Improved predictive models of peptide presentation on MHC I

Large surveys of peptides naturally presented on major histocompatibility class I (MHC I) proteins have enabled improved MHC I ligand prediction by dramatically expanding the available data for many MHC I alleles. However, it is unclear to what extent antigen processing signals can also be learned from these datasets. Here, we developed a predictor of antigen processing by training neural networks to discriminate mass spec-identified MHC I ligands from unobserved peptides, where both classes of peptides are predicted to be strong MHC I binders. The resulting predictor shows qualitative consistency with established preferences for the transporter associated with antigen processing, proteasomal cleavage, and endoplasmic reticulum aminopeptidases. When we combined the antigen processing predictor with a novel pan-allele MHC I binding predictor in a logistic regression model, the combination model significantly outperformed the two components alone as well as the NetMHCpan 4.0 and MixMHCpred 2.0.2 tools at predicting mass spec-identified MHC I ligands. Our predictors are implemented in the open source MHCflurry package, version 1.6.0 (github.com/openvax/mhcflurry).