Latent Models of Molecular Dynamics Data: Automatic Order Parameter Generation for Peptide Fibrillization

Area of Science:

• Computational chemistry
• Artificial intelligence
• Biophysics

Background:

• Variational autoencoders (VAEs) are powerful artificial neural networks for dimensionality reduction.
• Molecular dynamics (MD) simulations are crucial for understanding complex molecular behaviors like self-assembly.
• Amyloid aggregation is a significant process in various diseases, requiring accurate monitoring.

Purpose of the Study:

• To apply VAEs to MD simulations of peptide self-assembly.
• To develop an automatically learned order parameter for amyloid aggregation.
• To compare the VAE-derived order parameter with the traditional nematic order parameter.

Main Methods:

• Utilized VAEs to process high-dimensional data from MD simulations.
• Generated a single-valued order parameter by time-averaging latent parametrizations.
• Analyzed the VAE latent space to identify key internal coordinates influencing aggregation.

Main Results:

• The VAE-derived order parameter provided more detailed insights into aggregation mechanisms.
• VAE successfully identified fibril formation where the nematic order parameter failed.
• Latent space analysis allowed direct interpretation of the order parameter in terms of system behavior.

Conclusions:

• Latent models are effective for representing complex dynamic ensembles.
• VAEs offer a convenient dimensionality reduction tool for large-scale systems.
• This approach bypasses the need for researcher speculation in analyzing complex molecular transitions.