Graph neural networks (GNNs) for Protein-Ligand binding prediction

Drug discovery is a long process containing multiple complex pipelines that might take a decade to finish. With the rapid development of machine learning and deep learning, there is an increasing number of intermediate steps in drug discovery that can be solved by using artificial intelligence.

Protein-Ligand binding is one of the core problems in drug discovery, that aims to predict how a drug molecule interacts with a protein in a specific condition. Specifically, the task is to predict the 3D conformer of the molecule as well as its binding location on the surface of the protein.

In fact, both protein (Receptor) and drug molecule (Ligand) can be represented by graphs. For example, a drug molecule can be described as a graph of nodes as atoms and edges as atoms links between them; a protein can be considered as a graph of amino acids (residues) and edges are constructed based on the distance between them in 3D space. This allows Graph Neural Networks, a deep learning architecture, to be suitable for protein and molecule representations. Their representation can be used to predict how they interact with each other.