Computational prediction of TCR cross-reactivity: principles, challenges and translational opportunities

T-cell receptor (TCR) cross-reactivity, whereby a single TCR recognizes multiple peptide–MHC (pMHC) ligands, is essential for immune surveillance but also represents a major safety challenge for TCR-based therapeutics because of severe off-target toxicities. Rapid advances in immune-repertoire sequencing, structural biology, and immunopeptidomics have accelerated computational efforts to characterize cross-reactive recognition at scale. However, most existing methods are adapted from conventional TCR–pMHC specificity prediction and remain insufficiently aligned with the one-to-many, structurally plastic, and context-dependent nature of cross-reactivity. Here, we review the main data resources supporting TCR cross-reactivity modeling, including sequence- and structure-centric databases, and highlight persistent limitations, such as scarce explicitly annotated cross-reactive TCRs, strong biases toward viral epitopes and common HLA alleles, limited paired αβ-chain information, and a lack of rigorously validated negative examples. We then compare the major computational paradigms—sequence-based, structure-based, machine-learning, and multimodal approaches—with emphasis on their respective strengths and limitations in interpretability, generalization, and scalability. Finally, we discuss emerging translational applications and outline key priorities for the field, including dedicated datasets, rigorous benchmarking, and biologically grounded multimodal models to enable safer and more clinically actionable TCR-based immunotherapies.