, 944–945 (2000).
27. Barrick, J. E. & Lenski, R. E. Genome-wide mutational diversity in an evolving
population of Escherichia coli. Cold Spring Harb. Symp. Quant. Biol., doi:10.1101/
sqb.2009.74.018 (23 September 2009).
Vol 464 | 11 March 2010 | doi:10.1038/nature08691

LETTERS
Compensatory evolution in mitochondrial tRNAs
navigates valleys of low fitness
Margarita V. Meer1, Alexey S. Kondrashov2, Yael Artzy-Randrup2,3 & Fyodor A. Kondrashov1

A long-standing controversy in evolutionary biology is whether or
not evolving lineages can cross valleys on the fitness landscape that
correspond to low-fitness genotypes, which can eventually enable
them to reach isolated fitness peaks1–9. Here we study the fitness
landscapes traversed by switches between different AU and GC
Watson–Crick nucleotide pairs at complementary sites of mito-
chondrial transfer RNA stem regions in 83 mammalian species.
We find that such Watson–Crick switches occur 30–40 times more
slowly than pairs of neutral substitutions, and that alleles corres-
ponding to GU and AC non-Watson–Crick intermediate states
segregate within human populations at low frequencies, similar to
those of non-synonymous alleles. Substitutions leading to a Watson–
Crick switch are strongly correlated, especially in mitochondrial
tRNAs encoded on the GT-nucleotide-rich strand of the mitochon-
drial genome. Using these data we estimate that a typical Watson–
Crick switch involves crossing a fitness valley of a depth of about
1023 or even about 1022, with AC intermediates being slightly more
deleterious than GU intermediates. This compensatory evolution
must proceed through rare intermediate variants that never reach
fixation2. The ubiquitous nature of compensatory evolution in mam-
malian mitochondrial tRNAs and other molecules10,11 implies that
simultaneous fixation of two alleles that are individually deleterious
may be a common phenomenon at the molecular level.
When fitness is affected by complex interactions between alleles at
different loci, an evolving lineage must navigate an epistatic fitness
landscape1,12,13. The most basic mode of epistasis is the compensatory
interaction of two diallelic loci (sites) at which each allele can confer a
high fitness when combined with the suitable allele at the other
locus10. Nevertheless, even this basic mode of epistasis can lead to
fitness landscapes of two types (Fig. 1). The first is a continuous ridge
of high fitness that circumvents a maladapted genotype, commonly
referred to as a Dobzhansky–Muller incompatibility10,14,15. Evolution
along continuous fitness ridges has the key role in speciation16 and in
compensation of pathogenic mutations in proteins and tRNAs10,11.
The second is a pair of isolated fitness peaks separated by a fitness
valley, first considered by Sewall Wright in the context of his shifting-
balance model1,2.
The importance of crossing fitness valleys in evolution remains
unknown1–9, and we study this issue with regard to the evolution of
complementary mitochondrial tRNA stem sites. The evolution of
RNA molecules can proceed through switches between different
Watson–Crick pairs of nucleotides at a pair of interacting sites.
Four kinds of Watson–Crick switch are possible. The most common
is the switch between an AU and GC Watson–Crick pair, or vice versa,
with the intermediate state of AC or GU (AU–GC type). The three
other types of switch are AU–UA, GC–CG and AU–CG, with the
intermediate states being AA or UU, GG or CC, and AG or UC,
respectively. Watson–Crick switches involve compensation at the
1
Bioinformatics and Genomics Programme, Centre for Genomic Regulation, C/Dr Aiguader 88, Barcelona Biomedical Research Park Building, 08003 Barcelona, Spain. 2Life Sciences
Institute, Evolutionary Biology and Ecology Department, University of Michigan, Ann Arbor, Michigan 48109, USA. 3Howard Hughes Medical Institute and Department of Ecology and
Evolutionary Biology, University of Michigan, Ann Arbor, Michigan 49109, USA.
molecular level, because they maintain the high stability of the RNA
stem structures. Here we study whether or not such switches are also
compensatory at the level of fitness.
We obtained sequences, secondary structure predictions and
alignments for 1,825 mitochondrial tRNAs consisting of 22 non-
redundant sets from 83 mammalian species17, and subdivided them
into seven clades, each with an unambiguous phylogeny (see Methods).
For each clade we reconstructed the ancestral states for substitutions in
the stem secondary structures of mitochondrial tRNAs (mt-tRNAs)
and calculated the number of Watson–Crick switches for each of the
22 mt-tRNAs (see Methods and Supplementary Table 1). We recorded
a Watson–Crick switch when at a pair of interacting sites two taxa
harboured different Watson–Crick pairs of nucleotides and the switch
between them was caused by exactly two substitutions, as judged from
the reconstructed ancestral states (Fig. 2). Out of a total of 2,867 sub-
stitutions in mt-tRNA stems of the seven clades, we observed 1,532
Watson–Crick switches of all the four types, with a vast majority of
them belonging to the AU–GC type. Thus, at least 55% of all substitu-
tions observed in stem structures of mt-tRNAs are involved in Watson–
Crick switches. The prevalence of errors in our results caused by mul-
tiple substitutions at each site must be low, as demonstrated by a small
fraction (about 6% for the AU–GC type) of Watson–Crick switches
involving more than two substitutions (Table 1).
Continuous Isolated
fitness ridges fitness peaks
a c
AC GC AC GC
AU GU AU GU
b d
AC GC AC GC
AU GU AU GU
High fitness Low fitness
Figure 1 | Fitness landscapes resulting from an interaction between two
complementary sites in an RNA stem structure. A switch between AU and
GC Watson–Crick pairs can proceed on flat (a) or ascending (b) continuous
fitness ridges without a decrease in fitness, or through a fitness valley in
which both possible intermediate variants have reduced fitness by the same
(c) or a different (d) amount.

279
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS NATURE | Vol 464 | 11 March 2010

a GC Eubalaena australis
b GC Artibeus jamaicensis
GC GC
GC GC Eubalaena japonica
GC Mystacina tuberculata
GC Balaena mysticetus GC
GC GC Caperea marginata AU Chalinolobus tuberculatus
AU
AU Balaenoptera bonaerensis AU Pipistrellus abramus
GC
GC GC Balaenoptera acutorostrata GC
GC Rhinolophus monoceros
GC Megaptera novaeangliae GC
GC GC
GC Balaenoptera physalus GC Rhinolophus pumilus
GC
GC
GC Eschrichtius robustus GC Rousettus aegyptiacus
GC
GC Balaenoptera brydei GC
GC GC Pteropus dasymallus
GC GC Balaenoptera borealis
GC
GC Balaenoptera musculus GC Pteropus scapulatus

c AU Acinonyx jubatus d AU Hemiechinus auritus

AU AU
AU AU Felis catus
AU Erinaceus europaeus
AU
AU Herpestes javanicus
AU Canis familiaris AU Echinosorex gymnura
AU
AU Ursus americanus GC Mogera wogura
AU AU AU Ursus arctos ?? GC
AU
GC GC Talpa europaea
AU Ursus maritimus
AU AU Arctocephalus forsteri GC Urotrichus talpoides
AU
AU Eumetopias jubatus GC
AU GC Sorex unguiculatus
AU Odobenus rosmarus GC
AU
GC Episoriculus fumidus
GC Halichoerus grypus GC
GC
GC Phoca vitulina GC Crocidura russula

Figure 2 | Watson–Crick switches in mammalian mt-tRNAs. Examples Watson–Crick switches were detected on terminal (a) and internal
from mt-tRNATyr in Mysticeti (a), mt-tRNAPhe in Chiroptera (b), mt- (b–d) branches; however, those switches that spanned the root of the tree
tRNAHis in Carnivora (c) and mt-tRNAPro in Eulipotyphla (d). (d) were not considered in our analysis.

To address the key question of whether Watson–Crick switches
involve crossing fitness valleys, we studied substitutions and poly-
morphism of mt-tRNAs. We first investigated the evolutionary inter-
dependence of the two substitutions involved in an AU«GC
Watson–Crick switch. If the fitness landscape is a flat continuous
ridge (Fig. 1a), both substitutions are neutral and they are expected
to occur independently of each other, without clustering on the
phylogenetic tree18,19. The extent of clustering can be characterized
by the ratio of the numbers of terminal to intermediate states (the
terminal-to-intermediate ratio, TIR) in the last common ancestor
(LCA) of the species separated by a Watson–Crick switch. The
LCA of the species separated by an AU–GC switch can either be
identical to one of them—that is, to the terminal state (AU or
GC)—or it can coincide with the state intermediate between these
Watson–Crick pairs (AC or GU). When the LCA is identical to a
terminal state, two substitutions must have occurred in the same
lineage, whereas when the LCA corresponds to an intermediate state,
one substitution must have occurred on each of the two lineages
leading to the terminal states. If all substitutions were selectively
neutral, the TIR is expected to be 1:1 (LCA terminal:LCA intermedi-
ate) and would support the flat continuous ridge shown in Fig. 1a; a
TIR with more frequent LCA terminal states indicates clustering of
the two substitutions involved in a Watson–Crick switch and selec-
tion favouring at least the second of them18,19 (Fig. 1b–d).
For all AU«GC switches, we observed 568 cases in which the LCA
was identical to the terminal state and 157 cases in which the LCA was
identical to an intermediate state, resulting in a TIR of 3.6:1. In
addition to selection, the TIR can also be affected by unequal lengths
of the lineages of the two species, changes in the mutation rate, and
incorrect identification of the LCA state18,19. To ascertain the impact
of these factors, we assembled pseudo-Watson–Crick pairs by choos-
ing all random pairs of sites from stems of different mt-tRNAs and
applied a randomized sampling technique (see Methods). For these
randomly linked pairs of tRNA stem sites we obtained a TIR of 1:1.9
(1,248:2,316). A similar ratio of 1:2.2 (267,923:585,532) was obtained
for random pairs of fourfold synonymous sites of different mito-
chondrial protein-coding genes. Thus, a 3.6:1 TIR at interacting sites
exceeds the roughly 1:2 TIR observed in our randomization tests by
about sevenfold, and the difference is statistically significant
(P , 0.001, Fisher’s t-test).
Two kinds of fitness landscape can lead to such clustering of sub-
stitutions. First, an AU«GC switch can proceed, perhaps after a
change in the fitness landscape caused by evolution elsewhere in
the genome, on an ascending continuous fitness ridge such that the

Table 1 | Number of Watson–Crick switches in mammalian mt-tRNAs

Type of compensatory switch Terminal ancestral state Intermediate ancestral state Ancestral state not reconstructed
and tRNA
Ancestral state AU Ancestral state GC Ancestral state GU Ancestral state AC Ancestral state in More than two
root of the tree substitutions
GU-rich tRNAs AU«GC 103 68 68 1 173 13
AC-rich tRNAs AU«GC 239 158 31 57 356 38
AU«CG Ancestral state AU Ancestral state CG Ancestral state AU Ancestral state CG
4 2 0 2 6 39
CG«GC Ancestral state CG or GC Ancestral state CC Ancestral state GG
0 0 0 2 3
AU«UA Ancestral state AU or UA Ancestral state UU Ancestral state AA
27 5 1 24 22

280
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
fitness of the intermediate state GU or AC is no less than the fitnesses
of the initial Watson–Crick state (Fig. 1b). Second, an AU«GC
switch can proceed through a fitness valley (Fig. 1c, d), with the
two compensatory substitutions fixed together in the same haplo-
type2,3. These situations predict a different frequency of intermediate
states in Watson–Crick switches. Because GU pairs are more stable
than AC pairs in RNA stem structures20 and stability is a key factor
connecting mt-tRNAs to fitness21–23, GU pairs should on average
confer a higher fitness than AC pairs. Thus, if AU«GC switches
traverse a continuous ascending ridge (Fig. 1b), evolution should
occur preferentially through the GU intermediates and only rarely
through the AC intermediates, whereas AU«GC switches between
isolated peaks can occur through both GU or AC intermediates.
Because we use a phylogenetic approach to reconstruct the ances-
tral states, we cannot directly infer the intermediate state in a com-
pensatory switch (see Fig. 2 for examples). We therefore apply an
indirect approach by using the strong strand-specific mutation bias
present in mammalian mitochondria. In the GT-rich strand, the rates
of CRT and ARG transitions are increased about sixfold, and in the
AC-rich strand there is a reciprocal increase in the rates of GRA and
TRC transitions24. Accordingly, GU-rich mt-tRNAs (eight mt-
tRNAs encoded on the AC-rich strand) have a roughly sixfold higher
CRU and ARG mutation rate than the URC and GRA rate, and
this pattern is reversed in AC-rich mt-tRNAs (14 mt-tRNAs encoded
on the GT-rich strand). Thus, the mutational bias in AC-rich tRNAs
favours an AC intermediate in AU«GC switches, whereas a GU
intermediate is favoured by mutation in GU-rich tRNAs (Fig. 3).
Thus, the continuous ascending ridge model (Fig. 1b) predicts that
AU«GC switches should be more common in GU-rich tRNAs and
that the TIR should be similar in GU-rich and AC-rich tRNAs. In
contrast, in the isolated peaks model, which can involve both GU and
AC intermediates, only a marginally higher rate of AU«GC switches
in GU-rich tRNAs and a much higher TIR for AC-rich tRNAs2 are
expected.
The prevalence of AU«GC switches was roughly the same in the
two strands, with 1.6 switches in each stem site in GU-rich mt-tRNAs
and 1.4 in AC-rich mt-tRNAs (Table 1) rejecting the ascending con-
tinuous fitness ridge model. In contrast, the TIR was substantially
higher in 14 AC-rich tRNAs than in 8 GU-rich tRNAs (397:88 (4.5:1)
versus 171:69 (2.5:1), respectively), supporting the isolated peaks
model. This model was further supported by a 33:8 (4.1:1) TIR test
on non-AU«GC switches, in which the intermediate states are
expected to be less stable than the UG and AC pairs20.
As a further test of the applicability of the isolated peaks fitness
landscapes to mt-tRNAs, we applied Kimura’s model of compen-
satory evolution2 to data on the rate of Watson–Crick switches in
tRNAs and the rate of synonymous substitutions in mitochondrial
protein-coding genes. We found that the rates of AU«GC switches
in GU-rich and AC-rich tRNAs are 30-fold and 40-fold lower,
respectively, than the rate of synonymous evolution. Because Ne,
the effective population size, is about 10,000 in mammals, this
implies that the coefficient of selection against the intermediate states
is about 5 3 1023 and about 8 3 1023 in GU-rich and AC-rich
tRNAs, respectively (Figure 1 in ref. 2), as long as the strength of
selection is uniform at all sites. This difference between strands is to
GU
AU GC
AC
Figure 3 | Mutation strand bias in mammalian mitochondria. Mutations
with an increased rate in GU-rich (red) and AC-rich (blue) tRNAs are shown.
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
be expected if compensatory evolution in GU-rich tRNAs proceeds
mostly through less deleterious GU intermediate states.
We corroborated the estimated coefficients of selection against GU
and AC intermediate states obtained from Kimura’s model2 with data
on the frequency of GU and AC alleles from 5,004 complete human
mitochondrial genotypes. The AC and the GU alleles were substan-
tially rarer than synonymous alleles, with the AC alleles being slightly
rarer and the GU alleles slightly more common than non-synonymous
alleles (Fig. 4 and Supplementary Table 2). Because non-synonymous
substitutions in the mitochondria are under stringent selection25, both
GU and AC intermediates must also be substantially deleterious, with
AC being slightly more so.
Thus, the data suggest that Nes ? 1, where s is the coefficient of
selection against an intermediate allele, so that we can ignore random
drift and reverse mutations from rare deleterious alleles and use the
approximation of p 5 m/s to estimate selection from polymorphism
data, where m is the rate of mutation and p is the average frequency of
the segregating deleterious allele26. In mitochondria m has been esti-
mated as 2 3 1025 (ref. 27) with cumulative strand-specific rates of
TRC and GRA mutations as 2.8 3 1026 and 1.5 3 1025 in AC-rich
and GU-rich tRNAs, respectively24,28 (see Methods). We determined
p as 1.7 3 1023 and 2.0 3 1024 for segregating GU and AC alleles,
respectively, in GU-rich tRNAs, and 3.9 3 1024 and 1.4 3 1023 for
GU and AC alleles in AC-rich tRNAs. Combining the estimates of m
and p we obtain coefficients of selection against GU and AC inter-
mediates in stem structures of mt-tRNAs as about 8 3 1023 and
about 1022, respectively (Supplementary Table 3), in good agree-
ment with the estimates based on the rate of evolution. Moreover,
there is practically no overlap between the frequency distributions of
GU and AC alleles on the one hand, and of synonymous alleles on the
other, which rules out a substantial fraction of effectively neutral
intermediate states (Fig. 4). Data on eight other mammalian species,
a 1.0
AC alleles
GU alleles
Non-synonymous alleles
Fourfold synonymous alleles
Frequency
0.5
0
b 1.0
Frequency
0.5
0
0 1 2 3 4 5–10 >10
Number of individuals carrying the minor allele
Figure 4 | Polymorphism frequency distribution of minor alleles in the
human population. a, Derived alleles produced by high-rate mutation,
ARG and CRT on the heavy strand and GRA and TRC on the light strand.
b, Alleles produced by lower-rate mutations, ARG and CRT on the light
strand and GRA and TRC on the heavy strand.
281
LETTERS
in which the AC and GU alleles were also rarer than synonymous
alleles, confirm that GU and AC alleles are substantially deleterious
(Supplementary Table 4).
Taken in isolation, the decreased rate of AU«GC switches or the
low frequency of GU and AC alleles may also be consistent with
evolution on a continuous fitness ridge if we assume that at a majority
of sites GU and AC intermediates are very strongly deleterious and
switches occur mostly at a small fraction of sites at which such inter-
mediates are effectively neutral. However, the similarity of estimates
of selection coefficients against GU and AC intermediates obtained
from the application of the Kimura model2 to evolutionary data on
the one hand, and from frequencies of segregating alleles on the
other, strongly supports the hypothesis that AU«GC switches occur
through a fitness valley. Indeed, the similarity of these estimates is not
predicted by the continuous ascending fitness ridge model, because
values of selection coefficients for substitutions fixed by positive
selection must be very different from those characterizing segregating
polymorphisms29.
Taken together, the data on the rates of AU«GC switches, on the
frequencies of segregating GU and AC intermediate alleles and on the
clustering of substitutions involved in a Watson–Crick switch demon-
strate that evolving mt-tRNAs often cross fitness valleys of a substan-
tial depth of about 1023 or even about 1022. This selection against
intermediate alleles is due to the impact of altered tRNAs both on the
fitness of the organism and on the rate of propagation of a mitochon-
drion within the germ line25 and characterizes the efficiency of trans-
mission of a mitochondrial genotype between generations. Alleles
with such low fitnesses cannot be fixed in a population, and crossing
the deep valleys must occur without the assistance of genetic drift2,3.
Instead, compensatory evolution in mt-tRNAs involves a compen-
satory substitution emerging in the rare genotype carrying a segre-
gating strongly deleterious allele, making their simultaneous fixation
possible2,3. Thus, strongly deleterious alleles can serve as evolutionary
stepping stones by providing access to isolated fitness peaks. Whether
or not such stepping stones can be used to cross wider fitness valleys
remains an open question.
METHODS SUMMARY
Substitutions in stem structures of mt-tRNAs obtained from a mt-tRNA data-
base30 were reconstructed with the use of parsimony, maximum-likelihood and
Bayesian methods. A Watson–Crick switch was recorded when two different taxa
displaying one of the two Watson–Crick pairs were separated by exactly two
substitutions. The average frequency of segregating minor alleles in Watson–
Crick complementary stem sites was measured in 5,004 complete human mito-
chondrial genotypes. To estimate the strand-specific mutation rates we used the
total mutation rate of 2 3 1025 from ref. 27, an estimate of about 90% of all
mutations as transitions from ref. 28, and a combined bias of a sixfold increase in
ARG and CRT in the GT-rich strand and the same bias of reverse mutations in
the AC-rich strand24.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 3 June; accepted 16 November 2009.
Published online 24 February 2010.
1. Wright, S. in Proceedings of the Sixth International Congress of Genetics, Ithaca, New
York Vol. 1 (ed. Jones, D. F.) 356–366 (Brooklyn Botanic Garden, 1932).
2. Kimura, M. The role of compensatory neutral mutations in molecular evolution. J.
Genet. 64, 7–19 (1985).
3. Innan, H. & Stephan, W. Selection intensity against deleterious mutations in RNA
secondary structures and rate of compensatory nucleotide substitutions. Genetics
159, 389–399 (2001).
4. Crow, J. F. Mid-century controversies in population genetics. Annu. Rev. Genet. 42,
1–16 (2008).
5. Wade, M. J. & Goodnight, C. J. Wright’s shifting balance theory: an experimental
study. Science 253, 1015–1018 (1991).
6. Coyne, J. A., Barton, N. H. & Turelli, M. Is Wright’s shifting balance process
important in evolution? Evolution 54, 306–317 (2000).
282
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
7. Weinreich, D. M. & Chao, L. Rapid evolutionary escape by large populations from
local fitness peaks is likely in nature. Evolution 59, 1175–1182 (2005).
8. Chen, Y. & Stephan, W. Compensatory evolution of a precursor messenger RNA
secondary structure in the Drosophila melanogaster Adh gene. Proc. Natl Acad. Sci.
USA 100, 11499–11504 (2003).
9. Piskol, R. & Stephan, W. Analyzing the evolution of RNA secondary structures in
vertebrate introns using Kimura’s model of compensatory fitness interactions.
Mol. Biol. Evol. 25, 2483–2492 (2008).
10. Kondrashov, A. S., Sunyaev, S. & Kondrashov, F. A. Dobzhansky–Muller
incompatibilities in protein evolution. Proc. Natl Acad. Sci. USA 99, 14878–14883
(2002).
11. Kern, A. D. & Kondrashov, F. A. Mechanisms and convergence of compensatory
evolution in mammalian mitochondrial tRNAs. Nature Genet. 36, 1207–1212
(2004).
12. Maynard Smith, J. Natural selection and the concept of protein space. Nature 225,
563–564 (1970).
13. Poelwijk, F. J., Kiviet, D. J., Weinreich, D. M. & Tans, S. J. Empirical fitness
landscapes reveal accessible evolutionary paths. Nature 445, 383–386 (2007).
14. Dobzhansky, T. Studies on hybrid sterility. II. Localization of sterility factors in
Drosophila pseudoobscura hybrids. Genetics 21, 113–135 (1936).
15. Muller, H. J. Reversibility in evolution considered from the standpoint of genetics.
Biol. Rev. Camb. Philos. Soc. 14, 261–280 (1939).
16. Orr, H. A. The population genetics of speciation: the evolution of hybrid
incompatibilities. Genetics 139, 1805–1813 (1995).
17. Popadin, K. Y., Mamirova, L. A. & Kondrashov, F. A. A manually curated database
of tetrapod mitochondrially encoded tRNA sequences and secondary structures.
BMC Bioinformatics 8, 441 (2007).
18. Bazykin, G. A., Kondrashov, F. A., Ogurtsov, A. Y., Sunyaev, S. & Kondrashov, A. S.
Positive selection at sites of multiple amino acid replacements since rat-mouse
divergence. Nature 429, 558–562 (2004).
19. Donmez, N., Bazykin, G. A., Brudno, M. & Kondrashov, A. S. Polymorphism due to
multiple amino acid substitutions at a codon site within Ciona savignyi. Genetics
181, 685–690 (2009).
20. Davis, A. R. & Znosko, B. M. Thermodynamic characterization of single
mismatches found in naturally occurring RNA. Biochemistry 46, 13425–13436
(2007).
21. Kelley, S. O., Steinberg, S. V. & Schimmel, P. Functional defects of pathogenic
human mitochondrial tRNAs related to structural fragility. Nature Struct. Biol. 7,
862–865 (2000).
22. Wittenhagen, L. M. & Kelley, S. O. Impact of disease-related mitochondrial
mutations on tRNA structure and function. Trends Biochem. Sci. 28, 605–611
(2003).
23. McFarland, R., Elson, J. L., Taylor, R. W., Howell, N. & Turnbull, D. M. Assigning
pathogenicity to mitochondrial tRNA mutations: when ‘definitely maybe’ is not
good enough. Trends Genet. 20, 591–596 (2004).
24. Tanaka, M. & Ozawa, T. Strand asymmetry in human mitochondrial DNA
mutations. Genomics 22, 327–335 (1994).
25. Stewart, J. B. et al. Strong purifying selection in transmission of mammalian
mitochondrial DNA. PLoS Biol. 6, e10 (2008).
26. Sunyaev, S. et al. Prediction of deleterious human alleles. Hum. Mol. Genet. 10,
591–597 (2001).
27. Howell, N. et al. The pedigree rate of sequence divergence in the human
mitochondrial genome: there is a difference between phylogenetic and pedigree
rates. Am. J. Hum. Genet. 72, 659–670 (2003).
28. Malyarchuk, B. A., Rogozin, I. B., Berikov, V. B. & Derenko, M. V. Analysis of
phylogenetically reconstructed mutational spectra in human mitochondrial DNA
control region. Hum. Genet. 111, 46–53 (2002).
29. McDonald, J. H. & Kreitman, M. Adaptive protein evolution at the Adh locus in
Drosophila. Nature 351, 652–654 (1991).
30. Popadin, K. Y., Mamirova, L. A. & Kondrashov, F. A. A manually curated database
of tetrapod mitochondrially encoded tRNA sequences and secondary structures.
BMC Bioinformatics 8, 441 (2007).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank H. Innan, M. Laessig, R. Guigo, I. Povolotskaya,
D. Ivankov and M. Breen for thoughtful discussions and critical reading of the
manuscript.
Author Contributions M.V.M., Y.A.R. and F.A.K. obtained the initial evolutionary
and polymorphism data. A.S.K. and F.A.K. supplied the theoretical treatment of the
primary data. F.A.K. designed the study. All authors participated in writing the
paper.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to F.A.K.
(fyodor.kondrashov@crg.es).
doi:10.1038/nature08691
METHODS
Measuring evolution and polymorphism. For all stem structures of 22 mt-tRNAs
encoded in mammalian mitochondrial genomes we estimated the numbers of
substitutions on the phylogenetic trees for each of the seven clades, using eight
different models as implemented in the PAML31 package (see Supplementary
Information). Watson–Crick switches of one of four types, AU«GC, AU«UA,
GC«CG and AU«CG, were recorded when two different taxa displayed one of
the two Watson–Crick pairs and were separated by exactly two substitutions.
For randomization of Watson–Crick pairs in stem sites, we assembled pseudo-
Watson–Crick pairs by selecting all possible pairs of individual sites from stems
of different mt-tRNAs. Thus, a pseudo-Watson–Crick pair was composed of
sites from stems in mt-tRNAs such that each site in a pair was picked from a
different tRNA. We then obtained the TIR measurement in such pseudo-
Watson–Crick pairs on each of the seven phylogenetic trees in the same way
as for real Watson–Crick pairs, which thus reflected the random expectation of
substitutions occurring in two random sites on a given tree. The same procedure
was followed for obtaining pseudo-Watson–Crick pairs from synonymous sites
except that each site in a pseudo-Watson–Crick pair was chosen from fourfold
synonymous sites.
To measure the rate of polymorphism in tRNAs we excluded mt-tRNAThr
because of its uncharacteristically large rate of polymorphism32. We measured
the average frequency of segregating minor alleles in stem sites that had an AU or
GC Watson–Crick pair in humans and that underwent at least one Watson–
Crick switch in mammalian evolution. To estimate the rate of mutation rate we
used the total mutation rate of 2 3 1025 from ref. 27, an estimate of about 90% of
all mutations as transitions from ref. 28 and a combined bias of sixfold increase
in ARG and CRT in the GT-rich strand and the same bias of reverse mutations
in the AC-rich strand24. We did not use the model of ref. 3 because it assumes a
low rate of mutation, Nem = 1, which is not true of mammalian mitochondria27.
Obtaining sequences and phylogenetic reconstruction. We obtained secondary-
structure-based mt-tRNA alignments from ref. 30. The protein-coding genes were
obtained from GenBank33. The phylogeny for each taxon was constructed from
muscle34 protein-based codon alignments of 13 mitochondrial protein coding
genes by using MrBayes under the HKY model and a gamma-distribution of
substitution rates across sites for 106 iterations35. Our final data set included 12
species in the Carnivora, 9 in Chiroptera, 12 in Cetartiodactyla, 17 in Primata, 9 in
Rodentia and 12 in Eulipotyphla. Twelve Mysticeti species were placed into an
independent phylogeny because the exact location of whales in the Cetartiodactyla
and the monophyletic status of Odontoceti remain controversial. We used the
following references to check the consistency and determine the root of each
phylogenetic tree: Carnivora36,37, Chiroptera38,39, Cetartiodactyla40,41, Primata42–44,
Rodentia45 and Eulipotyphla46,47 and Mysticeti41,48,49. The phylogenetic trees not
displayed in the main text are shown below.
For polymorphism estimates the sequences of complete human mitochon-
drial genotypes were obtained from GenBank by using ‘‘Homo sapiens [orgn]
and complete genome’’ as a query with the Limits option set to mitochondrial
DNA in the Entrez retrieval system33. Genomes with undetermined bases (Ns) in
mt-tRNA sequences were discarded, leaving 5,004 genomes. We also studied
©2010 Macmillan Publishers Limited. All rights reserved
eight other mammalian species for which more than ten complete genomes were
available in GenBank.
31. Yang, Z. PAML: a program package for phylogenetic analysis by maximum
likelihood. Comput. Appl. Biosci. 13, 555–556 (1997).
32. Kivisild, T. et al. The role of selection in the evolution of human mitochondrial
genomes. Genetics 172, 373–387 (2006).
33. Benson, D. A., Karsch-Mizrachi, I., Lipman, D. J., Ostell, J. & Wheeler, D. L.
GenBank. Nucleic Acids Res. 34, D16–D20 (2006).
34. Edgar, R. C. MUSCLE: multiple sequence alignment with high accuracy and high
throughput. Nucleic Acids Res. 32, 1792–1797 (2004).
35. Ronquist, F. & Huelsenbeck, J. P. MrBayes 3: Bayesian phylogenetic inference
under mixed models. Bioinformatics 19, 1572–1574 (2003).
36. Flynn, J. J., Finarelli, J. A., Zehr, S., Hsu, J. & Nedbal, M. A. Molecular phylogeny of
the carnivora (mammalia): assessing the impact of increased sampling on
resolving enigmatic relationships. Syst. Biol. 54, 317–337 (2005).
37. Fulton, T. L. & Strobeck, C. Molecular phylogeny of the Arctoidea (Carnivora):
effect of missing data on supertree and supermatrix analyses of multiple gene
data sets. Mol. Phylogenet. Evol. 41, 165–181 (2006).
38. Springer, M. S., Teeling, E. C., Madsen, O., Stanhope, M. J. & de Jong, W. W.
Integrated fossil and molecular data reconstruct bat echolocation. Proc. Natl Acad.
Sci. USA 98, 6241–6246 (2001).
39. Teeling, E. C. et al. A molecular phylogeny for bats illuminates biogeography and
the fossil record. Science 307, 580–584 (2005).
40. Price, S. A., Bininda-Emonds, O. R. & Gittleman, J. L. A complete phylogeny of the
whales, dolphins and even-toed hoofed mammals (Cetartiodactyla). Biol. Rev.
Camb. Philos. Soc. 80, 445–473 (2005).
41. Agnarsson, I. & May-Collado, L. J. The phylogeny of Cetartiodactyla: the
importance of dense taxon sampling, missing data, and the remarkable promise of
cytochrome b to provide reliable species-level phylogenies. Mol. Phylogenet. Evol.
48, 964–985 (2008).
42. Murphy, W. J. et al. Molecular phylogenetics and the origins of placental
mammals. Nature 409, 614–618 (2001).
43. Schmitz, J., Roos, C. & Zischler, H. Primate phylogeny: molecular evidence from
retroposons. Cytogenet. Genome Res. 108, 26–37 (2005).
44. Matsui, A., Rakotondraparany, F., Munechika, I., Hasegawa, M. & Horai, S.
Molecular phylogeny and evolution of prosimians based on complete sequences
of mitochondrial DNAs. Gene 441, 53–66 (2009).
45. Horner, D. S. et al. Phylogenetic analyses of complete mitochondrial genome
sequences suggest a basal divergence of the enigmatic rodent Anomalurus. BMC
Evol. Biol. 7, 16 (2007).
46. Quérouil, S. et al. Phylogeny and evolution of African shrews (Mammalia:
Soricidae) inferred from 16s rRNA sequences. Mol. Phylogenet. Evol. 20, 185–195
(2001).
47. Douady, C. J. et al. Molecular phylogenetic evidence confirming the Eulipotyphla
concept and in support of hedgehogs as the sister group to shrews. Mol.
Phylogenet. Evol. 25, 200–209 (2002).
48. Hatch, L. T., Dopman, E. B. & Harrison, R. G. Phylogenetic relationships among the
baleen whales based on maternally and paternally inherited characters. Mol.
Phylogenet. Evol. 41, 12–27 (2006).
49. Xiong, Y., Brandley, M. C., Xu, S., Zhou, K. & Yang, G. Seven new dolphin
mitochondrial genomes and a time-calibrated phylogeny of whales. BMC Evol.
Biol. 9, 20 (2009).
Vol 464 | 11 March 2010 | doi:10.1038/nature08818

LETTERS
Sister chromosome pairing maintains heterozygosity
in parthenogenetic lizards
Aracely A. Lutes1,2, William B. Neaves1,3, Diana P. Baumann1, Winfried Wiegraebe1 & Peter Baumann1,2,4

Although bisexual reproduction has proven to be highly successful,
parthenogenetic all-female populations occur frequently in certain
taxa, including the whiptail lizards of the genus Aspidoscelis.
Allozyme analysis revealed a high degree of fixed heterozygosity
in these parthenogenetic species1,2, supporting the view that they
originated from hybridization events between related sexual species.
It has remained unclear how the meiotic program is altered to pro-
duce diploid eggs while maintaining heterozygosity. Here we show
that meiosis commences with twice the number of chromosomes in
parthenogenetic versus sexual species, a mechanism that provides
the basis for generating gametes with unreduced chromosome
content without fundamental deviation from the classic meiotic
program. Our observation of synaptonemal complexes and chias-
mata demonstrate that a typical meiotic program occurs and that
heterozygosity is not maintained by bypassing recombination.
Instead, fluorescent in situ hybridization probes that distinguish
between homologues reveal that bivalents form between sister chro-
mosomes, the genetically identical products of the first of two pre-
meiotic replication cycles. Sister chromosome pairing provides a
mechanism for the maintenance of heterozygosity, which is critical
for offsetting the reduced fitness associated with the lack of genetic
diversity in parthenogenetic species.
True parthenogenesis, characterized by the complete absence of
male contributions, has been described for various species of reptiles,
including whiptail lizards, geckos, blind snakes and rock lizards3.
Whiptail lizards of the genus Aspidoscelis, formerly part of the genus
Cnemidophorus4, are mostly native to the Southwestern United States
and Mexico, and about one-third of the more than 50 species re-
produce by obligate parthenogenesis.
Morphological, karyotypic and biochemical studies provided
strong evidence for hybrid origins of all parthenogenetic Aspidoscelis
species examined1,5,6. While hybridization between individuals from
mechanism to produce mature oocytes that carry the complete com-
plement of somatic chromosomes and maintain heterozygosity12. To
test this hypothesis, we set out to quantify the DNA content in
oocytes of the diploid parthenogenetic species A. tesselata and the
sexually reproducing control Aspidoscelis gularis. Extant A. gularis
and Aspidoscelis tigris are closely related to the individuals that hybri-
dized to generate the founding specimen of A. tesselata1,2,13,14.
We isolated germinal vesicles (GVs), the oocyte nuclei, and visualized
49,6-diamidino-2-phenylindole (DAPI)-stained chromosomes by two-
photon microscopy. Bivalents were readily observed in GVs from A.
gularis and A. tesselata, and their morphology was consistent with the
diplotene stage of prophase I (Fig. 1a, b). Visual inspection of three-
dimensional reconstructions of seven gularis and five tesselata GVs
revealed a larger number of bivalents in tesselata GVs compared to
gularis. Ambiguities in identifying the boundaries of individual chro-
mosomes prevented accurate counting of bivalents at this stage. Instead,
we quantified the volume occupied by chromosomes in each GV as an
indirect measure of DNA content (Fig. 1c; Supplementary Table 1).
a 1 2 3 4
b 1 2 3 4
c d 160
Chromosome volume

(arbitrary units)

distinct species can explain the initially high degree of heterozygosity 3,000
DNA content

120

across the genome, allozyme analysis demonstrated surprising per- 2,000 80

(µm3)

sistence of heterozygosity over many generations in several partheno- 40

1,000
genetic lineages of Aspidoscelis, including Aspidoscelis tesselata and 0
ris

ris

Aspidoscelis neomexicana6. The acceptance of skin grafts between indi-

ui
at

an

0
tig

la

ng
el

ic

1 2 3 4 5 6 7 1 2 3 4 5
gu

ss

ex

sa
A.

te

viduals of a parthenogenetic species7–10 and biochemical studies on A. tesselata

A.

A. gularis
om

ex
A.

A.
ne

several laboratory-reared lineages6 further support genetic uniformity.

A.

The mechanism that underlies clonal reproduction and fixed Figure 1 | Oocytes from parthenogenetic A. tesselata contain twice the
heterozygosity has been the topic of much speculation. Most varia- amount of chromosomal DNA compared to sexual A. gularis. a, DAPI-
tions of the meiotic program that would produce diploid oocytes by stained chromosomes in germinal vesicles (GVs) from A. gularis. Three-
skipping a division or by fusion of a haploid oocyte with a polar body dimensional projections of four GVs are shown; details on size and
quantifications are available as Supplementary Table 1. Scale bars, 10 mm.
cannot account for fixed heterozygosity unless recombination b, GVs from A. tesselata. c, Quantification of chromosome volumes.
between homologues is suppressed. On the basis of the exclusion d, Quantification of DNA content in somatic cells by flow cytometry.
of alternative models and the observation of large numbers of chro- Fluorescence intensities from biological triplicates of ,50,000 cells were
mosomes in two oocytes from Aspidoscelis uniparens11, premeiotic averaged and normalized against samples from A. gularis, which was set at
endoreplication of chromosomes was proposed as the most likely 100 to facilitate comparison.
1
Stowers Institute for Medical Research, Kansas City, Missouri 64110, USA. 2Department of Molecular and Integrative Physiology, University of Kansas Medical Center, Kansas City,
Kansas 66160, USA. 3University of Missouri Kansas City, School of Medicine, Kansas City, Missouri 64110, USA. 4Howard Hughes Medical Institute, Kansas City, Missouri 64110, USA.

283
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
Unlike measurements of fluorescence intensity, this approach is robust
against changes in staining efficiency or laser intensity fluctuations.
A. tesselata chromosomes occupied 2.24 6 0.18 times the volume of
the averaged A. gularis samples. While indicative of a twofold increase
in the DNA content of the prophase oocyte in the parthenogenetic
species, differences in genome size could also account for the increased
chromosome volume.
Although somatic cells from A. gularis and A. tesselata both harbour
46 chromosomes15, a direct comparison of genome sizes was needed
to inform our analysis. Taking advantage of the fact that reptilian
erythrocytes are nucleated, we subjected blood samples to flow cyto-
metry analysis and found that the nuclear DNA content in somatic
cells differed by less than 1% between the two species (Fig. 1d). For
comparison, samples from sexual diploid A. tigris and parthenogenetic
triploid Aspidoscelis exsanguis (A. exsanguis is the product of two con-
secutive hybridization events involving the sexual species Aspidoscelis
inornata, Aspidoscelis burti and A. gularis13) were also analysed, with the
latter showing an approximately 50% increase in DNA content, as
expected (Fig. 1d). Independent confirmation for a doubling in chro-
mosome number was obtained by examining GVs in late prophase. At
this stage, chromosomes are highly condensed and, consistent with a
doubling in chromosome number, we were able to distinguish 46
bivalents in A. tesselata GVs (Supplementary Fig. 1).
Entering meiosis with an 8n chromosome complement would
allow parthenogenetic animals to utilize the two normal meiotic
divisions to generate diploid gametes. However, the long-term main-
tenance of heterozygosity across the genome is only ensured if cross-
overs between homologues are suppressed. Two-photon imaging of
diplotene chromosomes from A. tesselata and another parthenogenetic
species, A. neomexicana, revealed no differences compared to sexual
controls besides the increased DNA content. Notably, bivalents
appeared to be connected by chiasmata in all samples, indicating that
crossing-over is not abandoned (Fig. 2a, b).
To further examine chromosome pairing, thin sections of ovaries
from A. tesselata, A. tigris and A. neomexicana were examined by
electron microscopy. Synaptonemal complexes, characterized by
a c
b
e
d
d e f Sister chromosome pairing
Figure 2 | Visualization of chiasmata and synaptonemal complexes in
parthenogenetic A. tesselata and sexual A. tigris. a, b, Projection of
bivalents from A. tigris (a) and A. tesselata (b) GVs in diplotene. c, Electron
microscopy image of A. tesselata GV in pachytene. Several sections of
synaptonemal complexes are visible. Close-up views of the two boxed areas
are shown in d and e. f, Close-up view of a synaptonemal complex from A.
tigris. Scale bars: 10 mm (a, b); 2 mm (c); 200 nm (d–f).
284
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
well-defined lateral and central elements, were observed in all species
examined, providing further support that a typical meiotic program
is underway (Fig. 2c–f, Supplementary Fig. 2). Based on the presence
of synaptonemal complexes in pachytene and chiasmata in diplotene,
we surmise that meiotic chromosome pairing and recombination are
not bypassed in parthenogenetic Aspidoscelis species.
The premeiotic doubling of chromosomes allows for bivalent
formation to occur either between homologues as in normal meiosis
or between sister chromosomes (Fig. 3). To distinguish between these
possibilities, we sought to identify probes that selectively recognize
one particular chromosome in a pair of homologues. We discovered
that 26 of the 46 A. tigris chromosomes, including all 22 macrochro-
mosomes and 4 microchromosomes, harbour large tracks of internal
telomeric repeats in addition to the signal at chromosome ends
(Fig. 4a). In contrast, staining metaphases of A. inornata chromo-
somes with a telomeric protein-nucleic acid probe only revealed
signal at the chromosome termini (Fig. 4b). Consistent with its
hybrid origin from these two sexual species1,2,5, A. neomexicana chro-
mosomes contained large internal repeats on 13 chromosomes,
allowing us to unambiguously identify 13 chromosomes inherited
from A. tigris in the original F1 hybrid (Fig. 4c). In the context of a
bivalent, hybridization signals on both sides indicate sister chro-
mosome pairing, whereas hybridization on only one side supports
homologue pairing.
To preserve the three-dimensional arrangements of chromosomes
in GVs and to provide better spatial resolution than commonly obtained
in chromosome spreads, we adapted the fluorescent in situ hybridiza-
tion (FISH) procedure to perform hybridization on intact GVs. At each
site where chromosome internal hybridization was detected, a signal
was observed on both sides of the bivalent (Fig. 4d, e). It is important
to note that sister chromatids resulting from the most recent round of
replication appear as one cytologically, as they are closely associated with
each other along their length during this stage of meiosis. The exclusive
presence of paired hybridization signals therefore strongly suggests that
bivalents are composed of sister chromosomes, not homologues. On the
basis of this experiment, we concluded that for the 13 chromosomes for
which the telomeric hybridization probe distinguishes sisters and
homologues in A. neomexicana, sister chromosome pairing is the rule.
Replication Pairing Exchange First division Second division
Canonical meiosis
Premeiotic endoreplication
Homologue pairing
or
Figure 3 | Meiosis in sexual and parthenogenetic Aspidoscelis species.
Top, in normal (canonical) meiosis, a single round of DNA replication is
followed by two consecutive divisions that result in a haploid gamete and
three polar bodies. Homologues are shown in red and blue. Recombination
generates chimaeric chromosomes. Bottom, premeiotic doubling of
chromosomes allows for pairing of homologous or sister chromosomes.
Homologue pairing and recombination result in some loss of heterozygosity
in the mature oocyte. Recombination between pairs of sister chromosomes
maintains heterozygosity at all loci.
NATURE | Vol 464 | 11 March 2010
a b c molecular markers in these studies precludes definitive conclusions,
d e
Figure 4 | Internal telomeric repeats distinguish homologues in A.
neomexicana and demonstrate sister chromosome pairing. a, A
CCCTAA(3) peptide nucleic acid probe (red) identifies chromosome termini
and large internal telomere repeat regions on metaphase spreads of A. tigris
chromosomes prepared from fibroblast cultures. DAPI-stained
chromosomes are shown in blue. b, Chromosome terminal signals, but no
internal telomeric repeats, are detected in A. inornata. c, Chromosomes
inherited from A. tigris but not from A. inornata are identified by internal
telomeric repeats in A. neomexicana. Boxed areas in a–c are magnified in the
bottom left corners of each panel. d, Projection of a subset of images from an
A. neomexicana GV visualized by confocal microscopy. DAPI-stained
chromosomes are shown in white and the telomeric probe in red. e, Close-up
of four representative areas showing paired fluorescence signals. The
differences in resolution stem from differences in projection angles.
Screening of several tri-, tetra- and hexanucleotide repeat probes
identified (CCAAGG)n as an additional marker for at least nine
chromosomes in A. neomexicana that are of A. tigris origin
(Supplementary Fig. 3a–c). When hybridized to diplotene chromo-
somes in acrylamide-embedded GVs, only paired signals were
observed (Supplementary Fig. 3d, e). In summary, two independent
probes enabled us to distinguish sister chromosomes from homolo-
gues, and for over 20 bivalents examined, pairing occurred exclu-
sively between sister chromosomes.
Entering meiosis with twice the usual number of chromosomes
allows parthenogenetic species to produce oocytes carrying the
complete somatic chromosome complement while preserving the
established meiotic program. There are two principal pathways by
which the premeiotic oocytes of a diploid species may acquire eight
rather than four sets of chromosomes. One is the process in which
chromosome duplication occurs without cytokinesis; this has been
termed endomitosis or endoreplication16. Alternatively, 8n germ cells
may arise by fusion of two cells either before or after the final pre-
meiotic doubling of chromosomes. There is ample precedent for
either mode of genome amplification in plants and animals, but
the regulatory mechanisms are largely unclear.
In sexual species, homologous chromosomes form bivalents, and
meiotic recombination promotes genetic diversity while ensuring
orderly segregation of chromosomes during the first meiotic division.
The same mechanism would result in loss of heterozygosity in par-
thenogenetic species, whereas formation of bivalents from genetically
identical sister chromosomes preserves heterozygosity. Interestingly,
this same variation of the meiotic program appears to enable par-
thenogenetic reproduction in widely diverged species. Premeiotic
doubling of chromosomes has been documented in triploid ambysto-
matid salamanders17 as well as a parthenogenetic grasshopper
(Warramaba virgo)18. In both cases, sister chromosome pairing was
suggested on the basis of bivalent morphology. Although the lack of
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
the striking parallels with whiptail lizards strongly indicate that a
common mechanism enables parthenogenetic reproduction in diverse
groups of animals. It seems likely that a relatively simple deviation
from the established program of oogenesis is sufficient to permit par-
thenogenesis. However, loss of heterozygosity, paternal inheritance of
centrosomes, and a requirement for fertilization in triggering comple-
tion of female meiosis are seemingly unconnected obstacles to par-
thenogenetic reproduction. A better understanding of the changes that
permit a small but diverse group of animals to reproduce without
males is clearly needed, and may well shed light on the overwhelming
success of sexuality.
METHODS SUMMARY
Laboratory colonies of Aspidoscelis species were from animals collected in Texas
and New Mexico under a permit from the New Mexico Department of Game and
Fish (permit numbers 3199 and 3395).
GV isolation and quantification of chromosome volume. Ovaries from adult
and sub-adult lizards were isolated, stained with 4’,6-diamidino-2-phenylindole
(DAPI) and imaged using a Zeiss LSM 510 system in two-photon excitation
mode. A non-parametric and unsupervised, automatic threshold selection
method developed by Otsu (see Methods) was used to obtain an unbiased
measurement of chromosome volumes.
Fluorescent in situ hybridization (FISH). Colcemid-treated embryonic fibro-
blasts were harvested, fixed and used to prepare metaphase spreads. FISH was
performed on dried coverslips using AlexaFluor-labelled peptide-nucleic acid
(PNA) and locked-nucleic acid (LNA) probes. Samples were imaged on a fluor-
escence microscope and images were analysed with AxioVision software. FISH on
meiotic chromosomes was performed after embedding GVs in an acrylamide gel.
Transmission electron microscopy. Ultrathin sections of epoxy-embedded
ovaries were collected on copper grids, stained with 2% uranyl acetate in 50%
methanol for 10 min, followed by 1% lead citrate for 7 min. Sections were photo-
graphed using a FEI transmission electron microscope at 80 kV.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 16 November 2009; accepted 6 January 2010.
Published online 21 February 2010.
1. Neaves, W. B. Adenosine deaminase phenotypes among sexual and
parthenogenetic lizards in the genus Cnemidophorus (Teiidae). J. Exp. Zool. 171,
175–183 (1969).
2. Neaves, W. B. & Gerald, P. S. Lactate dehydrogenase isozymes in parthenogenetic
teiid lizards (Cnemidophorus). Science 160, 1004–1005 (1968).
3. Vrijenhoek, R. C., Dawley, R. M., Cole, C. J. & Bogart, J. P. in Evolution and Ecology of
Unisexual Vertebrates Vol. 466 (eds Dawley, R. M. & Bogart, J. P.) 19–23 (New
York State Museum Bulletin, 1989).
4. Reeder, T. W., Cole, C. J. & Dessauer, H. C. Phylogenetic relationships of whiptail
lizards of the genus Cnemidophorus (Squamata: Teiidae): a test of monophyly,
reevaluation of karyotypic evolution, and review of hybrid origins. Am. Mus. Novit.
3365, 1–61 (2002).
5. Lowe, C. H. & Wright, J. W. Evolution of parthenogenetic species of
Cnemidophorus (whiptail lizards) in western North America. J. Ariz. Acad. Sci. 4,
81–87 (1966).
6. Dessauer, H. C. & Cole, C. J. Clonal inheritance in parthenogenetic whiptail lizards:
biochemical evidence. J. Hered. 77, 8–12 (1986).
7. Cordes, J. E. & Walker, J. M. Skin histocompatibility between syntopic pattern
classes C and D of parthenogenetic Cnemidophorus tesselatus in New Mexico. J.
Herpetol. 37, 185–188 (2003).
8. Cordes, J. E. & Walker, J. M. Evolutionary and systematic implications of skin
histocompatibility among parthenogenetic teiid lizards: three color pattern
classes of Aspidoscelis dixoni and one of Aspidoscelis tesselata. Copeia14–26
(2006).
9. Maslin, T. P. Skin grafting in the bisexual teiid lizard Cnemidophorus sexlineatus and
in the unisexual C. tesselatus. J. Exp. Zool. 166, 137–149 (1967).
10. Cuellar, O. & Smart, C. Analysis of histoincompatibility in a natural population of
the bisexual whiptail lizard Cnemidophorus tigris. Transplantation 24, 127–133
(1977).
11. Cuellar, O. Reproduction and the mechanism of meiotic restitution in the
parthenogenetic lizard Cnemidophorus uniparens. J. Morphol. 133, 139–165 (1971).
12. Neaves, W. B. Tetraploidy in a hybrid lizard of the genus Cnemidophorus (Teiidae).
Brev. Mus. Comp. Zool. 381, 1–25 (1971).
13. Dessauer, H. C. & Cole, C. J. in Evolution and Ecology of Unisexual Vertebrates Vol.
466 (eds Dawley, R. M. & Bogard, J. P.) 49–71 (New York State Museum Bulletin,
1989).
285
LETTERS
14. Parker, E. D. Jr & Selander, R. K. The organization of genetic diversity in the
parthenogenetic lizard Cnemidophorus tesselatus. Genetics 84, 791–805 (1976).
15. Wright, J. W. & Lowe, C. H. Evolution of the alloploid parthenospecies
Cnemidophorus tesselatus (Say). Mamm. Chromosome Newsl. 8, 95–96 (1967).
16. Edgar, B. A. & Orr-Weaver, T. L. Endoreplication cell cycles: more for less. Cell 105,
297–306 (2001).
17. Macgregor, H. C. & Uzzell, T. M. Jr. Gynogenesis in salamanders related to
Ambystoma jeffersonianum. Science 143, 1043–1045 (1964).
18. White, M. J. D., Cheney, J. & Key, K. H. L. A parthenogenetic species of
grasshopper with complex structural heterozygosity (Orthoptera: Acridoidea).
Aust. J. Zool. 11, 1–19 (1963).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank the staff of the Reptile Facility for their dedication to
animal husbandry; C. Painter and his colleagues at the New Mexico Department of
286
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
Game and Fish for scientific collection permits; F. Li for electron microscopy; the
Stowers Institute Microscopy, Cytometry and Molecular Biology Facilities for
support; and S. Hawley, J. Cole, C. Townsend and members of the Baumann and
Blanchette laboratories for discussions and comments. We also thank S. Hawley
for critical reading of the manuscript. This work was funded by the Stowers
Institute for Medical Research, and the Pew Scholars Program in the Biological
Sciences sponsored by the Pew Charitable Trusts. P.B. is an Early Career Scientist
with the Howard Hughes Medical Institute.
Author Contributions A.A.L., W.B.N., D.P.B. and P.B. contributed extensively to the
work presented in this paper. W.W. performed, and provided training in,
microscopy and image data quantification.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to P.B.
(peb@stowers.org).
doi:10.1038/nature08818
METHODS
Animals. Laboratory colonies of A. tigris, A. inornata, A. neomexicana, A. exsanguis
and A. tesselata were established from animals collected in Socorro, Sierra and
Otero Counties, New Mexico, under a permit from the New Mexico Department
of Game and Fish (permit numbers 3199 and 3395). A. gularis were collected in
Dickens County, Texas. Animals were propagated in our Reptile and Aquatics
Facility under conditions similar to a previously published description of captive
lizard husbandry19.
GV isolation and quantification of chromosome volume. Ovaries from adult
and sub-adult lizards were placed in PBS and GVs were isolated using jeweller’s
forceps. GVs were transferred with a pipette to glass-bottom dishes (MatTek)
containing 40 ng ml21 4’,6-diamidino-2-phenylindole (DAPI) in PBS, and
allowed to incubate at 4 uC overnight. The following day the GVs were imaged
using a LSM 510 META (Carl Zeiss Jena) system in two-photon excitation
mode20 equipped with a C-Apochromat 403, NA 1.2 water immersion lens
(Carl Zeiss Jena) optimized for ultraviolet and infrared transmission. Two-
photon excitation at 735 nm was used to avoid out-of-focus bleaching and all
fluorescence emission below 650 nm was collected. Images were cropped in
Adobe Photoshop to digitally remove the nuclear envelope.
To obtain an unbiased measurement of chromosome volumes, we used the
non-parametric and unsupervised, automatic threshold selection developed by
Otsu21. In this method, the grey-level histogram of an image suffices to find a grey
level threshold, which yields the optimal separation of the chromosomes from
the background without other a priori input. The Otsu method treats the nor-
malized histogram as a probability distribution of possible pixel values. Pixels are
dichotomized into two classes C0 and C1 (background and objects, in our case
chromosomes) by a threshold at level k. C0 denotes the class of pixels with values
[0, …, k], and C1 denotes the class of pixels with values [k 1 1, …, L], 0 being the
smallest and L the largest pixel value in the image. The goal is to determine the
threshold value k corresponding to the best class separation. Otsu showed that k
can be found by maximizing the between-class variance s2B ðk Þ for all k between 0
and L. s2B ðk Þ is defined as:
s2B ðk Þ~P0 (k)ðm0 (k){mG Þ2 zP1 (k)ðm1 (k){mG Þ2
Here P0,1 (k) is the probability that a pixel k is assigned to class C0 or C1, respec-
tively, m0,1 (k) represents the mean intensity of all the pixels assigned to class C0
or C1 (ref. 22), and mG is the average intensity of the whole image.
To calculate the threshold, we used the implementation of Otsu’s method by
C. Mei, A. Joshua and T. Collins as a plug-in for the open-source image proces-
sing package ImageJ23. The volume was determined with the commercial 3D
image processing software IMARIS 6.3 (Bitplane). This software bases its volume
calculation on an iso-surface rendered at an intensity value k, using the threshold
determined as described above. This protocol ensures an objective, reproducible,
and unbiased comparison of chromosome volumes measured for independent
samples.
Embryonic fibroblast cell culture. Cell cultures were modified from ref. 24.
Briefly, Aspidoscelis eggs were incubated at 29 uC for 30–40 days, sterilized in 90%
ethanol, 120 mM potassium iodide and 39 mM iodine, and opened under sterile
conditions in a laminar flow hood. Embryos were immediately decapitated and
cut into 1 cm or smaller pieces, rinsed with cold PBS and incubated with trypsin-
EDTA solution (T4049, Sigma) on a stir plate for 5 min at room temperature.
The supernatant was discarded and replaced with fresh PBS/trypsin, and incuba-
tion was continued for 15 min. The supernatant was decanted through sterile
cheesecloth into a 50-ml Falcon tube containing 2 ml ice-cold M199 cell culture
media (Sigma) supplemented with 20% fetal bovine serum, 50 mg ml21 gentamycin
(Sigma), glutamax (Invitrogen), MEM non-essential amino acids (Invitrogen),
MEM vitamin solution (Invitrogen), 56 U ml21 nystatin (Sigma), 100 U ml21
penicillin and 100 mg ml21 streptomycin (Sigma). Cell suspensions were kept on
ice while the remaining tissue was trypsin-treated for another 15 min and the
supernatants were combined. Cells were then pelleted, washed in M199 media plus
supplements and finally resuspended in 2–6 ml M199 media plus supplements and
seeded in 6-well dishes (Falcon, 353046). Cells were cultured at 30 uC, 5% O2 and
2% CO2, and split once they had reached 85–100% confluency.
Fluorescent in situ hybridization (FISH). Embryonic fibroblasts were treated
with 0.5 mg ml21 Karyomax colcemid (Invitrogen) at 50–70% confluency and
incubated for 3 h at 30 uC, 5% O2 and 2% CO2. The cells were harvested by
trypsin treatment and subjected to hypotonic swelling in 0.075 M KCl at 37 uC
for 10 min. The cells were then pelleted, washed twice and finally resuspended in
methanol:acetic acid fixative (3:1). Coverslips were cleaned with a 1:1 ethanol:ethyl
©2010 Macmillan Publishers Limited. All rights reserved
ether solution and cells were dropped onto the coverslips and immediately washed
liberally with fixative. Cover slips were then incubated on a heat block at 75 uC for
1 min. FISH was performed on dried coverslips as previously described25 using
either an AlexaFluor 543-labelled peptide-nucleic acid probe composed of
59-(CCCTAA)3-39or an AlexaFluor 488-labelled locked-nucleic acid (LNA) probe
composed of 59-(CCAAGG)2CC-39. Washes were modified for the LNA probe as
follows: three 2-min washes in 2X SSC, 0.05% Tween 20 at 37 uC, two 6-min washes
with 0.1X SSC at 55 uC, three 2-min washes with 2X SSC, 0.05% Tween 20 at room
temperature, followed by two 2-min washes with PBS. Samples were imaged on a
fluorescence microscope with a 1003, 1.4 NA Plan-Apochromat objective and
images were analysed with AxioVision software (Carl Zeiss Jena).
Ovaries were isolated from A. neomexicana and placed in Buffer A (15 mM
PIPES, pH 6.8, 20 mM NaCl, 60 mM KCl, 0.5 mM EGTA, 2 mM EDTA, 0.5 mM
spermidine, 1 mM DTT). GVs were then embedded in an acrylamide mixture
consisting of Buffer A, 5% acrylamide, 1 mM DTT, 15 mM sodium sulphite, and
11.5 mM sodium persulphate. Embedding chambers consisted of 22 3 55 mm
coverslips that had been siliconed to stainless steel washers. After the GV was
added to the washer, half of a 22 3 22 mm coverslip was dropped onto the
washer to seal the top and promote polymerization of the acrylamide. After
30–60 min, the coverslip was carefully lifted and the gel was washed 4 times
for 20 min with Buffer A on a shaking platform. Prehybridization was carried
out in 50% deionized formamide, 2X SSC for 80 min with three changes of the
prehybridization solution. Hybridization mixtures were as described for somatic
cells, and samples were incubated in a sealed chamber to minimize evaporation.
Slides/washers were placed on a PCR machine with the following program: non-
heated lid, 1 h at room temp, 30 min at 40 uC, 6 min at 94 uC and overnight at
37 uC. The following day, samples were washed four times in PBS with 0.1%
Tween 20, and three times in PBS and stained with DAPI (40 ng ml21) for at least
1 h at room temperature. Samples were imaged on an inverted confocal micro-
scope with a C-apochromat 403/1.20 W objective (Carl Zeiss Jena). The Alexa
Fluor 543-labelled peptide-nucleic acid probe (59-(CCCTAA)3-39) was excited at
561 nm and fluorescence emission was collected above 575 nm. The Alexa Fluor
488-labelled locked-nucleic acid probe (59-(CCAAGG)2CC-39) was excited at
488 nm and fluorescence emission from 505 to 550 nm was collected. DAPI-
stained chromosomes were visualized by excitation at 405 nm and collection
of fluorescence emission between 420 and 480 nm. To avoid emission cross-talk,
we switched the excitation between the different laser lines and collected data in
the appropriate detection channel only (multi-tracking).
Transmission electron microscopy. Ovaries were isolated from A. tigris and
A. neomexicana in PBS then transferred to a tube containing 70% hexane, 0.75%
paraformaldehyde (Electron Microscopy Sciences), and 0.125% Nonidet (US
Biological) in PBS. The tube was gently inverted 10 times then placed in a 20 uC
waterbath for 2.5 h, gently inverting the tube five times every 20 min. The ovaries
were washed three times with 0.2% Tween 20 in PBS and three times with PBS
(20 min each), followed by overnight fixation in 2.5% glutaraldehyde (Electron
Microscopy Sciences) in PBS at 4 uC. Fixed tissues were washed three times in PBS
and water, then post-fixed in aqueous 1% OsO4, 1% K3Fe(CN)6 for 10 min at
room temperature. Following 3 PBS washes, the tissue was dehydrated through a
graded series of 30–100% ethanol 100% propylene oxide and then infiltrated in
1:1 mixture of propylene oxide:Polybed 812 epoxy resin (Polysciences) for 1 h.
After two to three changes of 100% resin over 24 h, tissues were embedded in
moulds, cured at 37 uC overnight, followed by additional hardening at 65 uC for
two more days. Ultrathin (60 nm) sections were collected on copper grids, stained
with 2% uranyl acetate in 50% methanol for 10 min, followed by 1% lead citrate
for 7 min. Sections were photographed using a FEI transmission electron micro-
scope at 80 kV.
19. Townsend, C. R. Establishment and maintenance of colonies of parthenogenetic
whiptail lizards: Cnemidophorus spp. Int. Zoo. Yb. 19, 80–86 (1979).
20. Denk, W., Strickler, J. H. & Webb, W. W. Two-photon laser scanning fluorescence
microscopy. Science 248, 73–76 (1990).
21. Otsu, N. Threshold selection method from gray-level-histograms. IEEE Trans. Syst.
Man Cybern. SMC-9, 62–66 (1979).
22. Gonzalez, R. C. & Woods, R. E. Digital Image Processing 3rd edn, (Pearson/Prentice
Hall, 2008).
23. ImageJ (US National Institutes of Health, Bethesda, 1997–2009).
24. Moore, M. K., Work, T. M., Balazs, G. H. & Docherty, D. E. Preparation,
cryopreservation, and growth of cells prepared from the green turtle (Chelonia
mydas). Methods Cell Sci. 19, 161–168 (1997).
25. Sarthy, J., Bae, N. S., Scrafford, J. & Baumann, P. Human RAP1 inhibits non-
homologous end joining at telomeres. EMBO J. 28, 3390–3399 (2009).
Vol 464 | 11 March 2010 | doi:10.1038/nature08799

LETTERS
Systematic genetic analysis of muscle morphogenesis
and function in Drosophila
Frank Schnorrer1,2, Cornelia Schönbauer1, Christoph C. H. Langer1, Georg Dietzl2, Maria Novatchkova2,
Katharina Schernhuber2, Michaela Fellner2, Anna Azaryan2, Martin Radolf2, Alexander Stark2, Krystyna Keleman2
& Barry J. Dickson2

Systematic genetic approaches have provided deep insight into the screen (excluding for technical reasons those lines that were not
molecular and cellular mechanisms that operate in simple unicel- viable as homozygotes). For each of these lines, muscle organization
lular organisms. For multicellular organisms, however, the pleio- was visualized in live animals using a green fluorescent protein (GFP)
tropy of gene function has largely restricted such approaches to the marker that specifically labels the sarcomeric Z-line7. Defects were
study of early embryogenesis. With the availability of genome-wide readily observed for 190 genes, either in overall muscle morphology
transgenic RNA interference (RNAi) libraries in Drosophila1,2, it is (Fig. 2a and Supplementary Table 6) or sarcomeric organization
now possible to perform a systematic genetic dissection of any cell (Fig. 2b and Supplementary Table 7).
or tissue type at any stage of the lifespan. Here we apply these We distinguished three classes of defect in muscle morphology
methods to define the genetic basis for formation and function of (Fig. 2c–k): ‘split myofibril’ (53 genes), ‘missing muscles’ (8 genes)
the Drosophila muscle. We identify a role in muscle for 2,785 genes, and ‘rounded muscles’ (13 genes), in which muscles are either split
many of which we assign to specific functions in the organization of into thinner myofibrils, are missing or have the rounded appearance
muscles, myofibrils or sarcomeres. Many of these genes are phylo- characteristic of muscles that undergo normal morphogenesis but fail
genetically conserved, including genes implicated in mammalian to form stable attachments. The ‘split myofibril’ class includes several
sarcomere organization and human muscle diseases. signalling molecules, such as the FGF receptor heartless, Anxb11, the
Muscle biology is an attractive target for analysis by genome-wide actin regulator a-actinin and several RNA-binding proteins (for
transgenic RNAi. The basic cell and developmental biology of muscles example, CG5800; Fig. 2g). In these mutants, the muscles still generally
is largely conserved from insects to mammals3–6, and their multinuc- attach to the appropriate tendon cells. The ‘missing muscles’ class
lear architecture renders them inaccessible to conventional genetic includes splicing factors and His2A (Fig. 2f). The ‘rounded muscles’
mosaic strategies. To systematically disrupt gene functions exclusively class includes known muscle attachment factors such as integrins, ILK
in the muscles of intact animals, we crossed the muscle-specific Mef2-
GAL4 driver to each of the UAS-IR transgenic RNAi lines in our
genome-wide library1. Progeny from these crosses were assayed for a Phenotypic distribution
viability, posture, locomotion and flight. We screened a total of 17,759
RNAi lines, representing 10,461 distinct genes. A total of 3,909 lines
Lethal (1,963; 18.8%)
and 2,785 genes were scored as defective in one or more of these assays Adult lethal day 7 (69; 0.7%)
(Fig. 1, Supplementary Fig. 1 and Supplementary Tables 1 and 2). Flightless (316; 3.0%)
This screen identified 73 of 77 positive control genes (Supplemen- Locomotion (78; 0.7%)
tary Table 3), suggesting a false-negative rate of just 5%. To assess the Wing posture (13; 0.1%)
false-positive rate, we compiled a negative control list of 79 genes Weak flyer (276; 2.6%)
shown by classical genetic studies to have no function in muscles. Of Semi-lethal (70; 0.7%)
the 121 RNAi lines available for these genes, only one scored positive Wild type (7,676; 73.4%)
in our assays (Supplementary Table 4). We thus estimate that our n = 10,461
screen has a false-positive rate of 1.3% of genes (a false-discovery rate
of 5%). b Lethality stages
We selected 1,004 of the positive genes at random for validation
using a second, independent hairpin construct. These lines were Embryonic lethal (31; 1.6%)
obtained from a second RNAi library currently under construction Early larval lethal (213; 10.9%)
using the phiC31 site-specific integrase system (K.K. and B.J.D., Late larval lethal (343; 17.5%)
unpublished observations). A total of 874 genes (87.1%) were again Early pupal lethal (488; 24.9%)
positive (Supplementary Table 5), consistent with the estimated 5% Late pupal lethal (613; 31.2%)
false-discovery rate in our primary screen and a false-negative rate Pharate / adult lethal (211; 10.7%)
comparable to that observed with the first library (,5%). Undefined (64; 3.3%)
We began the detailed analysis of these muscle phenotypes by
examining the morphology of the larval body wall muscles n = 1,963
(Supplementary Fig. 1). Reasoning that defects in these muscles are Figure 1 | Phenotypic classification of primary screen. a, Distribution of
most likely to result in embryonic or larval lethality, we focused on genes into phenotypic classes in the primary muscle screen. b, Distribution
the 436 genes that fell into these phenotypic classes in the primary of the lethal stages for all essential genes.
1
Max-Planck-Institute of Biochemistry, Am Klopferspitz 18, 82152 Martinsried, Germany. 2Research Institute of Molecular Pathology (IMP), Dr. Bohr-Gasse 7, A-1030 Vienna, Austria.
287
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS NATURE | Vol 464 | 11 March 2010

a Muscle morphology b Sarcomere morphology

Fading Z (94; 21.6%)

Split (53; 12.2%) Spotty Z (50; 11.5%)
Missing (8; 1.8%) Clumpy Z (18; 4.1%)
Rounded (13; 3.0%) Undefined (8; 1.8%)
Wild type (362; 83.0%) Wild type (266; 61.0%)

n = 436 n = 436
c Wild type d Rounded

+ UAS-rhea-IR
e Wild type f Missing g Split

+ UAS-His2A:CG33865-IR UAS-CG5800-IR
h i j k

Wild type Rounded Missing Split

l Wild type m Spotty Z n Wild type o Spotty Z

+ UAS-bt-IR + UAS-how-IR
p Fading Z q Clumpy Z r Clumpy Z

UAS-Actn-IR UAS-flare-IR UAS-Bsg-IR

s t u v

Wild type Spotty Z Fading Z Clumpy Z

Figure 2 | Larval muscle screen. a, b, Muscle (a) and sarcomere
(b) morphology classes of the larval body muscles. c–k, Wild-type and
mutant phenotypes: normal muscles in wild-type embryos (c, h) compared
with rounded muscles in UAS-rhea-IR (TF40400) (d, i), wild-type L3 larval
muscles (e), missing muscles in UAS-CG33865-IR (TF39116) (f, j) and split
muscles in UAS-CG5800 (TF27519) (g, k). l–r, Sarcomere morphology
and rhea, the fly Talin orthologue (Fig. 2d), as well as the fly Parvin
orthologue.
We also defined three classes of defect in sarcomeric organization
(Fig. 2l–v): ‘fading Z’ (94 genes), ‘spotty Z’ (50 genes) and ‘clumpy Z’
(18 genes), in which the Z-lines are either reduced, discontinuous or
appear as large irregular aggregates, respectively. The ‘fading Z’ class
is exemplified by a-actinin (Actn; Fig. 2p), and the ‘spotty Z’ class by
how (Fig. 2o), which encodes an RNA-binding protein required for
muscle development8, and bent (Fig. 2m), which encodes a titin- or
projectin-like protein9. Most of the known sarcomeric components
(bt, sls, Mhc, actin, Tm2, TpnC47D, and TpnC73F) fell into the spotty
288
©2010 Macmillan Publishers Limited. All rights reserved
phenotypes: wild-type embryonic sarcomeres (l, s) compared with UAS-bt-
IR (TF46253) displaying spotty Z sarcomeres (m, t), and wild-type L3
sarcomeres (n) compared with spotty Z in UAS-how-IR (TF13756)
(o), fading Z in UAS-Actn-IR (TF7760) (p, u), and clumpy Z in UAS-flare
(TF22851) (q, v) and UAS-Bsg-IR (TF43307) (r). Scale bars, 100 mm (c, d) and
50 mm (e–g, i–r).
Z class. The ‘clumpy Z’ class includes flare (CG10724), the recently
identified fly orthologue of actin interacting protein 1 (AIP1), a
regulator of F-actin disassembly10 and Basigin (Bsg), a muscle trans-
membrane protein required for formation of the neuromuscular
junction11 (Fig. 2q, r). We confirmed efficient knockdown of several
of these proteins by antibody staining (Supplementary Fig. 2).
We performed a similar analysis of muscle morphology in adults,
using phalloidin staining to visualize the indirect flight muscles
(IFMs) for all RNAi lines that scored as flightless in the primary
screen (Supplementary Fig. 1 and Fig. 1a). The large regular structure
of the IFMs and their critical role in flight make them ideal models for
NATURE | Vol 464 | 11 March 2010 LETTERS

studying muscle structure and function12,13. Moreover, unlike the
larval body wall muscles, IFMs resemble vertebrate muscles in their
construction from multiple fibres, each composed of many myofi-
brils14. Defects in IFMs were observed in 196 of the 328 flightless
genes tested. We assigned each of these genes to one or more of nine
distinct phenotypic categories, based on defects in overall muscle
morphology (Fig. 3a and Supplementary Table 8), myofibril mor-
phology (Fig. 3b and Supplementary Table 9) or sarcomeric organi-
zation (Fig. 3c and Supplementary Table 10).
The two classes of defect in overall IFM morphology were ‘missing
IFMs’ (55 genes) and ‘irregular IFMs’ (12 genes). The ‘missing IFM’
class includes parkin (park) and flightin (fln) (Fig. 3e), both previously
associated with IFM degeneration15,16, as well as several transcription
factors that may contribute to the specification of individual muscles
(Supplementary Table 8). The ‘irregular IFM’ class includes flightless I
(fliI; Fig. 3f, g) and MICAL, which serve as positive controls17,18, as well
as CG8578, the mouse orthologue of which (LRRFIP2) interacts with
the leucine-rich repeats of mouse FliI19.
We defined four classes of myofibril defect: ‘degenerate’ (39
genes), ‘irregular’ (7 genes), ‘frayed’ (97 genes) and ‘trapezoid’ (29
genes). Most genes in the ‘degenerate’ myofibril class were also clas-
sified as ‘missing IFMs’. Distinct myofibrils are difficult to discern in
these lines, as is observed upon knockdown of the band 4.1 septate
junction protein encoded by coracle20 (Fig. 3l). The ‘irregular’ myofi-
bril class is characterized by misoriented and disorganized myofibrils,
as seen with fliI (Fig. 3j and Supplementary Fig. 3) or its interactor
CG8578 (Fig. 3o). FliI protein localizes to M- as well as Z-lines, both
of which are severely disrupted upon knockdown of fliI or CG8578
(Fig. 3o and Supplementary Fig. 3a, b), and in fliI mutants (Fig. 3n).
In the ‘frayed’ class, myofibrils are unusually thin and often frayed at
the edges, as for example upon knockdown of the actin regulator
coronin (Fig. 3i), and the M- and Z-lines are also thinner and appear
bent at the frayed edges of the myofibril (Supplementary Fig. 4d). The
‘trapezoid’ class is characterized by myofibrils with a zigzag-like
structure, usually with thick Z-lines and thinner actin filaments
towards the M-lines. One gene in this class is the adducin homologue
hts (Fig. 3k and Supplementary Fig. 4f, g). For 49 genes, mostly of the
‘trapezoid’ and ‘frayed’ classes, we also observed large actin aggre-
gates (Fig. 3m and Supplementary Table 9) that are reminiscent of so-
called zebra bodies21, possibly collapsed Z-lines, that are commonly
observed in human nemaline myopathies22.
Three further phenotypic classes were based on sarcomere organi-
zation: ‘no sarcomere’ (39 genes), ‘no M’ (12 genes) and ‘fuzzy Z’ (99
genes). The ‘no sarcomere’ set generally coincides with the degenerate
myofibril and missing IFM classes (Fig. 3s, w and Supplementary
Fig. 3e). The ‘no M’ class is characterized by sarcomeres with an
apparently normal Z-line, but no discernable M-line in phalloidin
stainings (Fig. 3q, u). This is seen upon knockdown of the obscurin
homologue unc-89 (Fig. 3q), and in unc-89 mutants (B. Bullard,
personal communication). Staining unc-89 knockdown IFMs with
the M-line marker anti-Mhc revealed that the M-line is indeed pre-
sent, but is significantly broadened and invaded by thin actin filaments
(Supplementary Fig. 5a, b). Knockdown of the potassium channel eag
leads to a more severe M-line phenotype with complete absence of

a Muscle morphology b Myofibril morphology c Sarcomere morphology

Degenerate (39; 12.0%)

No sarcomere (39; 12.0%)
Missing (55; 16.8%) Irregular (7; 2.1%)
No M (12; 3.7%)
Irregular (12; 3.7%) Trapezoid (29; 8.9%)
Fuzzy Z (99; 30.5%)
Wild type (261; 79.6%) Frayed (97; 29.9%)
Wild type (175; 53.9%)
Wild type (153; 47.1%)

n = 328 n = 325 n = 325

d Wild type e Missing IFM f Irregular IFM g Irregular IFM

UAS-CG1623-IR UAS-fln-IR UAS-fliI-IR fliI[3]/fliI[14]

h Wild type i Frayed j Trapezoid k Irregular

+ UAS-coro-IR UAS-hts-IR UAS-fliI-IR

l Degenerate IFM m Actin blob n Irregular o Irregular

UAS-cora-IR UAS-CG14260-IR fliI[3]/fliI[14] UAS-CG8578-IR

p Wild type q No M–line r Fuzzy Z s No sarcomeres

+ UAS-unc89-IR UAS-hts-IR UAS-CG13366-IR

t Z Z Z u Z Z Z v Z Z Z w

M M M M
Wild type No M–line Fuzzy Z No sarcomeres

Figure 3 | Adult muscle screen. a–c, Distribution of muscle (a), myofibril mutants (n) and UAS-CG8578-IR (TF35968) (o), degenerated IFMs in UAS-
(b) and sarcomere (c) morphology of the adult IFMs. d–g, Normal IFMs in cora-IR (TF9787) (l) and actin blobs in UAS-CG14260 (TF17452)
wild type (d), missing IFMs in UAS-fln-IR (TF46153) (e), irregular IFMs in (m). p–w, Normal sarcomere morphology in wild type (p), no visible M-line
UAS-fliI-IR (TF39528) (f) and fliI mutant (g). h–o, Normal myofibril in UAS-unc89-IR (TF29412) (q), fuzzy Z in UAS-hts-IR (TF29101) (r) and
morphology in wild type (h), frayed in UAS-coro-IR (TF44671) (i), trapezoid no sarcomeres in UAS-CG13366 (TF29606) (s), and the respective
in UAS-hts-IR (29101) (j), irregular in UAS-fliI-IR (TF39528) (k), fliI schematics (t–w). Scale bars, 100 mm (d–g), 10 mm (h–o), 5 mm (p–s).
289
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS NATURE | Vol 464 | 11 March 2010

a Protein Transcription, Metabolism,

G-protein coupled receptor protein signaling p.

Muscle synthesis splicing mitochondria

Monovalent inorganic cation transporter a.

DNA-directed RNA polymerase activity

Mitochondrial electron transport chain

Proteasome complex (sensu Eukarya)

Structural constituent of cytoskeleton

Transmembrane receptor activity

Transcription, DNA dependent

Transcription factor activity

Exosome (RNase complex)

Oxidative phosphorylation
Mitochondrial ribosome
Arp2/3 protein complex

Spliceosome complex
Translational initiation
Muscle development
Cell-matrix adhesion

Nucleic acid binding

Protein biosynthesis

Cell communication
Sensory perception
Muscle attachment

Muscle contraction
Actin cytoskeleton

Defence response
Lipid metabolism
Integrin complex

ATPase activity
Mitochondrion
RNA binding

Extracellular
Metabolism
Intracellular
Sarcomere

Translation

Ribosome
Myofibril

Nucleus
Mef2 wild type
Mef2 all positive
Mef2 all lethals
Mef2 embryo
Mef2 early larval
Mef2 late larval
Mef2 early pupal
Mef2 late pupal
Mef2 adult
Mef2 day7
Mef2 flightless
Larval muscle wild type
Larval muscle missing
Larval muscle split
Larval muscle rounded
Larval sarcomere wild type
Larval sarcomere clumpy Z
Larval sarcomere fading Z
Larval sarcomere spotty Z

–10

10
–9
–8
–7
–6
–5
–4
–3
–2
–1
0
1
2
3
4
5
6
7
8
9
Legend

Immunology

Metabolic
b c d

Skeletal
Muscle
Mef2 positives

Neuro
M. m.

A. m.

Eyes
C. e.

A. g.
H. s.
S. c.

Skin
A. t.

RNAi Muscle
phenotype expression Mef2 wildtype Mef2 wild type
n = 2,785 255 n = 504 Mef2 all positives Mef2 all positives
2,482 193
30 Mef2 all lethals Mef2 all lethals
18 26 Mef2 flightless Mef2 flightless
Larval sarcomere clumpy Z Larval sarcomere clumpy Z
33
Larval sarcomere fading Z Larval sarcomere fading Z
Larval sarcomere spotty Z Larval sarcomere spotty Z
Mef2
target
n = 107
–10 0 10 –5 0 5

Figure 4 | Bioinformatic analyses. a, GO-term enrichment in the various
phenotypic classes. b, Overlap of genes expressed in muscles, predicted Mef2
targets and positives in the Mef2 RNAi screen. Only genes tested in the RNAi
screen are included (n 5 10,460 for Mef2), and only Mef2 high-confidence
targets for which expression data were available are included (n 5 107).
c, d, Conservation across species (c) and link to clinical diseases (d) in Mef2
lethal as well as the sarcomere phenotypic sets. A. t., Arabidopsis thaliana; S.
c., Saccharomyces cerevisiae; M. m., Mus musculus; H. s., Homo sapiens; C. e.,
Caenorhabditis elegans; A. m., Apis mellifera; A. g., Anopheles gambiae.

Mhc at the M-line (Supplementary Fig. 5c). This may explain why eag
mutants display severe flight defects23. Surprisingly, we find upon
knockdown of the Par domain transcription factor pdp1 an overlap
of Z- and M-lines (Supplementary Fig. 5d), suggesting that mesoder-
mal pdp1 (ref. 24) regulates correct sarcomere assembly. The ‘fuzzy Z’
class is characterized by a broadening of the Z-line (Fig. 3r, v), and
generally coincides with the frayed and trapezoidal classes of myofibril
defect. We analysed RNA levels in isolated flight muscles for selected
genes of the ‘irregular’ myofibril class and the ‘no M’ class. This ana-
lysis confirmed a significant knockdown of the target messenger RNA.
In contrast, mRNA levels of predicted off-target genes (defined as
having at least one perfect 16 nucleotide match to the hairpin) were
not altered any more than mRNA levels of other randomly selected
genes (Supplementary Fig. 6).
We examined the gene ontology (GO) database (www.geneontology.
org) to assess the representation of various gene classes in each of our
phenotypic categories (Fig. 4a). GO terms related to muscles are all
significantly enriched in the set of Mef2-positive genes, particularly in
the embryonic lethals. We also note an enrichment of the GO terms
‘muscle attachment’ and ‘cell-matrix adhesion’ in the ‘rounded’ pheno-
type class, consistent with the failure to form a stable muscle attachment
in these lines. Another noticeable enrichment is the GO term ‘muscle
contraction’ within the ‘spotty Z’ class of sarcomeric defects, suggesting
that the actin defects in these lines correlate with impaired contractility.
290
©2010 Macmillan Publishers Limited. All rights reserved
Previous studies have sought to identify muscle genes systematically
either by expression profiling (http://www.fruitfly.org/cgi-bin/ex/insi-
tu.pl) or chromatin immunoprecipitation of Mef2-binding sites25, both
of which have been focused on embryos. More than half of the muscle-
expressed genes (285 of 504, P , 10247, Fig. 4b) and almost half of the
Mef2 targets (48 of 107, P , 1024, Fig. 4b) were functionally validated in
our screen. A total of 30 genes are positive in all three data sets, repre-
senting a set of Mef2 target genes with confirmed muscle expression and
function (Supplementary Table 11). Almost half of these had no func-
tional assignment before this study. Similar large-scale gene expression
and Mef2 target data are not yet available for later developmental stages,
but we note that our RNAi screen has assigned functions to 10 of 12
genes known to be differentially expressed in adult muscle precursors26
(Supplementary Table 11), and to 10 of 14 genes predicted to function in
terminally differentiated muscle27 (Supplementary Table 12).
Finally, we found that the genes with a muscle RNAi phenotype
are, on average, much better conserved than those with no phenotype
(Fig. 4c). Moreover, examination of the OMIM database (http://
www.ncbi.nlm.nih.gov/omim) revealed that these conserved genes
are enriched for genes implicated in human muscle diseases, but
not diseases that affect other tissues (Fig. 4d). Thus, our work not
only lays a foundation for a comprehensive analysis of muscle biology
in Drosophila, but also for the systematic analysis of gene function in
vertebrate muscles.
NATURE | Vol 464 | 11 March 2010
METHODS SUMMARY
All RNAi crosses to Mef2-GAL4 were performed at 27 uC and males were assayed
at day 7 of adult life blind to the genotype. Positives were also retested blind,
along with previously untested lines in the primary screen. Where multiple RNAi
lines for the same gene resulted in different phenotypes, the gene was assigned the
strongest of these phenotypes. Larval muscles were visualized with ZCL0663, a
GFP trap in CG6416 labelling the Z-line7; adult flight muscles were bisected and
stained with phalloidin.
Received 14 January; accepted 30 December 2009.
1. Dietzl, G. et al. A genome-wide transgenic RNAi library for conditional gene
inactivation in Drosophila. Nature 448, 151–156 (2007).
2. Matsumoto, A. et al. A functional genomics strategy reveals clockwork orange as
a transcriptional regulator in the Drosophila circadian clock. Genes Dev. 21,
1687–1700 (2007).
3. Schnorrer, F. & Dickson, B. J. Muscle building; mechanisms of myotube guidance
and attachment site selection. Dev. Cell 7, 9–20 (2004).
4. Hartenstein, V. in Muscle Development in Drosophila (ed. Sink, H.) 8–27 (Landes
Biosciences, 2006).
5. Taylor, M. V. in Muscle Development in Drosophila (ed. Sink, H.) 169–203 (Landes
Biosciences, 2006).
6. Buckingham, M. Myogenic progenitor cells and skeletal myogenesis in
vertebrates. Curr. Opin. Genet. Dev. 16, 525–532 (2006).
7. Quinones-Coello, A. T. et al. Exploring strategies for protein trapping in Drosophila.
Genetics 175, 1089–1104 (2007).
8. Zaffran, S., Astier, M., Gratecos, D. & Semeriva, M. The held out wings (how)
Drosophila gene encodes a putative RNA-binding protein involved in the control of
muscular and cardiac activity. Development 124, 2087–2098 (1997).
9. Ayme-Southgate, A., Vigoreaux, J., Benian, G. & Pardue, M. L. Drosophila has a
twitchin/titin-related gene that appears to encode projectin. Proc. Natl Acad. Sci.
USA 88, 7973–7977 (1991).
10. Ren, N., Charlton, J. & Adler, P. N. The flare gene, which encodes the AIP1 protein
of Drosophila, functions to regulate F-actin disassembly in pupal epidermal cells.
Genetics 176, 2223–2234 (2007).
11. Besse, F. et al. The Ig cell adhesion molecule Basigin controls
compartmentalization and vesicle release at Drosophila melanogaster synapses. J.
Cell Biol. 177, 843–855 (2007).
12. Vigoreaux, J. O. in Muscle Development in Drosophila (ed. Sink, H.) 143–156
(Landes Biosciences, 2006).
13. Vigoreaux, J. O. Genetics of the Drosophila flight muscle myofibril: a window into
the biology of complex systems. Bioessays 23, 1047–1063 (2001).
14. Bate, M. in The development of Drosophila melanogaster, Vol. 2 (eds Bate, M. &
Martinez-Arias, A.) 1013–1090 (Cold Spring Harbor Press, 1993).
15. Park, J. et al. Mitochondrial dysfunction in Drosophila PINK1 mutants is
complemented by parkin. Nature 441, 1157–1161 (2006).
16. Reedy, M. C., Bullard, B. & Vigoreaux, J. O. Flightin is essential for thick filament
assembly and sarcomere stability in Drosophila flight muscles. J. Cell Biol. 151,
1483–1500 (2000).
17. Campbell, H. D. et al. The Drosophila melanogaster flightless-I gene involved in
gastrulation and muscle degeneration encodes gelsolin-like and leucine-rich
repeat domains and is conserved in Caenorhabditis elegans and humans. Proc. Natl
Acad. Sci. USA 90, 11386–11390 (1993).
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
18. Beuchle, D., Schwarz, H., Langegger, M., Koch, I. & Aberle, H. Drosophila MICAL
regulates myofilament organization and synaptic structure. Mech. Dev. 124,
390–406 (2007).
19. Fong, K. S. & de Couet, H. G. Novel proteins interacting with the leucine-rich
repeat domain of human flightless-I identified by the yeast two-hybrid system.
Genomics 58, 146–157 (1999).
20. Lamb, R. S., Ward, R. E., Schweizer, L. & Fehon, R. G. Drosophila coracle, a member
of the protein 4.1 superfamily, has essential structural functions in the septate
junctions and developmental functions in embryonic and adult epithelial cells.
Mol. Biol. Cell 9, 3505–3519 (1998).
21. Nongthomba, U., Ansari, M., Thimmaiya, D., Stark, M. & Sparrow, J. Aberrant
splicing of an alternative exon in the Drosophila troponin-T gene affects flight
muscle development. Genetics 177, 295–306 (2007).
22. North, K. N., Laing, N. G. & Wallgren-Pettersson, C. Nemaline myopathy: current
concepts. The ENMC International Consortium and Nemaline Myopathy. J. Med.
Genet. 34, 705–713 (1997).
23. Homyk, T. & Sheppard, D. E. Behavioral mutants of Drosophila melanogaster. I.
Isolation and mapping of mutations which decrease flight ability. Genetics 87,
95–104 (1977).
24. Reddy, K. L. et al. The Drosophila PAR domain protein 1 (Pdp1) gene encodes
multiple differentially expressed mRNAs and proteins through the use of multiple
enhancers and promoters. Dev. Biol. 224, 401–414 (2000).
25. Sandmann, T. et al. A temporal map of transcription factor activity: mef2 directly
regulates target genes at all stages of muscle development. Dev. Cell 10, 797–807
(2006).
26. Butler, M. J. et al. Discovery of genes with highly restricted expression patterns in
the Drosophila wing disc using DNA oligonucleotide microarrays. Development
130, 659–670 (2003).
27. Arbeitman, M. N. et al. Gene expression during the life cycle of Drosophila
melanogaster. Science 297, 2270–2275 (2002).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank the Vienna Drosophila RNAi Center for transgenic
RNAi lines, and H. Bellen, B. Bullard, J. Saide and T. Volk for their gifts of various
antibodies. This work was supported by Boehringer Ingelheim GmbH, the
Max-Planck Society and a postdoctoral fellowship to F.S. from the Human Frontier
Science Program. Generation of a second set of RNAi lines was supported by a
grant from the European Union 7th Framework Programme. A.S. and M.N. were
supported by the Austrian Ministry for Science and Research (GEN-AU
Bioinformatics Integration Network).
Author Contributions F.S. designed and performed the screen and other
experiments, analysed the data and wrote the manuscript with B.J.D. C.S. and
C.C.H.L. analysed selected muscle phenotypes in detail. G.D., K.S., M.F. and A.A.
assisted in the primary screen, and G.D. also in the initial data analysis. M.R.
performed the RNA microarrays. M.N. and A.S. performed the bioinformatic
analyses. K.K. led the team that generated the second RNAi hairpin lines.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to F.S.
(schnorrer@biochem.mpg.de) or B.J.D. (dickson@imp.ac.at).
291
 Vol 464 | 11 March 2010 | doi:10.1038/nature08792

LETTERS
Telomere elongation in induced pluripotent stem cells
from dyskeratosis congenita patients
Suneet Agarwal1, Yuin-Han Loh1, Erin M. McLoughlin1, Junjiu Huang2,3, In-Hyun Park1, Justine D. Miller1,
Hongguang Huo1, Maja Okuka2, Rosana Maria dos Reis2, Sabine Loewer1, Huck-Hui Ng4, David L. Keefe2,
Frederick D. Goldman5, Aloysius J. Klingelhutz6, Lin Liu2,7 & George Q. Daley1,8

Patients with dyskeratosis congenita (DC), a disorder of telomere X-linked DC is caused by mutations in the dyskerin gene (DKC1)9
maintenance, suffer degeneration of multiple tissues1–3. Patient- which encodes an RNA binding protein whose inactivation destabilizes
specific induced pluripotent stem (iPS) cells4 represent invaluable the levels of TERC, resulting in shortened telomeres and premature
in vitro models for human degenerative disorders like DC. A cardinal senescence in patient cell lines10,11. We asked whether a DKC1 mutant
feature of iPS cells is acquisition of indefinite self-renewal capacity, fibroblast line (del37L, refs 9–11) could be reprogrammed and pro-
which is accompanied by induction of the telomerase reverse tran- pagated in a pluripotent state. Compared to normal cells, the repro-
scriptase gene (TERT)5–7. We investigated whether defects in telo- gramming efficiency of del37L cells was poor, yielding only 2–5 colonies
merase function would limit derivation and maintenance of iPS cells from 105 input cells with a delayed latency (Supplementary Table 1).
from patients with DC. Here we show that reprogrammed DC cells Nevertheless, DKC1 mutant iPS colonies showed all the hallmarks of
overcome a critical limitation in telomerase RNA component (TERC) pluripotency, including characteristic morphology (Fig. 1a), gene
levels to restore telomere maintenance and self-renewal. We dis- expression (Fig. 1b and Supplementary Fig. 2a), and formation of
covered that TERC upregulation is a feature of the pluripotent state, teratomas comprising all three embryonic germ layers (Fig. 1c). PCR
that several telomerase components are targeted by pluripotency- restriction fragment length polymorphism (RFLP) analysis for the
associated transcription factors, and that in autosomal dominant DKC1 mutation confirmed the del37L mutation in iPS lines, and karyo-
DC, transcriptional silencing accompanies a 39 deletion at the type analysis was normal (Fig. 1d and Supplementary Fig. 2b, c). These
TERC locus. Our results demonstrate that reprogramming restores data show that the somatic cells from patients with a genetic impair-
telomere elongation in DC cells despite genetic lesions affecting ment in telomere elongation can be reprogrammed to pluripotency.
telomerase, and show that strategies to increase TERC expression Despite induction of endogenous TERT and telomerase activity
may be therapeutically beneficial in DC patients. (Figs 1b, e), early passage del37L iPS cell lines displayed shorter
Reprogramming of somatic cells to a state of pluripotency is char- telomeres relative to the starting fibroblast population (Fig. 1f).
acterized by prolonged self-renewal, implying induction of telomere Addition of TERT to the reprogramming factors did not result in
maintenance mechanisms. Recently, it was reported that reprogram- telomere elongation in del37L mutant cells (Fig. 1f), unlike in normal
ming of mouse cells was accompanied by elongation of telomeres8, but cells (Supplementary Fig. 3), but did increase reprogramming effi-
given the significant differences in telomere length and telomerase regu- ciency (Supplementary Table 1; Supplementary Information). We
lation in mouse and human cells, we asked whether normal human cells obtained similar results with reprogramming of an independent
displayed telomere elongation after reprogramming. We generated iPS DKC1 mutant line (A386T, ref. 11) (Fig. 1g, Supplementary Fig. 4
cell lines by retroviral transduction of primary human fibroblasts with and Supplementary Table 1). Given the telomerase dysfunction and
the factors OCT4 (also known as POU5F1), SOX2, KLF4 and c-MYC shortened telomeres, we expected to observe limited passage of DKC1
(also known as MYC), confirmed the pluripotent phenotype by gene mutant iPS cells in culture. However, unlike the parental DKC1
expression and functional analysis4, and determined the mean terminal mutant fibroblasts, which senesced after 3–4 passages, we were able
restriction fragment (TRF) length of the donor fibroblast lines and to continuously culture the DKC1 mutant iPS cell lines. Compared to
corresponding iPS lines by Southern blot. We found that mean TRF the early passage cells, we found by TRF analysis that telomere length
length and total telomeric DNA was increased in iPS cell lines relative to in del37L iPS lines increased with continued passage (Fig. 2a). Con-
the parental fibroblasts (Supplementary Fig. 1a). Similar findings using sistent with this, despite numerous interval population doublings,
retrovirally-marked single-cell fibroblast sub-clones5 contradict the late passage del37L iPS lines had telomere lengths comparable to the
notion that cells with long telomeres are uniquely susceptible to repro- original fibroblast population by quantitative PCR12 (Fig. 2b). In a
gramming (Supplementary Fig. 1b). Induction of TERT expression and blinded assessment by the complementary method of quantita-
telomerase activity correlated with reprogramming to pluripotency as tive telomere fluorescence in situ hybridization, we confirmed that
shown previously5–7 (Supplementary Figs 1c–e). These data establish telomere length was shortened immediately after derivation but
that direct factor-based reprogramming of human somatic cells results increased over time (Fig. 2c). Late passage del37L iPS cells main-
in net telomere elongation. tained a characteristic morphology, normal karyotype, and the same
1
Division of Hematology/Oncology, Children’s Hospital Boston; Pediatric Oncology, Dana Farber Cancer Institute; Department of Biological Chemistry and Molecular Pharmacology,
Harvard Medical School; Division of Hematology, Brigham and Women’s Hospital; Harvard Stem Cell Institute; Manton Center for Orphan Disease Research; Boston, Massachusetts
02115, USA. 2Department of Obstetrics and Gynecology, MDC 3125, University of South Florida College of Medicine, 12901 Bruce B. Downs Boulevard, Tampa, Florida 33612, USA.
3
State Key Laboratory of Biocontrol, School of Life Sciences, Sun Yat-sen University, Guangzhou, 510275, China. 4Gene Regulatory Laboratory, Genome Institute of Singapore;
Department of Biological Sciences, National University of Singapore; 60 Biopolis Street, #02-01, Genome, Singapore 138672. 5Department of Pediatrics, The Children’s Hospital of
Alabama, 1600 7th Avenue South, ACC 512, Birmingham, Alabama 35233, USA. 6Department of Microbiology, University of Iowa, 2202 MERF, 375 Newton Road, Iowa City, Iowa
52242, USA. 7College of Life Sciences, Nankai University, Tianjin 300071, China. 8Howard Hughes Medical Institute, Children’s Hospital Boston, Karp Family Research Building 07214,
1 Blackfan Circle, Boston, Massachusetts 02115, USA.

292
©2010 Macmillan Publishers Limited. All rights reserved
 NATURE | Vol 464 | 11 March 2010 LETTERS

a b RT–PCR for pluripotency-associated Figure 1 | Derivation and characterization of

gene expression DKC1 mutant iPS cells. a, DKC1 del37L
fibroblasts and iPS colony. b, RT–PCR for
Endo OCT4
endogenous (Endo) OCT4, SOX2, TERT and
Endo SOX2 b-actin transcripts in del37L fibroblasts (Fib) and
Endo TERT
iPS cells derived with and without exogenous
Tert. c, Histology of del37L iPS1 teratomas.
β-Actin
d, PCR RFLP verification of del37L mutation in
DKC1 del37L fibroblast DKC1 del37L iPS Fib 1 2 3 4 5 6 7 iPS clones. e, Telomerase activity assay in DKC1
With TERT
mutant iPS lines (dash 5 background).
iPS
f, Telomere Southern blot in DKC1 del37L
c DKC1 del37L iPS 1 teratoma histology d PCR RFLP verification of DKC1 del37L fibroblasts and early passage iPS clones.
mutation g, Telomere Southern blot in DKC1 A386T
Wild type del37L fibroblasts and early passage iPS clones. kb,
Fib1 Fib2 Fib3 iPS1 iPS2 kilobases.
Xmnl – + – + – + – + – +

Retinal epithelium Muscle/bone Glandular epithelium

(ectoderm) (mesoderm) (endoderm)

e Telomerase activity assay f Telomere length analysis g

DKC1 del37L DKC1 A386T
2.5 kb 21
kb 21
2.0
A450 nm–690 nm

8.6
1.5 8.6
6.1
1.0 5.0 6.1
0.5 3.6 5.0

0 2.7 3.6
del37L del37L del37L A386T A386T
Fib iPS1 iPS2 Fib iPS2
Fib iPS1 iPS2 iPS4 iPS5 iPS6 iPS7 Fib iPS1 iPS2 iPS3
With TERT With TERT

clonal fingerprint as early passage cells, and reversion of the genetic of TERT alone results in telomere elongation in wild-type fibro-
mutation in DKC1 was excluded (Supplementary Fig. 5). These blasts13,14, whereas expression of both TERC and TERT is required
data show that even in cells carrying genetic lesions that reduce to restore telomere elongation in DKC1 mutant fibroblasts11 (Sup-
telomerase function, reprogramming restores telomere elongation plementary Fig. 6). We therefore investigated TERC levels in DC
and self-renewal. fibroblasts and iPS cells. By quantitative RT–PCR, we found that
Previous studies have shown that reduced TERC levels compromise TERC levels in DKC1 mutant fibroblasts were 10–15% of TERC levels
telomerase activity in DKC1 mutant fibroblasts11. Ectopic expression found in wild-type fibroblasts, consistent with previous reports10,11
a b Figure 2 | Telomere elongation in DKC1 mutant
MM Fib del37L iPS1 del37L iPS2 del37L cell line telomere
kb iPS cells. a, Telomere Southern blot of del37L
length by qPCR fibroblasts (Fib) and iPS clones 1 and 2 as a
8.6 6
7.4 function of passage. MM, molecular mass
5
Relative length

marker. b, Quantitative real-time PCR (qPCR)

6.1 4
analysis of telomere length in del37L fibroblasts
3 and iPS clones at indicated passages (p). Error
5.0
2 bars represent s.e.m. c, Quantitative fluorescence
4.2
1 in situ hybridization (qFISH) for telomere length
3.6 0 in del37L cells. a.f.u., arbitrary fluorescence units.
Fib p9 iPS1 p19 iPS2 p20 Inset, mean relative length values shown 6s.d.
2.7
Passage 6 12 14 16 19 5 11 13 15 18 20

c qFISH telomere length analysis 600

500 500
400 iPS1 p8 400 iPS1 p26
300 3.31 ± 1.76 5.37 ± 2.35
300
200 200
100 100
400
Frequency

Fib p9 0 0
300 0 1 3 5 7 9 11 13 15 17 0 1 3 5 7 9 11 13 15 17 19
5.18 ± 3.85
200
100
600
0 400
0 1 3 5 7 9 1113 15 17 19 500
Relative length (a.f.u.) 400 300 iPS2 p25
iPS2 p7
300 3.82 ± 2.53 6.75 ± 2.91
200
200
100 100

0 0
0 1 3 5 7 9 11 13 15 17 0 1 3 5 7 9 11 13 15 17 19

293
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS NATURE | Vol 464 | 11 March 2010

(Fig. 3a). Relative to parental fibroblasts from two patients with
different DKC1 mutations, we found TERC levels increased 6–8-fold
in the reprogrammed derivatives, approaching levels in normal fibro-
blasts (Fig. 3a). In normal iPS cells, we found that TERC levels were
approximately threefold higher than the fibroblasts from which they
were derived (Fig. 3a). We next examined a cell line from a patient
with autosomal dominant DC carrying a heterozygous 821 base pairs
(bp) deletion in the 39 region of the TERC locus (DCHSF1, refs 15, 16).
In these TERC1/2 fibroblasts TERC is limiting for telomere elonga-
tion, even in the presence of exogenous TERT16. We reprogrammed
the TERC1/2 patient fibroblasts and obtained several independent iPS
lines that all showed characteristics of pluripotency and maintained
the mutant genotype (Supplementary Fig. 7). TERC1/2 iPS cells also
showed a threefold induction of TERC relative to fibroblasts (Fig. 3a),
and displayed continuous self-renewal in contrast to the early
senescence seen in the parental fibroblasts16. These data demonstrate
that reprogramming of somatic cells is accompanied by upregulation
of endogenous TERC levels, and provide a mechanism for telomere
elongation and indefinite self-renewal in DC iPS cell lines (Sup-
plementary Fig. 6).
We found that human embryonic stem (ES) cells maintain elevated
TERC levels similar to those found in wild-type iPS cells (Fig. 3a), and
that TERC levels attained in iPS cell lines derived from DKC1 mutant,
TERC1/2 and wild-type cells correlated with their respective telo-
merase activity (Fig. 3b). Furthermore, TERC knockdown in DC iPS
lines compromised telomere maintenance (Supplementary Fig. 8a).
We found that the TERC locus is not amplified in DC iPS cells, and that
on differentiation of DC iPS cells, TERC expression reverted to the
pathologically low levels found in patient fibroblasts, and telomere
length decreased (Supplementary Fig. 8b–d). Collectively, these results
show that TERC levels are increased by reprogramming, and are a
dynamically regulated and reversible property of the pluripotent state.
TERC abundance is tightly regulated by several transcriptional and
post-transcriptional mechanisms17,18. To investigate transcriptional

a Quantitative RT–PCR for TERC b TRAP assay of iPS cells

3
Relative TERC level

Relative telomerase
2

activity ×103
2

1
1

0 0
Fib iPS Fib iPS Fib iPS Fib iPS hES del37L A386T TERC+/– Wild type
del37L A386T TERC+/– Wild type DKC1 mutant
c ChIP on TERC locus
BJ1 iPS 8 Anti-OCT4 hFib2 iPS
Anti-OCT4
Fold enrichment

4 Anti-GFP Anti-GFP
6

4
2
2

0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Anti-NANOG BJ1 iPS Anti-NANOG hFib2 iPS
Fold enrichment

8 Anti-GFP 16 Anti-GFP
6 12
4 8
2 4
0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
TERC
ARPM1
1 2 3 4 5 6 7 8
Human chromosome 3 821 bp deletion 1 kb

d Quantitative RT–PCR for DKC1 e

Dyskerin
4
Relative DKC1 level

Actin
Fib iPS1 iPS2 Fib iPS10 iPS12 Fib iPS1
3
DKC1 del37L TERC+/– Wild type
2 f
6 ChIP on DKC1 locus Anti-OCT4
Fold enrichment

5 Anti-NANOG
1 4
3
0 2
Fib iPS Fib iPS Fib iPS Fib iPS hES 1
del37L A386T TERC+/– Wild type 0
1 2 3 4
DKC1 mutant DKC1
Human chromosome X
1 23 4 1 kb

Figure 3 | Upregulation of TERC and DKC1 in iPS cells. a, qPCR deletion in TERC1/2 cells is indicated. GFP, green fluorescent protein.
measurement of TERC transcripts in DKC1 mutant, TERC1/2, and wild- d, qPCR measurement of DKC1 transcripts (as in a). e, Western blot of
type fibroblasts (Fib) and iPS cells, and human ES (hES) cells, normalized to dyskerin and actin in wild-type and DC fibroblasts and iPS clones. f, ChIP of
GAPDH and relative to wild-type fibroblasts. b, Quantitative TRAP assay in the DKC1 locus (see map) in wild-type iPS cells. All error bars represent
iPS cells. c, ChIP of the human TERC locus in WT iPS cells. The 821 bp s.e.m.
294
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 LETTERS

mechanisms of TERC upregulation, we performed chromatin immu-
noprecipitation (ChIP) in human iPS cells, and detected enhanced
binding of OCT4 and NANOG in the TERC locus (Fig. 3c). These
data were corroborated in murine ES cells, which likewise showed
enhanced binding of OCT4 and NANOG at the TERC locus
(Supplementary Fig. 9). However, we were unable to detect increased
levels of nascent TERC transcription by nuclear run-off assay in iPS
cells versus fibroblasts (Supplementary Fig. 10), indicating that
OCT4 and NANOG may affect transcriptional competence rather
than transcriptional rate of the TERC locus in pluripotent cells.
Additional mechanisms are required to explain the increased
steady-state levels of TERC. To investigate post-transcriptional
mechanisms of TERC regulation, we determined DKC1 levels, which
correlate with TERC levels in human cancer cells19. In all normal and
DC iPS lines tested, we found an increase in DKC1 transcript levels
and dyskerin protein relative to the fibroblasts (Figs 3d and e).
Moreover, we found that human ES cells, like iPS cells, maintain
higher levels of DKC1 (Fig. 3d), that DKC1 levels decreased with
differentiation of iPS cells into embryoid bodies (Supplementary
Fig. 11a), and that DKC1 knockdown caused a reduction in TERC
levels and compromised cellular viability (Supplementary Fig. 11b, c).
ChIP showed binding of OCT4 and NANOG to the DKC1 promoter
in pluripotent cells (Fig. 3f), and infection of fibroblasts with retro-
viruses encoding the four reprogramming factors increased DKC1
levels (Supplementary Fig. 11d). These data establish that several telo-
merase components (TERT, TERC, DKC1) are upregulated after
reprogramming, thereby accounting for increased telomerase activity
in the pluripotent state and explaining the restoration of telomere
maintenance in DC iPS cells.
In an autosomal dominant DC patient carrying an 821 bp deletion
in the 39 region of TERC, earlier studies have concluded that the hap-
loinsufficient phenotype is explained by impaired post-transcriptional
accumulation of TERC RNA and reduced assembly of the truncated
TERC RNA into telomerase holoenzyme20–22 (Supplementary Fig. 12).
Our ChIP data indicated OCT4 and NANOG binding in a region
downstream of TERC that overlaps the 821 bp deletion (Fig. 3c).
Moreover, we found that part of this region is predicted to have regu-
latory potential on the basis of comparative genomic sequence align-
ment (Supplementary Fig. 13a). We proposed that loss of this region
might compromise transcriptional activity of the deleted allele. We
used two distinct allele-specific methods to analyse the chromatin
configuration in TERC1/2 iPS cells, and in both cases found the
mutant TERC allele to be transcriptionally inactive (Fig. 4). First,
allele-specific ChIP showed a striking abrogation of H3K4me3 marks
and RNA polymerase II binding on the deleted allele (Fig. 4a and
Supplementary Fig. 13b, c). Second, allele-specific DNase I hypersensi-
tivity analysis showed pronounced nuclease accessibility at the TERC
promoter in normal cells (Fig. 4b) and on the normal allele in
TERC1/2 cells, but complete loss of promoter hypersensitivity on
the mutant TERC allele (Fig. 4c, d). Collectively, these results identify
a cis-element in the 39 region of the TERC locus that seems essential
for formation of a transcriptionally active chromatin structure, and

a Allele-specific ChIP in TERC+/– iPS cells b TERC DNAse I HS analysis in WT

iPS cells
0 DNase
Anti-H3K4me3 Anti-RNA pol II kb
5,000 14 9.0
12 WT allele
Fold enrichment

4,000
10 Mutant allele
3,000 8 4.5
6 II
2,000
4
1,000
2
2.6
0 0 I
Primers F/R1 F/R2 F/R1 F/R2
TERC

R2 R1 F
1.0

821 bp deletion HS sites II I

ARPM1

1 kb TERC Probe
N N

c Allele-specific DNase I HS in TERC+/– iPS cells d Allele-specific DNase I HS

in TERC+/– iPS cells
0 DNase kb 0
kb
3.8 3.6 del 0 DNase
WT kb
3.7
WT
2.9
del

2.0
IWT
1.2 IWT
1.0 X Idel

1.2 X Idel
0.6 WT
Probe 1 Probe 2
HS site I HS site I
TERC ARPM1 TERC
Probe 2 1 1 kb
Probe
1 kb P
B B B P
821 bp deletion 821 bp deletion

Figure 4 | Transcriptional silencing is associated with a 39 deletion in the HS analysis probing TERC locus Bsu36I (B) fragments in TERC1/2 iPS cells
TERC locus. a, ChIP using TERC1/2 iPS cells with allele-specific primers (F/ (probe 1 assesses wild-type (WT) parent allele and HS site I (IWT); probe 2
R1 for wild type; F/R2 for mutant). Error bars represent s.e.m. b, DNase I assesses both parent alleles (WT and del) but HS I only on the deleted allele
hypersensitivity (HS) analysis probing a TERC locus NdeI (N) fragment in (Idel)). d, As in c, with PflMI (P) digestion and indicated probe for
wild-type iPS cells, showing two discrete HS sites. c, Allele-specific DNase I simultaneous assessment of both WT and deleted parent alleles and HS sites.
295
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
indicate that mutations in this region can cause haplo-insufficiency in
autosomal dominant DC by diminishing TERC transcription (Sup-
plementary Fig. 12).
By reprogramming somatic cells from patients with the human
disease dyskeratosis congenita, we have discovered novel mechanisms
of regulation of telomerase activity in the pluripotent state, thus
illustrating the value of disease-specific iPS cells for basic and trans-
lational discovery. Moreover, we have shown that the RNA com-
ponent of telomerase is upregulated in the pluripotent state to a degree
sufficient to overcome limitations to telomere maintenance in
X-linked and autosomal dominant DC. Drugs that activate the
TERC locus may favour telomere maintenance over premature attri-
tion in stem cell compartments where TERT is present but telomerase
activity is limiting, such as haematopoietic stem cells23–27, and thus
might serve as therapeutic agents for bone marrow failure. Moreover,
we speculate that in DC patients certain cell types such as germ cells
and cancer cells, whose transcriptional programs share similarities
with pluripotent cells, may also upregulate TERC to permit germline
propagation of mutations and malignant proliferation.
METHODS SUMMARY
Fibroblast lines included GM01774 (DKC1 del37L)11, AG04646 (DKC1 A386T)11,
DCHSF1 (TERC1/2 821 bp deletion)16 and NHSF2 (wild type)16. The derivation,
characterization, culture and differentiation of iPS cells was performed as
described5. Verification of mutations in GM01774 was as described9. Telomere
length analysis was performed with the TeloTAGGG Telomere Length Assay kit
(Roche). iPS cells were cultured feeder-free on Matrigel (Becton Dickinson) for
telomerase and telomere quantitative real-time PCR (qPCR) assays. Telomere
qFISH and qPCR were performed as described12,28. Telomerase activity assays
were performed using the TeloTAGGG Telomerase PCR ELISA and PCR
ELISAPLUS kits (Roche). Detection of TERT, OCT4, SOX2, NANOG, and b-actin
transcripts by RT–PCR was as described5. TERC and DKC1 transcript levels were
measured using qPCR with quantification normalized to GAPDH. Primer
sequences are provided in Methods. Chromatin immunoprecipitation was per-
formed on BJ1, HFib2, NHSF2, and DCHSF1 iPS cells as described29. Antibodies
and primers sequences are provided in Methods. Western blot was performed on
total cellular lysates using antibodies against human dyskerin and actin (Santa
Cruz). DNase I hypersensitivity analysis was performed as described30. Primer
sequences for probes are provided in Methods.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 5 November; accepted 18 December 2009.
Published online 17 February 2010.
1. Kirwan, M. & Dokal, I. Dyskeratosis congenita: a genetic disorder of many faces.
Clin. Genet. 73, 103–112 (2008).
2. Mason, P. J., Wilson, D. B. & Bessler, M. Dyskeratosis congenita–a disease of
dysfunctional telomere maintenance. Curr. Mol. Med. 5, 159–170 (2005).
3. Calado, R. T. & Young, N. S. Telomere maintenance and human bone marrow
failure. Blood 111, 4446–4455 (2008).
4. Park, I. H. et al. Disease-specific induced pluripotent stem cells. Cell 134, 877–886
(2008).
5. Park, I. H. et al. Reprogramming of human somatic cells to pluripotency with
defined factors. Nature 451, 141–146 (2008).
6. Takahashi, K. et al. Induction of pluripotent stem cells from adult human
fibroblasts by defined factors. Cell 131, 861–872 (2007).
7. Yu, J. et al. Induced pluripotent stem cell lines derived from human somatic cells.
Science 318, 1917–1920 (2007).
8. Marion, R. M. et al. Telomeres acquire embryonic stem cell characteristics in
induced pluripotent stem cells. Cell Stem Cell 4, 141–154 (2009).
9. Heiss, N. S. et al. X-linked dyskeratosis congenita is caused by mutations in a
highly conserved gene with putative nucleolar functions. Nature Genet. 19, 32–38
(1998).
296
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
10. Mitchell, J. R., Wood, E. & Collins, K. A telomerase component is defective in the
human disease dyskeratosis congenita. Nature 402, 551–555 (1999).
11. Wong, J. M. & Collins, K. Telomerase RNA level limits telomere maintenance in
X-linked dyskeratosis congenita. Genes Dev. 20, 2848–2858 (2006).
12. Cawthon, R. M. Telomere measurement by quantitative PCR. Nucleic Acids Res.
30, e47 (2002).
13. Bodnar, A. G. et al. Extension of life-span by introduction of telomerase into
normal human cells. Science 279, 349–352 (1998).
14. Vaziri, H. & Benchimol, S. Reconstitution of telomerase activity in normal human
cells leads to elongation of telomeres and extended replicative life span. Curr. Biol.
8, 279–282 (1998).
15. Vulliamy, T. et al. The RNA component of telomerase is mutated in autosomal
dominant dyskeratosis congenita. Nature 413, 432–435 (2001).
16. Westin, E. R. et al. Telomere restoration and extension of proliferative lifespan in
dyskeratosis congenita fibroblasts. Aging Cell 6, 383–394 (2007).
17. Cairney, C. J. & Keith, W. N. Telomerase redefined: integrated regulation of hTR
and hTERT for telomere maintenance and telomerase activity. Biochimie 90,
13–23 (2008).
18. Yi, X., Tesmer, V. M., Savre-Train, I., Shay, J. W. & Wright, W. E. Both
transcriptional and posttranscriptional mechanisms regulate human telomerase
template RNA levels. Mol. Cell. Biol. 19, 3989–3997 (1999).
19. Montanaro, L. et al. Dyskerin expression influences the level of ribosomal RNA
pseudo-uridylation and telomerase RNA component in human breast cancer. J.
Pathol. 210, 10–18 (2006).
20. Fu, D. & Collins, K. Distinct biogenesis pathways for human telomerase RNA and
H/ACA small nucleolar RNAs. Mol. Cell 11, 1361–1372 (2003).
21. Marrone, A., Stevens, D., Vulliamy, T., Dokal, I. & Mason, P. J. Heterozygous
telomerase RNA mutations found in dyskeratosis congenita and aplastic anemia
reduce telomerase activity via haploinsufficiency. Blood 104, 3936–3942 (2004).
22. Trahan, C. & Dragon, F. Dyskeratosis congenita mutations in the H/ACA domain
of human telomerase RNA affect its assembly into a pre-RNP. RNA 15, 235–243
(2009).
23. Broccoli, D., Young, J. W. & de Lange, T. Telomerase activity in normal and
malignant hematopoietic cells. Proc. Natl Acad. Sci. USA 92, 9082–9086 (1995).
24. Chiu, C. P. et al. Differential expression of telomerase activity in hematopoietic
progenitors from adult human bone marrow. Stem Cells 14, 239–248 (1996).
25. Hiyama, K. et al. Activation of telomerase in human lymphocytes and
hematopoietic progenitor cells. J. Immunol. 155, 3711–3715 (1995).
26. Allsopp, R. C., Morin, G. B., DePinho, R., Harley, C. B. & Weissman, I. L. Telomerase
is required to slow telomere shortening and extend replicative lifespan of HSCs
during serial transplantation. Blood 102, 517–520 (2003).
27. Kirwan, M. et al. Exogenous TERC alone can enhance proliferative potential,
telomerase activity and telomere length in lymphocytes from dyskeratosis
congenita patients. Br. J. Haematol. 144, 771–781 (2009).
28. Liu, L. et al. Telomere lengthening early in development. Nature Cell Biol. 9,
1436–1441 (2007).
29. Loh, Y. H. et al. The Oct4 and Nanog transcription network regulates pluripotency
in mouse embryonic stem cells. Nature Genet. 38, 431–440 (2006).
30. Agarwal, S. & Rao, A. Modulation of chromatin structure regulates cytokine gene
expression during T cell differentiation. Immunity 9, 765–775 (1998).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements This work was funded by grants from the National Institutes
of Health (NIH) and the Manton Center for Orphan Disease Research (G.Q.D.);
NIH K08HL089150, Amy Clare Potter Fellowship and Manton Center for Orphan
Disease Research (S.A.); the Agency of Science, Technology and Research and the
Institute of Medical Biology, Singapore (Y.-H.L.); NIH R01AG0227388 (F.D.G. and
A.J.K.); and the James and Esther King Biomedical Research Program and MOST
973 project (2009CB941000) (D.L.K and L.L).
Author Contributions S.A. performed project planning, experimental work, data
interpretation and preparation of the manuscript. Y.-H.L., I.-H.P., J.H., E.M.M.,
J.D.M., R.M.R., M.O., H.H. and S.L. performed experimental work. H.-H.N., F.D.G.,
D.L.K., A.J.K., L.L. and G.Q.D. participated in project planning, data interpretation
and preparation of the manuscript.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare competing financial interests:
details accompany the full-text HTML version of the paper at www.nature.com/
nature. Correspondence and requests for materials should be addressed to G.Q.D.
(george.daley@childrens.harvard.edu).
doi:10.1038/nature08792
METHODS
Cell lines and iPS derivation, characterization and culture. Fibroblast lines and
iPS lines for BJ1 (ATCC); GM02416, GM04569, AG20446, GM04281, GM01390
(Coriell Cell Repository); and DH1cf subclones 16, 32 and 34 have been
described4,5,31. GM01774 (DKC1 del37L)11, AG04646 (DKC1 A386T)11, and
GM874 (ALT)32 were also obtained from Coriell Cell Repository. DCHSF1
(TERC1/2 821 bp deletion) and age-matched wild-type control cells NHSF2
have been described16. Derivation, characterization, culture, and differentiation
conditions of iPS cells have been described31.
Retroviral and lentiviral vectors. Retroviruses encoding TERT, OCT4, SOX2,
KLF4 and c-MYC were as described5. The DKC1 retroviral vector was created by
cloning a full-length DKC1 cDNA (OriGene) into pMIG-W (Addgene) and
retrovirus was produced as described31. The TERC lentiviral vector was created
by replacing GFP pHIV7/SF-GFP33 (a gift from J.-K. Yee) with cherry fluorescent
protein and inserting a U3 promoter plus TERC from pBABE-U3-TER11 (a gift
from K. Collins) at an upstream site. Virus was produced by the University of
Iowa Gene Transfer Vector Core.
Telomere length and telomerase assays. Using the TeloTAGGG Telomere
Length Assay kit (Roche), 1–2 mg of DNA for each sample was digested and
subjected to Southern analysis, and determination of mean terminal restriction
fragment (TRF) length34 was performed according to manufacturer’s instruc-
tions. Briefly, autoradiographs exposed in the linear range of the hybridization
signal for the lane(s) of interest were scanned with a densitometer. A volume
array consisting of 30–40 quadrants was generated across the entire lane for each
sample from 1–24 kb. After background subtraction, absorbance (Ai) was
obtained in each of these quadrants corresponding to different DNA lengths
(Li). Mean TRF was defined as S(Ai)/S( Ai/Li) for each sample. To determine
relative total telomeric DNA signal34 in iPS cells versus fibroblasts, the S(Ai) was
obtained for each lane and normalized by comparison to ethidium bromide
staining of the gel or hybridization to a probe targeting a diploid somatic locus.
The normalized, summated absorbance value for each iPS sample was divided by
that of the corresponding fibroblast line, yielding a relative signal ratio corres-
ponding to the quantitative difference in total telomeric DNA. Telomere quanti-
tative fluorescence in situ hybridization (qFISH) and quantitative real-time PCR
(qPCR) were performed as described12,28 in a blinded fashion. For telomere
qPCR and telomerase assays, iPS cells were cultured feeder-free on Matrigel
(Becton Dickinson). Telomerase activity assays were performed using the
TeloTAGGG Telomerase PCR ELISA and PCR ELISAPLUS kits (Roche).
Quantitative real-time PCR (RT–PCR). Quantitative RT–PCR was performed
using SYBR Green Master Mix (Stratagene). Several independent biological
samples were used where available for RNA isolation using the RNeasy kit
(Qiagen), and cDNA was prepared using the Superscript III First Strand
Synthesis kit (Invitrogen). Detection of endogenous TERT, OCT4, SOX2 and
b-actin transcripts by RT–PCR was as described5. TERC and DKC1 levels were
measured using qPCR, with quantification normalized to GAPDH, using primer
pairs TRC3F/3R for TERC and GAPDH0.45F/0.45R for GAPDH, as described35,
and DKC1F1: (59-ACAGGGTGAAGAGTTCTGGCACAT-39) and DKC1R:
(59-TGAAGGTGAGGCTTCCCAACTCAA-39) for DKC1.
Chromatin immunoprecipitation (ChIP). ChIP assay was performed on BJ131,
HFib231, NHSF2 and DCHSF1 iPS cells as described29, using anti-OCT4 (Santa
Cruz; sc-8628), anti-NANOG (R&D; AF1997), anti-H3K4me3 (Abcam; ab8580),
anti-RNA polymerase II (Abcam; ab817), anti-H3K27me3 (Millipore; 07-449)
and anti-GFP (Santa Cruz; sc-9996). ChIP assay was performed on murine E14 ES
cells using anti-OCT4 and anti-GFP antibodies above, and anti-NANOG (Cosmo
Bio; REC-RCAB0002P-F).
Primers spanning the human TERC locus, Figure 3c: 1S (59-CAGCACTTTGT
TCTGATGAAGCCATCCC-39), 1AS (59-GGGTCAAGGGTATCAATGCAGAG
GCTGAATAC-39), 2S (59-GCCTATGAATAGCACTGGGGTAGGT-39), 2AS
(59-GCACCAAGGAGTTCAACTTGACCTCA-39), 3S (59-GTAGGCCTCAAAC
TGGTAGGATGGT-39), 3AS (59-CCTCCTAGCTTAGGTTGTTGATAGC-39),
4S (59-ACCACTCCTCCCTTACACTTTGTATGACGG-39), 4AS (59-GGGCTT
CCTTTGTAAGGTCTGGAGTTGGTT-39), 5S (59-CTGGTCGAGATCTACC
TTGGGAGAA-39), 5AS (59-AGACGTGAAGGCACCTCCAAAGTCG-39), 6S
(59-AACCCCGGCTCACTGCCCATTCATTTT-39), 6AS (59-ATGCAGTTCGC
TTTCCTGTTGGTGG-39), 7S (59-TTTCTATCCTCTGCAGACCAGACGC-39),
7AS (59-CGAGACAAGATTCTGCTGTAGTCAG-39), 8S (59-TTCCTCAGGCCT
GTATCACATTTCA-39), 8AS (59-GGCCAAGAAACCCGATTGTCTAGAGA-39);
Figure 4a and Supplementary Fig. 13: DCHSF1 F (59-GCGAAGAGTTGG
GCTCTGTCA-39), DCHSF1 R1 (59-CACCAACAGGAAAGCGAACTGCAT-39),
DCHSF1 R2 (59-TCATCAGGATTCAGGCTATCACCC-39). Primers spanning the
human DKC1 locus, Figure 3e: 1S (59-TAAGAGAACTGAGAAGGCTGCG-39), 1AS
(59-ATGGCCACTCATGATGGTTTGGGATC-39), 2S (59-AGCAAAGGCCTTTCA
ACCTCTCCGAGC-39), 2AS (59-TAGTTGCTTCACCTCCGGGTTTAGCTC-39),
©2010 Macmillan Publishers Limited. All rights reserved
3S (59-AGAGGTACTGTTTACGGAGCGTTCAGC-39), 3AS (59-AATGCGAT
TTGCTGGCTCGGCCAGTA-39), 4S (59-CCGAGTTGCAAGAAAGTTCTAGAGG
CC-39), 4AS (59-GTAAGGCCAAATGTATGGGTCCCATG-39). Primers spanning
the mouse TERC locus, Supplementary Fig. 9: 1S (59-CCTCCCTCCACTCC
CATACAGAAGG-39), 1AS (59-CCAAGTCTCTTTGTCCCAACTCCTC-39), 2S
(59-CCAAATGTGACTCAGTCAATGGCACTCC-39), 2AS (59-GGAGGAAGTTT
GGGTTGTGCTCTGTA-39), 3S (59-ACCACATGTGCATGTTCCTGGAGCT-39),
3AS (59-GCCTCTTCAGTAGCCATCATGCCTAATG-39), 4S (59-TATGCCACTA
TGTGGCTTCCACAGAAGG-39), 4AS (59-CCTTTCCTTCCCTCCCTTCCGGT
AC-39), 5S (59-CATAGGGAGCTTCATCAGACTCAGTG-39), 5AS (59-GGCGAC
ATTTCTCAACCAGAAGACAG-39), 6S (59-TCCTTGGCTTCGGTGATGTTGAG
TTC-39), 6AS (59-CTAAGCCGGCACTCCTTACAAGGGA-39), 7S (59-TAAGACA
CCGAACACGGGGACCAGT-39), 7AS (59-AACGTCAGCGCAGGAGCTCCAGG
TT-39), 8S (59-GAGAAAAACAGCGGGCGGAGAACAA-39), 8AS (59-ACTGGC
TAGGAAGAGTGGGGAAGCG-39), 9S (59-CCCATCCCTTCCACACGTCAGTT
CT-39), 9AS (59-GCAGTAGTATCTCTCGGGTTGTCCTTCA-39), 10S (59-GAACT
TCACAATGACCATGAGCAGTCCC-39), 10AS (59-CCCAGAGGACATTCCTT
CTAGGTTCTCTGT-39), 11S (59-CCAAGGCTTTGTCACTGACTGCTCA-39),
11AS (59-GTAGCAAGCACAGCCACCGGTGTCT-39), 12S (59-GTTGGTAAGAT
GAAGCCATCAGCCTC-39), 12AS (59-GGATAGTTGCCAGGCCACAGATCA
AT-39), 13S (59-CTCTTTCTTGAATTGGACCGTGCAGG-39), 13AS (59-GCTTC
TGGTCAGGCAGCTCATCTAAT-39).
Western analysis. Total cellular lysates were subjected to SDS–PAGE and west-
ern transfer to PVDF membranes using standard procedures and analysed using
antibodies against human dyskerin (Santa Cruz; sc-48974) and actin (Santa
Cruz; sc-1615).
DNase I hypersensitivity (HS) analysis. DNase I HS analysis was performed as
described30. Briefly, nuclei were isolated from 5 3 107–1 3 108 iPS cells and
aliquots were subjected to treatment with various concentrations of DNase I
(Worthington), followed by lysis and phenol/chloroform extraction of DNA.
Approximately 25–50 mg of DNA was digested for each lane and subjected
to Southern analysis. Hybridization of PCR generated probes labelled with
32
P-dCTP (Ready-to-go DNA Labelling kit (-dCTP); G.E. Healthcare) was per-
formed using Rapid-Hyb buffer (G.E. Healthcare) according to manufacturer’s
instructions. Primers for probes were in Figure 4b, TERCNdeIF (59-TGT
GATACAGGCCTGAGGAA-39) and TERCNdeIR (59-ATGCTTGCCTGGA
TATCTGC-39); in Figure 4c TERCBsu1F (59-CAGGACTCGGCTCACACA
T-39), TERCBsu1R (59-CGATGACCATTAAAGGAACACA-39), TERCBsu2F
(59-GATCATCTGGGGGTAGTTGC-39) and TERCBsu2R (59-AAACTGAGGC
TTACTGAAGCTGA-39). TERCBsu2F/2R probe was used in Fig. 4d.
Verifications of mutations. Verification of mutations was performed by amp-
lifying genomic DNA: for GM01774 (DKC1 del37L) as described9, using primers
XAP101F25 (59-ATTGCCAGAAGAAGATGTAGCC-39) and XAP101R27
(59-TCTCTTCAGAGGATTTGAACC-39) followed by XmnI digestion; for
DCHSF1 (family DCR101)15 as described, using primers HTRF2 (59-CCTG
CCGCCTTCCACCGTTCATT-39) and HTRR3 (59-CATTACCAGCAACAGTG
GACTCT-39); for AG04646 (DKC1 A386T)11 using primers 4646F (59-CC
AGGGAGGAACCTTGTTCT-39) and 4646R (59-ACCTCCATGCTCACCTG
TTC-39) followed by BstNI digestion. The primers amplify a 360 bp product
flanking the mutation G1156A, resulting in loss of one of two BstNI sites.
Karyotype and DNA fingerprinting. Karyotyping and DNA fingerprinting were
performed in a blinded fashion by Cell Line Genetics.
Clonal integration analysis by Southern blot. OCT4, SOX2, KLF4, and c-MYC-
encoding MSCV-based retroviruses, each containing an IRES-GFP cassette and a
single NcoI restriction site, were used for reprogramming31. Genomic DNA from
iPS cells (10 mg) was digested with NcoI and subjected to Southern blot analysis
using a GFP probe to provide a viral integration fingerprint for each clone.
TERC and DKC1 shRNA knockdown. A duplex oligonucleotide containing an
shRNA targeting the template sequence (in bold) of human TERC36 (59-
GTCTAACCCTAACTGAGAACTCGAGTTCTCAGTTAGGGTTAGACTTTT-
T-39) was cloned into the pLKO.1-puro vector to create pLKO-SHTR. Several
pLKO.1-puro-based shRNA lentiviral constructs against DKC1 were obtained
from Open Biosystems, with the most effective sequence (59-GCTC
AGTGAAATGCTGTAGAA-39) targeting the 39UTR. pLKO.1-puro-control
consisted of scrambled shRNA sequences. Lentivirus containing knockdown
constructs was produced by co-transfection of 293T cells with pLKO.1 con-
structs, psPAX2 and pMD2.G using FuGENE. Supernatants were harvested,
filtered and frozen in aliquots on days 3–4 following transfection. Titres and
knockdown efficiency was determined by infection of 293T cells for 8–12 h using
varying quantities of viral supernatant and 10 mg ml21 of protamine sulphate.
After 24–36 h, infected cells were selected with 2 mg ml21 puromycin. RNA was
harvested after 36–48 h of selection and subjected to quantitative RT–PCR to
determine knockdown efficiency of DKC1 and TERC. Lentivirus constructs for
knockdown of DKC1 or TERC were introduced into iPS cells as follows. iPS
doi:10.1038/nature08792
cultures were harvested as single cells using Accutase (Stem Cell Technologies),
and 5 3 104 cells per well were plated in a 6-well plate on Matrigel (Becton
Dickinson) in mTESR medium (Stem Cell Technologies) in the presence of
10 mM Y-27632 (Sigma)37. After 24 h viral supernatants were added in various
quantities in the presence of 10 mg ml21 protamine sulphate in mTESR medium
for 6 h. After 48 h, infected cells were selected in 0.15 mg ml21 puromycin and
maintained thereafter in mTESR medium with Matrigel or in conventional hES
medium on DR4 puromycin-resistant feeder cells.
TERC copy number determination. TERC locus copy number in iPS relative to
parental fibroblasts was determined by performing qPCR on genomic DNA from
DCHSF1 (TERC1/2) and del37L DKC1 mutant fibroblasts and their iPS deriva-
tives. Primers for the TERC locus were TRC3F and TRC3R35 and for the b-actin
locus were ActinF (59- AACGGCAGAAGAGAGAACCAGTGA-39) and ActinR
(59-TTCTACAATGAGCTGCGTGTGGCT-39). Amplification of the TERC locus
was normalized to that of the b-actin locus for each sample, and results are
displayed as a ratio of normalized TERC signal for iPS cells relative to the parent
fibroblast. Error bars denote standard error from two biological replicates.
Nuclear runoff transcription assay. Nuclei from 5 3 107 to 1 3 108 wild-type
(NHSF2) fibroblasts or derivative iPS cells (NHSF2 iPS clone 1) were isolated
and 32P-labelled runoff transcripts were generated as described38. Labelled nas-
cent RNA transcripts were purified using TRIzol (Invitrogen) and hybridization
was performed using 1–2 3 106 c.p.m. ml21. Membranes containing hybridiza-
tion targets were prepared by slot-blot of 5 mg linearized plasmid DNA targets
©2010 Macmillan Publishers Limited. All rights reserved
containing full-length human TERC cloned into pUC19, a 439 bp fragment
consisting of exon 3 of human b-actin cloned into pUC19, or pUC19 vector
alone. Hybridization was performed using Rapid-Hyb buffer (G.E. Healthcare)
with pre-hybridization for 12–16 h followed by hybridization for 48 h.
31. Park, I. H., Lerou, P. H., Zhao, R., Huo, H. & Daley, G. Q. Generation of human-
induced pluripotent stem cells. Nature Protocols 3, 1180–1186 (2008).
32. Bryan, T. M., Englezou, A., Gupta, J., Bacchetti, S. & Reddel, R. R. Telomere
elongation in immortal human cells without detectable telomerase activity. EMBO
J. 14, 4240–4248 (1995).
33. Yam, P. Y. et al. Design of HIV vectors for efficient gene delivery into human
hematopoietic cells. Mol. Ther. 5, 479–484 (2002).
34. Harley, C. B., Futcher, A. B. & Greider, C. W. Telomeres shorten during ageing of
human fibroblasts. Nature 345, 458–460 (1990).
35. Atkinson, S. P., Hoare, S. F., Glasspool, R. M. & Keith, W. N. Lack of telomerase
gene expression in alternative lengthening of telomere cells is associated with
chromatin remodeling of the hTR and hTERT gene promoters. Cancer Res. 65,
7585–7590 (2005).
36. Li, S. et al. Rapid inhibition of cancer cell growth induced by lentiviral delivery and
expression of mutant-template telomerase RNA and anti-telomerase short-
interfering RNA. Cancer Res. 64, 4833–4840 (2004).
37. Watanabe, K. et al. A ROCK inhibitor permits survival of dissociated human
embryonic stem cells. Nature Biotechnol. 25, 681–686 (2007).
38. Greenberg, M. E. & Bender, T. P. in Current Protocols in Molecular Biology Ch. 4, Unit
4.10 (Wiley, 2007).
Vol 464 | 11 March 2010 | doi:10.1038/nature08783

LETTERS
The cells and peripheral representation of sodium
taste in mice
Jayaram Chandrashekar1{, Christina Kuhn2*, Yuki Oka1*{, David A. Yarmolinsky1{, Edith Hummler3,
Nicholas J. P. Ryba2 & Charles S. Zuker1{

Salt taste in mammals can trigger two divergent behavioural
responses. In general, concentrated saline solutions elicit robust
behavioural aversion, whereas low concentrations of NaCl are
typically attractive, particularly after sodium depletion1–5. Notably,
the attractive salt pathway is selectively responsive to sodium and
inhibited by amiloride, whereas the aversive one functions as a non-
selective detector for a wide range of salts1–3,6–9. Because amiloride
is a potent inhibitor of the epithelial sodium channel (ENaC), ENaC
has been proposed to function as a component of the salt-taste-
receptor system1,3,6–14. Previously, we showed that four of the five
basic taste qualities—sweet, sour, bitter and umami—are mediated
by separate taste-receptor cells (TRCs) each tuned to a single taste
modality, and wired to elicit stereotypical behavioural responses5,15–18.
Here we show that sodium sensing is also mediated by a dedicated
population of TRCs. These taste cells express the epithelial sodium
channel ENaC19,20, and mediate behavioural attraction to NaCl. We
genetically engineered mice lacking ENaCa in TRCs, and produced
animals exhibiting a complete loss of salt attraction and sodium taste
responses. Together, these studies substantiate independent cellular
substrates for all five basic taste qualities, and validate the essential
role of ENaC for sodium taste in mice.
Sodium is the major cation of extracellular fluids and an essential
component of every fluid compartment in the body. It is therefore not
surprising that animals have evolved dedicated salt-sensing systems,
including prominent detectors in the taste system1–3. These salt-
sensitive receptors are crucial for the acceptance of low concentrations
of sodium (for example, to satisfy the ‘salt appetite’)1,3 while simulta-
neously serving as a warning mechanism against hyper-salinity2,3, thus
helping to maintain ion and water homeostasis. For humans, the ‘taste
for salt’ also has direct bearing on excessive Na1-consumption, which
is believed to be a significant dietary risk factor in hypertension, par-
ticularly in the developed world21. In mice, the low-concentration,
and behaviourally ‘attractive’ salt-taste pathway has three salient
properties: it is activated at NaCl concentrations as low as 10 mM, it
is highly selective for sodium versus other cations, and it is blocked by
lingual application of the ion-channel inhibitor amiloride8,9,14,22. The
high-concentration (aversive) pathway, conversely, begins to be sig-
nificant only at concentrations greater than 150 mM NaCl, it is non-
selective for sodium (that is, other salts are equally effective), and it is
amiloride-insensitive9,14,22.
To explore the cellular basis for the taste of NaCl (that is, determine
whether the distinct physiological and behavioural responses are
mediated by the same or separate TRCs), we developed a new pre-
paration that allows functional imaging of TRCs in response to salt
1
Howard Hughes Medical Institute and Departments of Neurobiology and Neurosciences, University of California at San Diego, La Jolla, California 92093-0649, USA. 2National
Institute of Dental and Craniofacial Research, National Institutes of Health, Bethesda, Maryland 20892, USA. 3Pharmacology and Toxicology Department, Faculty of Biology and
Medicine, University of Lausanne, CH-1005 Lausanne, Switzerland. {Present addresses: Janelia Farm Research Campus, Howard Hughes Medical Institute, Ashburn, Virginia 20147,
USA (J.C.); Departments of Biochemistry and Molecular Biophysics and of Neuroscience, Howard Hughes Medical Institute, Columbia College of Physicians and Surgeons, Columbia
University, New York, New York 10032, USA (Y.O., D.A.Y., C.S.Z.).
*These authors contributed equally to this work.
©2010 Macmillan Publishers Limited. All rights reserved
stimulation. In essence, TRCs in fungiform papillae were loaded
with the calcium-sensitive dye calcium green23 in vivo, and then were
stimulated and imaged, ex vivo, with a regime that either preferentially
activated the low-concentration pathway (100 mM NaCl), or acti-
vated both the high- and low-concentration pathways (500 mM
NaCl). To separate the contribution of each of the two salt-sensing
systems at high-stimulus concentrations, we examined the salt res-
ponses in the presence and absence of 10 mM amiloride (Fig. 1).
Receptor cells that are only activated by high concentrations of salt
also respond to a wide range of non-sodium salts (for example, from
KCl to N-methyl-D-glucamine (NMDG)-Cl; Fig. 1 and Supplemen-
tary Fig. 1), and their activity is unaffected by the presence of amiloride
(Fig. 1b, c). In contrast, low-concentrations of NaCl activate a com-
pletely separate population of TRCs; these cells do not respond to
non-sodium salts (Fig. 1a, c, see also Supplementary Fig. 1), and their
responses are blocked by amiloride (Fig. 1c). These results demon-
strate the presence of two anatomically distinct salt-sensing systems,
and accordingly suggest that the appetitive and aversive behaviours
are mediated by non-overlapping populations of TRCs. As the TRCs
activated by low concentrations of NaCl are highly selective for
sodium salts, we consider them to be the dedicated sodium-sensing
system and thus are the subject of this study.
Because amiloride is an inhibitor of the epithelial sodium channel
(ENaC), ENaC has been proposed to be a potential component of the
salt-taste-receptor system1,6,8,10–12,14. The ENaC channel is made up of
three subunits (a, b and c), and has an important role in regulating
trans-epithelial transport of Na1 in a wide range of tissues, including
kidneys, airway cells of the lung, epithelial skin cells, and the ducts of
salivary and sweat glands19,20. Although the knockout of any ENaC
subunit is sufficient to completely abolish ENaC function20, conven-
tional ENaC knockouts die within a few days of birth20, precluding
their use in physiological and behavioural studies of taste.
To examine the role of ENaCs in the taste system, we used a floxed
ENaCa conditional knockout strategy (Scnn1aflox)24. In essence, we
generated animals in which ENaC function was selectively eliminated
in all differentiated TRCs by using the cytokeratin19 gene—a marker
for all mature taste cells (see Supplementary Fig. 2) to drive the
expression of Cre-recombinase in the taste system. To investigate
the taste responses of the conditional ENaCa knockout mice, we
recorded tastant-induced action potentials from nerves innervating
the taste cells of the tongue; this physiological assay monitors the
activity of the taste system at the periphery and provides a reliable
measure of TRC function9,18,25. In wild-type mice, NaCl elicits a dose-
dependent increase in action potentials in the chorda tympani nerve,
297
LETTERS
a c
NaCl NaCl KCl
[Low] (sodium selective) 100 mM 500 mM 500 mM
5
ΔF/F (%)
0
1%
2s
b NaCl NaCl KCl
100 mM 500 mM 500 mM
[High] (non-selective)
5
ΔF/F (%)
0
1%
2s
Figure 1 | Two classes of TRCs mediate distinct salt taste responses.
Fungiform taste buds loaded with the activity sensor calcium green respond
with high selectivity and specificity to different concentrations of salt. a, A
unique subset of TRCs (labelled as [low]) respond to low concentrations of
sodium chloride (100 mM) as well as higher concentrations (500 mM) but
not to other salts (KCl). Shown below the imaging data are individual traces
from four different TRCs depicting the kinetic and amplitude changes in
intracellular calcium levels after salt stimulation; calcium changes were
with a physiological response threshold of approximately 10 mM
(Fig. 2). Loss of ENaCa in the taste system does not affect responses
to four of the five basic taste qualities: sweet, bitter, umami and sour
stimuli (Fig. 2d and Supplementary Fig. 3). In contrast, ENaCa knock-
outs show a complete loss of the responses to low concentrations of
NaCl (Fig. 2). As would be expected if ENaC was the sodium sensor,
these animals are also missing all amiloride-sensitivity in their NaCl
responses (Fig. 2). Importantly, the knockout mice retain all responses
to non-sodium salts (Fig. 2d and Supplementary Fig. 3). These results
demonstrate that taste responses to salts are mediated by genetically
separable components.
Animals ranging from simple invertebrates to mammals readily
consume low to moderate concentrations of Na1, and actively seek it
under conditions of salt deprivation1,2,4,26. Therefore, we carried out
behavioural tests of salt consumption to examine the taste behaviour
of the ENaCa conditional knockout animals both under conditions
of salt depletion (to test attraction) and under water deprivation (to
test aversion)5. We reasoned that if ENaC encodes the principal
sodium taste sensor, it should mediate all attraction to salt, and
consequently, the knockout mice should have a total loss of beha-
vioural attraction to NaCl. Indeed, ENaCa knockout mice show no
significant attraction to salt, even under conditions in which control
animals have an extraordinary appetite for sodium (Fig. 3a). In con-
trast, the aversive responses to high concentrations of NaCl (and
KCl) are unaltered in the same knockout animals (Fig. 3b, c).
Notably, behavioural responses to sweet, sour, umami and bitter
298
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
9
ΔF/F (%)
[Low]
0
9
ΔF/F (%)
[High]
0
6
ΔF/F (%)
[High]
[Low]
0
NaCl NaCl KCl
500 mM 500 mM 500 mM
+ amiloride
[Low] cells 1%
[High] cells
1%
2s
pseudo-coloured as depicted. b, A different population of TRCs (labelled as
[high]) are activated only at increased NaCl concentrations (500 mM) and
are also stimulated by KCl; shown are individual traces for three different
TRCs. c, Amiloride selectively blocks [low] responses but has no effect on
[high] responses; shown are individual traces for four different TRCs; the
duration of tastant application is denoted by black bars. See Supplementary
Fig. 1 for a diagram of the preparation, quantifications and responses to
other salts.
tastants are indistinguishable from control animals (Fig. 3d). These
results validate ENaC as the mammalian taste receptor responsible
for behavioural acceptance of (and attraction to) NaCl.
Our previous studies have shown that sweet, bitter, umami and
sour tastes are mediated by independent populations of TRCs, each
tuned to a single taste modality5,15,17,18. If this labelled-line logic of
taste coding at the periphery extends to all five basic taste modalities,
then sodium taste should also be mediated by a unique population of
TRCs. Thus, we examined whether the amiloride-sensitive salt-
sensing cells indeed define a sub-population of TRCs separate from
sweet, bitter, sour and umami TRCs. We engineered mice expressing
Cre-recombinase under the control of the ENaCa gene, and then
crossed them to a floxed green fluorescent protein (GFP) reporter
line (Z/EG). To validate the fidelity of Cre expression in ENaCa-
expressing cells, we analysed progeny from four independent Cre-
driver founders and confirmed proper GFP reporter expression in
the airway cells of the lung as well as in the kidney cortical collecting
duct cells and distal convoluted tubules—well-characterized sites of
ENaCa expression19 (see Supplementary Fig. 4).
Co-labelling with the sweet/umami/bitter TRC marker, TrpM5 (ref.
17), demonstrated that ENaCa-expressing cells are distinct from
sweet-, bitter- or umami-TRCs (Fig. 4a and Supplementary Fig. 6).
In fungiform and palate taste buds co-localization with a sour cell
marker, Car4 (ref. 27), showed the presence of two populations of
TRCs: one exhibiting co-expression of ENaCa and Car4 (Fig. 4b),
and importantly, a second one expressing ENaCa, but not sour, sweet,
NATURE | Vol 464 | 11 March 2010 LETTERS

a NaCl (mM)
10 30 60 120 250 500 1,000
Amiloride – + – + – + – + – + – + – +
Control
ENaC-KO

45 s

b c d
60 Control Control 500 mM salt Control
ENaC-KO ENaC-KO ENaC-KO
Normalized response

80
80
40

40 40
20

0 0
0

Cs l
Cl

KCl
12 mM l
M l
l
Um eet

Bi i
So r
ur
C

-C

m KC
KC

am
tte
DG 2l
NM gC
Na

Sw
10 100 1,000 10 100 1,000

M
NaCl (mM) NaCl (mM)

0
0
25
Figure 2 | ENaC is necessary for high sensitivity taste responses to sodium responses of control and knockout animals; the coloured boxes are as in
salts. Conditional knockout of ENaCa in TRCs (ENaC-KO) abolishes a (n 5 4, mean 6 s.e.m., P , 0.001 for amiloride-sensitive responses of
responses to low NaCl concentrations, and eliminates amiloride-sensitivity. control and mutant groups at 30–1,000 mM NaCl). d, ENaC-KO mice retain
a, Integrated neural recordings from the chorda tympani nerve of normal normal responses to other salts or other taste qualities (n 5 4,
(control) and ENaC-KO mice in the presence (blue traces) or absence of mean 6 s.e.m.); see Methods for details of calculations, tastants used,
10 mM amiloride. Shaded boxes illustrate the amiloride-sensitive (pink) and concentrations, genotype of strains and abbreviations.
-insensitive (blue) components. b, c, Quantifications of integrated neural

a 15 bitter or umami markers (referred to as ‘ENaC-alone’ cells; Fig. 4b, c).

ENaC-KO
Control We hypothesized that the ENaC-alone cells are the bona-fide sodium
taste sensors, and that the expression of ENaCa in sour cells may just be
Salt-attraction lick ratio

a consequence (that is, non-functional) of a common lineage between

10
the cells mediating ionic tastes. Thus, we carried out further studies.
First, we generated animals lacking ENaCa solely in Car4-expressing
cells by using a sour-cell Cre driver to excise the conditional ENaCa
5 knockout allele. As expected, these mice show wild-type responses to
sweet, bitter, umami and sour stimuli. Notably, they have normal salt
responses that are indistinguishable from wild-type controls
(Supplementary Fig. 5), thus demonstrating that the ENaCa expres-
0 sion in the sour cells is in fact not required for salt taste. In a comple-
60 120 240 480
NaCl (mM) mentary study, we also generated mice entirely lacking sour-sensing
cells15; these animals show a total loss of sour sensing, yet they maintain
i
am

b c d
t

normal salt responses (Supplementary Fig. 5). Most critically, we

ee

r
ur
tte
Um
Sw

So
Bi

1 directly imaged salt and sour responses using our new peeled epithe-
Preference ratio

1
Salt-aversion

1.0 ENaC-KO
lium preparation. Indeed, there is total segregation of the cells respond-
lick ratio

Control
0.5
ing to salt (low and high concentrations) versus those responding to
Aversion ratio

0 0
0.5 1.0
NaCl (M)
Figure 3 | ENaC function in TRCs is required for behavioural attraction to
salt. Conditional knockout of ENaCa in TRCs selectively abolishes the
attractive taste of NaCl. a, After diuretic-induced Na1-depletion, ENaC-KO
mice show little or no preference for NaCl solutions relative to water, whereas
littermate controls exhibit very robust attractive responses (P , 0.02 between
control and mutant groups at 120–480 mM NaCl). b, c, In contrast, in salt-
aversion assays, water-deprived ENaC-KO mice and controls are
indistinguishable in their responses to NaCl (b) or KCl (c). d, Behavioural
responses to other taste qualities are unaffected in the ENaC-KO animals;
shown are means 6 s.e.m. (n $ 7); see Methods for further details.
0 acid stimulation (that is, sour cells never respond to salt stimuli13; see
0.5 Supplementary Fig. 1). Taken together, these results substantiate the
0.5 1.0 functional and anatomical segregation of sodium-sensing TRCs, and
1.0
KCl (M)
prove that all five basic taste modalities are mediated by separate and
dedicated receptor cells at the periphery.
An unusual feature of the physiology of sodium taste in mice has
been the observation that the back of the tongue (circumvallate
papillae) contains no sodium-selective (amiloride-sensitive) res-
ponses10,22,28, highlighting a strong topographic segregation (front
to back) in salt taste (see later). With the identity of the amiloride-
sensitive salt taste receptor at hand, we reasoned that it should now be
possible to explore the molecular basis of the absence of sodium
sensing at the back of the tongue. ENaC channels are composed of
299
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
a b c
Fungiform
ENaCα ENaCα
TrpM5 Car4 Car4
Palate
Figure 4 | ENaC defines a novel population of TRCs. Transgenic mice
expressing GFP under the control of ENaCa (see Methods for details) were
immunostained for markers of known classes of TRC. a, No overlap in the
expression of ENaCa (green) and TrpM5 (red; a marker of sweet, bitter and
umami TRCs) was observed in fungiform or palate taste buds. b, In contrast,
Car4-expressing sour cells (red) co-express ENaCa (green; compare red- and
yellow-labelled cells). However, ENaCa is also expressed in a unique subset
of ‘ENaC-alone’ TRCs (green-only cells in b and c); see Methods for details of
mice and the illustration in c. Scale bar, 10 mm. Note that because Car4-
expressing sour cells do not express the essential ENaCb subunit, they do not
respond to salt (see Supplementary Fig. 6).
three essential subunits (a, b and c), thus we anticipated that this
Na1- and amiloride-insensitivity could be easily explained if the
functional ENaC heterotrimeric channel was not found in the cir-
cumvallate papillae. Indeed, our results show that at the back of the
tongue, ENaCa and ENaCb subunits are found in completely non-
overlapping populations of TRCs (Supplementary Fig. 6). Therefore,
amiloride-insensitivity at the back of the tongue is due to the lack of
functional ENaC channels.
In this study we have shown that ENaC, first proposed to have a
role in salt taste more than 25 years ago8,12, functions as the sodium
taste receptor. We also demonstrated that sodium taste is mediated
by a dedicated population of TRCs separate from those mediating
sweet, umami, bitter and sour taste. Notably, the taste of sodium and
non-sodium salts are detected by genetically, pharmacologically and
physiologically distinguishable TRCs. The availability of two channels
(and cellular pathways) for salt-sensing endows animals with the
ability to distinguish sodium-containing salts from other salts; this
affords mammals with a powerful mechanism to select food sources
containing adequate sodium but at the same time to avoid ingesting
excessive amounts of salt.
The presence of salt shakers on dinner tables around the world
attests to the appetitive role of salt taste in the human diet. Indeed,
salt has been a food additive shared by humans for thousands of
years, with empires from the Roman (for salary) to the British (for
taxes) valuing it as a precious commodity. Does ENaC function in
human salt taste? Physiological recordings in non-human primates
have clearly demonstrated an amiloride-sensitive component in taste
responses to salt stimuli29,30. However, psychophysical experiments
in humans remain inconclusive7, with some reports of amiloride
altering salt taste7,12, and several failing to substantiate a significant
effect for amiloride in the perception of saltiness (reviewed in ref. 7).
Given the molecular similarities between mice and humans in all
other taste modalities16, a ‘human-specific’ molecular mechanism
for salt taste would be surprising. Perhaps more likely, the contri-
bution of ENaC to human salt taste may be masked as a result of
experience, exposure to salt, and diet. Future experiments studying
300
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
people subjected to controlled salt intake may help clarify the role, if
any, of ENaC in human taste.
METHODS SUMMARY
Transgenic animals and mouse strains. ENaCa-IRES-Cre, ENaCb-IRES-tTA and
cytokeratin19-IRES-Cre are bacterial artificial chromosome (BAC)-transgenics
engineered to express Cre-recombinase or the tetracycline-dependent transactiva-
tor (tTA) by inserting an IRES-Cre or IRES-tTA transgene 39 to the Scnn1a, Scnn1b
or Krt19 stop codon. Other strains have been described previously15,24.
Calcium imaging. Fungiform TRCs were pre-loaded in vivo with Calcium
Green-1 dextran 3kD (Invitrogen) by electroporating single taste buds. After
24–36 h, taste epithelium was enzymatically peeled11,18 and placed on a recording
chamber with the apical side of TRCs facing up (Supplementary Fig. 1). Taste
stimuli were delivered in artificial saliva by focal application. Changes in intra-
cellular calcium ([Ca21]i) were imaged using a 5-Live confocal microscope
(Zeiss) and the relative change in fluorescence (DF/F) from individual TRCs
analysed and pseudo-coloured as described previously23.
Nerve recording, behavioural and localization studies. All procedures were as
described previously5,17,18,25,27.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 4 November 2009; accepted 5 January 2010.
Published online 27 January 2010.
1. Contreras, R. J. in Neural Mechanisms in Taste (ed. Cagan, R. H.) 119–145 (CRC,
1989).
2. Duncan, C. J. Salt preferences of birds and mammals. Physiol. Zool. 35, 120–132
(1962).
3. Lindemann, B. Receptors and transduction in taste. Nature 413, 219–225 (2001).
4. Beauchamp, G. K., Bertino, M., Burke, D. & Engelman, K. Experimental sodium
depletion and salt taste in normal human volunteers. Am. J. Clin. Nutr. 51, 881–889
(1990).
5. Mueller, K. L. et al. The receptors and coding logic for bitter taste. Nature 434,
225–229 (2005).
6. Eylam, S. & Spector, A. C. Taste discrimination between NaCl and KCl is disrupted
by amiloride in inbred mice with amiloride-insensitive chorda tympani nerves.
Am. J. Physiol. Regul. Integr. Comp. Physiol. 288, R1361–R1368 (2005).
7. Halpern, B. P. Amiloride and vertebrate gustatory responses to NaCl. Neurosci.
Biobehav. Rev. 23, 5–47 (1998).
8. Heck, G. L., Mierson, S. & DeSimone, J. A. Salt taste transduction occurs through
an amiloride-sensitive sodium transport pathway. Science 223, 403–405 (1984).
9. Hettinger, T. P. & Frank, M. E. Specificity of amiloride inhibition of hamster taste
responses. Brain Res. 513, 24–34 (1990).
10. Doolin, R. E. & Gilbertson, T. A. Distribution and characterization of functional
amiloride-sensitive sodium channels in rat tongue. J. Gen. Physiol. 107, 545–554
(1996).
11. Kretz, O., Barbry, P., Bock, R. & Lindemann, B. Differential expression of RNA and
protein of the three pore-forming subunits of the amiloride-sensitive epithelial
sodium channel in taste buds of the rat. J. Histochem. Cytochem. 47, 51–64 (1999).
12. Schiffman, S. S., Lockhead, E. & Maes, F. W. Amiloride reduces the taste intensity
of Na1 and Li1 salts and sweeteners. Proc. Natl Acad. Sci. USA 80, 6136–6140
(1983).
13. Vandenbeuch, A., Clapp, T. R. & Kinnamon, S. C. Amiloride-sensitive channels in
type I fungiform taste cells in mouse. BMC Neurosci. 9, 1 (2008).
14. Yoshida, R. et al. NaCl responsive taste cells in the mouse fungiform taste buds.
Neuroscience 159, 795–803 (2009).
15. Huang, A. L. et al. The cells and logic for mammalian sour taste detection. Nature
442, 934–938 (2006).
16. Yarmolinsky, D. A., Zuker, C. S. & Ryba, N. J. Common sense about taste: from
mammals to insects. Cell 139, 234–244 (2009).
17. Zhang, Y. et al. Coding of sweet, bitter, and umami tastes: different receptor cells
sharing similar signaling pathways. Cell 112, 293–301 (2003).
18. Zhao, G. Q. et al. The receptors for mammalian sweet and umami taste. Cell 115,
255–266 (2003).
19. Canessa, C. M. et al. Amiloride-sensitive epithelial Na1 channel is made of three
homologous subunits. Nature 367, 463–467 (1994).
20. Hummler, E. & Beermann, F. Scnn1 sodium channel gene family in genetically
engineered mice. J. Am. Soc. Nephrol. 11, S129–S134 (2000).
21. Strazzullo, P., D’Elia, L., Kandala, N. B. & Cappuccio, F. P. Salt intake, stroke, and
cardiovascular disease: meta-analysis of prospective studies. BMJ. doi:10.1136/
bmj.b4567 (2009).
22. Ninomiya, Y. Reinnervation of cross-regenerated gustatory nerve fibers into
amiloride-sensitive and amiloride-insensitive taste receptor cells. Proc. Natl Acad.
Sci. USA 95, 5347–5350 (1998).
23. Oka, Y. et al. Odorant receptor map in the mouse olfactory bulb: in vivo sensitivity
and specificity of receptor-defined glomeruli. Neuron 52, 857–869 (2006).
24. Hummler, E., Merillat, A. M., Rubera, I., Rossier, B. C. & Beermann, F. Conditional
gene targeting of the Scnn1a (aENaC) gene locus. Genesis 32, 169–172 (2002).
NATURE | Vol 464 | 11 March 2010
25. Nelson, G. et al. An amino-acid taste receptor. Nature 416, 199–202 (2002).
26. Dethier, V. G. The taste of salt. Am. Sci. 65, 744–751 (1977).
27. Chandrashekar, J. et al. The taste of carbonation. Science 326, 443–445
(2009).
28. Frank, M. E. Taste-responsive neurons of the glossopharyngeal nerve of the rat.
J. Neurophysiol. 65, 1452–1463 (1991).
29. Hellekant, G., Danilova, V. & Ninomiya, Y. Primate sense of taste: behavioral and
single chorda tympani and glossopharyngeal nerve fiber recordings in the rhesus
monkey, Macaca mulatta. J. Neurophysiol. 77, 978–993 (1997).
30. Hellekant, G. & Ninomiya, Y. Bitter taste in single chorda tympani taste fibers from
chimpanzee. Physiol. Behav. 56, 1185–1188 (1994).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank W. Guo and A. Becker for generation and
maintenance of mouse lines, and K. Scott and members of our laboratories for
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
valuable comments. This research was supported in part by the intramural
research program of the NIH, NIDCR (N.J.P.R.). C.S.Z. is an investigator of the
Howard Hughes Medical Institute.
Author Contributions J.C. designed the study, carried out electrophysiological and
expression studies, analysed data and wrote the paper; C.K. designed and carried
out behavioural experiments and analysed expression in engineered and knockout
mice; Y.O. designed and carried out calcium imaging experiments and analysed
data; D.A.Y. carried out molecular studies and helped write the paper; E.H.
provided essential reagents; N.J.P.R. and C.S.Z. designed the study, analysed data
and wrote the paper.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare competing financial interests:
details accompany the full-text HTML version of the paper at www.nature.com/
nature. Correspondence and requests for materials should be addressed to C.S.Z.
(cz2195@columbia.edu).
301
doi:10.1038/nature08783
METHODS
Transgenic animals and mouse strains. The PKD2L1-IRES-Cre, ROSA-DTA,
Z/EG and ENaCa conditional knockout (Scnn1aflox/flox) strains have been
described previously15,24. ENaCa-IRES-Cre, ENaCb-IRES-tTA and cytokeratin19-
IRES-Cre are BAC-transgenics engineered to express Cre-recombinase or the
tetracycline dependent transactivator (tTA) by inserting an IRES-Cre or IRES-
tTA transgene 39 to the Scnn1a, Scnn1b or Krt19 stop codon. The ENaCb-IRES-
tTA transgenic mice also carried a TetO-sapphire (modified GFP) reporter in the
BAC-transgene. Four independent lines of ENaCa-IRES-Cre expressed the Cre
transgene appropriately in kidney and lung and labelled equivalent populations
of taste cells (Supplementary Figs 4 and 6). Similarly, three founder lines expressing
ENaCb-IRES-tTA showed equivalent tTA-expression in lung, kidney and taste
tissue. Mice were inter-crossed as described in the text to generate appropriate
genotypes for physiological, behavioural and anatomical experiments. Control
groups were littermates not carrying the IRES-Cre transgene and/or animals with
at least one copy of the wild-type Scnn1a allele. Wild type, heterozygous and the
conditional ENaCa taste knockouts had no obvious differences in size, weight,
fertility, life expectancy or food consumption.
Calcium imaging. Fungiform TRCs were pre-loaded in vivo with calcium green-1
dextran 3kD (Invitrogen) by electroporating single taste buds in the tongue of
anaesthetized mice using a 3.5 mA, 50 pulses s21 3 16 cycles regime. On average,
8–9 taste buds were loaded per animal. After 24–36 h of recovery, tongues were
removed and the taste epithelium enzymatically peeled as described previously11.
The epithelium was then placed on a recording chamber with the apical side of
TRCs facing up (Supplementary Fig. 1), ensuring that the integrity and polarity of
taste buds is maintained. The apical surface of the preparation was bathed in
artificial saliva at a constant flow rate of 4.5 ml min21, and taste stimuli were
delivered by focal application using a custom made dispensing pipette (800 mm
diameter). Tastant application was for 1 s, with a minimum of 10 s of artificial
saliva between stimuli. Changes in [Ca21]i were imaged using a 5-Live confocal
microscope (Zeiss) using a 340 C-Apochromat 1.20W objective; images were
captured at 4 Hz, and DF/F from individual TRCs analysed and pseudo-coloured
as described previously23.
Nerve recordings. Lingual stimulation and recording procedures were performed
as previously described18,25. All data analyses used the integrated response over a 25-s
period immediately after the application of the stimulus. Tastants used for nerve
recordings were: 3–30 mM acesulphameK (sweet); 1–100 mM monopotassium
©2010 Macmillan Publishers Limited. All rights reserved
glutamate plus 1.0 mM inosine monophosphate (umami); 1–10 mM quinine
hydrochloride (bitter); 1–50 mM citric acid (sour). The responses to 50 mM citric
acid were used to normalize responses to each experimental series in control and
ENaCa-KO (Fig. 2). To compute the amiloride-sensitive salt component (Fig. 2b),
the stimulation regime involved sequential applications of NaCl solutions
first without and then with amiloride (in the same experimental series). The
amiloride-insensitive component was defined as the response in the presence of
amiloride (Fig. 2c). The fraction of the response inhibited by amiloride was defined
as the amiloride-sensitive component (amiloride-sensitive component 5 response
without amiloride 2 response with amiloride; Fig. 2b). Responses in experiments
involving PKD2L1-IRES-Cre Rosa-DTA (PKD2L1-DTA) and PKD2L1-IRES-
Cre ENaCa flox/flox (PKD2L1-ENaC-KO) were normalized to responses obtained
with 30 mM acesulphameK (Supplementary Fig. 5). NaCl solutions used in dose-
response studies for measuring the amiloride-insensitive sodium responses (Fig. 2c)
included 10 mM amiloride. Differences between knockout and control responses
were analysed for statistical significance using an unpaired, two-tailed Student’s
t-test and 95% confidence limits.
Behavioural assays. Behavioural assays used a custom-made gustometer to
measure immediate lick responses as described previously17,18. For salt-attraction
assays, mice were injected with furosemide (50 mg kg21) and were placed on a
low sodium diet with unrestricted water for 16–20 h to deplete sodium before
testing5. For salt-aversion assays, mice were water deprived for 24 h before test-
ing5,17,18. Control tastants were 32 mM acesulphameK (sweet), 100 mM mono-
sodium glutamate plus 1 mM inosine monophosphate and 0.1 mM amiloride
(umami), 1 mM quinine sulphate (bitter), and 150 mM citric acid (sour).
Differences between knockout and control responses were analysed for statistical
significance using an unpaired, two-tailed Student’s t-test and 95% confidence
limits; for Supplementary Fig. 5b a one-way analysis of variance (ANOVA) with
Newman–Keuls posterior test was used to compare data sets.
Immunohistochemistry and cell labelling. Immunostaining, whole-mount
imaging (GFP) and in situ hybridization were performed as described previ-
ously17,18; animals were perfused with 4% paraformaldehyde and tissue was
post-fixed for 6–48 h to allow localization of GFP. Images were obtained using
a Leica SP2 TSC or a Zeiss 510 LSM meta confocal microscope. Anti-TrpM5 and
anti-Car4 antibodies were as described previously17,27. The illustration in Fig. 4c
was composed by converting the red and green channels in Fig. 4b to greyscale
and overlaying with the ENaC-alone (green only) cells using Adobe Photoshop.

LETTERS
B-cell-derived lymphotoxin promotes
castration-resistant prostate cancer
Massimo Ammirante1*, Jun-Li Luo2*, Sergei Grivennikov1, Sergei Nedospasov3 & Michael Karin1
Prostate cancer (CaP) progresses from prostatic intraepithelial
neoplasia through locally invasive adenocarcinoma to castration-
resistant metastatic carcinoma1. Although radical prostatec-
tomy, radiation and androgen ablation are effective therapies for
androgen-dependent CaP, metastatic castration-resistant CaP is a
major complication with high mortality2. Androgens stimulate
growth and survival of prostate epithelium and early CaP.
Although most patients initially respond to androgen ablation,
many develop castration-resistant CaP within 12–18 months2.
Despite extensive studies, the mechanisms underlying the emer-
gence of castration-resistant CaP remain poorly understood and
their elucidation is critical for developing improved therapies.
Curiously, castration-resistant CaP remains androgen-receptor
dependent, and potent androgen-receptor antagonists induce
tumour regression in castrated mice3. The role of inflammation
in castration-resistant CaP has not been addressed, although it
was reported that intrinsic NF-kB activation supports its growth4.
Inflammation is a localized protective reaction to injury or infec-
tion, but it also has a pathogenic role in many diseases, including
cancer5. Whereas acute inflammation is critical for host defence,
chronic inflammation contributes to tumorigenesis and metastatic
progression. The inflammation-responsive IkB kinase (IKK)-b and
its target NF-kB have important tumour-promoting functions
within malignant cells and inflammatory cells6. The latter, includ-
ing macrophages and lymphocytes, are important elements of the
tumour microenvironment7–9, but the mechanisms underlying
their recruitment remain obscure, although they are thought to
depend on chemokine and cytokine production10. We found that
CaP progression is associated with inflammatory infiltration and
activation of IKK-a, which stimulates metastasis by an NF-kB-
independent, cell autonomous mechanism11. Here we show that
androgen ablation causes infiltration of regressing androgen-
dependent tumours with leukocytes, including B cells, in which
IKK-b activation results in production of cytokines that activate
IKK-a and STAT3 in CaP cells to enhance hormone-free survival.
To determine whether IKK-b-driven NF-kB participates in the
development of castration-resistant CaP, we conditionally deleted
the Ikkb (also known as Ikbkb) gene in prostate epithelial cells of
TRAMP mice, in which CaP is induced by prostate-specific expres-
sion of simian virus 40 T antigen12. Counter to previous expecta-
tions4, IKK-b ablation in prostate epithelial cells had no effect on
genesis and progression of androgen-dependent CaP (Supplemen-
tary Fig. 1) or development of castration-resistant CaP after castra-
tion (Supplementary Fig. 2). To facilitate mechanistic analysis of
castration-resistant CaP development we used subcutaneous allo-
grafts of the mouse androgen-dependent CaP cell line myc-CaP,
which is derived from the FVB genetic background13. Silencing of
1
Laboratory of Gene Regulation and Signal Transduction, Department of Pharmacology and Cancer Center, School of Medicine, University of California, San Diego, 9500 Gilman Drive,
La Jolla, California 92093-0723, USA. 2Scripps Research Institute-Florida, Department of Cancer Biology, 130 Scripps Way, Jupiter, Florida 33458, USA. 3Laboratory of Molecular
Immunology, Engelhardt Institute of Molecular Biology, 119991, Moscow, Russia.
*These authors contributed equally to this work.
302
©2010 Macmillan Publishers Limited. All rights reserved
Vol 464 | 11 March 2010 | doi:10.1038/nature08782
IKK-b in myc-CaP cells did not affect primary tumour growth or
castration-resistant CaP re-growth in castrated FVB mice
(Supplementary Fig. 3).
By contrast, deletion of IKK-b in interferon-responsive cells upon
induction of Mx1-Cre expression prevented castration-induced
metastatic spread in TRAMP/IkkbF/F/Mx1-Cre mice (Supplementary
Fig. 4a). To narrow down the role of IKK-b to bone-marrow-derived
cells, we reconstituted irradiated FVB mice with bone marrow from
IkkbF/F/Mx1-Cre mice and inoculated the resulting IkkbF/F and IkkbD/D
(mice injected with poly(IC) to induce Mx1-Cre) chimaeras (Sup-
plementary Fig. 4b) with myc-CaP cells. Absence of IKK-b in bone-
marrow-derived cells had no impact on primary tumour growth, but
delayed castration-resistant CaP emergence after castration (Sup-
plementary Fig. 4c). A similar delay in castration-resistant CaP growth
was seen in castrated tumour-bearing mice treated with specific IKK-b
inhibitors14,15 (Supplementary Fig. 4d and data not shown).
Dependence of castration-resistant CaP emergence on IKK-b in
bone-marrow-derived cells suggested that androgen deprivation elicits
a tumour-associated inflammatory response. Castration of mice bear-
ing myc-CaP tumours resulted in CaP cell death, peaking within
1 week (Fig. 1a). Concurrently, the regressing tumours were infiltrated
with T and B lymphocytes, natural killer cells and myeloid cell types
(Fig. 1b and Supplementary Fig. 5). Infiltration was transient, declin-
ing by 2 weeks after castration. B- and T-lymphocyte infiltration was
also detected in 100% of human CaP samples (untreated patients with
Gleason scores of 6–8), but B cells were undetectable in normal
prostate or benign prostatic hyperplasia (Fig. 1c). The messenger
RNAs for many inflammatory chemokines were also upregulated in
the myc-CaP allografts, but no changes in androgen-receptor mRNA
expression were found (Supplementary Fig. 6a). These chemo-
kines may recruit lymphoid and myeloid cells into the regressing
tumour. Indeed, antibody-mediated inhibition of CXCL13, a B-cell
chemoattractant16, prevented castration-induced B-cell recruitment
(Supplementary Fig. 6b, c). Inflammatory cytokine mRNAs, including
interleukin (IL)-6, IL-12, TNF-a and lymphotoxin, were also upregu-
lated in the regressing myc-CaP allograft, but only lymphotoxin
expression was reduced upon CXCL13 inhibition (Supplementary
Fig. 7). Castration resulted in nuclear export of androgen receptor,
but after 3 weeks androgen receptor was nuclear again (Supplementary
Fig. 8), suggesting it is activated at late phases of castration-resistant
CaP growth despite androgen depletion.
STAT3 was proposed to promote activation of unliganded andro-
gen receptor17. Indeed, STAT3 was activated during emergence of
castration-resistant CaP, faster than androgen receptor (Supplemen-
tary Fig. 9). Mx1-Cre-mediated IKK-b deletion, which was nearly
complete in mature B and T lymphocytes (Supplementary Fig. 10a),
prevented STAT3 activation in regressing tumours, but did not affect
NATURE | Vol 464 | 11 March 2010 LETTERS

a Weeks after castration a b

Tumour volume (mm3)

Rag1 + B cells

p-STAT3
p
Normal Sham Week 1 Week 2 Week 3
Rag1 + B cells
1,000 WT + Rag1

TUNEL
500 ** WTWT +Ikkk F/F
W+T Ikk-β
Ikk
kk F//F
F
8 9 128 95 12 Apoptotic
cells per field

p
p-STAT3
b CD2R (T) CCR1 (Th1) CCR3 (Th2) 0
300 0 6 12 18 24 30 36 42 48
Relative mRNA

2,000 120
Days after castration
200
80
c
*** ** **
1,000 100 WT + Rag1
40

p-STAT3
0 0 0

p
p-STAT3
Normal Sham C1 C2 C3 Normal Sham C1 C2 C3 Normal Sham C1 C2 C3

CD8a (CTL) B220 (B) F4/80 (MAC)

12
Relative mRNA

3,000 10 3000
8 2000
2,000
*** 6 ** *** Rag1 + B Rag1 + T Rag1 + WT /
WT + Ikk-β∆∆
1,000 4 1000
2
0 0 0 Figure 2 | Role of B cells and IKK-b in STAT3 activation and castration-
Normal Sham C1 C2 C3 Normal Sham C1 C2 C3 Normal Sham C1 C2 C3
resistant CaP emergence. a, myc-CaP tumours were established in wild-
CD208 (DC) NK1.1 (NK) CXCR1 (Neut)
3.0 type mice reconstituted with bone marrow from Rag1–/– males (n 5 10) or in
800
Relative mRNA

2,000 2.5 Rag1–/– males. When tumours reached 1,000 mm3, mice were castrated.
1,600 600 2.0
1,200 *** 1.5 **
Three days before castration, Rag1–/– mice (n 5 10 per group) received
400
800 *** 1.0 purified splenic B or T cells through the tail vein. Tumour volume was
200 0.5
400 measured. Results are averages 6 s.e.m. P values were determined and are
0 0 0
Normal Sham C1 C2 C3 Normal Sham C1 C2 C3 Normal Sham C1 C2 C3 indicated as above. b, Tumours were removed from radiation chimaeras
c reconstituted with IkkbF/F, IkkbD/D or Rag1–/– bone marrow, 1 week after
castration. STAT3 phosphorylation was analysed by
Number of cells per field
CD20

600 CD4
immunohistochemistry. c, Tumour-bearing Rag1–/– males were injected
**
500 CD8 ** with wild-type splenocytes (Rag1 1 WT), or purified splenic B (Rag1 1 B) or
400 F4/80
CD20 T (Rag1 1 T) lymphocytes. One day later, mice were castrated and after
300
* 1 week tumours were removed and analysed for STAT3 phosphorylation.
DAPI

200
100
0
Normal Benign Malignant
Figure 1 | Androgen ablation induces tumour inflammatory infiltration.
Six-week-old FVB males (n 5 10) were inoculated with myc-CaP cells. When
tumours reached 1,000 mm3, mice were left untreated, castrated or sham-
operated. Tumours were collected when indicated for analysis. a, Paraffin-
embedded tumour sections were stained by TdT-mediated dUTP nick end
labelling (TUNEL) to determine apoptotic cell frequency (results are
averages, n 5 3). b, Total RNA was isolated from tumour samples and
expression of indicated cell-marker mRNAs was quantified and normalized
to that of cyclophilin A (norm, normal; C1, 2, 3, mice analysed 1, 2 or 3 weeks
after castration; sham, sham-operated). Results are averages 6 s.d. (n 5 10).
c, Frozen human prostate sections (normal tissue (n 5 3), prostatic
hyperplasia (n 5 3) and malignant CaP with Gleason scores 6–8 (n 5 10))
were stained with CD4, CD8 and CD20 antibodies and 49,6-diamidino-2-
phenylindole (DAPI) and analysed by immunofluorescent microscopy. The
histogram denotes average frequencies of indicated cell types (n 5 3 per
sample). P values were determined and are depicted as insignificant,
significant (*), very significant (**) or highly significant (***). Original
magnifications, 320 (a, c).
extracellular signal-regulated kinase (ERK) and AKT activation (Sup-
plementary Fig. 10b). Immunohistochemical analysis confirmed
STAT3 activation in CaP cells, which was inhibited by A490 (Sup-
plementary Fig. 10c), an inhibitor of STAT3 phosphorylation18 that
delayed appearance of castration-resistant CaP (Supplementary Fig.
10d) but did not inhibit IKK activation (Supplementary Fig. 11a).
Conversely, ML120B did not inhibit STAT3 activation (Supplemen-
tary Fig. 11b).
Ablation of bone-marrow-derived cell IKK-b did not prevent leu-
kocyte recruitment into regressing tumours (Supplementary Fig. 12)
but did inhibit cytokine induction (Supplementary Fig. 13). To investi-
gate the role of lymphocytes further, we used chimaeric mice generated
by transplantation of bone marrow from lymphocyte-deficient Rag1–/–
mice. Although primary tumour growth was identical in mice receiv-
ing wild-type or Rag1–/– bone marrow (Supplementary Fig. 14a),
castration-resistant CaP growth was significantly delayed in mice
receiving Rag1–/– bone marrow (Fig. 2a), but not in mice reconstituted
with bone marrow from Tcrb2/2/d2/2 mice (Supplementary Fig. 14b),
©2010 Macmillan Publishers Limited. All rights reserved
** Original magnifications, 320 (b) and 340 (c).
Normal Benign Malignant
which lack only mature T lymphocytes. Castration-resistant CaP
growth was delayed in mice reconstituted with bone marrow from
JH–/– mice (Supplementary Fig. 14c), which lack mature B cells, or upon
B-cell depletion with CD20 antibody19 (Supplementary Fig. 14d).
Reconstitution of Rag1–/– FVB mice with splenic B cells, but not T cells
of FVB mice, restored rapid castration-resistant CaP re-growth
(Fig. 2a). Primary tumours, isolated from Rag1–/– chimaeric mice
1 week after castration, did not show STAT3 activation (Fig. 2b), but
reconstitution with B cells, rather than T cells, restored castration-
induced STAT3 phosphorylation (Fig. 2c).
CaP allografts from castrated, but not sham-operated, mice exhibited
IKK-a nuclear translocation (Fig. 3a, b). Silencing of IKK-a in myc-CaP
cells using short interfering RNA (siRNA) (Supplementary Fig. 15a)
had little effect on primary tumour growth, but delayed emergence of
castration-resistant CaP (Fig. 3c). Nuclear translocation of IKK-a was
dependent on IKK-b in bone-marrow-derived cells and on B cells, but
not on T cells (Fig. 3d). IKK-a nuclear translocation parallels progres-
sion of human and murine CaP and coincides with primary tumour
infiltration with cells expressing IKK-a-activating cytokines, RANK
ligand and lymphotoxin-a11. Castration induced lymphotoxin-a and
lymphotoxin-b in regressing myc-CaP allografts, but did not alter
RANK ligand expression (Supplementary Fig. 7). Lymphotoxin expres-
sion in regressing tumours was absent in Rag1–/– mice (Supplementary
Fig. 16a), and flow cytometry localized it to tumour-infiltrating B cells
(TIBCs; Supplementary Fig. 16b). We characterized TIBCs by seven-
colour flow cytometry with several markers and a lymphotoxin-b
receptor-immunoglobulin fusion protein (LTbR-Ig) to detect lympho-
toxin. The typical TIBC was a conventional, mature B2 cell that
expressed lymphotoxin on its surface and was negative for B1 markers
(Supplementary Fig. 17). IKK-b deletion abolished lymphotoxin
expression by B cells (Fig. 4a), supporting the previously suggested20
role of NF-kB in lymphotoxin-a/b induction. To examine whether
lymphotoxin production by tumour-infiltrating lymphocytes stimu-
lates castration-resistant CaP growth, we transplanted bone marrow
from B-Ltb–/– or T-Ltb–/– mice, which lack lymphotoxin-b in either
B or T cells21, into lethally irradiated mice. Lymphotoxin-b ablation in B
303
LETTERS NATURE | Vol 464 | 11 March 2010

activation (Supplementary Fig. 19). Silencing of LTbR in Myc-CaP cells

14 d
7 d

y ate
a b

y te
da stra
(Supplementary Fig. 15b) also delayed growth of castration-resistant

am

da str
a
a
Sh

C
C
CaP (Fig. 4d). Exogenous lymphotoxin maintained myc-CaP growth

IKK-α
IKK-α
in the presence of flutamide, a clinically used androgen-receptor anta-
H3
C N C N C N
gonist3, in a manner dependent on IKK-a (Fig. 4e), whose nuclear
Sham Castrated translocation was lymphotoxin inducible (Fig. 4f).
c d Castration-resistant CaP is a major complication that limits the
success of androgen ablation therapy and is responsible for most

IKK-α
1,500 sc
mortality from prostate cancer2. Castration-resistant CaP was
Tumour volume (mm3)

IKK-α siRNA
1,000
studied mainly at the level of androgen-receptor function, the central
WT + IkkβF/F WT + Ikkβ∆ /∆ player in this process4. Our results suggest that an inflammatory
500
** response triggered by death of androgen-deprived primary cancer is

IKK-α
another important contributor to emergence of castration-resistant
0 CaP. In addition to dying CaP cells, critical participants in this res-
0 3 6 9 12 15 18 21 24 27 30 33 36 39
Days after castration ponse are tumour-infiltrating B cells, which produce lymphotoxin-a:b
Rag1 + B cells Rag1 + T cells
heterotrimers that stimulate LTbR on CaP cells to induce IKK-a nuc-
Figure 3 | Role of IKK-a in emergence of castration-resistant CaP. lear translocation and STAT3 activation, thereby enhancing androgen-
Tumour-bearing mice were castrated or sham operated as above. a, Tumours independent growth (Supplementary Fig. 20). Interference with any
were analysed 1 week later for nuclear IKK-a by immunohistochemistry. component of this response results in a significant and reproducible
b, Tumours removed at indicated times were divided into cytosolic (C) and
nuclear (N) fractions, and IKK-a and histone H3 distributions determined.
3- to 4-week delay in appearance of castration-resistant CaP. Although
c, Tumours were established using myc-CaP cells transduced with these inhibitory effects are not absolute, extrapolation from ‘mouse
lentiviruses expressing scrambled siRNA (sc) or IKK-a-specific siRNA. Mice time’ to ‘human time’ suggests that interventions that prevent lym-
were castrated as above and tumour volume was measured. Results are photoxin production or signalling may delay appearance of castration-
averages 6 s.e.m. (n 5 10). P values were determined and are indicated as resistant CaP in patients undergoing androgen ablation therapy by 2.3
above. d, Tumours were established in lethally irradiated FVB males to 3.1 years. Importantly, our results suggest that, at least for CaP, the
reconstituted with IkkbF/For IkkbD/D bone marrow or in Rag1–/– males inflammatory response elicited by the dying primary tumour contri-
reconstituted with either B or T cells. IkkbF/F and IkkbD/D chimaeras were butes to the failure rather than the previously proposed success of anti-
intraperitoneally injected three times with poly(IC) (250 mg) before
castration to delete IKK-b. One week after castration, tumour samples were
cancer therapy23. Although we have not determined how death of
analysed for IKK-a distribution by immunohistochemistry. Nuclear IKK-a androgen-deprived CaP triggers the inflammatory response described
results in punctate staining, whereas cytoplasmic IKK-a results in diffuse above, necrotic cell death releases mediators, such as HMGB1 (ref. 24)
staining. Original magnifications, 320 (a, d). and IL-1a (ref. 25), that activate IKK-b and NF-kB and stimulate
production of chemokines, one of which, CXCL13, recruits B cells
cells, but not in T cells, delayed growth of castration-resistant CaP into the regressing tumour. Notably, TIBCs were detected not only
(Fig. 4b) and abolished lymphotoxin-b expression within tumours in androgen-deprived mouse CaP, but also in human CaP. Although
but did not prevent B-cell or macrophage infiltration (Supplemen- B cells were reported to promote progression of skin carcinomas9 and
tary Fig. 18). Treatment of mice with the LTbR-Ig decoy22 was as effec- exert immunosuppressive effects through activation of inhibitory Fc
tive as B-cell-specific lymphotoxin-b ablation in delaying growth of receptors on myeloid cells26, the critical tumour-promoting B-cell
castration-resistant CaP (Fig. 4c) and prevented IKK-a and STAT3 function in our experimental model is production of lymphotoxin,

a b c
WT + T-Ltβ–/– hlgG
Tumour volume (mm3)

Tumour volume (mm3)

30 Ikkβ F/F WT + B-Ltβ–/– LTβR lg

1,000 1,000
Relative mRNA

∆ /∆
Ikkβ
20 **
500 ** **
500
10
*
0 0 0
LT-α LT-β 0 6 12 18 24 30 36 42 48 0 6 12 18 24 30 36 42 48
Days after castration Days after castration

d e f
sc sc siRNA + flutamide
Tumour volume (mm3)

50 sc siRNA + flutamide +
Number of cells ×10–5

LTβR siRNA
LT-α2β1
1,000 40 LT-α2β1 0 5 10 15 20 ng ml–1
IKK-α siRNA + flutamide
30 IKK-α siRNA + flutamide IKK-α
** + LT-α2β1
500 20
** IKK-γ
Histone H3
C N C N C N C N C N
10
0 0
0 6 12 18 24 30 36 42 48 0 5 10
Days after castration Days

Figure 4 | IKK-b-dependent lymphotoxin production by tumour-infiltrating expressing scrambled (sc) siRNA or lymphotoxin-b-specific siRNA. Mice
B cells stimulates IKK-a-dependent androgen-free survival. a, RNA from were castrated and tumour volume was measured. Results are
splenic B cells of IkkbF/Fand IkkbD/D mice was analysed for lymphotoxin averages 6 s.e.m. (n 5 10). e, Myc-CaP cells (previously infected with
(LT)-a and lymphotoxin-b expression as above. Results are averages 6 s.d. lentiviruses expressing scrambled or IKK-a siRNAs) were plated at 40%
(n 5 3). b, Lethally irradiated FVB males were reconstituted with bone confluency. After 6 h, the cells were cultured with or without flutamide
marrow from B-Ltb2/2 or T-Ltb2/2 mice (n 5 6 per group). After 8 weeks, (10 mM) in the absence or presence of lymphotoxin-a2b1, (LT-a2b1) and the
myc-CaP tumours were established, mice were castrated and tumour volume cell number determined. f, Myc-CaP cells were plated at 60% confluency.
was measured as above. Results are averages 6 s.e.m. c, FVB mice (n 5 6 After 12 h, cells were stimulated for 1 h with lymphotoxin-a2b1, collected,
each group) bearing myc-CaP tumours were castrated and given hIgG or divided into cytosolic (C) and nuclear (N) fractions and IKK-a and histone
LTbR-Ig (100 mg) every 5 days, starting 4 days before castration. Tumour H3 distributions determined. In a–e, P values were determined and are
volume was measured as above. Results are averages 6 s.e.m. d, Tumours indicated as above.
were established using myc-CaP cells transduced with lentiviruses
304
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
an IKK-a-activating cytokine27, which promotes survival of androgen-
deprived CaP. Another important function of TIBCs is activation of
STAT3, an anti-apoptotic and pro-tumorigenic transcription factor28.
Although the critical STAT3-activating cytokine in this system remains
to be identified, castration induces expression of STAT3-activating
IL-6 and IL-12 family members. Furthermore, CaP cells use autocrine
IL-6 to stimulate their progression29 and activated STAT3 promotes
ligand-independent androgen-receptor activation29. Lymphotoxin is
also involved in the aetiology of human CaP. An epidemiological study
revealed that reduced CaP risk due to consumption of non-steroidal
anti-inflammatory drugs, such as aspirin, is limited to men who
express a common polymorphic LTa allele that specifies high lympho-
toxin production30. Further work should examine the effect of LTa
polymorphism on the response to androgen ablation. Our results pre-
dict that individuals who produce high levels of lymphotoxin are more
likely to develop castration-resistant CaP and should therefore be the
main beneficiaries of anti-lymphotoxin therapy.
METHODS SUMMARY
Mice were handled according to institutional and National Institutes of Health
guidelines. Tumours were grown in FVB mice. Where indicated, lethally irra-
diated FVB mice were reconstituted with bone marrow from different strains
that were backcrossed into the FVB background for at least two generations. Ltb
knockout strains were, however, in the BL6 background, which does not elicit a
graft versus host response in FVB mice. Conditions for antibody used were
posted to http://biorating.com. Human material was obtained from the
Cooperative Human Tissue Network, along with pathology reports. Histology,
gene expression and cell signalling were analysed as described11,25.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 1 May; accepted 29 December 2009.
1. Isaacs, J. T. The biology of hormone refractory prostate cancer. Why does it
develop? Urol. Clin. North Am. 26, 263–273 (1999).
2. Gulley, J., Figg, W. D. & Dahut, W. L. Treatment options for androgen-
independent prostate cancer. Clin. Adv. Hematol. Oncol. 1, 49–57 (2003).
3. Tran, C. et al. Development of a second-generation antiandrogen for treatment of
advanced prostate cancer. Science 324, 787–790 (2009).
4. Jin, R. J. et al. The nuclear factor-kB pathway controls the progression of prostate
cancer to androgen-independent growth. Cancer Res. 68, 6762–6769 (2008).
5. Karin, M., Lawrence, T. & Nizet, V. Innate immunity gone awry: linking microbial
infections to chronic inflammation and cancer. Cell 124, 823–835 (2006).
6. Karin, M. Nuclear factor-kB in cancer development and progression. Nature 441,
431–436 (2006).
7. Balkwill, F., Charles, K. A. & Mantovani, A. Smoldering and polarized inflammation
in the initiation and promotion of malignant disease. Cancer Cell 7, 211–217
(2005).
8. Coussens, L. M. & Werb, Z. Inflammation and cancer. Nature 420, 860–867
(2002).
9. de Visser, K. E., Korets, L. V. & Coussens, L. M. De novo carcinogenesis promoted by
chronic inflammation is B lymphocyte dependent. Cancer Cell 7, 411–423 (2005).
10. Mantovani, A., Allavena, P., Sica, A. & Balkwill, F. Cancer-related inflammation.
Nature 454, 436–444 (2008).
11. Luo, J. L. et al. Nuclear cytokine activated IKKa controls prostate cancer
metastasis by repressing maspin. Nature 446, 690–694 (2007).
12. Bai, A., Higham, E., Eisen, H. N., Wittrup, K. D. & Chen, J. Rapid tolerization of
virus-activated tumor-specific CD81 T cells in prostate tumors of TRAMP mice.
Proc. Natl Acad. Sci. USA 105, 13003–13008 (2008).
13. Ellwood-Yen, K. et al. Myc-driven murine prostate cancer shares molecular
features with human prostate tumors. Cancer Cell 4, 223–238 (2003).
©2010 Macmillan Publishers Limited. All rights reserved
LETTERS
14. Park, B. K. et al. NF-kB in breast cancer cells promotes osteolytic bone metastasis
by inducing osteoclastogenesis via GM-CSF. Nature Med. 13, 62–69 (2007).
15. Wen, D. et al. A selective small molecule IkB kinase b inhibitor blocks nuclear
factor kB-mediated inflammatory responses in human fibroblast-like
synoviocytes, chondrocytes, and mast cells. J. Pharmacol. Exp. Ther. 317, 989–1001
(2006).
16. Legler, D. F. et al. B cell-attracting chemokine 1, a human CXC chemokine
expressed in lymphoid tissues, selectively attracts B lymphocytes via BLR1/
CXCR5. J. Exp. Med. 187, 655–660 (1998).
17. Chen, T., Wang, L. H. & Farrar, W. L. Interleukin 6 activates androgen receptor-
mediated gene expression through a signal transducer and activator of
transcription 3-dependent pathway in LNCaP prostate cancer cells. Cancer Res.
60, 2132–2135 (2000).
18. Eriksen, K. W. et al. Constitutive STAT3-activation in Sezary syndrome: tyrphostin
AG490 inhibits STAT3-activation, interleukin-2 receptor expression and growth
of leukemic Sezary cells. Leukemia 15, 787–793 (2001).
19. Hamel, K. et al. Suppression of proteoglycan-induced arthritis by anti-CD20 B cell
depletion therapy is mediated by reduction in autoantibodies and CD41 T cell
reactivity. J. Immunol. 180, 4994–5003 (2008).
20. Worm, M. M., Tsytsykova, A. & Geha, R. S. CD40 ligation and IL-4 use different
mechanisms of transcriptional activation of the human lymphotoxin a promoter
in B cells. Eur. J. Immunol. 28, 901–906 (1998).
21. Tumanov, A. V. et al. Dissecting the role of lymphotoxin in lymphoid organs by
conditional targeting. Immunol. Rev. 195, 106–116 (2003).
22. Lee, Y. et al. Recruitment and activation of naive T cells in the islets by
lymphotoxin b receptor-dependent tertiary lymphoid structure. Immunity 25,
499–509 (2006).
23. Ghiringhelli, F. et al. Activation of the NLRP3 inflammasome in dendritic cells
induces IL-1b-dependent adaptive immunity against tumors. Nature Med. 15,
1170–1178 (2009).
24. Scaffidi, P., Misteli, T. & Bianchi, M. E. Release of chromatin protein HMGB1 by
necrotic cells triggers inflammation. Nature 418, 191–195 (2002).
25. Sakurai, T. et al. Hepatocyte necrosis induced by oxidative stress and IL-1 a
release mediate carcinogen-induced compensatory proliferation and liver
tumorigenesis. Cancer Cell 14, 156–165 (2008).
26. Nimmerjahn, F. & Ravetch, J. V. Fc-receptors as regulators of immunity. Adv.
Immunol. 96, 179–204 (2007).
27. Bonizzi, G. et al. Activation of IKKa target genes depends on recognition of specific
kB binding sites by RelB:p52 dimers. EMBO J. 23, 4202–4210 (2004).
28. Kortylewski, M. & Yu, H. Stat3 as a potential target for cancer immunotherapy. J.
Immunother. 30, 131–139 (2007).
29. Smith, P. C., Hobisch, A., Lin, D. L., Culig, Z. & Keller, E. T. Interleukin-6 and
prostate cancer progression. Cytokine Growth Factor Rev. 12, 33–40 (2001).
30. Liu, X., Plummer, S. J., Nock, N. L., Casey, G. & Witte, J. S. Nonsteroidal
antiinflammatory drugs and decreased risk of advanced prostate cancer:
modification by lymphotoxin alpha. Am. J. Epidemiol. 164, 984–989 (2006).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank C. Sawyers for myc-CaP cells, L. Coussens for JH–/–
(FVB) mice, Y. X. Fu for LTbR-Ig fusion protein, R. Rickert for help with B-cell
phenotyping, H. Cheroutre for flow cytometer use and C. Ware for bone marrow.
M.A. was supported by Fondazione Italiana per la Ricerca sul Cancro and
American-Italian Cancer Foundation fellowships. J.-L.L. was supported by a Life
Science Research Fellowship. Work in M.K.’s laboratory was supported by grants
from the National Institutes of Health, the US Army Medical Research and Material
Command and Prostate Cancer Foundation. M.K. is an American Cancer Society
Research Professor.
Author Contributions M.A., J.-L.L. and M.K. designed the study; M.A., J.-L.L. and
S.G. performed research; S.N. provided lymphotoxin-deficient mice; M.A. and M.K.
analysed data and wrote the paper.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to M.K.
(karinoffice@ucsd.edu).
305
doi:10.1038/nature08782
METHODS
Mice and cell culture. IkkbF/F (BL6) mice were crossed to TRAMP (BL63129)
mice31 and PB-Cre4 (BL6)32 or Mx1-Cre (BL6) mice33 to generate TRAMP1/1/
IkkbF/F and IkkbF/F/PB-Cre41/2 and IkkbF/F/Mx1-Cre1/2 progeny that were
intercrossed with each other for six generations. After that, TRAMP1/1/
IkkbF/F mice were crossed to IkkbF/F/PB-Cre41/2 to generate TRAMP1/2/
IkkbF/F/PB-Cre41/2 and TRAMP1/2/IkkbF/F mice. TRAMP1/1/IkkbF/F mice
were crossed to IkkbF/F/Mx1-Cre1/2 to generate TRAMP1/–/IkkbF/F/Mx1-
Cre1/2 and TRAMP1/2/IkkbF/F mice. Only male littermates were used. FVB,
Tcrb–/–d–/– (BL6), Mx1-Cre and Rag1–/– (BL63129) mice were from the Jackson
Laboratory. JH–/– mice (FVB) were provided by L. Coussens (Cancer Research
Institute and Anatomic Pathology, University of California, San Francisco).
Bone marrow from B-Ltb–/– or T-Ltb–/– mice21 was provided by C. F. Ware
(La Jolla Institute for Allergy and Immunology). PB-Cre4 and TRAMP mice
were from the Mouse Models of Human Cancer Consortium. Mice were main-
tained under specific pathogen-free conditions, and experimental protocols were
approved by the University of California, San Diego Animal Care Program,
following National Institutes of Health guidelines. Radiation chimaeras were
generated as described34. In general, irradiated FVB mice were reconstituted with
bone marrow from different strains that had been backcrossed to the FVB back-
ground for at least two generations. However, in the case of B-Ltb–/– and T-Ltb–/–
mice, bone marrow donors were of the BL6 background, whose bone marrow did
not lead to a graft versus host response in irradiated FVB mice. Myc-CaP cells
derived from the FVB background were provided by C. Sawyers (University of
California, Los Angeles and Memorial Sloan Kettering Cancer Center)35 and
were cultured under standard conditions and confirmed to be mycoplasma free.
myc-CaP cells were injected subcutaneously into the flank of male FVB mice as
described35. Tumour growth was measured with a calliper. Surgical procedures
were as described35.
Human specimens. Anonymous human prostate, benign prostatic hyperplasia
and prostate cancer frozen sections were provided by the Cooperative Human
Tissue Network. Pathology reports were provided by Network for each sample.
CXCL13 and B-cell depletion and lymphotoxin inhibition. CXCL13 neutralizing
antibody was purchased from R&D and administered intraperitoneally at 200 mg
per mouse as described36. Anti-CD20 was provided by Genentech and was admi-
nistered intraperitoneally at 250 mg per mouse. LTbR-Ig fusion protein was a gift
from Y.-X. Fu (University of Chicago) and was administered as described22. hIgG
and mouse IgG2a were purchased from Sigma-Aldrich and were used as controls.
IKK-b inhibitors. ML120 was provided by Millennium and administered orally
as described37. IKK-b inhibitor IV was purchased from Calbiochem and was
injected through the tail vein as described14.
Histological procedures. Mouse prostate and CaP tissues and dissected meta-
static tumours were immersed in 10% neutral buffered formalin before section-
ing and embedding in paraffin. Sections were stained and processed as
described11, using haematoxylin and eosin stain, TUNEL assay kit or antibodies
for IKK-a (Imgenex), phospho-STAT3 (Cell Signaling) and CD19 (eBioscience)
as described38. Frozen sections of human and mouse origins were fixed in acetone
©2010 Macmillan Publishers Limited. All rights reserved
and processed as described11, using antibodies for androgen receptor (Santa
Cruz), B220 (BD Biosciences), CD20 (BD Biosciences), CD4 (BD Biosciences)
and CD8 (BD Biosciences).
Analysis of RNA and protein expression. Total tissue RNA was prepared using
RNAeasy (Qiagen). Quantitative PCR was performed as described38. Cells and
tumours were lysed and analysed by SDS–PAGE and immunoblotting38 with
antibodies to histone H3, a-tubulin, STAT3 (Santa Cruz Biotechnology), ERK,
phospho-ERK, AKT, phospho-AKT and phospho-STAT3 (Cell Signaling).
Nuclear extracts were prepared and analysed for NF-kB DNA binding as
described39.
Lentiviral and retroviral transduction. siRNAs to mouse IKK-a, IKK-b and
LTbR mRNAs were cloned into pLSLPw, provided by I. Verma (the Salk
Institute), and lentivirus stocks were prepared as described11. Virus-containing
supernatants were added to myc-CaP cells for 2 days with polybrene, and trans-
duced cells were selected in 5 mg ml21 puromycin (Invitrogen).
Leukocyte purification and flow cytometry. Peripheral blood mononuclear cells
were isolated by centrifugation on a double-layered Histopaque-Ficoll (GE
Lifescience) gradient. Splenic B and T lymphocytes were isolated by magnetic cell
sorting with CD4, CD8 or CD19 antibodies conjugated to magnetic beads. Tumour-
infiltrating leukocytes were stained with CD45, B220, LTbR-Ig, TCRb, Gr1, CD4 and
CD8 fluorescent antibodies, as well as Aqua LIVE/DEAD dye (Molecular Probes)
and analysed on a flow cytometer (Accuri C6 or Becton Dickinson LSR II).
Statistical analyses. Results are expressed as means 6 s.e.m. or s.d. Data were
analysed by Student’s t-test and Kaplan–Meier survival analysis using the
GraphPad Prism statistical program. Error bars depict the s.e.m. or s.d.
P values . 0.05 were considered insignificant, 0.01–0.05 were considered
significant, 0.001–0.01 were considered very significant and ,0.001 were con-
sidered highly significant.
31. Greenberg, N. M. et al. Prostate cancer in a transgenic mouse. Proc. Natl Acad. Sci.
USA 92, 3439–3443 (1995).
32. Wu, X. et al. Generation of a prostate epithelial cell-specific Cre transgenic mouse
model for tissue-specific gene ablation. Mech. Dev. 101, 61–69 (2001).
33. Kuhn, R., Schwenk, F., Aguet, M. & Rajewsky, K. Inducible gene targeting in mice.
Science 269, 1427–1429 (1995).
34. Kim, S. et al. Carcinoma-produced factors activate myeloid cells through TLR2 to
stimulate metastasis. Nature 457, 102–106 (2009).
35. Watson, P. A. et al. Context-dependent hormone-refractory progression revealed
through characterization of a novel murine prostate cancer cell line. Cancer Res.
65, 11565–11571 (2005).
36. Zheng, B. et al. CXCL13 neutralization reduces the severity of collagen-induced
arthritis. Arthritis Rheum. 52, 620–626 (2005).
37. Izmailova, E. S. et al. Use of molecular imaging to quantify response to IKK-2
inhibitor treatment in murine arthritis. Arthritis Rheum. 56, 117–128 (2007).
38. Luo, J. L., Maeda, S., Hsu, L. C., Yagita, H. & Karin, M. Inhibition of NF-kB in cancer
cells converts inflammation-induced tumor growth mediated by TNFa to TRAIL-
mediated tumor regression. Cancer Cell 6, 297–305 (2004).
39. Senftleben, U. et al. Activation by IKKa of a second, evolutionary conserved, NF-
kB signaling pathway. Science 293, 1495–1499 (2001).

LETTERS
JARID2 regulates binding of the Polycomb repressive
complex 2 to target genes in ES cells
Diego Pasini1,2{, Paul A. C. Cloos1,2*, Julian Walfridsson1,2*, Linda Olsson1,2, John-Paul Bukowski1,2,
Jens V. Johansen1, Mads Bak3, Niels Tommerup3, Juri Rappsilber4 & Kristian Helin1,2
The Polycomb group (PcG) proteins have an important role in
controlling the expression of genes essential for development,
differentiation and maintenance of cell fates1,2. The Polycomb
repressive complex 2 (PRC2) is believed to regulate transcriptional
repression by catalysing the di- and tri-methylation of lysine 27 on
histone H3 (H3K27me2/3)2. At present, it is unknown how the PcG
proteins are recruited to their target promoters in mammalian cells3.
Here we show that PRC2 forms a stable complex with the Jumonji-
and ARID-domain-containing protein, JARID2 (ref. 4). Using
genome-wide location analysis, we show that JARID2 binds to more
than 90% of previously mapped PcG target genes. Notably, we show
that JARID2 is sufficient to recruit PcG proteins to a heterologous
promoter, and that inhibition of JARID2 expression leads to a major
loss of PcG binding and to a reduction of H3K27me3 levels on target
genes. Consistent with an essential role for PcG proteins in early
development5–8, we demonstrate that JARID2 is required for the
differentiation of mouse embryonic stem cells. Thus, these results
demonstrate that JARID2 is essential for the binding of PcG proteins
to target genes and, consistent with this, for the proper differenti-
ation of embryonic stem cells and normal development.
PcG proteins are essential for embryonic development5–8. They
execute their repressive functions in two distinct multiprotein com-
plexes named PRC1 and PRC2 (refs 2, 3) and share a large number of
common target genes9–11. The recruitment of the PRC1 complex is to
a large extent dependent on the activity of the PRC2 complex12–14. In
contrast, the mechanism by which mammalian PRC2 is recruited to
target genes is still enigmatic3.
To identify proteins involved in PRC2 binding to DNA we decided
to purify proteins that associate with the PRC2 complex. We per-
formed a Flag-immunopurification on nuclear extracts prepared
from a HeLa cell line that stably expresses physiological levels of a
Flag–haemagglutinin (HA)–SUZ12 fusion protein (Supplementary
Fig. 1a) and identified proteins that co-purify with SUZ12 by mass
spectrometry. This analysis identified JARID2 (Fig. 1a) and the proteins
previously reported to be components of PRC2 (refs 12, 14–16).
JARID2 is a transcriptional repressor17 and a member of the Jumonji
C (JmjC) and ARID domain protein family18, which is essential for early
embryonic development4,19. Because the ARID domain of JARID2 can
bind directly to DNA17,20, we speculated that JARID2 could be involved
in the recruitment of PRC2 to target genes.
While purifying PRC2-associated proteins, simultaneous studies
in our laboratory were aimed at understanding the functional role of
JARID2. To this extent, we had generated a 293T cell line that stably
expresses a Flag–HA–JARID2 fusion protein to purify JARID2-
associated proteins. HA-Flag tandem immunopurification from
1
Biotech Research and Innovation Centre (BRIC), 2Centre for Epigenetics, University of Copenhagen, Ole Maaløes Vej 5, 2200 Copenhagen N, Denmark. 3Wilhelm Johannsen Centre
For Functional Genome Research, Department of Cellular and Molecular Medicine, University of Copenhagen, Blegdamsvej 3, 2200 Copenhagen N, Denmark. 4Wellcome Trust Centre
for Cell Biology, University of Edinburgh, Edinburgh EH9 3JR, UK. {Present address: European Institute of Oncology, Department of Experimental Oncology, Via Adamello 16, 20141
Milan, Italy.
*These authors contributed equally to this work.
306
©2010 Macmillan Publishers Limited. All rights reserved
Vol 464 | 11 March 2010 | doi:10.1038/nature08788
nuclear extracts demonstrated that all components of the PRC2 com-
plex co-purified with JARID2 (Fig. 1b). The specificity of these results
was validated by co-immunoprecipitations using the same cells as in
Fig. 1b (Supplementary Fig. 1b) or in independent experiments (Sup-
plementary Fig. 1c, d). Moreover, several other findings supported
the interaction between JARID2 and the PRC2 members. For
example, (1) Jarid2 expressed at sub-physiological levels in mouse
ES cells (Supplementary Fig. 1e) associates with components of the
PRC2 complex (Fig. 1c). (2) Endogenous Jarid2 associates with the
PRC2 complex (Fig. 1d). (3) The association between JARID2 and
PRC2 is stable at high salt concentrations (Supplementary Fig. 1f).
(4) Jarid2, Ezh2 and Suz12 co-elute in high molecular mass frac-
tions in embryonic stem (ES) cells (Supplementary Fig. 2a). (5)
Recombinant Flag–HA–JARID2 binds to EZH2, EED and SUZ12
in insect cells (Supplementary Fig. 2b, c). (6) Epitope-tagged JARID2
co-localizes with the PRC2 components EZH2 and SUZ12, and with
the PRC1 components RING1B and CBX8 in Polycomb bodies
(Supplementary Fig. 3a). Taken together, these results strongly indicate
that JARID2 interacts with the core of the PRC2 complex, and that
JARID2 is found in a stable complex with PRC2 in vivo.
We generated several JARID2 mutants (Fig. 2a), and identified
that a small region of JARID2 (amino acids 147–165) is required
for the interaction with PRC2 (Supplementary Fig. 3b). Notably, this
region is part of a previously mapped transrepression domain of
JARID2 (ref. 17) (Fig. 2a), suggesting that PRC2 binding could be
important for the repression function of JARID2.
Stable expression of wild-type or D147–165 Gal4–JARID2 fusion
proteins (Fig. 2b) in 293T cells that contain an integrated heterologous
Gal4/luciferase reporter construct, showed that wild-type JARID2, but
not the D147–165 mutant, induces efficient repression of luciferase
activity (Fig. 2c and Supplementary Fig. 3c). Notably, chromatin
immunoprecipitation (ChIP) analysis of the same cells showed that
the recruitment to the luciferase promoter of wild-type but not of
D147–165 Gal4–JARID2 is sufficient for the recruitment of SUZ12
and EZH2, and induces a strong enrichment of H3K27me3 and a loss
of H3K4me3 (Fig. 2d). Moreover, the expression of wild-type Gal4–
JARID2 but not D147–165, also led to the recruitment of RING1B to
the same promoter, supporting the model of PRC2 dependency for
PRC1 association to target genes (Fig. 2d)13. Together, these data
strongly indicate that JARID2 can actively recruit PRC2 and PRC1
activities to promoters, and that JARID2 association with the PRC2
complex is essential for the transcriptional repression mediated by
JARID2.
To determine whether Jarid2 could have a role in regulating PcG
association with target genes, we determined the DNA-binding sites
NATURE | Vol 464 | 11 March 2010 LETTERS

a Protein Mass (Da) Peptides Score c Inputs Flag-IP a b

PRC2 Repression Δ147–
SUZ12 83,744 58 3,490 E14 E14c6 E14 E14c6 binding domain Gal4–JARID2 – WT 165
EZH2 87,959 45 2,896 JmjN
JARID2 140,301 24 1,015 Jarid2 + 1 ARID JmjC zf 1246 Gal4
EZH1 87,325 21 991 (HA/Flag) JmjN
EED 50,907 17 701 + 1 JmjC zf 1246
Δ618–728 β-Tubulin
AEBP2 53,242 10 421
JmjN
RBBP4 47,780 9 406 Ezh2
RBBP7 48,132 7 299
+ 1 ARID zf 1246
Δ916–1031
HDAC2 66,294 6 207 JmjN
+ 84 ARID JmjC zf 1246
b Suz12 c
Protein Mass (Da) Peptides Score JmjN

Relative luciferase activity

JARID2 140,301 103 5,436 – 161 ARID JmjC zf 1246 1.2

SUZ12 83,744 41 2,172 1.0

EZH2 87,959 37 2,041 + 1 352
JmjN 0.8
EED 50,907 24 1,117
AEBP2 53,242 13 804
– 1 ARID JmjC zf 1246 0.6
Eed Δ84–160 0.4
EZH1 87,325 3 142 JmjN
RBBP7 48,132 6 306 + 1 ARID JmjC zf 1246 0.2
HDAC2 66,294 3 116 Δ106–120
JmjN
RBBP4 47,780 2 69 – WT Δ147–165
– 1 ARID JmjC zf 1246
Gal4–JARID2
IPs Δ147–165
d Ezh1
Input Ctrl Suz12 Ezh2
d 3.5 IgG Gal4 1.6 IgG SUZ12 4.0 IgG EZH2
Jarid2

Percentage of input

Percentage of input

Percentage of input
3.0 1.4 3.5
Ring1b 2.5 1.2 3.0
1.0 2.5
Suz12 2.0
0.8 2.0
1.5
0.6 1.5
1.0 0.4 1.0
Figure 1 | JARID2 is a component of the PRC2 complex. a, b, Summary of 0.5 0.2 0.5
the peptides identified by mass spectrometry from an anti-Flag purification
Gal4–JARID2 – WT Δ147– – WT Δ147– – WT Δ147–
of Flag–HA–SUZ12 stably expressed in HeLa cells (a) and from a Flag–HA 165 165 165
tandem purification of Flag–HA–JARID2 expressed in 293T cells (b). Mass,

Percentage of histone H3

Percentage of histone H3
0.2 IgG RING1B 35 IgG H3K27me3 80 IgG H3K4me3
observed peptides and Mascot score are indicated for each protein.
Percentage of input

c, Endogenous PRC2 components, but not the PRC1 component Ring1b, is
associated with ectopically expressed Flag–HA–JARID2. A representative
western blot, using antibodies to the indicated proteins, of lysates prepared
from E14 ES cells with stable sub-physiological expression of
Flag–HA–JARID2 (E14c6) (Supplementary Fig. 1d). Two per cent of input is
presented as loading control. d, Association of endogenous Jarid2 with the
endogenous PRC2 complex. A representative western blot using antibodies
specific for the indicated proteins of cell extracts prepared from mouse ES
cells. Antibodies specific for Suz12, Ezh2 and haemagglutinin (negative
control (ctrl)) were used for immunoprecipitations (IPs); where indicated
2% of input is presented as loading control.
of Jarid2 and Suz12 in mouse ES cells by ChIP followed by DNA
sequencing (ChIP-seq). This analysis showed that Jarid2 and Suz12
have highly comparable binding profiles at target gene loci (Fig. 3a, e
and Supplementary Fig. 4) and 90% of the nucleotides bound by
Jarid2 are also bound by Suz12 (Fig. 3b). In agreement with these
results, Jarid2 and Suz12 also share a large number of target genes,
which are almost identical to previously mapped genes for the PRC2
complex (Fig. 3c and Supplementary Fig. 5a)9,21. Consistent with the
overlap with PcG targets, the Jarid2 target genes also associate with
H3K27me3 (Fig. 3e and Supplementary Figs 4 and 5b), show some
overlap with H3K4me3 as reported for PcG proteins (Supplementary
Fig. 5c—so-called bivalent genes), but do not overlap with
H3K36me3 (Supplementary Fig. 5d) or H3K9me3 (data not shown).
Moreover, in agreement with being PcG target genes, Jarid2 binds to
a large number of genes that encode transcription factors and signal-
ling proteins that have essential roles in regulating proper develop-
ment (Supplementary Fig. 5e).
To test the specificity of the Jarid2 ChIP-seq and validate the Jarid2
target genes, we established ES cells stably expressing Jarid2 short
hairpin RNA (shRNA). The Jarid2 shRNA led to a significant reduc-
tion of Jarid2 expression without leading to detectable changes in
expression of the PRC2 complex (Supplementary Fig. 5f). As sum-
marized in Fig. 3d and Supplementary Table 1, and also shown for
selected genes in Fig. 3e and Supplementary Fig. 6, Jarid2 inhibition
led to a complete loss of gene-specific enrichment using the Jarid2
antibody. Notably, inhibition of Jarid2 expression also led to a more
than tenfold reduction of Suz12 binding to 90% of its target genes
©2010 Macmillan Publishers Limited. All rights reserved
30 70
25 60
20
50
0.1 40
15 30
10 20
5 10
Gal4–JARID2 – WT Δ147– – WT Δ147– – WT Δ147–
165 165 165
Figure 2 | JARID2-mediated transcriptional repression is dependent on
PRC2 binding. a, JARID2 amino acids 147–165, previously mapped to be
part of the transcriptional repression domain of JARID2, is required for
interaction with the PRC2 complex. Schematic representation of the
different JARID2 mutants used in Supplementary Fig. 3b, highlighting the
overlap between the previously mapped transcriptional repression domain
with the amino acids required for PRC2 binding. zf, zinc finger. b, Western
blot analysis of 293T cells with a stably integrated Gal4-luciferase reporter
construct, with or without expression of wild-type (WT) and D147–165
Gal4–JARID2 fusions. The indicated antibodies were used to detect the
expressed proteins. c, JARID2 amino acids 147–165 are required for JARID2
transcriptional repression. Average luciferase activity of different cell clones
expressing either wild-type or D147–165 Gal4–JARID2. Luciferase activity
was normalized to the amount of protein and presented as the fold difference
relative to 293T cells not expressing Gal4–JARID2. d, JARID2 amino acids
147–165 are required for the recruitment of PRC2 to DNA. Quantitative
ChIP analysis, by real-time qPCR, using primers specific for the luciferase
transcription start site of the cells presented in c. ChIP was performed using
the indicated antibodies. Purified rabbit IgG was a negative control.
Percentage of input indicates bound/input 3 100. All error bars denote s.d.,
n 5 3.
(Fig. 3d). ChIP analysis followed by real-time quantitative PCR
(qPCR) using short interfering RNA (siRNA) and shRNA sequences
targeting different regions of Jarid2 (Supplementary Fig. 6a, c) further
validated the ChIP-seq results (Supplementary Fig. 6b, d). Notably,
the inhibition of Jarid2 expression also led to a strong reduction of
Ezh2 and Suz12 binding to target genes and, consistent with this, to a
reduction of H3K27me3 levels at promoters (Fig. 3h and Sup-
plementary Fig. 6d). The expression of Ezh2, Suz12 and Eed (Sup-
plementary Figs 5f, 6a, c and 7a) was not affected by the inhibition of
Jarid2 expression; however, consistent with a role for Jarid2 in recruit-
ing the PRC2 complex to target genes, the global levels of H3K27me3
307
LETTERS NATURE | Vol 464 | 11 March 2010

a Peaks at b c d f Flag–JARID2 – – + – g 12 12
Fzd1
14
promoters WT IgG IgG IgG
Covered Reduced and/or lost Ezh2
Peaks not at Flag–JARID2 – – – + 10 Jarid2 10 Suz12 12
Not covered Not reduced ΔARID
promoters 10
Suz12 shRNA SCR Jarid2 Jarid2 Jarid2 8 8
100 100 4,568 100 8
90 90 2,697 6 6
169 90 Jarid2 6
70% 73% 89% 69% 90% 96% 98% 100%
80 80 4 4
80
Nucleotide coverage (%)

Percentage of input
4
70 70 2 2
70 2
Total peaks (%)

Flag

Total peaks (%)

60 60 1,054 60
70
50 50 50 Hoxd9
3 666 Ezh2 5 8 8
IgG IgG IgG
40 40 40 7 7
Jarid2 14 4 Jarid2 Suz12 Ezh2
30 30 1,296 6 6
Ezh2 30 Suz12
(ref. 21) 3 5 5
20 20 1,737 20
4 4
10 10 10 2
Eed 3 3
0 2 2
ChIP Jarid2 v Suz12 Not 10-fold 5-fold 3-fold Not 1
6

6
–1

–1
2× 2

2× 2

P<2.2×10–16 detect detect 1 1

2. uz1

2. arid
10

10
P< S

P< J

ChIP Suz12 Jarid2

β-Tubulin + – – – + – – –
SCR v Jarid2 shRNA shSCR + – – –
shJarid2 – + + + – + + + – + + +
HA–Flag–JARID2 WT – – + – – – + – – – + –
e H3K36me3
20 Bmp6 locus 20 Fzd1 locus 20 Fgf4 locus JARID2 ΔARID – – – + – – – + – – – +
(ref. 29) 10 10 10 h Fzd1
0 0 0 120 IgG Jarid2 120 IgG Suz12 120 IgG Ezh2 120 IgG K27me3
20 20 40
H3K4me3 100 100 100 100
(ref. 29) 10 10 20
0 0 0 80 80 80 80
20 20 20

Percentage of fold difference SCR v RNAi

H3K27me3 10
10 10 60 60 60 60
(ref. 29)
0 0 0
40 40 40 40 40 40 40
Ezh2
(ref. 21) 20 20 20
20 20 20 20
0 0 0
20 20 20
Jarid2
(shSCR) 10 10 10 Hoxd9
0 0 0 120 IgG Jarid2 120 IgG Suz12 120 IgG Ezh2 120 IgG K27me3
20 20 20
Jarid2 100 100 100 100
(shJarid2) 10 10 10
0 0 0 80 80 80 80
Suz12
50 50 50
(shSCR) 60 60 60 60
0 0 0
40 40 40 40
Suz12 50 50 50
(shJarid2) 20 20 20 20
0 0 0
Bmp6 Fzd1 Fgf4 shSCR + – + – + – + – + – + – + – + –
shJarid2 – + – – – + – – – + – – – + – –
38430000 38440000 4750000 4755000 4760000 152045000 152050000 152055000 shSuz12 – – – + – – – + – – – + – – – +

Figure 3 | JARID2 and PRC2 associate with a very similar set of genes in ES obtained from previous studies are indicated for comparison21,29. The y-axis
cells. a, Jarid2 and Suz12 preferentially associate with promoter regions, as denotes sequence tags read. f, Western blot analysis, using indicated
shown by the percentage binding obtained using ChIP-seq. False discovery antibodies, of mouse ES cells expressing non-specific (SCR) or Jarid2-
rate (FDR) , 0.1 from ChIP analyses of ES cells using Suz12- and Jarid2- specific shRNAs before and after transient expression of wild-type
specific antibodies. b, Nucleotide coverage of Jarid2 over Suz12 binding Flag–HA–JARID2 and DARID mutant. b-Tubulin is used as a loading
regions, showing a highly significant overlap between Jarid2- and Suz12- control. g, The ARID domain of JARID2 is required for JARID2 binding to
binding sites. c, Venn diagram showing overlap of Jarid2, Suz12 and Ezh2 target genes. ChIP analysis of Fzd1 and Hoxd9 promoters in the ES cells
target genes identified by ChIP-seq analysis in ES cells. d, The percentage of presented in f using indicated antibodies. Purified rabbit IgG was a negative
Suz12 and Jarid2 (right column) peaks that are not detected or reduced in control. h, JARID2 requires the PRC2 complex for efficient association to
intensity after ChIP-seq analysis in Jarid2-shRNA-expressing ES cells. A target genes. ChIP analysis of Fzd1 and Hoxd9 promoters in the ES cells
scrambled shRNA sequence (SCR) was used as negative control for Jarid2 presented in f and in Supplementary Fig 6b using indicated antibodies.
depletion. Different intensity cutoffs for Suz12 peaks are presented. The Purified rabbit IgG was a negative control. The percentage fold difference of
total number of reads present in each individual peak defines peak intensity. SCR versus RNA interference (RNAi) is calculated as the difference in
An equal number of aligned sequences was used to compare ChIP signals. percentage of the ChIP signal in ES cells expressing Jarid2-specific shRNA
e, Examples of ChIP-seq results for the indicated loci in ES cells expressing relative to cells expressing a non-specific (SCR) shRNA. All error bars denote
non-specific and Jarid2-specific shRNAs. Binding profiles of ChIP- s.d., n 5 3.
sequencing results for Ezh2, H3K27me3, H3K4me3 and H3K36me3

and H3K27me2 (Supplementary Fig. 7a) were decreased. In agree- regulates the binding of PRC2 to its target genes and that this is
ment with the notion that PRC2 activity is required for PRC1 asso- dependent on its ARID domain.
ciation to target genes, loss of Jarid2 also led to a major displacement To understand if JARID2 recruitment of PRC2 is a two-step process
of the core subunit of the PRC1 complex Ring1b (Supplementary or if the PRC2–JARID2 complex may be recruited as one entity, we
Fig. 6b, d). Taken together, these data demonstrate that JARID2 is tested whether JARID2 can associate with its target genes in the absence
required for the binding of the PRC2 and PRC1 complexes to their of the PRC2 complex. ChIP-qPCR analyses performed from Suz12 and
target genes. Jarid2 (control) shRNA-expressing ES cells (Supplementary Figs 7a
To test whether the ARID domain of Jarid2 is involved in the and 8a) show that inhibition of Suz12 expression led to a significant
binding of PRC2 to target genes, we transiently expressed wild-type decrease in Jarid2 association with its target genes (Fig. 3h and
and DARID JARID2 (Fig. 2a) in ES cells expressing Jarid2 shRNA Supplementary Fig. 8b). Notably, however, the inhibition of Suz12
(Fig. 3f). As shown in Fig. 3g and Supplementary Fig. 7b, the re- expression led to a significant downregulation of Jarid2 levels in the
expression of wild-type, but not DARID mutant JARID2, was able ES cells (Supplementary Fig. 8a), suggesting that the stability of
to partly rescue both JARID2 and PRC2 binding to target genes. Jarid2—as has been shown for members of the PRC2 complex7—is
Importantly, the loss of Jarid2 does not disrupt the formation of regulated by complex formation. In agreement with this notion, the
the core of the PRC2 complex or affect its enzymatic activity inhibition of Suz12 expression does not lead to a decrease in Jarid2
(Supplementary Fig. 7c). These results demonstrate that JARID2 messenger RNA levels (Supplementary Fig. 8c). Taken together, these
308
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010 LETTERS

a SCR shRNA Jarid2 shRNA b shJarid2

– +
c Jarid2 Fgf4 Oct4
(pluripotent) (pluripotent)

Fold difference
Jarid2 1.2 1.4
1.0 1.2 1.4
1.0 shCtrl
Day 2 Ezh2 0.8 1.0
0.6 0.8 shJarid2
0.6 0.6
Suz12 0.4 0.4
0.2 0.2 0.2
β-Tubulin Days 0 10 0 10 0 10
Day 7
Msi1 Nes Olig1 Olig2
(ectoderm) (ectoderm) (ectoderm) (ectoderm)

Fold difference
25 35 20 2,500
20 30 16 2,000
25
15 20 12 1,500
Day 10 10 15 8 1,000
5 10 4 500
5
0
Days 0 10 0 10 0 10 0 10
d 1,019 shCtrl shJarid2
Foxa2 Gata4 (endo/ Pax3 Hoxa4
815 (endoderm) mesoderm) 6,000 (mesoderm) 1,000

Fold difference
498 120 30
100 25 5,000 800
453 80 20 4,000 600
204 60 15 3,000
92 45 40 2,000 400
29 10
296 74 20 5 1,000 200
0 0 0
Upregulated genes Downregulated genes Days 0 10 0 10 0 10 0 10

Figure 4 | Jarid2 is required for ES cell differentiation. a, Jarid2 is required
for ES cell differentiation. Pictures of embryoid bodies formed from ES cells
expressing non-specific or Jarid2-specific shRNAs. Original magnification,
35. b, Western blot analysis of ES cells expressing non-specific and Jarid2-
specific shRNAs used in a and c. The western blots were probed with the
indicated antibodies. b-Tubulin is presented as a loading control.
c, Differentiation markers are not induced in ES cells expressing Jarid2
shRNA. qPCR analysis of the indicated genes during ES cell differentiation of
the cells presented in a. Oct4 is also known as Pou5f1. Error bars denote s.d.,
n 5 3. d, Downregulation of Jarid2 changes the gene expression pattern in ES
cells induced to differentiate. Venn diagrams of the genome-wide expression
profiles of protein-coding genes significantly altered 10 days after induction
of ES cells expressing control or Jarid2 shRNAs.

results indicate that JARID2 is in a stable complex with PRC2, and that
the JARID2–PRC2 complex is recruited to the PcG target genes as one
entity depending on the DNA-binding ARID domain of JARID2. This
suggestion is supported by the ChIP-seq data demonstrating that PRC2
is always present at the same genomic positions of JARID2.
PRC2 activity is not required for the proliferation of ES cells, but is
essential for the ability of ES cells to differentiate22–24. To determine
whether Jarid2 shares common biological functions with PcG proteins,
we challenged ES cells that stably express Jarid2-specific shRNAs
(Fig. 4b) to undergo differentiation by forming embryoid bodies in
suspension cultures. Although Jarid2-depleted ES cells showed no pro-
liferation defects and formed embryoid bodies with similar efficiency
as control cells (Fig. 4a, day 2), the prolonged differentiation of Jarid2-
depleted embryoid bodies was severely compromised when compared
to control cells (Fig. 4a, days 7 and 10, and Supplementary Fig. 9a,
day 7). The failure of proper differentiation shown by the Jarid2-
depleted embryoid bodies is consistent with the lack of transcriptional
activation of differentiation markers in the same cells (Fig. 4c and
Supplementary Fig. 9b). This result was further confirmed by the
genome-wide expression analyses showing that both the activation
and the repression of differentiation-regulated genes are strongly
compromised during the differentiation of Jarid2-depleted ES cells
(Fig. 4d). Together, these results indicate that Jarid2 and PcG proteins
share common biological functions, and demonstrate that Jarid2,
similar to PcG proteins, is required for correct differentiation of ES
cells.
The JARID2 protein is highly conserved throughout evolution
(Supplementary Fig. 10a) and belongs to a family of ARID- and
Jumonji-domain containing proteins together with JARID1A–D (also
known as KDM5A–D; Supplementary Fig. 11c). However, JARID2 is
only distantly related to members of the JARID1 family and shows
a low degree of similarity (<20%, Supplementary Fig. 11d, e), par-
ticularly within the highly conserved PRC2-binding region (<10%,
Supplementary Fig. 10a, b). The members of the JARID1 family func-
tion as specific H3K4me3/2 demethylases, and we previously reported
that the activity of one of the members of this family, JARID1A, is
involved in PcG-repressive functions but is not required for PcG
binding to target genes25. In contrast, JARID2 is probably not acting
as a hydroxylase and hence a histone demethylase, owing to the lack
of conservation of two essential histidine residues in the HXD/EXnH
metal-binding motif of the JmjC domain18 (Supplementary Fig. 10c–e).
Our data, together with the reported ability of the JARID2 ARID
domain to bind different DNA sequences17, demonstrate that JARID2
©2010 Macmillan Publishers Limited. All rights reserved
is a critical regulator of PRC2 binding to target genes. This is further
supported by the ability of JARID2 to directly recruit PRC2 activity
to an artificial promoter (Fig. 2d). Previous studies have identified
several DNA sequences binding to the ARID domain of JARID2
(ref. 17). The diversity of these sequences suggests that JARID2 might
have a broader specificity for DNA, and they could indicate that
the association of PRC2 with specific target genes is not only deter-
mined by JARID2, but also in combination with other transcription
factors and by the affinity of PRC2 for H3K27me2/3 (refs 26, 27).
Furthermore, the requirement of Jarid2 for ES cell differentiation,
but not for ES cell self-renewal, is consistent with the biological func-
tions of PcG proteins8,22–24 and the embryonic lethality observed in
Jarid2 knockout mice4,19, further illustrating the critical role of JARID2
in regulating PcG function.
METHODS SUMMARY
ES cells. ES cells were grown, manipulated and differentiated as previously
described22,28. In brief, undifferentiated ES cells were induced to differentiate
into embryoid bodies by hanging-drop formation, and were stimulated with
0.5 mM all-trans-retinoic acid (ATRA) between days 2 and 5 of differentiation.
Embryoid bodies were analysed at the indicated times.
Protein purification. For each purification, 100–200 mg of nuclear extracts were
used. In brief, nuclei were lysed in high salt buffer (300 mM NaCl) and incubated
overnight with M2 Flag-agarose beads (Sigma). Beads were washed six times in
high salt buffer and purified proteins were eluted with 0.5 mg ml21 of Flag
peptide. When required, eluted proteins were incubated overnight with protein
A-Sepharose beads cross-linked with haemagglutinin (12CA5)-specific antibodies,
washed six times in high salt buffer and purified proteins were eluted with
0.5 mg ml21 haemagglutinin peptide. Eluted proteins were identified by mass
spectrometry.
ChIP and ChIP-sequencing. ChIP experiments were performed as described28.
Chromatin immunoprecipitated DNA was prepared for sequencing following
the Illumina recommended procedure, and sequenced using the Genome
Analyser II (Illumina). Two mismatches within the first 32 bases of a read were
used as cutoff for mapping against the mouse mm9 genome. Enriched regions
were determined using CisGenome software and genome-wide analysis per-
formed using the Galaxy (http://galaxy.psu.edu/) application.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 28 July 2009; accepted 5 January 2010.
Published online 14 January 2010.
1. Schuettengruber, B., Chourrout, D., Vervoort, M., Leblanc, B. & Cavalli, G. Genome
regulation by polycomb and trithorax proteins. Cell 128, 735–745 (2007).
2. Schwartz, Y. B. & Pirrotta, V. Polycomb silencing mechanisms and the
management of genomic programmes. Nature Rev. Genet. 8, 9–22 (2007).
309
LETTERS
3. Bracken, A. P. & Helin, K. Polycomb group proteins: navigators of lineage
pathways led astray in cancer. Nature Rev. Cancer 9, 773–784 (2009).
4. Takeuchi, T., Kojima, M., Nakajima, K. & Kondo, S. jumonji gene is essential for the
neurulation and cardiac development of mouse embryos with a C3H/He
background. Mech. Dev. 86, 29–38 (1999).
5. Faust, C., Lawson, K. A., Schork, N. J., Thiel, B. & Magnuson, T. The Polycomb-
group gene eed is required for normal morphogenetic movements during
gastrulation in the mouse embryo. Development 125, 4495–4506 (1998).
6. O’Carroll, D. et al. The Polycomb-group gene Ezh2 is required for early mouse
development. Mol. Cell. Biol. 21, 4330–4336 (2001).
7. Pasini, D., Bracken, A. P., Jensen, M. R., Lazzerini Denchi, E. & Helin, K. Suz12 is
essential for mouse development and for EZH2 histone methyltransferase
activity. EMBO J. 23, 4061–4071 (2004).
8. Voncken, J. W. et al. Rnf2 (Ring1b) deficiency causes gastrulation arrest and cell
cycle inhibition. Proc. Natl Acad. Sci. USA 100, 2468–2473 (2003).
9. Boyer, L. A. et al. Polycomb complexes repress developmental regulators in
murine embryonic stem cells. Nature 441, 349–353 (2006).
10. Bracken, A. P., Dietrich, N., Pasini, D., Hansen, K. H. & Helin, K. Genome-wide
mapping of Polycomb target genes unravels their roles in cell fate transitions.
Genes Dev. 20, 1123–1136 (2006).
11. Lee, T. I. et al. Control of developmental regulators by Polycomb in human
embryonic stem cells. Cell 125, 301–313 (2006).
12. Cao, R. et al. Role of histone H3 lysine 27 methylation in Polycomb-group
silencing. Science 298, 1039–1043 (2002).
13. Cao, R., Tsukada, Y. & Zhang, Y. Role of Bmi-1 and Ring1A in H2A ubiquitylation
and Hox gene silencing. Mol. Cell 20, 845–854 (2005).
14. Kuzmichev, A., Nishioka, K., Erdjument-Bromage, H., Tempst, P. & Reinberg, D.
Histone methyltransferase activity associated with a human multiprotein complex
containing the Enhancer of Zeste protein. Genes Dev. 16, 2893–2905 (2002).
15. Czermin, B. et al. Drosophila enhancer of Zeste/ESC complexes have a histone H3
methyltransferase activity that marks chromosomal Polycomb sites. Cell 111,
185–196 (2002).
16. Müller, J. et al. Histone methyltransferase activity of a Drosophila Polycomb group
repressor complex. Cell 111, 197–208 (2002).
17. Kim, T. G., Kraus, J. C., Chen, J. & Lee, Y. JUMONJI, a critical factor for cardiac
development, functions as a transcriptional repressor. J. Biol. Chem. 278,
42247–42255 (2003).
18. Cloos, P. A., Christensen, J., Agger, K. & Helin, K. Erasing the methyl mark: histone
demethylases at the center of cellular differentiation and disease. Genes Dev. 22,
1115–1140 (2008).
19. Takeuchi, T. et al. Gene trap capture of a novel mouse gene, jumonji, required for
neural tube formation. Genes Dev. 9, 1211–1222 (1995).
20. Kim, T. G., Chen, J., Sadoshima, J. & Lee, Y. Jumonji represses atrial natriuretic
factor gene expression by inhibiting transcriptional activities of cardiac
transcription factors. Mol. Cell. Biol. 24, 10151–10160 (2004).
21. Ku, M. et al. Genomewide analysis of PRC1 and PRC2 occupancy identifies two
classes of bivalent domains. PLoS Genet. 4, e1000242 (2008).
22. Pasini, D., Bracken, A. P., Hansen, J. B., Capillo, M. & Helin, K. The Polycomb group
protein Suz12 is required for embryonic stem cell differentiation. Mol. Cell. Biol. 27,
3769–3779 (2007).
310
©2010 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 464 | 11 March 2010
23. Chamberlain, S. J., Yee, D. & Magnuson, T. Polycomb repressive complex 2 is
dispensable for maintenance of embryonic stem cell pluripotency. Stem Cells 26,
1496–1505 (2008).
24. Shen, X. et al. EZH1 mediates methylation on histone H3 lysine 27 and
complements EZH2 in maintaining stem cell identity and executing pluripotency.
Mol. Cell 32, 491–502 (2008).
25. Pasini, D. et al. Regulation of stem cell differentiation by histone
methyltransferases and demethylases. Cold Spring Harb. Symp. Quant. Biol. 73,
253–263 (2008).
26. Hansen, K. H. et al. A model for transmission of the H3K27me3 epigenetic mark.
Nature Cell Biol. 10, 1291–1300 (2008).
27. Margueron, R. et al. Role of the polycomb protein EED in the propagation of
repressive histone marks. Nature 461, 762–767 (2009).
28. Pasini, D. et al. Coordinated regulation of transcriptional repression by the RBP2
H3K4 demethylase and Polycomb-Repressive Complex 2. Genes Dev. 22,
1345–1355 (2008).
29. Mikkelsen, T. S. et al. Genome-wide maps of chromatin state in pluripotent and
lineage-committed cells. Nature 448, 553–560 (2007).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank K. Hansen for the RING1B antibody, J. Vikesaa for
bioinformatics support, and F. de Lima Alves for assistance during mass
spectrometry analyses. We thank L. Morey for critical reading of the manuscript,
and members of the Helin laboratory for discussions, technical advice and support.
D.P. was supported by a post-doctoral fellowship from the Danish Medical
Research Council, P.A.C.C. by a grant from the Benzon Foundation and J.W. by a
post-doctoral fellowship from NordForsk (Nordic Union). J.R. is a Senior Research
Fellow of the Wellcome Trust. The Wilhelm Johannsen Center is supported by the
Danish National Research Foundation and the Lundbeck Foundation. The work in
the Helin laboratory was supported by grants from the Danish National Research
Foundation, the Danish Cancer Society, the Novo Nordisk Foundation, the Danish
Medical Research Council, and the Danish Natural Science Research Council.
Author Contributions D.P. and K.H. conceived and designed the project. D.P.
performed experiments in Figs 1a, c, d, 2a, 3 and 4 and Supplementary Figs 1a, c–f,
2a, b, 3b and 4–9. P.A.C.C. designed and performed the experiments in Fig. 1b and
Supplementary Figs 1b and 3a. J.W. performed experiments in Fig. 2b–d,
Supplementary Figs 2c and 3a. J.-P.B. cloned JARID2. L.O. provided technical
support. M.B. and N.T. performed the Solexa sequencing. J.R. performed mass
spectrometry analyses. J.V.J. provided bioinformatics support. D.P. and K.H. wrote
the manuscript.
Author Information ChIP-seq data are available at the Gene Expression Omnibus
(GEO) under accession GSE19365. Reprints and permissions information is
available at www.nature.com/reprints. The authors declare competing financial
interests: details accompany the full-text HTML version of the paper at
www.nature.com/nature. Correspondence and requests for materials should be
addressed to K.H. (kristian.helin@bric.ku.dk).
doi:10.1038/nature08788
METHODS
Cloning and plasmid preparation. Human JARID2 and the different mutations
were PCR-amplified from Open Biosystems image clone 4520786, introduced
into the Gateway Entry vector pCR8/GW/TOPO (Invitrogen) following the
manufacturer’s protocol and sequenced verified. Different constructs were
subcloned in the desired vectors by Gateway technology (Invitrogen). pBABE-
Flag-HA SUZ12 was subcloned by EcoRI digestion from a previously described
pCMV-SUZ12 expression vector30. Three silent mutations in the targeting
sequence of wild-type JARID2 and D147–165 mutant were introduced by site-
directed mutagenesis to obtain an RNAi refractory JARID2 mRNA sequence.
Point mutations correspond to positions 2 and 4 (CGCRAGA, Arg 1050) and
position 10 (ATCRATA, Ile 1052) of the Jarid2 shRNA targeting sequence.
RNA interference. Stable shRNA experiments were performed as described28.
LKO.1 vectors expressing control (scrambled) or Jarid2-specific shRNA
sequences were purchased from Sigma. When required, the puromycin-selection
cassette of LKO.1 vectors was exchanged with a neomycin-selection cassette by a
BamHI-KpnI subcloning. Transient siRNA transfection was obtained with
Lipofectamine 2000 (Invitrogen) with a scrambled sequence (UNC) or Jarid2-
specific siRNA oligonucleotides purchased from Sigma. Jarid2 siRNA oligonu-
cleotide (Sigma catalogue SASI_Mm01_00093112) targets Jarid2 mRNA
sequence (NCBI accession NM_021878) around base pair (bp) position 1085.
Jarid2 shRNA LKO.1 lentiviral vector (TRCN0000096642) targets Jarid2 mRNA
with the following sequence 59-CCGCCATATAGCTAAGCCATT-39 starting
from base-pair position 3332. The Suz12 shRNA letiviral vector has been
described28.
Tissue culture. HeLa, HeLa S3, 293T, Flp-In T-REx 293 and U2OS cells were
grown in 10% FBS (Hyclone). Stable Flp-In T-REx 293 cell lines were generated
following manufacturer’s instructions. Doxycycline (200 ng ml21; Sigma) was
added for biochemical purification. ES cells (E14 clone) were cultured as
described22. Stable ES cell clones were generated by transient transfection
(Lipofectamine 2000, Invitrogen) of the expression constructs followed by selec-
tion with 2 mg ml21 of puromycin and single-cell clone isolation.
ES cells expressing Jarid2 shRNA from a LKO.1-neomycin vector were trans-
fected with pCAG-iresPRURO vectors expressing wild-type JARID2 and DARID
using Lipofectamine 2000. Cells were cultured for 48 h after transfection and
selected for a further 18 h with 5 mg ml21 puromycin. Fifty-six hours after trans-
fection samples were collected for western blot and ChIP analyses. Different ES
cell lines were induced to differentiate as previously described22.
Antibodies. Antibodies are listed in Supplementary Table 3.
Size-exclusion chromatography and immunoprecipitations. Nuclear extracts
were prepared as described28, lysed in high salt buffer (50 mM Tris-HCl, pH 7.6,
300 mM NaCl, 10% glycerol, 0.2% (v/v) Igepal CA 630) and fractionated on a
Superose 6 PC 3.2/30 (GE Healthcare). To test the stability of the Jarid2–PRC2
interaction, the salt concentration of the nuclear extracts was adjusted to the
indicated concentration and equilibrated for 2 h at 4 uC on a rotating platform.
Equilibrated extracts were incubated for 2 h with antibodies pre-coupled to
protein G Sepharose at 4 uC on a rotating platform. After six washes the immu-
noprecipitated proteins were analysed by western blots.
Recombinant proteins. Recombinant baculoviruses were generated as
described31. EZH2, EED and SUZ12 baculoviruses have been described7. Flag/
histidine-tagged JARID2 was expressed and purified by Flag-affinity chromato-
graphy essentially as described31.
©2010 Macmillan Publishers Limited. All rights reserved
Protein purifications and immunoprecipitations. Immunoprecipitations were
performed using 1 mg of cell extracts prepared in high salt buffer as described
earlier using the indicated antibodies. Complex purifications were carried out
from nuclear extracts prepared in high salt buffer as described earlier. Protein
extracts (100–200 mg) were incubated overnight at 4 uC on a rotating platform
with 200 ml packed M2-Flag agarose beads. Beads were washed six times in
high salt buffer and proteins were eluted with 0.5 mg ml21 of a Flag peptide
(DYKDDDDK). When indicated, Flag-eluted proteins were incubated with
200 ml of packed anti-haemagglutinin (12CA5) cross-linked protein-A
Sepharose beads overnight at 4 uC on a rotating platform, washed six times in
high salt buffer, and eluted with 0.5 mg ml21 of a haemagglutinin peptide
(YPYDVPDYA). Eluted proteins were analysed by nanoelectrospray tandem
mass spectrometry32.
ChIP and ChIP-sequencing. ChIP analyses were performed as described28. For
ChIP-seq, the DNA from three independent ChIP experiments was pooled and
prepared for sequencing with the Illumina ChIPSeq Sample Prep kit. DNA
libraries were quantified using a fluorometer (Fluoroscan) and diluted to
5–10 pM. Diluted libraries were used directly for cluster generation and sequencing
analysis using the Genome Analyser II (Illumina) following the protocol of the
manufacturer. Base calling and mapping to the mouse genome (mm9, July 2007
release) of the 36-bp sequences were done using Illumina’s analysis Pipeline v1.4
allowing up to two mismatches in the first 32 bases of a read. The enriched genomic
regions from our ChIP-seq experiments and the sequencing results from previ-
ously published ChIP-seq data sets were determined using CisGenome software33
by two-sample analysis using the IgG ChIP-signal as a negative control.
CisGenome two sample analysis determines regions where the ChIP signals are
enriched relative to the control ChIP (IgG) using a conditional binomial model
that involves 100-bp windows passing a 0.1 FDR cut-off. A detailed description of
CisGenome statistical analysis has been described33.
We have annotated 22,064 promoter regions as the 65 kb chromosomal regions
from the transcription start site of the refseq gene data set (mm9) extracted from the
UCSC database (http://genome.ucsc.edu) using Galaxy application (http://galaxy.
psu.edu). Target genes were defined as the promoter regions that overlapped with at
least one enriched genomic region in the different ChIP-seq data sets. All location
analyses were performed using the Galaxy application (http://galaxy.psu.edu).
P values were determined by chi-square test. The intensity of the ChIP enriched
regions is defined by the total number of reads included in each peak.
Real-time qPCR. qPCR analysis for expression and ChIP assays were carried out
using Applied Biosystems Real-Time PCR Systems. Primer sequences are listed
in Supplementary Table 3.
Gene expression analysis. Total RNA was extracted with the Qiagen RNAesy
Plus RNA extraction kit. RNA was hybridized on mouse Gene 1.0 ST arrays by
the RH Microarray Center at Rigshospitalet, Copenhagen, following Affymetrix
procedures and analysis. RNA normalized signals at day 0 were used as a baseline
for comparison.
30. Villa, R. et al. Role of the polycomb repressive complex 2 in acute promyelocytic
leukemia. Cancer Cell 11, 513–525 (2007).
31. Christensen, J. et al. RBP2 belongs to a family of demethylases, specific for tri-and
dimethylated lysine 4 on histone 3. Cell 128, 1063–1076 (2007).
32. Wilm, M. et al. Femtomole sequencing of proteins from polyacrylamide gels by
nano-electrospray mass spectrometry. Nature 379, 466–469 (1996).
33. Ji, H. et al. An integrated software system for analyzing ChIP-chip and ChIP-seq
data. Nature Biotechnol. 26, 1293–1300 (2008).
 CAREERS
PROSPECTS
Speak up
Peter Fiske argues that too many young scientists adopt a passive voice, to the detriment of their careers.
When I was in graduate school studying
geology and environmental sciences, many
of my professors insisted that we students
write our manuscripts in the passive voice:
“This was done” rather than “I did this”. They
reasoned that removing the agent from the
description of the action lent an objective
tone. As scientists, we stood apart from our
work and encouraged others to critique it
(rather than us).
Today, teachers are much more likely to
advocate use of the active voice
A. RUGGIERI/IMAGES.COM/CORBIS
in manuscripts. In general, the
active voice is more readable
and engaging, which can be
particularly helpful in light of
the sometimes turgid prose of
scientific papers.
In the years since I moved from
academic science to business,
young scientists have started to
adopt a more active tone in their
manuscripts. Still, the culture of
science still seems to encourage
a ‘passive voice’ in much of the
rest of their careers. Peers and
mentors often imply that they
should remain at the bench
rather than actively reach out to
potential sponsors, supporters
and collaborators. Young
scientists think that doing good
science will be enough to advance
their careers. “If you’re the best,
you’ll get a job” is the assurance
they receive.
Modesty or arrogance?
In life, as in writing, a passive
approach may be intended to
suggest a degree of objectivity.
Avoiding the appearance of ‘self-
promotion’ may seem desirable
for establishing credibility. But failure to make
any attempt to advocate for oneself and one’s
abilities can be misinterpreted as indifference,
even arrogance.
Collaboration is a case in point. Working
with researchers in other departments
and institutions is by far the most effective
way for young scientists to advance their
careers. Yet few PhD and postdoc advisers
encourage their charges to pursue any such
partnerships. Understandably, PhD advisers
may worry that their students will lose focus
by working outside their research group. But
some resist simply because letting students
engage in outside work would mean fewer
hands in the lab. PhD and postdoc advisers
have lots of influence, and this can lead to a
312
pattern of passivity among young scientists.
As a result, few of them pursue such
collaborations on their own.
This ‘passive voice’ also arises when
engaging in professional networking. Most
scientists recognize that networking with
colleagues has a crucial role in career
development. But rather than proactively
reaching out to others, many wait to
receive such gestures. Networking is both
acceptable and valued in the business world,
but many scientists and engineers view it
as ‘schmoozing’. Those who try to forge
close ties to their administration or financial
sponsors may earn a reputation for being
‘political’ among jealous colleagues. Younger
scientists witness this pathological behaviour
and quickly absorb the lesson: ‘true’ scientists
are above networking.
Although scientists may increasingly
write their papers in the active voice, the
way they promote their work often remains
passive. Scientists expect their publications
to communicate for them. A large body of
scholarly work certainly confers a degree of
authority and knowledge. But even the best-
written papers never completely capture the
passion and insight that led to their creation.
© 2010 Macmillan Publishers Limited. All rights reserved
NATURE|Vol 464|11 March 2010
Scientists must communicate about
their work — to other scientists, sponsors
of their research and the general public.
Active communication means more than
merely accepting invitations to give talks
at other institutions. An ‘active voice’
in communication means searching for
opportunities to give talks and lectures
— and seeking audiences that are outside
one’s immediate sphere of scientific influence
at, for example, science museums or local
civic organizations.
Public perceptions
Scientists may also have passive
attitudes towards political
advocacy. Many scientists are
incredulous at how little the
general public knows about
science and technology.
A survey carried out by the Pew
Research Center for the People
and the Press in Washington DC
last July found that 85% of US
scientists interviewed saw the
public’s lack of knowledge as a
major problem for science. Half of
them fault the public for having a
poor understanding of the pace of
scientific discovery.
But scientists do little to
address the gap in understanding.
Most think that their successes
in the lab are manifestly evident,
making education about the
value of their work unnecessary.
Few ever communicate with
their elected officials. With the
public footing most of the bill, this
misguided belief seems naive and
undermines those who campaign
for more funding.
In a global recession, taking a
passive career approach poses an increasing
disadvantage. Stronger competition for
a finite number of research positions will
favour those who combine outstanding work
with an ability to engage their communities
and constituencies. Excellent work is a
prerequisite for career progress, but is
not sufficient by itself. Broadcasting one’s
accomplishments and exercising the ‘active
voice’ in all aspects of one’s work is the best
way to earn notice, gain recognition and make
the public at large aware of the value of the
scientific enterprise. ■
Peter Fiske is chief technology officer of PAX
Water Technologies in San Rafael, California,
and author of Put Your Science to Work
(American Geophysical Union, 2001).
NATURE|Vol 464|11 March 2010
Lidia Brito, Mozambique’s former science minister,
Q&A now heads the science-policy division at the
United Nations Educational, Scientific and Cultural
Organization (UNESCO) in Paris.
What made you decide to
pursue science?
When I was 14 years old,
Mozambique became an
independent country. It was
a life-changing moment that
motivated me to contribute
by pursuing my interests in
chemistry — at the time,
I thought in industry. But
when I was 15, a teacher
shortage led to the closing
of the last years of high
school in all provinces. I was
fortunate to be able to go to a
special pre-university school
created by the government
in Maputo, the capital city.
At first I was encouraged to
study agricultural sciences,
but I did not like that at
all because I couldn’t see
a way to combine it with
chemical engineering. In my
second year, I jumped at the
opportunity to be among the
first class of foresters trained
in Mozambique under a new
forestry degree programme.
Did you immediately follow
an academic track?
No. After I finished my
undergraduate degree
at Eduardo Mondlane
University in Maputo, I was
recruited to stay and teach
there, but I was required
to work in forestry before I
could begin my academic
career. After getting some
field experience introducing
new sawmilling techniques to
improve the quality of wood
products, I took some time
off for my husband to focus
on his career before I finished
my MSc in wood sciences
and PhD in forest sciences in
the United States. I returned
to Mozambique and slowly
built an academic career
conducting research related
to biomass production and
forestry management.
Did you have a career
turning point?
Yes. The turning point in
my professional life was
developing leadership skills
as head of the department
of forestry. I learned that to
be a good manager I should
always approach my goals
with a critical eye, I should
ask the hard questions to
determine whether goals
are being reached and, if
necessary, redefine a policy
mission. That led me to a job
as deputy rector of academic
affairs at the university and a
completely different path.
Why do you think you are
effective in policy roles?
I have had the great fortune
of having worked with, and
learned from, wonderful
teams of people. I’ve
combined my life experiences
to sharpen a critical eye. Early
in my career I did field work
with industry while I was
still linked to the university
doing teaching
and research. As
a result, I got a
more practical
understanding of
what is at stake at
the production
level. I developed
the capacity to
simply ask the right questions,
which is very important
when you work in policy. In
the end, policy-makers must
create feedback mechanisms,
be open to change and look
at things with a critical eye. I
took all those lessons with me
when, along with colleagues,
I helped to establish
Mozambique’s Ministry of
Higher Education, Science
and Technology.
What job did you find the
most rewarding?
It’s difficult to choose. I
enjoyed life as an academic
working with students, but
working in the Mozambique
science ministry was very
special because it was a new
ministry. The team was small
but we were able to develop
new policies and mobilize
resources. It is very gratifying
to see that many of our
initial projects, such as the
establishment of the National
© 2010 Macmillan Publishers Limited. All rights reserved
Research Fund, a funding
agency, and programmes
such as the annual innovation
competitions or Science Fair,
are now flourishing pieces of
a much larger ministry.
What research is most
needed to inform future
science policy?
We need to better understand
how policy drives social
change. For example, we need
to know whether and how
certain policies really promote
basic values in society, such as
equality, inclusivity and access
to resources.
What is your top scientific
goal at UNESCO?
UNESCO has an important
mandate to use its focus on
science, communication
and culture to
offer relevant
policy advice.
Our challenge
is to identify
the triggers
and drivers of
development
— for example,
education initiatives — that
can really make a difference.
Clearly, the scientific
challenge for UNESCO is to
do more research on science-
policy development. For
example, we were recently
involved in a large research
project in Turkey and Malta
to study the effects of science
policies on equality in society.
We need to promote more
of this kind of ‘think tank’
approach to science-policy
development — preferably
through partnerships
with organizations that do
complementary work such as
the United States Agency for
International Development.
What was the best piece of
advice you were given?
People who are not afraid of
thinking differently from the
rest are the ones that come up
with alternative solutions to
old problems. ■ of the four programmes is 11 May and
Interview by Virginia Gewin
CAREERS
IN BRIEF
Better prospects
Demand for both permanent and contract
researchers is expected to increase this
year in the United Kingdom across
several sectors that employ scientists,
according to a survey of employers.
Slough-based recruitment company SRG
found that 29.5% of respondents expect
to boost numbers of permanent scientific
staff in 2010. Fields surveyed included
pharmaceuticals and biotechnology, oil
and petrochemicals, health care, research
institutes and government agencies.
About half of the employers in the oil, gas
and petrochemicals sector expect to have
open permanent research positions, and
none is forecasting cutbacks. More than
550 employers completed the survey.
Salary freeze
In 2009, some two-thirds of private and
public US universities gave no salary
increases to their chief executives,
academic deans and other senior
executives, and 2.1% cut salaries, according
to a survey released on 22 February. The
College and University Professional
Association for Human Resources,
based in Knoxville, Tennessee, polled
1,280 institutions for its Administrative
Compensation Survey Report.
Nevertheless, the median base salary
for the chief executive of a doctorate-
granting institution was up 7% from 2008
to US$375,000 in 2009. Respondents
predicted restricted hiring for the fiscal
year 2010, with fewer than 1% expecting
“significantly more” positions to be filled.
Wellcome translation
The Wellcome Trust is planning to
provide funding for 100 new PhD
positions in biomedical translational
research in UK and Irish universities.
The charitable trust, based in the United
Kingdom, aims to create four PhD
programmes, each lasting five years and
funding five new graduate students per
year, for research in areas highlighted by
the trust this year as key to its ten-year
strategic plan — applying genomics
and genetics to understanding disease,
brain and mental disorders, infectious
disease, ageing and chronic disease, and
links among environment, nutrition and
health. The closing date for preliminary
applications to create and direct one
winners will be announced in December.
313
FUTURES
The Omniplus Ultra
You can’t live without it.
Paul Di Filippo
Everyone wants an Omniplus Ultra, and I
am not immune to the urge. But of course
they are almost impossible to purchase —
for love or money.
Since their debut nine months ago at
the annual Consumer Electronics Show,
more than 40 million units have been sold
worldwide, exhausting the initial stockpile
but barely sating a fraction of consumer
demand. The Chinese factories that pro-
duce the Omniplus Ultra are tooling up as
fast as possible to make more, but retail-
ers cannot guarantee delivery any sooner
than six months. On eBay, each available
Omniplus Ultra, with a MSRP of $749.99,
sells for upwards of $5,000.
OmninfoPotent Corporation, the enig-
matic firm behind the Omniplus Ultra, has
leapt to the top of the NASDAQ exchange.
Its reclusive founders, Pine Martin and
Sheeda Waxwing, have vaulted into
the lower ranks of the Forbes 400.
Sales of the device are being cred-
ited with jump-starting the ailing
economy almost single-handedly.
The ad campaign for the Omni-
plus Ultra has already won six
Clios. The catchy theme music, O U
Kidz by the Black Eyed Peas, and the
images of average people of every
race, age, gender, nationality and creed
using their Omniplus Ultras to navigate
a plethora of life situations ranging from
sweetly comic to upliftingly tragic have
generated their own fan clubs, YouTube
mash-ups and punchlines for late-night
comedians. Allusions to the Omniplus
Ultra, as well as its invocation in meta-
phors, similes, rants, raves, jeremiads and
paeans, fill watercooler conversations and
the printed pages of the world’s magazines,
newspapers and blogs. The first instant
book on the Omniplus Ultra — Uber-
power! by Thomas Friedman and Charles
Stross — is due out any day.
I myself do not know anyone who actually
FUTURES
owns an Omniplus Ultra, but I’m dying to
see and handle one. But even 40 million
units, distributed across 7 billion people,
means that there is only one Omniplus
Ultra for every 175 citizens. Of course, the
gadgets are not seeded evenly around the
planet. They are concentrated in the hands
of relatively wealthy and elite consumers
and early adopters: circles I do not really
travel in, given my job in a Staples ware-
house and a set of friends whose familiarity
with the latest products of Silicon Valley
316
© 2010 Macmillan Publishers Limited. All rights reserved
generally extends no further than their TV
remotes.
So I must content myself with studying
the advertisements and the gadget-porn
reviews. These can’t say enough about the
life-changing capabilities of the Omniplus
Ultra, its potential to shatter all old para-
digms and to literally remake the world.
Publisher’s Weekly: After five centuries,
the printed book has found its worthi-
est successor in the Omniplus Ultra. The
future of reading is safely triumphant.
The Huffington Post: Opens new chan-
nels for the spread of democracy.
Boing-Boing: Coolest gadget since the
iPhone! The cold-laser picoprojector alone
is worth the cost.
Car and Driver: Jack the Omniplus Ultra
into your dash’s USB port and driving will
never be the same!
Entertainment Weekly: If you can’t
download your favourite show onto your
Omniplus Ultra, it’s not worth watching.
Variety: First flicks helmed with the
Omniplus Ultra to hit big screens soon!
Aerospace & Defense Industry Review:
Guaranteed to be standard equipment for
all future warriors.
Mother Jones: The Omniplus Ultra is the
greenest invention since the Whole Earth
Catalog.
BusinessWeek: Every chief executive will
benefit from having an Omniplus Ultra to
hand — and anyone without one will watch
competitors eat their lunch.
Rolling Stone: Elvis. The Beatles. The Sex
Pistols. The Omniplus Ultra. The sequence
is complete at last.
The more such talk I read, the longer
I drool over pictures of the sleekly tactile
Omniplus Ultra, with its customizable sexy
NATURE|Vol 464|11 March 2010
skins and ergonomically perfect controls,
and the more I lust to own one. Although
nothing in my condition has really changed,
and although I have enough money, love
and security, my life feels incomplete and
empty without an Omniplus Ultra.
But there was just no way for me to get
my hands on one.
Until I saw my boss’s boss’s boss walk
through the warehouse carrying one.
Then and there, I knew what I had to do.
As a low-level employee, I certainly
could not jump several levels of manage-
ment and directly approach my boss’s
boss’s boss and ask to fondle his Omniplus
Ultra. But I had a scheme.
It took me six frustrating weeks, but
at last I managed it. In a series of furtive
unauthorized forays into executive terri-
JACEY
tory, I caught the lucky Omniplus Ultra
owner in a lavatory break with his prized
possession carelessly left behind on
his desk.
That’s when I pulled the fire alarm.
While everyone else rushed outside,
I darted into the guy’s office, snatched
his Omniplus Ultra off the desk, and
sank down behind the furniture in the
knee well, out of sight.
With trembling hands I sought to
shuffle aside the protective wings of
the device, utilizing all the instructions
I had lovingly memorized, and expose
its intimate control and display surfaces
to my wanton gaze and lewd touch.
But I was doing something wrong! The
expected blossoming failed to happen.
Instead, after some fumbling, the unit
split open like a simple styrofoam clam-
shell container full of leftovers.
The interior gaped utterly vacant, except
for a simple piece of printed cardboard.
Dumbfounded, I removed the card-
board and read the message.
Dear Consumer: the Omniplus Ultra
is not what you need. You are already
everything you thought it could do. Pass
this message on as widely as you see fit.
Or not. Hopefully yours, Pine Martin and
Sheeda Waxwing, for the OmninfoPotent
Corporation.
I put the card back inside, resealed the
Omniplus Ultra, dropped it with a dull
thud on the desk, and joined all my peers
outside, waiting to resume our lives. ■
Paul Di Filippo’s new novel, Roadside
Bodhisattva, will be available in spring 2010.
He continues to review for various venues.
Join the discussion of Futures in Nature at
go.nature.com/QMAm2a