Active subspace learning for coarse-grained molecular dynamics.

Area of Science:

• Computational chemistry and biophysics
• Molecular dynamics simulations
• Machine learning for scientific discovery

Background:

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

Purpose of the Study:

• To introduce Active Subspace Coarse-Graining (ASCG), a novel framework for bottom-up coarse-graining.
• To develop a unified mathematical framework for defining coarse-grained mapping, interactions, and equations of motion.
• To demonstrate the accuracy and efficiency of ASCG on biomolecular systems.

Main Methods:

• Employed active subspace learning to identify key degrees of freedom from potential energy gradients.
• Developed a unified framework for simultaneous definition of mapping, interactions, and dynamics.
• Trained ASCG models using atomistic MD simulations of dialanine, Trp-cage, and chignolin.

Main Results:

• ASCG reduces solute dimensionality by over 90% while capturing many-body intramolecular effects.
• Achieved highly accurate free energy surfaces with low Jensen-Shannon divergences (as low as 0.034).
• Enabled significantly larger simulation time steps (up to 100 fs) and required only 100 ns of training data.

Conclusions:

• ASCG provides a robust, data-efficient method for learning coarse-grained representations of intramolecular forces.
• The framework departs from traditional particle-based models, offering a unified approach.
• Future work will focus on extending ASCG to handle intermolecular interactions using locality-aware representations.