Inferring Networks from Multiple Samples with Consensus LASSO

摘要

Networksareveryusefultoolstodeciphercomplexregulatoryrelationshipsbetweengenesinanorganism.Mostworkaddressthisissueinthecontextofi.i.d.,treatedvs.controlortime-seriessamples.However,manydatasetsincludeexpressionobtainedforthesamecelltypeofanorganism,butinseveralconditions.Weintroduceanovelmethodforinferringnetworksfromsamplesobtainedinvariousbutrelatedexperimentalconditions.Thisapproachisbasedonadoublepenalization:afirstpenaltyaimsatcontrollingtheglobalsparsityofthesolutionwhilstasecondpenaltyisusedtomakecondition-specificnetworksconsistentwithaconsensualnetwork.This“consensualnetwork”isintroducedtorepresentthedependencystructurebetweengenes,whichissharedbyallconditions.Weshowthatdifferent“consensus”penaltiescanbeused,someintegratingprior(e.g.,bibliographic)knowledgeandothersthatareadaptedalongtheoptimizationscheme.Inallsituations,theproposeddoublepenaltycanbeexpressedintermsofaLASSOproblemandhence,solvedusingstandardapproacheswhichaddressquadraticproblemswithL1-regularization.ThisapproachiscombinedwithabootstrapapproachandismadeavailableintheRpackagetherese1.Ourproposalisillustratedonsimulateddatasetsandcomparedwithindependentestimationsandalternativemethods.Itisalsoappliedtoarealdatasettoemphasizethedifferencesinregulatorynetworksbeforeandafteralow-caloriediet.