Variational embedding of protein folding simulations using Gaussian mixture variational autoencoders

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Molecular dynamics simulations generate high-dimensional data, challenging conventional analysis.
• Dimensionality reduction is crucial for extracting meaningful information from complex biomolecular simulations.

Purpose of the Study:

• To develop an unsupervised machine learning method for dimensionality reduction and clustering of biomolecular conformations.
• To analyze protein folding landscapes and identify metastable states.

Main Methods:

• Developed a Gaussian mixture variational autoencoder (GMVAE) model.
• Utilized Gumbel-softmax distribution for end-to-end differentiability.
• Applied GMVAE to analyze three long-timescale protein folding trajectories.

Main Results:

• GMVAE successfully learned a reduced representation of protein folding free energy landscapes.
• Identified highly separated clusters corresponding to metastable folding states.
• GMVAE embedding visualized protein folding funnels, with folded states at the bottom and unfolded states outside the path.

Conclusions:

• GMVAE offers an effective unsupervised approach for dimensionality reduction and clustering of biomolecular conformations.
• The latent space generated by GMVAE is suitable for kinetic analysis and building Markov state models.
• GMVAE-based kinetic analysis yields folding/unfolding timescales consistent with established methods.