Multiscale Reweighted Stochastic Embedding: Deep Learning of Collective Variables for Enhanced Sampling

Area of Science:

• Computational Chemistry
• Statistical Mechanics
• Machine Learning

Background:

• Enhanced sampling simulations are crucial for studying complex molecular systems.
• Identifying accurate collective variables (CVs) is essential for efficient free energy landscape exploration.
• Traditional methods often struggle to capture all relevant metastable states.

Purpose of the Study:

• To develop a novel machine learning method for automated collective variable (CV) discovery.
• To improve the efficiency and accuracy of enhanced sampling simulations.
• To enable the characterization of free energy landscapes in complex systems.

Main Methods:

• Propose multiscale reweighted stochastic embedding (MRSE), a novel deep learning approach.
• Utilize a parametric version of stochastic neighbor embedding with a deep neural network.
• Incorporate weight-tempered sampling, multiscale feature representation, and reweighting procedures.

Main Results:

• MRSE automatically learns low-dimensional collective variables (CVs) from high-dimensional simulation data.
• The method effectively balances equilibrium sampling with the exploration of high-free-energy metastable states.
• MRSE successfully characterized metastable states in the Müller-Brown potential, alanine dipeptide, and alanine tetrapeptide.

Conclusions:

• MRSE provides an effective automated approach for discovering collective variables (CVs) for enhanced sampling.
• The method enhances the ability to represent and sample complex free energy landscapes.
• MRSE offers a powerful tool for advancing molecular simulations and understanding chemical processes.