DeepDR：用于藥物反應(yīng)預(yù)測(cè)的深度學(xué)習(xí)庫(kù) DeepDR -a deep learning library for drug response prediction

Bioinformatics,2024,40(12),btae688

/10.1093/bioinformatics/btae688

AdvanceAccessPublicationDate:18November2024

ApplicationsNote

Dataandtextmining

Downloadedfrom

/bioinformatics/article/40/12/btae688/7903283bygueston26December2024

DeepDR:adeeplearninglibraryfordrugresponseprediction

ZhengxiangJiang

1

,

2

andPengyongLi

1

,

*

1SchoolofComputerScienceandTechnology,XidianUniversity,Xi’an,Shaanxi710126,China2SchoolofElectronicEngineering,XidianUniversity,Xi’an,Shaanxi710126,China

*Correspondingauthor.SchoolofComputerScienceandTechnology,XidianUniversity,266XinglongSectionofXifengRoad,Xi’an,Shaanxi710126,China.E-mail:lipengyong@

AssociateEditor:JonathanWren

Abstract

Summary:Accuratedrugresponsepredictioniscriticaltoadvancingprecisionmedicineanddrugdiscovery.Recentadvancesindeeplearning(DL)haveshownpromiseinpredictingdrugresponse;however,thelackofconvenienttoolstosupportsuchmodelinglimitstheirwidespreadapplication.Toaddressthis,weintroduceDeepDR,thefirstDLlibraryspecificallydevelopedfordrugresponseprediction.DeepDRsimplifiestheprocessbyautomatingdrugandcellfeaturization,modelconstruction,training,andinference,allachievablewithbriefprogramming.Thelibraryincorporatesthreetypesofdrugfeaturesalongwithninedrugencoders,fourtypesofcellfeaturesalongwithninecellencoders,andtwofusionmodules,enablingtheimplementationofupto135DLmodelsfordrugresponseprediction.WealsoexploredbenchmarkingperformancewithDeepDR,andtheoptimalmodelsareavailableonauser-friendlyvisualinterface.

Availabilityandimplementation:DeepDRcanbeinstalledfromPyPI

(/project/deepdr

).ThesourcecodeandexperimentaldataareavailableonGitHub

(/user15632/DeepDR

).

1Introduction

Precisionmedicineaimstodelivertailoredtherapiesforindividualtumorsatthemolecularlevel.Predictingdrugresponse(DR)

(Baptistaetal.2021

)remainsacomplexchallengewithinthisfield,reflectingtheintricaterelationshipbetweencancermulti-omicsinformationandtreatmenteffi-cacy.AccurateDRpredictioncouldsignificantlycontributetothedesignofpersonalizedtreatmentsandtheimprovementoftherapeuticoutcomes.Deeplearning(DL)

(LeCunetal.

2015

),amachinelearningapproach,hasdemonstratedcon-siderablepromiseinidentifyingcomplexpatternswithinbio-logicalinformation,includingcancermulti-omicsanddrugmolecules.ThispotentialhasspurreditsgrowingapplicationinDRmodeling,whereitisconsideredavaluabletoolforen-hancingunderstandingandpredictivecapabilities

(Lietal.

2021a

).However,despitethedevelopmentofnumerousmodelsinthisdomain,thereisstillalackofaunifiedandgeneralizedframeworkformodelconstructionandtraining.

CurrentDLapproachestoDRpredictiontypicallyuseastructuredmethodology,consistingofkeycomponentssuchasdrugmodeling,cellmodeling,andfusionmodulesforpredictiongeneration.Drugmodelingaimstoeffectivelyrepresentthechemicalpropertiesandpotentialbiologicaleffectsofdrugs.Thisisusuallyachievedbyrepresentingthemolecularstructureinformatsconducivetocomputationalprocessing,suchasmo-lecularfingerprints

(Lietal.2021a

),SMILES(SimplifiedMolecularInputLineEntrySystem)

(Liuetal.2019

),andmo-leculargraphs

(Liuetal.2020

),followedbylearningstructuralinformationthroughmodelslikeDeepNeuralNetworks

(DNNs)

(Chawlaetal.2022

),ConvolutionalNeuralNetworks(CNNs)

(Manicaetal.2019

),andGraphNeuralNetworks(GNNs)

(Zhangetal.2019

).Cellmodelinginvolvesprocessingbiologicaldatafromcells,includingtranscriptomics

(Chawla

etal.2022

),genomics

(Liuetal.2019

),andproteomics

(Matlocketal.2018

).DLtechniques,particularlyDNNs

(Chawlaetal.2022

),andCNNs

(Manicaetal.2019

),arelever-agedtolearnintricatepatternswithinthesefeatures.Thefusionmoduleintegratestheinsightsfromdrugandcellmodeling,us-ingDNNs

(Chawlaetal.2022

)orattentionmechanisms

(Sakellaropoulosetal.2019

),topredictdrugresponses.

DRpredictionmodelshaveabroadspectrumofapplica-tionsbeyondtheirprimaryfunction.Thesemodelscanbeutilizedtopredictthepharmacologicalpropertiesorbiologi-calactivityofmoleculesforvirtualscreeningandtoanalyzeomicsdataforcellclassification.TheversatilityofDLmod-elsrendersthemhighlyapplicableinarangeofcontexts.Forexample,clinicalresearchersinvestigatingtheimpactofge-neticvariationsondrugresponsesmightusethesemethodol-ogiestoanalyzegenomicdatafrompatientswithspecificdiseases.Similarly,computationalbiologistsaimingtodevelopadvancedpredictivemodelscanleveragediversedatasetstoexplorevariousmodelingarchitectures,therebyimprovingtheaccuracyofDRpredictions.However,imple-mentingthesemodelsrequiressubstantialexpertiseinDLandsignificantcodingefforts.Thetime-intensiveandcomplexityofadaptingtotheuniqueprogramminginterfa-cesofvariousopen-sourcetoolspresentnonnegligiblechal-lengerequiringresolution.

Received:9September2024;Revised:29October2024;EditorialDecision:11November2024;Accepted:13November2024。TheAuthor(s)2024.PublishedbyOxfordUniversityPress.

ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommonsAttributionLicense(

/licenses/by/4.0/

),whichpermitsunrestrictedreuse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.

2JiangandLi

Toaddressthechallengesabove,weintroduceDeepDR(DeepDrugResponse),aPython-basedDLlibrarydesignedforDRprediction.DeepDRincorporatesthreetypesofdrugfeaturesalongwithninedrugencoders,fourtypesofcellfea-turesalongwithninecellencoders,aswellastwofusionmodules.Thiscomprehensiveframeworksupportstheimple-mentationof135models,cateringtoclinicalresearchersandcomputationalbiologistswithlimitedprogrammingback-grounds.Inaddition,wedemonstratetheutilizationofDeepDRbyimplementingandvalidatingmultiplemodelsontheintegrateddatasets,whichhelpstoidentifythemosteffec-tivemodeling.Tofurthersupportresearchers,wedevelopavisualinterfacethatenablesuserswithoutprogrammingex-pertisetoutilizetheoptimalmodels.

2DeepDRlibrary

2.1Datasetframework

2.1.1Featurization

Drugfeaturization.DeepDRoffersthreemodalitiesofdrugfeatures:FP(MolecularFingerprints)(

Lietal.2021a

),SMILES(SimplifiedMolecularInputLineEntrySystem)

(Liu

etal.2019

),andmoleculargraphs

(Liuetal.2020

)(see

Fig.1B

).FParethebinaryvectorrepresentationsofmole-cules

(RogersandHahn2010

).SMILESprovidesaspecifica-tionforencodingmoleculesasstrings

(Weininger1988

).Graphsrepresentmoleculesbyabstractingatomsasnodesandchemicalbondsasedges

(Kearnesetal.2016

).Detailsareavailablein

SupplementaryTextS1

.

A

Cellfeaturization.DeepDRintegratesfourmodalitiesofcellfeatures:expressionprofile(EXP)

(Manicaetal.2019

),pathwayenrichmentscore(PES)

(Chawlaetal.2022

),muta-tionstatus(MUT)

(Liuetal.2019

),andcopynumbervaria-tion(CNV)

(Liuetal.2019

)(see

Fig.1B

).EXPreflectsthequantitativeexpressionlevelsofgenes

(Heller2002

).PESilluminatesthecombinatorialimplicationsamonggeneswithinspecificpathways

(Hnzelmannetal.2013

).MUTreferstothegeneticalterationsorvariationswithinspecificgenes(

Stensonetal.2017

).CNVrepresentsgenomicdele-tionsandduplicationsobservableatthesubmicroscopicscale

(Freemanetal.2006

).Giventhecomplexityofprocessinghigh-dimensionaldata,DeepDRprovidesfeaturesscreenedongenesubsetsinadditiontogenome-widefeatures

(Jiaetal.

2021

).Detailsareprovidedin

SupplementaryTextS2

.

2.1.2Datasetandsplitting

Downloadedfrom

/bioinformatics/article/40/12/btae688/7903283bygueston26December2024

DeepDRintegratestheCancerCellLineEncyclopedia(CCLE)

(Barretinaetal.2019

)andGenomicsofDrugSensitivityinCancer(GDSC)

(Yangetal.2016

),andallowsuserstousetheirowndatasets(see

SupplementaryTextsS3

andS4

).Themeasurementofdrugresponseisquantifiedus-ingseveralparameters:thenaturallogarithm-transformedIC50(HalfMaximalInhibitoryConcentration),AUC(AreaUndertheDose-responseCurve),andActArea(ActivityArea).Tosupportthevalidation,DeepDRincorporatesfourdatasetsplittingstrategies:commonrandom,leave-cell-out,leave-drug-out,andstrictsplit

(Manicaetal.2019

)(see

Fig.1C

).Theleave-cell-outsplitisdesignedtoeliminateanyoverlapofcellsbetweenthetraining,validation,andtestsets.Thisapproachaimstoreplicatethescenariowherethedrugresponseofnewcellstoexistingdrugsisevaluated.Similarly,theleave-drug-outsplitseekstoemulatetheresponseofknowncellstonoveldrugs,whilethestrictsplitisdesignedtosimulatetheresponseofnovelcellstonoveldrugs.

2.2ModelforDRprediction

DeeplearningDRpredictionmodelcanbeformulatedasencodingfordrugsandcellsandfusionofdrugandcellinfor-mation.Inlinewiththisframework,DeepDRhasdevelopedthreeintegralmodules:thedrugencoder,cellencoder,andfusionmodule.Thesecomponentsaredesignedtoprovidethefoundationfortheflexibleconstructionofpredictivemodelsofdrugresponse.Thefeaturesofdrugsandcellsareintroducedintotheencoder.Subsequently,theencodedinfor-mationisintegratedwithinthefusionmoduletogeneratethepredicteddrugresponse(see

Fig.1A

).

2.2.1Drugencoder

DeepDRintegratesnineencoderstailoredtoprocessdrugmoleculardata(see

Fig.1B

).Theseencodersincludethe

Drugencoder

Fusionmodule

Drugfeaturization

Cellencoder

IC50/AUC/ActArea

Valid

Test

Train

Cellfeaturization

B

C

rization

EXPPES

MUT/CNV

A1

A2

B3

C3

A1

A2

B3

C3

A1

A2

B3

C3

A1

A2

B3

C3

Leavecellout

Leavedrugout

Strict

Random

1.Drugfeaturization2.Cellfeatu

FP

SMILES

Graph

C1=C(C(=O)NC(=O)N1)F

FH

N

()

ONO

H

5.Fusionmodule

CNN

DNN

3.Drugencoder

GNNs

DNN

feature

Drug

GRU/LSTM

MHA

CNN

DAE

4.Cellencoder

Cellfeature

DNN

D

01fromDeepDRimportData,Model,CellEncoder,DrugEncoder,FusionModule

02data=Data.DrData(Data.DrRead.PairDef('CCLE','ActArea'),'EXP','Graph').clean()

03train_data,val_data,_=data.split('cell_out',fold=1,ratio=[0.8,0.2,0.0],seed=1)

04train_loader=Data.DrDataLoader(Data.DrDataset(train_data[0]),batch_size=64,shuffle=True)

05val_loader=Data.DrDataLoader(Data.DrDataset(val_data[0]),batch_size=64,shuffle=False)

06model=Model.DrModel(CellEncoder.DNN(6163,100),DrugEncoder.MPG(),FusionModule.DNN(100,768))

07result=Model.Train(model,epochs=100,lr=1e-4,train_loader=train_loader,val_loader=val_loader)

08data.pair_ls=[['CAL120','5-Fluorouracil'],['CAL51','Afuresertib']]

09result=Model.Predict(model=result[0],data=data)

E

Figure1.OverviewofDeepDRlibrary.(A)Thedrugandcellareprocessedthroughfeaturizationandencoder,andthenthedrugresponseisdecoded

usingthefusionmodule.(B)DeepDRprovidesdrugandcellfeaturization,encoder,andfusionmodule.(C)DeepDRprovidessplittingmethods,includingrandomsplit,leave-cell-outsplit,leave-drug-outsplit,andstrictsplit.(D)ProgrammingframeworkofDeepDRfordatasetloading,modelimplementation,training,andinference.(E)Leave-cell-outperformanceontheCCLEdataset.Usingsubsetmeansusingfeaturesscreenedonthegenesubset,rather

thangenome-widefeatures.Thevaluesinparenthesesarestandarddeviations.

Drugresponsepredictionlibrary3

DNN(DeepNeuralNetwork)leveragingmolecularfinger-prints,andarchitecturessuchasCNN(ConvolutionalNeuralNetwork)

(Liuetal.2019

),GRU(GatedRecurrentUnit)

(DeyandSalem2017

),andLSTM(LongShort-TermMemory)

(GravesandGraves2012

)thatarebasedonSMILESrepresentations.Inaddition,itfeaturesGCN(GraphConvolutionalNetwork)

(Zhangetal.2019

),GAT(GraphAttentionNetwork)

(Velickovicetal.2017

),MPG

(Lietal.

2021c

),AttentiveFP

(Xiongetal.2020

),andTrimNet(

Li

etal.2021b

)foranalyzingmoleculargraphs.TheDNNmod-uleencodesthedrugasasingularvector,whiletheotherarchitecturesproduceasequenceofvectors,witheachvectorcorrespondingtoaSMILEScharacteroranatomwithinthemoleculargraph.TheencodersbasedonSMILESandmolec-ulargraphsareintegratedwithanembeddinglayer,whichisinstrumentalingeneratingdensevectors.

2.2.2Cellencoder

Forcellmodeling,DeepDRintegratesnineencoders:DNNbasedonEXP,PES,MUT,orCNV

(Lietal.2021a

);CNNbasedonEXP,PES,MUT,orCNV

(Manicaetal.2019

);andDAE(DenoisingAutoencoder)basedonEXP

(Chenetal.

2022

)(see

Fig.1B

).TheDNNandCNNmodulesaredesignedtocompressthefeaturesofcellsintolow-dimensionalvectors,thusfacilitatingamorecompactandefficientrepresentationofthedata.TheDAE,ontheotherhand,isspecificallypre-trainedtofocusonminimizingthereconstructionlossofcellfeatures,utilizingthehiddenvectorsastheencodingvectorsforthecells.

2.2.3Fusionmodule

Intermsofintegratingdrugandcellinformation,DeepDRprovidestwomethods:aDNNbasedandanMHA(Multi-headAttention)-basedframework(see

Fig.1B

)

(Vaswani

etal.2017

,

Manicaetal.2019

).Thecellencoderisdesignedtoencodethecellasasinglevector,whilethedrugencoderencodesthedrugasasinglevectororseriesofvectors.WithintheDNN-basedframework,aseriesofvectorscanbecondensedintoasinglevectorthroughtechniquessuchasglobalaveragingormaximumpooling.Incontrast,theMHA-basedapproachcalculatesasfollows:

,、

Attention(Q;K;V)=softmaxV(1)

wherethecellvectorisactingasQ.Thedkisthedimensionofvectorsrepresentingthedrug,whichareconsideredasthematricesKandV.Thisleveragestheattentionmechanismtoeffectivelyextracttheinformationoncelldruginteractionsintoonevector.Botharchitecturesshareacommonprocesswherethevectorsforthedrugandcellareeitheraddedorconcatenated,followedbytheirintroductionintoasucces-sionoflinearlayersforthepredictionofdrugresponses.

3ProgrammingframeworkofDeepDR

DeepDRstreamlinestheDRpredictionworkflowintosevenmodularcomponents,eachthoughtfullystructuredasaclassorfunctiontoenhanceconvenience(see

Fig.1D

):(i)UseData.DrDatatoconstructdrugresponsedata,includingcell-drugpairs,correspondingdrugresponses,cellanddrugfeatures.(ii)Use.clean()and.split()tocleanandsplitdrugresponsedata.(iii)InstantiatethedatasetusingData.

DrDataset.(iv)UseData.DrDataLoadertoloadthedatasetformodeltrainingorvalidation.(v)ThenModel.DrModelisutilizedtoconstructtheDRpredictionmodel.(vi)ThemodelistrainedusingModel.Train,whichconcurrentlyevaluatesperformancetoensureefficacy.(vii)Finally,Model.Predictisdeployedtoforecastdrugresponses,leveragingtheknowl-edgegainedfromthetrainedmodel.DeepDRoffersthreekeymetrics:MeanSquaredError(MSE),R-squared(R2),andPearsonCorrelationCoefficient(PCC).

Downloadedfrom

/bioinformatics/article/40/12/btae688/7903283bygueston26December2024

4EstablishingbenchmarksviaDeepDR

Tobenchmarkdrugresponseprediction,weimplementedandevaluated16models,includingtCNNS

(Liuetal.2019

),Precily

(Chawlaetal.2022

),andDeepDSC

(Lietal.2021a

),alongwithother13novelmodels,onCCLEandGDSC2datasets.Weusedleave-cell-outandleave-drug-outsplittingstrategiestosplitthedatasetsintotraining,validation,andtestsets(8:1:1)usingthreerandomseeds.Eachmodelwastrainedfor100epochsusingtheMSElossfunction,withthelearningratetunedfrom{0.001,0.0001,0.00001}.Were-portthemeanandstandarddeviationofmodelperformanceacrossthethreeseeds.Ourfindings

(Fig.1E

and

SupplementaryTablesS1–S3

)highlightthreekeyobserva-tions:(i)optimalrepresentationsaregraphsfordrugsandex-pressionprofilesforcells.(ii)Predictingtheresponseofnoveldrugsisamoresignificantchallenge.(iii)Pre-trainingtechni-quesfacilitateaccuratepredictionofdrugresponse.Furtheranalysisandimplementationdetailscanbefoundin

SupplementaryTextsS5andS6

and

SupplementaryTables

S4–S7

.TheoptimalmodelsdevelopedwithDeepDRareavail-ableonavisualinterfaceat

https://huggingface.co/spaces/

user15632/DeepDR

.

Authorcontributions

ZhengxiangJiang(Methodology,Datacuration,Visualization,Writing—originaldraft,Writing—review&editing),PengyongLi(Conceptualization,Supervision,Investigation,Methodology,Writing—review&editing)

Supplementarydata

Supplementarydata

areavailableatBioinformaticsonline.Conflictofinterest:Nonedeclared.

Funding

ThisworkwassupportedinpartbytheNationalNaturalScienceFoundationofChina[62202353andU22A2037]andtheFundamentalResearchFundsfortheCentralUniversities.

Dataavailability

ThesourcecodeandexperimentaldataareavailableonGitHub:

/user15632/DeepDR

.InstallationofDeepDRinvolvessimplytyping“pipinstalldeepdr.”

References

BaptistaD,FerreiraPG,RochaM.Deeplearningfordrugresponsepre-

dictionincancer.BriefBioinform2021

;22:360–79.

4JiangandLi

BarretinaJ,CaponigroG,StranskyNetal.Addendum:thecancercell

lineencyclopediaenablespredictivemodellingofanticancerdrug

sensitivity.Nature2019

;565:E5–6.

ChawlaS,RockstrohA,LehmanMetal.Geneexpressionbasedinfer-

enceofcancerdrugsensitivity.NatCommun2022

;13:5680.

ChenJ,WangX,MaAetal.Deeptransferlearningofcancerdrug

responsesbyintegratingbulkandsingle-cellRNA-seqdata.Nat

Commun2022

;13:6494.

DeyR,SalemFM.Gate-variantsofgatedrecurrentunit(GRU)

neuralnetworks.In:2017IEEE60thInternationalMidwestSymposium

onCircuitsandSystems(MWSCAS).IEEE,2017

,1597–600.

FreemanJL,PerryGH,FeukLetal.Copynumbervariation:new

insightsingenomediversity.GenomeRes2006

;16:949–61.

GravesA,GravesA.LongShort-TermMemory.SupervisedSequence

LabellingwithRecurrentNeuralNetworks.NewYork,USA:

Springer,2012

,37–45.

HnzelmannS,CasteloR,GuinneyJ.GSVA:genesetvariationanalysis

formicroarrayandRNA-seqdata.BMCBioinformatics2013

;14:7–15.

HellerMJ.DNAmicroarraytechnology:devices,systems,andapplica-

tions.AnnuRevBiomedEng2002

;4:129–53.

JiaP,HuR,PeiGetal.Deepgenerativeneuralnetworkforaccurate

drugresponseimputation.NatCommun2021

;12:1740.

KearnesS,McCloskeyK,BerndlMetal.Moleculargraphconvolu-

tions:movingbeyondfingerprints.JComputAidedMolDes2016

;30:595–608.

LeCunY,BengioY,HintonG.Deeplearning.Nature2015

;521:436–44.

LiM,WangY,ZhengRetal.Deepdsc:adeeplearningmethodtopre-

dictdrugsensitivityofcancercelllines.IEEE/ACMTransComput

BiolBioinform2021a

;18:575–82.

LiP,LiY,HsiehC-Yetal.Trimnet:learningmolecularrepresentation

fromtripletmessagesforbiomedicine.BriefBioinform2021b

;22:bbaa266.

LiP,WangJ,QiaoYetal.Aneffectiveself-supervisedframeworkfor

learningexpressivemolecularglobalrepresentationstodrugdiscov-

ery.BriefBioinform2021c

;22:bbab109.

LiuP,LiH,LiSetal.Improvingpredictionofphenotypicdrugresponse

oncancercelllinesusingdeepconvolutionalnetwork.BMC

Bioinformatics2019

;20:408.

LiuQ,HuZ,Jiang