Fig. 1: Overview of deepAntigen.
From: Identifying T cell antigen at the atomic level with graph convolutional network
a–c deepAntigen workflow. a, deepAntigen takes the residue sequences of antigen, TCR β chain complementary determine region 3 (CDR3) and HLA as input. The input residue sequences are transformed into topological graphs, in which each node corresponds to an atom and each edge corresponds to chemical bond, through the atomic graph representation module (Fig. 1b). Next, crucial atoms for the interaction process can be identified through GCN layer and Scoring layer. Finally, multi-head attention module extends respective top-k atoms to form a k×k interaction map. deepAntigen_Seq aggregates interaction features of pairwise atoms from antigen and HLA/TCR to predict the interaction probability through a multilayer perceptron (MLP). b The generation process of atomic graph representation for a given residue sequence. c Fine-tuning deepAntigen_Atom using structural data after pretraining deepAntigen_Seq on the sequence-level binding data. Parameters of GCN in light blue region were frozen, while those in light pink region were fine-tuned. deepAntigen_Atom directly predicts the contact probability of each atom pairs through an MLP according the interaction map. d General comparison between deepAntigen and other state-of-the-art methods on diverse independent test dataset. #, the number of unique objects. The schematic diagram of the TCR-antigen-HLA complex in (a) was created in BioRender. Que, J. (2025) https://BioRender.com/hbixhe0.
