Supervised Consensus Scoring Methods for Structure-Based Drug Design

Molecular docking is widely used to discovery novel ligands in structure-based drug design (SBDD) [1]. Over the past 15 years, various docking programs have been developed. Generally, flexible docking programs, such as DOCK, AutoDock, FlexX, GOLD and Glide, have the ability to predict protein-ligand complex structures with reasonable accuracy and speed. These docking programs are aimed at predicting the binding mode of a ligand and discriminating between active and inactive compounds. They involve two computational steps: docking and scoring. In docking step, many ligand conformations, called decoys, are generated. Next, a scoring function is used to evaluate the protein-ligand affinity. The final docked conformations are selected according to their scores. Accordingly, scoring functions play important roles in SBDD [2, 3]. Although considerable efforts have been devoted to design scoring functions to describe protein-ligand interactions, it is pointed out that the major weakness of docking programs lies in scoring functions [2]. There are three groups in scoring functions: force field based methods, empirical scoring functions and knowledge-based potentials. Each scoring functions has specific advantages and disadvantages and there is no standard method. Recently, it is reported that consensus scoring method can improve the performance by compensating for the deficiencies of each scoring functions [2, 3]. However, conventional consensus scoring, such as rank-by-rank, rank-by-number, average rank, and linear combination of multiple scoring functions, does not address the problem of designing accurate scoring function essentially and improve the performance enough. In order to cope with the problem, we propose the supervised consensus scoring method (SCS) to incorporate free energy landscape between the protein and the ligand via the decoy structures using supervised learning. We show that SCS outperforms conventional consensus scoring in predicting binding mode and discuss why SCS improves the performance.