Sei sulla pagina 1di 138

1028 & 1074

CONTENTS
4 M A RC H 2 0 1 6 V O LU M E 3 5 1 I S S U E 6 2 7 7
Brain circuitry
and generous behavior

FEATURES ON THE COVER


1018 SLOW BURN Japanese media
Five years after the meltdowns tour the Fukushima
at the Fukushima atomic plant, Daiichi Nuclear
evacuees are weighing the prospect Power Station
of living near a nuclear disaster zone (Unit 4 reactor
for decades to come building shown
By D. Normile here). On 11 March,
Japan marks the
1020 TRIAL BY MELTDOWN fth anniversary
of the earthquake and tsunami that
At Fukushima, robots performed
triggered meltdowns at this site. Tokyo
poorly at rst but are now critical
Electric Power Company is in the early
players in the cleanup stages of a decades-long, $9 billion
By T. Hornyak

Downloaded from on March 3, 2016


decommissioning efort that is spurring
advances in robotics. Area residents,
1022 EPIDEMIC OF FEAR meanwhile, are wrestling with the
A bumper crop of thyroid abnormalities, possible health efects of long-term
including cancer, in Fukushima children exposure to low doses of radiation.
has perplexed scientists and alarmed See page 1018. Photo: Tomohiro
locals By D. Normile Ohsumi/Bloomberg via Getty Images

NEWS INSIGHTS 1030 SURPRISED BY SELECTIVITY


A bifunctional catalyst enables olen
synthesis from carbon monoxide and
hydrogen at high selectivity
IN BRIEF PERSPECTIVES By K. P. de Jong
REPORT P. 1065
1009 News at a glance 1024 LIBERATING FIELD SCIENCE
SAMPLES AND DATA 1032 AN RNA TWIST TO TH17 CELLS
IN DEPTH Promote reproducibility by moving A long noncoding RNA and RNA helicase
beyond available upon request
1012 FAST RADIO BURSTS TEASE constitute a new layer of T cell control
By M. McNutt et al.
ASTRONOMERS By M. K. Atianand and K. A. Fitzgerald
EDITORIAL P. 1005; EDITORS NOTE P. 1007
Conicting results suggest the potent
blasts come in two avors BOOKS ET AL.
By G. Schilling
1026 ELECTRONS GO WITH THE FLOW
IN EXOTIC MATERIAL SYSTEMS 1033 WHEN WE ARE NO MORE
Electronic hydrodynamic ow By A. S. Rumsey,
1013 PATIENT ZERO NO MORE
making electrons ow like a uidhas reviewed by C. B. Anderson
Sleuthing claries HIVs history
By J. Cohen been observed By J. Zaanen
REPORTS PP. 1055, 1058, & 1061 1034 STRANGE GLOW
By T. J. Jorgensen, reviewed by A. N. Creager
1014 SURVEY FRAUD TEST SPARKS BATTLE
Pew Research Center challenges 1028 WIRING THE ALTRUISTIC BRAIN
Communication between brain LETTERS
statistical test By J. Bohannon
regions uncovers hidden motives 1036 SEEING THE GRASSLANDS
1015 SCIENTISTS TO DRILL INTO for generous behavior By S. Gluth THROUGH THE TREES
DINOSAUR-KILLING BLAST and L. Fontanesi By S. DeWitt et al.
Cores from Chicxulub crater could REPORT P. 1074
illuminate how life returned after the 1036 RESPONSE
cataclysm By E. Hand 1029 A COPY-AND-PASTE GENE By W. Bond
REGULATORY NETWORK
1017 NSF MAKES A NEW BID Transposable elements provide a 1037 GOVERNMENT: PLAN FOR
TO BOOST DIVERSITY ready-made route to regulate complex ECOSYSTEM SERVICES
Initiative aims for fresh answers, gene networks By L. N. Joppa et al.
but researchers are still framing the By V. J. Lynch
questions By J. Mervis REPORT P. 1083 1037 TECHNICAL COMMENT ABSTRACT

1000 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
1078 1094
Targeting Ebola Cellular heterogeneity

1048 IMMUNOLOGY 1074 HUMAN ALTRUISM


RESEARCH Visualizing antibody afnity maturation
in germinal centers J. M. J. Tas et al.
The brains functional network
architecture reveals human motives
G. Hein et al.
IN BRIEF R EPORTS PERSPECTIVE P. 1028

1038 From Science and other journals ELECTRON TRANSPORT


1055 Negative local resistance caused 1078 EBOLA VIRUS
Isolation of potent neutralizing
by viscous electron backow in
RESEARCH ARTICLES antibodies from a survivor of the
graphene D. A. Bandurin et al.
2014 Ebola virus outbreak
1041 NEURAL COMPUTATION 1058 Observation of the Dirac uid and Z. A. Bornholdt et al.
Spiking neurons can discover the breakdown of the Wiedemann-
predictive features by aggregate-label Franz law in graphene J. Crossno 1083 IMMUNOGENOMICS
learning R. Gtig et al. Regulatory evolution of innate immunity
1061 Evidence for hydrodynamic electron through co-option of endogenous
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT:
ow in PdCoO2 P. J. W. Moll et al. retroviruses E. B. Chuong et al.
dx.doi.org/10.1126/science.aab4113
PERSPECTIVE P. 1029
PERSPECTIVE P. 1026

1042 CANCER BIOLOGY 1087 GENE EXPRESSION


1065 CATALYSIS
Activation of PKA leads to Expression homeostasis during DNA
Selective conversion of syngas to light
mesenchymal-to-epithelial transition olens F. Jiao et al. replication Y. Voichek et al.
and loss of tumor-initiating ability PERSPECTIVE P. 1030
D. R. Pattabiraman et al. 1090 TRANSCRIPTION
1068 QUANTUM COMPUTING Multiplexed protein-DNA cross-linking:
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT:
Realization of a scalable Shor algorithm Scrunching in transcription start site
dx.doi.org/10.1126/science.aad3680
T. Monz et al. selection J. T. Winkelman et al.

1043 STRUCTURAL BIOLOGY 1094 DNA REPAIR


1071 FLEXIBLE ELECTRONICS
Cryo-EM structure of a native, Highly stretchable electroluminescent Stochastic activation of a DNA damage
fully glycosylated, cleaved HIV-1 skin for optical signaling and tactile response causes cell-to-cell mutation
envelope trimer sensing C. Larson et al. rate variation S. Uphof et al.
J. H. Lee et al. PODCAST

DEPARTMENTS
1005 EDITORIAL
#IAmAResearchParasite
By Marcia McNutt
EDITORS NOTE P. 1007; PERSPECTIVE P. 1024

1007 EDITORS NOTE


Data sharing
By Marcia McNutt
EDITORIAL P. 1005; PERSPECTIVE P. 1024

1106 WORKING LIFE


Making a game of science
By Elizabeth Pennisi
1030
& 1065 Science Staf ............................................1002
New Products ............................................ 1101
Science Careers ....................................... 1102

SCIENCE (ISSN 0036-8075) is published weekly on Friday, except the last week in December, by the American Association for the Advancement of Science, 1200 New York Avenue, NW, Washington, DC 20005. Periodicals
mail postage (publication No. 484460) paid at Washington, DC, and additional mailing ofces. Copyright 2016 by the American Association for the Advancement of Science. The title SCIENCE is a registered trademark of the AAAS.
Domestic individual membership and subscription (51 issues): $165 ($74 allocated to subscription). Domestic institutional subscription (51 issues): $1522; Foreign postage extra: Mexico, Caribbean (surface mail) $55; other countries (air
assist delivery) $89. First class, airmail, student, and emeritus rates on request. Canadian rates with GST available upon request, GST #1254 88122. Publications Mail Agreement Number 1069624. Printed in the U.S.A.
Change of address: Allow 4 weeks, giving old and new addresses and 8-digit account number. Postmaster: Send change of address to AAAS, P.O. Box 96178, Washington, DC 200906178. Single-copy sales: $15.00 current issue,
$20.00 back issue prepaid includes surface postage; bulk rates on request. Authorization to photocopy material for internal or personal use under circumstances not falling within the fair use provisions of the Copyright Act is granted by
AAAS to libraries and other users registered with the Copyright Clearance Center (CCC) Transactional Reporting Service, provided that $35.00 per article is paid directly to CCC, 222 Rosewood Drive, Danvers, MA 01923. The identifcation
code for Science is 0036-8075. Science is indexed in the Readers Guide to Periodical Literature and in several specialized indexes.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1001


Published by AAAS
SENIOR EDITORIAL BOARD
1200 New York Avenue, NW, Washington, DC 20005 Gary King, Harvard University, Susan M. Rosenberg, Baylor College of Medicine,
Ali Shilatifard, Northwestern University Feinberg School of Medicine
Bateman House, 82-88 Hills Road, Cambridge, UK CB2 1LQ

BOARD OF REVIEWING EDITORS (Statistics board members indicated with S)


Adriano Aguzzi, U. Hospital Zrich Etienne Koechlin, Ecole Normale Suprieure
Takuzo Aida, U. of Tokyo Alexander Kolodkin, Johns Hopkins U.
Editor-in-Chief Marcia McNutt Leslie Aiello, Wenner-Gren Foundation Thomas Langer, U. of Cologne
Executive Editor Monica M. Bradford News Editor Tim Appenzeller Judith Allen, U. of Edinburgh Mitchell A. Lazar, U. of Pennsylvania
Sonia Altizer, U. of Georgia David Lazer, Harvard U.
Managing Editor, Research Journals Katrina L. Kelner Sebastian Amigorena, Institut Curie Thomas Lecuit, IBDM
Kathryn Anderson, Memorial Virginia Lee, U. of Pennsylvania
Deputy Editors Barbara R. Jasny, Andrew M. Sugden(UK), Valda J. Vinson, Jake S. Yeston Sloan-Kettering Cancer Center Stanley Lemon, U. of North Carolina at Chapel Hill
Meinrat O. Andreae, Max-Planck Inst. Mainz Ottoline Leyser, Cambridge U.
Paola Arlotta, Harvard U.
Wendell Lim, U.C. San Francisco
Research and Insights Johan Auwerx, EPFL
Marcia C. Linn, U. of California, Berkeley
SR. EDITORS Caroline Ash(UK), Gilbert J. Chin, Lisa D. Chong, Julia Fahrenkamp-Uppenbrink(UK), Pamela J. Hines, Stella M. David Awschalom, U. of Chicago
Jianguo Liu, Michigan State U.
Hurtley(UK), Paula A. Kiberstis, Marc S. Lavine(Canada), Kristen L. Mueller, Ian S. Osborne(UK), Beverly A. Purnell, L. Bryan Clare Baker, University of Cambridge
Jordi Bascompte, University of Zurich Luis Liz-Marzan, CIC biomaGUNE
Ray, Guy Riddihough, H. Jesse Smith, Jelena Stajic, Peter Stern(UK), Phillip D. Szuromi, Sacha Vignieri, Brad Wible, Nicholas S. Facundo Batista, London Research Inst. Jonathan Losos, Harvard U.
Wigginton, Laura M. Zahn ASSOCIATE EDITORS Brent Grocholski, Keith T. Smith ASSOCIATE BOOK REVIEW EDITOR Valerie B. Thompson Ray H. Baughman, U. of Texas, Dallas Ke Lu, Chinese Acad. of Sciences
ASSOCIATE LETTERS EDITOR Jennifer Sills CHIEF CONTENT PRODUCTION EDITOR Cara Tate SR. CONTENT PRODUCTION EDITORS Harry Jach, David Baum, U. of Wisconsin Christian Lscher, U. of Geneva
Lauren Kmec CONTENT PRODUCTION EDITORS Jeffrey E. Cook, Chris Filiatreau, Cynthia Howe, Barbara P. Ordway, Catherine Wolner SR. Carlo Beenakker, Leiden U. Laura Machesky, CRUK Beatson Inst.
Kamran Behnia, ESPCI-ParisTech for Cancer Research
EDITORIAL COORDINATORS Carolyn Kyle, Beverly Shields EDITORIAL COORDINATORS Joi S. Granger, Lisa Johnson, Anita Wynn PUBLICATIONS
Yasmine Belkaid, NIAID, NIH Anne Magurran, U. of St. Andrews
ASSISTANTS Aneera Dobbins, Jeffrey Hearn, Dona Mathieu, Le-Toya Mayne Flood, Shannon McMahon, Scott Miller, Caitlyn Phillips, Oscar Marin, CSIC & U. Miguel Hernndez
Philip Benfey, Duke U.
Jerry Richardson, Rachel Roberts(UK), Alice Whaley(UK), Brian White EXECUTIVE ASSISTANT Anna Bashkirova ADMINISTRATIVE SUPPORT May Berenbaum, U. of Illinois Charles Marshall, U. of California, Berkeley
Janet Clements(UK), Lizanne Newton(UK), Maryrose Madrid, John Wood(UK) Gabriele Bergers, U. of California, San Francisco C. Robertson McClung, Dartmouth College
Bradley Bernstein, Massachusettes General Hospital Graham Medley, U. of Warwick
Peer Bork, EMBL Tom Misteli, NCI
News Bernard Bourdon, Ecole Normale Suprieure de Lyon Yasushi Miyashita, U. of Tokyo
NEWS MANAGING EDITOR John Travis INTERNATIONAL EDITOR Richard Stone DEPUTY NEWS EDITORS Daniel Clery(UK), Robert Coontz, Chris Bowler, cole Normale Suprieure Mary Ann Moran, U. of Georgia
Ian Boyd, U. of St. Andrews
Elizabeth Culotta, David Grimm, David Malakoff, Leslie Roberts CONTRIBUING EDITOR Martin Enserink(Europe) SR. CORRESPONDENTS Emily Brodsky, U. of California, Santa Cruz
Richard Morris, U. of Edinburgh
Jeffrey Mervis, Elizabeth Pennisi NEWS WRITERS Adrian Cho, Jon Cohen, Jennifer Couzin-Frankel, Carolyn Gramling, Eric Hand, Alison Motsinger-Reif, NC State U. (S)
Ron Brookmeyer, U. of California Los Angeles (S) Thomas Murray, The Hastings Center
Jocelyn Kaiser, Catherine Matacic, Kelly Servick, Robert F. Service, Erik Stokstad(Cambridge, UK), Emily Underwood INTERNS Christian Bchel, U. Hamburg-Eppendorf
Daniel Neumark, U. of California, Berkeley
Hanae Armitage, Patrick Monahan, Nala Rogers CONTRIBUTING CORRESPONDENTS Michael Balter(Paris), John Bohannon, Warren Joseph A. Burns, Cornell U.
Kitty Nijmeijer, U. of Twente
Cornwall, Ann Gibbons, Mara Hvistendahl, Sam Kean, Eli Kintisch, Kai Kupferschmidt(Berlin), Andrew Lawler, Christina Carter Tribley Butts, U. of California, Irvine
Pr Nordlund, Karolinska Inst.
Gyorgy Buzsaki, New York U. School of Medicine
Larson(Beijing), Mitch Leslie, Charles C. Mann, Eliot Marshall, Virginia Morell, Dennis Normile(Shanghai), Heather Pringle, Tania Blanche Capel, Duke U. Helga Nowotny, European Research Advisory Board
Rabesandratana(London), Gretchen Vogel(Berlin), Lizzie Wade(Mexico City) CAREERS Donisha Adams, Rachel Bernstein(Editor) Mats Carlsson, U. of Oslo Ben Olken, MIT
COPY EDITORS Julia Cole, Dorie Chevlen, Jennifer Levin (Chief) ADMINISTRATIVE SUPPORT Jessica Adams Ib Chorkendorff, U. of Denmark Joe Orenstein, U. of California
David Clapham, Childrens Hospital Boston Berkeley & Lawrence Berkeley National Lab
Joel Cohen, Rockefeller U., Columbia U. Harry Orr, U. of Minnesota
Pilar Ossorio, U. of Wisconsin
Executive Publisher Rush D. Holt James J. Collins, MIT
Andrew Oswald, U. of Warwick
Robert Cook-Deegan, Duke U.
Interim Publisher Bill Moran Chief Digital Media Officer Rob Covey Lisa Coussens, Oregon Health & Science U. Steve Palumbi, Stanford U.
Alan Cowman, Walter & Eliza Hall Inst. Jane Parker, Max-Planck Inst.
BUSINESS OPERATIONS AND PORTFOLIO MANAGEMENT DIRECTOR Sarah Whalen PRODUCT DEVELOPMENT DIRECTOR Will Schweitzer PRODUCT Robert H. Crabtree, Yale U. of Plant Breeding Research
DEVELOPMENT ASSOCIATE Hal Moore BUSINESS SYSTEMS AND FINANCIAL ANALYSIS DIRECTOR Randy Yi MANAGER OF FULFILLMENT SYSTEMS Roberta Croce, Vrije Universiteit Giovanni Parmigiani, Dana-Farber Cancer Inst. (S)
Neal Hawkins SYSTEMS ANALYST Nicole Mehmedovich ASSISTANT DIRECTOR, BUSINESS OPERATIONS Eric Knott MANAGER, BUSINESS Janet Currie, Princeton U. John H. J. Petrini, Memorial Sloan-Kettering
OPERATIONS Jessica Tierney BUSINESS ANALYST Cory Lipman FINANCIAL ANALYST Robert Clark RIGHTS AND PERMISSIONS ASSISTANT Jeff L. Dangl, U. of North Carolina Cancer Center
Tom Daniel, U. of Washington Samuel Pfaff, Salk Institute for Biological Studies
DIRECTOR Emilie David PERMISSIONS ASSOCIATE Elizabeth Sandler RIGHTS, CONTRACTS, AND LICENSING ASSOCIATE Lili Kiser
Frans de Waal, Emory U. Joshua Plotkin, U. of Pennsylvania
MARKETING DIRECTOR Elise Swinehart ASSOCIATE DIRECTOR OF ACQUISITION AND RETENTION Julianne Wielga MARKETING ASSOCIATE Stanislas Dehaene, Collge de France Albert Polman, FOM Institute AMOLF
Robert Desimone, MIT
Elizabeth Sattler SR. MARKETING EXECUTIVE Jennifer Reeves ASSOCIATE DIRECTOR, CREATIVE SERVICES Tzeitel Sorrosa ART ASSOCIATE Claude Desplan, New York U.
Philippe Poulin, CNRS
Seil Lee JR. ART ASSOCIATE Kim Huynh ASSISTANT COMMERCIAL EDITOR Selby Frame MARKETING PROJECT MANAGER Angelissa Jonathan Pritchard, Stanford U.
Dennis Discher, U. of Pennsylvania
McArthur PROGRAM DIRECTOR, AAAS MEMBER CENTRAL Peggy Mihelich FULFILLMENT SYSTEMS AND OPERATIONS membership@aaas.org David Randall, Colorado State U.
Gerald W. Dorn II, Washington U. School of Medicine
Felix Rey, Institut Pasteur
MANAGER, MEMBER SERVICES Pat Butler SPECIALISTS Terrance Morrison, Latasha Russell MANAGER, DATA ENTRY Mickie Napoleoni Jennifer A. Doudna, U. of California, Berkeley
Bruce Dunn, U. of California, Los Angeles Trevor Robbins, U. of Cambridge
DATA ENTRY SPECIALISTS Brenden Aquilino, Fiona Giblin Jim Roberts, Fred Hutchinson Cancer Research Ctr.
William Dunphy, Caltech
PUBLISHER RELATIONS MANAGER Catherine Holland PUBLISHER RELATIONS, EASTERN REGION Keith Layson PUBLISHER RELATIONS, WESTERN Christopher Dye, WHO Barbara A. Romanowicz, U. of California,Berkeley
Todd Ehlers, U. of Tuebingen Amy Rosenzweig, Northwestern University
REGION Ryan Rexroth SALES RESEARCH COORDINATOR Aiesha Marshall MANAGER, SITE LICENSE OPERATIONS Iquo Edim SENIOR PRODUCTION
David Ehrhardt, Carnegie Inst. of Washington Mike Ryan, U. of Texas, Austin
SPECIALIST Robert Koepke SENIOR OPERATIONS ANALYST Lana Guz FULFILLMENT ANALYST Judy Lillibridge ASSOCIATE DIRECTOR, MARKETING Mitinori Saitou, Kyoto U.
Tim Elston, U. of North Carolina at Chapel Hill
Christina Schlecht MARKETING ASSOCIATES Thomas Landreth, Isa Sesay-Bah Gerhard Ertl, Fritz-Haber-Institut, Berlin Shimon Sakaguchi, Kyoto U.
Barry Everitt, U. of Cambridge Miquel Salmeron, Lawrence Berkeley National Lab
WEB TECHNOLOGIES SR. DEVELOPER Chris Coleman DEVELOPERS Dan Berger, Jimmy Marks, Ryan Jensen SR. PROJECT MANAGER Trista Smith
Ernst Fehr, U. of Zurich Jrgen Sandkhler, Medical U. of Vienna
PROJECT MANAGER Nick Fletcher Anne C. Ferguson-Smith, U. of Cambridge Alexander Schier, Harvard U.
Michael Feuer, The George Washington U. Vladimir Shalaev, Purdue U.
DIGITAL MEDIA DIRECTOR OF ANALYTICS Enrique Gonzales DIGITAL REPORTING ANALYST Eric Hossinger SR. WEB PRODUCER Sarah Crespi
Toren Finkel, NHLBI, NIH Robert Siliciano, Johns Hopkins School of Medicine
WEB PRODUCER Alison Crawford VIDEO PRODUCER Nguyen Nguyen SOCIAL MEDIA PRODUCER Brice Russ Kate Fitzgerald, U. of Massachusetts Denis Simon, Arizona State U.
Peter Fratzl, Max-Planck Inst. Uri Simonsohn, U. of Pennsylvania
DIRECTOR OF OPERATIONS PRINT AND ONLINE Lizabeth Harman DIGITAL/PRINT STRATEGY MANAGER Jason Hillman QUALITY TECHNICAL
Elaine Fuchs, Rockefeller U. Alison Smith, John Innes Centre
MANAGER Marcus Spiegler PROJECT ACCOUNT MANAGER Tara Kelly DIGITAL PRODUCTION MANAGER Lisa Stanford ASSISTANT MANAGER Daniel Geschwind, UCLA Richard Smith, U. of North Carolina (S)
DIGITAL/PRINT Rebecca Doshi SENIOR CONTENT SPECIALISTS Steve Forrester, Antoinette Hodal, Lori Murphy, Anthony Rosen Karl-Heinz Glassmeier, TU Braunschweig John Speakman, U. of Aberdeen
CONTENT SPECIALISTS Jacob Hedrick, Kimberley Oster Ramon Gonzalez, Rice U.
Allan C. Spradling, Carnegie Institution
Julia R. Greer, Caltech
DESIGN DIRECTOR Beth Rakouskas DESIGN EDITOR Marcy Atarod SENIOR DESIGNER Garvin Grulln DESIGNER Chrystal Smith of Washington
Elizabeth Grove, U. of Chicago
Jonathan Sprent, Garvan Inst. of Medical Research
GRAPHICS MANAGING EDITOR Alberto Cuadra SENIOR SCIENTIFIC ILLUSTRATORS Chris Bickel, Katharine Sutliff SCIENTIFIC ILLUSTRATOR Nicolas Gruber, ETH Zurich
Kip Guy, St. Judes Childrens Research Hospital Eric Steig, U. of Washington
Valerie Altounian SENIOR ART ASSOCIATES Holly Bishop, Nathalie Cary, Preston Huey PHOTOGRAPHY MANAGING EDITOR William Douthitt Paula Stephan, Georgia State U. and National Bureau
Taekjip Ha, U. of Illinois at Urbana-Champaign
PHOTO EDITORS Leslie Blizard, Christy Steele of Economic Research
Christian Haass, Ludwig Maximilians U.
Michael Hasselmo, Boston U. Molly Stevens, Imperial College London
DIRECTOR, GLOBAL COLLABORATION, CUSTOM PUBLICATIONS, ADVERTISING Bill Moran EDITOR, CUSTOM PUBLISHING Sean Sanders: 202-
Martin Heimann, Max-Planck Inst. Jena V. S. Subrahmanian, U. of Maryland
326-6430 ASSISTANT EDITOR, CUSTOM PUBLISHING Tianna Hicklin: 202-326-6463 ADVERTISING MARKETING MANAGER Justin Sawyers: Ira Tabas, Columbia U.
Yka Helariutta, U. of Cambridge
202-326-7061 science_advertising@aaas.org ADVERTISING SUPPORT MANAGER Karen Foote: 202-326-6740 ADVERTISING PRODUCTION OPERATIONS James A. Hendler, Rensselaer Polytechnic Inst. Sarah Teichmann, Cambridge U.
MANAGER Deborah Tompkins SR. PRODUCTION SPECIALIST/GRAPHIC DESIGNER Amy Hardcastle SR. TRAFFIC ASSOCIATE Christine Hall SALES Janet G. Hering, Swiss Fed. Inst. of Aquatic John Thomas, North Carolina State U.
COORDINATOR Shirley Young ASSOCIATE DIRECTOR, COLLABORATION, CUSTOM PUBLICATIONS/CHINA/TAIWAN/KOREA/SINGAPORE Ruolei Wu: Science & Technology Shubha Tole, Tata Institute of Fundamental Research
+86-186 0082 9345, rwu@aaas.org COLLABORATION/CUSTOM PUBLICATIONS/JAPAN Adarsh Sandhu + 81532-81-5142 asandhu@aaas.org EAST COAST/E. Kai-Uwe Hinrichs, U. of Bremen Christopher Tyler-Smith, The Wellcome Trust
David Hodell, U. of Cambridge Sanger Inst.
CANADA Laurie Faraday: 508-747-9395, FAX 617-507-8189 WEST COAST/W. CANADA Lynne Stickrod: 415-931-9782, FAX 415-520-6940 MIDWEST
David Holden, Imperial College Herbert Virgin, Washington U.
Jeffrey Dembksi: 847-498-4520 x3005, Steven Loerch: 847-498-4520 x3006 UK EUROPE/ASIA Roger Goncalves: TEL/FAX +41 43 243 1358 JAPAN Lora Hooper, UT Southwestern Medical Ctr. at Dallas Bert Vogelstein, Johns Hopkins U.
Katsuyoshi Fukamizu(Tokyo): +81-3-3219-5777 kfukamizu@aaas.org CHINA/TAIWAN Ruolei Wu: +86-186 0082 9345, rwu@aaas.org Tamas Horvath, Yale University Cynthia Volkert, U. of Gttingen
Raymond Huey, U. of Washington David Wallach, Weizmann Inst. of Science
WORLDWIDE ASSOCIATE DIRECTOR OF SCIENCE CAREERS Tracy Holmes: +44 (0) 1223 326525, FAX +44 (0) 1223 326532 tholmes@science-int. co.uk Fred Hughson, Princeton U. Ian Walmsley, U. of Oxford
CLASSIFIED advertise@sciencecareers.org U.S. SALES Tina Burks: 202-326-6577, Nancy Toema: 202-326-6578 EUROPE/ROW SALES Sarah Lelarge SALES Auke Ijspeert, EPFL Lausanne Jane-Ling Wang, U. of California, Davis (S)
ASSISTANT Kelly Grace JAPAN Hiroyuki Mashiki(Kyoto): +81-75-823-1109 hmashiki@aaas.org CHINA/TAIWAN Ruolei Wu: +86-186 0082 9345 rwu@ Stephen Jackson, USGS and U. of Arizona
David A. Wardle, Swedish U. of Agric. Sciences
aaas.org MARKETING MANAGER Allison Pritchard MARKETING ASSOCIATE Aimee Aponte Steven Jacobsen, U. of California, Los Angeles
David Waxman, Fudan U.
Kai Johnsson, EPFL Lausanne
Jonathan Weissman, U. of California, San Francisco
AAAS BOARD OF DIRECTORS, CHAIR Geraldine L. Richmond PRESIDENT Barbara A. Schaal PRESIDENT-ELECT Susan Hockfield TREASURER Peter Jonas, Inst. of Science & Technology (IST) Austria
Matt Kaeberlein, U. of Washington Chris Wikle, U. of Missouri (S)
David Evans Shaw CHIEF EXECUTIVE OFFICER Rush D. Holt BOARD Cynthia M. Beall, May R. Berenbaum, Carlos J. Bustamante, Ian A. Wilson, The Scripps Res. Inst. (S)
Stephen P.A. Fodor, Claire M. Fraser, Michael S. Gazzaniga, Laura H. Greene, Elizabeth Loftus, Mercedes Pascual William Kaelin Jr., Dana-Farber Cancer Inst.
Daniel Kahne, Harvard U. Timothy D. Wilson, U. of Virginia
SUBSCRIPTION SERVICES For change of address, missing issues, new orders and renewals, and payment questions: 866-434-AAAS (2227) or 202-326-6417, FAX 202-842- Daniel Kammen, U. of California, Berkeley Rosemary Wyse, Johns Hopkins U.
1065. Mailing addresses: AAAS, P.O. Box 96178, Washington, DC 20090-6178 or AAAS Member Services, 1200 New York Avenue, NW, Washington, DC 20005 Abby Kavner, U. of California, Los Angeles Jan Zaanen, Leiden U.
Masashi Kawasaki, U. of Tokyo Kenneth Zaret, U. of Pennsylvania School of Medicine
INSTITUTIONAL SITE LICENSES 202-326-6730 REPRINTS: Author Inquiries 800-635-7181 COMMERCIAL INQUIRIES 803-359-4578 PERMISSIONS 202-326-6765, V. Narry Kim, Seoul National U. Jonathan Zehr, U. of California, Santa Cruz
permissions@aaas.org AAAS Member Services 202-326-6417 or http://membercentral.aaas.org/discounts Joel Kingsolver, U. of North Carolina at Chapel Hill Len Zon, Childrens Hospital Boston
Science serves as a forum for discussion of important issues related to the advancement of science by publishing material on which a consensus has been reached as Robert Kingston, Harvard Medical School Maria Zuber, MIT
well as including the presentation of minority of conflicting points of view. Accordingly, all articles published in Scienceincluding editorials, news and comment, and book
BOOK REVIEW BOARD
reviewsare signed and reflect the individual views of the authors and not official points of view adopted by AAAS or the institutions with which the authors are affiliated.
David Bloom, Harvard U., Samuel Bowring, MIT, Angela Creager, Princeton U.,
INFORMATION FOR AUTHORS See pages 624 and 625 of the 5 February 2016 issue or access www.sciencemag.org/authors/science-information-authors Richard Shweder, U. of Chicago, Ed Wasserman, DuPont

1002 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
ED ITORIAL

#IAmAResearchParasite

I
n the midst of steady progress in policies for data to implement, and it takes more than community ac-
sharing, a recent editorial expressed a contrar- ceptance of the practice. Agencies supporting research
ian view.* The authors described the concern of in oceanography have long funded data and sample
some scientists about the rise of an underclass of repositories and have encouraged data and sample de-
research parasites who exploit data sets that are position by making new awards contingent on compli-
collected and curated by others. Even worse, these ance. Repositories are instrumental in setting formats
parasites might use such data to try to disprove for data, so much so that standard programs and apps
the conclusions posited in the datas original source accept and output data in the standard format. The ma- Marcia McNutt is
studies. The editorial raised the points of how anyone rine community supports data professionals who are
Editor-in-Chief,
not involved in the original study could use the data responsible for the quality control of data collected on
Science Journals
without misrepresenting ships and from other major
it, and the danger of per- observing programs.
haps arriving at erroneous Often overlooked is the
conclusions. The editorial importance of community-

Downloaded from on March 3, 2016


advised instead that data established metadata, so
sharing be implemented by that those not involved in
involving the authors of the the original research will
original study as coauthors know what the data mean.
in follow-up research. The As an example, in an ocean-
research community im- ographic temperature data
mediately took to Twitter base, the community had to
under the hashtag #IAmA- agree on what T(0) meant.
ResearchParasite to voice Was it temperature at atmo-
opposition to the editorial. spheric pressure? Tempera-
Much of what we know ture at the sea surface?
about the large-scale fea- Communities must dis-
tures of this planet is appar- courage low-quality data
ent thanks to widespread collection. A well-attended
data-sharing practices and poster presentation at
the early establishment of one prominent scientic
data banks in the geosci- meeting some years ago
ences. Aspects such as de- compared the crossover er-
termining the shape of the
There are costs rors (mists) of nontime-
ocean oor, ocean chemis- for re-collecting data dependent measurements
try, the internal structure (such as depth soundings)
of Earths deep interior, the
for new uses. from ships tracks where
physics and chemistry of they intersected in the
the atmosphere, and many other topics could not have worlds oceans. Any discrepancy at a crossing could be
been ascertained from a single investigators eld pro- attributed to poor data quality control on either ship,
PHOTOS: (INSET) HARDIE/ALAMY; (TOP) STACEY PENTLAND PHOTOGRAPHY

gram. One meta-analysis I published on the South Pa- but with thousands of crossings, institutions with sys-
cic beneted from observations of my own and those tematically more mists than others stand out. The
of others, including the 18th-century British explorer results did not escape the attention of the funding agen-
Captain James Cook. Involving Cook as a coauthor on cies that support ship time.
my paper was clearly not an option, any more than it There are costs to implementing data reuse, but there
would have been feasible or desirable to include the are also costs for irreproducible research and for re-
dozens of others, living or dead, who had contributed to collecting data for new uses. And no amount of funding
the data repository. Many elds, including the biomedi- can reconstruct lost ephemeral or time-dependent
cal sciences, are now beneting from meta-analyses of phenomena for which the data were not well curated.
data to better understand the big picture. No more excuses: Lets step up to data sharing.
Efective data sharing is not trivial or inexpensive Marcia McNutt

*D. L. Longo, J. M. Drazen, N. Engl. J. Med. 374, 276 (2016). 10.1126/science.aaf4701

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1005


Published by AAAS
ED ITORS N OTE

Data sharing
The Science family of journals is committed to sharing and nal data as rapidly and widely as possible, includ-
data relevant to public health emergencies, and there- ing with public health and research communities and
fore we are signatories to, and wholeheartedly endorse, the World Health Organisation.
the following statement by funders and journals.*
We urge other organisations to make the same
Statement on Data Sharing in Public Health commitments.
Emergencies This commitment is in line with the consensus state-
The arguments for sharing data, and the consequences ment agreed at a WHO expert consultation on data Marcia McNutt is
of not doing so, have been thrown into stark relief by sharing last year whereby researchers are expected to
Editor-in-Chief,
the Ebola and Zika outbreaks. share data at the earliest opportunity, once they are
Science Journals
In the context of a public health emergency of inter- adequately controlled for release and subject to any
national concern, there is an imperative on all parties safeguards required to protect research participants
to make any information available that might have and patients.
value in combatting the crisis.

Downloaded from on March 3, 2016


We are committed to working in partnership to en- Signatories to the Statement
sure that the global response to public health emergen- Academy of Medical Sciences, UK
cies is informed by the best available research evidence Bill and Melinda Gates Foundation
and data, as such: Biotechnology and Biological Sciences Research Council
(BBSRC)
Journal signatories will make all content concern- Bulletin of the World Health Organization
ing the Zika virus free to access. Any data or preprint Canadian Institutes of Health Research
deposited for unrestricted dissemination ahead of Centers for Disease Control and Prevention
submission of any paper will not pre-empt its publica- Chinese Academy of Sciences
tion in these journals. Chinese Centre for Disease Control and Prevention
Funder signatories will require researchers undertak- Department of Biotechnology, Government of India
ing work relevant to public health emergencies to set Department for International Development (DFID)
in place mechanisms to share quality-assured interim Deutsche Forschungsgemeinschaft (DFG)
eLife
Faculty of 1000 (F1000)
Aedes aegypti mosquito Fondation Mrieux
Fundao Oswaldo Cruz (Fiocruz)
Institut Pasteur
Japan Agency for Medical Research and
Development (AMED)
The JAMA Network
The Lancet
Mdecins Sans Frontires/Doctors Without Borders
(MSF)
National Academy of Medicine
National Institutes of Health, USA
National Science Foundation
The New England Journal of Medicine (NEJM)
PLOS
Science Journals
PHOTOS: CDC/JAMES GATHANY/SCIENCE SOURCE

South African Medical Research Council


Springer Nature
UK Medical Research Council
Wellcome Trust
ZonMwThe Netherlands Organisation for Health
Research and Development
Marcia McNutt

Published online 10 February 2016. *Updates to signatories can be found at


www.wellcome.ac.uk/News/Media-ofce/Press-releases/2016/WTP060169.htm. 10.1126/science.aaf4545

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1007


Published by AAAS
Precision medicine is not a replacement for

NEWS making sure people have just basic health care.
President Obama, 25 February, at a White House summit marking 1 year
of the Precision Medicine Initiative, when asked what the initiative can do to
address opioid drug abuse in rural communities. http://scim.ag/PMI1year

AROUND THE WORLD
IN BRIEF
Average innovator? Not Steve Jobs
| If you think the
WA S H I N GT O N , D.C .
archetypal U.S. researcher-entrepreneur
Rewilding led to hybrid orangutans is a young white college dropout building
a startup in his garage, then think again,
argues a new study of the demographics
of U.S. innovators in the information, life,
and materials sciences. The face behind
the next big thing could easily belong
to a middle-aged male Ph.D. from India
or China, working at a large rm, con-

Downloaded from on March 3, 2016


clude researchers from the Information
Technology and Innovation Foundation
in Washington, D.C., and George Mason
University in Fairfax, Virginia. The team
used patents and awards to identify and
send surveys to more than 6000 U.S. inno-
vators. Of the 923 who responded, 46% are
immigrants or the children of immigrants.
More than half hold doctorates and work
at rms with more than 500 employees.
The median age is 47, and just 12% are
women. Fewer than 8% belong to a minor-
ity group. http://scim.ag/avginnovator

Many orangutans at Camp Leakey in Borneo now carry a cocktail of genes from diferent subspecies.
Budget crunch at Berkeley

F
rom 1971 to 1985, primatologists have released more than | The University
B E R K E L E Y, CA L I F O R N I A

90 rehabilitated orangutansmost conscated from the illegal pet of California (UC), Berkeley, may disband
the universitys College of Chemistry to
tradeinto the wilds of Borneo. At the time, orangutans were con- help cope with a cash crunch. Faced with
sidered a single species; however, they are now thought to be two $150 million in debt, as well as at income
separate species with at least three reproductively distinct sub- from tuition and rising costs, the university
species. Those early rewildings, a new study shows, inadvertently is considering saving money by closing
the college and absorbing its departments
introduced the wrong subspecies into local populations. Primatologist
PHOTOS: (LEFT TO RIGHT) GRAHAM L BANES; KEEGAN HOUSER/UC BERKELEY

into other university colleges. No decisions


Graham Banes of the University of Cambridge in the United Kingdom have yet been made; so far, more than 2250
and colleagues analyzed 44 years of data from Borneos Camp Leakey
and also used fecal samples to study the genetics of the camps ape
population. Two reintroduced females were not from the local sub-
species; one has at least 14 descendants, whereas the other has had few
ofspring, many with health issues that could be the result of mixing
populations. Meanwhile, an estimated nine males from mixed-species
unions have carried their cocktail of genes into the wilds of Tanjung
Puting National Parkwith unknown repercussions. Given the num-
ber of orangutan reintroductions and the likelihood that other wildlife
sanctuaries have also reintroduced the wrong species, this may be the
tip of the iceberg, the researchers wrote last week in Scientic Reports.
Latimer Hall at UC Berkeleys College of Chemistry.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1009


Published by AAAS
NEWS | I N B R I E F

problems, and eye damage, researchers


Softball-sized Rafesia reported last week at a special session
consueloae is the of the Conference on Retroviruses and
smallest species in the Opportunistic Infections in Boston.
giant fower family. Another group reported that many men
harbor Ebola RNA in their semen after
they recover, and for far longer than
imaginedwhich suggests that fresh
outbreaks caused by sexual transmission
can remain a threat for nearly a year.
And a third team of scientists presented
disappointing results from a clinical trial
of what was once considered the most
promising Ebola treatment: an antibody
cocktail called ZMapp. The study, started
late in the epidemic, enrolled fewer people
than planned, and failed to provide sta-
tistically signicant results. Although the
conference focuses on HIV/AIDS, meeting
organizers included the Ebola ndings
because of overlap in the HIV and
Corpse fowers cousin smells like coconut Ebola research communities and because
of the magnitude of the epidemic.

T
he worlds largest and perhaps stinkiest fower has a newly discoveredand
http://scim.ag/_Ebolanews
much smaller and much more pleasant-smellingcousin. The genus Rafesia
has produced some of the largest fowers on Earth; the largest known individual
fower is about a meter across and was produced by R. arnoldii, which some call
the corpse fower for its distinctive odor (and not to be confused with another
Scientists decry spending plan
NEW DELHI | Science spending in India is
giant fower commonly called a corpse fower, Amorphophallus titanium). Now, in the
mountains of the Philippiness Luzon Island, scientists have found R. consueloae, a slated to rise by 11% in the 201617 scal
parasitic fowering plant (like its larger cousins) that has a far less ofensive smell. year, according to the budget proposal
It measures only 10 centimeters across, making it the smallest of the giants, they presented this week by Prime Minister
reported last week in PhytoKeys. Narendra Modi. But ination, projected
to be 5% in the coming year, will consume
much of the increase, many scientists say
and far greater investments will be needed
people have signed a petition asking UC patents Illumina has exclusively licensed to revamp crumbling infrastructure. We
Berkeley Chancellor Nicholas Dirks to scrap from the University of Washington and the have very few laboratories and institutions
the idea of disbanding the school, which is University of Alabama, Birmingham. comparable to the best in the world, says
home to 101 faculty and 1492 students and C.N.R. Rao, a chemist and adviser to the
postdocs. Thirteen of its faculty have won previous prime minister, Manmohan Singh.
Nobel Prizes. Ebola risk, ZMapp limits revealed We need to provide much more support
BOSTON | Worrisome ndings about the for improving this state. In the new budget,
201316 Ebola epidemic in West Africa are agricultural research gets a 19% boost,
Nanopore pioneer sued continuing to emerge. Many survivors of and earth sciences and renewable energy
| The rst
S A N D I E G O, CA L I F O R N I A the disease have experienced long-term increase by 16% each. Space research and

PHOTOS: (TOP TO BOTTOM) BIODIVERSITY RESEARCH LABORATORY/UP BIOLOGY; OCHA


company to bring nanopore DNA sequenc- efects, including joint pain, neurological atomic energy would get increases of just
ing to market is facing charges of patent 6.6% and 5.1%, respectively. The Council for
infringement from genetic sequencing Scientic and Industrial Research, which
giant Illumina. The company led a law- runs a network of 38 national labs, will
suit against U.K.-based Oxford Nanopore get a 4.6% increase, in line with orders it
Technologies last week, claiming that the received last year to self-nance half of its
nanopore platforms it has made available budget within the next 2 to 3 years.
to researchers since 2014 use patented http://scim.ag/Indiascience
technology Illumina controls. Hailed as
a revolution in cheap and portable DNA
analysis (Science, 19 February, p. 800), NEWSMAKERS
nanopore sequencing reads long strands
of DNA by measuring changes in electrical Surgeon facing dismissal
current as diferent nucleotides of a strand Pressure on surgeon Paolo Macchiarini
pass through a bacterial pore. Oxford hasnt is increasing as the Karolinska Institute
revealed what type of pore is inside its (KI) in Stockholm says it will try to cut
devices, but Illumina contends that they ties with him before his contract runs
rely on the Mycobacterium smegmatis porin Ebola survivors, like this mother and daughter, face out in November. Macchiarini, a surgeon
(Msp). The Msp pore is described in two possible long-term symptoms. who implanted articial tracheae into

1010 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
CANCER TARGETS
IN YOUR HANDS
patients at KI and elsewhere, is under a on 24 February. Cardiologist Robert Calif
cloud of controversy after colleagues and has faced opposition from several sena-
media reports questioned the ethics of the tors since his nomination in September
operations and the accuracy of papers he 2015. Some labeled him an industry insider Our new innovative panel of
published about their success (Science, for his role in overseeing large company- PrecisA Monoclonals targets
5 February, p. 546). Meanwhile, The Lancet funded trials at Duke University Medical Epithelial to Mesenchymal
published a letter on 23 February detailing Center in Durham, North Carolina. Others, proteins furthering the
problems with a 2011 paper by Macchiarini meanwhile, used the nomination as a
that described early success with the rst chance to take the agency to task over vari- understanding of cancer
patient to receive the polymer trachea ous issues not related to Calif, including progression.
seeded with stem cells. The journal has FDAs response to the epidemic of opioid
previously declined to amend or retract the abuse. The opposition was blunted by a Atlas Antibodies. The No.1 in IHC
paper. http://scim.ag/Macchiarini procedural vote last week that limited oor
debate, and the Senate approved the nomi-
nation by a vote of 89 to four. Calif
Calif confrmed as FDA head will take the helm as FDA faces pressure
President Barack Obamas new pick to from Congress to speed the approval
head the Food and Drug Administration of medical products to patients.
(FDA) was conrmed by the U.S. Senate http://scim.ag/_Calif

1/3v AD
HERE

Linda, Scientic Support Manager


Return of the monarchs

T
he number of monarch butterfies wintering in Mexico is up for the second year in
a rowbut their famous migration remains imperiled, researchers said in a press
conference in Mexico City 26 February. Every year, millions of butterfies journey
from their breeding grounds in the U.S. Midwest and Canada to their winter
habitat in the oyamel fr and pine forests of central Mexico. Since 1993, research-
MADE IN SWEDEN
PHOTO: FRANS LANTING MINT IMAGES/NEWSCOM

ers have estimated the butterfies numbers based on the total forest area they
inhabit. Monarchs covered 4 hectares of forest this winter, according to a December
2015 survey released last week. Thats nearly four times as much area as last year
(1.13 hectares), and suggests an ongoing rebound from the all-time recorded low of
0.67 hectares in the winter of 201314. But monarch numbers remain low compared
with the late 1990s and early 2000s, when the butterfies often covered 8 hectares
or more. Their main threat is still the loss of milkweed in the United States (Science,
23 January 2015, p. 357), says Omar Vidal, the director general of the World Wildlife
Fund Mexico in Mexico City, which administers the annual count.

Learn more today at atlasantibodies.com


SCIENCE sciencemag.org

Published by AAAS
NEWS

IN DEP TH

Young pulsars like the one


at the heart of the Crab
Nebula (shown in this artists
conception) may spark
repeating fast radio bursts.

RADIO ASTRONOMY

Downloaded from on March 3, 2016


Fast radio bursts tease astronomers
Conicting results suggest the potent blasts come in two avors
By Govert Schilling rst one in 2007, in archival data collected tions with the Japanese 8.2-meter Subaru
in 2001 by the 64-meter Parkes Observa- Telescope on Mauna Kea, Hawaii, revealed

T
he more radio astronomers learn tory radio telescope in Australia. Since then, a dim elliptical galaxy some 6 billion light-
about the potent bursts of radio 16 more of the mysterious signals have been years from Earthapparently FRB 150418s
waves called fast radio bursts (FRBs), detected, most at Parkes, but no one knew host galaxy.
the deeper their confusion becomes. what to make of them. Theories about their Not everyone is convinced, however. On
Last week, a scientic paper sug- originincluding speculative ideas about 26 February, in a note on the arXiv pre-
gested that these milliseconds-long black holes turning inside outvastly out- print server, astrophysicists Peter Williams
radio bursts result when superdense burnt- number the discoveries themselves. and Edo Berger of Harvard University pro-
out stars called neutron stars collide and One thing astronomers do agree on is that posed that the fading radio source could be
perish in remote galaxies. But this week, FRBs originate in remote galaxies. Thats a naturally varying signal from the distant
another paper reports multiple bursts from because longer radio waves from the bursts galaxys nucleusand unrelated to the FRB.
an FRBsomething merging neutron stars arrive slightly later than shorter waves do. The researchers say new radio observations
cannot explain. That so-called dispersion results from an they carried out on 27 and 28 February with
These are some papers! says astro- interaction between the radio waves and the Very Large Array radio telescope on the
physicist Chryssa Kouveliotou of George tenuous clouds of electrons in space. The Plains of San Agustin in New Mexico sup-
Washington University in Washington, D.C. space between the stars in our own Milky port that interpretation.
The rst one already presented a major dis- Way galaxy contains too few electrons to But if Keane and his colleagues are cor-
covery, showing that the progenitor of the cause the large dispersion that astronomers rect about the extremely remote location, the
FRB was destroyed in the process. Now, the see, so the bursts must have originated most plausible source of the burst may be
second one has a completely diferent result: much farther away. two ultracompact neutron stars coalescing
Not only did the progenitor survive, but it Last weeks paper, which Evan Keane of into a black holean event energetic enough
kept giving. Its extremely interesting. the Square Kilometre Array Organisation, in to explain the bursts estimated power of 1035
FRBs last on the order of a millisecond Manchester, U.K., and colleagues published watts and its long radio afterglow, presum-
and pop up all over the sky, probably thou- in the 26 February issue of Nature, appeared ably from a cooling reball. Such neutron
sands of times per day across the observable to prove that FRBs are extremely remote, star mergers are also the prime candidate
universe. Without knowing their distance, one-time cataclysms. On 18 April 2015, for explaining another cosmic mystery, short
IMAGE: DAVID A. AGUILAR/NASA/ESA

its impossible to say how energetic they the Parkes telescope detected a new milli- gamma ray bursts: eeting blasts of radiation
are. Unfortunately for researchers, they are second burst, FRB 150418, in the constella- that can outshine entire galaxies for a few
extremely hard to detect, mainly because tion Canis Major. Just a few hours later, the seconds. No gamma rays have been detected
most radio telescopes can view only min- six-dish Australia Telescope Compact Array from FRB 150418, Keane says, but that could
ute patches of sky. David Narkevic, a gradu- in Narrabri found a faint source of radio be because the radiation is focused in a tight
ate student working with radio astronomer waves at the same location, slowly fading beam not aimed at Earth.
Duncan Lorimer at West Virginia University over 6 dayspresumably the afterglow of But a one-time catastrophe cant explain
in Morgantown, serendipitously found the a catastrophic event. Subsequent observa- the phenomenon reported in this weeks

1012 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
NEWS

issue of Nature by Laura Spitler of the Max INFECTIOUS DISEASE


Planck Institute for Radio Astronomy in
Bonn, Germany; Jason Hessels of the Uni-
versity of Amsterdam and ASTRON, the
Netherlands Institute for Radio Astronomy;
Patient Zero no more
and colleagues. They reobserved the re- Sleuthing claries HIVs history
gion of the sky where the 305-meter Are-
cibo Observatory radio telescope in Puerto By Jon Cohen, in Boston proven to be the culprit in 1984, researchers
Rico had detected FRB 121102 in November analyzed stored blood samples from gay and

A
2012. In 3 hours of follow-up observations new study pinpoints with greater bisexual men collected in the late 1970s in
from May and June 2015, Paul Scholz, a precision than ever before when and San Francisco, California, and New York City
postdoctoral student at McGill University where HIV entered the United States and found that the men already carried anti-
in Montreal, Canada, identied two more and sparked North Americas AIDS bodies to the virus. Worobey plucked the vi-
bursts on 17 May and another eight on epidemic. It also puts to rest a media- rus from eight of those samples, which dated
2 June. Many others may have gone un- fueled theory about a peripatetic from 1978 and 1979, and obtained nearly
noticed. The additional radio blips leave no ight attendant with a very active sex life: the complete genetic sequences of each one.
doubt that the object they come from is not so-called Patient Zero, who after his death Jafe, who had helped piece together Dugass
destroyed in the process, Lorimer says: Its became infamous for supposedly bringing connection to a cluster of early AIDS cases
basically impossible for a cataclysmic event the virus into the country and seeding the in California and New York, supplied a 1983
to produce repeating bursts. infection in diferent locales. sample of the ight attendants blood, which
So if this cant be a neutron star merger, In And the Band Played On, the 1987 also yielded a near-complete HIV genome.
what are the alternatives? One once- best-selling account of the U.S. epidemics HIV mutates at a constant rate each time
promising candidateso-called giant ares early years, journalist Randy Shilts wrote it copies itself, so Worobey could use the
on young, highly magnetized neutron stars in detail about Patient diferences among the
known as magnetarshas fallen from favor, Zero, a Canadian named sequences to create a
Kouveliotou says. No known magnetar in our Gatan Dugas who died family tree of the viruses
own galaxy has produced more than one gi- from AIDS 3 years before and estimate when each
ant are in almost half-a-century of observ- the book was published. isolate emerged. His
ing. A stronger possibility, Kouveliotou and [T]heres no doubt that work suggests the U.S.
the authors agree, is that the repeating bursts Gatan [sic] played a epidemic most likely
of FRB 121102 are irregular giant explosions key role in spreading the began in New York City
on a pulsar: a rapidly spinning neutron star. new virus from one end around 1970, when the
The pulsar at the heart of the famous Crab of the United States to real index case brought
Nebula (which also sends out regular weaker the other, Shilts wrote. in a virus that closely
pulses of radio waves) is known to produce He did not say Dugas matched the sequences
incidental giant pulses. introduced the disease of older HIVs isolated
The bottom line, Lorimer says, is that to North America, but from people in Haiti
the results in the two papers immediately subsequent media ac- and a few other Carib-
divide the sample of fast radio bursts into counts, including one in bean countries. Although
two classes. Similar things have happened Time magazine, did make his sample size is small,
before in astronomy: In the 1930s, astrono- that claimand vilied Worobey said the proba-
mers discovered that supernova explosions Dugas. A front-page New Gatan Dugas, infamously, and wrongly, bility that New York City
come in several varieties, and in the early York Post headline went labeled Patient Zero. was the origin of the U.S.
1990s, they realized that diferent mecha- so far as to proclaim, epidemic is very, very
nisms produce long and short gamma ray THE MAN WHO GAVE US AIDS. HIV/ high indeed. He estimates the virus reached
bursts. In fact, says Kouveliotou, the cur- AIDS researchers have long been skeptical San Francisco around 1975.
rent situation with FRBs is starting to make about the Patient Zero claim. On the family tree of the early U.S. iso-
me feel like some 24 years ago, when I was Last week at the Conference on Retro- lates, Dugass HIV genome fell in about the
working on gamma ray bursts and then viruses and Opportunistic Infections here, middle. Theres nothing special about his
magnetars popped in to add to the mystery Michael Worobey, an evolutionary bio- genome, Worobey said. Although Dugas
and the confusion. Its all very exciting. logist at the University of Arizona in Tuc- had been to Haitiin 1977there is no sign
Lorimer expects FRBs to become a ma- son, solidly debunked that claim, saying that he was either the index patient or key
jor focus of astrophysical research. To nd Dugas was far from what epidemiologists to spreading the virus around the country.
more of them, astronomers need wide-eld refer to as the index patient in the United Worobey notes that when CDC researchers
instruments like the Canadian Hydrogen States. Worobeys new work will help put were piecing together how the rst AIDS
Intensity Mapping Experiment, under con- the myth of Patient Zero to rest, says cases were linked to each other, they origi-
struction in British Columbia, or ASTRONs epidemiologist Harold Jafe of the Centers nally referred to Dugas not by the number
upgraded Westerbork Synthesis Radio Tele- for Disease Control and Prevention (CDC) 0 but by the letter O, because he was
PHOTO: WIKIMEDIA COMMONS

scope, which will become operational later in Atlanta, who helped unravel the early from outside California. Later, for reasons
this year. To capture other signals from each spread of the disease. that remain murky, Dugas morphed into Pa-
radio source before they fade, astronomers The rst AIDS cases surfaced in ve gay tient Zero.
also need to build systems for quickly alert- men in Los Angeles, California, and were So rather than Dugas sparking the U.S.
ing other telescopes. Lorimer says: Theres reported in CDCs Morbidity and Mortality epidemic, his case, historically speaking, re-
an awful lot of work ahead of us. Weekly Report in June 1981. After HIV was ally is just an oh.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1013


Published by AAAS
NEWS | I N D E P T H

Surveys, like this one collecting family information in


Guatemala, may often be marred by fabricated data,
a new statistical test suggests.

exactly the appropriate method for detect-


ing fabrication, Kuriakose says.
After he left Pew, Robbins became director
of Arab Barometer, which surveys opinion
across the Arab world. He tested data from
that project and found signs of fabrication,
which subsequent investigation conrmed.
Wondering about the scale of the problem,
Robbins and Kuriokose rened their test
and applied it to 1008 international surveys,
including many from Pew, though they only
reported the aggregate results. Their test
agged 17% as likely to contain a signicant
portion of fabricated dataand the gure
shot up to 26% for surveys done in the de-
veloping world.
We found out about this study and were
very alarmed, says Courtney Kennedy, di-
SCIENTIFIC INTEGRITY rector of survey research for Pew. She and

Downloaded from on March 3, 2016


her colleagues used the same test on Pews

Survey fraud test sparks battle surveys and found that a certain share
failed by Robbins and Kuriakoses stan-
dards, Kennedy says, but follow-up investi-
Pew Research Center challenges statistical test gation left the organization convinced only
a handful were suspicious.
By John Bohannon positives. The organization has gone so far Late last year, after the pair submitted a
as to request Robbins and Kuriakose de- paper on their method and ndings to the

W
ith grants from charitable dona- sist from publishing their original analy- peer-reviewed Statistical Journal of the
tions, the Pew Research Center sis, which is now in press. Pews actions IAOS, one of the elds leading publications,
funds dozens of surveys a year, on are pretty disappointing, says Kuriakose, they received an email from Kennedy and
everything from Americans views now a research scientist at SurveyMonkey other top ofcials at Pew. We strongly sug-
about science and religion to what in Palo Alto, California. This problem isnt gest that you retract the paper, the email
people in African countries see as going to just go away. stated, as we believe the analysis is severely
national priorities and problems. So its not Robbins and Kuriakose originally met underspecied and will give both survey
surprising that Pew wouldnt take kindly at Pew, which is based Washington, D.C. vendors and contractors a false metric for
to the suggestion that some, perhaps more Michael was doing methodology work on identifying fraud. Kennedy calls the letter
than a quarter, of its international surveys Pews international surveys and we con- appropriate because our organizations
are marred by data fabrication. Indeed, the nected about data quality, Kuriakose says. reputation is on the line. You cant make
organization is vigorously challenging that Surveying communities in the developing cavalier claims like that.
claimfrom two former Pew researchers. world often requires face-to-face inter- Kuriakose and Robbins did not withdraw
At a meeting on survey data fabrication views, done house-by-house in dangerous their paper. It was accepted in December
in Washington, D.C., last week, the pair, environments. To avoid risk or save time, 2015 and is in press, and at last weeks meet-
Michael Robbins and Noble Kuriakose, pre- Robbins says, interviewers sometimes re- ing they lobbed a second bombshell: their
sented an update on a newly developed sta- sort to curbstoning: sitting on the curb analysis of Pew surveys. Kennedy responded
tistical test that has been roiling the survey and inventing survey responses, often by with an attack on the tests methodology.
research community for the past year. When duplicating answers. For example, she points out, it does not
they apply the test to more than 1000 public To detect such fabrication, their test looks account for the number of questions on a
data sets from international surveys, about for highly similar responses from multiple survey, the number of respondents, or other
one in ve of the surveys fail, indicating a respondents. How similar is too similar? Af- factors that can skew the results. (Pew has
high likelihood of fabricated data. At the ter running a simulation of data fabrication posted its analysis online.)
meeting, they debuted an analysis focusing scenarios, they settled on 85% as the cutof. Before the meeting, co-organizer Steven
on 309 of Pews international studies that In a 100-question survey of 100 people, for Koczela, president of the MassINC Poll-
found a failure rate of 30%. example, fewer than ve people would be ing Group in Boston and a previous survey
Robbins and Kuriakose have uncovered expected to have identical answers on 85 of research leader for the U.S. State Depart-
PHOTO: MALCOLM LINTON

a massive problem, says Michael Spagat, the questions if the data are genuine. The ment told Science that the case laid out by
an economist at Royal Holloway, Univer- 85% rule isnt appropriate for all kinds of Kuriakose and Robbins seems unassail-
sity of London, who has investigated high- surveys, the pair acknowledgessome types able. Now that both sides have hashed it out
prole cases of possible survey data fab- tend to produce more uniform answers. But publicly, the community is digging into the
rication in war zones. But Pew ofcials for the large-scale opinion surveys typically issue, he says. I am encouraged by the sense
dismiss the test, saying it is prone to false carried out in the developing world this is of momentum and common purpose.

1014 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
PLANETARY SCIENCE

Scientists to drill into dinosaur-killing blast


Cores from Chicxulub crater could illuminate how life returned after the cataclysm
By Eric Hand meters, says David Smith, the IODP op-
erations manager at the British Geological Probing ground zero

T
his month, a drilling platform will Survey in Edinburgh, U.K.
In April, scientists will drill into Chicxulub crater,
rise in the Gulf of Mexico, but it Although this is the rst ofshore at- where an asteroid impact 66 million years ago led
wont be aiming for oil. Scientists tempt to drill into the crater, roughnecks to one of Earths biggest mass extinctions. They
will try to sink a diamond-tipped bit have sunk wells into it on landeven before hope to reach a buried peak ring, Earths only
into the heart of Chicxulub crater scientists knew a crater was there. In the preserved example.
the buried remnant of the asteroid 1950s, geologists for Pemex, Mexicos na-
Buried treasure
impact 66 million years ago that killed of tional oil company, conducted gravity and
Ofshore from Progreso, Mexico, scientists will drill
the dinosaurs, along with most other life magnetic surveys of the Yucatn Peninsula
into the craters peak ring, partially seen in geophysical
on the planet. They hope that the retrieved and were intrigued to see underground cir- remote sensing data (below). Onshore wells have been
rock cores will contain clues to how life cular structurespossible oil traps. They drilled into the crater before, but few were cored and
came back in the wake of the cataclysm, drilled several exploratory wells but lost none reached the peak ring.
and whether the interest when they
crater itself could got volcanic rocks in- U.S.
have been a home stead of oil-bearing

Downloaded from on March 3, 2016


MEX.
for novel microbial sediments. When Crater rim
life. And by drilling they found the igne-
into a circular ridge ous rocks, they said, Area in Peak ring
detail
inside the 180-kilo- Oh, this is a volcanic
meter-wide crater center, says Alan
rim, scientists hope Hildebrand, a geolo-
to settle ideas about gist at the University Drill site
how such peak of Calgary in Canada. Progreso
rings, hallmarks of In 1980, however,
the largest impact Nobel laureate Luis
CREDITS: (PHOTO) DETLEV VAN RAVENSWAAY/SCIENCE SOURCE; (DIAGRAM) V. ALTOUNIAN/SCIENCE; (MAP) SEAN GULICK, UNIVERSITY OF TEXAS

craters, take shape. Alvarez and others


Chicxulub is the called attention to a
50 km
only preserved struc- thin layer of iridium Previous wells
ture with an intact Artists reconstruction of Chicxulub crater soon possible material from
peak ring that we after impact, 66 million years ago. an asteroidfound Making the mounds
can get to, says Uni- all over the world in One model for peak ring formation predicts that
versity of Texas, Austin, geophysicist Sean rocks from the time of the dinosaur extinc- impact shocks would make rocks behave like fuids.
Gulick, cochief scientist for the $10 mil- tions. It was the signature, they said, of a Deep crustal rocks could end up on top of rocks of
lion project, sponsored by the International previously unsuspected cause of the extinc- shallower origin.
Ocean Discovery Program (IODP) and the tions: a giant impact. In 1991 Hildebrand and
International Continental Scientic Drill- colleagues ngered the village of Chicxulub 1. Postimpact excavation and beginning of uplift.
ing Program. All the other ones are either as the site of the cataclysm, nding quartz
on another planet, or theyve been eroded. crystals shocked by an impact in samples Ejecta curtain
At the end of March, a specially equipped from the Pemex wellssamples that had sat
vessel will sail from the Mexican port of around for more than a decade. Some peo-
Impact
Progreso to a point 30 kilometers ofshore. ple are a little embarrassed about that these Upper crust melt
There, in water 17 meters deep, the boat days, he says.
will sink three pylons and raise itself above The data from the Pemex wells were Lower crust
the waves, creating a stable platform. By spotty, and so scientists have always wanted
1 April, the team plans to start drilling, to go back for a detailed look at the impact 2. Central uplift
quickly churning through 500 meters of and its aftermath, says cochief scientist becomes unstable.
limestone that were deposited on the sea Joanna Morgan of Imperial College Lon-
oor since the impact. After that, the drill- don. It seems like a lifetimes ambition
ers will extract core samples, in 3-meter- coming true, says Morgan, who rst pro-
long increments, as they go deeper. For posed a scientically cored well to the
2 months, they will work day and night in IODP in 1998. Although ofshore drilling
an attempt to go down another kilometer, is expensive, she says that working at sea 3. Uplift collapses to form peak ring.
looking for changes in rock types, cata- means the team will face fewer hassles Peak ring
loging microfossils, and collecting DNA with environmental permitting and wont
samples (see gure, p. 1016). Weve got have to cope with the Yucatns poor roads.
one shot to try and get this down to 1500 In 2005, Morgan and Gulick led a $2 mil-

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1015


Published by AAAS
NEWS | I N D E P T H

lion remote-sensing campaign that used ed the oceans, contributing to the extinc- ter impact. Above that would be sediments,
small seismic explosions to help illuminate tions, so the drill team will look at whether since hardened into rock, that were swept
the subterranean structures and pinpoint seaoor animals just after the impact were in as the ocean rushed into the vast new de-
the best spot to reach the peak ring. species that tolerate low pH. pression. The impact layer may be capped
As the drill approaches the crater, Just above the crater lies an impact layer, by hardened deposits of ash that persisted
800 meters down, scientists expect to nd 100 meters or more thick, that would have in the atmosphere for weeks or more before
fewer species of the shell-producing ani- been deposited in the weeks after the cata- falling out.
mals that make up the limestone, because clysm. At its base, scientists expect to nd For many of the IODP scientists, the
life was just recovering from the impact. a hodgepodge of chunks of bedrock blasted main event will be reaching the peak ring.
Some scientists think that carbon dioxide up by the impact and once-molten rock that Peak rings abound on the moon, Mercury,

CREDITS: (DIAGRAM) V. ALTOUNIAN/SCIENCE; (PHOTOS) CHRIS JENSON, DES; DAVE SMITH, ECORD; MICHAEL RYMER; DAVID KRING; GEOLOGICAL SOCIETY OF AMERICA; ARTOGRAPHY/SHUTTERSTOCK; VILAX/SHUTTERSTOCK; HINRICH BAESEMANN/ALAMY
released by the impact would have acidi- fell back into the crater in the minutes af- and Mars. But on Earth, there are just two
craters larger than Chicxulub that should
also have peak rings: the 2-billion-year-
Going deep old Vredefort crater in South Africa, and
Using a three-legged lift boat to create a stable drilling platform, scientists will drill 17 m the 1.8-billion-year-old Sudbury crater in
1500 meters deep, and back in time. Some layers took millions of years to form, whereas Canadayet they are so old that the peak
others were laid down in minutes. rings have eroded away.
The IODP team wants to test a leading
model for peak ring formation, in which
granite from Earths depths rebounds after a
major impact, like water struck by stone, to
Drill site form a central tower, taller than the crater
rim. In minutes, the tower would collapse
and collide with material slumping in from
the rims to form the peak ring. Conrma-
tion for the model could come from nding
1 rocks out of order: deep rocks, probably
Peak ring Crater rim 500 m
granite, brought up in the central tower, ly-
ing atop originally shallower younger rocks.
Theyre going to test whether our numeri-
1 Switching of the bits (500 m)
At frst, no cores will be taken in layers of limestone.
cal models are making any sense or not,
After casing the hole with steel, scientists will switch says Jay Melosh, a planetary scientist at
2 550 m
from a tungsten-carbide roller cutter to a diamond-tipped Purdue University in West Lafayette, Indi-
system that can retrieve 3-meter-long cores. ana, who helped develop the model.
The team is interested not just in the
2 PaleoceneEocene Thermal Maximum (550 m) structure of the peak ring rocks but in what
A time 55 million years ago when Earth was life they might host. Remote sensing has al-
600 m
5oC warmer. The warmth may have triggered algal
ready suggested that the peak ring is less
blooms that died and fell to the sea Poor. These might
3 dense than expected for a granitea sign
show up in the rock record as black shales amid
the limestone. that the rocks are porous and fractured in
places. It is possible that these fractures, in
3 Life returns (550650 m) 4 650 m the wake of the impact, were lled with hot
The scientists will look for life coming back after the uids. Those will be preferred spots for mi-
impact, as the shelled animals that make up limestone crobes to grow, but it depends whether the
return and evolve into new species. Moving downhole fractures have energy and nutrients, says
and back in time to the impact, the size and number
Charles Cockell, an astrobiologist on the
of species of forams and coccolithophores is
expected to shrink. 700 m IODP team at the University of Edinburgh.
When the drill bit encounters mineral veins
4 Impact layer (650800 m) or other fracture zones in the peak ring,
Rocks at the base may contain chunks of rock and Cockell and his colleagues will take a sub-
impact melt. The ocean rushing in to fll the crater hole
core from the core: a biopsy on the geopsy.
could have deposited overlying sediment, perhaps 750 m
containing glassy impact spherules. Settling on top They will count and culture any microbes
could be fne ash containing crystals of shocked quartz. still living in the fractures, and sequence
DNA to look for the genes responsible for
5 Peak ring (8001500 m) metabolic pathways.
If formation models are correct, peak ring rocks Those genes might show that peak ring
5 800 m
probably volcanic granitescould be sitting out of microbesdescendants of those that lived
order. Deeper granites, with darker minerals, could rest after the impactderive their energy not
on top of granites with lighter minerals. 6 from carbon and oxygen, like most mi-
6 Microbial life (8001500 m)
crobes, but from iron or sulfur deposited
by hot uids percolating through the frac-
The peak ring is expected to be fractured and flled with
mineral veins that were once part of a vast hydrothermal tured rock. And that would mean the im-
system. Scientists will look for DNA evidence that 1500 m pact crater, harbinger of death, was also a
chemosynthetic microbes live in the cracks. habitat for life.

1016 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
SCIENTIFIC WORKFORCE

NSF makes a new bid to boost diversity


Initiative aims for fresh answers, but researchers are still framing the questions
By Jefrey Mervis more community-based science programs needs to be based on good theory. The
or boosting the number of underrepre- agency convened the meeting to help cre-

I
n what could be the signature program sented minorities earning Ph.D.s in STEM ate a science-based approach to reach con-
of her 6-year term, National Science (science, technology, engineering, and clusions on what works and what doesnt
Foundation (NSF) Director France mathematics) elds. work, said Mumpower, who last year re-
Crdova last week launched a 5-year, The INCLUDES launch generated some turned to Texas A&M University, College
$75 million initiative aimed at creating positive buzz at the 2526 February work- Station. But for the moment we are forced
a more diverse technical community. shop, held next door to NSF headquarters to rely on intuition and seat-of-the-pants
Its a laudable goal, say researchers who in northern Virginia. The 60 participants eforts.
attended an NSF-organized workshop on have extensive experience in both the re- INCLUDES traces its roots to a 2012
the science of broadening participation search and practice of broadening partici- report by an NSF advisory committee on
held just 2 days after the new program pation, and the fact that there were three equal opportunity that called for a bold
was announced. But it may be out of reach, times as many women of color in the room new initiative. Last summer Crdova as-
they add, without a greater federal invest- as white men reected the inclusiveness sembled an internal working group to esh
ment in understanding the problem. that NSF aims to foster. Fealing, a former out the idea and build support for it both
Despite decades of work, scholars are NSF program manager, said she hoped inside and outside the agency.

Downloaded from on March 3, 2016


still struggling to understand Ira Harkavy, a social scientist
why women, many minori- at the University of Pennsylvania
ties, people with disabilities, and chair of the advisory commit-
Too little diversity in science
and other groups are under- tee, says hes thrilled that Cr-
A new National Science Foundation initiative aims to fnd ways to broaden par-
represented in U.S. science and ticipation of women, Hispanics, African-Americans, and other minority groups dova has embraced the initiative.
engineering. NSF may be put- underrepresented in the U.S. science and engineering workforce. The committee plans to keep a
ting the cart before the horse close eye on INCLUDES, he adds,
in funding potential solutions 80 with an upcoming report on how
without knowing the underly- to ensure that the results of IN-
70
ing factors, says Kaye Husbands CLUDES wont simply become
Fealing, an economist at Georgia 60 fodder for scholarly articles and
Institute of Technology in At- future grant applications.
50
Percentage

lanta, who organized the work- Crdova ofered few specics


shop and says she welcomes the 40 in an interview this week. While
new program. Crdova promises we will continue to fund the pro-
30
that the initiative will break new grams we already haveand I
ground. Im not interested in 20 want to underline that, because
funding more of the same, she they are really good programs
10
told Science. we are looking for something
Since arriving at NSF in March 0 very diferent here. She cited a
2014, Crdova has spoken re- White Asian Hispanic African- Other All long-running NSF program to
American women
peatedly about why broadening help women succeed in academic
participation is not just the right Those with S&E jobs Those with S&E degrees U.S. population scienceknown as ADVANCE
thing to do, but also how diver- as an example of the limitations
sity fuels better research and bolsters the some INCLUDES pilot projects will lead to of current approaches.
U.S. economy. The new initiativedubbed standard NSF grants that strengthen the I was an ADVANCE [principal investi-
DATA: DATA FROM 2013, SCIENCE AND ENGINEERING INDICATORS 2016.

INCLUDES (Inclusion across the Nation science of broadening participation. gator] at Purdue University, she recalled.
of Communities of Learners of Under- But there were also expressions of con- But as hard as we were working and as
represented Discoverers in Engineering cern, starting with how INCLUDES will efective as the program was locally, the
and Science)will give 40 researchers difer from NSFs existing $750-million-a- nationwide problem was not getting much
$300,000 each this year to test their year portfolio of dozens of programs that better. And with the countrys changing
ideas on fostering inclusion. Next year, a pursue similar goals, and how it might af- demographics, one could argue that the
second competition will award $12.5 mil- fect them. Researchers wondered how NSF gap between those with access to STEM
lion each over 5 years to ve larger teams, would measure the success or failure of IN- and those who dont is actually increasing.
called alliances. CLUDES projects. There were also worries Theres an old saying that if you keep do-
NSF wants proposals that lay out a that a 15 April deadline for preliminary ing the same thing to achieve a diferent so-
convincing strategy for scaling up eforts proposals is too soon. lution, youre not going to make progress,
through partnerships with industry, local In his opening remarks at the workshop, Crdova added. And the smaller amount of
and state governments, schools, and non- Jeryl Mumpower, former director of NSFs money [for INCLUDES] is because we need
prot groups. The NSF solicitation ofers division of social and economic sciences, to test new approaches. And if they are suc-
the hypothetical examples of sponsoring cited a broader concern: Good policy cessful, we can build on that investment.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1017


Published by AAAS
FEATURES

Downloaded from on March 3, 2016


A
long, grinding struggle back to
normal is underway at Fuku-

SLOW BURN
shima. As workers make progress
in cleaning up contaminated land
surrounding the reactor, evacuees
are grappling with whether to
return to homes sealed of since
the accident. The power plant it-
self remains a dangerous disaster
Five years after the meltdowns zone, with workers just beginning the com-
plex, risky job of locating the melted fuel
at the Fukushima atomic plant, evacuees and guring out how to remove it.
The magnitude 9.0 earthquake that struck
face the prospect of living near a nuclear northeastern Japan on 11 March 2011 and the
40-meter tsunami that followed left 15,893
disaster zone for decades to come dead and 2572 missing, destroyed 127,290
buildings, and damaged more than a mil-
By Dennis Normile lion more. It also triggered the meltdowns at
the Fukushima Daiichi Nuclear Power Plant
and the evacuation of 150,000 people from

1018 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
NEWS | F U K U S H I M A L E G AC Y

A crane lifts a panelpart of a diation exposure levels, they feel, are still too TEPCO has found ways to divert ground-
temporary canopyto gauge high. There has been no education regard- water from the site, cutting inltration to
the effects of dust suppression ing radiation, says Katsunobu Sakurai, the about 150 tons per day. Now its about to
inside the wreckage of Fukushima mayor of Minamisoma, where 14,000 people freeze out the rest. Borrowing a technique
Daiichis Unit 1 reactor building. were evacuated after the accident. Its dif- for making temporary subsurface barriers
cult for many people to make the decision during tunnel construction, a contractor
to return without knowing what these radia- has driven 1500 pipes 30 meters down to
tion levels mean and what is safe, he says. bedrock, creating something akin to an un-
Some citizen groups are suing the national derground picket fence encircling the four
government and Tokyo Electric Power Com- crippled reactor units. Brine chilled to 30C
pany (TEPCO), the Fukushima plants owner, circulating in the pipes will freeze the soil
over plans to end compensation payments between the pipes; the frozen wall should
for those who choose not to return home. keep groundwater out and contaminated
Highly contaminated areas close to the nu- water in. TEPCO was planning to start the
clear plant will remain of limits indenitely. operation shortly after Science went to press.
Conditions at the plant are
really stable, the plant man-
ager, Akira Ono, recently told
reporters. Radioactivity and
heat from the nuclear fuel have
fallen substantially in the past
5 years, he says. But cleanup is
of to a slow start, hampered
by sketchy knowledge of where
the nuclear fuel is located. Last
year managers agreed to a road
map for decommissioning the
site over the next 30 to 40 years
that calls for removing melted
nuclear fuel masses and demol-
ishing the plants four reactor
halls at a cost that could top $9
billion. TEPCO intends to start
removing nuclear debris from
the reactors in 2021.
Ono puts the decommission-
ing at around 10% complete.
One big hurdle was cleared in
December 2014, when crews re-
moved the last of 1535 fuel rods
stored in the Unit 4 spent fuel The Fukushima plant is crowded with 10-meter-tall tanks storing
pool. At the time of the accident, tainted water used to cool melted nuclear fuel masses and
some feared that cooling water groundwater that infiltrated the sitesome 750,000 tons in all.
had drained out of the pool and
within 20 kilometers of the nuclear plant as exposed the fuel to air, which might have led The most daunting task is recovering the
well as from areas beyond that were hard hit to overheating and melting. It hadnt, but fuel debris. TEPCO modeling and analyses
by fallout. the fuel remained a threat. suggest that most, if not all, of the fuel in
Now, the nuclear refugees face a dilemma: The biggest challenge at present, Ono says, the Unit 1 reactor melted, burned through
How much radiation in their former homes is contaminated water. Cooling water is con- the reactor pressure vessel, dropped to the
is safe? In a herculean efort, authorities tinuously poured over the melted cores of bottom of the containment vessel, and per-
have so far scooped up some 9 million cubic units 1, 2, and 3 to keep the fuel from over- haps ate into the concrete base. Units 2 and
meters of contaminated soil and leaves and heating and melting again. The water drains 3 sufered partial meltdowns, and some fuel
washed down buildings and roadways with into building basements, where it mixes may remain in the cores.
the goal of reducing outdoor radiation ex- with groundwater. To reduce the amount of To try to conrm the location and con-
PHOTOS: THE ASAHI SHIMBUN VIA GETTY IMAGES

posure to 0.23 microsieverts per hour. Last contaminated water seeping into the ocean, dition of the melted fuel, the International
September, the government began lifting TEPCO collects and stores it in 10-meter-tall Research Institute for Nuclear Decommis-
evacuation orders for the seven municipali- steel tanks. They now ll nearly every corner sioning, set up by TEPCO and other enti-
ties wholly or partly within 20 kilometers of of the grounds, holding some 750,000 tons ties, has been probing the reactors innards
the plant. As the work progresses, authori- of water. The government is evaluating ex- with muons. Wispy cousins of the electron,
ties expect that 70% of the evacuees will be perimental techniques for cleansing the wa- muons are generated by the trillions each
allowed to return home by spring 2017. ter of a key radioisotope, tritium (Science, 12 minute when cosmic rays slam into the up-
But evacuees are torn over safety and December 2014, p. 1278). Ono says a solution per atmosphere. A few muons are absorbed
compensation issues. Many claim they are is sorely needed before the plant runs out of or scattered, at a rate that depends on a ma-
being compelled to go home, even though ra- room for more tanks. terials density. Because uranium is denser

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1019


Published by AAAS
NEWS | F E AT U R E S | F U K U S H I M A L E G AC Y

than steel or concrete, muon imaging can


potentially locate the fuel debris.
In February 2015, a group at Japans High
Energy Accelerator Research Organization
in Tsukuba supplied two van-sized muon
detectors, which TEPCO placed adjacent to
the Unit 1 reactor at ground level. After a
TRIAL BY
month of collecting muons, the detectors
conrmed there was no fuel left in the core.
Because they were positioned at ground
level, the devices could not image the re-
actor building basements and so could not
pin down where the fuel is or its condition.
MELTDOWN
At Fukushima, robots performed poorly at first
TEPCO plans to use robots to map the loca-
tion of the fuel debris so it can develop a but are now critical players in the cleanup
strategy for removing it (see story, right).
A second team has developed detectors By Timothy Hornyak, in Tokyo task of decontaminating and dismantling
that observe muons before and after they the reactors. Roboticists are making halt-

W
pass through an object of interest, promis- atch Hondas Asimo humanoid ro- ing progress in developing machines for
ing a more precise picture of reactor interi- bot up close, and you may fall un- specic tasks, such as decontaminating and
ors. For Fukushima, the researchersfrom der its spell. As it walks gracefullyremoving melted nuclear fuel masses, but
Los Alamos National Laboratory in New across a stage, sometimes opening they know that their creations need to be
Mexico and Japans Toshiba Corp.built bottles or serving tea, the elegant adaptable. We must prepare for unforeseen

Downloaded from on March 3, 2016


mammoth detectors, 7 meters across, mechatronic man seems at home situations beyond the scope of expectation
which they intended to place outside Unit anywhere. So after a 9.0-magnitude earth- and imagination, says Satoshi Tadokoro, a
2. That work has been postponed because quake and tsunami waves struck Japans roboticist at Tohoku University in Sendai,
TEPCO decided to rst send a robot into Fukushima Daiichi plant on 11 March 2011, Japan, who led development of Quince.
the containment vessel; high radiation lev- it was no wonder that one of Asimos fans
els have delayed that plan. Our current tweeted, Cant Asimo be dispatched to sur- THE STRUGGLES of Japanese roboticists
task is to reduce that exposure, Ono says, vey the interior of Unit 4, where radiation is came despite a national program in the
using robotic oor and wall scrubbers in too high for human workers? 1980s to develop robotics for nuclear power
the area workers need to access to deploy If only. In the early days of the crisis, en- plants. One of the rst was AMOOTY, which
the robot. gineers were desperate to learn about the Toshiba and University of Tokyo research-
While the authorities struggle to clean up damaged reactors cores and the radiation ers built in 1985. The radiation-hardened
the site and resettle residents, some locals levels inside the buildings, data that robots crawler could climb stairs and manipulate
are judging safety for themselves. In 2014, a should have been able to provide. Alas, objects in mock power plants. But in 1999,
group of enterprising high school students neither Asimodesigned to when an accident at a ura-
in Fukushima city, outside the evacuation navigate sedate ofcesnor nium reprocessing facility
zone, launched an international radiation- any other of Japans vaunted in Tokaimura killed two
dosimetry project. Some 216 students and robots was up to the chal- We must prepare workers and contaminated
teachers at six schools in Fukushima Prefec- lenge of navigating Fukushi- the plant, AMOOTY was
ture, six elsewhere in Japan, four in France, mas complex, debris-strewn
for unforeseen still considered too experi-
eight in Poland, and two in Belarus wore do- interiors. The plants op- situations beyond mental to use.
simeters for 2 weeks while keeping detailed erator, Tokyo Electric Power In the years after To-
diaries of their whereabouts and activities. Company (TEPCO), had to
the scope of kaimura, Japan developed
I wanted to know how high my exposure turn to U.S. military-grade expectation and other nuclear disaster ro-
dose was and I wanted to compare that dose robots, such as iRobots bots and imported Menhir,
with people living in other places, explains 510 PackBot, to get its rst imagination. a large mobile unit built by
Haruka Onodera, a member of Fukushima glimpses inside the facility. Satoshi Tadokoro, Frances Cybernetix thats
High Schools Super Science Club, which Only in June 2011, 2 months Tohoku University equipped with radiation
conceived the project. The students pub- later, did a modied Japa- shielding, cameras, and a
lished their ndings last November in the nese rescue robot called Quince enter one of manipulator. But well before the Fukushima
Journal of Radiological Protection. Their the ruined complex buildings. disaster, TEPCO and other industry testers
conclusion: High school students in Fuku- When Fukushima occurred, I was as- judged them to be too big, slow, and inef-
shima [Prefecture] do not sufer from signif- tounded that Japan had no robots to help fective. Government funding ran out, and
icantly higher levels of radiation than those out in any signicant way, says Frederik Menhir is now on display at Tohoku Univer-
living elsewhere, Onodera says. Schodt near San Francisco, California, au- sity. The other machines were mothballed
Thats good news for Fukushima city thor of Inside the Robot Kingdom: Japan, or cannibalized for parts, according to The
residents, perhaps, but cold comfort to dis- Mechatronics, and the Coming Robotopia. Asia-Pacic Journal: Japan Focus; mis-
placed people now weighing the prospect of It practically brought me to tears. placed faith in the safety of nuclear plants
moving back to homes closer to the shat- Five years after the accident, however, and lack of long-term funding scuppered
tered nuclear plant. robots are nally ready to enter the ruined the program.
reactors en masse. They are now expected After the catastrophe of 11 March 2011,
With reporting by Timothy Hornyak. to play an essential part in the daunting the rst task for robots was to survey dam-

1020 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
age, radiation, and variables such as tem- actor building. More sophisticated versions, sion. Honda developed a robot based on
perature and humidity in areas that were dubbed Rosemary and Sakura, were also Asimos joint-stabilization technology that
too hazardous for workers. The PackBot sent into the reactor buildings. Sakura acts can extend 7 meters vertically to inspect
droids explored the ground oors of the as a communications relay, and Rosemary is upper nooks. Every robot needed difers
Unit 1 and Unit 3 reactor buildings, and equipped with a U.K.-developed system that according to its purpose and the damage,
found maximum radiation levels of tens of combines radiation meters, a sheye cam- says Tomohisa Ito, a spokesperson for the
millisieverts (mSv) per hour. Workers ex- era, and a laser rangender to produce 3D International Research Institute for Nuclear
posed to such levels for less than a workday radiation maps. Decommissioning here, a consortium of nu-
would exceed their emergency safety limit By now, nearly a dozen robots have been clear plant companies that aims to develop
for an entire year, 250 mSv. The radiation developed to get closer looks at the heart of new technologies for cleaning up the Fuku-
and temperature maps made by the Pack- the plant. Some oat or swim through pools shima plant.
Bots and two iRobot 710 Kobra bots allowed that have formed in the buildings bowels Robots slow start at Fukushima holds
TEPCO workers to plot paths for quick for- because of the constant need to inject wa- a broader lesson, says Gill Pratt in Bos-
ays into the reactor buildings with the least ter to cool the damaged cores. Two snake- ton, an engineer who led the Defense Ad-
radiation exposures, says iRobot Vice Presi- like robots crawled through a pipe leading vanced Research Projects Agency Robotics
dent Tim Trainer in Bedford, Massachusetts. into the 48-meter-tall primary containment Challenge from 2012 to 2015 before joining
But the complex structure of the reactor vessel in the Unit 1 reactor to ascertain the Toyota to head its new articial intelligence
buildings ground and upper oors, with state of melted fuel masses. Although one lab. Roboticists, he says, need to develop
many staircases and thick concrete walls, got stuck, the machines returned valuable emergency robotic equipment that can be
challenged robot mobility and wireless com- video and dose information, according to deployed immediately without the need
munications. The iRobot bots couldnt climb TEPCO. In the dark and vaporous interior, for additional training or adaptation. The
slippery stairs or turn corners easily. Once it the robots measured radiation in one area great lesson of Fukushima, Pratt says, is
was clear that more capable machines were as high as 25 sieverts per hourenough to that disasters are often fast moving and dif-
needed, TEPCO and the government con- kill a person in minutes. cult to predict events, where the window
tacted roboticists at Chiba Institute of Tech- Makers are now developing robots that of time for efective intervention is small.
nology and Tohoku University. They hastily can tackle specic decommissioning chores. In the immediate aftermath of a future
modied Quince, a survey droid on caterpil- For example, Toshiba has developed ma- nuclear accident, robots, and their masters,
lar treads that climbs stairs and debris, by chines that decontaminate surfaces with will have to be far nimbler. But at least at
equipping it with two cameras, a dosimeter, blasts of dry ice, inspect vent pipes for Fukushima, they are evolving.
and a power and communications cable leaks, and cut and remove debris covering
that stretched hundreds of meters. Quince fuel rod assemblies in the Unit 3 building, Timothy Hornyak is a freelance reporter
explored the upper oors of the Unit 2 re- which was damaged by a hydrogen explo- based in Tokyo.

Into the hot zone


At the stricken Fukushima nuclear power plant, robots are performing an increasing variety of tasks in areas that are
too radioactive for people to work in. Here are a few of the top performers.

Quince, Rosemary, and Sakura Quadruped inspection robot ASTACO-SoRa Raccoon


TOHOKU UNIVERSITY/CHIBA KODAI TOSHIBA HITACHI GE NUCLEAR ENERGY ATOX
ILLUSTRATION: K. SUTLIFF/SCIENCE

A modifed earthquake rescue robot, With its four multijointed legs, this Whereas most robots at the Fukushima Working with a relay robot that
Quince was the frst Japanese bot to 65-kilogram droid has a mobility plant have been used for surveys, this wrangles its water hoses, Raccoon is a
enter the nuclear plants wrecked advantage over crawling robots: It can 2.5-ton shovel-wielding robot has done nuclear-hardened maid that has been
buildings. Smaller and nimbler than its climb stairs and access hard-to-reach yeomans work removing debris such scrubbing the foors of Unit 2. Its
U.S. military counterparts, Quince and crannies. The quadruped inspection as sheet metal, fallen ducts, concrete decontamination work has helped
its successors Rosemary and Sakura robot verifed that there were no leaks chunks, and nitrogen canisters from tamp down radiation levels to the point
have gathered data for radiation maps, in the Unit 2 primary containment reactor Units 1 and 3. Its two arms are that human workers can access the
sampled airborne radioactive particles, vessel, the crucible holding the equipped with swappable manipulators interior for decommissioning tasks.
and monitored radiation dose rates. partially melted down reactor. such as grippers and rotating blades. Raccoon doesnt do walls or windows.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1021


Published by AAAS
NEWS | F E AT U R E S | F U K U S H I M A L E G AC Y

EPIDEMIC OF FEAR
A bumper crop of thyroid abnormalities in Fukushima children,
including cancer, has perplexed scientists and alarmed locals

By Dennis Normile interpretations. Many acknowledge that over, the day after the meltdowns, Japanese
baseline data from noncontaminated areas authorities evacuated some 150,000 people

T
he March 2011 meltdowns at the Fu- were needed from the outset and that the living within 20 kilometers of the plant,
kushima Daiichi Nuclear Power Plant public should have been better educated to and a week later they started screening for
caused extensive human suffering understand results and, perhaps, to accept contaminated food. In addition, Fukushima
evacuations, emotional trauma and watchful waiting as an alternative to imme- residents were ofered iodine tablets after
premature deaths, disrupted jobs diate surgery. But most also say the ndings the accident to block absorption of any ra-
and schooling. What they have not hint at a medical puzzle: Why are thyroid dioactive iodine that managed to nd its
caused, so far, is radiation-related illness abnormalities so common in children? The way into the food supply.
among the general public, and few special- surprising results of the screening, Wil- In 2013, WHO estimated that the 12 to
ists expect dramatic increases in cancers or liams says, show that many more thyroid 25 millisieverts (mSv) of exposure in the

Downloaded from on March 3, 2016


other ailments. The reactors spewed just a carcinomas than were previously realized rst year after the accident in the hard-
tenth of the radiation emitted by the Cher- must originate in early life. est hit areas might result in minuscule in-
nobyl disaster, winds blew much of that out creases in cancer rates. (Worldwide, people
to sea, and evacuations were swift. Yet one MEMORIES OF CHERNOBYL got Japanese receive on average 2.4 mSv per year from
wave of illness has been linked to the disas- authorities worrying about thyroid cancer. background radiation; a medical chest x-
terthe ironic result of a well-intentioned The fallout from that April 1986 accident ray delivers about 0.1 mSv.) WHO noted
screening program. included radioactive iodine, which settled that females have a 0.75% lifetime risk of
Months after the disaster, Fukushima across swathes of Belarus, developing thyroid cancer;
Prefecture set about examining the thy- Russia, and Ukraine, con- it estimated that the highest
roids of hundreds of thousands of children taminating pastures grazed exposures in the Fukushima
and teens for signs of radiation-related
cancers. The screening efort was unprec-
edented, and no one knew what to expect.
by dairy cows. Children who
drank the tainted milk accu-
mulated the radioactive io-
300,476
Number of Fukushima
area raised that risk by an
additional 0.5%.
The initial round of thy-
So when the rst round of exams started dine in their thyroids. (Adult residents 18 and under roid screening, started in
turning up thyroid abnormalities in nearly thyroids absorb less iodine.) whose thyroid screening late 2011, was simply to pro-
half of the kids, of whom more than 100 A 2006 World Health Orga- results were available as vide baseline data, as any
were later diagnosed with thyroid cancer, a nization (WHO) study found Science went to press. radiation-induced tumors
restorm erupted. that in the most contami- were not expected to emerge
One result, says Kenji Shibuya, a public
health specialist at University of Tokyo, was
overdiagnosis and overtreatment, leading
nated areas, there had been
about 5000 thyroid cancer
cases among those who
50%
Approximate fraction of
for at least 4 years. Children
with nodules larger than 5.0
mm or cysts bigger than 20.1
dozens of children to have their thyroids were under 18 at the time of mm underwent a second,
those screened with solid
removed, perhaps unnecessarily. Activists the accident, though the re- more detailed examination
nodules or fluid-filled cysts
trumpeted the ndings as evidence of the port noted that more cases on their thyroids.
and, if necessary, ne needle
dangers of nuclear power. The large num- could emerge over time. The aspiration. After the initial
ber of abnormalities appearing so soon af- United Nations in 2006 at- screening, children will have
ter the accident would indicate that these
children almost certainly received a very
high dose of thyroid radiation from inhaled
tributed 15 childhood thyroid
cancer deaths to Chernobyl.
Caught early, the cancer is
110
Number of thyroid cancer
their thyroids examined ev-
ery 2 years until age 20 and
every 5 years after that.
and ingested radioactive iodine, antinu- almost always cured by re- cases identified by Results were released as
clear crusader Helen Caldicott wrote in a moval of the thyroid gland. December 2014 as a result screening progressed, and
post on her homepage. With that in mind, Japa- of the screening. right from the start there
Scientists emphatically disagree. The nese authorities set out to were surprisingly high rates
evidence suggests that the great majority screen the thyroids of all 368,651 Fuku- of abnormalities. Findings from the initial
and perhaps all of the cases so far discov- shima residents who were 18 and under at round of screening, completed in April
ered are not due to radiation, says Dill- the time of the accident (Science, 5 August 2015 and released in August 2015, showed
wyn Williams, a thyroid cancer specialist 2011, p. 684). Most experts were not antici- that nearly 50% of the 300,476 subjects had
at University of Cambridge in the United pating a bumper crop of thyroid problems. solid nodules or uid-lled cysts on their
Kingdom. In journal papers and in a series For starters, the potential radiation expo- thyroids. Smaller studies elsewhere had
of letters published last month in Epidemi- sure of Fukushima residents was slight hinted that tiny thyroid cysts and nodules
ology, scientists have attacked the alarmist compared with Chernobyl victims. More- were common in all ages. But specialists

1022 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
wrote in Epidemiology. In a letter
to Epidemiology, Tsuda claims to
have addressed the screening efect
by adjusting the number of cancer
cases to account for the lag time
between when an ultrasound ex-
amination would catch the cancers
and when they could be clinically
identied. He did not address other
criticisms or respond to repeated
requests for comment from Science.
Although many scientists dis-
agree with the spin Tsuda and activ-
ists have put on the ndings, they
endorse the screening efort. A thy-
roid screening program would be
expected to save lives by detecting
cancers early, whether or not the
cancers were caused by radioactiv-
ity, says Timothy Jorgensen, a radia-
tion health physicist at Georgetown
University in Washington, D.C.
Yet it has become clear the public
and even many doctors did not have
the background to put the results in
perspective. Even though the vast
majority of thyroid abnormalities
are safe to ignore, nding small le-
sions causes patients anxiety, says
Seiji Yasumura, vice director of the
Fukushima Prefecture Health Man-
In Iwaki, a town south of the Fukushima nuclear plant, a doctor conducts a thyroid examination on 4-year-old Maria agement Survey. Virtually all of
Sakamoto. Scientists are puzzled over a high number of thyroid abnormalities observed so soon after the accident. those diagnosed with thyroid cancer
have had the glands removed, even
did not know whether the frequency [in the ditional clinical examinations of patients though accumulating evidence suggests
Fukushima results] was high or low, says who have lumps or symptoms. It is inap- in many cases it might have been better
Noboru Takamura, a radiation health scien- propriate to compare the data from the to wait, the University of Tokyos Shibuya
tist at the Atomic Bomb Disease Institute at Fukushima screening program with cancer adds. Careful observation would be the
Nagasaki University in Japan. registry data from the rest of Japan where best option.
As the number of conrmed cancers rose, there is, in general, no such large-scale South Korea ofers a cautionary tale. In
worries grew about a link to radiationand screening, Richard Wakeford, an epidemi- 1999, the South Korean government initi-
those concerns gained a high-prole pro- ologist at the University of Manchester in ated a health program in which care pro-
ponent. In 2013, Toshihide Tsuda, an en- the United Kingdom, wrote on behalf of 11 viders ofered ultrasound thyroid screening
vironmental epidemiologist at Okayama members of a WHO expert working group for a small additional feeand thyroid can-
University in Japan, started presenting on Fukushima health consequences. Theirs cer diagnoses exploded. In 2011, the rate of
analyses at international conferences was one of seven letters to Epidemiology thyroid cancer diagnosis was 15 times what
claiming the number of thyroid cancers in published online last month that blasted it was in 1993, yet there was no change in
the Fukushima screening was unusually Tsudas methodology and conclusions. thyroid cancer mortality, Heyong Sik Ahn
high. Last October, he published his results To see what comparable screening would of Korea University in Seoul and colleagues
online in Epidemiology, concluding that nd in a population not exposed to radia- reported in The New England Journal of
the rst round of screening indicated can- tion, Takamuras team used the Fukushima Medicine in November 2014. Virtually all
cer incidence rates ranging from 0 to 605 survey protocol to examine 4365 children those diagnosed underwent total or par-
cases per million kids, depending on loca- aged 3 to 18 from three widely separated tial thyroid removal. Most required lifelong
tion, but overall an approximately 30-fold prefectures. They found similar numbers thyroid-hormone replacement therapy. To
increase over the normal childhood cancer of nodules and cysts and one cancer, for a stem this epidemic, Ahn and others dis-
rate. That claim fed alarming headlines. prevalence of 230 cancers per million peo- courage routine thyroid cancer screening.
Other scientists were swift and severe in ple, as they reported in Scientic Reports Williams says the evidence suggests that
PHOTO: TORU HANAI/REUTERS

their criticism. A fundamental error, accord- in March 2015. Other Japanese studies thyroid growths among children are far
ing to several epidemiologists, is that Tsuda reported thyroid cancer rates of 300, 350, more common than previously thought
compared the results of the Fukushima and even 1300 per million. The prevalence and must be considered normal. The Fu-
survey, which used advanced ultrasound of thyroid cancer detected by advanced kushima survey, he says, promises a better
devices that detect otherwise unnotice- ultrasound techniques in other areas of understanding of the origins and develop-
able growths, with the roughly three cases Japan does not difer meaningfully from ment of such growths and may lead to bet-
of thyroid cancer per million found by tra- that in Fukushima Prefecture, Takamura ter treatment protocols.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1023


Published by AAAS
Turning the tap of

INSIGHTS
electron fow p. 1026
From syngas to olefns in
one step p. 1030

Downloaded from on March 3, 2016


PERSPECTIVES

RESEARCH INTEGRITY

Liberating eld science samples and data


Promote reproducibility by moving beyond available upon request

By Marcia McNutt,1* Kerstin Lehnert,2 duct research and analysis, as well as to the they even address such access, commonly
ILLUSTRATION: D. BONAZZI/@SALZMANART

Brooks Hanson,3 Brian A. Nosek,4 Aaron code used to analyze and process data and do so by simply noting data and samples
M. Ellison,5 John Leslie King6 samples, is a fundamental requirement for available upon request. They lag behind
transparency and reproducibility. The eld some laboratory sciences in making data

T
ransparency and reproducibility en- sciences (e.g., geology, ecology, and samples available to the broader re-
hance the integrity of research re- POLICY and archaeology), where each search community. It is time for this to
sults for scientic and public uses study is temporally (and often change. We discuss cultural, nancial, and
and empower novel research ap- spatially) unique, provide exemplars for the technical barriers to change and ways in
plications. Access to data, samples, importance of preserving data and samples which funders, publishers, scientic societ-
methods, and reagents used to con- for further analysis. Yet eld sciences, if ies, and others are responding.

1024 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
Repeating a study from start to nish National Science Foundations PASSCAL fying to journals completion of deposition
using new samples and equivalent pro- (Portable Array Seismic Studies of the Con- requirements. Because deposition alone does
cedures under identical conditions is the tinental Lithosphere) and OOI (Ocean Ob- not guarantee that data or samples are dis-
ideal. This may be practical in laboratory servatory Initiative) projects are examples coverable or usable, machine-readable, qual-
sciences but is rarely possible in eld sci- of programs that are acclimating commu- ity-controlled, public metadata would help
ences. Objects of study might be ephemeral nities to more progressive data-sharing researchers nd, understand, and use the
(Superstorm Sandy), exceptionally rare policies, demonstrating research gains, and resources. Automated services can promote
(Dreadnoughtus), or forever changing (suc- undermining arguments against open data. cost-efective and data-conscious research
cession in a forest or how climate afects a Initiatives such as OpenAIRE and the cultures, while providing incentives to data
prairie ecosystem). Nevertheless, transpar- Center for Open Science are advancing collectors by demonstrating impact through
ency and reproducibility have substantial principles and guidelines for transparent metrics of views, downloads, and data and
value for eld sciences. Independent analy- research and publication. More than 500 sample uses. The Open Science Framework
sis of original data can uncover statistical journals have indicated intent to review (http://osf.io/) connects repositories so au-
or coding errors, data selection bias, or the TOP (Transparency and Openness Pro- thors can nd domain-specic repositories
problems with observations that are too motion) (5) publication guidelines for po- and provides links to general-purpose reposi-
good to be true. Original analyses may be tential implementation. However, journals tories for unusual classes of data (7).
augmented with new techniques to test that require that data and samples be de- Repositories can expand assignments of
novel questions. Data and samples can be posited upon publication may sufer if jour- digital object identiers (DOIs) to data sets
combined across studies for more precise nal staf lack resources to verify deposition, to aid citation. International initiatives, in-
or generalizable tests [e.g., the Paleobio- evaluate metadata, or check code accessi- cluding DataCite (www.datacite.org) and
logical Database (1) or PetDB database (2)]. bility of code or other materials. the IGSN e.V. (www.igsn.org, International
Such eforts must recognize that Geo Samples Numbers), develop
motivations for promoting trans- and promote common conventions
parency and reproducibility vary for unambiguous identication and
by stakeholder. Researchers want All should credit data creators and citation of data and samples to en-
to produce knowledge in new di- accelerate recognition of the value of hance usability, to guard against
rections and to get credit for their loss, and to provide credit to cre-
contributions. Funders want to see data in thesystem. ators. Publishers can adopt resolu-
greater value from their investment. tions such as the Joint Declaration
Journals want to facilitate reproducible sci- Funding agencies and journals can guide of Data Citation Principles (8) that specify
ence. Repositories want to support their expectations and set requirements, but top- data citation in the references. Journals can
communities and streamline data ow. down mandates alone are unlikely to foster publish descriptions of data sets and meth-
needed cultural changes in scientic com- ods. All should credit data creators and ac-
FUNDING, PUBLISHING, AND CULTURE munities. Research culture prioritizes pub- celerate recognition of the value of data in
CHANGE. Transparency and reproducibil- lications, innovation, and insight, which the scientic reward system.
ity in scientific research require invest- puts data stewardship and reuse far down Despite many eforts, there remains wide-
ment. Quality control is costly, and even the list. Data professionals in large proj- spread disagreement regarding data and
large projects can have difficulty curating ects too often are invisible team members. sample availability and metadata, as well as
data to make them useful for others. For To change this, bottom-up approaches are uneven sample deposition across the eld
example, many data sets in ecology and needed to earn community buy-in and ef- sciences. Journals might use DOIs for data or
evolution publications are incomplete fect culture change to recognize research- accession numbers for curated samples (e.g.,
or inaccessible (3). Funders that support ers who develop and curate original data IGSNs), but these are not used routinely.
domain-specific data and sample reposito- (6): Principal investigators can lead by Journals that ensure precise descriptions of
ries do so because they advance science by example; universities can ofer education variables, data, or sample provenance and
facilitating preservation and reuse, as well in data and samples management; and that provide details concerning collecting
as support data professionals who help scholarly societies can recognize excellence permits and requirements or restrictions for
with collection, management, and curation. in data and samples stewardship in their reuse can help. However, such information
U.S. and UK funding agencies require awards, selection of fellows, and leader- must come from those closest to the data and
publicly supported researchers to provide ship. Development of data journals with samplesthe researchers themselves. Scien-
data management plans and use open re- citable output can provide data experts ac- tists and technical staf must know how to
positories for data and samples. Investiga- knowledged leadership roles. create such information and be supported by
tors are often allowed reasonable amounts funding agencies to do so. Repositories could
of time between collection and deposition REPOSITORIES, CITATIONS, AND CURA- provide user-friendly software and training
(e.g., up to 2 years) for quality control or TION. Although general-purpose reposito- to eld teams to overcome resistance to shar-
publication priority, but making data and ries are useful for data that do not fit easily ing data and samples.
samples available sooner can advance within a specific domain, the value of pre- Scientic societies can work with stake-
science more rapidly [e.g., (4)]. The U.S. served data is best maximized by discipline- holders to set guidelines for provenance
specific repositories. They can provide of metadata; access; and security or ethi-
1
Editor-in-Chief, Science, Washington, DC 20005 USA.
services informed by community priorities, cal restrictions (e.g., for protecting human
2
Columbia University, Palisades, NY 10964 USA. 3American which can promote cultural change. subjects or rare and endangered samples);
Geophysical Union, Washington, DC 20009 USA. 4University Repositories can improve quality control and cooperate with all journals, not just
of Virginia and Center for Open Science, Charlottesville, VA and compliance, but progress will be faster those they own or sponsor. Community
22903 USA. 5Harvard University, Petersham, MA 01366 USA.
6
University of Michigan, Ann Arbor, MI 48109 USA. if repositories move beyond curation and partnerships [e.g., COPDESS (Coalition for
*E-mail: mmcnutt@aaas.org provide value-added services, such as certi- Publishing Data in the Earth and Space Sci-

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1025


Published by AAAS
INSIGHTS | P E R S P E C T I V E S

ences) (9)] can provide discipline-specic P H YS I C S


metadata and efective quality control with
secure badging for journal verication. The
Research Data Alliance (RDA) is developing
approaches and infrastructure for the pub-
Electrons go with the ow
lishing community.
Finally, not every sample can be saved.
Museums and other special-purpose re-
in exotic material systems
positories (e.g., ice-core labs) face resource Electronic hydrodynamic owmaking electrons
and space limitations. Curators must decide ow like a uidhas been observed
what to keep. Samples supporting peer-re-
viewed publications should have priority.
Digitized samples and collection informa- By Jan Zaanen the space they are moving in is made in-
tion or other metadata will facilitate remote homogeneous by, for example, putting

T
examination. Digital catalogs can provide urn a switch and the light goes on. the water in a pipe such that the overall
persistent access to metadata on samples The laymans perception is that this momentum relaxes and the kinetic energy
used in publications. These should include is like opening a tap so that the wa- turns into heat. Electrons in solids, how-
information on access linked to publica- ter starts running. But this analogy is ever, move in a background of static ions,
tions via resolvable unique identiers such misleading. The ow of water is gov- breaking this translational invariance, and
as the IGSN. The System for Earth Sample erned by the theory of hydrodynam- imperfections occur even in the most per-
Registration (SESAR), iDigBio, and Cyverse ics, whereby the behavior of the uid does fect periodic crystals. It is now a matter of
provide examples of metadata proles. not require knowledge of the motions of numbers. Could it be the case that an indi-
By working together, stakeholders can individual molecules. Electrical currents in vidual electron can lose its momentum be-

Downloaded from on March 3, 2016


create a virtuous cycle of increasing data solids, however, are formed from electrons. cause of scattering from the ionic disorder
and sample accessibility. The days when sci- In metals, these do not collide with each before it meets another electron (Knudsen
entists held on to samples and data hoping other, but they do scatter from lattice im- ow) (see the gure, left panel), or will
to squeeze out one more publication are end- perfections. The resulting Knudsen ow the electron uid equilibrate rst through
ing. Sharing can be more productive than of electrons is reminiscent of the avalanche many electron-electron collisions without
hoarding when researchers get credit for use of balls cascading through a dense forest of noticing the imperfections (hydrodynamic
of their data or samples. The citation advan- pins, as in a Pachinko machine. On pages regime) (right panel)?
tage for papers with open data (10) suggests 1058, 1055, and 1061 of this issue, evidence To better understand the situation, we
that stakeholders help themselves by pro- is presented that electrons can actually must invoke quantum physics. On the
moting transparency and reproducibility. yield to the laws of hydrodynamics (13). microscopic scale, electrons in solids are
What is additionally surprising is that these strongly interacting, but quantum many-
REFERENCES AND NOTES
observations are in agreement with math- body systems submit to the principle of
1. Paleobiological Database, https://paleobiodb.org/#/.
2. D. B. Pet, http://www.earthchem.org/petdb.
ematical techniques borrowed from string
3. D. G. Roche, L. E. B. Kruuk, R. Lanfear, S. A. Binning, PLOS theory (4). These techniques have been ap-
Biol. 13, e1002295 (2015). plied to describe strongly interacting forms
4. D. Lindenmayer, G. E. Likens, Ecol. Soc. Am. Bull. 94, 338
(2013).
of quantum matter, predicting that they hydrodynamic ows
should exhibit hydrodynamic ows (5).
5. B. A. Nosek, G. Alter, G. C. Banks, D. Borsboom, S. D.
Bowman, S. J. Breckler, S. Buck, C. D. Chambers, G. Chin, G. The experiments have been made possi- are much richer than the
Christensen, M. Contestabile, A. Dafoe, E. Eich, J. Freese, R.
Glennerster, D. Gorof, D. P. Green, B. Hesse, M. Humphreys,
ble by progress in new materials and nano- difusive currents that
fabrication techniques. Two of the papers
J. Ishiyama, D. Karlan, A. Kraut, A. Lupia, P. Mabry, T. A.
Madon, N. Malhotra, E. Mayo-Wilson, M. McNutt, E. Miguel, report on complementary aspects of the have been the traditional
E. L. Paluck, U. Simonsohn, C. Soderberg, B. A. Spellman,
J. Turitto, G. VandenBos, S. Vazire, E. J. Wagenmakers, R.
electron hydrodynamics in graphene (1, 2). mainstay of solid-state
The third paper deals with an oxide mate-
Wilson, T. Yarkoni, Promoting an open research culture.
Science 348, 14221425 (2015). rial that exhibits highly surprising trans- electronics.
6. J. A. Mills et al., Trends Ecol. Evol. 30, 581 (2015). port properties. By conning the electrical
7. S. Buck, Science 348, 1403 (2015). currents to nanoscale pipes, hydrodynamic renormalization, in which the electrons be-
8. Data Citation Synthesis Group, Joint Declaration of ow is demonstrated (3). havior is dependent on the scale at which
Data Citation Principles, M. Martone, Ed. (FORCE11,
San Diego CA; 2014); www.force11.org/group/
The ow of substances is governed by the system is observed. In conventional
joint-declaration-data-citation-principles-fnal. simple conservation laws: Matter, energy, metals, the renormalized electrons in-
9. B. Hanson, K. Lehnert, J. Cutcher-Gershenfeld, Eos 96, and electrical charge are naturally con- creasingly ignore each other as the energy
10.1029/2015EO022207 (2015). served, while in a perfectly homogeneous decreases. On the macroscopic scale, the
10. H. A. Piwowar, T. J. Vision, PeerJ. 1, e175 (2013).
space the velocity of an aggregate of matter electrons behave like the individual balls of
ACKNOWLEDGMENTS is not changing either; that is, momentum the Pachinko machine. However, it might
These recommendations resulted from a workshop supported is also conserved. A classical uid, such as well happen that the efects of the interac-
by the Laura and John Arnold Foundation. Attendees who water, looks like a dense trafc of colliding tions increase as the energy decreases (giv-
helped shape these recommendations include P. G. Brewer,
B. Clement, C. Duke, D. Erwin, L. Ferguson, R. Freckleton,
water molecules exchanging momentum ing rise to a complex quantum soup), and
A. Friedlander, T. Gardner, B. Grocholski, L. Hale, S. Hageman, at a very high rate. However, their com- until now we did not have the mathematical
J. Kratz, W. J. Kress, M. Morovati, M. Parsons, K. Ram, bined momentum does not change unless tools to describe transport in the resulting
M. Ramamurthy, H. J. Smith, A. Sugden, M. Uhen, J. VanDecar,
T. Vision, N. Wigginton, and R. Wakimoto.
highly collective quantum state. Recently, it
Instituut-Lorentz for Theoretical Physics, Leiden University, has been shown that the mathematical ma-
10.1126/science.aad7048 Leiden, Netherlands. E-mail: jan@lorentz.leidenuniv.nl chinery developed by string theorists can

1026 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
Distinguishing diferent fow regimes. (Left) In conventional metals, the fow of electrical current is due to electrons (balls) moving independently as a consequence of
quantum physics while scattering against crystal imperfections (bumpers). (Right) In normal fuids such as water, the molecules collide with each other, equilibrating in a
macroscopic fuid that is described by the theory of hydrodynamics. Electrons in particular solids that form strongly interacting quantum systems are also found to exhibit
hydrodynamic transport properties (13).

be used to describe general features of such induced with the particles and antiparticles traditional mainstay of solid-state elec-
quantum soups (4), predicting that at nite moving in the same direction; this current tronics. It will therefore be interesting
temperature they have a strong tendency is sensitive to momentum conservation (8). to see whether behavior associated with
to form hydrodynamic uids. Evidence for The data signatures of hydrodynamical hydrodynamics (shocks, turbulence, etc.)
such behavior has surfaced in seemingly un- behavior can then be discerned, which on can be incorporated into electronic device
related areas of physics (4): the quark-gluon closer inspection appear to be consistent technology. There is also the possibility of
plasma created in high-energy colliders, with the notion that these are rooted in the using hydrodynamic ow as a diagnostic to
and the unitary fermion gas of cold-atom strongly interacting quantum nature of this detect whether the states of strongly inter-
physics. These ndings raised the question electron system (9). acting quantum matter predicted by theory
of whether such behavior could also occur Bandurin et al. (2) present complemen- are actually realized in nature. Although
for electrons in metals (5). tary evidence. They exploited qualitative these applications represent tantalizing
Graphene is a strong candidate to meet diferences in the ow patterns associated possibilities, it will be a grand challenge
the required conditions. An advantage of with hydrodynamic versus Knudsen ows. for experimentalists to tame these usually
this material is the remarkable perfec- They manufactured a device to probe the difcult materials to a degree that they can
tion of its crystal structure. Its electrons current patterns that are formed upon in- be subjected to the controlled nanofabrica-
are also known to mimic the behavior of jecting a current through a narrow nozzle tion required to detect the hydrodynamic
the relativistic fermions of high-energy in a graphene sheet. These turn out to be ow.
physics (6). Considering their many-body impossible to explain on the basis of Knud-
REFERENCES
physics, the electrons in graphene exhibit sen ow, but are in accord with the whirl-
1. J. Crossno, J. K. Shi, K. Wang, X. Liu, A. Harzheim, A. Lucas,
the special renormalization property that ing patterns formed by a hydrodynamic S. Sachdev, P. Kim, T. Taniguchi, K. Watanabe, T. A. Ohki,
the interactions are marginally irrelevant, description (10). K. C. Fong, Science 351, 1058 (2016).
meaning that at zero temperature the elec- Moll et al. (3) studied PdCoO3, a material 2. D. A. Bandurin, I. Torre, R. Krishna Kumar, M. Ben Shalom,
A. Tomadin, A. Principi, G. H. Auton, E. Khestanova, K. S.
trons behave as independent particles, but as disorder-free as graphene but entirely
NovoseIov, I. V. Grigorieva, L. A. Ponomarenko, A. K. Geim,
upon raising energy or temperature the diferent in other respects. Although oxide M. Polini, Science 351, 1055 (2016).
interactions grow rapidly, tending toward metals are strongly interacting electron 3. P. J. W. Moll et al., Science 351, 1061 (2016).
the quantum soup of string theory (7). systems, this one is known to be a weakly 4. J. Zaanen, Y.-W. Sun, Y. Liu, K. Schalm, Holographic Duality
ILLUSTRATION: PETER AND RYAN ALLEN/SECOND BAY STUDIOS

in Condensed Matter Physics (Cambridge Univ. Press,


To observe the predicted hydrodynamic interacting metal. But its transport prop- 2015).
behavior, Crossno et al. (1) exploited the erties are anomalousit is exceptionally 5. S. A. Hartnoll, P. K. Kovtun, M. Mueller, S. Sachdev, Phys.
property of the graphene electron system conductive. Moll et al. owed electrons in Rev. B 76, 144502 (2007).
that it is like a Dirac vacuum with no long conduction channels, or pipes, with 6. A. H. Castro Neto et al., Rev. Mod. Phys. 81, 109 (2009).
7. M. Mueller, J. Schmalian, L. Fritz, Phys. Rev. Lett. 103,
net density of particles. Upon raising tem- variable widths. In hydrodynamic ow, the 025301 (2009).
perature, negatively and positively charged resistance is determined by the channel 8. M. S. Foster, I. L. Aleiner, Phys. Rev. B 79, 085415 (2009).
particles and antiparticles are thermally width in such a way that it can be sharply 9. A. Lucas, J. Crossno, K. C. Fong, P. Kim, S. Sachdev, http://
arxiv.org/abs/1510.01738 (2015).
excited, but they occur in precisely the same distinguished from the Knudsen ow (11).
10. L. Levitov, G. Falkovich, http://arxiv.org/abs/1508.00836
amount. In an electrical eld, the particles It is a remarkable coincidence that (2015).
and antiparticles move in opposite direc- three groups have presented independent 11. M. J. M. de Jong, L. W. Molenkamp, Phys. Rev. B 51, 13389
tions, with the efect that the whole sys- evidence for electron hydrodynamics. Such (1995).
tem does not move. However, by applying hydrodynamic ows are much richer than
a temperature diference, a heat current is the difusive currents that have been the 10.1126/science.aaf2487

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1027


Published by AAAS
INSIGHTS | P E R S P E C T I V E S

NEUROSCIENCE Group 1 Group 2


Motive-

Wiring the induction


manipulation
Induced empathy Induced reciprocity

altruistic Same increase in brain activation

brain ACC

Communication between
Motive vs.
brain regions uncovers control

hidden motives for fMRI data VS

generous behavior
AI
By Sebastian Gluth and Laura Fontanesi

C
ontrary to classical economic supposi- Empathy network Reciprocity network
Brain
tion (1), understanding peoples prefer- functional
ences and decisions is not as simple as network VS VS

Downloaded from on March 3, 2016


observing their actions. Indeed, there fMRI data,
are many reasons for behaving altruis- DCM analysis
tically, such as being moved by some- ACC AI ACC AI
ones sufering (empathy) or feeling obliged
to return a favor (reciprocity) (2, 3). One of Motive vs.
the major challenges for social psychologists control Same increase in altruistic behavior
and neuroscientists is to characterize the dif- Behavioral data
ferent motives underlying our interactions
with other people. On page 1074 in this issue, The altruistic brain. By eliciting diferent motives, the same increase in brain activation and in altruistic behavior is
Hein et al. (4) show that knowing how dis- observed, but the functional coupling between activated brain regions difers. Continuous and dashed arrows indicate
tinct areas in the human brain communicate positive and negative connections, respectively. Thicker arrows indicate stronger connections. ACC, anterior cingulate
with each other can tell us why someone be- cortex; AI, anterior insula; VS, ventral striatum.
haves altruistically.
To induce diferent motives for altruism, brains were scanned by functional mag- ticipants to their motive-induction treatment
Hein et al. developed a clever experimental netic resonance imaging (fMRI)which uses (see the gure). Hence, the reasons for being
design in which a participant rst interacts changes in local blood ow as an indicator gracious to someone appear to be hidden in
with two partners (who received specic in- of neural activity changeswhile they chose how cortical and subcortical structures com-
structions from the authors) in one of two how to allocate the money. As expected, par- municate with each other.
scenarios. In one experimental group, a par- ticipants sacriced more money to the empa- The study of Hein et al. is an intriguing
ticipant observes one partner receiving pain- thy or reciprocity partner than to the control example of the richness and power of fMRI
ful shocks, thereby eliciting an empathic partner. Critically, this increased altruism did data, when we do not simply look for ac-
response in the participant (empathy group). not difer between the groups, so that the hid- tivated blobs in brain images, but investi-
In the other group, a participant observes a den motive could not be accurately inferred gate the ow of information between brain
partner sacricing money to save the partici- from behavior alone. Hein et al. analyzed the areas. In this spirit, the authors report that
pant from receiving painful shocks, thereby neural data using a conventional contrast of the functional networks of the empathy and
eliciting a desire in the participant to return brain activation during altruistic decisions the control condition are very similar to each
the kind behavior (reciprocity group). In toward the empathy/reciprocity partner ver- other but dissimilar to the reciprocity net-
both groups, the participant is also paired sus the control partner. In line with previous work. This indicates that empathy-motivated
with a second partner who serves as a con- work (5, 6), a brain network including ante- altruism is perhaps (phylogenetically) more
trolthat is, a second person who does not rior cingulate cortex (ACC), anterior insula common than reciprocal altruism.
receive painful shocks in the empathy group, (AI), and ventral striatum (VS) was identied. Hein et al. provide yet another striking
or who does not sacrice money to relieve But again, this analysis did not reveal any dif- nding. After categorizing participants as
the participants shock treatment in the reci- ferences between the two motives. However, either selsh or prosocial on the basis of
ILLUSTRATION: K. SUTLIFF/SCIENCE

procity group. when the authors examined how these brain their behavior toward the control partner,
Following this phase of motive induction, areas interact with each other during altruis- they nd that only selsh participants in-
all participants performed a money alloca- tic decisions using dynamic causal modeling crease their donations toward the empathy
tion task. They chose between maximizing a (DCM) (7), they observed a distinction. The partner, whereas only prosocial participants
partners monetary payof by reducing their connectivity patterns difered remarkably be-
own (prosocial behavior) or holding on to the tween empathy and reciprocity, and, by using
Department of Psychology, University of Basel,
money at a cost to the other person (selsh a novel DCM-based classication technique Missionsstrasse 62a, 4055 Basel, Switzerland. E-mail:
behavior). During this task, the participants (8), the authors successfully categorized par- sebastian.gluth@unibas.ch

1028 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
act (even) more generously toward the reci- GENETICS
procity partner. Provided that this result can
be established as robust and generalizable
to nonlaboratory contexts, it bears high rel-
evance for social policy and for treatments
A copy-and-paste gene
of antisocial behavior: Whether we want to
increase altruism in an already cooperative
group or in individuals with decits in pro-
regulatory network
social skills, we should focus on emphasizing Transposable elements provide a ready-made route
reciprocity or empathy, respectively. to regulate complex gene networks
The study raises many new questions, in-
cluding whether these connectivity patterns
are specic to pain-related altruism or gen- By Vincent J. Lynch whether cis-regulatory elements originate
eral to any altruistic behavior. Furthermore, more rapidly through concerted processes.

C
it remains open whether there is a causal hanges in gene regulatory networks On page 1083 of this issue, Chuong et al.
relationship between the proper function- (GRNs) underlie many phenotypic dif- provide evidence that concerted processes,
ing of these networks and altruismin other ferences between species. However, involving endogenous retroviruses (ERVs),
words, whether we can alter altruistic behav- the mechanisms of GRN evolution which are remarkably abundant in mamma-
ior by disrupting specic connections within are still being debated (15). Explain- lian genomes, have contributed to the evo-
the networks. In addition, how (diferently ing how GRNs originate, diversify, and lution of the regulatory systems that control
motivated) altruistic decisions develop over maintain their identities despite regulatory the mammalian immune system (6).
time appears to be a particularly pressing element turnover is essential for develop- Transposable element (TE)mediated
question, given that Hein et al. did not model ing mechanistic explanations for the origin, origination of cis-regulatory elements is an

Downloaded from on March 3, 2016


the decision process itself. Sequential sam- diversication, and conservation of homolo- attractive alternative to the de novo appear-
pling models, which describe decisions as a gous characters between species. Among the ance of such elements. TEs provide a mech-
noisy process of accumulating evidence up major outstanding questions in GRN evolu- anism to rapidly distribute nearly identical
to a response threshold, can explain exactly tion is whether individual cis-regulatory el- copies of regulatory elements across the
how and when decisions are made (9). These ements arise de novo through the gradual genome that can respond to the same in-
models have a long and successful history in accumulation of mutations that increase put signals (79) rather than requiring that
accounting for perceptual and preferential the regulatory potential of existing DNA or each gene in a GRN evolve cis-regulatory
decision-making, such as when we react to a Transcription
red or green trafc light or when we choose TG-1
factors
between ice cream avors, respectively. But TG-2
their application to social interactions is
TG-3
just emerging (10). Because sequential sam-
pling models take not only choices but also
De novo model Alternate model
response times into account and break the
decision process down into its elements (e.g.,
response conict, decision urgency), they New transcription
would ofer an in-depth comparison of em- factor binding sites
pathy- and reciprocity-based altruistic behav-
ior. Moreover, given that sequential sampling TG-1
model predictions can be used to inform
New CRE
DCM-based analyses of fMRI data (11), a TG-2
closer inspection of the process that leads to
selsh or prosocial actions is a promising av- TG-3
enue for future research.
REFERENCES AND NOTES New CRE
1. P. Samuelson, Economica 5, 61 (1938). New PPI
2. E. Fehr, U. Fischbacher, Nature 425, 785 (2003).
3. C. D. Batson, The Altruism Question: Toward a Social
Psychological Answer (Erlbaum, Hillsdale, NJ, 1991).
4. G. Hein et al., Science 351, 1074 (2016). TG-1
5. K. L. Phan, C. S. Sripada, M. Angstadt, K. McCabe, Proc.
Natl. Acad. Sci. U.S.A. 107, 13099 (2010).
6. T. Singer et al., Nature 439, 466 (2006).
TG-2
7. K. J. Friston et al., NeuroImage 19, 1273 (2003).
8. K. H. Brodersen et al., PLOS Comput. Biol. 7, e1002079
ILLUSTRATION: K. SUTLIFF/SCIENCE

(2011).
9. P. L. Smith, R. Ratclif, Trends Neurosci. 27, 161 (2004). TG-3
10. C. A. Hutcherson, et al., Neuron 87, 451 (2015).
Plug-and-play gene regulatory network formation. The de novo model of GRN formation requires that each gene in
11. S. Gluth et al., Neuron 86, 1078 (2015).
the network (TG-1 to TG-3) independently evolve a new cis-regulatory element (CRE) that can bind to the same sets of
ACKNOWLEDGMENTS transcription factors (red and blue circles). However, individual transcription factor binding sites may not be sufcient to
This work is supported by a Swiss National Science Foundation activate target gene expression, requiring the evolution of additional binding sites for cooperating transcription factors or
grant (100014_153616) to both authors.
novel protein-protein interactions (PPIs) to recruit more transcription factors and cofactors. Alternatively, TEs can generate
10.1126/science.aaf4688 new regulatory networks by donating already functional cis-regulatory elements to neighboring genes.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1029


Published by AAAS
INSIGHTS | P E R S P E C T I V E S

elements capable of responding to the ing have diverged within mammals, poten- C H E M I ST RY
same stereotyped signals de novo (see the tially reecting lineage-specic adaptations
gure). The de novo origination model, for
example, would require the independent
evolution of the same transcription factor
to host-pathogen interactions. Chuong et
al. found that 27 TE families were enriched
within IFNG-responsive cis-regulatory ele-
Surprised by
binding site(s) in many genes distributed
across the genome, whereas TEs can donate
the same sets of transcription factor bind-
ments. Remarkably, 20 of these TE families
originated from long terminal repeat (LTR)
promoter regions of ERVs. These data sug-
selectivity
ing sites to nearby genes upon their integra- gest that many TE-derived IFNG-responsive A bifunctional catalyst
tion into the genome. cis-regulatory elements arose from ancient enables olen synthesis
Although TE-mediated rewiring of GRNs retroviral infections.
provides a simple and straightforward model Previous studies have implicated ERVs from carbon monoxide and
of gene regulatory evolution, and numerous in the genesis of cis-regulatory elements hydrogen at high selectivity
cis-regulatory elements for individual genes important for the evolution of mamma-
in mammalian genomes are derived from lian pregnancy (12), placentation (13), and
TEs (10), few studies have demonstrated that most notably the core regulatory networks By Krijn P. de Jong
TEs actually have globally rewired GRNs in embryonic stem cells (14). ERVs are nor-

L
(11). Among the most notable barriers to de- mally repressed in somatic cells by histone ower olens, particularly ethylene
veloping mechanistic explanations for the modications and methylation, suggesting (C2H4), propylene (C3H6), and butylenes
role of TEs in the origin and diversication that ERVs escape silencing in some con- (C4H8), are important intermediates
of GRNs is a lack of model networks that are texts, such as the placenta and embryonic in the manufacture of products such
both amenable to detailed functional studies stem cells, which appear to have pervasive as plastics, solvents, paints, and medi-
and that evolve rapidly enough that TEs can expression of ERVs. Chuong et al., for ex- cines. They are produced worldwide

Downloaded from on March 3, 2016


be caught in the act of remodeling the regu- ample, speculate that IFNG-responsive cis- in amounts exceeding 200 million tons per
regulatory elements in LTRs are remnants of year (see the photo) (1), mostly from crude
ancient retroviral enhancers that used host oil. More recent approaches use methanol or
signaling to promote viral transcription and synthesis gas (syngas; a mixture of carbon
TEs provide a mechanism replication; thus, ERVs may have evolved to monoxide and hydrogen) as feedstocks, but
to rapidly distribute nearly be derepressed upon IFNG stimulation. This capital investments are high and/or selectivi-
conjecture suggests an answer to a nagging ties to lower olens limited. A bifunctional
identical copies of regulatory question: What is it about TEs that predis- catalyst reported by Jiao et al. on page 1065
elements poses them to act as tissue-specic regula- of this issue (2) enables the direct conversion
tory elements? In this case, ERVs may be of synthesis gas to lower olens with a sur-
predisposed to act as regulatory elements in prisingly high selectivity.
latory landscape. The mammalian immune IFNG-responsive cells because they already Synthesis gas is an important intermediate
system is an ideal model in which to explore possess functional IRF and STAT binding in the chemical industry for the production
the molecular mechanisms that underlie sites. More generally, TEs may be biased in of ammonia, hydrogen, and methanol. Any
GRN evolution because numerous genomic the kinds of transcription factor binding carbon-containing feedstock such as coal,
and experimental resources have been de- sites they contain because of their own biol- natural gas, or biomass can be converted
veloped for mammals. In addition, immune ogy, leading to domestication as regulatory via steam reforming or gasication into a
responses evolve fast enough that evolution- elements in cell types that already express mixture of carbon monoxide and hydrogen.
ary changes can occur between relatively those transcription factors. Although more In previous studies, lower olens have been
closely related species, yet slow enough that detailed experimental approaches will be produced in one step from synthesis gas with
the functional signicance of TE-derived required to reconstruct the exact mecha- selectivities of 50 to 60% (1, 3), a process re-
regulatory elements can be inferred and ex- nisms by which TEs rewire gene regulatory ferred to as FTO (Fischer-Tropsch to olens).
perimentally validated. networks, the experimental framework pro- Methanol can also be converted to olens
Proinammatory cytokines, such as vided by Chuong et al. will lead the way. with the help of zeolite catalysts (4). Using
interferon- (IFNG), are essential signaling this MTO (methanol to olens) process, high
REFERENCES
molecules in the innate immune response yields of ethylene and propylene have been
1. S. B. Carroll, Cell 134, 25 (2008).
that are released upon infection and that 2. S. B. Carroll, PLOS Biol. 3, e245 (2005). obtained, but catalyst deactivation poses
regulate a battery of downstream immu- 3. B. Prudhomme, N. Gompel, S. B. Carroll, Proc. Natl. Acad. challenges. Recently, the MTO technology
nity factors called IFN-stimulated genes Sci. U.S.A. 104, 8605 (2007). was implemented at industrial scale in China,
4. G. P. Wagner, V. J. Lynch, Curr. Biol. 20, R48 (2010).
(ISGs). ISGs are regulated by cis-regulatory 5. V. J. Lynch, G. P. Wagner, Evolution 62, 2131 (2008). using coal as feedstock for the production of
elements containing binding sites for in- 6. E. B. Chuong, N. C. Elde, C. Feschotte, Science 351, 1083 methanol. Since the shale-gas revolution in
terferon regulatory factor (IRF) and signal (2016). the United States, methanol production has
7. R. J. Britten, E. H. Davidson, Science 165, 349 (1969).
transducer and activator of transcription 8. E. Davidson, R. Britten, Science 204, 1052 (1979).
become economically much more attractive.
(STAT) transcription factors that are acti- 9. B. McClintock, Science 226, 792 (1984). In the coming decade, transport of an addi-
vated by IFN signaling. Although the innate 10. R. Rebollo, M. T. Romanish, D. L. Mager, Annu. Rev. Genet. tional 30 million tons per annum of metha-
46, 21 (2012).
immune response is conserved as a process, 11. F. S. J. de Souza, L. F. Franchini, M. Rubinstein, Mol. Biol.
nol from the United States to China for use in
the specic genes activated by IFNG signal- Evol. 30, 1239 (2013). MTO technology in Asia is foreseen (5). How-
12. V. J. Lynch et al., Cell Rep. 10, 551 (2015).
13. E. B. Chuong et al., Nat. Genet. 45, 325 (2013).
Department of Human Genetics, The University of Chicago, 14. G. Kunarso et al., Nat. Genet. 42, 631 (2010). Inorganic Chemistry and Catalysis, Debye Institute for
920 East 58th Street, CLSC 319C, Chicago, IL 60637, USA. Nanomaterials Science, Utrecht University, Utrecht,
E-mail: vjlynch@uchicago.edu 10.1126/science.aaf2977 Netherlands. E-mail: k.p.dejong@uu.nl

1030 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
Large-scale olefn production. At
this chemical plant in Ludwigshafen,
Germany, steam cracking is used to
make olefns from crude oil.

ever, methanol needs to be produced from hydrocarbons. On metal or metal carbide lyst and an MTO catalyst. However, as noted
methane, making olen production more surfaces, CHx species polymerize, resulting by the authors, previous attempts to com-
complex and expensive. Direct conversion in a statistical distribution of hydrocarbon bine methanol synthesis and MTO catalysts
of methane to chemicals (including lower chain lengths called the Anderson-Schulz- mainly yielded parafn and aromatics. Fur-
olens) seems attractive, but product yields Flory (ASF) distribution. This statistical dis- thermore, they show that when methanol is
have been limited (6) or process conditions tribution of products leads to a maximum fed to the MSAPO catalyst as an intermedi-
harsh (7). lower-olen yield of around 60% (3). By using ate, catalyst stability does not reach the >100
Jiao et al. now report selectivities for lower a metal oxide catalyst to activate CO, the au- hours found for the OX-ZEO catalyst.
olens of about 80% of total hydrocarbons, thors prevent polymerization of CHx species, Most important, by disconnecting CO
produced directly from synthesis gas at 17% allowing CO insertion followed by ketene for- activation and C-C bond formation in hy-
CO conversion under industrially attrac- mation. Mass spectrometry data show that drocarbons, the authors circumvent ASF
tive conditions400C, 25 bar, and H2/CO ketene is indeed present in the gas phase. The distribution in the FTO process. This leads
volume ratio of 1.5and catalyst lifetimes authors use ketene from the decomposition to the surprisingly high 80% selectivity to
exceeding 100 hours. Their bifunctional cata- of acetic acid anhydride to provide evidence lower olens.
lyst, OX-ZEO, consists of a metal oxide (Zn- for olen formation over the zeolite from this The research reported by Jiao et al. should
CrOx) and a zeolite (MSAPO). CO and H2 are intermediate. Well-known MTO catalysts (8) be of interest to both academia and industry.
activated on the metal oxide surface, leading did not perform well as the second compo- It remains to be shown whether ketene or the
to formation of CH2 species and subsequently nent of the bifunctional catalyst; the authors thermodynamically more stable methanol is
of ketene (CH2CO). The latter difuses via the therefore tailored the acidity of the MSAPO the key intermediate in the new process. The
gas phase to the zeolite and is converted to zeolite, using temperature-programmed de- bifunctional nature of the catalyst invites
lower olens (see the gure). sorption of ammonia as an indicator for acid structure performance studies including the
The authors provide extensive experimen- site strength. efects of proximity of oxide and acid func-
tal and theoretical data in support of this The main components of the OX-ZEO tions (9). The new process could become a
mechanism for converting synthesis gas to catalyst resemble a methanol synthesis cata- serious competitor for industrial processes
such as FTO (3) and MTO (4, 8).
CREDITS: (PHOTO) BASF/FLICKR; (ILLUSTRATION) P. HUEY/SCIENCE

C2H4
C3H6 REF ERENCES AND NOTES

2 CO C4H8 1. H. M. Torres Galvis, K. P. de Jong, ACS Catal. 3, 2130 (2013).


Difusion 2. F. Jiao et al., Science 351, 1065 (2016).
+ +
CO2 CO CH2CO n CH2CO n CO 3. H. M. Torres Galvis et al., Science 335, 835 (2012).
H2
4. P. Tian, Y. Wei, M. Ye, Z. Liu, ACS Catal. 5, 1922 (2015).
5. A. H. Tullo, Chem. Eng. News 92, 12 (11 August 2014).
6. R. Horn, R. Schlgl, Catal. Lett. 145, 23 (2015).
7. X. Guo et al., Science 344, 616 (2014).
CH2 CH2 H+ H+ H+ 8. U. Olsbye et al., Angew. Chem. Int. Ed. 51, 5810 (2012).
9. J. Zecevic et al., Nature 528, 245 (2015).

ACKNOWL EDGMENTS
Supported by the European Research Council (EU FP7 ERC
Advanced Grant no. 338846).
Metal oxide (ZnCrOx) Zeolite (MSAPO)
Two steps in one. Schematic representation of the OX-ZEO catalyst and the reaction scheme proposed by Jiao et al. 10.1162/science.aaf3250

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1031


Published by AAAS
INSIGHTS | P E R S P E C T I V E S

I M M U N O L O GY cartilage-hair hypoplasia (CHH), an early-


onset genetic disease characterized by skel-

An RNA twist to TH 17 cells etal malformations and immune dysfunction


(9). Using clustered regularly interspaced
short palindromic repeats (CRISPR)Cas9-
A long noncoding RNA and RNA helicase based genome editing, mice were generated
that expressed a mutant form of Rmrp bear-
constitute a new layer of T cell control ing a mutation corresponding to that found
in CHH patients. Cells from these mice had
By Maninjay K. Atianand and and autoimmune encephalomyelitis. The impaired Rmrp functions in the context of
Katherine A. Fitzgerald study furthers our understanding of RORt- DDX5-RORtdependent regulation of TH17
dependent gene regulation and the coactiva- cell efector functions.

T
helper lymphocytes play essential tor functions of DDX5, which also cooperates The discovery of Rmrp as a functional
roles in the adaptive immune system. with Runt-related transcription factor 2 partner of DDX5 and a regulator of RORt
They come in distinct types dened (Runx2), a key transcription factor important target genes reveals a new layer of regulation
by unique transcriptional programs in osteoblast diferentiation (7). of TH17 cell function. However, important
that control their development and Determining how DDX5 cooperates with questions arise regarding the signals that
functions. Among these, T helper RORt began with the observation that the trigger formation of the DDX5-Rmrp com-
17 (TH17) cells are important in protect- RNA helicase activity of DDX5 was essential plex. DDX5 and Rmrp are both abundantly
ing mucosal surfaces against fungal and for inducing RORt-regulated gene expres- expressed in TH17 cells, as well as in many
bacterial infections. In addition, TH17 cells sion. Wild-type DDX5, but not a helicase- other cell types. The precise molecular sig-
contribute to the pathogenesis of multiple dead DDX5 mutant, rescued the expression nals that trigger the assembly of the complex
autoimmune diseases (1). A recent study (2) of IL-17a and IL-17f in DDX5-decient T cells in TH17 cells therefore remain to be dened.

Downloaded from on March 3, 2016


adds yet another layer of complexity to the under TH17-polarizing conditions. RNA se- It is likely that specic environmental cues
biology of these complex cellsan RNA he- quencing was used to screen for endogenous present at mucosal surfaces unleash this
licase and a long noncoding RNA (lncRNA) RNAs associated with DDX5 in TH17 cells. complex in a specic manner under TH17-
that act together to control polarizing conditions. Com-
their efector functions. mensal bacteria or their
Diferentiation of nave T TH17 products could provide such
helper cells into distinct popu- a signal. It will also be impor-
lations depends on a cytokine tant to determine how DDX5
milieu that governs the ex- Rmrp and Rmrp interface with other
pression of lineage-dening Stat3, Irf4, factors such as signal trans-
transcription factors (3). The DDX5 Batf, Blimp-1 ducer and activator of tran-
induced expression of the Signal scription 3 (Stat3), interferon
?
nuclear hormone receptor RORt regulatory factor 4 (Irf4), basic
retinoic acid receptorrelated IL-17a/f, IL-22 leucine zipper transcription
orphan nuclear receptor factor, ATF-like (Batf), and B
gamma t (RORt) drives TH17 lymphocyteinduced matura-
cell diferentiation and func- Host defense tion protein1 (Blimp-1), all of
Autoimmune disease
tions by controlling the expres- Colitis
which control RORt target
sion of interleukin-17a (IL-17a), gene expression in TH17 cells
IL-17f, and IL-22 cytokines, A TH17 cell efector program. In response to infection, the DDX5-Rmrp complex is (10, 11). It will also be inter-
among others (4, 5). Using recruited to RORt-occupied DNA to drive the expression of genes that promote host esting to assess the potential
mass spectrometry, the recent defense and autoimmunity. The signals that trigger this pathway are not yet clear. contribution of DDX5-Rmrp
study identied a coactivator complexes in controlling IL-17
for RORt called DEAD-box protein 5 (DDX5; This revealed RNA component of mitochon- gene transcription in other immune cells,
also known as p68). DDX5 belongs to a large drial ribonuclease (RNAse) protein complex such as innate lymphoid cells. The identica-
family of more than 70 evolutionarily con- (Rmrp) as a major constituent of the DDX5 tion of DDX5 and Rmrp raises the intriguing
served RNA helicases that hydrolyze adenos- complex. Through elegant biochemical, possibility that additional RNA helicaseln-
ine 5-triphosphate to unwind RNA (6). These genome-wide, and in vivo studies, Rmrp cRNA complexes await discovery that might
helicases are involved in diverse aspects of was shown as essential for the assembly and function similarly to control gene expression
RNA biogenesis, including gene transcrip- recruitment of DDX5-RORt complexes to in other T helper cell subsets.
tion, splicing, and messenger RNA decay. regulatory regions of RORt target genes in
REFERENCES
In TH17 cells, DDX5 mediates the transcrip- TH17 cells (see the gure).
1. D. D. Patel, V. K. Kuchroo, Immunity 43, 1040 (2015).
tion of nearly 40% of RORt target genes Rmrp is a lncRNA (a nonprotein-coding 2. W. Huang et al., Nature 528, 517 (2015).
by binding to regulatory regions of these RNA larger than 200 nucleotides) that 3. S. Nakayamada et al., Curr. Opin. Immunol. 24, 297 (2012).
ILLUSTRATION: P. HUEY/SCIENCE

4. I. I. Ivanov et al., Cell 126, 1121 (2006).


genes. DDX5 was also found to contribute to functions in mitochondrial RNA process- 5. T. Korn et al., Annu. Rev. Immunol. 27, 485 (2009).
RORt-mediated inammatory pathologies ing (MRP; hence, it is also known as RNase 6. P. Linder, E. Jankowsky, Mol. Cell. Biol. 12, 505 (2011).
in a mouse model of T cellinduced colitis MRP) and maturation of 5.8S ribosomal RNA 7. E. D. Jensen et al., J. Cell. Biochem. 103, 1438 (2008).
8. C. L. Hsieh et al., Genomics 6, 540 (1990).
(8). Rmrp also functions as a precursor to mi- 9. A. N. Martin, Y. Li, Cell Res. 17, 219 (2007).
Program in Innate Immunity, Division of Infectious Diseases croRNAs. Interestingly, genetic mutations in 10. M. Ciofani et al., Cell 151, 289 (2012).
and Immunology, Department of Medicine, University of Mass- 11. R. Jain et al., Immunity 44, 131 (2016).
achusetts Medical School, 364 Plantation Street, Worcester, MA this lncRNA had previously been linked to a
01605, USA. E-mail: kate.ftzgerald@umassmed.edu rare autosomal recessive disorder known as 10.1126/science.aaf4691

1032 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
INSIGHTS

B O OKS et al .

BIG DATA

The future
of memory
A deluge of digital data
threatens to destroy our
collective identity
By Cliford B. Anderson

A
surprising fact about professional ar-
chivists and librarians is their unsenti-
mental readiness to turn away books,
documents, and other tidbits of infor-

Downloaded from on March 3, 2016


mation, an act known as appraisal.
Archivists are skilled at making these
evaluations because they are charged with
curating coherent collections rather than as-
sembling hodgepodges of stuf. Opened in 1878, the Peabody Library Building in Baltimore houses more than 300,000 titles.
The art of appraisal makes sense when
the material in question must be stored in By contrast, digital artifacts like websites after all, among other things, a moral act.
a physical space. Archivists would be over- change constantly. And if you unplug a data We are what we know. However, as Rumsey
whelmed if they tried to keep every item de- center, you stand to lose everything. notes, forgetting also serves a critical pur-
livered to their doorsteps. But what about Rumsey tells her story through a series of pose, freeing up our minds to incorporate
in the digital world? If technological ad- vignettes, focusing on signicant episodes new experiences.
vances allow us to keep every bit of digital in cultural history. She sketches the devel- Is there a danger that computers may
data, should we? How would the resulting opment of writing in the ancient Near East, supplant us as our cognitive systems strug-
proliferation of information afect our abil- the organization of libraries in the Hellenis- gle to keep up with all the data we now
ity to make sense of it all? And if we come tic world, the invention of movable type and produce? Rumsey is skeptical; she regards
to depend on this digital record, how can the rise of the rst print natives, and the brains and computers as diferent and
we ensure that it persists into the future? In Jefersonian dream of a universal library. largely complementary. Drawing on con-
When We Are No More, Abby Smith Rumsey, These inection points paved the way for temporary neuroscience, she contends that,
a professional historian and consultant to our current understanding of the material unlike computers, the brain relies on afec-
libraries and other memory organizations, world as one that encodes its own history. tive, emotional experiences of the world to
explores these questions through the wide Nature is the ultimate archive, the most organize facts into patterns and to produce
lens of human history. complete set of records about the past, the narratives about future courses of action.
Rumsey adopts an evolutionary perspec- Universal Library itself, she writes, and sci- Rumsey also issues a warning against the
tive on the function of memory in society, ence is the ultimate library card. temptation to overoptimize our preserva-
arguing that our ability to transmit infor- The digital age has vastly increased our tion strategies. Although we may consider
mation by encoding it in physical media ability to collect and recollect data, al- certain forms of information unimportant
is what enables us not only to develop a though perhaps at the cost of our personal or useless at present, that same information
sense of the past but also to meet future memories. Our reliance on smart devices, may turn out to be vital in the future. She
challenges. However, the externalization of which summon facts on demand, decreases points, for example, to information about
memory has its downside. We become reli- our need to recall facts unaided. In part, climate change that scholars have retrieved
PHOTO: MATTHEW PETROFF/WIKIMEDIA COMMONS

ant on the material artifacts of culture. This this is an old worry. Socrates fretted about from the antiquarian logbooks of naval
dependence becomes fraught as we move how the technology of writing afected wis- voyages.
to more ephemeral media. Clay tablets, dom. Committing something to memory is, The goal of When We Are No More is not
although not capable of recording much to have the nal word on the subject of in-
information, have survived over millennia. formation management but to challenge
Burn down a Mesopotamian archive and us to consider more seriously how the con-
the red tablets become more durable. When We Are No More sequences of our current data deluge will
How Digital Memory Is inuence society moving forward. In that,
Shaping Our Future Rumsey succeeds admirably.
The reviewer is Director of Scholarly Communications,
Vanderbilt University, Nashville, TN 37240, USA. Abby Smith Rumsey
E-mail: cliford.anderson@vanderbilt.edu Bloomsbury, 2016. 246 pp. 10.1126/science.aaf1188

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1033


Published by AAAS
INSIGHTS | B O O K S

HISTORY OF SCIENCE
Strange Glow

Risky business The Story of Radiation


Timothy J. Jorgensen
Princeton University Press,
Hard-won lessons shed light on the promise 2016. 506 pp.

and perils of radiation and nuclear power


By Angela N. Creager medical practitioner at the turn of the cen- exposure to most people. Especially worri-
tury pioneered the use of x-rays to treat some to Jorgensen is the recent popularity

I
n 1953, President Dwight D. Eisen- cancer. of whole-body CT scans, which can expose
hower launched his Atoms for Peace In chapter 7, Jorgensen recounts in detail individuals to 20,000 times the radiation
initiative, aimed at increasing energy the dreadful injuries caused by the atomic dose of an x-ray to the armroughly the
production, improving agriculture, and weapons detonated over Hiroshima and same as experienced by the lower-dose Japa-
advancing health care. The following Nagasaki. He also notes the deadly conse- nese atomic bomb survivors. In addition,
year, the U.S. Pacic atomic test Castle quences of the shock waves and restorms many physicians are unaware of how many
Bravo showered a Japanese shing ves- from those blasts, which together repre- false positives are generated by medical tests
sel with radioactive fallout, resulting in 23 sented 85% of the energy released. Those based on x-rays (such as mammograms) (1).
cases of radiation sickness (including one victims didnt live to tell their tales. But if Such results can precipitate further tests and
death) and a boatload of contaminated they had, we might have a fuller understand- surgeries, which are at best unnecessary and
tuna. These eventseven more than the use ing of the horrors of atomic warfare. at worst bring their own complications.
of nuclear weapons at the end of World War Most Americans are more afraid of nu- Ionizing radiation is now the single best-

Downloaded from on March 3, 2016


IIlaunched an ongoing public debate over clear power plants than stored nuclear war- studied carcinogen on Earth, in no small
whether the risks of radiation exposure heads. In chapter 17, Jorgensen illustrates measure due to the long-term studies of sur-
outweigh the benets of nuclear vivors from Hiroshima and Naga-
technology. For many, the verdict saki. Even so, uncertainty bedevils
is still out. Or, more to the point, many of the best risk assessments
it is hard to reach a verdict even about dose estimates, particularly
today, because unbiased, compre- when it comes to exposures from
hensible information on radia- environmental contamination.
tion risk is hard to come by. For the most part, I nd Jor-
Timothy J. Jorgensens book gensens recounting of radiations
Strange Glow: The Story of Radi- uses and their hazards clarifying,
ation lls this gap, using history if not reassuring. But I wonder
to explain how we encounter ra- about his assumption that indi-
diation and how it afects us. His viduals can make choices about
goal is to provide readers with their exposures. As Ulrich Beck
the basic facts so that they can observed in his landmark book
make decisions about the risks Risk Society, social status is now
they want to live with. reected in the distribution of risk
Jorgensen, a radiation biologist rather than (simply) the distribu-
at Georgetown University, notes Data collected from atomic bomb survivors help scientists estimate the cancer tion of wealth (2). Not everyone in
that quantitative approaches to risk associated with radiation-based medical treatments and diagnostics. the world gets to choose how they
educating the public about radia- encounter radioactivity, which ar-
tion have failed. He believes that people learn why these fears may be misplaced. Power gues for the importance of just policies as
through stories more easily than through reactors are not without hazards, but they well as informed consumers.
statistics. The book begins by comparing ra- cannot explode in the same way as atomic Although I doubt that the risks of radia-
dioactivity to electricity, which at the time of weapons are designed to do. Jorgensen re- tion will ever seem as manageable as those
Edison was not well understood by the public counts chillingly many near accidents dur- of electricity, I am ultimately convinced by
and was feared for its deadliness. Over time, ing the Cold War, including one example in Jorgensens approach. His book gave me a
PHOTO: MARTIN SCHUTT/PICTURE-ALLIANCE/DPA/AP IMAGES

however, people began to see how the dan- which a B-52 bomber broke open in midair more commonsense understanding of the
gers might be managed and were ultimately over a farm in North Carolina, releasing risks associated with radiation exposure
persuaded that the potential benets out- two hydrogen bombs. Although both bombs than either my scientic training or my his-
weighed the risks. He believes this illustrates remained intact, the ensuing investigation torical research on this topic (3).
that efective education as well as experience revealed that all but one of the seven steps
REFERENCES
can change risk perception. required to detonate had occurred. So long
1. G. Gigerenzer, Risk Savvy: How to Make Good Decisions
Jorgensen is a good storyteller, with an as nuclear warheads remain in the custody (Viking, New York, 2014).
eye for compelling details. In chapter 6, for of fallible people, their accidental detona- 2. U. Beck, Risk Society: Towards a New Modernity, Mark
example, he explains how a homeopathic tion is by far the greatest large-scale radia- Ritter, Transl. (Sage, Los Angeles, 1986).
3. A. N. H. Creager, Life Atomic: A History of Radioisotopes
tion-related threat to humankind. in Science and Medicine (University of Chicago Press,
The reviewer is at the Department of History,
Aside from this catastrophic possibil- Chicago, 2013).
Princeton University, Princeton, NJ 08544, USA. ity, routine medical usage of x-rays delivers
E-mail: creager@princeton.edu the greatest amount of man-made radiation 10.1126/science.aaf3339

1034 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
INSIGHTS

Bonds Perspective and Veldman et al. (10) 2. S. Minnemeyer, L. Laestadius, N. Sizer, C. Saint-Laurent,
LET TERS define grassland biomes broadly, whereas P. Potapov, A world of opportunity (World Resources
Institute, Washington, DC, 2011).
the Atlas defines forest landscapes as 3. L. Laestadius et al., Unasylva 238 62, 47 (2011/2).
Edited by Jennifer Sills those with a potential tree canopy cover 4. L. Laestadius et al., Science 347, 6227 (2015).
as low as 10%, consistent with the widely 5. ITTO and IUCN, Restoring forest landscapes, ITTO
Technical Series No. 23 (2005).
accepted FAO definition (11) but at a much 6. IUCN and World Resources Institute, A guide to the
lower level of resolution (1 by 1 km). As a
Seeing the grasslands result, some grassland biomes identified by
Restoration Opportunities Assessment Methodology
(ROAM) (IUCN and WRI, Gland, Switzerland, 2014);
http://cmsdata.iucn.org/downloads/roam_handbook_
through the trees Bond and by Veldman et al. overlap with
forest landscapes identified by our Atlas.
lowres_web.pdf.
7. R. Winterbottom, Restoration: Its about more than
IN HIS PERSPECTIVE Ancient grasslands These mosaic areas require individual just the trees (2014); www.wri.org/blog/2014/05/
restoration-it%E2%80%99s-about-more-just-trees.
at risk (8 January, p. 120), W. J. Bond assessment, including identifying the natu- 8. IUCN, Restoring the worlds ecosystems: More than
makes a compelling argument for the need ral level of tree cover before deciding on planting trees (2015); www.iucn.org/about/work/
to identify, study, and preserve ancient the appropriate management approach. programmes/forest/?22210/restoring-the-worlds-
ecosystems-more-than-planting-trees.
grassland ecosystems because of their Further research and collaboration are 9. W. A. Hoffman et al., Ecol. Lett. 15, 759 (2012).
ecological importance. We strongly agree. needed to explore areas where forests 10. J. W. Veldman et al., BioScience 65, 1011 (2015).
However, we disagree with his claim that and grasslands overlap. Agreement 11. FAO, FRA 2015 Terms and Definitions, Forest Resources
Assessment Working Paper 180 (2012).
our Atlas of Forest Landscape Restoration on terminology, refinement of maps,
Opportunities (13) promotes activities and identification of the appropriate
that threaten these grasslands by target- management policies and practices for
Response
ing them for afforestation or reforestation these complex ecosystems are among the DEWITT ET AL. claim that the Forest
beyond their natural level of tree cover. outcomes we can achieve through col- Landscape Restoration (FLR) plan (13)
Bond claims that the Atlas is linked laboration. We invite Bond to work with would not return tree cover beyond

Downloaded from on March 3, 2016


to a global plan to refor- what would be ecologically
est degraded lands to offset appropriate. However, the
anthropogenic CO2emissions. labeling system of the Atlas of
This indicates a fundamental Forest Landscape Restoration
misunderstanding of the con- Opportunities stymies this goal.
cept of Forest and Landscape The Atlas defines a landscape as
Restoration (FLR) and of the degraded if it is subject to any
Atlas. As we have previously process that reduces the volume
indicated (4), FLR is not a for- and canopy cover of trees.
est plantation project and is Consequently, misclassifications
not focused solely on offset- of ancient ecosystems, main-
ting CO2 emissions (though tained by fire and/or herbivory,
sequestering carbon is one of are inevitable. Under these
many benefits of FLR). FLR assumptions, the FLR handbook
does not call for increasing tree would mark such areas as suit-
cover beyond what would be able for reforestation.
ecologically appropriate for a It has long been assumed that
particular location, nor should tropical grassy ecosystems were
it cause any loss or conversion originally forests degraded by
of natural forests, grasslands, or Wildebeests graze in Serengeti National Park, Tanzania. cutting, and especially burning
other ecosystems. (4, 5). However, as shown by
Rather, FLR is a process that seeks us on these objectives, with the end goal the studies I reviewed in my Perspective,
to regain the ecological integrity and of leveraging the principles of FLR to both evidence of fire does not mean that the
enhance human well-being in deforested preserve critical ecosystems and improve landscape originated as forest. In striking
or degraded forest landscapes (5). The FLR local livelihoods. contrast to closed forests, ancient grassy
approach includes a wide range of sustain- Sean DeWitt,1* James Anderson,1 ecosystems are maintained, not degraded,
able land-management practices applicable Chetan Kumar,2 Lars Laestadius,4 by fire and herbivory. Yet they are misclas-
to both forest and nonforest ecosystems Stewart Maginnis,2 S. Minnemeyer,1 sified by the Atlas as degraded forest. For
(57). A fundamental first step in the FLR Peter V. Potapov,3 Katie Reytar,1 example, the western Serengeti and Kruger
process involves understanding the eco- Carole Saint-Laurent2 National Park are mapped as deforested
system at the landscape level, including its 1
World Resources Institute, Washington, DC 20002, instead of as ancient savannahs, and the
historical and cultural values, before mak- USA. 2International Union for Conservation of natural species-rich montane grasslands of
Nature, 1196, Gland, Switzerland. 3Department of
ing decisions on a restoration approach South Africa and southern Brazil are both
Geographical Sciences, University of Maryland,
PHOTO: FRANS LANTING/CORBIS

(8). FLR under this definition would College Park, MD 20742, USA. 4Silver Spring, mapped as deforested and/or degraded.
promote exactly the same conservation of MD 20901, USA. The Atlas label of forest for any
ancient grasslands that Bond advocates. *Corresponding author. E-mail: sdewitt@wri.org landscape with more than 10% tree canopy
Bonds claims also highlight disagree- is also misleading. Restoration strate-
REFERENCES
ment over how certain ecosystems should gies differ for forests and grasslands. The
1. World Resources Institute, Atlas of forest landscape
be classified. Grasslands and sparse forests restoration opportunities (World Resources Institute,
FLR handbooks promote fire suppression
are not mutually exclusive, often occur- Washington, DC, 2014); www.wri.org/applications/maps/ as one of the tools to restore forests and
ring together in mosaic ecosystems (9). flr-atlas/#. landscapes (6, 7). Although appropriate

1036 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS
for closed forests, fire suppression can such policies may be most effective when
be disastrous for ancient grassy systems incorporated into existing decision-making TECHNICAL COMMENT
defined, inappropriately, as forests based frameworks. ABSTRACTS
on the tree cover definition (8). A full res- As members of the Ecological Society
toration program would not only recognize of Americas (ESAs) Steering Committee Comment on Estimating the
ancient grasslands but also promote fire on the Intergovernmental Platform for reproducibility of psychological
and herbivory to help maintain them. It Biodiversity and Ecosystem Services science
would call for repeal of fire suppression (IPBES) (2), we urge U.S. federal agencies Daniel T. Gilbert, Gary King, Stephen
legislation that inhibits grassland fire to consider how this recent mandate pro- Pettigrew, Timothy D. Wilson
management and endorse active burning vides opportunities to leverage the global A paper from the Open Science
programs along with livestock farming as impact of IPBES while achieving national Collaboration (Research Articles, 28
suitable land practices. policy objectives. IPBES, which was estab- August 2015, aac4716) attempting to
I appreciate DeWitt et al.s call for col- lished with support from the United States replicate 100 published studies suggests
laboration with experts in open grassy but not mentioned in the administrations that the reproducibility of psychological
ecosystems. Laestadius et al. (3) also memo, provides scientific assessments of science is surprisingly low. We show that
noted the need for wider consultation. If the state of biodiversity and ecosystem ser- this article contains three statistical errors
the grassy biomes and the processes key vices (3). At its fourth plenary, one month and provides no support for such a conclu-
to their ecological integrity are formally before the administrations deadline, sion. Indeed, the data are consistent with
included in the restoration agenda, we will, IPBES will likely approve the first of its the opposite conclusion, namely, that the
at last, have put a century of misunder- global ecosystem service assessments. reproducibility of psychological science is
standing of tropical vegetation behind us. We encourage U.S. federal agencies to quite high.
William Bond familiarize themselves with these assess- Full text at http://dx.doi.org/10.1126/science.
South African Environmental Observation Network, ments and the ongoing IPBES work aad7243
National Research Foundation, Claremont, 7735, program. Creating a cross-agency policy
South Africa and Department of Biological Sciences,
agenda for understanding, monitoring, and Response to Comment on Estimating
University of Cape Town, Rondebosch, 7701,
South Africa. E-mail: William.Bond@uct.ac.za managing ecosystem services in the United the reproducibility of psychological
States would reduce the risk of scattered science
REFERENCES and inconsistent national-level policy Christopher J. Anderson, tepn Bahnk,
1. World Resources Institute, Atlas of forest landscape
mechanisms and would align U.S. policy Michael Barnett-Cowan, Frank A. Bosco,
restoration opportunities (World Resources Institute,
Washington, DC, 2014); www.wri.org/applications/maps/ and IPBES goals. Seeking input from the Jesse Chandler, Christopher R. Chartier,
flr-atlas/#. countrys scientific community would help Felix Cheung, Cody D. Christopherson,
2. S. Minnemeyer, L. Laestadius, N. Sizer, C. Saint-Laurent, build a secure foundation for these policies Andreas Cordes, Edward J. Cremata,
P. Potapov, A world of opportunity (World Resources
Institute, Washington, DC, 2011). and offer U.S. scientists an opportunity to Nicolas Della Penna, Vivien Estel,
3. L. Laestadius et al., Unasylva 238 62, 47 (2011/2012). contribute their knowledge to the scientific Anna Fedor, Stanka A. Fitneva,
4. A. M. Aubrville, Unasylva 1, 5 (1947). foundations upon which effective environ- Michael C. Frank, James A. Grange,
5. J. P. H. Acocks, Veld Types of South Africa (Botanical
Research Institute, ed. 3, 1953). mental policy rests. Joshua K. Hartshorne, Fred Hasselman,
6. ITTO and IUCN, Restoring forest landscapes, ITTO Lucas N. Joppa,1* James W. Boyd,2 Felix Henninger, Marije van der Hulst,
Technical Series No. 23 (2005). Clifford S. Duke,3 Stephanie Hampton,4 Kai J. Jonas, Calvin K. Lai, Carmel A.
7. IUCN and World Resources Institute, A guide to the
Restoration Opportunities Assessment Methodology
Stephen T. Jackson,5 Katharine L. Levitan, Jeremy K. Miller, Katherine S.
(ROAM) (IUCN and WRI, Gland, Switzerland, 2014); Jacobs,6 Karim-Aly S. Kassam,7 Harold Moore, Johannes M. Meixner, Marcus
http://cmsdata.iucn.org/downloads/roam_handbook_ A. Mooney,8 Laura A. Ogden,9 Mary R. Munaf, Koen I. Neijenhuijs, Gustav
lowres_web.pdf
Ruckelshaus,10 Jason F. Shogren11 Nilsonne, Brian A. Nosek, Franziska
8. C. L. Parr et al., Trends Ecol. Evol. 29, 205 (2014).
1
Microsoft Research, Redmond, WA 98052, USA. Plessow, Jason M. Prenoveau, Ashley A.
2
Resources for the Future, Washington, DC 20036, Ricker, Kathleen Schmidt, Jeffrey R. Spies,
USA. 3Ecological Society of America, Washington,
Government: Plan for DC 20036, USA. 4Washington State University,
Center for Environmental Research, Education
Stefan Stieger, Nina Strohminger, Gavin
B. Sullivan, Robbie C. M. van Aert,
ecosystem services and Outreach, Pullman, WA 99164, USA. 5United
States Geological Survey, Southwest Climate
Marcel A. L. M. van Assen, Wolf
Vanpaemel, Michelangelo Vianello,
Science Center, Tucson, AZ 85721, USA. 6University
NATURAL AND MANAGED ecosystems of Arizona, Center for Climate Adaptation Science Martin Voracek, Kellylynn Zuni
provide food, water, and other valuable and Solutions, Tucson, AZ 85721, USA. 7Cornell Gilbert et al. conclude that evidence
services to human societies. Unnoticed University, Department of Natural Resources, from the Open Science Collaborations
Ithaca, NY 14850, USA. 8Stanford University,
by many in the scientific community, the Department of Biological Sciences, Stanford, CA Reproducibility Project: Psychology
values associated with ecosystem services 94305, USA. 9Dartmouth University, Department indicates high reproducibility, given the
have been integrated into U.S. government of Anthropology, Hanover, NH 03755, USA. 10The study methodology. Their very optimis-
Natural Capital Project, Stanford University,
policy. A recent administration memo (1) Stanford, CA 94305, USA. 11University of Wyoming,
tic assessment is limited by statistical
put U.S. federal agencies on notice: The College of Business, Laramie, WY 82071, USA. misconceptions and by causal inferences
clock is ticking to integrate ecosystem *Corresponding author. from selectively interpreted, correlational
services into their planning and decision- E-mail: lujoppa@microsoft.com data. Using the Reproducibility Project:
making. By 30 March 2016, agencies are REFERENCES Psychology data, both optimistic and
to describe approaches for conduct- 1. S. Donovan, C. Goldfuss, J. Holdren, Incorporating pessimistic conclusions about reproduc-
ing decision-relevant and scale-specific Ecosystem Services into Federal Decision Making ibility are possible, and neither are yet
ecosystem-services assessments, as well as (2015); www.whitehouse.gov/sites/default/files/omb/ warranted.
memoranda/2016/m-16-01.pdf.
plans for effective monitoring and evalu- 2. ESA, IPBES (http://esa.org/ipbes/). Full text at http://dx.doi.org/10.1126/science.
ation. The administration stresses that 3. IPBES (http://ipbes.net/). aad9163

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1037


Published by AAAS
R ES E A RC H

referred us to another of his projects that did. The


T E C H N I C A L CO M M E N T Many Labs project (MLP) (9) involved 36 inde-
pendent laboratories that attempted to replicate
each of 16 original psychology studies, resulting
PSYCHOLOGY in 574 replication studies. These replication studies,
like OSCs replication studies, did not always use

Comment on Estimating original populations and procedures, so their data


allow us to estimate the amount of error that
sampling and infidelity together introduce. To
the reproducibility of make this estimate, we simply treated each of the
studies reported by MLP as an original effect

psychological science and then counted how many of the remaining


replications of that particular study observed
that original effect. This analysis revealed that
Daniel T. Gilbert,1* Gary King,1 Stephen Pettigrew,1 Timothy D. Wilson2 when infidelities were allowed, only 65.5% of the
replication effects fell within the confidence in-
A paper from the Open Science Collaboration (Research Articles, 28 August 2015, tervals of the original effects. Applying this esti-
aac4716) attempting to replicate 100 published studies suggests that the reproducibility of mate to OSCs data produces a sobering conclusion:
psychological science is surprisingly low. We show that this article contains three If every one of the 100 original studies that OSC
statistical errors and provides no support for such a conclusion. Indeed, the data are attempted to replicate had described a true ef-
consistent with the opposite conclusion, namely, that the reproducibility of psychological fect, then more than 34 of their replication studies
science is quite high. should have failed by chance alone. [All informa-

T
tion and code necessary to replicate our results
he replication of empirical research is a crit- marks assume that the one and only way in which are archived in Dataverse (10).] The bottom line

Downloaded from on March 3, 2016


ical component of the scientific process, and OSCs replication studies differed from the orig- is that OSC allowed considerable infidelities that
attempts to assess and improve the repro- inal studies is that they drew new samples from introduced random error and decreased the rep-
ducibility of science are important. The Open the original population. In fact, many of OSCs rep- lication rate but then compared their results to
Science Collaboration (OSC) (1) conducted lication studies differed from the original studies a benchmark that did not take this error into
a large-scale, collaborative effort to obtain an ini- in other ways as well. account.
tial estimate of the reproducibility of psychologi- For example, many of OSCs replication studies Second, we will discuss the issue of power.
cal science by attempting to replicate 100 original drew their samples from different populations OSC attempted to replicate each of 100 studies
studies that had been published in one of three than the original studies did. An original study that just once, and that attempt produced an un-
top-tier psychology journals in 2008. Depending measured Americans attitudes toward African- settling result: Only 47% of the original studies
on the criterion used, only 36 to 47% of the orig- Americans (3) was replicated with Italians, who were successfully replicated (i.e., produced effects
inal studies were successfully replicated, which do not share the same stereotypes; an original that fell within the confidence interval of the
led many to conclude that there is a replication study that asked college students to imagine being original study). In contrast, MLP attempted to
crisis in psychological science (2). Here, we show called on by a professor (4) was replicated with replicate each of its studies 35 or 36 times and
that when these results are corrected for error, participants who had never been to college; and then pooled the data. MLPs much more pow-
power, and bias, they provide no support for this an original study that asked students who com- erful method produced a much more hearten-
conclusion. In fact, the data are consistent with mute to school to choose between apartments that ing result: A full 85% of the original studies were
the opposite conclusion, namely, that the repro- were short and long drives from campus (5) was successfully replicated. What would have hap-
ducibility of psychological science is quite high. replicated with students who do not commute to pened to MLPs heartening result if they had
First, we will discuss the issue of error. If an school. Whats more, many of OSCs replication used OSCs method? Of MLPs 574 replication
original study reports a true effect, and if a rep- studies used procedures that differed from the studies, only 195 produced effects that fell within
lication study uses the original procedures with original studys procedures in substantial ways: the confidence interval of the original, published
a new sample of subjects drawn from the orig- An original study that asked Israelis to imagine study. In other words, if MLP had used OSCs
inal population, the replication study will some- the consequences of military service (6) was rep- method, they would have reported an unsettling
times fail to replicate the original effect because licated by asking Americans to imagine the con- replication rate of 34% rather than the hearten-
of sampling error alone. If all 100 of the orig- sequences of a honeymoon; an original study that ing 85% they actually reported. (A similar result
inal studies examined by OSC had reported true gave younger children the difficult task of locating occurs when we limit our analysis to those MLP
effects, then sampling error alone should cause targets on a large screen (7) was replicated by replication studies that had sample sizes at least
5% of the replication studies to fail by producing giving older children the easier task of locating as large as the original studies.) Clearly, OSC used
results that fall outside the 95% confidence in- targets on a small screen; an original study that a method that severely underestimates the actual
terval of the original study and 8% to fail by showed how a change in the wording of a chari- rate of replication.
producing results that are not also statistically table appeal sent by mail to Koreans could boost Third, we will discuss the issue of bias. The
significant (with the same sign). OSC used the response rates (8) was replicated by sending foregoing analyses generously assume that infi-
latter figure as the benchmark to which the ac- 771,408 e-mail messages to people all over the delities are a source of random error that are
tual replication failure rate in their data was world (which produced a response rate of essen- equally likely to increase or decrease the likeli-
compared. Neither of these figures provides an tially zero in all conditions). hood of successful replication. Is this assumption
appropriate benchmark, however, because both All of these infidelities are potential sources of true, or were the infidelities in OSCs replication
assume that sampling error is the only source of random error that the OSCs benchmark did not studies more likely to decrease than to increase the
error in the data. In other words, these bench- take into account. So how many of their replica- likelihood of successful replication? Answering this
tion studies should we expect to have failed by question requires an indicator of the fidelity of
1
chance alone? Making this estimate requires having each replication study, which OSC attempted
Harvard University, Cambridge, MA, USA. 2University of
Virginia, Charlottesville, VA, USA.
data from multiple replications of the same orig- to provide. Before conducting each replication
*Corresponding author. E-mail: gilbert@wjh.harvard.edu inal study. Although OSC did not collect such data, study, OSC asked the authors of the original
Authors are listed alphabetically. the corresponding author of OSC, Brian Nosek, study whether they endorsed the methodological

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1037-b


R ES E A RC H | T E C H N I C A L CO M M E N T

protocol for the to-be-attempted replication. Only have found that 59.7% [95% confidence interval failure. As a result, OSC seriously underestimated
69% of the original authors did. Although en- (CI): 47.5%, 70.9%] were replicated successfully. In the reproducibility of psychological science.
dorsement is an imperfect indicator that may fact, we estimate that if all the replication studies
overestimate the fidelity of a replication study had been high enough in fidelity to earn the en- REFERENCES AND NOTES
(e.g., some of the original authors may have dorsement of the original authors, then the rate 1. Open Science Collaboration, Science 349, aac4716 (2015).
knowingly endorsed low-fidelity protocols and of successful replication would have been 58.6% 2. B. Carey, Psychologys fears confirmed: Rechecked studies
dont hold up. New York Times (27 August 2015), p. A1.
others may have discovered that the replication (95% CI: 47.0%, 69.5%) when controlling for rel- 3. B. K. Payne, M. A. Burkley, M. B. Stokes, J. Pers. Soc. Psychol.
studies were low fidelity only after they were evant covariates. Remarkably, the CIs of these es- 94, 1631 (2008).
completed) or may underestimate the fidelity timates actually overlap the 65.5% replication rate 4. J. L. Risen, T. Gilovich, J. Pers. Soc. Psychol. 95, 293307
of a replication study (e.g., endorsement deci- that one would expect if every one of the original (2008).
5. E. J. Masicampo, R. F. Baumeister, Psychol. Sci. 19, 255260
sions may be influenced by original authors studies had reported a true effect. Although that (2008).
suspicions about the weakness of their studies seems rather unlikely, OSCs data clearly provide 6. N. Shnabel, A. Nadler, J. Pers. Soc. Psychol. 94, 116132 (2008).
rather than by the fidelity of the replication no evidence for a replication crisis in psycholog- 7. V. LoBue, J. S. DeLoache, Psychol. Sci. 19, 284289 (2008).
protocol), it is nonetheless the best indicator of ical science. 8. M. Koo, A. Fishbach, J. Pers. Soc. Psychol. 94, 183195 (2008).
9. R. A. Klein et al., Soc. Psychol. 45, 142152 (2014).
fidelity in OSCs data. So what does that indi- We applaud efforts to improve psychological 10. D. T. Gilbert, G. King, S. Pettigrew, T. D. Wilson, Replication
cator indicate? science, many of which have been careful, respon- data for Comment on Estimating the reproducibility of
When we compared the replication rates of the sible, and effective (11), and we appreciate the effort psychological science, Harvard Dataverse, V1;
endorsed and unendorsed protocols, we discovered that went into producing OSC. But metascience http://dx.doi.org/10.7910/DVN/5LKVH2 (2016).
11. J. P. Simmons, L. D. Nelson, U. Simonsohn, Psychol. Sci. 22,
that the endorsed protocols were nearly four times is not exempt from the rules of science. OSC used 13591366 (2011).
as likely to produce a successful replication (59.7%) a benchmark that did not take into account the
as were the unendorsed protocols (15.4%). This multiple sources of error in their data, used a AC KNOWLED GME NTS
strongly suggests that the infidelities did not just relatively low-powered design that demonstrably We acknowledge the support of NSF Grant BCS-1423747 to T.D.W.
introduce random error but instead biased the underestimates the true rate of replication, and and D.T.G.

replication studies toward failure. If OSC had lim- permitted considerable infidelities that almost 26 October 2015; accepted 28 January 2016
ited their analyses to endorsed studies, they would certainly biased their replication studies toward 10.1126/science.aad7243

1037-b 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H

tive sample of studies in high-impact journals.


TECHNICAL RESPONSE Had Gilbert et al. selected the similar Many Labs
3 study (6) instead of ML2014, they would have
arrived at a more pessimistic conclusion: a 30%
PSYCHOLOGY overall replication success rate with a multisite,
very high-powered design.
That said, Gilbert et al.s analysis demonstrates
Response to Comment on that differences between laboratories and sam-

Estimating the reproducibility of


ple populations reduce reproducibility according
to the CI measure. Also, some true effects may
exist even among nonsignificant replications (our

psychological science additional analysis finding evidence for these


effects is available at https://osf.io/smjge). True
effects can fail to be detected because power cal-
Christopher J. Anderson,1* tpn Bahnk,2 Michael Barnett-Cowan,3 Frank A. Bosco,4 culations for replication studies are based on
Jesse Chandler,5,6 Christopher R. Chartier,7 Felix Cheung,8 Cody D. Christopherson,9 effect sizes in original studies. As OSC2015 dem-
Andreas Cordes,10 Edward J. Cremata,11 Nicolas Della Penna,12 Vivien Estel,13 onstrates, original study effect sizes are likely
Anna Fedor,14 Stanka A. Fitneva,15 Michael C. Frank,16 James A. Grange,17 inflated due to publication bias. Unfortunately,
Joshua K. Hartshorne,18 Fred Hasselman,19 Felix Henninger,20 Marije van der Hulst,21 Gilbert et al.s focus on the CI measure of re-
Kai J. Jonas,22 Calvin K. Lai,23 Carmel A. Levitan,24 Jeremy K. Miller,25 producibility neither addresses nor can account
Katherine S. Moore,26 Johannes M. Meixner,27 Marcus R. Munaf,28 for the facts that the OSC2015 replication effect
Koen I. Neijenhuijs,29 Gustav Nilsonne,30 Brian A. Nosek,31,32 Franziska Plessow,33 sizes were about half the size of the original
Jason M. Prenoveau,34 Ashley A. Ricker,35 Kathleen Schmidt,36 Jeffrey R. Spies,31,32 studies on average, and 83% of replications elic-
ited smaller effect sizes than the original studies.

Downloaded from on March 3, 2016


Stefan Stieger,37 Nina Strohminger,38 Gavin B. Sullivan,39 Robbie C. M. van Aert,40
The combined results of OSC2015s five indica-
Marcel A. L. M. van Assen,40,41 Wolf Vanpaemel,42 Michelangelo Vianello,43
tors of reproducibility suggest that, even if true,
Martin Voracek,44 Kellylynn Zuni45
most effects are likely to be smaller than the
original results suggest.
Gilbert et al. conclude that evidence from the Open Science Collaborations Reproducibility
Gilbert et al. attribute some of the failures to
Project: Psychology indicates high reproducibility, given the study methodology. Their
replicate to low-fidelity protocols with meth-
very optimistic assessment is limited by statistical misconceptions and by causal
odological differences relative to the original, for
inferences from selectively interpreted, correlational data. Using the Reproducibility
which they provide six examples. In fact, the
Project: Psychology data, both optimistic and pessimistic conclusions about
original authors recommended or endorsed three
reproducibility are possible, and neither are yet warranted.
of the six methodological differences discussed

A
cross multiple indicators of reproducibil- of other ML2014 studies of the same phenome- 1
Russell Sage College, Troy, NY, USA. 2University of
ity, the Open Science Collaboration (1) non and concluded that this reflects the max- Wrzburg, Wrzburg, Germany. 3University of Waterloo,
(OSC2015) observed that the original result imum reproducibility rate for OSC2015. Their Waterloo, Ontario, Canada. 4Virginia Commonwealth
was replicated in ~40 of 100 studies sam- analysis using ML2014 is misleading and does University, Richmond, VA, USA. 5University of Michigan, Ann
pled from three journals. Gilbert et al. (2) not apply to estimating reproducibility with Arbor, MI 48104, USA. 6Mathematica Policy Research,
Washington, DC, USA. 7Ashland University, Ashland, OH,
conclude that the reproducibility rate is, in fact, OSC2015s data for a number of reasons. USA. 8Michigan State University, East Lansing, MI, USA.
as high as could be expected, given the study First, Gilbert et al.s estimates are based on 9
Southern Oregon University, Ashland, OR, USA. 10University
methodology. We agree with them that both pairwise comparisons between all of the repli- of Gttingen, Institute for Psychology, Gttingen, Germany.
11
methodological differences between original cations within ML2014. As such, for roughly half University of Southern California, Los Angeles, CA, USA.
12
Australian National University, Canberra, Australia.
and replication studies and statistical power of their failures to replicate, replications had 13
Technische Universitt Braunschweig, Braunschweig,
affect reproducibility, but their very optimistic larger effect sizes than original studies, whereas Germany. 14Parmenides Stiftung, Munich, Germany.
15
assessment is based on statistical misconceptions just 5% of OSC2015 replications had replication Queens University, Kingston, Ontario, Canada. 16Stanford
and selective interpretation of correlational data. CIs exceeding the original study effect sizes. University, Stanford, CA, USA. 17Keele University, Keele,
Staffordshire, UK. 18Boston College, Chestnut Hill, MA, USA.
Gilbert et al. focused on a variation of one of Second, Gilbert et al. apply the by-site varia- 19
Radboud University Nijmegen, Nijmegen, Netherlands.
OSC2015s five measures of reproducibility: how bility in ML2014 to OSC2015s findings, thereby 20
University of Koblenz-Landau, Landau, Germany. 21Erasmus
often the confidence interval (CI) of the original arriving at higher estimates of reproducibility. Medical Center, Rotterdam, Netherlands. 22University of
study contains the effect size estimate of the rep- However, ML2014s primary finding was that Amsterdam, Amsterdam, Netherlands. 23Harvard University,
Cambridge, MA, USA. 24Occidental College, Los Angeles, CA,
lication study. They misstated that the expected by-site variability was highest for the largest USA. 25Willamette University, Salem, OR, USA. 26Arcadia
replication rate assuming only sampling error is (replicable) effects and lowest for the smallest University, Glenside, PA, USA. 27University of Potsdam,
95%, which is true only if both studies estimate the (nonreplicable) effects. If ML2014s primary find- Potsdam, Germany. 28University of Bristol, Bristol, UK.
29
same population effect size and the replication has ing is generalizable, then Gilbert et al.s analysis 30
Vrije Universiteit Amsterdam, Amsterdam, Netherlands.
Karolinska Institutet, Stockholm University, Stockholm,
infinite sample size (3, 4). OSC2015 replications did may leverage by-site variability in ML2014s larger Sweden. 31Center for Open Science, Charlottesville, VA, USA.
not have infinite sample size. In fact, the expected effects to exaggerate the effect of by-site variabil- 32
University of Virginia, Charlottesville, VA, USA. 33Harvard
replication rate was 78.5% using OSC2015s CI ity on OSC2015s nonreproduced smaller effects, Medical School, Boston, MA, USA. 34Loyola University,
measure (see OSC2015s supplementary informa- thus overestimating reproducibility. Baltimore, MD, USA. 35University of California, Riverside, CA,
USA. 36Wesleyan University, Middletown, CT, USA. 37University
tion, pp. 56 and 76; https://osf.io/k9rnd). By this Third, Gilbert et al. use ML2014s 85% repli- of Konstanz, Konstanz, Germany. 38Yale University, New Haven,
measure, the actual replication rate was only cation rate (after aggregating across all 6344 par- CT, USA. 39Coventry University, Coventry, UK. 40Tilburg
47.4%, suggesting the influence of factors other ticipants) to argue that reproducibility is high University, Tilburg, Netherlands. 41Utrecht University, Utrecht,
than sampling error alone. when extremely high power is used. This inter- Netherlands. 42University of Leuven, Leuven, Belgium.
43
University of Padova, Padova, Italy. 44University of Vienna,
Within another large replication study, Many pretation is based on ML2014s small, ad hoc Vienna, Austria. 45Adams State University, Alamosa, CO, USA.
Labs (5) (ML2014), Gilbert et al. found that sample of classic and new findings, as opposed to *Authors are listed alphabetically. Corresponding author. E-mail:
65.5% of ML2014 studies would be within the CIs OSC2015s effort to examine a more representa- nosek@virginia.edu

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1037-c


R ES E A RC H | T E C H N I C A L RE S P O N S E

by Gilbert et al., and a fourth (the racial bias study evidence in OSC2015 countering their interpreta- testing to determine why. When results do not
from America replicated in Italy) was replicated tion, such as evidence that surprising or more differ, it offers some evidence that the finding is
successfully. Gilbert et al. also supposed that non- underpowered research designs (e.g., interaction generalizable. OSC2015 provides initial, not defin-
endorsement of protocols by the original authors tests) were less likely to be replicated. In sum, itive, evidencejust like the original studies it
was evidence of critical methodological differ- Gilbert et al. made a causal interpretation for replicated.
ences. Then they showed that replications that OSC2015s reproducibility with selective inter-
were endorsed by the original authors were more pretation of correlational data. A constructive step REFERENCES AND NOTES
likely to be replicated than those not endorsed forward would be revising the previously non- 1. Open Science Collaboration, Science 349, aac4716 (2015).
(nonendorsed studies included 18 original au- endorsed protocols to see if they can achieve en- 2. D. T. Gilbert, G. King, S. Pettigrew, T. D. Wilson, Science 351,
thors not responding and 11 voicing concerns). dorsement and then conducting replications with 1037 (2016).
3. G. Cumming, R. Maillardet, Psychol. Methods 11, 217227 (2006).
In fact, OSC2015 tested whether rated similarity the updated protocols to see if reproducibility 4. G. Cumming, J. Williams, F. Fidler, Underst. Stat. 3, 299311
of the replication and original study was corre- rates improve. (2004).
lated with replication success and observed weak More generally, there is no such thing as exact 5. R. A. Klein et al., Soc. Psychol. 45, 142152 (2014).
relationships across reproducibility indicators replication (810). All replications differ in innu- 6. C. R. Ebersole et al., J. Exp. Soc. Psychol. 65 (2016);
https://osf.io/csygd.
(e.g., r = 0.015 with P < 0.05 criterion; supplemen- merable ways from original studies. They are con- 7. A. Dreber et al., Proc. Natl. Acad. Sci. U.S.A. 112, 1534315347
tary information, p. 67; https://osf.io/k9rnd). ducted in different facilities, in different weather, (2015).
Further, there is an alternative explanation for with different experimenters, with different com- 8. B. A. Nosek, D. Lakens, Soc. Psychol. 45, 137141 (2014).
the correlation between endorsement and repli- puters and displays, in different languages, at dif- 9. Open Science Collaboration, Perspect. Psychol. Sci. 7, 657660
(2012).
cation success; authors who were less confident ferent points in history, and so on. What counts 10. S. Schmidt, Rev. Gen. Psychol. 13, 90100 (2009).
of their studys robustness may have been less as a replication involves theoretical assessments
likely to endorse the replications. Consistent with of the many differences expected to moderate a AC KNOWLED GME NTS
the alternative account, prediction markets ad- phenomenon. OSC2015 defined (direct) replica- Preparation of this response was supported by grants from the
ministered on OSC2015 studies showed that it is tion as the attempt to recreate the conditions Laura and John Arnold Foundation and the John Templeton
Foundation.
possible to predict replication failure in advance believed sufficient for obtaining a previously ob-
based on a brief description of the original find- served finding. When results differ, it offers an 2 December 2015; accepted 28 January 2016
ing (7). Finally, Gilbert et al. ignored correlational opportunity for hypothesis generation and then 10.1126/science.aad9163

1037-c 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H

fectively adjust their number of output spikes.


RESEARCH ARTICLE SUMMARY Because a neurons discrete number of spikes
does not provide a direction of gradual im-
provement, I derived the multi-spike tempotron
NEURAL COMPUTATION
learning rule in an abstract space of con-
tinuous spike threshold variables. In this
Spiking neurons can discover space, changes in synaptic efficacies are di-
rected along the steepest path, reducing the

predictive features by discrepancy between a neurons fixed biolog-


ical spike threshold and the closest hypothet-
ical threshold at which the neuron would fire
aggregate-label learning a desired number of spikes. With the result-
ing synaptic learning rule, aggregate-label
Robert Gtig learning enabled simple neuron models to
solve the temporal credit assignment prob-
INTRODUCTION: Opportunities and dangers remain silent otherwise. Hence, the number of lem. Neurons reliably identified all clues whose
can often be predicted on the basis of sensory output spikes of this neuron should be pro- occurrences contributed to a delayed feedback
clues. The attack of a predator, for example, portional to the number of times that the clue signal. For instance, a neuron could learn to
may be preceded by the sounds of breaking occurred. The reversal of this observation is respond with different numbers of spikes to
twigs or whiffs of odor. Life is easier if one the core hypothesis of this study: A neuron can individual clues without being told how many
learns these clues. However, this is difficult identify an unknown clue when it is trained to different clues existed, when they occurred,
when clues are hidden within distracting fire in proportion to the clues number of oc- or how much each one of them contributed to
streams of unrelated sensory activity. Even currences. This aggregate-label hypothesis the feedback. This learning was robust to high

Downloaded from on March 3, 2016


worse, they can be separated from the events entails that when a neuron is trained to match levels of feedback and input noise and performed
that they predict by long and variable delays. its number of output spikes to the magnitude well on a connected speech-recognition task.
To discover those clues, a learning procedure of a feedback signal, it will identify a set of Aggregate-label learning enabled populations
must bridge the gap between the short epochs clues within its input activity whose occur- of neurons to solve unsupervised learning tasks
within which clues occur and the time when rences predict the feedback. This learning by relying on internally generated feedback
feedback arrives. This temporal credit-assignment requires neither knowledge of the time nor
signals that amplified cor-
problem is a core challenge in biological and of the absolute number of individual clues. ON OUR WEB SITE relations between the neu-
machine learning. Read the full article rons output spike counts.
RESULTS: To implement aggregate-label learn- at http://dx.doi. These self-supervised net-
RATIONALE: A neural detector of a sensory ing, I calculated how neurons should modify org/10.1126/ works discovered reoccur-
clue should fire whenever the clue occurs but their synaptic efficacies in order to most ef- science.aab4113 ring constellations of input
..................................................
patterns even if they were
rare and distributed over spatial and tempo-
ral scales that exceeded the receptive fields of
individual neurons. Because learning in self-
supervised networks is driven by aggregate
numbers of feature occurrences, it does not
require temporal alignment of the input activ-
ities of individual neurons. When competitive
interactions between individual neurons were
mediated through the internal feedback circuit,
the formation of feature maps was possible
even when the features asynchrony incapac-
itated lateral inhibition.

CONCLUSION: Aggregate-label learning solves


the long-standing question of how neural sys-
tems can identify features within their input
activity that predict a delayed feedback. This
solution strongly enhances the known learn-
ing capabilities of simple neural circuit models.
Because the feedback can be external or inter-
nal, these enhancements apply to supervised
and unsupervised learning. In this framework,
both forms of learning converge onto the same
rule of synaptic plasticity, inviting future research
Membrane potential traces of a model neuron before and after learning to detect reward
predictive sensory clues. Before learning, top trace; after learning, second through fifth traces
on how they cooperate when brains learn.

from top; clues, colored squares. Each clue occurrence is represented as a spike pattern within
E-mail: guetig@em.mpg.de
the neurons input activity (raster plot). After learning, the number of output spikes (vertical Cite this article as R. Gtig, Science 351, aab4113 (2016).
deflections) elicited by each clue encodes the clues contribution to a delayed reward. DOI: 10.1126/science.aab4113.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1041


R ES E A RC H

the steepest path toward the desired number of


RESEARCH ARTICLE output spikes, and an approximation that is bio-
logically more plausible.

NEURAL COMPUTATION The multi-spike tempotron


The minimum disturbance principle (17) posits

Spiking neurons can discover that a neural implementation of aggregate-label


learning should use a synaptic learning rule that
adjusts a neurons synaptic efficacies in the di-
predictive features by rection along which its number of output spikes
changes most rapidly. Naive gradient-based ap-

aggregate-label learning proaches to find this direction fail in spiking neu-


rons because of the discrete nature of the number
of output spikes. Its derivatives with respect to a
Robert Gtig neurons synaptic efficacies are zero almost every-
where and infinite at points where a new spike
The brain routinely discovers sensory clues that predict opportunities or dangers. However, appears or an existing one disappears. To cir-
it is unclear how neural learning processes can bridge the typically long delays between cumvent this problem, stochastic approaches have
sensory clues and behavioral outcomes. Here, I introduce a learning concept, aggregate- been tried with mixed success (11, 18, 19). Here, I
label learning, that enables biologically plausible model neurons to solve this temporal introduce an alternative approach that is based on
credit assignment problem. Aggregate-label learning matches a neurons number of output a continuous deterministic objective function for
spikes to a feedback signal that is proportional to the number of clues but carries no spiking neurons, the spike-threshold-surface. For
information about their timing. Aggregate-label learning outperforms stochastic each synaptic configuration and presynaptic activ-
reinforcement learning at identifying predictive clues and is able to solve unsegmented ity pattern, the spike-threshold-surface is defined
speech-recognition tasks. Furthermore, it allows unsupervised neural networks to discover as the mapping between the neurons firing thresh-
reoccurring constellations of sensory features even when they are widely dispersed across old and the number of elicited output spikes
space and time. (materials and methods, spike-threshold-surface)

T
(fig. S1). On the basis of this curve, the discrete num-
he fitness of an individual depends on its can provide the direction of synaptic changes in ber of elicited output spikes is replaced with the
ability to recognize circumstances that re- spiking neurons has remained elusive (10). Appli- distance between the fixed biological firing thresh-
quire specific actions to procure reward or cations of stochastic reinforcement learning have old of a neuron and the closest hypothetical
avoid punishment. The approach of a pre- resulted in slow learning and limited performance threshold at which the neuron would fire the
dator, for example, may be preceded by one (3, 4, 1113). Other, non-neuromorphic solutions desired number of spikes. When this distance
or more brief sensory clues, such as the sounds of rely on complex iterative optimization algorithms is zero, the neuron responds to the input pattern
breaking twigs or whiffs of odor. Humans and such as expectation-maximization (14). Their bio- with the desired number of spikes. The multi-spike
other animals are very good at learning such logically plausible implementation has remained tempotron is defined as the gradient-based learn-
associations, even when clues are hidden within challenging. Here, I show that the temporal credit ing rule that approaches the desired number of
distracting streams of unrelated sensory activity assignment problem can be solved by training a spikes along the (locally) steepest path (materials
of similar structure, and the delays between clues neuron to match its number of output spikes to and methods, multi-spike tempotron) (fig. S1D). I
and outcomes are variable and orders of mag- the magnitude of a feedback signal. implemented the multi-spike tempotron by cal-
nitudes longer than the durations of individual culating the exact gradient (materials and meth-
clues. To discover those clues, a learning pro- Results ods, *-gradient). I also approximated the learning
cedure must bridge the gap between the short Aggregate-label learning rule with a biologically established form of synaptic
epochs within which clues occur and the long A reliable clue detector should fire at least one plasticity that is driven by correlations between
delays after which feedback arrives. This tem- action potential whenever the clue occurs but must presynaptic firing and postsynaptic depolarization.
poral credit-assignment problem is a core chal- remain silent otherwise. Hence, the number of Specifically, in this correlation-based learning
lenge in biological and machine learning (14). spikes elicited within a given sensory episode should each synapse maintains a local trace of eligibility
Often, the assignment of credit to a particular grow with the number of clue occurrences. Because by correlating its own activity with the postsynaptic
sensory clue is made even harder by the fact we have only aggregate feedback information, membrane potential (materials and methods,
that the learner must not only discover when I used the total number of spikes when construct- correlation-based learning). Beyond these eligi-
but also in which sensory channel a predictive ing an objective function to drive learning (15). bilities, this learning does not require a memory
clue can be found. This is the spatial credit as- Here, I introduce the learning concept of aggregate- of past inputs or membrane potential values.
signment problem. label learning: The temporal credit assignment
To forecast opportunities or dangers, a neural problem is solved by a broad class of synaptic Learning predictive sensory clues
feature detector should be active whenever a pre- learning rules that credit individual synapses I tested aggregate-label learning within a fam-
dictive sensory clue occurs but remain silent other- according to their ability to reduce the mismatch ily of generic feature learning tasks for single
wise. If the timing of the clues is known, synaptic between the number of output spikes and the mag- neurons. Each task consisted of a fixed set of
learning rules (58) that resemble perceptron nitude of the aggregate feedback. This feedback brief (50 ms long) activity patterns that repre-
learning (9) can realize such desired responses by is only assumed to be proportional to the num- sented responses of a neurons afferent periphery
strengthening synapses that depolarize the cell ber of clues but provides no information about their to occurrences of distinct features within an or-
at the time of a clue and weakening those that do timing or absolute numbers. Aggregate-label learn- ganisms sensory environment (Fig. 1A). In each
otherwise. However, when the timing of the clues ing entails that neurons can learn to respond syn- trial, a random number of feature activity pat-
is unknown, an appropriate objective function that chronously to predictive sensory clues from delayed terns was embedded within a long stream of
feedback signals (16). I implemented aggregate- random background activity at random times
Max Planck Institute of Experimental Medicine, Hermann-
label learning with two specific synaptic learning (materials and methods, embedded features).
Rein-Strasse 3, 37075 Gttingen, Germany. rules: the multi-spike tempotron, a gradient-based Each feature was either a clue or a distractor.
E-mail: guetig@em.mpg.de learning rule that adjusts synaptic efficacies along The task of the neuron was to signal all clues by

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-1


R ES E A RC H | R E S EA R C H A R T I C LE

firing a specific number of action potentials when- based approximation of the mutli-spike tempotron (Fig. 2A). It also tolerates a wide range of clue
ever the corresponding feature activity pattern resulted in a similar learning performance (fig. S3A). frequencies (Fig. 2B). The learning is robust to
occurred within its input stream. In contrast, the Aggregate-label learning is not limited to a par- substantial correlations between clue and distrac-
neuron had to remain silent during intervals of ticular gain between a clues number of occur- tor occurrences (materials and methods, correlated
background activity or when encountering a dis- rences and the feedback signal. When the desired feature occurrences) (Fig. 2C) and reproduces the
tractor. In a first test, I trained the multi-spike number of output spikes corresponded to five phenomenon of blocking (2) when correlations
tempotron to detect a single clue (out of 10 fea- times the number of the clues, the neuron learned are high (fig. S2D). When compared with a recent
tures) by matching each trials aggregate label to to fire a burst of five spikes in response to each neural implementation of reinforcement learning
the clues number of occurrences. All remaining clue (Fig. 1B, third trace; fig. S2A; and table S1). (4), aggregate-label learning required much fewer
features were distractors. Whenever the cell fired Aggregate-label learning also enables neurons to synapses and learning trials to solve a feature
fewer or more spikes than there were clues, the decompose the feedback signal and identify the learning task that was similar (although substan-
learning rule adjusted the neurons synaptic effi- individual values of multiple sensory clues whose tially simpler) to the ones studied here (materials
cacies along the direction of the steepest path occurrences are mixed within a sensory scene (Fig. and methods, reinforcement learning comparison
toward the synaptic configuration that elicited 1, B, fourth and fifth traces, and C, right; fig. S2B; task) (table S2) (20). Last, when sensory clues were
one more or one less spike, respectively (materials and table S1). This problem arises in many etho- not discrete but instead drawn from a continu-
and methods, multi-spike tempotron, and Eq. 7). logical settings (2) whenever feedback signals in- ous feature dimension, aggregate-label learning
Neural responses (materials and methods, neural corporate contributions from several sources, such enabled neurons to learn continuous tuning curves
responses) to random background activity and to as food quality being associated with different (supplementary text S1 and fig. S4). The outcome
all distractor features declined rapidly, and the color and odor clues. Our neural implementation of seeing a face, for instance, is likely to depend
cell learned to fire exactly one spike whenever of aggregate-label learning remains efficient even on its direction of gaze (ranging from looking
the clue occurred (Fig. 1, B, second trace, and C, when applied to long sensory episodes dominated straight at us to looking away) or its emotional
left; and fig. S2A) (response statistics are provided by background activity (Fig. 2A and fig. S2C) or expression (between angry and friendly).
in table S1). I have checked that the correlation- when the number of distractor features is high
Noisy feedback
In many learning tasks, the relation between clues
and feedback can be unreliable. For instance,
sensory clues can be occluded or feedback signals
corrupted by noise. To test aggregate-label learn-
ing in such conditions, I modified the desired
number of spikes on each trial by adding or sub-
tracting (with equal probabilities) random Poisson
noise. I found that neurons continued to reliably
identify sensory clues even when the feedback noise
was high (Fig. 2D). However, increasing levels
of noise caused a gradual decline in the average
number of spikes elicited by the clues in favor of
spikes elicited by random background activity.
This behavior reflects the inability of the identified
clues to fully account for the (corrupted) feedback
signal and could be interpreted as the neurons
attempt to search for a better predictor within
the background activity. However, when the teacher
is wrong, definitions of good and bad learning
become circumstantial.
Prompted by its robustness to noise in the above
scenario, I next considered whether aggregate-label
learning can be extended to a well-studied prob-
abilistic feature learning task. In the weather
prediction task (21), relations between clues and
feedback are not deterministic, but rather, clue
occurrences increase the probability of a binary
reward at the end of a trial (materials and meth-
Fig. 1. Supervised learning of sensory clues. (A) Input activity (left raster plot) and schematic of a ods, probabilistic reward contingencies). In addi-
postsynaptic readout neuron (right). For visibility, only spikes (black ticks) of 10% of the neurons 500 tion to introducing feedback noise through
synaptic afferents are shown. Colored rectangles depict occurrences of Nf = 10 distinct sensory features probabilistic rewards, this scenario also relaxes
(color coded) whose corresponding 50-ms activity patterns have been embedded within a stream of ran- our previous assumption that the feedback is
dom background activity. (B) Voltage traces of the readout neuron before (top trace) and after being trained proportional to the number of clue occurences.
to fire one (second trace) or five (third trace) spikes in response to a single clue (dark blue rectangles), Instead, here the number of clue occurences is
respectively. For the fourth and fifth traces, the neuron was trained to respond to five clues out of the 10 related to the reward (probability) through a
features (colored rectangles). The neuron either had to fire one spike in response to each of the clues saturating nonlinearity whose steepness is de-
(fourth trace) or one, two, three, four, and five spikes in response to the dark blue, light blue, green, yellow, termined by the clues reward contingency. The
and red clues, respectively (fifth trace). (Right) Enlargements of the underlined (gray horizontal lines) higher the contingency, the more step-like this
clue intervals. (C) Neural responses (materials and methods, neural responses) as a function of learning mapping becomes. Because of this tasks stochas-
cycles (each containing 100 trials) for the two tasks underlying the second (left) and fifth (right) voltage tic nature, the relation between clues and rewards
traces in (B), respectively. Colored lines show the mean numbers of spikes (right y axis) elicited by each has to be acquired and maintained across trials
feature [color coded as in (A)], and gray lines show the mean responses to background activity (left for example, the absence of a reward on a single
y axis). Population learning curves are provided in fig. S3, left. trial should not cause the immediate unlearning

aab4113-2 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

CT (cycles)

Rate (Hz)
Spikes

Spikes
CT
and MI Feedback noise Cycles

Fig. 2. Robustness of aggregate-label learning. (A to C) Mean convergence percentile of 1000 independent simulations. (D) Learned responses of a
times (CT) of a neuron that was trained to fire one spike in response to a single neuron that was trained to fire five spikes in response to a single clue cmean
f 5
clue for different task parameters. (A) For a fixed mean feature count cmeanf 5 as function of the average noise that corrupted the feedback signal on each
as a function of the mean trial duration Ttrial (blue circles) and for a fixed mean trial. Neural responses were averaged over the last 100 of 1000 cycles of
trial duration Ttrial = 50 s as a function of the number of features Nf (black training and converted to the expected number of spikes that were elicited by
squares). Solid lines depict exponential fits with slopes of 1.23 and 0.07, occurences of the clue (black), the nine distractors (gray), or the 7.5 s of
respectively. (B) For fixed mean trial durations Ttrial = 5 s (red) and Ttrial = 10 s background activity (white) during each trial. All measurements were averaged
(blue) as a function of the mean feature count cmean f . (Inset) The expected over 1000 independent simulations. (E) Mean convergence times in the prob-
number of clue encounters before convergence in the Ttrial = 10 s condition as a abilistic single-clue detection task as a function of the clues reward con-
function of cmean
f (same interval as in the main data) normalized to the value at tingency p. Colors (cool to warm) depict rising noise levels in the input activity
cmean
f = 0.1 (37.33). (C) For a fixed mean trial duration Ttrial = 5 s and a fixed patterns implemented by an increasing spike deletion probability of 0, 0.1, 0.2,
mean feature count cmean f 5 as a function of feature correlations r (blue 0.3, 0.4, and 0.5. (F) Neural responses as in Fig. 1C, but for the probabilistic
curve) or the corresponding normalized mutual information MI between the reward scenario. Three clues had reward contingencies of p1 1 (dark blue), p2
aggregate-label and the clue identity (red circles).The red line shows an inverse 0.8 (light blue), and p3 0.6 (green), respectively, and on each trial, a quarter of
power-law relationship, with a nearly quadratic exponent of 2.05. Shaded all input spikes were randomly deleted. Except the black line in (A), Nf = 10
areas in (A), (B), and (C) correspond to regions between the 10th and 90th throughout this figure.

of the corresponding clue. I therefore modified of words or phonemes within a known text. How- I repeated the comparison with a strongly re-
the learning scheme so that each rewarded trial ever, their temporal alignment with a correspond- duced training set in which target digits, if present,
invoked a learning step toward increasing the ing speech signal has been notoriously difficult occurred multiple times. Indeed, the perform-
neurons output spike number regardless of its and requires complex optimization algorithms (14). ance gap between the two models widened to
response. In contrast, unrewarded trials triggered I tested whether word counts are sufficient 47%. In contrast, the difference fell to 16% when
a learning step to reduce the number of output to successfully apply aggregate-label learning trials with multiple target digit occurrences were
spikes only in the case of a false alarm,, when the to continuous speech recognition problems. I excluded. This remaining difference shows that
neuron had fired at least one spike. This asym- addressed this question by extending a recent the ability of the multi-spike tempotron learning
metric learning continues to strengthen the neu- neural model of spoken digit recognition (mate- rule to enforce a single output spike, as opposed
rons responses to the sensory clues even after they rials and methods, auditory front-end) (23) from to realizing at least one, resulted in an advantage
have been identified, increasing their robustness. binary to multi-spike tempotrons. In contrast to over the binary tempotron, even when the labels
In addition, selectivity of the learned neural re- the previously used TI46 database, which is lim- of a task were binary. I expect that the advan-
sponses is enhanced by synaptic competition (22) ited to utterances of single digits, I used the tages of aggregate-label learning will become
that was induced here by bounding the overall TIDIGITS speech corpus (24), which also contains even more pronounced in applications to neural
strength of the neurons synaptic efficacies through sequences of digits of variable length (ranging models of phoneme detection that will encounter
divisive normalization (materials and methods, from 2 to 7). With all utterances of single digits higher numbers of target feature occurrences.
probabilistic reward contingencies). The resulting reserved to test performance, training was based
learning reliably identified predictive clues, even if on unsegmented digit sequences only. Specifi- Self-supervised neural networks
their reward contingencies were far from perfect cally, I used the number of occurrences of a sin- Consider the above family of feature learning
and, in addition, all activity patterns were corrup- gle target digit as aggregate-label to train neurons tasks, however, without any feedback (25, 26).
ted by random deletions of presynaptic spikes on to discriminate between this target and the re- How can neurons still discover synaptic activ-
each trial (Fig. 2E). I checked that the learned maining distractor digits. I found that multi- ity patterns that represent repeated occurrences
neural responses were hardly affected by the pres- spike tempotrons reliably learned to fire a single of sensory features in the environment? In gen-
ence of false rewards that were not predicted by spike whenever their target digit occurred and to eral, structure is discriminated from noise through
any of the available clues. The main effect of such remain silent otherwise (Fig. 3A). Across all digits, correlations. Indeed, existing models of unsuper-
false rewards was to slow down the clue acquisi- the majority of simulations achieved zero errors vised learning in spiking neural networks have
tion (fig. S5A). When the reward probability re- over the test set (Fig. 3B), with an average test er- typically relied on spatiotemporally local corre-
sulted from the combination of several different ror of 0.0055 (materials and methods, TIDIGITS- lations between the pre- and postsynaptic ac-
clues, the relative strengths of the average learned task). For comparison with the binary tempotron tivity of individual neurons to detect structure
responses reflected each clues individual predic- model (12), I also evaluated the test error under a within a networks input activity (27)for instance,
tive strength (Fig. 2F and fig. S5, B and C). Thus, binary decision rule, which only required the neu- through spike-timingdependent synaptic plastic-
learned responses of multi-spike tempotrons with ron to fire at least one spike when the target digit ity (25, 26, 28). Can aggregate-label learning
stochastic feedback resembled recordings in the occurred within its input. With this looser perform- enable populations of neurons to detect structure
lateral intraparietal cortex of rhesus monkeys per- ance measure, the test error of the multi-spike across much larger scales in space and time by
forming a similar task (21). tempotron reduced to 0.0021, outperforming the amplifying correlations between the neurons
binary tempotron by 23%. I next exploited that output spike counts? In contrast, neurons that
Application to speech recognition ~20% of the training sequences in which a tar- respond to random inputs typically generate un-
In machine learning approaches to automatic get digit occurred contained multiple occurrences. correlated numbers of output spikes. I tested
speech recognition, it is easy to count occurrences To expose the role of aggregate-label learning, this hypothesis by modeling a neural processing

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-3


R ES E A RC H | R E S EA R C H A R T I C LE

connections between the processing layers of


several self-supervised networks (materials and
methods, lateral inhibition) ensured that each of
the networks converged to a nonoverlapping re-
gion of feature space (fig. S8, A and B). This
mechanism did not require fine tuning of the
strength of inhibition (fig. S8B) and accommo-
dated different numbers of participating networks
by automatically increasing each networks se-
lectivity when the number of networks grew
(fig. S8C).
Fig. 3. Application of aggregate-label learning to continuous speech recognition. (A) Spectrogram Lateral inhibition shaped the neural activity
of one example utterance of the TIDIGITS database (file 7zzz9z6a.wav of male speaker ka) containing profiles that emerged from the unsupervised learn-
the spoken digit sequence seven, zero, zero, zero, nine, zero, six. Vertical lines show the output ing dynamics by introducing a complex interfer-
spike times of four neurons that were trained with the occurrence counts of the digits seven (dot- ence between feedforward inputs and recurrent
dashed), zero (solid), nine (dotted), and six (dashed). Each of the neurons responds to the ocurrence inhibitory activity. The complexity of these highly
of its target digit by firing a single spike. (B) Histogram of individual test errors over the isolated digits nonlinear interactions makes it difficult (presum-
subset of the database. The histogram comprises error counts of 10 independent simulations for each of ably also for evolutionary processes) to control
the 242 task conditions (11 target digits 22 dialects). Colors depict contributions from individual digits the exact shape of the emerging neural tuning
[starting with zero (darkest blue) followed by oh, numerical values increase with color temperature]. profiles. In addition, this tuning remains depen-
dent on the continued presence of the recurrent
layer as a population of multi-spike tempotrons the individual receptive fields of different neurons inhibition even after the learning has converged.
that were driven by a common sensory periph- on the basis of their common occurrence counts, Last, because direct lateral inhibition operates
ery. Learning in this population was controlled irrespective of their timing. In particular, features on the fast time scale of the processing layer, it
by a feedback signal that was generated inter- can occur asynchronously across the processing requires the tight temporal alignment of all sen-
nally by a single downstream supervisor unit layer. To demonstrate this capability, I implemented sory feature streams. This conflicts with the in-
(Fig. 4A) and signaled a common desired num- a self-supervised network architecture in which herent capability of self-supervised networks to
ber of output spikes to all processing layer each processing layer neuron received individual discover temporal structures that extend over
neurons. To amplify correlations between the input activity (Fig. 4D): For each neuron, sensory the much slower time scale of the supervisory
processing layer responsesto drive neurons to features were represented by individual activity feedback circuit. These three drawbacks can be
respond to structure within their input activity patterns that occurred at independent random resolved by replacing the direct interactions be-
the common feedback signal was based on the times within independent random background tween processing layer neurons with indirect
average number of spikes fired within the pro- activity. Such individual input architectures mimic interactions through a more refined supervisory
cessing layer during the preceding sensory episode. neural integration stages of segregated or even circuit. Specifically, I assume that instead of the
This self-supervision entailed that all processing multimodal sensory input streamssuch as visual, common supervisory signal introduced above, a
neurons that fired more spikes than the popula- auditory, and olfactorywhose activity might not group of supervisory units maps the activity of
tion average implemented a learning step toward be temporally aligned. Self-supervised learning the processing

layer onto an array of supervisory
a reduced spike response. Vise versa, neurons functioned also within this network architecture signals whose components i specify the in-
that fired fewer spikes than their peers aver- (Fig. 4B, inset). Detailed analyses (supplemen- dividual desired spike counts of each processing
age response adjusted their synaptic efficacies tary text S2) showed that self-supervised learning neuron (Fig. 5A). In its simplest form, these labels
toward an increased output spike count. If by is sensitive to the number and the timing of spikes are given by a linear projection,
chance a subset of neurons responded to occur- within a features activity pattern (fig. S6, C and X
Np

rences of the same sensory feature, their elevated E). It can tolerate high levels of input noise (fig.
i S ij aj 1
joint weight within the average population re- S6F) and operate over a wide range of processing
j1
sponse biased the remaining neurons to search layer sizes (fig. S6D). Furthermore, the learning
for a feature that occurred the same number of mechanism can be implemented with various which has a straightforward neural interpretation
times. This learning will come to an end when functional forms of the internal feedback signal, (materials and methods, neuron specific feedback
all processing layer neurons generate an equal including a neural implementation in which the signals in self-supervised networks) and suffices to
number of output spikes (materials and meth- feedback is driven by output spikes of a neural expose the basic principle underlying the proposed
ods, zero labels, correlations in synaptic efficacies supervisor (fig. S6G). I have ensured that self- mechanism. Nevertheless, nonlinearities within
and synaptic learning noise)in particular, when supervised learning was unaffected when synaptic this supervisory pathway, such as the saturation
neurons cease to respond to random background plasticity in the processing layer used the ap- of individual feedback signals, can have important
activity and respond with the same number of proximate, correlation-based learning rule (fig. functional consequences (compare with Fig. 6).
spikes to all sensory features. Indeed, when ap- S3C). Last, self-supervised learning functioned In the above equation, Np is the number of proc-
plied to the embedded feature detection task, pro- analogously when the unknown features were essing layer neurons, and the component S ij of
cessing layer neurons of self-supervised networks not discrete but drawn from continuous feature the Np Np supervisory matrix specifies how
converged (materials and methods, convergence dimensions (fig. S7). much the activity aj of the jth processing neuron
in unsupervised learning tasks) to subsets of the contributes to the supervisory signal i of the ith
recurring features and encoded each of their oc- Neural interactions mediated through processing neuron. The regulation of this cou-
currences by firing the same number of spikes internal feedback signals pling between the supervisory circuit and the proc-
while remaining silent otherwise (Fig. 4, B and C; An important question in unsupervised neural essing layer is assumed to take place over much
population data is provided in fig. S6A). The learning is what mechanisms of interactions can slower (developmental or evolutionary) time scales
ability of aggregate-label learning to overcome ensure the formation of neural maps in which than that of learning in the processing layer.
long feedback delays permits the networks su- different neurons represent different features. One In any unsupervised neural network, the shape
pervisory circuit to integrate the activity of the intuitive and long established mechanism for of the emerging neural activity profiles depends
processing layer over long time windows. Within competitive interactions between neurons is lat- on the networks architecture. In self-supervised
these windows, features are bound together across eral inhibition (27). Indeed, recurrent inhibitory networks, this dependence is particularly simple:

aab4113-4 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

Fig. 4. Self-supervised neural networks. (A) Network scheme with shared simualtions) of shared [(A), dotted lines] and individual [(D), solid lines] input
input architecture. A layer of processing neurons (blue) receives input activity architectures for tasks with 1 (dark blue) and 10 (light blue) features. The open
from a common input layer (gray). These multi-spike tempotrons project to a bar represents tasks in which the independent features of each processing-
single supervisor unit (red), which feeds back a common feedback signal (red layer neuron were not ordered with respect to their spike counts (supplemen-
lines) to each neuron of the processing layer. (B) Learning curves of a shared tary text S2) (C) Example voltage traces of all 10 processing-layer neurons
input network showing the evolution of the processing-layer responses to the (color coded) at the end (500 cycles) of the simulation shown in (B). Although
background activity (top, black line) and to each of the individual features individual voltage traces have remained diverse, all neurons have learned to
[bottom; color code as in (C)]. Full colors denote the mean responses over all respond the dark-blue feature. (Right) The 10 traces within one of the feature
10 processing-layer neurons of the network, and light tones depict responses of occurrences (gray horizontal line). (D) Network scheme as in (A), but with
individual cells. (Inset) The mean convergence times (over 1000 independent individual input architecture.

The learning dynamics in self-supervised net- and B). When exposed to sensory features from a gregate rewards and the integrated neural re-
works reaches a fixed point when all supervisory continuous dimension, self-supervised networks sponses. The existence of such reward prediction
signals match the activities of the recipient pro- with saturating supervisory signals (materials and errors, such as within responses of midbrain dopa-
cessing layer neurons so that all discrepancies methods, continuous feature maps) are capable mine neurons, has long been established (2, 29). I
between feedback signals and neural responses of generating continuous sensory feature maps have shown that a simple binary readout (increase
vanish. This is the case when the supervisory cir- without requiring lateral inhibition or temporal versus decrease) of such error signals (2, 29, 30)
cuit projects the processing layer activities onto alignment of incoming sensory features (materials is sufficient to drive aggregate-label learning. At

themselveswhen, mathematically, a is an and methods, continuous feature maps) (Fig. 6 the core of aggregate-label learning lies the chal-
eigenvector of the supervisory matrix S with and fig. S9C). lenge to credit the right synapses when a neu-
eigenvalue 1. This observation immediately ex- rons discrete number of output spikes requires
plains the behavior of the self-supervised net- Discussion adjustment. The gradient-based multi-spike tem-
works with a single supervisor that corresponds I have demonstrated that aggregate-label learn- potron learning rule establishes a normative solu-
to a uniform supervisory matrix with constant ing enables simple neural architectures to solve tion of this fundamental problem. Thus, it provides
S ij 1=Np . With uniform supervision, the only the temporal credit-assignment problem under- an important frame of reference for experimental
eigenvector of S with nonzero eigenvalue is also lying learning tasks with delayed feedback. This measurements of synaptic plasticity. Its essential
uniform, so that convergence requires all neurons approach remains robust even when input sig- capabilities can be reproduced by an approxima-
to respond with the same number of spikes to nals are corrupted by noise or when the feedback tion (materials and methods, correlation-based
each input pattern. Although uniform supervi- is stochastic and binary. Aggregate-label learning learning) (fig. S3) that is based on temporal cor-
sion offers the simplest neural implementation, outperforms reinforcement learning and can solve relations between presynaptic activity and post-
neural circuits can use multiple supervisor units connected-speech recognition tasks without tem- synaptic depolarization (12, 31, 32). Although
as an efficient mechanism to stabilize more com- porally annotated training data. Although organ- performing well on easier tasks, this simple ap-
plex neural activity profiles across feedforward isms rarely have access to the timing of an proximation becomes unstable for difficult tasks in
sensory processing stages. In fact, with appro- unknown clue, a clues number of occurrences which it requires early stopping or stabilization
priate supervisory connections the learned neu- is often related to the strength or likelihood of for instance, through divisive weight normaliza-
ral responses to a discovered sensory feature can the feedback signal that it predicts. Hence, the tion (fig. S3B). I expect that refinements, such as
be shaped almost arbitrarily (materials and meth- concept applies to a large class of ethological nonlinear eligibilities or the incorporation of
ods, supervisory matrices, neural population tun- learning scenarios. The learning requires neural corrections from elicited postsynaptic spikes (33),
ing for discrete features) (Fig. 5B and fig. S9, A representations of the discrepancy between ag- will further improve the performance and stability

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-5


R ES E A RC H | R E S EA R C H A R T I C LE

of neural implementations of the multi-spike


tempotron. I have focused on the computation-
ally hardest scenario and applied aggregate-label
learning to single-step prediction problems (1),
within which no feedback at all is available until
the end of a trial. Extensions of the concept to
multistep prediction problemswithin which
some feedback can also arrive throughout a trial,
such as in the context of temporal difference
learning (1)will further boost learning perform-
ance. These extensions will need to elaborate how
the neural feedback circuit should segment the
learning in time when consecutive feedback sig-
nals become available. One possibility is that the
circuit inititiates a learning step whenever the
dicrepancy between the available external feed-
back and its internal prediction reaches a critical
value. The neural implementation of aggregate-
label learning can be interpreted as a continuous
(temporal) form of multiple-instance learning [re-
viewed in (34)], establishing an unexplored cross-
link between neurobiological and machine learning.
The concept of aggregate labels gives rise to a
Fig. 5. Feedback-mediated neural interactions. (A) Network scheme (Fig. 4D) of a generalized self- form of unsupervised learning that enables pop-
supervised network in which each processing layer neuron receives an individual feedback signal (red ulations of neurons to discover reoccurring con-
lines) from a dedicated supervisor unit (red circles). (B) Learned neural responses (open circles, y axis) of stellations of input patterns even if they are rare
50 processing layer neurons (x axis) in a single feature task (Nf = 1). Responses where measured after and distributed over spatial and temporal scales
200 cycles of unsupervised learning (materials and methods, neural population tuning for discrete that exceed the receptive fields of single neurons.
features) with two different supervisor matrices. (Top) The supervisory matrix (inset) stabilized a generic This detection of global structure through the
population tuning profile v whose individual components were drawn from a uniform distribution (ma- amplification of correlations is related to the
terials and methods, supervisory matrices).The red line shows a two-parameter (baseline and scale) fit of long established statistical technique of canoni-
v with R2 = 0.9973. (Bottom) The circulant supervisory matrix (inset) stabilized a sinusoidal neural tuning cal correlation analysis (35, 36). Implemented
profile. The red line shows a first-order harmonic fit (baseline, amplitude, and phase) with R2 = 0.9987. within an information maximization approach,
Background intervals were T 2 s long, and mean feature counts were cmean f 25. the underlying principle has previously enabled
abstract neural networks with analog (sigmoi-
dal) units to detect instantaneous spatial struc-
ture across neural populations (36, 37). Through
aggregate-label learning, this principle could be
realized in the time domain and implemented
within a spiking neural network. The resulting
self-supervised learning is consistent with observed
increases in response coherence (within local groups
of neurons) accompanying the emergence of di-
rection selectivity in early cortical development
(38) and captures the processs sensitivity to stim-
ulus statistics (39). Self-supervised networks are
robust to extensive temporal misalignments be-
tween feature occurrences across their process-
ing layers. Overcoming the inability of direct
lateral inhibition to introduce competitive inter-
actions between neurons that are processing
asynchronous features, I have demonstrated that
such interactions can be straightforwardly medi-
ated through internal supervisory signals, offer-
ing a simple yet powerful mechanism with which
Fig. 6. Emergence of continuous feature maps. (A and B) Learned neural responses when the gen- to mold the emerging neural maps. After con-
eralized self-supervised network from Fig. 5B, bottom, was driven with features from a continuous di- vergence, the supervisory circuit can disengage
mension (materials and methods, continuous feature dimensions). The emergence of these continuous without affecting the learned neural responses.
maps required the supervisory signals to be saturating (materials and methods, continuous feature maps) By detecting common numbers of occurrence,
to enforce an even distribution of neural responses across the processing layer. (A) Example output spike self-supervised networks realize a general prin-
raster of a processing layer (after learning) when driven by a probe trial containing 21 features (colored ciple for invariance learning that encompasses
rectangles) embedded within background activity at evenly spaced times (x axis).The features evenly span (as a special case) the recently proposed principle
the feature dimension (color bar). (B) Mean neural responses (gray scale) of all 50 processing-layer of temporal contiguity (40, 41). Nevertheless,
neurons (y axis) to feature from the continuous dimension (x axis) for a single network (left) and the mean constrains on the time evolution of the internal
map (right) over 1000 independent simulations. Phases and directions of individual maps (fig. S9C) were supervisory signals could be considered to fur-
aligned before averaging. Background intervals were T 1 s long, and mean feature counts were ther extend the window of temporal contiguity
cmean
f 2. beyond the time frame of the supervisory signaling.

aab4113-6 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

The mechanisms through which self-supervision sampling (with replacement) from a uniform of the feature increased the reward probability.
segments learning in time, such as through thresh- distribution over {1, 2, ..., N}. On trials that contain only a single featureonly
olding the total number of processing layer spikes, one of the ci = 1 while the others are zerothe
will be important constituents of its specific neu- Correlated feature occurrences reward probability reduces to pR = pi. In fig. S5A,
ral implementations. The structural and compu- In Fig. 2C and fig. S2D, homogeneous pairwise I test the effect of false rewards. These were gen-
tational simplicity of self-supervised networks correlations 0 r 1 between the numbers of erated by additional features with nonzero reward
permits a number of neural implementations at occurrences Ci of all features i = 1, ..., Nf were contingencies, however, which were invisible
different spatial scales. I have shown that the introduced through a common source Y. Specif- to the neuron because their occurrences did not
internal feedback signal can be realized by the ically, on each trial each features number of occur- contribute to the neurons input spike train. In
spiking activity of a simple model neuron, dem- rences was realized through the sum of two Poisson learning tasks with probabilistic reward contin-
onstrating that modulatory feedback of neural random variables, Ci = Y + Xi. With the mean of the gencies, synaptic potentiation was invoked after
origin could be provided to local as well as dis- common Y set to rcmean f and the means of the each rewarded trial irrespective of the neurons
tributed (multimodal) processing layers. Alterna- individual Xi set to 1 rcmean
f , the resulting Ci response. Specifically, the neuron performed a
tively, I speculate that a local supervisory signal were Poisson distributed with mean cmean f and multi-spike tempotron learning step toward in-
could be provided by astrocytes that have been pairwise correlations Corr[Ci, Cj] = r for i j. The creasing the current response by one spike. In
implicated in modulating synaptic plasticity and mutual information (MI) between the labeled num- contrast, depression was dependent on the neurons

also monitor the activity levels of neighboring bers of occurrences C that is, C C1 ; :::; CNf response, and a learning step toward decreasing
neural circuits (42). Last, the computational sim- and the corresponding aggregate-label and the neurons response by one spike was only invoked
ilarity between a two-layer neural network and the identity of the clue is given by the reduction in when the neuron fired at least one spike during a
dendritic nonlinearities that converge onto a soma the clues uncertainty, MI Hclue HcluejC (ma- nonrewarded trial (false alarm). I used divisive
(43) invites the intriguing thought that the prin- terials and methods: mutual information be- normalization to bound the resulting learning
ciple of self-supervision could be used by a neu- tween feedback signal and clue identity). dynamics: Whenever, after a learning step, the

rons somatic compartment to guide its dendritic Euclidean norm jw j of the synaptic efficacies
branches in a search for structure within the Reinforcement learning comparison task surpassed the fivefold multiple of its initial value

cells presynaptic activity (44). To compare aggregate-label learning with the jw ini j, all efficacies were multiplied by 5jw ini j=jw j.
performance of a recent implementation of rein-
Materials and methods forcement learning (4), I studied a feature learn- Continuous feature dimensions
Embedded features ing task in which no background activity was To mimic sensory features from a continuous
High-dimensional spike patterns of N = 500 present and the desired response to each feature dimension I modeled a sensory encoding stage
sensory afferents were modeled as Poisson point was binary: at least one spike versus no spike. In that mapped each feature parameter a 0; 2p to
processes with an average firing rate of rf = 5 Hz. addition, the duration Tf = 500 ms of the Nf = 10 a feature activity pattern of Tf = 50 ms duration.
Each multineural input spike pattern consisted Poissonian feature activity patterns with mean This deterministic map was constructed by spe-
of random background activity of duration T , firing rates of rf = 6 Hz was much longer. Each cifying how the feature spike trains of each indi-
which was interleaved by a random number of trial consisted of a random sequence of all fea- vidual afferentthe number and timing of spikes
feature activity patterns. Feature activity pat- tures. To mimic the feedback delay of 1350 ms within the 50 ms feature intervaldepended on
terns were Tf = 50 ms long and remained fixed (4), each input activity pattern was partitioned the feature parameter a. For each afferent, this
across all trials of a given learning task. In tasks into 10 segments of 1350 ms duration whose start- mapping was defined through interpolation be-
that mimicked learning of discrete sensory fea- ing times were locked to the beginning of each of tween a set of spike train templates. The number
tures, Nf independent realizations of feature the 10 features. Half of the features were targets of templates assigned to each afferent was a
activity patterns were generated from the back- and the neuron required to respond with at least random number m = 1 + G with G drawn from a
ground firing statistics. For each trial, the num- one spike within the 500 ms feature interval, Poisson distribution with mean g. Each of these
ber of each features occurrences was drawn whereas the other half were distractors, for which m templates was randomly assigned to a fixed
from a Poisson distribution with mean cmean f . the neuron had to remain silent. After a miss, the position ai i 1; :::; m within the feature di-
Thus, the average total number of features neuron was trained to increase its number of mension, drawn from a uniform distribution
within each trial was given by Nf cmean
f , and the output spikes over the preceding 1350-ms input over [0, 2p]. Each template i stored a spike count
mean trial duration was activity pattern by one spike, and after a false ci drawn from the background statistics and a
alarm, the neuron was trained to decrease its num- vector of spike times drawn from a uniform dis-
T T Nf cmean Tf 2
f
ber of output spikes. The convergence criterion tribution over [0, Tf]. The number of spike times
For each feature occurrence of a given trial, an required 90% correct feature responses (4). stored in each template was equal to the maxi-
occurrence time was drawn from a uniform mal spike count of the afferent over all of its m
distribution over 0; T . After sorting all feature Probabilistic reward contingencies templates. For any feature parameter a 0; 2p,
occurrence times, the input activity pattern was In Fig. 2, E and F, and fig. S5 I study an adapted a vector of spike times was given by linear inter-
constructed by successively inserting the feature version of the weather prediction task (21), in polation between the two spike time vectors of
activity patterns into the background. Overlap which the occurrences of sensory features incre- the adjacent templates k and k + 1 with ak <
was avoided by delaying all background spikes ment the probability pR that a binary reward is a < ak1 . For afferents with only one spike train
after a feature occurrence and all subsequent fea- delivered at the end of a trial. In each trial, the template, m = 1, the spike train remained un-
ture occurrence times by Tf whenever a feature reward probability was given by changed for all a and did not carry any infor-
activity pattern was inserted. "X Nf # mation about the feature parameter. If the two
adjacent templates had the same spike counts
Fixed feature activity spike counts pR 1 exp ci log1 pi 3
if ck = ck+1a was mapped to a spike train of ck
i1
In fig. S6, D to F, I studied the convergence times spikes that occurred at the first ck components of
of self-supervised networks with feature activity where the sum runs over all existing sensory fea- the interpolated spike time vector. If the spike
patterns whose total spike counts were fixed tures and ci 0 denotes their independent Pois- counts were different, Dck = |ck ck+1| > 0, an
instead of drawn from Poisson count statistics. sonian occurrence counts (with mean cmean
f 0:1). additional set of transition values aj j 1; :::; Dck
Specifically, for each feature activity pattern the The reward contingency of each feature i was from a uniform distribution over ak ; ak1  was
predetermined number of spikes was randomly controlled by the parameter pi 0; 1, which de- stored with the templates. These transition values
distributed across all N afferents by randomly termined by how much each individual occurrence determined the positions of incremental changes

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-7


R ES E A RC H | R E S EA R C H A R T I C LE

in the number of spikes as a transitioned through matrix. During training, feature parameters were 0 , which maps each threshold value to
ak < a1 < a2 < ::: < aDck < ak1 . The appear- rounded to a discrete grid that matched the the number of elicited output spikes, STS
ance and disappearance of spikes along the column resolution of the image. (fig. S1C). The STS can be characterized by the de-
feature dimension was ordered according to a creasing sequence of critical threshold values *k :
last-in-first-out principle, so that the most Input noise
recently added spike was always the first to dis- I used two types of input noise: First, random *k supf : STS kg; k 6
appear when the spike count began to decrease. deletion of input spikes that mimicked synaptic
The only internal parameter of the above encod- transmission failures (Fig. 2, E and F, and figs. The critical threshold *k denotes the threshold
ing stage is g, which controls the density of inde- S4F, S5, and S6F), and second, presynaptic spike value (supremum) at which the number of out-
pendent spike train templates along the feature time jitter (Fig. 3 and fig. S4E). Spike deletion put spikes jumps from k 1 to k. For thresholds
dimension and thus determines how quickly the was parameterized through a probability pdel, between two critical values, the STS is constant,
activity patterns of two feature parameters become which determined for each spike of an input ac- * < < *k k. Because a neuron re-
STSk1
uncorrelated as their distance along the dimension tivity pattern (including background spikes) the mains silent if its firing threshold lies above the
grows (fig. S4A). Except in fig. S4A where I varied independent probability that it was not transmitted maximum postsynaptic voltage Vmax, the STS is
g, I used g = 5 throughout the present work. to the postsynaptic compartment. Spike jitter zero for thresholds above Vmax, STS( > Vmax) =
was realized by adding independent Gaussian 0. The first output spike is generated when the
Encoding schemes noise with zero mean and standard deviation s threshold equals to Vmax, *k Vmax . In general,
In fig. S4, D to F, I compare the learning per- to each spike time of the input activity. Neural each *k corresponds to a voltage value described
formance for different encoding stages. In addi- responses were measured with the same noise by Eq. 5 and hence is a function of the neurons

tion to the above encoding in which both spike settings that were used during training. synaptic efficacies w and differentiable with re-
counts and spike times carry information about spect to them. When I use the STS for the defi-
the feature parameter a, I implemented three Neuron model nition of gradient-based synaptic learning rules, I
additional alternatives: In a count-based encod- The multi-spike tempotron was implemented as ignore the possible existence of singular points in
ing, the feature parameter a only determined the current-based leaky integrate-and-fire neuron mod- the space of synaptic efficacies where a required
deterministic number of spikes that each affer- el with reset. Specifically, the neurons postsynaptic *k and/or its derivatives might not existfor ex-
ent contributed to the feature activity pattern, membrane potential V(t) was given by integration ample, when two adjacent critical thresholds co-
but their timing was drawn independently for of exponentially decaying currents from N = 500 alesce and the output spike count of a neuron
each feature realization. In a variation of this synaptic afferents, yielding jumps by two spikes at a single critical value. I
scheme, these deterministic spike counts were have not encountered such pathologies in any of
X
N X  j
interpreted as the mean values of random Pois- V t Vrest wi K tti the simulations underlying the present work.
son spike count variables that were redrawn for i1 tij <t
each realization of a feature activity pattern and j
! Multi-spike tempotron
X ttspike
the corresponding number of spikes placed ran- exp 5 The goal of multi-spike tempotron learning is to
domly within the feature interval. In contrast, j
tm modify a neurons synaptic efficacies so that it
tspike <t
the spike-timingbased encoding stage was con- fires a desired number of output spikes in re-
strained so that the spike counts of all spike train Here, tij denote the arrival times of individual sponse to a given input spike pattern. A simpli-
j
templates of a given afferent remained constant presynaptic spikes of the ith afferent, and tspike fied version of this problem has been solved with
along the entire feature dimension. Here, only are the times of output spikes elicited in the the binary tempotron (12), which is a supervised
the relative timing of spikes within the feature postsynaptic compartment. Each input spike at synaptic learning rule that can train a neuron to
activity pattern, but not their numbers, carried time tij contributes a postsynaptic potential respond to a given input spike pattern by either
information about the feature parameter. (PSP), whose shape is given by the kernel firing at least one spike or remaining silent.
h    i
t tj t tj Because of the original tempotrons binary deci-
Continuous tuning curves K t tij Vnorm exp tm i exp ts i and
sion, the learning could be formulated as a
Random periodic target tuning curves of the form whose peak amplitude is determined by the gradient-based learning rule operating on the max-
afferents individual synaptic efficacy wi. Nor-
X
O
malization of K to unit amplitude resulted from
imum postsynaptic potential: increasing it toward
the cells firing threshold when the neuron should
fO a A0 b Bk coska fk 4
setting Vnorm = h[h/(h 1)]/(h 1) with h = tm/ts. fire at least one spike, and lowering it below the
k1
The temporal shape of PSPs is governed by the firing threshold when the neuron should remain
were generated by drawing the amplitudes Bk from integration time constant of the postsynaptic silent. However, this reliance on the maximal
a uniform distribution over [0, 1] and the phases membrane, tm = 20 ms, and the decay time con- postsynaptic potential as its objective function
fk from a uniform distribution over [0, 2p]. The stant of synaptic currents, ts = 5 ms. PSPs are prevents the binary tempotron learning rule from
parameter O determined the order of the high- causal; K t tij is defined to vanish for t < tij . controlling the number of spikes beyond one, when
est nonzero Fourier coefficient in the above ex- Whenever V(t) crosses the firing threshold , the the nonlinearity of spike generation renders the
pansion. For each realization, the baseline A0 neuron emits an output spike, and the voltage maximum postsynaptic potential insensitive
and the global scale b were set so that the min- resets to Vrest = 0 by the last term in Eq. 5. I used to the number of elicited spikes. To overcome this
imum and maximum of fO were fOmin 0 and = 1 for the fixed biological firing threshold of problem, the multi-spike tempotron implements
fOmax 10, respectively. In fig. S4C, I used the the neuron. Numerical simulations of the neuron a gradient-based learning rule whose objective
gray-level values of Leonardo da Vincis Mona model were event-driven and based on exact is to shape the neurons STS by modifying the
Lisa as target tuning curves. The image data was solutions of Eq. 5. In simulations underlying positions of the critical thresholds *k (Eq. 6).
based on the file Mona_Lisa,_by_Leonardo_ table S2, I also tested readout neurons with N = Although many objective functions can be defined
da_Vinci,_from_C2RMF_retouched.jpg downloaded 80 and N = 20 synaptic afferents. on the basis of the STS, I focus here on the sim-
from http://commons.wikimedia.org/wiki in a plest form in which each learning step operates
resolution of 687 by 1024 pixels. A 100- by 100- Spike-threshold-surface along the gradient of only a single *k with k ;
pixel region was extracted, converted to gray- To formulate a continuous objective function for a that is, the neurons synaptic efficacies are changed

scale, and rescaled so that the minimum value neurons discrete number of spikes, I defined, for by D wrw *k after each error trial.
was 0 and the maximum 10. Each target tuning a given input spike pattern, the spike-threshold- To generate the desired number of output
function corresponded to a row of the resulting surface (STS) of a neuron as the function STS : spikes d in response to a given input spike pattern,

aab4113-8 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

a neurons STS must be aligned with its firing spikes, and negative updates applied to trials tum heuristic was not engaged in self-supervised
threshold so that d1 * < < d* . If this con- in which the neuron fired too many. networks, in which I set m = 0.
dition is violated on a given trialif the neuron
misclassifies its input spike patternan abso- Noninteger labels in supervised Initialization
lute learning rule would decrease d1 * through learning tasks Before learning, the initial synaptic efficacies

D ww d1* whenever the neuron fired more In tasks involving supervised learning of contin- w ini of all multi-spike tempotrons were initial-
spikes than required. If the neuron had fired less uous tuning curves (fig. S4), aggregate-labels were ized so that the neurons fired with ~5 Hz when
spikes than desired, absolute learning would in- not restricted to integer values (supplementary driven by the 5-Hz Poisson background activity.

crease d* through D ww d* . Such absolute materials, equation 41) and could not be directly Specifically, after efficacies were drawn from a
learning (fig. S2C) would require the supervi- interpreted as the desired number of output spikes Gaussian distribution with zero mean and a stan-
sory signal to provide the exact value of desired d. Although technically the relative learning rule dard deviation of 0.01, neurons were trained
output spike count. In contrast, a simpler, biolog- defined in Eq. 7 does not require d to be an integer, with 1-s-long trials of background activity with
ically more plausible, and from a incremental learn- I probabilisticly mapped noninteger labels to integer random Poisson labels with a mean of 5. During
ing perspective perhaps also more intuitive learning values d, being either the nearest lower L or higher L+ this initialization, the momentum heuristic was
rule is based on a binary feedback signal that only integer adjacent to . Specifically, on each trial was disengaged (mini = 0), and the learning rate was
specifies whether the neuron should increase mapped to d = L with probability p L and to set to lini = 103. Initialization proceeded over
or decrease its current output spike count o: d = L+ with probability 1 p L . As a result, independent blocks of 100 trials and terminated
( the average integer label hdi pL 1 pL when the mean firing rate of the neuron over the
lw *o if o > d L L L L matched the nonin- last block exceeded 5 Hz. Subsequently, the pa-
Dw 7
lw *o1 if o < d teger aggregate. rameters l and m were set to their simulation
values, and previous synaptic changes (materials
where the learning rate l > 0 controls the step Learning rate and annealing and methods, momentum heuristic) were ini-

size of individual updates. The present work is In simulations of the multi-spike tempotron, I tialized to zero (D wprevious 0). The above ini-
based on this relative multi-spike tempotron used a learning rate of l = 105 (Eq. 7) in all su- tialization was important for self-supervised
learning rule (Eq. 7) (materials and methods, *- pervised learning tasks with discrete target features. networks to avoid the trivial fixed point at which
gradient). This value resulted in near optimal convergence all processing layer neurons remain silent. When
times in the single-clue detection tasks, but the self-supervised networks were studied with noisy
Correlation-based learning minimum was shallow (fig. S2C, inset). In su- input spike trains (fig. S6F), the spike deletion prob-
In correlation-based learning (12), the synaptic pervised learning tasks of continuous tuning curves ability pdel was also applied to the Poisson back-
credit assignment of the multi-spike tempotron (fig. S4), learning convergence was improved by the ground activity used during the initialization, but
learning rule is approximated on the basis of the use of a simple annealing schedule: If the fraction the target spike rate remained at 5 Hz. For super-
correlation between presynaptic activity and post- of error trials within the past 2500 trials failed to vised learning tasks, initialization was not crucial
synaptic voltage. Given the ith synapses pre- decrease relative to the preceding 2500 trial in- and had only appreciable effects on the conver-
synaptic spike train with spike times tij , its eligibility terval, the learning rate l was reduced by a mul-
X gence times when convergence was fast (fig. S2C).
tiplicative factor of 0.9. Correspondingly, a higher
Ei V ji was defined by the sum over all
tij initial learning rate of l = 104 was used in these Neural responses
individual correlations V ji t j dtV tK t tij , simulations. In all self-supervised network simu- The mean response Rmean y of a neuron to a feature
i
where V(t) is the postsynaptic voltage given by lations with a single supervisor unit, the learning activity pattern y during and after learning was
Eq. 5 and K is the PSP kernel. In neurons, eleg- rate was l = 5 104 and remained fixed (Fig. 4 measured through dedicated batches of probe trials
ibilities of this or similar form could be realized and figs. S6, S7, and S8). In simulations of self- during which the learning rule was not engaged.
on the basis of intracellular calcium signals, supervised networks with generalized supervisory The index y refers either to a particular feature
which are sensitive to the coincidence of pre- architectures, I lowered the learning rate to l = 1 f f1; :::; Nf g from the set of discrete features or
and postsynaptic activity through the voltage 105 during the critical and consolidation phases to the feature activity pattern of a particular fea-
dependence of synaptic N-methyl-D-aspartate of the discrete feature learning task (Fig. 5 and fig. ture parameter a 0; 2p when feature activity
(NMDA) receptors. It is well established that S9, A and B) and used l = 5 105 when studying patterns belonged to a continuous manifold. Each
the induction of long-term synaptic changes the formation of continuous maps (Fig. 6 and fig. probe trial consisted of a Tprobe = 2 s segment of
reqires these calcium signals to reach specific S9C). For correlation-based learning (Eq. 8), I used a background activity within which a central gap of
plasticity induction thresholds (31, 45, 46). Corre- constant learning rate of lV 1  105 for super- 50 ms was inserted. For each probe trial, the out-
spondingly, I defined correlation-based learn- vised learning and lV 5  104 in self-supervised put spike count of a neuron was measured twice:
ing so that all synapses whose eligibility reached networks. In the simulations summarized in once giving s , with the central gap being empty,
a plasticity induction threshold y on a given table S2, I used the optimized learning rates and once giving sy, with the feature activity pat-
error trial became elegible to undergo poten- given there. tern placed within the gap. The response Rmean y
tiation or depression. To ensure sufficient spe- was defined as the mean difference in the number
cificity of this learning, the induction threshold Momentum heuristic of output spikes elicited in these two conditions,
should be adjusted so that on average, only a Following previous studies (12, 23), I used a Rmean
y hsy s i, with the average taken over 1000
limited fraction of synapses undergo change. momentum heuristic (47) to accelerate conver- independent probe trials. The embedding of the
In the present work, I have simplified this gence in all supervised learning tasks. Specifi- feature activity pattern within background activity
adjustment of y by automatically selecting the cally, synaptic updates combined the learning ensured that the neurons responses were measured

10% most eligible synapses on each error trial. rule correction D w of the current error trial (Eq. 7) in the same conditions as present during learning.
Specifically, all synapses whose V i exceeded with a fraction of the previous synaptic change The averaging captured the stochasticity of the
previous
the 9th, (the highest) decile D9 were updated: Dw , so that D wcurrent D w mD wprevious . If neurons responses due to the random background
on a given error trial the individual learning rule activity that surrounded the feature activity pattern

TlV if Ei > D9 correction of an efficacy i f1; :::; N g was zero or additional input noise, whose statistics were kept
DwTi 8
0 if Ei D9 (Dwi = 0), neither the synapse nor its previous identical to the training conditions of a given task.
change were updated. I used m = 0.99 for the Responses to background activity Rmean were de-
Positive updates applied to trials in which the momentum parameter, implementing a decaying fined as the mean spike rate elicited in the absence
neuron fired less than the desired number of trace of former synaptic changes (12). The momen- of the feature activity pattern, Rmean
hs i=Tprobe .

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-9


R ES E A RC H | R E S EA R C H A R T I C LE

Response variability (table S1) was quantified In addition, for all feature activity patterns f, the spike tempotron learning rule that was used in
through the corresponding standard deviations deviations between the individual responses of the supervised tasks. However, in contrast to re-
y STDsy s  and R STDs =Tprobe .
Rstd std
each processing layer neuron and the population ceiving an external feedback signal, self-supervised
mean had to lie below 5%, ji Rmean f Rf j < networks generated their training labels inter-
Neural tuning curves 0:05Rintf , when Rf 1 or j i Rf Rf j < 0:05
int mean
nally as a function of the spiking activity with-
The tuning of neurons that were exposed to fea- when Rint f 0. in their processing layers. After each learning
tures from a continuous dimension (Fig. 6 and When comparing convergence times of net- trial, the components of the activity vector
N
figs. S4, S7, S8, and S9C) was characterized by works with varying processing-layer sizes and a 0 p consisted of the Np individual spike
measuring the neural responses to 100 probe fea- also different levels of input noise, the above counts ai(i = 1, ..., Np) that were elicited in the
tures that were linearly spaced along the con- convergence criterion was not suitable. Instead, processing layer throughout a given learning
tinuous feature dimension. The mean responses successful detection of the single feature in the trial. In self-supervised networks with a single
shown in fig. S7A are the average of these re- tasks of fig. S6F was based on a population cri- supervisor (Fig. 4 and figs. S3C, S6, S7, and S8),

sponses over the networks processing layer. Tuning terion that required the total number of spikes eli- each activity vector a was mapped to a common
width in figs. S7, B and C, and S8C were defined Xin response to the
cited within the processing layer label through a scalar feedback function
Np N
as the fraction of the continuous feature dimen- L: 0 p 0 with a L a . I X defined L to
feature activity pattern S f i sf i s to Np
sion over which the processing layers average i1 be the mean spike count, La i
ai =Np ,
response exceeded 0.5 spikes per feature. X
exceed the population response to the 2-s back-
Np
except in fig. S6G, where I also tested self-
supervised learning with nonlinear feedback
Convergence in supervised tasks ground activity interval S i s of the
i1 functions that realized the mode and the median
Convergence times for the supervised single- probe trials. Specifically, I required that over of the processing-layer spike counts. When L
feature learning tasks in Fig. 2, A to C and E, and 1000 independent probe trials, the minimum implemented the mean spike count, noninteger
fig. S2C were defined as the number of learning of S f was greater than the maximum of S . The labels were rounded to the nearest integers. In fig.
cycles required until neural responses converged comparison of the above two convergence mea- S6G, I also tested the learning performance under
to firing one spike in response to the clue and sures in the absence of noise is shown in fig. S6D. a neural implementation of the scalar feedback
remaining silent in response to distractors and Although both measures give similar values for function L, defined as the normalized spike count
background activity. Specifically, neural responses small populations, the strictness of the second, of a supervisor neuron. Specifically, the same
were evaluated after each cycle of 100 indepen- population-based convergence criterion does not integrate-and-fire model that was used for the
dent learning trials. Convergence required all increase with growing network size, so that con- multi-spike tempotrons of the processing layer,
distractor feature and background responses to vergence times do not grow. but with fast synaptic time constants of ts = 0.5 ms,
satisfy Rmeandistractor < 0:01 and R
mean
< 0:01=Tf , In fig. S7, B, C, and E, I report the tuning sta- received excitatory projections from all processing-
respectively. In addition, in deterministic feedback tistics of self-supervised networks that were ex- layer neurons. These connections had fixed strong
tasks (Fig. 2, A to C, and fig. S2C) convergence posed to feature activity patterns from a continuous individual efficacies of w = 1.01 so that in essence,
required the mean discrepancy between the clue feature dimension. For continuous dimensions, every processing-layer spike caused a spike in the
responses and the desired single spike to fall below the computational costs of the above convergence supervisor neuron. As a result, when normalized
hjstarget s 1ji < 0:01. In probabilistic feedback criteria that required an ongoing evaluation of by the number of processing-layer neurons Np,
tasks (Fig. 2E), convergence required the mean the processing layers responses were not prac- its spike count approximated the mean spike
clue > 1. Mean
clue response to exceed one spike, Rmean tical. As an efficient albeit indirect alternative, I count within the processing layer.
convergence times report the number of cycles defined convergence on the basis of spike count
after which the criterion was first satisfied, aver- coherence (SCC), which is the agreement between Neuron-specific feedback signals in
aged over 1000 independent simulations. the spike counts of all processing layer neurons self-supervised networks
during successive learning cycles of 100 trials (fig. I generalized the concept of self-supervised neural
Convergence in unsupervised tasks S7D, bottom). Specifically, convergence required networks by replacing the scalar feedback function
N N
Depending on the learning task, I used different that throughout a cycle, an average of 99% of the L by a vector valued projection L : 0 p 0 p with

definitions of convergence for self-supervised net- processing-layer neurons fired in agreement with a La , whose components specify an in-
works. In Fig. 4B, inset, and fig. S6, C, E, and G, I the mode value of the processing-layer spike counts dividual label for each processing-layer neuron.
studied the convergence times of networks with a that the peak in the processing layers spike Having in mind a neural implementation of
fixed number of Np = 10 processing layer neurons count histogram would on average contain over this feedback function through a back-projecting

that were engaged in discrete feature detection 99% of the individual neural responses. The tuning layer of Ns supervisor units, I considered

L to

tasks without input noise. In these simulations, I of multiple self-supervised networks with recurrent be a linear projection of the form La Sa
used a direct convergence criterion that required inhibition (fig. S8) was evaluated after a fixed (Eq. 1). In addition, all labels were rounded to
convergence of all individual processing-layer neu- simulation length of 1000 cycles. the nearest non-negative integers. In the context
rons. As in the supervised tasks, learning proceeded of a neural implementation, the Np Np su-
in cycles of 100 learning trials that were interleaved Temporal alignment of sensory features pervisor matrix S can be thought of as the
by measurements of the neural responses, giving To highlight the capability of self-supervised net- product of two sets of connections, S BF : A
the individual mean feature i R mean
f f f1; :::; Nf g works to identify structure within their input Ns Np matrix F denoting feed-forward con-
and background i R mean responses of each pro- activity across long temporal distances, I did not nections of neural activity from the processing
cessing layer neuron i f1; :::; Np g (materials and align the random occurrence times of feature to the supervisor layer, and a Np Ns matrix B
methods, neural responses). On the basis of these, activity patterns in simulations of individual in- denoting feedback connections of supervisory
I defined the population feature responses Rf as put networks. However, as motivated in the text, signals from the supervisor units to the process-
X responses across the processing
the average feature
Np
I did enforce such alignment in simulations of ing neurons. In this interpretation, the number
networks with recurrent lateral inhibition be- of supervisory units required for the implemen-
layer Rf i Rf
mean
=Np and its rounded
i1 tween multiple self-supervised groups (fig. S8). tation of a particular projection S is given by
integer value as Rf roundRf . Convergence
int
its rank.
required the background responses of all individual Scalar feedback signals in
processing neurons to fall below i R mean < 0:05 self-supervised networks Zero labels
Hz, while at least one of the rounded population In self-supervised networks, all processing-layer The learning dynamics of self-supervised net-
feature responses Rint
f had to be greater than zero. neurons implemented the same relative multi- works has a trivial fixed point at which all

aab4113-10 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E


processing-layer neurons remain silent (a fore converging onto a sensory feature was lim- was neutral (lhom = 1 and linhom = 1), and the
0) for all input spike patterns. To prevent con- ited for a broad range of task parameters that I processing layer ceased responding to background
vergence to this state, I enforced that any given explored in this study, the problem could be alle- activity. In the following critical period, which
processing-layer neuron would refrain from learn- viated by introducing an additive Gaussian learn- lasted only a single cycle, the supervisory signal-
ing on an individual trial when its feedback ing noise at each synaptic update. Specifically, ing amplified the desired mode of the processing
was zero. after any learning step invoked in a processing- layer responses by operating with linhom = 4.
layer neuron, all synaptic efficacies received ad- During a subsequent consolidation phase, the
Stability of self-supervised ditive noise, so that supervisory circuit returned to linhom = 1, and
learning dynamics the responses stabilized. Fit parameters between
w w x 9
When feedback signals were implemented as the learned processing layer feature responses
the rounded mean of the processing-layer ac- with individual xi drawn independently from a i Rf
mean
i 1; :::; Np and the inhomogeneous
tivity, the resulting self-supervised learning dy- Gaussian distribution with zero mean and stan- mode are shown in fig. S9B. Specifically, each

namics was susceptible to a runaway instability dard deviation sw = 103. With such synaptic learn- random tuning profile v (fig. S9B, top) was fitted
in which all processing-layer responses continued ing noise, which also highlights the robustness (least squares) to the learned response profile by
to grow in response to background activity. This of the present learning scheme, neural responses optimizing a baseline b and an amplitude a
was in contrast to the stable learning dynamics remained diverse even over extended learning X Np h
parameter, minimizing Rmean b
i1 i f
that emerged under the nonlinear scalar feed- periods. For comparability, the synaptic update i2
back functions that implemented the mode or noise was also applied to individual input archi- avi . Because of the two first-order harmonic
median because of the well-known robustness of tectures in Fig. 4 and figs. S3C, S6 and S7, al- eigenvectors of the circulant S, harmonic fits
these measures to outliers. Although this in- though by construction these architectures could (fig. S9B, bottom) required an additional phase
stability could be effectively countered by a small not develop correlations between the efficacies of parameter f so that the corresponding fits
bias in the rounding operation that mapped the their processing-layer neurons (fig. S6B). minimized the discrepancies between i Rmean f
mean processing-layer spike counts to integer- and b + acos[2p(i 1)/Np + f].
valued labels, I followed an alternative approach: Supervisory matrices
Whenever a neurons output exceeded its label, I studied supervisor-mediated interactions be- Continuous feature maps
the multi-spike tempotron learning step was fol- tween processing-layer neurons with two types The emergence of continuous maps required
lowed by an additional attenuation of all weights of supervisory matrices S that stabilized differ- longer periods of amplifying supervisory sig-
by a factor of 0.995. In addition to stabilizing ent tuning profiles across the processing-layer naling. Specifically, I used the circulant supervi-
the learning dynamics, this attenuation step population. In the first type, S corresponded to sory matrix (materials and methods, supervisory
substantially accelerated the convergence of the the sum of two matrices S S hom linhom S v , matrices) with a strong amplification of the first
processing-layer neurons to stop firing in response each of which contributed one eigenvector with harmonic mode, whose eigenvalue was set to
to background activity. It was implemented in all nonzero eigenvalue, lhom and linhom, respective- linhom = 20. Task parameters were as in lateral
reported simulations of self-supervised networks, ly, to S. The constant matrix S hom with ele- inhibition case, T 1 s and cmeanf 2. A sat-
except in the critical and consolidation phases of ments S hom ij 1=Np was normalized by the uration of each neurons individual feedback
Figs. 5B and 6 and fig. S9. processing-layer size Np so that the eigenvalue signal at D = 2 spikes per trial resulted in the
of the homogeneous eigenvector became one, distribution of neural activity across the pro-
Lateral inhibition lhom = 1. This component of the supervisory cir- cessing layer. Specifically, each neurons indi-
Lateral inhibition between multiple self-supervised cuit ensured that the learning dynamics pre- vidual label (Eq. 1) was mapped to by the
networks was introduced by recurrent inhibitory served the mean processing layer activity. In saturating nonlinearity
T
connections between all processing neurons that contrast, the second contribution S v v v with h i
did not belong to the same self-supervised net-

eigenvector v is used to stabilize an arbitrary D 1 exp=D 10

work. Specifically, each multi-spike tempotron in activity profile v with zero mean. Specifically, in
these networks received inhibitory afferents, with the simulations of Fig. 5B, top, and fig. S9, A, In addition, this saturation of neural feedback

a fixed efficacy wI < 0, that were driven by the top, and B, top, the components of v were drawn signals prevented the unbounded growth of
output spikes of all processing-layer neurons of randomly from a uniform distribution, central- neural responses within the processing layer
the other networks. ized by subtraction of their mean and scaled so due to the strongly amplifying harmonic mode.

that each v had unit norm. In the second type of Results in Fig. 6 and fig. S9C were obtained by
Correlations in synaptic efficacies and supervisory matrices (Figs. 5B, bottom, and 6 and running the above self-supervised learning
synaptic learning noise figs. S9, A, bottom, B, bottom, and C, S was con- dynamics for 1000 cycles. For each simulation,
In biological implementations, neurons in the structed as a circulant matrix whose only nonzero the learned neural response map (Fig. 6B, left)
processing layer would neither be identical in eigenvalues lhom and linhom corresponded to the was characterized by the dependence of the
terms of their morphologies or biophysical pro- homogeneous (one eigenvector) and the first- center response (across the processing layer)
perties, nor would they receive exactly identical order harmonic modes (two eigenvectors), respec- on the continuous stimulus parameter. Specif-
input spike trains. However, as a worst-case tively. Specifically, I set the first row of S to S 1;j ically, for each stimulus parameter aj = 2pj/100
scenario I also studied this limiting case in which lhom 2linhom cos2pj=Np =Np where j = 1, ..., with j = 0, ..., 99 of the learned response map,
the self-supervised learning dynamics of the Np, which fixed the circulant matrix S. the center of the processing layers response
shared input architectures had a pathological was measured by the weighted circular mean
Neural population tuning for "
fixed point at which all neurons converged to  
identical synaptic efficacies. In this case, the discrete features ^ 1 X Np
2pk
maj atan2 k Raj
mean
sin ;
learning would stop even if the processing-layer To demonstrate that the steady-states of the Np k1 Np
neurons had failed to lock onto a sensory fea- self-supervised learning dynamics can be con-
#
ture and instead responded to the random back- trolled by the structure of the supervisory matrix 1 XNp

k R aj cos2pk=Np 11
mean
ground activity. Indeed, in the shared input S, I used a triphasic learning schedule that con- Np k1
scenario, correlations between the synaptic ef- sisted of an initialization period, a critical period,
ficacies within the processing layer grew during and a consolidation period. Specifically, each and converted back to the processing layer loca-
^
learning (fig. S6B). Although the probability that simulation started by a brief initialization period tions (neuron id) of the center responses id aj
^
a network would enter the pathological state be- of 50 cycles, during which the supervisory matrix Np maj =2p. These center responses were fitted

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-11


R ES E A RC H | R E S EA R C H A R T I C LE

by a straight line whose y-intercept (phase), slope, the remaining 55 connected digit sequences from where m denotes the number of output spikes
and coefficient of determination (R2) values were each speaker. The lengths of these sequences were that the neuron fires before t*, tsj < t * for
used to characterize the learned maps over 1000 variable, comprising two, three, four, five, and j f1; :::; mg. By definition of the neuron model
independent simulations (fig. S9C). seven digits (11 each). The TIDIGITS database and the critical threshold *, all spike times tsj
comprises 22 distinct dialect groups (five to six and t* satisfy
Auditory front-end speakers each), which were used separately in * V t * V tsj 15
Following the auditory brain stem model used the present application and resulted in 22 11 =
in (23), speech samples from the TIDIGITS data- 242 individual tasks. Generalization performance so that for each afferent i f1; :::; N g
base were converted to spike patterns through was improved by training margins and spike time d d d
the following sequence of processing steps: Audio noise (23), which I optimized over a three- *i * V t * V tsj 16
dwi dwi dwi
files (20 kHz sampling frequency) were con- dimensional grid spanning Gaussian spike time

verted into spectrograms by using the specgram jitters s f0; 5; 10g ms and independent spike where we introduced the abbreviation *i for
function of the matplotlib python package, with threshold margins k = {0, 0.1, 0.2, 0.3, 0.4, 0.5}. the ith component of the wanted gradient.
a sliding window of 512 samples and a temporal The two margin parameters refer to a positive Because * depends on the synaptic efficacy wi
resolution of 1 ms. These spectrograms were margin, = 1 + k+, when at least one output spike also through the previous spike times tsj < t *
smoothed in frequency space into 16 overlapping was required, and a negative margin, = 1 k (Eq. 14), its derivative with respect to wi is given
channels between 360 and 8000 Hz. Specifically, when the neuron should remain silent. For each by
the signal in each channel corresponded to the of the 3 6 6 different parameter combina-
dot product between the spectrogram and a tri- tions, the performance was measured by the mean @ Xm
@ d j
*i V t * j
V t * ts 17
angular frequency filter. Each of the 16 filters minimal error (over 1000 learning cycles) over 10 @wi j1 @t s dwi
spanned three adjacent values of the logarithmic independent simulations. The histogram in Fig. 3B
(Mel-scale) frequency list (360, 500, 658, 837, reports the test errors of all 10 simulations for the where we dropped the vanishing
1039, 1269, 1528, 1822, 2155, 2531, 2957, 3439, best parameter combination for each digit and
@ d
3984, 4602, 5301, 6092, 6987, and 8000) and dialecta total of 10 11 22 = 2420 values. V t * t* 0 18
@t * dwi
peaked at the center value. The resulting fre- Accommodating the dimensions of the front-end,
quency averaged spectrogram S was normalized the number of afferents in this application of the The above relation holds because V(t*) is either a
to its global maximum, logarithmicly rescaled by multi-spike tempotron was set to N = 496. All other local maximum with @V(t*)/@t* = 0 or t* is the
S = log(S + e) log(e) with e = 105, and again parametersexcept the learning rate, which was time of an inhibitory input spike whose arrival
peak normalized. Last, the signal in each channel increased to l = 5 105remained unaltered. time does not depend on wi. Similarly, for each
of S was smoothed in time by a Gaussian filter k f1; :::; mg
with a time constant of 20 ms. For spike gen- *-gradient
eration, the final signals were rescaled to the in- The multi-spike tempotron learning rule re- d @
V tsk V tsk
terval [0, 1] by subtraction of the global minimum quires the evaluation of the gradient w * of dwi @wi
and subsequent division by the global maximum. the appropriate critical threshold * in the Xk
@ d j
V tsk t 19
Spikes were generated by threshold crossings. space of synaptic efficacies w . After an error trial, j dwi s
j1 @ts
Specifically, in each frequency band 15 thresh- * is determined numerically in two steps: First,
old values (0.0625, 0.125, 0.1875, 0.25, 0.3125, the * is bracketed by interval halving until the from which we obtain
"
0.375, 0.4375, 0.5, 0.5625, 0.625, 0.6875, 0.75, 0.8125, extra spike that is elicited when the spike thres- d k 1 @

0.875, and 0.9375) triggered onset (crossings from hold corresponds to the lower bound can be ts *i V tsk
dwi V tsk @wi
below) and offset (crossings from above) spikes, analytically associated with a local subthreshold
#
respectively, and fed these into separate output voltage maximum vmax() when is set to the X
k1
@ d j
k
channels. One additional threshold placed at upper bound. In the second step, * is deter- V t s t 20
j1 @ts
j dwi s
1 fed into a single output channel. In total, each mined numerically (with accuracy 1013) as the
speech sample was converted into a spike pat- root of the analytic expression for [ vmax()]. with the time derivatives
tern over 16 (2 15 + 1) = 496 afferents that At the critical spike threshold *, the post- @ 

traced the envelope of the spectrotemporally synaptic voltage V(t) (Eq. 5) that is elicited by V tsk V t k 21
@t tts
smoothed spectrogram. the presynaptic spike pattern with spike times
tij is given by evaluated from the left, before the spike reset.
TIDIGITS-task
  To solve Eq. 17 for *i , we refactor the right
X t tsj
In Fig. 3, I used continuous digit sequences from V t Vo t * exp 12 hand side of Eq. 20 into the form
the TIDIGITS database to train multi-spike tem- j
tm d k 1
ts <t
t *i Ak Bk  22
potrons with aggregate-label learning to detect dwi s V t k
one out of the 11 English digit words zero, oh, where s

one, two, three, four, five, six, seven, where the coefficients Ak and Bk are given by
X
N X
eight, and nine, which served as the clue. For Vo t wi K ttij 13 the recursive equations
each training example, the neurons supervisory i1 tij <t Xk1
Aj @
signal consisted only of the aggregate-label, the Ak 1 j j
V tsk 23
number of times that the clue occurred within is the unreseted subthreshold voltage and the j1 V ts @ts

the speech sequence. The neurons task was to sum over tsj in Eq. 12 runs over all output spikes and
learn to fire one spike in response to each clue elicited before time t. We assume (materials and @ X
k1
Bj @
occurrence. I divided the speech data from all methods, spike-threshold-surface) that for any Bk V tsk j
V tsk 24
@wi j1 V tsj @ts
111 male speakers (21 to 70 years of age) of the given * there exists a unique t* so that
database into a training and a test set. The test With the analogous definitions
set consisted of the isolated digit subset of the * V t *
X  
database and contained 22 utterances of isolated m
t * tsj X
m
Aj @
digits (two productions of each digit) from each Vo t * * exp 14 A* 1 V t * 25
j1
tm j1 V tsj @ts
j
speaker. In contrast, I defined the training set as

aab4113-12 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

and where the conditional entropy 6. R. V. Florian, The chronotron: A neuron that learns to fire
temporally precise spike patterns. PLOS ONE 7, e40233
@ X
m
Bj @ X (2012). doi: 10.1371/journal.pone.0040233; pmid: 22879876
B* V t * j
V t * 26 HtjC pC HtjC c 34 7. Y. Xu, X. Zeng, S. Zhong, A new supervised learning algorithm
@wi j1 V tsj @ts for spiking neurons. Neural Comput. 25, 14721511 (2013).
C
doi: 10.1162/NECO_a_00450; pmid: 23517101
for t*, insertion of Eq. 22 into Eq. 17 yields quantifies the average uncertainty in the target 8. R. M. Memmesheimer, R. Rubin, B. P. Olveczky, H. Sompolinsky,
Learning precisely timed spikes. Neuron 82, 925938 (2014).
B feature identity after a labeled occurrence count
doi: 10.1016/j.neuron.2014.03.026; pmid: 24768299
*i * 27 vector has been provided.
A* 9. F. Rosenblatt, Principles of Neurodynamics: Perceptrons and
For a specific realization of the labeled occur- the Theory of Brain Mechanisms (Spartan Books, 1962).

To calculate A* and B*, we consider all times tx rence count vector c , the entropy HtjC c 10. R. Gtig, To spike, or when to spike? Curr. Opin. Neurobiol.
at which the voltage reaches the spike thresh- lnm only depends on the number of features m 25, 134139 (2014). doi: 10.1016/j.conb.2014.01.004;
pmid: 24468508
old, when tx fts1 ; ts2 ; :::; tsm ; t * g. At these times, whose occurrence counts are equal to the label. 11. H. S. Seung, Learning in spiking neural networks by
Eq. 12 reduces to Hence, reinforcement of stochastic synaptic transmission. Neuron 40,
Vo tx X X
10631073 (2003). doi: 10.1016/S0896-6273(03)00761-X;
V tx 28
Nf
pmid: 14687542
C tx HtjC pC HtjC c pmlnm 35 12. R. Gtig, H. Sompolinsky, The tempotron: A neuron that learns
m1
C spike timing-based decisions. Nat. Neurosci. 9, 420428
with (2006). doi: 10.1038/nn1643; pmid: 16474393
  where p(m) can be written as the total probability 13. I. R. Fiete, M. S. Fee, H. S. Seung, Model of birdsong learning
X tx tsj
Ctx 1 exp 29 X based on gradient estimation by dynamic perturbation of
tm pm pXt kpmjXt k 36 neural conductances. J. Neurophysiol. 98, 20382057 (2007).
tsj <tx doi: 10.1152/jn.01311.2006; pmid: 17652414
k0
14. L. R. Rabiner, A tutorial on hidden Markov models and selected
and gives the derivatives Because all individual feature counts Xi for applications in speech recognition. Proc. IEEE 77, 257286

@ 1 @ i f1; :::; Nf g are iid, we can define (1989). doi: 10.1109/5.18626


V tx Vo tx 15. At first sight, it seems that losing the clue timing eliminates
@wi Ctx @wi almost all the necessary feedback information, and asking for a
X pk pXi k specific number of spikes seems irrelevant because we do not
1
K tx tij 30 1 rcmean k exp1 rcmean  really care about the total number of spikes if we only get at
C tx j f f
37 least one spike for each clue. However, a varying spike count
ti <tx k! per clue would indicate that the detector is operating in a
h i regime in which noise could suppress or add spikes, which
tx tsk and obtain implies that the detector would not function reliably when the
@ Vo tx exp tm
V tx for tks < t x 31 ! clues are embedded in background activity. This suggests that
@tsk Ct2x tm Nf 1 Nf m
the number of spikes could serve as an objective function to
pmjXt k pm1
k 1 pk 38 stabilize an effective synaptic configuration and might also
m1 suffice for learning such a configuration from random initial
and
 conditions.
1 @ so that 16. Because a neurons memory for previous inputs decays rapidly
V tx Ctx Vo tx (neural membrane time constants rarely exceed a few tens of
Ct2x @tx !
  Nf 1 X m
milliseconds), it has to fire rather promptly in response to a
Vo tx X tx tsj pm p 1 pk Nf m 39 clue, if at all.
exp 32 m 1 k0 k 17. B. Widrow, M. A. Lehr, 30 years of adaptive neural networks:
tm j tm
ts <tx Perceptron, Madaline, and backpropagation. Proc. IEEE 78,
and 14151442 (1990). doi: 10.1109/5.58323
18. D. Barber, Learning in spiking neural assemblies in Advances
!
Mutual information between feedback X
Nf
Nf 1
in Neural Information Processing Systems, S. Becker, S. Thrun,
It;C ln Nf ln m  K. Obermayer, Eds. (MIT Press, 2002), vol. 15, pp. 149156.
signal and clue identity
m1
m1 19. J. P. Pfister, T. Toyoizumi, D. Barber, W. Gerstner, Optimal
In the discrete feature detection task, the sensory X

Nf m
spike-timing-dependent plasticity for precise action potential
environment is composed of Nf distinct features, pm
k 1 pk 40 firing in supervised learning. Neural Comput. 18, 13181348
k0 (2006). doi: 10.1162/neco.2006.18.6.1318; pmid: 16764506
one of which is randomly selected as target fea- 20. Although a population of 135 neurons (with 80 synapses
ture t f1; :::; Nf g. Hence, a priori the uncertainty In Fig. 2C, I use the normalized mutual informa- each) required more than 3000 trials of reinforcement
about the target feature identity is given by the tion, It;C It;C =ln Nf , which measures the learning (4), the single integrate-and-fire neuron underlying
the present study required an average of eight or four trials,
entropy Ht = lnNf. On a given learning trial, the fraction of bits of the target feature uncertainty
when learning was implemented with the multi-spike
occurrence counts of all sensory features i f1; that can on average be inferred from a given trial. tempotron or the correlation-based approximation (discussion
:::; Nf g are realizations ci of Poisson random The pairwise correlations r only enter the mutual and materials and methods, correlation-based learning),
variables Ci, each with mean rate cmean f . Homo- information through the probabilities pk of the respectively. In fact, based on aggregate labels the task
could be readily solved by a single neuron with only
geneous pairwise correlations of strength r 0; 1 individual count variables Xi.
20 synapses (table S2). Similarly strong performance
between these occurrences counts are introduced differences between other neural implementations of
by implementing each individual feature count RE FERENCES AND NOTES gradient-based and reinforcement learning (11) have been
Ci = Y + Xi as the sum of a common count var- 1. R. S. Sutton, Learning to predict by the methods of temporal reported (12) in the context of a binary classification task.
21. T. Yang, M. N. Shadlen, Probabilistic reasoning by neurons.
iable Y and a feature specific variable Xi, with differences. Mach. Learn. 3, 9 (1988).
2. W. Schultz, P. Dayan, P. R. Montague, A neural substrate of Nature 447, 10751080 (2007). doi: 10.1038/nature05852;
mean rates rcmeanf and 1 rcmean
f , respectively. pmid: 17546027
prediction and reward. Science 275, 15931599 (1997).
On each learning trial, information about the doi: 10.1126/science.275.5306.1593; pmid: 9054347 22. L. F. Abbott, S. B. Nelson, Synaptic plasticity: Taming the
target feature identity is provided by a label that 3. E. M. Izhikevich, Solving the distal reward problem through beast. Nat. Neurosci. 3 (suppl.), 11781183 (2000).
linkage of STDP and dopamine signaling. Cereb. Cortex 17, doi: 10.1038/81453; pmid: 11127835
specifies the occurrence count of the target fea-
24432452 (2007). doi: 10.1093/cercor/bhl152; pmid: 17220510 23. R. Gtig, H. Sompolinsky, Time-warp-invariant neuronal
ture, = ct. To quantify the amount of information 4. J. Friedrich, R. Urbanczik, W. Senn, Spatio-temporal credit processing. PLOS Biol. 7, e1000141 (2009). doi: 10.1371/
that the resulting labeled vector of occurrence assignment in neuronal population learning. PLOS Comput. journal.pbio.1000141; pmid: 19582146

counts C c1 ; c2 ; :::cNf ; l provides about the Biol. 7, e1002092 (2011). doi: 10.1371/journal.pcbi.1002092; 24. R. G. Leonard, G. Doddington, TIDIGITS, vol. LDC93S10
target feature identity, we calculate the mutual pmid: 21738460 (Linguistic Data Consortium, 1993).
5. F. Ponulak, A. Kasiski, Supervised learning in spiking 25. T. Masquelier, R. Guyonneau, S. J. Thorpe, Spike timing
information neural networks with ReSuMe: Sequence learning, dependent plasticity finds the start of repeating patterns
classification, and spike shifting. Neural Comput. 22, 467510 in continuous spike trains. PLOS ONE 3, e1377 (2008).
It;C Ht HtjC ln Nf HtjC 33 (2010). doi: 10.1162/neco.2009.11-08-901; pmid: 19842989 doi: 10.1371/journal.pone.0001377; pmid: 18167538

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aab4113-13


R ES E A RC H | R E S EA R C H A R T I C LE

26. B. Nessler, M. Pfeiffer, L. Buesing, W. Maass, Bayesian 34. J. Amores, Multiple instance classification: Review, taxonomy 44. M. Larkum, A cellular mechanism for cortical associations: An
computation emerges in generic cortical microcircuits through and comparative study. Artif. Intell. 201, 81105 (2013). organizing principle for the cerebral cortex. Trends Neurosci. 36,
spike-timing-dependent plasticity. PLOS Comput. Biol. 9, doi: 10.1016/j.artint.2013.06.003 141151 (2013). doi: 10.1016/j.tins.2012.11.006; pmid: 23273272
e1003037 (2013). pmid: 23633941 35. H. Hotelling, Relations between two sets of variates. Biometrika 45. J. A. Cummings, R. M. Mulkey, R. A. Nicoll, R. C. Malenka, Ca2+
27. R. Linsker, Perceptual neural organization: Some approaches 28, 321377 (1936). doi: 10.1093/biomet/28.3-4.321 signaling requirements for long-term depression in the
based on network models and information theory. Annu. Rev. 36. S. Becker, G. E. Hinton, Self-organizing neural network that hippocampus. Neuron 16, 825833 (1996). doi: 10.1016/
Neurosci. 13, 257281 (1990). doi: 10.1146/ discovers surfaces in random-dot stereograms. Nature 355, S0896-6273(00)80102-6; pmid: 8608000
annurev.ne.13.030190.001353; pmid: 2183677 161163 (1992). doi: 10.1038/355161a0; pmid: 1729650 46. R. C. Malenka, R. A. Nicoll, Long-term potentiationA decade
28. R. Gtig, R. Aharonov, S. Rotter, H. Sompolinsky, Learning 37. S. Becker, Mutual information maximization: models of cortical of progress? Science 285, 18701874 (1999). doi: 10.1126/
input correlations through nonlinear temporally asymmetric self-organization. Network 7, 731 (1996). doi: 10.1088/ science.285.5435.1870; pmid: 10489359
Hebbian plasticity. J. Neurosci. 23, 36973714 (2003). 0954-898X/7/1/003 47. J. Hertz, A. Krough, R. G. Palmer, Introduction to the Theory of
pmid: 12736341 38. Y. Li, S. D. Van Hooser, M. Mazurek, L. E. White, D. Fitzpatrick, Neural Computation (Westview Press, 1991).
29. W. Schultz, Updating dopamine reward signals. Curr. Opin. Experience with moving visual stimuli drives the early
Neurobiol. 23, 229238 (2013). doi: 10.1016/ development of cortical direction selectivity. Nature 456, AC KNOWLED GME NTS
j.conb.2012.11.012; pmid: 23267662 952956 (2008). doi: 10.1038/nature07417; pmid: 18946471 It is a pleasure to thank W. Denk, D. Hansel, and P. Dayan for
30. V. Pawlak, J. R. Wickens, A. Kirkwood, J. N. Kerr, Timing is 39. S. D. Van Hooser et al., Initial neighborhood biases and the quality very insightful discussions throughout all stages of this work
not everything: Neuromodulation opens the STDP gate. Front. of motion stimulation jointly influence the rapid emergence of and L. Abbott, J. Hillmann, A. Schaefer, S. Shamma, T. Sharpee,
Synaptic Neurosci. 2, 146 (2010). doi: 10.3389/ direction preference in visual cortex. J. Neurosci. 32, 72587266 and F. Wolf for helpful comments on this manuscript. I thank
fnsyn.2010.00146; pmid: 21423532 (2012). doi: 10.1523/JNEUROSCI.0230-12.2012; pmid: 22623671 D. Fliegner and U. Haase for outstanding technical support and the
31. A. Artola, S. Brcher, W. Singer, Different voltage-dependent 40. L. Wiskott, T. J. Sejnowski, Slow feature analysis: Unsupervised Max Planck Society for making this research possible.
thresholds for inducing long-term depression and long-term learning of invariances. Neural Comput. 14, 715770 (2002).
potentiation in slices of rat visual cortex. Nature 347, 6972 doi: 10.1162/089976602317318938; pmid: 11936959 SUPPLEMENTARY MATERIALS
(1990). doi: 10.1038/347069a0; pmid: 1975639 41. N. Li, J. J. DiCarlo, Unsupervised natural experience rapidly
www.sciencemag.org/content/351/6277/aab4113/suppl/DC1
32. D. E. Feldman, The spike-timing dependence of plasticity. alters invariant object representation in visual cortex. Science
Materials and Methods
Neuron 75, 556571 (2012). doi: 10.1016/j.neuron.2012.08.001; 321, 15021507 (2008). pmid: 18787171
Supplementary Text S1 and S2
pmid: 22920249 42. R. Min, M. Santello, T. Nevian, The computational power of
Figs. S1 to S9
33. V. Pawlak, D. S. Greenberg, H. Sprekeler, W. Gerstner, astrocyte mediated synaptic plasticity. Front. Comput. Neurosci. 6,
Tables S1 and S2
J. N. Kerr, Changing the responses of cortical neurons 93 (2012). doi: 10.3389/fncom.2012.00093; pmid: 23125832
References (4856)
from sub- to suprathreshold using single spikes in vivo. 43. P. Poirazi, T. Brannon, B. W. Mel, Pyramidal neuron as two-
eLife 2, e00012 (2013). doi: 10.7554/eLife.00012; layer neural network. Neuron 37, 989999 (2003). 26 April 2015; accepted 20 January 2016
pmid: 23359858 doi: 10.1016/S0896-6273(03)00149-1; pmid: 12670427 10.1126/science.aab4113

aab4113-14 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H

strategy that has been used previously is the


RESEARCH ARTICLE SUMMARY development of agents that show specific
or preferential cytotoxicity toward TICs. In
the current study, we have embraced an al-
CANCER BIOLOGY
ternative strategy that is designed to induce
TICs to undergo a mesenchymal-to-epithelial
Activation of PKA leads to transition (MET). This induced differen-
tiation approach would trigger cells to exit

mesenchymal-to-epithelial transition the more mesenchymal tumor-initiating state

ON OUR WEB SITE



and enter into an epithe-
lial non-stemlike state.
and loss of tumor-initiating ability Read the full article
at http://dx.doi.
In principle, this transi-
tion would make the cells
Diwakar R. Pattabiraman, Brian Bierie, Katharina Isabelle Kober, Prathapan Thiru, org/10.1126/ more vulnerable to con-
science.aad3680 ventional cytotoxic treat-
Jordan A. Krall, Christina Zill, Ferenc Reinhardt, Wai Leong Tam, Robert A. Weinberg* ..................................................
ments and thereby reduce
INTRODUCTION: Tumor-initiating cells (TICs) the TIC state has presented an attractive op- the likelihood of metastasis and clinical
have emerged in recent years as important portunity for drug development using agents relapse.
targets for cancer therapy owing to their el- that preferentially target more mesenchymal
evated resistance to conventional chemother- carcinoma cells, rather than their epithelial RESULTS: To identify agents that might in-
apy and their tumor-initiating ability. Although counterparts, in an effort to eliminate TICs. duce an MET in mesenchymal mammary
their mode of generation and biological Adenosine 3,5-monophosphate (cAMP) is a epithelial cells, we performed a screen for

Downloaded from on March 3, 2016


properties have been explored in a diverse ar- second messenger that transmits intracellu- compounds that stimulate transcription
ray of cancer types, our understanding of lar signals through multiple downstream effec- of CDH1, which encodes E-cadherin, a key
the biology of TICs remains superficial. The tors; the most well studied of these is protein epithelial protein. Through this screen,
epithelial-to-mesenchymal transition (EMT) is kinase A (PKA). In this study, we explore the compounds that activate adenylate cyclase
a cell-biological program that confers mesen- role of PKA in determining the epithelial versus (cholera toxin, CTx; and forskolin, Fsk) were
chymal traits on both normal and neoplastic mesenchymal properties of mammary epithe- identified as key inducers of the epithelial state.
epithelial cells, which enables both to acquire lial cells and how this signaling pathway affects We found that mesenchymal cells treated with
stemlike properties. In the case of carcinoma the tumor-initiating ability of transformed either CTx or Fsk differentiated into benign
cells, entrance into a more mesenchymal state cells. epithelial derivatives that had lost their abil-
is associated with elevated resistance to a va- ity to effectively initiate tumors and that
riety of conventional chemotherapeutics. This RATIONALE: At least two approaches might were more susceptible to conventional che-
association between the EMT program and be taken to target mesenchymal TICs. One motherapeutic agents in vitro. Further inter-
rogation revealed that these agents elevated
the intracellular levels of cAMP, which in
turn activates PKA. PHF2, a histone H3
with acetylated lysine 9 (H3K9) histone
demethylase and PKA substrate, was found
to be essential for the cAMP-induced MET.
By studying the genome occupancy of PHF2
and the epigenomic state of the cells before
and after PKA activation, we determined
that PHF2 promotes the demethylation and
derepression of epithelial genes that ulti-
mately contribute to acquisition of an epi-
thelial state.

CONCLUSION: We conclude that PKA par-


ticipates in the differentiation of TICs by
enforcing residence in the epithelial state
and preventing or reversing the EMT pro-
gram. Our study reveals a new direction for
targeting the TIC population. We propose
that pharmacological induction of epige-
netic reprogramming of these cells could
promote their differentiation to a more
Induction of the MET as a potential cancer therapy. TICs have mesenchymal attributes that epithelial state and increase their suscep-
tibility to conventional chemotherapeutic

contribute to their ability to seed new tumors. Treatment of TICs with compounds that increase
cAMP levels (e.g., CTx and Fsk) activates PKA. This leads to epigenetic reprogramming through drugs.
subsequent activation of the histone demethylase PHF2, a PKA substrate, which in turn
promotes differentiation of the cells into a more epithelial state, accompanied by a loss of their The list of author affiliations is available in the full article online.
*Corresponding author. E-mail: weinberg@wi.mit.edu
tumor-initiating ability. Drugs targeting various steps of this signaling pathway might be Cite this article as D. R. Pattabiraman et al., Science 351,
developed into a differentiation-based cancer therapy for certain breast cancers. aad3680 (2016). DOI: 10.1126/science.aad3680

1042 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H

Because PKA is initially assembled as a hetero-


RESEARCH ARTICLE tetrameric holoenzyme, its activity depends on
cAMP binding to its two regulatory subunits,
which leads to the release of active catalytic sub-
CANCER BIOLOGY units and the phosphorylation of a diverse array
of substrates (12).

Activation of PKA leads to In previous work, PKA has been shown, under
some conditions, to promote an EMT; PKA was
shown to regulate the transcription factor Snail
mesenchymal-to-epithelial transition in one study; and another study demonstrated
that hypoxia-inducible factor 1a (HIF-1a) could

and loss of tumor-initiating ability regulate transcription of PRKACA under hypoxic


conditions (13, 14). However, PKA signaling has
been shown to favor the epithelial state, but the
Diwakar R. Pattabiraman,1 Brian Bierie,1 Katharina Isabelle Kober,1 mechanistic understanding of this phenomenon
Prathapan Thiru,1 Jordan A. Krall,1 Christina Zill,1 Ferenc Reinhardt,1 is limited. One report showed that schwannomas
Wai Leong Tam,1,2,3 Robert A. Weinberg1,4,5* in mice without Prkar1a (encoding the PKA reg-
ulatory subunit) exhibited loss of vimentin and
The epithelial-to-mesenchymal transition enables carcinoma cells to acquire gain of cytokeratins and E-cadherin (15), whereas
malignancy-associated traits and the properties of tumor-initiating cells (TICs). TICs have another study revealed inhibition of the forma-
emerged in recent years as important targets for cancer therapy, owing to their ability to tion of mesoderm-derived structures in Prkar1a
drive clinical relapse and enable metastasis. Here, we propose a strategy to eliminate null mice (16). A recent study reported that de-
mesenchymal TICs by inducing their conversion to more epithelial counterparts that have letion of the Gas subunit repressed the activity of
lost tumor-initiating ability. We report that increases in intracellular levels of the second PKA, which limited the proliferative potential of
messenger, adenosine 3,5-monophosphate, and the subsequent activation of protein epithelial hair follicle stem cells (17). Neverthe-
kinase A (PKA) induce a mesenchymal-to-epithelial transition (MET) in mesenchymal less, the connection of PKA signaling to TICs and
human mammary epithelial cells. PKA activation triggers epigenetic reprogramming of the stemlike state is poorly understood, and the
TICs by the histone demethylase PHF2, which promotes their differentiation and loss of exploitation of this pathway as a differentiation-
tumor-initiating ability. This study provides proof-of-principle for inducing an MET as based cancer therapy has not been explored.
differentiation therapy for TICs and uncovers a role for PKA in enforcing and maintaining
the epithelial state. Identification of agents that induce an
MET in mammary epithelial cells

T
Human breast cancers are characterized by cells
umor-initiating cells (TICs), also known as enchymal attributes, which suggests that they that show various degrees of epithelial and mes-
cancer stem cells, are defined operationally have passed, at least partially, through an EMT enchymal properties, as revealed by the expres-
by their ability to seed new tumors when (79). This association between the EMT pro- sion pattern of markers, such as cytokeratins and
implanted in appropriate hosts. They have gram and the TIC state has presented an at- vimentin (fig. S1). Almost 85% of the carcinomas
emerged in recent years as important tar- tractive opportunity for drug development using we examined showed varied expression patterns
gets for cancer therapy, owing to their elevated agents that preferentially target more mesenchy- of cytokeratins, which indicated that the loss of
resistance to conventional chemotherapy and mal carcinoma cells, rather than their epithelial epithelial properties is a commonly occurring
their tumor-initiating abilitythe latter allows counterparts, in an effort to eliminate TICs. event. Notably, ~10% of the carcinomas we exam-
them to metastasize and to drive clinical relapse At least two approaches might be taken to ined exhibited high degrees of intratumoral hete-
(1, 2). Although their mode of generation and target mesenchymal TICs. One strategy would be rogeneity, created in part by the presence of
biological properties have been explored in a to develop agents that show specific or preferen- subpopulations of neoplastic cells that have both
diverse array of cancer types (3), our understand- tial cytotoxicity toward TICs (1). In this study, we epithelial and mesenchymal properties. These
ing of the biology of TICs remains superficial. have embraced an alternative strategy that is are reminiscent of cells that have undergone an
Cytotoxic therapies designed specifically to elim- designed to induce TICs to exit the more mes- EMT, which resemble TICs that have a higher
inate TICs might be targeted, for example, to enchymal tumor-initiating state and enter into tumor-initiating propensity and an increased re-
interdict the signaling pathways that are used an epithelial non-stemlike state. Such induced sistance to chemotherapy (18). To model the be-
preferentially or uniquely by these cells (4). At differentiation should, we reasoned, place cells havior of these subpopulations of carcinoma cells
present, however, the nature of such TIC-specific in a state where they would become more vul- from human basal-like breast cancers, we used
signaling pathways remains to be fully elucidated. nerable to conventional cytotoxic treatments. immortalized human mammary epithelial cells
The epithelial-to-mesenchymal transition (EMT) Accordingly, we screened for agents that could (HMLE cells) (19), which display an epithelial mor-
is a cell-biological program that confers mesen- induce a mesenchymal-to-epithelial transition phology; express E-cadherin at adherens junctions;
chymal traits on both normal and neoplastic (MET) and, thereby, uncovered the central role and have low levels of mesenchymal markers,
epithelial cells (5). In addition, activation of an of adenosine 3,5-monophosphate (cyclic AMP such as vimentin and fibronectin. They also ex-
EMT program enables both classes of cells to or cAMP) and its downstream target, protein hibit a CD44lo/CD24hi cell surface marker pheno-
acquire stemlike properties (6, 7). Indeed, TICs kinase A (PKA), in governing the transition of type that is characteristic of previously reported
from several carcinoma types have distinct mes- cells from the mesenchymal to the epithelial state. cells that lack stemlike properties (non-CSCs)
cAMP is a second messenger that transmits (20). We also used their spontaneously arising
1
Whitehead Institute for Biomedical Research, Cambridge,
intracellular signals when certain hormones and mesenchymal derivatives, termed NAMEC8 (N8)
MA 02142, USA. 2Genome Institute of Singapore, 60 Biopolis neurotransmitters interact with receptors on the cells (21). Relative to their HMLE counterparts,
Street, Singapore. 3Cancer Science Institute of Singapore, 14 plasma membrane (10). cAMP regulates multiple N8 cells express mesenchymal markers, such as
Medical Drive, Singapore. 4Department of Biology, downstream effectors; the first of these to be vimentin and fibronectin, as well as the EMT-
Massachusetts Institute of Technology, Cambridge, MA
02139, USA. 5Ludwig Center for Molecular Oncology at MIT,
identified and the most well studied is PKA (11), inducing transcription factors Snail and Zeb1 at
Cambridge, MA 02142, USA. which plays numerous roles in various cell types higher levels; lack expression of E-cadherin at
*Corresponding author. E-mail: weinberg@wi.mit.edu and operates in several subcellular locations (11). prominent cell junctions; and display a CSC-like

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aad3680-1


R ES E A RC H | R E S EA R C H A R T I C LE

CD44hi/CD24lo cell surface marker profile (Fig. 1, encodes E-cadherin, a key epithelial protein, in cells with the characteristic cobblestone mor-
A to C). They also have a greater propensity to N8 cells. As a reporter for the activity of the CDH1 phology of epithelial cells; such cells acquired the
form mammospheres (Fig. 1, D and E), which is gene, we constructed a lentiviral vector that ex- expression of E-cadherin at adherens junctions
often used as an in vitro surrogate assay for the presses a portion of the CDH1 promoter fused to along with a loss of mesenchymal markers, such
stemness of mammary epithelial cells. They are luciferase (fig. S2A). We performed a screen using as vimentin (Fig. 1, A and C). Also, the cell surface
more efficient at migration through a trans-well a 400-compound library for agents that were able marker expression profile of the N8 cells switched
membrane and invasion through a Matrigel- to induce the CDH1-driven luciferase expression in from a stemlike CD44hi/CD24lo to a nonstem
coated Boyden chamber membrane (Fig. 1, F and N8 cells (fig. S2B). Most striking was the behavior CD44lo/CD24hi phenotype after this treatment
G); both in vitro assays represent models of of forskolin (Fsk), an adenylate cyclase activator (20) (Fig. 1B). These shifts were accompanied
cancer cell invasiveness in vivo. N8 cells are also that induced a 40-fold increase in luciferase activ- by a 100-fold increase in CDH1 mRNA levels, as
more resistant to treatment with chemother- ity (fig. S2C). Another adenylate cyclase activator, well as decreases in the mRNA levels of Snail,
apeutic drugs, such as doxorubicin and paclitaxel cholera toxin (CTx), was also able to induce an in- Twist1, and Zeb1 EMT-inducing transcription
(Fig. 1, H and I), as shown previously (21). Hence, crease in luciferase activity (fig. S2D), which sug- factors (EMT-TFs) to 25%, 20%, and 14% , re-
two cell types represent epithelial and mesen- gested that activation of adenylate cyclase could spectively, of the N8 cells before the transition
chymal derivatives of mammary epithelial cells induce the acquisition of epithelial properties. (fig. S3, A and B). Treatment of N8 cells with
of common origin that were used to model the either CTx or Fsk resulted in a near-complete loss
two cell states and how they affect tumor initia- Fsk or CTx and the induction of an MET of mammosphere-forming ability (Fig. 1, D and
tion and progression. in mammary epithelial cells E), as well as their ability to migrate and invade
To search for agents that can induce an MET, We found that treatment of N8 cells in mono- (Fig. 1, F and G). There were no significant dif-
we performed a screen to identify compounds layer culture with either CTx or Fsk for a period ferences in the rates of proliferation between the
that could induce transcription of CDH1, which of 14 days induced the formation of islands of N8 cells and their CTx- and Fsk-treated derivatives

Fig. 1. Induction of an MET on treatment of N8 cells with CTx or Fsk. (P < 0.05, n = 4) and (G) to invade in transwell assays (P < 0.05, n = 4) and to
Mesenchymal N8 cells acquire an epithelial morphology as adjudged by their acquire increased sensitivity to treatment with (H) doxorubicin and (I) paclitaxel
morphology (A), loss of a stemlike CD44hi/CD24lo profile to assume a pre- (P < 0.05, n = 4). (J) Heat map of mRNA-Seq data, which demonstrates
dominantly CD44lo/CD24hi profile (B), and expression of E-cadherin at cell similarity in gene expression between HMLE, N8, and N8-CTx cells. Data
junctions and loss of vimentin (C). Reverted N2-CTx and N3-Fsk cells lose their (means SD) in (E), (F), and (G) were analyzed by Students t test; (H) and (I)
ability to form (D and E) mammospheres (P < 0.05, n = 4), (F) to migrate were analyzed by Bonferroni correction. All scale bars, 25 mm.

aad3680-2 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

(fig. S3C). Of additional interest, withdrawal of mined by differential gene expression (Fig. 1J (25); cyclic nucleotidegated ion channels that
CTx after 14 days of treatment led to cell popu- and tables S1 and S2) and principal component are primarily found in cells of the kidney, heart,
lations that continued to reside in an epithelial analyses (fig. S3D), the N8-CTx cells assume a testis, and central nervous system (26); and the
state for >2 months in culture. gene expression profile that is almost completely most commonly studied downstream effector,
Reversion to an epithelial state, ostensibly sim- converted to that of the epithelial HMLE cells PKA (11). To delineate the downstream pathways
ilar to that of HMLE cells, rendered the N8 cells 8 and is significantly different from the mesenchy- that are activated in response to increase in cAMP
times as sensitive to killing by doxorubicin [low- mal N8 cells (Fig. 1J). Gene set enrichment anal- levels, we treated N8 cells with two cAMP analogs
ered the median inhibitory concentration (IC50) yses showed that the changes in gene expression 8-bromoadenosine 3,5-cyclic monophosphate
from 1.39 mM to 0.159 mM] and 13 times as sen- from N8 to the N8-CTx cells are highly similar (8-Br-cAMP), which is known to preferentially
sitive to paclitaxel (lowered the IC50 from 4.79 mM to several previously published EMT and MET activate PKA (27), or 8-(4-chlorophenylthio)-2-O-
to 0.35 mM) (Fig. 1, H and I). Also, the induced gene signatures (2224) (fig. S3E). Taken togeth- methyladenosine-3,5-cyclic monophosphate
MET resulted in increased sensitivity to a range er, these observations demonstrated a transition (8-CPT-2Me-cAMP), which is a selective activator of
of chemotherapeutic drugs and inhibitors in- of the N8 cells from a mesenchymal-like state to the exchange proteins activated by cAMP (EPACs)
cluding methotrexate, 90-kD heat shock protein a bona fide epithelial state, which rendered these (28). As was seen with Fsk or CTx, treatment with
(HSP90) inhibitors, proteasome inhibitors, and epi- cells more sensitive to a variety of drugs with 8-Br-cAMP was also able to induce an MET in N8
dermal growth factor receptormitogen-activated potentially important therapeutic implications. cells, whereas 8-CPT-2Me-cAMP treatment had
protein kinase (EGFR-MAPK) pathway inhib- no effect on their mesenchymal properties (Fig.
itors, as observed when we screened against Effects of Fsk and CTx on intracellular 2B). This allowed us to conclude that PKA, rather
two small-molecule libraries (Selleck Anticancer cAMP levels and PKA than the cAMP-activated exchange proteins,
Compound Library and Enzo Kinase Inhibitor To confirm that both Fsk and CTx were working was more likely to play a central role in the MET
Library) (fig. S4). Hence, the induction of an MET through alteration of cAMP levels, we measured process.
rendered the N8 cells more sensitive to a range of the levels of this second messenger in both HMLE Knockdown of the catalytic subunit of PKA
drugs and inhibitors, which points to its utility as and N8 cells using liquid chromatographymass using two different small hairpin RNAs (shRNAs)
a means of overcoming therapeutic resistance. It spectrometry (LC-MS). Treatment with CTx re- (fig. S5A) abrogated the CTx-induced MET pro-
also reinforces the notion that mesenchymal cells sulted in a six- to eightfold increase in the intra- cess in N8 cells, as demonstrated by their inabil-
are more resistant to a range of cytotoxic agents. cellular levels of cAMP, which could be dampened ity to develop a clear epithelial morphology; to
We then performed mRNA sequencing (mRNA- by exposure to SQ22536, an inhibitor of adenylate acquire junctional E-cadherin; and to shed mes-
Seq) to compare the global gene expression profiles cyclase, the enzyme responsible for the formation enchymal markers, such as fibronectin (Fig. 2,
of the mesenchymal N8 and the reverted N8-CTx of cAMP (Fig. 2A). C and D). Moreover, treatment of these PKA-
cells in order to view the transcriptional changes The major downstream targets of cAMP are knockdown cells with CTx failed to induce an ef-
that occur after the induction of MET. As deter- exchange proteins activated by cAMP (EPAC1/2) fective transition from the CD44hi/CD24lo stemlike

Fig. 2. cAMP increases activate PKA, which is both necessary and suffi- for induction of an MET. Knockdown of either PRKACA or PRKACB prevents
cient for the induction of an MET in N8 cells. (A) Mass-spectrometry CTx from inducing an MET in N8 cells as observed by changes in (C) mor-
measurement of cAMP levels in N8 cells that have been treated with CTx or phology, (D) immunofluorescence for E-cadherin and vimentin, and (E) CD44/
Fsk alone and in combination with adenylate cyclase inhibitor SQ22536 CD24 status. (F) Morphological changes of N8 cells undergoing an MET upon
(means SD, P < 0.05, n = 3). (B) Treatment of N8 cells with either 8-CPT-2me- ectopic expression of an active mutant of PKA (caPKA). Data in (A) were ana-
cAMP or 8-Br-cAMP to identify downstream pathways that are responsible lyzed using the Students t test. All scale bars, 25 mm.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aad3680-3


R ES E A RC H | R E S EA R C H A R T I C LE

state to the CD44lo/CD24hi non-stemlike state, cally engineered MMTV-PyMT transgenic mouse been previously reported to be PKA substrates
which was otherwise observable in the absence of model of breast cancer, in which the expression that are retained in the cytoplasm after phos-
PKA knockdown (Fig. 2E). These results further of the oncogene is driven by the mouse mam- phorylation (41) and found no retention of any
reinforced the important role of PKA in the MET mary tumor virus promoter (33). Finally, we note of the Gli proteins in the cytoplasm after treat-
process. that others have recently reported that Fsk pro- ment with CTx or Fsk (fig. S9C). These obser-
Then, we tested whether PKA activity, inde- motes the maintenance of an epithelial mor- vations suggest that the Gli proteins may not
pendent of cAMP, was sufficient to induce an phology in primary human mammary epithelial play a role in the observed PKA-induced MET.
MET. Thus, we ectopically expressed a doxycycline- cells, the absence of which led spontaneously to Having explored the two most commonly re-
inducible constitutively active, cAMP-independent, acquisition of mesenchymal attributes, such as ported nuclear substrates of PKA, we then fo-
constitutively active mutant form of PKA (caPKA) down-regulation of E-cadherin expression and cused on PHF2, a histone H3 with acetylated
(29) in N8 cells and found that it was capable of up-regulation of mesenchymal markers (34). lysine 9 (H3K9) histone demethylase, which is
inducing a reversion to the epithelial state in 7 to Taken together, these data signify the impor- known to become activated upon phosphoryl-
10 days (Fig. 2F). Hence, it appears as though tance of PKA signaling in maintaining epithe- ation by PKA (42). We found that knockdown
PKA is both necessary and sufficient to induce lial characteristics in a variety of normal and of PHF2 expression in N8 cells using either of
the acquisition of epithelial properties in the neoplastic epithelial cells. These data give an in- two shRNAs (fig. S5B) phenocopied the effects
N8 cells. dication that these responses might be a general of PKA knockdown in that it prevented CTx-
We tested the role of CTx or Fsk in inducing an property of cAMP-induced activation of PKA in induced MET (Fig. 3, C and D). In contrast,
epithelial state in other cell systems to assess the the reversal of phenotypes created by activation knockdown of PHF2 did not alter the ability of
generality of our observations. Removing CTx of an EMT program. HMLE cells to undergo an EMT (Fig. 3, F and
from the standard culture medium of MCF10A Although CTx was able to induce entrance of G), which indicated that this enzyme, although
immortalized human mammary epithelial cells the N8 cells and a range of other cell systems into necessary for induction of an MET, apparently
(30), caused them to acquire mesenchymal prop- a stably maintained epithelial state, there were plays no role in the reverse processthe EMT
erties, to lose cell-cell adherens junctions, to lose some models in which neither CTx nor Fsk was which suggested that it is specifically impor-
their characteristic cobblestone morphology, and able to do so, namely, the MDA-MB-231 and tant for the derepression of silenced epithelial
to gain a CD44hi/CD24lo cell surface marker pro- SUM159 human breast cancer cell lines, amongst genes through its function as a H3K9 histone
file. They also lost E-cadherin expression and ex- others. These cell lines are maintained in the demethylase.
hibited an increase in expression of Zeb1, Vim, and mesenchymal state through the deletion or sta- PHF2 can be phosphorylated by PKA at four
FN1 (fig. S6, A to E). Readdition of CTx or forced ble silencing of several key epithelial genomic serine residues in its C terminus (42) (Fig. 3E).
expression of caPKA led to the reacquisition of loci, which includes the repression of E-cadherin Accordingly, we engineered a phospho-mimetic
epithelial features (fig. S6, A to E). Moreover, the through strong DNA promoter hypermethylation form of PHF2 in which all four of these serines
MCF10A cells that lost epithelial properties upon (35). Hence, although the observed effects of PKA were replaced by aspartate residues. Although
CTx withdrawal were 4 times as resistant to treat- activation are applicable to some breast cancer expression of this mutant in N8 cells was not suf-
ment with doxorubicin and so extended our ob- lines and other carcinomas, they are not univer- ficient on its own to induce an MET, the phospho-
servations made in N8 cells that the mesenchymal sal and depend instead on the specific genetic or mimetic PHF2 was able to accelerate the rate of
variants were more resistant to conventional epigenetic state of the cells. CTx-induced transition from the mesenchymal to
chemotherapeutic agents (fig. S6F). the epithelial state from 14 days to 7 days (Fig. 3,
We then proceeded to test the role of CTx and/or Essential role of PKA-induced activation H and I). Hence, although PHF2 is required for
Fsk in a series of other cell lines. MCF7-Ras hu- of PHF2 in MET the acquisition of epithelial traits, it appears to be
man breast cancer cells (31) can be induced to PKA is known to act on many substrates in both only one of the effectors of the PKA operating
undergo an EMT through the ectopic expression the cytoplasm and nucleus (36). Treatment of during induction of epithelial transition.
of EMT-inducing transcription factors, such as HMLE and N8 cells with CTx resulted in an To test whether PHF2 can be directly phos-
Slug. Cotreatment of the cells undergoing an EMT immediate increase in the presence of both iso- phorylated by PKA in our system, we performed
with CTx led to a 48-hour delay in the acquisition forms of the catalytic subunit in the nucleus an immunoprecipitation of PHF2 followed by
of mesenchymal morphology and CD44hi cell (fig. S9, A and B), which suggested that PKA immunoblotting using an antibody that recog-
surface marker expression (fig. S7A). Similarly, might be regulating nuclear substrates after ac- nizes the phospho-PKA substrate motif. As shown
the ability of HMLE-Ras cells to undergo an EMT tivation by cAMP. The most well-studied sub- in Fig. 3J, 24 hours after treatment of N8 cells
upon ectopic expression of Zeb1 was also ham- strate of PKA, CREB1, translocates to the nucleus with CTx, phosphorylation of PHF2 by PKA can be
pered upon cotreatment with CTx (fig. S7B). upon phosphorylation by PKA at Ser133, there- observed, which provides evidence that PKA phos-
PANC1 pancreatic adenocarcinoma cells undergo after altering the transcription of hundreds of phorylates PHF2 in the N8 cells. Together, these
an EMT upon treatment with transforming growth target genes (37). In fact, about 300 distinct results suggest an important role for PHF2 as a
factorb1 (TGF-b1) for 48 hours (32). Cotreatment physiologic stimuli have been described in the PKA substrate in the induction of an MET.
of PANC1 cells undergoing an EMT with either literature that can induce CREB Ser-133 phos-
CTx or Fsk delayed the ability of TGF-b1 to in- phorylation (38). It was, therefore, not surpris- PKA-induced activation of PHF2 and the
duce an EMT by 48 to 72 hours, which enabled ing that CREB was already phosphorylated and epigenetic reprogramming of
the temporary retention of epithelial properties present in the nucleus of the N8 cells even mesenchymal cells
(fig. S7, C and D). before their treatment with either CTx or Fsk The H3K9me2 and H3K9me3 marks have been
Treatment with CTx or Fsk induced the acqui- (Fig. 3A). Note that knockdown of CREB1 by associated with repression of gene transcription
sition of epithelial properties in a range of cell the use of at least two shRNAs (fig. S5C) did not (43). Given the previously reported role of PHF2
lines that have mesenchymal traits, including the affect the ability of CTx to induce an MET in as an H3K9me2/3 demethylase, we performed
Hs578T triple-negative breast cancer cell line the N8 cells (Fig. 3B). Moreover, loss of CREB1 chromatin immunoprecipitation followed by deep
(fig. S8A), the NCI-H596 lung adenosquamous alone induced a partial MET in N8 cells (Fig. sequencing (ChIP-Seq) using antibodies against
carcinoma cell line (fig. S8B), and the mesenchy- 3B), consistent with previous reports of its role histone 3 with trimethylated lysine 9 (H3K9me3)
mal EpCAMlo CD24lo fraction of the EF021 ovarian in the induction of an EMT (39, 40). On the and H3K9me2 marks to observe the presence of
carcinoma cell line (fig. S8C). Induction of epi- basis of these observations, we conclude that these marks in untreated N8 cells as well as CTx-
thelial properties was also observed in PB3 cells CREB1 did not play an important role in the treated counterparts in which PHF2 is active. In
(fig. S8D), which constitute an aggressive cell line PKA-induced MET. We then assessed the lo- addition we also performed ChIP-Seq for PHF2,
isolated from mammary tumors of the geneti- calization of Gli1, Gli2, and Gli3, which have comparing genome-wide occupancy in N8 cells

aad3680-4 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

to the N8-CTx cells. We did so in order to mon- As seen in Fig. 4A, there was a striking inverse previously reported, PHF2 appears to occupy the
itor PHF2-associated alterations that might en- correlation at specific loci of the presence of promoter region of genes where it recognizes the
able phenotypic shifts from the mesenchymal PHF2 with the repressive H3K9me2 or H3K9me3 H3K4me3 histone mark (Fig. 4B) (9). Interest-
to epithelial states, including shifts that might marks. This suggests that presence of this de- ingly, the total H3K9me3 counts (>4-fold enrich-
relieve the H3K9-mediated silencing of epithelial methylase may, on it own, suffice to relieve ment above control) in N8-CTx cells was almost
genes. histone-mediated transcriptional silencing. As half of the total counts of the same mark in N8

Fig. 3. The PKA substrate PHF2, but not CREB1, is necessary for the MET-
inducing properties of CTx. Activation state of CREB1 as measured by levels of p-
CREB1 across HMLE and N8 cells that have been treated with CTx or Fsk (A). Loss
of CREB1 through shRNA-mediated knockdown induces a partial METand permits
CTx-mediated complete MET as shown by changes in morphology and immuno-
fluorescence (B). shRNA-mediated knockdown of PHF2 prevents CTx from inducing
an MET, which prevents changes in (C) morphology and immunofluorescence-based
detection of E-cadherin and fibronectin expression as well as (D) blocking a shift
from the CD44hi/CD24lo state to the CD44lo/CD24hi state. Expression of a PHF2 phos-
phomimetic where the C-terminal serines were modified to aspartate (E) accelerated
the MET transition by 5 days, as observed by changes in immunofluorescence (F) and
quantitative EMT marker analysis by qPCR (G). Effects of shRNA-mediated
knockdown of PHF2 on the ability of HMLE cells to undergo an EMT upon ectopic expression of Zeb1 (H and I) (qPCR data, means SD, P < 0.05, n = 3).
Immunoprecipitation of PHF2 followed by immunoblotting with a phospho-PKA substrate antibody showed phosphorylation of PHF2 by PKA 24 hours after
treatment of N8 cells with CTx (J). (I) was analyzed using the Students t test. All scale bars, 25 mm.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aad3680-5


R ES E A RC H | R E S EA R C H A R T I C LE

cells (35,455 versus 18,675). Similarly, the total activation of PKA in vitro. We transplanted at growing tumor suffices on its own to reduce the
H3K9me2 counts in N8-CTx cells were also less limiting dilutions the neoplastic, RAS-transformed tumor-initiating properties of its cells, as indicated
than a half of that in the N8 cells (1295 vs 473). derivatives of HMLE, N8, and the reverted N8-CTx by their subsequent inability to propagate upon
As shown in the representative circos plots, cells, termed HMLE-Ras, N8-Ras, and N8-CTx-Ras, secondary transplantation.
these data indicate a widespread loss of H3K9- into the mammary fat pads of nonobese diabetic
mediated repression of genomic regions upon severe combined immunodeficient (NOD/SCID) Discussion
treatment of N8 cells with CTx and subsequent mice. As anticipated, the frequency of TICs in the Cyclic AMP and its main effector, PKA, have been
activation of PHF2 (Fig. 4C). N8-Ras cells was far greater than in the HMLE-Ras studied for four decades in a variety of cell-
We then sorted for genomic regions present in cell population, in this case 100-fold as high. Note biological and physiologic settings, where its
the N8-CTx but not N8 cells that contained PHF2 that the N8-CTx-Ras cells were as inefficient at effects in activating a number of distinct, tissue-
binding and lacked repressive H3K9me2/3 marks tumor-initiation as the HMLE-Ras cells (Fig. 5A). specific traits have been repeatedly documented
(table S3). This provided us with a list of genomic The primary tumors that arose upon orthotopic (11). A role that it might play in governing the
regions that were relieved of H3K9me2/3-mediated mammary stromal fat pad implantation of N8- epithelial cell state and thus suppressing en-
silencing in the N8-CTx cells, as compared to the Ras tumors spawned 20 to 30 micrometastases trance into the alternative mesenchymal state
N8 cells, owing to PHF2 occupancy. To ensure in the lungs by 12 weeks after implantation. This in breast cancers has not been described. The
that these changes were specific for the loss of property was lost upon induction of an MET by present work makes it clear that this second
PHF2, we performed ChIP-Seq for H3K9me2/3 CTx treatment before transplantation (Fig. 5C messenger and its main effector, PKA, play a
and PHF2 in CTx-treated N8 cells that had an and fig. S11A), which nevertheless formed primary key role in determining this epithelial versus
shRNA against PHF2 preventing the MET (table tumors of comparable size (Fig. 5B). Moreover, this mesenchymal balance of mammary epithelial
S4). These cells that remained morphologically confirmed previous observations that the pheno- cells, as well as epithelial cells of other tissues.
mesenchymal also demonstrated an epigenetic typic state of these cells before neoplastic trans- Indeed, in light of these results, it becomes plau-
profile more like that of N8 cells with an overlap formation strongly influenced their behavior after sible that maintenance of the residence of cells
of 11,807 peaks compared with the reverted N8- transformation. in an epithelial state depends on tonic elevated
CTx cells, which had an overlap of 6864 peaks. To better mimic a clinical scenario, we next levels of intracellular cAMP. In retrospect, it
Hence, the list of altered genomic regions out- asked how the induction of an MET would affect now seems likely that the use of cholera toxin
lined in table S3 represents genes that were preestablished tumors derived from mesenchy- as an ingredient in the tissue culture medium
relieved of H3K9-mediated repression upon CTx- mal N8 cells. Although we wished to pharmaco- of various epithelial cell types [including cells
induced activation of PHF2. This suggests that logically treat mice that already had established of the epidermis, mammary gland, and bronchus
PHF2 activity could be directly responsible for N8-Ras tumors, CTx is too toxic to be administ- (46, 47)] was motivated by the observation that
the derepression of these genes that are charac- ered systemically, and the rapid clearance and loss of such cells in culture was accompanied
teristic of the epithelial cell state. In addition, poor pharmacodynamics of Fsk made it difficult by an overgrowth of fibroblast-like cells (46).
the expression values of genes that correspond to study its effects upon systemic administration. These results collectively indicate a role for
to these genomic loci were also measured in re- Such difficulties in treating mice with PKA ago- PKA in the differentiation of TICs by enforcing
verted N8-CTx (epithelial) and parental N8 (mesen- nists have also been reported previously (44, 45). residence in the epithelial state and preventing
chymal) cells by RNA-seq, which verified that gain For this reason, we focused our efforts on study- or reversing the EMT program. Although PKA
of PHF2 occupancy and loss of H3K9 marks did ing the proof-of-principle effects of PKA activa- can act via a large number of substrates, we iden-
indeed lead to increased expression (table S3). tion in vivo using the doxycycline-inducible tified PHF2 as an important downstream effector
Several genes that play a major role in the version of constitutively active PKA (caPKA). of PKA that mediates the induction of epithelial
phenotype and profile of cells in the epithelial Thus, we induced expression of the caPKA in characteristics through epigenetic reprogram-
state were activated by CTx treatment. The list of already formed N8-Ras tumors of 5-mm diame- ming to a chromatin state that is more favorable
genes that were relieved of silencing when treated ter (Fig. 5D). On visual inspection, the tumors for residence in the epithelial state. We find that
with CTx includes CDH1 and CDH3 (among oth- that had been exposed for 14 days to doxycycline activating this histone demethylase enables PKA
er cadherin genes) that code for E-cadherin and contained pasty, fluid-filled necrotic cores when to induce the transcription of genes that play a
P-cadherin (fig. S10A), respectively, which are es- compared with the tumors that had never been role in the entrance into and maintenance of resi-
sential components of adherens junctions and exposed to doxycycline: The latter were solid dence in the epithelial state.
hallmark proteins of basal epithelial cells; KRT8 with a hard center of viable cells. Tumors from The EMT program is known to represent one
and KRT18 (fig. S10B), whose gene products are mice that received doxycycline weighed less than defined route for the generation of both normal
characteristic components of the cytoskeleton of those that did not receive any (Fig. 5F). More- and neoplastic epithelial stem cells (6, 7, 48). The
epithelial cells; and AJAP1 and CLDN4 (fig. S10, C over, those tumors in which expression of caPKA observations that PKA-induced activation of
and D), which specify genes coding for constit- had been induced developed a more differen- PHF2 can either reverse or curtail this program
uents of adherens and tight junctions that are tiated histomorphology as revealed by hematox- present an opportunity to exploit such a mech-
formed by epithelial but not mesenchymal cells. ylin and eosin (H&E) staining of tumor sections anism for therapeutic gain. Indeed, the differen-
Other regions include the TP63 gene (fig. S10E), (fig. S11, B and C). When tumors were harvested tiation of TICs through the induction of an MET
whose product is a hallmark transcription factor and subjected to fluorescence-activated cell sort- is an attractive propositionone that could be
of basal mammary epithelial cells, and ITGB2, ing (FACS) analysis, the doxycycline-treated tu- pursued through the induced increase of intra-
ITGB6 (fig. S10F), and ITGB8, which code for in- mors showed a decrease in expression of the CD44 cellular cAMP levels, activation of PKA, or ac-
tegrins that are typically expressed on epithelial cell surface marker associated with the stemlike tivation of PHF2. Nonetheless, it is likely that
cells. These observations reveal a mechanism by population (20), in contrast to the untreated tu- many such approaches will result in widespread
which activation of this demethylase enables the mors (fig. S11D). side-effect toxicities, owing to the multitude of
transcription of genes that induce the acquisition Secondary transplantation of cells isolated from signaling pathways that are activated down-
of epithelial properties, ultimately defining the the doxycycline-exposed tumors at limiting di- stream of cAMP increase (11). Specific activa-
state of the cells. lutions revealed a 20-fold loss of tumor-initiating tion of the PHF2 histone-modifier enzyme may
ability (Fig. 5E), which showed that activation serve as a means of derepressing genes that are
Activation of PKA and the of PKA induces differentiation of TICs and di- essential for the differentiated epithelial state
differentiation of TICs in vivo minishes their ability to subsequently seed new without eliciting many of the toxicities of induced
We tested the tumor-initiating ability of cells tumors. This result demonstrates that constitu- cAMP increases. Along the same lines, identification
that have been induced to undergo an MET by tive expression of PKA for a 14-day period in a of a histone methyltransferase that counteracts

aad3680-6 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

Fig. 4. Activation of PHF2 leads the


epigenetic reprogramming of mesen-
chymal cells. Genome-wide occupancy of
H3K9me2, H3K9me3, and PHF2 marks
shows the inverse correlation between the
presence of the histone marks and the
demethylase (A), which interacts mainly
with the promoter and the first intronic
region of genes (B). Circos plots of
representative chromosomes 5 and 8
show widespread changes in the
H3K9me2 and H3K9me3 profiles (C).

PHF2 function may also provide an attractive through epigenetic rewiring that ultimately insulin at 10 mg/ml. Cells were treated with ei-
target for therapeutic inhibition, a strategy that results in their differentiation and increased ther 100 ng/ml of CTx (Calbiochem, USA; 227036),
has proven successful in the case of DOT1L inhi- susceptibility to conventional chemotherapeu- which was replenished every 2 days, or 1 mM Fsk
bition against mixed lineage leukemia (MLL) tic drugs. (Tocris Biosciences, USA; 1099), which was replen-
driven leukemias (49). The role of the G9a histone ished daily over a period of 14 to 16 days. Cells were
methyltransferase in establishing the H3K9me2 Materials and methods split to a ratio of 1:6 every 2 to 3 days during the
mark for repression of the CDH1 promoter in Cell culture and treatments treatments. MCF10A and EF021 cells were a gift
breast cancer cells has been reported previous- HMLE, NAMEC8, and all derived cell lines were from N. Kalaany and R. Drapkin, respectively.
ly (50). grown in Mammary Epithelial Cell Growth Me-
This study provides mechanistic insight into dium medium (Lonza, USA); MCF10A cells were Screening
the benefits of targeting such an enzyme in ep- grown in Dulbeccos minimum essential medium; For the CDH1 reporter screen, 500 N8 cells bearing
ithelial tumors, which prevents the constituent nutrient mixture F-12 (DMEM/F12) containing 5% wild-type (wt) CDH1 promoter luciferase were
cells from undergoing an EMT and thereby ac- horse serum (Sigma, USA; H0146); epidermal growth seeded into 384-well plates in a volume of 40 ml.
quiring aggressive characteristics, including in- factor, 20 ng/ml; hydrocortisone, 0.5 mg/ml; CTx, Twenty-four hours later, 100 nl of each compound
creases in the numbers of TICs. The pathways 100 ng/ml; and insulin, 10 mg/ml. EF021 and H596 (200 mM stock) were added using a CyBio liquid
explored in this study provide insight into the cells were grown in Roswell Park Memorial Insti- handler, which resulted in a final screen concen-
functions of PKA in the induction of an MET tute (RPMI) medium containing 10% fetal bovine tration of 0.5 mM. Four days later, the plates were
and the differentiation of the more aggressive serum. MCF7Ras cells were grown in DMEM con- read for either firefly luciferase activity (Pierce,
TICs within a tumor. This study reveals a new taining 10% fetal bovine serum. Hs578T were grown 16177) or CellTiter-Glo (Promega, USA; G7572).
direction for targeting the TIC population in DMEM containing 10% fetal bovine serum and The Enzo compound library (plate A and plate

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aad3680-7


R ES E A RC H | R E S EA R C H A R T I C LE

B; 451 compounds, including repeats) was ob- (DAPI) (Life Technologies, USA; D1306) was used After 12 to 24 hours, the cells on the upper sur-
tained from the Koch Institute Screening Facility to exclude dead cells. Cells were sorted on BD face of the filters were removed with a cotton
at MIT. Firefly luciferase and CellTiter-Glo assays FACSAria SORP and analyzed on BD LSRII, using swab. For visualization, cells on lower filter sur-
were performed in triplicates. BD FACSDiva Software (BD Biosciences, USA). faces were fixed and stained with a Diff-Quick
The vulnerabilities of the reverted cells were Antibodies used were against CD44-PE-Cy7 (Bio- staining kit (Dade Behring/Siemens, Germany).
assessed by screening against the Selleck anti- legend, USA; 103029); against CD24-PE (BD Bio- Three to five fields per filter were counted. Data
cancer compound library (400 compounds) and sciences, USA; 555428); against CD45-Pacific Blue are presented as migrated cells per field.
the Enzo kinase library (80 compounds) at the (Biolegend, USA; 103125); and against CD31-
Koch Institute Screening Facility at MIT. N8 or Pacific Blue (Biolegend, USA; 102421). RNA preparation and polymerase chain
N8-CTx cells (1000 each) were seeded in 384-well reaction analysis
plates in a volume of 50 ml. Twenty-four hours Mammosphere and tumorsphere culture Total RNA was isolated using the RNeasy Plus
later, 50 nl of each compound were added to Mammosphere culture was performed as previ- Mini kit (Qiagen, USA; 74136) and reverse tran-
assay a 5-point dose response. Three days later, ously described (51). Cells (1000) were seeded per scription was performed with the High Capacity
the plates were read for CellTiter-Glo, and assays well of a 96-well Corning Ultra-Low attachment RNA-to-cDNA kit (Life Technologies, USA; 4387406),
were performed in duplicate. plate (Corning, USA; CLS3474) in replicates of both according to the manufacturers proto-
10; sphere numbers were counted between days cols. A cDNA sample prepared from 1 mg total
Flow cytometry 8 to 12. RNA was used for quantitative reverse tran-
Cells were prepared according to standard proto- scription polymerase chain reaction (RT-PCR).
cols and suspended in 2% inactivated fetal bovine Migration and invasion assays The PCR reactions were performed with the
serum with phosphate-buffered saline (IFS/PBS). The Cells (25,000) were seeded into 24-well cell cul- Fast SYBR Green Master Mix (Life Technologies;
fluorescent stain 4,6-diamidino-2-phenylindole ture inserts with 8 mm pores (BD Falcon, USA). 4385612), data collection and data analysis were

Fig. 5. PKA-induced MET is sufficient to


deplete the tumor-initiating ability of N8-Ras
cells in vivo. (A) Differences in tumor-initiating
ability of HMLE-Ras, N8-Ras, and N8-CTx-Ras
cells upon transplantation with limiting dilution
into NOD/SCID mice. Tumors that arose from
transplantation of 2 106 cells were of similar
size (B) with only the N8-Ras cells able to form
micrometastases (C). (Each dot represents one
mouse; data analyzed using Students t test;
P < 0.05, n = 10.) (D) Experimental outline to test
the tumor-initiating ability of N8-Ras cells upon
transient in vivo expression of PKA showing a
(E) 20-fold decrease in tumor-initiating ability
after secondary transplantation with (F) no sig-
nificant differences in the tumor volume but a
decrease in tumor mass. (Each dot represents one mouse; data analyzed using Students t test; P < 0.05,
n = 10.)

aad3680-8 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

performed on the ABI7900 machine (Applied aqueous lysis buffer (50 mM Tris pH 7.5, 150 mM face metastases were counted with a fluorescent
Biosystems, USA) by using the SDS2.0 and RQ NaCl, 10 mM EDTA pH 8.0, 0.2% sodium azide, microscope.
manager software. The thermal-cycling parame- 50 mM NaF, and 0.5% NP40) containing com-
ters for the PCR were as follows: 95C for 5 min, plete miniprotease inhibitor cocktail (Roche, Chromatin IP followed by sequencing
followed by 45 cycles each of 95C for 10 s, 49C USA; 04693159001) and stored at 80C. The ChIP for PHF2, H3K9me2, and H3K9me3 was
for 7 s, and 72C for 25 s. The relative mRNA proteins names, sources, and dilutions are shown carried out using the SimpleChIP Plus Enzymatic
quantity was normalized against the relative in table S3. After thawing, they were centrifuged Chromatin IP Kit (Cell Signaling Technology,
quantity of HPRT1 mRNA in the same sample. at top speed on a benchtop centrifuge at 4C for USA; 9005) and the protocols within. The PHF2
The primer sequences in a 5 to 3 orientation are 20 min, and the supernatant was assayed for pro- rabbit monoclonal antibody (Cell Signaling Tech-
shown in table S1. tein concentration with Bradford Reagent (Bio-Rad; nology, USA; 3497) was used at 1:25 per IP; the
500-0006). Of the total protein, 30 mg were sep- H3K9me2 mouse monoclonal (Abcam, USA;
Immunofluorescence (cultured cells) arated by SDSpolyacrylamide gel electrophore- ab1220); and the H3K9me3 rabbit polyclonal
Cells were cultured on dishes containing cover- sis (SDS-PAGE) on NuPage gels (Invitrogen, USA) antibodies (Abcam, ab8898) were used at 1:50
slips for 2 to 3 days, after which coverslips were and transferred to Hybond-P polyvinylidene di- (10 mg) per IP. The ChIP DNA was used to pre-
washed in cold PBS, fixed in 4% paraformaldehyde fluoride membrane (GE Healthcare, USA). Mem- pare libraries for sequencing, which was car-
for 10 min at 4C and permeabilized in 0.2% branes were probed with specific primary antibodies ried out in the Genome Technology Core at the
TritonX in PBS for 2 min. Cells were then washed and antibody-protein complex detected by horse- Whitehead Institute.
in PBS, blocked for 1 hour at room temperature radish peroxidase (HRP)conjugated secondary
in PBS containing 3% normal horse serum (Vector antibodies and SuperSignal West Dura Extended Library preparation for sequencing
Labs, USA; S-2000). Fixed cells were then incu- Duration Substrate (Life Technologies, USA; 34075). To prepare libraries for RNA-Seq, the TruSeq
bated with the primary antibody in PBS contain- stranded mRNA protocol was followed to prep
ing 1% bovine serum albumen (BSA) solution Immunoprecipitation the libraries as described in the kit (Illumina,
overnight at 4C. Cells were washed in PBS three N8 cells in 15-cm dishes were scraped and har- USA; RS-122-2101) manual. To prepare libraries
times, and secondary antibody was added in PBS vested in 0.5 ml of ice-cold immunoprecipitation for the ChIP-Seq, the TruSeq ChIP protocol was
containing 1%BSA solution for 1 to 2 hours at (IP) buffer (Cell Signaling Technology, USA) con- followed as described in the kit (Illumina, USA;
room temperature in the dark. Cells were washed taining protease and phosphatase inhibitors (Roche, IP-202-1012) manual.
three times in PBS and were incubated for 2 min Germany) after which they were sonicated with
in DAPI solution, after which they were washed three pulses on ice. Lysates were spun down at Deep sequencing and data analysis
in PBS and mounted with a drop of Prolong Gold 14,000g for 10 min, and supernatants were col- Libraries were pooled together and sequenced
antifade reagent (Life Technologies, USA; P36961) lected for IP. PHF2-specific antibody (Cell Sig- on the HiSEq 2500 sequencer using the standard
and placed on coverslips. Slides were viewed on a naling Technology, USA; 3497) was added at 1:25 sequencing protocols. Images analysis and base
PerkinElmer Ultraview Spinning Disk Confocal and incubated overnight at 4C rotating. The calling was done using the Standard Illumina
imager and analyzed using Volocity software. following morning, 40 ml of magnetic protein pipeline, and then demultiplexed into FASTQ
A/G beads (Thermo Scientific, USA; 26162) were files. RNASeq paired-end reads from Illumina 1.5
Immunofluorescence added and incubated for 30 min on a rotator, after encoding were aligned using TopHat (version
(tissue microarrays) which beads were washed five times by using a 2.0.13) (52) to the human genome (GRCh37) with
Slides were rehydrated by incubating in Histo- magnetic separator with cold IP buffer. Beads Ensembl annotation (GRCh37.75) in gtf format.
clear solution twice for 5 min each, followed by were then boiled in lithium dodecyl sulfate Differential expression was assayed using HTSeq
incubation in 100% ethanol twice for 5 min each, buffer and run on an SDS-PAGE gel followed count (53) and DESeq (54). ChIPSeq data were
in 95% ethanol twice for 5 min each, 70% ethanol by immunoblotting. aligned to the human genome (GRCh37) using
twice for 5 min each, once in 35% ethanol for Bowtie2 (version 2.2.5) (55), base encoding as
5 min, and in water for 5 min. Pressure cooker RNAi-mediated knockdown above, and peaks were called using MACS2 (ver-
mediated heat-induced epitope retrieval was To generate shRNA-expressing plasmids, double- sion 2.1.0.20150420) (56) with nomodel option,
carried out in 250 ml of unmasking buffer con- stranded oligonucleotides (oligos) encoding the and fragment length was determined by strand
taining sodium citrate at pH 6. After retrieval, desired shRNA were cloned into a Tet-pLKO-puro cross-correlation (using phantompeakqualtools;
slides were blocked for 30 min in PBS containing lentiviral vector (Addgene, plasmid 21915). In the https://code.google.com/p/phantompeakqualtools/).
3% normal horse serum after which they were absence of doxycycline, shRNA expression is re- Differential binding was determined by using
incubated with primary antibody in blocking pressed by constitutively expressed TetR protein. MACS bdgdiff tool. Peaks were annotated using
solution overnight at 4C. Slides were washed With the addition of doxycycline to the growth Cis-regulatory Element Annotation System (CEAS)
twice with PBS and incubated with secondary medium, shRNA expression is triggered, which (57), and ChIPSeq data profiles were viewed in
antibody at room temperature for 1 hour in the results in target gene knockdown. The cloning ngsplot (58). Overlap between peaks, and with
dark. After two PBS washes, 20 ml of mounting vector has a 1.9-kb stuffer that is released by di- expression data, was determined using bedtools
medium was added, then slide contents were gestion with Age I and Eco RI. shRNA oligos are (59). ChIPSeq data profiles were viewed in ngsplot
topped with coverslips, and stored in the dark for cloned into the Age I and Eco RI sites in place of (58) and the Integrative Genomics Viewer (60).
24 hours before imaging. A table of the antigens the stuffer. PKA hairpins are shown in tables S4 RNA-seq and ChIP-seq data have been submitted
with source, host, and dilution is shown in table S2. to S6. to GEO under the generic stream encapsulation
(GSE) ID GSE74883.
Proliferation assays Animal studies
To measure rate of proliferation, 1000 cells were Research involving animals complied with pro- LC/MS-based metabolite profiling
seeded onto a 96-well plate in quadruplicate. tocols approved by the MIT Committee on LC/MS analyses were conducted on a QExactive
Proliferation was measured using CyQuant (Life Animal Care. For tumor studies, cells suspended benchtop Orbitrap mass spectrometer equipped
Technologies, USA; C7026), according to the manu- in 15 ml 30% Matrigel (GFR)/PBS mix (BD Bio- with an Ion Max source and a HESI II probe,
facturers protocols. sciences; 356230) were injected into the ingui- which was coupled to a Dionex UltiMate 3000
nal mammary gland fat pads of age-matched high-performance liquid chromatography sys-
Protein extraction and Western blotting female NOD/SCID mice (Jackson Laboratory). tem (Thermo Fisher Scientific, San Jose, CA).
To obtain protein extracts, cells were washed with Mice were killed after 10 weeks or when tu- External mass calibration was performed using
chilled PBS and scraped from culture dishes in mors reached a diameter of >1 cm. Lung sur- the standard calibration mixture every 7 days.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aad3680-9


R ES E A RC H | R E S EA R C H A R T I C LE

Polar metabolites were extracted using 1 ml of carcinoma. PLOS ONE 6, e16466 (2011). doi: 10.1371/journal. unregulated biological activity. Proc. Natl. Acad. Sci. U.S.A. 89,
ice-cold 80% methanol with 10 ng/ml phenylalanine- pone.0016466; pmid: 21304586 47264730 (1992). doi: 10.1073/pnas.89.10.4726;
9. N. Ahmed, K. Abubaker, J. Findlay, M. Quinn, Epithelial pmid: 1584809
d8 or phenylalanine-13C9-15N as an internal 30. H. D. Soule et al., Isolation and characterization of a
mesenchymal transition and cancer stem cell-like phenotypes
standard. After 10 min of using a vortex and facilitate chemoresistance in recurrent ovarian cancer. Curr. spontaneously immortalized human breast epithelial cell line,
centrifugation for 10 min at 10,000g, both at 4C, Cancer Drug Targets 10, 268278 (2010). doi: 10.2174/ MCF-10. Cancer Res. 50, 60756086 (1990). pmid: 1975513
samples were dried in a centrifugal evaporator. 156800910791190175; pmid: 20370691 31. C. L. Sommers, A. Papageorge, G. Wilding, E. P. Gelmann,
J. A. Beavo, L. L. Brunton, Cyclic nucleotide research still Growth properties and tumorigenesis of MCF-7 cells
Dried samples were stored at 80C and then 10.
transfected with isogenic mutants of rasH. Cancer Res. 50,
expanding after half a century. Nat. Rev. Mol. Cell Biol. 3,
resuspended in 100 ml water; 2.5 ml of each sam- 710718 (2002). doi: 10.1038/nrm911; pmid: 12209131 6771 (1990). pmid: 2403419
ple was injected onto a ZIC-pHILIC 2.1 150 mm 11. K. Taskn, E. M. Aandahl, Localized effects of cAMP mediated 32. V. Ellenrieder et al., Transforming growth factor beta1
(5-mm particle size) column (EMD Millipore). by distinct routes of protein kinase A. Physiol. Rev. 84, 137167 treatment leads to an epithelial-mesenchymal
(2004). doi: 10.1152/physrev.00021.2003; pmid: 14715913 transdifferentiation of pancreatic cancer cells requiring
Buffer A was 20 mM ammonium carbonate, 0.1% extracellular signal-regulated kinase 2 activation. Cancer Res.
12. S. S. Taylor, R. Ilouz, P. Zhang, A. P. Kornev, Assembly of
ammonium hydroxide; buffer B was acetonitrile. allosteric macromolecular switches: Lessons from PKA. Nat. 61, 42224228 (2001). pmid: 11358848
The chromatographic gradient was run at a flow Rev. Mol. Cell Biol. 13, 646658 (2012). doi: 10.1038/nrm3432; 33. C. T. Guy, R. D. Cardiff, W. J. Muller, Induction of mammary
rate of 0.150 ml/min as follows: 0 to 20 min, lin- pmid: 22992589 tumors by expression of polyomavirus middle T oncogene: A
13. M. R. MacPherson et al., Phosphorylation of serine 11 and transgenic mouse model for metastatic disease. Mol. Cell. Biol.
ear gradient from 80% to 20% B; 20 to 20.5 min, 12, 954961 (1992). doi: 10.1128/MCB.12.3.954; pmid: 1312220
serine 92 as new positive regulators of human Snail1 function:
linear gradient from 20% to 80% B; 20.5 to 28 min, Potential involvement of casein kinase-2 and the cAMP- 34. J. R. Linnemann et al., Quantification of regenerative potential
hold at 80% B. The column oven was held at 25C. activated kinase protein kinase A. Mol. Biol. Cell 21, 244253 in primary human mammary epithelial cells. Development 142,
The mass spectrometer was operated with the (2010). doi: 10.1091/mbc.E09-06-0504; pmid: 19923321 32393251 (2015). doi: 10.1242/dev.123554; pmid: 26071498
35. A. Hollestelle et al., Distinct gene mutation profiles among
spray voltage set to 3.0 kV, the heated capillary 14. D. Shaikh et al., cAMP-dependent protein kinase is essential for
luminal-type and basal-type breast cancer cell lines. Breast
hypoxia-mediated epithelial-mesenchymal transition,
held at 275C, and the HESI probe held at 350C; Cancer Res. Treat. 121, 5364 (2010). doi: 10.1007/s10549-
migration, and invasion in lung cancer cells. Cell. Signal. 24,
the sheath gas flow was set to 40 units, the aux- 009-0460-8; pmid: 19593635
23962406 (2012). doi: 10.1016/j.cellsig.2012.08.007;
36. J. B. Shabb, Physiological substrates of cAMP-dependent
iliary gas flow was set to 15 units, and the sweep pmid: 22954688
protein kinase. Chem. Rev. 101, 23812412 (2001).
gas flow was set to 1 unit. To measure cAMP, a 15. K. S. Nadella et al., Targeted deletion of Prkar1a reveals a role
doi: 10.1021/cr000236l; pmid: 11749379
for protein kinase A in mesenchymal-to-epithelial transition.
positive targeted SIM (tSIM) scan was performed 37. G. A. Gonzalez, M. R. Montminy, Cyclic AMP stimulates
Cancer Res. 68, 26712677 (2008). doi: 10.1158/0008-5472. somatostatin gene transcription by phosphorylation of CREB at
at a resolution of 70,000, an automatic gain con- CAN-07-6002; pmid: 18413734 serine 133. Cell 59, 675680 (1989). doi: 10.1016/0092-8674
trol (AGC) target of 1 105, and the maximum 16. P. S. Amieux et al., Increased basal cAMP-dependent protein (89)90013-5; pmid: 2573431
injection time at 250 ms. The tSIM window kinase activity inhibits the formation of mesoderm-derived 38. M. Johannessen, M. P. Delghandi, U. Moens, What turns CREB
was set to a width of 1.0 m/z and centered at structures in the developing mouse embryo. J. Biol. Chem. 277, on? Cell. Signal. 16, 12111227 (2004). doi: 10.1016/
2729427304 (2002). doi: 10.1074/jbc.M200302200; j.cellsig.2004.05.001; pmid: 15337521
330.05980 m/z, corresponding to the [M+H]+ pmid: 12004056 39. D. Wu et al., cAMP-responsive element-binding protein
ion of cAMP. To monitor other endogenous polar 17. R. Iglesias-Bartolome et al., Inactivation of a Ga(s)-PKA tumour regulates vascular endothelial growth factor expression:
metabolites and the internal standard, the tSIM suppressor pathway in skin stem cells initiates basal-cell Implication in human prostate cancer bone metastasis.
scans were interspersed with positive and nega- carcinogenesis. Nat. Cell Biol. 17, 793803 (2015). Oncogene 26, 50705077 (2007). doi: 10.1038/sj.onc.1210316;
doi: 10.1038/ncb3164; pmid: 25961504 pmid: 17310988
tive mode scans in the range of 70 to 1000 m/z, 18. X. Li et al., Intrinsic resistance of tumorigenic breast cancer 40. R. Singh, B. S. Shankar, K. B. Sainis, TGF-b1-ROS-ATM-CREB
with the resolution set to 70,000, the AGC target cells to chemotherapy. J. Natl. Cancer Inst. 100, 672679 signaling axis in macrophage mediated migration of human
at 106, and the maximum injection time at 80 ms. (2008). doi: 10.1093/jnci/djn123; pmid: 18445819 breast cancer MCF7 cells. Cell. Signal. 26, 16041615 (2014).
Relative quantitation of polar metabolites was per- 19. B. Elenbaas et al., Human breast cancer cells generated by doi: 10.1016/j.cellsig.2014.03.028; pmid: 24705025
oncogenic transformation of primary mammary epithelial cells. 41. T. Sheng, S. Chi, X. Zhang, J. Xie, Regulation of Gli1 localization
formed with XCalibur QuanBrowser 2.2 (Thermo Genes Dev. 15, 5065 (2001). doi: 10.1101/gad.828901; by the cAMP/protein kinase A signaling axis through a site
Fisher Scientific) by using a 5 parts per million pmid: 11156605 near the nuclear localization signal. J. Biol. Chem. 281, 912
mass tolerance and referencing an in-house li- 20. M. Al-Hajj, M. S. Wicha, A. Benito-Hernandez, S. J. Morrison, (2006). doi: 10.1074/jbc.C500300200; pmid: 16293631
brary of chemical standards. M. F. Clarke, Prospective identification of tumorigenic breast 42. A. Baba et al., PKA-dependent regulation of the histone lysine
cancer cells. Proc. Natl. Acad. Sci. U.S.A. 100, 39833988 demethylase complex PHF2-ARID5B. Nat. Cell Biol. 13,
Statistical analysis (2003). doi: 10.1073/pnas.0530291100; pmid: 12629218 668675 (2011). doi: 10.1038/ncb2228; pmid: 21532585
21. W. L. Tam et al., Protein kinase C a is a central signaling node 43. P. Hublitz, M. Albert, A. H. F. M. Peters, Mechanisms of
Data are presented as means SD. A Students t and therapeutic target for breast cancer stem cells. Cancer transcriptional repression by histone lysine methylation. Int. J.
test (two-tailed) was used to compare two groups Cell 24, 347364 (2013). doi: 10.1016/j.ccr.2013.08.005; Dev. Biol. 53, 335354 (2009). doi: 10.1387/ijdb.082717ph;
pmid: 24029232 pmid: 19412890
(P < 0.05 was considered significant) unless 22. T. T. Onder et al., Loss of E-cadherin promotes metastasis via 44. M. P. Saunders et al., A novel cyclic adenosine monophosphate
otherwise indicated. multiple downstream transcriptional pathways. Cancer Res. 68, analog induces hypercalcemia via production of 1,25-
36453654 (2008). doi: 10.1158/0008-5472.CAN-07-2938; dihydroxyvitamin D in patients with solid tumors. J. Clin.
pmid: 18483246 Endocrinol. Metab. 82, 40444048 (1997). doi: 10.1210/
RE FE RENCES AND N OT ES 23. D. Anastassiou et al., Human cancer cells express Slug-based jcem.82.12.4410; pmid: 9398710
1. P. B. Gupta et al., Identification of selective inhibitors of cancer epithelial-mesenchymal transition gene expression signature 45. D. J. Propper et al., Phase I study of the novel cyclic AMP
stem cells by high-throughput screening. Cell 138, 645659 obtained in vivo. BMC Cancer 11, 529 (2011). doi: 10.1186/1471- (cAMP) analogue 8-chloro-cAMP in patients with cancer:
(2009). doi: 10.1016/j.cell.2009.06.034; pmid: 19682730 2407-11-529; pmid: 22208948 Toxicity, hormonal, and immunological effects. Clin. Cancer
2. J. Chen et al., A restricted cell population propagates 24. E. Charafe-Jauffret et al., Gene expression profiling of breast Res. 5, 16821689 (1999). pmid: 10430069
glioblastoma growth after chemotherapy. Nature 488, cell lines identifies potential new basal markers. Oncogene 25, 46. J. Taylor-Papadimitriou, P. Purkis, I. S. Fentiman, Cholera
522526 (2012). doi: 10.1038/nature11287; pmid: 22854781 22732284 (2006). doi: 10.1038/sj.onc.1209254; toxin and analogues of cyclic AMP stimulate the growth of
3. A. Kreso, J. E. Dick, Evolution of the cancer stem cell model. pmid: 16288205 cultured human mammary epithelial cells. J. Cell. Physiol.
Cell Stem Cell 14, 275291 (2014). doi: 10.1016/ 25. J. de Rooij et al., Epac is a Rap1 guanine-nucleotide-exchange 102, 317321 (1980). doi: 10.1002/jcp.1041020306;
j.stem.2014.02.006; pmid: 24607403 factor directly activated by cyclic AMP. Nature 396, 474477 pmid: 6248570
4. D. R. Pattabiraman, R. A. Weinberg, Tackling the cancer stem (1998). doi: 10.1038/24884; pmid: 9853756 47. H. Green, Cyclic AMP in relation to proliferation of the
cellswhat challenges do they pose? Nat. Rev. Drug Discov. 13, 26. M. C. Broillet, S. Firestein, Cyclic nucleotide-gated channels. epidermal cell: A new view. Cell 15, 801811 (1978).
497512 (2014). doi: 10.1038/nrd4253; pmid: 24981363 Molecular mechanisms of activation. Ann. N. Y. Acad. Sci. 868, doi: 10.1016/0092-8674(78)90265-9; pmid: 83196
5. B. De Craene, G. Berx, Regulatory networks defining EMT 730740 (1999). doi: 10.1111/j.1749-6632.1999.tb11352.x; 48. W. Guo et al., Slug and Sox9 cooperatively determine the
during cancer initiation and progression. Nat. Rev. Cancer 13, pmid: 10414360 mammary stem cell state. Cell 148, 10151028 (2012).
97110 (2013). doi: 10.1038/nrc3447; pmid: 23344542 27. R. B. Meyer Jr., J. P. Miller, Analogs of cyclic AMP and cyclic doi: 10.1016/j.cell.2012.02.008; pmid: 22385965
6. S. A. Mani et al., The epithelial-mesenchymal transition GMP: General methods of synthesis and the relationship of 49. S. R. Daigle et al., Selective killing of mixed lineage leukemia
generates cells with properties of stem cells. Cell 133, 704715 structure to enzymic activity. Life Sci. 14, 10191040 (1974). cells by a potent small-molecule DOT1L inhibitor. Cancer Cell
(2008). doi: 10.1016/j.cell.2008.03.027; pmid: 18485877 doi: 10.1016/0024-3205(74)90228-8; pmid: 4362776 20, 5365 (2011). doi: 10.1016/j.ccr.2011.06.009;
7. A. P. Morel et al., Generation of breast cancer stem cells 28. J. M. Enserink et al., A novel Epac-specific cAMP analogue pmid: 21741596
through epithelial-mesenchymal transition. PLOS ONE 3, e2888 demonstrates independent regulation of Rap1 and ERK. Nat. Cell 50. C. Dong et al., G9a interacts with Snail and is critical for Snail-
(2008). doi: 10.1371/journal.pone.0002888; pmid: 18682804 Biol. 4, 901906 (2002). doi: 10.1038/ncb874; pmid: 12402047 mediated E-cadherin repression in human breast cancer.
8. C. Chen et al., Evidence for epithelial-mesenchymal transition 29. S. A. Orellana, G. S. McKnight, Mutations in the catalytic J. Clin. Invest. 122, 14691486 (2012). doi: 10.1172/JCI57349;
in cancer stem cells of head and neck squamous cell subunit of cAMP-dependent protein kinase result in pmid: 22406531

aad3680-10 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S E ARCH | R E S E A R C H A R T I C L E

51. G. Dontu et al., In vitro propagation and transcriptional profiling genomic databases. BMC Genomics 15, 284 (2014). Foundation, Singapore (NRF-NRFF2015-04); and National Medical
of human mammary stem/progenitor cells. Genes Dev. 17, doi: 10.1186/1471-2164-15-284; pmid: 24735413 Research Council, Singapore (NMRC/TCR/007-NCC/2013). This
12531270 (2003). doi: 10.1101/gad.1061803; 59. A. R. Quinlan, I. M. Hall, BEDTools: A flexible suite of utilities for research was supported by the Ludwig Center for Molecular Oncology
pmid: 12756227 comparing genomic features. Bioinformatics 26, 841842 at MIT (R.A.W.), Breast Cancer Research Foundation (R.A.W.), Samuel
52. D. Kim et al., TopHat2: Accurate alignment of transcriptomes (2010). doi: 10.1093/bioinformatics/btq033; pmid: 20110278 Waxman Cancer Research Foundation (R.A.W.) and NIH R01-CA078461
in the presence of insertions, deletions and gene fusions. 60. J. T. Robinson et al., Integrative genomics viewer. Nat. (R.A.W.). R.A.W. is an American Cancer Society and D. K. Ludwig
Genome Biol. 14, R36 (2013). doi: 10.1186/gb-2013-14-4-r36; Biotechnol. 29, 2426 (2011). doi: 10.1038/nbt.1754; Foundation Cancer Research Professor. R.A.W. is a shareholder in and
pmid: 23618408 pmid: 21221095 the chairman of the Scientific Advisory Board of Verastem, Inc., a
53. S. Anders, P. T. Pyl, W. Huber, HTSeqa Python framework to work company focused on developing drugs to treat cancer by the targeted
with high-throughput sequencing data. Bioinformatics 31, 166169 killing of cancer stem cells. The Whitehead Institute and the authors
ACKN OWLED GMEN TS
(2015). doi: 10.1093/bioinformatics/btu638; pmid: 25260700 (R.A.W., D.R.P., B.B., and W.L.T.) have filed a patent application
54. S. Anders, W. Huber, Differential expression analysis for We thank the Keck Microscopy Facility, Metabolite Profiling Core
(PCT/US2015/028239) that covers methods of differentiating cancer
sequence count data. Genome Biol. 11, R106 (2010). Facility, Genome Technology Core, Bioinformatics and Research
stem cells through the induction of a MET and the use of these methods
doi: 10.1186/gb-2010-11-10-r106; pmid: 20979621 Computing Core, and Proteomics Core Facility at the Whitehead
for treating cancer. RNA-Seq and ChIP-Seq data from this study
55. B. Langmead, S. L. Salzberg, Fast gapped-read alignment with Institute; the Koch Institute Swanson Biotechnology Center (SBC),
have been deposited at GEO under accession number GSE74883.
Bowtie 2. Nat. Methods 9, 357359 (2012). doi: 10.1038/ specifically the Histology Facility and the High-Throughput Screening
nmeth.1923; pmid: 22388286 Facility; J. Benson (SBC) for providing the compound library;
56. Y. Zhang et al., Model-based analysis of ChIP-Seq (MACS). D. Bachovchin (Broad Institute) for assistance with automation; and SUPPLEMENTARY MATERIALS
Genome Biol. 9, R137 (2008). doi: 10.1186/gb-2008-9-9-r137; T. DiCesare for assistance with scientific illustration. We thank
www.sciencemag.org/content/351/6277/aad3680/suppl/DC1
pmid: 18798982 A. Dongre, S. Iyer, and J. Frse for technical assistance; all members of
Figs. S1 to S11
57. H. Shin, T. Liu, A. K. Manrai, X. S. Liu, CEAS: Cis-regulatory the Weinberg laboratory for helpful discussions; and A. Lambert,
Tables S1 to S10
element annotation system. Bioinformatics 25, 26052606 K. Krishnan, and S. Rajavasireddy for critical reading of the manuscript.
References
(2009). doi: 10.1093/bioinformatics/btp479; pmid: 19689956 D.R.P is supported by a C. J. Martin Overseas Biomedical Fellowship
58. L. Shen, N. Shao, X. Liu, E. Nestler, ngs.plot: Quick mining and from the National Health and Medical Research Council of Australia 3 September 2015; accepted 7 January 2016
visualization of next-generation sequencing data by integrating (NHMRC APP1071853). W.L.T. is supported by the National Research 10.1126/science.aad3680

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 aad3680-11


R E SE A R CH

SOSIP.664 structures (49), most of the C termi-


RESEARCH ARTICLES nus of HR2 is helical until residue 664. However,
hydrogen deuterium exchange mass spectrome-
try (HDXMS) studies (21) have demonstrated
STRUCTURAL BIOLOGY that the C-terminal region of HR2 in BG505
SOSIP.664 has a flexible topology. Whereas the

Cryo-EM structure of a native, fully C-terminal region of gp41 is observed in our JR-FL
EnvDCT structure, the micelle-embedded MPER
and TM just downstream of HR2 were both un-
glycosylated, cleaved HIV-1 resolved (Fig. 1A; fig. S1, D and E; and fig. S2C).
Crystal structures of MPER peptideFab complexes

envelope trimer have also shown that MPER can adopt different
conformations (2224).

Jeong Hyun Lee, Gabriel Ozorowski, Andrew B. Ward* Model building of newly resolved regions
We used the BG505 SOSIP.664 and PGT151 Fab
The envelope glycoprotein trimer (Env) on the surface of HIV-1 recognizes CD4+ T cells and x-ray structure coordinates [PDB IDs 4TVP (6)
mediates viral entry. During this process, Env undergoes substantial conformational and 4NUG (16), respectively] as starting models
rearrangements, making it difficult to study in its native state. Soluble stabilized trimers for building and refinement (figs. S3, S5F, and S6
have provided valuable insights into the Env structure, but they lack the hydrophobic and table S1). The fusion peptide (FP, residues
membrane proximal external region (MPER, an important target of broadly neutralizing 512 to 527) and HR1N (548 to 568) (Fig. 1C and fig.
antibodies), the transmembrane domain, and the cytoplasmic tail. Here we present (i) a S4) regions of gp41 were both resolved in the cur-
cryogenic electron microscopy (cryo-EM) structure of a clade B virus Env, which lacks only rent structure. The HR1N helix in interfaces 2 and
the cytoplasmic tail and is stabilized by the broadly neutralizing antibody PGT151, at a 3 was tilted by ~24 away from the center of the

Downloaded from on March 3, 2016


resolution of 4.2 angstroms and (ii) a reconstruction of this form of Env in complex with trimer and ~26 to the right in comparison with
PGT151 and MPER-targeting antibody 10E8 at a resolution of 8.8 angstroms. These interface 1 (Fig. 2A), when viewed normal to the
structures provide new insights into the wild-type Env structure. threefold axis of the trimer. The HR1N and HR1C

T
form a helix-turn-helixtype motif but contain
he HIV-1 envelope glycoprotein (Env) houses to make the sample amenable for cryogenic electron residues with high helical propensity in the turn
the receptor binding site and fusion machin- microscopy (cryo-EM) (fig. S1, A and B), resulting region (Fig. 2A and fig. S4). The HR1N region in
ery to infect target cells. The intrinsic in- in a 4.2 resolution reconstruction (Fig. 1, A and the unliganded protomer is less well defined than
stability of and glycosylation on Env have B, and figs. S1 to S3). Similar to our negative-stain the liganded HR1N helices (fig. S5D), as would be
made solving a high-resolution structure reconstructions (16), PGT151 Fab bound in an asym- expected for a conformationally variable segment.
a daunting task. Low-resolution tomographic metric manner, with a maximum of two Fabs per In the BG505.664 x-ray structure, the N termi-
reconstructions of Env on the viral surface have trimer. Because the three gp140 interfaces of the nus of the FP (residues A512 to A518) is disordered
described the overall shape of the trimer (1, 2), Env trimer were nonequivalent, we hereafter and appears to project into the solvent. Similarly,
and more recently, structures of an engineered, refer to them as interfaces 1, 2, and 3 (Fig. 1A). at interface 1 of the cryo-EM model, residues A512
soluble clade A BG505 SOSIP.664 trimer have The clade B JR-FL EnvDCT shares a similar to L520 are unresolved, but a density adjacent to
been solved at high resolution (39). The BG505 topology to BG505 SOSIP.664, despite the lack of HR1N could be attributed to the FP (fig. S7A); this
SOSIP.664 trimer interacts preferentially with broad- stabilizing mutations and the difference in sub- suggests that the hydrophobic N terminus of the
ly neutralizing antibodies (bnAbs) but not with types (68.5% sequence identity) (fig. S4 and fig. FP may be inserted into the trimer core in wild-
nonneutralizing antibodies (10) and has promising S5, A and E). Differences were observed at the type Env, as observed in another class I viral fu-
immunogenic properties (11, 12). Although these trimer apex and the N-terminal region of heptad sion protein, influenza hemagglutinin (25). At
data suggest that this soluble trimer recapitulates repeat 1 (HR1N) of gp41. In the JR-FL trimer, the the PGT151 liganded interfaces, there are changes
native Env, it is not known what effect the sta- inter-V1/V2 loop region of the trimer apex is more in both the antibody and the FP. The majority of
bilizing mutations or the lack of the membrane loosely associated than in the unliganded BG505 the Fab heavy- and light-chain residues adopt the
proximal external region (MPER) and transmem- SOSIP.664 [Protein Data Bank identifier (PDB ID) same conformation as when the Fab is unlig-
brane domain (TM) have on the Env structure. 4ZMJ] (fig. S5B). This phenomenon is consistent anded, except for the heavy-chain complementarity-
with fluorescence resonance energy transfer studies determining region 3 (CDRH3) of PGT151 (P100a
BG505 SOSIP.664 shares a highly of viral Env (17), as well as studies of clades B and C to Y100l), which is substantially different in the
similar architecture with wild-type Env SOSIP.664 trimers in which loop mobility and two structures (fig. S7B). Upon PGT151 binding,
We studied the JR-FL Env strain with the cyto- flexibility have been observed (18, 19). Despite this the entire FP is resolved (fig. S7C), where it pro-
plasmic tail (CT) deleted (hereafter referred to as weaker interaction, the V3 loop in all three pro- jects away from the trimer and is sequestered in
EnvDCT). In some isolates, the deletion of the CT tomers remains in contact with the base of V2 on a pocket formed between the PGT151 Fab CDRH2,
has been shown to increase the exposure of non- the adjacent protomer (figs. S5B and S6A), and CDRH3, and CDRL3 (L, light chain) loops via
neutralizing epitopes (13), but the deletion of CT therefore it probably confers the majority of the hydrophobic and backbone interactions (Fig. 2B
in JR-FL does not abolish the ability of the trimer stability at the trimer apex. and fig. S7D). Because the FP is pulled away from
to fuse and infect (14, 15). Our previously described In the published BG505 SOSIP.664 structures, the trimer core in both BG505 SOSIP.664 and
protocol (16) for extracting the complex formed by HR1N does not adopt a regular secondary struc- PGT151-bound Env, which stabilizes the pre-
JR-FL EnvDCT and the bnAb PGT151 was modified ture (49), whereas in our JR-FL EnvDCT model, fusion trimer, this conformation of the FP may
this region is helical (fig. S5D). We attribute this counterintuitively contribute to trimer stability.
Department of Integrative Structural and Computational difference to the I559P mutation in SOSIP that dis-
Biology, Center for HIV/AIDS Vaccine Immunology and rupts the propensity of the HR1 peptide to form Newly revealed glycans in the wild-type
Immunogen Discovery, International AIDS Vaccine Initiative an extended and stable a-helix during fusion (20). Env trimer
Neutralizing Antibody Center, and Collaboration for AIDS
Vaccine Discovery, The Scripps Research Institute, La Jolla,
In contrast, the SOS disulfide bond does not cause The Env trimers in our studies contain fully pro-
CA 92037, USA. any major conformational differences relative to cessed native glycans. Similar to the two other
*Corresponding author. E-mail: abward@scripps.edu the wild-type structure (fig. S5C). As in the BG505 cryo-EM structures (5, 9), at least the two core

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1043


R ES E A RC H | R E S EA R C H A R T I C LE S

Fig. 1. Cryo-EM reconstruction of JR-FL EnvDCT. Interface 1


(A) Reconstruction of JR-FL EnvDCT in complex 180
180 90
90
with PGT151 Fab at 4.2 resolution, segmented to
highlight densities corresponding to gp120 (yellow),
gp41 (blue), PGT151 Fab (pink), and the micelle
surrounding the MPER and TM domain (gray). The
three possible PGT151 binding sites are labeled as 2.5 nm
interface 1 (unliganded), interface 2, and interface
3. (B) Model of the EnvDCTectodomain, colored as
in (A).The Fab light and heavy chains are colored in
pink and magenta, respectively. Glycans are shown
as spheres, with the gp120 and gp41 glycans shown
in light and dark green, respectively. (C) Simplified
cartoon of gp41. Most of the portion of HR1 that
does not have a regular secondary structure in
SOSIP trimers (residues 548 to 568) is here re-
vealed to be an a-helix. To distinguish this region
from the central HR1 helix (residues 571 to 593), we
call these two helices HR1N and HR1C, respectively.
The complete HR1 spans residues 534 to 593.The
cartoon cylinder and loops are colored according
to the sequence shown at the bottom.
90

N-acetylglucosamine (GlcNAc) moieties are visi- a highly conserved CDRH2 region of the PGT151 proximal gp120 glycans (fig. S11C), suggesting
ble at the majority of glycosylation sites, except family and also extends far enough to make po- that it makes considerably fewer contacts relative
in disordered peptide regions such as V1 and tential contacts with the FP of the adjacent gp41 to N448 or N241, although is positioned to restrict
V4 (fig. S6, B to D, and table S1). The glycans in (Fig. 3, A and C; fig. S8; and fig. S12, A and B). accessibility to the PGT151 epitope.
the PGT151 epitope are ordered, allowing us to The N637 glycan is a tetra-antennary glycan In TZM-bl neutralization assays, the IC50 (half-
resolve four highly branched glycans at N241 (Fig. 3, B and D, and fig. S9), and the core tri- minimal inhibitory concentration) improved slight-
and N448 in gp120 and N611 and N637 in gp41. saccharide interacts along the length of CDRH3 ly when the N241 or N448 glycan was knocked
Gp120 glycans have been studied extensively, (Fig. 3B). The interactions are probably backbone- out in JR-FL pseudovirus (26), corroborating our
but much less is known about gp41 glycans. In mediated, because the PGT151-family CDRH3 se- structural observations. In the same assay, the
our structure, glycans at positions N625 and N616 quences in this stretch are variable (fig. S12, A and presence of either the N611 or the N637 glycan
were only resolved up to the core GlcNAc residues C). The Man(a13) branch projects between the alone in gp41 was sufficient for neutralization,
(table S1). The ordered glycans at positions N611 and heavy chain and light chain, with the LacNAc albeit with decreased potency (26). In JR-FL, the
N637 that interact with PGT151 were built as com- (b14) unit primarily interacting with CDRH1 reduced neutralization potency due to the knock-
plex glycans with N-acetyllactosamine (LacNAc) and the LacNAc(b12) branch interacting with the out of N637 was recovered by a second, simulta-
branching, which is consistent with published gly- C terminus of CDRL2. The GlcNAc(b16)Man(a1 neous knockout of the N448 glycan. This effect is
can array binding data (Fig. 3, A to D, and figs. S8 6) branch interacts with another well-conserved not observed for BG505 pseudovirus, which nat-
and S9) (26). The base GlcNAc residues of the region in CDRL2 (Fig. 3B and fig. S12, A and C). urally lacks the N241 glycan. Furthermore, in
N611 and N637 glycans were core-fucosylated Previous low-resolution modeling and glycan BG505 pseudovirus, the N637A glycan knockout
(Fig. 3, A and B, and figs. S8 and S9), which is knockout neutralization assays allowed us to pre- does not cause as large a decrease in neutraliza-
consistent with glycan array data demonstrating dict that glycans from N262 and N448 in one tion potency as in JR-FL (26). We therefore hypoth-
fucose-dependent binding differences by some gp120 protomer, and from N276 in the adjacent esize that the N448 glycan imposes a steric hurdle
clonal relatives of PGT151 (26) and with binding gp120 protomer on a BG505 trimer, could inter- for PGT151 binding, but only in the presence of
studies (fig. S10). act with the PGT151 Fab (16). In this study, we the highly conserved N241 glycan (Fig. 4A). The
The N611 glycan is minimally a tri-antennary gly- observed a density for a glycan at N241 (glycosyl- N241 glycan probably limits the range of motion
can, with the two LacNAc units of the mannose(a1 ation site ~96% conserved, not present in BG505), of the N448 glycan, which in turn limits access to
6) [Man(a16)] arm resolved. The LacNAc(b16) which, along with the N448 glycan, extends toward the PGT151 epitope, in a similar manner to steric
branch packs against the heavy-chain framework FWRH3 (Fig. 3E and fig. S11, A and B). The N276 restriction of epitopes by glycans that has been
region 3 (FRWH3). The LacNAc(b12) interacts with glycan is the least resolved of the four PGT151- observed with other bnAbs (2729). It has also

1044 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H | R ES E A R C H A R T I C L E S

been suggested that the presence or absence of by the TM, but MPER bound to 10E8 and the ture and emphasizes the flexibility in this region
one glycan can affect the conformational space HR2-MPER interface could be visualized (Fig. (Fig. 5C and fig. S15A).
that can be occupied by adjacent glycans (30). 5B and fig. S15A). The HR2-MPER connectivity 10E8 recognizes Env primarily via CDRH3 con-
This type of mechanism illustrates the complex was substantially different from the crystal struc- tacts with MPER helix 672 to 683, with additional
nature of Env surface accessibility and the dif-
ficulty of determining complete epitopes outside
of high-resolution structures. Lastly, the D2/D3
24 26 HR1N
branch of the very highly conserved N262 oligo- 40
mannose glycan is in close proximity to the FWRL3
of PGT151; this suggests interactions with the D2
arm, although it is unclear whether this glycan is
hindering or enhancing PGT151 binding (fig. S11D).
Trimers of different genotypes may contain
different glycoforms, especially in gp41 (31), and
variation at these sites may be responsible for a
neutralization plateau (26). Our model illustrates
that the PGT151 epitope extends beyond the gp41 Fusion Peptide
glycans and includes gp120 glycans that may also Fig. 2. Conformational changes induced by
contribute to incomplete neutralization. Overall, PGT151 binding. (A) Compared with interface 1 (gray),
PGT151 is highly glycan-dependent, with up to six HR1N in interfaces 2 (blue) and 3 (teal) is shifted
glycans in JR-FL (N241, N262, N276, N448, N611, about 24 outwards toward the Env surface and 26 toward gp120 of the same protomer (yellow surface).
and N637) surrounding the antibody (Fig. 4A). The position of the I559 residue is shown in bright yellow. (B) The FP (teal) is inserted into a hydrophobic
pocket formed by the PGT151 CDR loops (CDRH2, magenta; CDRH3, purple; CDRL3, orange). The hydro-
Mechanism of trimer stabilization phobic aromatic residues in these CDRs are shown as sticks.
Unlike any other bnAb identified so far, including
gp41-gp120 interface antibodies such as 35O22
(fig. S13A) (32), PGT151 stabilizes the metastable
prefusion Env for prolonged periods (16). Con-
tact with both N611 and N637 glycans probably
prevents the HR2 helix from progressing to the 60
postfusion conformation (Fig. 4B). PGT151 and its
clonal relatives all have a long CDRL1 of 16 res-
idues and a conserved N28 (D28 in PGT154 to
-158; fig. S12A) at the CDRL1 tip that interacts
with a conserved region at the end of HR1N (se-
quence conservation: Q551 to Q552, 100%; N553,
58%; N554, 99%). Additionally, CDRH3 is wedged
between the HR1N and the FP-proximal HR1 re-
gion. In this manner, the CDRL1 and CDRH3 cap
the end of a short helix to prevent it from ex-
tending into a longer helix and thereby thwart the
60
transition to a postfusion conformation (Fig. 4C).
PGT151 has an unusual binding stoichiometry of
two Fabs per trimer, even though there is no ob-
vious steric barrier to the binding of the third Fab
(fig. S13B). Rather, binding of two PGT151 anti-
bodies appears to have an allosteric effect, altering
the conformation of the third binding site (Fig. 4D).

10E8-bound conformation of Env


To visualize the MPER, we added MPER-specific
bnAb 10E8 to the complex and analyzed the JR-FL
EnvDCTPGT151 Fab10E8 Fab complex by means 4
of cryo-EM (fig. S14, A and B). Although 10E8 had
some stabilizing effect on the MPER, we still ob-
served flexibility in this region (Fig. 5A and fig.
S14C), which is consistent with a previously solved
crystal structure of the MPER peptide10E8 Fab N-Acetylglucosamine
complex (Fig. 5C) (22). Despite variation in the (GlcNAc)
Mannose (Man)
protein conformation and binding occupancies
Galactose (Gal)
on the trimer, we obtained an 8.8 resolution Fucose (Fuc)
reconstruction from a subset of the JR-FL EnvDCT
PGT151 Fab10E8 Fab complexes (Fig. 5A and fig. Fig. 3. Glycan structures on the Env trimer. (A) The glycan at position N611 makes extensive contacts
S14D). This reconstruction revealed that the with PGT151 Fab.The glycan residues are colored according to the diagram in (C) and (D). (B) As in (A), but
center of the trimer at the base is empty, similar for the N637 glycan. (C) The glycans modeled at N611 and (D) N637. Only sugar moieties resolved in the
to the JR-FL EnvDCTPGT151 structure. Again, current structure are shown. (E) The N241 and N448 glycans (different shades of green) are in close
we did not observe a three-helix bundle formed proximity to the CDRL3 and FWRH3 of PGT151.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1045


R ES E A RC H | R E S EA R C H A R T I C LE S

Fig. 4. The complete PGT151 epitope. (A) A mod-


el of PGT151-glycan interactions. Glycans from up
to four different subunits (two from gp120 and two
from gp41) of two protomers of the trimer can lock
the Fab in its bound form (left). Some of the gly-
cans bind PGT151 with high affinity (black arrows),
but there are numerous steric barriers that need to
be overcome (red arrows). Glycans N241 and N448
probably have an inhibitory effect on PGT151 bind-
ing by influencing the conformations of each other.
The lack of the N241 glycan (right) alleviates steric
pressure by N448 (blue arrow). Different gp120 sub-
units are shown in shades of yellow, gp41 subunits
in shades of blue, Fab light and heavy chains in pink
and magenta, gp120 glycans in green, and gp41
glycans in dark green. (B) When PGT151 is bound
to glycans at N611 and N637, HR2 is locked in a
bent conformation and therefore cannot undergo
conformational changes into the extended post-
fusion form. Colors are as in (A). (C) PGT151 CDRL1
(yellow) and the N637 glycan fucose (Fuc, dark
green) interact with a glutamic acid and asparagine-
rich region of HR1N on the adjacent gp41 (Q551 to
N554). The CDRH3 (purple) inserts between HR1N
and the FP, and these interactions cap HR1N to lock
gp41 in the prefusion conformation.The interacting
residues in the Fab and HR1N are shown in orange.
Only the core Man(Fuc)GlcNAc2 residues are shown
for the N637 glycan for clarity. In (B) and (C), red
arrows with crosses show that gp41 regions are blocked
from undergoing conformational changes. (D) A mea-
surement of the inter-gp41 distances in PGT151-bound
JR-FL (left) compared with the unliganded BG505
trimer (right).The unliganded BG505 trimer measures
~37 between the Ca of N628 and N637 (residues
shown in yellow) on the adjacent protomer, whereas
Interface 1
the distance between the same two residues mea-
44.2 36.6
sures ~35 at the PGT151-liganded interfaces. How-
ever, the inter-gp41 distance at interface 1 (~44 ) is
~9 farther apart in comparison with the liganded
interfaces, indicating that the trimer becomes asym-
metric in the PGT151-bound form. The three shades
of blue differentiate the three gp41 monomers.

36.6 36.6

Interface 2
34.7

interactions between FWRH3 and the gp41- fusion intermediates (3739), but although 10E8 embedded or during some transient exposure when
gp120 interface (Fig. 6, A and B). Moreover, N88 potently neutralizes virus after CD4 attachment the trimer is lifted off the membrane surface.
and N625 glycans are positioned to sterically (33), it also neutralizes the prefusion trimer (22, 32). In the negative-stain reconstruction of 10E8 in
block 10E8 binding. The N88 glycan, built as a Other studies have suggested that the CD4-bound complex with MPER-containing BG505 SOSIP.683
ManGlcNAc2 in the 4.2 model, clashes with form of Env may be lifted from the membrane and (fig. S15, B and C), the Fab does not induce an
FWRH1, and a complete glycan at N625 would would therefore provide greater accessibility to opening of the trimer, as it does when soluble
also clash with the 10E8 Fab constant region membrane proximal epitopes (32). CD4 binding CD4 or CD4-induced antibodies are bound (2).
(Fig. 6, A and B). The N625Q mutation increases opens Env via a rotation in gp120, which moves Although the 2D class averages of 10E8-bound
maximum neutralization of JR2, and glycoforms the N88 glycan farther away from the base of the BG505 SOSIP.683 trimers display distinguish-
affect the degree of neutralization plateauing (33). trimer (Fig. 6C). Nuclear magnetic resonance struc- able Fab densities (fig. S15B), the particles refine
In the trimer structure, both the N88 and N625 tures of MPER peptides, as well as our EnvDCT poorly in the 3D reconstruction, and the 10E8 Fab
glycans are accessible for glycan processing and model (fig. S16), suggest that the ground-state binds at a different angle than in EnvDCT (fig.
have been predicted to be complex (3436). Thus, MPER epitope is embedded in the membrane (15). S15C), which again is consistent with MPER flex-
these glycans could restrict access to the MPER Thus, although MPER is difficult to access, MPER ibility. Without the TM, however, the Fab ap-
epitope. Most MPER antibodies bind to gp41 antibodies can either bind MPER while membrane- proaches from the bottom of the trimer, which

1046 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H | R ES E A R C H A R T I C L E S

Fig. 5. JR-FL EnvDCT bound to PGT151 and 10E8.


(A) Cryo-EM reconstruction of JR-FL EnvDCT (gray)
in complex with both PGT151 (pink) and the MPER-
binding antibody 10E8 (blue) at 8.8 resolution
(left).The JR-FL EnvDCTPGT151 reconstruction (low
passfiltered to 8.8 ) is shown on the right for
comparison.The reconstructions indicate that when
10E8 is bound, the trimer is lifted off the membrane
(red arrow), suggesting a conformational change in
the MPER and TM. (B) The Env HR2-MPER connec-
tivity in the 10E8-bound form is modeled into the
EM density. (C) A comparison of the position of
residues 659 to 670 in the two asymmetric units of
the 10E8-bound MPER peptide x-ray model (dark
and light gray), superimposed on the complete Env
model (teal), in which the primary MPER epitope
(residues 671 to 685) is shown in yellow. This
N-terminal segment exhibits different conforma-
tions in the two asymmetric units. D664 is colored
in red as a point of reference. 10E8 is shown as the
white surface.

Fig. 6. 10E8 contact analysis in the context of


the Env ectodomain. (A and B) A model of the
10E8 epitope in the context of the intact Env trimer.
The Fab constant region (dark and light gray
surface) and the nearby Env gp120 (white) are
also shown. On the left in (A) and the middle-
lower right in (B) is the gp41 in which the 10E8 Fab
makes primary interactions with MPER residues
671 to 685 (yellow). Additional contacts could be
made with the HR2 and C-terminal region of the FP
in the adjacent gp41 (teal), as well as regions in
gp120 (white). These additional contacts to Env
within a 4 radius of 10E8 are colored red. Many
of these interactions are probably mediated by
FWRH3 (orange). The model also demonstrates
that the N88 (A) and N625 (B) glycans could
sterically obstruct 10E8 binding (red circles). The
glycans modeled here are ManGlcNAc2 for N88 and
GlcNAc for N625 (table S1), but they are expected
to be larger in native Env. (C) Glycans at N88 and
N625 sterically hinder 10E8 binding to the trimer
(left, red arrows). Binding of 10E8 (left, purple) or
CD4 (center right) lifts the MPER up from the mem-
brane relative to the ground state (center left). In
the CD4-bound conformation, the opening of the
trimers results in rotation of the gp120s, moving N88 away from the 10E8 binding site and relieving some steric hindrance (right, blue arrow). Colors are as in
Fig. 4A. The red-orange areas in the cartoons on the right indicate the CD4-bound state.

would be impossible in the context of the viral the HR2 C terminus and the micelle in our 10E8- plex with PGT151 Fab at 4.2 resolution, demon-
membrane. Cross-linking the ectodomain of full- bound structure (Fig. 5A) probably represents the strating the structure not only of wild-type Env
length Env also has no effect on MPER-bnAb binding, optimal display of the membrane-anchored MPER but also of a type I viral fusion protein with an
despite the reduced CD4 binding (40). Together, epitope with all additional constraints in place. intact TM, which unexpectedly was found to be
these observations indicate that (i) MPER is The epitope of 10E8, and perhaps of other MPER flexible. The PGT151 epitope includes the FP and
largely inaccessible on the viral membrane, with antibodies, is therefore more complex than pre- an extensive network of primary and secondary
the membrane imposing a steric hurdle for the viously thought, involving elements from multi- glycan interactions that stabilize the prefusion
Fab approach angle, and (ii) 10E8 binds a tran- ple gp41 protomers, as well as from gp120. conformation of the Env trimer. The MPER ap-
siently exposed MPER, which is not necessariliy pears to be sequestered in the detergent micelle
the CD4-bound conformation of Env nor a fusion- Discussion in the unliganded state, but it is outside the mi-
transition conformation, in contrast with previous Here we present the cryo-EM reconstruction of a celle in the 10E8-bound structure, suggesting a
observations (41). The appearance of a gap between cleaved wild-type JR-FL EnvDCT trimer in com- dynamic topology. This propertyin combination

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1047


R ES E A RC H | R E S EA R C H A R T I C LE S

with steric constraints from gp120, gp41, and glycans Bank with accession codes PDB-5FUU, EMD-3308, EMD-3309, and conducted at the cryogenic electron microscopy facility at The
at N88 and N625effectively shields the con- EMD-3312. This work was supported by the NIH (grant UM1 Scripps Research Institute. This is manuscript number 29175 from
AI100663), the International AIDS Vaccine Initiative (IAVI) the Scripps Research Institute.
served MPER. Thus far, MPER peptide vaccines, Neutralizing Antibody Consortium through the Collaboration for
though immunogenic, produce nonneutralizing AIDS Vaccine Discovery (grants OPP1084519 and OPP1115782), SUPPLEMENTARY MATERIALS
antibodies, probably because of the lack of the and the California HIV/AIDS Research Program Dissertation Award
www.sciencemag.org/content/351/6277/1043/suppl/DC1
additional constraints provided by the trimer and (to J.H.L.). This work was partially funded by IAVI with the
Materials and Methods
generous support of the U.S. Agency for International Development
membrane (42, 43); our model suggests that the (USAID), the Ministry of Foreign Affairs of the Netherlands, and the
Figs. S1 to S16
minimalistic MPER epitope peptide presentation Table S1
Bill and Melinda Gates Foundation; a full list of IAVI donors is
References (4464)
may not be the most ideal strategy to elicit MPER available at www.iavi.org. The contents of this manuscript are the
bnAbs. Overall, our data indicate that Env is a responsibility of the authors and do not necessarily reflect the 13 August 2015; accepted 3 February 2016
views of USAID or the U.S. government. The EM work was 10.1126/science.aad2450
pliable structure in which several of the protein-
protein interfaces can be remodeled, making it a
difficult moving target for the immune system.

RE FE RENCES AND N OT ES IMMUNOLOGY


1. J. Liu, A. Bartesaghi, M. J. Borgnia, G. Sapiro, S. Subramaniam,
Nature 455, 109113 (2008).
2. E. E. Tran et al., PLOS Pathog. 8, e1002797 (2012).
3. J. P. Julien et al., Proc. Natl. Acad. Sci. U.S.A. 110, 43514356
(2013).
Visualizing antibody affinity
4. J. P. Julien et al., Science 342, 14771483 (2013).
5. D. Lyumkis et al., Science 342, 14841490 (2013).
6. M. Pancera et al., Nature 514, 455461 (2014).
maturation in germinal centers
7. Y. Do Kwon et al., Nat. Struct. Mol. Biol. 22, 522531 (2015).
8. F. Garces et al., Immunity 43, 10531063 (2015).
Jeroen M. J. Tas,1* Luka Mesin,1* Giulia Pasqual,1 Sasha Targ,1 Johanne T. Jacobsen,1
9. J. H. Lee, N. de Val, D. Lyumkis, A. B. Ward, Structure 23, Yasuko M. Mano,1 Casie S. Chen,1 Jean-Claude Weill,2 Claude-Agns Reynaud,2

Downloaded from on March 3, 2016


19431951 (2015). Edward P. Browne,3,4 Michael Meyer-Hermann,5,6 Gabriel D. Victora1
10. R. W. Sanders et al., PLOS Pathog. 9, e1003618 (2013).
11. R. W. Sanders et al., Science 349, aac4223 (2015).
12. P. Dosenovic et al., Cell 161, 15051515 (2015). Antibodies somatically mutate to attain high affinity in germinal centers (GCs). There,
13. J. Chen et al., Science 349, 191195 (2015). competition between B cell clones and among somatic mutants of each clone drives an
14. L. G. Abrahamyan et al., J. Virol. 79, 106115 (2005). increase in average affinity across the population. The extent to which higher-affinity cells
15. Z. Y. Sun et al., J. Mol. Biol. 426, 10951108 (2014).
eliminating competitors restricts clonal diversity is unknown. By combining multiphoton
16. C. Blattner et al., Immunity 40, 669680 (2014).
17. J. B. Munro et al., Science 346, 759763 (2014).
microscopy and sequencing, we show that tens to hundreds of distinct B cell clones seed
18. P. Pugach et al., J. Virol. 89, 33803395 (2015). each GC and that GCs lose clonal diversity at widely disparate rates. Furthermore, efficient
19. J. P. Julien et al., Proc. Natl. Acad. Sci. U.S.A. 112, 1194711952 affinity maturation can occur in the absence of homogenizing selection, ensuring that
(2015). many clones can mature in parallel within the same GC. Our findings have implications for
20. R. W. Sanders et al., J. Virol. 76, 88758889 (2002).
development of vaccines in which antibodies with nonimmunodominant specificities must
21. M. Guttman et al., Structure 22, 974984 (2014).
22. J. Huang et al., Nature 491, 406412 (2012). be elicited, as is the case for HIV-1 and influenza.
23. R. Pejchal et al., J. Virol. 83, 84518462 (2009).

T
24. S. Bryson, J. P. Julien, R. C. Hynes, E. F. Pai, J. Virol. 83,
1186211875 (2009).
he average affinity of specific antibodies in- containing up to a few thousand B cells that
25. I. A. Wilson, J. J. Skehel, D. C. Wiley, Nature 289, 366373 creases dramatically over the course of an emerge in multiple copies within secondary
(1981). immune response (1, 2). This phenomenon lymphoid organs upon infection or immunization
26. E. Falkowska et al., Immunity 40, 657668 (2014). is known as affinity maturation and is the (37). Each GC is traditionally viewed as arising
27. F. Garces et al., Cell 159, 6979 (2014).
28. D. Sok et al., Sci. Transl. Med. 6, 236ra63 (2014).
result of a Darwinian process that alternates from a handful of independent founder clones,
29. J. H. Lee et al., Nat. Commun. 6, 8167 (2015). stochastic somatic hypermutation (SHM) of immu- many of which are lost as affinity-based compe-
30. R. Derking et al., PLOS Pathog. 11, e1004767 (2015). noglobulin (Ig) genes with the selection and clonal tition leads to progressive focusing of the reper-
31. L. K. Pritchard et al., J. Virol. 89, 69526959 (2015). expansion of B cells that have acquired affinity- toire on the most successful progeny (which we
32. J. Huang et al., Nature 515, 138142 (2014).
33. A. S. Kim, D. P. Leaman, M. B. Zwick, PLOS Pathog. 10,
enhancing mutations (2). Successive iterations of refer to as homogenizing selection) (8). Although
e1004271 (2014). this cycle drive an increase in the overall affinity it is inducible under experimental settings (9), the
34. J. M. Cutalo, L. J. Deterding, K. B. Tomer, J. Am. Soc. of antibodies over time, which is essential for extent to which homogenizing selection takes
Mass Spectrom. 15, 15451555 (2004). their effectiveness in curtailing and preventing place during normal immunization or infection
35. E. P. Go et al., J. Virol. 89, 82458257 (2015).
36. E. P. Go et al., J. Virol. 85, 82708284 (2011). infection. has been difficult to determine, mostly due to
37. E. T. Crooks et al., Hum. Antibodies 14, 101113 (2005). Affinity maturation takes place in germinal technical limitations in the methods used to
38. R. M. Cardoso et al., Immunity 22, 163173 (2005). centers (GCs), defined microanatomical clusters measure clonal diversity within individual GCs
39. G. Frey et al., Proc. Natl. Acad. Sci. U.S.A. 105, 37393744
(2008).
(1014).
40. D. P. Leaman, J. H. Lee, A. B. Ward, M. B. Zwick, J. Virol. 89,
67256745 (2015). 1
Whitehead Institute for Biomedical Research, Cambridge, Measuring clonal diversity in GCs using
41. J. Chen et al., J. Virol. 88, 12491258 (2014). MA 02142, USA. 2Institut Necker-Enfants Malades, INSERM multiphoton microscopy
42. J. Guenaga et al., PLOS ONE 6, e16074 (2011). U1151-CNRS UMR 8253, Sorbonne Paris Cit, Universit
43. G. Ofek et al., Proc. Natl. Acad. Sci. U.S.A. 107, 1788017887 Paris Descartes, Facult de Mdecine-Site Broussais, 75014
As an approach to estimating GC clonal diversity
(2010). Paris, France. 3Koch Institute for Integrative Cancer by imaging, we made use of a brainbow allele
Research, Massachusetts Institute of Technology (MIT), for multicolor fate mapping to permanently tag
ACKN OW LEDG MEN TS Cambridge, MA 02142, USA. 4Broad Institute of Harvard and individual B cells and their progeny with differ-
We thank C. Blattner for the PGT151 and JR-FL EnvDCT plasmids, MIT, Cambridge, MA 02142, USA. 5Department of Systems
J. Torres and N. Overney for helping with plasmid preparations, Immunology and Braunschweig Integrated Centre of
ent combinations of fluorescent proteins (15). This
T. Nieusma for technical assistance, A. Sarkar and L. Kong for Systems Biology, Helmholtz Centre for Infection Research, method, as implemented in the Rosa26Confetti
advice on glycan modeling, and I. A. Wilson for helpful comments Inhoffenstrabe7, 38124 Braunschweig, Germany. 6Institute allele (16), relies on stochastic Cre-mediated re-
and discussion. The data from this study are tabulated in the main for Biochemistry, Biotechnology and Bioinformatics, combination to commit cells to expression of one
paper and in the supplementary materials. The cryo-EM Technische Universitt Braunschweig, Braunschweig,
reconstructions of JR-FL EnvDCTPGT151 and JR-FL EnvDCT Germany.
of four possible fluorescent proteins, generating
PGT151 Fab10E8 Fab and the model of JR-FL EnvDCTPGT151 *These authors contributed equally to this work. Corresponding 10 different color combinations when two alleles
have been submitted to the PDB and the Electron Microscopy Data author. E-mail: victora@wi.mit.edu are recombined in homozygous mice (fig. S1, A

1048 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H | R ES E A R C H A R T I C L E S

A pLN, unimmunized B pLN, 20d post-imm. C mLN, unimmunized E Mx1-Confetti G PA/sequencing


100 100

% Dominant clone
% Dominant color
75 75

50 50

25 25

0 0
D Skin-draining LN, 15d post-immunization
F PA/sequencing (PAGFP-tg)
CD35/xfer B CD35/xfer B CD35/xfer B
Pre-PA Post-PA

Confetti Confetti Confetti Phycoerythrin (FDC) Photoactivated cells


Sort photoactivated
GC B cells
(CD19+CD38-
FAShiIgD-) from
each fragment,
sequence VDJH
to determine
clonal identity

Fig. 1. Visualizing clonal expansions in GCs using a brainbow allele. (A) tification of data as in (D). Each symbol represents one GC. Graph shows per-
pLN of an unimmunized Rosa26Confetti/Confetti.Mx1-Cre mouse, imaged by centage of cells expressing the most abundant color combination. Data pooled
multiphoton microscopy. (B) pLN of a mouse immunized 20 days previously from four mice, two independent experiments. (F) Quantifying GC clonality by
with 10 mg CGG in alum in the hind footpad. The image shows a cluster of photoactivation (PA). Photoactivatable-GFP-transgenic mice were immunized
similar-colored cells corresponding to a GC (dashed line), as evidenced by the in the footpad with 10 mg CGG in alum and imaged 15 days later. FDC networks
presence of tingible body macrophages (arrowheads). (C) Single-colored GC were labeled with phycoerythrin immune complexes. (Top) Images of a single
(dashed line) in a mesenteric lymph node (mLN) from an unimmunized mouse. GC within a pLN, before and after photoactivation. Scale bar, 100 mm. (Bottom)
(D) GCs in draining LN of mice immunized subcutaneously with CGG in alum Single pLN containing two photoactivated GCs (arrowheads) dissected into
15 days before imaging. The location of the GC dark zone (dashed line) was two fragments, each of which is separately processed for sorting of PA+ GC
determined by injection of labeled antibody to CD35 and surface-labeled nave B cells and Igh sequencing. Scale bar, 500 mm. (G) Quantification of clonal
B cells (top panels; fluorescence is from Alexa 633 label). For each GC, Confetti dominance in multiple GCs. Data obtained as in (F), with clonal identity as-
colors (bottom panels) were imaged independently and used for quantifi- signed based on Igh sequence. Each symbol represents one GC, with two GCs
cation. Confetti colors are as shown in fig. S1B. GC identity is confirmed by sequenced per LN (full clonality charts in fig. S2). The percentage of cells
presence of tingible-body macrophages (arrowheads). Scale bars, 100 mm. belonging to the most abundant clone is given. Data are from five mice, three
Second-harmonic generation from collagen fibers is shown in blue. (E) Quan- independent experiments.

and B). Because recombination stops upon cessa- in Peyers patches (fig. S1C) of unimmunized mice, also revealed the frequent presence of clones that
tion of Cre activity, selective proliferation of an where GCs form spontaneously in response to in- were shared between two individual GCs in the
individual GC B cell would lead to the appear- testinal microbiota. Thus, clonal expansions within same LN, indicative of synchronous origin (colored
ance of clusters of daughter cells of the same GCs can be readily detected by multicolor fate slices in fig. S2). We conclude that GCs display
color. mapping, even in cases where the driving anti- variable degrees of clonal dominance, even when
We generated Rosa26Confetti/Confetti mice carry- gen is unknown. induced synchronously by immunization. Although
ing the Mx1-Cre transgene, which triggers Confetti Quantification of cell colors in GC dark zones predominantly monoclonal GCs do exist, these are
recombination spontaneously during early life at 15 days after immunization (Fig. 1D) revealed a relatively rare at the time point assayed.
(17). Multiphoton imaging of popliteal lymph wide range of color dominances (from 20.2 to
nodes (pLN) in nave Mx1-Confetti mice showed 89.7%; median, 44.0%) and a relative paucity of Early GCs are highly diverse
that, in the absence of immunization, cells of dif- predominantly single-colored GCs [6 of 40 (15%) The diversity of clonal dominance levels among
ferent colors are intermixed, as expected from the GCs with color dominance > 70%] (Fig. 1E). To mature GCs led us to question the generalizabil-
polyclonal nature and migratory behavior of nave validate these estimates, we used in situ photo- ity of reports proposing that GCs are seeded
lymphocytes (Fig. 1A and movie S1). In contrast, activation followed by fluorescence-activated cell pauciclonally (11, 12, 19). To address this, we gen-
subcutaneous immunization with alum-adsorbed sorting (9, 18) to obtain single B cells from indi- erated mice expressing the photoactivatable-GFP
chicken gamma globulin (CGG) induced the ap- vidual GCs, the clonal identity of which we then (green fluorescent protein) transgene along with
pearance in the draining lymph node (LN) of determined by Igh mRNA sequencing (Fig. 1F). Cre recombinase driven by the endogenous Aicda
clusters of cells of the same color, suggestive of This approach again showed varying degrees of locus (AicdaCre) (20), expressed in activated B cells,
monoclonal expansion (Fig. 1B and movies S2 clonal dominance among individual GCs [3 of 12 and a Rosa26lox-stop-lox-tdTomato reporter. At 6 days
and S3). Such clusters could also be found in (25%) GCs with dominance > 70%; range, 22.5% after immunization with CGG in alum, early
mesenteric LNs (Fig. 1C), and even occasionally to 87.5%; median, 43.0%] (Fig. 1G and fig. S2). It GCs were identifiable in these mice as clusters of

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1049


R ES E A RC H | R E S EA R C H A R T I C LE S

AicdaCre/+.Rosa26Stop-tdTomato/+ Observed
6 days after immunization with CGG-alum 3 Chao1 estimator (95%CI)
10
Photoactivation Microdissection ACE estimator

Distinct clones/GC
Pre Post Pre Post
10 2
+
Sort tdTomato
photoactivated
GC B cells from 10 1
each fragment,
sequence VDJH
10 0
GC: 1 2 1 2 1 2 1 2
(no. cells) (40) (41) (34) (35) (69) (75) (44) (77)
pLN #1 pLN #2 pLN #3 pLN #4 Mouse: 1 2 3 4

103
GC1 GC1 GC1 GC1 *
(27/40) (25/34) (40/67) (36/44)

Chao1 estimator
(normalized)
102

101

GC2 GC2 GC2 GC2


(35/41) (23/35) (36/75) (46/77) 100

VA

VA
H
G

O
C

P-
N
Fig. 2. Clonal diversity in early GCs. (A) Photoactivation of single early different mice in three independent experiments. (C) Estimation of total clonal
GC clusters. Photo activatable-GFP-transgenic.AicdaCre.Rosa26lox-stop-lox-tdTomato richness in individual GCs using the Chao1 and ACE estimators. Graphs show
mice were immunized in the footpad with 10 mg CGG in alum and imaged 6 days observed clonal richness [from (B)] and total richness according to the
later. FDC networks were marked by injection of labeled antibody to CD35. Left indicated estimator. (D) Estimated clonal richness (Chao1) in GCs elicited by
panels show images of a single tdTomato+ cluster (arrowheads) within a pLN various antigens. Mice were immunized with 10 mg of the indicated antigen and
before and after photoactivation. Scale bar, 200 mm. Right panels show then imaged/photoactivated as in (A). Each symbol represents one GC; bar
dissection of a single pLN with two photoactivated GCs (arrowheads) into two indicates median. CGG, chicken gamma globulin; OVA, chicken ovalbumin; HA,
fragments, each of which is separately processed for cell sorting. Scale bar, influenza hemagglutinin (H3); NP-OVA, 4-hydroxy-3-nitrophenylacetyl-OVA.
500 mm. (B) Pie charts showing clonal diversity in early GCs. Each slice rep- For comparison purposes, estimates are normalized by interpolation to the size
resents one clone. Colored slices represent clones that were found in both GCs of the smallest sample (34 cells). Numbers for CGG GCs in (D) differ from
(upper and lower pie charts) from the same pLN. Numbers in the center of those in (C) due to normalization. Further details are in fig. S3. *P < 0.05,
each chart are number of clones observed/total number of cells sequenced. Kruskall-Wallis test with Dunns post-test. All other comparisons were not
Clonal identity was assigned based on Igh sequence. Pairs are from four significant.

tdTomato+ cells within follicular dendritic cell bor]. Similar experiments using other antigens the early GC or by strong expansion of single SHM
(FDC) networks. We photoactivated two such showed that, although the extent of early GC variants over cells of the same and of different
clusters per LN and sorted photoactivated GC diversity can vary depending on the immunizing clones. To investigate these dynamics, we generated
B cells from each cluster for Igh sequencing (Fig. antigen, high diversity is not a unique feature of mice in which recombination of Rosa26Confetti was
2A and fig. S3A). Early GCs were highly and uni- the response to CGG (Fig. 2D and fig. S3, C to E). driven by the tamoxifen-inducible AicdaCreERT2
formly polyclonal, with 23 to 46 unique variable/ Diversity was also not the result of nonspecific allele, specifically expressed in activated B cells
diversity/joining (VDJ) rearrangements detected recruitment of B cells to early GCs (figs. S3C (22). We thus delayed recombination until after
when sequencing 34 to 77 single cells per GC (Fig. and S4). We conclude that early GCs are highly GC formation, allowing different members of the
2B). Extrapolation of these numbers using the diverse, containing tens to hundreds of distinct same clone to express different color combina-
Chao1 and abundance coverage-based estimator clones, depending on the antigen used for im- tions, each giving origin to an independent early
(ACE) estimators of species richness (21) gave munization. The progression from uniform di- GC lineage. In this system, GCs dominated by
rough estimates of GC clonality that ranged versity (Fig. 2) to variable degrees of clonal the descendants of a single SHM variant would
from ~50 to ~200 clones per GC (median Chao1, dominance (Fig. 1) in the CGG response implies be identifiable as having resolved to a single
102 clones) for the four pairs of GCs analyzed that individual GCs display different rates of dominant color (Fig. 3A).
(Fig. 2C). Overall mutation and class-switching homogenizing selection acting subsequently to Administration of tamoxifen to Rosa26Confetti/Confetti.
levels were low at this time point (fig. S3B), con- this coalescence. AicdaCreERT2/+ (AID-Confetti) mice at 5 days after
firming that these were indeed early GCs. We CGG immunization triggered recombination, of-
were again able to identify B cell clones that were Extent of homogenizing selection in ten of a single Confetti allele, in ~40 to 50% of GC
shared between two GCs in the same pLN (col- individual GCs B cells (fig. S5, A and B) by the end of the 3- to
ored slices, Fig. 2B) [mean 15.8% (SD 6.4) of clones Clonal dominance in GCs may arise by parallel se- 5-day period during which tamoxifen is active
found in one GC were also found in its neigh- lection of multiple members of a clone present in (23). Recombination thus ends near the onset

1050 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H | R ES E A R C H A R T I C L E S

Before After After Fig. 3. Kinetics of color dominance in


CGG-alum Tmx Image
tamoxifen tamoxifen resolution individual GCs. (A) Graphic representation
of the experimental protocol. AID-Confetti
Day: -5 0 3 5 7 9 11 15...
mice were immunized in the footpad with
Tmx activity 10 mg CGG in alum, treated with tamoxifen
(tmx) 5 days later, and imaged at the
AicdaCreERT2/+.Rosa26Confetti/Confetti indicated time points. Tamoxifen triggers
Day 3 post-tmx (8 post-CGG) Day 11 post-tmx (16 post-CGG) recombination of one or both Confetti
alleles in individual GC B cells, independent
i. (0.14) i. (0.71) i. of clonal origin. (B to E) Whole lymph node
(large panel; scale bar, 500 mm) and higher-
i. magnification images (side panels; scale
bar, 100 mm) showing GCs at different times
after tamoxifen administration. Cell colors
as in fig. S1B. Second-harmonic generation

ii. (0.09) ii. (0.19) from collagen fibers is shown in blue.


Numbers in parentheses are the normalized
ii.
ii. dominance score (NDS). [(B) to (D)] AID-
Confetti mice imaged at the time points
indicated. (E) Wild-type (WT) recipients of
1 to 2 104 adoptively transferred AID-
Confetti-B1-8 Igl+ B cells, immunized with
Day 15 post-tmx (20 post-CGG) i. NP-OVA as in (A) and imaged 11 days post-
Summary, days 3-15
** tmx. Higher-magnification panels show
i. (0.79) iii. (0.19) 1.00 independently acquired images of the GCs
i. **
indicated in the overview panel. (F) Quanti-
0.75
ii. fication of data as shown in (B) to (E). Bars
NDS

0.50 represent the median. (G) Divergence index


0.25 for AID-Confetti and AID-Confetti-B1-8
ii. (0.52) iii. mice at days 3 and 15 post-tmx. Bars
0.00 represent the median. A green dotted line is
iv. iv. (0.32) 3 5 7 9 11 15
Days post tamoxifen placed at the median of the day 15 AID-
Confetti-B1-8 data for reference. (H) Quan-
AID-Confetti AID-Confetti-B1-8
tification of GC selection in AID-Confetti
Divergence Friend mice infected with FV as detailed in fig. S7.
IghB1-8hi/+.AicdaCreERT2/CreERT2.Rosa26Confetti/Confetti virus Graph shows NDS and divergence index at
index
Day 11 post-tmx (16 post-NP-OVA) (d10 post-tmx) day 10 post-tmx (day 30 after infection).
**** Bars represent the median. For panels (F) to
i. (0.16) iii. (0.31) *
1.25 1.25
i. (H), each symbol represents one GC. Data
****
1.00 1.00 are pooled from 2 to 6 replicate
experiments. For NDS quantification, we
0.75 0.75 exclude GCs with density of fluorescent cells
iii.
ii. 0.50 0.50 below 0.4 cells per 100 mm2 (equivalent to
~40% of cells having recombined a Confetti
ii. (0.23) iv. 0.25 0.25
iv. (0.27) allele; see the supplementary text). *P <
0.00 0.00 0.05; **P < 0.01; ****P < 0.0001,
Mann-Whitney U test.
ce
15

15
3

en
D

Days post-tmx
N

rg
ive
D

of T celldriven selection, at the beginning of inance score (NDS) for each GC, which approxi- GCs at days 11 and 15 after tamoxifen remained
the second week of the response (24) (Fig. 3A). mates (25) the fraction of cells belonging to a single close to day 3 levels. Evaluation of color distribu-
GCs analyzed 3 days after tamoxifen treatment lineage originating at the time of recombination tion in exceptional LNs in which GCs could still
(8 days after immunization) displayed roughly (fig. S5D). Kinetic analysis was extended to 15 days be found at very late time points showed that low-
equal proportions of CFP+, YFP+, and RFP+ after tamoxifen treatment (20 days after immuni- dominance GCs were present up to the very end
(cyan, yellow, and red fluorescent protein, re- zation), near the end of the CGG-aluminduced of the response (fig. S6, A to C). Therefore, GC
spectively) cells and a lower proportion of GFP+ GC response in most LNs. Median dominance in- selection does not inevitably lead to high color
and doubly recombined cells (Fig. 3B), as expected creased monotonically, from 0.16 at day 3 to 0.39 dominance within the lifetime of the response to
from the original description of the Confetti al- at day 15 after tamoxifen. GCs with NDS of 0.5 or CGG-alum.
lele (16). Only residual, if any, GCs containing re- more (corresponding to ~50% of all GC cells dis- To determine the extent to which progression
combined cells could be detected in unimmunized playing the same color combination) appeared as of color dominance depends on affinity-based
or alum-only control mice (fig. S5C). early as day 9 after tamoxifen, and a single color competition, we generated an adoptive transfer
Color distribution changed over time, with pro- could account for close to 80% of GC B cells at model in which wild-type hosts received B cells
gressive emergence of GCs dominated by a single day 11 and virtually all cells in a GC 4 days later from AicdaCreERT2/CreERT2.Rosa26Confetti/Confetti.
color/combination (Fig. 3, B to G). To quantify (Fig. 3F). However, progression among GCs varied IghB1-8hi/+ (AID-Confetti-B1-8) donors (in which
color dominance, we calculated a normalized dom- markedly, as dominance in the lowest-scoring Igl+ B cells have uniformly high affinity to the

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1051


R ES E A RC H | R E S EA R C H A R T I C LE S

Fig. 4. Clonal relationships 200m pLN section Ig sequencing Legend:


among cells obtained from
IgG1 cell Unmutated RFP
GCs with high or low color
IgM cell ancestor YFP
dominance. (A) Method used
to obtain Ig sequences. (B) Igh Single-cell Putative precursor 3 = CFP
index sorting CFP/RFP
sequence relationship among n n identical cells
B cells from two pairs of 1 mutation
individual GCs from two pLNs Clonal
4 mutations burst
of different mice, obtained n n mutations (n >4)
10 days after tamoxifen Vibratome Manual
administration (15 days after sectioning dissection
1.2U 1.4U
immunization), as described
in Fig. 3. Each panel GC1 (0.71) 1.1U
GC2 (0.22) 3
2 4
contains: (Top left) Multi- 3 5
Lymph node #1

1.3U 2
photon image (scale bar, 15 2
100 mm; cell colors as in fig. 1.1M
3
6 1.4M 2
S1B; second-harmonic gener- 3 5
ation from collagen fibers is 7 7
2 6 1.2M
shown in blue; number in 3
6
parentheses is the normalized 2
dominance score (NDS).
(Bottom left) Clonal
52 9 5
5
13 66 2 3
6
distribution pie chart (with
1.3M
clones represented in gray-
scale in the inner ring and 2.1U 2.1M6
Confetti colors in the outer GC1 (0.86) GC2 (0.31) 3
ring; number of cells 3 8
2.1M3
sequenced is indicated in the
Lymph node #2

2.1M1 5 7
center). (Right) Trees repre- 8 2 3 5
2 17 3 2 2 3
senting the phylogeny of Ig 2 3 2 6
heavy-chain V-segment 3 2 2 2.1M4
7
sequences within each clone 5
2
(symbols according to the 2 2
2
2
legend in the top right corner).
Dashed lines within phyloge- 70 63 2.1M7
2.1M2
nies indicate multiple variants 2.1M5
distanced the same number
of mutations from the origi-
nating node. IDs of variants for which affinity was measured in Fig. 5 are indicated by black lines. For each LN, GC1 and GC2 were considered as displaying high and
low color dominance, respectively.

NP hapten and do not undergo affinity matura- bution (at day 15 after tamoxifen, 25.7% of AID- by viral infection, and thus likely represents an
tion due to lack of a functional Aicda allele). Confetti GCs scored below the median of the AID- intrinsic property of GC evolution.
4-Hydroxy-3-nitrophenylacetyl-ovalbumin (NP-OVA) Confetti-B1-8 distribution) (Fig. 3G). We conclude
induced GCs in recipient mice approached high that individual GCs are highly heterogeneous Homogeneous GCs are the product of
color dominance only rarely and with much de- with respect to selection: Although a fraction of clonal bursts
layed kinetics (Fig. 3, E and F, and fig. S6D). these structures become heavily dominated by To investigate the underlying clonal structure of
Therefore, differences in affinity between competing cells of one color in a matter of dayssuggesting the variation in homogenizing selection rates
B cells likely contribute toward the rapid rise in strong expansion of the descendants of a single among GCs, we determined the Igh gene sequences
dominance observed in a fraction of GCs. To SHM variant arising at or after the onset of GC of fluorescent B cells isolated from GCs with dif-
quantify the effect of neutral competition selectionothers deviate from baseline color dis- ferent degrees of color dominance. To achieve this,
defined as the change in lineage abundances over tribution at levels that do not exceed those attained we dissected pLNs from immunized AID-Confetti
time due to stochastic factors, in the absence of by neutral competition. mice into fragments containing single GCs using
differences in affinityin AID-Confetti-B1-8 GCs, We extended our findings to a different anti- vibratome sectioning guided by multiphoton mi-
we calculated a divergence index, which computes genic system by measuring clonal dominance in croscopy (Fig. 4A and fig. S8). From each LN, we
the difference between the expected and observed GCs elicited by infection with Friend retrovirus sorted cells from one high-dominance GC and from
proportions of all 10 colors in each GC (fig. S5E) (FV) (26). We treated infected AID-Confetti mice a neighboring low-dominance GC for Ig sequencing.
and is thus more sensitive to small changes in with tamoxifen at 20 days after infection (an We obtained Igh sequences from 52 to 74 single
color abundance than the NDS. This metric showed early time point in the delayed GC response to cells per GC from three pairs of pLN GCs harvested
marked divergence from baseline distribution in FV) and imaged spleen slices 10 days later (fig. 10 days after tamoxifen treatment (15 days after
this population (Fig. 3G and fig. S6D), highlight- S7, A and B). A wide range of NDS and diver- immunization with CGG-alum) (Fig. 4B and fig.
ing the effect on GC selection of factors unrelated gence scores were also observed in FV-induced S9A). Comparison of SHM levels between high- and
to affinity. Moreover, the least divergent GCs in GCs (Fig. 3H and fig. S7C). Therefore, heteroge- low-dominance GCs from the same LN (fig. S9B)
polyclonal and SHM-proficient AID-Confetti mice neity in the outcome of selection is a common and between Aicda-sufficient and haploinsufficient
remained relatively close to baseline color distri- property of GCs induced by model antigens and GCs (fig. S9C) indicated that heterogeneity was

1052 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H | R ES E A R C H A R T I C L E S

1.2
LN#1 1.2 1.2
LN#2
1.1M 1.3M
GC1

OD450 (CGG)
1.1U 1.3U

OD450 (IgY)
OD450 (IgY)
0.8 1.2M 0.8 0.8
1.2U
1.3M GC2 2.1U, common
0.4 1.3U 0.4 0.4 2.1M1, GC1
1.4M 2.1M2-7, GC2
1.4U
0.0 0.0
0.01 0.1 1 10 100 1000 0.01 0.1 1 10 100 1000 0.01 0.1 1 10 100 1000
Fab (nM) Fab (nM) Fab (nM)

LN#1 LN#2
0.8 0.8 0.8

Binding to IgY (nm)


Binding to IgY (nm)

Binding to IgY (nm)


0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0


0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Time (s) Time (s) Time (s)

10-7 GC1 2.1U GC2 2.1M3


Igk 119C (WT)
0.8
Igk 119G (mut)
Binding to IgY (nm)

LN #2 Clone 2.1 (Igk)


K D (M)

4 2
10-8 0.6 5 5
2 3
2 23 4 2
0.4 2 2
4 2.1M6 2
10-9 3
0.2 4 2 2.1M4
2.1M1
2. M1
2. M2
2. M3
2. M4
2. 5
2. M6
7
2. 1U

1M

1M

2.1M5
1
1
1
1

1
2.

0.0 2.1M2
2.1M7
Fold-change over UA

15 0 500 1000 1500


Time (s)
10

5
Legend:
C119=C Putative precursor
1
C119>G (Ala>Gly) RFP YFP CFP
2. M1
2. M2
2. 3
2. M4
2. M5
2. M6
7
2. 1U

C119>A (Ala>Asp) Other symbols as in Fig. 4


1M

1M
1
1

1
1
1
2.

Fig. 5. Affinity maturation in GCs with high or low color dominance. from clone 2.1, either unmutated (gray lines) or with replacement of a single
Affinity measurements for reconstructed Fabs derived from B cell clones/variants Igk nucleotide [C119G (AlaGly), black lines]. (G) Igk sequence relationships
indicated in Fig. 4B. (A) Binding of Fabs cloned from LN1 to IgY (right) or CGG among B cells from clone 2.1 from LN2/GC1 (left) or LN2/GC2 (right). Symbols
(left), measured by enzyme-linked immunosorbent assay. (B) Biolayer inter- according to the legend below the figure and in Fig. 4B. Cloned Fab IDs are
ferometry for Fabs cloned from LN1 binding to IgY. (C and D) As in (A) and (B), indicated by black lines. Although not all cells yielded both Igk and Igh
respectively, but using Fabs cloned from LN#2. (E) Affinity for IgY among var- sequences, clonal relationships are drawn from all available data, and therefore
iants of clone 2.1 (Fig. 4B) from LN2/GC1 (blue) and LN2/GC2 (orange), shown exact correspondence between the trees in (G) and in Fig. 4B is not expected.
as KD (top) and fold change over UA (bottom). The unmutated ancestor (2.1U) Biolayer interferometry was performed with Fabs at 20, 40, 80, and 160 nM.
is shown in gray. Open bars have the WTnucleotide (C) in the 119 position; closed (D) shows only the 160-nM measurement. Reported affinities are the average
bars have the C119G (AlaGly) mutation. (F) Affinity of Fabs reconstructed of two measurements fitted globally at the 20- to 160-nM range.

not the consequence of asynchronous GC forma- panded over a short period of time, leading to a larger clone (accounting for 59% of all recombined
tion or of substantially impaired AID activity, extensive loss of clonal diversity concomitant cells), within which multiple colors were repre-
respectively. Most cells of the dominant color in with broad diversification of the expanded var- sented. This expanded clone carried heavy- and
the three high-dominance GCs (Fig. 4B and fig. iant. On the other hand, low-dominance GCs light-chain V(D)J rearrangements identical to the
S9A, left panels) derived from a single expanded (Fig. 4B and fig. S9A, right panels) were more clo- dominant clone in the single-colored GC sequenced
clone. In all three GCs, dominance could be read- nally diverse than their high-dominance neighbors. from the same LN (LN2/GC1), indicative of a com-
ily attributed to the selective expansion of a single Low color dominance is unlikely to be the result mon cell of origin. In contrast to the clonal bursts
SHM variant, three to five heavy-chain variable of failure to identify a dominant nonfluorescent of single-colored GCs, however, this multicolored
segment (VH) mutations (four to nine total muta- lineage, because fluorescent and nonfluorescent expansion branched off from the Igh UA itself
tions) distanced from the unmutated ancestor B cells obtained from two independently sequenced and developed in parallel along several distinct
(UA) (dashed boxes in Fig. 4B and fig. S9A). Thus, low-dominance GCs were largely clonally related lineages. Thus, expansion of this clone began in
high-dominance GCs are predominantly mono- (fig. S10). Although low-dominance GCs consisted the pre/early GC period, before the end of the ta-
clonal and appear to result from clonal bursts, predominantly of small independent expansions, moxifen pulse, and subsequent competition in the
in which a single SHM variant is heavily ex- one of three GCs sequenced (LN2/GC2) contained mature GC failed to focus on a single dominant

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1053


R ES E A RC H | R E S EA R C H A R T I C LE S

SHM variant. Together, these data indicate that, of distinct B cell clones; (ii) individual GCs ap- 7. G. D. Victora, M. C. Nussenzweig, Annu. Rev. Immunol. 30,
although clonal dominance in GCs can arise by proach homogeneity at different rates, leading to 429457 (2012).
8. G. D. Victora, L. Mesin, Curr. Opin. Immunol. 28, 9096 (2014).
parallel expansion of members of the same clone, variable loss of diversity; (iii) homogeneous GCs 9. G. D. Victora et al., Cell 143, 592605 (2010).
loss of diversity is greatest in GCs in which strong are the product of clonal bursts that lead to rapid 10. J. Jacob, R. Kassir, G. Kelsoe, J. Exp. Med. 173, 11651175
clonal bursts rapidly expand and diversify single expansion and further diversification of SHM (1991).
SHM variants. variants with improved affinity; and (iv) affinity 11. F. G. Kroese, A. S. Wubbena, H. G. Seijen, P. Nieuwenhuis,
Eur. J. Immunol. 17, 10691072 (1987).
maturation can take place in the presence or ab-
Affinity maturation in the absence of sence of such bursts and thus does not necessar-
12. Y. J. Liu, J. Zhang, P. J. Lane, E. Y. Chan, I. C. MacLennan,
Eur. J. Immunol. 21, 29512962 (1991).
homogenizing selection ily require radical loss of clonal diversity. 13. R. Kppers, M. Zhao, M. L. Hansmann, K. Rajewsky, EMBO J.
To investigate the relationship between clonal We speculate that affinity maturation arises 12, 49554967 (1993).
14. M. Ziegner, G. Steinhauser, C. Berek, Eur. J. Immunol. 24,
bursting and affinity maturation, we measured through a balance between slow but steady elim- 23932400 (1994).
the affinity of recombinant Fab fragments derived ination of lower-affinity clones and variants, 15. J. Livet et al., Nature 450, 5662 (2007).
from B cells originating from high- and low- punctuated by sporadic clonal bursts of various 16. H. J. Snippert et al., Cell 143, 134144 (2010).
dominance GCs (Fig. 4B), as well as from their magnitudes that generate a wealth of SHM var- 17. Materials and methods are available as supplementary
materials on Science Online.
deduced or sequenced UAs. Fab affinities varied iants of a clone at the expense of clonal diversity. 18. Z. Shulman et al., Science 341, 673677 (2013).
markedly between clones, from undetectable to This balance would ensure that the polyclonality 19. J. Jacob, J. Przylepa, C. Miller, G. Kelsoe, J. Exp. Med. 178,
low nanomolar dissociation constants (KD) (Fig. of the GC response is maintained at the same 12931307 (1993).
5). As expected, the two Fabs obtained from time as very successful clones can be heavily di- 20. D. F. Robbiani et al., Cell 135, 10281038 (2008).
21. N. J. Gotelli, R. K. Colwell, in Biological Diversity: Frontiers in
SHM variants that underwent clonal bursts versified and exported to effector or memory fates, Measurement and Assessment, A. E. Magurran, B. J. McGill,
had higher affinity for the immunizing antigen generating the range of GC outcomes observed by Eds. (Oxford University Press, Oxford, 2010), pp. 3954.
than their respective UAs [5.3-fold increase (1.0 us and others (14, 27). Although it cannot be ruled 22. I. Dogan et al., Nat. Immunol. 10, 12921299 (2009).
107 to 1.9 108) for clone 1.1 and 14.5-fold in- out that differences in outcomes between GCs 23. M. Jarjour et al., J. Exp. Med. 211, 11091122 (2014).
24. D. Dominguez-Sola et al., Nat. Immunol. 13, 10831091 (2012).
crease (4.1 108 to 2.9 109) for clone 2.1] (Fig. may be related to factors such as initial clonal 25. See the supplementary materials for mathematical validation
5). However, affinity maturation was also evident composition or timing of the appearance of key of the NDS.
in Fabs cloned from low-dominance GCs. Of the mutations, the finding that GCs can deviate sub- 26. S. Nair et al., Retrovirology 8, 76 (2011).
three clones obtained from LN1/GC2, SHM var- stantially from baseline color distribution even 27. M. D. Radmacher, G. Kelsoe, T. B. Kepler, Immunol. Cell Biol.
76, 373381 (1998).
iants of two clones (1.3 and 1.4) showed increased when all cells have equal affinity and cannot 28. A. D. Gitlin, Z. Shulman, M. C. Nussenzweig, Nature 509,
affinity over their UAs, whereas the affinity of the undergo SHM (Fig. 3, F and G, and fig. S6D) sug- 637640 (2014).
third clone (1.2) remained below detection level gests that stochastic factors unrelated to affinity 29. A. D. Gitlin et al., Science 349, 643646 (2015).
(Fig. 5). Clone 1.4 showed a gain in affinity of can strongly influence GC selection. A system in 30. J. Jacob, G. Kelsoe, J. Exp. Med. 176, 679687 (1992).
31. J. Faro, M. Or-Guil, BMC Bioinformatics 14 (Suppl 6), S8
12-fold (2.4 107 to 2.0 108), notably larger which B cell selection is driven by dynamic en- (2013).
than the 5.4-fold change observed for clone 1.1 counters with a limiting number of T helper cells 32. G. D. Victora, P. C. Wilson, Cell 163, 545548 (2015).
(associated with a clonal burst in the neighbor- would be especially sensitive to such effects (9, 27). 33. F. Klein et al., Science 341, 11991204 (2013).
ing single-colored GC). We also determined the The explosive B cell proliferation that can be driv-
AC KNOWLED GME NTS
affinity of six additional variants of clone 2.1 en by such encounters would provide a mechanis-
We thank M. Carroll, A. Schmidt, and S. Harrison (Harvard Medical
from LN2/GC2 (which shared its UA with the tic basis for the clonal burst phenomenon (28, 29). School), H. Wardemann (German Cancer Research Center), and
dominant clone of LN2/GC1 and is thus directly The finding that early GCs can contain a large K. Hasenkrug (National Institute of Allergy and Infectious Diseases)
comparable between GCs). Gains in affinity were diversity of clones is in contrast to the widely for essential reagents, and R. K. Colwell (University of Connecticut)
evident in all of these variants, ranging from 3.5- held view that these structures form pauciclonally and S. Binder (Helmholtz Centre for Infection Research) for
to 13.8-fold (KD 1.2 108 and 3.0 109, respec- (1012, 30). Although this discrepancy may be help with statistical analysis. We are especially thankful to the
late Herman N. Eisen (19182014) for the lengthy discussions
tively) (Fig. 5, C to E). Although the highest of partly attributable to methodological issues (31), of affinity maturation on which the present study is based. The
these was close in magnitude to the 14.5-fold Kppers et al. (13) reported higher clonal diver- data presented in this manuscript are tabulated in the main
change seen in variant 2.1M1 (associated with the sity in one of two GCs originating from different paper and in the supplementary materials. Antibody sequences
are also available from the GenBank database, accession
clonal burst in LN2/GC1), the increase in affinity reactive human lymph nodes and speculated that numbers KU613419 to KU615568. Aicda-CreERT2 mice are
among GC2 Fabs was in general lower (median pauciclonal GCs were likely the product of exten- available from C. A. Reynaud and J. C. Weill under a material
increase of 6.8-fold). Thus, affinity maturation sive selection acting on originally polyclonal early transfer agreement with INSERM. This work was supported by
can occur in the absence of clonal bursts, although GCs. Our current data strongly support this view. NIH grant 5DP5OD012146 (G.D.V.), with additional support
from Human Frontier Science Program Grant RGP0033/2015
limited data from the two LN2 GCs suggests that Our findings have implications for the design (G.D.V. and M.M.H.), German Federal Ministry of Education
selection of high-affinity mutants in these cases of vaccines against highly variable pathogens such and Research within the Measures for the Establishment of
may not be as efficient. as influenza and HIV, where broadly protective Systems Medicine, eMed project SYSIMIT, FKZ: 01ZX1308B
Finally, detailed analysis of the SHM trajec- antibodies targeting conserved, nonimmuno- (M.M.H.), Swiss National Science Foundation Postdoctoral
Mobility Fellowship and Cancer Research Institute Irvington
tory of variant 2.1M1 showed that this burst was dominant epitopes must be elicited (32, 33). Under- Postdoctoral Fellowship (G.P.), and Norwegian Research
associated with a somatic mutation in Igk standing whether GC competition restricts the Council Fri Prosjektsttte (FRIPRO) mobility grant (J.T.J.). The
(C119G, AlaGly) that alone increased the Fab appearance of such clones, and how clonal bursts Octet RED96 Bio-Layer Interferometry System was supported by
affinity by 7.6-fold (Fig. 5, E to G). This mutation may promote the somatic diversification required NIH grant S10 OD016326 to the MIT Biophysical
Instrumentation Facility.
was also found to have occurred independently for broad neutralization to emerge, may prove key
at least twice in LN2/GC2, but without the ac- to the successful generation of broadly protective
SUPPLEMENTARY MATERIALS
companying burst. Thus, identical mutations oc- antibodies by vaccination.
www.sciencemag.org/content/351/6277/1048/suppl/DC1
curring within the same clone may still have Materials and Methods
distinct outcomes when in different GCs, again RE FERENCES AND NOTES Supplementary Text
suggesting a role for stochastic factors in GC 1. H. N. Eisen, G. W. Siskind, Biochemistry 3, 9961008
Figs. S1 to S11
selection. Tables S1 and S2
(1964).
Movies S1 to S3
2. H. N. Eisen, Cancer Immunol. Res. 2, 381392 (2014).
Conclusions 3. I. C. MacLennan, Annu. Rev. Immunol. 12, 117139 (1994).
Database S1
References (3455)
4. K. Rajewsky, Nature 381, 751758 (1996).
Using a combination of imaging methods and 5. C. D. Allen, T. Okada, J. G. Cyster, Immunity 27, 190202 1 September 2015; accepted 2 February 2016
single-cell sequencing, we show that (i) early GCs (2007). Published online 18 February 2016
can be highly diverse, containing up to hundreds 6. D. M. Tarlinton, Immunol. Cell Biol. 86, 133138 (2008). 10.1126/science.aad3439

1054 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


R ES E A RC H

decays only T 1, slower than ee , which should


REPORTS in principle allow the hydrodynamic description
to apply over a certain temperature range, until
other phonon-mediated processes become im-
ELECTRON TRANSPORT portant. So far, there has been little evidence for
hydrodynamic electron transport. An exception

Negative local resistance caused is an early work on 2D electron gases in ballistic


devices ( ~ W) made from GaAlAs heterostruc-
tures (15). These devices exhibited nonmonotonic
by viscous electron backflow changes in differential resistance as a function of a
large applied current I, which was used to increase

in graphene the electron temperature (making ee short) while


the lattice temperature remained low (allowing
long ). The nonmonotonic behavior was attri-
D. A. Bandurin,1 I. Torre,2 R. Krishna Kumar,1,3 M. Ben Shalom,1,4 A. Tomadin,5 buted to the Gurzhi effect, a transition between
A. Principi,6 G. H. Auton,4 E. Khestanova,1,4 K. S. Novoselov,4 I. V. Grigorieva,1 Knudsen (ee ) and viscous electron flows (7, 15).
L. A. Ponomarenko,1,3 A. K. Geim,1* M. Polini7* Another possible hint about electron hydrody-
namics comes from an explanation (16) of the
Graphene hosts a unique electron system in which electron-phonon scattering is extremely Coulomb drag measured between two graphene
weak but electron-electron collisions are sufficiently frequent to provide local equilibrium above sheets at the charge neutrality point (CNP) (17).
the temperature of liquid nitrogen. Under these conditions, electrons can behave as a viscous Here we address electron hydrodynamics by
liquid and exhibit hydrodynamic phenomena similar to classical liquids. Here we report using a special measurement geometry (Fig. 1)
strong evidence for this transport regime.We found that doped graphene exhibits an anomalous that amplifies the effects of the shear viscosity n
(negative) voltage drop near current-injection contacts, which is attributed to the formation and, at the same time, minimizes a contribution

Downloaded from on March 3, 2016


of submicrometer-size whirlpools in the electron flow.The viscosity of graphenes electron liquid from ballistic effects that can occur not only in
is found to be ~0.1 square meters per second, an order of magnitude higher than that of honey, the Knudsen regime but also in viscous flows in
in agreement with many-body theory. Our work demonstrates the possibility of studying graphene. A viscous flow can lead to vortices ap-
electron hydrodynamics using high-quality graphene. pearing in the spatial distribution of the steady-
state current (Fig. 1, A and B). Suchpelectron

whirlpools have a spatial scale Dn nt, which

T
he collective behavior of many-particle sys- such as those involving impurities and phonons. depends on electron-electron scattering through
tems that undergo frequent interparticle The above inequalities are difficult to meet exper- n and on the electron systems quality through t
collisions has been studied for more than imentally. At low temperatures (T), ee varies ap- (18). To detect the whirlpools, electrical probes
two centuries and is routinely described proximately T 2, reaching a micrometer scale should be placed at a distance comparable to Dn.
by the theory of hydrodynamics (1, 2). The at liquid helium T (14), which necessitates the By using single- and bi-layer graphene (SLG and
theory relies only on the conservation of mass, use of ultraclean systems to satisfy ee . At higher BLG, respectively) encapsulated between boron
momentum, and energy and is highly successful T, electron-phonon scattering rapidly reduces . nitride crystals (1921), we were able to reach a
in explaining the response of classical gases and However, for two-dimensional (2D) systems in Dn of 0.3 to 0.4 mm thanks to the high viscosity of
liquids to external perturbations that vary slowly which acoustic phonon scattering dominates, graphenes Fermi liquid and its high carrier
in space and time. More recently, it has been
shown that hydrodynamics can also be applied
to strongly interacting quantum systems, includ-
ing ultrahot nuclear matter and ultracold atomic 60
Fermi gases in the unitarity limit (36). In prin-
xx (mS)

ciple, the hydrodynamic approach can also be 40


used to describe many-electron phenomena in
20 100 K
condensed matter physics (713). The theory be- 200 K
300 K
comes applicable if electron-electron scattering 0
provides the shortest spatial scale in the prob- -3 -2 -1 0 1 2 3
lem, so that ee W, , where ee is the electron-
stray-current contribution
electron scattering length, W is the characteristic 10 100 K
200 K 100 K
sample size, nFt is the mean free path, nF is the 300 K 200 K
RV ()

300 K
Fermi velocity, and t is the mean free time with 5
respect to momentum-nonconserving collisions,
0
1
School of Physics and Astronomy, University of Manchester,
Oxford Road, Manchester M13 9PL, UK. 2National Enterprise -5
-3 -2 -1 0 1 2 3
for nanoScience and nanoTechnology, Scuola Normale 12 -2
Superiore, I-56126 Pisa, Italy. 3Department of Physics, n (10 cm )
Lancaster University, Lancaster LA14YB, UK. 4National
Graphene Institute, University of Manchester, Manchester Fig. 1. Viscous backflow in doped graphene. (A and B) The calculated steady-state distribution of a
M13 9PL, UK. 5National Enterprise for nanoScience and
current injected through a narrow slit for (A) a classical conducting medium with zero n and (B) a viscous
nanoTechnology, Istituto Nanoscienze-Consiglio Nazionale
delle Ricerche and Scuola Normale Superiore, I-56126 Pisa, Fermi liquid. (C) Optical micrograph of one of our SLG devices. The schematic explains the measurement
Italy. 6Institute for Molecules and Materials, Radboud geometry for vicinity resistance. The top gate electrode appears in white and the mesa, which is etched in
University, NL-6525 AJ Nijmegen, Netherlands. 7Istituto encapsulated graphene and not covered with a metal, appears in purple. Mixed colors at the periphery are
Italiano di Tecnologia, Graphene Labs, Via Morego 30,
areas of metallic contacts on top of the mesa. (D and E) Longitudinal conductivity sxx and Rv as a function
I-16163 Genova, Italy.
*Corresponding author. E-mail: geim@man.ac.uk (A.K.G.); of n, induced by applying gate voltage. The dashed curves in (E) show the contribution expected from
marco.polini@iit.it (M.P.) classical stray currents in this geometry (18). I = 0.3 mA; L = 1 mm; W = 2.5 mm.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1055


R ES E A RC H | R E PO R TS

Fig. 2. Vicinity resist-


ance maps. (A and B)
Rv(n,T) for SLG and
BLG, respectively. The
black curves indicate
zero Rv. For each n
away from the CNP,
there is a wide range
of T over which Rv is
negative. For the SLG
device, W = 2.5 mm
and L = 1 mm; for the
BLG device, W =
2.3 mm and L = 1.3 mm.
All measurements for BLG presented in this work were taken with zero displacement between the graphene layers (18).

mobility m, even at high T. Such a large Dn, which Fig. 3. Whirlpools in


is unique to graphene, nevertheless necessitates electron flow. (A to C)
submicron resolution to probe the electron back- Calculated J(r) and (r)
flow. To this end, we fabricated multiterminal for a geometry similar to
Hall bars with narrow (~0.3 mm) and closely spaced that shown in Fig. 1C, with
(~1 mm) voltage probes (Fig. 1C and fig. S1). De- the green bars indicating
tails of the device fabrication are given in (18). voltage contacts. Dn =
All our devices were first characterized in the 2.3, 0.7, and 0 mm for (A),
standard geometry by applying I along the main (B), and (C), respectively.
channel and using side probes for voltage mea- Vortices are evident in the
surements. The typical behavior of longitudinal top right corners of
conductivity sxx at a few characteristic values of (A) and (B), where the
T is shown in Fig. 1D. At liquid helium T, the de- current flow is in the
vices exhibited m ~ 10 to 50 m2 V1 s1 for carrier direction opposite to that
concentrations n over a wide range of the order in (C), which shows the
of 1012 cm2, and m remained above 5 m2 V1 s1 case of zero viscosity.
up to room T (fig. S2). Such values of m allow In each panel, the current
ballistic transport with > 1 mm at T < 300 K. At streamlines also change
T 150 K, ee decreases to 0.1 to 0.3 mm over the from white to black to
same range of n (figs. S3 and S4) (22, 23). This indicate that the current
allows the essential condition for electron hydro- density |J(r)| is lower to
dynamics (ee W, ) to be satisfied within this the right of the injecting
temperature range. If one uses the conventional contact.
longitudinal geometry of electrical measurements,
viscosity has little effect on sxx (figs. S5 to S7), -I/0 I/0 2 m

essentially because the flow in this geometry is
uniform, whereas the total momentum of the
moving Fermi liquid is conserved in electron- ventional current flow (Fig. 1A) and negative for (15) and used the current I to increase the elec-
electron collisions (18). The only evidence for viscous backflow (Fig. 1B). Figure 1E shows ex- tron temperature. In this case, Vv changed its
hydrodynamics that we could find in the longi- amples of Rv for the same SLG device as in Fig. sign as a function of I from positive to negative to
tudinal geometry was the Gurzhi effect that ap- 1D, and other SLG and BLG devices exhibited positive again, reproducing the behavior of Rv with
peared as a function of the electron temperature, similar behavior (18). Away from the CNP, Rv is increasing T of the cryostat (fig. S10). Comparing
which is controlled by applying large I, similar to negative over a wide range of intermediate T, figs. S8 and S10, it is clear that the vicinity ge-
the observations in (15) (fig. S8). despite an expected substantial offset due to stray ometry strongly favors the observation of hydro-
To look for hydrodynamic effects, we used the currents. Figure 2 details our observations fur- dynamic effects: The measured vicinity voltage
geometry shown in Fig. 1C. In this setup, I is ther by showing maps of Rv (n,T) for SLG and changed its sign, whereas in the standard geom-
injected through a narrow constriction into the BLG. The two Fermi liquids exhibited somewhat etry, the same viscosity led only to relatively
graphene bulk, and the voltage drop Vv is mea- different behavior, reflecting their different elec- small changes in dV/dI. We also found that the
sured at the nearby side contacts located at the tronic spectra, but Rv was negative over a large magnitude of negative Rv decayed rapidly with L
distance L ~ 1 mm away from the injection point. range of n and T for both. Two more Rv maps are (fig. S11), in agreement with the finite size of
These can be considered as nonlocal measure- provided in fig. S9. In total, seven multiterminal electron whirlpools.
ments, although the stray currents are not ex- devices with W ranging from 1.5 to 4 mm were Negative resistances can in principle arise from
ponentially small (dashed curves in Fig. 1E). To investigated, showing vicinity behavior that was other effects, such as single-electron ballistic trans-
distinguish from the proper nonlocal geometry highly reproducible both for different contacts port (ee ) or quantum interference (18, 20, 24).
(24), we refer to the linear-response signal mea- on the same device and for different devices, in- The latter contribution is easily ruled out, because
sured in our geometry as vicinity resistance, Rv = dependently of their W, although we note that the quantum corrections rapidly wash out at T > 20 K
Vv/I. The idea is that, in the case of a viscous backflow was more pronounced for devices with and have a random sign that rapidly oscillates
flow, whirlpools emerge as shown in Fig. 1B, and the highest m and lowest charge inhomogeneity. as a function of magnetic field. Also, our numeri-
their appearance can then be detected as sign The same anomalous vicinity response was cal simulations using the Landauer-Bttiker for-
reversals of Vv, which is positive for the con- also evident when we followed the method of malism and the realistic device geometry showed

1056 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

Fig. 4. Viscosity of the


vorticity satisfies the equation wr D2n 2 wr,
Fermi liquids in graphene.
where Dn plays the role of a diffusion constant.
(A and B) Solid curves show
The current I injects vorticity at the source con-
values of n for SLG and BLG,
tact, which then exponentially decays over the
respectively. Dashed curves
length scale Dn. For n = 0.1 m2 s1 [as estimated
represent calculations based
in (25)] and t = 1.5 ps (fig. S2), we find that Dn
on many-body diagrammatic
0.4 mm, in qualitative agreement with the mea-
perturbation theory (no
surements in fig. S11.
fitting parameters). The gray
Lastly, we can combine the measurements of Rv
shaded areas indicate
and resistivity rxx with the solution of Eqs. 1 and 2
regions around the CNP
in Fig. 3 to extract the kinematic viscosity for SLG
where our hydrodynamic
and BLG. Because the observed Gurzhi effect in rxx
model is not applicable (18).
is small at low currents (fig. S6), we can use rxx =
1/s0 = m/(ne2t) to determine t(n,T) (18). Further-
more, for the experimentally relevant values of
Dn, we find that Rv is a quadratic function of Dn
RV b aD2n s1
0 4
where a and b are numerical coefficients de-
pendent only on the measurement geometry and
boundary conditions, and b describes the contri-
bution from stray currents (fig. S14). For the
specific device in Fig. 3, we determined that a =
0.29 mm2 and b = 0.056, and this allows us to
estimate Dn(n,T) from measurements of Rv. For
the known t and Dn, we find that nn; T D2n =t.
The applicability limits of this analysis are dis-
cussed in (18), and the results are plotted in Fig. 4
for one of our devices. The figure shows that, at
carrier concentrations of ~1012 cm2, the Fermi
liquids in both SLG and BLG are highly viscous,
that no negative resistance could be expected for momentum-nonconserving processes that are with n 0.1 m2 s1. In comparison, liquid honey
the vicinity configuration in zero magnetic field parameterized by the scattering time t(n,T). has typical viscosities of ~0.002 to 0.005 m2 s1.
(19, 21). Nonetheless, we carefully considered the Equations 1 and 2 can be solved numerically Figure 4 also shows the results of fully inde-
possibility of any accidental spillover of single- (18), and Fig. 3 shows examples of spatial dis- pendent microscopic calculations of n(n,T), which
electron ballistic effects into the vicinity geom- tributions of (r) and J(r). For experimentally were carried out by extending the many-body
etry of our experiment. The dependences of the relevant values of Dn, a vortex appears in the vi- theory of (25) to the case of 2D electron liquids
negative vicinity signal on T, n, I, and the device cinity of the current-injecting contact. This is hosted by encapsulated SLG and BLG. Within
geometry allowed us to unambiguously rule out accompanied by the sign reversal of (r) at the the range of applicability of our analysis in Fig. 4
any such contribution (18). For example, the single- vicinity contact on the right of the injector, which (n ~ 1012 cm2), the agreement in absolute values
electron ballistic phenomena should become is positive in Fig. 3C (no viscosity) but becomes of the electron viscosity is good, especially taking
more pronounced for longer (that is, with de- negative in Fig. 3, A and B. Our calculations into account that no fitting parameters were used
creasing T or electron temperature and with in- for this geometry show that Rv is negative for in the calculations. Because the strong inequality
creasing n), in contrast to the nonmonotonic Dn 0.4 mm (18). Because both t and n decrease ee required by the hydrodynamic theory can-
behavior of Vv. with increasing T, Dn also decreases, and stray not be reached even for graphene, it would be
Turning to theory, we can show that negative currents start to dominate the vicinity response unreasonable to expect better agreement (18). In
Rv arises naturally from whirlpools that appear at high T. This explains why Rv in Figs. 1 and 2 addition, our analysis does not apply near the
in a viscous Fermi liquid near current-injecting becomes positive close to room T, even though our CNP, because the theory neglects contributions
contacts. As discussed in (18), electron transport hydrodynamic description has no high-temperature from thermally excited carriers, spatial charge in-
for sufficiently short ee can be described by the cutoff (18). Despite positive Rv values, the viscous homogeneity, and coupling between charge and
hydrodynamic equations contribution remains considerable near room T energy flows, which can play a substantial role at
(Fig. 1D and fig. S12). At low T, the electron sys- low doping (16, 18). Further work is needed to
Jr 0 1
tem approaches the Knudsen regime, and our understand electron hydrodynamics in the inter-
and hydrodynamic description becomes inapplicable mediate regime ee and, for example, to ex-
because ee ~ (18). In the latter regime, the whirl- plain ballistic transport ( > W) in graphene at
s0
fr D2n 2 Jr Jr 0 2 pools should disappear and Rv should become high T in terms of suitably modified hydrodynamic
e positive (fig. S13), in agreement with the experi- theory. The naive single-particle description that
where J(r) is the (linearized) particle current ment and our numerical simulations based on is routinely used for graphenes ballistic phenom-
density, and (r) is the electric potential in the 2D the Landauer-Bttiker formalism. ena even above 200 K (19, 21) cannot be justified;
plane. If Dn 0, Eq. 2 yields Ohms law eJ(r) = The numerical results in Fig. 3 can be under- a more complete theory is needed to describe the
s0E(r) with a Drude-like conductivity s0 ne2t/m, stood if we rewrite Eqs. 1 and 2 as injection of a collimated electron beam into a
where e and m are the electron charge and the strongly interacting 2D liquid. As for experimen-
s0
effective mass, respectively (E, electric field). The Jr fr nD2n  wr 3 tal approaches, the highly viscous Fermi liquids
three terms in Eq. 2 describe (i) the electric force e in graphene and their accessibility offer a promis-
generated by the steady-state charge distribu- where wr n1  Jr wr^ z is the vor- ing opportunity to use various scanning probes
tion in response to the applied current I, (ii) the z is the unit vector perpendicular to the
ticity (^ for visualization and further understanding of
viscous force (1, 2), and (iii) friction caused by graphene plane) (2). Taking the curl of Eq. 3, the electron hydrodynamics.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1057


R ES E A RC H | R E PO R TS

RE FE RENCES AND N OT ES 13. B. N. Narozhny, I. V. Gornyi, M. Titov, M. Schtt, A. D. Mirlin, AC KNOWLED GME NTS
1. L. D. Landau, E. M. Lifshitz, Fluid Mechanics (Pergamon Press, Phys. Rev. B 91, 035414 (2015). This work was supported by the European Research Council,
1987). 14. G. F. Giuliani, G. Vignale, Quantum Theory of the Electron Liquid the Royal Society, Lloyds Register Foundation, the Graphene
2. G. K. Batchelor, An Introduction to Fluid Dynamics (Cambridge (Cambridge Univ. Press, 2005). Flagship, and the Italian Ministry of Education, University and
Univ. Press, 1967). 15. M. J. M. de Jong, L. W. Molenkamp, Phys. Rev. B 51, Research through the program Progetti Premiali 2012 (project
3. B. V. Jacak, B. Mller, Science 337, 310314 (2012). 1338913402 (1995). ABNANOTECH). D.A.B. and I.V.G. acknowledge support from the
4. C. Cao et al., Science 331, 5861 (2011). 16. J. C. W. Song, L. S. Levitov, Phys. Rev. Lett. 109, 236602 Marie Curie program SPINOGRAPH (Spintronics in Graphene).
5. E. Elliott, J. A. Joseph, J. E. Thomas, Phys. Rev. Lett. 113, (2012). A.P. received support from the Nederlandse Wetenschappelijk
020406 (2014). 17. R. V. Gorbachev et al., Nat. Phys. 8, 896901 (2012). Organisatie. R.K.K. received support from the Engineering and
6. T. Schfer, D. Teaney, Rep. Prog. Phys. 72, 126001 (2009). 18. Supplementary materials are available on Science Online. Physical Sciences Research Council.
7. R. N. Gurzhi, Sov. Phys. Usp. 11, 255270 (1968). 19. A. S. Mayorov et al., Nano Lett. 11, 23962399
8. A. O. Govorov, J. J. Heremans, Phys. Rev. Lett. 92, 026803 (2011).
(2004). 20. L. Wang et al., Science 342, 614617 (2013). SUPPLEMENTARY MATERIALS
9. R. Bistritzer, A. H. MacDonald, Phys. Rev. B 80, 085109 21. T. Taychatanapat, K. Watanabe, T. Taniguchi, P. Jarillo-Herrero,
www.sciencemag.org/content/351/6277/1055/suppl/DC1
(2009). Nat. Phys. 9, 225229 (2013).
Supplementary Text
10. M. Mller, J. Schmalian, L. Fritz, Phys. Rev. Lett. 103, 025301 22. Q. Li, S. Das Sarma, Phys. Rev. B 87, 085406 (2013).
Figs. S1 to S14
(2009). 23. M. Polini, G. Vignale, http://arxiv.org/abs/1404.5728
References (2637)
11. A. V. Andreev, S. A. Kivelson, B. Spivak, Phys. Rev. Lett. 106, (2014).
256804 (2011). 24. D. A. Abanin et al., Science 332, 328330 (2011). 14 July 2015; accepted 23 December 2015
12. A. Tomadin, G. Vignale, M. Polini, Phys. Rev. Lett. 113, 235901 25. A. Principi, G. Vignale, M. Carrega, M. Polini, http://arxiv.org/ Published online 11 February 2016
(2014). abs/1506.06030 (2015). 10.1126/science.aad0201

ELECTRON TRANSPORT those applicable to its nonrelativistic counter-


parts. A number of unusual properties have been

Observation of the Dirac fluid and the predicted, including nearly perfect (inviscid) flow
(19) and a diverging thermal conductivity, which

Downloaded from on March 3, 2016


results in the breakdown of the Wiedemann-Franz
breakdown of the Wiedemann-Franz (WF) law at finite temperature (5, 6).
Away from the CNP, graphene has a sharp Fermi

law in graphene surface, and the standard FL phenomenology holds.


By tuning the chemical potential, we are able to
measure thermal and electrical conductivity in
Jesse Crossno,1,2 Jing K. Shi,1 Ke Wang,1 Xiaomeng Liu,1 Achim Harzheim,1 both the DF and the FL in the same sample. In a
Andrew Lucas,1 Subir Sachdev,1,3 Philip Kim,1,2* Takashi Taniguchi,4 Kenji Watanabe,4 FL, the relaxation of heat and charge currents is
Thomas A. Ohki,5 Kin Chung Fong5* closely related, as they are carried by the same
quasi-particles. The WF law (20) states that the
Interactions between particles in quantum many-body systems can lead to collective behavior electronic contribution to a metals thermal con-
described by hydrodynamics. One such system is the electron-hole plasma in graphene near ductivity ke is proportional to its electrical con-
the charge-neutrality point, which can form a strongly coupled Dirac fluid. This charge-neutral ductivity s and temperature T, such that the
plasma of quasi-relativistic fermions is expected to exhibit a substantial enhancement of the Lorenz ratio L satisfies
 
thermal conductivity, thanks to decoupling of charge and heat currents within hydrodynamics. ke p2 kB 2
Employing high-sensitivity Johnson noise thermometry, we report an order of magnitude L L0 1
sT 3 e
increase in the thermal conductivity and the breakdown of the Wiedemann-Franz law in the
thermally populated charge-neutral plasma in graphene. This result is a signature of the Dirac where e is the electron charge, kB is the Boltz-
fluid and constitutes direct evidence of collective motion in a quantum electronic fluid. mann constant, and L0 is the Sommerfeld value
derived from FL theory. L0 depends only on

U
fundamental constants, not specific details of the
nderstanding the dynamics of many inter- interacting quantum systems, from high-mobility system such as carrier density or effective mass.
acting particles is a formidable task in phys- electrons in conductors (2) to cold atoms (3) and As a robust prediction of FL theory, the WF law
ics. For electronic transport in matter, strong quark-gluon plasmas (4). Hydrodynamic effects has been verified in numerous metals (20). At
interactions can lead to a breakdown of the are expected to greatly modify transport coef- high temperatures, the WF law can be violated
Fermi liquid (FL) paradigm of coherent ficients compared with their FL counterparts, as due to inelastic electron-phonon scattering or
quasi-particles scattering off of impurities. In has been argued for strongly interacting massless bipolar diffusion in semiconductors, even when
such situations, provided that certain conditions Dirac fermions in graphene at the charge-neutrality electron-electron interactions are negligible (21).
are met, the complex microscopic dynamics can point (CNP) (58). In recent years, several nontrivial violations of
be coarse-grained to a hydrodynamic description Many-body physics in graphene is interesting the WF lawall of which are related to the emer-
of momentum, energy, and charge transport on because of electron-hole symmetry and a linear gence of non-FL behaviorhave been reported
long length and time scales (1). Hydrodynamics dispersion relation at the CNP (9, 10). Together in strongly interacting systems such as Luttinger
has been successfully applied to a diverse array of with the vanishing Fermi surface, the ultra- liquids (22), metallic ferromagnets (23), heavy fer-
relativistic spectrum leads to ineffective screening mion metals (24), and underdoped cuprates (25).
(11) and the formation of a strongly interacting Owing to the strong Coulomb interactions be-
1
Department of Physics, Harvard University, Cambridge, MA quasi-relativistic electron-hole plasma known as tween thermally excited charge carriers, the WF
02138, USA. 2John A. Paulson School of Engineering and
Applied Sciences, Harvard University, Cambridge, MA 02138,
a Dirac fluid (DF) (12). The DF shares many fea- law is expected to be violated at the CNP in a DF.
USA. 3Perimeter Institute for Theoretical Physics, Waterloo, tures with quantum critical systems (13): most An electric field drives electrons and holes in op-
Ontario N2L 2Y5, Canada. 4National Institute for Materials importantly, the electron-electron scattering time posite directions; collisions between them intro-
Science, Namiki 1-1, Tsukuba, Ibaraki 305-0044, Japan. is fast (1417) and well suited to a hydrodynamic duce a frictional dissipation, resulting in a finite
5
Quantum Information Processing Group, Raytheon BBN
Technologies, Cambridge, MA 02138, USA.
description. Because of the quasi-relativistic na- conductivity even in the absence of disorder (26).
*Corresponding author. E-mail: pkim@physics.harvard.edu (P.K.); ture of the DF, this hydrodynamic limit is de- In contrast, a temperature gradient causes elec-
kc.fong@bbn.com (K.C.F.) scribed by equations (18) quite different from trons and holes to move in the same direction,

1058 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

Te was monitored independently of the lattice


temperature through the Johnson noise power
emitted at 100 MHz, with a 20-MHz bandwidth
defined by an inductor-capacitor matching net-
work. We designed our JNT setup to be operated
over a wide temperature range, from 3 to 300 K
(35). With a precision of ~10 mK, we measured
small deviations of Te from Tbath (i.e., DT Tbath ).
In this limit, the temperature of the graphene
lattice is well thermalized to the bath (32), and
our JNT setup allows us to sensitively measure
the electronic cooling pathways in graphene.
When the temperature is low enough, electron
and lattice interactions are weak (35, 36), and
most of the Joule heat generated in graphene
escapes via direct diffusion to the contacts (21).
As the temperature increases, electron-phonon
scattering becomes appreciable, and thermal tran-
sport becomes limited by the electron-phonon
coupling strength (3638). The onset temperature
of appreciable electron-phonon scattering, Tel-ph,
Fig. 1. Temperature- and density-dependent electrical and thermal conductivity. (A) Resistance (R) depends on the sample disorder and device geo-
versus gate voltage (Vg) at various temperatures. kW, kilohm. (B) Electrical conductivity (blue) as a metry: Tel-ph ~ 80 K (35, 36, 39, 40) for our samples.
function of the charge density set by the back gate for different bath temperatures. The residual carrier Below this temperature, the electronic contribu-
density at the neutrality point (green) is estimated by the intersection of the minimum conductivity with tion of the thermal conductivity can be obtained
a linear fit to log(s), away from neutrality (dashed gray lines). Curves have been offset vertically such that from P and DT using the device dimensions (21).
the minimum density (green) aligns with the temperature axis to the right. Solid black lines correspond to Figure 1C shows ke(Vg) plotted alongside the
4e2/h. At low temperatures, the minimum density is limited by disorder (charge puddles). Above Tdis ~ 40 K simultaneously measured s(Vg) at various fixed
(yellow shaded area), thermal excitations begin to dominate, and the sample enters the nondegenerate bath temperatures. Here, for a direct quantitative
regime near the neutrality point. (C to E) Thermal conductivity (red points) as a function of (C) gate voltage comparison based on the WF law, we plot the
and [(D) and (E)] bath temperature, compared to with the WF law, sTL0 (blue lines). At low temperature scaled electrical conductivity as sT L0 in the same
and/or high doping jmj kB T, we find the WF law to hold. This is a nontrivial check on the quality of our units as ke. In a FL, these two values will coincide,
measurement. In the nondegenerate regime jmj < kB T, the thermal conductivity is enhanced and the in accordance with Eq. 1. At low temperatures (T <
WF law is violated. Above T ~ 100 K, electron-phonon coupling becomes appreciable and begins to dominate Tdis ~ 40 K), where the puddle-induced density
thermal transport at all measured gate voltages. At this temperature, the yellow shaded background ends. All fluctuations dominate, we find that ke sT L0 ,
data from this figure are taken from sample S2 [inset in (E)]. monotonically increasing as a function of carrier
density with a minimum at the neutrality point,
confirming the WF law in the disordered regime.
thus inducing an energy current that grows un- the charge carrier density n = ne nh, where ne As T increases (T > Tdis), however, we begin to
impeded by interparticle collisions as the total and nh are the electron and hole densities, re- observe violation of the WF law. This violation ap-
momentum is conserved. The thermal conduc- spectively (21). All measurements were performed pears only close to the CNP, with the measured
tivity is therefore limited by the rate at which in a cryostat to control the bath temperature Tbath. thermal conductivity maximized at n = 0 (Fig. 1C).
momentum is relaxed by residual impurities. Figure 1A shows the resistance R versus Vg mea- The deviation is the largest at 75 K, where ke is
Realization of the DF in graphene requires sured at various fixed temperatures for a repre- more than an order of magnitude larger than the
that the thermal energy be larger than the local sentative device [see (21) for all samples]. From value expected for a FL. The nonmonotonicity of
chemical potential m(r), defined at position r: this, we used the known sample dimensions to ke(T) is consistent with acoustic phonons relaxing
kB T jmrj. Impurities cause spatial variations estimate the electrical conductivity s (Fig. 1B). At momentum more efficiently than impurities as T
in the local chemical potential, and even when the CNP, the residual charge carrier density nmin increases (41). For T 100K in our samples, activ-
the sample is globally neutral, it is locally doped can be estimated by extrapolating a linear fit of ation of optical phonons introduces an additional
to form electron-hole puddles with finite m(r) log(s) as a function of log(n) out to the minimum electron-phonon cooling pathway (35), and the
(2730). At high temperatures, formation of the conductivity (34). At the lowest temperatures, nmin measured thermal conductivity is larger than ke.
DF is complicated by phonon scattering, which saturates to ~8 109 cm2. Extraction of nmin by This non-FL behavior quickly disappears as jnj in-
can relax momentum by creating additional in- this method prompts overestimation of the charge- creases; ke returns to the FL value and restores the
elastic scattering channels. This high-temperature puddle energy, consistent with previous reports (31). WF law. In fact, away from the CNP, the WF law
limit occurs when the electron-phonon scattering Above the disorder temperature scale Tdis ~ 40 K, holds for a wide temperature range, consistent
rate becomes comparable to the electron-electron nmin increases as Tbath is raised, which suggests with previous reports (35, 36, 39) (Fig. 1E). For this
scattering rate. These two temperatures set the that thermal excitations begin to dominate and FL regime, we verify the WF law up to T ~ 80 K.
experimental window in which the DF and the the sample enters the nondegenerate regime near Our observation of the breakdown of the WF
breakdown of the WF law can be observed. the CNP. law in graphene is consistent with the emer-
To minimize disorder, we used monolayer gra- Electronic thermal conductivity was measured gence of the DF. Figure 2 shows the full density
phene samples encapsulated in hexagonal boron using high-sensitivity Johnson noise thermometry and temperature dependence of the experimen-
nitride (31). All devices used in this study have (JNT) (32, 35). We applied a small bias current tally measured Lorenz ratio, highlighting the
two terminals, so as to keep a well-defined temper- through the sample, thus injecting a Joule heating presence of the DF. The blue region denotes
ature profile (32), with contacts fabricated using power P directly into the electronic system and L L0 , suggesting that the carriers in graphene
the one-dimensional edge technique (33) to mini- inducing a small difference between the temper- exhibit FL behavior. The WF law is violated in
mize contact resistance. We employed a back-gate ature of the graphene electrons and that of the the DF (yellow-red region), with a peak Lorenz
voltage Vg applied to the silicon substrate to tune bath: DT Te Tbath . The electron temperature ratio 22 times larger than L0 . The green dotted

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1059


R ES E A RC H | R E PO R TS

Fig. 2. Breakdown of the WF law 100 20 mined for any given sample. For simplicity, we
in the DF regime. The Lorenz ratio
assume that we are well within the DF limit,
90
is shown as a function of the charge
where lm and H are approximately independent
80 16 of n. We fit Eq. 2 to the experimentally measured
carrier density and bath temper-
ature. Near the CNP and for tem- 70 Ln for all temperatures and densities in the
peratures above the disorder 12 DF regime to obtain lm and H for each sample

L / L0
Tbath (K)
60 (Fig. 3C). lm is estimated to be 1.5, 0.6, and
(charge puddle) regime but below
the onset of electron-phonon cou- 50 0.034 mm for samples S1, S2, and S3, respec-
8 tively. For the system to be well described by
pling, the Lorenz ratio is measured 40
to be an order of magnitude
hydrodynamics, lm should be long compared
30 4 with the electron-electron scattering length of
greater than the FL value of
1 (blue). The WF law is observed to 20 ~0.1 mm that is expected for the DF at 60 K (19).
hold outside of the DF regime. All
This is consistent with the pronounced signa-
10 0 tures of hydrodynamics in S1 and S2, but not in
data from this figure are taken from 15 10 5 0 5 10 15
sample S1. Green dotted lines show
S3, where only a glimpse of the DF appears in
n (109 cm-2) this more disordered sample. We also observe in S1
the corresponding nmin(T) for this sample.
that Ln dips substantially below L0 : its mini-
mum is L 0 =3. L < 1 occurs in Eq. 2 for n n0.
Fig. 3. Disorder in the DF. (A) The inset to Fig. 3C shows the fitted enthalpy
Minimum carrier density as a function density as a function of temperature compared
of temperature for all three samples. At with that expected in clean graphene (dashed line)
low temperatures, each sample is (19), excluding renormalization of the Fermi ve-
limited by disorder. At high temper- locity. In the cleanest sample, H varies from 1.1
atures, all samples become limited by to 2.3 eV/mm2 in the hydrodynamic regime. This
thermal excitations. Dashed lines are a enthalpy density corresponds to ~20 meV or ~4kBT
guide for the eye. (B) Lorentz ratio of per charge carrierabout a factor of 2 larger than
all three samples as a function of bath the model calculation without disorder (19). The
temperature. The largest WF violation is sharp temperature and impurity dependence ob-
seen in the cleanest sample. (C) The served in L is a prediction of these hydrodynamic
gate dependence of the Lorentz ratio is models. These effects and the magnitude of L are
well fit to the hydrodynamic theory of inconsistent with alternative models for WF vio-
(5, 6). Fits of all three samples are lations, including bipolar diffusion in graphene
shown at 60 K. All samples return to (21, 42). Furthermore, recent experiments report
the FL value (black dashed line) at high monotonic behavior in thermopower as a func-
density. The inset graph shows the tion of T (43), implying that phonon drag is not
fitted enthalpy density as a function of responsible for the peak in ke that we observe as
temperature and the theoretical value a function of T.
in clean graphene (black dashed line). To fully incorporate the effects of disorder, a
The schematic inset illustrates the hydrodynamic theory that treats inhomogeneity
difference between heat and charge nonperturbatively is necessary (41, 44). The en-
current in the neutral Dirac plasma. thalpy densities reported here are larger than the
theoretical estimation obtained for disorder-free
graphene, consistent with the picture that chem-
ical potential fluctuations prevent the sample from
lines show the corresponding nmin(T ) for this relativistic hydrodynamic theory of the DF, incor- reaching the Dirac point. Although we find thermal
sample; the DF is found within this regime, indi- porating the effects of weak impurity scattering conductivity well described by (5, 6), electrical
cating the coexistence of thermally populated (5, 6, 18) conductivity increases more slowly than expected
electrons and holes. Disorder and electron-phonon LDF away from the CNP, a result consistent with hydro-
scattering serve as lower and upper limits, respec- L 2 dynamic transport in a viscous fluid with charge
1 n=n0 2 2
tively, on the temperature range over which the puddles (41).
DF can be observed. In a hydrodynamic system, the ratio of shear
We investigate the effect of impurities on hydro- viscosity h to entropy density s is an indicator of the
dynamic transport by comparing the results ob- where strength of the interactions between constituent
tained from samples with varying disorder. Figure particles. It has been suggested that the DF can
H vF lm
3A shows nmin as a function of temperature for LDF 3 behave as a nearly perfect fluid (19): h/s approaches
three samples used in this study. nmin(T = 0) is T 2 smin Kovtun et al.s conjecture that h=s==kB 1=4p
estimated as 5 109, 8 109, and 10 109 cm2, and (where is Plancks constant divided by 2p) for a
respectively, in samples S1, S2, and S3. All devices strongly interacting system (45). A nonperturbative
H smin
show qualitatively similar DF behavior: For sam- n20 4 hydrodynamic framework can be employed to es-
ples S1, S2, and S3, the largest values of L=L0 e2 vF lm timate h (41). A direct measurement of h is of great
measured in the DF regime are 22, 12, and 3, interest.
respectively (Fig. 3B). Cleaner samples not only Here, vF is the Fermi velocity in graphene, smin is Beyond a diverging thermal conductivity and
have a more pronounced peak but also have a the electrical conductivity at the CNP, H is the an ultralow viscosity, other peculiar phenomena
narrower density dependence (Fig. 3C), as pre- fluid enthalpy density, lm is the momentum re- are expected to arise in this plasma. The massless
dicted (5, 6). laxation length from impurities, and n0 is the nature of the Dirac fermions is expected to result
More quantitative analysis of Ln in our ex- density scale over which L varies appreciably. in a large kinematic viscosity, despite a small shear
periment can be performed by employing a quasi- Two parameters in Eq. 2, lm and H , are undeter- viscosity h. Observable hydrodynamic effects have

1060 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

also been predicted to extend into the FL regime ACKN OWLED GMEN TS from the Elemental Strategy Initiative conducted by the Ministry
(46). The study of magnetotransport in the DF will We thank M. Foster, D. Efetov, and G.-H. Lee for helpful discussions. of Education, Culture, Sports, Science and Technology, Japan.
T.T. acknowledges support from a Grant-in-Aid for Scientific
lead to further tests of hydrodynamics (5, 18). The major experimental work at Harvard University is supported
Research (grant 262480621) and a grant on Innovative Areas
by the U.S. Department of Energy (grant DE-SC0012260) and
at Raytheon BBN Technologies by Internal Research and Nano Informatics (grant 25106006) from the Japan Society for
Development. J.C. acknowledges support from the Function the Promotion of Science. T.A.O. and K.C.F. acknowledge
RE FE RENCES AND N OT ES support from Raytheon BBN Technologies. This work was
Accelerated nanoMaterial Engineering (FAME) Center, sponsored
1. L. P. Kadanoff, P. C. Martin, Ann. Phys. 24, 419469 by Semiconductor Research Corporation MARCO and Defense performed, in part, at the Center for Nanoscale Systems (CNS),
(1963). Advanced Research Projects Agency. K.W. is supported by Army a member of the National Nanotechnology Infrastructure
2. M. J. M. de Jong, L. W. Molenkamp, Phys. Rev. B 51, Research Office (ARO) Multidisciplinary University Research Network, which is supported by the NSF under award
1338913402 (1995). Initiative (MURI) (grant W911NF-14-1-0247). J.K.S. is supported by no. ECS-0335765. CNS is part of Harvard University.
3. C. Cao et al., Science 331, 5861 (2011). ARO (grant W911NF-14-1-0638) and the Agency for Science,
4. E. Shuryak, Prog. Part. Nucl. Phys. 53, 273303 (2004). Technology and Research (A*STAR). P.K. acknowledges partial
SUPPLEMENTARY MATERIALS
5. M. Mller, L. Fritz, S. Sachdev, Phys. Rev. B 78, 115406 support from the Gordon and Betty Moore Foundations EPiQS
(2008). Initiative (grant GBMF4543) and the Nano Material Technology www.sciencemag.org/content/351/6277/1058/suppl/DC1
6. M. Foster, I. Aleiner, Phys. Rev. B 79, 085415 (2009). Development Program through the National Research Foundation Materials and Methods
7. S. S. Apostolov, A. Levchenko, A. V. Andreev, Phys. Rev. B 89, of Korea (grant 2012M3A7B4049966). A.L. and S.S. are Figs. S1 to S8
121104 (2014). supported by the NSF under grant DMR-1360789, the Templeton Table S1
8. B. Narozhny, I. Gornyi, M. Titov, M. Schtt, A. Mirlin, Phys. Rev. Foundation, and MURI grant W911NF-14-1-0003 from ARO. References (4759)
B 91, 035414 (2015). Research at the Perimeter Institute for Theoretical Physics is
9. K. S. Novoselov et al., Nature 438, 197200 (2005). supported by the Government of Canada through Industry 16 July 2015; accepted 6 January 2016
10. Y. Zhang, Y. W. Tan, H. L. Stormer, P. Kim, Nature 438, Canada and by the Province of Ontario through the Ministry of Published online 11 February 2016
201204 (2005). Research and Innovation. K.W. and T.T. acknowledge support 10.1126/science.aad0343
11. D. A. Siegel, W. Regan, A. V. Fedorov, A. Zettl, A. Lanzara, Phys.
Rev. Lett. 110, 146802 (2013).
12. D. E. Sheehy, J. Schmalian, Phys. Rev. Lett. 99, 226803
(2007).
13. S. Sachdev, B. Keimer, Phys. Today 64, 29 (2011).
ELECTRON TRANSPORT
14. C. H. Lui, K. F. Mak, J. Shan, T. F. Heinz, Phys. Rev. Lett. 105,

Downloaded from on March 3, 2016


127404 (2010).
15. M. Breusing, C. Ropers, T. Elsaesser, Phys. Rev. Lett. 102,
086809 (2009).
Evidence for hydrodynamic electron
16. K. J. Tielrooij et al., Nat. Phys. 9, 248252 (2013).
17. J. C. Johannsen et al., Phys. Rev. Lett. 111, 027403 (2013).
18. S. A. Hartnoll, P. Kovtun, M. Mller, S. Sachdev, Phys. Rev. B
flow in PdCoO2
76, 144502 (2007).
19. M. Mller, J. Schmalian, L. Fritz, Phys. Rev. Lett. 103, 025301 Philip J. W. Moll,1,2,3 Pallavi Kushwaha,3 Nabhanila Nandi,3
Burkhard Schmidt,3 Andrew P. Mackenzie3,4*
(2009).
20. N. W. Ashcroft, N. D. Mermin, Solid State Physics (Brooks Cole,
ed. 1, 1976).
21. Materials and methods are available as supplementary Electron transport is conventionally determined by the momentum-relaxing scattering of
materials on Science Online. electrons by the host solid and its excitations. Hydrodynamic fluid flow through
22. N. Wakeham et al., Nat. Commun. 2, 396 (2011).
23. R. P. Smith et al., Nature 455, 12201223 (2008).
channels, in contrast, is determined partly by the viscosity of the fluid, which is governed
24. M. A. Tanatar, J. Paglione, C. Petrovic, L. Taillefer, Science 316, by momentum-conserving internal collisions. A long-standing question in the physics
13201322 (2007). of solids has been whether the viscosity of the electron fluid plays an observable role in
25. R. W. Hill, C. Proust, L. Taillefer, P. Fournier, R. L. Greene, determining the resistance. We report experimental evidence that the resistance of
Nature 414, 711715 (2001).
26. L. Fritz, J. Schmalian, M. Mller, S. Sachdev, Phys. Rev. B 78,
restricted channels of the ultrapure two-dimensional metal palladium cobaltate (PdCoO2)
085416 (2008). has a large viscous contribution. Comparison with theory allows an estimate of the
27. S. Adam, E. H. Hwang, V. M. Galitski, S. Das Sarma, Proc. Natl. electronic viscosity in the range between 6 103 kg m1 s1 and 3 104 kg m1 s1,
Acad. Sci. U.S.A. 104, 1839218397 (2007). versus 1 103 kg m1 s1 for water at room temperature.
28. J. Martin et al., Nat. Phys. 4, 144148 (2008).

I
29. Y. Zhang, V. W. Brar, C. Girit, A. Zettl, M. F. Crommie, Nat.
Phys. 5, 722726 (2009). n a quantum fluid without an associated lat- lisions in the electron fluid (electron-impurity, nor-
30. J. Xue et al., Nat. Mater. 10, 282285 (2011).
tice, such as 3He, the momentum of the fluid mal electron-phonon, Umklapp electron-electron,
31. C. R. Dean et al., Nat. Nanotechnol. 5, 722726
(2010). is conserved except where it interacts with and Umklapp electron-phonon) relax momentum,
32. K. C. Fong, K. C. Schwab, Phys. Rev. X 2, 031006 the walls of a channel through which it is flow- taking the fluid far from the hydrodynamic limit.
(2012). ing. As the temperature decreases and the At least some of these momentum-relaxing col-
33. L. Wang et al., Science 342, 614617 (2013).
quasiparticle-quasiparticle mean free path with- lisions are inevitable in any real material. Strictly
34. N. J. G. Couto et al., Phys. Rev. X 4, 041019 (2014).
35. J. Crossno, X. Liu, T. A. Ohki, P. Kim, K. C. Fong, Appl. Phys. in the fluid increases (because of the decrease of speaking, momentum of the electron fluid can
Lett. 106, 023121 (2015). its quasiparticle scattering rate), interactions with never be conserved, even in a bulk sample for which
36. K. C. Fong et al., Phys. Rev. X 3, 041008 (2013). the walls become more probable and the viscosity boundary scattering is negligible. However, this
37. A. C. Betz et al., Phys. Rev. Lett. 109, 056805
and flow resistance increase. This is intuitively at does not mean that the electronic viscosity plays
(2012).
38. C. B. McKitterick, D. E. Prober, M. J. Rooks, Phys. Rev. B 93, odds with the behavior seen for electrons moving no role in determining electrical resistance. A
075410 (2016). in a crystalline lattice, whose flow resistance de- pragmatic benchmark is whether momentum-
39. S. Yien, A. R. Champagne, Nano Lett. 14, 289293 creases as increases. The resolution of this appar- conserving processes are faster or slower than
(2014).
ent paradox is that coupling to the lattice and its momentum-relaxing ones. If the electron fluids
40. A. Laitinen et al., Nano Lett. 14, 30093013 (2014).
41. A. Lucas, J. Crossno, K. C. Fong, P. Kim, S. Sachdev, Phys. Rev. excitations means that the large majority of col- momentum is relaxed slowly, it can be thought of
B 93, 075426 (2016). as being quasi-conserved, and hydrodynamic sig-
42. H. Yoshino, K. Murata, J. Phys. Soc. Jpn. 84, 024601 1 natures might be observable (19).
(2015). Laboratory for Solid State Physics, ETH Zrich, CH-8093
43. F. Ghahari et al., http://arxiv.org/abs/1601.05859 Zrich, Switzerland. 2Department of Physics, University of The search for hydrodynamic effects in elec-
(2016). California, Berkeley, CA 94720, USA. 3Max Planck Institute trons in solids has been given extra impetus by
44. A. Lucas, New J. Phys. 17, 113007 (2015). for Chemical Physics of Solids, 01187 Dresden, Germany.
4
the introduction of the holographic correspon-
45. P. K. Kovtun, D. T. Son, A. O. Starinets, Phys. Rev. Lett. 94, Scottish Universities Physics Alliance, School of Physics and
Astronomy, University of St. Andrews, St. Andrews KY16
dence to condensed matter physics (10). This
111601 (2005).
46. A. Principi, G. Vignale, M. Carrega, M. Polini, http://arxiv.org/ 9SS, UK. technique introduced the concept of a mini-
abs/1506.06030 (2015). *Corresponding author. E-mail: andy.mackenzie@cpfs.mpg.de mum viscosity, argued to be applicable to strongly

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1061


R ES E A RC H | R E PO R TS

Fig. 1. FIB-prepared devices of demonstrated by direct nonequilibrium meas-


PdCoO2 crystals. (A) The crystal used urements (12).
for our channel-thinning experiment, To investigate whether a hydrodynamic con-
Thinning cuts
after the first two processing steps. An tribution to electrical transport can be clearly
initial channel 120 mm wide has been separated from the more standard contribu-
reprocessed to produce a narrower tions from momentum-relaxing processes, we
conducting channel of width 60 mm. The sought a material in which momentum-relaxing
same channel was subsequently scattering is anomalously suppressed. The ma-
reprocessed seven further times, terial that we chose was PdCoO2, a layered metal
narrowing it by approximately half each with a number of unusual properties (1321). Its
time, until it was 0.7 mm wide. electronic structure is deceptively simple, with
(B) A meander channel processed in a one highly dispersive band, dominantly of Pd
second crystal for use in a search for the 4d/5s character, crossing the Fermi level (2226).
Shubnikhovde Haas effect. Its Its Fermi volume corresponds to one electron
approximate width is 6 mm in the per formula unit to high accuracy (18), and the
narrowed parts; thickness is 17 mm. Its ratio of in-plane to out-of-plane resistivity is
resistance in a magnetic field of 14 T at a approximately 103, justifying the use of a two-
temperature of 2 K is 5.7 103 ohms. dimensional approximation in treating the in-
plane properties.
100 m The electrical conductivity of PdCoO2 is re-
Ag epoxy
markable. At room temperature, its resistivity is
just 2.6 106 ohmcm, 30% lower per carrier than
that of elemental copper. Below 15 K, the resis-
Fig. 2. Effect of channel
tivity is essentially independent of temperature;
width W on magnetotran- A in the best single crystals, the resistivity is below
sport. (A) Magnetoresistance
1 108 ohmcm (18). This striking behavior
data from the sample shown
might be attributable to phonon drag, in which
in Fig. 1A taken at 2 K after
the phonons follow the electrons into an out-of-
successive channel-thinning
equilibrium distribution when an electric field is
steps. (B) Magnetoresistance
applied. In PdCoO2, the activation temperature
data from the sample shown
for Umklapp electron-phonon processes is at least
in Fig. 1B taken at 2 K.
160 K, unusually high for a metal (18).
(C) Bmax [indicated by arrows
These properties make PdCoO2 a good candi-
in (A)] varies as a(kF/We) as
date for a search for hydrodynamic effects. Be-
the device shown in Fig. 1A is B low 15 K, momentum-relaxing processes are far
successively thinned. Both the
slower than those observed in most metals. Bet-
functional form and prefactor
ter still, if phonon drag is indeed taking place,
are in good agreement with
the normal electron-phonon processes that usu-
theory and with previous
ally impede electrons in solids from approaching
measurements on semi-
the hydrodynamic limit are now helping that
conductor wires.
process instead, because they contribute a source
of momentum-conserving scattering. An impor-
tant independent hint that the balance between
momentum-conserving and momentum-relaxing
scattering is both unusual and favorable for the
observation of hydrodynamic effects comes from
C the ratio of the mean free paths deduced, respec-
tively, from the resistivity, which is sensitive to
processes that efficiently relax momentum, and
from analysis of the de Haasvan Alphen (dHvA)
effect, whose amplitude is sensitive to a wider
range of scattering processes. The inverse of that
ratio is 5 to 10%, an unusually small value ap-
proximately an order of magnitude lower than
that seen in ordinary metals (18).
In a purely hydrodynamic fluid, flow resistance
in channels is determined entirely by momentum-
relaxing boundary scattering, the efficiency of
which is determined by the fluids viscosity. Bound-
interacting fluids as diverse as the quark-gluon that unambiguously separates the two effects. ary scattering also contributes to the resistance in
plasma and cold atomic gases (11). Hydrody- In a pioneering experiment, unusual current- thin wires or channels of metals in which hydro-
namic effects have also been postulated to be voltage relationships in a semiconductor wire dynamic effects play no observable role, but the
at the root of the temperature-linear resistiv- were convincingly ascribed to hydrodynamic standard theory of this phenomenon is well es-
ity of the high-temperature superconductors effects (3), but that avenue of research has not tablished (27). We therefore set out to construct
(6, 7), but because momentum-relaxing scattering been widely pursued, even though the large dif- a series of PdCoO2 wires, study how their resist-
is strong in those materials, it is difficult to ference between transport and electron-electron ance varied with the channel width, and investi-
perform an analysis of the experimental data scattering rates in semiconductors was later gate whether there were observable deviations

1062 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

from the predictions of standard theory. The wires Fig. 3. Shubnikhovde Haas
were produced from flux-grown single crystals oscillations. (A and B) SdH
by focused ion beam (FIB) etching (Fig. 1A). Six oscillations from the pat-
such devices were made and shown to have con- terned meander track shown
sistent properties (28). We discuss the nature of in Fig. 1B. (C) The frequencies
the boundaries produced by FIB processing in of the SdH oscillations
(28) and show that damage is restricted to ap- extracted by Fourier analysis
proximately 20 nm from the edge of the channel. of the data in (A).
Taking all uncertainties into account, the un-
damaged channel width can be determined to
an accuracy of 80 nm. For the main experiment,
measurements were done on a single successively
etched wire from the same crystal, to remove as
many experimental uncertainties as possible. A
second crystal (Fig. 1B) was etched into a mean-
der channel ideally suited for a measurement of
the Shubnikhovde Haas (SdH) effect. For each
experiment, we studied the magnetoresistance of
the wire in magnetic fields B ranging from 14 T
to +14 T. We also fabricated a multicontact device
to verify that our data are length-independent at
constant width (28).
Data obtained in the channel-narrowing ex-
periment are shown in Fig. 2A for widths rang- Fig. 4. Hydrodynamic
ing from 60 mm to 0.7 mm at a measurement effect on transport.
temperature of 2 K. Data from the meander chan- (A and B) The measured
nel are shown in Fig. 2B. Consistent with previous resistivity of PdCoO2
measurements on other single crystals (18), the channels normalized to
resistivity in zero field for the 60-mm sample is that of the widest channel
9 109 ohmcm. The value of the momentum- (r0), plotted against the
relaxing mean free path MR is a crucial param- inverse channel width
eter in the analysis of the resistance of restricted 1/W multiplied by the bulk
channels; the single-band electronic structure momentum-relaxing mean
and well-known Fermi surface shape and volume free path MR (solid black
(16, 18) allow an accurate calculation of MR = circles). Blue solid line:
18.5 + 1.5 mm (28). The wire widths W used for Prediction of a standard
the experiment therefore cover the range 0.3 Boltzmann theory includ-
MR/W 26, enabling study of the crossover be- ing boundary scattering
tween a nearly bulk regime and one in which a but neglecting momentum-
sample dimension falls far below the bulk mean conserving collisions. Red
free path. line: Prediction of a model
We adopt the usual solid-state physics conven- that includes the effects of
tion of describing the transport properties of our momentum-conserving
channels in terms of the resistivity r. For a chan- scattering (see text). Error
nel of width W, length L, thickness T, and bars stem from uncertain-
resistance R, r RTW/L. Conceptually, r is a ties of sample dimension
bulk property of the material, so in the absence (28). (C) Predictions of the
of boundary effects it should be independent of hydrodynamic theory over a
W. In contrast, at low fields, the overall channel wide range of parameter
resistivity r increases by more than an order of space.
magnitude as the wire width is decreased (Fig.
2). Because this involves repeated exposure to
ion beam etching, it is natural to wonder whether
this trend is caused by beam damage increasing
the scattering in the bulk of the wire. However,
extending the data to higher fields proves that
this is not the case. First, we note that at high cause the helical pitch of the motion of the (31). The data shown in Fig. 2C therefore provide
fields, the resistivity is similar at all widths above drifting electrons becomes so tight that boundary further evidence that the overall scale of the
0.7 mm. Second, the pronounced maximum seen scattering is suppressed, and eventually the bulk, resistivity is increasing because of additional
at fields Bmax in the magnetoresistance for W width-independent resistivity is recovered at boundary scattering rather than additional bulk
30 mm is a well-known phenomenon from the high fields. Specifically, Bmax = a(kF /We) (where scattering. Very low field peaks in the mag-
study of narrow conducting channels for which is Plancks constant divided by 2p, kF is the Fermi netoresistance can still be seen in the 30-mm
the bulk mean free path is on the order of the wave vector, and e is the electronic charge) with channel, providing further direct geometrical
channel width or larger (2931). For each chan- constant a = 0.55 has been reported for re- evidence that MR is long. Further evidence
nel width, the rise in the magnetoresistivity at stricted channels of two-dimensional electron that extremely high bulk purity is retained after
low fields is stopped when the cyclotron orbit gas (30), and a similar functional form with a = ion beam etching comes from the data from the
radius falls to less than the channel width, be- 0.9 was observed in recent work on graphene meander sample. Its Bmax is consistent with the

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1063


R ES E A RC H | R E PO R TS

measured width of 6 mm (Fig. 2, B and C), and at prediction of the standard theory. At its other REFERENCES AND NOTES
high fields, clear SdH oscillations are seen, with limit of strong momentum-conserving scatter- 1. R. N. Gurzhi, Zh. Eksp. Teor. Fiz. 44, 771 (1963) [Sov. Phys. 17,
frequencies in agreement with those seen in ing (black line), the prediction for r/r0 is ap- 521 (1963)].
2. R. N. Gurzhi, Usp. Fiziol. Nauk 94, 689 (1968) [Sov. Phys. Usp.
dHvA oscillations from bulk crystals (18). High proximately quadratic as a function of MR/W. In 11, 255 (1968)].
frequencies are particularly hard to see by the this regime, viscous effects dominate, and the 3. L. W. Molenkamp, M. J. M. de Jong, Phys. Rev. B 49,
SdH effect; the 30-kT frequencies shown in Fig. W 2 dependence of r is equivalent to the W 3 50385041 (1994).
3B are among the highest ever reported in SdH prediction for flow resistance that is obtained 4. M. J. M. de Jong, L. W. Molenkamp, Phys. Rev. B 51,
1338913402 (1995).
measurements. from the purely hydrodynamic Navier-Stokes 5. B. Spivak, S. A. Kivelson, Ann. Phys. 321, 20712115
The data presented in Figs. 2 and 3 provide equation (28). (2006).
strong evidence that boundary rather than bulk Our value for MC/MR (red line in Fig. 4C) sits 6. R. A. Davison, K. Schalm, J. Zaanen, Phys. Rev. B 89, 245116
scattering dominates the rise in resistivity seen between these limiting cases. Because viscosity is (2014).
7. S. A. Hartnoll, Nat. Phys. 11, 5461 (2015).
as we reduce the channel width. As discussed inversely proportional to MC, the initial ~W 2 rise 8. A. V. Andreev, S. A. Kivelson, B. Spivak, Phys. Rev. Lett. 106,
above, boundary scattering is expected even in is steeper than for MC/MR = 0.01. As MR/W rises 256804 (2011).
the absence of hydrodynamic effects as the chan- toward 10, MC is no longer much less than W, and 9. R. Mahajan, M. Barkeshli, S. A. Hartnoll, Phys. Rev. B 88,
nel width falls to less than the mean free path the Navier-Stokes prediction evolves smoothly to 125107 (2013).
10. S. A. Hartnoll, P. K. Kovtun, M. Mller, S. Sachdev, Phys. Rev. B
and the system enters the ballistic transport re- a solution in which viscous effects are important 76, 144502 (2007).
gime. The relevant theory (4, 27) can be expressed but the channel constriction is such that we leave 11. P. K. Kovtun, D. T. Son, A. O. Starinets, Phys. Rev. Lett. 94,
in a useful dimensionless form, shown as a blue the purely hydrodynamic regime. Physically the 111601 (2005).
line in Fig. 4. If r is normalized to the bulk resis- system is in a hybrid situation in which tradi- 12. C. P. Weber et al., Nature 437, 13301333 (2005).
13. R. D. Shannon, D. B. Rogers, C. T. Prewitt, Inorg. Chem. 10,
tivity r0 of an infinitely wide sample and plotted tional impurity and boundary scattering mix with 713718 (1971).
against MR/W, the prediction has no free param- viscosity-stimulated boundary scattering to pro- 14. M. Hasegawa, I. Inagawa, M. Tanaka, I. Shirotani, H. Takei, Solid
eters. At MR/W = 25, r is calculated to be 10.3 r0. duce the overall evolution of resistivity with State Commun. 121, 203205 (2002).
Our measured value for r/r0 is more than 50% channel width. Fig. 4C also shows that the hy- 15. H. Takatsu et al., J. Phys. Soc. Jpn. 76, 104701
(2007).
larger than this prediction, and our data also drodynamic prediction is insensitive to the precise 16. H.-J. Noh et al., Phys. Rev. Lett. 102, 256404 (2009).
show a functional form that is at odds with the choice of MC/MR. Our choice of 0.1 was not the 17. H. Takatsu, S. Yonezawa, S. Fujimoto, Y. Maeno, Phys. Rev.
simple theory. result of fitting, but simply an estimate motivated Lett. 105, 137201 (2010).
18. C. W. Hicks et al., Phys. Rev. Lett. 109, 116401
To examine whether the large deviations of by the ratio of scattering rates deduced from
(2012).
the data from the predictions of standard trans- measurements of resistivity and the dHvA effect. 19. H. Takatsu et al., Phys. Rev. Lett. 111, 056601 (2013).
port theory are linked to electronic hydrodynam- Choices that were larger or smaller by a factor of 20. R. Daou, R. Frsard, S. Hbert, A. Maignan, Phys. Rev. B 91,
ics, we carefully studied the predictions of a 2 would give a similar level of agreement with 041113 (2015).
21. N. Kikugawa et al., http://arxiv.org/abs/1412.5168
more sophisticated theory that takes momentum- the data. (2015).
conserving scattering into account (4). Originally The data and predictions shown in Fig. 4 and 22. M. Hasegawa et al., Mater. Trans. 42, 961964 (2001).
formulated to analyze the current-dependent hy- discussed in (28) provide strong evidence that 23. V. Eyert, R. Frsard, A. Maignan, Chem. Mater. 20, 23702373
drodynamic signatures reported in (3), the the- we have observed a substantial hydrodynamic (2008).
24. K. Kim, H. C. Choi, B. I. Min, Phys. Rev. B 80, 035116
ory encodes momentum-relaxing scattering via contribution to electrical transport in a bulk (2009).
the role of impurities and momentum-conserving material. Further analysis of the theory in its 25. K. P. Ong, J. Zhang, J. S. Tse, P. Wu, Phys. Rev. B 81, 115120
scattering via normal electron-electron scattering Navier-Stokes limit allows a quantitative esti- (2010).
processes. In PdCoO2, the scale of momentum- mate of the electronic viscosity itself as a func- 26. K. P. Ong, D. J. Singh, P. Wu, Phys. Rev. Lett. 104, 176601
(2010).
conserving electron-electron scattering is un- tion of the hydrodynamic contribution to r/r0. 27. C. W. J. Beenakker, H. van Houten, Solid State Phys. 44, 1228
certain because of Fermi surface faceting (26), As a result, we are able to estimate the dynamic (1991).
and phonons dragged out of equilibrium are viscosity h of the electronic fluid in PdCoO2 as 28. See supplementary materials on Science Online.
also likely to be making a contribution to the lying in the range between 6 103 kg m1 s1 and 29. E. Ditlefsen, J. Lothe, Philos. Mag. 14, 759773
momentum-conserving processes. In this sense, 3 104 kg m1 s1. For comparison, the h values (1966).
30. T. J. Thornton, M. L. Roukes, A. Scherer, B. P. van de Gaag,
even this more sophisticated theory is oversim- of two well-known fluids, water at room tem- Phys. Rev. Lett. 63, 21282131 (1989).
plified, and it would be dangerous to use it to pre- perature and liquid nitrogen at 75 K, are 1 31. S. Masubuchi et al., Phys. Rev. Lett. 109, 036601
dict temperature-dependent transport in PdCoO2. 103 kg m1 s1 and 1 104 kg m1 s1, respectively. (2012).
However, its results at any fixed temperature de- Although the bounds we can place on the elec- 32. J. A. N. Bruin, H. Sakai, R. S. Perry, A. P. Mackenzie, Science
339, 804807 (2013).
pend only on the ratio of a momentum-conserving tronic viscosity of PdCoO2 are not very precise, we
mean free path MC to the momentum-relaxing stress that this is an issue of theory, not experi- AC KNOWLED GME NTS
mean free path MR, and the microscopic origin ment. We believe that there is such a large hydro- We thank J. C. Davis, C. Geibel, A. G. Green, S. A. Hartnoll,
of the scattering that produced that ratio is not dynamic contribution to our data that they contain C. W. Hicks, S. A. Kivelson, Y. Maeno, T. Oka, J. W. Orenstein,
relevant. The theory is therefore a useful guide all the information required to estimate the and S. H. Simon for stimulating discussions and the UK
Engineering and Physical Sciences Research Council
to the consequences of including hydrodynamic viscosity precisely, and we hope that our experi- for financial support. The FIB work was supported by the
effects in experiments performed at constant ment motivates further work on this issue. As SCOPE-M center for electron microscopy at ETH Zrich.
temperature, such as the one shown in Fig. 4, A discussed above, PdCoO2 is a weakly scattering We thank P. Gasser, J. Reuteler, and B. Batlogg for
and B. For any value of MC/MR, it predicts a system, so it is not surprising that a comparison of FIB support. The data and/or materials supporting
this publication can be accessed at http://dx.doi.org/10.17630/
unique functional form and overall magnitude our estimate of h with measurements of the en- cdc6c488-2ea4-440e-a68d-e2a426709171.
for r/r0 versus MR /W, with no free fitting tropy density s (15, 18) yields h /s 106 /kB, far
parameters. from a proposed minimum viscosity limit (11).
As can be seen from the red lines in Fig. 4, A It will also be interesting to reexamine a pos- SUPPLEMENTARY MATERIALS
and B, the hydrodynamic theory for MC/MR = 0.1 sible role of hydrodynamic effects in explaining www.sciencemag.org/content/351/6277/1061/suppl/DC1
produces an excellent match to our data. In Fig. 4C the resistivity in systems in which the momentum- Materials and Methods
Figs. S1 to S6
we show how the predictions of the theory change conserving scattering is extremely strong (6, 7, 32). References (3339)
as a function of MC/MR, marking the blue and red In principle, a range of viscosities is to be ex-
lines of Fig. 4, A and B, on the contour plot for ref- pected in different electronic fluids, and even the 24 June 2015; accepted 23 December 2015
erence. For large MC/MR (i.e., weak momentum- attainment of turbulent electronic flow may be Published online 11 February 2016
conserving scattering), it limits rapidly to the possible. 10.1126/science.aac8385

1064 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

CATALYSIS which separates CO activation and CC coupling


onto two different types of active sites with com-
plementary properties. CO and H2 are activated
Selective conversion of syngas to over a partially reduced oxide (ZnCrOx) surface,
whereas CC coupling is controlled within the

light olefins confined environment of zeolite pores with acidic


sites. As a result, a C2C4 selectivity up to 94%
(including 80% C2=C4= and 14% C2oC4o) of all
Feng Jiao,1* Jinjing Li,1* Xiulian Pan,1* Jianping Xiao,1 Haobo Li,1 Hao Ma,1 hydrocarbons (carbon atombased) was achieved
Mingming Wei,1 Yang Pan,2 Zhongyue Zhou,2 Mingrun Li,1 Shu Miao,1 Jian Li,1 with only 2% methane at a CO conversion of 17%.
Yifeng Zhu,1 Dong Xiao,1 Ting He,1 Junhao Yang,1 Fei Qi,2 Qiang Fu,1 Xinhe Bao1 This C2C4 selectivity is far beyond the maximum
predicted by the ASF model in FTS.
Although considerable progress has been made in direct synthesis gas (syngas) conversion The composite catalyst contained an oxide
to light olefins (C2=C4=) via Fischer-Tropsch synthesis (FTS), the wide product distribution (ZnCrOx) that exhibits a typical spinel structure
remains a challenge, with a theoretical limit of only 58% for C2C4 hydrocarbons. We (fig. S1) and a mesoporous SAPO zeolite (MSAPO)
present a process that reaches C2=C4= selectivity as high as 80% and C2C4 94% at exhibiting CHA structure with a hierarchical pore
carbon monoxide (CO) conversion of 17%. This is enabled by a bifunctional catalyst texture (figs. S1 and S2 and table S1) (9). Figure 1A
affording two types of active sites with complementary properties. The partially reduced shows that C2=C4= selectivity reached 74% at CO
oxide surface (ZnCrOx) activates CO and H2, and CC coupling is subsequently conversion of 16% over a catalyst with a mass ratio
manipulated within the confined acidic pores of zeolites. No obvious deactivation is of ZnCrOx/MSAPO = 1.4, under reaction condi-
observed within 110 hours. Furthermore, this composite catalyst and the process may allow tions of H2/CO = 1.5, pressure 2.5 MPa, and space
use of coal- and biomass-derived syngas with a low H2/CO ratio. velocity 4800 ml/hgcat. A higher H2/CO ratio be-
nefits CO conversion, which rose to 30% at H2/

F
CO = 3, for instance, whereas a higher space velocity

Downloaded from on March 3, 2016


ischer-Tropsch synthesis (FTS) has played and ruthenium, have been tested (6). However, favored the selective formation of olefins. Selec-
an important role as a gas-to-liquid tech- application of this technology is still limited by tivities varied in the range from 67 to 80% for
nology, producing synthetic lubricants and low olefin selectivity and high methane selectiv- C2=C4= and 81 to 94% for C2C4 in the studied
synthetic fuels from coal, natural gas, or bio- ity, as well as severe carbon deposition. These H2/CO (0.5 to 3.0) and space velocity ranges
mass since it was developed almost a cen- drawbacks arise from the FTS reaction mecha- (1285 to 7714 ml/hgcat) and different ratios of
tury ago. It is also the only effective technology to nism, which is generally accepted to proceed via ZnCrOx/MSAPO. This C2=C4= selectivity is high-
date for direct conversion of synthesis gas (syngas) surface polymerization of CHx (x = 1, 2, or 3)i.e., er than the best value reported for FTTO (61%) (4)
to light olefinsi.e., C2=C4=olefins contain- addition of CHx monomer units to the adsorbed and exceeds the maximum predicted for C2C4
ing two to four carbon atoms, so called Fischer- alkyl species on open metal surfaces. Thus, a wide hydrocarbons according to the ASF distribution
Tropsch to olefins (FTTO) (15). FTS has been distribution of hydrocarbons with different chain in typical FTS, as depicted in Fig. 1B. Both CH4
under extensive study for more than 50 years, and lengths is produced. It can be described by the and C5+ selectivities were <5% (Fig. 1A), promi-
a variety of metal catalysts, including iron, cobalt, Anderson-Schulz-Flory (ASF) model (7, 8), which nently lower than those in FTTO (in a range of 10
predicts that selectivity of C2 to C4 hydrocarbons to 40% for CH4 and >10% for C5+) (4). In contrast,
1 (C2C4), including C2=C4= olefins and C2oC4o the C1 and C5+ products were difficult to suppress
State Key Laboratory of Catalysis, 2011-Collaborative
Innovation Center of Chemistry for Energy Materials, Dalian paraffins, does not exceed 58% (6). simultaneously in FTTO; causing one to decline
Institute of Chemical Physics, Chinese Academy of Sciences, Thus, the key challenge of selective forma- caused the other to increase (1, 7). The reactivity
Zhongshan Road 457, Dalian 116023, China. 2National tion of light olefins from syngas lies in precise of this composite catalyst showed good repro-
Synchrotron Radiation Laboratory, University of Science and
control of CC coupling while suppressing over- ducibility, with CO conversion and selectivity
Technology of China, Hefei 230029, China.
*These authors contributed equally to this work. Corresponding hydrogenation and methane formation. We pre- fluctuating within 4 and 5%, respectively, among
author. E-mail: panxl@dicp.ac.cn (X.P.); xhbao@dicp.ac.cn (X.B.) sent here a process named OX-ZEO (Oxide-Zeolite), 11 tests (fig. S3). Furthermore, it delivered rather

Fig. 1. Catalytic process of OX-ZEO. (A) CO conversion and product distribution at different H2/CO ratios in syngas over a catalyst with a mass ratio of ZnCrOx/
MSAPO = 1.4 at a space velocity of 4800 ml/hgcat. (B) Hydrocarbon distribution in OX-ZEO in comparison to that reported for FTTO (4) and that in FTS predicted
by the ASF model at a chain growth probability of 0.46, with the yellow bar representing selectivity of C2C4 hydrocarbons. (C) A stability test of a composite
catalyst with ZnCrOx/MSAPO ratio = 0.9 at 6828 ml/hgcat and H2/CO of 2.5.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1065


R ES E A RC H | R E PO R TS

Fig. 2. Bifunctionality of the composite ZnCrOx/MSAPO catalyst and sequence and separated by quartz wool; Mode 4, MSAPO and ZnCrOx well
investigation of the reaction intermediate. (A) Reaction results over the mixed. (B) In situ NAP-XPS Zn3d spectra of ZnCrOx under different conditions.
catalysts with the two functionalities packed in different modes under the same (C) In situ NAP-XPS C1s spectra of ZnCrOx exposed to H2, CO, and H2 se-
conditions. The picture on the right side displays the catalyst beds after re- quentially, and then again to CO atmosphere under different conditions. (D) In
action, representing Mode 1 to 4 from left to right: Mode 1, the catalyst contains situ study of syngas conversion over ZnCrOx by SVUV-PIMS at hu = 9.72 eV.The
ZnCrOx only; Mode 2, MSAPO packed below ZnCrOx, separated by an inert insets display the signals of m/z = 42.01 (ketene) and m/z = 42.05 (propene)
layer of quartz wool; Mode 3, MSAPO and ZnCrOx packed in an alternating detected at hu = 9.72 and 11.40 eV, respectively.

stable performance (Fig. 1C). The total C2C4 69% containing 23% C2=C4=, whereas CH4 se- zeolite (fig. S4) (9), which lends further support
selectivity remained >90%, and that of C2=C4= lectivity dropped to 26%. to the above hypothesis that the composite cata-
~78% during a 110-hour test at 400C, 2.5 MPa, This result shows that reaction intermediates lyst is bifunctional and that the reaction involves
and a space velocity of 6828 ml/hgcat. generated over the oxides must have transported intermediate transport in gas phase.
Although composite catalysts containing metal in the gas phase toward the active sites of MSAPO, Activation of the CO bond has been reported
oxides such as Cr2O3-ZnO and CuO-ZnO, and ze- where they were converted to C2C4 hydrocarbons, previously for surfaces promoted with oxygen
olites such as ZSM-5, Y, and b have been attempted rather than being hydrogenated to CH4 over the vacanciesfor example, Fe3O4 (16), CoOx/TiO2 (17),
previously for syngas conversion (1015), the prod- oxide or in gas phase. Thus, any measures that Cu/ZnO (0001) (18), and ZnGa2O4 (19). However,
ucts were mainly dimethylether (10), liquefied could facilitate the transport of intermediates in little is known about subsequent formation of
petroleum gas (C3oC4o) (1113), or gasoline (14, 15). gas phase should benefit selective formation of CC and CH bonds in the presence of H2. Den-
We attribute the efficiency of the OX-ZEO process olefins, for example, increasing space velocities, sity functional theory (DFT) calculations using the
to the bifunctionality of the catalyst, with two types as demonstrated in Fig. 2A. In addition, Mode 3, ZnCr2O4 spinel (111) surface as a model reveal that
of active sites exhibiting complementary and com- where the oxides and MSAPO were packed in an it is reducible, resulting in a surface with a num-
patible properties. In the absence of MSAPO alternating sequence and thus the transport dis- ber of oxygen vacancies (figs. S5 and S6) (9). In
(corresponding to Mode 1 of Fig. 2A), syngas was tance was reduced, also leads to enhanced C2=C4= situ near-ambient pressure x-ray photoelectron
converted mainly to CH4 (selectivity 53%), and selectivity (65% at 7714 ml/hgcat). We observed an spectroscopy (NAP-XPS) of ZnCrOx surface (Fig. 2B)
selectivity to C2C4 hydrocarbons was only 38% even higher selectivity of C2=C4= (80%), with revealed a new signal with a lower Zn3d binding
over ZnCrOx. Upon combination with MSAPO, overall C2C4 selectivity reaching 94% and only energy upon exposure to H2 and CO, compared
which was packed below the oxide and separated 2% CH4 over the well-mixed composite catalyst with the initial oxidized surface. CO can be activated
by a layer of inert quartz wool (Mode 2), the prod- (Mode 4). Furthermore, CO conversion could be on such a surface, leading to formation of CO2,
ucts shifted and C2C4 selectivity increased to tuned by varying the relative mass ratio of oxide/ detected by time-of-flight mass spectrometry

1066 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

Fig. 3. The role of MSAPO in OX-ZEO.


(A) Catalytic conversion of ketene to
olefins over MSAPO at 400C, with the
effluent monitored by SVUV-PIMS.The
inset shows the results of the blank ex-
periment with ketene fed to the same
reactor containing no catalyst. Spectra
are recorded at 10.70 eV. (B) Ratio of
olefin/paraffin as a function of the NH3
desorption peak temperature from the
medium-strength acidic sites of MSAPO.
The dashed line is only for the purpose
of guiding the eyes, and the inset dis-
plays a typical NH3-TPD profile.

(TOF-MS) (fig. S7) (9) and surface *C species in MSAPO alone as a methanol-to-olefin (MTO) cat- SAPO-34 exhibiting NH3 desorption at ~394C
absence of hydrogen, which is evidenced by a alyst also deactivated quickly within 30 hours, had a C2=C4= selectivity of 43% and an olefin/
strong C1s signal emerging at 284.7 eV, even at even at a methanol partial pressure as low as paraffin ratio (C2=C4=/C2oC4o) of 0.9. The olefin/
350C (curve II in Fig. 2C). This C1s signal 5 mbar, and 0.18 mol C/gcat was converted before paraffin ratio increased with the decreasing NH3
remained even after heat treatment at 300C deactivation. By contrast, the composite catalyst desorption temperature and reached 4.7 for a
under ultrahigh vacuum (fig. S8) (9), but it remark- delivered rather stable performance in OX-ZEO. MSAPO sample with NH3 desorption peak at
ably attenuated in H2 (curve III) and finally van- No obvious degradation of CO conversion was 350C (Fig. 3C). In contrast, a composite catalyst
ished at 450C (curve IV). Analysis of the effluent observed for operation as long as 650 hours (fig. with almost no medium strength acidic sites had
by TOF-MS shows that *C species were removed S13), and 9.6 mol C/gcat was converted within this a similar product distribution, as the catalyst con-
via hydrogenation forming hydrocarbons (fig. S7, period of time. The total coke deposit was only tained no zeolite (fig. S17 and table S3) (9). Thus, it
b to d) (9). Thus, the precursor of hydrocarbons, 11 weight % (wt %) (fig. S14) (9). In contrast, fast is reasonable to assume that there may exist an
CHx species formed over ZnCrOx, are likely the deactivation remains a major issue for MTO in optimum acidity with NH3 desorption temper-
intermediates in the OX-ZEO process. that coke deposits can mount up about 10 wt % ature <350C for olefin formation. However, it is
To detect the intermediate in gas phase, we within 15 min (25). Furthermore, very little water still a challenge to synthesize zeolites with a weaker
turned to the highly sensitive synchrotron-based is produced in OX-ZEO, in contrast to the MTO acidity but without reducing the number of active
vacuum ultraviolet photoionization mass spectrom- process, where two moles of H2O are produced sites (necessary to achieve a reasonable conver-
etry (SVUV-PIMS) (9), which has been widely for each mole of ethylene (assumed as the only sion). In addition, the preliminary results in fig.
used in the field of combustion to detect active product) from methanol (9). In addition, O from S18 (9) show that the distribution of different-sized
intermediates and radicals with low concentra- CO is removed mainly by producing CO2 with a hydrocarbons might be tunable through the shape
tions in gas phase (20, 21). It provides a tunable selectivity of 40 to 45% in OX-ZEO, whereas lit- selectivity of zeolites, because a larger pore gener-
photon energy, soft ionization, and superior signal- tle CO2 forms in MTO. The above results show ally yields higher hydrocarbons. However, further
to-noise ratio. The schemes of SVUV-PIMS and the that reaction via methanol may not be the dom- elucidation of the relation between the structure
reactor are shown in fig. S9. When syngas was fed inating pathway in OX-ZEO, although it cannot (pore size/acidity) and the activity/selectivity will
into the reactor in the absence of catalysts (a blank be excluded completely, whereas ketene likely require much more sophisticated experiments be-
experiment), no other signal was observed. In the plays an important role. cause the acidity frequently varies simultaneously
presence of ZnCrOx, a signal of mass/charge Such CH2 species are very active and readily with the pore structure and crystallinity.
ratio (m/z) = 42.01 appeared in the effluents at a react with CO in the presence of CO, forming a The capability of the partially reduced oxide in
photon energy hu = 9.72 eV (Fig. 2D) in addition relatively less reactive ketene (26, 27), which is activating CO but incapability of catalyzing sur-
to stable hydrocarbon products. Further varying detectable by SVUV-PIMS. Thus, CH2 species are face polymerization of CHx makes it possible to
the photon energy (fig. S10) (9) confirmed that likely the primary intermediates. By forming manipulate CC coupling with confined acidic
this signal was unambiguously attributed to ke- ketene, the reaction pathway of surface poly- zeolite pores. Thus, the CO conversion and selec-
tene CH2CO, considering the m/z ratio and the merization of CHx is blocked, hence circumventing tivity can be tuned at the same timei.e., CO con-
ionization energy (2224). In comparison, pro- the ASF limits of FTS. Subsequently, in the presence version is manipulated via the surface structure
pene (m/z = 42.05) was only distinguishable at of confined acidic environment of zeolite pores, of the oxides and the ratio of oxides/zeolite,
hu > 9.73 eV (fig. S10e) (9), and it became prom- CH2CO can be converted to olefins (Fig. 2A). This whereas the olefin selectivity is controlled by the
inent at hu = 11.40 eV (the black signal in the inset was validated by feeding ketene directly as the properties of zeolites, particularly the pore struc-
of Fig. 2D). Additional signals in Fig. 2D were reactant to the MSAPO catalyst, with the effluents ture and acidity. These findings open up a new
assigned to stable products of C4H8 and C6H6, monitored by the online SVUV-PIMS. Ketene was avenue for development of syngas-to-olefin tech-
and their derivatives from stepwise addition of synthesized via pyrolysis of acetic anhydride nologies, which may allow utilization of coal- and
CH2 monomers up to m/z = 126. following the reported procedure (fig. S15) (24, 28). biomass-derived syngas with a low H2/CO ratio.
In addition to ketene and the stable products, Comparison of the results in the blank reactor and
signals of methanol (m/z = 32) and its dissociated in the presence of the catalyst (Fig. 3A and fig.
product methoxyl group (m/z = 31) were also S15) demonstrates the capability of MSAPO cata- REFERENCES AND NOTES

detected at a high hu (11.40 eV) in the presence of lyzing conversion of ketene to light olefins. 1. R. Snel, Catal. Rev. Sci. Eng. 29, 361445 (1987).
2. Y. D. Xu, D. W. Goodman, Catal. Lett. 24, 3135 (1994).
ZnCrOx catalyst. However, the reaction did not Figure 3B and fig. S16 (9) demonstrate that the 3. Z. Yang, S. Guo, X. Pan, J. Wang, X. Bao, Energy Environ. Sci. 4,
appear to go via methanol in this OX-ZEO pro- product selectivity can be modulated by the medi- 45004503 (2011).
cess because the yield of C2=C4= dramatically um strength acidity, characterized by temperature- 4. H. M. Torres Galvis et al., Science 335, 835838 (2012).
5. X. Chen, D. Deng, X. Pan, Y. Hu, X. Bao, Chem. Commun.
dropped to 3% within 22 hours when feeding programmed-desorption (TPD) of NH3, with the (Camb.) 51, 217220 (2015).
methanol (50 mbar in He) directly to the com- peak maximum located in the range from 350 to 6. H. M. Torres Galvis, K. P. de Jong, ACS Catal. 3, 21302149
posite catalyst (fig. S11). Figure S12 shows that 410C. For example, a commercially available (2013).

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1067


R ES E A RC H | R E PO R TS

7. R. A. Friedel, R. B. Anderson, J. Am. Chem. Soc. 72, 12121215 19. C. Jia et al., Phys. Chem. Chem. Phys. 16, 75387547 (2014). (no. 2013CB933100), the Strategic Priority Research Program of
(1950). 20. F. Qi, Proc. Combust. Inst. 34, 3363 (2013). the Chinese Academy of Sciences (grant XDA09030101), and
8. I. Puskas, R. S. Hurlbut, Catal. Today 84, 99109 (2003). 21. J. D. Savee et al., Science 347, 643646 (2015). Dalian Institute of Chemical Physics Fundamental Research
9. Supplementary materials are available on Science Online. 22. NIST Chemistry WebBook, http://webbook.nist.gov/ Program for Clean Energy (DICP M201308). The Advanced Light
10. G. Yang, N. Tsubaki, J. Shamoto, Y. Yoneyama, Y. Zhang, chemistry/mw-ser.html. Source is supported by the Director, Office of Science, Office of
J. Am. Chem. Soc. 132, 81298136 (2010). 23. T. A. Cool et al., J. Chem. Phys. 119, 83568365 (2003). Basic Energy Sciences, of the U.S. Department of Energy under
11. K. Fujimoto, H. Saima, H. O. Tominaga, J. Catal. 94, 1623 24. B. Yang et al., Int. J. Mass Spectrom. 309, 118128 (2012). contract no. DE-AC02-05CH11231. A Chinese patent and an
(1985). 25. D. Chen, K. Moljord, A. Holmen, Micro. Meso.Mater. 164, international patent application under the Patent Cooperation
12. Q. Ge, X. Li, H. Kaneko, K. Fujimoto, J. Mol. Catal. Chem. 278, 239250 (2012). Treaty are pending.
215219 (2007). 26. J. Chanmugam, M. Burton, J. Am. Chem. Soc. 78, 509519
13. Y. Yu et al., React. Kinet. Mech. Catal. 112, 489497 (2014). (1956).
SUPPLEMENTARY MATERIALS
14. J. Erea, J. M. Arandes, J. Bilbao, A. T. Aguayo, H. I. de Lasa 27. Z. H. Zhang, Y. Zhang, J. B. Wang, ACS Catal. 1, 16211630
(2011). www.sciencemag.org/content/351/6277/1065/suppl/DC1
Ind. Eng. Chem. Res. 37, 12111219 (1998).
28. J. W. Williams, C. D. Hurd, J. Org. Chem. 5, 122125 (1940). Materials and Methods
15. P. Mohanty, K. K. Pant, J. Parikh, D. K. Sharma, Fuel Process.
Figs. S1 to S18
Technol. 92, 600608 (2011).
ACKN OWLED GMEN TS Tables S1 to S4
16. Y. Tamaura, M. Tahata, Nature 346, 255256 (1990).
References (29, 30)
17. G. Melaet et al., J. Am. Chem. Soc. 136, 22602263 (2014). This work was financially supported by the National Natural
18. J. Xiao, T. Frauenheim, J. Phys. Chem. Lett. 3, 26382642 Science Foundation of China (grant nos. 21425312, 21321002, and 3 January 2016; accepted 22 January 2016
(2012). 91545204), the Ministry of Science and Technology of China 10.1126/science.aaf1835

QUANTUM COMPUTING ax mod N can be extracted from a number of


measurements not increasing with the size of the
number to be factored.
Realization of a scalable What are the requirements and challenges of
implementing Shors algorithm? We first focus

Shor algorithm on the period register and subsequently address

Downloaded from on March 3, 2016


modular exponentiation in the computational
register. Factoring N, an n = log2(N)-bit num-
Thomas Monz,1* Daniel Nigg,1 Esteban A. Martinez,1 Matthias F. Brandl,1 ber (with the quantity in brackets rounded up
Philipp Schindler,1 Richard Rines,2 Shannon X. Wang,2 to next integer number), requires a minimum
Isaac L. Chuang,2 Rainer Blatt1,3 of n qubits in the computational register (to
store the results of ax mod N) and generally
Certain algorithms for quantum computers are able to outperform their classical counterparts. about 2n qubits in the period register (9, 10).
In 1994, Peter Shor came up with a quantum algorithm that calculates the prime factors Thus, even a seemingly simple example, such as
of a large number vastly more efficiently than a classical computer. For general scalability of factoring 15 (an n = 4-bit number), requires 3n =
such algorithms, hardware, quantum error correction, and the algorithmic realization itself 12 qubits. These qubits then have to be mani-
need to be extensible. Here we present the realization of a scalable Shor algorithm, as proposed pulated with high-fidelity gate operations. Given
by Kitaev. We factor the number 15 by effectively employing and controlling seven qubits the current state-of-the-art control over quantum
and four cache qubits and by implementing generalized arithmetic operations, known as systems (11), such an approach would probably
modular multipliers. This algorithm has been realized scalably within an ion-trap quantum yield an unsatisfactory performance. However,
computer and returns the correct factors with a confidence level exceeding 99%. a full quantum implementation of this part of the
algorithm is not necessary. As noted by Kitaev

S
(12), if only the classical information of the QFT
hors algorithm for factoring integers (1) is that the number we want to factor is N = 15. We (such as the period r) is of interest, 2n qubits
one example in which a quantum computer pick a random number a2; N 1 (the base)say, subject to a QFT can be replaced by a single qubit.
(QC) outperforms the most efficient known a = 7. We evaluate whether the greatest common Still, this approach requires qubit recycling (spe-
classical algorithms. Experimentally, its im- divisor gcd(a, N) = 1; if not, a factor is already de- cifically, in-sequence single-qubit readout and
plementation is highly demanding (27) be- termined. This is the case for a = {3, 5, 6, 9, 10, 12}. state reinitialization) paired with feed-forward be-
cause it requires both a sufficiently large quantum Next, we calculate the modular exponentiation havior to compensate for the reduced system size.
register and high-fidelity control. Such challenging ax mod N for x = 0, 1, 2 and find its period r: the In the following, Kitaevs QFT will be referred
requirements raise the question of whether opti- first value of x > 0 such that ax mod N = 1. Given to as KQFT(M). It replaces a QFT acting on M
mizations and experimental shortcuts are possi- r, finding the factors of N requires calculating the qubits with a semiclassical QFT acting repeatedly
ble. Optimizations, especially system-specific or greatest common divisors of ar/2 1 and N, which on a single qubit. Similar applications of Kitaevs
architectural optimizations, are certainly possible, is efficiently possible with a classical approach approach to a semiclassical QFT in quantum al-
but for a demonstration of Shors algorithm in a for instance, using Euclids algorithm. For our ex- gorithms have been investigated (1315). For the
scalable manner, special care must be taken to not ample (N = 15, a = 7), the modular exponentiation implementation of Shors algorithm, Kitaevs ap-
oversimplify the implementationfor instance, yields 1, 7, 4, 13, 1,, which has a period of 4. The proach provides a reduction from the previous n
by employing knowledge about the solution be- greatest common divisors of ar/2 1 = 74/2 1 = computational qubits and 2n QFT qubits (in total,
fore the actual experimental application (8). {48, 50} and N = 15 are {3, 5}, the nontrivial factors 3n qubits) to only n computational qubits and
How does Shors algorithm work? First, we of N. In this example, the cases a = {4, 11, 14} have 1 KQFT(2n) qubit (in total, n + 1 qubits).
consider a classical factoring recipe, assuming period r = 2 and require a single multiplication The second key ingredient of Shors algorithm
step (a2 mod N = 1), which is considered an easy and a notably more challenging aspect than the
1
Institut fr Experimentalphysik, Universitt Innsbruck,
case (8). Note that the periodicity for a chosen a QFTis modular exponentiation, which admits
Technikerstrae 25, A-6020 Innsbruck, Austria. 2Center for cannot be predicted. the following general simplifications.
Ultracold Atoms, Massachusetts Institute of Technology, 77 How can this recipe be implemented in a QC? 1) Considering Kitaevs approach (Fig. 1), the
Massachusetts Avenue, Cambridge, MA 02139, USA. 3Institut A QC also has to calculate ax mod N in a com- input state j1i (in decimal representation) is sub-
fr Quantenoptik und Quanteninformation, sterreichische
Akademie der Wissenschaften, Otto-Hittmair-Platz 1, A-6020
putational register for x = 0, 1, 2 and then ex- ject to a conditional multiplication based on the
Innsbruck, Austria. tract r. Using the quantum Fourier transform most significant bit k of the period register. At
*Corresponding author. E-mail: thomas.monz@uibk.ac.at (QFT) applied to the period register, the period of most, there will be two results after this first step.

1068 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

It follows that, for the very first step, it is sufficient


to implement an optimized operation that con-
k
ditionally maps j1ija2 mod N i. Considering
the importance of a high-fidelity multiplication
(with its performance being fed forward to all
subsequent qubits), this efficient simplification
improves the overall performance of experimen-
tal realizations.
2) Subsequent multipliers can similarly be re-
placed with maps by considering only possible
outputs of the previous multiplications. However,
using such maps will become intractable, as the
number of input and output states to be consid-
ered grows exponentially with the number of
steps: After n steps, 2n > N possible outcomes
need to be considered, a numerical task as chal-
lenging as factoring N by classical means. Thus,
controlled full modular multipliers should be im-
plemented. Figure 2 shows the experimentally
obtained truth table for the modular multiplier 2
mod 15 [see also (16) for modular multipliers with
bases {7, 8, 11, 13}]. These quantum circuits can be
efficiently derived from classical procedures by
using a variety of standard techniques for revers-
ible quantum arithmetic and local logic optimi-
zation (17, 18).
3) The very last multiplier allows one more
simplification: Considering that the results of the
modular exponentiation are not required for Shors
algorithm (as only the period encoded in the pe-
riod register is of interest), the last multiplier only
has to create the correct number of correlations
between the period register and the computation
register. Local operations after the conditional
(entangling) operations may be discarded to fa- Fig. 1. Quantum circuits. Diagrams of Shors algorithm for factoring N = 15, using a generic textbook
cilitate the final multiplication without affecting approach (A) compared with Kitaevs approach (B) for a generic base a. (C) The actual implemen-
the results of the implementation. tation for factoring 15 to base 11, optimized for the corresponding single-input state. Here qi corre-
4) In rare cases, certain qubits are not subject sponds to the respective qubit in the computational register. (D) Kitaevs approach to Shors algorithm
to operations in the computation. Thus, these for the bases {2, 7, 8, 13}. Here, the optimized map of the first multiplier is identical in all four cases, and
qubits can be removed from the algorithm entirely. the last multiplier is implemented with full modular multipliers, as depicted in (E). In all cases, the
For large-scale quantum computation, opti- single QFT qubit is used three times, which, together with the four qubits in the computation register,
mization steps 1, 3, and 4 will only marginally totals seven effective qubits. (E) Circuit diagrams of the modular multipliers of the form a mod N for
affect the performance of the implementation. bases a = {2, 7, 8, 11, 13}.
These steps represent merely a small subset of
the entire computation, which mainly consists
of the full modular multipliers. Thus, the re-
alization of these modular multipliers is a core
requirement for the implementations of a scal-
able Shor algorithm.
Furthermore, Kitaevs approach requires in-
sequence measurements, qubit recycling to reset
the measured qubit, feed-forward behavior of gate
settings on the basis of previous measurement
results, and controlled quantum operations
tasks that have not been realized in a combined
experiment to date.
We demonstrate these techniques in our real-
ization of Shors algorithm in an ion-trap QC, with
five 40Ca+ ions in a linear Paul trap. The qubit is
encoded in the ground state S1=2 m 1=2 j1i
and the metastable state D5=2 m 1=2 j0i
(where m denotes the Zeeman sublevel) [for more
details, see (16, 19)]. Unitary operations, illustrated
in Fig. 1, are decomposed into primitive components, Fig. 2. Truth table. Experimentally obtained truth table of the controlled 2 mod 15 multiplier. (A) With
such as two-target controlled-NOT (C-NOT) and C- the control-qubit being in state 0, the truth table corresponds to the identity operation. (B) When the
SWAP gates (or gates with global symmetries, control qubit triggers the multiplication, the truth table illustrates the multiplication of the input state
such as the four-target C-NOT gate employed with 2 mod 15. The mean fidelity with respect to the expected output state is 48(5)%.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1069


R ES E A RC H | R E PO R TS

Fig. 3. Experimental find-


ings. Results and correct
order-assign probability for
the different implementa-
tions to factor N = 15. Three-
digit results (in decimal rep-
resentation) of Shors
algorithm for the different
bases. The ideal data (red)
for period {2, 4} are shown
adjacent to the raw data
(blue). The squared statisti-
cal overlap is larger than
90% for all cases.

here), from which an adaptation of the gradient- a continued fraction expansion applied to x/2k, a 11. J. Stajic, Science 339, 1163 (2013).
ascent pulse engineering algorithm (20) can ef- good approximation of the ideal case when k, the 12. A. Y. Kitaev, http://arxiv.org/abs/quant-ph/9511026
(1995).
ficiently derive an equivalent sequence of laser number of qubits, is sufficiently large. In our
13. R. B. Griffiths, C. S. Niu, Phys. Rev. Lett. 76, 32283231 (1996).
pulses acting on only the relevant qubits. The realizations with bases a = {2, 7, 8, 11, 13}, the 14. S. Parker, M. B. Plenio, Phys. Rev. Lett. 85, 30493052 (2000).
problem with this approach is that the resulting probabilities (and their error estimates in paren- 15. M. Mosca, A. Ekert, in Quantum Computing and Quantum
sequence generally includes operations acting theses) to obtain output states that allow the der- Communications, vol. 1509 of Lecture Notes in Computer
on all qubits. Implementing the optimized three- ivation of the correct period are {56(2), 51(2), Science, C. P. Williams, Ed. (Springer, 1999), pp. 174188.
16. Supplementary materials are available on Science Online.
qubit operations for a five-ion string therefore re- 54(2), 47(2), 50(2)}%. Thus, to obtain a confidence
17. V. Vedral, A. Barenco, A. Ekert, Phys. Rev. A 54, 147153 (1996).
quires decoupling the remaining qubits from the level of >99% for the periodicity, the experiment 18. R. Van Meter, K. M. Itoh, Phys. Rev. A 71, 052320 (2005).
computation space. We spectroscopically decouple has to run about eight times. 19. P. Schindler et al., New J. Phys. 15, 123012 (2013).
qubits by transferring any information from jSi We have presented the realization of Kitaevs 20. V. Nebendahl, H. Hffner, C. F. Roos, Phys. Rev. A 79, 012312
to jD i D5=2 m 5=2 and from jDi to jS i vision of Shors algorithm based on scalable build- (2009).
21. H. Dehmelt, Bull. Am. Phys. Soc. 20, 60 (1975).
S1=2 m 1=2. Here, the subspace fjS i; jD ig ing blocks with three-digit resolution to factor N =
22. M. Riebe et al., Nature 429, 734737 (2004).
serves as a readily available quantum cache to 15, using bases {2, 7, 8, 11, 13}. To do this, we suc- 23. D. J. Wineland et al., J. Res. Natl. Inst. Stand. Technol. 103, 259
store and retrieve quantum information for the cessfully employed a semiclassical QFT combined (1998).
purpose of facilitating quantum computations. with single-qubit readout, feed-forward behavior, 24. I. Marzoli, J. I. Cirac, R. Blatt, P. Zoller, Phys. Rev. A 49,
Finally, to complete the toolbox necessary for and qubit recycling. Compared with the traditional 27712779 (1994).
25. J. Chiaverini et al., Science 308, 9971000 (2005).
Kitaevs approach to Shors algorithm, we also algorithm, our realization of Shors algorithm re-
26. D. Kielpinski, C. Monroe, D. J. Wineland, Nature 417, 709711 (2002).
implement (i) single-qubit readout, by encoding all duces the required number of qubits by nearly a
other qubits in the fjDi; jD ig subspace and sub- factor of 3. Furthermore, the entire quantum reg- AC KNOWLED GME NTS
sequent electron shelving (21) on the S1=2 P1=2 ister has been subject to the computation in a We acknowledge support from the Austrian Science Fund (FWF),
transition; (ii) feed-forward behavior, by storing black-box fashion. Employing the equivalent of through the SFB FoQus (FWF project no. F4002-N16); the European
Commission (AQUTE), the NSF Interdisciplinary Quantum Information
counts detected during the single-qubit readout a quantum cache by spectroscopic decoupling fa-
Science and Engineering (iQuISE) Integrative Graduate Education and
(22) in a classical register and subsequent condi- cilitated the derivation of the necessary pulse se- Research Traineeship (IGERT); and the Institut fr Quantenoptik und
tional laser pulses; and (iii) state reinitialization, quences to achieve high-fidelity results. We envision Quanteninformation. E.A.M. is a recipient of a DOC Fellowship of the
using optical pumping for the ion, and Raman that our scalable algorithm implementation will Austrian Academy of Sciences. This research was funded by the Office
cooling (23, 24) for the motional state of the ion be combined with a scalable trap architecture (26) of the Director of National Intelligence (ODNI), Intelligence Advanced
Research Projects Activity (IARPA), through Army Research Office
string. The pulse sequences and additional infor- and quantum error correction to enable arbitrary grant W911NF-10-1-0284. All statements of fact, opinion, or conclusions
mation on the implementation of the modular long quantum computation. contained herein are those of the authors and should not be construed
multipliers are available in (16). as representing the official views or policies of IARPA, the ODNI, or
The measurement results for base a = {2, 7, 8, RE FERENCES AND NOTES
the U.S. government. T.M., D.N., and P.S. developed the research, on
the basis of theoretical ideas derived with R.R., S.X.W., and I.L.C; T.M.,
11, 13} with period r = {4, 4, 4, 2, 4} are shown in 1. P. W. Shor, in Proceedings of the 35th Annual Symposium on D.N., E.A.M., M.F.B., P.S., and S.X.W. performed the experiments; T.M.
Fig. 3. To quantify the performance of the imple- Foundations of Computer Science (IEEE Computer Society and D.N. analyzed the data; T.M., D.N., E.A.M., P.S., M.F.B., and R.B.
mentation, previous realizations focused mainly Press, 1994), pp. 124134. contributed to the experiment; T.M., D.N., R.R., M.F.B., I.L.C., and R.B.
2. A. Politi, J. C. F. Matthews, J. L. OBrien, Science 325, 1221 (2009). wrote the manuscript; and all authors contributed to discussions about
on the squared statistical overlap (SSO) (25), the 3. E. Martn-Lpez et al., Nat. Photonics 6, 773776 (2012). the results and the manuscript. We declare no competing financial
classical equivalent to the Uhlmann fidelity (10). 4. E. Lucero et al., Nat. Phys. 8, 719723 (2012). interests.
Although we achieved an SSO of {0.968(1), 0.964(1), 5. C. Y. Lu, D. E. Browne, T. Yang, J. W. Pan, Phys. Rev. Lett. 99,
0.966(1), 0.901(1), 0.972(1)} for the case of a = 250504 (2007).
6. B. P. Lanyon et al., Phys. Rev. Lett. 99, 250505 (2007). SUPPLEMENTARY MATERIALS
{2, 7, 8, 11, 13}, we argue that this does not answer 7. L. M. K. Vandersypen et al., Nature 414, 883887 (2001).
the question What is the period? Shors algo- 8. J. A. Smolin, G. Smith, A. Vargo, Nature 499, 163165 (2013).
www.sciencemag.org/content/351/6277/1068/suppl/DC1
Supplementary Text
rithm allows one to deduce the period with high 9. E. G. Rieffel, W. H. Polak, Quantum Computing: A Gentle
Figs. S1 and S2
probability from a single-shot measurement, as Introduction (Scientific and Engineering Computation) (The MIT
Tables S1 and S2
Press, ed. 1, 2011).
the output of the QFT (x) is, in the exact case, a References (27, 28)
10. M. A. Nielsen, I. L. Chuang, Quantum Computation and
ratio of integers, where the denominator gives Quantum Information (Cambridge Series on Information and 25 November 2015; accepted 1 February 2016
the desired period. This period is extracted by using the Natural Sciences, Cambridge Univ. Press, ed. 1, 2004). 10.1126/science.aad9480

1070 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

FLEXIBLE ELECTRONICS field; unlike current-driven LEDs, which require


lithography to form p-n junctions, this material
system can be processed using replica molding.
Highly stretchable Application of an AC electric field causes lumi-
nescence within the semiconducting phosphor at

electroluminescent skin for optical wavelength centers corresponding to the dopants


in the ZnS lattice. Green and blue centers are
typically produced using low [~0.01 weight % (wt
signaling and tactile sensing %)] and high (~0.1 wt %) concentrations of Cu,
whereas yellow is produced using Mn (~1 wt %)
C. Larson,1* B. Peele,1* S. Li,2* S. Robinson,2 M. Totaro,3 L. Beccai,3
(25). White light can be achieved using combina-
tions of these dopants.
B. Mazzolai,3 R. Shepherd1,2
The HLEC (Fig. 1B) is a five-layer structure con-
sisting of an electroluminescent dielectric layer
Cephalopods such as octopuses have a combination of a stretchable skin and color-
that is sandwiched between two electrodes and
tuning organs to control both posture and color for visual communication and disguise.
encapsulated in low elastic modulus (E ~ 30 kPa)
We present an electroluminescent material that is capable of large uniaxial stretching
(26) silicone (Ecoflex 00-30, Smooth-on Inc.). Our
and surface area changes while actively emitting light. Layers of transparent hydrogel
hydrogel electrodes are designed with a balance
electrodes sandwich a ZnS phosphor-doped dielectric elastomer layer, creating thin
of high mechanical toughness, low volatility, and
rubber sheets that change illuminance and capacitance under deformation. Arrays of
low electrical resistance under deformation (fig.
individually controllable pixels in thin rubber sheets were fabricated using replica
S1 and data table S1). Aqueous lithium chloride
molding and were subjected to stretching, folding, and rolling to demonstrate their use
(LiCl) is used as the ionic conductor because of
as stretchable displays. These sheets were then integrated into the skin of a soft
its high conductivity (~10 S m1), ionic strength,
robot, providing it with dynamic coloration and sensory feedback from external and
and hygroscopic nature, whereas polyacrylamide

Downloaded from on March 3, 2016


internal stimuli.
(PAM) is used as the elastomeric matrix because

B
of its high toughness (27) and optical transpar-
iological systems employ a host of strat- that are used in soft robotics to mimic the move- ency. Electrodes are synthesized by first dissolv-
egies for visual display and camouflage. ments of animals. ing acrylamide monomer (AAm), polyacrylamide,
Cephalopods, for example, can mimic their Biological skin also enables animals to sense and N,N-methylenebisacrylamide crosslinker in
environment by changing skin color and their environments. A number of approaches aqueous LiCl and casting the solution onto an
texture, as well as posture (1). Recent devel- have been used to create pressure-sensitive elec- ultraviolet (UV)ozonetreated silicone (Ecoflex
opments in soft robotics (2, 3), bioinspired design tronic skins, including arrays of organic field- 00-30) substrate. The aqueous PAM-AAm solu-
(4, 5), and stretchable electronics (6) reveal strate- effect transistors (FETs) deposited on flexible tion is then crosslinked under UV light (28),
gies that enable us to engineer some of the func- parylene-polyamide substrates (17, 18) and in- producing a highly stretchable and transparent
tions of cephalopod skin synthetically. For example, side stretchable rubber (19), as well as thin Au electrode. The EL layer is formed by mixing
microfluidic networks filled with liquid dyes have films and liquid metal embedded in polydi- commercially available phosphor powders (Global
been used as active camouflage and displays for methylsiloxane (PDMS) (20, 21). More recently, Tungsten & Powders) (25 mm, ~8% by volume)
soft mobile robots, giving them the ability to dielectric elastomer transducers (DETs), which into silicone (Ecoflex 00-30) and then molding
change their appearance via color, texture, and are stretchable capacitors composed of highly the dispersion into a 1-mm-thick sheet. Fi-
luminescence (7). More recently, electro-mechano- extensible ionic hydrogels, have been used. These nally, we bond the EL layer between the two
chemically responsive films were exploited to hydrogels are intrinsically soft, highly transpar- electrode-patterned silicone substrates and en-
render fluorescent patterns under the control of ent in the visible spectrum (extinction coefficient capsulate the capacitor in an insulating layer
electric fields (8), and adaptive optoelectronic mext ~ 106 mm1) (22), can exhibit very high ulti- of silicone.
camouflage systems have been used to mimic the mate strain (eult ~2000%) and toughness (U ~ The stress-strain curves of the HLEC and its
visual appearance of cephalopod skin (9). Another 9 kJ m2) (23), and have relative changes in resis- silicone-containing layers (Ecoflex and Ecoflex-
approach is the use of active display technologies, tivity with strain that are orders of magnitude EL composite) are all coincident, whereas the
such as polymeric light-emitting devices (PLEDs) less than those of electrodes based on percolated elastic modulus of the hydrogel is two orders of
and organic light-emitting diodes (OLEDs), which networks of conductive particles (such as metal magnitude lower, allowing the HLEC to stretch
use stretchable transparent electrodes based on nanoparticles, carbon powder, or nanotubes) (24). freely without delaminating. Mechanical testing
indium tin oxide (ITO) films (10), graphene (11), Presently, soft robots are primarily used be- data (Fig. 1C and data table S2) and images
single- or multi-walled carbon nanotubes (SWNTs cause their low mechanical compliance enables (Fig. 2A, data table S3, and movie S1) show the
or MWNTs) (12, 13), polyethylene-dioxythiophene: safe human-robot interaction; however, their po- excellent adhesion between the layers. The HLEC
polystyrene-sulfonate (PEDOT:PSS) (14), or other tential is limited by a lack of suitable electronics achieved a mean strain of 487 59% (SD), as
percolated networks of conductive colloids or that can stretch continuously with their bodies. measured at five locations across the width of
nanowires (15). Despite the broad applicability of No soft robot can dynamically display informa- the illuminated section, with portions exceed-
LED-based systems for consumer displays, their tion on its body, and there are relatively few ing 500% before the external copper leads lost
electrical function is limited to ultimate strains, examples that can sense external and internal contact with the hydrogel electrodes. For these
eult < 120% (16), well below the ultimate strain of stimuli. Here we present a hyperelastic light- tests, the HLECs were operated at 700 Hz under
elastomers (such as silicones; eult ~400 to 700%) emitting capacitor (HLEC) that enables both a nominal electric field of ~25 kV cm1, with a
light emission and touch sensing in a thin rub- power consumption of 0.2 W and a luminous effi-
1
Department of Mechanical and Aerospace Engineering,
ber sheet that stretches to >480% strain (Fig. 1A). cacy of 43.2 millilumens per watt (mlm W1) (28).
Cornell University, Ithaca, NY 14853, USA. 2Department of These HLECs are composed of ionic hydrogel We used this same replica molding technique to
Materials Science and Engineering, Cornell University, Ithaca, electrodes and composites of doped ZnS phos- form an 8-by-8 array of 4-mm pixels (Fig. 3A).
NY 14853, USA. 3Center for Micro-BioRobotics@SSSA, phors embedded in a dielectric matrix of silicone This HLEC display can undergo many deforma-
Istituto Italiano di Tecnologia, Viale Rinaldo Piaggio 34,
I-56025 Pontedera, Italy.
elastomer. We used electroluminescent (EL) phos- tion modes, including stretching, rolling, folding,
*These authors contributed equally to this work. Corresponding phor powders that emit light via excitations with- and wrapping (Fig. 3, B to E, and movie S2). Dyna-
author. E-mail: rfs247@cornell.edu in intrinsic heterojunctions under an AC electric mic control of the pixels is shown in Fig. 3, F to I.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 107 1


R ES E A RC H | R E PO R TS

In addition to emitting light, the HLEC also is stretched under uniaxial (Fig. 2B and data S4) early as the sample is stretched (eq. S11). For
serves as a dielectric elastomer sensor (DES), due and biaxial (fig. S2 and data S5) tension (28). biaxial test conditions, we observe that the rela-
to its construction as a parallel-plate capacitor. We model the capacitance by expressing A and d tive change in capacitance follows C/C0 = l4
Changes in the electrode area (A) and separation in terms of the principal stretches, l1, l2, and l3, (eq. S12); however, at higher strains, the mea-
distance (d) cause the capacitance (C) to change which represent the axial, transverse, and out- sured values are slightly lower, due to a decrease
according to C/C0 Ad1, allowing the HLEC to of-plane orientations, respectively (supplementary in the permittivity of the dielectric (24).
sense deformations from pressure and stretch- text). For uniaxial boundary conditions, we ob- The illuminance of the HLEC also increases as
ing. The capacitance of the HLEC changes as it serve that the relative capacitance increases lin- the device is stretched. We attribute this change

Fig. 1. HLEC. (A) Image of the HLEC conforming to the end of a pencil. (B) Exploded view of the HLEC showing its five-layer structure consisting of a
~1-mm-thick electroluminescent layer (ZnS-Ecoflex 00-30) that is sandwiched between two PAM-LiCl hydrogel electrodes and encapsulated in Ecoflex 00-30.
(C) Stress-stretch curves of Ecoflex 00-30, the electroluminescent layer, and the composite device. The hydrogel data are shown in the inset because of its
much lower elastic modulus.

Fig. 2. The capacitive and luminescent behavior of the HLEC display under uniaxial stretching.
(A) A nominal electric field of ~25 kV cm1 was applied to the HLEC at the start of the uniaxial test. Five
lengths were measured using image analysis software to obtain l1 across the width of the illuminated
portion of the tensile bar. We report the mean and standard deviation of those measurements. At an
engineering strain (grip to grip) of 395%, we measured the mean strain of the illuminated portion to be
487%, with a range of 420 to 549%. (B) The capacitance of the HLEC as a function of its uniaxial stretch
(number of samples, n = 4). (C) The relative illuminance of the HLEC versus its uniaxial stretch (n = 4),
plotted alongside predicted values (supplementary text).

1072 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

to two interrelated phenomena: (i) the increase luminescence measurements in Fig. 2C (data skin, whereas the bottom two are used for pneu-
in electric field (E) as d decreases and (ii) the table S7) matic actuation (Fig. 4A). Inspired by architec-
decrease in areal number density of phosphor I tures developed for mobile soft robots (30), our
l1 l2 1 exp5:681 l3 
1=2
particles (h) as A increases. Starting with the Alfrey- 1 pneumatic actuator uses a series of inflatable
I0
Taylor equation (eq. S13, fig. S3, and data table chambers embedded in silicone, with a bottom
S6) (29), we predict the scaling law in Eq. 1 by To demonstrate the ability to monolithic- layer composed of an inextensible fiber-elastomer
expressing E/E0 as a function of the principal ally integrate the HLEC into soft systems, we composite (28). The inextensible layer induces a
stretches and by correcting for the change in h embedded three HLEC panels in a crawling net bending moment as the pneumatic chambers
with stretching (h/h0 A0/A) (supplementary soft robot by bonding six layers together. The are inflated; the resulting curvature is exploited
text). The predicted trend is shown alongside top four layers make up the electroluminescent to create an undulating gait.

Fig. 3. Multipixel electroluminescent displays fabricated via replica molding. The device measures 5 mm thick, with each of the 64 pixels measuring
4 mm. We show the devices in various states of deformation and illumination: (A) undeformed, (B) stretched, (C) wrapped around a finger, (D) folded, (E) rolled,
(F to H) with subsets of pixels activated, and (I and J) subsets of pixels activated while being deformed.

Fig. 4. HLEC skins endow soft robots with the ability to sense their actu- a negligible change in the capacitive response of the sensors over a period of
ated state and environment and communicate optically. (A) Schematic of a 120 hours. (E) Array of three HLEC panels, each emitting a different wavelength
three-chambered soft robot. A series of three independently actuated pneumatic through selective doping of the EL phosphor layer. Each HLEC panel is activated
chambers is embedded between the HLEC skin (top) and a strain-limiting layer independently. (F) An undulating gait is produced by pressurizing the chambers in
(bottom). (B) Capacitance plotted versus the actuation amplitude, defined as the sequence along the length of the crawler. This sequence produces forward
relative change in deflection between the uninflated and fully inflated states locomotion at a speed of ~4.8 m hour1 (~32 body lengths hour1). As each
(number of samples, n = 5). (C) A firm finger press induces an ~25% increase pneumatic chamber is pressurized, the outer electroluminescent skin is stretched,
in capacitance. (D) Change in capacitance versus applied pressure. We observed increasing the electric field across the EL layer and thus the luminescence.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1073


R ES E A RC H | R E PO R TS

The crawling robot uses its HLEC skin to sense 2. F. Ilievski, A. D. Mazzeo, R. F. Shepherd, X. Chen, 26. R. F. Shepherd, A. A. Stokes, R. M. D. Nunes, G. M. Whitesides,
its physical state and environment (i.e., proprio- G. M. Whitesides, Angew. Chem. Int. Ed. Engl. 50, 18901895 Adv. Mater. 25, 67096713 (2013).
(2011). 27. Y. Bai et al., Appl. Phys. Lett. 105, 151903 (2014).
ception and exteroception). The capacitance of the 3. D. Rus, M. T. Tolley, Nature 521, 467475 (2015). 28. Materials and methods are available as supplementary
HLEC changes with pneumatic actuation (Fig. 4B 4. M. J. Spenko et al., J. Field Robot. 25, 223242 (2008). materials on Science Online.
and data S8) and externally applied pressure (Fig. 5. E. Kreit et al., J. R. Soc. Interface 10, 20120601 (2012). 29. G. F. Alfrey, J. B. Taylor, Proc. Phys. Soc. B 68, 775784 (1955).
4, C and D, and data table S9) (10). Actuation of 6. J. A. Rogers, T. Someya, Y. Huang, Science 327, 16031607 30. R. F. Shepherd et al., Proc. Natl. Acad. Sci. U.S.A. 108,
(2010). 2040020403 (2011).
the three underlying pneumatic chambers results 7. S. A. Morin et al., Science 337, 828832 (2012). 31. D. H. Kim et al., Science 333, 838843 (2011).
in capacitance changes (D C) of up to 1000% when 8. Q. Wang, G. R. Gossweiler, S. L. Craig, X. Zhao, Nat. Commun. 32. J. Kim et al., Nat. Commun. 5, 5747 (2014).
the chambers are fully inflated. Additionally, each 5, 4899 (2014).
HLEC panel is largely decoupled from the state of 9. C. Yu et al., Proc. Natl. Acad. Sci. U.S.A. 111, 1299813003 AC KNOWLED GME NTS
(2014). Data reported in the paper are included in the supplementary
the surrounding pneumatic chambers (fig. S4 and 10. P. E. Burrows et al., Displays 22, 6569 (2001). materials. This work was supported by the Army Research Office
data table S11) (28). The ability to identify the 11. T. H. Han et al., Nat. Photonics 6, 105110 (2012). (grant no. W911NF-15-1-0464), the Air Force Office of Scientific
actuated state of the robot using the capacitive 12. T. Sekitani et al., Nat. Mater. 8, 494499 (2009). Research (grant no. FA9550-15-1-0160), the NSF MRSEC program
sensor readings enables proprioception. To dem- 13. M. K. Shin et al., Adv. Mater. 22, 26632667 (2010). (DMR-1120296), and an NSF Graduate Research Fellowship
14. M. S. White et al., Nat. Photon. 7, 811816 (2013). (grant no. DGE-1144153). The hyperelastic electroluminescent
onstrate the tactile sensing capabilities of the elec- 15. L. Hu, H. S. Kim, J. Y. Lee, P. Peumans, Y. Cui, ACS Nano 4, capacitors presented in this work have been filed under a
tronic skin, we pressed each of the HLEC panels 29552963 (2010). provisional patent application, no. 62/250,172 for Stretchable
on the robot and measured the capacitive response 16. J. Liang, L. Li, X. Niu, Z. Yu, Q. Pei, Nat. Photon. 7, 817824 (2013). Electroluminescent Devices. The listed inventors are Chris Larson,
(Fig. 4C). A firm finger press resulted in a ~25% 17. T. Someya et al., Proc. Natl. Acad. Sci. U.S.A. 102, 1232112325 Shuo Li, Bryan Peele, Sanlin Robinson, and Robert Shepherd.
(2005).
increase in capacitance. The relative capacitance 18. K. Takei et al., Nat. Mater. 9, 821826 (2010).
versus applied pressure, ranging from 0.9 to 19. T. Someya et al., Proc. Natl. Acad. Sci. U.S.A. 101, 99669970 SUPPLEMENTARY MATERIALS
30.9 kPa, remained nearly constant over a period (2004).
www.sciencemag.org/content/351/6277/1071/suppl/DC1
of 120 hours (Fig. 4D). Arrays of these tactile sen- 20. Y. L. Park, B. R. Chen, R. J. Wood, IEEE Sens. J. 12, 27112718
Materials and Methods
(2012).
sors enable exteroception in soft robotic systems. 21. D. P. J. Cotton, I. M. Graz, S. P. Lacour, IEEE Sens. J. 9,
Supplementary Text
An array of three HLEC panels patterned into Figs. S1 to S7

Downloaded from on March 3, 2016


20082009 (2009).
Table S1
the three-chambered crawling robot enables eight 22. C. Keplinger et al., Science 341, 984987 (2013).
Reference (33)
distinct illuminated states (Fig. 4E). The embedded 23. J. Y. Sun et al., Nature 489, 133136 (2012).
Movies S1 to S3
24. J. Y. Sun, C. Keplinger, G. M. Whitesides, Z. Suo, Adv. Mater.
HLEC remains functional as the robot is actuated 26, 76087614 (2014).
Data Tables S1 to S11 (single Excel workbook)
through its crawling sequence (Fig. 4F and movie 25. A. Kitai, Luminescent Materials and Applications (Wiley, West 5 May 2015; accepted 4 February 2016
S3). During actuation, the embedded HLEC un- Sussex, UK, 2008), pp. 249268. 10.1126/science.aac5082
dergoes stretches of l1 = 2.63 and l2 = 2.42 in the
longitudinal (front to rear) and transverse (side to
side) directions, respectively, to produce a ~635%
increase in the skins surface area (fig. S5). Sim- HUMAN ALTRUISM
ilar to the single-panel HLEC (movie S1), the
luminescence of the embedded skin increases
during actuation as its thickness is decreased.
Integrating these highly stretchable and com-
The brains functional network
pliant displays into soft actuators enables two
new capabilities in soft electronics: (i) displays
architecture reveals human motives
that actively change their shape and (ii) robots
that actively change their color. Using replica Grit Hein,1* Yosuke Morishima,1,2,3 Susanne Leiberg,1 Sunhae Sul,4 Ernst Fehr1
molding, we fabricated a multipixel array of in-
dividually addressable HLECs, and we used the Goal-directed human behaviors are driven by motives. Motives are, however, purely mental
same process to monolithically integrate these constructs that are not directly observable. Here, we show that the brains functional
displays into a soft robot capable of changing network architecture captures information that predicts different motives behind the same
posture. The HLEC array imparts both dynamic altruistic act with high accuracy. In contrast, mere activity in these regions contains no
coloration and the potential for feedback control, information about motives. Empathy-based altruism is primarily characterized by a
which would be useful in epidermal electronics positive connectivity from the anterior cingulate cortex (ACC) to the anterior insula (AI),
(31) and robotics (32). Although the luminous ef- whereas reciprocity-based altruism additionally invokes strong positive connectivity from
ficacy of our HLEC (43.2 mlm W1) is not as high the AI to the ACC and even stronger positive connectivity from the AI to the ventral
as that of commercial AC powder electrolumine- striatum. Moreover, predominantly selfish individuals show distinct functional
scent devices (~4 lm W1) (32), it can be greatly architectures compared to altruists, and they only increase altruistic behavior in response
improved by tuning the materials system and to empathy inductions, but not reciprocity inductions.

T
device architecture (such as higher-transmissivity
encapsulation layers, reduced thickness, and opti- he theory of revealed preference (1) pro- evant information about motives (6, 7). There-
mized particle size). For applications requiring vides the choice-theoretic foundations for fore, human motives have been typically inferred
higher display resolution, HLECs could be made modern economics. In this view, prefer- from individuals behavior by assuming that dif-
compatible with photolithography and other ences cannot be identified independently ferent motives lead to different behaviors.
microfabrication techniques by using photo- of behavior, and motives play no causal role
1
polymerizable polymers. These techniques would in economists explanatory toolboxa view that Laboratory for Social and Neural Systems Research,
Department of Economics, University of Zurich, Switzerland.
also allow us to decrease the thickness of the is in direct contradiction to the neuroeconomic 2
Division of Systems Neuroscience of Psychopathology,
electroluminescent layer, thereby reducing the approach (24). In psychology, motives are also Translational Research Center, University Hospital of
voltage required to power the HLEC. considered to be independent drivers of goal- Psychiatry, University of Bern, Switzerland. 3Japanese
directed human behavior (5). Motives are, however, Science and Technology Agency, PRESTO, Japan.
4
Department of Psychology, Pusan National University,
RE FE RENCES AND N OT ES mental constructs that are not directly observable Pusan, South Korea.
1. A. Barbosa, J. J. Allen, L. Mthger, R. T. Hanlon, Proc. R. Soc. and frequently not even accessible introspectively, *Corresponding author. E-mail: grit.hein@econ.uzh.ch (G.H.);
London Ser. B 279, 8490 (2012). meaning that asking people does not provide rel- ernst.fehr@econ.uzh.ch (E.F.)

1074 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

Here we ask whether different motives have a make inferences about motives, could we still iden- partner) receiving pain shocks in a number of
distinct neurophysiological representation that tify and predict their motives merely on the basis trials, a situation known to elicit an empathic re-
is generalizable across individuals. That is, even of their functional neural network architecture? sponse (18, 19). The reciprocity motive is defined
if we had no information about individuals be- We tackled this question in the context of human as the desire to reciprocate perceived kindness
haviors or if these behaviors would not allow us to altruistic decisions (814). Subjects participated with a kind behavior (13, 14). Therefore, in the
in an allocation task in which they could make reciprocity induction group, we activated the
selfish or altruistic decisions. We studied the role reciprocity motive by instructing one of the con-
of two key motives for altruistic behaviorsthe federates (the reciprocity partner) to give up
empathy motive and the reciprocity motive, two money in several trials to save the subject from
important drivers for human altruism (814). We painful shocks (13, 14, 20). No motives were
induced these motives experimentally in two dif- induced toward the respective second partner
ferent groups of subjects, i.e., subjects were ran- (baseline partner), who played the same role
domly assigned to either the empathy induction in both the empathy- and the reciprocity induc-
group or the reciprocity induction group. After tion group [for details of motive induction, see
the motive inductions, subjects participated in supplementary materials (21)]. It is important to
the allocation task in which they could allocate stress that the subjects received painful shocks
money to other individuals at a cost to them- not only in the reciprocity induction group but
selves. All subjects faced the same allocation task also in the empathy induction group, and that
regardless of the previous motive-induction group. the number of painful shocks was identical across
Therefore, their underlying motive cannot be conditions (21). This feature has two advantages.
inferred from the mere fact that they behave First, by equalizing the shock frequency across con-
altruistically. Can we now predict the induced ditions, we can be sure that the two motive in-
motive solely on the basis of the subjects func- ductions contain the same number of aversive
Fig. 1. Example trial in the allocation task in tional neural architecture? events. Second, the application of painful shocks
which subjects could allocate money to them- We used dynamic causal models (DCMs) of to the subject in the empathy condition is likely
selves (indicated by the green bars in the sec- functional magnetic resonance imaging data to enhance the ability to empathize with the em-
ond screen) and one of their partners (indicated (1517) collected during the allocation task and pathy partner because subjects know how the
by the blue bars in the second screen). Subjects used the estimated DCM parameters to predict shock feels. Finally, to assess the success of the
could either maximize their own monetary payoff subjects hidden motivational state with ma- motive inductions, the subjects also completed
(selfish decision) or maximize the paired partners chine learningan approach known as genera- emotion ratings where they indicated in each trial
monetary payoff by giving up their own payoff (al- tive embedding (17). More specifically, the DCM how they felt.
truistic decision). They faced several trials during the analyses of subjects brain data during altruistic In the allocation task, subjects were in the
allocation task that differed in the subjects cost of decision-making gave us information about indi- scanner. In each trial, they allocated money be-
maximizing the partners payoff (21). In each trial, viduals network architecture in the different mo- tween themselves and one of the partners (Fig. 1).
subjects from the empathy induction group were tive conditions. These parameters then provided They could choose between maximizing the other
paired with the empathy partner (empathy condition) the raw material for our predictions and for persons monetary payoff by reducing their own
or the baseline partner (baseline condition). Likewise, the mechanistic insights that follow from our monetary payoff (altruistic behavior), or maxi-
in each trial, subjects from the reciprocity induction examination (17). mizing their own payoff (selfish behavior) at a
group were paired with a reciprocity partner (reci- Both in the empathy and the reciprocity in- cost to the partner. Depending on the type of
procity condition) or the baseline partner (baseline duction group, subjects were paired with two partner subjects faced in the allocation task,
condition). The arrow in the first screen indicates the partners (confederates of the experimenter), who there were three conditionsthe empathy condi-
partner with whom the subject was paired in a trial. were sitting on either side of the subject. In the tion, the reciprocity condition, and the baseline
The choice problems in the empathy, the reciprocity, empathy induction group, the subject repeatedly condition. Because neither reciprocity nor empathy
and the baseline conditions were identical. observed one of the confederates (the empathy motives were induced in the baseline condition

80 0
% Accuracy of classification

(1) AI VS
75 -0.5
(2) ACC AI
70 -1
ACC AI
Log P-value

(3) AI ACC P < 0.05


65
-1.5
(4) ACC input 60 P < 0.01
-2
(5) VS input 55
VS 50 -2.5
(6) AI input
45 -3
(7) ACC VS P < 0.001
40 -3.5
0 1 2 3 1 2 3 4 5 6 7 1 2 3 4 5 6 7
Classification weight Network components Network components

Fig. 2. DCM-based classification of different motives that resulted in the motive. (C) Accuracy for each level of the stepwise classification, based on the
same altruistic decision. (A) Dynamic causal model (DCM) that was used to order of classification weights in (B). (D) Log P-value of stepwise classification,
investigate the impact of the empathy and the reciprocity motive on brain based on the order of classification weights in (B). AI, anterior insula; ACC,
connectivity during altruistic decisions. Straight arrows indicate interregional anterior cingulate cortex; VS, ventral striatum. Network components: 1 = AI VS;
connections (four components), and dashed arrows represent stimulus inputs 2 = plus ACCAI; 3 = plus AIACC; 4 = plus ACC input; 5 = plus VS input; 6 =
from outside of the network (three components). (B) Discriminative compo- plus AI input; 7 = plus ACCVS. The highest accuracy is reached when only the
nents. The figure visualizes the weights with which each of the seven network first three of the seven connectivity and input parameters are used to classify
components contributed to the classification of the empathy and reciprocity the empathy and the reciprocity motive [(A), highlighted in red].

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1075


R ES E A RC H | R E PO R TS

(i.e., vis--vis the baseline partner), the behavior reciprocity partner was significantly higher than ner, and the respective baseline partner from in-
in this condition measures subjects raw or toward the baseline partner [motive induction ver- dividual regions of interest (ROIs) in left AI, left
homegrown unconditional altruism. This al- sus baseline, F(1,32) = 12.5, P = 0.001; empathy ver- VS, and ACC (table S2).
truism is unconditional in the sense that the sus baseline, t(17) = 2.5, P = 0.022; reciprocity The DCM analyses were based on the anatom-
baseline partner did not do anything during the versus baseline, t(15) = 2.7, P = 0.017]. There was, ical model shown in Fig. 2A, which is charac-
motive induction on which the subjects could however, no significant difference in the increase terized by seven components: three inputs (dashed
condition their behavior during the allocation of altruistic decisions between the two motive in- arrows) and four interregional connectivities (solid
task. Homegrown unconditional altruism can ductions [empathy versus baseline compared to arrows). Because most VS projections target the
play a role in all three conditions, but the in- reciprocity versus baseline, F(1,32) = 0.2, P = 0.64]. cortex via the thalamus (26, 27), ascending pro-
duced empathy motive plays an additional role In addition, the results of a Bayesian analysis show jections from VS were not included in the anato-
in the empathy condition and the reciprocity that the null hypothesis that the two motive in- mical model. We used Bayesian model averaging
motive plays an additional role in the reciproc- ductions cause the same increase in altruism (28) to determine the DCM parameters for the
ity condition. relative to baseline is more than five times more seven components for each subject. For a subject
Empathy means that if an individual observes likely (84% versus 16%) than the hypothesis of a of the empathy group, for example, we calculated
someone else in pain, the individual also feels or differential increase (21). (i) how activation in ACC, AI, and VS during
shares that pain (18). Stronger empathizing Our imaging analyses focused on altruistic de- altruistic decisions is changed as a result of im-
during the empathy induction should therefore cisions during the allocation task. First we used a pulses from outside the network under the em-
induce worse feelings (due to shared pain), conventional general linear model (GLM) analysis pathy and under the baseline conditions (dashed
which should then lead to a stronger empathy for a whole-brain search of regions with signifi- arrows in Fig. 2A), and (ii) how activation in every
motive for altruistic behavior. We indeed ob- cantly different activity during altruistic decisions (target) region is changed by the level of activation
served that the worse a subject felt when seeing driven by empathy versus altruistic decisions in the other regions during altruistic decisions
the empathy partner in pain, the more frequent driven by reciprocity. There were no significant under the empathy and the baseline conditions
the altruistic decision toward this person in the differences even at the very liberal threshold of (effective connectivity, solid arrows in Fig. 2A).
subsequent allocation task [r(18) = 0.51, P = Puncorrected < 0.05. This suggests that empathy- The same computations were also done for sub-
0.03]. Because the reciprocity partner is willing driven and reciprocity-driven altruism activate jects of the reciprocity group. We thus obtained
to incur cost to remove painful shocks, the partner similar brain regions. We therefore compared 14 DCM parameters per subject, seven for altruistic
is likely to be perceived as kind, which should brain activations under altruistic decisions in both decisions under the respective motive-induction
result in positive feelings and higher frequencies motive-induction conditions with the activations condition (empathy or reciprocity), and seven for
of altruistic decisions toward the reciprocity part- involved in altruistic decisions in the baseline altruistic decisions under the baseline condition.
ner. We found indeed that the better a subject condition. The results revealed a network con- We then subtracted the individual DCM param-
felt after the reciprocity partners decision during sisting of left anterior insula (AI), left ventral eters of the baseline condition from the DCM
motive induction, the more altruistic the decisions striatum (VS), and anterior cingulate cortex (ACC) parameters of the motive-induction condition. The
toward this person in the subsequent allocation [P < 0.05; family-wise error (FWE) corrected; see resulting seven D-DCM parameters reflect the in-
task [r(16) = 0.57, P = 0.021]. Reciprocity in- table S1 for details]regions that were reported dividual pattern of neural connectivity specific for
duction also induced significantly higher likabil- by previous studies on the reciprocity (2224) and altruistic decisions driven by empathy or reciprocity.
ity ratingswhich were collected after scanning the empathy motive (18, 19, 25). In a third step, we submitted these individual
(How much do you like the other person? 9 = very In a second step, we determined the pattern D-DCM parameters to a classification algorithm
much to 1 = not at all)for the reciprocity part- of neural connectivity within this network for (support vector machine, SVM) to test if the in-
ner as compared to the baseline partner [t(15) = each subject, using DCM. We extracted the time dividual patterns of brain connectivity can be
3.24, P = 0.005]. The overall frequency of altruistic series of activations during altruistic decisions used to detect the specific motive for altruism
decisions toward the empathy partner and the toward the empathy partner, the reciprocity part- that the experimenter induced (17). This approach
Average DCM parameters for the ACC AI

1.5
connection under the baseline condition

0.09 0.03 0.17


0.09 1
0.15 0.12
ACC AI ACC AI ACC AI 0.5
0.26

-0.08 -0.06 0.34 0


0 10 20 30 40
VS VS -0.5
VS
-1

-1.5

-2
Average number of altruistic decisions
under the baseline condition

Fig. 3. Average best models and mean DCM parameters. Best models t(1,32) = 0.57, P = 0.58; AIVS, t(1,32) = 1.03, P = 0.31; ACC input, t(1,32) =
and mean DCM parameters for (A) Altruistic decisions driven by the empathy 0.55, P = 0.59]. In (A) to (C), dashed arrows indicate direct inputs to an area;
motive, (B) altruistic decisions driven by homegrown unconditional altruism solid arrows indicate directed interregional connectivities. AI, anterior insula;
in the baseline condition, and (C) altruistic decisions driven by the reciprocity ACC, anterior cingulate cortex; VS, ventral striatum. (D) Correlation between
motive. For visualization purposes, the baseline data for subjects from the the individual parameters of the ACCAI connection under baseline condi-
empathy and the reciprocity group were pooled in (B), because there were no tions and the individual average number of altruistic decisions toward the
significant differences in the corresponding best-model parameters [ACCAI, baseline partner.

1076 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

significantly predicted the subjects induced al- the classification algorithm. This analysis did These findings also raise the question whether
truistic motives, i.e., whether subjects went through not yield classification accuracy significantly the differential network components present in
the empathy or the reciprocity induction condi- above chance (classification accuracy = 55.2%, the reciprocity condition are related to particular
tion (classification accuracy = 68.4%, P = 0.016). P = 0.3). It was also not possible to classify the psychological features. Inspired by previous evi-
The computation of classification accuracy with motives based on the increases in the frequency dence that has linked the social evaluation of
the associated P-value is based on the posterior of altruistic decisions in the two motive-induc- other individuals to activation in ventral striatal
probability of balanced accuracy as in (17). We tion conditions relative to the baseline condi- regions (31, 32), we correlated subjects likability
also tested if the two motives can be classified on tion [classification accuracy = 41%, P = 0.93; ratings of the reciprocity partner with the individ-
the basis of conventional functional activations. see also (29)]. ual strengths of the AIVS connectivity. There
We extracted the corresponding beta values from To better understand which DCM parameters was a significant positive correlation [r(16) = 0.59,
the traditional GLM analysis (empathy versus jointly enabled the distinction between the em- P = 0.016], whereas no such correlation was present
baseline condition; reciprocity versus baseline pathy and the reciprocity motive, we calculated in the empathy and the baseline conditions [em-
condition) from the same ROIs that were used the average classification weights with which each pathy condition: r(18) = 0.29, P = 0.25; baseline
for the DCM analyses and submitted them to of the seven different network components con- condition: r(16) = 0.001, P = 0.99].
tributed to the classification (Fig. 2B), and the There is, however, also a common network
stepwise classification performance of the sequen- featurethe ACCAI connectivitythat is pres-
tial combination of network components (Fig. 2, C ent in all three conditions. We thus hypothe-
and D). The strongest classification weight was sized that this component of the network might
attached to the connection from AI to VS, followed reflect basic prosocial motivation. To test this
by the directional connection from the cingulate conjecture, we correlated individuals frequency
cortex to the insular cortex and vice versa (see of altruistic decisions toward their baseline part-
red arrows in Fig. 2A). The combination of these ner with their DCM parameters of the ACCAI
three network components was sufficient for connection. The results show a significant posi-
classifying the induced motives behind subjects tive correlation [r (34) = 0.4, P = 0.017] (Fig. 3D).
altruistic decisions with an accuracy of 77% (P = The stronger a persons connectivity from ACC to
0.0007; Fig. 2, C and D). AI, the higher the baseline level of altruism. In-
To identify the motive-specific functional net- deed, if we divide our sample of subjects in pro-
work architectures, we selected the best average social and selfish individuals on the basis of a
network for altruistic decisions under the empa- median split in the frequency of altruistic de-
thy, reciprocity, and baseline conditions. The cisions in the baseline condition, we find that
selection procedure was based on 28 different prosocial individuals display a positive connec-
models (fig. S3). To define this model space, we tivity from ACC to AI of 0.16, whereas selfish
used the criteria of anatomical plausibility, net- individuals show a negative connectivity of 0.17
work coherence, and functional plausibility [see a difference that is highly significant [F(1,32) =
methods section in supplementary materials (21) 9.49, P = 0.004].
for details]. We used random-effect Bayesian If selfish and prosocial subjects display differ-
model selection (30) to select the models whose ent network architectures, they may also respond
structure and patterns of inputs and effective differently to the empathy and the reciprocity in-
connectivity fit best with the neural processes duction. We sorted the selfish and prosocial in-
evoked by altruistic decisions under the differ- dividuals by the respective motive-induction
ent conditions (21). condition; this resulted in nine prosocial sub-
Figure 3 shows the best models of neural con- jects in both the empathy and the reciprocity
nectivity during empathy-driven, reciprocity-driven, condition, nine selfish subjects in the empathy
and homegrown altruism in the baseline condi- condition, and seven selfish subjects in the reci-
tion (see figs. S3 and S4 for details). A comparison procity condition. Figure 4A shows that the in-
between Fig. 3, A and B, shows a marked simi- duction of the empathy motive significantly
larity in the functional network architecture of increased altruistic decisions in selfish indi-
empathy-driven and homegrown altruism. Both viduals, whereas there was no such effect after
models show a positive connectivity between ACC the induction of reciprocity. In contrast, the
Fig. 4. Differential impact of the empathy and and AI, and a slightly negative connectivity be- induction of reciprocity resulted in a further
the reciprocity motive on individuals with pre- tween AI and VS, with no significant differences enhancement of altruistic behavior in prosocial
dominantly selfish and prosocial preferences. in the respective DCM parameters [ACC to AI, individuals, whereas the empathy induction had
(A) Change in the frequency of altruistic decisions t(1,17) = 0.24, P = 0.8; AI to VS, t(1,17) = 0.66, no effect on these individuals. Thus, the two
after the empathy and the reciprocity induction, com- P = 0.52; ACC input, t(1,17) = 0.89, P = 0.34]. types of subjects respond very differently to the
pared to baseline, in selfish and prosocial individ- This contrasts sharply with the best model for two motive inductions [motive induction indi-
uals. The empathy induction only increases the reciprocity-driven altruism (Fig. 3C). In this model vidual type, F(1,30) = 8.8, P = 0.006].
frequency of altruistic decisions in more selfish there is, first, a strong bidirectional projection We next compared the differences in brain con-
subjects, whereas the reciprocity induction only between AI and ACC. Second, there is a strong nectivity between the motive-induction and the
increases the frequency of altruistic decisions in positive connectivity between AI and VS, which baseline conditions. Empathy induction increased
subjects with more prosocial preferences. (B) Change significantly differs from the negative AI-to-VS effective connectivity from ACC to AI in selfish
in brain connectivity from the ACC to the AI after connectivity under the baseline [t(1,15) = 2.8, P = subjects, while there was no such effect after the
the empathy and reciprocity induction, compared 0.015] and empathy conditions [t(1,32) = 2.91, P = selfish subjects received a reciprocity induction
to baseline, in selfish and prosocial individuals. 0.006; a(Bonferroni-corrected) < 0.025]. There were no and in prosocial individuals [motive induction
(C) Change in brain connectivity from the AI to the significant differences with regard to the con- individual type, F(1,30) = 4.1, P = 0.05] (Fig. 4B).
VS after the empathy and reciprocity induction, com- nectivity from ACC to AI [reciprocity versus base- In contrast, the reciprocity induction led to an
pared to baseline, in selfish and prosocial individuals. line, t(1,15) = 0.03, P = 0.9; reciprocity versus enhancement of neural connectivity from AI to
Error bars indicate SEM. empathy, t(1,32) = 0.28, P = 0.78]. VS in prosocial individuals, which is not the case

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1077


R ES E A RC H | R E PO R TS

after the empathy induction and in selfish types 4. E. Fehr, A. Rangel, J. Econ. Perspect. 25, 330 (2011). decisions in the motive-induction conditions, the behavioral
[motive induction individual type, F(1,30) = 5. K. Lewin, Field Theory in Social Science (Harper, New York, 1951). classification becomes marginally significant (classification
6. R. B. Zajonc, Am. Psychol. 35, 151175 (1980). accuracy of 64.2%, P = 0.051). However, if we perform the
4.9, P = 0.034] (Fig. 4C). 7. R. E. Nisbett, T. D. Wilson, Psychol. Rev. 84, 231259 (1977). same classification analysis with connectivity datai.e., in
Motives are purely mental constructs that are 8. C. D. Batson, The Altruism Question: Toward a Social- addition to the -DCM parameters we also use the level of the
not directly observable. Here we show, however, Psychological Answer (Erlbaum, Hillsdale, NJ, 1991). DCM parameters in the baseline condition for classification
that distinct motives have a distinct neurophys- 9. E. Fehr, U. Fischbacher, Nature 425, 785791 (2003). purposesthe classification accuracy increases even to 83%,
10. J. Moll et al., Proc. Natl. Acad. Sci. U.S.A. 103, 1562315628 (2006). P = 0.00004. Thus, the classification based on brain
iological representation in the brain. Although 11. W. T. Harbaugh, U. Mayr, D. R. Burghart, Science 316, connectivity data clearly outperforms the behavior-based
the empathy and the reciprocity motive increase 16221625 (2007). classification (see also supplementary materials).
the frequency of altruistic acts by the same amount 12. Y. Morishima, D. Schunk, A. Bruhin, C. C. Ruff, E. Fehr, Neuron 30. K. E. Stephan, W. D. Penny, J. Daunizeau, R. J. Moran,
relative to the baseline condition, they are asso- 75, 7379 (2012). K. J. Friston, Neuroimage 46, 10041017 (2009).
13. M. Rabin, Am. Econ. Rev. 83, 12811302 (1993). 31. M. R. Delgado, R. H. Frank, E. A. Phelps, Nat. Neurosci. 8,
ciated with different patterns of brain connectivity 14. A. Falk, U. Fischbacher, Games Econ. Behav. 54, 293315 (2006). 16111618 (2005).
that enabled us to predict the different motives 15. K. J. Friston, L. Harrison, W. Penny, Neuroimage 19, 12731302 32. A. Tusche, T. Kahnt, D. Wisniewski, J. D. Haynes, Neuroimage
with relatively high accuracy. We predicted each (2003). 72, 174182 (2013).
subjects induced motive with a classifier whose 16. K. E. Stephan et al., Neuroimage 42, 649662 (2008).
17. K. H. Brodersen et al., PLOS Comput. Biol. 7, e1002079 (2011). AC KNOWLED GME NTS
parameters were not influenced by that subjects 18. T. Singer et al., Science 303, 11571162 (2004). We thank K. E. Stephan for useful comments on early versions of the
brain data (nor by that subjects behavioral data). 19. G. Hein, G. Silani, K. Preuschoff, C. D. Batson, T. Singer, Neuron manuscript and K. Treiber and S. Klein for assistance with data
Instead, the parameters of the classifier were solely 68, 149160 (2010). collection. We also thank two anonymous referees for their helpful
informed by other subjects brain data. This means 20. R. A. Emmons, M. E. McCullough, The Psychology of Gratitude comments. The paper is part of the advanced European Research
(Oxford Univ. Press, 2004). Council grant on the Foundations of Economic Preferences and the
that the motive-specific brain connectivity patterns 21. Methods and materials, supplementary analyses, supplemen- Synergia grant of the Swiss National Science Foundation on the
are generalizable across subjects. The distinct and tary figures, and supplementary tables are available as neuroeconomics of value-based decision making (CRSII3_141965).
across-subjectgeneralizable neural representa- supporting material on Science Online. The data are stored on the server of the Laboratory for Social and
tion of the different motives thus provides evi- 22. D. Tomlin et al., Science 312, 10471050 (2006). Neural Systems Research at the University of Zurich.
23. J. K. Rilling et al., Neuropsychologia 46, 12561266 (2008).
dence for a distinct neurophysiological existence 24. K. L. Phan, C. S. Sripada, M. Angstadt, K. McCabe, Proc. Natl. SUPPLEMENTARY MATERIALS
of motives.

Downloaded from on March 3, 2016


Acad. Sci. U.S.A. 107, 1309913104 (2010).
www.sciencemag.org/content/351/6277/1074/suppl/DC1
The findings also provide mechanistic insights 25. C. Lamm, J. Decety, T. Singer, Neuroimage 54, 24922502 (2011).
Materials and Methods
into the neural underpinnings of important al- 26. S. N. Haber, B. Knutson, Neuropsychopharmacol 35, 426 (2010).
Figs. S1 to S4
27. A. Parent, L. N. Hazrati, Brain Res. Brain Res. Rev. 20, 91127 (1995).
truistic motives and how motive inductions change 28. J. A. Hoeting, D. Madigan, A. E. Raftery, C. T. Volinsky, Stat. Sci.
Tables S1 and S2
the underlying neural network. In particular, pre- References (3337)
14, 382 (1999).
dominantly selfish individuals were character- 29. If we use both the number of altruistic decisions in the baseline 15 June 2015; accepted 22 January 2016
ized by a low or even negative connectivity from condition and the increase in the frequency of altruistic 10.1126/science.aac7992
ACCAI in the baseline condition, whereas pre-
dominantly prosocial individuals displayed a pos-
itive connectivity between these regions. However,
when we induce the empathy motive, the selfish, EBOLA VIRUS
but not the prosocial, types become more altruistic
and show a substantial increase in ACCAI con-
nectivity. Thus, after the empathy induction, selfish Isolation of potent neutralizing
individuals resemble homegrown unconditional
altruists in terms of both brain connectivity and
altruistic behavior. This contrasts with the effect
antibodies from a survivor of the 2014
of inducing the reciprocity motive, which ren-
ders the prosocial, but not the selfish, types more
Ebola virus outbreak
altruistic and increases only the prosocial types
AIVS connectivity. Zachary A. Bornholdt,1 Hannah L. Turner,2 Charles D. Murin,1,2 Wen Li,3 Devin Sok,1
We obtain these mechanistic insights because Colby A. Souders,4 Ashley E. Piper,5 Arthur Goff,5 Joshua D. Shamblin,5
the inputs into the support vector machine are Suzanne E. Wollen,5 Thomas R. Sprague,5 Marnie L. Fusco,1 Kathleen B. J. Pommert,1
not merely brain activations but small brain Lisa A. Cavacini,4 Heidi L. Smith,4 Mark Klempner,4 Keith A. Reimann,4
models of how relevant brain regions interact Eric Krauland,3 Tillman U. Gerngross,3 Karl D. Wittrup,3 Erica Ollmann Saphire,1
with each other (i.e., functional neural architec- Dennis R. Burton,1,6 Pamela J. Glass,5 Andrew B. Ward,2 Laura M. Walker3*
tures). Thus, by correctly predicting the induced
motives, we simultaneously determine those Antibodies targeting the Ebola virus surface glycoprotein (EBOV GP) are implicated in
mechanistic models of brain interaction that best protection against lethal disease, but the characteristics of the human antibody response
predict the motives. And it is these models that to EBOV GP remain poorly understood. We isolated and characterized 349 GP-specific
deliver the mechanistic insights into brain func- monoclonal antibodies (mAbs) from the peripheral B cells of a convalescent donor who
tion and how changes in brain function relate survived the 2014 EBOV Zaire outbreak. Remarkably, 77% of the mAbs neutralize live
to behavioral changes due to motive inductions. EBOV, and several mAbs exhibit unprecedented potency. Structures of selected mAbs in
Our study, therefore, also demonstrates how mere complex with GP reveal a site of vulnerability located in the GP stalk region proximal to the
prediction and insights into the mechanisms viral membrane. Neutralizing antibodies targeting this site show potent therapeutic
that underlie psychological concepts (such as efficacy against lethal EBOV challenge in mice. The results provide a framework for the
motives) can be simultaneously achieved if func- design of new EBOV vaccine candidates and immunotherapies.
tional neural architectures are the inputs for the

I
prediction.
n recent years, Ebola virus (EBOV) outbreaks trimer is the sole known target for protective
RE FE RENCES AND N OT ES
have increased in frequency, duration, and antibodies and is currently the primary target for
1. P. A. Samuelson, Economica 15, 243253 (1948).
2. P. W. Glimcher, A. Rustichini, Science 306, 447452 (2004).
geographical spread, underscoring the need antiviral vaccines and therapies (2, 3). A small num-
3. A. Rangel, C. Camerer, P. R. Montague, Natl. Rev. 9, 545556 for pre- and post-exposure treatments (1). The ber of protective monoclonal antibodies (mAbs)
(2008). membrane-anchored EBOV glycoprotein (GP) to GP have been isolated from immunized mice,

1078 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

and recent structures of these antibodies in com- mAbs, 349 bound to EBOV GP in preliminary domainfree GP1 core and glycan cap sequence
plex with GP have illuminated key sites of vulner- binding screens (table S1). Analysis of the heavy- (fig. S5) (17, 18). This analysis revealed that 39%
ability on the EBOV glycoprotein (37). However, and light-chain variable regions (VH and Vk, re- of GPDmuc-reactive mAbs failed to bind to sGP,
only a small number of mAbs to GP have been spectively) revealed that the anti-GP repertoire 2% bound with similar apparent affinity to both
isolated from human EBOV survivors (810), and was highly diverse, containing 294 independent GPDmuc and sGP, and 59% reacted with both
therefore the characteristics of the human antibody clonal lineages (fig. S3A and table S2). This result proteins but bound with higher apparent affinity
response to EBOV GP remain largely undefined. contrasts with previously described anti-HIV and to sGP (Fig. 1, B and C, and table S3). The latter
In this study, we aimed to comprehensively anti-influenza repertoires, which show a signifi- result is consistent with previous studies showing
profile the human B cell response to EBOV GP by cantly higher degree of clonal restriction (11, 14). that sGP is secreted in large quantities during
cloning an extensive panel of mAbs to GP from Comparison to nonGP-reactive antibodies (15) re- natural infection and may behave as an antigenic
the peripheral B cells of a convalescent donor vealed that the EBOV GPspecific repertoire was decoy by redirecting the immune response toward
(subject 45) who survived the 2014 EBOV Zaire skewed toward Ig light-chain kappa (Igk) versus epitopes that are either inaccessible on surface
outbreak. Three months after primary infec- Ig light-chain lambda (Igl) and longer heavy- GP or shared between the two proteins (17, 19).
tion, the donor plasma showed strong immuno- chain complementarity-determining region 3 To further define the epitopes targeted by the
globulin G (IgG) binding reactivity to EBOV GP (CDRH3) lengths (fig. S3, B and C, and table S2). mAbs to GP, we performed competitive binding
and potent neutralizing activity, suggesting that Similar biases have also been observed in HIV- experiments (20). We first tested the 321 mAbs
this donor had mounted a robust neutralizing 1infected patient repertoires (11, 12). VH and Vk for competition with two well-characterized murine
antibody (Nab) response to GP by this time point germline gene usage in the GP-specific repertoire mAbs, 1H3 and 13C6, that recognize overlapping
(fig. S1, A and B). To assess the magnitude of the was similar to nonGP-specific repertoires (15, 16) epitopes in the glycan cap (4). The vast majority
B cell response to EBOV GP, B cells were stained (fig. S3, D and E, and table S2). As expected for of sGP cross-reactive binders competed with one
with a fluorescently labeled EBOV GP ectodomain antibodies derived from IgG+ B cells, almost all of or both mAbs, suggesting that they also bind with-
(GPDTM) (4) and analyzed by flow cytometry. the GP-specific clones were somatically mutated, in the glycan cap (Fig. 2A and fig. S6A). We next
Approximately 3% of IgG+ B cells were specific with an average of 5.1 and 2.7 nucleotide sub- tested the GP-specific mAbs for competition with
for GPDTM (fig. S2), which is comparable to the stitutions in VH and VL, respectively (fig. S3F and KZ52, a human antibody that binds at the inter-
percentage of circulating antigen-specific periph- table S2). face of GP1 and GP2 (6, 8). Approximately half of
eral B cells observed during chronic HIV infec- To map the antigenic specificities of the mAbs the GP-specific binders competed with KZ52 (Fig.
tion and after primary dengue infection (11, 12). to GP, we produced 321 IgGs in larger quantities 2A and fig. S6B), suggesting that this antigenic
Cognate antibody heavy- and light-chain pairs were and performed biolayer interferometry (BLI) bind- site is a common target for antibodies elicited
rescued from 420 individual GPDTM-reactive ing experiments with several GP variants. We first by natural EBOV infection, at least for the donor
B cells by single-cell polymerase chain reaction tested binding to EBOV GPDTM and a mucin-like studied. Because KZ52 has been shown to ex-
and were subsequently cloned and expressed domain deletion construct (GPDmuc) (6). Unex- hibit specificity for Zaire GP (6), we next tested
as full-length IgGs in an engineered strain of pectedly, only two mAbs failed to bind to GPDmuc, selected KZ52 competitors for cross-reactivity with
Saccharomyces cerevisiae (13). Of the 420 cloned indicating that less than 1% of the GP-specific Sudan (SUDV) GP and Bundibugyo (BDBV) GP.
antibody response in this donor is directed against Similar to KZ52, most of these mAbs did not
1
Department of Immunology and Microbial Science, The epitopes within or dependent on the mucin-like show broad species cross-reactivity (Fig. 2B and
Scripps Research Institute, La Jolla, CA 92037, USA. domain (Fig. 1A and table S3). About 30% of the fig. S7A). However, in contrast to KZ52 and other
2
Department of Integrative Structural and Computational mAbs showed increased binding responses and well-characterized GP base binders (4), most of
Biology, The Scripps Research Institute, La Jolla, CA 92037,
USA. 3Adimab, Lebanon, NH 03766, USA. 4MassBiologics,
faster association rates to GPDmuc as compared the KZ52 competitor mAbs failed to react with a
University of Massachusetts Medical School, Boston, MA to GPDTM (fig. S4), suggesting that these mAbs minimal thermolysin-processed GP core, in which
02126, USA. 5U.S. Army Medical Research Institute of probably recognize epitopes that are partially oc- both the mucin domain and glycan cap regions
Infectious Diseases, Frederick, MD 21702, USA. 6Ragon cluded by the mucin-like domain. We next tested had been proteolytically removed (GPCL) (4) (Fig.
Institute of Massachusetts General Hospital, Massachusetts
Institute of Technology, and Harvard University, Cambridge,
the mAbs for binding to a secreted GP isoform, 2C). Thus, this class of antibodies appears to tar-
MA 02142, USA. sGP, which is expressed as a disulfide-linked GP1 get specific epitopes that either directly overlap
*Corresponding author. E-mail: laura.walker@adimab.com dimer containing the majority of the mucin-like with the KZ52 epitope or are sterically inhibited

Fig. 1. Antigen-binding properties of mAbs to GP. (A) Apparent binding mined by BLI measurements. (C) Pie chart summarizing antibody binding profiles.
affinities of GP-specific IgGs to Zaire GPDTM and Zaire GPDmuc constructs as Cross-reactive mAbs refer to those that bind to both GP and sGP. N.B., nonbinder;
determined by BLI measurements. Newly discovered mAbs to GP are shown as W.B., weak binder. IgG equilibrium dissociation constants were calculated for mAbs
red circles. KZ52 IgG (yellow diamond), 13C6 IgG (green triangle), 1H3 IgG (orange with BLI responses >0.1 nm. MAbs with BLI responses <0.05 nm were desig-
square), and 2G4 IgG (purple hexagon) are included for comparison. (B) Apparent nated as N.B.s; MAbs with BLI responses between 0.05 and 0.1 nm were desig-
binding affinities of GP-specific IgGs to Zaire sGP and Zaire GPDmuc as deter- nated as W.B.s. All data are representative of two or more independent experiments.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1079


R ES E A RC H | R E PO R TS

Fig. 2. Epitope mapping. (A) Percentage of sGP-reactive and sGPnon-reactive mAbs directed
against each antigenic site on EBOV GP. Epitope binning was performed using a previously de-
scribed yeast-based competition assay (20). (B) Percentage of selected KZ52 competitors that
cross-react with SUDV GP and BDBV GP. Binding cross-reactivity was assessed by enzyme-linked immunosorbent assay (ELISA). (C) ELISA binding of selected KZ52
competitors to a minimal GP core that contains deletions in the mucin-like domain and glycan cap (GPCL). ELISA binding is expressed as the optical density at 405 nm
(OD405) reading at a concentration of 0.2 mg/ml. (D) Percentage of selected KZ52 noncompetitors that cross-react with SUDV GP and BDBV GP. Binding cross-reactivity
was assessed by ELISA. (E) Summary of the antigenic sites targeted by the mAbs to GP. All data are representative of two or more independent experiments.

Fig. 3. Neutralizing activity of mAbs to GP. (A) Per-


centage of mAbs in each competition group that reached
PRNT50 or PRNT80 at concentrations 50 mg/ml. The
total number of mAbs tested from each competition group
is shown at the top of the corresponding bar. (B) PRNT50
and PRNT80 values of selected mAbs from each com-
petition group. KZ52 IgG is included for comparison (green
inverted triangles). Red bars indicate median PRNT50 and
PRNT80 values. Neutralization assays were performed
using a live virus plaque reduction assay. PRNT50 and
PRNT80 values represent the concentration of IgG re-
quired to reduce viral infectivity by 50 and 80%, respec-
tively. All data are representative of two independent
experiments.

1080 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

by the KZ52 Fab. To estimate the number of dif- We next measured the neutralizing activity of portion of NAbs, with the KZ52 and 13C6/1H3
ferent antigenic sites recognized by the remain- the B cellderived mAbs using a live virus plaque competition groups containing the highest pro-
ing GP-specific mAbs, we performed competitive reduction neutralization (PRNT) assay. Because of portion of NAbs (Fig. 3A and table S5). The latter
binding experiments with four high-affinity mAbs the large number of mAbs and the high-throughput result was unexpected, because 13C6 and 1H3
from the panel (ADI-15974, ADI-15933, ADI-15810, nature of our study, initial neutralization screen- only weakly neutralize in the absence of comple-
and ADI-15983) that did not show significant ing was performed using a single concentration ment (7, 21). We next performed neutralization
competition with KZ52 or with each other (table of purified IgG (table S5 and fig. S9). Remark- titration experiments in order to evaluate neu-
S4). Eighty percent of the nonKZ52-competitive ably, 77 and 63% of the mAbs reduced viral in- tralization potency. These results showed that
GP-specific mAbs bound to epitopes overlapping fectivity by 50 and 80% (PRNT50 and PRNT80), several NAbs, particularly those in the ADI-15974
that of ADI-15974 or ADI-15810 (Fig. 2A and fig. respectively, at concentrations 50 mg/ml (Fig. 3A and KZ52 competition groups, exhibited extraor-
S6B). This group of mAbs also showed signifi- and table S5). Control experiments with yeast- dinary potency. Half of the NAbs tested from
cantly broader cross-species GP-binding reactiv- produced and CHO-produced (CHO, Chinese ham- the ADI-15974 competition group, and two of
ity than the KZ52 competitors (Fig. 2D and fig. ster ovary cells) IgGs demonstrated that functional the NAbs tested from the KZ52 competition group,
S7B). Overall, these data show that the anti-GP activity is probably not affected by the host pro- neutralized with PRNT50 values 0.05 mg/ml (Fig.
repertoire in this patient is primarily composed duction system (fig. S8). Analysis of neutralizing 3B and table S5). In contrast, the majority of 13C6
of clones that target four non-overlapping anti- activity by a competition group revealed that the and/or 1H3 competitor mAbs neutralized with
genic sites on EBOV GP (Fig. 2E). majority of competition groups contained a pro- relatively modest potency, with PRNT50 values
averaging 3.3 mg/ml. We conclude that the GP-
specific antibody repertoire in the donor studied
contains a high proportion of NAbs, the most po-
tent of which bind to epitopes overlapping those
of KZ52 or ADI-15974.
To structurally define the epitopes recognized
by the NAbs, we used single-particle electron mi-
croscopy (EM) to examine five potent NAbs, rep-
resenting each of the four major competition
groups, in complex with fully glycosylated EBOV
GPDTM. These NAbs included ADI-15731 (a 13C6
competitor), ADI-15734 and ADI-15762 (KZ52 com-
petitors), ADI-15758 (an ADI-15974 competitor),
and ADI-15859 (an ADI-15810 competitor). We were
able to obtain negative-stain two-dimensional
(2D) class averages for all five complexes of Fabs
bound to EBOV GPDTM (fig. S10) and 3D recon-
structions for four of the Fab:EBOV GPDTM com-
plexes at 18 to 24 resolution (Fig. 4 and fig. S11).
Fig. 4. Negative-stain In agreement with the competitive binding data,
electron microscopy of the EM reconstruction of ADI-15731 showed that
Fab:EBOV GPDTM complexes. this NAb binds within the glycan cap, with a foot-
(A) A structure-based (PDB print approximately between the epitopes recog-
3CSY and 3S88) (5) surface nized by 13C6 and 1H3 and with a similar angle of
representation of the EBOV GP approach (Fig. 4 and fig. S12A). We next examined
trimer. The mucin domain the two KZ52 competing NAbs, ADI-15734 and
(gray), glycan cap domain of ADI-15762. As anticipated, ADI-15734 bound to
GP1 (aqua green), GP1 core EBOV GPDTM at the GP1/GP2 interface, slightly
(blue), GP2 (light blue), fusion adjacent to the KZ52 epitope and at a similar
loop region of GP2 (pink), and angle of approach (Fig. 4 and fig. S12C). In con-
the stalk/HR2 region (orange) trast, ADI-15762 actually binds within the glycan
have been mapped onto the cap, but with a shallow binding angle that prob-
structure. The residues that ably sterically occludes the KZ52 epitope (Fig. 4
make up the trimeric body and and fig. S12B). Last, we determined the structure
the stalk region of the EBOV GP of EBOV GPDTM in complex with ADI-15758 (an
are displayed on the left. The ADI-15974 competitor), one of the most potent
mucin domains are modeled NAbs described in this panel. The EM reconstruc-
only as spheres, because they tion shows that ADI-15758 binds to a region prox-
are largely unstructured and imal to the viral membrane, distal to all previously
poorly defined (27). Residues described epitopes, and below the body of the
613 to 637 corresponding to the trimeric EBOV GP structure (Fig. 4 and fig. S12D).
stalk/HR2 region were modeled Although this region has not yet been structurally
in silico using threefold characterized at high resolution in the pre-fusion
symmetry and peptide structure GP context, it corresponds to the a-helical heptad
prediction for the HR2 region repeat 2 (HR2; residues 613 to 637) defined in the
(28). (B) Corresponding 3D reconstructions of four Fab:EBOV GPDTM complexes are shown in trans- post-fusion conformation (22). Docking of the
parent surface representation (gray) with the model from (A) fitted in the density. Additionally, structural EBOV GP crystal structure into the reconstruction
models for each Fab variable region were generated using the ROSIE server (29, 30) and then fitted into suggests that the ADI-15758 epitope is within the
the density maps as surface representations. Each structure is shown as side (left) and top (right) views, C-terminal 24 residues of GP2 contained in the
with the exception of ADI-15758, which is shown from the bottom up, respective to the viral membrane. EBOV GPDTM construct (6). Three Fab molecules

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1081


R ES E A RC H | R E PO R TS

Fig. 5. Therapeutic efficacy of NAbs against MA-EBOV. Kaplan-Meier survival curves for ADI-15974 competitor NAbs (A), KZ52 competitor NAbs (B), 13C6
competitor NAbs (C), and NAbs targeting undefined epitopes (D). Mice were infected with 100 PFU of MA-EBOV and treated intraperitoneally with a single
dose of the indicated mAbs at 2 dpi (dotted black line). Negative control mice were treated with phosphate-buffered saline. MAb 2G4 is included for com-
parison. Data are representative of one experiment with 10 mice per group.

could be visualized in the 2D class averages (fig. tition group were also highly effective in pro- clones that primarily target three non-overlapping
S10), suggesting that the HR2 region may exist as tection, with survival rates ranging between 60 antigenic sites on the GP spike: the glycan cap;
a three-helix bundle in the pre-fusion GP struc- and 100% (Fig. 5B and table S6). With the excep- the GP1/GP2 interface; and the stalk , inclusive of
ture (6). Additionally, although we were not able tion of ADI-15818 (a KZ52 competitor), all of the the HR2 helices (fig. S14). A substantial fraction
to generate a 3D reconstruction of ADI-15859 (an NAbs in these two competition groups showed of the mAbs cloned from GP-specific B cells show
ADI-15810 competitor) bound to EBOV GPDTM, greater therapeutic efficacy than 2G4, which only neutralizing activity, demonstrating that at least
the negative-stain 2D class averages indicate that provided 40% protection under these conditions. in this donor, NAb responses can develop rela-
this mAb also binds within the GP stalk region. In contrast to the ADI-15974 and KZ52 compet- tively early after EBOV infection. The most po-
Collectively, these data suggest that the GP stalk itor NAbs, the NAbs targeting the glycan cap tently neutralizing and therapeutically effective
region containing the HR2 helices is an accessi- (13C6/1H3 competitors) and undefined epitopes mAbs in our panel target the GP1/GP2 interface
ble antigenic region targeted by NAbs, a propor- generally showed little to no therapeutic efficacy and the GP stalk region, suggesting that these
tion of which exhibit remarkable neutralization (Fig. 5, C and D, fig. S13, and table S6). Only ADI- epitopes may be promising targets for rational
potency. 16037 (a 13C6 competitor NAb) provided potent vaccine design. In addition, the observation that
Finally, we sought to determine whether cer- protection, yielding 80% survival and 7% average EBOV escape variants can emerge after treatment
tain NAbs to GP showed greater in vivo efficacy weight loss. The remaining NAbs in these groups with the MB-003 antibody cocktail highlights the
than others. For this experiment, several NAbs yielded 50% survival, which in most cases was need for protective mAbs that target new anti-
were selected from each competition group and not a statistically significant increase in protec- genic sites, such as those described here target-
evaluated for post-exposure therapeutic effica- tion over the negative control (Fig. 5, C and D, and ing the GP stalk (25, 26).
cy against lethal EBOV challenge in a murine table S6). This result is consistent with previous
infection model (23). Groups of mice were chal- studies showing that mAbs targeting the glycan REFERENCES AND NOTES

lenged with a target dose of 100 plaque-forming cap generally do not afford significant protection 1. P. Roddy, Viruses 6, 36993718 (2014).
2. A. Marzi et al., Science 349, 739742 (2015).
units (PFU) of mouse-adapted EBOV (MA-EBOV), when administered at 1 to 2 dpi (3, 24). In sum- 3. X. Qiu et al., Nature 514, 4753 (2014).
followed by a single 100 mg dose of mAb at 2 days mary, most of the NAbs targeting the GP stalk 4. C. D. Murin et al., Proc. Natl. Acad. Sci. U.S.A. 111, 1718217187
post-infection (dpi). The previously described neu- region (ADI-15974 competitors) and the GP1/GP2 (2014).
tralizing mAb 2G4, a component of the ZMapp interface (KZ52 competitors) provided significant 5. J. M. Dias et al., Nat. Struct. Mol. Biol. 18, 14241427 (2011).
6. J. E. Lee et al., Nature 454, 177182 (2008).
cocktail, was also included for comparison (3). post-exposure protection against lethal EBOV chal- 7. G. G. Olinger Jr. et al., Proc. Natl. Acad. Sci. U.S.A. 109,
Of significance, all of the ADI-15974 competitor lenge, whereas NAbs targeting the glycan cap 1803018035 (2012).
NAbs (GP stalk binders) provided significant post- (13C6/1H3 competitors) and undefined regions 8. T. Maruyama et al., J. Virol. 73, 60246030 (1999).
exposure protection, with survival rates ranging generally showed little to no therapeutic efficacy 9. W. B. Oswald et al., PLOS Pathog. 3, e9 (2007).
10. A. I. Flyak et al., Cell 164, 392405 (2016).
from 60 to 100% and average weight loss ranging under these conditions. 11. J. F. Scheid et al., Nature 458, 636640 (2009).
between 8 and 10% (Fig. 5A, fig. S13, and table We have shown that the human B cell response 12. M. Beltramello et al., Cell Host Microbe 8, 271283 (2010).
S6). Five out of the six NAbs in the KZ52 compe- to EBOV GP is composed of a broad diversity of 13. Y. Xu et al., Protein Eng. Design Select. 26, 663670 (2013).

1082 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

14. J. Wrammert et al., Nature 453, 667671 (2008). 30. A. Sivasubramanian, A. Sircar, S. Chaudhury, J. J. Gray, U19AI109762, Consortium for Immunotherapeutics Against
15. S. D. Boyd et al., J. Immunol. 184, 69866992 (2010). Proteins 74, 497514 (2009). Viral Hemorrhagic Fevers. E.O.S. was also supported by
16. K. J. Jackson et al., Immunogenetics 64, 314 (2012). R01AI067927. C.D.M. was supported by a predoctoral fellowship
17. A. Sanchez, S. G. Trappier, B. W. Mahy, C. J. Peters, S. T. Nichol, ACKN OWLED GMEN TS from NSF. This study was supported in part by U.S. NIH grants
Proc. Natl. Acad. Sci. U.S.A. 93, 36023607 (1996). We thank T. Boland and M. Vasquez for assistance with antibody U19 AI109762 and R01 AI067927 awarded to E.O.S. Research
18. S. K. Gire et al., Science 345, 13691372 (2014). sequence analysis, C. Williams and S. M. Eagol for assistance was funded in part by the Defense Advanced Research Projects
19. G. S. Mohan, W. Li, L. Ye, R. W. Compans, C. Yang, with figure preparation, R. Pejchal for providing helpful comments Agency (DARPA-BAA-13-03). D.R.B. and D.S. acknowledge support
PLOS Pathog. 8, e1003065 (2012). on the manuscript, and M. Haynes for assistance with flow from Center for HIV/AIDS Vaccine Immunology and Immunogen
20. D. R. Bowley, A. F. Labrijn, M. B. Zwick, D. R. Burton, cytometry. All of the IgGs were sequenced by Adimab's Molecular Discovery Grant UM1AI100663. This is manuscript no. 29237
Protein Eng. Design Select. 20, 8190 (2007). Core and produced by the High Throughput Expression group. from The Scripps Research Institute. Opinions, interpretations,
21. J. Audet et al., Sci. Rep. 4, 6881 (2014). BLI binding experiments were performed by Adimab's protein conclusions, and recommendations are those of the authors
22. W. Weissenhorn, A. Carf, K. H. Lee, J. J. Skehel, D. C. Wiley, analytics group. The ZMapp cocktail mAb 2G4 was generously and are not necessarily endorsed by the U.S. Army.
Mol. Cell 2, 605616 (1998). provided by Mapp Biopharmaceutical. The data presented in
23. M. Bray, K. Davis, T. Geisbert, C. Schmaljohn, J. Huggins, this manuscript are tabulated in the main paper and in the SUPPLEMENTARY MATERIALS
J. Infect. Dis. 179, S248S258 (1999). supplementary materials. GenBank accession numbers for the www.sciencemag.org/content/351/6277/1078/suppl/DC1
24. X. Qiu et al., PLOS Negl. Trop. Dis. 6, e1575 (2012). antibody variable-region gene sequences reported in this study Materials and Methods
25. J. R. Kugelman et al., Cell Rep. 12, 21112120 (2015). can be found in table S7. The cryo-EM maps have been deposited to Figs. S1 to S14
26. X. Qiu et al., Sci. Transl. Med. 4, 138ra81 (2012). the Electron Microscopy Data Bank (accession numbers EMD-6586, Tables S1 to S7
27. T. Hashiguchi et al., Cell 160, 904912 (2015). EMD-6587, EMD-6588, and EMD-6589). E.O.S., Z.A.B., M.L.F., References (3139)
28. Y. Shen, J. Maupetit, P. Derreumaux, P. Tuffry, J. Chem. K.B.J.P., A.B.W., H.L.T., and C.D.M. acknowledge support from
Theory Comput. 10, 47454758 (2014). the NIH/National Institute of Allergy and Infectious Diseases 7 October 2015; accepted 8 February 2016
29. S. Lyskov et al., PLOS One 8, e63906 (2013). Center for Excellence in Translational Research Grant 10.1126/science.aad5788

IMMUNOGENOMICS investigate whether TE-derived regulatory ele-


ments influence biological outcomes.

Regulatory evolution of innate To explore the influence of TEs on IFNG-

Downloaded from on March 3, 2016


inducible regulatory networks, we examined
their contribution to IRF1 and STAT1 binding
immunity through co-option of sites with the use of published chromatin immu-
noprecipitation sequencing (ChIP-seq) data for

endogenous retroviruses three human cell lines treated with IFNG: K562
myeloid-derived cells, HeLa epithelial-derived
cells, and primary CD14+ macrophages (16, 17).
Edward B. Chuong, Nels C. Elde,* Cdric Feschotte* Our initial analysis revealed 27 TE families en-
riched within IFNG-induced binding peaks in at
Endogenous retroviruses (ERVs) are abundant in mammalian genomes and contain least one of the data sets examined (18) (table S1
sequences modulating transcription. The impact of ERV propagation on the evolution of gene and fig. S1, A and B) and included TEs previously
regulation remains poorly understood. We found that ERVs have shaped the evolution of a predicted to be cis-regulatory elements (11, 19).
transcriptional network underlying the interferon (IFN) response, a major branch of innate These sequences contained evolutionarily young
immunity, and that lineage-specific ERVs have dispersed numerous IFN-inducible enhancers to ancient TE families, of which the majority (20 of
independently in diverse mammalian genomes. CRISPR-Cas9 deletion of a subset of these 27) originated from long terminal repeat (LTR) pro-
ERV elements in the human genome impaired expression of adjacent IFN-induced genes and moter regions of ERVs (Fig. 1A). These data suggest
revealed their involvement in the regulation of essential immune functions, including that ERVs, which arose from ancient retroviral in-
activation of the AIM2 inflammasome. Although these regulatory sequences likely arose in fections and currently constitute 8% of the human
ancient viruses, they now constitute a dynamic reservoir of IFN-inducible enhancers fueling genome (20), represent a source of novel binding
genetic innovation in mammalian immune defenses. sites bound by IFNG-inducible transcription factors.

C
We next investigated whether these ERVs may
hanges in gene regulatory networks under- specific cis-elements capable of rewiring regu- contribute to IFNG-inducible regulation of adja-
lie many biological adaptations, but the latory networks, but the adaptive consequences cent cellular genes. ERVs bound by STAT1 and/or
mechanisms promoting their emergence of this process for specific physiological func- IRF1 in CD14+ macrophages were strongly en-
are not well understood. Transposable ele- tions remain largely unexplored. riched near ISGs (binomial test, P = 1.4 1087;
ments (TEs), including endogenous retro- We investigated the evolution of gene regula- Fig. 1B and fig. S2), determined from a matched
viruses (ERVs), have been proposed to facilitate tory networks induced by the proinflammatory RNA-seq data set (table S2) (18, 21). A complemen-
regulatory network evolution because they con- cytokine interferon-g (IFNG). Interferons are pro- tary approach using the genomic regions enrich-
tain regulatory elements and can amplify in num- inflammatory signaling molecules that are re- ment of annotations tool (GREAT) (22) revealed
ber and/or move throughout the genome (13). leased upon infection to promote transcription enrichment of CD14+ STAT1-bound and/or IRF1-
Genomic studies support this model (4), reveal- of innate immunity factors, collectively defined bound ERVs near genes annotated with immune
ing that a substantial fraction of TE-derived as IFN-stimulated genes (ISGs) (13). ISGs are reg- functions (fig. S3, A and B). These findings sug-
noncoding sequences evolve under selective con- ulated by cis-regulatory elements that are bound gest a potentially widespread role for ERVs in
straint (3, 5), are frequently bound by tran- by IRF (interferon regulatory factor) and STAT the regulation of the human IFNG response.
scription factors (610), and often exhibit cell (signal transducer and activator of transcription) MER41 is an endogenized gammaretrovirus
typespecific chromatin states consistent with transcription factors upon activation of IFN sig- that invaded the genome of an anthropoid pri-
regulatory activity (11, 12). These observations naling pathways (13). Although innate immune mate ancestor ~45 to 60 million years ago with
implicate TEs as a potential source of lineage- signaling pathways are conserved among mam- 7190 LTR elements, from six subfamilies (MER41A,
mals, the transcriptional outputs of these path- B, C, D, E, and G), now fixed in the human genome
Department of Human Genetics, University of Utah School of ways differ across species (14, 15), likely reflecting (fig. S4A). Our analysis revealed the primate-
Medicine, Salt Lake City, UT 84112, USA.
lineage-specific adaptation in response to inde- specific MER41 family of ERVs as a source of
*Corresponding author. E-mail: nelde@genetics.utah.edu (N.C.E.);
cedric@genetics.utah.edu (C.F.) These authors contributed equally pendent host-pathogen conflicts. Thus, these path- IFNG-inducible binding sites (fig. S4B), with near-
to this work. ways provide useful systems that allow us to ly 1000 copies in humans (N = 962) bound by

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1083


R ES E A RC H | R E PO R TS

STAT1 and/or IRF1 in at least one cell type vates an inflammatory response (24). AIM2 is binding motifs (Fig. 2D). These binding sites are
(table S3 and fig. S4C). In CD14+ macrophages, IFNG-inducible in humans but is constitutively conserved across anthropoid primates (fig. S7A),
STAT1-bound MER41 elements exhibited stereo- transcribed in mice (24). In humans, MER41.AIM2 and IFNG-inducible reporter activity is conserved
typical induction of histone H3 Lys27 (H3K27) appears to provide the only STAT1 binding site across orthologous MER41.AIM2 elements cloned
acetylation upon IFNG stimulation, a hallmark within 50 kb of the AIM2 gene, and the element from chimpanzee, rhesus macaque, and marmo-
of cis-regulatory enhancer activity (23) (Fig. 1C). gained H3K27 acetylation upon IFNG stimula- set (Fig. 2D). We also confirmed that orthologs
Consistent with the idea that this ERV family tion (Fig. 2A). Therefore, the regulation of AIM2 of AIM2 were all IFNG-inducible in primary fi-
affects IFNG-inducible regulation, MER41B se- has undergone evolutionary divergence across broblasts from these species (fig. S7B). These re-
quences were identified as enriched within STAT1 mammalian lineages, which in turn suggests that sults establish MER41.AIM2 as an IFNG-inducible
ChIP-seq peaks in IFNG-stimulated HeLa cells the transposition of MER41 upstream of AIM2 enhancer and suggest that it was co-opted for
(19). A tandem pair of predicted STAT1 binding may have conferred regulation by IFN signaling AIM2 regulation in an ancestor of anthropoid
sites coincided with STAT1 ChIP-seq peak local- in anthropoid primates. primates.
ization (Fig. 1D). These sites also occur in the We used the CRISPR-Cas9 system to delete The binding of AIM2 to cytoplasmic double-
ancestral (consensus) sequence of the MER41B the MER41.AIM2 element in HeLa cells (fig. S6) stranded DNA from intracellular bacteria and
subfamily (Fig. 1D) but not in the MER41A sub- (18). Cells homozygous for the MER41.AIM2 de- viruses promotes the assembly of a molecular
family, which is characterized by a 43base pair letion (DMER41.AIM2) failed to express AIM2 upon platform known as an inflammasome, which ini-
(bp) deletion that has eliminated these binding IFNG treatment, in contrast to control cells in tiates pyroptotic cell death by cleaving and activat-
sites (fig. S5). MER41A sequences showed no which AIM2 transcript levels were robustly in- ing caspase-1 (25). To test whether MER41.AIM2
enrichment within IFNG-inducible binding sites, duced by IFNG (Fig. 2B). IFNG-induced AIM2 pro- is required for this response to infection, we in-
despite otherwise sharing 99% sequence identity tein levels were undetectable in DMER41.AIM2 cells fected DMER41.AIM2 cells with vaccinia virus
with MER41B (figs. S4B and S5). Together, these (Fig. 2C), thus demonstrating that MER41.AIM2 (VACV) for 24 hours and assayed secretion of
data suggest that many MER41 elements are is necessary for endogenous IFNG-inducible regu- the active cleaved form of caspase-1 (subunit p10)
directly bound by STAT1 upon IFNG treatment, lation of AIM2. as the readout of inflammasome activity. Se-
likely owing to the presence of ancestral STAT1 We further delineated the regulatory activity creted levels of activated caspase-1 were markedly
binding motifs within their LTR sequences. of MER41.AIM2 by means of luciferase reporter reduced in DMER41.AIM2 cells relative to wild-
Next, we focused on the MER41.AIM2 ERV, assays (18). MER41.AIM2 was sufficient to drive type cells, and caspase-1 activation was restored
which is located 220 bp upstream of the gene IFNG-inducible reporter expression in HeLa cells, by transient transfection with an AIM2 over-
Absent in Melanoma 2 (AIM2), an ISG that en- and this activity was significantly diminished by expression construct [pCMV-AIM2 plasmid (Fig.
codes a sensor of foreign cytosolic DNA and acti- point mutations ablating the predicted STAT1 2E)]. Collectively these experiments demonstrate

Fig. 1. Dispersion of IFNG-inducible regulatory


elements by ERVs. (A) Age distribution (left) and
enrichment within ChIP-seq data sets (right) of
27 TE families that were enriched within binding sites
for IFNG-stimulated cells (18). Estimated primate/
rodent divergence time (82 million years ago) is
from (34). (B) Frequency histogram of absolute
distances from each ERV to the nearest ISG, for
CD14+ cells. The background expectation is from
the genome-wide ERV distribution (18). Statistical
significance of the observed enrichment within the
first 10 kb of the nearest ISG was assessed by
binomial test. (C) Heat map of CD14+ ChIP-seq
signals centered across STAT1 peak summits with-
in MER41B elements. Bottom metaprofiles rep-
resent average normalized ChIP signal across bound
elements. (D) Schematic of the MER41B LTR con-
sensus sequence.Triangles indicate gamma activated
site (GAS; TTCNNNGAA, where N = any nucle-
otide) motifs predicted to bind STAT1 in response
to IFNG (13). Heat map depicts the presence of
GAS motifs across 728 extant STAT1-bound MER41B
copies in HeLa cells (18). Bottom metaprofile rep-
resents average presence of STAT1 motifs relative
to the MER41 consensus sequence, overlain with
normalized STAT1 ChIP-seq density across the
same elements.

1084 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

that MER41.AIM2 is likely a necessary element we used CRISPR-Cas9 to generate genomic of these MER41 elements did not completely abol-
of the inflammatory response to infection. deletions of MER41.APOL1, MER41.IFI6, and ish IFNG-induced transcript levels of these genes.
The dispersion of cis-regulatory elements prop- MER41.SECTM1 in HeLa cells (figs. S8 and S9). This difference may be due to additional STAT1
agated by the same TE family might facilitate Upon treatment with IFNG, each mutant cell line binding sites located near these genes (Fig. 3A).
the recruitment of multiple genes into the same exhibited significantly decreased transcript lev- In such cases, MER41 elements may contribute
regulatory network (3). We identified three ad- els of the corresponding ISG relative to wild-type regulatory robustness as partially redundant or
ditional MER41 elements within 20 kb of APOL1, levels (Fig. 3B), indicating that these MER41 ele- shadow enhancers (29).
IFI6, and SECTM1, which all are involved in human ments had also been co-opted as IFNG-inducible ERVs related to the primate-specific MER41
immunity (2628) (Fig. 3A). As with MER41.AIM2, enhancers. However, in contrast to AIM2, deletion family (MER41-like) have been identified in

Fig. 2. A MER41 element is essential for AIM2 inflammasome activation. (A) Genome browser view of AIM2. ChIP-seq tracks are normalized per million
reads. The uniqueness track displays genome-wide short-read alignability. (B) Quantitative polymerase chain reaction (qPCR) of AIM2 levels in wild-type and
DMER41.AIM2 HeLa cells after 24 hours of IFNG treatment. (C) Western blot of AIM2 in wild-type and DMER41.AIM2 cells after IFNG treatment. (D) Luciferase
reporter assays of MER41.AIM2, MER41.AIM2 with mutations in the predicted STAT1 sites, and primate orthologs of MER41.AIM2 (see fig. S7A). (E) Western
blot of caspase-1 from supernatants of wild-type and DMER41.AIM2 cells infected with vaccinia virus (18). *P < 0.05, Students t test. Error bars denote SD.

Fig. 3. Multiple MER41 ele-


ments have been co-opted to
regulate the IFNG response.
(A) Genome browser views of
MER41 elements located near
APOL1, IFI6, and SECTM1.
ChIP-seq data are depicted as
normalized signal per million
reads. (B) qPCR of each gene
comparing IFNG-inducible levels
in wild-type HeLa cells and
MER41 deletion mutants.
*P < 0.05, Students t test. Error
bars denote SD.

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1085


R ES E A RC H | R E PO R TS

Fig. 4. IFNG-inducible ERVs are pervasive in mammalian genomes.


(A) A consensus mammalian species phylogeny overlain with boxplots
(median and 25th/75th percentiles) depicting the estimated age of
MER41-like amplifications (18). My, million years ago; triangles depict
conserved GAS motifs. (B) Luciferase reporter assays of MER41-like LTR
consensus sequences from cow and dog (18). (C) Heat map of ChIP-seq
signals centered on STAT1 peak summits within muroid-specific RLTR30B
elements. Columns depict STAT1 ChIP-seq data from mouse bone marrow
derived macrophages (BMM) that were either untreated or treated with
IFNB or IFNG. Only RLTR30B elements that are bound by STAT1 upon IFNG
treatment are shown. Bottom metaprofiles represent average normalized ChIP signal across bound elements. (D) Rodent phylogeny overlain with a
boxplot depicting the amplification of RLTR30B, as in (A). ISRE denotes interferon-stimulated response element motif (TTTCNNTTTC) predicted to bind
STAT1 in response to IFNB (13). (E) Luciferase reporter assay of RLTR30B consensus sequence, as in (B). [Time-calibrated phylogenies in (A) and (D) are
from (34).] *P < 0.05, Students t test. Error bars denote SD.

most major mammalian lineages (30), raising the of RLTR30B also provides IFNG-inducible en- 3. C. Feschotte, Nat. Rev. Genet. 9, 397405 (2008).
possibility of similar contributions to immune reg- hancer activity in HeLa cells (Fig. 4E). GREAT 4. R. Rebollo, M. T. Romanish, D. L. Mager, Annu. Rev. Genet. 46,
2142 (2012).
ulation. Further analysis, including cross-species analysis also revealed significant enrichment of
5. C. B. Lowe, G. Bejerano, D. Haussler, Proc. Natl. Acad. Sci. U.S.A.
genomic alignments, confirmed that multiple mouse STAT1-bound ERVs near functionally an- 104, 80058010 (2007).
mammalian lineages were independently colo- notated immunity genes (fig. S11B). 6. T. Wang et al., Proc. Natl. Acad. Sci. U.S.A. 104, 1861318618
nized by related MER41-like gammaretroviruses Together, our findings reveal IFN-inducible (2007).
~50 to 75 million years ago (table S4). Remark- enhancers introduced and amplified by ERVs in 7. G. Kunarso et al., Nat. Genet. 42, 631634 (2010).
8. D. Schmidt et al., Cell 148, 335348 (2012).
ably, we found that the tandem STAT1 binding many mammalian genomes. On occasion, these 9. E. B. Chuong, M. A. K. Rumi, M. J. Soares, J. C. Baker, Nat.
motifs present in anthropoid MER41 are con- elements have been co-opted to regulate host Genet. 45, 325329 (2013).
served in MER41-like relatives found in lemuri- genes encoding immunity factors. Although we 10. J. H. Notwell, T. Chung, W. Heavner, G. Bejerano, Nat.
formes, vesper bats, carnivores, and artiodactyls have shown that ERVs play a functional role Commun. 6, 6644 (2015).
11. P.-. Jacques, J. Jeyakani, G. Bourque, PLOS Genet. 9,
(Fig. 4A and fig. S10), which suggests that they regulating innate immune pathways in human e1003504 (2013).
might also have dispersed IFN-inducible enhancers HeLa cells, further studies will be necessary to 12. V. Sundaram et al., Genome Res. 24, 19631976
in the genomes of these species. Consistent with extend our findings to primary hematopoietic (2014).
this prediction, we found that reconstructed an- cells and other species such as mouse. We spec- 13. L. C. Platanias, Nat. Rev. Immunol. 5, 375386 (2005).
14. L. B. Barreiro, J. C. Marioni, R. Blekhman, M. Stephens, Y. Gilad,
cestral (consensus) sequences of MER41-like LTRs ulate that the prevalence of IFN-inducible en- PLOS Genet. 6, e1001249 (2010).
from dog and cow can drive robust IFNG-inducible hancers in the LTRs of these ancient retroviruses 15. K. Schroder et al., Proc. Natl. Acad. Sci. U.S.A. 109, E944E953
reporter activity in HeLa cells (Fig. 4B). is not coincidental, but may reflect former viral (2012).
These results suggest that ERVs may have in- adaptations to exploit immune signaling pathways 16. M. B. Gerstein et al., Nature 489, 91100 (2012).
17. Y. Qiao et al., Immunity 39, 454469 (2013).
dependently expanded the IFN regulatory network promoting viral transcription and replication (32). 18. See supplementary materials on Science Online.
in multiple mammalian lineages. To further inves- Indeed, several extant viruses, including HIV, 19. C. D. Schmid, P. Bucher, PLOS ONE 5, e11425 (2010).
tigate this possibility, we analyzed a STAT1 ChIP- possess IFN-inducible cis-regulatory elements (33). 20. E. S. Lander et al., Nature 409, 860921 (2001).
seq data set of IFNG- and IFN-b (IFNB)stimulated It would be ironic if viral molecular adaptations 21. X. Su et al., Nat. Immunol. 16, 838849 (2015).
22. C. Y. McLean et al., Nat. Biotechnol. 28, 495501
primary macrophages from mouse (31), a spe- had been evolutionarily recycled to fuel innova- (2010).
cies that lacks MER41-like elements but harbors tion and turnover of the host immune repertoire. 23. R. Ostuni et al., Cell 152, 157171 (2013).
a diverse repertoire of lineage-specific ERVs (30). Regardless of how these sequences originated, our 24. V. Hornung et al., Nature 458, 514518 (2009).
Our analysis revealed a muroid-specific endog- study illuminates how selfish genetic elements 25. T. Fernandes-Alnemri et al., Nat. Immunol. 11, 385393
(2010).
enous gammaretrovirus named RLTR30B en- have contributed raw material that has been re-
26. L. Vanhamme et al., Nature 422, 8387 (2003).
riched for both IFNG- and IFNB-inducible purposed for cellular innovation. 27. K. Meyer et al., Sci. Rep. 5, 9012 (2015).
STAT1 binding events (Fig. 4C and fig. S11A), 28. T. Wang et al., J. Leukoc. Biol. 91, 449459 (2012).
which coincide with overlapping motifs cor- 29. M. Lagha, J. P. Bothma, M. Levine, Trends Genet. 28, 409416
RE FERENCES AND NOTES (2012).
responding to both IFNG- and IFNB-induced
1. R. J. Britten, E. H. Davidson, Science 165, 349357 30. W. Bao, K. K. Kojima, O. Kohany, Mob. DNA 6, 11
STAT1 binding sites located in the 5 end of the (1969). (2015).
LTR consensus sequence (Fig. 4D). Reporter 2. B. McClintock, Proc. Natl. Acad. Sci. U.S.A. 36, 344355 31. S.-L. Ng et al., Proc. Natl. Acad. Sci. U.S.A. 108, 2117021175
assays revealed that the consensus sequence (1950). (2011).

1086 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

32. R. E. Randall, S. Goodbourn, J. Gen. Virol. 89, 147 (2008). J. Carleton, and K. Cone for technical assistance, and D. Hancks SUPPLEMENTARY MATERIALS
33. M. Sgarbanti et al., J. Virol. 82, 36323641 (2008). and J. F. McCormick for their critical input. Supported by a www.sciencemag.org/content/351/6277/1083/suppl/DC1
34. R. W. Meredith et al., Science 334, 521524 (2011). Pew Charitable Trusts award and NIH grants GM082545 and Materials and Methods
GM114514 (N.C.E.) and by NIH grants GM112972 and GM059290 Tables S1 to S6
ACKN OW LEDG MEN TS (C.F.). E.B.C. is a Howard Hughes Medical Institute postdoctoral Figs. S1 to S11
Accession numbers for the published data sets analyzed in fellow of the Jane Coffin Childs Fund. N.C.E. is a Pew Scholar References (3549)
this study are available in the supplementary materials. We in the Biomedical Sciences and Mario R. Capecchi Endowed
thank all members of the Elde and Feschotte labs for insightful Chair in Genetics. The authors declare no financial conflicts 30 September 2015; accepted 2 February 2016
discussions. We thank A. Kapusta, A. Lewis, D. Downhour, of interest. 10.1126/science.aad5497

GENE EXPRESSION lished data set describing how individual dele-


tions of 165 chromatin-associated factors affect
the genome-wide expression profile (13). Deleting
Expression homeostasis during a factor that limits transcription from replicated
DNA will increase gene expression in proportion
DNA replication to the time at which the gene is replicated in S
phase, so that early-replicated genes will increase
in expression more than genes replicated late. We
Yoav Voichek,* Raz Bar-Ziv,* Naama Barkai therefore searched for mutants in which gene
expression levels were (on average) negatively cor-
Genome replication introduces a stepwise increase in the DNA template available for related with gene replication timing (Fig. 3A). Of
transcription. Genes replicated early in S phase experience this increase before late- the three mutants showing the strongest correla-
replicating genes, raising the question of how expression levels are affected by DNA tion between gene expression and replication
replication. We show that in budding yeast, messenger RNA (mRNA) synthesis rate is timing, two were involved in H3 acetylation: the

Downloaded from on March 3, 2016


buffered against changes in gene dosage during S phase. This expression homeostasis acetyltransferase Rtt109 and its histone chaperone
depends on acetylation of H3 on its internal K56 site by Rtt109/Asf1. Deleting these cofactor Asf1 (1416). A similar effect was detected
factors, mutating H3K56 or up-regulating its deacetylation, increases gene expression in in expression data from fission yeast deleted of
S phase in proportion to gene replication timing. Therefore, H3K56 acetylation on newly the Asf1 paralogue (fig. S4A) (17, 18). The third
deposited histones reduces transcription efficiency from replicated DNA, complementing candidate, Tos4, is a less-characterized putative
its role in guarding genome stability. Our study provides molecular insight into the transcription factor (19). All three genes increase
mechanism maintaining expression homeostasis during DNA replication. in expression during G1, just before DNA repli-

T
cation (20).
he synthesis of mRNA depends on protein with DNA replication timing (Fig. 1, A and B, The correlation between gene expression and
factors binding to the DNA template. Dur- and figs. S1 and S2). Further, the synthesis rates replication timing were highly significant in all
ing the cell cycle, DNA dosage increases of early-replicating genes increased by only ~20% three mutants, but the difference in expression
at discrete times in S phase, whereas cell relative to late-replicating genes, significantly between early- and late-replicated genes was
volume increases continuously, introduc- less than the ~70% increase in DNA content small. This is expected, as measurements were
ing considerable temporal variations in DNA (Fig. 1A). We also examined cells arrested in the taken in asynchronous cultures in which only a
concentration. How these variations in DNA beginning of S phase after 3-hour treatment minority of cells are in S phase. To amplify the
level affect mRNA synthesis was examined in with HU. Despite the stable increase in DNA difference in mRNA levels between early- and
classical studies (1). In bacteria, mRNA produc- content of early-replicated genes, their expres- late-replicated genes, we profiled gene expres-
tion follows gene dosage, so that the expression sion increased by a mere 5% relative to that of sion in all three mutants synchronized by HU
of each gene increases rapidly after its replica- late nonreplicated genes, suggesting that buf- (Fig. 3B). In both Drtt109 and Dasf1 cells (but
tion (24). By contrast, experiments in eukary- fering under this S phasearrested condition is not Dtos4), expression levels correlated with DNA
otic cells, ranging from yeast to mammals (57), even stronger than in cycling cells (Fig. 1C and content, with genes that were already replicated
indicate a limited dependency of gene expression fig. S8D). Taken together, our results are consist- showing 24 to 28% increase in expression relative
on DNA dosage, prompting the hypothesis that ent with previous studies showing that during to the nonreplicated genes. We next measured
transcription of newly replicated DNA is tran- S phase, DNA dosage has a limited influence on the mRNA levels and synthesis rates in Drtt109
siently repressed (8). mRNA synthesis rates. cells progressing synchronously through S phase.
We extended previous studies, which mea- In contrast to mRNA levels, the binding of Indeed, expression and synthesis rates of early-
sured total mRNA synthesis (9, 10), or focused RNA polymerase II to DNA did correlate with replicated genes increased transiently during mid
on individual genes (11, 12) by directly compar- DNA content in HU-arrested cells and after re- S phase relative to late-replicating genes (Fig. 3C
ing the expression of early- versus late-replicating lease into S phase. Still, the increase in PolII bin- and fig. S5). Therefore, Rtt109 is required for
genes during S phase. If replicated loci produce ding to replicated genes (30%) was lower than buffering mRNA synthesis during DNA repli-
more mRNA than unreplicated ones, then ex- expected by the increase in DNA content (Fig. 2A cation (Fig. 3C, red line).
pression of genes that replicate early should in- and fig. S3). In HU-arrested cells, early-replicated A similar loss of buffering was observed for
crease relative to the expression of late-replicating genes were depleted of elongating PolII (Fig. 2B). Dasf1 (Fig. 3D) and also for Dtos4 cells (fig. S7).
genes during S phase (fig. S1A). In contrast, we However, this difference was specific to HU ar- The latter is particularly notable because Dtos4
find that the relative expression of early- versus rest and disappeared upon release, before the com- did not abrogate buffering in HU-arrested con-
late-replicating genes remained relatively con- pletion of replication. Therefore, reduced PolII ditions, consistent with an additional buffering
stant in budding yeast, progressing synchronously binding to replicated DNA may partially account mechanism acting upon HU arrest. To examine
through S phase after release from a-factor or for the buffering of gene expression, with addi- whether the three candidates act through the same
hydroxyurea (HU) arrest and did not correlate tional differences in elongation capacity that in- pathway, we measured gene expression in the
creases buffering in HU-arrested cells. double deletions Drtt109 Dasf1 and Drtt109Dtos4.
Department of Molecular Genetics, Weizmann Institute of
We hypothesized that chromatin regulators The increased expression of early-replicated genes
Science, Rehovot 76100, Israel.
*These authors contributed equally to this work. Corresponding may suppress transcription from replicated DNA. was similar to the single deletions for the two
author. E-mail: naama.barkai@weizmann.ac.il To identify such factors, we examined a pub- pairs, suggesting that the three genes function

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1087


R ES E A RC H | R E PO R TS

Fig. 1. mRNA synthesis rate is buffered against changes in DNA con-


tent during S phase. (A and B) Expression during S phase. Average increase
in DNA content (black), expression levels (blue), and mRNA synthesis rate
(red) of early-replicating genes relative to late-replicating ones, for cells
progressing synchronously through S phase, after release from a-factor (A) or
HU (B) arrest. Synchronized progression along the cell cycle was verified by
total DNA content profile using fluorescence-activated cell sorting (right
panel) and average expression (relative to G1) of cell-cycle genes (top panel)
(see also table S4). In (A), G1/S transition is defined as time-point 0. (C) Ex-
pression in HU-arrested cells. Genes were ordered by their time of replication,
taken from Yabuki et al. (29). Shown is the log2 relative DNA content (black)
and gene expression (blue) normalized by G1 (a-factor arrest), averaged over a
500-gene sliding window.

Fig. 2. RNA polymerase binding increases on


replicated genes. (A) PolII binding during S phase.
The ratio between PolII binding of early- versus late-
replicating genes (purple) as measured by chromatin
immunoprecipitation sequencing (ChIP-Seq). A line
is presented to guide the eye, omitting the 10-min
time point. DNA content (black) and gene expression
(blue) were sampled in the same experiment and
shown also in Fig. 1B. See fig. S3A for other anti-
bodies. (B) PolII phosphorylation on Ser2 and Ser5.
The log2 ratio of Ser2P (initiating PolII) versus Ser5P
(elongating PolII) abundance on early- (red) and late-
(green) replicated genes in HU-arrested cells and after
their release into S phase as measured by ChIP-Seq.
Error bars represent standard error.

through the same pathway (Fig. 3, D and E, and to alanine, KA) or constant acetylation (lysine tween late- and early-replicated genes. Consist-
figs. S6A and S7). to glutamine, KQ) (22). Mutating K9 did not ently, buffering was also lost upon overproducing
Rtt109 acetylates histone H3 on two residues, affect buffering of early-replicated gene expres- the H3K56ac-specific histone deacetylases, Hst3
K56 and K9, and Asf1 is required for both func- sion in HU-arrested cells (Fig. 4A). In contrast, and Hst4 (23) (Fig. 4B and fig. S8). Therefore,
tions (21). To differentiate which of these resi- the relative expression of early-replicating genes H3K56 acetylation is required for the reduced
dues is responsible for the reduced transcription was significantly higher in cells mutated for K56 gene expression from newly replicated DNA.
efficiency of replicated DNA, we considered mu- (average 34%, compared with a 5% increase in In summary, we find that Rtt109/Asf1-dependent
tants in which K56 or K9 were replaced by resi- wild-type cells). As expected, both modifications H3K56ac suppresses transcription from newly rep-
dues that mimic constant nonacetylation (lysine (KA and KQ) eliminated the asymmetry be- licated DNA during S phase, thereby maintaining

1088 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

Fig. 4. H3K56 acetylation is required for ex-


pression homeostasis during S phase. (A) Ex-
pression in HU-synchronized cells of H3K56 or
H3K9 point mutations. The indicated strains were
HU-arrested for 3 hours. Shown is the increase in
expression of replicated genes (early) relative to
nonreplicated (late) ones for the indicated strains
(top). Correlations of gene expression with the rep-
lication timing are also shown (bottom). (B) Expres-
sion during S phase in a strain overexpressing Hst4:
the average increase in gene expression levels of
early-replicating genes relative to late-replicating
ones for cells overexpressing Hst4, progressing
synchronously through S phase after release from
a-factor arrest (see also table S4). A milder over-
expression of Hst3 led to an intermediate effect
(fig. S8C).

expression homeostasis during this time when the


DNA dosage of different genes transiently differs.
H3K56 is an internal site that is acetylated on
newly synthesized histones before incorporation
onto DNA (15, 24) (fig. S8E). Previous studies as-
Fig. 3. Rtt109 and Asf1 are required for expression homeostasis during S phase. (A) Expression in sociated this modification with active transcription
asynchronous cultures. Each dot represents a single mutant. Correlation between replication timing and of specific genes (25), showing that it promotes
change in gene expression is shown on the y axis, and the difference in the expression of early-and late- nucleosome disassembly (26). H3K56ac, however,
replicating genes is shown on the x axis. The two measures are expected to be correlated. Expression is primarily a marker of replicated DNA during
data from a data set by Lenstra et al. (13). (B) Expression in HU-synchronized cells. Increase in the S phase (27) (fig. S11), when it promotes nucleo-
expression of replicated genes (early) relative to nonreplicated (late) ones for the indicated strains (top), some assembly and guards genome stability
in cells arrested with HU for 3 hours. Correlations of gene expression with replication timing is also (14, 16, 28). Our study ascribes a complemen-
shown (bottom). (C) Expression during S phase in Drtt109. Same as in Fig. 1A for Drtt109 after a-factor tary role to H3K56ac in maintaining expression
synchronization. Average DNA content from Drtt109 and wild-type (WT) are plotted along the time homeostasis during S phase.
course (top right) (DNA content from WT same as in Fig. 1A). (D and E) Expression during S phase in
mutants. The average increase in expression levels (blue) of early-replicating genes relative to late-
REFERENCES AND NOTES
replicating ones is depicted for the indicated mutants progressing through S phase after release from
1. S. Marguerat, J. Bhler, Trends Genet. 28, 560565 (2012).
a-factor arrest. The maximal increase in the expression of early- versus late-replicated genes is 2. M. G. Chandler, R. H. Pritchard, Mol. Gen. Genet. 138, 127141
summarized in (E) over 5 and 4 repeats for WT and Drtt109, respectively (see also figs. S7B and S8D). (1975).

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1089


R ES E A RC H | R E PO R TS

3. M. B. Schmid, J. R. Roth, J. Bacteriol. 169, 28722875 17. C. Heichinger, C. J. Penkett, J. Bhler, P. Nurse, EMBO J. 25, AC KNOWLED GME NTS
(1987). 51715179 (2006). We thank D. Rosin and M. Krupkin for technical assistance and
4. C. Sousa, V. de Lorenzo, A. Cebolla, Microbiology 143, 18. K. Yamane et al., Mol. Cell 41, 5666 (2011). J. Karagiannis for yeast strains. We also thank members of our
20712078 (1997). 19. C. E. Horak et al., Genes Dev. 16, 30173033 (2002). laboratory and B. Z. Shilo for fruitful discussions and comments on
5. S. G. Elliott, C. S. McLaughlin, Proc. Natl. Acad. Sci. U.S.A. 75, 20. P. T. Spellman et al., Mol. Biol. Cell 9, 32733297 the manuscript. This work was supported by the European
43844388 (1978). (1998). Research Council, Israel Science Foundation, and Binational
6. D. Killander, A. Zetterberg, Exp. Cell Res. 38, 272284 (1965). 21. E. Radovani et al., Eukaryot. Cell 12, 654664 (2013). Science Foundation. RNA-seq, ChIP-seq, and DNA-seq data are
7. S. E. Pfeiffer, J. Cell. Physiol. 71, 95104 (1968). 22. J. Dai et al., Cell 134, 10661078 (2008). deposited in the European Nucleotide Archive (ENA) under
8. S. G. Elliott, Mol. Gen. Genet. 192, 204211 (1983). 23. I. Celic et al., Curr. Biol. 16, 12801289 (2006). accession number PRJEB11977.
9. R. S. Fraser, F. Moreno, J. Cell Sci. 21, 497521 (1976). 24. C. A. Davey, D. F. Sargent, K. Luger, A. W. Maeder,
10. R. S. Fraser, P. Nurse, Nature 271, 726730 (1978). T. J. Richmond, J. Mol. Biol. 319, 10971113 (2002).
11. S. Yunger, L. Rosenfeld, Y. Garini, Y. Shav-Tal, Nat. Methods 7, 25. F. Xu, K. Zhang, M. Grunstein, Cell 121, 375385 (2005). SUPPLEMENTARY MATERIALS
631633 (2010). 26. S. K. Williams, D. Truong, J. K. Tyler, Proc. Natl. Acad. Sci. www.sciencemag.org/content/351/6277/1087/suppl/DC1
12. O. Padovan-Merhar et al., Mol. Cell 58, 339352 (2015). U.S.A. 105, 90009005 (2008). Materials and Methods
13. T. L. Lenstra et al., Mol. Cell 42, 536549 (2011). 27. H. Masumoto, D. Hawke, R. Kobayashi, A. Verreault, Nature Figs. S1 to S11
14. R. Driscoll, A. Hudson, S. P. Jackson, Science 315, 649652 436, 294298 (2005). Tables S1 to S4
(2007). 28. M. Clemente-Ruiz, R. Gonzlez-Prieto, F. Prado, PLOS Genet. 7, References (3061)
15. J. Han, H. Zhou, Z. Li, R.-M. Xu, Z. Zhang, J. Biol. Chem. 282, e1002376 (2011).
1415814164 (2007). 29. N. Yabuki, H. Terashima, K. Kitada, Genes Cells 7, 781789 28 July 2015; accepted 19 January 2016
16. Q. Li et al., Cell 134, 244255 (2008). (2002). 10.1126/science.aad1162

TRANSCRIPTION terminants for TSS selection, using a library


containing all 47 (~16,000) sequences at posi-
tions 4 to 10 base pairs (bp) downstream of the
Multiplexed protein-DNA 10 element of a consensus bacterial promoter
(MASTER-N7; Fig. 1A and fig. S1) (2). Here, to

Downloaded from on March 3, 2016


cross-linking: Scrunching in define effects on TSS selection of sequences out-
side the TSS region, we analyzed a template li-
brary containing all 410 (~1,000,000) sequences
transcription start site selection at positions 1 to 10 bp downstream of the 10
element, extending the randomized sequence
Jared T. Winkelman,1,2,3,4* Irina O. Vvedenskaya,1,3* Yuanchao Zhang,1,5* to include the discriminator (710), located
Yu Zhang,2,3* Jeremy G. Bird,1,2,3 Deanne M. Taylor,1,5,6,7 Richard L. Gourse,4 between the TSS region and the 10 element
Richard H. Ebright,2,3 Bryce E. Nickels1,3
(MASTER-N10; Fig. 1A and fig. S1). Results of
MASTER-N10 analysis reveal that the discrimi-
In bacterial transcription initiation, RNA polymerase (RNAP) selects a transcription start
nator affects TSS selection (Fig. 1, B to D; figs. S2
site (TSS) at variable distances downstream of core promoter elements. Using next- to S4; and table S1). Changes in the discriminator
generation sequencing and unnatural amino acidmediated protein-DNA cross-linking, we change TSS selection by up to ~3 bp (Fig. 1D and
have determined, for a library of 410 promoter sequences, the TSS, the RNAP leading-edge
fig. S4) and change the mean TSS, averaged over
position, and the RNAP trailing-edge position. We find that a promoter element upstream the ~16,000 templates analyzed for each of the
of the TSS, the discriminator, participates in TSS selection, and that, as the TSS changes, 64 discriminator sequences, by ~1 bp (Fig. 1B).
the RNAP leading-edge position changes, but the RNAP trailing-edge position does not Discriminators having a purine at each position
change. Changes in the RNAP leading-edge position, but not the RNAP trailing-edge
(RRR), particularly GGG, favor TSS selection at
position, are a defining hallmark of the DNA scrunching that occurs concurrent with RNA upstream-shifted positions, whereas discrimina-
synthesis in initial transcription. We propose that TSS selection involves DNA scrunching tors having a pyrimidine at each position (YYY),
prior to RNA synthesis. particularly CCT, favor TSS selection at downstream-
shifted positions (modal TSS for RRR, 7 bp down-
stream of 10 element; modal TSS for YYY, 8 bp

D
uring bacterial transcription initiation, RNA scription start site (TSS), and aligns the TSS template- downstream of 10 element; Fig. 1, B and C, and
polymerase (RNAP) holoenzyme binds to strand nucleotide with the RNAP active center fig. S2B). Results from MASTER-N10, where the
promoter DNA through sequence-specific (1). There is variability in the position of the TSS discriminator is GTG, match results from MASTER-
interactions with core promoter elements, relative to core promoter elements (26). The N7, where the discriminator is GTG, demonstrat-
unwinds a turn of promoter DNA to form mechanistic basis for this variability has remained ing the reproducibility of the approach (Fig. 1, C
an RNAP-promoter open complex (RPo) with an unclear. In addition, although DNA-sequence and D, and fig. S2B). We conclude that the dis-
unwound transcription bubble, selects a tran- determinants for TSS selection within the TSS criminator is a determinant of TSS selection.
region have been defined (2), it has remained A conserved region of transcription initiation fac-
1
Department of Genetics, Rutgers University, Piscataway, NJ unclear whether there also are DNA-sequence tor s, s region 1.2 (s1.2), makes sequence-specific
08854, USA. 2Department of Chemistry and Chemical determinants for TSS selection outside the TSS protein-DNA interactions with the nontemplate
Biology, Rutgers University, Piscataway, NJ 08854, USA.
3
Waksman Institute, Rutgers University, Piscataway, NJ 08854,
region. strand of the discriminator in the transcription
USA. 4Department of Bacteriology, University of Wisconsin- To investigate whether there are sequence de- bubble in RPO (7, 8). These interactions confer
Madison, Madison, WI 53705, USA. 5Department of Biomedical terminants for TSS selection outside the TSS re- specificity for GGG (79). To determine whether
and Health Informatics, Childrens Hospital of Philadelphia, gion, we applied a next-generation-sequencing sequence-specific s1.2-discriminator interactions
Philadelphia, PA 19104, USA. 6Department of Pediatrics,
Perelman School of Medicine, University of Pennsylvania,
approach that enables comprehensive analysis of affect TSS selection, we used MASTER-N10 to
Philadelphia, PA, 19104, USA. 7Department of Obstetrics, sequence determinants during transcription: mas- compare wild-type s to a s derivative having ala-
Gynecology and Reproductive Sciences, Rutgers Robert Wood sively systematic transcript end readout (MASTER) nine substitutions that disrupt sequence-specific
Johnson Medical School, New Brunswick, NJ 08901, USA. (2). MASTER entails generating transcripts from discriminator-s1.2 interactions: s1.2-mut (7, 11).
*These authors contributed equally to this work. Present address:
Chinese Academy of Sciences, Shanghai, 200032, China.
a library of bar-coded randomized sequences and The results show that disrupting s1.2-discriminator
Corresponding author. E-mail: bnickels@waksman.rutgers.edu sequencing transcript ends (fig. S1) (2). In pre- interactions markedly alters TSS selection for tem-
(B.E.N.); ebright@waksman.rutgers.edu (R.H.E.) vious work, we defined TSS-region sequence de- plates containing a GGG discriminator, resulting

1090 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

in a downstream shift in mean TSS (Fig. 1E). We with position 7 generally being preferredand It has been hypothesized that TSS selection at
conclude that s1.2-discriminator interactions are that discriminator sequence affects TSS selection. positions downstream of the modal TSS (gen-
a determinant of TSS selection. These results imply that RPo can accommodate erally position 7; Fig. 1C and fig. S2A) involves
The results in Fig. 1 and figs. S2 to S4 show that ~17 (5 bp 3.4 /bp) variation in the position transcription-bubble expansion (scrunching; fig.
TSS selection can occur at any of five positions down- of the TSS and that discriminator sequence af- S5), and TSS selection at positions upstream of
stream of the 10 elementi.e., positions 6 to 10, fects the ability to accommodate this variation. the modal TSS involves transcription-bubble con-
traction (anti-scrunching; fig. S5) (2, 12). Accord-
ing to this hypothesis, RPo generally contains a
13-bp transcription bubble that places position
7 in the RNAP active-center initiating nucleoside
triphosphate (NTP) site (i site) and position 8 in
MASTER-N7 the RNAP active-center NTP addition site (i+1
site; fig. S5, TSS = 7). For TSS selection to occur
at positions 8, 9, or 10, it is hypothesized that the
MASTER-N10 downstream DNA duplex is unwound by 1, 2, or
3 bp; the unwound DNA is pulled into and past
the RNAP active center, and the unwound DNA
mean modal is accommodated as bulges within the transcription
discriminator
TSS TSS bubble, yielding a scrunched complex (fig. S5, TSS =
7.31 7 8, 9, or 10). For TSS selection to occur at position 6,
7.56 7 it is hypothesized that the opposite occurs: Down-
7.57 7 stream DNA is rewound by 1 bp, downstream DNA
7.94 8 is extruded from the RNAP active center, and the
extrusion of DNA is accommodated by stretching
7.40 7 DNA within the transcription bubble, yielding
7.65 7 an anti-scrunched complex (fig. S5, TSS = 6). Two
7.80 8 lines of evidence support this hypothesis: Single-
molecule fluorescence resonance energy transfer
results suggest that transcription-bubble size in
10
RPo can vary (12); and negative supercoiling,
which provides a driving force for transcription-
mean TSS MASTER-N7

9 bubble expansion, favors TSS selection at down-


discriminator sequence

stream positions (2). However, direct evidence


(with GTG)

8
for this hypothesis has not been reported.
7 Transcription-bubble expansion (scrunching)
occurs in initial transcription, where it is coupled
6
to RNA synthesis (1315) (fig. S5). A hallmark of
5 scrunching during initial transcription is that the
RNAP trailing edge remains stationary relative
4
to DNA, whereas the RNAP leading edge moves
4 5 6 7 8 9 10 relative to DNA (1315) (fig. S5). Here, we inves-
mean TSS MASTER-N10 tigated whether this hallmark of scrunching is a
(subsets with GGG,GTG,CCT)
feature of TSS selection. We used unnatural amino
acidmediated protein-DNA photocross-linking
8
to define the RNAP trailing-edge position and
RNAP leading-edge position in RPo on a MASTER-
N10 library (MASTER-N10-XL; Fig. 2A and fig. S6).
RNAP- wt
mean TSS

To perform MASTER-N10-XL, we incorporated the


photoactivatible unnatural amino acid p-benzoyl-
L-phenylalanine (Bpa; fig. S6) (15, 16) at specific
GGG sites at the RNAP trailing edge and RNAP lead-
ing edge; formed RPO between the Bpa-containing
7 RNAP derivatives and the MASTER-N10 library;
and irradiated complexes with ultraviolet light to
7 8 7 8
mean TSS mean TSS induce cross-linking between Bpa and adjacent
(bp downstream of -10 element) RNAP-1.2 mut
DNA nucleotides (Fig. 2A and fig. S6). We then
Fig. 1. Sequences upstream of TSS region affect TSS selection. (A) Promoter sequences mapped positions of cross-links on each of the 410
analyzed in MASTER-N7 (2) and MASTER-N10. Promoter 35, 10, and discriminator elements are sequences by primer extension and high-throughput
indicated. Green, randomized nucleotides. (B) Effect of discriminator on position of TSS (numbered sequencing (Fig. 2A). The method defines the
in base pairs downstream of 10 element). Data show means and 99.9% confidence intervals for mean RNAP trailing-edge position, the mean RNAP
each of the 64 discriminator sequences (~16,000 templates analyzed for each discriminator). Green, leading-edge position, and the mean distance be-
GGG and other RRR discriminators; blue, CCT and other YYY discriminators; red, GTG discriminator. tween them (RNAP trailing-edge/leading-edge dis-
(C) Mean and modal TSS. *, GTG data from MASTER-N7; **, GTG data from MASTER-N10. (D) Upstream tance; Fig. 2 and table S2).
and downstream shifts in TSS selection with the ~16,000 GGG and ~16,000 CCT discriminators (green The results show that changes in the discrim-
and blue, respectively) relative to the ~16,000 GTG discriminators (red). (E) Effect of s1.2-discriminator inator change the RNAP trailing-edge/leading-edge
interactions on TSS selection (downstream shift in mean TSS for ~16,000 GGG-discriminator templates distance in the same manner that changes in the
on replacement of s by s1.2 mutant). discriminator change TSS selection (Figs. 1 and 2

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1091


R ES E A RC H | R E PO R TS

trailing edge leading edge discriminator trailing-edge/leading-edge


distance (mean, bp)
Bpa N15 Bpa N15
barcode barcode 25.1
25.4
25.9
UV-irradiate UV-irradiate
25.2
25.4
Bpa N15 Bpa N15
barcode barcode 25.7

denature denature
400
primer extension primer extension

number of sequences
discriminator sequence
sequencing sequencing
# reads

mean trailing-edge/ leading-edge distance

discriminator 0
23 24 25 26 27 28
trailing-edge/leading-edge
distance (mean, bp)
Fig. 2. Sequences upstream of TSS region affect RNAP trailing-edge/leading-edge
distance. (A) MASTER-N10-XL. (B) Effect of discriminator on RNAP trailing-edge/leading-
edge distance in RPo (symbols as in Fig. 1). (C) RNAP trailing-edge/leading-edge distances. 26

trailing-edge/leading-edge
(D) Decreases and increases in RNAP trailing-edge/leading-edge distance with the ~16,000

distance (mean, bp)


GGG- and ~16,000 CCT-discriminator templates (green and blue, respectively) relative

RNAP- wt
to the ~16,000 GTG-discriminator templates (red). Dashed lines indicate mean trailing-edge/
leading-edge distances. (E) Effect of s1.2-discriminator interactions on RNAP trailing-edge/
leading-edge distance (increase in RNAP trailing-edge/leading-edge distance for ~16,000
GGG-discriminator templates on replacement of s by s1.2 mutant).
25 GGG

25 26 25 26
RNAP-1.2 mut
trailing-edge/leading-edge trailing-edge/leading-edge
distance (mean, bp) distance (mean, bp)

Fig. 3. As TSS changes, RNAP leading-edge po-


sition changes, but RNAP trailing-edge posi-
tion does not change. (A) RNAP trailing-edge
position (left; slope ~0) and RNAP leading-edge
position (right; slope ~1) as a function of mean
TSS for each of the 64 discriminator sequences
(~16,000 templates analyzed for each discrimina-
tor; green, RRR discriminators; blue, YYY discrim-
inators). (B) Interpretation: Changes in TSS selection
result from changes in DNA scrunching. Gray, RNAP;
yellow, s; blue, 10 element nucleotides; purple,
discriminator nucleotides; i and i+1, NTP binding
sites; red, Bpa and nucleotide cross-linked to Bpa;
boxes, DNA nucleotides (nontemplate-strand nucleo-
tides above template-strand nucleotides; nucleotides
downstream of 10 element numbered). Scrunching
is indicated by bulged-out nucleotides. Anti-scrunching
is indicated by a stretched nucleotide-nucleotide
linkage.

1092 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

with the RNAP active center, facilitating down-


stream TSS selection; and transcription-bubble
contraction (anti-scrunching) places upstream DNA
in contact with the RNAP active center, facilitat-
ing upstream TSS selection. According to this
model, the discriminator alters TSS selection
by altering the energy landscape describing
the ensemble of transcription-bubble sizes in
RPo (Fig. 4). The scrunching that occurs in TSS
selection is mechanistically analogous to the
scrunching that occurs during initial transcrip-
tion (1315) (fig. S5). However, scrunching in TSS
selection occurs before NTP binding and nucle-
otide incorporation and, in the absence of an
additional energy source, is driven by energy
available from the thermal bath and therefore
Fig. 4. Crystal structures define paths of DNA nontemplate strands with representative RRR and YYY limited to ~1 to 3 bprather than occurring
discriminators. (A) Crystal structure of an initiation complex with RRR discriminator and nontemplate- after NTP binding and nucleotide incorpora-
strand length corresponding to TSS at position 7 (RPo-GGG-7). Left, simulated annealing |Fo Fc| omit tion, being driven by a combination of thermal
map contoured at 2.3s and atomic model for interactions of RNAP with DNA nontemplate strand. Right, energy, NTP binding, and nucleotide incorpo-
schematic path of DNA. Gray, RNAP; yellow, s; purple, discriminator nucleotides; pink, nontemplate-strand ration, and being able to span tens of base
nucleotides downstream of discriminator; green, omit map. Red box, third nucleotide of discriminator. pairs. We suggest that, in the presence of an
(B) Crystal structure of an initiation complex with YYYdiscriminator and nontemplate-strand length corres- additional energy source, scrunching in TSS
ponding to TSS at position 8 [RPo-CCC-8; symbols as in (A)]. Nucleic-acid scaffolds used for crystal- selection could access a larger range of TSS
lization included short oligoribonucleotides (see Methods). (C and D) Interpretation: The different position positions, and, in particular, we speculate that
of the third nucleotide of the discriminator in the structure with a YYYdiscriminator increases the distance this occurs with transcription factor IIH (TFIIH)
between the third nucleotide of the discriminator and downstream duplex DNA, accommodating an dependent adenosine 5-triphosphate (ATP) hy-
additional nucleotide in the connector. drolysis as the additional energy source in the
long-range TSS scanning observed with eu-
karyotic RNAP II in some species (17).
and fig. S7). Changes in the discriminator change strand length corresponding to its modal TSS at
the RNAP trailing-edge/leading-edge distance position 8, we determined crystal structures of
by up to ~3 bp (fig. S8) and change the RNAP representative initiation complexes of each: RPO- REFERENCES AND NOTES

trailing-edge/leading-edge distance, averaged over GGG-7 and RPO-CCC-8 (Fig. 4, figs. S10 and S11, 1. E. F. Ruff, M. T. Record Jr., I. Artsimovitch, Biomolecules 5,
10351062 (2015).
the ~16,000 templates analyzed for each of the and table S3). For RPo-GGG-7, the first nucleotide 2. I. O. Vvedenskaya et al., Mol. Cell 60, 953965 (2015).
64 discriminator sequences, by ~1 bp (Fig. 2B and of the discriminator is unstacked and inserted 3. W. Jeong, C. Kang, Nucleic Acids Res. 22, 46674672
fig. S7). The RNAP trailing-edge/leading-edge into a pocket in s1.2, and the next six nucleotides (1994).
distance is shortest for RRR, especially GGG, dis- of the nontemplate-strand form a continuous stack 4. J. Liu, C. L. Turnbough Jr., J. Bacteriol. 176, 29382945
(1994).
criminators and longest for YYY, especially CCT, (Fig. 4A and fig. S10) (8). In contrast, for RPo- 5. K. A. Walker, R. Osuna, J. Bacteriol. 184, 47834791
discriminators (Fig. 2, B to D). Disruption of s1.2- CCC-8, the third nucleotide of the discriminator (2002).
discriminator interactions results in a marked in- occupies a different position, being unstacked, ro- 6. D. E. Lewis, S. Adhya, Mol. Microbiol. 54, 692701
crease in RNAP trailing-edge/leading-edge distance tated by ~180, and inserted into a pocket in s (2004).
7. S. P. Haugen et al., Cell 125, 10691082 (2006).
for templates containing a GGG discriminator region 2 (s2; Fig. 4B and figs. S10 and S11). As a 8. Y. Zhang et al., Science 338, 10761080 (2012).
(Fig. 2E). result of the difference in the position of the third 9. A. Feklistov et al., Mol. Cell 23, 97107 (2006).
We next compared effects of discriminator nucleotide of the discriminator, the distance be- 10. A. A. Travers, J. Bacteriol. 141, 973976 (1980).
sequence on TSS selection (Fig. 1) to effects of tween the third nucleotide of the discriminator 11. S. P. Haugen, W. Ross, M. Manrique, R. L. Gourse, Proc. Natl.
Acad. Sci. U.S.A. 105, 32923297 (2008).
discriminator sequence on RNAP trailing-edge and the downstream duplex is ~4-5 greater in 12. N. C. Robb et al., J. Mol. Biol. 425, 875885 (2013).
position and RNAP leading-edge position (Figs. 2 RPo-CCC-8 than in RPo-GGG-7, and the DNA seg- 13. A. Revyakin, C. Liu, R. H. Ebright, T. R. Strick, Science 314,
and 3 and fig. S9). The results show that, as the ment between the third nucleotide of the discrim- 11391143 (2006).
position of the TSS changes by 1 bp, the RNAP inator and the downstream duplex accommodates 14. A. N. Kapanidis et al., Science 314, 11441147
(2006).
leading-edge position changes by 1 bp (Fig. 3A an additional nucleotide (Fig. 4B and fig. S10). 15. J. T. Winkelman et al., Mol. Cell 59, 768780 (2015).
and fig. S9, right), but the RNAP trailing-edge po- Two factors account for the difference in the po- 16. J. W. Chin, A. B. Martin, D. S. King, L. Wang, P. G. Schultz,
sition does not change (Fig. 3A and fig. S9, left). sition of the third nucleotide of the discriminator Proc. Natl. Acad. Sci. U.S.A. 99, 1102011024 (2002).
Thus, TSS selection exhibits a defining hallmark in the RPO-CCC-8: (i) weak stacking between YY, 17. F. M. Fazal, C. A. Meng, K. Murakami, R. D. Kornberg,
S. M. Block, Nature 525, 274277 (2015).
of scrunching: namely, the RNAP leading edge as compared to RR, at the second and third nu-
moves, whereas the RNAP trailing edge does not cleotides of the discriminator (Fig. 4B and fig. AC KNOWLED GME NTS
move (Fig. 3B). The results provide support for S10); and (ii) the ability of Y at the third nucleo- Work was supported by NIH grants GM37048 (R.L.G.),
the hypothesis that flexibility in TSS selection is tide of the discriminator to fit in a pocket in s2 GM041376 (R.H.E.), GM088343 (B.E.N.), and GM115910 (BEN).
mediated by scrunching/anti-scrunching, and that that has size and hydrogen-bonding character Protein Data Bank accession codes are 5E17 and 5E18.
the discriminator affects TSS selection by modu- specific for Y (Fig. 4B and fig. S11).
lating the extent of scrunching/anti-scrunching. Our results indicate that flexibility in TSS se- SUPPLEMENTARY MATERIALS

To define the path of the DNA nontemplate- lection occurs through changes in scrunching/ www.sciencemag.org/content/351/6277/1090/suppl/DC1
Materials and Methods
strand in RPO having an RRR discriminator and anti-scrunching in RPo. We propose that RPo Figs. S1 to S11
a nontemplate-strand length corresponding to uses thermally driven DNA fluctuations to access Tables S1 to S4
its modal TSS at position 7, and to define the an ensemble of transcription-bubble sizes (Fig. References (1823)
path of the DNA nontemplate-strand in RPO 3 and fig. S5). Transcription-bubble expansion 21 October 2015; accepted 25 January 2016
having a YYY discriminator and a nontemplate- (scrunching) places downstream DNA in contact 10.1126/science.aad6881

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1093


R ES E A RC H | R E PO R TS

DNA REPAIR the quantitative agreement between the percen-


tages of cells with a delayed response and with
zero Ada molecules. Consequently, the delay before
Stochastic activation of a DNA response activation should match the time until
the first random expression event occurs in these

damage response causes cell-to-cell cells. Indeed, the distribution of Ada copies before
damage was very close to a Poisson distribution
(Fig. 1G) with an average production rate of
mutation rate variation one molecule per cell cycle (fig. S6), and the late-
responding cells also activated the response with
Stephan Uphoff,1,2* Nathan D. Lord,2 Burak Okumus,2 Laurent Potvin-Trottier,2,3
a Poisson rate of once per cell cycle (Fig. 1F).
These findings also mean that most cells re-
David J. Sherratt,1 Johan Paulsson2*
liably launch the response with just one or two
Ada molecules to sense the damage and to induce
Cells rely on the precise action of proteins that detect and repair DNA damage. However, gene
ada expression (Fig. 2A). We indeed observed dis-
expression noise causes fluctuations in protein abundances that may compromise repair. For
tinct single-molecule signatures: The rates of Ada
the Ada protein in Escherichia coli, which induces its own expression upon repairing DNA
production displayed staircase patterns with equi-
alkylation damage, we found that undamaged cells on average produce one Ada molecule per
distant states during response activation and de-
generation. Because production is stochastic, many cells have no Ada molecules and cannot
activation at low MMS concentrations (Fig. 2B and
induce the damage response until the first expression event occurs, which sometimes delays
fig. S7), indicative of discrete production and loss
the response for generations. This creates a subpopulation of cells with increased mutation
events of the meAda molecules that control
rates. Nongenetic variation in protein abundances thus leads to genetic heterogeneity in the
Ada expression. To further confirm the low num-
population. Our results further suggest that cells balance reliable repair against toxic side
bers, we titrated meAda using promoter sites on
effects of abundant DNA repair proteins.
a low copy-number plasmid (15), which markedly

Downloaded from on March 3, 2016


T
decreased steady-state Ada induction, as expected
he integrity of the genome is constantly 20% of the cells did not respond at all, even at (fig. S8). Furthermore, the discrete production
threatened by DNA damage. Most damage saturating doses of MMS (Fig. 1, B and C). Quan- rate steps disappeared when meAda abundance
events are reversed by active repair systems, titatively similar results were obtained with a was increased using high MMS concentrations
but the ones that escape repair can cause transcriptional fluorescent reporter in cells with (fig. S7).
cell death or mutations. An intriguing ques- untagged Ada (fig. S2), which showed that the Because failure to trigger the adaptive response
tion is what causes those failures. Specifically, protein fusion did not affect the observations. seems to be the result of a complete lack of Ada
the classic perspective suggests that failures to To visualize the dynamics of the process, we molecules in a fraction of cells, it should be possible
repair reflect the intrinsic error rate of the repair monitored Ada-mYPet abundance in real time in to reduce this fraction with a slight increase in
enzymes, for example, because of the random a microfluidic device that allows imaging of single the average abundance of Ada. Specifically, for
search for lesions (1, 2). Alternatively, most failures cells over tens of generations during constant DNA many distributions (including the Poisson) the
could occur in an error-prone subpopulation of damage treatment (fig. S3 and movies S1 and S2) probability mass in the tails depends sensitive-
cells (3, 4) in which repair is compromised by (12, 13). At low-to-intermediate MMS concentra- ly on the average. We therefore moderately in-
fluctuations in the abundances of the repair tions (<200 mM MMS), cells showed random un- creased Ada numbers per cell either by inhibiting
proteins (57). synchronized pulses of Ada expression (Fig. 1D). cell divisionkeeping the concentrations con-
To distinguish between these possibilities we The pulse frequency increased proportionally to stant (16)or by expressing additional unlabeled
quantitatively analyzed, with single-molecule res- the MMS concentration (fig. S4), as expected Ada from the PAda promoter on a very low copy-
olution in single cells, the adaptive response that when triggering is limited by the probability that number plasmid (MiniF; ~2 copies per cell). In
protects Escherichia coli against the toxic and Ada finds a lesion. At higher MMS concentra- both cases, we observed the predicted uniform
mutagenic effects of DNA alkylation damage tions, most cells rapidly induced a persistent MMS response and disappearance of the late-
(8). The Ada protein functions not only in the and uniform response (Fig. 1E). However, 20 to responding cell subpopulation (Fig. 2, C and D).
direct repair of alkylated DNA but also as the 30% of cells were lagging even at saturating MMS These observations raise the question of why
transcriptional activator of the adaptive response and triggered the response after exponentially the native ada gene is expressed at such low basal
(Fig. 1A) (9, 10). Specifically, ada expression is distributed delays with an average of one gen- amounts. Following the fates of single cells over
induced by methylated Ada (meAda) after ir- eration time (Fig. 1F and fig. S5). Some cells thus time showed that a failure to activate the adaptive
reversible methyl transfer from DNA phospho- failed to respond for several generations. response during MMS treatment lowered the via-
triester and O6MeG lesions onto cysteine residues To identify the molecular determinants of this bility of those cells, as expected (Fig. 3A and fig. S9).
of Ada. Because Ada is present in low numbers heterogeneity, we measured the Ada abundance However, the moderate overexpression of Ada re-
before damage, this positive-feedback gene reg- before MMS treatment. Ada-mYPet was undetec- sulted in severe toxicity of MMS treatment (Fig. 3A
ulation may amplify stochastic fluctuations and table over the autofluorescence background of and fig. S10) (14, 17) and caused spontaneous
create cell-to-cell heterogeneity in the repair sys- cells, which suggested that absolute amounts were triggering of the response in the absence of MMS
tem (2, 11). on the order of a few molecules per cell. We (Fig. 3B and fig. S10), something we never observed
We imaged the endogenous expression of a therefore turned to single-molecule microscopy at native ada expression (Fig. 3B and fig. S3). The
functional Ada-mYPet fluorescent protein fusion to directly count individual proteins in live cells extremely low abundance of Ada can thus be ad-
(fig. S1), in cells treated with methyl methane- (Fig. 1G and fig. S6). The abundance of Ada was vantageous to the population as a whole, which
sulfonate (MMS) (Fig. 1B). We observed a strong extremely low: The observed population average implies that the repair system faces a trade-off to
and uniform expression of Ada in most cells, but was 1.4 0.1 molecules per cell (SEM) and 20 to repair exogenous alkylation damage without in-
30% of the cells did not contain a single Ada mol- troducing harmful effects. In fact, given the low
1
Department of Biochemistry, University of Oxford, Oxford ecule. Because the ada gene is strictly autoregula- numbers of molecules, the ada regulation is re-
OX1 3QU, UK. 2Department of Systems Biology, Harvard tory, i.e., it can only be induced by the Ada protein markably precise: First, the Poisson distribution
Medical School, Boston, MA 02115, USA. 3Biophysics Ph.D.
Program, Harvard Medical School, USA.
(810, 14), cells with zero Ada molecules should be before damage shows an almost complete absence
*Corresponding author. E-mail: stephan.uphoff@bioch.ox.ac.uk unable to trigger the adaptive response, despite of gene expression bursts or extrinsic noise (Fig.
(S.U.); johan_paulsson@harvard.edu (J.P.) great amounts of damage. This is supported by 3C and fig. S6), in stark contrast to the regulation

1094 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

of most genes studied (5, 6, 7, 18). This can be control response deactivation after the damage has while also imaging Ada-mYPet in the same live cells
explained by a short half-life and inefficient trans- been repaired (21). Indeed, removal of MMS caused (Fig. 4). Without MMS treatment, the apparent
lation of ada mRNAs (19, 20), as well as the all cells to switch off the adaptive response uni- mutation frequency was low (fig. S13), and most
tendency of Poisson noise to dominate at very formly, and Ada was diluted because of cell growth MutS molecules were mobile (average 6% bound)
low abundances. Second, a dual reporter assay (5) (Fig. 3F). (fig. S14). MMS treatment of Ada-deficient cells
that simultaneously monitors expression of the The total number of Ada molecules directly de- (Dada) increased both the mutation frequency (fig.
endogenous PAda ada-mYPet and an ectopic PAda termines a cells repair capacity: each Ada mol- S13) and MutS binding (56% bound) (Fig. 4 and
cfp insertion (Fig. 3, D and E, and fig. S11) showed ecule can only act once to remove one mutagenic fig. S14). MMS treatment of wild-type cells resulted
that both the activation time after MMS treatment O6MeG lesion (10). Furthermore, a lack of Ada in highly variable amounts of bound MutS mol-
and the subsequent expression dynamics were repair capacity cannot be compensated for by the ecules between cells. This variation could be entirely
closely correlated between the two genes, with DNA mismatch repair pathway, because unre- explained by the heterogeneity in Ada expression
little uncorrelated noise that would indicate tran- paired O6MeG lesions miscode for T instead of (Fig. 4 and fig. S15): MutS binding was increased
scriptional bursting. Considering the central role C. This leads to futile mismatch repair cycles, only in the subpopulation of cells with low Ada ex-
of meAda in ada regulation (810, 14), these expres- which eventually cause stable mutations during pression (30% bound), whereas cells with abundant
sion dynamics likely reflect fluctuations in meAda the next round of replication (17). We therefore Ada retained low MutS activity (10% bound). Sto-
numbers. Indeed, the normalized standard devia- tested whether heterogeneity in Ada concentra- chastic activation of the adaptive response therefore
tion was inversely proportional to the square root tions affects mutation rates. To directly measure leads to an error-prone cell subpopulation that
of the expected average number of DNA damage genomic mutation rates in single cells, we used does not efficiently repair DNA alkylation dam-
sites, quantitatively consistent with the simplest the DNA mismatch recognition protein MutS as age and accumulates mutations.
model, where varying damage levels determine a marker for labeling nascent mutations (22). Spe- We found that a cells fate after DNA damage
meAda abundances that then reliably control ada cifically, photoactivated single-molecule tracking (23) can be accurately predicted by the presence or
expression (fig. S12). Third, ada transcription acti- allowed us to classify individual MutS-PAmCherry absence of a single protein molecule. The re-
vation is inhibited by unmethylated Ada. This may fusion proteins as DNA-bound or mobile (24, 25), sulting cell heterogeneity increases the chance

Fig. 1. Stochastic gene expression delays Ada response activation in a log(1-CDF) of response delay times for the last 30% of cells to activate Ada-
cell subpopulation. (A) Methylation of Ada N- and C-terminal domains func- mYPet expression after MMS treatment in the microfluidic chip. Different MMS
tions as a damage sensor, turning Ada into an autoregulatory activator of concentrations in colors as in main plot. Straight lines on log scale reflect ex-
genes involved in DNA alkylation repair. (B) Ada-mYPet fluorescence (yellow) ponential distributions as generated by a Poisson process; the slope cor-
in cells treated with 10 mM MMS for 1 hour. Constitutive mKate2 serves as responds to the average-delay time constant. Gray area: Poisson process with
fluorescent cell marker (gray). Scale bar, 5 mm. (C) Percentage of cells that a rate of 1 0.1 per generation. Main plot: Average delay time constants from
activated Ada-mYPet expression after 1 hour in MMS. (Inset) Histogram of the inset data (SEM). (G) Single-molecule counting of Ada-mYPet without
Ada-mYPet fluorescence per cell with 10 mM MMS. (D and E) Time traces of MMS. Example cell shown. Poisson model was generated using measured
Ada-mYPet fluorescence in single cells treated with 50 mM and 750 mM MMS production rate of 1 molecule per generation. Note that the actual value may
(added at time 0). Example cells in yellow; time in units of average generation be closer to 1.2 because of delayed maturation of mYPet (see supplementary
times (42 min) throughout. (F) (Inset) Transformed cumulative distribution materials).

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1095


R ES E A RC H | R E PO R TS

Fig. 2. Single-molecule trigger of the Ada response. (A) Stochastic ex- Histograms show number of frames spent in the expression rate states. Losses
pression and random segregation of molecules at cell division creates a can occur because of rare meAda degradation or by segregation at cell division.
subpopulation of cells with zero Ada molecules which therefore fails to auto- At very low numbers, all meAda molecules should sometimes remain in the
induce the adaptive response. (B) Sections of time traces showing distinct same cell, maintaining expression rates, as observed. (C) Uniform Ada-mYPet
steps in Ada-mYPet expression rates during response activation upon 200 mM induction when cell division was inhibited with cephalexin before MMS treatment
MMS treatment, deactivation after MMS removal, and stochastic activation and (orange). (D) Uniform accumulation of endogenous Ada-mYPet with additional
deactivation transitions with 100 mM MMS. Vertical lines indicate cell divisions. MiniF plasmid carrying PAda ada (green). Scale bars, 5 mm.

Fig. 3. High precision of the Ada response. (A) Cell fates after treatment with activation for endogenous ada-mYPet and ectopic PAda cfp are closely cor-
10 mM MMS for 1 hour: Percentages of cells failing to recover growth during related. Each dot represents one cell. (Inset) Example expression-rate time
time-lapse microscopy without MMS for 3 hours (SEM). Cells were distinguished traces with simultaneous activation of both genes. (E) Example time traces
if they had activated (Ada on) or failed the response (Ada off). (B) Percentages of showing correlated expression-rate fluctuations of the dual reporter genes and
cells spontaneously triggering Ada-mYPet expression without MMS (SEM). simultaneous response deactivation after MMS removal. (F) Deterministic re-
(C) Fano factors (variance/mean) for Ada-mYPet without MMS, using single- sponse deactivation: Time traces after MMS removal at time 0 (average: yellow).
molecule counting data from Fig. 1G (SEM bootstrapped). Cells grouped by The dilution model (circles) has an exponential decay constant equal to the
size. Expression bursting would give Fano factors above Poisson limit of 1. average generation time. (Inset) Narrow distribution of delay times from MMS
(D) Dual reporter assay: Delay times between MMS addition and response removal until response is deactivated (dotted line threshold).

1096 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE


RE S EAR CH | R E P O R T S

17. D. Fu, J. A. Calvo, L. D. Samson, Nat. Rev. Cancer 12, 104120


(2012).
18. H. Maamar, A. Raj, D. Dubnau, Science 317, 526529
(2007).
19. J. A. Bernstein, A. B. Khodursky, P.-H. Lin, S. Lin-Chao,
S. N. Cohen, Proc. Natl. Acad. Sci. U.S.A. 99, 96979702
(2002).
20. G.-W. Li, D. Burkhardt, C. Gross, J. S. Weissman, Cell 157,
624635 (2014).
21. B. M. Saget, G. C. Walker, Proc. Natl. Acad. Sci. U.S.A. 91,
97309734 (1994).
22. M. Elez et al., Curr. Biol. 20, 14321437 (2010).
23. S. Manley et al., Nat. Methods 5, 155157 (2008).
24. S. Uphoff, R. Reyes-Lamothe, F. Garza de Leon, D. J. Sherratt,
A. N. Kapanidis, Proc. Natl. Acad. Sci. U.S.A. 110, 80638068
(2013).
25. Y. Liao, J. W. Schroeder, B. Gao, L. A. Simmons, J. S. Biteen,
Proc. Natl. Acad. Sci. U.S.A. 112, E6898E6906 (2015).
26. J.-W. Veening, W. K. Smits, O. P. Kuipers, Annu. Rev. Microbiol.
62, 193210 (2008).

AC KNOWLED GME NTS


Fig. 4. Increased binding of mismatch recognition protein MutS in cells with delayed Ada response. We thank R. Reyes-Lamothe, U. Alon, J.-Y. Bouet, A. Kapanidis, C. Lesterlin,
Photoactivated single-molecule tracking of MutS-PAmCherry and Ada-mYPet fluorescence in single cells A. Upton, and P. Zawadzki for reagents and discussions. We thank
C. Saenz and the Microfluidics Core Facility at Harvard Medical School.
treated with 10 mM MMS for 1 hour. (A) Tracks of bound (red) and mobile MutS (blue). Cell outlines drawn;
Microscopy at Micron Oxford was supported by a Wellcome Trust
scale bars, 2 mm. (B) Percentage of bound MutS molecules versus Ada-mYPet fluorescence per cell. Native Strategic Award (091911) and Medical Research Council grant (MR/
strain with (yellow) and without MMS (black); Dada with MMS (gray). K01577X/1). S.U. is funded by a Sir Henry Wellcome Fellowship by the
Wellcome Trust and a Junior Research Fellowship at St Johns
College, Oxford. J.P., N.D.L., L.P-T., and B.O. are funded by NIH grant
of genetic adaptation in a hypermutagenic sub- 2. S. Uphoff, A. N. Kapanidis, DNA Repair (Amst.) 20, 3240 GM095784. L.P.-T. acknowledges fellowship support from the
population of cells without jeopardizing the genetic (2014). Natural Sciences and Engineering Research Council of Canada
3. R. S. Galhardo, P. J. Hastings, S. M. Rosenberg, Crit. Rev. (NSERC) and the Fonds de recherche du QubecNature et
integrity in the majority of the population during Biochem. Mol. Biol. 42, 399435 (2007). technologies. D.J.S. is funded by a Wellcome Trust Investigator
stress (3, 26). However, our observations that high 4. A. Marusyk, V. Almendro, K. Polyak, Nat. Rev. Cancer 12, Award (099204/Z/12Z). The primary data described in the manuscript
Ada expression is toxic and that cells appear to 323334 (2012). is available upon request. Author contributions: S.U. conceived the
minimize the heterogeneity in several ways suggest 5. M. B. Elowitz, A. J. Levine, E. D. Siggia, P. S. Swain, Science study, generated cell strains, and designed and performed experiments
297, 11831186 (2002). and analysis. S.U., D.J.S., and J.P. interpreted the data. N.L. and
that this is not an adaptive bet-hedging strategy, 6. I. Golding, J. Paulsson, S. M. Zawilski, E. C. Cox, Cell 123, L.P.-T. developed the microfluidic imaging methods. B.O. developed
but rather a side effect of maximizing short-term 10251036 (2005). the single-molecule counting method. S.U., D.J.S., and J.P. wrote the
fitness: Because proteins with the capacity to modify 7. P. J. Choi, L. Cai, K. Frieda, X. S. Xie, Science 322, 442446 manuscript. The authors declare competing financial interests. A U.S.
DNA can be detrimental, cells may be forced to (2008). Patent Application 20150247790 entitled Microfluidic assisted cell
8. L. Samson, J. Cairns, Nature 267, 281283 (1977). screening was filed on behalf of B.O., J.P., and co-workers by the
express them in low amounts, such that random 9. P. Landini, M. R. Volkert, J. Bacteriol. 182, 65436549 President and Fellows of Harvard College.
fluctuations are unavoidable. Mutations can then (2000).
result from stochastic variation in the concentra- 10. B. Sedgwick, Nat. Rev. Mol. Cell Biol. 5, 148157 (2004). SUPPLEMENTARY MATERIALS
tions of DNA repair proteins. Just as genetic het- 11. U. Alon, Nat. Rev. Genet. 8, 450461 (2007). www.sciencemag.org/content/351/6277/1094/suppl/DC1
12. P. Wang et al., Curr. Biol. 20, 10991103 (2010).
erogeneity can cause phenotypic heterogeneity, Materials and Methods
13. T. M. Norman, N. D. Lord, J. Paulsson, R. Losick, Nature 503,
Figs. S1 to S15
the reverse is thus also true. 481486 (2013).
Table S1
14. D. E. Shevell, P. K. LeMotte, G. C. Walker, J. Bacteriol. 170,
Movies S1 and S2
52635271 (1988).
RE FE RENCES AND N OT ES References (2735)
15. R. C. Brewster et al., Cell 156, 13121323 (2014).
1. N. M. Kad, B. Van Houten, Prog. Mol. Biol. Transl. Sci. 110, 124 16. J. C. W. Locke, J. W. Young, M. Fontes, M. J. Hernndez Jimnez, 8 July 2015; accepted 5 February 2016
(2012). M. B. Elowitz, Science 334, 366369 (2011). 10.1126/science.aac9786

SCIENCE sciencemag.org 4 MARCH 2016 VOL 351 ISSUE 6277 1097


PNAS Simplies
Online Submission.
Authors can now submit
to PNAS using the fast and
easy Express Submission.
When you choose Express Submission for both Direct
Submissions and for PNAS Plus, a single PDF le
containing manuscript text and any supporting
information may be uploaded, saving you time
and effort. To submit to PNAS, please visit
www.pnascentral.org.

PNAS
PNAS is committed to transparency in its editorial
review process. FAQs for authors and other resources
explaining the editorial process are available at
www.pnas.org
www.pnas.org/site/authors/index.xhtml
A M E R I CA N ASSO C I AT I O N FOR THE ADVANCEMENT OF SCIENCE

Open Minds

The American Association for the Advancement of Science (AAAS) is the worlds most trusted resource
for science across disciplines and a leader in empowering scientifc information in the public sphere.
By joining, you are bridging a gap between science and society. You are strengthening an organization
that develops evidence-based policy making, enhances science education, and communicates unbiased
science to the public. Your member benefts include:

51 weekly issues of Science, the leading Advocacy, community outreach,


journal of original science research and volunteer opportunities
Career resources and Timely science and technology news
networking opportunities Access to research dating back to 1880
Member-only webinars, videos, blogs, Discounts on journals, meetings,
online courses, podcasts, and more products, and services

Visit www.aaas.org/join. Together, we can make a diference.


SERVING SOCIETY THROUGH

SCIENCE POLICY

To make decisions, societies rely on knowledge and multiple


perspectives. Policies both within and outside science should be
informed by the best available evidence. Science alone cannot
translate knowledge into viable policy options; other factors
include societal norms and cultural values.

One aspect of science policy is advancing the practice of science:


How do we continually improve scientifc methods, mechanisms,
and outputs to increase the value of science to society?

Another aspect of science policy is providing the best available


scienti?c evidence to policymakers, community leaders, and
citizens. How can we most efectively provide scienti?c evidence
that informs policy? What are the roles for scientists as individuals
and for the scienti?c enterprise as a whole in the policy landscape?
What emerging and established areas of science are essential to
future policy decisions?

Call for Session Proposals


Session proposals for the 2017 AAAS Annual Meeting are now being
solicited. To submit a proposal, visit aaas.org/meetings.
The deadline for submission is April 22, 2016.

aaas.org/meetings

AMERICAN ASSOCIATION FOR THE ADVANCEMENT OF SCIENCE


CHEMISTRY

EXPERTISE
EMD MILLIPORE + SIGMA-ALDRICH
have joined forces to simplify your world
so you can focus with confidence on
tackling life sciences biggest challenges.

MilliporeSigma now offers an expanded Visit us at


global network that extends across ACS Spring Booth #919
more than 66 countries, a world-class March 13-17, San Diego CA
e-commerce platform that simplifies
transactions, and a combined portfolio
that has grown to include over 300,000 Learn more at
products from Catalytic Screening to sigma-aldrich.com/acs
Chromatography and more.

MilliporeSigma is a
business of Merck KGaA,
Darmstadt, Germany

Copyright 2016 EMD Millipore Corporation. All Rights Reserved. MilliporeSigma is a registered trademark of Merck KGaA, Darmstadt, Germany and its affiliates. 84384
2016 Global Conference
Connecting Science, Law,
and Opportunity

May 57, 2016


Monona Terrace
Convention Center
Madison, Wisconsin

REGISTER NOW

Visit biotechandthelaw.org

Cambridge Healthtech Institutes Fifteenth Annual REGISTER EARLY FOR


MAXIMUM SAVINGS!
MENTION KEYCODE 1620L35

Building a global network


for precision medicine by
uniting the Bio-IT community
APRIL 5 - 7, 2016 | Seaport World Trade Center | Boston, MA

CONFERENCE TRACKS PLENARY KEYNOTE SPEAKERS


1 Data & Storage Management 8 Data Visualization Howard Jacob, Ph.D.
& Exploration Tools Executive Vice President for Medical Genomics and
2 Data Computing Chief Medical Genomics Offcer, HudsonAlpha
9 Pharmaceutical R&D
3 Software Applications & Services Informatics Heidi L. Rehm, Ph.D., FACMG,
4 Cloud Computing 10 Clinical Genomics Chief Laboratory Director, Laboratory for Molecular Medicine,
Partners Healthcare Personalized Medicine; Clinical Director,
5 Bioinformatics 11 Open Source Innovations Broad Institute Clinical Research Sequencing Platform; Associate
Professor of Pathology, Brigham & Womens Hospital and Harvard Medical School
6 Next-Gen Sequencing Informatics 12 Cancer Informatics
Yaron Turpaz, Ph.D., MBA
7 Clinical Research & 13 Data Security
Translational Informatics Chief Information Offcer,
Human Longevity, Inc.

Enabling Technology. Leveraging Data. Transforming Medicine. Bio-ITWorldExpo.com


New Promotion March 1 June 30, 2016

Macro-offers for Microbiologists


Premium solutions for your microbiological processes now save up to 25 %!
Our new Eppendorf Advantage > Purchase any Eppendorf Xplorer or
special ofers include Eppendorf Xplorer plus electronic pipette with
premium products ideally suited 20 % of
for the most important steps of > Save 20 % on the epMotion 96
your microbiology workow. > Get the Eppendorf HeatSealer S100
> Buy a Mastercycler nexus gradient with 20 % savings
PCR cycler and save 25 %

www.eppendorf.com/advantage
Eppendorf, the Eppendorf logo, epMotion, Eppendorf Xplorer, and Mastercycler are registered trademarks of Eppendorf AG, Hamburg, Germany.
Eppendorf Advantage is a trademark of Eppendorf AG, Hamburg, Germany. U.S. Design Patents are listed on www.eppendorf.com/ip.
Ofers may vary by country. All rights reserved, including graphics and images. Copyright 2016 by Eppendorf AG.
LIFE SCIENCE TECHNOLOGIES
TISSUE ANALYSIS
In Search of

Explore one of the greatest mysteries of all time!


On the 80th Anniversary of
Amelias World Flight & Disappearance!
June 21July 11, 2017 On July 2, 1937,
Amelia Earhart and her
navigator Fred Noonan
Immunohistochemistry went missing on one of
the last legs of their round the
for the 21st Century world flight after they departed
New Guinea for Howland Island. Our colleagues at
Immunohistochemistry (IHC) dates back decades, the International Group for Historic Aircraft Recovery
but that doesnt mean its stagnant. Automated (TIGHAR), believe they turned their Electra south and
sample processors, digital slide scanners, image landed at Nikumaroro Island. On board the Reef
analysis software, and more have given the aged Endeavour, well sail north from Fiji for 1,000 miles,
procedure a decidedly 21st century air. stopping at Rotuma Island en route to Niku. We will
join TIGHAR scientists in their research at Niku and
look for artifacts which may indicate Amelia and
See the full story on page 1098.
Fred did land there! From $8,995 + air.
Upcoming Features For a detailed brochure, call (800) 252-4910
General Lab EquipmentApril 29 All prices are per person twin share + air
MicroscopyMay 13
Exosomes/MicrovesiclesJune 10 BETCHART EXPEDITIONS Inc.
17050 Montebello Rd, Cupertino, CA 95014
Email: AAASInfo@betchartexpeditions.com
Produced by the Science/AAAS Custom Publishing Office www.betchartexpeditions.com

myIDP:
A career plan customized
for you, by you.

For your career in science, theres only one

Features in myIDP include:


Exercises to help you examine your skills, interests, and values
A list of 20 scientic career paths with a prediction of which ones best t your skills and interests
A tool for setting strategic goals for the coming year, with optional reminders to keep you on track
Articles and resources to guide you through the process
Options to save materials online and print them for further review and discussion
Recommended by
Ability to select which portion of your IDP you wish to share with advisors, mentors, or others
leading professional
societies and the NIH A certicate of completion for users that nish myIDP.
Visit the website and start planning today!
myIDP.sciencecareers.org
In partnership with:
LIFE SCIENCE TECHNOLOGIES Produced by the Science/AAAS Custom Publishing Office

TISSUE ANALYSIS

Requiring from four to eight hours, the process is tedious


. - - +  , 
   - ,--+ 
slide, the general process includes antigen retrieval,
incubation with primary and secondary antibodies, washing,
and development with a chromogenic reagent. The data are
then interpreted using a light microscope.
The best way I can describe IHC is, its so easy a high
 --   -  +    - 
Experimental Pathology Laboratory at the National Insti-
tutes of Health and editor-in-chief of the Journal of Histo-
chemistry and Cytochemistry.
The problem is, each of these steps can be tweaked, and
, -- ! ,+,- --" , 
of biology, researchers dont always document their meth-
-#   -,- , -, +  "

Immunohistochemistry any event, the resulting images are typically read by eye,
an inherently subjective approach. As a result, variability

for the 21st Century from day to day and lab to lab can be considerable. But
researchers have developed strategies for locking down that
variability and improving other IHC pain points as well.
Immunohistochemistry (IHC) dates back decades, but that
doesnt mean its stagnant. Automated sample processors, Automated sample preparation
digital slide scanners, image analysis software, and more One approach to decreasing IHC variability is using auto-
have given the aged procedure a decidedly 21st century mationemploying systems that will treat every slide pre-
&,  
  cisely the same, while at the same time freeing technicians

I
and pathologists to tend to other tasks. What we talk about
mmunohistochemistry is all about spatial relationships.  -+,  +,- -, .-, -, # 
If researchers want to know whether or not a given 
- , -   - $-% -
piece of tissue expresses a particular protein, they can -,   -  .- &-,   
use mass spectrometry or Western blotting. But to de- situ hybridization (FISH) staining.
termine where in the tissue that protein is found, they have The NMCSD pathology lab has purchased systems to au-
to visualize it in its cellular context. tomate two stages of the IHC process, says Gates: Sample
Thats where IHC comes in. Employing antibodies as processing systems from Polaris and Leica dehydrate and
+,-. -,+  
   -+,-  
 , ,  ' +     (  ,Ventana
of thin slivers of tissuea tumor biopsy, for example system from Roche automates the IHC staining itself.
stained to display the expression and spatial distribution of For those with considerably lighter workloadsup to
one or more proteins. about 100 assays a weekthe SNAP i.d. 2.0 IHC system
It answers the fundamental question of localization, of from MilliporeSigma provides a vacuum manifold to increase
 , -, 
-
 -+, , ,
 
   # , +,-. -   

 -
 -, 
,-      says Long. Researchers can process up to 24 slides simul-
content development team at MilliporeSigma (a life science taneously as reagents are removed from all slides at once
business of Merck KGaA in Darmstadt, Germany). via vacuum suction, he says.
Spatial data can be used by researchers, for instance, to " -  -   -) ,- ,- , . .   *
, + -  , -  -   for clinical laboratories. But he notes that most academics
can use the data to grade tumors or assess a patients likely engaged in basic research have little need for such power.
response to a therapy. After all, automation is just another word for robotics, he ex-
The IHC laboratory at the Naval Medical Center San Di- plains. If youre just an investigator who needs to do a few
ego (NMCSD) sees some 18,000 surgical cases a year, ac- '+ ,  -,-,    ,
---.-, -, 
cording to Director Greg Gates. Each yields anywhere from -,- . 
PHOTO: VETPATHOLOGIST/SHUTTERSTOCK.COM

2 to over 100 slides, depending on the size of the sample


and the complexity of the case, but on average, Gates says, Digital pathology
each case generates about 4 slides apiecesome 72,000 Another transformative development in IHC is the advent
 - + --
 - , , .-  of so-called digital pathology. Instead of viewing and
80% of cases can be diagnosed from histologic staining analyzing IHC slides under a microscope, digital pathology
(such as hematoxylin and eosin) alone, estimates Andreas systems scan the stained slides in their entirety, storing
Hoel, director of marketing for IHC at Agilent Technolo- the resulting whole-slide images in a digital format that can
gies. But the remainder require more extensive evaluation. then be viewed and manipulated on a computer.

Upcoming Features
General Lab EquipmentApril 29 MicroscopyMay 13 Exosomes/MicrovesiclesJune 10

1098 sciencemag.org/custom-publishing SCIENCE


Produced by the Science/AAAS Custom Publishing Office LIFE SCIENCE TECHNOLOGIES
TISSUE ANALYSIS

%
3
2
 2
13

3 1
Conway, senior product manager for image analysis and
pathology imaging at Leica Biosystems, which sells both
 3 '
1(21'
 
111
the Aperio brand. For one thing, whole-slide imaging facili-
tates medical education by ensuring that students have ac-
cess to interesting and rare cases they might not otherwise
 1
1  
 )33
2 3 
process of sharing data with remote colleagues and obtain-
ing second opinionsa process called telepathology.
The University of Pittsburgh Medical Center (UPMC),
213
12
 23
3 2   3
13-
searchers can upload slides for evaluation by the hospitals
Digital pathology systems scan the stained slides in their
pathologists for between $50 and $150 per slide. Anthony
33 *3
+ ,222
1%
1
13,
- entirety, storing the resulting whole-slide images in a
thology at UPMC, who coauthored a 2012 study document- digital format that can then be viewed and manipulated
ing the universitys experience with telepathology, says on a computer.
he has evaluated samples from as far away as Australia

10
$3
&3 13
321 2(3
13 Researchers have used Open Slide to build other applica-
would be otherwise. tions as well, such as telepathology portals and an educa-
In one case, Demetris remotely evaluated a liver biopsy of tional resource called SlideTutor. For his part, Satya has
an elderly patient that the primary physician suspected was used it to implement his own search tool, called Diamond,
 1(

321 323 
3
 3 with which pathologists can ask such questions as, Among
a tumor, so he asked the on-site physician to perform and all patients seen at this hospital between this date and this
upload some follow-up stains for T and B lymphocytes and date who were taking this medication, do you see any [pa-
Epstein-Barr virus (EBV) nucleic acids. In just a few days, he tients] where the following properties are similar to what I
says, the proper diagnosis emerged: EBV-related lymphoma. see in the slide? OpenSlide, he says, greatly increases the
- 2 ' 
1 
2 
 32
 number of degrees of freedom you have as a designer.
and retrieval by providing an easily accessible record of
the slide (as opposed to physical slides), and it facilitates Image analysis
data analysis and subsequent reanalysis. You can go 0123 132 3

3 2 2 3 
 3 32 -

&

1
1

1323 


1
&13 form sophisticated image analyses. Given a whole-slide
questions, Demetris says. It actually becomes a reusable image, these software packages identify cells and sub-
data resource. cellular compartments (usually the cytoplasm, nucleus, and
1
32132
 
11



 2 cell membrane), outputting such values as cell counts and
Hamamatsu, Olympus, Philips, Roche, and Zeiss, among staining intensity.
others. But because each system uses a proprietary image The Halo software package from Indica Labs, for in-
format, researchers are generally restricted in what they can stance, includes modules for counting cells expressing one
do with the images, especially if they have scanners from or two proteins in the nucleus or cytoplasm, counting cells
multiple vendors. positive for membrane staining, assessing in situ hybridiza-
Mahadev Satyanarayanan (Satya), the Carnegie Group tion, and more. Researchers can even assess spatial rela-
Professor of Computer Science at Carnegie Mellon Uni- 321  
1
 132
3

versity1,33 ,11 
1


3
1. - For example, how many CD8-positive lymphocytes are
ence with that problem. Several years ago, Satya was work- within 50, 100, or 200 microns of the nearest tumor cells?
ing with pathologists to translate an image search algorithm In one recent study, a research team led by Antoni Ribas
he had developed to analyze digital slide images. He very
33 13 2
21
201  
232
quickly hit a wallthere was no way to search across dif- 
13   2 33 3 11332-
ferent image formats. The only software around that could partments in myeloma tumors. They observed a positive
interpret those formats was proprietary software, he says. 2
321313  12 !"  

PHOTO: SCIENCE PHOTO/SHUTTERSTOCK.COM

*2*
3
3
2&233 3
 23 13  at the tumor margin with patient responses to antibodies
formats, codifying that knowledge in OpenSlide, an open- targeting the programmed death-1 receptordata sug-
source C programming library. OpenSlide, he explains, func- 31 3
33
 2 321
3 
 2
tions like a computer printer driver, freeing programmers such therapeutics by selecting the patients most likely to
from needing to understand how a given scanner works. 1323 
You just say print, and its assumed that the device driver 

1
13122

#23-
underneath will provide the appropriate translation to the ware for analyzing photomicrographs has been available
 2 1
23
3 13/*
  2 
- 2 
$%  1 3
32&1 21
mers using OpenSlide can simply direct the software to read gigabyte-sized images. As a result, the software must be

 3 23
1 32&12 23
3
33 specially optimized to squeeze as much power as cont.>

SCIENCE sciencemag.org/custom-publishing 1099


LIFE SCIENCE TECHNOLOGIES Produced by the Science/AAAS Custom Publishing Office

TISSUE ANALYSIS

Featured Participants uous variable (say, 01000)a change that should allow
researchers to identify more subtle biomarkers.
Agilent Technologies Philips Healthcare
www.agilent.com www.philips.co.uk/
Yet according to David Rimm, professor of pathology at
healthcare/solutions/ Yale University and director of Yale Pathology Tissue Ser-
Carnegie Mellon pathology vices, the traditional method of IHC isnt really compatible
University with continuous variables, and thus isnt truly quantitative.
www.cmu.edu University of Pittsburgh IHC, he says, is not linear and its not reproducible.
Medical Center
Hamamatsu Photonics www.upmc.com In traditional IHC, he explains, a secondary antibody con-
www.hamamatsu.com/jp/ jugated to horseradish peroxidase converts a chromogenic
en/5007.html Ventana Medical Systems substrate into a dark brown or red precipitate. But as light
www.ventana.com cannot easily penetrate that materialwhat Rimm calls
Indica Labs
www.indicalab.com Yale Pathology Tissue mudthe method is at best semiquantitative.
Services /0112 00  
 
 0 
 0
Laboratory of Pathology, medicine.yale.edu/    
00011
 
  
NIH pathology/research/ (QIF). With such a strategy, small changes in protein
ccr.cancer.gov/laboratory- tissueservices
of-pathology abundance lead to comparable changes in light output,
Zeiss and coupled with whole-slide scanning and digital image
Leica Biosystems www.zeiss.com analysis, that technique can provide a powerful platform
www.leicabiosystems.com for IHC analysis.
MilliporeSigma Additional Resources Breast cancer biopsies, for instance, are commonly
www.emdmillipore.com tested for human epidermal growth factor receptor 2 (HER2)
 expression to determine whether the HER2-targeting drug
Naval Medical Center San  
 00 0  0 00  
Diego
www.med.navy.mil/sites/ OpenSlide teins extracellular domain, yet most of the antibodies used
nmcsd openslide.org for HER2 IHC in the clinic target the proteins intracellular
domain, notes Rimm.
Olympus Life Science SlideTutor
slidetutor.upmc.edu In 2014 and 2015, Rimm and his team used QIF to assess
Solutions

 0 0 00 0
www.olympus-lifescience
.com/en tibodies targeting the HER2 intra- or extracellular domains.
As it turns out, they could. Using the chromogenic sub-
strate 3,3-diaminobenzidine and IHC, other teams found
possible from a computers hardware. In an analysis of    0 0 
0 
50 whole-slide images, for instance, total Halo process- cytoplasmic domain antibody and the extracellular domain
ing time ranged from 21.7 hours with one CPU core to 0.9 antibodystaining intensity with both antibodies was iden-
hours (64 cores). 0 

0  0  1 !
"#$ 
!
1 & 10  
  0 
0 -*0 %#$ &0 2'/% 0   (
ens (recently acquired by MedImmune), PerkinElmer, and  
 10  ) 2 0 *
Visiopharm. Alternatively, Anne Carpenter of the Broad Insti- from trastuzumab [Herceptin].
tute of the Massachusetts Institute of Technology (MIT) and )
 10 10  

0  
10
  +. *+. *!  ence between a successful drug and a failure, says Rimm.
a pair of open-source software packages that accomplish Referring to the staining categories used today, he says,
many of the same tasks. Science isnt ordinalits continuous. And its possible
In theory, such tools could make life simpler in the clinic, that the value of many potential biomarkers will only be-
& 0  
0   0 
0    1
0 *   &0 0
regions of interest. However, says Lillard, image analysis Who knows how many other markers are out there that
tools are used mainly in research and drug development ultimately failed because there was not acceptance of the
rather than clinical laboratories, a result of regulatory ob- methods required to measure them quantitatively, and you
stacles. But that could change, she says. At the moment, needed to be above a certain level before the drug was ef-
researchers are using Halo and related software to identify fective? asks Rimm.
interesting biomarkers that correlate with drug response or Thats not to say every histopathologist needs to switch
survival. But to be of use, they will have to translate those to QIF and digital scanning. Most of what pathologists do
discoveries from bench to bedside. The hope is that those & +0  0  
00  /0112 ,*
will eventually transition into the clinic. & 
 2+ *& 
 &  
70 years, with no apparent end in sight.
    
   
Thanks to digital pathology and image analysis software,      
     
pathologists can score staining intensity not in relatively
crude bins (0, 1+, 2+, 3+) as is done today, but as a contin- DOI: 10.1126/science.opms.p1600103

1100 sciencemag.org/custom-publishing SCIENCE


Produced by the Science/AAAS Custom Publishing Office LIFE SCIENCE TECHNOLOGIES
NEW PRODUCTS: TISSUE ANALYSIS

Stem Cell Counting Technology Monoclonal Antibody


Since the beginnings of mammalian tis- A new immunohistochemistry (IHC)-
sue stem cell biology, counting has been )**+, )
* +)-
the bane of both basic research and clini- bit monoclonal antibody (mAb) is now
cal stem cell studies. Biomarkers, such as available for use in biomedical research.
#&&     ,) 
**+

,+, 

   
 $

&% activation, C10orf54, GI24, SISP1) is a

&  %  $# !" negative checkpoint control protein that

&& &$ $#  +)
 )  ) ,) 
Corporation, provides precise determina- response. Its role in negative checkpoint
tions of the number, viability, and self- control has made it an important target
renewal pattern (asymmetric versus sym-  ,,+ ,,) 
metric) of stem cells in complex tissue cell that seeks to develop better therapeutic
preparations, including both research and options by dismantling the mechanisms

# $
 $# !"
&- tumors use to generate and thrive in their
ogy uses computer simulation to extract Pathology Imaging System own immunosuppressive microenviron-
tissue stem cell properties from specially  (& 
&&)&  
 !"#) 
$%
%  % #
%##- ()'& & '&)) #64953 is validated using Cell Signaling
$# !"  

& 
- & 




  ,,$
+ ,+,

))+
,+ -
nology has many exciting applications, seven-color multiplexing and visualiza- , %) &')(
+
including determining the number of stem tion capabilities are designed to enable blot research applications.
cells in experimental samples, determin- pathologists and oncologists conduct- Cell Signaling Technology
ing the stem cell dose for transplantation ing research to gain a deeper level of For info: 877-616-2355
treatments, identifying agents that can understanding of disease mechanisms www.cellsignal.com
improve the vitality of transplanted tissue related to new cancer immunotherapy
stem cells, and identifying agents that are &''()& 



& 
 ChIP-Validated Antibodies
toxic to tissue stem cells. &&
& 
&')'
 A new range of chromatin immunopre-
Asymmetrex immune cells in situ )(&  " cipitation (ChIP)-validated antibodies are
For info: 617-990-6819 '&(&! %%#


 - 1#. . .
1  $. 
www.asymmetrex.com tions. It incorporates 10x whole slide .
/ " . .. 1 


imaging and the Phenochart whole slide essential for the success of a ChIP assay.
  

     viewer, allowing researchers to anno- . 1


0 1 . 
  '  tate and navigate slides with interactive 1. /01 .
11%
is designed for identifying and isolating interfaces to better identify regions of Antibodies from other applications do
$ &
 
&$ (
)(& 
' (&&  - 1.#.
#10&# "
1/
is used to label cancer stem cells with a
 ) '"  &'&  - $.  . .. 1 
#  10
&#'&
$  %'

' &
  & )&& & )) 
10
 1    0/011 1

distinguish cancer stem cells in live cell multiple biomarkers and reveals spatial and their results, safe in the knowledge
%'%$ '
&
%   $ ' )"  &  &)($))&
- that a ChIP-validated antibody was at the
was previously impossible. Aldehyde sist researchers with better, faster deci- heart of their assay. Chromatrap currently
&'$
$ $ &

)
)  ( #(
 1'0
.0.1"& 
10$. .-
cell marker enzyme, causes the Al- inForm image analysis software, which  /10.
 . 0.  1$"10
&$'#'&
 $& & &)& &  
&& 

$ ..
1#. 
spectrum, allowing the cancer cells to ')'& )
 

 (&  0
 . 1 
  
 . ..-
  $
'$  ' &- system provides researchers the power  1/010.# .0!(%/1
&
'&#'&
  
 of multiplexed biomarker imaging and %/1)!(*+()"
transgenic animals, and reporter assays. & & &&

  &&  &(  !( !,. (."01.0./
& '
&%'&&
"  &   &)($) validated antibodies are available in 100-
#'&
  $  ! ' PerkinElmer . $1
.. 1 0. 11-
identify positive cells in an otherwise For info: 877-754-6973 ml, delivered overnight.
&#'&
$ &' www.perkinelmer.com Chromatrap
EMD Millipore For info: +44-1978-666222
For info: 800-645-5476 www.chromatrap.com
www.emdmillipore.com/stemcells

Electronically submit your new product description or product literature information! Go to www.sciencemag.org/about/new-products-section for more information.
Newly offered instrumentation, apparatus, and laboratory materials of interest to researchers in all disciplines in academic, industrial, and governmental organizations
are featured in this space. Emphasis is given to purpose, chief characteristics, and availability of products and materials. Endorsement by Science or AAAS of any
products or materials mentioned is not implied. Additional information may be obtained from the manufacturer or supplier.

SCIENCE sciencemag.org/custom-publishing 1101


want new
technologies?
antibodies
apoptosis
biomarkers
cancer
cytometry
data
diseases
DNA
epigenetics
genomics
immunotherapies
medicine
microbiomics watch our webinars
microfuidics Learn about the latest breakthroughs, new
microscopy technologies, and ground-breaking research
in a variety of felds. Our expert speakers
neuroscience explain their quality research to you and
proteomics answer questions submitted by live viewers.
sequencing
toxicology VIEW NOW!
transcriptomics webinar.sciencemag.org

Brought to you by the Science/AAAS


Custom Publishing Ofce

@SciMagWebinars
Whats in Your Sample?
Find Your Answers!

rndsystems.com/
Immunoassays
online @sciencecareers.org

College of Medical, Veterinary and Life Sciences


Research Institute of Infection, Immunity and Inflammation
(Centre for Virus Research)

Professor Chair in Infectious Diseases (Virology)


Theres only one Ref: M00659
The MRC-University of Glasgow Centre for Virus Research (CVR) is a distinctive

Science Careers international research centre with a critical mass of researchers dedicated to the
multidisciplinary study of human and animal viruses. We are seeking a distinguished
Advertising researcher to establish a programme in the general area of emerging viruses
complementing the existing strengths of the CVR, either in basic, clinical or translational
For full advertising details, go to research. Research programmes focused on any aspect of emerging virus infections,
ScienceCareers.org and click including virus genomics, evolution and virus-host interactions are welcome. The
For Employers, or call one of appointee will have a distinguished track record of research achievement, an international
our representatives. reputation and a demonstrated ability to attract national and international funding. They
will take on a research leadership role within and beyond the CVR.
Tracy Holmes
The CVR has a collegiate and collaborative environment and it is well positioned to offer
Worldwide Associate Director
an attractive start-up package and resources.
Science Careers
Phone: +44 (0) 1223 32()2) For further information, please visit www.cvr.ac.uk and www.facebook.com/
THE AMERICAS centreforvirusresearch. Informal enquiries should be directed to Prof Massimo Palmarini
(massimo.palmarini@glasgow.ac.uk). For further details and to submit an electronic
E-mail: advertise@sciencecareers.org application online visit: www.gla.ac.uk/about/jobs/
Fax: +1 (202) 28+ (742 Closing date 31 March 2016
Tina Burks The University has been awarded the Athena SWAN Institutional Bronze Award.
Phone: +1 (202) 32( ()77 The University is committed to equality of opportunity in employment.
The University of Glasgow, charity number SC004401.
Nancy Toema
Phone: +1 (202) 32( ()78
Online Job Posting Questions www.glasgow.ac.uk
Phone: +1 (202) 312 (37)
EUROPE /I NDIA/ A USTRALIA/
NEW ZEALAND / R EST OF WORLD
E-mail: ads@science-int.co.uk JOB FOCUS: CHEMISTRY POSITIONS OPEN
Fax: +44 (0) 1223 32()32 PRINCETON UNIVERSITY DEPARTMENT
OF CHEMISTRY
Sarah Lelarge
Phone: +44 (0) 1223 32()27 The Princeton University Department of Chemistry
seeks two postdoctoral research associates for ultrafast
Kelly Grace laser experimental positions in the group of Professor
Phone: +44 (0) 1223 32()28 Herschel Rabitz. The first research position involves
Online Job Posting Questions
ultrafast laser control in biology of multiple optogenetic
switches and fluorescent protein sensors in vitro and The University of Minnesota is seeking an Executive
Phone: +44 (0) 1223 32()28 Director who will oversee the Bell Museum of Natural
in vivo. The second position concerns exploring the
History and Planetarium s (BMNHP) mission through
JAPAN systematics of laser control of chemical reactions. Ap-
plicants with experience in ultrafast lasers and ideally its personnel, research/programs and facilities. The
Katsuyoshi Fukamizu (Tokyo) BMNHP mission is to ignite curiosity and wonder, ex-
quantum control are particularly encouraged to apply.
E-mail: kfukamizu@aaas.org Applications must be submitted electronically to website: plore our connections to nature and the universe, and
Phone: +81 3 321+ )777 https://jobs.princeton.edu (Requisition No. 1600157) create a better future for our evolving world. For more
and submit Curriculum Vitae, list of publications, and information about this position and to apply, go to the
Hiroyuki Mashiki (Kyoto) University of Minnesota employment website and search
E-mail: hmashiki@aaas.org contact information for three references. Ph.D. required.
These positions are subject to the University_s back- for requisition #307546.
Phone: +81 7) 823 110+
ground check policy.
CHINA / KOREA /S INGAPORE/ Princeton University is an Equal Opportunity Employer and
TAIWAN /THAILAND all qualified applicants will receive consideration for employment
Ruolei Wu
without regard to race, color, religion, sex, sexual orientation, gender
identity, national origin, disability status, protected veteran status, or
Phone: +8( 18( 0082 +34)
any other characteristic protected by law.
E-mail: rwu@aaas.org
Regulatory Pharmaceutical Analysis Program
All ads submitted for publication must comply with OFFICE OF REGULATORY AFFAIRS (ORA)
applicable U.S. and non-U.S. laws. Science reserves
the right to refuse any advertisement at its sole
Download the The FDA_s Office of Regulatory Affairs is seeking
Ph.D. level scientists with advanced experience in chem-
discretion for any reason, including without limitation ical analysis of complex pharmaceutical products to
for offensive language or inappropriate content,
and all advertising is subject to publisher approval.
work at one of Office of Regularoty Affairs field
laboratories located in Jamaica, New York or Irvine,
Science encourages our readers to alert us to any ads
that they feel may be discriminatory or offensive. Job App California. Candidates must have expertise either in
mass spectrometry or NMR spectroscopySalary range:
$75,000 - $120,000 per year and will be commensurate
SEARCH JOBSS with experience. Interested persons can can send cur-
riculum vitae or resume to email: ORSJOBSEARCH@fda.
ON THE GO! hhs.gov Attention: Gail Tarry.

ScienceCareers.org apps.sciencemag.org
g.org
online @sciencecareers.org
MY WORK IS A
STORY OF
SUSTAINABILITY.
I help save the environment. Im part of a group that works to improve
processes for our customers, including replacing hazardous chemicals
with more environmentally friendly alternatives. Its rewarding to be part
of a team thats having a real impact by promoting smarter, safer chemistries.

Our innovations have even earned Thermo Fisher Scientic the

featured employer
Environmental Protection Agencys Presidential Green Chemistry Challenge
Award. With the largest R&D budget in the industry, we are fully committed
to promoting innovative thinking and pursuing a sustainable future.

Matt
Scientist

When youre ready for sustainable success in your career, youll discover
that, at Thermo Fisher Scientic, each one of our 50,000 extraordinary minds
has a unique story to tell. And we all contribute to a singular mission
enabling our customers to make the world healthier, cleaner and safer.

What story will you tell?

Explore our opportunities at


jobs.thermofsher.com and join our team.
online @sciencecareers.org

Advance
IN SEARCH OF your career
REMARKABLE SCIENTISTS with expert
Molecular Technologist, Houston, Texas advice from
Science
Do you have a PhD in microbiology or biochemistry with a strong
Careers.
desire to change the world to biofuels? Then wed like you to help us
pilot an exciting new project.
Joining our in-house microbial biofuels team, youll be working with
transformable host organisms to efciently produce different biofuel
components at both lab and pilot scale. Youll need experience in
one or more of the following areas: metabolic pathway improvement
via transformation, microbial/pathway evolution, transformation of
non-standard hosts, solid or liquid fermentation (batch or continuous
ow at pilot scale).
Find out more and apply online at www.shell.com/careers. In the
Students and Graduates section, click on Shell Recruitment Day
and then create an account. When entering your contact details, Download Free Career
select Other in response to Where you found out about this Shell Advice Booklets!
opportunity, click Next and then type in Biofuels. ScienceCareers.org/booklets

DISCOVER WHAT YOU CAN ACHIEVE AT SHELL

Minority, Female, Disabled and Veteran EEO/AA Employer.

GRADUATE PROGRAMS

Sandia National Laboratories is seeking a Co-Director for the Center for


Integrated Nanotechnology (CINT), who will also lead the Integrated
Nanotechnologies Department.
The Integrated Nanotechnologies Department, within CINT, oversees the
operations of several departments and provides oversight in the areas of Basic
Energy Sciences and Materials Science.
CINT is a Department of Energy/Office of Science Nanoscale Science
Research Center jointly supported by Sandia National Labs and Los Alamos
National Lab, operating as a national user facility devoted to establishing
nanoscale science by developing the scientifc principles that govern the
design, performance and integration of nanoscale materials. With sites in
both Albuquerque, NM and Los Alamos, NM, CINT provides world-class
facilities and expertise to the broad scientifc community with the mission to
support and accelerate the progress of nanoscience. Proposal-based access is
offered to tools and expertise focused around CINT scientifc thrust areas:
nanoscale electronics, mechanics and systems, nanophotonics and optical
nanomaterials, soft and biological nanomaterials, theory and simulation of
nanoscale phenomena.
Job Summary: The Senior Manager will work closely with management
colleagues at Los Alamos to:
Provide leadership to ensure the success of the overall scientifc and user
program of CINT
Maintain and grow strong and effective partnership between the Laboratories
Grow the impact of CINT to the missions of DOE and the two host
Laboratories
Develop and execute CINTs scientifc vision and strategy as it relates to the
broader set of sponsors beyond BES
Serve as the primary point of contact for CINTs external advisory bodies
The advancement plan for this role is excellent, as the Co-Director is expected
to succeed the Director at a future point.
Apply by completing an online application at: http://www.sandia.gov/
careers. Click on View all jobs and enter Job Number 652668 into the
Keyword search.
online @sciencecareers.org
The University and its STEMM disciplines all hold silver or bronze Athena SWAN charters

CROSS THE
BOUNDARIES
TO NEW
HORIZONS
The Living Systems Institute (LSI) is a major new fundamental understanding into predictive capability, to
initiative by the University of Exeter (Global Top 100, secure human health and wellbeing.
Times Higher Education World University Rankings)
designed to promote interdisciplinary research A total of twenty appointments will be made at a variety
that transcends the boundaries of mathematics, of career stages, based on experience and potential.
engineering and the biological, medical and physical A demonstrable interest in the fundamental principles
sciences. Its mission is to analyse living cells and that govern how living-systems operate and how they
organisms as integrated systems, with the aim of breakdown in disease is essential, as is a commitment to
unravelling how diseases disrupt biological processes working across traditional discipline boundaries. We are
at the molecular, cellular and organismal level. particularly keen to recruit investigators with expertise in
mathematics, physics, chemistry or engineering, as well
The LSI will open in autumn 2016 under the direction as in the life sciences.
of eminent geneticist, Philip Ingham FRS. Our vision is
to build a team of researchers who will work together Candidates for Full and Associate Professorial
across disciplines to identify new ways to diagnose, appointments will be leaders in their eld of research
treat and cure disease in plants and animals, especially as evidenced by outstanding publication and funding
humans. Our goal is to identify common signatures of track records. Early career stage investigators will be
disease, irrespective of host or the precise mechanism considered for appointment to xed term (ve year)
of disease pathogenesis (i.e. whether caused by genetic Fellowships with the opportunity for progression to
changes or environmental agents). Group Leader and a proleptic permanent appointment.
All appointees will be highly innovative and collaborative
We are now seeking to appoint ambitious investigators with strong interdisciplinary research interests and
committed to working in this interdisciplinary manner, activities that provide an excellent t for collaboration
who will explore the basis of disease and translate within the LSI and beyond.

Professors in Living Systems (Ref: P00582) Senior Lecturers in Living Systems (Ref: P00584)
Associate Professors in Living Systems (Ref: P00583) Research Fellows in Living Systems (Ref: P00585)

The closing date for applications is 11 April 2016


Interviews are expected to take place between 31 May and 3 June 2016.
For more information go to www.exeter.ac.uk/livingsystems/recruitment
For an informal discussion about opportunities within LSI please contact:
infolsi@exeter.ac.uk

The University of Exeter is an equal opportunity employer which is Positive about Disabled People. Whilst all applicants will
be judged on merit alone, we particularly welcome applications from groups currently underrepresented in the workforce.
WORKING LIFE
By Elizabeth Pennisi

Making a game of science

W
hen Ariel Marcy was 7 years old, her physician father gave her an electronic game to help
her understand what he did all day. Playfully exploring the human body as a child, she
says, set her on the path to where she is today: working to inspire the next generation
with her own science games and beeng up her credentials by pursuing a biology Ph.D. at
the University of Queensland in Brisbane, Australia. One of her creations, Go Extinct!, a
board game in which players explore evolutionary trees, is already in a second printing. To
fashion this dual career in science and game design, Marcy had to think beyond the standard routes
in science training and take the initiative to make her own way.

As an undergraduate student in beginning of this new game, she

Downloaded from on March 3, 2016


biology at Stanford University in teamed up with several schools to
Palo Alto, California, Marcy came get students to test early versions.
under the wing of evolutionary With the design in hand, she
biologist Elizabeth Hadly, whose turned to the crowdfunding site
outreach work made Marcy keenly Kickstarter and raised more than
aware of the value of doing more $16,000, enough to pay for an artist
than just research. Eager to branch and print 1000 games. Eight months
out from her science studies and after the rst printing, the game was
broaden her skills, Marcy went on to sold out and she had 3000 more
take design classes and build games printed. Shes now working on a new
in a computer science course. As she game, Suddenly Cute, in which
was nishing her bachelors degree, players learn how small changes
she didnt want to head down the early in a vertebrates embryonic
expected path of graduate school. development can lead to big difer-
Instead, she realized that she could ences in adults.
combine her loves of evolutionary Even with Go Extinct! under her
biology and building games to be- She could combine her loves belt, she found she needed more cre-
come an educational game designer. of evolutionary biology and dentials to make it in the game de-
While working as a human biology sign world, so her pendulum swung
instructor, she made her rst foray building games. back toward research. Now a Ph.D.
into the game world. She developed a student, she plans to scale back her
digital game to help students master cell signaling pathways, game design eforts, temporarily. She sees her Ph.D. as im-
with the aid of lots of books and advice from game designers. portant for a long-term career in science outreach and educa-
At the same time, she recalls, I was noticing that my stu- tion. As a person advocating for science and science careers,
dents were having a tough time wrapping their heads around it gives me more authority and shows more integrity for me
evolutionary trees. To make the concept easier to grasp, she to have actually pursued the advanced degree required for
put her design skills and gaming experience to work. She the discipline, she says.
spent hours combing through the scientic literature to de- Her experience designing games has also helped prepare
velop Go Extinct! and took an early prototype to the Society her for academic work. I have a better understanding of
of Vertebrate Paleontology annual meeting to get feedback what the public thinks biology is, she says, which will help
from researchers. Her biology training helped with more than me in communicating my research. Shes learned that she
ILLUSTRATION: ROBERT NEUBECKER

just the content. Both research and game development re- needs to work on listening, too. Its really easy to come up
quire a lot of experimentation and testing to see what works, with an idea that you are really in love with, she says. The
Marcy explains. In addition, success in either arena requires most challenging part for me is listening to feedback and
working with a lot of people and taking advantage of their acting on it. But receiving criticism and feedback is a huge
expertise. She also built on past experienceand missteps, part of science.
a tactic learned from lab work. She only tried out her rst
game, Cancer Avenger, on student players after launching Elizabeth Pennisi is a senior correspondent for Science. Send
it, when it was too late to x design aws. So from the very your story to SciCareerEditor@aaas.org.

1106 4 MARCH 2016 VOL 351 ISSUE 6277 sciencemag.org SCIENCE

Published by AAAS

Potrebbero piacerti anche