A phage-display library of random peptides is a combinatorial experimental technique that can be harnessed for studying antibody–antigen interactions. In this technique, a phage peptide library is scanned against an antibody molecule to obtain a set of pe
correspondingtouandvisshorterthanadefaultcutoffdis-tanceof4s[asinRef.(15,16)].Mappingapeptidetothesurfacegraph
Givenaquerypeptidesequence,ouraimisto ndthepaththatyieldsthealignmentwithmaximalweight(alignmentscore).LetQ¼(qP¼(p1,...,q)asimplek)beaquerypeptideoflengthkand1,...,pkpathinGwherepi2V.Leth(qqueryi,presiduei)denoteanaminoacidsimilarityscorebetweenaqiandgraphvertexpi(seesection‘Scoringaminoacidsimilarities’below).TheweightofthealignmentbetweenQandPisthesummationoftheaminoacidsimilar-ityscoresbetweenqueryresiduesandgraphverticesXkWðQ PÞ¼
hðqi piÞ:
i¼1
Anaf nity-selectedpeptide,althoughlinear,adoptsadistinct3Dconformation.Itisthuspossiblethatnotallresiduesinsuchapeptidecorrespondtoresiduesontheantigen.Hence,residuesofthepeptideareallowedtobeunmatchedtographvertices(i.e.gaps).Speci cally,foreachunmatchedresidueagappenaltydDisaddedtothealignmentscore.Inaddition,weallowforcasesinwhichonlyasubsetofpeptideresiduesarematched,i.e.nogappenaltyisgiventounmatchedresiduesateitherendofthepeptide.Hence,thealgorithmperformsalocal,ratherthanaglobal,alignment.
Ideally,allpossiblepathsinGshouldbescannedandthepathwiththeoptimalalignmentdetected.However,theenumerationoverallpossiblesimplepathsinGiscom-putationallyintractableforrealistic-sizeproblems.Thisintractabilitystemsfromtherequirementthatapathshouldnotcontaincycles,i.e.eachgraphvertexshouldappearatmostonceinthealignment.Toaddressthisconstraint,wedevelopedadynamicprogrammingbasedalgorithm,whichreliesonthecolor-codingtechniqueofAlonetal.(17).Colorcodingisanef cienttechniquefordetectingsimplepathsinagraph.Itproceedsbyassigningeveryvertexv2Vacolorc(v)drawnuniformlyatrandomfromthesetC¼{1,...,k},wherekisthelengthofthequerypeptide.Givensuchacoloredgraph,adynamicprogrammingscheme(detailedbelow)isusedto ndthehighestscoringcolorfulpath(spanningkdistinctcolors).However,sincethecoloringisrandom,thereisnoguaranteethatthebestalignment(theonethatwegloballyaimto ndregardlessofthecolor-ing)correspondstoapathofdistinctcolors.Thus,manyrandomcoloringtrialsareneeded.Inanygiveniteration,theprobabilitythattheoptimalpathiscolorfulisk!/kk.Hence,theminimalnumberoftrialsneededinordertoreceivethebestpathwithprobabilityabovepis½logð1ÀpÞ/logð1Àk!/kkÞ .Inallrunsconductedherepwassetto0.95toensurethatthebestpathisfoundwithahighprobability.
Givenacoloredgraph,adynamicalgorithmisusedto ndtheoptimalalignedpathofkdistinctcolors.LetW(i,j,S)denotethemaximumweightofanalignmentforthe rstiresiduesinthequerythatendsatvertexj2VandvisitsavertexofeachcolorinS,whereSisasubsetofthecolors.W(i,j,S)iscomputedrecursively
NucleicAcidsResearch,2006,Vol.00,No.00
3
asfollows:
8
<W½iÀ1 m SÀcðjÞ þhðqi jÞðm jÞ2E
Wði j SÞ¼maxm2V:
WðiÀ1 j SÞþdD
W½1 j cðjÞ ¼maxfhðq1 jÞ 0g
ThemaximalweightofthealignmentisThecorrespondingalignmentisobtainedthroughj2maxV S C
Wðk j SÞ.
standarddynamicprogrammingbacktracking.Therunningtimeofeachcoloringtrialdependsonthelengthofthepeptideandthesizeofthesurfacegraph,andisOð2kjEjÞ:Asimilardynamicprogrammingschemewasusedin(18)forqueryingpathwaysinaprotein–proteininteractionnetwork.
Thedynamicprogrammingdetailedaboveconsidersthewholespaceofdifferent{i,j,S}values.Practically,therun-ningtimemaybereducedconsiderablybyapplyingabranch-and-boundtechnique.Thefull{i,j,S}matrixis rst lledforsmallvaluesofi.Thedynamicalgorithmproceedsonlyforentriesthatcantheoreticallyresultinhigherscoringalign-mentsthanthebestalignmentproducedinpreviouscoloringtrials.Wehavefoundthatintroducingthisbranch-and-boundadditionresultsin$10-foldspeed-upoftherunningtime.AssigningaP-valuetoapeptide-pathalignment
Theprobabilityofobtainingamatchwithaspeci cscoreatrandom(P-value)isusedintheproceedingclusteringstageandisgeneratedasfollows.First,foreachpeptideanempiricalbackgrounddistributionofalignmentscoresiscon-structedbygeneratingasetofmrandomsequencesequalinlengthtothatofthegivenpeptide.Theaminoacidsofeachsuchsequencearedrawnwithprobabilitiesderivedfromtheirfrequenciesinthesurfaceoftheantigen.Thus,thispro-cessapproximatesthegenerationofmrandompaths(inallrunsconductedm¼106).Eachrandomsequenceisthenalignedtothegivenpeptide.AP-valueforeachalignedpathcanbeobtainedfromthisempiricaldistribution.How-ever,inordertoconstructanaccuratedistribution,anexceed-inglylargenumberofrandomsequencesisneeded.Thedistributionoflocalalignmentscoresisthusapproximatedusinganextremevaluedistribution,whoseparametersare t-tedfromtheempiricaldistributionusingthemethodofmoments(19).
Clusteringpathstopredictedepitopes
Onceeachpeptideisalignedtoahighestscoringpath,thequestionishowtocombinethesepathstoobtainapredictedepitope.Wede netwopathstobeconnectediftheyshareatleastoneresidue.Thisde nesclustersofconnectedpaths.ThescoreofeachsuchclusteristhesumofÀlog(P-value)ofthepathswithinit.Weusethisscoreratherthanthesumofalignmentscoressinceinsomedatasetsnotallpeptidesareequalinlengthandthustheiralignmentscoresarenotcomparable.
Thedefaultclusteringstepviewsthepeptidesasasetofuniquesequences.Obviously,differentpeptidesbindtheantibodywithdifferentaf nities,whichcouldbere ectedbydifferentfrequenciesofisolationforeachpeptide.Thistypeofinformationcanbeintegratedintheclusteringstep
35, No. 1 69–78 doi:10.1093/nar/gkl975 Epitope mapping using combinatorial phage-display libraries: a graph-based algorithm Itay Mayrose, Tomer Shlomi...
35, No. 1 69–78 doi:10.1093/nar/gkl975 Epitope mapping using combinatorial phage-display libraries: a graph-based algorithm Itay Mayrose, Tomer Shlomi...
热门文档