文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > DOI 10.1093bioinformaticsbtg1024 Predicting
侵权投诉

DOI 10.1093bioinformaticsbtg1024 Predicting

Vol. 19 Suppl. 1 2003, pages i183–i189

O.D.Kingetal.

The180MIPSphenotypesareorganizedasahierarchy,withmorespeci cphenotypesdescendingfrommoregeneralphenotypes.WeassociatedwitheachphenotypejanindicatorrandomvariableXj,withXj(i)=1ifgeneiisannotatedashavingphenotypejandXj(i)=0otherwise.

TheGOattributesareorganizedasadirectedacyclicgraph(DAG)(likeahierarchy,butinwhichattributesmayhavemultipleparents,notjustmultiplechildren)consist-ingofthreebranches:cellularcomponent,molecularfunc-tion,andbiologicalprocess.TheGOconsistsofroughly13000terms,3051ofwhichwereassociatedwithatleastoneyeastgene.Weexcludedannotationsforthethreeat-tributes‘biologicalprocessunknown’,‘molecularfunc-tionunknown’,and‘cellularcomponentunknown’,andweassociatedwitheachoftheother3048attributesjanindicatorrandomvariableXj,withXj(i)=1ifgeneiisannotatedashavingattributejandXj(i)=0otherwise.Notethatthe lesfromMIPSandSGDusuallycontainexplicitannotationsonlyatthemostspeci clevelsthataresupportedbytheliterature.Inde ningXjwealsoincludethoseannotationslogicallyimpliedbythestructureoftheMIPShierarchyandtheGODAG,sothatXj(i)=1ifgeneiisexplicitlyannotatedwithtermjoranydescendantoftermjintheMIPShierarchyorGODAG.Wemadepredictionsforonlythe130mostspeci cphenotypes,althoughweallowedall180phenotypes,aswellasthe3048GOattributes,tobeusedaspredictors.LetX~jdenotethevectorconsistingofallthoserandomvariablesXkforwhichk=jandknotanancestorofjintheMIPShierarchy,andletX~j(i)denotethevectorofvaluesoftheserandomvariablesforthegenei.Foreachofthe130mostspeci cphenotypesjwebuiltadecisiontreeforpredictingXjfromX~j,andweusedthisdecisiontreetocompute

q(i,j)=Pr(Xj=1|X~j=X~j(i))

foreachgenei.(Notethatwhenmakingapredictionaboutwhetheragenehasacertainphenotypej,wedonotusetheancestorsofjintheMIPShierarchyaspredictors.)Thescoreq(i,j)maybeinterpretedastheprobabilitythatarandomlyselectedgeneisannotatedashavingphenotypej,giventhatitsotherannotations,ignoringthoseforphenotypejanditsancestorsintheMIPShierarchy,agreewiththoseforgenei.

whetherthegeneisannotatedashavingphenotypej.Thetestsarealloftheform‘Isthegeneannotatedashavingfunctionorphenotypeattributek?’forsomeGOattributekorMIPSphenotypek=j,withknotanancestorofjintheMIPShierarchy.Whichtestisapplieddependsontheresultofprevioustests—hencethetreestructure.

Weconstructedthedecisiontreeforphenotypejgreedily,bystartingwithallgenesginthetrainingsetinasinglerootnode,andthenrecursivelysplittingeachnodeNbytestingontheattributekforwhichtheinformationgainforphenotypejismaximal.

Ifwetestonattributek,splittingNintotwonodesN0andN1whereNt={g∈N:Xk(g)=t},thentheinformationgainisde nedtobe

HN(Xj) Pr(g∈N0|g∈N)HN0(Xj)

Pr(g∈N1|g∈N)HN1(Xj).

HereHN(Xj)istheentropyofXjatnodeN,whichisde nedtobe pNlog(pN) (1 pN)log(1 pN),wherepNistheprobabilitythatageneg∈GatanodeNisannotatedashavingj(seee.g.CoverandThomas,1991).Asin(NiblettandBratko1986),weusedtheestimate

pN=

#{g∈N:Xj(g)=1}+mp(j)

,

#{g∈N}+m

wherep(j)isthefractionofthegenesintheentiretrainingsetthatareannotatedashavingphenotypej,andmisanadjustableparameter.Thetermmp(j)isusedasapseudocount—asmallsample-sizeregularizationterm,withaninterpretationasapriorprobabilityinaBayesianframework(seee.g.EwansandGrant,2001)—withmbeingthetotalnumberofpseudocounts;wesetm=2.Weused#{g∈Nt}/#{g∈N}asanestimateforPr(g∈Nt|g∈N)fort=0andt=1,againfollowing(NiblettandBratko,1986).

WhennotestatanodeNprovidesapositiveinforma-tiongain,thenodeisnotsplit,butbecomesaleaf.Itisla-belledwiththeestimatepNoftheprobabilitythatageneatnodeNhasphenotypej,asde nedabove.

Atreegrowninthismannerwillusuallyover tthetrainingdata,andconsequentlyperformpoorlyontheheld-outtestdata.Astandardwayofcombatingthisistopruneawaysomeofthebranchesafterthetreeisgrown.WeusedtheBayesianInformationCriterion

BIC= 2lnPr(data|model)+(lnM)K,whichisasymptoticallyequivalenttotheMinimumDescriptionLength(MDL)(Schwartz1978),formodelselectionduringpruning(seee.g.FriedmanandGold-szmidt,1996).HereKisthenumberoffreeparametersinthemodel(whichinourcasecoincideswiththenumberofleavesinthedecisiontree),andMisthenumberof

Decisiontrees

(ThepresentationinthissectionisadaptedfromKingetal.,2003.)

See(Breimanetal.,1984)or(Quinlan,1993)foranoverviewofdecisiontreesandtheirapplications.Forourpurposes,thedecisiontreeforphenotypejprescribesasequenceofteststoapplytoagenetoaidinpredicting

i184

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email