Enhanced prediction of antigen and antibody binding interface using ESM-2 and Bi-LSTM

The binding interface between antigens and antibodies is pivotal in humoral immune responses and provides crucial effective defense against pathogens and exogenous threats. Existing predictive computational methodologies, including structure-based and sequence-based approaches, offer valuable insights but face challenges such as unknown antigen structures and reliance on manually curated features. Most current methods primarily predict antigen epitope, often neglecting the specific molecular epitope-paratope interactions essential for immune efficacy. In this study, we introduce a novel approach EPP (Epitope-Paratope Predictor), using the ESM-2 protein language model as a feature encoder and a Bi-LSTM network to predict epitope-paratope interactions. Our method processes antigen and antibody sequences as inputs, leveraging a novel dataset strategy and encoding protein representations to enhance prediction accuracy. The results demonstrate a significant improvement in prediction accuracy compared to existing methods, highlighting the importance of protein feature encoder and temporal dependencies within sequences. The model's performance in different antigen clusters is analyzed, while those predictions are compared with that from AlphaFold3 and Dock method. Our method validation shows superior performance in recognizing distinctive epitopes of the same antigen when bound to different antibodies. This approach offers a new strategy for an in-depth understanding of antigen-antibody interactions, essential for an array of pioneer projects, such as structure-guided design and affinity maturation for precision antibodies targeting a given epitope.