A Pan-Specific GRU-Based Recurrent Neural Network for Predicting HLA-I-Binding Peptides

Human leukocyte antigens (HLAs) play a critical role in human-acquired immune responses by the recognition of non-self-peptides derived from exogenous bacteria, fungi, virus, and so forth. The accurate prediction of HLA-binding peptides is thus extremely useful for the mechanistic research of cell-mediated immunity and related epitope-based vaccine design. In this work, a simple pan-specific gated recurrent unit (GRU)-based recurrent neural network model was successfully proposed for predicting HLA-I-binding peptides. In comparison with the available six allele-specific, four pan-specific, and two ensemble-based prediction models, the GRU model achieves the highest area under the receiver operating characteristic curve (AUC) scores for 21 of 64 entries of the test benchmark datasets. Besides, the GRU model also achieves satisfactory performance on other 24 entries, of which the AUC scores differ by less than 0.1 from the highest scores. Overall, taking the advantages of the GRU network and auto-embedding techniques into account, the established pan-specific GRU model is more simple and direct and shows satisfactory prediction performance for HLA-I-binding peptides with varying lengths.