Abstract: Class I MHC proteins present short peptides to T cells of the immune system. Vaccines incorporating peptides from viruses or containing mutations present in the cancer genome represent a low-cost, broadly applicable type of immunotherapy which can sensitize the immune system to disease, break tolerance, and confer long-lasting immunologic memory with fine specificity. Predicting which candidate neoantigens are likely immunogenic is imprecise, limiting applicability in treating spontaneous disease. We previously demonstrated prediction of immunogenicity from physicochemical characteristics of the peptide in its structural context are more accurate than predictions from sequence-based tools. Our estimation of these characteristics, however, is limited by available three-dimensional structures and the accuracy of computational structural predictions.
In this study, we identified a group of mutant peptides presented in a cancer model and rationalized their immunogenicity with molecular modeling, as predicted peptide binding affinity alone could not explain reactivity. After solving crystal structures of two peptide-MHCs, we developed a method to improve peptide-MHC structural modeling. When evaluated on 93 crystal structures of peptides bound to the prevalent human class I MHC, HLA-A*02:01, our method demonstrated significantly higher accuracy than conventional model selection. These results are of interest for improving our ability to identify “differences from self” MHC-presented peptides exhibit, but are also of translational interest in producing realistic structural and energetic signatures for predicting immunogenicity.