Co-Supervised Pre-Training of Pocket and Ligand

Abstract

Can we inject the pocket-ligand complementarity knowledge into the pre-trained model and jointly learn their chemical space? Pre-training molecules and proteins have attracted considerable attention in recent years, while most of these approaches focus on learning one of the chemical spaces and lack the consideration of their complementarity. We propose a co-supervised pre-training (CoSP) framework to learn 3D pocket and ligand representations simultaneously. We use a gated geometric message passing layer to model 3D pockets and ligands, where each node’s chemical features, geometric position, and direction are considered. To learn meaningful biological embeddings, we inject the pocket-ligand complementarity into the pre-training model via ChemInfoNCE loss, cooperating with a chemical similarity-enhanced negative sampling strategy to improve the representation learning. Through extensive experiments, we conclude that CoSP can achieve competitive results in pocket matching, molecule property prediction, and virtual screening.