Fig. 5. RACER-m’s performance on differentiating strong binders from point-mutant weak binders.

(A) Distribution of z scores from strong binders of 1E6 TCR and weak binders from point mutagenic screen. (B) ROC curve for RACER-m classification performance using the strong and point-mutant weak binders for 1E6 TCR. (C) Comparison of RACER-m and NetTCR-2.0 in classification of strong and point-mutant weak binders from ATLAS dataset. Here, RACER-m predictions used the known crystal structure selected by the sequence similarity calculation results as a representative template for threading each test case.