cel

Article

MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling

Graphicalabstract

virtualmF

GigaTMEtranslatesH&EimagesintomFimagesacross21proteinchannels

CD8PD-1

TryptasePHH3

CD16CD14CD138

Transgelin

CD11CActin

CD20CD34

caspase3

T-betCK

DAPICD68CD3PD-L1

ki67

CD4

virtualpopulation

H&Eimagespopulation-scaleReal-worldEvidence

>

>

Months

LongitudinalAnalysis

306cancersubtypes

BiomarkerDiscovery

14,256patients,from51hospitals

Highlights

Authors

JeyaMariaJoseValanarasu,HanwenXu,NaotoUsuyama,...,CarloBifulco,

ShengWang,HoifungPoon

Correspondence

carlo.bifulco@

(C.B.),

swang@

(S.W.),

hoifung@

(H.P.)

Inbrief

GigaTIMEleveragesmultimodalAIto

generatevirtualmultiplex

immunoluorescence(mIF)profilesfromstandardH&Eslides,enabling

comprehensivetumorimmune

microenvironmentmodelingacrossa

large(>14,000)anddiversepatient

population.Thisvirtualapproachunlocksnewopportunitiesforlarge-scaleclinicaldiscoveriesthatwerepreviouslyhinderedbythescarcityofmIFdata.

?GigaTIMEusesmultimodalAItotranslateH&Epathologyslidestospatialproteomics

?GigaTIMEgeneratesavirtualpopulationwithcellstatesfromroutineH&Eslides

?Virtualpopulationenableslarge-scaleclinicaldiscoveryandpatientstratification

?Virtualpopulationrevealsnewspatialandcombinatorialproteinactivationpatterns

Valanarasuetal.,2026,Cell189,1–15

January22,2026?2025TheAuthor(s).PublishedbyElsevierInc.

/10.1016/j.cell.2025.11.016

cellress

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

cel

cellress

OPENACCESS

Article

MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling

JeyaMariaJoseValanarasu,

1,

7

HanwenXu,

1,

2,

7

NaotoUsuyama,

1,

7

Chanwoo

Kim,1,

2,

7

CliffWong,

1

PenielArgaw,

1

RacheliBenShimol,

4

AngelaCrabtree,

4

KevinMatlock,

4

AlexandraQ.Bartlett,

4

JaspreetBagga,

1

YuGu,

1

ShengZhang,

1

TristanNaumann,

1

BernardA.Fox,

4

BillWright,

5

AriRobicsek,

5,

6

BrianPiening,

3,

4

CarloBifulco,

3,

4,

8,

*

ShengWang,

1,

2,

8,

*

andHoifungPoon

1,

8,

9,

*

1MicrosoftResearch,Redmond,WA,USA

2SchoolofComputerScienceandEngineering,UniversityofWashington,Seattle,WA,USA

3ProvidenceGenomics,Portland,OR,USA

4EarleA.ChilesResearchInstitute,ProvidenceCancerInstitute,Portland,OR,USA

5ProvidenceResearchNetwork,Renton,WA,USA

6ArclightIntelligence,Newton,MA,USA

7Theseauthorscontributedequally

8Seniorauthor

9Leadcontact

*Correspondence:

carlo.bifulco@

(C.B.),

swang@

(S.W.),

hoifung@

(H.P.)

/10.1016/j.cell.2025.11.016

SUMMARY

Thetumorimmunemicroenvironment(TIME)criticallyimpactscancerprogressionandimmunotherapyresponse.Multipleximmunoluorescence(mIF)isapowerfulimagingmodalityfordecipheringTIME,butitsapplicabilityislimitedbyhighcostandlowthroughput.WeproposeGigaTIME,amultimodalAIframeworkforpopulation-scaleTIMEmodelingbybridgingcellmorphologyandstates.GigaTIMElearnsacross-modaltranslatortogeneratevirtualmIFimagesfromhematoxylinandeosin(H&E)slidesbytrainingon40millioncellswithpairedH&EandmIFdataacross21proteins.WeappliedGigaTIMEto14,256patientsfrom51hos-pitalsandover1,000clinicsacrosssevenUSstatesinProvidenceHealth,generating299,376virtualmIFslidesspanning24cancertypesand306subtypes.Thisvirtualpopulationuncovered1,234statisticallysig-nificantassociationslinkingproteins,biomarkers,staging,andsurvival.Suchanalyseswerepreviouslyinfea-sibleduetothescarcityofmIFdata.Independentvalidationon10,200TCGApatientsfurthercorroboratedourfindings.

INTRODUCTION

Thetumorimmunemicroenvironment(TIME)playsacriticalroleincancerprogression,inluencingtumorgrowth,invasion,metastasis,andresponsetocancertherapiesbyaffectingtumorimmunesurveillanceandevasion.

1–3

TheTIMEisahighlycom-plexspatialecosystemconsistingofcancercells,anddiversenon-malignantcelltypes,includingimmunecells,cancer-asso-ciatedfibroblasts(CAFs),endothelialcells(ECs),pericytes,andothercelltypes,embeddedinanalteredextracellularmatrix.

4

Immunohistochemistry(IHC)visualizestheactivationofaspe-cificprotein,whichoffersanimportanttoolforunveilingkeycellstatesinTIMEstudy.Forexample,PD-L1IHCstainingiden-tifiestheactivationofthesuppressivetumorimmunecheckpointPD-L1,whichisawidelyusedpredictorforresponsetocheck-pointinhibitortherapies.AcriticallimitationofIHCisthattheac-tivationsareevaluatedoneproteinatatime,eachonaseparatetissuesample.Thisshortcomingisparticularlyproblematicfortumormicroenvironmentmodeling,wherethestudyofthecom-

plexinterplayamongtumorandimmunecellsrequiresthesimul-taneousevaluationofavarietyofproteins.Multipleximmunolu-orescence(mIF)hasemergedasapowerfulalternative,enablingco-localized,multi-channelproteinprofilingonthesametissuewhilepreservingthespatialarchitecture.

5–8

Despitethepromise,mIFremainsconsiderablyexpensiveforlarge-scalestudy,duetothesubstantialcostsofreagents,specializedequipment,andcomputationalinfrastructure,com-binedwiththelabor-intensiveworklowsforstaining,imaging,anddataprocessing.

8

,9

Consequently,existingmIFdatasetsareextremelyscarce,whichsignificantlylimittheirapplicabilityinclinicaldiscoveryandtranslation.Bycontrast,hematoxylinandeosin(H&E)imagesareroutinelygeneratedinclinicalwork-lowsatlowcostforstudyingtissuestructureandcellmorphology.

10

,

11

WhileanH&Eimagedoesnotexplicitlyrevealcellstates,thespatialconfigurationofcellsitrevealscanshedlightontheirindividualstates.Suchpatternsmightnotbeobviousforhumanexpertsbutarepotentiallydiscernibleusingadvancedmultimodal

AI.12–16

Recentadvancesinfoundation

Cell189,1–15,January22,2026?2025TheAuthor(s).PublishedbyElsevierInc.1ThisisanopenaccessarticleundertheCCBYlicense(

/licenses/by/4.0/

).

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

eacellress

OPENACCESS

cel

Article

modelsfurtheramplifythispotential,asAIhasdemonstratedsu-periorperformancebypretrainingfromalargecollectionofpa-thologyimages.

17–19

SuchadvancesinAIsuggesttheplausibilityoflearningpathologicalfeaturesindicativeofspatiallyresolvedproteinactivations.

Inthispaper,weproposeGigaTIME,amultimodalAIframe-workthatenablespopulation-scaleTIMEstudybylearningtogeneratevirtualmIFimagesfromreadilyavailableH&Eimages.GigaTIME’scross-modaltranslatorwaspretrainedusing40millioncellswithpairedH&EandmIFslidesforthesametis-sues.WeappliedGigaTIMEtogeneratevirtualmIFimagesforcancerpatientsfrom51hospitalsandoverathousandclinicsacrosssevenUSstatesatProvidenceHealth,yielding299,376virtualmIFwhole-slideimagesacross21proteinchan-nels.Thisenabledustocreatealargeanddiversevirtualpop-ulationof14,256patientsacross24cancertypesand306can-cersubtypes,withvirtualmIFslidesandkeyclinicalattributessuchasbiomarkers,staging,andsurvivalinformation.Inturn,thisenablesustoconductpopulation-scaleclinicaldiscovery,identifying181,453,and600statisticallysignificantTIMEpro-tein-biomarkerassociationsatthepan-cancer,cancer-type,andsubtypelevels,respectively,alladjustedformultipletesting,formingacomprehensiveprotein-biomarkerassocia-tioncollectionforTIME.Thevirtualpopulationalsoenabledpa-tientstratificationpredictiveofstagingandsurvival,withtheGigaTIMEsignatureintegratingallvirtualproteinactivationsoutperformingindividualvirtualproteinsinpatientstratification,underscoringtheimportanceofmultiplexedanalysis.Wedemonstratedhowavirtualpopulationcansupportspatiallyinformedmetrics,suchassharpnessandentropy.Wefurtheridentifiedcombinatorialproteinchannelsthatrevealedsyner-gisticassociations.Forexample,ourGigaTIMEanalysisfoundthatthecombinationofCD138andCD68outperformeachpro-teinaloneinpredictingallrepresentativeclinicalbiomarkers,withstatisticallysignificantdifferencefor13outof20bio-markers,suggestingtheircombinedroleinantibody-mediatedtumormechanisms.

ToevaluateGigaTIME’sgeneralizability,weappliedittogenerateanindependentvirtualpopulationof10,200tumorsfromtheCancerGenomeAtlas(TCGA).Weobservesignifi-cantconcordanceacrossthevirtualpopulationsfromProvi-denceandTCGA,whichattainedaSpearmancorrelationof0.88forvirtualproteinactivationsacrosscancersubtypes.ThetwopopulationsalsouncoverasignificantoverlapofTIMEprotein-biomarkerassociations(Fisher’sexacttest;p<2×10?9).Ontheotherhand,theProvidencevirtualpop-ulationyielded33%moresignificantassociationsthanTCGA,highlightingthevalueoflargeanddiversereal-worlddataforclinicaldiscovery.

GigaTIMEoffersapromisingAIframeworkforscalingtumormicroenvironmentmodelingbylearningtotranslatereadilyavail-ableH&EimagesintohighlyinformativevirtualmIFimages,thuspavingthewayforadvancingprecisionimmuno-oncologythroughpopulation-scaleTIMEanalysisanddiscovery.Tofacil-itatefutureresearchintumormicroenvironmentmodeling,wewillreleaseourpretrainedGigaTIMEmodel,aswellasourin-housedatasetof40millioncellsfromde-identifiedpairedH&EandmIFslidesacross21proteinchannels.

Table1.TIMEmarkersusedinourstudyandtheircellular

expression

TIMEmarker

Typicalcellularexpression

CD4

Thelper

CD138

Plasmacells

CD68

Macrophages

CD14

Monocytes

Tryptase

Mastcell

Caspase3

Cellsundergoingapoptosis

PHH3

Mitoticspindle

Ki67

Proliferatingcells(normalandtumor)

T-bet

NKcells,Tcells(CD3+,ofCD4+orCD8+lineages)

Transgelin

Fibroblasts;smoothmuscle

CD3

AllTcells

CK

Epithelialcelllineage

Actin

Myofibroblast

DAPI

Nuclearmarker

CD8

CytotoxicTcells

CD20

Bcells

CD16

CD16a:NKcells;CD16b:neutrophils¯ophages

CD34

Vascularmarker(alsoHematopoieticprecursors,capillaryendothelium)

PD-1

Tcells,Bcells

PD-L1

Antigenpresentingcells(APCs),tumorcells,vascularendothelium

CD11c

Dendriticcells(antigenpresenting)

Tablelistingalltumorimmunemicroenvironment(TIME)markersincludedinouranalysis,alongwiththeirpredominantcellularlocalizationandbiologicalfunction.

RESULTS

GigaTIMEgeneratesavirtualpopulationofmultipleximmunoluorescence

GigaTIMEenablesthegenerationofalargeanddiversevirtualpopulationofmIFimagesfrompopulation-scaleH&Eslides.Wefirstexperimentallyacquired441mIFimagesfrom21H&E-stainedslidesacross21proteinchannels(see

STARMethods

;FigureS1;

Table1

).ThesepairedH&EandmIFslideswereprocessedthroughacomputationalpipelinethatincludesimageregistrationandcellsegmentation,resultinginadatasetcomprising40millioncellswithpairedH&EandmIFslides(

Figures1

Aand

S2

).

Wedividedthepaireddataintotraining,development,andheld-outtestsets(see

STARMethods

).TotranslateH&EimagesintomIFones,GigaTIMEwastrainedonthetrainingpaireddatausingapatch-basedencoder-decoderarchitecturebuiltonNes-tedUNet.

20

ThemodelinputsanH&Eimagepatchandoutputs21correspondingmIFpatches,oneforeachproteinchannel.Thesechannel-specificpatchesaresubsequentlystitchedtogethertoreconstructthewhole-slidemIFimages,enablingspatiallyresolved,slide-levelproteinactivationprofiling.

2Cell189,1–15,January22,2026

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

Article

ll

OPENACCESS

C

A

B

Figure1.GigaTIMEenablespopulation-scaletumorimmunemicroenvironmentanalysis

(A)GigaTIMEinputsahematoxylinandeosin(H&E)whole-slideimageandoutputsmultipleximmunoluorescence(mIF)across21proteinchannels.ByapplyingGigaTIMEto14,256patients,wegeneratedavirtualpopulationwithmIFinformation,leadingtopopulation-scalediscoveryonclinicalbiomarkersandpatientstratification,withindependentvalidationonTCGA.

(B)Circularplotvisualizingatumorimmunemicroenvironment(TIME)spectrumencompassingtheGigaTIME-translatedvirtualmIFactivationscoresacrossdifferentproteinchannelsatthepopulationscale,whereeachchannelisrepresentedasanindividualcircularbarchartsegment.TheinnercircleencodesOncoTree,whichclassifies14,256patientsinto306subtypesacross24cancertypes.Theoutercirclegroupstheseactivationsbycancertypes,allowingvisualcomparisonacrossmajorcategories.

(C)Scatterplotcomparingthesubtype-levelGigaTIME-translatedvirtualmIFactivationsbetweenTCGAandProvidencevirtualpopulations.Eachdotdenotestheaverageactivationscoreofaproteinchannelamongalltumorsofacancersubtype.

SeealsoFiguresS1,

S2

,

S3

,and

S4

.

Cell189,1–15,January22,20263

4Cell189,1–15,January22,2026

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

eacellress

OPENACCESS

cel

Article

Effectively,givenaproteinchannel,GigaTIMEoutputsabinaryclassificationlabelforeachpixelindicatingwhetheritisacti-vatedforthegivenproteinchannel.Wecanthentallythecountofactivatedpixels,aswellasactivationdensityscore(proportionofactivatedpixels),foranyimagepatchorwholeslide.

WethenappliedGigaTIMEtoalargeanddiversereal-worldda-tasetconsistingof14,256whole-slideH&Eimagescollectedfrom51hospitalsandover1,000clinicsacrosssevenUSstatesatProvidenceHealth,encompassing24cancertypesand306can-cersubtypes(

FigureS3

).Usingourtrainedmodel,wegenerated299,376virtualmIFwhole-slideimagesforthesepatients.Thisen-ablesustoconstructalarge-scaleanddiversemultimodalvirtualpopulationwithH&EandvirtualmIFimages,alongwithclinicalat-tributes,suchasbiomarkers,staging,andsurvivalstatus.

Asaproofofconcept,wequantifiedaproteinactivationden-sityscoreforeachmIFimage,definedastheproportionofacti-vatedpixels.Thesescoreswereaggregatedacrosstumorsofthesamesubtypeusingmeanpooling,yieldingaspectrumofmIF-basedTIMEfeaturesacrosscancersubtypes(

Figure1

B).Toevaluatetherobustnessofourapproach,wefurtherappliedGigaTIMEto10,200tumorsfromTCGA,generating214,200vir-tualmIFwhole-slideimagesacross21channels(

FigureS4

).Ahighdegreeofconcordancewasobservedbetweentheaggre-gatedactivationscoresderivedfromtheProvidence-basedandTCGA-basedvirtualpopulations(

Figure1

C),underscoringthegeneralizabilityandreliabilityofGigaTIME.

GigaTIMEtranslateswhole-slideH&EimagestomIF

WefirstevaluatedGigaTIME’sperformanceintranslatingstan-dardH&Ewhole-slideimagesintomIFimages.Tobenchmarkthistranslationtasksystematically,wecomparedGigaTIMEwithCycleGAN,

21

awidelyusedimagetranslationmodelfrequentlyappliedinvirtualstainingtasks.

22

Evaluationwascon-ductedusingthreestandardmetricsacrossdifferentlevelsofgranularity(pixel,cell,andslide)tocapturebothlocalandglobaltranslationfidelity.WealsoevaluatedanaverageactivationbaselinebyfirstcomputingtheaveragepositivityrateforeachproteinchannelinthegoldmIFdata,andthenusingthatasthesamplingprobabilitytogeneratevirtualpositivepixelsforthegivenchannel.Atthepixellevel,wetreatedthetranslationproblemasasegmentationtaskandusedtheDicescoretoeval-uatefine-grainedspatialconcordancebetweenthemeasuredmIFandGigaTIME-translatedvirtualmIF.Duringevaluation,weskippedallblankregionsatthe8×8leveltominimizepoten-tialriskforthemtoinlatethescores.

GigaTIMEsignificantlyoutperformedCycleGANon15outof21proteinchannels(nostatisticallysignificantdifferencefortheremaining6channels),highlightingtheimportanceofacquiringpairedH&E-mIFdataandlearningdirectlyfromsuchpaireddata(

Figure2

A).Wealsoobservedthattheaverageacti-vationbaselineperformedverypoorly.Forexample,thetestDicescoreforaverageactivationbaselinewasonly0.12forDAPI,comparedto0.72attainedbyGigaTIME.TheDicescoresforaverageactivationbaselineinotherproteinchannelswereevenworse.

Toassessperformanceatthecelllevel,weusedan8×8-pixelwindowandcountedthenumberofactivatedpixelsineachwin-dowforeachchannel.Pearsoncorrelationwasthencomputed

betweenpredictedandground-truthcountsacrossthewin-dows.GigaTIMEachievedsignificantlyhighercorrelationsthanCycleGAN,whichperformedclosetorandom,suggestingthatCycleGANfailstorecovercoherentcell-levelpatterns(

Figure2

B).Inparticular,GigaTIMEachievedaPearsoncorrela-tionof0.59ontheDAPIchannel,whichbindsDNAandmarksnuclei,indicatingitscapabilityinidentifyingandlocalizingindi-vidualcells.Bycontrast,CycleGANattainsaPearsoncorrelationof0.03.Theaverageactivationbaselineisevenworse,attainingzeroornegativescores(

FigureS8

).

TheseresultsverifiedthatGigaTIMEindeedlearnednon-trivialgeneralizableinformationabouttissueandproteinstructures.Forevaluatingglobalspatialpatterns,weimplementedaslide-levelmetricinspiredbyimmunoscore,

23

avalidatedscoringmethodthatwasoriginallydevelopedforcolorectalcancerandhasbeenusedinassociationstudiesofpatientoutcomesandtreatmentresponse.Weimplementedaconceptuallysimilarversionbycomputingtheratioofactivatedpixelswithineach256×256patch,andthencalculatedtheSpearmancorrelationforpatch-wiseactivationratiosbetweenthemeasuredmIFandGigaTIME-translatedvirtualmIF.Thismetricprovidesabroaderviewthatcomplementspixel-andcell-levelevaluations.Giga-TIMEachievedaSpearmancorrelationof0.98forDAPIand0.56forallchannels,whereasCycleGANyieldednear-zerocor-relationforallchannels(

Figures2

Cand

S5

),validatingthene-cessityoftrainingonpairedH&EandmIFdataforaccuratecross-modaltranslation.Finally,qualitativecomparisonsonrepresentativewhole-slideimagepatchesfurtherillustratetheagreementbetweenthemeasuredmIFandtheGigaTIME-trans-latedvirtualmIF(

Figure2

D).

TofurtherassessthegeneralizabilityofGigaTIME,wecon-ductedadditionalanalysesbasedonnewlygeneratedpaireddatatodirectlyevaluatesupervisedvirtualmIFgenerationonun-seencancertypes.Specifically,wecuratedtumorcoresfromtis-suemicroarrays(TMAs)ofbreastandbrain,whichwerenotincludedinthetrainingdata.Thesecohortsencompassawidespectrumofhistologicalsubtypesandclinicalstages.Inadditiontothedifferencesincancertypes,stages,andhistologies,thenewmIFdatacomprisemanysmall,cylindricaltissuesamplesseparatedbyblankregions,whichwerequitedifferentfromtheoriginalwholeslideswithlargecontiguoustissueregions.Despiteallthesemajordifferences,GigaTIMEhasdemonstratedremarkablegeneralizabilityinDicescoresandPearsoncorrela-tions(

FiguresS11

and

S12

).ItalsocontinuedtooutperformCycleGANandrandomimputationbyawidemargin.Wealsoobservedthattheorderingoftranslationqualityisbroadlyconservedacrossthecancertypes.

Wealsoprovidestratifiedresultsin

FigureS13

basedonsub-cellularlocalization:nuclear(DAPI,PHH3,Ki-67,andT-bet),sur-face(CD4,CD138,CD11c,CD14,CD3,CD8,CD20,CD16,CD34,PD-1,andPD-L1),andcytoplasmic(CD68,tryptase,caspase-3,transgelin,cytokeratin,andactin).Generally,nuclearproteinshavebettertranslationqualitythancytoplasmicandsurfaceproteins.Thismakessenseasnuclearproteins,suchasDAPIandKi-67,generallyexhibitcompactstructureswithclearhistologicalboundaries.Bycontrast,cytoplasmicandsur-faceproteinsoftenexhibitdiffuseorheterogeneousspatialpat-terns,makingtheminherentlymorechallengingtopredict.

Cell189,1–15,January22,20265

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

Article

ll

OPENACCESS

A

B

C

D

Figure2.GigaTIMEenablestranslationfromH&EtomIFimages

(AandB)BarplotcomparingGigaTIMEandCycleGANonthetranslationperformanceintermsofDicescore(A)andPearsoncorrelation(B)witherrorbarsdenotingstandarddeviationandsignificancelevelsfromtwo-tailedttestsrepresentedby*p≤0.05,**p≤0.01,and***p≤0.001.

(C)ScatterplotscomparingtheactivationdensityofthetranslatedmIFandtheground-truthmIFacrossfourchannels.

(D)QualitativeresultsforasampleH&Ewhole-slideimagefromourheld-outtestsetwithzoomed-invisualizationsofthemeasuredmIFandGigaTIME-translatedmIFforDAPI,PD-L1,andCD68channels.

Seealso

FiguresS5

,

S8

,

S9

,

S10

,

S11

,

S13

,and

S14

.

Virtualpopulationenableslarge-scalediscoveryofprotein-biomarkerassociations

AftervalidatingGigaTIME’sperformanceintranslatingwhole-slideH&EimagesintomIFimages,wenextexaminedhowthevirtualmIFpopulationcanenablelarge-scaleclinical

discovery.Specifically,ourvirtualpopulationidentified1,234statisticallysignificantassociationsbetween21GigaTIME-translatedvirtualproteinchannelsand20clinicalbiomarkersatpan-cancer,cancer-type,andcancer-subtypelevels(

Figure3

A).Statisticalsignificancewasdeterminedby

6Cell189,1–15,January22,2026

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

ll

OPENACCESS

Article

A

C

B

D

E

F

G

H

Figure3.GigaTIMEidentifiesnovelTIMEproteinvs.biomarkerassociationsatpan-cancer,cancer-type,andcancer-subtypelevels

(A)GigaTIMEgeneratesavirtualpopulationof14,256withvirtualmIFbytranslatingavailableH&EimagestomIFimages,enablingpan-cancer,cancer-type,andcancer-subtypelevelsofbiomedicaldiscovery.

(legendcontinuedonnextpage)

Cell189,1–15,January22,20267

Pleasecitethisarticleinpressas:Valanarasuetal.,MultimodalAIgeneratesvirtualpopulationfortumormicroenvironmentmodeling,Cell

(2026),

/10.1016/j.cell.2025.11.016

cel

Article

cellress

OPENACCESS

comparingproteinactivationdensitiesbetweenbiomarker-positiveandbiomarker-negativegroupsusingttest,withpvaluesadjustedformultiplehypothesistesting(see

STAR

Methods

).

Atthepan-cancerlevel,weidentified175significantprotein-biomarkerassociations(

Figure3

B).Manyofthesefindingsaresupportedbyexistingliterature.Forinstance,bothhightu-mormutationalburden(TMB-H)andmicrosatelliteinstability-high(MSI-H)genotypesexhibitedstrongassociationswithincreasedactivationinTIME-relatedchannelssuchasCD138,CD20,CD68,andCD4,consistentwiththewell-describedef-fectsofantigenic-mediatedimmuneactivation.

24

,

25

Amonggenomicalterations,KMT2Dmutationsshowstrongpositivecorrelationswithmultipleimmunemarkers,includingCD3,CD8,andCD20,suggestingKMT2Dmutationsareassociatedwithenhancedimmuneinfiltrationatapan-cancerlevel.

26

Conversely,KRASmutationswerenegativelyassociatedwithmarkersofimmuneinfiltration,suchasCD3andCD8atthepan-cancerlevel,relectinganimmune-excludedphenotype.Interestingly,KRASmutationsalsoassociatednegativelywithPD-L1expression,despiteestablishedmechanismsbywhichKRASsignalingpromotesPD-L1expressionviatheERKpathway.

27

,

28

Furtherinvestigationisneededtodeterminewhethersuchcorrelationsarecausalormerelystemfromcon-foundingeffects.

TheclinicallyreportedPD-L1biomarker,assessedviaIHC,waspositivelycorrelatedwithvirtualPD-L1channelactivation,indicatingstrongconcordancebetweenvirtualpopulation-basedpredictionsandclinicallyobservedproteinexpression.Moreover,PD-L1IHCshowednegativeassociationswithseveralTIMEmarkers,includingCD3,CD8,andCD20,relectingtheroleofthePD-1/PD-L1checkpointinestablishingimmuneblockade.

29

Additionally,PD-L1expressionwasinverselycorre-latedwithproliferationmarkers(Ki-67andPHH3)andtheapoptosismarkercaspase3.OneexplanationforthesefindingsisthatPD-L1expressionisassociatedwithlessproliferativetu-morstatesthatdemonstrateresistancetoapoptosis.Alterna-tively,itisalsopossiblethatimmunecellsaresimplylessprolif-erativewhenthereishighPD-L1.

Atthecancer-typelevel,weobservedahighnumberofas-sociationsinbrain(

Figure3

C),lung(

Figure3

D),andbowelcan-cers(

Figure3

E),whicharealsothemostrepresentedcancertypesinourdataset(

FigureS3

).Specifically,GigaTIMEidenti-fied64protein-biomarkerassociationsinbraincancer,137inlungcancer,and175inbowelcancer.Manyoftheseassocia-tionsw