Fig. 5: The transfer-learning framework and validation performance of ImmuneApp-Neo for immunogenicity prediction.
From: ImmuneApp for HLA-I epitope prediction and immunopeptidome analysis
a The development of ImmuneApp-Neo with a transfer-learning strategy for immunogenicity prediction. ImmuneApp-Neo was trained on new curated immunogenicity data by retraining the last three fully connected layers of the mixed prediction model ImmuneApp-MA, which outputs the neoepitope immunogenicity score. b PPVn was calculated for all benchmark methods, including ImmuneApp-Neo, ImmuneApp-MA, ImmuneApp-EL, PRIME 1.0&2.0, NetMHCpan-4.1, MHCflurry 2.0, MixMHCpred 2.1&2.2, HLAthena, MHCnuggets-2.4, and TransPHLA, as the fraction of neoepitopes that are immunogenic within the top n predictions (value of n ranges from 1 to 349). c Mean PPVn with a 95% confidence interval (CI) for all methods are shown. It summarizes the PPVn curves for all valid choices of n (n = 349). d, e Mean AUROC (d) and AUPRC (e) were calculated for all benchmark methods for neoepitope immunogenicity prediction. Bars represent means and error bars are 95% CIs. Source data are provided as a Source Data file.
