Statistically Optimal Force Aggregation for Coarse-Graining Molecular Dynamics

Area of Science:

• Computational chemistry
• Molecular dynamics simulations
• Machine learning in science

Background:

• Coarse-grained (CG) models enable simulating large molecular systems beyond atomistic methods.
• Training accurate CG models is crucial but challenging.
• Current methods often map all-atom forces to CG representations inefficiently.

Purpose of the Study:

• To develop a statistically efficient and accurate method for learning coarse-grained force fields.
• To address limitations in existing force mapping techniques for CG model training.
• To improve the accuracy of machine-learned CG models.

Main Methods:

• Developed an optimization statement for force mappings in CG model training.
• Implemented and tested optimized force mapping strategies.
• Utilized all-atom molecular dynamics data for training CG force fields.
• Applied the method to miniproteins like chignolin and tryptophan cage.

Main Results:

• Demonstrated that optimized force maps lead to substantially improved CG force fields.
• Showcased the statistical inefficiency and potential incorrectness of common mapping methods, especially with constraints.
• Validated the improved CG force fields using benchmark molecular systems.
• Released the developed method as open-source code.

Conclusions:

• Optimized force mapping is essential for accurate machine-learned coarse-grained models.
• The proposed method enhances CG force field learning from existing simulation data.
• This advancement facilitates more reliable and accurate simulations of large molecular complexes.