Establishment And Validation Of A Web-based Tool For Prediction And Design Of HLA Class â…  Supertype Specific Epitopes

Major histocompatibility complex (MHC) molecules bind short peptides derived fromantigens and present these peptides on the surface of antigen-presenting cells (APCs) forrecognition by T-cell receptors (TCRs). T-cell epitopes, which are peptides that bind toMHC molecules and trigger an immune response, are important for studying the specificityof the immune system. Identification of these epitopes allows the development ofdiagnostics and peptide-based vaccines against viral infections certain tumors and evenautoimmune diseases. Binding of peptides to MHC molecules is the most selective step anda prerequisite for T cell recognition; however, identifying peptides that bind to MHCs is anexpensive, laborious and time-consuming process. Computational prediction methodsenable the systematic screening and identification of candidate T-cell epitopes from largesets of protein antigens and effectively reduce the number of peptides that must besynthesized and assayed.The MHC genes in humans are named human leucocyte antigens (HLAs), Most HLAclass I molecules can be assigned to supertypes based on their overlapping peptide-bindingspecificities, and peptides that bind to a given HLA class I molecule frequently bind tomultiple HLA molecules belonging to the same supertype. These promiscuously bindingpeptides, called supertype-specific binders, have great potential in the development ofvaccines with broad and unbiased coverage across the human population. However, eachmolecule in a supertype has unique, allele-specific peptide-binding preferences, and the factthat a peptide can bind to a given allele does not necessarily indicate whether it can alsobind to another allele within the same supertype. As a result, identifying peptides that bindto HLA class I molecules has been turned to distinguishing the cross-binding abilities ofpeptides to alleles within the same supertype. Understanding the contribution of residues at the peptide to the HLA supertype-specific cross-binding should provide importantinformation for the design of epitope-based vaccines with high population coverage.The generation of any prediction method depends on the experimentally verified data.In this study, we traversed the available MHC-peptide databases online, and extractedHLA-I-peptide binding data. The peptides that were related to HLA molecules of the samesupertype were integrated into one group where peptides were annotated by the bindinginformation for alleles of the same supertype, and A HLA class I supertype-specific binderdatabase (HLASSB_DB) were generated. HLASSB_DB contained18157records aboutpeptides cross-binding to HLA molecules belonging to the same supertype, covering38alleles in10supertypes. This database provide three different retrieval methods: peptidesequence, supertype and corresponding alleles, and protein sequence.The performance of in silico methods models was determined by the properties of thepeptides in the training datasets. However, the available peptide data contain significantbiases because scientists often prefer to verify the binding affinity of potential binders thatshare the binding motif of the allele under study. We investigated the effect of the peptidesin the training dataset on the performance of the generated model. Our result indicated thatmodels trained on imbalanced datasets with a relative paucity of non-motif sharing peptideshad a low performance for biological dataset (identifying binders from a given antigen).Moreover, we found there was a discrepancy between the classification of binders frombiological data and classification of binders from super-motif-sharing peptides: the existingof non-motif-containing non-binders would improve the predictive performance forbiological dataset but reduce the performance for motif-sharing peptides (distinguishing thecross-binding abilities of peptide to HLA molecules in the same supertypes). Most HLAclass I molecules could be assigned to supertypes based on their overlappingpeptide-binding specificities, therefore, we proposed a supertype-based method for themodeling of the HLA class I-peptide binding: candidates of peptides binding to alleles in agiven supertypes were screened using the super-motifs, and then the peptides binding tospecific allele in the supertype were predicted by the model trained on thesuper-motif-sharing peptides. The efficacy of this supertype-based method was examined inthree matrix-based methods and one machine learning method for20alleles in HLAsupertype A1, A2, A3, A24, B44and B7. Evaluations on several benchmark datasets indicated that the supertype-based method achieved remarkable success in improving theprediction of HLA-binding peptides. This supertype-based method was universal for theclassification methols.Applying the supertype-based method to a multiplicative average relative bindingmatrix (mARB)-based method, we build a web server for prediction and design of HLAclass I supertype-specific binders (HLASSB_PreD). Comparative study based onbenchmark datasets indicated that HLASSB_PreD outperformed the common binaryclassification methods (SVMHC RANKPEP YKW). Moreover, the generated predictivemodels for alleles in the same supertype shared a conversed threshold, which enableHLASSB_PRED to analysis the contribution of residue at each position of a given peptideto the supertype-specific cross-binding and guide the design of HLA supertype-specificbinders with high cross-binding ability. We defined the HLA supertype-specificcross-binding motifs for supertype A2and A3using the generated mARB matrices, andvalidated these motifs in binding peptides with various cross-binding abilities. Evaluationon the Poly-A analogues and epitope HBsAg18-27point mutational analogues indicated theHLASSB_PreD obtained a high predictive accuracy93%for Poly-A analogues and100%for HBsAg18-27analogues, which further validated the performance of HLASSB_PreD indesin of the supertype-spcific binders. We screened the HLA-A3supertype specificepitopes of HBcAg and HBsAg using HLASSB_PreD and obtained five candidates. Theresults of ELISPOT assay indicated that each of these peptids could induce the immuneresponse of the PBMC with at least two HLA-A3allele retrictions, which validate theability of HLASSB_PreD to identify supertype-specific epitopes from a given antigen.HLA molecules exhibit a high frequency of polymorphisms. Each encoded HLAmolecule binds to a distinct set of peptides, but binding preferences of most alleles have notyet been experimentally characterized. Thus, prediction of peptides bound to HLAmolecules with only a small number of experimentally obtained binders, called pan-specificprediction have received intense interest recently. Pan-specific methods use experimentaldata of multiple HLA alleles (source alleles) as input and attempt to predict binders of notonly the input alleles but also the alleles with very few or even no known binders (targetalleles). In this study, we proposed a supertype-based modular concept, generating thepan-specific binding matrices using the binding matrices of well-studied alleles in supertype A2and A3according the similarity of the position modular, and pan-specificprediction models for35alleles in supertype A2and37alleles in supertype A3weregenerated. The results of leave-one (allele)-out cross-validation indicate the generatedpans-specific prediction models were appropriate for allele with only a small number ofexperimentally obtained binders. We screened the HLA-A*3303epitopes of HBcAg usingpan-specific prediction method and obtained10candidates. The results of ELISPOT assayindicated that7of10peptids could induce the immune response of the PBMC withHLA-A*3303restriction, which validate the ability of the pan-specific model to identifyepitopes from a given antigen.Our results and the methods proposed in this study should provide important insightsinto how the prediction methods could be improved. The generated database (HLASSB_DB)and prediction servers (HLASSB_PreD) are free for users, both of which could be accessedvia http://www.immunoinformatic.net. We believe that our study should facilitate thedevelopment of the epitope-based vaccines, especially the vaccines with high populationcoverage.