Active subspace learning for coarse-grained molecular dynamics.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Atomistic molecular dynamics (MD) simulations are computationally expensive.
• Coarse-graining (CG) methods simplify complex systems but often require separate parameterization for mapping, interactions, and dynamics.
• Developing systematic, data-driven CG methods is crucial for studying large biomolecular systems.

Purpose of the Study:

• To introduce Active Subspace Coarse-Graining (ASCG), a novel framework for systematic bottom-up coarse-graining.
• To develop a unified approach that simultaneously defines CG mapping, effective interactions, and equations of motion.
• To demonstrate the efficiency and accuracy of ASCG for biomolecular simulations.

Main Methods:

• Employed active subspace learning to identify optimal projections of atomistic degrees of freedom.
• Utilized these projections to define CG variables capturing dominant collective motions.
• Derived effective CG forces and noise terms directly from potential energy gradients.
• Applied the ASCG method to biomolecules: dialanine, Trp-cage, and chignolin.

Main Results:

• Achieved accurate free energy surface recapitulation (Jensen-Shannon divergence as low as 0.034).
• Significantly reduced system dimensionality (>90%) and eliminated solvent degrees of freedom.
• Enabled larger integration timesteps (up to 100 fs), 4-10x greater than conventional CG methods.
• Demonstrated accuracy with minimal training data (100 ns).

Conclusions:

• ASCG provides a robust, data-efficient, and interpretable framework for learning complete CG representations.
• The unified mathematical framework eliminates the need for separate parameterization schemes.
• ASCG represents a significant departure from traditional particle-based CG models, offering improved computational efficiency.