Predicting MHC class I epitopes in large datasets

Background: Experimental screening of large sets of peptides with respect to their MHC binding capabilities is still very demanding due to the large number of possible peptide sequences and the extensive polymorphism of the MHC proteins. Therefore, there is significant interest in the development of computational methods for predicting the binding capability of peptides to MHC molecules, as a first step towards selecting peptides for actual screening.Results: We have examined the performance of four diverse MHC Class I prediction methods on comparatively large HLA-A and HLA-B allele peptide binding datasets extracted from the Immune Epitope Database and Analysis resource (IEDB). The chosen methods span a representative cross-section of available methodology for MHC binding predictions. Until the development of IEDB, such an analysis was not possible, as the available peptide sequence datasets were small and spread out over many separate efforts. We tested three datasets which differ in the IC₅₀ cutoff criteria used to select the binders and non-binders. The best performance was achieved when predictions were performed on the dataset consisting only of strong binders (IC₅₀ less than 10 nM) and clear non-binders (IC₅₀ greater than 10,000 nM). In addition, robustness of the predictions was only achieved for alleles that were represented with a sufficiently large (greater than 200), balanced set of binders and non-binders.Conclusions: All four methods show good to excellent performance on the comprehensive datasets, with the artificial neural networks based method outperforming the other methods. However, all methods show pronounced difficulties in correctly categorizing intermediate binders.