TopoMHC: Sequence–Topology Fusion for MHC Binding

Ciyuan Yu, Hongzong LI, Jiahao MA, Shiqin Tang, Ye-Fan Hu, Jian-Dong Huang

Keywords: immunogenicity prediction, major histocompatibility complex, peptide representation learning, statistical topology, persistent homology, protein language models, cross-modal learning, vaccine design

Abstract: Accurate prediction of peptide immunogenicity, particularly the binding affinity to major histocompatibility complex (MHC) molecules, is critical for vaccine design and immunotherapy. Existing approaches are predominantly sequence-based and often overlook structural variability and topological organization, which restricts predictive reliability. In this work, we introduce a multi-modal framework that integrates sequence embeddings from a pre-trained protein language model (e.g., ESM-C) with topology-informed descriptors derived from peptide conformations. We generate peptide conformers using molecular dynamics simulations and RDKit-based methods, and from these conformations we compute persistent homology invariants, Betti numbers, geometric statistics, and residue connectivity measures. These topological features are then fused with sequence embeddings through a cross-attention mechanism, allowing the model to capture both local sequence patterns and global conformational organization. Extensive experiments demonstrate consistent improvements over conventional structure-based and sequence-only baselines, establishing state-of-the-art performance in peptide immunogenicity prediction.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 83