Figure 4. Disentangling TCR-facing and MHC-facing effects in the 1E6–ALW mutational scan.

(a) AlphaFold3 model of the WT ALW peptide bound to HLA-A*02:01, showing MHC-facing versus TCR-facing orientations. (b–c) Representative Pro7 substitutions (Pro7 →Leu in panel b; Pro7 →Ser in panel c) illustrating how mutations reshape peptide accommodation within the MHC groove while leaving the TCR-facing surface essentially unchanged. (d) RACER-m predictions after ALW-restricted training enriched with 18 AlphaFold3-derived peptide–mutant structures; mutational effects at MHC-facing residues (Leu2, Pro7, Ala9, Ala10) are interpreted using NetMHCpan 4.1, as these positions predominantly modulate peptide–MHC stability rather than TCR contact. (e–h) Position-specific correlations between experimental TNF responses and NetMHCpan 4.1 predictions for Leu2 (e), Pro7 (f), Ala9 (g), and Ala10 (h), confirming that experimental sensitivity at these sites arises primarily from altered peptide–MHC binding. (i) Integrated correlation combining RACER-m predictions for TCR-facing positions with NetMHCpan 4.1 predictions for MHC-facing positions, demonstrating substantially improved overall agreement once peptide–MHC energetics are incorporated.