MultiMHC: a multiscale learning framework integrating sequence and residue contact topology for peptide–MHC binding prediction

Abstract

Abstract Accurate prediction of peptide–major histocompatibility complex (pMHC) binding is central to understanding antigen presentation and identifying T-cell epitopes in cancer immunotherapy. Although recent methods have achieved notable progress, most approaches rely predominantly on sequence-based features and only partially capture structural determinants of pMHC recognition. In this study, we present MultiMHC, a multiscale deep learning framework that integrates complementary sequential and biologically-informed residue contact topology for pMHC binding prediction. MultiMHC combines a recurrent neural network to model global sequence context with a graph-based representation to encode residue contact patterns derived from known pMHC binding interfaces. Features from these two pathways are subsequently fused to generate binding predictions across peptides of varying lengths. Benchmark evaluations show that MultiMHC achieves performance comparable to existing state-of-the-art methods, with particularly strong performance in ranking-based assessments. Moreover, evaluation on the Tumor Neoantigen Selection Alliance dataset showed that MultiMHC binding scores can enrich experimentally validated T-cell epitopes among top-ranked candidates, supporting its potential utility as an upstream binding-based filtering step in neoantigen discovery workflows. Together, these results highlight the value of incorporating residue interaction patterns and lightweight structural information into pMHC modeling and establish MultiMHC as a robust and extensive framework for antigen presentation analysis.