Deep learning path-like collective variable for enhanced sampling molecular dynamics.

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Enhanced sampling techniques are crucial for exploring complex free energy landscapes.
• Existing reaction path collective variables have limitations in defining complex pathways.

Purpose of the Study:

• Introduce a novel path-like collective variable, DeepLNE (deep-locally non-linear-embedding).
• Address limitations of current methods for exploring reactive pathways.

Main Methods:

• DeepLNE is inspired by locally linear embedding and trained on reactive trajectories.
• Utilizes a differentiable generalized autoencoder combining neural networks and k-nearest neighbor selection.
• Automatically selects neighbor search metrics and learns paths without manual landmark selection.

Main Results:

• DeepLNE closely approximates ideal reaction coordinates in toy models (Müller-Brown potential, alanine dipeptide).
• Accelerates transitions and estimates folding free energy in RNA tetraloop molecular dynamics simulations.

Conclusions:

• DeepLNE offers an effective and automated approach for defining reaction coordinates.
• Demonstrates utility in accelerating molecular dynamics simulations and free energy estimation for complex systems.