决策树是通过递回分割

AFive-GeneSignatureandClinicalOutcomeinNon–Small-CellLungCancerFrom:nengljmed356;1january4,2007By:Hsuan-YuChen,M.Sc.,Sung-LiangYu,Ph.D.,etalReporter:R6謝廣宇Background•Currentstagingmethodsareinadequateforpredictingoutcome•developeda5-genesignature•non–small-celllungcancer(NSCLC)—mostcommoncauseofdeathfromcancerworldwide•relapseratewithearly-stageNSCLCis40%within5yearsafterpotentiallycurativetreatment1.決策樹是通過遞迴分割（recursivepartitioning）建立而成，遞迴分割是一種把資料分割成不同小的部分的疊代過程cDNAmicroarrayschemacDNA晶片製造原理MicroarrayanalysisOperationPrinciple:Samplesaretaggedwithflourescentmaterialtoshowpatternofsample-probeinteraction(hybridization)Microarraymayhave60KprobeMicroarrayProcessingsequenceFrom:[e-mail]•Gene-expressionprofilingbymeansofmicroarraysandreverse-transcriptasepolymerasechainreaction(RT-PCR)•examinedgeneexpressionin125surgicalspecimensofNSCLC,usingmicroarraysandreal-timeRT-PCRtoidentifyagenesignaturecorrelatedwithclinicaloutcomeMethods•computer-generatedrandomnumberstoassignspecimensfrom185consecutivepatientsformicroarrayanalysis•frozenspecimensoflung-cancertissuefrom125randomlyselectedpatientswhounderwentsurgicalresectionofNSCLCatTaichungVeteransGeneralHospitalbetweenDecember1999andDecember2003------five-generisk-predictionmodelusinganindependentcohortof60randomlyselectedpatients•125specimens--60wereadenocarcinomas,52weresquamous-cellcarcinomas,and13wereothertypesofcancer•notreceivedadjuvantchemotherapy•672genesassociatedwithinvasiveactivity,identifiedinapreviousstudy---rearrayedinduplicateonanylonmembrane•4μgoftotalRNAfromeachspecimen•validatelevelsofexpressionofgenesfoundonmicroarrayanalysisRT-PCRwasperformedon16genesandacontrolgeneforTATA-box–bindingprotein(TBP),withuseofspecificTaqManprobesandprimersets---transcriptswereamplifiedwithreagent(TaqManOne-StepRT-PCRMasterMixReagent,AppliedBiosystems)andasequencedetectionsystem(ABIPrism7900HT,AppliedBiosystems)•Toreducebackgroundnoisebackgroundintensityvaluesoflessthan3000wereassignedvalueof3000log-transformedtoabase-2scale•Geneswithcoefficientsofvariationoflessthan3%wereexcludedfromfurtheranalyses•thegene-expressionintensityvaluesweretransformedtoordinalcodingvaluesbyrankingoflevelofgeneexpressionamongthe485genesin125patients(60,625observations)•intensityvaluewascodedas1-2-3-4accordingto0-25%-50%-75%-100%•HazardratiosfromunivariateCoxregressionanalysistodeterminewhichgenesassociatedwithdeathfromanycauseorrecurrenceofcancer•Protectivegenesweredefinedwithahazardratiofordeathof1;riskgenesweredefinedwithahazardratiofordeathof1•univariateCoxproportional-hazardsregressionanalysis•Genessignificantlycorrelatedwithsurvival---alinearcombinationofgene-expressioncodingvaluesweightedbyregressioncoefficientstocalculateariskscoreforeachpatient•patient’sriskscorewascalculatedassumoflevelsofexpressionofeachgene,asmeasuredbymicroarrayanalysis,multipliedbycorrespondingregressioncoefficients•high-riskgenesignatureoralow-riskgenesignature,with50thpercentile(median)oftheriskscoreasthethresholdvalue(median,4.9;range,1.3to21.9)---thresholdvaluetoreflectalmosthalfofpatientswithearlystageNSCLCrelapsewithin5yearsafterpotentiallycurativesurgery,eliminateeffectofextremevalues•levelsofexpressionof16genesconfirmedbyRT-PCRandindexedbySpearman’srank-correlationtestfurtheridentified5genessignificantlyassociatedwithsurvival•recursive-partitioningdecisiontree&Avadissoftware(StrandGenomic)ahigh-riskgenesignatureoralow-riskgenesignature•Kaplan–Meiermethodtoestimateoverallsurvivalandrelapse-freesurvival•MultivariateCoxproportionalhazardsregressionanalysiswithstepwiseselectionevaluateindependentprognosticfactorswithsurvival,5-genesignature,age,sex,tumorstage,andhistologiccharacteristicsascovariates•Tofurthervalidateourmodel,weappliedittomicroarraydatafrom86patientswithNSCLC,reportedbyBeeretal•fivegenes(andtheircorrespondingAffymetrixprobesets)wereDUSP6(X93920_at),MMD(X85750_at),STAT1(M97936_at),ERBB3(S61953_at),andLCK(M26692_s_at);thecontrolgenewasTBP(X54993_s_at)•maximumfollow-uptimeforsurvivalanalysisinourstudywas62months,weused5-yearsurvivaldatafor86patientsResults•16genescorrelatedwithdeathfromanycause:4wereprotectivegenes(hazardratio1)and12wereriskgenes(hazardratio1)•dual-specificityphosphatase6(DUSP6)•monocyte-to-macrophagedifferentiationassociatedprotein(MMD)•signaltransducerandactivatoroftranscription1(STAT1)•v-erb-b2avianerythroblasticleukemiaviraloncogenehomolog3(ERBB3)•lymphocyte-specificproteintyrosinekinase(LCK)•5-genesignaturestronglyassociatedwithoverallsurvival(sensitivity,98%;specificity,93%;positivepredictivevalue,95%;negativepredictivevalue,98%;andoverallaccuracy,96%)•AmongpatientswithstageIIdisease,overallsurvivaldidnotdiffersignificantlybetweenhigh-riskandlow-riskgenesignatures,probablyowingtosmallnumberofpatientsDiscussion•NSCLCisaheterogeneousdiseasebasedonclinicalandpathologicalfindingsmayhavelimitofusefulnessforpredictingoutcomes,butmolecularmethodsaddvalue•useofmicroarraysinclinicalpracticelimitedbythelargenumberofgenesintheanalysis,complicate