Two for One: Diffusion Models and Force Fields for Coarse-Grained Molecular Dynamics.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Coarse-grained (CG) molecular dynamics is crucial for simulating large biological systems.
• Accurate CG force fields are essential but challenging to develop.
• Current methods often require complex force inputs during training.

Purpose of the Study:

• To develop a novel method for learning CG force fields without requiring force inputs.
• To leverage score-based generative models, specifically diffusion models, for CG force field learning.
• To demonstrate the efficacy of this approach in simulating biological systems.

Main Methods:

• Training a diffusion generative model on protein structures from molecular dynamics simulations.
• Utilizing the score function of the trained model as an approximate CG force field.
• Applying the learned force field to simulate CG molecular dynamics.

Main Results:

• The learned score function effectively approximates a CG force field.
• The method enables CG molecular dynamics simulations without explicit force inputs.
• Improved performance in simulating protein systems up to 56 amino acids.
• Accurate reproduction of CG equilibrium distributions.
• Preservation of protein folding dynamics observed in all-atom simulations.

Conclusions:

• Score-based diffusion models offer a simplified and effective approach to learning CG force fields.
• This method advances the simulation capabilities of CG molecular dynamics for biological processes.
• The approach holds promise for studying larger and more complex biological systems efficiently.