Structural classification of peptide/HLA complexes enables accurate high-throughput modeling

Abstract

Abstract The class I major histocompatibility complex (MHC-I) proteins display 8 to 15 residue peptides derived from endogenous proteins on the cell surface for immune surveillance. Accurate structural modeling of peptides bound to HLA (the human MHC) proteins has been mired by the inability to capture the central residues of the peptide. Here, an analysis of high-resolution X-ray crystal structures revealed that diverse peptide conformations arise as a result of local changes in the backbone configuration. As a result, we established an internal-coordinate based system to assess structural similarity between peptide backbones and found that peptide/HLA complexes, encompassing a wide range of allotypes and peptide sequences, present a discrete set of peptide backbone conformations. Leveraging these representative backbones and a regression model trained on a physically realistic scoring function, we developed RepPred, a structural modeling approach which predicts near native nonamer/HLA-A*02:01 complexes with high accuracy and throughput. Our method significantly outperformed five state-of-the-art peptide/HLA modeling approaches on the basis of structural accuracy. Thus, the structural information obtained from RepPred can be used to address long-standing immunology challenges such as predicting peptide immunogenicity. Supported by grants from NIH (R01AI143997, R35GM125034, U01DK112217).