Fig. 1: The framework, model architecture, and utility of ImmuneApp for HLA-I antigen prediction and immunopeptidome analysis.

a The deep learning framework for prediction of human leukocyte antigen (HLA) class I antigen presentation. ImmuneApp took encoded matrixes of the peptides sourced from mass spectrometry-eluted HLA ligands or peptides with binding affinity (BA) measurements, and pseudo-sequences of HLA alleles on the BLOSUM50 substitution matrix. Then, the input matrixes were fed into a convolutional neural network (CNN) and long short-term memory (LSTM) with attention modules for training. Features obtained from different parts of the neural network were retrieved from various layers, and then they were combined. A probability of the likelihood of antigen presentation in the setting of certain HLA class I alleles is produced by the output layer that implements a sigmoid nonlinear transformation. b ImmuneApp provides various presentation prediction capabilities, including eluted ligand (EL) likelihood estimate, in vitro BA measurements, and immunogenicity prediction. c ImmuneApp provides one-stop analysis, statistical reports, and visualization for immunopeptidomics data, such as quality control, binding annotations, HLA assignment, motif discovery and decomposition, and antigen presentation prediction on a sample-specific basis.