Open-ComBind: Harnessing unlabeled data for improved binding pose prediction

Accurate determination of a ligand binding pose is crucial for in silico drug discovery. Molecular docking is a widely used method for the prediction of non-covalent binding of a protein and ligand system. However, most molecular docking pipelines consider only one ligand at a time, even though ligands may share binding interactions when bound to the same receptor. Open-ComBind, an open-source version of the ComBind molecular docking pipeline, addresses this issue by incorporating information from multiple ligands, including those without known bound structures, to improve pose selection. Poses are simultaneously predicted for a docking ligand of interest as well as a set of helper ligands. First, we generate distributions of feature similarities between ligand pose pairs, comparing near-native poses with all sampled docked poses. These distributions capture the likelihood of observing similar features, such as hydrogen bonds or hydrophobic contacts, in different pose configurations. We then combine these similarity distributions with a per-ligand docking score to improve overall pose selection by 8.5% and 4.5% for high-affinity and congeneric series helper ligands, respectively. Open-ComBind reduces the average RMSD of ligands in our benchmark dataset by 8.2%. We investigate the significance of the inter-molecular features as well as the number of helper ligands to achieving accurate pose selection. We provide Open-ComBind as an easy-to-use command line and Python API to increase pose prediction performance at www.github.com/drewnutt/open-combind.