Molecular latent space simulators.

Area of Science:

• Computational Chemistry and Molecular Dynamics
• Biophysics and Protein Folding
• Machine Learning and Deep Learning Applications

Background:

• Molecular dynamics (MD) simulations are limited to millisecond timescales due to small integration time steps.
• Existing methods like Markov state models (MSMs) and equation-free approaches enable longer timescales but lack continuous atomistic trajectory reconstruction.
• Discretization of configurational space and inability to reconstruct molecular configurations hinder the generation of continuous atomistic trajectories.

Purpose of the Study:

• To develop a novel method, latent space simulators (LSS), for learning kinetic models from molecular dynamics data.
• To enable the generation of continuous atomistic simulation trajectories at significantly reduced computational cost.
• To accurately reproduce molecular structure, thermodynamics, and kinetics for systems like protein folding.

Main Methods:

• Training three deep learning networks to learn slow collective variables.
• Propagating system dynamics within the learned slow latent space.
• Generatively reconstructing molecular configurations from the latent space representation.

Main Results:

• Demonstrated LSS on the Trp-cage miniprotein system, producing novel ultra-long synthetic folding trajectories.
• Achieved accurate reproduction of atomistic molecular structure, thermodynamics, and kinetics.
• Reduced computational cost by six orders of magnitude compared to traditional MD simulations.

Conclusions:

• Latent space simulators (LSS) offer a computationally efficient method for generating long-timescale molecular dynamics trajectories.
• The approach enables significantly improved sampling and reduced statistical uncertainties in thermodynamic and kinetic estimates.
• LSS provides a powerful tool for studying complex molecular systems and protein folding dynamics.