Using Time Dependent Rate Analysis to Evaluate the Quality of Machine Learned Reaction Coordinates for Biasing and...

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Biophysics

Background:

• Accurate reaction coordinates (RCs) are crucial for molecular kinetics.
• Quantitative metrics for evaluating RC quality are limited.
• The Exponential Average Time-dependent Rate (EATR) method offers a dimensionless gamma (γ) metric.

Purpose of the Study:

• To evaluate the utility of the gamma (γ) metric for assessing reaction coordinate (RC) quality in molecular dynamics simulations.
• To demonstrate that gamma (γ) can guide the iterative refinement of RCs using the State Predictive Information Bottleneck (SPIB) approach.
• To establish a practical criterion for selecting RCs that yield accurate kinetic predictions.

Main Methods:

• Utilized the dimensionless gamma (γ) metric from the EATR method.
• Employed the iterative State Predictive Information Bottleneck (SPIB) approach to approximate RCs.
• Evaluated six protein-ligand dissociation systems.
• Computed gamma (γ) values and mean accelerated times (τ) across systematically scanned fitting parameters.

Main Results:

• Demonstrated that gamma (γ) increases towards 1 as the reaction coordinate (RC) quality improves.
• Observed a consistent inverse correlation between gamma (γ) and mean accelerated times (τ), where τ decreases with improved RCs.
• Showcased that gamma (γ) effectively indicates when an RC's utility for kinetic prediction is enhanced through iterative updates.

Conclusions:

• The gamma (γ) metric serves as a practical and reliable criterion for evaluating reaction coordinate (RC) quality.
• Gamma (γ) provides valuable guidance for selecting SPIB-derived RCs that lead to quantitative kinetic predictions.
• This work establishes a quantitative framework for assessing and improving RCs in molecular dynamics simulations.