Predicting Antibody-Antigen Interactions with Structure-Aware LLMs: Insights from SARS-CoV-2 Variants

Predicting antibody-antigen interactions is a critical step in developing new therapeutics to defend against viral infections. However, measuring the extent of these interactions in vitro is costly and time-consuming. With the increased availability of experimental data, predictive methods using machine learning, particularly large language models (LLMs), have emerged as a powerful alternative to wet lab experiments. Here we focus on antibodies targeting SARS-CoV-2 variants, given the abundance of data on this highly contagious virus and the impact of COVID-19 on human life. The objective of this work is to predict the binding and the neutralizing properties of SARS-CoV-2 antibodies. While there are many studies on predicting binding, to the best of our knowledge, we are the first to address the problem of predicting the neutralizing properties of SARS-CoV-2 antibodies. Here we propose a new classifier that combines LLMs with structural information. Extensive experimental results show our method (i) achieves high prediction accuracy (especially for closely related antigen variants) and (ii) outperforms other classifiers in the literature on the prediction of antibody-antigen binding.