EpiTox: A Multi-Modular Framework for Population-Aware Off-Target Prediction Highlighting MAGEA3 Cross-Reactivity

Targeting intracellular tumor antigens presented by peptide human leukocyte antigen (pHLA) complexes offers a promising immunotherapeutic strategy for patients with limited treatment options. However, development of pHLA-targeted drugs such as T-cell receptor (TCR) mimic antibodies (TCRm) and TCR-based therapeutics remains challenging due to severe offtumor and off-target toxicities arising from incomplete pHLA off-target profiling. The fatal neurological toxicities observed in the MAGEA3 TCR-T cell trial, potentially linked to unanticipated cross-reactivity with EPS8L2- and MAGEA12-derived peptides, underscore the urgent need for more comprehensive safety assessment. While current preclinical de-risking methods utilize sequence similarity searches, they lack the layered integration of genetic context, HLA binding profiles, and structured risk assessment needed to comprehensively evaluate peptide cross-reactivity. To address these limitations, we developed EpiTox, a computational multi-modular platform that systematically identifies and evaluates pHLA off-targets. EpiTox integrates proteome-wide sequence similarity, genetic context, HLA binding profiles, and layered risk assessment to provide a holistic evaluation of potential cross-reactive peptides. Using MAGEA3 as a model, EpiTox identified known toxic off-targets and predicted novel epitopes. Notably, an anti-MAGEA3 therapeutic TCRm antibody bound several newly predicted off-targets, including three SNP-derived peptides. By enabling comprehensive and predictive preclinical safety screening, Epi-Tox supports safer TCRm drug development and may help prevent future clinical failures.