Transformer-Genetic Algorithm Co-Optimization for Neoantigen Prediction in Cancer Immunotherapy

Personalized neoantigen vaccine development requires accurate predictions of binding affinity, immunogenicity, and peptide stability across diverse HLA profiles. This paper presents a novel multi-task learning transformer architecture augmented with evolutionary optimization for comprehensive neoantigen prioritization. The proposed framework integrates enhanced positional encoding and biological attention mechanism targeting peptide anchor residues. We trained and validated our model on 30,450 neoantigen candidates from 50 melanoma patients, incorporating multi-omics data including somatic mutations, gene expression, copy number variations, and methylation profiles. The transformer architecture simultaneously predicts three critical neoantigen properties while a genetic algorithm optimizes vaccine candidate selection for population coverage and clinical outcomes. Experimental results demonstrate state-of-the-art performance with ROC AUC scores of 0.983, 0.980, and 0.962 for binding affinity, immunogenicity, and stability prediction, respectively, achieving an overall mean AUC of 0.975. Clinical validation shows significant survival correlation (r = 0.316, p = 0.028) and optimized vaccine candidates achieve $90$ HLA population coverage with genetic algorithm scores reaching 0.940. The proposed framework establishes a new benchmark for AI-driven neoantigen vaccine design with immediate translational applications in precision oncology.