Fig. 3. Prediction performance on weak binders generated by mismatching peptides with TCRs.

(A) Receiver operating characteristic (ROC) curves for RACER-m classification performance on differentiating weak binders generated by mismatching peptides from NLV, TAX, FLU, and 1E6 clusters with MART-1 TCRs from MART-1 strong binders with the same set of TCRs. (B) ROC curves for RACER-m classification performance on distinguishing MART-1 strong binders from mismatched weak binders generated by pairing MART-1–specific peptides with TCRs from NLV, TAX, FLU, and 1E6 clusters. (C) When TAX A6 TCR is paired with MART-1 peptide ELAGIGILTV, the z score of the mismatched TCR-pMHC pair (triangle) resembles the values from the strong binders (violin shape) formed by the same peptide and TCR RD1-MART1^High and its point mutants, which was engineered from A6. In the reverse scenario, TCR RD1-MART1^High shows lower z score (cross) than TAX strong binders (violin shape) when paired with TAX specific peptide LLFVYPVYV.