Computational modeling demonstrates accurate TCR-epitope immunogenicity prediction from sequencing data, highlighting an end-to-end framework for personalized cancer vaccine design.
Key Points
To develop an integrated, context-aware deep learning framework that directly predicts neoantigen immunogenicity and T-cell receptor binding from clinical sequencing data.
Integrated variant detection, HLA typing, pMHC affinity prediction, and TCR mining into a unified pipeline.
Implemented a hybrid transformer-convolutional neural network architecture with Bi-LSTM sequence features, gated physicochemical fusion, and explicit cross-attention.
Evaluated predictive accuracy using a strict TCR-split cross-validation strategy against six existing tools and validated on a patient-specific cancer cohort.
Achieved a mean AUROC of 0.7856 and AUPRC of 0.7932 under strict TCR-split conditions, outperforming six state-of-the-art predictors by 4% to 5%.
Demonstrated robustness against structural hard negatives and imbalanced datasets.
Successfully recovered 18 of 34 clinically validated high-affinity neoepitopes from a patient-specific cancer cohort.