European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84

Contents lists available at ScienceDirect

European Journal of Pharmaceutics and Biopharmaceutics

journal homepage: www.elsevier.com/locate/ejpb

Research paper

Accurate prediction of vaccine stability under real storage conditions and T

during temperature excursions
Didier Clénet
Formulation & Stability Platform, Bioprocess R&D Department, Sanofi Pasteur, 1541 Avenue Marcel Mérieux, 69280 Marcy L'Étoile, France

A R T I C L E I N F O A B S T R A C T

Keywords: Due to their thermosensitivity, most vaccines must be kept refrigerated from production to use. To successfully
Vaccine stability carry out global immunization programs, ensuring the stability of vaccines is crucial. In this context, two im-
Shelf life predictions portant issues are critical, namely: (i) predicting vaccine stability and (ii) preventing product damage due to
Stability modeling excessive temperature excursions outside of the recommended storage conditions (cold chain break). We applied
Cold chain break
a combination of advanced kinetics and statistical analyses on vaccine forced degradation data to accurately
Storage temperature
describe the loss of antigenicity for a multivalent freeze-dried inactivated virus vaccine containing three var-
iants. The screening of large amounts of kinetic models combined with a statistical model selection approach
resulted in the identification of two-step kinetic models. Predictions based on kinetic analysis and experimental
stability data were in agreement, with approximately five percentage points difference from real values for long-
term stability storage conditions, after excursions of temperature and during experimental shipments of freeze-
dried products. Results showed that modeling a few months of forced degradation can be used to predict various
time and temperature profiles endured by vaccines, i.e. long-term stability, short time excursions outside the
labeled storage conditions or shipments at ambient temperature, with high accuracy. Pharmaceutical applica-
tions of the presented kinetics-based approach are discussed.

1. Introduction
The thermal sensitivity of formulated biologics, including vaccines,
is a critical factor impacting product quality and potency, hence in-
fluencing their worldwide distribution [1,2]. The thermal stability of
vaccines is of concern to the vaccine industry, health authorities, gov-
ernment institutions, and philanthropic organizations attempting to
increase the distribution of vaccines to people living in countries with
poor infrastructure, unreliable transportation and substandard storage
facilities for the preservation of vaccines requiring refrigeration [3].
World Health Organization (WHO) sets guidelines for the stability
evaluation of vaccines [4] and recommends conducting regular and
accelerated stability studies. These guidelines aim to provide a frame-
work for shelf life and storage conditions, monitor vaccine stability in
the post licensure period and support manufacturing changes by de-
monstrating consistency of vaccine batches. Such guidelines can be
used for the selection of optimal stabilizing conditions, shelf life esti-
mation, temperature excursion modeling, and investigation of changes
to manufacturing process that may potentially alter vaccine stability
[5]. Many different stability indicating assays can be used, including
viral titer, immunochemical assays, liquid chromatography and gel
electrophoresis [6,7]. An often used parameter is shelf life, this refers to
the time period during which a drug product remains capable of ac-
ceptable performance. Several shelf life estimation methods exist, those
described in International Conference on Harmonisation (ICH) guide-
lines [8] are generally based on linear or nonlinear regression and
statistical modeling through poolability tests. Focusing on the fact that
an agreement does not currently exist, an investigation of current sta-
tistical methods for estimating shelf life, based on stability data, was
conducted by a Product Quality Research Institute (PQRI) working
group [9]. Additionally, to improve ICH procedure, the application of
quantile regression or mixed model tolerance interval methods and
batch effects, was proposed [10,11].
Kinetics is typically used to estimate vaccine degradation rates
through accelerated stability programs which expose products to tem-
peratures greater than those recommended for vaccine storage (typi-
cally 5 °C, 25 °C, 37 °C). During kinetic analysis of the data, very often
the simplified kinetics models, such as zero- or first order mechanisms,
are considered. Such models fail to correctly describe the complicated
course of decomposition of biological materials, which frequently show
complex and multi-step degradation behavior [12–15]. The rate con-
stant derived from simplified models are often of little value during
kinetic workflow; for example the determination of the kinetic values of
pre-exponential factor and activation energy from the Arrhenius plot

E-mail address: didier.clenet@sanofi.com.

https://doi.org/10.1016/j.ejpb.2018.01.005
Received 28 July 2017; Received in revised form 29 November 2017; Accepted 8 January 2018
Available online 16 January 2018
0939-6411/ © 2018 The Author. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/BY-NC-ND/4.0/).
D. Clénet
[7,16]. Superior kinetic description of the experimental decomposition
data of biologicals [17,18] was obtained with kinetic parameters de-
rived from two-step models, these better mimic the complicated de-
composition of the investigated samples. Also of great use is the Šesták-
Berggren approach [19], being a phenomenological kinetic model
which allows kinetic analysis, even of complex reactions [14,20].
In this report, we applied a general procedure combining advanced
kinetic and statistical analyses of vaccine forced degradation data to
identify kinetic models which best describe loss of antigenicity for three
variants of inactivated viruses in freeze-dried form. Kinetic analysis was
applied to predict the long-term vaccine degradation of the freeze-dried
products, either under isothermal or specific temperature excursions.
Additionally, the antigenicity recoveries predicted by kinetic models
were compared with the experimental data of tested samples.
Pharmaceutical applications of the presented kinetics-based approach
are discussed.
2. Material and methods
2.1. Materials
A multivalent freeze-dried inactivated vaccine by Sanofi Pasteur
was used for this study; it is composed of three different viral variants
referred as A, B, C. To allow the detection of the antigenicity recovery
losses during temperature excursions outside of the refrigerated storage
condition (cold chain break), a thermosensitive formulated product was
selected. Final bulk products were prepared using an appropriate sta-
bilizer containing a mixture of excipients (amino acids, sugar, surfac-
tant and chelator) in a HEPES buffer to maintain pH at 7. All reagents
were purchased from Sigma-Aldrich (St. Louis, MO, USA). 0.5 mL of
final bulk products were dispensed in 3 mL type I glass vials partially
stoppered with septum closure and subsequently freeze-dried, then
fully stoppered and crimped with aluminum cap. Dried products were
investigated with low residual moisture content (0.7%) measured by
coulometric titration (Karl Fisher).
Antigenicity was tested by enzyme-linked immunosorbent assay
(ELISA) at different time intervals.
2.2. Stability monitoring and antigen content by ELISA
Product stability was studied by storing freeze-dried formulations at
5 °C, 25 °C, 37 °C and 45 °C into temperature-controlled incubators for
up to 6 months. This forced degradation plan was carefully designed to
obtain experimental datasets able to be fitted by the software, with a
minimum of 20 data points in total obtained at three or more storage
temperatures, with significant degree of reaction reached for the more
drastic condition. Hence, time intervals for the forced degradation plan
were chosen, as follows: 0, 30, 90 and 180 days at 5 °C; 7, 14, 30, 60, 90
and 180 days at 25 °C; 1, 3, 7, 14, 30, 60 and 180 days at 37 °C, and 1, 3,
7, 14, 30 and 90 days at 45 °C.
During studies of the influence of temperature excursions, the two
following temperature programs were applied to freeze-dried products
before analysis by ELISA: (i) 1 month at 5 °C, followed by 9 months at
25 °C and then returned to 5 °C for 2 months; (ii) 1 month at 5 °C, fol-
lowed by 6 months at 37 °C and then returned to 5 °C for 4.5 months. To
evaluate the impact of experimental shipments under uncontrolled cold
chain conditions, product samples kept for 534 days (∼1.5 year) at 5 °C
were shipped from Lyon (France) to Toronto (Canada) at ambient
temperature. Another set of samples was also sent from Lyon (France)
to Barcelona (Spain). Real temperature profiles were recorded with
temperature monitoring devices (TempTale®4 USB Monitors, Sensitech
Inc., Beverly, USA) placed into the box containing the vaccines. The
temperature profiles recorded continuously during shipment were used
for predictions. All samples came back to France and were kept at 5 °C
before ELISA analyses.
The antigen titer of vaccine samples was determined by a variant-
77
European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84
specific sandwich ELISA. Freeze-dried products were stored at 5 °C
before analysis and reconstituted extemporaneously prior to testing.
Three vials were analyzed for each kinetic time point and a mean of
these independent dosage values was displayed. Results are system-
atically expressed as a percentage of the mean value at T = 0 and re-
ferred to as antigen recovery. To estimate prediction accuracy, per-
centage point was used as the arithmetic difference between
antigenicity recoveries, experimentally determined and predicted.
2.3. Advanced kinetic analyses
A general stability modeling procedure using AKTS-Thermokinetics
software [21] (version 4.1, AKTS AG, Advanced Kinetics and Tech-
nology Solutions, Siders, Switzerland) was already described in detail
[20] and applied to predict stability for various products constituting
vaccines [14]. Briefly, according to a least-squares regression analysis,
the software uses ‘one-step’ and ‘two-step’ kinetic models based on the
truncated Šesták-Berggren equation (‘two-step’ in Eq. (1) to determine
the function(s) which best fits the experimental data. Hence, reaction
rate was described as follows [20]:
dα E E
= A1 exp ⎛− 1 ⎞ (1−α )n1α m1 + A2 exp ⎛− 2 ⎞ (1−α )n2α m2
dt ⎝ RT ⎠ ⎝ RT ⎠ (1)
With α: the reaction progress, A: pre-exponential factor, E: activation
energy, n: reaction order, m: a parameter introduced to take into ac-
count the possible autocatalytic behavior of reaction (1) and (2), for the
first and second step, respectively. During evaluation one, two or more
reaction stages may be considered, depending on the complexity of the
reaction. For a two-step reaction, consecutive and competitive type
models were screened. Competitive two-step models use one more ad-
justable parameter than consecutive two-step model, with the con-
tribution of the reaction rate of the first and second stage to the overall
reaction rate. The quality of regression fit is quantified according to RSS
(Residual Sum of Squares) value. The combination of different kinetic
models best describing the reaction course is identified with software
according to the higher Akaikie Information Criteria and Baysian In-
formation Criteria weighted scores, referred to as wAIC and wBIC, re-
spectively. In kinetic analysis, when applying the different models de-
scribing the reaction course, an important issue is to find the best
method for their discrimination [22]. The criteria for comparing models
applied in this study were based on the information theory introduced
by Akaike [23] and its Bayesian counterpart [24,25]. Both these criteria
indicate not only which model is more likely to better fit the analyzed
data, but also allow quantification of likelihood. An example of criteria
application, using limited experimental points for the prediction of
thermal stability of biological materials, is described in detail elsewhere
[20]. Finally, by using the selected model, the software was employed
to run bootstrap analysis (resampled 1000 times with replacement) and
determine the confidence intervals (CI), in the form of the upper and
lower 95 percentiles (predictive band at 95%), which were evaluated
for each of the fitted parameters [20]. After determination of the best
kinetic models the software was able to simulate the reaction progress
during arbitrarily chosen aging time and under any temperature pro-
files. We have used this approach to continuously predict degradation
progresses of variants when vaccines were kept under isothermal con-
ditions (long-term stability), during temperature excursions and during
exposures under real atmospheric temperature profiles (e.g. shipments
at ambient temperature).
3. Results
3.1. Thermal degradation of virus-based vaccine described by two-step
kinetic models
Lots of models, starting from the simplest ones (zero and first-order)
and ending with more complex were screened to fit the six months
D. Clénet European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84

Fig. 1. Kinetic modeling of vaccine antigenicity. Antigenicity predictions are displayed as lines for variants A, B and C at 5 °C (blue), 25 °C (red), 37 °C (green) and 45 °C (pink) based on
one-step models (first-order and zero-order models displayed as solid lines and dashed lines, respectively) in left column and based on two-step models (competitive and consecutive
models are displayed as solid and dashed lines, respectively) in right column. ELISA data used for kinetic modeling are displayed as filled circles. Additional experimental data obtained
later in time, for another vaccine batch, are displayed as open triangles (blue: 5 °C, red: 25 °C after 2.5 months).

experimental data. The kinetic parameters (A, E, n, m) were system- models were preferred to consecutive models for two-step type kinetics.
atically adjusted during the fitting procedure, comprising ‘one-step’ or Consequently, the most appropriate kinetic models for variants A, B,
‘two-step’ kinetic models. The software compared and ranked the and C were competitive two-step models, mathematically described as
models from the worst to the best, for each variant. One-step models follows:
such as zero- and first-order kinetics did not correctly fit experimental Variant A:
data, especially at temperatures higher than 5 °C for variant A and C
dα 83.2E3 119.3E3
⎞ (1−α1) 4 + 0.591∗1.68E 31 exp ⎛− ⎞ (1−α2)2
(Fig. 1, left). Better fits were obtained by using two-step models for all dt
= 0.409∗1.08E 8 exp ⎛−
⎝ RT ⎠ ⎝ RT ⎠ (2)
variants (Fig. 1, right). Based on ranking statistical scores wAIC and
wBIC, competitive two-step model containing six adjustable kinetic Variant B:
parameters were selected, since they provided the best quality of fit dα 72.0E3 197.0E3
= 0.115∗8.89E 6 exp ⎛− ⎞ (1−α1) + 0.885∗5.13E 25 exp ⎛− ⎞ (1−α2)3
(low RSS value) and best wAIC and wBIC scores. It was confirmed for all dt ⎝ RT ⎠ ⎝ RT ⎠ (3)
three serotypes, with wAIC and wBIC scores at least equal to 15%
Variant C:
(Tables S1–S3), in contrast to the other scores, close to zero, obtained
with one-step and consecutive two-step models (Tables S1–S3). Firstly, dα
= 0.327∗4.79E13 exp ⎛−
111.3E3
⎞ (1−α1) + 0.673∗3.48E 20 exp ⎛−
164.1E3
⎞ (1−α2)
this means that the adjustable kinetic parameters used with one-step dt ⎝ RT ⎠ ⎝ RT ⎠ (4)
models were not enough to fit datasets, and secondly, that competitive where for example, in Eq. (2), the values 0.409 and 0.591 represent the

78
D. Clénet
contribution of the reaction rates of the first and second stages of the
overall reaction rate, 1.08E8 and 1.68E31 depict the values of the pre-
exponential factors A1 and A2 for first- and second steps, 83.2E3 and
119.3E3 represent the values of the activation energy E1 and E2, ex-
ponents 2 and 4 show the value of the reaction order in the first- and
second step, respectively. For all three variants, such contribution was
higher for the second stage of the reaction rate. Activation energies
obtained for the second stage were also higher than the activation en-
ergies of the first stages (e. g. in Eq. (2), 119.3 kJ/mol for the second
stage vs. 83.2 kJ/mol for the first stage). Finally, a similar observation
could be made regarding pre-exponential factors with higher values for
the second stage as compared to the first stage (e. g. in Eq. (2), 1.68E31
for the second stage vs. 1.08E8 for the first stage). All three variants
were described with competitive two-step models containing fixed
order, from n = 1 to n = 4 (Eqs. (2)–(4)).
Based on experimental data generated for up to 6 months, the re-
spective models were used for the simulation of stability for 3 years at
5 °C, 25 °C, 37 °C and 45 °C (Fig. 2, left). For the three variants, loss of
antigenicity exhibited first a rapid initial decay followed by a gradual
decrease (Fig. 2, left). The initial loss of antigenicity occurred with a
magnitude depending on incubation temperatures and variants. At 5 °C,
the first step of the degradation process reached an antigenicity re-
covery of around 85%, 90% and 72% after approximately 1 year,
0.5 year and 1.5 years for variants A, B and C, respectively (Fig. 2, left).
Then, the second step of the degradation process was predicted to be
more gradual and slowly reached an antigenicity recovery around 80%,
88% and 67% for variants A, B and C, respectively after 3 years at 5 °C
(Fig. 2, left). This two-step degradation behavior looked more pro-
nounced with higher storage temperatures and depending on variants.
As an example, 50% of antigenicity recovery was predicted to be
reached after 2 years at 25 °C for variant A, 1.5 years for variant C and
for variant B 3 years and beyond (data not shown), making this ob-
viously the more thermostable variant (Fig. 2, left). Incubation at high
temperatures (37 °C, 45 °C) rapidly induced loss of antigenicity of sev-
eral tens of percents (Fig. 2, left). This phenomenon was especially
pronounced for variant C, obviously the more thermosensitive variant.
Focusing on 5 °C, the experimental data generated for 6 months are
inside the predictive band determined by bootstrap analysis (CI 95% –
Fig. 2, right). The same was observed for 2.5 months data originating
from another batch (Fig. 2, right, open triangles). Similarly, further
experimental data (obtained later) not included in kinetic modeling,
collected after 1.8 years at 5 °C for all variants and after 2.8 years for
variant A, were mostly contained within the predictive bands (Fig. 2,
right). Only one of the three data points was outside the predictive band
for variant B. Quantitatively, predictions were performed for each
variant based on their respective two-step models (Eqs. (2)–(4)), as-
suming an aging of 1.8 years at 5 °C. Antigenicity recoveries, which
were predicted at 82%, 89% and 68% for variants A, B and C, respec-
tively, were later confirmed by real stability data, with approximately
five percentage points difference from experimental values (Table 1).
3.2. Vaccine stability predicted during excursions of temperature
3.2.1. Confronting vaccine stability predictions with real data obtained after
experimental temperature excursions for months at 25 °C and 37 °C
Based on kinetic modeling performed with 6 months forced de-
gradation data, selected ‘two-step’ models (Eqs. (2)–(4)) were applied
to temperature excursion to determine if antigenicity of variants could
be adequately predicted. Since the recommended vaccine storage con-
dition is 5 °C ± 3 °C, excursions to higher temperatures (9 months at
25 °C and 6 months at 37 °C) were applied to assess the degree of de-
gradation obtained at other arbitrarily selected storage conditions and
verify the simulated predictions. Antigenicity changes under these
temperature excursions were simulated on the basis of the selected
model. Gradual decreases were predicted for all variants during ex-
posure at elevated temperatures, except for variant B at 25 °C (Fig. 3).
79
European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84
In that particular case, no significant loss of antigenicity was predicted
when compared to sample kept at 5 °C (Fig. 3, left, middle), indicating a
lower thermal sensitivity for variant B, compared to variants A and C. In
the same condition, at 25 °C, antigenicity evolution were predicted in
two stages for variant A and C, with a rapid initial decrease followed by
a more gradual loss of antigenicity (Fig. 3, left, top and bottom). It
should be noted that variant C exhibited a faster initial degradation
profile when exposed at elevated temperature (Fig. 3, bottom), in-
dicating a particular thermosensitivity for this variant. Finally, ex-
posure at 37 °C induced pronounced loss of antigenicity for all variants,
with two-step profiles (Fig. 3, right). Quantitatively, antigenicity re-
coveries were predicted at 72%, 86% and 58% for variants A, B and C,
respectively when vaccine was exposed for 9 months at 25 °C (Table 2).
Those predictions were later confirmed experimentally, with less than
five percentage points difference from experimental values (Table 2). In
addition, predictions based on two-step kinetic models indicated that
exposure of vaccine for 6 months at 37 °C induced a loss of antigenicity
recoveries up to 44%, 60% and 24% for variants A, B and C, respec-
tively. Experimental data obtained by ELISA were included in the
predictive band (CI 95%), confirming agreement between antigenicity
recovery predictions and experimental data with not more than five
percentage points difference (Table 2).
3.2.2. Confronting vaccine stability predictions with real data obtained after
shipment of products at ambient temperatures
Recorded oscillations around ambient temperature indicated
minimum/maximum temperatures at 21 °C/27 °C during shipment to
Toronto and at 4 °C/48 °C during shipment to Barcelona (Fig. 4). Gra-
dual decreases of antigenicity for all the three variants were predicted
by the software (Eqs. (2)–(4)) depending on the recorded temperature
profiles (i.e. for shipment in Barcelona – Fig. 4). Predictions identified
variant B as the most stable and variant C as the least stable. Quanti-
tatively, antigenicity recoveries after the Lyon – Barcelona round trip
were predicted at 81%, 88% and 66% for variants A, B and C, respec-
tively (Table 3). Those predictions were later confirmed experimen-
tally, with approximate five percentage points difference from experi-
mental values (Table 3). Interestingly, similar antigenicity recoveries
were obtained after the return flight between Lyon and Toronto (82%,
89% and 67% for variants A, B and C, respectively – Table 3).
To assess the impact of maintaining freeze-dried products at am-
bient temperature for an extended period of time, we designed a case
study using simulations of storage at ambient temperature in Kenya
(Mombassa) and Argentina (Buenos Aires). The software mimicked the
real atmospheric temperature profiles in both locations, with oscilla-
tions between 23 °C and 32 °C for Mombassa, and between 4 °C and
30 °C for Buenos Aires (Fig. S1). The impact of such storage outside the
standard refrigerated conditions on vaccine antigenicity was predicted
for 2 years for variant A, B and C as a rapid initial loss of antigenicity,
followed by a gradual decrease (Fig. S1). The first degradation step
predicted a loss of antigenicity of variant A to approximately 79% and
72% in Buenos Aires and Mombassa, respectively (Fig. S1), and a gra-
dual decrease to approximately 72% and 64% after 2 years, in Buenos
Aires and Mombassa, respectively (Fig. S1). For comparison, anti-
genicity recovery of variant A kept under refrigerated conditions
(2–8 °C) for 2 years was predicted at 82%. For variant B, the first pre-
dicted degradation step amounted to approximately 90% in both
countries, followed by a gradual decrease to approximately 87% and
79% after 2 years in Buenos Aires and Mombassa, respectively (Fig. S1).
Finally, the predictions for variant C indicated a loss of antigenicity up
to approximately 65% in both countries for the first rapid step, and to
approximately 60% and 40% after 2 years in Buenos Aires and
Mombassa, respectively (Fig. S1).
4. Discussion
In this study, an experimental vaccine used as complex biomaterial
D. Clénet European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84

Fig. 2. Long-term prediction of vaccine antigenicity based on two-step kinetic models. Left column: Antigenicity predictions are displayed as lines for variants A, B and C for 3 years at
5 °C (blue), 25 °C (red), 37 °C (green) and 45 °C (pink). ELISA data used for kinetic modeling are displayed as filled circles. Right column: At 5 °C, antigenicity predictions (lines) are
shown with predictive band representing 95% CI (dotted lines). Experimental data used for kinetic modeling are displayed as filled circles. Additional experimental data obtained later in
time and not used for kinetic modeling are displayed as open circles, after 1.8 and 2.8 years for variant A and after 1.8 years for variants B and C. Experimental data obtained for another
vaccine batch are also displayed as open triangles.

Table 1
Comparison of antigenicity predicted by kinetic models (one-step and two-step) with experimental data (average of three independent ELISA results ± standard deviation) for vaccine
containing variants A, B and C kept under refrigerated conditions (2–8 °C). Differences between prediction and experimental antigenicity recovery are displayed (percentage point).

Time under refrigerated conditions Antigenicity recovery (%)

Two-step model One-step model (first-order)

Variant Experimental data ± SD (%) Prediction (%) Percentage point (%) Prediction (%) Percentage point (%)

1.8 years A 87 ± 2 82 −5 97 10
B 94 ± 8 89 −5 99 4
C 74 ± 8 68 −6 99 25

2.8 years A 79 ± 8 79 0 94 15

80
D. Clénet European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84

Fig. 3. Vaccine antigenicity predicted from kinetic models during temperature excursions outside the cold-chain. Left column: Antigenicity predictions are displayed as lines for variant
A, B and C for customized temperature excursions for 9 months in incubator set at 25 °C (red, solid line) and for 12 months under refrigerated conditions (2–8 °C, blue and dotted line).
Experimental data determined at the end of temperature excursions and not used for kinetic modeling are displayed as open red circles. Right column: Antigenicity predictions for
variants A, B and C for customized temperature excursions for 6 months in incubator set at 37 °C (red, solid line) and for 12 months under refrigerated conditions (2–8 °C, blue and dotted
line). Experimental data determined at the end of temperature excursions and not used for kinetic modeling are displayed as open red circles.

and containing three variants (variant A, B and C) was thermally
stressed in a forced degradation program comprising of four incubation
temperatures ranging between 5 °C and 45 °C. Vaccine antigenicity was
periodically evaluated by ELISA up to 6 months for each variant. No or
partial fit of experimental data was shown when zero- and first-order
kinetics were applied, leading to long-term antigenicity predictions
overestimated by up to 25%. Software such as ASAPprime® [26,27] and
ScienTek® [28] are not able to correctly describe the course of the
complex degradations of viruses studied here. AKTS-Thermokinetics
software [21] however, applies simultaneous combination of two-step

Table 2
Comparison of antigenicity predicted by two-step models with experimental data (average of three independent ELISA results ± standard deviation) for vaccine containing variants A, B
and C, subjected to customized temperature excursions at 25 °C and 37 °C. Differences between prediction and experimental antigenicity recovery are displayed (percentage point).

Shipments and storage conditions Antigenicity recovery (%)

Two-step model

Variant Experimental data ± SD (%) Prediction (%) Percentage point (%)

1 month in a cold chamber (2–8 °C), followed by 9 months in an incubator set at 25 °C, then A 68 ± 6 72 4
returned to the cold chamber for 2 months B 83 ± 14 86 3
C 56 ± 4 58 2

1 month in a cold chamber (2–8 °C), followed by 6 months in an incubator set at 37 °C, then A 44 ± 2 47 3
returned to the cold chamber for 4.5 months B 60 ± 4 62 2
C 24 ± 5 19 5

81
D. Clénet
Fig. 4. Vaccine antigenicity predicted from kinetic models during experimental shipment
of vaccine containing variants A, B and C performed at ambient temperature (red, solid
lines, top). Temperature oscillation (4–48 °C) suffered by vaccine during shipment is
displayed as red, solid line (bottom). For comparison, antigenicity predictions for samples
kept under refrigerated conditions (real temperature oscillation as blue line, bottom) are
displayed for variants A, B and C as blue and dotted line (top).
models enabling the consideration of an unlimited amount of kinetic
models, for the best description of the reaction course. When complex
degradation progresses were observed experimentally, it was demon-
strated that the application of wAIC and wBIC scores allowed determi-
nation of whether one or two-steps model was more likely to be correct
[20]. These criteria take into account not only the quality of fit, such as
the sum of residual squares (RSS), but also the number of experimental
points available and model parameters used. Additionally, they de-
termine which model is more likely to be correct, and quantify how
much more likely. Hence, possible errors of over-fitting were avoided,
and it was statistically confirmed that the two-step models better de-
scribe the loss of antigenicity of variants A, B and C. Thus, exhibiting
highest statistical scores (wAIC and wBIC), competitive two-step models
were selected for each variant (see Eqs. (2)–(4) and Tables S1–S3),
describing a rapid initial loss of antigenicity followed by a gradual
decrease.
Two-phase degradation profiles have already been described during
accelerated stability studies of several viruses (measles vaccine [29],
attenuated RBOK vaccine [30], canine distemper vaccine [31], ALVAC
poxvirus [14], and a freeze-dried live, attenuated virus vaccine [15]).
For these enveloped viruses, from different viral families, two steps in
the degradation process is also described (initial rapid drop followed by
a long gradual decrease phase), regardless of the readout used (in-
fectious tiers, antigenic titers etc.) [15,29,32]. One hypothesis put
forward in such forced virus degradation studies is that this biphasic
behavior entails the presence of at least two components with different
inactivation kinetics which are affected by the thermal treatment
[29,32] (e.g. epitope unfolding, capsid integrity loss, major virus con-
formational reorganization, key protein functionality loss or any
adaptive viral strategy). It is important to mention that many chemical
degradations and physical transitions occur during virus thermal
Table 3
Comparison of antigenicity predicted by two-step models with experimental data (average of 3 independent ELISA results ± standard deviation) for vaccine containing variants A, B and
C, subject to experimental shipments performed at ambient temperature. Differences between prediction and experimental antigenicity recovery are displayed (percentage point).
Storage conditions
Min. – Max. temp. exposure
during shipment (°C)
534 days in a cold chamber (2–8 °C) then 10 days travel at 21–27
ambient temperature during shipment from Lyon to Toronto
534 days in a cold chamber (2–8 °C) then 10 days travel at 4–48
ambient temperature during shipment from Lyon to
Barcelona
82
European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84
degradation. The selection of two-step reactions model is purely based
on the statistical analysis and goodness of fit with two major pathways
of phenomenological kinetic model of the reaction rate, which are most
likely built up from the sum of multiple reactions. At that point, it is
necessary to clearly underline that this result does not have a real
mechanistic basis, but it rather only depicts a phenomenological
mathematical model applied for fitting the reaction course. Kinetic
models mathematically described in Eqs. (2)–(4) for variant A, B and C
respectively, showed a higher contribution of the second step (gradual
decrease of antigenicity) as compared to the first step (rapid loss of
antigenicity). This is particularly marked for variant B, indicating that
stability is mainly governed by the second step with a contribution of
88.5%, as compared with the first step at 11.5% (Eq. (3). This very slow
rate of antigenicity decrease predicted by the kinetic model was found
to be in agreement with real stability data presented in this study
(Figs. 2 and 3, Tables 1 and 2). Variant B was shown to be the most
thermostable vaccine variant.
Conversely, results obtained demonstrated that kinetic models de-
scribing degradation changes of variants A, B and C are able to accu-
rately predict vaccine stability under any kind of storage conditions,
with just a few percentage points difference from experimental values
(Tables 1–3). First of all, long-term stability at usual recommended
storage temperature for freeze-dried product (5°C ± 3 °C) was pre-
dicted with success for all variants, with less than 6% as differences
between predicted and experimental values (Table 1). The use of the
Monte Carlo method, such as bootstrap analysis, provides accurate
predictions. It was already shown that the bootstrap CI gives a tight and
realistic estimate of the true CI [33]. One of the main interests of
bootstrap CI determination is that this approach takes into account the
scattering of experimental data without imposing a Gaussian distribu-
tion around predicted values, and that it does not need to know the
standard deviation of the analytical technique used. For the prediction
of degradation processes, the lower 95 percentile (predictive band at
95%), obtained with bootstrap analysis, can advantageously be used to
determine product shelf life, i.e. time to reach the lowest acceptable
specification value. This approach could improve applicable standards
used to determine shelf life, such as compliance with the acceptance
criterion as the time associated with the last measurement within the
specification [4].
Our results confirm that the impact of experimental shipments with
rupture in the cold chain and short time excursions above 40 °C can be
predicted with high accuracy. During shipments to Lyon-Barcelona and
Lyon-Toronto, predictions in term of antigenicity losses were confirmed
by experimental ELISA data, with no more than 6% expressed as dif-
ferences between predicted and experimental values (Table 3). Inter-
estingly, even if temperature excursions endured by freeze-dried pro-
ducts during our experimental shipments were different, including
different gaps between the minimum and the maximum exposure
temperatures, similar predictions were obtained regarding the level of
Antigenicity recovery (%)
Two-step model
Variant Experimental data ± SD (%) Prediction (%) Percentage point
(%)
A 86 ± 3 82 −4
B 90 ± 7 89 −1
C 69 ± 18 67 −2
A 80 ± 6 81 1
B 94 ± 4 88 −6
C 69 ± 9 66 −3
D. Clénet
degradation of variants. This observation could be explained by the fact
that shipments were performed after 1.5 years of storage at 5 °C. In-
deed, after this period of time, the initial rapid loss of antigenicity was
passed and samples were in the second step of the degradation process,
engaging thermo-resistant behavior with a gradual decrease of anti-
genicity. As shipments were executed during the stable phase, it could
be assumed that variants were less impacted by temperature oscillation
profiles, which could explain why antigenicity losses were similar de-
spite such different shipment conditions (by car to Barcelona and by
plane to Toronto). In addition, the simulation indicated that these
shipments, performed at ambient temperature, did not dramatically
impact antigenicity level when compared to shipments performed
under controlled refrigerated conditions at 5 °C. Thus, predicted anti-
genicity values only differed by 2.5% between experimental shipments
Lyon-Barcelona performed under controlled cold chain versus ambient
temperature. This result illustrated how kinetic-based modeling could
help to recommend storage conditions for vaccine shipments and to
evaluate the quality of vaccine under any temperature exposure.
Kinetic-based modeling approach presented in this work can ad-
vantageously be used to evaluate impact of storage temperature on
vaccine quality. It was illustrated by simulating antigenicity losses for
variant A, B and C during 2 years under different storage conditions (i.e.
ambient temperature in Mombassa, in Buenos Aires and under re-
frigerated conditions). Depending on variants, antigenicity losses
reached from 10% to 60% for a vaccine kept at ambient temperature,
whereas for vaccines kept in refrigerated conditions the losses were
limited to approximately 10% to 30%. One can take advantage of such
long-term predictions to justify, document and recommend the storage
temperature of products. Improvements in knowledge of vaccine sta-
bility behavior, and their thermosensitivity level could certainly have
economic impacts and financial gain.
It was stated that the inherent temperature sensitivity of vaccines is
particularly significant for live virus-based vaccines that require careful
control of the storage cold chain [34]. As the cold chain in the last mile
is particularly demanding, a controlled temperature chain (CTC) label
[35] was recently implemented to ensure that vaccines can withstand a
temperature excursion up to 40 °C for up to four days. The kinetic-based
approach described here could easily, and advantageously, be used to
predict degradation profiles of vaccines during storage at 40 °C for four
days. In addition, kinetic modeling is able to predict the impact of any
kind of temperature excursions, independent of their complexity pro-
vided that temperature was accurately recorded and used for predic-
tion. Hence, new ways of supply chain management could emerge. The
accurate prediction of changes in vaccine properties during shipment
would require appropriate controlled temperature monitoring which, in
turn, allows the use of real temperature profiles in the kinetic analysis.
Hence, electronic time-temperature devices would be able to inform, in
real time, the level of vaccine degradation provided that specific and
convenient kinetic models such as the ones described in this study could
be integrated into compact and connected devices, suggesting real-time
tracking of temperature excursions could be achieved. This could be of
major interest, knowing that maintaining the cold chain for storing and
transporting vaccines is often difficult for health workers in poor
countries, especially in remote areas. The WHO has stated that as many
as 25% of all vaccine products reach their destination in a degraded
state [3] and it was demonstrated that major providers expose vaccine
products to inappropriate temperatures [36]. Considering that the cold
chain, from production to use of vaccines, amounts to about 80% of the
total cost of vaccination programs [37], supply chain product man-
agement integrating kinetic modeling could significantly reduce fi-
nancial losses due to cold chain breaks.
Finally, the experimental and thermosensitive virus-based vaccine
used for this study, can be considered as one of the most complex
biomaterials with multi-step degradation processes. Therefore, it is ra-
tional to assume that stability behavior of other biologicals (e.g. pep-
tides, viral vectors, monoclonal antibodies, antibody-drug conjugates)
83
European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84
and adjuvants (e.g. emulsions, immunomodulators) could also be suc-
cessfully predicted. Such forecasts may significantly support the for-
mulation, process development and help in optimization of supply
chain procedures.
5. Conclusion
A kinetic-based modeling approach was studied to evaluate in-
activated virus-based vaccine stability, estimate impact of temperature
excursions outside recommended storage conditions and predict the
long-term stability of a vaccine containing three variants (variant A, B
and C), formulated as a freeze-dried product. Based on two-step kinetic
models, losses of virus antigenicity were accurately predicted for freeze-
dried product in isothermal long-term storage conditions, as well as in
real temperature profiles (i.e. temperature excursions during ship-
ments, maintaining vaccine products at ambient temperature in dif-
ferent climatic zones). Based on experimental data collected during
several months, these results uncover the potential of a kinetic-based
modeling approach to (i) predict shelf life of vaccines, (ii) recommend
appropriate storage conditions, and (iii) evaluate the impact of tem-
perature excursions (cold chain breaks), with an error amounting to
approximately five percentage points. From these results, it seems
reasonable to assume that development of recommendations, based on
advanced kinetics and stability modeling approaches can be a very
useful tool. It can also support the conclusions gained from procedures
described in international guidelines [8,38,39]. The application of ad-
vanced kinetic software may have a significant impact on shortening
time-consuming shelf life determinations, evaluating impact of tem-
perature excursions on vaccine efficacy, and evaluating the pharma-
ceutical product stability of vaccines, monoclonal antibodies, antibody-
drug conjugates, peptides and viral vectors.
Acknowledgments
The author would like to express his cordial thanks to Marie-Julie
Diana, Elisabeth Vialon and Romain Pizatto for ELISA analyses, Rayane
Seba, Laure Cigolotti and Manvi Hasija for their technical support
during shipments of samples and Aure Saulnier for fruitful discussions
and critical scientific review of the manuscript. Editorial assistance with
the preparation of the manuscript was provided by Rebecca Hornby,
inScience Communications, Springer Healthcare. Funding for this as-
sistance was provided by Sanofi Pasteur.
Didier Clenet is an employee of Sanofi Pasteur.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in the
online version, at https://doi.org/10.1016/j.ejpb.2018.01.005.
References
[1] D. Chen, D. Kristensen, Opportunities and challenges of developing thermostable
vaccines, Expert Rev. Vaccines 8 (2009) 547–557.
[2] D.T. Brandau, L.S. Jones, C.M. Wiethoff, J. Rexroad, C.R. Middaugh, Thermal sta-
bility of vaccines, J. Pharm. Sci. 92 (2003) 218–231.
[3] A. Black, E-Logistics in Cold Chain Management, <http://www.samedanltd.com/
magazine/11/issue/17/article/899> (accessed on 05/22/2017).
[4] World Health Organization, WHO/BS/06.2049, Guidelines on Stability Evaluation
of Vaccines, 2006.
[5] T.L. Schofield, Vaccine stability study design and analysis to support product li-
censure, Biologicals 37 (2009) 387–396.
[6] M. Hasija, L. Li, N. Rahman, S.F. Ausar, Forced degradation studies: an essential tool
for the formulation development of vaccines, Vaccine Dev. Ther. 3 (2013) 11–33.
[7] J. Zhang, E. Pritchard, X. Hu, T. Valentin, B. Panilaitis, F.G. Omenetto, D.L. Kaplan,
Stabilization of vaccines and antibiotics in silk and eliminating the cold chain, Proc.
Natl. Acad. Sci. U.S.A. 109 (2012) 11981–11986.
[8] International Conference on Harmonisation (ICH) Q1E, CPMP/ICH/420/02,
Evaluation of stability data, current Step 4 version dated 6 February 2003 by the
ICH Steering Committee, Evaluation of stability data, 2003.
[9] R. Capen, D. Christopher, P. Forenzo, C. Ireland, O. Liu, S. Lyapustina, J. O’Neill,
D. Clénet
N. Patterson, M. Quinlan, D. Sandell, J. Schwenke, W. Stroup, T. Tougas, On the
shelf life of pharmaceutical products, AAPS PharmSciTech 13 (3) (2012) 911–918.
[10] W. Stroup, M. Quinlan, Alternative shelf life estimation methodologies, JSM
Proceedings, Biopharmaceutical Section, American Statistical Association,
Alexandria, VA, 2010, pp. 2056–2066.
[11] M. Quinlan, W. Stroup, J. Schwenke, D. Christopher, Evaluating the performance of
the ICH guidelines for shelf-life estimation, J. Biopharm. Stat. 23 (2013) 881–896.
[12] V. Kayser, N. Chennamsetty, V. Voynov, B. Helk, K. Forrer, B.L. Trout, Evaluation of
a non-Arrhenius model for therapeutic monoclonal antibody aggregation, J. Pharm.
Sci. 100 (2011) 2526–2542.
[13] A. Oliva, A. Santovena, M. Llabres, J.B. Farina, Stability study of human serum
albumin pharmaceutical preparations, J. Pharm. Pharmacol. 51 (1999) 385–392.
[14] D. Clenet, F. Imbert, P. Probeck, N. Rahman, S.F. Ausar, Advanced kinetic analysis
as a tool for formulation development and prediction of vaccine stability, J. Pharm.
Sci. 103 (2014) 3055–3064.
[15] L. Hansen, J.V. Renterghem, R. Daoussi, C. Vervaet, J.P. Remon, T.D. Beer,
Spectroscopic evaluation of a freeze-dried vaccine during an accelerated stability
study, Eur. J. Pharm. Biopharm. 104 (2016) 89–100.
[16] R.K. Evans, D.K. Nawrocki, L.A. Isopi, D.M. Williams, D.R. Casimiro, S. Chin,
M. Chen, D.M. Zhu, J.W. Shiver, D.B. Volkin, Development of stable liquid for-
mulations for adenovirus-based vaccines, J. Pharm. Sci. 93 (2004) 2458–2475.
[17] M.A. Watzky, A.M. Morris, E.D. Ross, R.G. Finke, Fitting yeast and mammalianprion
aggregation kinetic data with the Finke-Watzky two-step model of nucleation and
autocatalytic growth, Biochemistry 47 (2008) 10790–10800.
[18] A.M. Morris, M.A. Watzky, R.G. Finke, Protein aggregation kinetics, mechanism and
curve-fitting: a review of the literature, Biochim. Biophys. Acta 1794 (2009)
375–397.
[19] J. Šesták, G. Berggren, Study of the kinetics of the mechanism of solid-state reac-
tions at increasing temperature, Thermochm. Acta 3 (1971) 1–13.
[20] B. Roduit, M. Hartmann, P. Folly, A. Sarbach, R. Baltensperger, Prediction of
thermal stability of materials by modified kinetic and model selection approaches
based on limited amount of experimental points, Thermochim. Acta 579 (2014)
31–39.
[21] AKTS-Thermokinetics Software, <http://www.akts.com>.
[22] J.B. Johnson, K.S. Omland, Model selection in ecology and evolution, Trends Ecol.
Evol. 19 (2004) 101–109.
[23] H. Akaike, A new look at the statistical model identification, IEEE Trans. Automatic
Control 19 (1974) 716–723.
[24] K.P. Burnham, D.R. Anderson, Model Selection and Multimodel Inference: A
Practical Information – Theoretic Approach, second ed., Springer, New York, 2002.
[25] D.K. Mishra, K.D. Dolan, L. Yang, Bootsrap confidence intervals for the kinetic
84
European Journal of Pharmaceutics and Biopharmaceutics 125 (2018) 76–84
parameters of gegradation of anthocyanins in grape pomase, J. Food Process Eng.
34 (2011) 1220–1233.
[26] K.C. Waterman, The application of the accelerated stability assessment program
(ASAP) to quality by design (QbD) for drug product stability, AAPS PharmSciTech.
12 (2011) 932–937.
[27] R. Waterman, J. Lewis, K.C. Waterman, Accelerated stability modeling for peptides:
a case study with bacitracin, AAPS PharmSciTech 18 (5) (2017) 1692–1698.
[28] ScienTek®, <http://www.stabilitysystem.com/statistical-data-analysis.htm>.
[29] L.M.C. Allison, G.F. Mann, F.T. Perkins, A.J. Zuckerman, An accelerated stability
test procedure for measles vaccines, J. Biol. Stand. 9 (1981) 185–194.
[30] J.C. Mariner, J.A. House, A.E. Sollod, C. Stem, M. Van Den End, C.A. Mebus,
Comparison of the effect of various chemical stabilizers and lyophilization cycles on
the thermostability of a vero cell-adapted rinderpest vaccine, Vet. Microbiol. 21 (3)
(1990) 195–209.
[31] P.M. Precausta, D. Simatos, M. Le Pemp, B. Devaux, F. Kato, Influence of residual
moisture and sealing atmosphere on viability of two freeze-dried viral vaccines, J.
Clin. Microbiol. 12 (4) (1980) 483–489.
[32] F. Higashikawa, L.J. Chang, Kinetic analyses of stability of simple and complex
retroviral vectors, Virology 280 (1) (2001) 124–131.
[33] D.K. Mishra, K.D. Dolan, L. Yang, Bootstrap confidence intervals for the kinetics
parameters of degradation of anthocyanins in grape pomace, J. Food Process Eng.
34 (2011) 1220–1233.
[34] U. Kartoglu, J. Milstien, Tools and approaches to ensure quality of vaccines
throughout the cold chain, Expert Rev. Vaccines 13 (7) (2014) 843–854.
[35] S. Zipurskya, M. Harouna Djingarey, J.C. Lodjo, L. Olodo, S. Tiendrebeogo,
O. Ronveaux, Benefits of using vaccines out of the cold chain: Delivering Meningitis
A vaccine in a controlled temperature chain during the mass immunization cam-
paign in Benin, Vaccine 32 (2014) 1431–1435.
[36] D.R. Levinson, Vaccines for children program: vulnerabilities in vaccine manage-
ment. Department of Health and Human Services, Office of Inspector General OEI-
04-10-00430 (2012), available from: http://www.oig.hhs.gov/oei/reports/oei-04-
10-00430.pdf.
[37] Making vaccines not need refrigeration, http://blog.pnas.org/2013/04/making-
vaccines-not-need-refrigeration/, posted on April 16 (2013) by PNAS (accessed on
05/22/2017).
[38] International Conference on Harmonisation (ICH) Q5C, CPMP/ICH/138/95,
Quality of Biotechnological Products: Stability Testing of Biotechnological/
Biological Products, 1996.
[39] European Medicines Agency, EMA/CHMP/BWP/354898/2008, Guideline on the
requirements for quality documentation concerning biological investigational
medicinal products in clinical trials, 2012.