Unsupervised learning of progress coordinates during weighted ensemble simulations: Application to millisecond...

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Simulating rare events like protein folding is computationally challenging.
• Traditional methods struggle with long timescales and complex conformational landscapes.
• Unsupervised learning offers potential for identifying key molecular coordinates.

Purpose of the Study:

• To develop and validate a deep learning-enhanced West-East (WE) method for efficient rare event sampling.
• To improve the accuracy and speed of estimating protein folding rates.
• To advance unsupervised learning techniques for identifying slow collective variables in molecular dynamics.

Main Methods:

• A convolutional variational autoencoder was used to model system conformations in a latent space.
• Outliers in the latent space were identified "on-the-fly" to enhance sampling efficiency.
• Simulations utilized discrete-state synthetic molecular dynamics trajectories and a Markov state model.
• The method was applied to simulate a millisecond protein folding process.

Main Results:

• The DL-enhanced WE method demonstrated a >3-fold increase in efficiency for estimating folding rate constants.
• Successful simulation of a millisecond protein folding process was achieved.
• The approach effectively identified and utilized outliers to guide the sampling process.

Conclusions:

• Deep learning-based outlier detection significantly enhances the efficiency of the WE method for rare event sampling.
• This work represents a substantial advancement in unsupervised learning for identifying slow coordinates in molecular simulations.
• The developed method holds promise for accelerating the study of complex biological processes.