Reengineering the antigen optimization process for superior neoantigen vaccine design

Identifying effective neoantigen sequences is essential for enhancing anti-tumor immunity. However, the vast sequence space (>10 ⁹ possible peptides) and limited accuracy of existing immunogenicity predictors hinder efficient vaccine design for patient-specific human leukocyte antigens (HLAs). We present AlphaVacc, a deep reinforcement learning framework that integrates Monte Carlo Tree Search with a Transformer-based network to optimize antigenic peptides. AlphaVacc outperforms previous generative models in binding-affinity prediction. Experimental validation of 12 AlphaVacc-generated variants of the BING-4 peptide confirmed that 11 showed increased HLA-A*02:01 binding and 7 elicited significant T cell responses. Further testing of 16 single-mutation peptides confirmed computational predictions for 15 candidates, exhibiting a remarkable success rate. AlphaVacc thus provides a powerful tool for designing neoantigen-based cancer vaccines and may accelerate personalized immunotherapies.