Benchmarking TCR-pMHC Recognition Geometry

T-cell receptor (TCR) recognition of peptide-MHC (pMHC) complexes drives adaptive immunity, but experimental structure determination is slow. Accurate prediction of these complexes would accelerate TCR engineering and vaccine design. Current benchmarks score predicted complexes with general distance measures such as RMSD and DockQ. On 87 complexes released after the AlphaFold3 training cutoff, across six architectures, these measures proved largely redundant, sharing Spearman correlations of 0.73 to 0.96. However, the angles describing the receptor's orientation over its target sit outside that group and are only weakly related to each other. Structural templates, which let a model copy solved structures, improve distance measures by 8 to 13\% but the angles by only 2 to 4\%. We argue that for AI to meaningfully advance structural immunology, benchmarks must move beyond global similarity. Future models should be evaluated on the precise interaction geometries and specific functional interfaces that dictate immune recognition.