Multi-strategy embedded framework for neoantigen vaccine maturation

Neoantigen-based vaccines enable personalized cancer immunotherapy, but most generative design methods rely on sequence information and overlook the three-dimensional dynamics of peptide-HLA (pHLA) binding. We present NEOM, a generative neoantigen maturation framework that integrates adaptive Markov chain Monte Carlo sampling with explicit pHLA structural modeling to evolve a single input peptide into diverse candidates with enhanced HLA class I binding. NEOM comprises five modules: “policy”, “structure”, “evaluation”, “selection” and “filter” that enable precise, interpretable, and customizable optimization. Applied to clinically derived and synthetic peptides, NEOM generated more high-quality candidates than baseline models. In silico filtering identified numerous peptides with strong predicted binding. Free energy perturbation (FEP) analysis further narrowed 38 candidates to six peptides with improved binding affinities. Finally, MHC tetramer exchange assays and flow cytometry validated five peptides as promising candidates. By unifying sequence exploration with structure-guided evaluation in an efficient pipeline, NEOM provides a streamlined platform for designing personalized peptide vaccines.