SE3Bind: SE(3)-equivariant model for antibody-antigen binding affinity prediction

Abstract

Predicting antibody-antigen binding affinity is critical for therapeutic development, but machine learning-based approaches to the problem are typically hampered by the small amount of available structural and affinity data. We introduce SE3Bind, an SE(3)-equivariant architecture trained on two related tasks: re-docking of an antibody structure to its matching antigen structure, and antibody-antigen binding free energy prediction. Both tasks encourage the model to learn an energy function formulated in terms of scalar and vector fields associated with each protein. Under a stringent training/validation split of the data based on antigen sequence similarity, SE3Bind demonstrates an ability to generalize to out-of-distribution examples and achieves a performance comparable to that of existing models evaluated under similar conditions.