Universal Optimizations of Scoring Functions for Virtual Screening

Abstract

Structure-based virtual screening is gaining popularity in drug discovery. A number of molecular docking programs and scoring functions have been developed in the community, but they had not fulfilled the demands for the improved accuracy, yet. In order to improve the accuracy, the consensus scoring method has been developed. It combines docking scores from various scoring functions without considering characteristics of the docking scores. In this study, we adopted the concepts of the consensus scoring, and improved the docking score from each docking programs, DOCK, FRED or GOLD, for virtual screening. Instead using simple sum of score components in those docking scores, weight factors of the score components were introduced and adjusted for better predictions of active ligands during virtual screening. Several optimization processes were tested to find the best optimization methods of the docking scores using a wide variety of 113 target proteins with over 2000 diverse decoys. Finally, the optimizations improved the chance to discover the active ligands by up to 52.4% (e.g. from 36.8% to 56.1% using GOLD) for the test set. Additionally, the combination of the optimized scores using GOLD and FRED improved success rate in the test set by 77.2%, and approximately 70% of ligands for target proteins were predictable in the test set with 20 times enrichment.

Discovery of Novel Antimicrobial Agents Targeting the Bacterial RNA Polymerase by High-Throughput Virtual Screening

“…Then, optimized scores were calculated from the docking scores obtained by those docking programs according as the equations previously developed (Eq. 1 and 2) [6].…”

“…Previously, we developed the optimized docking scores for general usage by using various types of target proteins [6]. A total of 113 ligand-protein complexes were obtained from the Protein Data Bank (PDB).…”

“…In our previous study, the combination of FRED and GOLD made a good combination for the virtual screenings [6]. Thus, we used those two docking programs for this study.…”

“…We also prepared two other unrelated proteins for negative controls (PDB IDs: 1AAQ and 8GCH). Those were simply selected from first and last entries in the dataset previously used for the score optimization [6].…”

“…Since we found that the combination of FRED and GOLD yielded the best results in our previous optimizations [6], we used both FRED and GOLD for this study. The screening library of 120,727 entries from ZINC [8] was docked to the four PDB protein entries above by FRED and GOLD.…”