MachineIntelligenceResearch

20(1),February2023,57-78DOI:10.1007/s11633-022-1361-0

MachineLearningforBrainImagingGenomicsMethods:AReview

Mei-LingWang1,2WeiShao1,2Xiao-KeHao3Dao-QiangZhang1,2

1CollegeofComputerScienceandTechnology,NanjingUniversityofAeronauticsandAstronautics,Nanjing211106,China

2KeyLaboratoryofPatternAnalysisandMachineIntelligence,MinistryofIndustryandInformationTechnology,Nanjing211106,China

3SchoolofArtificialIntelligence,HebeiUniversityofTechnology,Tianjin300401,China

Abstract:Inthepastdecade,multimodalneuroimagingandgenomictechniqueshavebeenincreasinglydeveloped.Asaninterdiscip-linarytopic,brainimaginggenomicsisdevotedtoevaluatingandcharacterizinggeneticvariantsinindividualsthatinfluencephenotyp-icmeasuresderivedfromstructuralandfunctionalbrainimaging.Thistechniqueiscapableofrevealingthecomplexmechanismsbymacroscopicintermediatesfromthegeneticleveltocognitionandpsychiatricdisordersinhumans.Itiswellknownthatmachinelearn-ingisapowerfultoolinthedata-drivenassociationstudies,whichcanfullyutilizeprioriknowledge(intercorrelatedstructureinforma-tionamongimagingandgeneticdata)forassociationmodelling.Inaddition,theassociationstudyisabletofindtheassociationbetweenriskgenesandbrainstructureorfunctionsothatabettermechanisticunderstandingofbehaviorsordisorderedbrainfunctionsisex-plored.Inthispaper,therelatedbackgroundandfundamentalworkinimaginggenomicsarefirstreviewed.Then,weshowtheunivari-atelearningapproachesforassociationanalysis,summarizethemainideaandmodellingingenetic-imagingassociationstudiesbasedonmultivariatemachinelearning,andpresentmethodsforjointassociationanalysisandoutcomeprediction.Finally,thispaperdiscussessomeprospectsforfuturework.

Keywords:Brainimaginggenomics,machinelearning,multivariateanalysis,associationanalysis,outcomeprediction.

Citation:M.L.Wang,W.Shao,X.K.Hao,D.Q.Zhang.Machinelearningforbrainimaginggenomicsmethods:Areview.MachineIntelligenceResearch,vol.20,no.1,pp.57-78,2023.

/10.1007/s11633-022-1361-0

1Introduction

Inrecentyears,withthedevelopmentofcognitiveneuroscience,neuroimaginghasbroughtnewvitalitytothestudyoftheworkingmechanismofthehumanbrain.Atthesametime,withthedevelopmentofnoninvasivebrainimagingtechnology,researchershopetogainnewinsightsintotheimagingcharacteristicsandmolecularmechanismsofthebrain,aswellastheirimpactonnor-malanddisorderedbrainfunctionandbehavior.Com-monlyusedbrainimagingtechniquesincludestructuralmagneticresonanceimaging(sMRI),functionalmagneticresonanceimaging(fMRI),diffusiontensorimaging(DTI),andpositronemissiontomographyimaging(PET).Inaddition,withthedevelopmentofgenetictech-nology,researcherscanidentifygeneticmarkersassoci-atedwithneurologicalandpsychiatricdiseasesfromamorerefinedmolecularlevel(suchassinglenucleotidepolymorphisms(SNPs)).

Withrecenttechnologicaladvancesinacquiringmul-

Review

ManuscriptreceivedApril16,2022;acceptedAugust1,2022

RecommendedbyAssociateEditorJun-ZhouHuang

oInstituteofAutomation,ChineseAcademyofSciencesandSpringer-VerlagGmbHGermany,partofSpringerNature2023

timodalbrainimagingdataandhigh-throughputgenom-icsdata,brainimaginggenomicsisemergingasarapidlygrowingresearchfield.HaririandWeinberger[

1

]proposedtheconceptofimaginggenomicsorimaginggenetics,whichperformsintegrativestudiesthatanalysegeneticvariations,suchasSNPs,aswellasepigeneticandcopynumbervariations(CNVs),molecularfeaturescapturedbyvariousomicsdata,andbrainimagingquantitativetraits(QTs),coupledwithotherbiomarker,clinical,andenvironmentaldata[

2

,

3

].

Asanemergingdatascience,brainimaginggenomicshasachievedrapidgrowth,whichisgreatlyattributedtothepublicavailabilityofvaluableimagingandgenomics

datasets.Duetotheopen-sciencenatureoftheAlzhei-

mer'sDiseaseNeuroimagingInitiative(ADNI)project[

4

],hundredsofpublicationsusingADNIimaginggenomicsdatahavebeenproducedinthepastdecade,yieldingin-novativemachinelearningmethodsandnovelbiomedicaldiscoveries.SimilartotheADNI,anincreasingnumberoflandmarkstudiesareproducingbigdata,includingmulti-dimensionalimagingandomicsmodalities,makingthemavailabletotheresearchcommunity.TheseincludetheEnhancingNeuroImagingGeneticsthroughMetaAna-lysis(ENIGMA)Consortium[

5

],PhiladelphiaNeurodevel-

58

opmentalCohort(PNC)[

6

]andParkinson'sProgressionMarkersInitiative(PPMI)[

7

].

Brainimaginggenomicsmainlyusesbrainimagingtechnologytoevaluatethegeneticinfluenceonindividu-alsbyusingbrainstructureandfunctionasphenotypes,andexploreshowgenesaffecttheneuralstructureandfunctionofthebrain,aswellastheresultingneurologic-alpathology.Studyingtheassociationbetweengeneticsandbrainstructureandfunction,andbuildingavisiblebridgebetween“genesandbrain”,canbetterrevealthepathogenesisofneuropsychiatricdiseases[

8

-

10

].Imagingge-nomicscanalsoidentifybiologicalindicatorsoren-dophenotypesofabraindisease,whichprovidesamoreaccuratemethodforpredictinganddiagnosingthedis-ease.Specifically,mostresearchersconsiderSNPsasgen-otypedataforassociationanalysis.Intheacquisitionofendophenotypicdata,researchersmostlyusebrainima-gingdata(i.e.,MRI)inclinicforanalysis.Forexample,sMRI,animagingtechniquethatmeasuresthestructuralorganizationofthebrain,canquantifyabnormalitiesinmorphology(i.e.,graymattervolume).fMRIscanshavebeenshowntobeeffectiveinrevealingfunctionalcon-nectivitypatternsofthebrain.Basedondifferentmodal-itiesofbrainimagingtechnology,atpresent,imagingge-nomicsmainlyfocusesontheassociationanalysisbetween

geneSNPsandbrainstructure,function,andconnectiv-ity[

11

-

14

].

Earlyimaginggenomicsapproachesconsistedofuni-variatepairedstatisticalanalysismethods,wheremul-tipletestsareemployedtofindtheassociationbetweenSNPsorgenesandcomplexdiseasesormeasurablequant-itativetraits(QTs).Genome-wideassociationstudy(GWAS)usesthewholegenomehigh-throughputsequen-cingtechnologytoclassifythesequencevariationinthegenomeoftheresearchobject,andfinallyselectssignific-

antSNPsviathebiostatisticsmethodsandbioinformat-icsmethods

[

15

].SincethefirstGWASresearchpaperonage-relatedmaculardegenerationpublishedinSciencein2005

[

16

],thismethodhasbeenusedintheanalysisofpsy-

chiatricdisorders[

17

].GWAShasplayedagreatroleinthestudyofimaginggenetics,buttherearealsosomeprob-lems,suchasstrictmultiplecorrection,sothatmanysmalleffectvariantscannotpassthecorrectionlevel.Inaddition,GWAScanonlyobtainasingledegreeofassoci-ationbetweengeneticvariationandtraits,andcannotwellexplainthecomplexmolecularmechanismsofthebrain.

Inrecentyears,withtherapiddevelopmentofma-chinelearninginacademiaandindustry,researchershavetriedtousethesedataanalysistoolstosolvesomeprob-lemsinmanyfields.Intheassociationanalysisofima-ginggenetics,inadditiontounivariatestatisticalanalys-is,themultivariatemachinelearningmodelisthemostwidelyused,andithasidentifieddisease-sensitiveima-gingandgeneticbiomarkers.Internationally,someschol-

MachineIntelligenceResearch20(1),February2023

arshavealsowrittenareviewofrelatedmethodsinima-ginggenetics.Forexample,Medlandetal.[

18

]haveraisedtheproblemsandchallengesofusingtraditionalunivari-atestatisticalmodelstoprocesslarge-scalegenome-widebrainimagingassociationanalysis,reviewingthere-searchresultsindifferentcentraldatabases.LiuandCal-houn[

19

]summarizedtheapplicationofothermultivariatemethodssuchasindependentcomponentanalysisinima-ginggenetics.Thompsonetal.[

20

]focusedontheassoci-ationanalysisbetweengeneticsandbrainstructurecon-nectivityandfunctionalnetworks.Basedontheabovere-viewworks,thisarticleisdevotedtoprovidingcompre-

hensiveandup-to-datecoverageofmachinelearning

methodsinbrainimaginggenomics.

Fig.

1

isadoptedtopresentaschematicofthetopicscoveredinbrainima-ginggenomics.Oneofthemaingoalsofimaginggenom-icsbasedonmachinelearningistorealizeassociationanalysisstudiesforunderstandingmechanismsandpath-ways.Wegrouptheseimaginggenomicsbasedonma-chinelearningmethodsintotwocategories.Thefirstcat-egorymainlyusesregressionmodelstoidentifycomplexmulti-SNPand/ormulti-QTassociations.Mostofthere-gressionmodelsdiscussedearliercanbedescribedusingtheregularizedlossfunctionframework.Asparsity-indu-cingregularizationtermisoftenincludedinthesemodels.Themotivationsaretwofold.First,itisreasonabletohy-pothesizethatonlyasmallnumberofmarkersarerelev-antintheresultingimaginggenomicsassociation.Thesparsitytermcanhelpidentifytheserelevantmarkers.Second,thesparsityconstraintcanreducethemodelcomplexityandsubsequentlyreducetheriskofoverfit-ting.Inadditiontoregressionmodels,anothercategoryofprominentmethodsdevelopedforbrainimaginggenom-icsstudiesarecorrelationmodels,suchassparsecanonic-alcorrelationanalysis(SCCA)[

21

-

23

]andparallel-inde-pendentcomponentanalysis(pICA)[

24

,

25

].Similartotheregressionmodeldiscussedearlier,thesparsityisencour-agedinthesecorrelationmodelstoreducemodelcom-plexityandtheriskofoverfitting,aswellasidentifyrel-evantbiomarkers.Overall,thisarticleisfocusedonthethreetypesoflearningproblemsasfollows.First,wewillshowthelimitationsoftheunivariateimaginggeneticsassociationanalysisandshowtheunivariatelearningap-proachesforcorrelationanalysis.Second,wewillpresenttheproblemofmultivariateimaginggeneticsassociationanalysisandsummarizethemainideaandmodellingingenetic-imagingassociationstudiesbasedonmultivariatemachinelearning.Third,wewillreviewmethodsthatareusedtopredictanoutcomeofinterestbycombiningbothimagingandgenomicsdata,andmethodsforjointassoci-ationanalysisandoutcomeprediction.Finally,someun-solvedproblemsingeneticimagingandfutureresearchdirectionsareprospected.

2Univariateanalysismethod

Thestatisticalanalysisofsingle-geneticvariablesusu-

Case-controlstudy-qualitative

trait

Univariateanalysismethod

Associationanalysisinimaginggenomics

Populationbased

study-qualitative

trait

Outcomeprediction

Multivariategenetic-univariateimagingregression

Multivariateanalysismethod

Multivariateimaging-univariategeneticregression

Multivariateregression

Sparsemulti-geneticmulti-imagingregression

Multivariatecorrelation

Regularizedsparsecanonicalcorrelationanalysis

M.L.Wangetal./MachineLearningforBrainImagingGenomicsMethods:AReview59

Chi-squaretestlogisticregression

Univariategenetic-univariateimaginganalysis

Analysisofvariancelinearregression

(Multi-genetics)Structuredsparselinearregression

(Longitudinal/Multimodalimaging)Sparsemulti-taskregression

Multivariategenetic-

multivariate

imaging

Fig.1Schematicoftopicscoveredinbrainimaginggenomics.Thegoalistopresentassociationanalysisinimaginggeneticsbasedonmachinelearning.

allyadoptsthePearson'schi-squaretestfortheexperi-mentalgroupandthecontrolgroupasthealleledetec-tionmethod.Thatis,toconfirmwhetherthelocusisas-sociatedwithageneticriskfactorbyanalysingwhethertherearestatisticaldifferencesbetweenthecorrespond-inggenomiclociofagroupofpatientswithvariousdis-easesandagroupofnormalcontrols.Imaginggeneticsanalysisbasedonunivariatestatisticalmethodscanuse

linearregressionandanalysisofvariancemodelsasallele

associationanalysismethods

[

26

].Inaddition,forthemul-tipleunivariatemodels,firstly,p×qlinearregressionmodels(yj=βjkxk,wherepisthegenefeaturedimen-sionandqistheimagingfeaturedimension)arefitted.Then,p×qnullhypotheses(H0:βjk=0)aretested.Fi-nally,thep-valuesaresortedtoselectthesmallerp-val-ues.Forexample,in2009,Potkinetal.[

27

]performedagenome-wideassociationstudy(GWAS)onpatients,nor-malcontrols,andimagingphenotypes.Thatis,theeffectofSNPsonquantitativephenotypesofbrainareascanbecalculatedbyageneralizedlinearmodel,whichiscon-structedbyimagingphenotypes,diseasediagnosisandgenedata.Theexpressionisasfollows:

Y=b0+b1SNP+b2APOEe4+b3gender+b4age+

b5diagnosis+b6SNP×diagnosis+?(1)

whereYdenotestheneuroimagingQT,birepresentsthecoefficientofeachvariable,andSNP×diagnosisrepresentstheinteractionrelationship.Thep-value

obtained

istheassociationresultbetweenSNPandQT[

27

].

Intheunivariateimaginggeneticsassociationanalysis,

accordingtodifferentscales[

28

],wesummarizeasfollows:forthegeneticlevel,itincludes1)candidategenetics/SNPs[

29

-

32

],2)relatedbiologicalfunctionscharacteristicpathways/networks[

33

-

35

],3)wholegenome[

27

,

36

-

39

].Forthebrainimaginglevel,itincludes1)individualregionsofinterest(ROI)[

27

,

29

,

33

,

36

],2)multipleROI[

30

,

34

,

37

],3)wholebrain[

31

,

32

,

35

,

38

,

39

].WhetheritistheassociationanalysisbetweencandidategeneticlocusSNPandneuroimaging[

40

](cerebrospinalfluid[

41

],cognitivescore[

42

],andanyotherQT),ortheassociationanalysisbetweenwholegenomeandneuroimagingoreventheassociationanalysisbetweenwholegenomeandsmallervoxel-wisebrainimaging,linearregressionandanalysisofvariancecansolvetheproblemsofimaginggeneticsassociationanalysisatdifferentscales.Inaddition,someresearchers

havereleasedrelevantstatisticalanalysissoftware,suchasPlink

[

43

].

GWASgeneticstatisticalanalysisneedstofindtheassociationwithdiseasephenotypesfrommillionsoreven

tensofmillionsofSNPs.AlthoughtheBonferronicorrec-

tioncanbeusedtostrictlycontrolthesignificance

[

44

,

45

],thisstrategywillleadtomanysmalleffectvariationsthatcannotpassthecorrectionlevel,andmultiplesuchsmalleffectvariationsmayacttogethertohaveagreatimpactonthetraits.Theapplicationofunivariateanalysismeth-odsinimaginggeneticshasamoreintuitiveexplanation,andcansimplyandquicklydetecttheassociationbetweenasingleSNPandasingleQT.However,duetothehigh-dimensionalcharacteristicsofdatavariables,alargenumberofmultiplecomparisonseventuallymakethestatisticaltestresultsnotsignificant,andtheabovetestmethodisbasedonastricthypothesis.Thatis,ge-neticlociorimagingcharacteristicvariablesarestatistic-

60

allyindependent,whiletheimportantinformationoftheassociationbetweenvariablesisignored.Therefore,forhigh-dimensionalfeatures,theunivariateapproachstillhassomelimitationsindealingwiththeproblemofima-ginggeneticsassociationanalysis.

3Multivariateanalysismethod

Followingtheunivariatevoxel-wisegenome-wideasso-ciationanalysis(vGWAS)[

39

],Hibaretal.[

46

,

47

]proposedamultivariatevoxel-wisegene-wideassociationstudy(vGe-neWAS),whichsolvestheproblemofvariablecollinear-itybyprincipalcomponentsregression(PCReg)toallSNPsinagenome.Specifically,principalcomponentana-lysis(PCA)wasfirstusedtoobtainthemutuallyortho-gonalfactorsthatmaximizethevarianceontheSNPre-gressionvariableset.Then,thestandardpartialF-test

wasusedontheseorthogonalfactors.Finally,following

therelatedworkproposedbySteinetal.

[

39

]in2010,thesamegeneticandbrainimagingdatasetwereusedtogroupSNPsanddetecttheassociationbetweengroupedSNPswithvoxel-wiseimaging.Experimentalresultsshowthatthismethodachievesbetterassociationperformanceandreducesthenumberofstatisticaltests.Therefore,inordertoenhancetheabilitytodetecttheassociationbetweengeneticsandquantitativetraits(QTs),somere-searchershaveusedmultivariatemethodstoaddresstheassociationofmulti-geneticormulti-locuscombinedef-fectsinimaginggenetics[

19

,

48

].Recently,researchonma-chinelearningbasedimaginggeneticshasattractedmuchattention,whichaimstoidentifytheassociationbetweengeneticsandimagingfeaturesbyusingregressionmodels.Wecanusedifferentcriteriatodividethesemethodsin-toregressionmodels(includingmultivariategenetic-uni-variateimagingregression,multivariateimaging-univari-ategeneticregression,andmultivariategenetic-multivari-ateimagingregression)andcorrelationmodels(i.e.,mul-tivariategenetic-multivariateimagingcorrelation).Inthenextsubsection,severalclassicandstate-of-the-artassoci-ationmodelswillbeintroducedbytheabovedivisionstrategy.

3.1Regressionmodels

3.1.1Multivariategenetic-univariateimagingreg-ression

Weusuallyuseasparseregularizedregressionmodel

torealizemultivariategenetic-univariateimagingregres-sion.Themainmotivationistwofold.First,assumingthatonlyafewmarkersareassociatedwithimaginggen-omics,sparsetermsassisttoidentifytheserelatedmark-

ers.Second,sparseconstraintscanreducethe

complexity

ofthemodelandtheriskofoverfitting.In[

49

,

50

],re-gressionmodelsbasedonL1normpenaltyconstraintshavebeensuccessfullyappliedtomultivariategeneticdataanalysis.TheyaimtoidentifysparseSNPlocithat

MachineIntelligenceResearch20(1),February2023

arehighlyassociatedwithspecificbrainregions.Thesemodelsprovideageneraltechnicalframeworktodealwiththesmallsampleregressionproblemofdetectingandidentifyinghigh-dimensionalgeneticSNPs.However,theconstraintsbasedontheL1normdonotfullycon-siderthestructuralrelationshipbetweenthefeaturevari-ables,thereforetheoptimalregressionresultscannotbeachievedintheory.Consideringthespatialstructurerela-tionshipbetweenSNPfeatures,Silveretal.[

51

-

53

]pro-posedthegroupsparsemodelorfusionsparsemodeltoselectSNPlociinthesamegrouporadjacentfeaturevariables,andthemodelsbasedongroupsparsityorfu-sionsparseareasfollows:

(2)

(3)

miwn∥y?Xw∥+λ√ji)w

miwn∥y?Xw∥+λ|wi?wj|

wherethesetwoequationsareutilizedforidentifyingasetofSNPsfromXandpredictingasingleimagingphenotypey.In(2),wjinthegroupsparsitytermrepresentsalltheSNPlocifeaturesbelongingtothegroupG(i),andthegoalistocontroltheselectedlocitoincludethecharacteristicsofclustering.Forexample,therewillbealinkagedisequilibrium(LD)effect

[

54

]betweengeneloci,thatis,SNPslinkedondifferentgeneswillappearinthesameLDblocknonrandomly.ThisprovidesdomainknowledgeforthefeatureselectionmodelbasedongroupsparsitysothatSNPsinthesameLDgroupcanbedetectedsimultaneously.In(3),thefusionLassotermcancontroltheweightcontributionofadjacentpositionfeatureswiandwjtobeassimilaraspossible,thatis,thefeaturevariablesselectedbythefusionLassotermhavespatialcontinuity.TheempiricalstudywasperformedonanADNIsample.

Inaddition,thereisnotonlyaflatspatialrelation-shipbetweenSNPloci,butalsoahierarchicalrelation-shipintheactualgenestructure.Forexample,inacer-tainpathway,theinteractionofspecificgenelocicanaf-fectproteinsynthesisandfunctionaltransformation,andsomeSNPlociunderthesamegenealsohavecertaincor-relations(suchasLD).Therefore,makingfulluseofthepriorknowledgeofthishierarchicalstructuretoperformtheimaginggeneticanalysiswilloftenreducetheerrorin

theregressionanalysisandlearnmoreexplanatoryfea-

turepatterns

[

55

-

57

],asshownin

.

Fig

2

.Asshown,themodelusesatree-guidedsparselearning(TGSL)methodtoidentifytheassociationbetweengenotypeandpheno-type.Whenconstructingatreestructure,theSNPlociareusedasleafnodes,theLDblockandthegeneblockareusedasintermediatenodes,andallgenesinthepath-wayareusedasthefinalrootnodes.Thestructuretreehasdlayersandeachlayerhasninodes.Thenodeofthe

Candidateset

Rootnode

IntervalnodesGenes

IntervalnodesLDblocks

SNPs

Leafnodes

Feature

d

Time

T

Pathway

M.L.Wangetal./MachineLearningforBrainImagingGenomicsMethods:AReview61

ImagingQT

SVR

prediction

model

Fig.2Tree-guidedsparseregressionmodel,whichaimstoidentifyasetofSNPsforpredictingasingleimagingpheno-type[

55

-

57

].

i-thlayeris{G,···,G,···,Gi},andthetree-guidedsparseregressionmodelisasfollows:

(4)

miwn∥y?Xw∥+λα∥wG∥2

whichalsoaimstoidentifyasetofSNPsforpredictingasingleimagingphenotypey.αistheweightofanynode

Gpredefinedaccordingtopriorknowledge.wGisthe

weightofanynodeGinthelearnedtreestructure.Itisworthnotingthatwhentheweightofanodeiszero,itschildnodesareallzero,thatis,allthefeaturesofthesubtreehavenothingtodowiththeregressiontaskandarenotselected.ComparedwiththetraditionalLassomethod,theSNPsobtainedbytheoptimizationofthemodelhavesmallererrorsinpredictingthegraymattervolumeofthebrain,andtheseSNPlociassociatedwithMRIbrainregionshaveahierarchicalclustering.TheempiricalstudywasperformedonanADNIsampletoidentifysparseSNPpatternsattheblockleveltobetterguidethebiologicalinterpretation.

3.1.2Multivariateimaging-univariategeneticreg-ression

Inresearchonmachinelearningbasedimaginggenet-ics,mostoftheworkshavefocusedondiscoveringanddetectingmultivariateSNPlociassociatedwithimagingphenotypes.However,fewstudieshaveexploredhowSNPvalueschangewhenphenotypicmeasurementvari-ableschange,thatis,usingmultivariateimagingtore-gressunivariategeneticfeatures.Forexample,Shenetal.[

38

]proposedatask-relatedtimeseriesmultivariatesparseregressionmodelbasedonthegroupstructurein-formationbetweenpredictionvariables.Themodelisasfollows:

「

TdlT

(5)

miW