Best practices for constructing, preparing, and evaluating protein-ligand binding affinity benchmarks [Article v0.1].

Area of Science:

• Computational chemistry
• Drug discovery
• Molecular modeling

Background:

• Free energy calculations are vital for structure-enabled drug discovery.
• Current benchmarking practices lack standardization and suffer from data quality issues.
• Existing benchmarks may not accurately predict real-world performance.

Purpose of the Study:

• To establish guidelines for creating high-quality, standardized benchmark datasets for evaluating computational methods.
• To develop best practices for preparing benchmark inputs and analyzing results.
• To provide a standardized benchmark set (PLBenchmarks) and assessment toolkit (arsenic).

Main Methods:

• Curating experimental data for benchmark set construction.
• Developing best practices for benchmark input preparation.
• Implementing statistically sound analysis of prediction results.
• Introducing the PLBenchmarks dataset and arsenic toolkit.

Main Results:

• Guidelines for data curation, input preparation, and analysis are presented.
• A standardized, curated, and versioned benchmark set (PLBenchmarks) is released.
• An open-source toolkit (arsenic) for standardized assessments is provided.
• The work addresses limitations in existing benchmark reports.

Conclusions:

• Standardized, high-quality benchmarks are essential for reliable assessment of computational methods in drug discovery.
• The PLBenchmarks set and arsenic toolkit offer a community resource for method evaluation.
• These guidelines and tools will improve the accuracy and applicability of free energy calculations and machine learning predictions.