Vectorwww.topit.me

RESEARCHOpenAccessIdentificationofconformationalB-cellEpitopesinanantigenfromitsprimarysequenceHifzurRahmanAnsari,GajendraPSRaghava*AbstractBackground:OneofthemajorchallengesinthefieldofvaccinedesignistopredictconformationalB-cellepitopesinanantigen.
Inthepast,severalmethodshavebeendevelopedforpredictingconformationalB-cellepitopesinanantigenfromitstertiarystructure.
ThisisthefirstattemptinthisareatopredictconformationalB-cellepitopeinanantigenfromitsaminoacidsequence.
Results:AllSupportvectormachine(SVM)modelsweretrainedandtestedon187non-redundantproteinchainsconsistingof2261antibodyinteractingresiduesofB-cellepitopes.
Modelshavebeendevelopedusingbinaryprofileofpattern(BPP)andphysiochemicalprofileofpatterns(PPP)andachievedamaximumMCCof0.
22and0.
17respectively.
Inthisstudy,forthefirsttimeSVMmodelhasbeendevelopedusingcompositionprofileofpatterns(CPP)andachievedamaximumMCCof0.
73withaccuracy86.
59%.
WecompareourCPPbasedmodelwithexistingstructurebasedmethodsandobservedthatoursequencebasedmodelisasgoodasstructurebasedmethods.
Conclusion:ThisstudydemonstratesthatpredictionofconformationalB-cellepitopeinanantigenispossiblefromisprimarysequence.
ThisstudywillbeveryusefulinpredictingconformationalB-cellepitopesinantigenswhosetertiarystructuresarenotavailable.
AwebserverCBTOPEhasbeendevelopedforpredictingB-cellepitopehttp://www.
imtech.
res.
in/raghava/cbtope/.
BackgroundAregionorsegmentofanantigen,recognizedbyaspeci-ficantibodyorB-celliscalledantigenicregionorB-cellepitope.
TheseB-cellepitopescanbecategorizedintotwoclasses,continuousanddiscontinuous.
Acontinuous/lin-earepitopeisasegmentofconsecutiveresiduesinthepri-marysequencewhileadiscontinuous/conformationalepitopeisabunchofresiduesofanantigenthatarefarawayfromeachotherintheprimarysequencebutarebroughttospatialproximityasaresultofpolypeptidefolding.
ItisalsoknownthatmostoftheB-cellepitope(~90%)areconformationalepitope.
Bothtypesofepitopesplayanimportantroleinthepeptide-basedvaccinesanddiseasediagnosis[1,2].
Oneofthebeautiesofimmunesys-temisthatitrecognizestheforeignproteins/antigensandgeneratespecificantibodyagainsttheseantigens.
Thispotentialofimmunesystemhasbeenexploitedbyresearchersfordesigningsubunitvaccines[3,4].
Inthepostgenomicerawherealargenumberofpathogenshavebeencompletelysequenced,itiscrucialtoidentifyB-cellepitopeorhereaftercalledantibodyinteractingresiduesinanantigenforthedesignofsubu-nitvaccinesagainstthesepathogens.
Inthepastseveralexperimentaltechniqueshavebeendevelopedformap-pingantibodyinteractingresiduesonanantigenthatincludesidentificationofinteractingresiduesfromstructureofantibody-antigencomplexes[5].
Oneofthepopularapproachesisoverlappingpeptidesynthesiscov-eringtheentireantigensequence,whichidentifiesmainlysequentialepitopes[6].
Mappingofantibodyinteractingresidueshasbeenseverelyhamperedbythecostlyandtimetakingprocessof3Dstructuredetermi-nation.
Manytools,coveringcompilation,visualizationandpredictionofBandTcellepitopeshavebeendevel-oped[7].
Despiteofmajorityofepitopesbeingconfor-mational,mostofthecomputationalmethodsanddatabasescenteredatthesequentialepitopes[8-10].
Linearepitopepredictionmethodscanbecategorizedintophysico-chemicalproperty[11],HMM[12]and*Correspondence:raghava@imtech.
res.
inBioinformaticsCenter,InstituteofMicrobialTechnology,Sector39-A,Chandigarh,IndiaAnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6IMMUNOMERESEARCH2010AnsariandRaghava;licenseeBioMedCentralLtd.
ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommonsAttributionLicense(http://creativecommons.
org/licenses/by/2.
0),whichpermitsunrestricteduse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.
ANNbased[13].
Manymethodsareavailableforanti-bodyinteractingresiduesidentificationifantigen'soritshomolog'stertiarystructureisknownwhichinitselfisabiglimitation.
Thesearebasedonfeatureslikeflexibil-ity,solventaccessibility[14,15]andaminoacidpropen-sityscales[16].
Earlierresearcherscreatedabenchmarkdatasetfromthe3DPDBstructuresandevaluatedsev-eralstructure-basedprotein-proteinbindingsitepredic-tionmethodswhichincludedpopularCEP[15]andDiscoTope[16]forpredictingimmunogenicregions[17].
Theyoptedthedefinition,thatepitopeconsistofantigenresiduesinwhichanyatomoftheantigenresi-dueisseparatedfromanyantibodyatombyadistanceof≤4.
Theyfoundthattheperformanceofallmeth-odsweremediocreandnomethodcouldachieveAreaundercurve(AUC)greaterthan0.
7.
Inadditiontotheseabunchofimprovedmethodshavebeendevel-opedforthepredictionofantibodyinteractingresiduesiftertiarystructureofantigenisknown[18-23].
Insum-mary,oneneedstodeterminestructureofantigenusingcrystallographyinordertoidentifyantibodyinteractingresiduesinantigen.
Theexperimentaltechniqueslikecrystallographyareexpensiveandtimeconsumingwhereasfunctionalassaysarenotreliableenough[5].
Thusthereisneedtodevelopalternatetechniqueforpredictingantibodyinteractingresiduesinaprotein.
Inthisstudyattempthasbeenmadetopredictanti-bodyinteractingresiduesinanantigenfromitsprimarysequence.
Firstwecreatedthepatternsofdifferentwin-dowlengthsfromthecorrespondingaminoacidsequencesthenusedthestandardbinaryandphysico-chemicalprofilesofpatterns.
Wehaveintroducedforthefirsttimetheconceptofcompositionprofileofpat-tern(CPP)generatedthroughslidingwindowwherethecentralresidueisantibodyinteracting.
ThesefeatureswereusedtodevelopSVMbasedmodelstopredictantibodyinteractingresidueswithhighaccuracy.
MethodsDefinitionofantibodyinteractingresiduesorepitopeTherearemanylevelsofantigen-antibodyinteractionsonecanobtainfromPDBstructures.
Amongtheseinteractionswedefinedantibodyinteractingresidueasaresidueofantigenwhichisatleastoneatomseparatedfromanantibodyatomby4distance.
Weborrowedthisdefinitionfrombenchmarkpaper[17]inordertocompareourmodelswithexistingmethods.
DatasetsMaindatasetWeobtained526antigenicsequencescombinedfromIEDBdatabaseandbenchmarkdataset[9,17].
SequenceredundancywasremovedusingprogramCDHIT[24]at40%cutoff.
Finallywegot187antigenswherenotwosequenceshavemorethan40%sequenceidentity.
Theseantigenshave2261antibodyinteractingor2261residuesarepartofconformationalB-cellepitopeand107414aminoacidresidueswerenon-antibodyinteractingfromthesameantigensequences.
BenchmarkDatasetInadditiontomaindataset,wealsoevaluateourmodelsonbenchmarkdataset[17]whichcontains161proteinchainsfrom144antigen-antibodycomplexstructures.
Finallywegotnon-redundantsetof52antigenchainswherenotwosequenceshavemorethan40%sequenceidentity.
Thisbenchmarkdatasetof52antigenscontains858antibodyinteractingand9366non-antibodyinter-actingresidues.
CreationofpatternsItisknownthatthefunctionofaresidueisnotsolelydeterminedbyitselfbutinfluencedbyitsneighboringresidues[25-27].
Thuswegeneratedoverlappingpatternsofdifferentwindowsizesfrom5to21aminoacidsforeachantigeninthedatasets.
Apatternisassignedaspositiveifitscentralresidueinteractswiththeantibody;elseitisassignedasnegative(Figure1).
Thisisthestan-dardprocedureusedforassigningpatterns,whichhavebeenusedinnumberofmethodslikepredictionofNADinteractingresidues[26],DNA,RNAbindingsitesinpro-teins[27],cleavagesites[28]andsignalpeptides[29].
Inordertocreateapatternfortheterminalresidues,weadded(L-1)/2numberofdummyresidue'X'onbothsidesoftheproteinsequence(Lislengthoftheproteinsequence)fore.
g.
forwindowsize17weadded8'X'.
RealisticandbalancelearningInordertodeveloppredictionmethodoneneedstogenerateoverlappingpatternsforeachantigeninadata-set;onepatternforeachresidue.
Itwillproducetwotypesofpatternspositiveandnegative,positivepatternshaveantibodyinteractingcentralresidue.
Thesepatternsareusedtotrainmachine-learningtechniquesfordevel-opingmodels.
InreallifeonlyfewresiduesinanantigenarerecognizedbyantibodyorB-cellreceptor.
Thismeansthatthenumberofnegativepatternswillbemuchhigherthanpositivepatternsinourtrainingdata-set;for2261positivepatternstherewere107414nega-tivepatterns.
Thiscreatestwoproblems;i)poorperformanceofmodelsduetoimbalancedsetofpat-ternsandii)trainingofmodelsistimeconsumingandCPUintensive.
Thusinthisstudywehaveusedtwopat-ternsetsforlearningourmodels;i)realisticsetofpat-ternsthatincludesallnegativepatternsandii)balancesetofpatternshavingequalnumberofpositiveandnegativepatterns.
Incaseofbalanceset,werandomlypickedupequalnumberofnegativesfromnegativepatternset.
AnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page2of9DerivationoffeaturesfrompatternsBinaryprofileofpatterns(BPP)Eachpatternwasconvertedintobinaryprofile,whereanaminoacidwasrepresentedbyavectorofdimension21(e.
g.
Alaby1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0).
Apat-ternofwindowlengthWwasrepresentedbyavectorofdimensions21xW(Additionalfile1,TableS1).
Thebin-aryprofilehasbeenusedinanumberofexistingmeth-ods[30,31].
Physico-chemicalprofileofpatterns(PPP)Asaminoacids'physico-chemicalpropertiescontributeinthedeterminationofitsstructureandfunction,weselectedfivepropertiestestedbyothers[32].
TheseareGranthampolarity[33],Karplus-Schulzflexibility[34],Kolaskarantigencity[35],Parkerhydrophobicity[36]andPonnuswamipolarityindex[37].
Physico-chemicalprofileofpatternsissimilartotheBPP,theonlydiffer-enceliesinthepropertiesofaminoacids.
Hereeachaminoacidisrepresentedbyavectorof5i.
e.
eachpat-ternconvertedintoavectorsizeof5xW.
ForexampleAlaisrepresentedas[pHydrophobicity,pFlexibility,pPolarity_Grantham,pPolarity_Ponnuswami,pAntigene-city]correspondingtodifferentpropertyvalues(Addi-tionalfile1,TableS2).
Compositionprofileofpatterns(CPP)Inthepastresearchershaveexploitedaminoacidcom-positionofproteinsformanybiologicalproblemslikesub-cellularlocalizationandclassificationofproteins[38,39].
Insteadofcalculatingcompositionofantigensequence,weintroducedconceptofcompositionofpat-terns.
Theaminoacidcompositionofpatternswascal-culatedusingthefollowingequation.
compiRNi()=*100Wherecomp(i)isthepercentcompositionofaresi-dueoftypei;Riisnumberofresiduesoftypei,andNisthetotalthenumberofresiduesinthepattern.
SupportVectorMachines(SVM)InthepastSVMhadbeenusedinanumberofbiologi-calproblems,fromclassificationtofunctionalpredictionFigure1Featureextractionfora19windowlengthpattern.
Antibodyinteractingresiduesaremarkedinrede.
g.
S/T,PositivepatternshadedingreenwhereSisatthecenterwith9neighboringresiduesoneitherside,otheroverlappingnegativepatternsareshowninblue.
a)Creationof19windowoverlappingpatternsfromaminoacidsequence,b)generationofbinaryprofileofpattern(BPP),c)generationofphysico-chemicalprofile(PPP)andd)generationofcompositionprofileofpattern(CPP).
AnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page3of9ofproteins[40-42].
Inthepresentstudy,wehavedevel-opedaSVMmodelusingapowerfulpackageSVM_lighthttp://svmlight.
joachims.
org/,forpredictingantibodyinteractingresiduesinproteins.
Cross-validationtechniqueTherearemanytechniquesforevaluatingtheperformanceofmodelslikeleave-one-outorjack-knifetest,n-foldcrossvalidationetc[43].
Thoughjackknifetestisthebestamongcross-validationtechniques[44],itistimeconsum-ingandCPUintensivetechnique[40,45].
Inordertosavetimeandresourcesweusedwidelyacceptable5-foldcross-validationtechnique.
Inthistechniquedataisran-domlydividedintofiveequalsetsofwhichfoursetsareusedfortrainingandtheremainingfifthsetfortesting.
Thisprocessisrepeatedfivetimesinsuchawaythateachsetisusedoncefortesting.
Finalperformanceistheaver-ageofperformancesachievedonthefivesets.
PerformanceMeasuresTheperformanceofvariousmodelsdevelopedinthisstudywascomputedbyusingthreshold-dependentaswellasthreshold-independentparameters.
Inthreshold-dependentparametersweusedsensitivity(Sen),Specifi-city(Spe)orpercentcoverageofnon-interactingresi-dues,overallaccuracy(Acc)andMatthew'scorrelationcoefficient(MCC)usingfollowingequations.
SensitivityTPTPFN=+*100SpecificityTNTNFP=+*100AccuracyTPTNTPTNFPFN=++++*100MCCTPTNFPFNTPFNTNFPTPFPTNFN=**++++()()[TP=truepositive;FN=falsenegative;TN=truenegative;FP=falsepositive]WecreatedROC(receiveroperatingcurve)forallofthemodelsinordertoevaluateperformanceofmodelsusingthresholdindependentparameters.
ROCplotswithAreaundercurve(AUC)werecreatedusingSPSSstatisticalpackage.
ResultsAnalysisofantibodyinteractingresiduesInordertounderstandwhethercertaintypesofaminoacidsarepreferredinantibodyinteractions,wecomparedthecompositionofantibodyinteractingandnon-interactingresiduesinantigens.
AsshowninFigure2,certaintypesofresidueslikeCystein,Aspartate,Gluta-mate,Lysine,Asparagine,Glutamine,Arginine,Trypo-phanandTyrosinearepreferredinantibodyinteractions.
Mostofthesearepolarandchargedresidues.
Inordertounderstandthepreferenceofinteractionindepth,wecreated2SampleLogos[46]fordifferentproperties.
Itwasobservedthatcharged,hydrophilic,surfaceexposedandflexibleresiduesaremoreabundantinconforma-tionalB-cellepitopes(Additionalfile1,FiguresS1,S2,S3,S4,andS5).
SVMModelsbasedonBPPandPPPFirst,SVMbasedmodelshavebeendevelopedusingbinaryprofileofpatternswherepatternisrepresentedbyavectorofdimensionsNx21(Nislengthofpattern).
InordertooptimizetheperformanceofSVMmodels,wedevelopedSVMmodelsusingpatternsofwindowlength5to21.
Itwasobservedthatmodelsperformbet-terforwindowsize13,wherewegotmaximumMCC0.
22withaccuracyof60.
84%(Table1).
Weselectedmodelswithminimumdifferencebetweensensitivityandspecificity.
Varyingthekernelparameterscouldnotenhancetheperformanceofmodelsandresultswerejustbetterthanrandom.
DetailperformanceofBPPbasedSVMmodelforwindowlength13atdifferentthresholdsisshowninAdditionalfile1,TableS3.
Itwasobservedthataminoacidshavingcertaintypesofphysico-chemicalpropertiesarepreferredinantibodyinteractions(Additionalfile1,FiguresS1,S2,S3,S4,andS5).
ThuswedevelopedSVMbasedmodelsusingPPPandobservedbestperformanceforpatternlengthof15residues.
AsshowninTable2,wegotmaximumMCC0.
17withaccuracy58.
31%.
Thetrendandperfor-manceofSVMmodelsbasedonBPPandPPPissimilar.
DetailperformanceofPPPbasedSVMmodelforwin-dowlength15atdifferentthresholdsisshowninAddi-tionalfile1,TableS4.
OverallperformanceofPPPbasedmodelisslightlypoorerthanBPPbasedmodel(Additionalfile1,TablesS3andS4).
Allmodelsweretrainedandtestedonmaindatasetusingbalancesetofpatterns.
SVMModelusingCompositionProfileofPatterns(CPP)Tounderstandtheantibodyinteractingpatternsbetter,wecomputedandcomparedaminoacidcompositionofpositiveandnegativepatterns.
AsshowninAdditionalfile1,FigureS6,compositionprofileofpositiveandnegativepatternsaredifferent.
Thismeansthatpositiveandnegativepatternscanbediscriminatedfromtheiraminoacidcomposition.
Basedonthisobservation,wedevelopedSVMmodelsforpredictingantibodyinteract-ingresiduesinproteinsusingcompositionprofileofAnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page4of9patterns(CPP).
TheperformanceofCPPbasedSVMmodelshavebeenshowninTable3.
ItissurprisingthatsimplecompositionbasedmodeloutperformsBPPandCPPbasedmodels.
WeachievedmaximumMCC0.
73withaccuracy86.
59%atwindowlength19.
DetailperformanceofCPPbasedSVMmodelforwindowlength19isshowninAdditionalfile1,TableS5.
Theperformanceimprovedsignificantlyforalmostallwin-dowsizesascomparedtobinaryorphysico-chemicalproperties.
AsshowninFigure3,weachievedareaundercurve(AUC)0.
90whichissignificantlybetterthanAUCachievedusingBPPandPPPbasedmodels.
Allmodelsweredevelopedfrommaindatasetusingbal-ancesetofpatternsandevaluatedusingfive-foldcross-validationtechnique.
Figure2Comparisonofaminoacidcompositionofantibodyinteractingresidues(B-cellepitope)andnon-interactingresidues(non-epitope).
Table1TheperformanceofBPPbasedSVMmodeldevelopedusingdifferentwindowlengthsfrom5to21residuesWindowsizeKernelparametersThr*SenSpeAccMCC5t2g0.
01j1c100.
158.
3858.
5558.
470.
177t2g0.
01j1c10.
155.
8759.
8157.
840.
169t2g0.
01j1c10.
155.
6658.
8557.
260.
1511t2g0.
001j1c10061.
5556.
9959.
270.
1913t2g0.
1j1c1062.
5859.
0960.
840.
2215t2g0.
1j1c10059.
9357.
6358.
780.
1817t2g0.
001j1c10058.
3757.
1857.
780.
1619t2g0.
001j1c100.
152.
9263.
7858.
350.
1721t2g0.
001j1c10059.
6957.
2258.
450.
17*(Thr-Threshold,Sen-Sensitivity,Spe-Specificity,Acc-Accuracy,MCC-Matthew'scorrelationcoefficient).
Table2TheperformanceofPPPbasedSVMmodeldevelopeddifferentwindowlengthsfrom5to21residuesWKernelparametersThr*SenSpeAccMCC5t2g0.
00001j1c10-0.
353.
9559.
6256.
780.
147t2g0.
00001j1c100.
155.
8258.
0356.
930.
149t2g0.
00001j1c10054.
5655.
8455.
20.
111t2g0.
00001j1c100.
152.
362.
4857.
390.
1513t2g0.
00001j1c100.
155.
1160.
3757.
740.
1615t2g0.
00001j1c10056.
5760.
0658.
310.
1717t2g0.
00001j1c10060.
1955.
7757.
980.
1619t2g0.
00001j1c10057.
8254.
1555.
980.
1221t1d1057.
3158.
3257.
810.
16AnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page5of9ComparisonwithexistingmethodsInordertovalidateourobservations,wedevelopedandevaluatedourmodelsonbenchmarkdataset;adatasetusedinthepasttobenchmarkearliermethods.
Allwin-dowsizepatternsweremadeuniqueanddividedintorealisticandbalancesetofpatterns.
Realisticsetofpat-ternsrepresentsthereal-lifesituationwherenoninter-actingresiduesaremuchhigherthaninteractingresidues.
Wetrainedandtestedourmodelsonbench-markdatasetusingbalancesetofpatternsandachievedMCC0.
13and0.
72forBPPandCPPrespectively(Table4).
TheseresultsdemonstratesthatCPPbasedmodelsarealsoeffectiveonbenchmarkdataset.
Inordertomakeevaluationmorerealistic,wealsotrainedandtestedourmodelsusingrealisticsetofpatternsbasedonBPPandachievedMCC0.
06and0.
44forBPPandCPPrespectively.
MCCdecreaseswhenweusedrealisticsetofpatternsinsteadofbalancesetofpatternsbutaccuracywasnearlythesameinbothcases.
Inordertocompareperformanceofourmodelwithexistingmeth-odswealsomeasuredperformanceintermofAUC.
Figure4showstheROCplotofourmodelsonbench-markdataset,weachievedAUC0.
56,0.
570.
89formod-elsbasedonBPP,PPPandCPPrespectively.
TheseresultsdemonstratethatCPPbasedmodelsaremoreaccuratethanothermodels.
AUCwasmorethan0.
85forbothsetofpatterns,realisticandbalance(Figure4).
Wecomparedperformanceofourmodelwithexistingmethods(Table5)andobservedthatourmodelisasgoodasanyothermethod.
Thismeansourmodelmaycomplementexistingmethodsandcanbeusedwhenstructureoftheantigenisnotavailable.
ImplementationAuser-friendlywebserver'CBTOPE'wasdevelopedforthepredictionofantibodyinteractingresiduesorB-cellconformationalepitopes.
TheserverisdevelopedusingCGI-Perlscript,HTMLandinstalledonaSunServer(420E)underUNIX(Solaris7)environment.
Theusermaysubmittheaminoacidsequence(s)in'FASTA'for-mat.
Theservergeneratesthe19windowpatternsofallsubmittedsequences,calculatesaminoacidcompositionandpredictsantibodyinteractingresidues.
Theoutputistheaminoacidsequencemappedwithaprobabilityscalerangingfrom0to9foreachaminoacid.
0indi-catestherarestchanceofbeingthatresidueinaB-cellepitopeand9asthemostprobable.
Wesuggestthatforhighspecificity(highconfidence)prediction,usershouldselectthehigherthresholdvaluebutcompromisingthesensitivityofprediction.
However,formaximumpredic-tionofantibodyinteractingresiduesusershouldoptlowerthreshold.
Thereisalwaysinterplaybetweensen-sitivityandspecificity.
Thedefaultthresholdwassetat-0.
3asatthisvalue,sensitivityandspecificitywasfoundequalduringthedevelopment.
Web-serverisfreelyavailableathttp://www.
imtech.
res.
in/raghava/cbtope.
DiscussionIthasbeenagreatchallengefortheacademicianstodevisealgorithmsandmethodsfortheidentificationandmappingofpotentialB-cellepitopesfromanantigensequence.
MuchefforthasbeenputintryingtopredicttheconformationalB-cellepitope.
PreviousmethodspredictconformationalB-cellepitopeswithreasonablyhighaccuracy,thelimitationofthesemethodsisthattheyrequiretertiarystructureoftheantigen.
Experi-mentaltechniquelikeX-raycrystallographyusedfordeterminingstructureofaproteiniscostly,tediousandtimeconsuming.
Tothebestofauthor'sknowledgeTable3TheperformanceSVMmodelsdevelopedusingcompositionprofileofpatternsatdifferentwindowlengthsWindowsizeKernelparametersThr*SenSpeAccMCC5t2g0.
001j1c1061.
7558.
1159.
930.
27t2g0.
001j1c10068.
3562.
265.
270.
319t2g0.
001j1c10073.
4567.
2170.
330.
4111t2g0.
01j1c1-0.
182.
0877.
2679.
670.
5913t2g0.
01j1c10-0.
182.
5784.
1783.
370.
6715t2g0.
01j1c1-0.
179.
9690.
3185.
140.
7117t2g0.
01j1c1-0.
180.
6990.
185.
40.
7119t2g0.
01j1c1-0.
183.
1390.
0686.
590.
7321t2g0.
01j1c1-0.
183.
6288.
9686.
290.
73Figure3TheperformanceofSVMmodelsdevelopedusingcomposition,binaryandphysic-chemicalpropertyprofile.
AnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page6of9thereisnomethodwhichcanpredictconformationalB-cellepitopesinanantigeninabsenceoftertiarystruc-ture.
ThereisaneedtodevelopmethodsforpredictingconformationalB-cellepitopesinanantigenfromitsprimarysequence.
ThisstudydescribesthemethodCBTOPEdevelopedforpredictingconformationalepi-topesofantibodyinteractingresiduesinantigens.
Inordertocompareperformanceofourmodelswechoseabenchmarkdataset,whichwasusedtoevaluatetheperformanceofstructurebasedmethods.
InordertoincreasethedataweincludeddatafromIEDBdatabase.
WepresumedthattheantibodyinteractingresiduesaretheconformationalB-cellepitoperesidues.
Weusedtra-ditionalfeaturesofbinaryandphysico-chemicalprofilesofpatterns,evaluatedby5-foldcrossvalidationwhileusingSVMasaclassifier.
PerformancewasverypoorinBPPmodelsduetothefactthatfor21xWvectorsizeonlyWvaluesrepresent1,therestallare0sothenoiseismoreinBPPmodel.
PPPmodelalsocouldnotper-formwellalthoughitwasearlierusedforlinearandstructurebasedconformationalB-cellepitopeprediction.
Fromthepreliminaryanalysisofthecompo-sitionand2samplelogoplotsofpositiveandnegativepatterns,itwasclearthatthereissignificantdifferenceinthecompositionandsurfacepropensitiesofcertainresidueswhichcanbeexploitedtodiscriminatethepat-terns.
Finallyweusedforthefirsttime,inourstudysimpleaminoacidcompositionmodelofpatterns(CPP)withvectorsizeof20whichwasevaluatedontwodif-ferentdatasets.
TheperformanceimprovedsignificantlyanditisinterestingtonotethatitcanbeusedforthepredictionofconformationalB-cellepitopesdespitethefactthatinCPPmodelwelosttheaminoacidorderinformationunlikeBPP.
Thisproblemmaybeequatedtothesub-cellularlocalizationofproteinswhereinitwasobservedthatsimpleaminoacidcompositionmodelperformbetterthanotherfeatures.
Butunlikesub-cellularlocalizationweexploitedcompositionofpatternsinsteadofwholeproteinsequence.
ItshouldbenotedthatdespitethepredictionofantibodyinteractingorindividualB-cellepitoperesidues,beingasequencebasedmethodandthelackof3Dstructuralinput,CBTOPEcannotassistindeterminingthenumberanddistanceneededtomakeanepitopesegmentintheanti-gensequence.
Thisinformationcanbeobtainedbymappingofthepredictedresiduesonthemodeledstructure.
Wehopethatthepresentmodelisuniqueinitskindandwillcomplimenttheavailablestructurebasedmethodsusedforthepredictionofantibodyinter-actingresiduesorconformationalB-cellepitopes.
ConclusionWeshowedthatsimpleantigensequencecanbeusedforthepredictionofconformationalB-cellepitopesandnostructureorhomologyisrequired.
Weintroducedforthefirsttimeconceptoflocalaminoacidcompositionofanti-gen.
WeshowedthatourCPPcompositionbasedSVMmodeloutperformedotherstructuremethodswithbettersensitivityandAUConthesamebenchmarkdataset.
AdditionalmaterialAdditionalfile1:AdditionalfileforCBTOPE.
Additionalfile1containingBPPandPPPmatrixanddetailedthreshold-wiseresultsofselectedwindowsandkernels.
Table4TheperformanceofBPPandCPPbasedSVMmodelonBenchmarkdataset,developedusingbalanceandrealisticsetofpatternsTypeofPatternsetModelSVMparametersThr*SenSpeAccMCCRealisticBPPt2g0.
001j10c10-0.
250.
4960.
2859.
490.
06CPPt2g0.
001j10c10-0.
380.
4184.
6484.
300.
44BalanceBPPt2g0.
01j1c100.
161.
3151.
2256.
270.
13CPPt2g0.
01j1c10082.
3689.
4285.
890.
72Modelsweredevelopedusingwindowsize19.
Figure4TheperformanceofSVMmodelsonBenchmarkdatasetasshownbyROCplot.
AnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page7of9AcknowledgementsTheauthor'sarethankfultotheCouncilofScientificandIndustrialResearch(CSIR)andDepartmentofBiotechnology(DBT),GovernmentofIndiaforfinancialassistance.
HifzurRahmanAnsariisaSeniorResearchFellowandfinanciallysupportedbyCSIR.
Authors'contributionsHRAcarriedoutthedataanalysisandinterpretation,developedcomputerprograms,wrotethemanuscriptanddevelopedtheweb-server.
GPSRconceivedandcoordinatedtheproject,guideditsconceptionanddesign,helpedintheinterpretationofdata,refinedthedraftedmanuscriptandgaveoverallsupervisiontotheproject.
Bothauthorsreadandapprovedthefinalmanuscript.
CompetinginterestsTheauthorsdeclarethattheyhavenocompetinginterests.
Received:20May2010Accepted:20October2010Published:20October2010References1.
GershoniJM,Roitburd-BermanA,Siman-TovDD,TarnovitskiFreundN,WeissY:Epitopemappingthefirststepindevelopingepitope-basedvaccines.
BioDrugs2007,21:145-156.
2.
PomesA:RelevantBcellepitopesinallergicdisease.
IntArchAllergyImmunol2010,152:1-11.
3.
AlmagroJC:Identificationofdifferencesinthespecificity-determiningresiduesofantibodiesthatrecognizeantigensofdifferentsize:implicationsfortherationaldesignofantibodyrepertoires.
JMolRecognit2004,17:132-143.
4.
MacCallumRM,MartinAC,ThorntonJM:Antibody-antigeninteractions:contactanalysisandbindingsitetopography.
JMolBiol1996,262:732-745.
5.
VanRegenmortelMH:Structuralandfunctionalapproachestothestudyofproteinantigenicity.
ImmunolToday1989,10:266-272.
6.
FrankR:TheSPOT-synthesistechnique.
Syntheticpeptidearraysonmembranesupports–principlesandapplications.
JImmunolMethods2002,267:13-26.
7.
XingdongY,XinglongY:Anintroductiontoepitopepredictionmethodsandsoftware.
ReviewsinMedicalVirology2009,19:77-96.
8.
SahaS,RaghavaGP:SearchingandmappingofB-cellepitopesinBcipepdatabase.
MethodsMolBiol2007,409:113-124.
9.
VitaR,ZarebskiL,GreenbaumJA,EmamiH,HoofI,SalimiN,DamleR,SetteA,PetersB:Theimmuneepitopedatabase2.
0.
NucleicAcidsRes2010,38:D854-862.
10.
SahaS,RaghavaGP:PredictionmethodsforB-cellepitopes.
MethodsMolBiol2007,409:387-394.
11.
SahaS,RaghavaGP:BcePred:PredictionofcontinuousB-cellepitopesinantigenicsequencesusingphysico-chemicalproperties.
ICARIS,LNCS2004,3239:197-204.
12.
LarsenJE,LundO,NielsenM:ImprovedmethodforpredictinglinearB-cellepitopes.
ImmunomeRes2006,2:2.
13.
SahaS,RaghavaGP:PredictionofcontinuousB-cellepitopesinanantigenusingrecurrentneuralnetwork.
Proteins2006,65:40-48.
14.
NovotnyJ,HandschumacherM,HaberE,BruccoleriRE,CarlsonWB,FanningDW,SmithJA,RoseGD:Antigenicdeterminantsinproteinscoincidewithsurfaceregionsaccessibletolargeprobes(antibodydomains).
ProcNatlAcadSciUSA1986,83:226-230.
15.
Kulkarni-KaleU,BhosleS,KolaskarAS:CEP:aconformationalepitopepredictionserver.
NucleicAcidsRes2005,33:W168-171.
16.
HasteAndersenP,NielsenM,LundO:PredictionofresiduesindiscontinuousB-cellepitopesusingprotein3Dstructures.
ProteinSci2006,15:2558-2567.
17.
PonomarenkoJV,BournePE:Antibody-proteininteractions:benchmarkdatasetsandpredictiontoolsevaluation.
BMCStructBiol2007,7:64.
18.
SweredoskiMJ,BaldiP:PEPITO:improveddiscontinuousB-cellepitopepredictionusingmultipledistancethresholdsandhalfsphereexposure.
Bioinformatics2008,24:1459-1460.
19.
MoreauV,FleuryC,PiquerD,NguyenC,NovaliN,VillardS,LauneD,GranierC,MolinaF:PEPOP:computationaldesignofimmunogenicpeptides.
BMCBioinformatics2008,9:71.
20.
HuangY,BaoY,GuoS,WangY,ZhouC,LiY:Pep-3D-Search:amethodforB-cellepitopepredictionbasedonmimotopeanalysis.
BMCBioinformatics2008,9:538.
21.
HuangJ,GutteridgeA,HondaW,KanehisaM:MIMOX:awebtoolforphagedisplaybasedepitopemapping.
BMCBioinformatics2006,7:451.
22.
BublilEM,FreundNT,MayroseI,PennO,Roitburd-BermanA,RubinsteinND,PupkoT,GershoniJM:StepwisepredictionofconformationaldiscontinuousB-cellepitopesusingtheMapitopealgorithm.
Proteins2007,68:294-304.
23.
PonomarenkoJ,BuiH-H,LiW,FussederN,BourneP,SetteA,PetersB:ElliPro:anewstructure-basedtoolforthepredictionofantibodyepitopes.
BMCBioinformatics2008,9:514.
24.
LiW,GodzikA:Cd-hit:afastprogramforclusteringandcomparinglargesetsofproteinornucleotidesequences.
Bioinformatics2006,22:1658-1659.
25.
GarnierJ,GibratJF,RobsonB:GORmethodforpredictingproteinsecondarystructurefromaminoacidsequence.
MethodsEnzymol1996,266:540-553.
26.
AnsariHR,RaghavaGP:IdentificationofNADinteractingresiduesinproteins.
BMCBioinformatics2010,11:160.
27.
KumarM,GromihaMM,RaghavaGP:PredictionofRNAbindingsitesinaproteinusingSVMandPSSMprofile.
Proteins2008,71:189-194.
28.
BhasinM,RaghavaGP:Pcleavage:anSVMbasedmethodforpredictionofconstitutiveproteasomeandimmunoproteasomecleavagesitesinantigenicsequences.
NucleicAcidsRes2005,33:W202-207.
29.
ChouKC,ShenHB:Signal-CF:asubsite-coupledandwindow-fusingapproachforpredictingsignalpeptides.
BiochemBiophysResCommun2007,357:633-640.
30.
XiaoX,WangP,ChouKC:GPCR-CA:AcellularautomatonimageapproachforpredictingG-protein-coupledreceptorfunctionalclasses.
JComputChem2009,30:1414-1423.
31.
XiaoX,ShaoS,DingY,HuangZ,ChouKC:Usingcellularautomataimagesandpseudoaminoacidcompositiontopredictproteinsubcellularlocation.
AminoAcids2006,30:49-54.
32.
RubinsteinND,MayroseI,MartzE,PupkoT:Epitopia:aweb-serverforpredictingB-cellepitopes.
BMCBioinformatics2009,10:287.
33.
GranthamR:Aminoaciddifferenceformulatohelpexplainproteinevolution.
Science1974,185:862-864.
Table5OverallperformanceofstructurebasedandCBTOPEalgorithmsonbenchmarkdatasetEvaluationparameterProMatePSI-PREDbestpatchPatchDockbestmodelClusPro(DOT)bestmodelCEPDiscoTope(-7.
7)CBTOPE*(ThisStudy)Sen*0.
090.
330.
430.
450.
310.
420.
801-Spe0.
080.
140.
110.
070.
220.
210.
15PPV0.
100.
190.
260.
390.
110.
160.
31Acc0.
840.
820.
850.
890.
740.
750.
84AUC0.
510.
600.
660.
690.
540.
600.
89*(Thr-Threshold,Sen-Sensitivity,Spe-Specificity,Acc-Accuracy,PPV-positivepredictivevalue)AnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page8of934.
KarplusPA,SchulzGE:PredictionofChainFlexibilityinProteins-AtoolfortheSelectionofPeptideAntigens.
Naturwissenschafren1985,72:212-213.
35.
KolaskarAS,TongaonkarPC:Asemi-empiricalmethodforpredictionofantigenicdeterminantsonproteinantigens.
FEBSLett1990,276:172-174.
36.
ParkerJM,GuoD,HodgesRS:Newhydrophilicityscalederivedfromhigh-performanceliquidchromatographypeptideretentiondata:correlationofpredictedsurfaceresidueswithantigenicityandX-ray-derivedaccessiblesites.
Biochemistry1986,25:5425-5432.
37.
PonnuswamyPK,PrabhakaranM,ManavalanP:Hydrophobicpackingandspatialarrangementofaminoacidresiduesinglobularproteins.
BiochimBiophysActa1980,623:301-316.
38.
KaundalR,RaghavaGP:RSLpred:anintegrativesystemforpredictingsubcellularlocalizationofriceproteinscombiningcompositionalandevolutionaryinformation.
Proteomics2009,9:2324-2342.
39.
BhasinM,RaghavaGP:GPCRpred:anSVM-basedmethodforpredictionoffamiliesandsubfamiliesofG-proteincoupledreceptors.
NucleicAcidsRes2004,32:W383-389.
40.
ChenC,ChenL,ZouX,CaiP:PredictionofproteinsecondarystructurecontentbyusingtheconceptofChou'spseudoaminoacidcompositionandsupportvectormachine.
ProteinPeptLett2009,16:27-31.
41.
ChenJ,LiuH,YangJ,ChouKC:PredictionoflinearB-cellepitopesusingaminoacidpairantigenicityscale.
AminoAcids2007,33:423-428.
42.
YangZR:Biologicalapplicationsofsupportvectormachines.
BriefBioinform2004,5:328-338.
43.
ChouKC,ZhangCT:Predictionofproteinstructuralclasses.
CritRevBiochemMolBiol1995,30:275-349.
44.
ChouKC,ShenHB:Cell-PLoc:apackageofWebserversforpredictingsubcellularlocalizationofproteinsinvariousorganisms.
NatProtoc2008,3:153-162.
45.
ChouKC,ShenHB:Anewmethodforpredictingthesubcellularlocalizationofeukaryoticproteinswithbothsingleandmultiplesites:Euk-mPLoc2.
0.
PLoSONE2010,5:e9931.
46.
VacicV,IakouchevaLM,RadivojacP:TwoSampleLogo:agraphicalrepresentationofthedifferencesbetweentwosetsofsequencealignments.
Bioinformatics2006,22:1536-1537.
doi:10.
1186/1745-7580-6-6Citethisarticleas:AnsariandRaghava:IdentificationofconformationalB-cellEpitopesinanantigenfromitsprimarysequence.
ImmunomeResearch20106:6.
SubmityournextmanuscripttoBioMedCentralandtakefulladvantageof:ConvenientonlinesubmissionThoroughpeerreviewNospaceconstraintsorcolorgurechargesImmediatepublicationonacceptanceInclusioninPubMed,CAS,ScopusandGoogleScholarResearchwhichisfreelyavailableforredistributionSubmityourmanuscriptatwww.
biomedcentral.
com/submitAnsariandRaghavaImmunomeResearch2010,6:6http://www.
immunome-research.
com/content/6/1/6Page9of9