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

Area of Science:

• Computational chemistry
• Structural biology
• Bioinformatics

Background:

• Machine learning, particularly random forest, is increasingly used for developing protein-ligand scoring functions.
• Random forest-based scoring functions (RFbScore) show promise in correlating with experimental binding data but often underperform in docking and screening power tests.
• Traditional scoring functions may lack the accuracy needed for comprehensive drug discovery pipelines.

Purpose of the Study:

• To develop a novel protein-ligand scoring function that simultaneously improves scoring, docking, and screening capabilities.
• To address the limitations of existing random forest-based scoring functions in predicting binding affinity and screening efficacy.
• To introduce a new framework for parameterization and feature selection using random forest for enhanced protein-ligand interaction prediction.

Main Methods:

• Developed a Δvina RF parameterization and feature selection framework based on random forest algorithms.
• Integrated 20 descriptors with the AutoDock Vina score to create the Δvina RF20 scoring function.
• Evaluated the performance of Δvina RF20 using the CASF-2013 and CASF-2007 benchmarks for scoring, docking, and screening power tests.

Main Results:

• The developed Δvina RF20 scoring function demonstrated superior performance across all power tests (scoring, docking, and screening) compared to classical scoring functions.
• Δvina RF20 achieved better results on both the CASF-2013 and CASF-2007 benchmark datasets.
• The new scoring function effectively balances scoring accuracy with docking and screening efficiency.

Conclusions:

• The Δvina RF20 scoring function represents a significant advancement in protein-ligand docking and screening.
• This novel approach overcomes the trade-off between scoring power and predictive performance in docking and screening.
• The Δvina RF20 scoring function and its associated code are publicly available for further research and application.