Fitting coarse-grained models to macroscopic experimental data via automatic differentiation

Area of Science:

• Computational chemistry
• Molecular modeling
• Biophysics

Background:

• Developing physics-based models for molecular simulation involves fitting numerous parameters to experimental data.
• Traditional methods are often piecemeal, difficult to reproduce, and lead to fragmented models.

Purpose of the Study:

• To establish a systematic framework for fitting coarse-grained molecular models to macroscopic experimental data.
• To leverage automatic differentiation for efficient parameter optimization and sensitivity analysis.
• To demonstrate broad applicability across diverse biomolecular systems.

Main Methods:

• Utilized automatic differentiation for low-variance gradient estimates.
• Optimized structural, mechanical, and thermodynamic properties of a DNA force field.
• Applied methods across various simulation techniques and timescales (micro- to milliseconds).
• Adapted multitask learning for simultaneous constraint imposition.

Main Results:

• Successfully optimized a DNA force field for multiple properties.
• Demonstrated efficient sensitivity analyses providing insights into parameter behavior.
• Showcased applicability to RNA and DNA-protein hybrid models.
• Achieved accurate simultaneous imposition of multiple constraints.

Conclusions:

• The developed framework enables transparent, reproducible, and community-driven force field development.
• This systematic approach accelerates progress in molecular modeling and simulation.
• Automatic differentiation is key to efficient and insightful model parameterization.