An attention network for predicting T cell receptor-peptide binding can associate attention with interpretable protein structural properties

Understanding how a T cell receptor (TCR) recognizes its specific ligand peptide is crucial for gaining insight into biological functions and disease mechanisms. Despite its importance, experimentally determining TCR-peptide interactions is expensive and time-consuming. To address this challenge, computational methods have been proposed, but they are typically evaluated by internal retrospective validation only, and few have incorporated and tested an attention layer from language models into structural information. Therefore, in this study, we developed a machine learning model based on a modified version of the Transformer, a source-target-attention neural network, to predict TCR-peptide binding solely from the amino acid sequences of the TCRs complementarity-determining region (CDR) 3 and the peptide. This model achieved competitive performance on a benchmark dataset of TCR-peptide binding, as well as on a truly new external dataset. Additionally, by analyzing the results of binding predictions, we associated the neural network weights with protein structural properties. By classifying the residues into large and small attention groups, we identified statistically significant properties associated with the largely attended residues, such as hydrogen bonds within the CDR3. The dataset that we have created and our models ability to provide an interpretable prediction of TCR-peptide binding should increase our knowledge of molecular recognition and pave the way to designing new therapeutics.