A Standardized Benchmark for Machine-Learned Molecular Dynamics Using Weighted Ensemble Sampling.

Area of Science:

• Computational Biology
• Biophysics
• Molecular Modeling

Background:

• Molecular dynamics (MD) methods, including machine-learned models, are rapidly evolving.
• Lack of standardized validation tools hinders objective comparison of simulation approaches.
• Inconsistent metrics, insufficient sampling, and absent benchmarks impede progress.

Purpose of the Study:

• Introduce a modular benchmarking framework for systematic evaluation of protein MD methods.
• Address the need for standardized, reproducible, and objective comparisons of simulation techniques.
• Facilitate rigorous benchmarking across the molecular simulation community.

Main Methods:

• Utilize weighted ensemble (WE) sampling via WESTPA with TICA-derived progress coordinates for efficient conformational space exploration.
• Employ a flexible propagator interface supporting various simulation engines (classical force fields, ML models).
• Implement a comprehensive evaluation suite with over 19 metrics and visualizations.

Main Results:

• Developed and validated a framework for protein MD method evaluation.
• Generated a dataset of nine diverse proteins extensively simulated.
• Demonstrated utility by comparing classic MD with ML models (CGSchNet).

Conclusions:

• The framework standardizes evaluation protocols for molecular dynamics simulations.
• Enables direct, reproducible comparisons across different MD approaches.
• Provides a foundation for consistent and rigorous benchmarking in the field.