制藥科學中的AI Artificial Intelligence in Pharmaceutical Sciences

JournalPre-proofs

ArtificialIntelligenceinPharmaceuticalSciences

MingkunLu,JiayiYin,QiZhu,GaoleLin,MinjieMou,FuyaoLiu,ZiqiPan,

NanxinYou,XichenLian,FengchengLi,HongningZhang,LingyanZheng,

WeiZhang,HanyuZhang,ZihaoShen,ZhenGu,HonglinLi,FengZhu

PII:

S2095-8099(23)00164-9

DOI:

/10.1016/j.eng.2023.01.014

Reference:

ENG1255

Toappearin:

Engineering

ReceivedDate:

30September2022

RevisedDate:

11December2022

AcceptedDate:

6January2023

Pleasecitethisarticleas:M.Lu,J.Yin,Q.Zhu,G.Lin,M.Mou,F.Liu,Z.Pan,N.You,X.Lian,F.Li,H.

Zhang,L.Zheng,W.Zhang,H.Zhang,Z.Shen,Z.Gu,H.Li,F.Zhu,ArtificialIntelligenceinPharmaceutical

Sciences,Engineering(2023),doi:

/10.1016/j.eng.2023.01.014

ThisisaPDFfileofanarticlethathasundergoneenhancementsafteracceptance,suchastheadditionofacoverpageandmetadata,andformattingforreadability,butitisnotyetthedefinitiveversionofrecord.Thisversionwillundergoadditionalcopyediting,typesettingandreviewbeforeitispublishedinitsfinalform,butweareprovidingthisversiontogiveearlyvisibilityofthearticle.Pleasenotethat,duringtheproductionprocess,errorsmaybediscoveredwhichcouldaffectthecontent,andalllegaldisclaimersthatapplytothejournalpertain.

?2023PublishedbyElsevierLtd.onbehalfofChineseAcademyofEngineering.

1

Research

SmartProcessManufacturing—Review

ArtificialIntelligenceinPharmaceuticalSciences

MingkunLua,c,JiayiYina,QiZhua,GaoleLina,MinjieMoua,FuyaoLiua,ZiqiPana,NanxinYoua,XichenLiana,FengchengLia,HongningZhanga,LingyanZhenga,c,WeiZhanga,HanyuZhanga,ZihaoShenb,d,ZhenGua,

HonglinLib,d,e,*,FengZhua,c,*

aTheSecondAffiliatedHospital,ZhejiangUniversitySchoolofMedicine&CollegeofPharmaceuticalSciences,ZhejiangUniversity,Hangzhou310058,ChinabShanghaiKeyLaboratoryofNewDrugDesign,EastChinaUniversityofScienceandTechnology,Shanghai200237,China

cInnovationInstituteforArtificialIntelligenceinMedicineofZhejiangUniversity,Alibaba–ZhejiangUniversityJointResearchCenterofFutureDigitalHealthcare,Hangzhou330110,ChinadInnovationCenterforAIandDrugDiscovery,EastChinaNormalUniversity,Shanghai200062,China

eLingangLaboratory,Shanghai200031,China

*Correspondingauthors.

E-mailaddresses:

hlli@

(H.Li),

zhufeng@

(F.Zhu).

ARTICLEINFO

Articlehistory:

Received

Revised

Accepted

Availableonline

Keywords:

Artificialintelligence

Machinelearning

Deeplearning

Targetidentification

Targetdiscovery

Drugdesign

Drugdiscovery

2

ABSTRACT

Drugdiscoveryanddevelopmentaffectsvariousaspectsofhumanhealthanddramaticallyimpactsthepharmaceuticalmarket.However,investmentsinanewdrugoftengounrewardedduetothelongandcomplexprocessofdrugresearchanddevelopment(R&D).Withtheadvancementofexperimentaltechnologyandcomputerhardware,artificialintelligence(AI)hasrecentlyemergedasaleadingtoolinanalyzingabundantandhigh-dimensionaldata.ExplosivegrowthinthesizeofbiomedicaldataprovidesadvantagesinapplyingAIinallstagesofdrugR&D.Drivenbybigdatainbiomedicine,AIhasledtoarevolutionindrugR&D,duetoitsabilitytodiscovernewdrugsmoreefficientlyandatlowercost.ThisreviewbeginswithabriefoverviewofcommonAImodelsinthefieldofdrugdiscovery;then,itsummarizesanddiscussesindepththeirspecificapplicationsinvariousstagesofdrugR&D,suchastargetdiscovery,drugdiscoveryanddesign,preclinicalresearch,automateddrugsynthesis,andinfluencesinthepharmaceuticalmarket.Finally,themajorlimitationsofAIindrugR&Darefullydiscussedandpossiblesolutionsareproposed.

1.Introduction

Inthepastfewdecades,thepharmaceuticalindustryhasbeenlimitedbytheextentofcutting-edgeresearchinpharmaceuticalsciences,becausethedevelopmentofnewdrugsisalongandcomplexprocessaccompaniedbyhighrisksandhighcosts[1,2].Inotherwords,thecurrentfieldofdrugresearchanddevelopment(R&D)requiressignificantproductivityimprovementstoshortenthecycletimeandcostofdrugdevelopment[3].Technologiessuchasnetworkpharmacology,RNA-sequencing(RNA-seq),high-throughputscreening(HTS),orvirtualscreening(VS)haveallacceleratedthediscoveryofnewtargets,aswellasnewdrugstosomeextent[4–9].Nevertheless,thesetechnologieshaverarelybeensignificantcontributorstothecurrentprocessofnewdrugdiscovery.Thus,thereisanurgentneedfornewtechnologytodrivethedevelopmentofnewdrugs.

Asthecomputingpowerofdevicesgrows,artificialintelligence(AI)hasbeenusedinmanyrealcases,suchasinimageclassificationandspeechrecognition,duetoitsabilitytolearn,process,andpredictmassiveamountsofinformation[10–12].Atpresent,afteralongperiodofdataaccumulation,incombinationwiththedevelopmentofhigh-throughputRNA-seqtechnology,massiveamountsofbiomedicaldatahavebeencollected[13–18].Biomedicaldata,whichhasahighlevelofheterogeneityandcomplexity,comesfromavarietyofsources,includingomicsdatafromdifferentplatforms,experimentaldatafrombiologicalorchemicallaboratories,datageneratedbypharmaceuticalcompanies,publiclydisclosedtextualinformation,andmanuallycollateddatafrompubliclyavailabledatabases[19–22].AIcanbeusedtolearnthepotentialpatternsinthesevastamountsofbiomedicaldata,therebybringingnewopportunitiesandchallengestothepharmaceuticalsciencesandindustries.

TheAlphaFold2systemusedAIinthecriticalassessmentofproteinstructureprediction14(CASP14)competitionandoutperformedothersinaccuratelypredictingthethree-dimensional(3D)structuresofproteins[23].Similarly,intheOpen-GraphBenchmarkLarge-ScaleChallenge(OGB-LSC)competition,agraphneuralnetwork(GNN)combinedwithatransformermodelwonthetoprankinpredictingthemolecularpropertiescalculatedbymeansofdensityfunctionaltheory(DFT),whichisdifficultandhighlytime-consumingusingtraditionalmethods[24].ThesecompetitionsdemonstratedthestrongabilityofAItoanalyzebiologicalorchemicaldata.Duetoitspowerfulcapabilitytoutilizerelatedbiomedicaldatatounderstandcomplexbiologicalsystemsandchemicalreactionspaces[25,26],AIhashadarevolutionaryimpactonallstagesofdrugR&D,includingnotonlyresearchonproteinsandsmallmoleculesbutalsotheassisteddesignofclinicaltrialsandpost-marketsurveillance[27].Furthermore,inpharmaceuticalcompanies,manystate-of-the-art(SOTA)AImodelshavebeenadoptedindiversepipelinestoshortentheR&Dcycletimeanddecreasecosts[28–30].

AItechniquesinthiscontextmainlyinvolvemachinelearning(ML)anddeeplearning(DL).BothMLandDLalgorithmsareinvolvedintargetdiscoveryandvalidation[31],drugdiscoveryanddesign[32],andpreclinicaldrugresearch[33],wheretheyareusedtoanalyzedifferentdatacharacteristicsindifferentformats.Afteradrugcandidateisenrolledinaclinicaltrial[34],DLplaysapivotalroleinassistinginthedesignoftheclinicaltrialandinsupervisingandanalyzingdatafromtheclinicalphaseIV[33].Approveddrugshaveastrongimpactonmanufacturing[35]andthemarketeconomy,andDLcanplayapartintheseareasaswell.Therefore,inthisreview,wepresentacomprehensiveoverviewofmostaspectsoftheuseofAIinthepharmaceuticalsciences.WefocusonhowAIcanbeusedtopromotetargetdiscoveryanddrugdiscovery(asshowninFig.1)andreflectonhowtofurtheracceleratethedevelopmentofthisfield.

3

Fig.1.SummaryofAIapplicationsinthepharmaceuticalsciences.ADMET:absorption,distribution,metabolism,excretion,andtoxicity.

2.BasicconceptsofAIanditsscopeofapplication

AIwasfirstproposedattheDartmouthConferencein1956andwasdefinedasanalgorithmthatgivesmachinestheabilitytoreasonandperformfunctions[36].Fromperceptualmachinestosupportvectormachines(SVM)andartificialneuralnetworks(ANNs),thedevelopmentofAIhasgonethroughseveralupsanddowns,andiscurrentlyflourishingthankstothehardwaresupportthatisnowavailable.BothMLandDLfallunderthecategoryofAI;strictlyspeaking,DLcanbeplacedwithinthecategoryofML.However,ourdiscussionofMLinthisreviewonlyconcentratesontraditionalMLmethods,suchasrandomforest(RF)andSVMs.

2.1.Thebigdataera

Inthecurrentbigdataera,giganticamountsofbiologicalandclinicaldatahavelaidafoundationfortheapplicationofAIinthefieldofmedicalandpharmaceuticalresearch.AlthoughAIhasbeensuccessfullyandeffectivelyappliedinmultipleaspectsofthedrugR&Dprocess,thequantityandqualityofmedicaldatahavebecomeoneofthemainobstaclestothedevelopmentofAIinthepharmaceuticalsciences.Thusfar,pharmaceuticaldatabaseswithdetailedandstructuredbigdataproposedbymedicinalresearchersworldwideareplayingakeyroleinpromotingAIapplicationsinmedicalandpharmaceuticalresearch.

Forexample,thetherapeutictargetdatabase(TTD)includesthemostcomprehensiveinformationaboutknownand

4

Proteins

Genes

Drugs/drug

targets

Diseases

RCSB

PDB

PRIDE

UniProt

InterPro

VARIDT

Ensembl

UCSC

Genome

GEO

GenBank

RefSeq

EA

TTD

ChEMB

L

PubChe

m

DrugBank

DrugMAP

DTC

PHARO

S

TCGA

DisGenNET

ClinVar

OMIM

PDBcontains3Dstructuraldataoflargebiologicalmolecules,suchasproteinsandnucleicacids

PRIDEisapublicdatarepositoryforproteomics,includingproteinandpeptideidentifications,post-translationalmodificationsandsupportingspectralevidence

UniProtisaproteindatabasecontainingproteinsequences,functionalinformation,andanindexofresearchpapersInterProprovidesfunctionalanalysisofproteinsbyclassifyingthemintofamiliesandpredictingdomainsandimportantsitesVARIDTprovidescomprehensivedataonallaspectsofdrugtransporters’variability

Ensemblprovidescentralizedgenomicdataandpowerfulfunctionalitiessuchasgeneannotationandregulatoryfunctionpredictions

TheUCSCGenomebrowseroffersaccesstogenomesequencedatafromavarietyofvertebrateandinvertebratespeciesandmajormodelorganisms

TheGEOisadatabaserepositoryofhigh-throughputgeneexpressiondataandhybridizationarrays,chips,andmicroarraysGenBankisanannotatedcollectionofallpubliclyavailableDNAsequences

RefSeqprovidesseparateandlinkedrecordsforthegenomicDNA,genetranscripts,andcorrespondingproteinsformultipleorganisms

EAcollectsbaselinegeneexpressiondatafordifferentspeciesandcontexts,andcontainsdifferentialstudiesreportingexpressionchangesundertwodifferentconditions

TTDincludesthemostcomprehensiveinformationaboutknownandexploredtherapeuticproteinandnucleicacidtargetsChEMBLisamanuallycuratedlibraryofbioactivecompoundswithdrug-likeproperties

PubChemcoverscollectiveinformationonchemicalmoleculesandtheiractivitiesinresponsetobiologicalassaysDrugBankcombinescomprehensivedrugtargetinformationwithspecificdrugdata

DrugMAPprovidesacomprehensivelistofinteractingmoleculesfordrugs/drugcandidates,includinginformationondifferentialexpressionpatterns

DTCenablestheexplorationofbioactivitydata,theprocessingofnewbioactivitydata,anddatacurationinordertoimprovetheunderstandingofDTIs

PHAROSprovidesacomprehensive,integratedknowledgebaseforthedruggablegenome

TCGAhasover2.5petabytesofgenomic,epigenomic,transcriptomic,andproteomicdatarelatedtothecancergenomeDisGenNETcontainslarge,publiclyavailablecollectionsofgenesandvariantsassociatedwithhumandiseasesClinVarisapublicarchiveofreportsonrelationshipsamonghumanvariationsandphenotypes,withsupportingevidenceOMIMisanonlinecatalogofhumangenesandgeneticdisorders

[43]

[44]

[18]

[45]

[46,4

7]

[48]

[49]

[50]

[51]

[52]

[53]

[37]

[54]

[17]

[55]

[56]

[57]

[58]

[59]

[60]

[61]

[62]

exploredtherapeuticproteinandnucleicacidtargets,thetargeteddisease,pathwayinformation,andthecorrespondingdrugsdirectedateachofthesetargets.Itprovidesdetailedknowledgeofthefunctionsoftargets,aswellastheirsequence,3Dstructures,ligand-bindingproperties,relevantenzymes,andcorrespondingdruginformation[37].PubChem[17]providescollectiveinformationofchemicalmoleculesandtheiractivitiesinresponsetobiologicalassays,includingmolecularstructure,identifiers,physicochemicalproperties,patentinformation,andmoleculartoxicity.Somepopulardatabasesaimedatvariouspharmaceuticalissueshavebeenproposedandarefrequentlyused;theseplaysignificantrolesinpromotingtheapplicationofAIinmedicalandpharmaceuticalresearch[38–42].Summarizingvariouspopularpharmaceuticaldatabases,Table1[17,18,37,43–62]providesbriefinformationonpopularpharmaceuticaldatabases,categorizedintoprotein-related,gene-related,drug-related,anddisease-relateddatabases.

Table1

Pharmaceuticaldatabasesfocusingonproteins,genes,drugs/drugtargets,anddiseases.

FocusDatabaseDescriptionRefs.

PDB:proteindatabank;PRIDE:proteomicsidentificationdatabase;GEO:geneexpressionomnibus;EA:expressionatlas;DTC:drugtargetcommons;DTIs:drug–targetinteractions;TCGA:thecancergenomeatlas;OMIM:onlinemendelianinheritanceinman.

5

2.2.MLandDL

Unliketraditionalcomputerprogrammingcalculations,MLandDLcanlearnpotentialpatternsfromtheinputdatawithoutexplicitprogramming.Theyarenotlimitedbytheformatoftheinputdata,whichisbroadandcanincludetext,images,sound,andmore(alltypesofdatathatcanbeencoded)[63].Similartothehumanlearningmodel,MLandDLcangraduallyrecognizedifferentfeaturesofthedata,inferthepatternslyingwithin,andupdatetheirmodelparametersthroughcontinuousiterationsuntilavalidmodelisformed.

Accordingtotheapplicationscenarios,themodelscanbecategorizedintoregressionmodelsandclassificationmodels.Thedifferencebetweenclassificationandregressiontasksliesmainlyinwhetherthetypeofoutputvariableiscontinuousordiscrete.ChengandNg[64]appliedMLapproachestopredictthebiologicalactivityofper-andpolyfluorinatedalkylsubstances(PFAS)withanoutputofcontinuousvalues,andthisstudyisatypicalregressiontask.Hongetal.[65]builtaDLmodeltopredictwhetheraproteininabacteriumisoftheT4SEtype,withanoutputofdiscretevalues(e.g.,0/1),andthisstudyisatypicalclassificationtask.

Dependingonthetypeoflearningalgorithmrequiredtosolvetheproblem,modelsareconceptualizedintothreecategories:supervisedlearning,unsupervisedlearning,andreinforcementlearning.Supervisedlearningisalabeled-data-drivenprocessthattrainsamodelontherelationshipbetweeninputanditsprespecifiedoutputinordertopredictthecategoriesorcontinuousvariablesoffutureinput.Incomparison,unsupervisedmethodsareusedforidentifyingpatternsinunlabeleddatasetsandexploringadataset’spotentialstructurestoallowclusteringofthedataforfurtheranalysis.Inaddition,semi-supervisedlearningispart-waybetweensupervisedandunsupervisedlearning;itacceptsonlypartofthelabeleddatatodevelopatrainingmodelandisusedasapotentialsolutionforproblemsthatlackhigh-qualitydata[66].Reinforcementlearningperformsmodelconstructionthroughconstantinteractivelearning,relyingonpenaltiesforfailureorrewardsforsuccess.

2.3.IntroductiontodifferenttypesofML/DL-basedalgorithms

MLandDLmethodshavebeensuccessfullyappliedtosolverelevantbiomedicalproblems,withtheadoptedmodelingapproachvaryingfordifferentproblemsoreventhesameproblems.Forexample,smallmoleculesusedtobecharacterizedasengineeredfeaturesfordirectloadinginseveralMLmethodstopredicttheproperties;however,morerecently,GNNscanalsobeutilizedtodescribesmallmoleculesforpredictionsofproperties[67].Determiningthefunctionannotationsofproteinsisessentialfortheselectionofdruggableproteinsaspotentialtargets.Maxatetal.[68]conductedaconvolutionalneuralnetwork(CNN)toannotatethegeneontologyannotation(GOA)ofproteins.Nadavetal.[69]builtarecurrentneuralnetwork(RNN)forproteinfunctionannotations,andXiaetal.[70]combinedbothaCNNandRNNtopredictthegeneontology(GO)labelofproteins.

MLbuildsaspecialalgorithm—notaspecificalgorithm—thatfocusesonthefeaturesofthedataandtransformsthemintoknowledgethatmachinescanreadtoprovidehumanswithnewinsights.Variouscommonalgorithmsexistforresearcherstochoosefrom.Thena?veBayes(NB)algorithmisaprobabilistic-basedclassifierbasedonBayes’theoremandindependenceassumptionsbetweenfeatures;itisasimpleandintuitivealgorithm[71].AnRFalgorithmconstructsasetofunrelateddecisiontreesthatformawholehierarchicalstructure;undermodelconstruction,eachtreeisindividuallyresponsibleforacorrespondingproblem[72].Thefinaldecisionisbasedonthemajorityvotesofthedecisiontrees.Modelsthatmakedecisionsbasedonthisapproacharealsocommonlyreferredtoasensemblemodels.eXtremegradientboosting(XGBOOST)isascalableMLalgorithmbasedongradientboosting,whichisalsoanensemblemodel[73].Multi-layerperceptron(MLP)canbeviewedasadirectedgraphconsistingofmultiplenodelayers,eachfullyconnectedtothenextlayer,sothatitmapsasetofinputvectorstoasetofoutputvectors.SVMisoneofthemostwidelyappliedMLalgorithms.Anoptimalhyperplaneisusedtoclassifysamples,whichareobtainedbymaximizingthemarginsbetweendifferentclassesinaspecificdimensionalspace,withthedimensionalitybeingdeterminedbythenumberoffeatures[74].K-nearestneighbor(KNN)isregardedas“l(fā)azylearning”thatclassifiesthesampleaccordingtoonlyafewneighboringsampleswhendistinguishingbetweencategories[75].Inadditiontotheabovemethods,severalotherMLmethodssuchasprincipalcomponentanalysis(PCA),partialleast-squares(PLS),lineardiscriminantanalysis(LDA),andlogisticregression(LR)havebeenappliedinbiomedicaldataprocesses[76,77].

DLispopularduetoitspowerfulgeneralizationandfeature-extractioncapabilities;itslearningandpredictionprocessisend-to-end.UnlikethetraditionalMLprocess(whichoftenconsistsofmultipleindependentmodules),DLobtainstheoutputdata(output-end)directlyfromtheinputdata(input-end)duringthemodeltrainingprocessandcontinuouslyadjustsandoptimizesthemodelbasedontheerrorbetweentheoutputandthetruevalue,untilitmeetstheexpectedresult.Adeepneuralnetwork(DNN)isafeed-forwardneuralnetworkconsistingofdenselyconnectedinput,hidden,andoutputlayers.Itachievesthefeaturelearningofinputdatabysimulatingnonlineartransformationsbetweenneurons,witheachlayerconsistingofvariousneurons[78].ACNNisafeed-forwardneuralnetworkthatconsistsofconvolutional(featureextraction)andpooling(dimensionalityreduction)layers.Theconvolutionalandpoolinglayershelptoextractalltheinformationinadatasetwithout

6

consumingtoomuchtimeandcomputationalresources[79].AnRNNisaclassofANNinwhichlinkednodesformadirectedorundirectedgraphalongatemporalsequence.AnRNNincludesafeedbackcomponentthatallowssignalsfromonelayertobefedbacktothepreviouslayer.Itistheonlyneuralnetworkwithinternalmemory,whichhelpstoaddressthedifficultyoflearningandstoringlong-terminformation[80].AGNNisaconnectivitymodelthatderivesthedependenciesinagraphbymeansofinformationtransferbetweennodesinthenetwork[81,82].AGNNupdatesthestateofanodeaccordingtoneighborsofthenodeatanydepthfromthenode;thisstateisabletorepresentthenodeinformation.TheneuralnetworkarchitecturesofthefournetworksdescribedaboveareshowninFig.2.

Anautoencoder(AE),whichconsistsofanencoderandadecoder,isusedtolearnefficientencodingsofinputdata.Theencoding,whichisgeneratedbyfeedinginputtotheencoder,regeneratestheinputbythedecoder.AnAEisusuallyusedfordatacompressionanddimensionalityreductionthroughtherepresentationmethods(i.e.,theencoding)ofasetofdata[83].Agenerativeadversarialnetwork(GAN)iscomposedoftwounderlyingneuralnetworks:ageneratorneuralnetworkandadiscriminatorneuralnetwork.Theformerisusedtogeneratecontent,whilethelatterisusedtodiscriminatethegeneratedcontent[84].Modelscanalsobeusedincombinationtosolveawiderrangeofproblems.Forexample,agraphconvolutionnetwork(GCN)extendsconvolutionaloperationsfromtraditionaldata(e.g.,images)tographdata[85].

Fig.2.SchematicnetworkarchitecturesforaDNN,GNN,CNN,andRNN.

Whenamodelfailstolearntheunderlyingpatternsindatafeatureseffectivelyandlosestheabilitytogeneralizetonewdata,suchaproblemiscalledmodelunderfitting[86].Incontrast,overfittingoccurswhenthemodelistrainingandnoisein

7

thedatafittedasarepresentativefeatureresultinginpoorpredictionsfornewdata[87].Comparedwithunderfitting,modeloverfittingismoredifficulttodealwith.Modelsoftenbecomeoverfittedduetobeingoverlycomplexorbecauseofanunderrepresentationofdata.Adatasetusedforamodelisoftendividedintoatrainingset,validationset,andtestset.Thesesetsarerespectivelyusedformodeltraining,modeladjustment,andmodelevaluation.Toputitsimply,amodelthatworksbadlyonboththetrainingandtestsetsisanunderfittedmodel,whileamodelthatworkswellonthetrainingsetbutbadlyonthetestsetisanoverfittedmodel.Typicalwaystosuppressoverfittingincluderegularization,dataaugmentation[88],dropout[89],earlystopping,ensemblelearning,andamongothermethods.

Researchersencounteredunderfittingandoverfittingproblems,usingonlyonemodeloftraditionalepidemicmodelsorMLmodels,whenpredictingthelong-termtrendsofthecoronavirusdisease2019(COVID-19)pandemic.Toaddresstheseissues,Sunetal.[90]proposedanewmodelcalleddynamic-susceptible-exposed-infective-quarantined(D-SEIQ).TheD-SEIQmodelcanaccuratelypredictthelong-termtrendsofCOVID-19outbreaksbyappropriatelymodifyingthesusceptible-exposed-infective-recovered(SEIR)modelandintegratingML-basedparameteroptimizationunderreasonableepidemiologyconstraints.

Differentmodelshavedifferentevaluationcriteria.Inregressionmodels,commonlyusedevaluationcriteriaincludemeansquarederror(MSE),rootMSE(RMSE),andRsquared.Inclassificationmodels,themorecommonlyusedcriteriaarerecall,precision,andF1score.Thereceiveroperatingcharacteristic(ROC)curveandprecision-recallcurve(PRC)arethemostcommonlyusedevaluationcriteriainclassificationmodels,withROCcurvestakingintoaccountbothpositiveandnegativecasestoassesstheoverallperformanceofthemodel,whilePRCsfocusmoreonpositivecases[91].

2.4.Abriefdescriptionofmoleculerepresentationasmodelinput

Overtime,theaccumulationofdataonsmallmoleculesandproteinshasresultedinanextremelylargedataresource.Databasesofmolecularsequences,structures,physicochemicalproperties,andsoforthhavebeencollectedandorganizedbydifferentorganizationsandcontainagreatdealofknowledgeandinformation.However,thedifferentsourcesandformatsofthedatamakeitdifficulttointegratethecorrelateddatafrommultipleheterogeneoussources.Therefore,itisparticularlyimportanttoadoptsuitablemethodstorepresentmoleculesinanappropriatewayandtominethecrucialinformationinthedataonmoleculesbymeansofAI[92].CurrentAIalgorithmsarehighlydependentonthequalityofthedata;thus,whenperformingmodelconstruction,itisnecessarytounifytheinputformatofmolecules,suchasbyrepresentingsmallmoleculesandproteinsasmodel-readablevectorsormatrices.

Atpresent,therepresentationofsmallmoleculesisgenerallydoneusingoneoffourmainapproaches.Thefirstapproachinvolvesknowledge-basedrepresentation.MoleculardescriptorsandmolecularfingerprintsbasedonhumanaprioriknowledgearewidelyusedinvariousMLorDLalgorithms[93].Thesecondapproachinvolvesdirectrepresentationbasedonimages.CNNshavenowbeenusedtolearnrulesfromtwo-dimensional(2D)digitalimages.A2DchemicaldigitalgridofamoleculecanbedirectlyusedasinputtoallowaCNNmodeltolearnthepropertiesofthemolecule[94].Thethirdapproachisstring-basedrepresentation.Forexample,atypicalcanonicalsimplifiedmolecular-inputline-entrysystem(SMILES)representssmallmoleculesintheformofstrings.Thus,CNNsandRNNscanbefurtherusedtolearnmolecularembeddingsfromthestringrepresentationsofchemicalstructures[95–97].Thefourthapproachinvolvesgraph-basedfeaturerepresentation.Representationmethodsbasedongraphconvolutionorgraphattentionhavebeenwidelyusedtoexplorethefeaturerepresentationofsmallmolecules.Inthesemethods,atomsandbondsareconsideredtobenodesandedges,respectively,whilenewmolecularrepresentationsareobtainedduringthecontinuousupdatingofinformationatindividualnodes.Graph-basedrepresentationshaveachievedoutstandingperformanceinavarietyofpharmaceuticallearningtasks[98,99].

Proteinrepresentationmethodscanbebasicallyclassifiedintofourcategories:representationbasedonintrinsicpropertiesofsequences,representationbasedonphy