Figure 3. Enhanced training-set composition improves RACER-m predictions for the 1E6–ALW single–mutation landscape.

(a) Crystallographic reference structure of the 1E6 TCR–pMHC complex (PDB: 3UTT), with the ALW peptide shown in the HLA-A*02:01 binding groove and the central Gly4, Pro5, and Asp6 residues highlighted to delineate the dominant TCR-facing hot-spot region. (b) Predicted disruption heatmap (Δz=z_mut−z_WT)) when RACER-m is trained exclusively on the eight 1E6–ALW crystal structures (ALW). (c) Predictions from the model trained on the 69 ATLAS TCR–pMHC crystal structures augmented with AlphaFold3-generated strong-binding peptide–mutant complexes (ATLAS–AF). (d) Predictions from the combined antigen-specific and structure-augmented configuration, in which RACER-m is trained on the eight 1E6–ALW crystal structures together with AlphaFold3-derived peptide–mutant complexes (ALW–AF). (e) Global correlation between experimental TNF responses and RACER-m predictions computed using the ALW–AF model, summarizing the quantitative agreement achieved after antigen restriction and peptide-side structural augmentation. Across all training setups, the characteristic Gly4–Pro5–Asp6 hot spot is consistently recovered, while incorporation of AlphaFold-derived peptide variants enhances agreement at several flanking positions.