Fig. 3

Optimise TRAP architecture. A Performance of pre- and post-padding strategies. ROC-AUC values of respective deep learning models with pre- vs. post-padding strategies for pathogenic (i) and self-antigens (ii). Each point represents the ROC-AUC value from one round of 10-fold cross-validations. B Example of n-grams found in both 9aa and 10aa peptides, coloured by the start position of the respective n-gram in the peptides. C Comparing amino acid embedding strategies. ROC-AUC values of a single dense classification prediction using different encoding strategies, including one-hot encoding (OHE), amino acid descriptors (aaDescriptors) and embeddings from protein transformer-based PLMs on pathogenic peptides (i) and self-antigens (ii). Each point represents the ROC-AUC value from one round of 10-fold cross-validations. D Comparing the performance of different deep learning architectures. ROC-AUC values comparing the performance of different machine learning and deep learning architectures using embeddings from 5 protein transformer-based PLMs for pathogenic (i) and self-peptides (ii). XGBOOST: an extreme gradient boosting. BiRNN: bidirectional recurrent neural network. BiLSTM: bidirectional Long short-term memory. 1DCNN: 1-dimensional convolutional neural network. 2DCNN: 2-dimensional CNN. Each point represents ROC-AUC value from one round of 10-fold cross-validations. E Schematic diagram of the model incorporating MHC binding and hydrophobicity. F Adding hallmarks of the immunogenicity. ROC curve comparing the performance of models after adding MHC binding rank score and/or hydrophobicity to peptide sequence-based 1D CNN model for pathogenic (i) and self-peptides (ii)