Abstract
Accurate prediction of T cell receptor (TCR) reactivity is a long-standing goal in immunology. Here, we asked whether functional validation of TCR-peptide-major histocompatibility complex (MHC) (TCR-pMHC) pairs at scale alters performance estimates of TCR-pMHC reactivity prediction models. We developed a TCR rapid assembly platform (T-RAP) that allowed the generation of large and uniform TCR libraries. T-RAP enabled evaluation of TCR signaling or MHC multimer binding of thousands of previously reported TCR-pMHC pairs under standardized conditions. In this setting of systematic standardized evaluation, only ∼50% of these TCRs showed the previously reported TCR reactivity. Notably, AlphaFold3 structural predictions showed good performance in identifying reactive TCR-pMHC pairs within the set of TCRs that experimentally validated without any task-specific training. AlphaFold3 structural predictions could likewise be used to shortlist TCRs reactive to patient-specific cancer neoantigens. In silico prediction of TCR-pMHC reactivity thus has greater feasibility than previously assumed, with implications for basic research and clinical application.