Generative AI-enabled neoantigen vaccine engineering: From tumor antigen discovery to personalized construct design and translational validation.

Abstract

Neoantigen vaccines have rekindled interest in therapeutic cancer vaccination, yet their clinical efficacy remains constrained by imperfect antigen prioritization, incomplete modeling of immunogenicity, tumor heterogeneity, and immune evasion mechanisms. Current computational pipelines are dominated by discriminative models that rank pre-existing mutant peptides based on features related to HLA binding and antigen presentation. Although these approaches have improved candidate prioritization, their ability to optimize antigen selection and vaccine constructs across multiple determinants, including presentation, recognition potential, and translational feasibility, remains limited. Generative artificial intelligence offers a complementary, design-oriented framework that can explore and iteratively optimize biological sequence space under explicit constraints, rather than merely scoring predefined candidates. In this review, we discuss how generative AI may broaden the computational scope of tumor-anchored antigen discovery, support tumor-context-integrated optimization of neoantigen candidates, facilitate the engineering of multi-epitope and mRNA vaccine constructs, and incorporate tumor-specific constraints such as antigen-presentation defects, clonal architecture, and the state of the tumor immune microenvironment. Within this framework, generative models are considered components of an AI-assisted integrated workflow rather than substitutes for tumor-derived evidence, established prediction tools, or experimental validation. We further discuss the role of immunopeptidomics-guided calibration and iterative validation in improving biological realism and translational relevance. A longer-term frontier is immunopeptidomics-anchored synthetic immunogen design, in which validated tumor-presented immunogenic peptides may serve as templates for designing neoepitope mimetics or heteroclitic peptide analogs with improved HLA compatibility, pHLA stability and T-cell priming capacity. Finally, we examine current bottlenecks, including limited functionally validated immunogenicity datasets, uncertain generalizability, experimental validation burden, and the emerging regulatory demand for interpretability and traceability. At present, generative AI should be viewed as a promising design-enabling biotechnology platform whose clinical value remains to be established through prospective comparison with standard neoantigen pipelines.