Deep learning methods for predicting binding affinity of antibody-antigen: a comparative study of embedding techniques and model architectures

Abstract

Antibody-antigen interactions are critical to immune responses and play a pivotal role in therapeutic development, diagnostics, and vaccine design. Accurate prediction of binding affinity between antibodies and antigens is essential for accelerating antibody discovery and advancing our understanding of immune system dynamics. This study explores deep learning methods for predicting antibody-antigen binding affinity, focusing on employing various deep learning architectures and evaluating different antibody-antigen representation techniques by leveraging transfer learning with large language models. A comprehensive comparative analysis is conducted to assess the performance of various embedding techniques and deep learning architectures, including transformer-based models, in modelling the binding affinity between antibody-antigen pairs. The research aims to establish fundamental benchmarks through unified datasets and standardized evaluation metrics to provide a solid foundation for future research in this area. Results from the proposed models demonstrate superior performance, surpassing existing state-of-the-art approaches in binding affinity prediction. By offering valuable insights into the effective application of deep learning to this complex task, this work lays the groundwork for future innovations in immunology, therapeutic design, and vaccine development.