Scaling transferable coarse-graining with mean force matching.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Coarse-grained molecular dynamics (CGMD) offers computational efficiency but often lacks accuracy.
• Machine-learned potentials (MLPs) can bridge this gap, matching atomistic accuracy.
• Developing MLPs for CGMD faces data scaling challenges with traditional objectives.

Purpose of the Study:

• To introduce and validate mean force matching (MFM) as a superior objective for training CGMD models.
• To demonstrate MFM's ability to overcome data scaling limitations in MLP development.
• To establish MFM's thermodynamic consistency and accuracy for protein simulations.

Main Methods:

• Implementing mean force matching for training thermodynamically consistent CGMD potentials.
• Systematically analyzing and reducing noise in the objective function.
• Benchmarking MFM against other common objectives using thermodynamic consistency as the primary metric.

Main Results:

• MFM requires 50× fewer training samples than conventional methods.
• MFM achieves higher accuracy in the potential of mean force for unseen proteins.
• Noise reduction in the objective function enables scalable machine learning architectures for CGMD.

Conclusions:

• Mean force matching provides a computationally efficient and accurate approach to developing transferable CGMD models.
• MFM significantly reduces the data requirements for training MLPs in coarse-grained simulations.
• This work enables the creation of highly accurate and scalable CGMD models for complex systems.