Are predefined decoy sets of ligand poses able to quantify scoring function accuracy?

Area of Science:

• Computational chemistry
• Drug discovery
• Structural biology

Background:

• Selecting appropriate docking programs and scoring functions is a challenge in structure-based drug discovery.
• Existing studies often use predefined decoy sets to compare scoring function performance.
• The reliability of decoy sets for evaluating true pose prediction performance is questioned.

Purpose of the Study:

• To investigate whether predefined decoy sets can reliably evaluate and rank scoring functions for pose prediction.
• To rigorously compare the pose prediction performance of different scoring functions through extensive sampling.
• To identify the strengths and weaknesses of specific scoring functions based on active site properties.

Main Methods:

• Assessed scoring function performance using both rescoring of predefined decoy sets and direct docking calculations.
• Employed extensive sampling of the search space in docking calculations for a more rigorous evaluation.
• Analyzed scoring function performance on subsets of complexes categorized by active site properties.

Main Results:

• Predefined decoy sets accurately reflected rescoring performance but not overall pose prediction accuracy.
• Direct docking calculations revealed performance degradation for scoring functions when not just rescoring.
• Identified relative strengths and weaknesses of ChemPLP, GoldScore, and Astex Statistical Potential.
• Could not determine the cause for the poor performance of all tested functions on this specific test set.

Conclusions:

• Predefined decoy sets are insufficient for a comprehensive evaluation of scoring function pose prediction capabilities.
• Extensive sampling via docking calculations is necessary for a more accurate assessment of pose prediction performance.
• Further investigation is needed to understand the reasons behind the observed poor performance on the tested dataset.