Learning data-efficient coarse-grained molecular dynamics from forces and noise.

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Molecular dynamics (MD) simulations are crucial for understanding biomolecular functions.
• All-atom models are computationally expensive, limiting their application to large systems.
• Machine-learned coarse-grained (MLCG) models offer a computationally efficient alternative, but require extensive training data.

Purpose of the Study:

• To develop a novel framework for training MLCG models that reduces the reliance on large atomistic datasets.
• To integrate generative diffusion model principles with traditional force-matching techniques.
• To enable the construction of accurate and stable MLCG force fields with significantly lower computational cost.

Main Methods:

• Unification of MLCG model training with generative diffusion model principles.
• Integration of traditional force-matching with denoising objectives to recover molecular ensemble distributions.
• Validation across diverse protein folds and scales.

Main Results:

• Accurate high-dimensional distributions of molecular ensembles were recovered.
• Physically consistent and stable force fields were constructed.
• Atomistic data requirements for training were reduced by up to two orders of magnitude.

Conclusions:

• The developed framework substantially lowers the computational cost of constructing accurate MLCG models.
• This approach broadens the applicability of MLCG models to large biomolecular systems.
• It establishes a significant bridge between molecular dynamics simulations and modern generative learning techniques.