The landscape of T cell epitope immunogenicity in sequence space

The existence of population-wide T cell immunity is widely recognized for multiple pathogen-derived immunodominant epitopes, despite the vast diversity and individualized nature of T cell receptor (TCR) repertoire. We thus hypothesized that population-wide epitope immunogenicity could be probabilistically defined by exploiting public TCR features. To gain a proof-of-concept, here we describe a machine learning framework yielding probabilistic estimates of immunogenicity, termed \"immunogenicity scores\", by utilizing features designed to mimic thermodynamic interactions between peptides bound to major histocompatibility complex (MHC) and TCR repertoire. Immunogenicity score dynamics among observed and computationally simulated single amino acid mutants delineated the landscape of position- and residue-specific mutational impacts, and even quantitatively estimated escaping potentials of known epitopes with remarkable positional specificity. This study illustrates that the population-wide aspect of adaptive immunity is predictable via non-individualized approach, possibly indicating antigen-guided convergence of human T cell reactivity.