Figure 1. Workflow, training data, and rescue screen for 1E6–ALW using energy-based model.

Training data: a curated set of 69 TCR–pMHC-I crystal structures is used to train the residue–residue energy matrix (γ in Eq. 1) and to define a decoy-normalized z-score scale for binding. Scoring peptide mutants: single–amino-acid substitutions are introduced into the ALW peptide and evaluated on distance-weighted contact maps with the trained model, yielding z-score distributions that distinguish weak vs. strong binders (top center). Template-based modeling: for each peptide mutant, a structural template is chosen by a composite sequence-similarity score (peptide, CDR3α, CDR3β), the complex is threaded and built in MODELLER, and energies are computed on the same decoy-normalized scale (bottom center). Example views compare the WT HLA-A*02-restricted 1E6–ALW structure (PDB: 3UTT) with a designed complex, highlighting interaction loss and potential new contacts after peptide mutation (top right). Compensatory screen: exhaustive single-site substitutions across 1E6 CDR3α/CDR3β are co-threaded with ALW peptide mutants classified as disrupted and subsequently re-scored; pairs that exceed the WT reference margin ($z\ge {\mu }_{z\mathrm{W}\mathrm{T}}-{\sigma }_{\mathrm{W}\mathrm{T}}$) are retained and ranked as rescue candidates (right).