although these findings require further explo-ration in larger clinical cohorts. Additional studiesof clinical response and resistance to immunecheckpoint inhibitorsmay benefit from integratingexome and transcriptome sequencing data toinform the relative contributions of tumor im-munogenicity and host immune infiltration indetermining clinical benefit.

REFERENCES AND NOTES

1. C. Robert et al., N. Engl. J. Med. 364, 2517–2526 (2011).2. F. S. Hodi et al., N. Engl. J. Med. 363, 711–723 (2010).3. J. Larkin et al., N. Engl. J. Med. 373, 23–34 (2015).4. M. A. Postow et al., N. Engl. J. Med. 372, 2006–2017 (2015).5. D. Schadendorf et al., J. Clin. Oncol. 33, 1889–1894 (2015).6. M. Maio et al., J. Clin. Oncol. 33, 1191–1196 (2015).7. B. M. Carreno et al., Science 348, 803–808 (2015).8. M. M. Gubin et al., Nature 515, 577–581 (2014).9. M. Yadav et al., Nature 515, 572–576 (2014).10. T. N. Schumacher, R. D. Schreiber, Science 348, 69–74 (2015).11. N. van Rooij et al., J. Clin. Oncol. 31, e439–e442 (2013).12. D. B. Johnson et al., Cancer Immunol. Res. 3, 288–295 (2015).13. A. Snyder et al., N. Engl. J. Med. 371, 2189–2199 (2014).14. M. S. Rooney, S. A. Shukla, C. J. Wu, G. Getz, N. Hacohen, Cell

160, 48–61 (2015).15. S. D. Brown et al., Genome Res. 24, 743–750 (2014).16. S. Spranger, R. Bao, T. F. Gajewski, Nature 523, 231–235 (2015).17. E. M. Van Allen et al., Nat. Med. 20, 682–688 (2014).18. Materials and methods are available as supplementary

materials on Science Online.19. K. Cibulskis et al., Nat. Biotechnol. 31, 213–219 (2013).20. E. Hodis et al., Cell 150, 251–263 (2012).21. E. A. Eisenhauer et al., Eur. J. Cancer 45, 228–247 (2009).22. D. Schadendorf et al., Nat. Rev. Dis. Prim. 2015, 15003 (2015).23. N. A. Rizvi et al., Science 348, 124–128 (2015).24. D. T. Le et al.., N. Engl. J. Med. 372, 2509–2520 (2015).25. T. N. Schumacher, C. Kesmir, M. M. van Buuren, Cancer Cell 27,

12–14 (2015).26. The Cancer Genome Atlas Network, Cell 161, 1681–1696 (2015).27. G. F. Hofbauer et al., J. Immunother. 27, 73–78 (2004).28. P. Sharma, J. P. Allison, Science 348, 56–61 (2015).29. S. Kreiter et al., Nature 520, 692–696 (2015).30. R. F. Wang, X. Wang, A. C. Atwood, S. L. Topalian,

S. A. Rosenberg, Science 284, 1351–1354 (1999).31. C. Linnemann et al., Nat. Med. 21, 81–85 (2015).32. J. D. Wolchok et al., Clin. Cancer Res. 15, 7412–7420 (2009).33. K. Cibulskis et al., Bioinformatics 27, 2601–2602 (2011).34. J. T. Robinson et al., Nat. Biotechnol. 29, 24–26 (2011).

ACKNOWLEDGMENTS

This work was supported by NIH U54 HG003067. All sequencing dataare available in dbGap, accession number phs000452.v2.p1.We thank

the Dermatologic Cooperative Oncology Group (DeCOG; Germany)for clinical accrual and sample coordination, S. Bahl (Broad) forproject management effort, and the Broad Genomics Platform forpiloting dual extraction and RNA sequencing from formalin samples.S.M.G. is supported by a fellowship from Janssen and is a paid adviserfor Merck, Bristol-Myers Squibb, Roche, and Novartis. C.L. is a paidadviser for Merck, Bristol-Myers Squibb, Roche, and Novartis. L.A.G.is a consultant for Novartis, Foundation Medicine, and BoehringerIngelheim, a paid adviser for Warp Drive, has ownership interestin Foundation Medicine, and holds a Commercial Research Grantfrom Novartis. E.M.V.A. is a paid adviser for Syapse and aconsultant for Roche Ventana. E.M.V.A., D.M., B.S., D.S., andL.A.G. designed the study, performed the analyses, and wrote themanuscript. S.A.S., D.M., E.M.V.A. and C.J.W. performed HLAtyping and neoantigen binder analyses. E.M.V.A., S.G., and L.A.G.

performed sequencing. B.S., C.B., L.Z., A.S., M.H.G.F., S.M.G., J.U.,J.C.H. B.W., K.C.K., C.L., P.M., R.G., R.D., and D.S. developedthe clinical samples cohort. U.H. performed pathology reviewon the samples.

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/350/6257/207/suppl/DC1Materials and MethodsFigs. S1 to S5Tables S1 to S5References (35–41)

13 July 2015; accepted 27 August 2015Published online 10 September 201510.1126/science.aad0095

ONCOGENE SIGNALING

Identification of an oncogenicRAB proteinDouglas B. Wheeler,1,2,3* Roberto Zoncu,4 David E. Root,3

David M. Sabatini,3,5,6,7,8† Charles L. Sawyers1,8†

In a short hairpin RNA screen for genes that affect AKTphosphorylation, we identified theRAB35 small guanosine triphosphatase (GTPase)—a protein previously implicated inendomembrane trafficking—as a regulator of the phosphatidylinositol 3′-OH kinase (PI3K)pathway. Depletion of RAB35 suppresses AKTphosphorylation in response to growth factors,whereas expression of a dominant active GTPase-deficient mutant of RAB35 constitutivelyactivates the PI3K/AKTpathway. RAB35 functions downstream of growth factor receptors andupstream of PDK1 and mTORC2 and copurifies with PI3K in immunoprecipitation assays.Twosomatic RAB35 mutations found in human tumors generate alleles that constitutively activatePI3K/AKTsignaling, suppress apoptosis, and transform cells in a PI3K-dependent manner.Furthermore, oncogenic RAB35 is sufficient to drive platelet-derived growth factor receptora toLAMP2-positive endomembranes in the absence of ligand, suggesting that there may be latentoncogenic potential in dysregulated endomembrane trafficking.

The phosphatidylinositol 3′-OH kinase (PI3K)/AKT pathway is a key regulator of cellsurvival, proliferation, and growth and iscommonly activated in human cancers byreceptor tyrosine kinase (RTK) signaling

(1, 2). To identify previously unknown regulatorsof PI3K/AKT signaling, we performed an RNAinterference (RNAi) screen using lentivirusesthat express short hairpin RNAs (shRNAs) (3)targeting genes coding for all known G proteins

SCIENCE sciencemag.org 9 OCTOBER 2015 • VOL 350 ISSUE 6257 211

Fig. 3. Immune microenvironmentcytolytic and immune activity cor-relates with response to ipilimu-mab. (A) Patients who achieved clinicalbenefit from immune checkpoint block-ade therapy had significantly higherlevels of tumor cytolytic activity thanthose who had minimal or no benefitfrom ipilimumab (P = 0.039). (B)Patients who achieved clinical benefitfrom ipilimumab therapy had signifi-cantly higher levels of expressionimmune checkpoint receptors thanthose who did not (CTLA-4: P = 0.033,PD-L2: P = 0.041). One point is notshown because of an outlying highCTLA-4 expression value in a nonres-ponder patient (>50 reads per kilobaseper million mapped reads). (C)Response to ipilimumab did not corre-late with expression of or mutations inHLA alleles (P > 0.05 for all). Asterisks (*) indicate P < 0.05.

RN

A e

xpre

ssio

n (R

PK

M)

30

20

10

0

60

40

20

0

Cytolytic activity

CTLA-4 PD-L2 HLA-A HLA-B HLA-C

RN

A e

xpre

ssio

n (R

PK

M)

RN

A e

xpre

ssio

n (R

PK

M)

Clinical benefitLong-term survival with no clinical benefit

Minimal or no clinical benefit

15

10

5

0

PD-L1PD-1

RESEARCH | REPORTS

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

and lipid or protein kinases. Because AKT phos-phorylation is a faithful indicator of PI3K activ-ity, we used an immunofluorescent approach toquantitatively measure phosphorylation of AKTat serine 473 (S473) (4) in HeLa cells, which dis-play robust growth factor–induced AKT phos-phorylation and lack mutations in core PI3Ksignaling pathway genes (Fig. 1A). We screened

a collection of 7450 shRNAs in an arrayed formatin triplicate to identify kinases or guanosine tri-phosphatases (GTPases) whose depletion could al-ter PI3K-dependent AKT phosphorylation.To identify shRNAs that altered AKT phos-

phorylation (i.e., “hits”), we translated the phospho-AKT signal in each screened well to a z score (5).shRNAs targeting genes that positively regulateAKT phosphorylation (pAKT)—RICTOR, PIK3CA,PIK3CB, AKT1, and AKT3—all diminished AKTphosphorylation and scored as “pAKT-low” hits(Fig. 1B and tables S1 to S3). Conversely, shRNAstargeting genes whose protein products inhibitAKT signaling—such as RAPTOR and RHEB—elevated phospho-AKT levels and scored as“pAKT-high” hits. Thus, our screen successfullyidentified known regulators of PI3K/AKT signal-ing, as well as a number of genes not previouslyimplicated in regulation of the pathway (table S4).We prioritized hits for genes that had no reportedlink to PI3K/AKT signaling (table S5) and searchedoncogenomic databases to identify those with

somatic mutations reported in human cancers(6, 7). This left us with a list of 29 genes (table S6),26 of which were pAKT-low hits. Next, we groupedthe 26 pAKT-low hits by their annotated func-tions from the literature and found that the best-represented functional groupwas the RABGTPases,which are well known to regulate intracellularprotein trafficking (Fig. 1C). This finding sug-gests that dysregulated protein trafficking by mu-tant RAB proteins could play a role in oncogenesis.Of the five RAB proteins that appeared in our

list in table S6 (see also Fig. 1C), we pursued RAB35because, in addition to meeting all of the criteriamentioned previously, it is a well-studied GTPasefor which reagents and tools are readily availa-ble. RAB35 is a regulator of cytoskeletal organi-zation and trafficking at the recycling endosome(8–11). Endomembrane trafficking by RAB35 isregulated by the DENND1 family of RAB35guanine nucleotide exchange factors (9, 12–14)and by the TBC1D10 family of RAB35 GTPaseactivating proteins (15–17). Because RAB35 was

212 9 OCTOBER 2015 • VOL 350 ISSUE 6257 sciencemag.org SCIENCE

Fig. 1. An arrayed lentiviral loss-of-function RNA interference screenidentifies known and previously unrecognized regulators of thePI3K/AKT axis. (A) An immunofluorescent assay for phospho-S473 AKTquantitatively reflects PI3K activity. HeLa cells grown in 384-well plates wereserum-starved overnight, treated with PIK-90 and insulin as indicated, andthen fixed and processed for immunofluorescence with a whole-cell stain (red)and an anti–phospho-S473 AKTmonoclonal antibody (green) (top). Signalswere imaged and quantified with a near-infrared scanner, and the phospho-AKTsignal intensity was divided by the whole-cell stain signal intensity to calculate anormalized phospho-AKTvalue for each well (bottom). (B) Averaged (n = 3 replicates) phospho-AKT z scores of screened shRNAs. HeLa cells were plated in 384-well plates, transduced with lentiviral reagents expressing control shRNAs or shRNAs targeting human kinases and GTPases, and then processed as in (A).Phospho-AKT z scores for each shRNAwere calculated, averaged, and arranged as indicated. (C) Functional grouping of the 26 genes whose knockdown loweredAKTphosphorylation, that are not known to regulate PI3K/AKTsignaling, and that are mutated in human cancers. Numbers indicate the number of genes in eachfunctional group. MAPK, mitogen-activated protein kinase; JNK, c-Jun N-terminal kinase; cAMP, adenosine 3′,5′-monophosphate. (D) Representation of somaticmutations in RAB35 in cancer found in oncogenomic databases. Mutations predicted by the cBioPortal for Cancer Genomics to have a medium effect on proteinfunction are noted in black,whereasmutations predicted to have a high effect are noted in red.Table S7 contains a full list of somaticmutations of RAB35 found incancers. aa, amino acids. (E) shRNAs targetingRAB35deplete RAB35 protein levels and inhibit AKTphosphorylation. HeLa cells were transducedwith lentivirusesexpressing the indicated shRNAs, starved overnight and treated with serum as indicated, and lysed; cell lysates were analyzed by immunoblotting. Red asterisksindicate shRNA “hits” from our screen. FBS, fetal bovine serum.

1Human Oncology and Pathogenesis Program, MemorialSloan Kettering Cancer Center (MSKCC), New York, NY10065, USA. 2Weill Cornell/Rockefeller University/SloanKettering Tri-Institutional MD-PhD Program, New York, NY10021, USA. 3Broad Institute of MIT and Harvard,Cambridge, MA 02142, USA. 4Department of Molecular andCell Biology, University of California at Berkeley, Berkeley, CA94720, USA. 5Whitehead Institute for Biomedical Research,Cambridge, MA 02142, USA. 6Department of Biology,Massachusetts Institute of Technology (MIT), Cambridge, MA02142, USA. 7David H. Koch Institute for Integrative CancerResearch at MIT, Cambridge, MA 02142, USA. 8HowardHughes Medical Institute, Chevy Chase, MD 20815, USA.*Present address: Dana-Farber Cancer Institute, Boston, MA02215, USA. †Corresponding author. E-mail: sawyersc@mskcc.org (C.L.S.) or sabatini@wi.mit.edu (D.M.S.)

RESEARCH | REPORTS

a robust hit in our screen (fig. S1A), has not beenimplicated as a regulator of PI3K/AKT signaling,and has somatic mutations that have been re-ported in human cancers mapping to residues con-served in other small GTPases (Fig. 1D and tableS7), we prioritized RAB35 for further study.Next, we confirmed that depletion of RAB35

with the same shRNAs that scored in the orig-inal screen suppressed AKT phosphorylation toa similar extent as depletion of the mTORC2component RICTOR (Fig. 1E). Furthermore, ad-ditional shRAB35 hairpins not present in theoriginal screen also suppressed AKT phospho-rylation, as well as the AKT substrate FOXO1/3A(fig. S1B). Thus, multiple independent shRNAstargeting RAB35 consistently impair AKT phos-phorylation in HeLa cells.To ensure that the effect of RAB35 knock-

down on AKT signaling was not cell-type specific,we depleted RAB35 protein in human embryonickidney (HEK) 293 E cells and in murine NIH-3T3cells (Fig. 2, A and B). Knockdowns of RAB35 withtwo different shRNAs diminished AKT phos-

phorylation at S473 in response to serum stim-ulation in both human and mouse cells and didso more potently than RICTOR. Furthermore, thePI3K-dependent phosphorylation of FOXO1/3Aand the SGK1 substrate NDRG-1 was also de-creased in cells depleted of RAB35. We expandedthis analysis to seven additional cancer cell lineswith different mutational backgrounds (onco-genic PI3KCA or KRAS, deleted PTEN, mutatedNF2 or TP53) and generally found reduced AKTphosphorylation after RAB35 depletion, withsome modest differences that are likely attri-butable to distinct hairpins in distinct cellularcontexts (fig. S2, A to C). Furthermore, AKT phos-phorylation at the PDK1 phosphorylation siteT308 was consistently reduced by RAB35 knock-down, suggesting that RAB35 may function up-stream of PDK1. Collectively, these data indicatethat RAB35 is broadly necessary for the efficientactivation of PI3K/AKT signaling in response toserum stimulation.To explore whether RAB35 is sufficient to ac-

tivate the PI3K/AKT pathway, we generated cell

lines that stably expressed either wild-type RAB35(RAB35wt) or the dominant active GTPase-deficient,guanosine triphosphate (GTP)–bound RAB35Q67→L67 (Q67L) mutant (RAB35Q67L). Cells ex-pressing the wild-type RHEB GTPase (RHEBwt),which regulates mTORC1, served as a control.Stable expression of either RHEBwt or RAB35wt

did not alter phosphorylation of AKT or extra-cellular signal–regulated kinase (ERK) in responseto growth factor signaling (Fig. 2C and fig. S3).However, expression of the active RAB35Q67L mu-tant rendered AKT phosphorylation (at both T308and S473) constitutively elevated and refractoryto growth factor deprivation but did not affectERK phosphorylation. FOXO1/3A phosphoryl-ation was also consistently elevated in serum-deprived cells expressing RAB35Q67L. Therefore,the expression of GTP-bound RAB35 is sufficientto activate PI3K/AKT signaling in cells, even inthe absence of growth factors.The fact that RAB35Q67L results in constitu-

tive phosphorylation of T308 as well as S473 sug-gests that RAB35 may function upstream of the

SCIENCE sciencemag.org 9 OCTOBER 2015 • VOL 350 ISSUE 6257 213

Fig. 2. RAB35 is necessary and sufficient for serum-induced PI3K/AKTsignaling and associates with PI3Ka. (A and B) Depletion of RAB35 inhibitsserum-induced activation of AKT. (A) HeLa and HEK-293E cells were trans-duced with lentiviral reagents expressing the indicated shRNAs. Cells wereserum-starved overnight, treated as indicated with FBS, and lysed; cell lysateswere then analyzed by immunoblotting. (B) Murine NIH-3T3 cells were serum-starved then treated as indicated with bovine calf serum (BCS) and analyzedas in (A). (C) GTPase-deficient RAB35Q67L activates PI3K/AKTsignaling. HeLa

and HEK-293E cells were stably transduced with lentiviruses expressingFLAG-RHEBwt, -RAB35wt, or -RAB35Q67L. Cells were serum-starved overnight,treated as indicated, and analyzed as in (A). (D) RAB35 associates with PI3Kin a nucleotide-dependent manner. Cells stably expressing FLAG-RHEBwt,-RAB35wt, -RAB35Q67L, or -RAB35S22N were lysed, and immunoprecipitations(IPs) were performed using the indicated antibodies. Immunoprecipitates andtotal cell lysates were then analyzed by immunoblotting with the indicatedantibodies. n.s., nonspecific band.

RESEARCH | REPORTS

T308 kinase PDK1 (18, 19). However, attributionof RAB35 function to PDK1 versus mTORC2activation is challenging because the phosphoryl-ation states of T308 and S473 are interdependent(20–22). To address this issue, we took advantageof cells that stably express alleles of murine AKT1(mAkt1) in which either T308 or S473 were mu-tated to phosphomimetic aspartate residues (23)(fig. S4, A and B). As expected, mTOR blockadewith the TORC1/2 inhibitor torin1 or blockadeby RICTOR depletion (resulting in loss of mTORC2function) suppressed phosphorylation at bothsites in mAkt1wt-expressing cells (fig. S4C). InmAkt1S473D-expressing cells, which lack a regu-lated mTORC2 site, these same interventions ei-ther had no effect on T308 phosphorylation(RICTOR depletion) or enhanced T308 phos-phorylation (torin 1).Having confirmed the ability of this system

to distinguish direct effects on T308 versus S473,we examined the impact of RAB35 knockdown.Unlike depletion of RICTOR or treatment with

torin 1, depletion of RAB35 decreased T308 phos-phorylation inmAkt1wt‑ andmAkt1S473D-expressingcells. RAB35 depletion also decreased S473 phos-phorylation inmAkt1wt- andmAKTT308D-expressingcells (fig. S4C). These data suggest that RAB35signals to AKT by functioning upstream of bothPDK1 and mTORC2.We reasoned that if RAB35 acts upstream of

both PDK1 and mTORC2, it may be control-ling their common regulator, PI3K. Althoughwe could not detect an interaction between en-dogenous RAB35 and PI3K, we found that stablyexpressed RAB35, but not RHEB, copurified withPI3K when the kinase was immunoprecipitatedby its p85 subunit (Fig. 2D). Interestingly, moreFLAG-tagged RAB35 (FLAG-RAB35) copurifiedwith PI3K in immunoprecipitates from cells ex-pressing constitutively GTP-bound RAB35Q67L

than from cells expressing RAB35wt. Conversely,we found that the same experiment performedusing lysates from cells stably expressing a con-stitutively guanosine diphosphate–bound allele of

RAB35 (RAB35S22N) did not recover any FLAG-RAB35. Thus, the physical association of RAB35and PI3K is nucleotide-dependent. We alsoasked whether RAB35 is necessary for PI3K ki-nase activity in vitro. As expected, PI3K immu-nopurified from cells depleted of PI3Ka hadsignificantly reduced in vitro kinase activity to-ward phosphatidylinositol 4,5-bisphosphate (fig.S5A). PI3K purified from cells depleted of RAB35had a ~50% reduction in PI3K kinase activity invitro. Further, we found that depletion of RAB35did not reduce the amount of p110a recovered inp85 immunoprecipitates, which suggests thatRAB35 depletion does not disrupt the p85-p110ainteraction (fig. S5B). Thus, RAB35 copurifies withPI3K in a nucleotide-dependent manner and maybe required for optimal PI3K enzymatic function.Given that RAB35 functions upstream of PI3K,

we investigated whether RAB35 acts downstreamof any particular growth factor receptor. Deple-tion of RICTOR, PI3Ka, or RAB35 in HeLa cellsreduced S473 phosphorylation of AKT in response

214 9 OCTOBER 2015 • VOL 350 ISSUE 6257 sciencemag.org SCIENCE

Fig. 3. RAB35 mutants identified in human tumors activate PI3K signal-ing, suppress apoptosis, and are transforming in vitro. (A) Two mutationsin RAB35 that code for amino acid changes A151Tand F161L were identified inthe MSKCC cBio and COSMIC data sets. Alignment of RAB35 with KRAS wasperformed using ClustalW2. (B) Mutant RAB35 alleles from human tumors canactivate PI3K/AKTsignaling. NIH-3T3 mouse fibroblasts stably expressing theindicated RAB35 alleles were serum-starved for 4 hours, left unstimulatedor treated with BCS, and lysed; cell lysates were analyzed by immunoblot-ting for the indicated proteins. (C and D) Expression of RAB35 mutants sup-presses apoptosis. (C) NIH-3T3 cells stably expressing the indicated proteinswere plated and treated as indicated for 4 hours with or without serum-

containing medium. Cells were then lysed, and the cell lysates were immuno-blotted for the indicated proteins. (D) NIH-3T3 cells were treated as in (C)and trypsinized, and the number of viable cells was counted. Cell counts werenormalized to the number of viable cells for each cell line from non–serum-starvedconditions. Graph bars represent mean normalized cell count; error bars indicateSD (n = 3 replicates); and asterisks indicate statistical significance (P < 0.05), asdetermined by a Student’s t test. (E) RAB35 mutants can transform cells in vitroin a PI3K-dependent manner. NIH-3T3 cells stably expressing the indicated RAB35alleles, HA-p110aH1047R, or myristoylated FLAG-AKT1myr were cultured for 4 weeksin medium containing 2% BCS with either dimethyl sulfoxide (DMSO) or 250 nMGDC-0941 (a pan-PI3K inhibitor), then fixed, stained with crystal violet, and imaged.

RESEARCH | REPORTS

to treatment with either insulin-like growthfactor 1 (IGF-1), epidermal growth factor (EGF),platelet-derived growth factor AA (PDGF-AA), orvascular endothelial growth factor (VEGF) (fig.S6A). RAB35 depletion did not dramatically alterphosphorylation of ERK or tyrosine phosphoryl-ation of receptors for IGF, EGF, PDGF, or VEGF(IGFR, EGFR, PDGFR, or VEGFR, respectively)(fig. S6, B and C). Collectively, these data suggesta model in which RAB35 functions upstream ofPI3K and downstream of growth factor RTKs toactivate AKT.One criterion for prioritization of genes on

our hit list (table S6) was evidence for mutationin human cancer databases (table S7). For RAB35,missense mutations have been reported at two

residues that are conserved in RAS-like GTPases,A151T and F161L. We noticed that the A151T andF161L mutations in RAB35 were markedly sim-ilar to two well-documented KRAS mutations(A146T and F156L) that have been previouslyidentified in human tumor samples (Fig. 3A)(24, 25). Although they are not canonical activ-ating mutations, stable expression of either ofthese two KRAS mutants was sufficient to activ-ate ERK signaling and transform NIH-3T3 cellsin vitro. We therefore asked whether the two sim-ilar mutations in RAB35 might also be gain-of-function mutations that activate RAB35 signaling.To test this possibility, we generated NIH-3T3

cells stably expressing either RAB35wt, GTPase-deficient RAB35Q67L, or the RAB35A151T and

RAB35F161L alleles reported in human tumors.As expected from our earlier experiments inhuman cells, expression of RAB35wt did not ac-tivate AKT phosphorylation upon serum depriva-tion, whereas expression of the GTPase-deficientRAB35Q67L did (Fig. 3B). Notably, expression ofthe two naturally occurring RAB35A151T andRAB35F161L mutants also elevated AKT phospho-rylation levels during serum deprivation, in-dicating that both are gain-of-function allelessufficient to activate PI3K/AKT signaling.Because PI3K/AKT signaling inhibits apopto-

sis, we next examined whether cells expressingmutant RAB35 alleles are resistant to apoptosistriggered by growth factor withdrawal. Four hoursof serum deprivation of NIH-3T3 cells stably

SCIENCE sciencemag.org 9 OCTOBER 2015 • VOL 350 ISSUE 6257 215

Fig. 4. Expression of dominant active RAB35Q67L activatesPI3K signaling via PDGFRa. (A) Stable expression of dominantactive RAB35Q67L activates the PDGFR. HEK-293E cells stably ex-pressing the indicated FLAG-tagged GTPases were serum-starved,treated as indicated, lysed, and analyzed by immunoblotting. (B)Pharmacologic inhibition of PDGFRa inhibits PI3K/AKT activity incells expressing dominant active RAB35Q67L. Cells stably expressingRAB35Q67Lwere treated with DMSO, a pan-PI3K inhibitor (PI3Ki) (GDC-0941,0.5 mM), a PI3Ka-specific inhibitor (PI3Kai) (BYL719, 1 mM), a PDGFRb inhibitor(CP-673451, 1 mM), or a PDGFRa/b inhibitor (crenolanib, CP-868596, 1 mM) inserum-free media overnight then lysed and analyzed as in (A). (C) PDGFRa istrafficked to LAMP2-positive endomembranes after stimulation with PDGF.HEK-293E cells stably expressing FLAG-RAB35wt were transfected withcDNA expressing PDGFRa-myc, serum-starved and then treated with PDGF-AA(100 ng/ml) for 15 min, and processed for immunofluorescence with anti-mycand anti-LAMP2 antibodies. (D) Expression of RAB35Q67L traffics PDGFRa toLAMP2-positive endomembranes in the absence of PDGF. HEK-293E cellsstably expressing FLAG-RAB35Q67L were transfected with cDNA coding for

PDGFRa-myc and then treated and processed for immunofluorescence as in(C). Scale bars in (C) and (D) represent 10 mm. (E) Cartoon depicting thetrafficking of PDGFRa to LAMP2-positive endomembranes in cells expressingRAB35Q67L. In cells expressing RAB35wt, PDGFRa is activated at the cellmembrane by PDGF-AA ligand and then internalized to EEA1-positive andLAMP2-positive endomembranes, where liganded PDGFRa drives PI3K/AKTsignaling. Dashed lines reflect the fact that endomembrane-based signalingmay not be as prominent as plasma membrane signaling in normal states(left). In cells expressing GTPase-deficient, GTP-bound RAB35Q67L, unli-ganded PDGFRa is internalized to EEA1- and LAMP2-positive compartments,where it drives constitutive activation of PI3K/AKT signaling (right).

RESEARCH | REPORTS

expressing RAB35wt were sufficient to elevatecleaved levels of the apoptotic markers PARPand caspase 3 (Fig. 3C). In comparison, cellsstably expressing one of the three RAB35 mu-tants or oncogenic p110aH1047R had decreasedlevels of cleaved PARP and cleaved caspase 3after serum withdrawal. Further, although cellsexpressing RAB35wt died in response to serumdeprivation, cells expressing mutant alleles ofRAB35 or p110aH1047R had significantly improvedcell viability when deprived of growth factors(Fig. 3D). Taken together, these data suggestthat mutant RAB35 proteins can mitigate celldeath in response to growth factor deprivation.To determine whether the mutant alleles of

RAB35 are oncogenic, we examined their activityin a standard NIH-3T3 cell focus-formation assay—an in vitro model of density-independent growththat often correlates with tumorigenicity (26).NIH-3T3 cells expressing oncogenic p110aH1047R,myristoylated AKT1 (AKT1myr), or any of thethree mutant alleles of RAB35, but not RAB35wt,formed foci (Fig. 3E and fig. S7). Moreover, wefound that when cultured in medium contain-ing 250 nM of the pan-class I PI3K inhibitorGDC-0941 (27), cells expressing RAB35 mutantsand p110aH1047R were unable to form foci. Notsurprisingly, PI3K blockade did not inhibit thegrowth of cells transduced with AKT1myr. Thus,the expression of the RAB35 mutants identifiedin human cancers can transform NIH-3T3 cellsin vitro in a PI3K-dependent manner.Because RAB proteins regulate endomembrane

trafficking, we considered whether our evidenceimplicating RAB35 in AKT activation, downstreamof RTKs and upstream of PI3K, is linked to thismore established function of RAB proteins. Thefact that many RTKs undergo ligand-stimulatedinternalization and recycling through endosomes,while remaining competent to signal (28–32), sug-gests that RAB35 may function in this process. Inconsidering this possibility, we noted highly ele-vated levels of tyrosine phosphorylation of thePDGFRa/b—but not of IGF1R, fibroblast growthfactor receptor, VEGFR, or MET—in cells expres-sing the constitutively active RAB35Q67L allele(Fig. 4A and fig. S8). This increase in PDGFRa/bphosphorylation was markedly reduced by treat-ment of cells with crenolanib (CP-868596), aPDGFR inhibitor that affects both PDGFRa andPDGFRb with similar potency (33–35) (Fig. 4Band fig. S9). Notably, crenolanib treatment alsodiminished phosphorylation of AKT and the down-stream substrate FOXO1/3A, suggesting that theRAB35Q67L allele activates AKT through PDGFR.Treatment with a more b-selective PDGFR in-hibitor (CP-673451) only moderately affected AKTphosphorylation, implicating PDGFRa as themost likely RAB35Q67L target. The pan-PI3K in-hibitor (GDC-0941) and the PI3Ka-specific inhib-itor (BYL719) both suppressed AKT phosphorylationbut not PDGFRa/b phosphorylation, providingfurther evidence that PDGFR/RAB35Q67L functionsupstream of PI3K.We postulated that RAB35Q67L might initiate

PDGFR-mediated AKT activation by alteringPDGFRa localization. To explore this possibility,

we transfected cells that stably expressed eitherRAB35wt or RAB35Q67L with cDNA encoding formyc-tagged PDGFRa and studied the localiza-tion of tagged PDGFRa in the presence or absenceof the PDGFRa ligand PDGF-AA (fig. S10). Asexpected, in cells that expressed RAB35wt, stim-ulation with PDGF-AA relocalized PDGFRa intopunctate intracellular structures. Notably, PDGFRawas constitutively localized to punctate structuresin RAB35Q67L-expressing cells, even in the absenceof PDGF-AA stimulation. This effect was specific toPDGFRa because internalization of EGFR was notaltered by RAB35Q67L expression (fig. S11). By sim-ultaneously visualizing PDGFRa-myc with the ly-sosomal protein LAMP2 or the early endosomemarker EEA1, we next explored whether thepunctate localization of PDGFRa in RAB35Q67L-expressing cells was the same as that seen in thenormal physiologic context of PDGFRa activationby PDGF-AA. In both PDGF-AA-stimulated RAB35wt

cells and unstimulated RAB35Q67L-expressing cells,transfected PDGFRa-myc colocalized with endog-enous LAMP2 (Fig. 4, C and D, and fig. S12) andEEA1 (fig. S13). We also found that FLAG-RAB35is present in endomembranes that contain LAMP2(fig. S14A) but not in those that contain EEA1(fig. S14B). This suggests that PI3K activity down-stream of RAB35Q67L may originate on LAMP2-positive membranes. Finally, because it has beenestablished that PI3K activity is required for theinternalization and trafficking of PDGFRs (36, 37),we asked whether the localization of PDGFRa-myc in cells expressing RAB35Q67L was a result ofthe increased PI3K signaling in these cells. Al-though PI3K was required for ligand-inducedPDGFRa internalization in cells that expressRAB35wt (fig. S15A), treatment of cells expres-sing RAB35Q67L with a pan-PI3K inhibitor didnot diminish the localization of PDGFRa-myc toLAMP2-positive compartments (fig. S15B). PI3Kinhibitor treatment resulted in changes in LAMP2vesicle morphology that may merit further inves-tigation (fig. S15, A and B). Together, these dataestablish that activated PDGFRa is normallytrafficked through LAMP2-positive endomem-branes and that constitutively active RAB35 likelydrives unliganded PDGFRa to this location, whereit is active and can signal to PI3K/AKT (Fig. 4E).Further study is required to determine whethermutant RAB35 diverts newly synthesized PDGFRafrom the cell membrane to LAMP2-positive endo-membranes or prevents the exit of existing PDGFRafrom LAMP2 compartments.Here we demonstrate a previously unappreciated

role for the small GTPase RAB35 in regulatinggrowth factor signaling to PI3K/AKT. We alsocharacterize two RAB35 missense mutations foundin human tumors that confer gain of function instandard transformation assays, displaying phe-notypes similar to well-characterized oncogenicKRAS alleles. Searches of existing human tumorsequencing databases reveal that RAB35 muta-tions are rare (table S7) and are therefore unlike-ly to appear on current lists of human cancer genesthat rely solely on statistical methods to distin-guish between driver and passenger mutations.Nonetheless, the conserved location of the RAB35

mutations to analogous residues in KRAS andtheir oncogenic phenotype in functional assaysprovide strong evidence that these are true driv-er mutations and are perhaps clinically actionableon the basis of their sensitivity to PI3K inhibition.Considering the broad role of RAB family proteinsin endomembrane trafficking, it is intriguing tospeculate that the oncogenic RAB35 phenotypereported here may be a consequence of dysregu-lated membrane trafficking. Our studies showingthat the constitutively active RAB35Q67L allele issufficient to drive PDGFRa to LAMP2-positiveendomembranes, in a pattern that mirrors thatseen with native PDGF-AA ligand, are consistentwith this model. Alternatively, these mutationsmay be neomorphic, conferring a novel molecularfunction to RAB35 that results in signaling in aRAS-like manner. The fact that other RAB pro-teins (RAB1B, RAB39, RABL3, RASEF) emergedas hits in our screen and have rare missensemutations reported in human tumors (table S6)begs a broader examination of this question.

REFERENCES AND NOTES

1. R. J. Shaw, L. C. Cantley, Nature 441, 424–430 (2006).2. L. C. Cantley, Science 296, 1655–1657 (2002).3. J. Moffat et al., Cell 124, 1283–1298 (2006).4. Single-letter abbreviations for the amino acid residues are as

follows: A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His;I, Ile; K, Lys; L, Leu; M, Met; N, Asn; P, Pro; Q, Gln; R, Arg; S, Ser;T, Thr; V, Val; W, Trp; and Y, Tyr.

5. See supplementary materials and methods for a detailedexplanation of z score calculation and hit criteria.

6. E. Cerami et al., Cancer Discov. 2, 401–404 (2012).7. S. A. Forbes et al., Nucleic Acids Res. 39 (suppl. 1),

D945–D950 (2011).8. D. Dambournet et al., Nat. Cell Biol. 13, 981–988 (2011).9. P. D. Allaire et al., Mol. Cell 37, 370–382 (2010).10. C. E. Chua, Y. S. Lim, B. L. Tang, FEBS Lett. 584, 1–6

(2010).11. J. Zhang, M. Fonovic, K. Suyama, M. Bogyo, M. P. Scott,

Science 325, 1250–1254 (2009).12. A. L. Marat, M. S. Ioannou, P. S. McPherson, Mol. Biol. Cell 23,

163–175 (2012).13. A. L. Marat, P. S. McPherson, J. Biol. Chem. 285, 10627–10637

(2010).14. S. Yoshimura, A. Gerondopoulos, A. Linford, D. J. Rigden,

F. A. Barr, J. Cell Biol. 191, 367–381 (2010).15. C. Hsu et al., J. Cell Biol. 189, 223–232 (2010).16. G. Patino-Lopez et al., J. Biol. Chem. 283, 18323–18330

(2008).17. M. Chaineau, M. S. Ioannou, P. S. McPherson, Traffic

14, 1109–1117 (2013).18. J. R. Bayascas, D. R. Alessi, Mol. Cell 18, 143–145 (2005).19. A. Mora, D. Komander, D. M. van Aalten, D. R. Alessi, Semin.

Cell Dev. Biol. 15, 161–170 (2004).20. R. M. Biondi, A. Kieloch, R. A. Currie, M. Deak, D. R. Alessi,

EMBO J. 20, 4380–4390 (2001).21. M. P. Scheid, P. A. Marignani, J. R. Woodgett, Mol. Cell. Biol.

22, 6247–6260 (2002).22. D. D. Sarbassov, D. A. Guertin, S. M. Ali, D. M. Sabatini, Science

307, 1098–1101 (2005).23. D. D. Sarbassov et al., Mol. Cell 22, 159–168 (2006).24. J. W. Tyner et al., Blood 113, 1749–1755 (2009).25. M. Janakiraman et al., Cancer Res. 70, 5901–5911 (2010).26. G. J. Clark, A. D. Cox, S. M. Graham, C. J. Der, Methods

Enzymol. 255, 395–412 (1995).27. A. J. Folkes et al., J. Med. Chem. 51, 5522–5532 (2008).28. M. D. Chamberlain et al., Cell. Signal. 22, 1562–1575 (2010).29. A. V. Vieira, C. Lamaze, S. L. Schmid, Science 274, 2086–2089

(1996).30. J. C. Chow, G. Condorelli, R. J. Smith, J. Biol. Chem. 273,

4672–4680 (1998).31. C. L. Howe, J. S. Valletta, A. S. Rusnak, W. C. Mobley, Neuron

32, 801–814 (2001).

216 9 OCTOBER 2015 • VOL 350 ISSUE 6257 sciencemag.org SCIENCE

RESEARCH | REPORTS

32. S. Kermorgant, D. Zicha, P. J. Parker, EMBO J. 23, 3721–3734(2004).

33. N. L. Lewis et al., J. Clin. Oncol. 27, 5262–5269(2009).

34. M. C. Heinrich et al., Clin. Cancer Res. 18, 4375–4384(2012).

35. J. Dai et al., Clin. Cancer Res. 19, 6935–6942 (2013).36. M. Joly, A. Kazlauskas, F. S. Fay, S. Corvera, Science 263,

684–687 (1994).37. M. Joly, A. Kazlauskas, S. Corvera, J. Biol. Chem. 270,

13225–13230 (1995).

ACKNOWLEDGMENTS

We thank all of the members of the Sabatini and Sawyerslaboratories for critical readings of this manuscript and helpfuldiscussions. We also thank N. D. Novikov, M. A. Domser, K. C. Wood,

D. J. Spencer, T. R. Peterson, C. Thoreen, S. Kinkel, M. J. Evans,R. Bose, P. Watson, O. S. Andersen, S. C. Blanchard, T. E. McGraw,N. Rosen, and A. Smogorzewska for helpful discussions; S. Silverfor assistance with the RNAi screens; V. Boyko and Y. Romin of theMSKCC Molecular Cytology Core Facility for assistance withconfocal microscopy; A. Kazlauskas for helpful discussions aboutPDGFR biology; and the late A. Hall for helpful input about GTPasebiology. Funds for the RNAi screen were provided by the StarrCancer Consortium Award I2-A117. D.B.W. was supported by theDepartment of Defense Prostate Cancer Research Program via theOffice of the Congressionally Directed Medical Research ProgramsPredoctoral Prostate Cancer Training Award PC094483, as well asby NIH Medical Scientist Training Program grant GM07739. R.Z.was supported by the Glenn Foundation for Medical Research, theNIH Director’s New Innovator Award 1DP2CA195761-01, and thePew-Stewart Scholarship for Cancer Research. This work was

partially supported by NIH grants CA103866 and AI47389 (toD.M.S.) and CA155169 and CA092629 (to C.L.S.). Data from theRNAi screen described in this manuscript can be found in thesupplementary materials.

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/350/6257/211/suppl/DC1Materials and MethodsFigs. S1 to S15Tables S1 to S7References (38–55)

14 December 2014; accepted 24 August 2015Published online 3 September 201510.1126/science.aaa4903

VIROLOGY

Dengue subgenomic RNA bindsTRIM25 to inhibit interferonexpression for epidemiological fitnessGayathri Manokaran,1,2,3 Esteban Finol,1,3,4 Chunling Wang,5 Jayantha Gunaratne,3,6

Justin Bahl,7 Eugenia Z. Ong,1 Hwee Cheng Tan,1 October M. Sessions,1 Alex M. Ward,1

Duane J. Gubler,1 Eva Harris,5 Mariano A. Garcia-Blanco,1,8,9 Eng Eong Ooi1,3,10,11*

The global spread of dengue virus (DENV) infections has increased viral genetic diversity,some of which appears associated with greater epidemic potential. The mechanismsgoverning viral fitness in epidemiological settings, however, remain poorly defined. Weidentified a determinant of fitness in a foreign dominant (PR-2B) DENV serotype 2 (DENV-2)clade, which emerged during the 1994 epidemic in Puerto Rico and replaced an endemic (PR-1)DENV-2 clade.The PR-2B DENV-2 produced increased levels of subgenomic flavivirus RNA(sfRNA) relative to genomic RNA during replication. PR-2B sfRNA showed sequence-dependent binding to and prevention of tripartite motif 25 (TRIM25) deubiquitylation, which iscritical for sustained and amplified retinoic acid–inducible gene 1 (RIG-I)–induced type Iinterferon expression. Our findings demonstrate a distinctive viral RNA–host proteininteraction to evade the innate immune response for increased epidemiological fitness.

Dengue outbreaks emerge when DENVs areintroduced into populationswith lowherdimmunity.However, several outbreakshaveensued when new DENV strains emergedand displaced endemic strains, albeit of the

same serotype and genotype (1–6), suggesting thatgenetic variation altered viral fitness in epidemi-ological settings.

To gain insights into how DENV genome var-iation affects its epidemiological phenotype, weexamined aDENV-2 clade replacement event thatcoincided with an epidemic of severe dengue inPuerto Rico in 1994 (1, 7, 8). Phylogenetic analysisof the complete coding sequences of DENV-2 iso-lated from Puerto Rico and neighboring countriesshowed that threedistinct cladeswithin theAsian/American genotype circulated in 1994 (fig. S1), con-cordant with previous findings (1). Clade PR-1 wasprevalent from 1986 to 1995. Clade PR-2 containedviruses in two subclades, both first isolated in 1994:PR-2A persisted at low levels until 1998, whereasPR-2B spread and replaced PR-1 to become thedominant DENV-2 in Puerto Rico from 1995 to2007. Immune escape is an unlikely explanationfor PR-2B emergence, as both PR-1 and PR-2Bbelong to the Asian/American genotype. Fur-thermore, prM and E amino acid sequenceanalysis showed weak bootstrap values (fig. S2),indicating low levels of antigenic differencesbetween the clades. Instead, PR-2B emergenceand subsequent replacement of PR-1 virusescould be due to differences in epidemiologicalfitness.

To identify the genomic variations that segre-gated theseDENVs into different clades andmod-ulated epidemiological fitness, we applied theShimodaira-Hasegawa (SH) test (9) that comparesthe topology of region-specific with full-genometrees. This approach identifiedNS1, NS3,NS5, andthe 3′ untranslated region (3′UTR) sequence var-iations as potentially critical in segregatingDENVsinto PR-1 and PR-2B (figs. S3 to S5).NS1, NS3, and NS5 are DENV replication com-

plex components that also regulate the innateimmune response (10). Less is known about the3′UTR. During flaviviral replication, the cellular5′-to-3′ exoribonuclease digests uncapped genomesbut stalls at pseudoknot RNA structures in the3′UTR. The resultant 0.3- to 0.5-kb subgenomicflavivirus RNA (sfRNA) regulates pathogenicityin both mammalian and mosquito cells, likelythrough mechanisms and structural elementswithin its noncoding RNA that remain to be fullydefined (11–15). Recently, we showed that sfRNAcould interact with proteins to inhibit transla-tion of interferon-stimulated genes (16). Becausetwo substitutions identified in the PR-2B 3′UTRwere located where pseudoknots could form(fig. S6) (17–19), we tested for a possible role forsfRNA in modulating fitness in our viruses.To measure viral fitness in vitro, we performed

DENV infection assays in human hepatoma(HuH-7) cells. We expected PR-2B DENVs to rep-licate more efficiently than PR-1 viruses. Paradox-ically, however, more viral progeny (Fig. 1A) andgenomic RNA (gRNA) (Fig. 1B) were produced byPR-1 than PR-2B viruses at 24 hours postinfection(hpi). The difference in sfRNA levels betweenPR-1 and PR-2B viruses was smaller, although sta-tistically significant (Fig. 1C), resulting in highersfRNA:gRNA ratios in PR-2B as compared toPR-1 viruses (Fig. 1D). While the absolute valuesof the sfRNA:gRNA ratios should be interpretedwith caution, our data clearly show that theseratios are markedly higher for the epidemicPR-2B viruses.Increasedproduction of sfRNA relative to gRNA

in an epidemic clade of viruses is not specific tothe PR-2B viruses. In Nicaragua, NI-2B DENV-2emerged abruptly in 2005, displaced the endemicNI-1DENV-2, and caused increased rates of severedisease across several epidemic seasons (5). SH teston these Asian/American genotype DENV-2 iso-lates similarly identified variations in NS1, NS5,

SCIENCE sciencemag.org 9 OCTOBER 2015 • VOL 350 ISSUE 6257 217

1Program in Emerging Infectious Diseases, Duke–NationalUniversity of Singapore Graduate Medical School, Singapore.2Defence Medical and Environmental Research Institute, DSONational Laboratories, Singapore. 3Yong Loo Lin School ofMedicine, National University of Singapore, Singapore. 4SwissTropical and Public Health Institute, Universität Basel,Switzerland. 5Division of Infectious Diseases and Vaccinology,School of Public Health, University of California, Berkeley,CA, USA. 6Institute of Molecular and Cell Biology, Singapore.7Center for Infectious Diseases, The University of TexasSchool of Public Health, Houston, TX, USA. 8Department ofBiochemistry and Molecular Biology, The University of TexasMedical Branch, Galveston, TX, USA. 9Department ofMolecular Genetics and Microbiology and Center for RNABiology, Duke University, Durham, NC, USA. 10Saw SweeHock School of Public Health, National University ofSingapore, Singapore. 11Singapore–Massachusetts Institute ofTechnology Alliance in Research and Technology InfectiousDisease Interdisciplinary Research Group, Singapore.*Corresponding author. E-mail: engeong.ooi@duke-nus.edu.sg

RESEARCH | REPORTS

DOI: 10.1126/science.aaa4903, 211 (2015);350 Science

et al.Douglas B. WheelerIdentification of an oncogenic RAB protein

This copy is for your personal, non-commercial use only.

clicking here.colleagues, clients, or customers by , you can order high-quality copies for yourIf you wish to distribute this article to others

  here.following the guidelines

can be obtained byPermission to republish or repurpose articles or portions of articles

  ): October 8, 2015 www.sciencemag.org (this information is current as of

The following resources related to this article are available online at

http://www.sciencemag.org/content/350/6257/211.full.htmlversion of this article at:

including high-resolution figures, can be found in the onlineUpdated information and services,

http://www.sciencemag.org/content/suppl/2015/09/02/science.aaa4903.DC1.html can be found at: Supporting Online Material

http://www.sciencemag.org/content/350/6257/211.full.html#relatedfound at:

can berelated to this article A list of selected additional articles on the Science Web sites

http://www.sciencemag.org/content/350/6257/211.full.html#ref-list-1, 29 of which can be accessed free:cites 52 articlesThis article

http://www.sciencemag.org/cgi/collection/medicineMedicine, Diseases

subject collections:This article appears in the following

registered trademark of AAAS. is aScience2015 by the American Association for the Advancement of Science; all rights reserved. The title

CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience

on

Oct

ober

9, 2

015

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om