Peptides binding to HLA molecules elicit specific T cell immune responses and are useful in the development of peptide vaccines and therapeutics. Thus, prediction of HLA-binding peptide is critical. Here, a prediction model is described. Many statistical and molecular mechanics models for HLA peptide binding prediction have been developed and tested during the last decade. However, their efficiency and HLA diversity coverage are far from satisfactory. Analysis of structural data revealed that there are polymorphic pockets in the HLA peptide binding groove that can accommodate the anchor residues of the peptide. The residues that form the pocket determine the geometry and chemical properties of these structural pockets and thus determine the antigen peptides that would be preferentially bound. This accounts for the differential ability of different alleles to bind a variety of peptides. Thus, the peptide-binding groove of the HLA is essentially an exchange or shuffling of pockets among different allele. Based on this analysis, a novel predictive model for HLA peptide binding was developed. The model was extensively cross-validated using peptide binding data. This model is superior to many existing methods because of its potential application to any given HLA allele whose sequence is clearly defined.