Social Influence Bias: A Randomized Experiment
Lev MuchnikScience 341
This copy is for your personal, non-commercial use only.
, you can order high-quality copies for your
If you wish to distribute this article to others
colleagues, clients, or customers by
can be obtained by
Permission to republish or repurpose articles or portions of articles
following the guidelines
The following resources related to this article are available online at
www.sciencemag.org (this information is current as of
August 8, 2013
including high-resolution figures, can be found in the online
Updated information and services,
on August 8, 2013
version of this article at:
can be found at:
, 20 of which can be accessed free:
cites 41 articles
(print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. 2013 by the American Association for the Advancement of Science; all rights reserved. The title
registered trademark of AAAS.
BVOCs. The contribution from hundreds of indi-
3. M. O. Andreae, P. J. Crutzen, Science 276, 1052–1058 (1997).
16. See supplementary materials on Science Online.
vidually minor compounds is important. Future
4. M. S. Jang, N. M. Czoschke, S. Lee, R. M. Kamens, Science
17. A. Lee et al., J. Geophys. Res. Atmos. 111, 21
298, 814–817 (2002).
research is required to determine if this is also
5. A. H. Goldstein, I. E. Galbally, Environ. Sci. Technol. 41,
true for ecosystems that emit larger quantities of
Acknowledgments: This work was supported by the
isoprene and/or monoterpenes.
6. M. R. Kurpius, A. H. Goldstein, Geophys. Res. Lett. 30,
Netherlands Organization for Scientific Research (NWO)
The minor species with lower fluxes and con-
under the ALW-Middelgroot program (grant 834.08.002),
7. P. Di Carlo et al., Science 304, 722–725 (2004).
the California Air Resources Board (Award 06-329),
centrations are not in current BVOC emission
8. R. Holzinger, A. Lee, K. T. Paw, A. H. Goldstein, Atmos.
the Citrus Research Board, and the University of Utrecht
models, but their sizes, chemical formulae, and
Chem. Phys. 5, 67–75 (2005).
short stay fellowship program. We thank the Gorden
sum suggest that they should be important for
9. T. M. Ruuskanen et al., Atmos. Chem. Phys. 11, 611–625
Ranch for hosting this study and for the assistance provided
SOA formation and regional photochemistry.
by the University of California Lindcove Research and
10. M. Hallquist et al., Atmos. Chem. Phys. 9, 5155–5236
Their presence in the atmosphere may also ac-
count for a sizable amount of the missing OH
11. T. Karl et al., Science 330, 816–819 (2010).
chemical reactivity and O
3 chemical loss observed
12. J.-H. Park et al., Atmos. Chem. Phys. 13, 1439–1456
in plant environments.
13. A. Jordan et al., Int. J. Mass Spectrom. 286,
Materials and Methods
References and Notes
14. M. Graus, M. Müller, A. Hansel, J. Am. Soc. Mass
1. A. Guenther et al., J. Geophys. Res. Atmos. 100,
Spectrom. 21, 1037–1044 (2010).
15. R. Holzinger, A. Kasper-Giebl, M. Staudinger, G. Schauer,
2. W. L. Chameides, R. W. Lindsay, J. Richardson,
T. Rockmann, Atmos. Chem. Phys. 10, 10111–10128
11 January 2013; accepted 2 July 2013
C. S. Kiang, Science 241, 1473–1475 (1988).
Social Influence Bias:
ment on true quality is nearly impossible in ob-servational data (21–27). For example, popularproducts may be popular because of the irrational
A Randomized Experiment
effect of past positive ratings, or alternatively, the
on August 8, 2013
best products may become popular because theyare of the highest quality. We must distinguish
Lev Muchnik,1 Sinan Aral,2* Sean J. Taylor3
these explanations to determine the extent towhich social influence creates irrational herding.
Our society is increasingly relying on the digitized, aggregated opinions of others to
We therefore designed and analyzed a large-
make decisions. We therefore designed and analyzed a large-scale randomized experiment
scale randomized experiment to quantify the
on a social news aggregation Web site to investigate whether knowledge of such aggregates
effects of social influence on users' ratings and
distorts decision-making. Prior ratings created significant bias in individual rating behavior,
discourse on a social news aggregation Web
and positive and negative social influences created asymmetric herding effects. Whereas
site, where users contribute news articles and
negative social influence inspired users to correct manipulated ratings, positive social
discuss them. Users of the site that we studied
influence increased the likelihood of positive ratings by 32% and created accumulating
write comments in response to posted articles,
positive herding that increased final ratings by 25% on average. This positive herding
and other users can then "up-vote" or "down-vote"
was topic-dependent and affected by whether individuals were viewing the opinions of friends
these comments, yielding an aggregate current
or enemies. A mixture of changing opinion and greater turnout under both manipulations
rating for each posted comment equal to the num-
together with a natural tendency to up-vote on the site combined to create the herding
ber of up-votes minus the number of down-votes.
effects. Such findings will help interpret collective judgment accurately and avoid social
Users do not observe the comment scores before
influence bias in collective intelligence in the future.
clicking through to comments—each impression of
a comment is always accompanied by that com-
Werelyonratingscontributedbyothers Collective intelligence has recently been her- ment'scurrentscore,tyingthecommenttothe
to make decisions about which hotels,
alded as a harbinger of accelerated human po-
score during users' evaluation—and comments
books, movies, political candidates,
tential (6). But, social influence on individuals'
are not ordered by their popularity, mitigating
news, comments, and stories are worth our time
perceptions of quality and value could create
selection bias on high (or low) rated comments.
and money (1). Given the widespread use and
herding effects that lead to suboptimal market
Similar scoring mechanisms are widely used on
economic value of rating systems (2–4), it is im-
outcomes (7, 8); rich-get-richer dynamics that
the Web to reward users for supplying insightful
portant to consider whether they can successfully
exaggerate inequality (9–12); a group think men-
or interesting analysis, while penalizing those post-
harness the wisdom of crowds to accurately ag-
tality that distorts the truth (13); and measurable
ing irrelevant, redundant, or low-quality comments.
gregate individual information. Do they produce
disruptions in the wisdom of crowds (14). If
The vast majority of interuser relations occur on
useful, unbiased, aggregate information about
perceptions of quality are biased by social influ-
the Web site, in contrast to Web sites whose mem-
the quality of the item being rated? Or, as sug-
ence, attempts to aggregate collective judgment
bers also interact offline. The data therefore pro-
gested by the experiments of Salganik et al. (5),
and socialize choice could be easily manipulated,
vide a unique opportunity to comprehensively
are outcomes path dependent, yielding different
with dramatic consequences for our markets, our
study social influence bias in rating behavior.
aggregate ratings for items of equivalent quality?
politics, and our health.
Over 5 months, 101,281 comments submitted
The recent availability of population-scale
on the site were randomly assigned to one of
data sets on rating behavior and social commu-
three treatment groups: up-treated, down-treated,
1School of Business Administration, The Hebrew University of
nication enable novel investigations of social
or control. Up-treated comments were artificially
Jerusalem, Mount Scopus, Jerusalem, 91905 Israel. 2MIT Sloan
School of Management, 100 Main Street, Cambridge, MA
–20). Unfortunately, our under-
given an up-vote (a +1 rating) upon the com-
20142, USA. 3NYU Stern School of Business, 44 West 4th
standing of the impact of social influence on col-
ment's creation, whereas down-treated comments
Street, New York, NY 10012, USA.
lective judgment is limited because empirically
were given a down-vote (a –1 rating) upon the
*Corresponding author. E-mail: email@example.com
distinguishing influence from uninfluenced agree-
comment's creation. Users were unaware of the
www.sciencemag.org SCIENCE VOL 341 9 AUGUST 2013
Probability to Up-vote
Probability to Down-vote
Fig. 1. Effect of manipulation on voting behavior. The positively manipulated
rater random effects. (C) The mean final scores of positively manipulated, nega-
treatment group (up-treated), the negatively manipulated treatment group (down-
tively manipulated, and control group comments are shown with 95% confidence
treated), and the control group (dotted line) are shown. The probabilities to up-vote
intervals inferred from Bayesian linear regression of the final comment score with
(A) and down-vote (B) positively manipulated, negatively manipulated, and control
commenter random effects. Final mean scores on this Web site are measured as
group comments are shown by the first unique viewer; 95% confidence intervals
the number of up-votes minus the number of down-votes. We discuss the impli-
are inferred from Bayesian logistic regression with commenter, rater, and commenter-
cations of this measurement in greater detail in the supplementary materials.
on August 8, 2013
manipulation and unable to trace votes to any
Fig. 2. Effects of topic on
particular user. As a result of the randomization,
herding. Mean final scores
comments in the control and treatment groups
of positively manipulated and
were identical in expectation along all dimen-
control group comments are
sions that could affect users' rating behavior
shown with 95% confidence
except for the current rating. This manipulation
intervals inferred from Bayesian
created a small random signal of positive or neg-
linear regression of the fi-
ative judgment by prior raters for randomly se-
nal comment score with com-
menter random effects across
lected comments that have the same quality in
the seven most active topic
expectation, enabling estimates of the effects of
categories on the site, or-
social influence holding comment quality and all
dered by the magnitude of
other factors constant. The 101,281 experimental
the difference between the
comments (of which 4049 were positively treated
mean final score of positive-
and 1942 were negatively treated to reflect the
ly manipulated comments
natural proportions of up- and down-votes on
and the mean final score of
the site) were viewed more than 10 million times
control comments in each
and rated 308,515 times by subsequent users.
We may sample users multiple times, and
nonrandom heterogeneity may exist in users'comment-generating process, their selection ofcomments to vote on, and in relationships be-
tween commenters and raters. We therefore es-timated hierarchical Bayesian models of votingbehavior and specified commenter, rater, andcommenter-rater pair random effects; i.e., theconfidence intervals are based on repeated re-
all users' ratings for comments in the three groups
the control group (P = 1.0 × 10–6) (Fig. 1A). Up-
sampling, creating a distribution of parameter
treated comments were not down-voted signifi-
estimates from which the 95% confidence bands
Figure 1A shows the immediate up-vote
cantly more or less frequently than the control
are derived (see materials and methods in the
and down-vote probabilities for the first viewer
group, so users did not tend to correct the upward
of comments in each of the three categories.
manipulation. In the absence of a correction, pos-
We first compared the probabilities that com-
Up-votes were 4.6 times as common as down-
itive herding accumulated over time.
ments in each group would be up-voted or down-
votes on this site, with 5.13% of all comments
The positive manipulation created a posi-
voted by the first viewer after the manipulation.
receiving an up-vote by the first viewer of the
tive social influence bias that persisted over our
These probabilities measure the immediate effect
comment and only 0.82% of comments receiv-
5-month observation window, generating accu-
of current ratings on users' rating behavior. We
ing a down-vote by the first viewer. The up-vote
mulating herding effects that increased comments'
then analyzed comments' long-run ratings dis-
treatment significantly increased the probabil-
final mean ratings by 25% relative to the final
tributions and final mean scores by aggregating
ity of up-voting by the first viewer by 32% over
mean ratings of control group comments (c2 test;
9 AUGUST 2013 VOL 341 SCIENCE www.sciencemag.org
categories with significant differences in control
and treatment ratings and those with no significantdifferences had similar levels of activity. There wasno significant negative herding in any category.
Friendship also moderated the impact of
social influence on rating behavior (Fig. 3, Aand B). The Web site has a feature wherebyusers can indicate that they "like" or "dislike"other users, forming "friends" and "enemies"social preference graphs. Unsurprisingly, friendsof the commenter were more likely to up-vote acomment than those who disliked him or her(9.2% versus 2.7%, c2 test; P = 2.7 × 10–49)[compare the average (dotted line) in Fig. 3A tothe average (dotted line) in Fig. 3B]. Friendsalso tended to herd on current positive ratings
Fig. 3. Effects of friendship on rating behavior. The figure shows the probability of a friend (A) and
(friends' probability to up-vote a positively ma-
enemy (B) of the commenter to up-vote positively manipulated, negatively manipulated, and control group
nipulated comment: 0.122 versus friends' prob-
comments. Friends and enemies are defined as users who had previously clicked a button on the Web site
ability to up-vote a control comment: 0.092; c2
labeling the commenter as someone they "liked" or "disliked," respectively.
test; P = 1.4 × 10–2) and to correct commentswith negatively manipulated ratings (friends'probability to up-vote a negatively manipulatedcomment: 0.176 versus friends' probability to
up-vote a control comment: 0.092; c2 test; P =4.0 × 10–12) (Fig. 3A), mirroring the cooperationfound in human social networks (28). By con-
on August 8, 2013
trast, enemies of the commenter were not sus-ceptible to social influence. Enemies' ratings wereunaffected by our treatments, possibly becauseof the small sample of potential first ratings byenemies (though there are a substantial numberof enemies in the community, they are less active)(Fig. 3B).
Finally, social influence in ratings behavior
did not affect discourse in our setting during the5-month observation period. Neither the positive
nor the negative manipulation affected the aver-age number of replies (Fig. 4A) or the averagedepth of the discussion tree created in response to
Fig. 4. Effects on subsequent discourse. The figure displays the average number of responses (A)
a comment (Fig. 4B).
and the average depth of the discussion tree (B) that developed in response to positively manipulated,
Several data-generating processes could ex-
negatively manipulated, and control group comments; 95% confidence intervals are inferred from
plain our findings. A selection effect could in-
Bayesian linear regressions with author random effects.
spire different populations of voters to turn out
to rate the item (selective turnout)—for example,
P = 2.3 × 10–11) (Fig. 1C), and Kolmogorov-
10–3) (Fig. 1B). However, this effect was offset
if the negative manipulation inspired voters who
Smirnov (K-S) tests showed that the final score
by a larger correction effect. The probability that
tend to down-vote (negative voters) to vote in
distribution of up-treated comments was signif-
a down-treated comment would subsequently be
higher proportion. Alternatively, prior ratings
icantly shifted toward higher scores (K-S test
up-voted was 0.099, whereas the probability that
could bias users' voting behavior by changing
statistic: 0.083; P = 1.2 × 10–23). Comments in
a control comment would be up-voted was sig-
their opinions about comment quality and there-
the up-treated group were also significantly more
nificantly different at 0.054 (c2 test; P = 1.0 ×
fore their votes (opinion change). We analyzed
likely than those in the control group to accu-
10–30) (Fig. 1A). This correction neutralized so-
changes in turnout (the likelihood of rating rather
mulate exceptionally high scores. Up-treated com-
cial influence in the ratings of negatively manip-
than just viewing comments) and changes in pos-
ments were 30% more likely to reach or exceed
ulated comments, and their final mean ratings
itivity (the proportion of positive ratings) across
a score of 10 (6.4% versus 4.9% in the control
were not statistically different from the control
subgroups in our study population to identify var-
group, c2 test; P = 2.0 × 10–5). The small ma-
group's final mean ratings (Fig. 1C).
iance in our results explained by turnout effects
nipulation of a single random up-vote when the
We next estimated changes in the final mean
and opinion change, respectively. We divided
comment was created resulted in significant-
score for up-treated comments compared to con-
raters and commenters, on the basis of their rating
ly higher accumulated ratings due to social
trol comments in the seven most active topic
history, along four subgroup dimensions by their
categories on the Web site. We found significant
positivity (proportion of positive votes), the com-
Positive and negative social influence created
positive herding effects for comment ratings in
menters' quality (prior scores of their comments),
asymmetric herding effects. The probability of
"politics," "culture and society," and "business,"
the frequency with which a rater rated a particular
down-treated comments receiving subsequent
but no detectable herding behavior for comments
commenter, and whether raters were friends or
down-votes was 0.014, whereas the probability
in "economics," "IT," "fun," and "general news"
enemies with commenters. We then compared treat-
of control comments receiving a down-vote was
(Fig. 2). These differences are not due to the fre-
ment effects with expected voting behavior in
0.007; a significant difference (c2 test; P = 1.1 ×
quency of commenting in these categories, as
these subgroups. Our analysis revealed several
www.sciencemag.org SCIENCE VOL 341 9 AUGUST 2013
Fig. 5. Effects on turnout versus positivity. The figure
displays treatment effects of the negative and positive
manipulation on turnout (the likelihood of rating) andpositivity (the proportion of positive votes) for negative(positive) raters, defined as active raters (with at least
100 votes on control comments) who display less than
(greater than) the median proportion of positive votes
on control comments. Results displayed in red are sta-tistically significant at the 95% level, whereas results
displayed in gray are not. (A) Treatment effect of thenegative manipulation on positivity; (B) treatment effectof the positive manipulation on positivity; (C) treatment
effect of the negative manipulation on turnout; and (D)treatment effect of the positive manipulation on turnout.
In each panel, the first estimate is the treatment effect
on negative raters, the second estimate is the treatment
effect on positive raters, and the third estimate is theratio of the treatment effects on negative/positive raters.
The treatment effects displayed are odds ratios of the
treatment effect on the treated compared to the control
Treatment Effect (odds ratio)
for negative and positive raters. The negative manipu-lation created countervailing opinion change for positive
and negative raters, increasing the positivity of negativeraters and decreasing the positivity of positive raters,whereas both treatments increased turnout uniformly
across both subgroups.
on August 8, 2013
behavioral mechanisms that together explain the
bined with a general preference for positivity
References and Notes
experimental results (see "Estimating Behavioral
on the site, created a tendency toward positive
1. J. A. Chevalier, D. Mayzlin, J. Mark. Res. 43, 345–354
Mechanisms" in the supplementary text).
ratings under both treatments. Prior work has
2. D. G. Pope, J. Health Econ. 28, 1154–1165
First, both treatments increased turnout
shown a preference for positivity in ratings in
(by 82 and 28% for first viewers of down- and
other contexts (1), which suggests that these re-
3. C. Dellarocas, Manage. Sci. 49, 1407–1424
up-treated comments, respectively), but neither
sults will generalize. Together, these findings sug-
created differential turnout for different types of
gest that a mixture of changing opinion and the
4. L. Cabral, A. Hortacsu, "The dynamics of seller reputation:
Theory and evidence from eBay." National Bureau
voters. This suggests that selection, or differential
natural tendency to up-vote together with greater
of Economic Research, working paper w10363
turnout by voter type (e.g., selecting different
turnout under both manipulations, combine to
proportions of positive or negative voters, or fre-
create the herding effects we see (see "Estimat-
5. M. J. Salganik, P. S. Dodds, D. J. Watts, Science 311,
quent or infrequent voters), cannot explain our
ing Behavioral Mechanisms" in the supplemen-
6. A. W. Woolley, C. F. Chabris, A. Pentland, N. Hashmi,
results. Second, we found statistically significant
T. W. Malone, Science 330, 686–688 (2010).
opinion change in two of the four subgroup di-
Our results demonstrate that whereas positive
7. S. Bikhchandani, D. Hirshleifer, I. Welch, J. Polit. Econ.
mensions. The positive treatment created a sys-
social influence accumulates, creating a tendency
100, 992 (1992).
tematic increase in the proportion of positive
toward ratings bubbles, negative social influence
8. D. Hirshleifer, in The New Economics of Human
ratings for raters with little prior experience rating
is neutralized by crowd correction (29). Our find-
Behaviour, M. Tommasi, K. Ierulli, Eds. (Cambridge Univ.
Press, Cambridge, UK, 1995), pp. 188–215.
the particular commenter whose comment was
ings suggest that social influence substantially
9. R. K. Merton, Social Theory and Social Structure (Free
manipulated (a 7% increase in the ratio of pos-
biases rating dynamics in systems designed to
Press, New York, 1968).
itive to negative votes and a 50% increase in the
harness collective intelligence. Future research
10. H. A. Simon, Biometrika 42, 425 (1955).
probability of up-votes for these raters) and no
that further explores the mechanisms driving
11. A.-L. Barabási, R. Albert, Science 286, 509–512
decrease in positivity in any subgroup. This im-
individual and aggregate ratings—especially in
12. J.-P. Onnela, F. Reed-Tsochas, Proc. Natl. Acad. Sci. U.S.A.
plies that positive opinion change explains part
in vivo social environments—will be essential to
107, 18375–18380 (2010).
of the variation in positive herding. The nega-
our ability to interpret collective judgment ac-
13. I. L. Janis, Groupthink: Psychological Studies of Policy
tive treatment, by contrast, created countervailing
curately and to avoid social influence bias in
Decisions and Fiascoes (Cengage Learning, Boston, MA,ed. 2, 1982).
opinion change for positive and negative raters,
collective intelligence. We anticipate that our ex-
14. J. Lorenz, H. Rauhut, F. Schweitzer, D. Helbing, Proc.
cancelling out any aggregate-level evidence of
periment will inspire more sophisticated analyses
Natl. Acad. Sci. U.S.A. 108, 9020–9025 (2011).
opinion change (Fig. 5). This in part explains
of social influence bias in electoral polling, stock
15. C. Danescu-Niculescu-Mizil, G. Kossinets, J. Kleinberg,
why we find no aggregate trend in either direc-
market prediction, and product recommendation
L. Lee, How opinions are received by online communities:A case study on amazon.com helpfulness votes, in
tion for the negative treatment. Third, both treat-
and that these results will be used to adapt rating
Proceedings of the 18th International Conference on
ments created a uniform increase in turnout across
and review technologies to account for social
World Wide Web (Association for Computing Machinery,
voter types. This overall increase in turnout, com-
influence bias in their outputs.
New York, 2009), pp. 141–150.
9 AUGUST 2013 VOL 341 SCIENCE www.sciencemag.org
16. C. A. Hidalgo, R. A. Castro, C. Rodriguez-Sickert, New J.
23. S. Aral, Mark. Sci. 30, 217–223 (2011).
data that we analyzed is publically available, except the
Phys. 8, 52 (2006).
24. S. Aral, D. Walker, IEEE Intell. Syst. 26, 91–96
treatment assignment and the random identifier that links a
17. E. Bakshy, B. Karrer, L. A. Adamic, Social influence and
deidentified user to a vote, and therefore does not contain any
the diffusion of user created content, in Proceedings of
25. S. Aral, D. Walker, Manage. Sci. 57, 1623–1639
other information that cannot be obtained by crawling the
the 10th ACM Conference on Electronic Commerce
Web site. The randomized testing that the Web site performed
(Association for Computing Machinery, New York, 2009),
26. S. Aral, D. Walker, Science 337, 337–341 (2012).
is covered by the Web site's terms of service. Opt-in
27. C. R. Shalizi, A. C. Thomas, Sociol. Methods Res. 40,
permissions were granted by the users when they registered
18. J. Leskovec, A. Singh, J. Kleinberg, Patterns of influence
for the Web site. No data on nonregistered users was collected.
in a recommendation network, in Advances in Knowledge
28. C. L. Apicella, F. W. Marlowe, J. H. Fowler, N. A.
Discovery and Data Mining, Lecture Notes in Computer
Christakis, Nature 481, 497–501 (2012).
Science (Springer, Berlin, Heidelberg, 2006), vol. 3918,
29. D. Centola, R. Willer, M. Macy, Am. J. Sociol. 110,
19. F. Wu, B. A. Huberman, Internet and Network Economics
Materials and Methods
5385, 334–341 (2008).
Acknowledgments: We thank D. Eckles and D. Watts for
20. N. Archak, A. Ghose, P. G. Ipeirotis, Manage. Sci. 57,
invaluable discussions. This work was supported by a Microsoft
Research faculty fellowship (S.A.) and by NSF Career Award
21. C. F. Manski, Rev. Econ. Stud. 60, 531–542
0953832 (S.A.). The research was approved by the New York
University institutional review board. There are legal obstacles
22. S. Aral, L. Muchnik, A. Sundararajan, Proc. Natl. Acad.
to making the data available and revealing the name of the
14 May 2013; accepted 16 July 2013
Sci. U.S.A. 106, 21544–21549 (2009).
Web site, but code is available upon request. All of the user
Pluripotent Stem Cells Induced
ical substitutes of Oct4 in the absence of trans-genes. We found that VC6T plus FSK (VC6TF)induced some GFP-positive clusters expressing
from Mouse Somatic Cells
E-cadherin, a mesenchyme-to-epithelium transi-tion marker, reminiscent of early reprogramming
by Small-Molecule Compounds
by transcription factors (10, 11) (Fig. 1C and fig.
S3). However, the expression of Oct4 and Nanogwas not detectable, and their promoters remained
Pingping Hou,1* Yanqin Li,1* Xu Zhang,1,2* Chun Liu,1,2* Jingyang Guan,1* Honggang Li,1*
on August 8, 2013
hypermethylated, suggesting a repressed epige-
Ting Zhao,1† Junqing Ye,1,2† Weifeng Yang,3† Kang Liu,1† Jian Ge,1,2† Jun Xu,1† Qiang Zhang,1,2†
netic state (fig. S3).
Yang Zhao,1‡ Hongkui Deng1,2‡
To identify small molecules that facilitate late
reprogramming, we used a doxycycline (DOX)–
Pluripotent stem cells can be induced from somatic cells, providing an unlimited cell
inducible Oct4 expression screening system,
resource, with potential for studying disease and use in regenerative medicine. However,
adding DOX only in the first 4 to 8 days (6). Small-
genetic manipulation and technically challenging strategies such as nuclear transfer
molecule hits, including several cAMP agonists
used in reprogramming limit their clinical applications. Here, we show that pluripotent
(FSK, Prostaglandin E2, and Rolipram) and epi-
stem cells can be generated from mouse somatic cells at a frequency up to 0.2% using a
genetic modulators [3-deazaneplanocin A (DZNep),
combination of seven small-molecule compounds. The chemically induced pluripotent
5-Azacytidine, sodium butyrate, and RG108], were
stem cells resemble embryonic stem cells in terms of their gene expression profiles, epigenetic
identified in this screen (fig. S4 and table S1B).
status, and potential for differentiation and germline transmission. By using small molecules,
To achieve complete chemical reprogramming
exogenous "master genes" are dispensable for cell fate reprogramming. This chemical
without the Oct4-inducible system, these small
reprogramming strategy has potential use in generating functional desirable cell types
molecules were further tested in the chemical re-
for clinical applications.
programming of OG-MEFs without transgenes.
When DZNep was added 16 days after treatment
Pluripotentstemcells,suchasembryonic ized. Moreover, their effects on inhibiting and withVC6TF(VC6TFZ),GFP-positivecellswere Downloaded from
stem cells (ESCs), can self-renew and
activating the function of specific proteins are
obtained more frequently by a factor of up to 65
differentiate into all somatic cell types.
often reversible and can be finely tuned by varying
than those treated with VC6TF, forming compact,
Somatic cells can be reprogrammed to become
the concentrations. Here, we identified small-
epithelioid, GFP-positive colonies without clear-
pluripotent via nuclear transfer into oocytes or
molecule combinations that were able to drive the
cut edges (Fig. 1, D and E, and fig. S5). In these
through the ectopic expression of defined factors
reprogramming of mouse somatic cells toward
cells, the expression levels of most pluripotency
(1–4). However, exogenous pluripotency-associated
marker genes were elevated but were still lower
factors, especially Oct4, are indispensable for es-
To identify small molecules that facilitate cell
than in ESCs, suggesting an incomplete repro-
tablishing pluripotency (5–7), and previous repro-
reprogramming, we searched for small molecules
gramming state (fig. S6). After switching to 2i-
gramming strategies have raised concerns regarding
that enable reprogramming in the absence of Oct4
medium with dual inhibition (2i) of glycogen
the clinical applications (8, 9). Small molecules
using Oct4 promoter-driven green fluorescent pro-
synthase kinase-3 and mitogen-activated protein
have advantages because they can be cell perme-
tein (GFP) expression (OG) mouse embryonic
kinase signaling after day 28 posttreatment, cer-
able, nonimmunogenic, more cost-effective, and
fibroblasts (MEFs), with viral expression of Sox2,
tain GFP-positive colonies developed an ESC-like
more easily synthesized, preserved, and standard-
Klf4, and c-Myc. After screening up to 10,000 small
morphology (domed, phase-bright, homogeneous
molecules (table S1A), we identified Forskolin
with clear-cut edges) (Fig. 1F) (12, 13). These col-
(FSK), 2-methyl-5-hydroxytryptamine (2-Me-
onies could be further cultured for more than 30
College of Life Sciences and Peking-Tsinghua Center for Life
Sciences, Peking University, Beijing 100871, China. 2School
5HT), and D4476 (table S1B) as chemical "sub-
passages, maintaining an ESC-like morphology
of Chemical Biology and Biotechnology, Peking University
stitutes" for Oct4 (Fig. 1, A and B, and figs. S1
(Fig. 1, G and H). We refer to these 2i-competent,
Shenzhen Graduate School, Shenzhen 518055, China. 3Beijing
and S2). Previously, we had developed a small-
ESC-like, and GFP-positive cells as chemically
Vitalstar Biotechnology Co., Ltd., Beijing 100012, China.
molecule combination "VC6T" [VPA, CHIR99021
induced pluripotent stem cells (CiPSCs).
*These authors contributed equally to this work.
(CHIR), 616452, Tranylcypromine], that enables
Next, we optimized the dosages and treatment
†These authors contributed equally to this work.
reprogramming with a single gene, Oct4 (6). We
duration of the small molecules and were able
‡Corresponding author. E-mail: firstname.lastname@example.org(H.D.); email@example.com (Y.Z.)
next treated OG-MEFs with VC6T plus the chem-
to generate 1 to 20 CiPSC colonies from 50,000
www.sciencemag.org SCIENCE VOL 341 9 AUGUST 2013
January - March 2013 If you met Leanne Hosking, you'd probably never guess she is battling a neuromuscular disease. She is a lively 32 year old woman who works full-time, exercises daily, and enjoys traveling the world. But this wasn't the case three 1000 John R, Suite 111 years ago when she suffered from severe muscle weakness and fatigue. She couldn't drive at night and found it difficult to complete everyday tasks like getting
SMITHS DETECTION Technical Information Application Brief AB-055LIGHT MICROSCOPY, FT-IR MICROSPECTROSCOPY AND X-RAYDIFFRACTION:COMPLIMENTARY TECHNIQUES FOR SOLID STATE DRUG ANALYSIS The analysis of active pharmaceutical ingredients (APIs) in thesolid state is critically important in drug development andquality assurance, because many drugs are dispensed assolids. Indeed, the APIs may exist in different polymorphiccrystalline or amorphous states, as hydrates or solavates, oreven as salts with various counterions. When a drug substanceis formulated, knowing and controlling the composition andstate of the API is necessary to produce an effective product.Additionally, there are legal implications related to patentclaims and intellectual property that can translate into millions