Fig. 2: ImmuScope improves prediction of CD4+ T cell epitope and antigen presentation.
From: Self-iterative multiple-instance learning enables the prediction of CD4+ T cell immunogenic epitopes
a, Box plot of AUCs on the CD4+ epitope benchmark. The P values were calculated using a two-sided Wilcoxon signed-rank test to compare ImmuScope with existing methods (NetMHCIIpan-4.3, P = 1.9 × 10−19; NetMHCIIpan-4.2, P = 2.2 × 10−19; MixMHC2pred-2.0, P = 3.3 × 10−22; n = 824). Box centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; dashed line, random; ****P < 0.0001. b, Comparisons of ImmuScope with MixMHC2pred-2.0 (left), NetMHCIIpan-4.2 (middle) and NetMHCIIpan-4.3 (right) in terms of AUC. The diagonal line indicates identical AUC values. The percentage in the lower-right corner shows the proportion of alleles for which ImmuScope outperforms the comparator, whereas marker size denotes the number of data points for each allele. Kernel density estimation curves along the top and right margins depict the marginal distributions of AUCs for each method, with the marked numbers indicating the AUC values at the highest density. c, Performance comparison across the SA, SA + MA-MIL and ImmuScope-EL models: AUPR (left) and PPV (right). The P values were calculated by the two-sided Wilcoxon signed-rank test (n = 58): SA versus SA + MA-MIL (AUPR, P = 5.3 × 10−11; PPV, P = 4.3 × 10−11) and SA + MA-MIL versus ImmuScope-EL (AUPR, P = 3.5 × 10−11; PPV, P = 3.5 × 10−11). Box centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; ****P < 0.0001. d, AUPR and PPV of ImmuScope-EL by stratifying different HLA loci. The bars represent the mean by 1,000 bootstrap iterations, and the error bars indicate the 95% CIs. e, UMAP visualization of instance embeddings from the SA and ImmuScope-EL models on the test set. In c and d, each data point represents the performance of the corresponding MHC allele.
