Antibody-antigen binding affinity prediction through the use of geometric deep learning : a framework for binding affinity prediction with graph neural networks

Abstract

Antibodies are an integral part of our body’s immune system due to their ability to trigger immune responses or render exogenous proteins ineffective. The mechanism of binding of antibodies has been studied thoroughly in order to select promising antibodies or even design new ones. Even though a variety of knowledge-primed methods (e.g. force fields) have been adopted for affinity prediction, the challenge of accurately predicting binding affinity between antibodies and antigens remains. In this thesis, we propose a purely data-driven approach to predict antibody-antigen binding affinity using geometric deep learning methods. Our research aims to evaluate the effectiveness of graph neural networks (GNN) and compare it to a state-of-the-art force field-based method. In order to achieve this, available crystallized antibody-antigen complexes are converted to graph structures and used to train GNN-based learning methods. In addition, given the scarce availability of training data for the antibody-antigen affinity prediction problem, we explore the potential of transfer learning to improve predictive performance (e.g. through the inclusion of general PPI complexes). The implementation of our designed GNN provides a PyTorch framework for generic binding affinity prediction using graph-like structures. The trained GNN outperforms the force field baseline on a diverse set of antibody-antigen complexes by showing robust results across low- and high-quality structures. The implemented transfer-learning techniques did not result in significant performance improvements, although numerous such techniques remain to be explored. The effectiveness of GNNs for affinity prediction and their end-to-end differentiability highlights their potential for studying the mechanisms of antibody binding. These properties further allow applications in the improvement and de novo design of experimental and therapeutic antibodies.