Evaluating The Global Performace of Protein-Ligand Docking Prediction

The global ligand-protein redocking problem remains unsolved. Identifying the specific ligand confirmations and target sites in a protein, ab initio, remains a substantial challenge. Nearly three decades on, several models have been developed to computationally evaluate suggested binding conformations of small molecules to a protein of interest. Approaches have mainly focused on two methods: knowledge-based scoring functions finding trends in datasets available from deposited experimental structures in the Protein Data Bank, and physics-based energy considerations based on electrostatic interactions, Van der Waal forces and stereochemical constraints. Modern approaches often also leverage machine learning to guide sampling and scoring such as AutoDock. Herein, we constructed and refined a dataset of high-resolution x-ray crystal structures on which exhaustive sampling of potential protein-ligand surfaces were formed. Evaluating current approaches revealed difficulties in identifying appropriate binding sites, driving the need for improved scoring functions to appropriately distinguish a true global minimum from distinct or false local minima. Proposals for scoring function improvement are also presented.