Modeling and enhanced sampling of molecular systems with smooth and nonlinear data-driven collective variables

Area of Science:

• Computational Chemistry
• Molecular Dynamics Simulations
• Biophysics

Background:

• Collective variables (CVs) are crucial for analyzing molecular conformations and free energy landscapes in simulations.
• Existing nonlinear manifold learning methods for identifying CVs lack the required differentiability for enhanced sampling techniques.
• There is a need for systematic methods to identify effective CVs for complex molecular systems.

Purpose of the Study:

• To introduce a novel methodology for constructing smooth, nonlinear, data-driven collective variables (SandCV).
• To enable the use of nonlinear manifold learning outputs in enhanced sampling methods.
• To improve the exploration of free energy landscapes in molecular dynamics.

Main Methods:

• Developed a methodology to build smooth and nonlinear data-driven collective variables (SandCV) from nonlinear manifold learning outputs.
• Utilized an ensemble representative of molecular flexibility to generate SandCV.
• Integrated SandCV with the adaptive biasing force (ABF) method for enhanced sampling.

Main Results:

• Demonstrated the effectiveness of SandCV using alanine dipeptide as a benchmark molecule.
• Showed that SandCV can be non-intrusively combined with existing enhanced sampling methods.
• Illustrated improved exploration of poorly sampled regions in molecular ensembles using SandCV-enhanced simulations.
• Confirmed the transferability of SandCV from simple (vacuum) to complex (explicit water) systems.

Conclusions:

• SandCV provides a differentiable mapping from high-dimensional configurations to low-dimensional collective variables.
• The method facilitates enhanced sampling and exploration of free energy landscapes in molecular dynamics.
• SandCV offers a transferable and effective approach for analyzing molecular flexibility and behavior in complex systems.