Multi-pathway feature-level interpretability in MHC-i antigen presentation via concept-based modeling

Accurate prediction of MHC-I antigen presentation is central to neoantigen discovery and immunotherapy development. Although recent deep-learning based predictors achieve high accuracy, most operate as black-box models, which limits interpretability, and consequently trustworthiness of these predictors. We present MHCCBM, a gray-box framework that decomposes antigen presentation into a set of pathway-level intermediate concepts. Each step, i.e. proteasomal cleavage, TAP transport, peptide-MHC binding affinity, and chaperone dependency is treated as an independent predictive module, allowing different computational or experimental estimators to be substituted without altering the overall model. This architecture naturally supports representational multimodality, permitting the integration of sequence-based, structure-informed, or empirical predictors. In our reference implementation, ESM-2-derived models estimate peptide processing and chaperone dependency, while peptide-MHC binding affinity is predicted using MHCflurry. The concept outputs are combined using logistic regression. Our accuracy is comparable to that of state-of-the-art accuracy while providing MHC-I pathway level interpretability consistent with known cellular mechanisms.