Aglycosylphosphatidylinositol(GPI)anchorisacommonbutcomplexC-terminalpost-translationalmodificationofextracellularproteinsineukaryotes.HereweinvestigatetheproblemofcorrectlyannotatingGPI-anchoredproteinsforthegrowingnumberofsequencesinpublicdatabases.Wedevelopedacomputa-tionalsystem,calledFragAnchor,basedonthetandemuseofaneuralnet-work(NN)andahiddenMarkovmodel(HMM).Firstly,NNselectspotentialGPI-anchoredproteinsinadataset,thenHMMparsesthesepotentialGPIsig-nalsandrefinesthepredictionbyqualitativescoring.FragAnchorcorrectlypredicted91%ofalltheGPI-anchoredproteinsannotatedintheSwiss-Protdatabase.Inalarge-scaleanalysisof29eukaryoteproteomes,FragAnchorpredictedthatthepercentageofhighlyprobableGPI-anchoredproteinsisbetween0.21%and2.01%.ThedistinctivefeatureofFragAnchor,comparedwithothersystems,isthatittargetsonlytheC-terminusofaprotein,makingitlesssensitivetothebackgroundnoisefoundindatabasesandpossibleincompleteproteinsequences.Moreover,FragAnchorcanbeusedtopredictGPI-anchoredproteinsinalleukaryotes.Finally,byusingqualitativescoring,thepredictionscombinebothsensitivityandinformationcontent.Thepredictorispubliclyavailableathttp://navet.ics.hawaii.edu/～fraganchor/NNHMM/NNHMM.html.
