Modeling antigen presentation and immune recognition with Restricted Boltzmann Machines

Immune recognition of infected and malignant cells requires presentation on their surface of antigens (i.e. short peptides) by human leukocyte antigen class I (HLA-I) proteins, which are coded by one of the most polymorphic alleles in the human genome. We first introduce RBM-MHC, a method for prediction of HLA-presented antigens based on the statistical physics framework known as Restricted Boltzmann Machine (RBM) to infer generative models from amino acid sequence data. RBM-MHC can be trained on custom and newly available samples with no or a small amount of HLA annotations, it ensures improved predictions for rare HLA alleles and matches state-of-the-art performance for well characterized alleles while being less data-demanding. RBM-MHC is shown to be a flexible and easily interpretable method that can be used as predictor of cancer neo-antigens and viral epitopes, as tool for feature discovery, and to reconstruct peptide motifs presented on specific HLA molecules. Next we extend the RBM approach to modeling the complementary process of immune recognition of presented antigens. The approach is able to discriminate responses specific to different antigens and highlights amino acid patterns that are central to such specificity at the molecular level.