Predicting the interactions between T cell receptors (TCR) and peptide-MHC complexes (pMHC) can enhance understanding of the immune response against cancer and enable the development of novel immunotherapies. Several computational methods have been proposed to predict TCR-pMHC interactions but have so far failed to show generalizability to epitopes not seen during training.
Here, Regina Barzilay and Tyler Jacks aim to develop novel machine learning methods for predicting the binding of any TCR-pMHC pairing, regardless of an epitope’s presence in the training data. They will study the effectiveness of general protein binding models towards predicting TCR-pMHC pairing and the benefits of in-domain TCR-pMHC training data. They aim to demonstrate the utility of the model in an experimental workflow by accelerating the discovery of tumor reactive T cells. This is typically a laborious process that involves staining T cells, isolation of the population of interest, and subsequent single cell capture and sequencing of the population. Instead, they propose to leverage the model to identify the reactive TCR’s by using single cell RNA sequencing of the processed tumor alone, thus skipping staining and sorting altogether. These models will be made publicly available and open-sourced to the research community.