Comparative assessment of scoring functions on a diverse test set

Area of Science:

• Computational chemistry and cheminformatics
• Structural biology and drug discovery
• Bioinformatics and computational drug design

Background:

• Scoring functions are crucial for evaluating protein-ligand interactions in structure-based drug design.
• Numerous scoring functions exist, implemented in commercial software or developed by academic groups.
• A comprehensive benchmark is needed to assess their performance.

Purpose of the Study:

• To comparatively assess the performance of 16 popular scoring functions.
• To evaluate functions based on "docking power", "ranking power", and "scoring power".
• To provide an updated benchmark for scoring function evaluation in drug design.

Main Methods:

• Selected 195 diverse protein-ligand complexes from the PDBbind database.
• Evaluated 16 scoring functions independently of molecular docking or virtual screening.
• Assessed "docking power" using a root-mean-square deviation cutoff of 2.0 Å.
• Analyzed "ranking power" and "scoring power" using experimental binding constants.

Main Results:

• Six scoring functions achieved >70% success rate for "docking power"; consensus schemes improved this to >80%.
• X-Score, DrugScore(CSD), DS::PLP, and SYBYL::ChemScore were top performers for "ranking" and "scoring" power.
• Correlation coefficients between computed and experimental binding constants ranged from 0.545 to 0.644.
• Performance varied across different protein-ligand complex types and evaluation aspects.

Conclusions:

• No single scoring function consistently outperforms others across all evaluated aspects.
• The choice of scoring function should be tailored to the specific application and purpose.
• This study provides valuable insights for selecting appropriate scoring functions in drug design.