简介:Ithasbeenshownthattheprogressinthedeterminationofmembraneproteinstructuregrowsexponentially,withapproximatelythesamegrowthrateasthatofthewater-solubleproteins.Inordertoinvestigatetheeffectofthis,ontheperformanceofpredictionalgorithmsforbothα-helicalandβ-barrelmembraneproteins,weconductedaprospectivestudybasedonhistoricalrecords.WetrainedseparatehiddenMarkovmodelswithdifferentsizedtrainingsetsandevaluatedtheirperformanceontopologypredictionforthetwoclassesoftransmembraneproteins.Weshowthattheexistingtop-scoringalgorithmsforpredictingthetransmembranesegmentsofα-helicalmembraneproteinsperformslightlybetterthanthatofβ-barreloutermembraneproteinsinallmeasuresofaccuracy.Withthesamerationale,ameta-analysisoftheperformanceofthesecondarystructurepredictionalgorithmsindicatesthatexistingalgorithmictechniquescannotbefurtherimprovedbyjustaddingmorenon-homologoussequencestothetrainingsets.Theupperlimitforsecondarystructurepredictionisestimatedtobenomorethan70%and80%ofcorrectlypredictedresiduesforsinglesequencebasedmethodsandmultiplesequencebasedones,respectively.Therefore,weshouldconcentrateoureffortsonutilizingnewtechniquesforthedevelopmentofevenbetterscoringpredictors.