Being able to identify regions within or around proteins, to which ligands can potentially bind, is an essential step to develop new drugs. Binding site identification methods can now profit from the availability of large amounts of 3D structures in protein structure databases or from AlphaFold predictions. Current binding site identification methods rely on geometric deep learning, which takes geometric invariances and equivariances into account. Such methods turned out to be very beneficial for physics-related tasks like binding energy or motion trajectory prediction. However, their performance at binding site identification is still limited, which might be due to limited expressivity or oversquashing effects of E(n)-Equivariant Graph Neural Networks (EGNNs). Here, we extend EGNNs by adding virtual nodes and applying an extended message passing scheme. The virtual nodes in these graphs both improve the predictive performance and can also learn to represent binding sites. In our experiments, we show that VN-EGNN sets a new state of the art at binding site identification on three common benchmarks, COACH420, HOLO4K, and PDBbind2020.
Forschungs-