MORE | Spring 2023

Few Shot Learning for TCR-Epitope Binding Affinity Prediction

Project aims to develop a state-of-the-art TCR-epitope binding affinity prediction model for out-of-sample epitopes. The plan is to fine-tune a pre-trained amino-acid embedding model to predict the similarity between two TCRs, which can indicate their likelihood to bind to the same epitope, using L1 and MSE loss. To generate the training set, TCR-epitope pairs will be collected and provide a support set of TCRs binding to an unseen epitope. The research will optimize model performance by experimenting with different numbers of dense layers on top of each embedding network. Initial experiments using Siamese networks with positive and negative pairs at a 1:1 ratio achieved a baseline accuracy of 62%. Researchers will focus on improving model performance using techniques like early stop, batch normalization, and powerful neural network structures such as LSTM and Transformers. This research has the potential to impact clinical applications such as immunotherapies and vaccine development.