nuTCRacker: Predicting the recognition of HLA-I-peptide complexes by αβTCRs for unseen peptides.

The ability to predict which antigenic peptide(s) the αβTCR of a given CD8+ T-cell clone can recognise would represent a quantum leap in the understanding of T-cell repertoire selection and development of targeted cell-mediated immunotherapies. Current methods fail to make accurate predictions for antigenic peptides not present in the training dataset. Here, we propose a novel deep learning method called nuTCRacker that makes accurate predictions for a subset of unseen peptides, with an AUC > 0.7 for around a third of peptides evaluated using a large dataset compiled from curated public resources. An additional evaluation was undertaken using a small cellula-validated dataset of αβTCR peptides associated with cancer. Our analysis suggests that it is possible to make useful predictions for an unseen peptide provided the training dataset contains: many samples with the same HLA class I molecule as that bound to the peptide; at least one peptide that is similar to the target peptide; and a small number of αβTCRs that are similar to those bound to the unseen peptide of interest.