Fig. 5: Immunogenicity benchmarking and analysis of melanoma neoantigens and SARS-CoV-2 epitopes.
From: Self-iterative multiple-instance learning enables the prediction of CD4+ T cell immunogenic epitopes
a, Receiver operating characteristic curves of ImmuScope-IM and other methods on the immunogenicity benchmark. b, AUCs of ImmuScope-IM and other methods on the immunogenicity benchmark. The P values were calculated by the two-sided Wilcoxon signed-rank test to compare ImmuScope-IM with existing methods (HLAIImaster, P = 2.2 × 10−7; TLimmuno2, P = 1.4 × 10−7, n = 62). Box centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; points, data points; ****P < 0.0001. c, Pairwise comparison of AUCs between HLAIImaster and ImmuScope across different MHC-II alleles. d, Predictive analysis of melanoma neoantigen presentation based on ImmuScope-EL. e, Structural conformation of EDIL3290-304 epitopes bound to HLA-DPA1*01:03/DPB1*02:01 on the mutation predicted by AlphaFold3 (average predicted local distance difference test = 92.5; interface-predicted template modelling score = 0.92). The mutated residue is highlighted in red. The dashed lines indicate the hydrogen bonds, and interaction sites within 4 Å are displayed in dark blue. f, SHAP interpretation of p.P298F impact on EDIL3290-304 presentation by HLA-DPA1*01:03/DPB1*02:01. g, Predictive analysis of the immunogenicity of melanoma neoantigens based on ImmuScope-IM. h, Bar plots of AUCs on the SARS-CoV-2 immunogenic epitope benchmark. The bars represent the mean AUCs by 1,000 bootstrap iterations, and the error bars indicate the 95% CIs. i, Predicted binding peptide motifs for HLA-DRB1*01:01 as determined by the antigen presentation (ImmuScope-EL) and immunogenicity (ImmuScope-IM) models.
