Predictions of Hot Spot Residues at Protein-Protein Interface s Using Support Vector Machines 论文 计算生物学
PredictionsofHotSpotResiduesatProtein-ProteinInterfacesUsingSupportVectorMachines
StefanoLise,DanielBuchan,MassimilianoPontil,DavidT.Jones*
DepartmentofComputerScience,UniversityCollegeLondon,London,UnitedKingdom
Abstract
Protein-proteininteractionsarecriticallydependentonjustafew‘hotspot’residuesattheinterface.Hotspotsmakeadominantcontributiontothefreeenergyofbindingandtheycandisrupttheinteractionifmutatedtoalanine.Here,wepresentHSPred,asupportvectormachine(SVM)-basedmethodtopredicthotspotresidues,giventhestructureofacomplex.HSPredrepresentsanimprovementoverapreviouslydescribedapproach(Liseetal,BMCBioinformatics2009,10:365).Itachieveshigheraccuracybytreatingseparatelypredictionsinvolvingeitheranarginineoraglutamicacidresidue.Thesearetheaminoacidtypesonwhichtheoriginalmodeldidnotperformwell.WehavethereforedevelopedtwoadditionalSVMclassifiers,specificallyoptimisedforthesecases.HSPredreachesanoverallprecisionandrecallrespectivelyof61%and69%,whichroughlycorrespondstoa10%improvement.Animplementationofthedescribedmethodisavailableasawebserveratdoc.guandang.net.
Citation:LiseS,BuchanD,PontilM,JonesDT(2011)PredictionsofHotSpotResiduesatProtein-ProteinInterfacesUsingSupportVectorMachines.PLoSONE6(2):e16774.doi:10.1371/journal.pone.0016774
Editor:CollinStultz,MassachusettsInstituteofTechnology,UnitedStatesofAmericaReceivedOctober23,2010;AcceptedDecember30,2010;PublishedFebruary28,2011
Copyright:ß2011Liseetal.Thisisanopen-accessarticledistributedunderthetermsoftheCreativeCommonsAttributionLicense,whichpermitsunrestricteduse,distribution,andreproductioninanymedium,providedtheoriginalauthorandsourcearecredited.
Funding:ThisworkwasfundedbygrantreferenceBB/E017452/1,fromtheBiotechnologyandBiologicalSciencesResearchCouncil(BBSRC).Thefundershadnoroleinstudydesign,datacollectionandanalysis,decisiontopublish,orpreparationofthemanuscript.CompetingInterests:Theauthorshavedeclaredthatnocompetinginterestsexist.*E-mail:d.jones@cs.ucl.ac.uk
Introduction
Alaninescanningmutagenesisisapowerfulexperimentalmethodologyforinvestigatingthestructuralandenergeticcharacteristicsofproteincomplexes[1].Individualamino-acidsaresystematicallymutatedtoalanineandchangesinfreeenergyofbinding(DDG)measured.Asalanineaminoacidsdonothaveaside-chainbeyondtheb-carbon,thisprocedureineffectteststheimportanceofindividualside-chaingroupsforcomplexformation,providingamapoftheso-calledfunctionalepitope.Resultsfromanumberofexperimentsindicatethatonlyasmallsubsetofcontactresiduescontributesignificantlytothebindingfreeenergy.Theseresidueshavebeentermed‘hotspots’andifmutatedtheycandisrupttheinteraction.Forthemajorityofinterfaceresiduesinstead,theeffectofanalaninemutationisminimal[2].
HotspotsaretypicallydefinedasthoseresiduesforwhichDDG§2kcal=mol.Inrecentyears,severalcomputationalapproacheshavebeendevelopedtoidentifythematprotein-proteininterfaces[3–16].Accuratepredictivemodelsprovideavaluablecomplementtoexperimentalstudiesandaddtoourunderstandingofthefactorsthatinfluenceaffinityandspecificityinprotein-proteininterfaces.Inaddition,theycanhaveimportantapplicationsinthefieldofdrugdiscovery.Anumberofrecentstudieshavebeensuccessfulindeveloping(drug-like)smallmoleculesthatbindathotspotsandinhibitcomplexformation[17].Reliablehotspotspredictionscouldthereforerepresentthefirststepinrationaldrugdesignprojects[18].
Inapreviouswork,wepresentedamachinelearningstrategytoidentifyhotspotresiduesinprotein-proteininterfaces,giventhestructureofthecomplex[12].Weconsideredthebasicenergetictermsthatcontributetohotspotinteractions,i.e.vanderWaals
PLoSONE|doc.guandang.net
1
potentials,solvationenergy,hydrogenbondsandCoulombelectrostatics,andtreatedthemasinputfeaturesofaSupportVectorMachine(SVM)classifier.Wefoundthatthemethodcouldpredicthotspotswithoverallgoodaccuracy,comparingfavourablytootheravailableapproaches.However,bygroupingmutationsaccordingtotheaminoacidtype,weobservedthatinsomecasestheSVMmodeldidnotperformtoowell,forexampleonpredictionsinvolvingarginineorglutamicacidresidues.
Inthispaper,wereportthedevelopmentofHSPred,ahotspotpredictionmethodthataimstoovercomethelimitationshighlightedabove.Forthispurpose,wehaveintegratedtheoriginalapproachwithtwoadditionalSVMclassifiers,specificallybuiltformutationsinvolvingArgandGluresidues.Thetwoadditionalmodelsaretrainedonthesamedatasetasthe‘general’modelbutarebiasedtoperformwellonArgandGluduetoadifferentchoiceofinputfeatures.Employingastrictcross-validationscheme,weshowthatthisstrategyleadstoasignificantimprovementoverthepreviousversionofthemethod.WefurthervalidatetheresultsbyapplyingHSPredtoanexternaltestcase,whichisnotpartoftheoriginaldataset.
ResultsandDiscussion
Theproblemwehaveinvestigatedisthepredictionofhotspotresiduesataprotein-proteininterfaceusingamachinelearningapproach.Asinputvariables,wehaveconsideredbasicenergyterms(vanderWaals,hydrogenbond,electrostaticanddesolva-tionpotentials)calculatedfromthecomplexstructure.Wehavedistinguishedcontributionsfromdifferentstructuralregionsinthecomplex,leadingto3distincttypesofinteractions:side-chaininter-molecular,environmentinter-molecularand

side-chain
February2011|Volume6|Issue2|e16774
热门文档