Canonical correlation analysis for RNA-seq co-expression networks

RNA-SEQ

Publishedonline4March2013NucleicAcidsResearch,2013,Vol.41,No.8e95

doi:10.1093/nar/gkt145

Canonical

co-expression典型的correlationanalysisforRNA-seqnetworks

ShengjunHong1,2,XiangningChen3,LiJin1andMomiaoXiong1,2,*

1StateKeyLaboratoryofGeneticEngineeringandMOEKeyLaboratoryofContemporaryAnthropology,SchoolofLifeSciencesandInstitutesofBiomedicalSciences,FudanUniversity,Shanghai200433,China,2DivisionofBiostatistics,HumanGeneticsCenter,UniversityofTexasSchoolofPublicHealth,Houston,TX77030,USAand3VirginiaInstituteforPsychiatricandBehavioralGenetics,VirginiaCommonwealthUniversity,Richmond,VA23298-0126,USA

ReceivedNovember28,2012;RevisedFebruary1,2013;AcceptedFebruary13,2013

ABSTRACTtheidenti?edgeneticvariationinassociationstudiesDigitaltranscriptomehasstillbeenlimited(1).Geneexpressionvariationmaysequencing转analysisbynext-generation

discovers录组substantialmRNAsigni?cantlycontributetophenotypevariation(2).Gene

Variationingeneexpression大量的variants.

underliesexpressionanalysesareimportantsourcestostudy

gicalprocessesandholds成manybiolo-

a为key基础tounravellingfunctionofgeneticvariationandareincreasingly

acquiringanimportantroleinunravellingmechanismofmechanismofcommondiseases.However,解决thecomplextraits.Therapidly解决

currentmethodsforconstructionsequencing特性developednext-generationtechnologieshavebeenbecomingthenetworksusingoverallgene构造ofco-expression

expressionareorigin-platformofchoiceforgeneexpressionpro?ling.

allydesignedformicroarrayexpressiondata,andRNA-seqforexpressionpro?lingofferscomprehensivetheyoverlookalargenumberofpictureoftranscriptome提供合的expressions.忽略variationsingene

Touseinformationonexon,genomicplatforms.RNA-seq转录组andissuperior

hasmadea优number秀的to综microarrayofsigni?cant

positionallevelandallele-specificexpressions,qualitativeandquantitative

develop位置wenovelcomponent-basedsion定性improvementsongeneexpres-

analysis的and定provides量的multiplelayersofresolutions

andbivariatecanonical组件methods,single

correlationanalysis,forcon-andtranscriptomecomplexity:theexpression分at辨exon,率

structionofco-expressionnetworkswithRNA-seqsingle-nucleotidepolymorphism(SNP)andpositionaldata.Toevaluatetheperformanceofourlevel;splicing;

forco-expression评估methodsnetworkinferencetheentire拼接post-transcriptionalRNAeditingacross

gene;isoformandallele-speci?cexpressions

(ASE)(1,3–5).data,theyareappliedtolungsquamous推论withRNA-seq

cellcancerVariationincomplexphenotypesisnotcausedbya

expressiondatafromTCGAdatabaseandoursinglegeneactingasamarker,butbyasetofinteracted两bipolardisorderandschizophreniagenesthatareoftenorganizedintovarioustypesofbio-The级的preliminary扰乱的results精神分裂症RNA-seqstudy.logicalnetworks(6).Geneco-expressionnetworksareco-expression初步的demonstratethatthenetworksconstructedbycanonicaloftenusedtoextractimportantinformationcorrelationanalysisandRNA-seqdataproviderichgroupsofco-regulated提取about

genesthatplayacentralrolein

geneticandmolecularinformationtogaininsightregulatoryprocesses.Co-expressionnetworksareabletointobiologicalprocessesanddiseasemechanism.洞察comprehensivelycapturetherelationshipsofindividualOurnewmethodssubstantiallyoutperformcomponentsofthe获transcriptome得perturbedbyenviron-currentstatisticalmethods大量的forco-expression胜过thements(7);hence,theyprovideapowerful扰乱tooltogainnetworkconstructionwithmicroarrayexpressionnewinsightsintothefunctionofgenes,biologicaldataorRNA-seqdatabasedonoverallgeneexpres-processes,theglobalstructureofthetranscriptomeand

mechanismofcomplexdiseases(6,8–11).

sionlevels.Traditionalstatisticalmethodsforconstructionof

co-expressionnetworks,suchasweightedco-expression

INTRODUCTIONnetworks,mutualinformationrelevance

anceselection共同andsparsegraphical关联networks,covari-

model,and共变异

Despitegreatprogressingeneticstudiesofcomplexcorrelationmethodsaremainly图解designedformicroarray局partial部的diseaseshasbeenmade,informationonthefunctionofexpressiondata(12–15).Allthesemethodsuseasingle*Towhomcorrespondenceshouldbeaddressed.Tel:+17135009894;Fax:+17135000900;Email:momiao.xiong@http://wendang.chazidian.com

ßTheAuthor(s)2013.PublishedbyOxfordUniversityPress.

ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommonsAttributionNon-CommercialLicense(http://wendang.chazidian.com/licenses/by-nc/3.0/),whichpermitsunrestrictednon-commercialuse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.Downloaded from http://wendang.chazidian.com/ at Harbin Medical Uniersity on May 1, 2013

RNA-SEQ

e95NucleicAcidsResearch,2013,Vol.41,No.8

valueofsummarizingstatistictorepresentgeneexpressionlevelandoverlookallinformationonexpressiondiffer-enceinexons,genomicpositionandalleles.Therefore,althoughRNA-seqdramaticallyincreasesthelevelofbio-logicaldetails(11),westillusethetraditionalstatisticalmethodsforco-expressionnetworkinference,whicharedesignedformicroarrayexpressiondataanddonotef?-cientlyusealloftheinformationcontainedinRNA-seqdata.Tofullyusethecomprehensiveinformationofthetranscriptomeandcaptureexpressionvariationatthelevelofexon,chromosomalposition,alleleandsplicingisoformswhichareprovidedbyRNA-seq,developmentofpowerfulcomputationaltoolsforexpressiondataanalysisishighlydesirable(16).

Inthisarticle,wedevelopcomputationalmethodstoaddresschallengesarisingfromco-expressionnetworkin-ferencewithRNA-seqdata.Toexploreobservedexpres-sionvariationinexonsoringenomicpositionacrossthegenes,weuseanationanalysis(CCA)普ordinary通的

singlevariatecanonicalcorrel-thatquanti?esthebetweenalinearcombinationof量the化

correlation

expressionsatexonlevelsorpositionlevelsinonegeneandanothersuchcom-binationofexpressionsinasecondgenetoconstructco-expressionnetworks.Speci?cally,theexpressionlevelateachexon,orexpressionlevelateachgenomicposition,willbeconsideredasvariables.Theexonexpressionsorgenomicpositional-levelexpressionsoftwogenesformtwolargesetsofvariables.Wewishtostudythoselinearcombinationsofvariablesmosthighlycorrelated.ThegoalofCCAistoseeklinearcombinationsoftwosetsofvariablesthatmaximizethecorrelationbetweentwosetsofvariables.Toachievethis,we?rstidentifythepairoflinearcombinationsthathavethelargestcorrel-ation.Next,weidentifythepairoflinearcombinationshavingthelargestcorrelationamongallpairsuncorrelatedwiththeinitiallyselectedpair,andsoon.Therefore,CCAmeasurestheco-expressionbetweentwogenesthatcantakegenomicpositionandallelelevelsofexpressionsintoconsideration.TomodelASE,wedevelopbivariateCCAtoconstructco-expressionnetworkswithASEdata,allowinglevelsofASEtovaryacrossSNPsandtoconsidercomplicatedpatternsofASEbecauseofallele-speci?csplicingandalternativetranscriptionstartsites(2).BivariateCCAconsidertwosetsofvectorsofmeasurements.Twoallele-speci?cexpressionsateachSNPformavectorfortheSNP.BivariateCCAistoseekafewlinearcombinationsofvectorswithtwoalleles’expressionsthathavethelargestcorrelations.Therefore,twovariateCCAmeasuretheco-expressionsbetweentwogenesthatcanconsiderexpressionsoftwoallelesateachSNP.ToevaluatetheperformanceofCCAforco-expressionnetworkinferencewithRNA-seqdata,theCCAforco-expressionnetworkconstructionisappliedtolungsquamouscellcancer(LUSC)expressiondatafromTCGAandabipolardisorderandschizophre-niaRNA-seqstudy.We?ndthatCCAforco-expressionnetworkconstructionwithRNA-seqdatasubstantially

outperformsthecurrentstatisticalmethods大量的

co-expression胜过

fornetworkconstructionwithmicroarrayexpressiondataoroverallgeneexpressiondata.Aprogramforimplementing实施

thedevelopedCCAfor

PAGE2OF15

co-expressionnetworkconstructioncanbedownloadedfrombioconductor(http://wendang.chazidian.com/)andourlocalwebsitehttp://www.sph.uth.tmc.edu/hgc/faculty/xiong/index.htm.MATERIALSANDMETHODSDataaccess

TheTCGARNA-seqdatasetsarepubliclyavailablefromtheTCGAwebsite(https://tcga-data.nci.nih.gov/tcga/).PathwaysareavailablefromKEGGdatabase(17,18)(http://www.genome.jp/kegg/http://wendang.chazidian.comAmethodforconstructionofgeneco-expressionnetworks

Ageneco-expressionnetworkisconsideredasanundir-ectedgraph,whereageneisrepresentedasanodeandeachedgeconnectingtwonodesisregardedastheco-expressionrelationshipofthetwoconnectedgenes.Constructionofco-expressionnetworksisoftencarriedoutbydetectingthepairwisecorrelationofgeneco-expression.TheCCAistoseekmaximizationofthecorrelationbetweentwolinearcombinationofthevariablesinthedatasets(19).Supposethatwehavepexonsorpositionsinonegeneandqexonsorpositions

inanothergene.Letxð1Þ

jdenotetheexpressionofthej-thexonorthenumberofreadsatthej-thgenomicposition

withinthe?rstgene.Wecansimilarlyde?nexð2Þ

forthesecondgene.LetXð1Þ¼½xð1Þ1,...,xð1ÞT

andj

Xð2Þ¼½xð2Þð2ÞT

p??

we1,...,xassumethatq??.Fortheconvenienceofpresenta-tion,p q.Let

X¼??Xð1Þ ??Æ

Xð2ÞandƼcovðX,XÞ¼Æ1112

Æ:21Æ22Constructionofco-expressionnetworksisimplementedbyseekingmaximizationofcorrelationÞcoef?cientsbetweenlinearcombinationU¼aTXð1forthe?rstgeneandlinearcombinationV¼bTXð2Þ:maxa,bcorrðU,VÞ¼aTÆ12b

aTÆ11abTÆ22b

ð1Þ

Solutionstotheoptimizationproblemeigenvalues??21!??2the2!ÁÁÁ!??2

(1)arethe

oftheRayleighpandtheircorrespondingeigenvectorsofquotientmatrix:

R¼ÆÀ1=2À1

À1=211Æ12Æ22Æ21Æ11:

InRNA-seqdata,weobserveeithermultipleexon

expressionsorsetsofnumberofreadsatgenomicpositionlevelsacrosstwogenes.Theexon-levelorgenomicpositional-levelexpressionsformtwosetsofvari-ablesortwovectorsofvariables.Canonicalcorrelationbetweentwogenesisto?ndthepairoflinearcombin-ationsofthevariablesdeterminedbyaandbsuchthattheircorrelationismaximized.The?rstpairoflinearcombinationiscalledthe?rstpairofcanonicalvariables.Theirlargestcorrelationiscalledthe?rstcanonicalcor-relations.Next,weidentifythepairoflinearcombinations

Downloaded from http://wendang.chazidian.com/ at Harbin Medical Uniersity on May 1, 2013

RNA-SEQ

PAGE3OF15

thathavethelargestcorrelationamongallpairsuncorrelatedwiththeinitiallyselectedpairandcalledthesecondpairofcanonicalvariables,andsoon.The?rstcanonicalcorrelationisequaltothesquarerootofthelargesteigenvalue??1ofthematrixR,thesecondcanonicalcorrelationisequaltothesquarerootofthelargesteigenvalue??2ofthematrixR,andsoon.LetebetheeigenvectorofthematrixRassociatedwiththejeigenvalue??j.Then,thevectorsofcoef?cientsaandbaregivenby

aj¼ÆÀ1

ÆÀ1=2À1=211Æ1222ej,andbj¼Æ22ej,j¼1,...,p:

Ordinarycorrelationcoef?cientcanonlymeasurethelinearrelationshipbetweentwovariables.Whenitisappliedtoquantifyco-expressionbetweentwogenes,exon-levelexpressionsorgenomicpositional-levelexpres-sionsneedtobeaggregatedintooverallexpressions.Theexon-levelandgenomicpositional-levelexpressionvariationinformationcannotbepreservedintheordinarycorrelation.Canonicalcorrelationisextensionofordinarycorrelationbetweentwovariablestocanonicalcorrelationbetweentwosetsofvariables.Therefore,canonicalcorrelationformeasuringco-expressionbetweentwogenesinRNA-seqdatacanmoreaccuratelyquantifythelinearrelationshipbetweentwogenesthantheordinarycorrelation.

LetPkbetheP-valuepoftheteststatistic(20)

Tk¼À½nÀðp+qÞ??P

logð1À??^2i¼k+1

iÞwithdistribution

2ðpÀkÞðqÀkÞ,wherenissamplesizefortestingthenull

hypothesisH0:??k¼...¼??p¼0.Weassignaweighttotheedgeconnectingtwogenes:

P

p??iIðlogPiÞ

w¼i¼1ð2Þ

IðlogPiÞ

i¼1

whereIðlogPÞ¼

??

0P>0:05

ÀlogPP 0:05

.Whenthedenom-inatoriszero,theweightiszero.

ThemethodfordeterminingthethresholdforretaininganedgebytheCCAmethodishardthresholdmethod(8).We?rstrankededgesbytheirweightsfromthelargesttothesmallest.Wethenselectedgesbypre-determinednumberofedgesorpercentageofedges.Inthisarticle,weselected5%ofedgeswithtopweights.

Theedgeswithranklargerthanthresholdareretainedinthenetwork.

Ingeneral,wehavemultiplecanonicalvariablesandcanonicalcorrelations.Tofullyusecanonicalcorrelationstocharacterizetherelationshipsbetweentwosetsofvariables,weintroducetheweightw.Thelargerthecontributiontotheedgeweight,thelargertheeigenvalueorcanonicalcorrelation.Ifthei-thcanonicalcorrelationisnotsigni?cant(P>0.05),itscontributiontotheedgeweightwillbesmall.Therefore,IðlogðpÞÞisgivenvalueof0.Theedgeweightde?nedbyEquation(2)canfullyusecanonicalcorrelationinformationtomeasurethedegreesofco-expressionoftwogenes.NucleicAcidsResearch,2013,Vol.41,No.8e95

BivariateCCAforconstructionofco-expressionnetworks

withASEdata

WedevelopnovelbivariateCCAforconstructionof

co-expressionnetworkswithASEdata.Letxð1Þjandxð2Þj

bethenumberofreadsofthemajorandminoralleleatthej-theSNPinthegene,respectively.Wecansimilarly

de?neyð1Þð2Þð1Þð2Þ

jandyjforanothergene.LetX¼½x1,x1,...,

xðp1Þ,xðð1Þð2Þp2Þ??TandY¼½y1,y1,...,yðq1Þ,yðq2Þ??T

.De?nelinear

combinationsU¼aTXandV¼bTY,wherea¼½??ð1Þ

1,??ð2Þ1,...,??ðp1Þ,??ðp2Þ??Tandb¼½??ð1Þð2Þ1,??1,...,??ðq1Þ,??ðq2Þ??T.Theselinearcombinationscanberewrittenas

U¼½??ð1Þ??TXð1Þ+½??ð2Þ??TXð2ÞandV¼½??ð1Þ??TYð1Þ+½??ð2Þ??TYð2Þ,

where??ð1Þ¼½??ð1Þ1,...,??ðp1Þ??T,??ð2Þ¼½??ð2Þ1,...,??ðp2Þ??T,??ð1Þ

¼½??ð1Þ,...,??ðq1Þ??T,??ð2Þ¼½??ð2Þ1,...,??ðq2Þ??T,Xð1Þ¼½xð1Þ11,...,

xðp1Þ??TXð2Þ¼½xð2Þ1,...xðp2Þ??T,Yð1Þ¼½yð1Þ1,...,yð1ÞT

andYð2Þ¼½yð2Þq??

1,...,yð2ÞT

De?nethecovarianceq??.

matrices:

Ƽ??Æ ÆxxÆ

Æxyyx

yy

,where

Æ??

Æxð1Þxð1ÞÆxð1Þxð2Þ

xx¼Æ,

xð2ÞxðÆ??1ÞÆxð2Þxð2Þ

Æð1Þð1ÞÆð1Þð2Þ

xy¼ÆTxyxy

yx¼ÆandðÆ??x2Þy

ð1Þ

Æxð2Þyð2ÞÆyð1Þyð1ÞÆyð1Þyð2Þ

yy¼

Æ:yð2Þyð1ÞÆyð2Þyð2Þ

TheCCAseekstomaximize

max??,??corrðU,VÞ¼??TÆxy??

ð3where??¼????TÆxx????TÆyy??

Þ??ð1Þ ????

ð1Þ ??ð2Þand??¼??ð2Þ

.Thesolutionstotheoptimizationproblem(3)aretheeigenvectorsofthematrixwiththeeigenvalues??1!??2!...!??2p:

R¼ÆÀ1=2À1À1=2

xxÆxyÆyyÆyxÆxx:

Ourformulationconsidersthecorrelationbetweenthe

expressionsoftwoalleles.Ifwedonottaketheircorrel-ationsintoaccount,thetwovariateCCAwillbecometwosinglevariateCCA.Again,letPTkbek¼À½nÀðp+qÞ??Pthe2pP-valueofthei¼k+1logð1À??^teststatistic(20)

2iÞ

withdistribution 2ð2pÀkÞð2qÀkÞ,wherenissamplesizefortestingthenullhypothesisH0:??k¼...¼??2p¼0.Weassignaweighttotheedgeconnectingtwogenes:

P

2p??iIðlogPiÞ

w¼

i¼1P2pð4Þ

IðlogPiÞ

i¼1

Downloaded from http://wendang.chazidian.com/ at Harbin Medical Uniersity on May 1, 2013

RNA-SEQ

e95NucleicAcidsResearch,2013,Vol.41,No.8PAGE4OF15whereIðlogPÞ¼??0P>0:05

ÀlogPP 0:05.Whenthedenom-SupplementaryFigureS1.Itwasclearlyshownthatthe

inatoriszero,theweightiszero.accuracyofCCAmethodforconstructionofco-SimilartosingleCCAforconstructionofco-expressionexpressionnetworkswasmuchhigherthanthatbynetworks,afterweranktheweights,wealsouserankGLASSOunderalldifferentnetworksizes.Thenweproceduretoprunethenetworks.studiednon-smallcelllungcancerpathwayinKEGG

withLUSCdataset.Afterdiscardingtheisolated

GraphicalLASSOinthepathwayandmatching抛弃nodes

totheTCGALUSC

RNA-seqlevel3data,weincluded44genesinthe

Sparseundirectedgraphicalmodelscanbeestimatedbyanalysis.Theconstructedco-expressionnetworksforthetheuseofL1(LASSO:leastabsoluteshrinkageandnon-smallcelllungcancerpathwaybytheCCAandselectionoperator)regularization(21).WeassumethatGLASSO(21)methodwereshowninFigure1.Edgestheoverallexpressionsofgeneshaveamultivariatewithredcolourwereintheco-expressionnetworkcon-Gaussiandistributionwithmean??andcovariancestructedonlybyCCA.EdgeswithbluecolourwereinmatrixÆ.Itisshowntheco-expressionnetworksconstructedbybothCCAinversematrixÆÀ1thatiftheijthecomponentoftheiszero,thenvariablesiandjareandGLASSO.Edgeswith蓝cyancolourwereintheconditionallyindependent,giventheothervariables.co-expressionnetworksconstructed绿色onlybyGLASSO.Therefore,co-expressionnetworkscanbeconstructedbyItconsistedoffourpathways:ErbBsignallingpathway,estimatingtheinverseofcovariancematrixÆÀ1throughMAPKsignallingpathway,PI3KpathwayandL细apoptosis胞1regularization.pathway.Figure1showedthatEGF(epidermalgrowth凋亡

factor)–EGFR–PI3K/Akt–apoptosissignal表皮pathwayand

MAPK(Raf–MEK–ERK)signalpathwaywereinthe

RESULTSco-expressionnetworkconstructedbytheCCAmethod.CanonicalcorrelationanalysisforconstructionofHowever,EGF–EGFR–PI3KpathwayandMAPKco-expressionnetworkswithexon-levelexpressiondatapathwayconnectionswerenotintheco-expressionAgeneco-expressionnetworkisrepresentedasanundir-networkconstructedbyGLASSO.Recentstudiesun-ectedgraph,whereeachnodedenotes未受指导的covered发现(22)thattheEGFstimulates刺激theproductionofpreciselyageneexpressionpro?le,表ageneormore

and示eachedgeconnect-白interleukin(IL)-8fromlungcancercells,whichinturn

ing精确two的nodesindicatessigni?cantco-expressionrelation-activates细胞介素EGFRandsignallingpathwayofPI3K/Akt.shipsofthetwogenes.ToexploreexonswithvaryingPI3K/Aktpathwayactivationplaysa重crucial要的roleinlungexpressioninformation,wedevelopedaCCA-basedcancerdevelopmentandproliferation.

wasinvolvedingenetranscription,增殖Raf–MEK–ERK

regulationofcellmethodforconstructionofco-expressionnetworks.Thesurvivalandangiogenesisandwasassociatedwithlungcanonicalcorrelationsmeasurethestrengthofassociationmetastasis(23).血管生成

betweenthetwosetsofexonexpression.ToillustratetheNetwork转移topologyplaysanimportantroleintheapplicationofCCAforconstructionofco-expressionfunctionandinformationprocessingofbiologicalnetworkswithexonexpressiondata,theCCAwasnetworks(24).AssortativityandcentralizationaretwoappliedtoLUSCRNA-seqdatasetfromTCGAimportanttopology协measures调集中化database.LUSCconsistsof242samples(225caseisapreference拓扑for学ofnetworks.Assortativityanetwork’snodestoattachtosamplesand17controlsamples).Here,weonlyusedatasimilar倾nodes.向other

Assortativityismeasuredbytheassort-belongingtothecasegroup.TheCCAmethodwasativitycoef?cientthatisde?nedasthePearsoncorrelationcomparedwithagraphicLASSO(GLASSO)(21)coef?cientofdegreebetweenpairsoflinkednodes(25).methodthatusedoverallgeneexpressiontoconstructIftheassortativitycoef?cientwaspositive,thenetworkco-expressionnetwork(fordetail,see‘Materialsandwassaidtobeassortative.Ontheotherhand,iftheMethods’section).Asimplerank-basedcut-offmethodassortativitycoef?cientwasnegative,thenetworkwaswasusedtoprunethenetwork(fordetail,see‘Materialsrecognizedasdisassortative.WeobservethatinandMethods’删section).减networks,highly异connected配的social

nodestendtobeconnectedWe?rstrandomlyselectedTheTCGARNA-seqlevel3withotherhighdegreenodes.Theassortativitycoef?cientdatawith50,200,400,600and1000genes,resultingininsocialnetworksis,hence,positive(25).However,?veRNA-seqdatasets.WethenappliedGLASSOandtechnologicalandbiologicalnetworkstypicallyshowCCAtothese?vesampledRNA-seqdatasetstoconstructthathigh-degreenodestendtoattachtolow-degreeco-expressionnetworksthatwerecalledtheoriginalnodes.Theirassortativitycoef?cientisnegative.For

co-expressionnetworksasthebasistoevaluatetheper-randomnetworks,theassortativitycoef?cienttendedtoformanceofCCAandGLASSOforconstructionofbenearlyzero(Wikipedia,thefreeencyclopedia).Thecen-co-expressionnetworks.Weusedbootstrapping

re-samplingRNA-seqdata1000timesfrom引导程fortralityofavertexeach序of?verelativeimportance顶点withinagraphthatdeterminesthe

ofavertexwithinthegraphisan

RNA-seqdatasets.TheCCAandGLASSOmethodsimportantconceptinnetworktheory.Centralizationiswereappliedtothere-sampledRNA-seqdatatoconstructbasedonthe概念conceptofcentrality.Itisde?nedasco-expressionnetworksfortestingtheaccuracy(howmeasuringthesumindifferencesincentralitybetweenmanyedgesintheoriginalnetworksarereservedinthethemostcentralnodeinanetworkandallothernodes.reconstructedco-expressionnetworksfrom保the留re-sampledItattemptstoquantifythelevelofanetworkabouthowRNA-seqdatasets).Theresultswereshownincentralizeditwasaroundparticularnodes(26).Table1Downloaded from http://wendang.chazidian.com/ at Harbin Medical Uniersity on May 1, 2013

RNA-SEQ

PAGE5OF15NucleicAcidsResearch,2013,Vol.41,No.8e95Figure1.Thesharednetworkstructurebynon-smalllungcancerpathwayinKEGGandreconstructedco-expressionnetworksusingtheCCAandGLASSOmethods.

Table1.Topologypropertyofco-expressionnetworksgeneratedbyexpressionswasmorebiologicallymeaningfulthanthatbyCCA,GLASSO,randomandKEGGoverallgeneexpressions.

MethodAssortativityCentralizationTofurtherevaluatetheperformanceoftheCCAfor

Mean(SD)Mean(SD)constructionofco-expressionnetworks,weappliedboth

CCAandGLASSOmethodstoanotherTCGARNA-seq

CCAÀ0.3937(0.0407)0.7606(0.0078)

GLASSOÀ0.0482(0.0868)0.6297(0.0458)dataset(uterinecorpusendometrioidcarcinoma)where

RandomÀ0.0489(0.0064)0.5666(0.0033)416casesamplesofUCECwereusedinthestudy.KEGGÀ0.23050.7257Thereconstructedpathwayfortheendometrial

bytheCCAandGLASSOmethod子宫were内膜cancer

shownin

SupplementaryFigureS2.Fromthis?gure,itwas

showedthatassortativityandcentralizationofshownthatonlytheCCAmethoddetectedtheMAPKco-expressionnetworksconstructedbytheCCA,(Raf–MEK–ERK)signalpathwayintheco-expressionGLASSO,randomselectionandtruenon-smallcelllungnetwork.Wealsopresentthenetworktopologycompari-cancerpathwayinKEGG,wherestandarddeviationsonoftworeconstructednetworkaswellasrandomcalculatedby1000re-sampling.Weobserved偏差was

thatnetworkandtrue

topologyoftheco-expressionnetworkconstructedbytheSupplementary子endometrioid

Table宫内膜carcinomapathwayin

S1.样的Thistableshowedthatthe

GLASSO(overallgeneexpression)wasclosetotheco-expressednetworkreconstructedbytheCCAmethodrandomnetwork;however,topologyoftheco-expressionwasclosertothepathwayinKEGG,butthetopologyofnetworkconstructedwasclosetothestructureoftrueco-expressionnetworkinferredbytheoverallexpressionsnon-smallcelllungcancerpathwayinKEGG.Table1andGLASSOwasmoresimilartotherandomone.indicatedthattheco-expressionnetworksconstructedby

RNA-seqexonexpressionswerehighlycentralizedandCCAforco-expressionnetworkconstructionwithdisassortative,whichwereinherenttopologyposition-levelRNA-seqdata

thebiologicalnetworks.This固有的featuresinfurtherdemonstratedthatThecurrentmethodsforco-expressionnetworkconstruc-theco-expressionnetworkconstructedbyRNA-seqexontionwithRNA-seqdataareto崩collapse溃倒塌the

position-levelDownloaded from http://wendang.chazidian.com/ at Harbin Medical Uniersity on May 1, 2013