TCR-pMHC Binding Specificity Prediction From Structure Using Graph Neural Networks

The mapping of T-cell-receptors (TCRs) to their cognate peptides is crucial to improving cancer immunotherapy. Numerous computational methods and machine learning tools have been proposed to aid in the task. Yet, accurately constructing this map computationally remains a difficult problem. Most prior work has sought to predict TCR-peptide-MHC (TCR-pMHC) binding specificity by analyzing the amino acid sequences of the TCRs and peptides. However, recent advancements in crystallography, cryo-EM, and in silico protein modeling have provided researchers with the necessary data to analyze the 3D structures of TCRs, peptides, and MHCs. Current research suggests that information contained in the 3D structure of the TCRs and pMHCs can explain instances of TCR specificity that are not explained by sequence alone. As protein structure data continues to become more accurate and easier to obtain, structure-based methodologies for predicting TCR-pMHC binding will become increasingly important. We present STAG, a novel graph-based machine learning architecture for predicting TCR-pMHC binding specificity using 3D structure data. We show that STAG achieves comparable or better performance than existing methods while utilizing only spatial and physicochemical features from modeled protein structures.