Enhancing Molecular Docking Performance with a GNN-Based Algorithm Selection Model
Molecular docking is a critical process in drug discovery and design with various algorithms to predict how small molecules, or ligands, interact with proteins. The present study introduces a novel use of using Graph Neural Network (GNN) to build effective Algorithm Selection (AS) models for automatically choosing suitable molecular docking algorithms concerning specific scenarios. This method addresses the limitations highlighted by the no free lunch theorem, which states that no single algorithm can outperform all the others universally. We build graphs for molecules by their 3D structures with atoms to be node and bonds to be edges, and set the attributes of nodes and bonds with the atomic numbers and bond types. By using an existing, complex GNN network with a combination of GCN, GAT and GIN, we successfully train an Algorithm Selection (AS) model to predict the (near-)optimal docking algorithm for different docking situations. The trained model shows improvement for accuracy in terms of RMSD. Also, it reaches ∼25% accuracy for predicting the best algorithm and 50% for predicitng near best, within a 1.5 Å margin.
