PMGen: From Peptide-MHC Prediction to Neoantigen Generation

Major histocompatibility complexes (MHCs) present peptides to T cells, playing a central role in adaptive immunity against cancer and infections. Accurate structural modeling of peptide-MHC (pMHC) complexes is essential for developing personalized immunotherapies. Existing tools often focus on either MHC class I or II, support limited peptide lengths, and address isolated tasks such as pMHC binding prediction or structure modeling. We introduce Peptide MHC Generator (PMGen), an integrated framework for predicting, modeling, and generating peptides across both MHC classes and a broad range of peptide lengths. PMGen enhances AlphaFold predictions by incorporating anchor residue information into structural module. PMGen achieves high structural accuracy (mean peptide core RMSD: 0.54 A for MHC-I, 0.33 A for MHC-II), improving over the state of the art for both MHC classes. PMGen can also generates diverse, high-affinity peptides with minimal structural deviation and accurately models the impact of single-point mutations in benchmark neoantigen-wild-type pairs. PMGen is a versatile tool for advancing personalized immunotherapy and is freely available at https://github.com/soedinglab/PMGen.