Fig. 6

Benchmark TRAP performance and use it to identify tumour antigens in glioblastoma. A, B ROC curve comparing the performance of TRAP with existing models, such as IEDB, iPred, NetTepi, PRIME, Repitope and DeepImmuno for self- (A) and pathogenic peptides (B). Due to the nature of these models, IEDB, iPred, NetTepi PRIME and Repitope were predicted in an HLA-agnostic manner (i.e. by peptide sequence only) (i), and DeepImmuno predicted in an HLA-restrictive manner (i.e. by peptide-HLA) using HLA-balanced data (ii). C ROC curve comparing the performance of TRAP and Repitope that have been trained using 1511 Non-Wuhan SARS-CoV-2 peptides by 10-fold cross-validations. D ROC curve comparing the performance of Non-Wuhan SARS-CoV-2 trained TRAP and Repitope models in predicting 66 Wuhan SARS-CoV-2 peptides. E–J Application of TRAP in identifying glioblastoma neoepitopes. E Confusion matrix of GBM cancer neoepitope prediction using self-antigen TRAP model in reference to T-cell assay readout. F ROC curve of TRAP performance on GBM dataset. G Distribution of MCDropout values for GBM peptides predicted to be non-immunogenic. H The proportion of GBM epitopes (Positive) and non-epitopes (Negatives) predicted Positive or Negative with high or low confidence based on MCDropout. The confidence has been determined by the self-antigen out-of-distribution linear regression classifier at threshold 0.76. I. ROC curve of TRAP performance taking confidence into account, in which peptides that were predicted Negative with low confidence were removed to be considered as the potential epitope candidates. J Confusion matrix of TRAP prediction after taking the confidence into account, in which peptides that were predicted Negative with low confidence were included as Positive