Genetic algorithm optimization in drug design QSAR

MolDiversDOI10.1007/s11030-010-9234-9COMPREHENSIVEREVIEWGeneticalgorithmoptimizationindrugdesignQSAR:Bayesian-regularizedgeneticneuralnetworks(BRGNN)andgeneticalgorithm-optimizedsupportvectorsmachines(GA-SVM)MichaelFernandez·JulioCaballero·LeydenFernandez·AkinoriSaraiReceived:14May2009/Accepted:25January2010©SpringerScience+BusinessMediaB.V.2010AbstractManyarticlesin“insilico”drugdesignimple-mentedgeneticalgorithm(GA)forfeatureselection,modeloptimization,conformationalsearch,ordockingstudies.SomeofthesearticlesdescribedGAapplicationstoquan-titativestructure–activityrelationships(QSAR)modelingincombinationwithregressionand/orclassificationtechniques.WereviewedtheimplementationofGAindrugdesignQSARandspecificallyitsperformanceintheoptimizationofrobustmathematicalmodelssuchasBayesian-regularizedartificialneuralnetworks(BRANNs)andsupportvectormachines(SVMs)ondifferentdrugdesignproblems.ModeleddatasetsencompassedADMETandsolubilityproperties,cancertargetinhibitors,acetylcholinesteraseinhibitors,HIV-1pro-teaseinhibitors,ion-channelandcalciumentryblockers,andantiprotozoancompoundsaswellasproteinclasses,func-tional,andconformationalstabilitydata.TheGA-optimizedpredictorswereoftenmoreaccurateandrobustthanprevi-ouspublishedmodelsonthesamedatasetsandexplainedmorethan65%ofdatavariancesinvalidationexperiments.Inaddition,featureselectionoverlargepoolsofmoleculardescriptorsprovidedinsightsintothestructuralandatomicpropertiesrulingligand–targetinteractions.M.Fernandez(B)·A.SaraiDepartmentofBioscienceandBioinformatics,KyushuInstituteofTechnology(KIT),680-4Kawazu,Iizuka820-8502,Japane-mail:[e-mail]rmaticaySimulacionMolecular,UniversidaddeTalca,2Norte685,Casilla721,Talca,Chilee-mail:[e-mail]omputingCenter—CentroNacionaldeSupercomputación,NexusIIBuildingc/JordiGirona,29,08034Barcelona,SpainKeywordsDrugdesign·Enzymeinhibition·Featureselection·Insilicomodeling·QSAR·Review·SAR·Structure–activityrelationshipsListofabbreviationsADMETAbsorption,distribution,metabolism,excretionandtoxicityADAlzheimer’sdiseaselogSAqueoussolubilityANNsArtificialneuralnetworksBRANNsBayesian-regularizedartificialneuralnetworksBRGNNsBayesian-regularizedgeneticneuralnetworksBBBBlood–brainbarrierCoMFAComparativemolecularfieldanalysisCGConjugatedGradientGAGeneticalgorithmGA-PLSGeneticalgorithm-basedpartialleastsquaresGA-SVMGeneticalgorithm-optimizedsupportvectormachinesGNNGeneticneuralnetworksGSRGeneticstochasticresonanceHIAHumanintestinalabsorptionPPBRHumanplasmaproteinbindingrateLogPLipophilicityLHRHLuteinizinghormone-releasinghormoneMMPMatrixmetalloproteinaseMTMitochondrialtoxicityMLRMultiplelinearregressionMT−NegativemitochondrialtoxicityNNEsNeuralnetworkensemblesEVANormalcoordinateeigenvalueBIOOralbioavailability123MolDiversPLSPartialleastsquaresP-gpP-glycoproteinPCCPhysicochemicalcompositionMT+PositivemitochondrialtoxicityPC-GA-ANNPrincipalcomponent-geneticalgorithm-artificialneuralnetworkPCsPrincipalcomponentsPPRProjectionpursuitregressionQSARQuantitativestructure–activityrelationshipQSPRQuantitativestructure–propertyrelationshipRBFRadialBasicFunctionSOMsSelf-organizedmapsSRStochasticresonanceSVMsSupportvectormachinesTrb1Thyroidhormonereceptorb1TdpTorsadesdepointesVKCsVoltage-gatedpotassiumchannelsIntroductionOneofthemainchallengesintoday’sdrugdesignisthediscoveryofnewbiologicallyactivecompoundsonthebasisofpreviouslysynthesizedmolecules.Quantitativestructure–activityrelationship(QSAR)isanindirectligand-basedapproachwhichmodelstheeffectofstructuralfeaturesonbiologicalactivity.Thisknowledgeisthenemployedtoproposenewcompoundswithenhancedactivityandselec-tivityprofileforaspecifictherapeutictarget[1].QSARmethodsarebasedentirelyonexperimentalstructure–activ-ityrelationshipsforenzymeinhibitororreceptorligands.Incomparisontodirectreceptor-basedmethods,whichincludemoleculardockingandadvancedmoleculardynamicssimu-lations,QSARmethodsdonotstrictlyrequirethe3D-struc-tureofatargetenzymeorevenareceptor–effectorcomplex.Theyarecomputationallynotdemandingandallowestab-lishingan“insilico”toolfromwhichbiologicalactivityofnewlysynthesizedmoleculescanbepredicted[1].Three-dimensional-QSAR(3D-QSAR)methods,espe-ciallycomparativemolecularfieldanalysis(CoMFA)[2]andComparativeMolecularSimilarityIndicesAnalysis,(CoMSIA)[3]arenowadaysusedwidelyindrugdesign.Themainadvantagesofthesemethodsarethattheyareapplicabletoheterogeneousdatasets,andtheybringa3D-mappeddescriptionoffavorableandunfavorableinterac-tions,accordingtophysicochemicalproperties.Inthissense,theyprovideasolidplatformforretrospectivehypothesesbymeansoftheinterpretationofsignificantinteractionregions.However,somedisadvantagesofthesemethodsarerelatedtothe3Dinformationandalignmentofthemolecularstruc-tures,sincethereareuncertaintiesaboutdifferentbindingmodesofligands,anduncertaintiesaboutthebioactivecon-formations[4].CoMFAandCoMSIAhaveemergedasthe3D-QSARmethodsmostembracedbythescientificcommunitytoday;however,currentarticlesonQSARencompasstheuseoftoomanyfor