Mutational binding assessment for TCR engineering on experimental and AlphaFold 3 scaffolds

Machine learning-based binding affinity prediction can support T-cell receptor (TCR) engineering, but its reliability remains uncertain when experimental structures of TCRs bound to peptide–major histocompatibility complexes (pMHCs) are unavailable and predicted structures must be used instead. Here, we examined how AlphaFold 3 (AF3) scaffold choice and conformational sampling affect Δ Δ G -based prioritization of TCR–pMHC complementarity-determining region (CDR) mutations. Bound–unbound structural comparisons showed that pMHC binding is associated mainly with local CDR remodeling, especially in CDR3 loops, whereas TCR framework regions remain comparatively stable. AF3 generally recovered bound-like TCR conformations, but downstream Δ Δ G prediction performance depended strongly on both scaffold source and model architecture. These scaffold-dependent differences indicate that experimental structures remain the most reliable input when available, although such structures are rarely accessible in prospective TCR design settings. Among AF3-based inputs, Prompt-DDG and RDE-PPI were relatively robust to wild-type predicted scaffolds, whereas BA-DDG and H3-DDG benefited more from mutant-specific AF3 scaffolds evaluated by reverse-mutation scoring. Across four templates, AF3-mutant scaffolds gave the highest mean top-10 recovery among the AF3-based inputs for all four predictors. These results support using experimental structures when available and suggest using AF3-mutant scaffolds for recovering top-ranked CDR candidates when only predicted structures are available.