Using classifiers to understand coarse-grained models and their fidelity with the underlying all-atom systems

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning applications

Background:

• Coarse-grained (CG) molecular dynamics models use effective Hamiltonians, often optimized against atomistic simulation data.
• Current validation methods for CG models rely on low-dimensional statistics, which may not fully capture model inaccuracies.

Purpose of the Study:

• To introduce a novel framework for estimating high-dimensional errors in CG models.
• To enhance the validation process of CG molecular dynamics simulations using explainable AI.

Main Methods:

• Employing classification as a variational method to estimate high-dimensional error.
• Utilizing Shapley additive explanations (SHAP), a form of explainable machine learning, for model interpretation.
• Demonstrating the approach with two coarse-grained protein models.

Main Results:

• The proposed method successfully estimates high-dimensional errors in CG models.
• Explainable machine learning techniques provide insights into model discrepancies.
• The framework offers a more rigorous validation compared to traditional low-dimensional statistics.

Conclusions:

• Classification and explainable AI offer a powerful approach to quantitatively assess CG model accuracy.
• This framework can improve the reliability and trustworthiness of molecular dynamics simulations.
• The method holds potential for verifying the accurate propagation of atomistic effects, like allostery, in CG models.