Analysis and sampling of molecular simulations with adversarial autoencoders

Area of Science:

• Computational Chemistry
• Machine Learning
• Biophysics

Background:

• Designing collective variables for molecular simulations is challenging, often requiring expert knowledge.
• Machine learning, particularly artificial neural networks, offers potential solutions for automating this process.

Purpose of the Study:

• To introduce and validate the use of adversarial autoencoders (AAEs) for designing collective variables.
• To demonstrate the effectiveness of AAE-derived latent space coordinates for molecular analysis and enhanced sampling.

Main Methods:

• Utilized an adversarial autoencoder architecture to encode molecular simulation data into a latent space.
• Employed an adversarial game to control the distribution of the latent space.
• Applied the method to alanine dipeptide and tryptophan cage miniprotein trajectories.

Main Results:

• The latent space coordinates effectively served as collective variables for analysis and sampling.
• Demonstrated efficient visualization of the conformational landscapes for both systems.
• Showcased accelerated folding of the tryptophan cage using metadynamics with AAE-derived variables.

Conclusions:

• Adversarial autoencoders provide a novel and effective method for automated collective variable discovery.
• This approach significantly enhances the analysis and sampling efficiency in molecular simulations.
• The developed technique shows promise for studying complex biomolecular systems.