Vol. 19 Suppl. 1 2003, pages i183–i189
Predictingphenotypefrompatternsofannotation
samplesinthedata-set(whichinourcaseisthenumberofgenesinthetrainingset).The rsttermmeasuresthegoodness-of- tofthemodeltothedata,andthesecondtermpenalizesmodel-complexity.Weprunedthetreeinabottom-upfashion,startingattheleavesandworkingtowardtheroot,pruningawayanybranchwhoseremovalcausedthetree’sBICscoretodecrease.IncomputingtheBICscorewetreatedthegenesasindependent,sothatthelikelihoodPr(data|model)factoredastheproductofthelikelihoodforeachgene.(Thismaynotbestrictlytrue,duetohomologybetweengenesforexample.)
Thescoreq(i,j)wasthenjustpN,whereNistheleafthatgeneiendsupatinthedecisiontreeforphenotypej.OnenotabledifferencebetweenourapproachandtheapproachusedinClareandKing(2002)topredictgenefunctionisthattheybuiltasingledecisiontreetopredictwhatcombinationoffunctionsagenehas,whereaswebuildaseparatedecisiontreeforpredictingeachphenotype.
RESULTS
Cross-validation
LetGdenotethesetof6898yeastgeneslistedintheSGDandTthesetof130most-speci cMIPSyeastphenotypes.Weassessedourdecisiontreesusing10-foldcross-validation.WerandomlypartitionedthesetGofyeastgenesinto10setsofequalsize(±1).Foreachofthe10setsofgenes,webuilt130decisiontreesusingtheremainingninesets(combined)astrainingdata.Thenforeachgeneiintheheld-outset,weusedthesedecisiontreestocomputeq(i,j)foreachphenotypejinT.
Thescoresq(i,j)foreachofthe10foldsofthecross-validationwerepooledtogether,andforeachthresholdt∈[0,1]wecomputedthetrue-positiverate
TPt=
#{(i,j)∈G×T:q(i,j)≥t&Xj(i)=1}
#{(i,j)∈G×T:Xj#{(i,j)∈G×T:q(i,j)≥t&Xj(i)=0}
.
#{(i,j)∈G×T:Xjandthefalse-positiverateFPt=
Ofthe6898×130=896510examples(i,j)inthesetG×T,3029werepositive(i.e.hadXj(i)=1)andtheremaining893481werenegative.Thus,forexample,inthetopgraphinFigure1,whenusingMIPSphenotypescombinedwithGOattributesaspredictors,atthepointontheROCcurvewherethetrue-positiverateis0.1,303ofthe3029positiveexamplesarecorrectlyclassi edassuch,and307ofthe893481negativeexamplesaremisclassi ed.
Thereisthepossibilityofsomecircularityinthiscross-validation,sincecertainGOannotationsareinformedbyphenotypeannotations.Whenpredictingwhethergeneihasphenotypej,wedonotlookatwhethergeneiisannotatedashavingphenotypejoranyoftheancestorsofphenotypejintheMIPShierarchy,sinceannotationsfortheseancestralphenotypesmightbederivedfromanannotationforphenotypej,whichwearetryingtopredict.Similarly,theremightbeGOannotationsforgeneithatarederivedfromanannotationforphenotypej,andusingtheseaspredictorsforphenotypejcouldgivearti ciallyin atedperformance.
SGDcuratorsusetenevidencecodeswhenassigningGOannotations(seedoc.guandang.net).Phenotype-derivedSGDGOannotationsshouldusuallyhaveevidencecodeIMP(inferredfrommutantphenotype),TAS(traceableauthorstatement),NAS(non-traceableauthorstatement),IC(inferredbycurator),orNR(notrecorded),asopposedtoothercodessuchasIPI(inferredfromphysicalinteraction).Toreducethein uenceofphenotype-derivedGOannotations,were-ranthecross-validation,thistimemakingpredictionsonlyforthe4835geneswithnoannotationsoftypeIMP,TAS,NASorNR.(Thisleft1318positiveexamplesand627232negativeexamples.)TheresultingROCcurves(thelowertwographsinFigure1)showonlyaslightreductioninpredictionperformance.(Notealsothatinthetopgraph,circularityisnotanissuewhenweuseonlyMIPSphenotypesaspredictors.)
Figure1showsReceiverOperatingCharacteristic(ROC)curves,plottingTPtversusFPt.TodemonstratethevalueofusingMIPSphenotypescombinedwithGOattributesaspredictors,wehaveincludedROCcurvesfordecisiontreesinwhichonlyGOattributeswereusedaspredictorsandinwhichonlyMIPSphenotypeswereusedaspredictors.WehavealsoincludedtheROCcurveforamodelinwhichallphenotypesandattributesaretreatedasmutuallyindependent,sothatq(i,j)isjustthefractionofthegenesinthetrainingsetthatareannotatedashavingphenotypej.(Thisgiveshigherscoresforpredictionsofmorecommonphenotypes,independentofgene.)
Literature-basedassessment
Whilethecross-validationperformedabovedemonstratesthatgene-phenotypeassociationsmayoftenbepredictedaccuratelyonthebasisofotherannotations,thiswouldbeoflittleuseifallofthephenotypesassociatedwitheachgenewerealreadyknown.Butthisisalmostcertainlynotthecase,andweoperateunderthepremisethatthosegenesiandphenotypesjforwhichq(i,j)islargearegoodcandidatesforbeinggenuinelyassociated,eveniftheassociationisnotlistedinMIPS.Whilethepresentmethodisnottheidealwaytodealwithmissingphenotypicdata,asdiscussedinKingetal.(2003)ithasthevirtueofbeingcomputationallytractable,andisformallyidenticaltoatreatmentofmissingdatathathas
i185
21 no. 16 2005, pages 3422–3423 doi:10.1093/bioinformatics/bti553 ACT...J. Mol. Biol., 157, 105–132. Lupas,A. et al. (1991) Predicting ...
23 no. 5 2007, pages 597–604 doi:10.1093/bioinformatics/btl660 Prote...In this article, we compared the performance of CRFs in predicting protein ...
BIOINFORMATICS ORIGINAL PAPER doi10.1093bioinformaticsbtm358 Sequence analysis ...et al. (2005) Predicting protein function from sequence and structural data...
21 no. 8 2005, pages 1644–1652 doi:10.1093/bioinformatics/bti103 Predicting gene function through systematic analysis and quality assessment of high-...
BIOINFORMATICS ORIGINAL PAPER doi10.1093bioinformaticsbtm083 Data and text mining Data redu Motivation: Data reduction of liquid chromatography-mass ...
BIOINFORMATICS ORIGINAL PAPER doi10.1093bioinformaticsbtm044 Systems biology Simulating Eps Motivation: Epstein-Barr virus (EBV) infects greater than 90 % ...
BIOINFORMATICS APPLICATIONS NOTE doi10.1093bioinformaticsbtk005 Gene expression_专业资料。Summary: EDGE (Extraction of Differential Gene Expression) is an open...
BIOINFORMATICS ORIGINAL PAPER doi10.1093bioinformaticsbtm397 Systems biology An error model Motivation: Quantitative experimental data is the critical bottleneck...
25 no. 22 2009, pages 2897–2905 doi:10.1093/bioinformatics/btp537 De ...ed in this study could possibly aid in predicting ncRNAs in eukaryotic ...
23 ISMB/ECCB 2007, pages i159–i166 doi:10.1093/bioinformatics/btm208 ...4 CONCLUSION Predicting interactions between host and pathogen proteins is an ...
热门文档