SwiftMHC: A High-Speed Attention Network for MHC-Bound Peptide Identification and 3D Modeling

Cancer immunotherapies show promise in eliminating tumors, but identifying tumor peptides that bind patient MHC proteins to trigger immune responses remains challenging. The vast peptide-MHC diversity makes experimental identification costly, emphasizing the need for computational predictions.

To address scalability, we introduce SwiftMHC, an attention network for identifying MHC-bound peptides and generating all-atom 3D structures simultaneously. SwiftMHC processes cases in 0.01 to 2.2 seconds on a single A100 GPU, overcoming the speed bottleneck of structure-based approaches. It rivals or surpasses state-of-the-art tools, including an AlphaFold-based approach, in binding affinity prediction and 3D modeling.

SwiftMHC showcases the advantages of task-specific AI trained on physics-derived synthetic data for speed and precision. Optimized for HLA-A*02:01 9-mers, it can be adapted to other MHC alleles. Its speed and accuracy may enable the community to identify peptides distinct from self-peptides at T-cell exposed surfaces, ensuring safer therapies, expanding immunotherapy targets, and accelerating development.