Improving scoring-docking-screening powers of protein-ligand scoring functions using random forest

Abstract

The development of new protein-ligand scoring functions using machine learning algorithms, such as random forest, has been of significant interest. By efficiently utilizing expanded feature sets and a large set of experimental data, random forest based scoring functions (RFbScore) can achieve better correlations to experimental protein-ligand binding data with known crystal structures; however, more extensive tests indicate that such enhancement in scoring power comes with significant under-performance in docking and screening power tests compared to traditional scoring functions. In this work, to improve scoring-docking-screening powers of protein-ligand docking functions simultaneously, we have introduced a Δ_vina RF parameterization and feature selection framework based on random forest. Our developed scoring function Δ_vina RF₂₀ , which employs 20 descriptors in addition to the AutoDock Vina score, can achieve superior performance in all power tests of both CASF-2013 and CASF-2007 benchmarks compared to classical scoring functions. The Δ_vina RF₂₀ scoring function and its code are freely available on the web at: https://www.nyu.edu/projects/yzhang/DeltaVina. © 2016 Wiley Periodicals, Inc.

Keywords: docking; machine learning; protein-ligand binding affinity; random forest; scoring function.

Figure 1

Performance of 22 scoring functions in (A) scoring power measured by Pearson’s R, (B) ranking power in terms of high-level success rate and (C) docking power measured by the success rate when the best-scored pose is considered to match the native pose in CASF-2013 benchmark. Δ_vinaRF₂₀ is colored in red and AutoDock Vina is colored in green. All results colored in blue are obtained from reference^[13].

Figure 2

Performance of 18 scoring functions in (A) scoring power measured by Pearson’s R, (B) ranking power in terms of high-level success rate and (C) docking power measured by the success rate when the best-scored pose is considered to match the native pose in CASF-2007 benchmark. Δ_vinaRF₂₀ is colored in red and AutoDock Vina is colored in green. All results colored in blue are obtained from reference^[12].

Figure 3

Performance of 22 scoring functions in screening power measured by (A) enrichment factor and (B) success rate at top 1% level in CASF-2013 benchmark. Δ_vinaRF₂₀ is colored in red and AutoDock Vina is colored in green. All results colored in blue are obtained from reference^[13].

Figure 4

CASF-2013 benchmark test performance of AutoDock Vina (colored in green), scoring function developed with RF approach (colored in blue) using experimental data alone and the same twenty features in Δ_vinaRF₂₀ scoring function, and scoring functions developed with RF approach and Δ_vinaRF approach (colored in red) using the same twenty features and the same training set for the development of the Δ_vinaRF₂₀ scoring function. (A) Scoring power; (B) Docking power; (C) Screening power. Each set is run 10 times with different random seed for random forest and calculated by averaging over 10 performances except AutoDock Vina. The AutoDock Vina performance is also indicated by dashed line.