Abstract
T-cell cross-reactivity occurs when distinct peptide-MHC complexes are recognized by the same T-cell population, a phenomenon with important implications for vaccine design, cancer immunotherapy, and off-target immune responses. Since experimental structural determination of peptide-MHC complexes remains costly and time-consuming, computational modeling offers a practical alternative for investigating candidate cross-reactive epitopes. However, pMHC modeling methods are usually evaluated by structural similarity to crystallographic references rather than by their ability to reproduce known immunological cross-reactivity patterns. In this study, we compared DockTope, APE-GEN 2.0, AlphaFold3, and PANDORA according to their capacity to reproduce experimentally described cross-reactivity webs among HLA-A*02:01-restricted nonameric epitopes. Modeled pMHC structures were visualized in GRASP 2.0, and RGB values representing electrostatic potential in TCR-contacting regions were extracted to generate pairwise electrostatic-distance matrices. Depending on dataset structure, method performance was evaluated using ROC-AUC, relative compactness scores, or normalized discounted cumulative gain. Across the majority of the article-derived epitope sets, DockTope showed the strongest overall agreement with expected cross-reactivity behavior, achieving the highest composite score and best mean rank. APE-GEN 2.0 showed intermediate and dataset-dependent performance, while AlphaFold3 performed strongly in positive-only compactness datasets but less consistently when discrimination from non-reactive comparators was required. PANDORA showed the weakest overall agreement under this workflow. These findings suggest that pMHC modeling tools differ substantially in their ability to preserve TCR-facing electrostatic relationships relevant to cross-reactive recognition. Benchmarking pMHC models against immunological behavior, rather than structural accuracy alone, may improve computational prioritization of cross-reactive epitopes.