Fig. 3: Performance of retrained TCR–epitope prediction models.
From: Assessment of computational methods in predicting TCR–epitope binding recognition
a, Epitope (and corresponding TCR) counts across antigen groups from different datasets. YFV, yellow fever virus; HBV, hepatitis B virus; HHV, human herpesvirus; HIV, human immunodeficiency virus. b, Performance of retrained CDR3β-only models on seen-epitope test. c, AUPRC comparison of models (n = 24) from b across AS/PS/HS negatives. d, Performance of retrained CDR3β-only models on seen-epitope independent test. e, AUPRC comparison of models (n = 24) from d across AS/PS/HS negatives. f, Performance of retrained CDR3β-only models in unseen-epitope test. g, AUPRC comparison of models (n = 21) from f across AS/PS/HS negatives. h,i, Performance on seen-epitope CDR3β + others test data of retrained CDR3β + others models (h) and CDR3β-only models (i). j, AUPRC comparison between CDR3β-only (n = 24, i) and CDR3β + others (n = 7, h) models. k,l, Performance on seen-epitope CDR3β + others independent test data of retrained CDR3β + others models (k) and CDR3β-only models (l). m, AUPRC comparison between CDR3β-only (n = 24, l) and CDR3β + others (n = 7, k) models. n,o, Performance on unseen-epitope CDR3β + others independent test data of retrained CDR3β + others models (n) and CDR3β-only models (o). p, AUPRC comparison between CDR3β-only (n = 21, o) and CDR3β + others (n = 3, n) models. Dot plots (b,d,h,i,k l) show per-antigen AUPRC, with adjacent heatmaps showing overall AUPRC, ordered by AS-based AUPRC. Heatmaps (f,n,o) show epitope-level AUPRC, categorized by antigen group and ordered by AS-based AUPRC. Colored dots (c,e,g) represent individual model AUPRC, black dots indicate the mean and error bars represent the mean ± s.d. Box plots (j,m,p) show mean (center line), first and third quartiles (box) and minimum and maximum values within 1.5 × interquartile range (whiskers). All P values were from two-sided Wilcoxon rank-sum tests.
