Autoencoders for dimensionality reduction in molecular dynamics: Collective variable dimension, biasing, and...

Area of Science:

• Biochemistry
• Molecular Biology
• Computational Biophysics

Background:

• Heat shock protein 90 (Hsp90) is a crucial molecular chaperone.
• Hsp90 regulates client protein folding and activation via ATP hydrolysis.
• The N-terminal domain (NTD) is the active site of Hsp90.

Purpose of the Study:

• To characterize the dynamic behavior of the Hsp90 NTD.
• To develop and apply autoencoder-learned collective variables (CVs) for Hsp90 NTD dynamics.
• To investigate the optimal dimensionality for CVs in molecular dynamics simulations.

Main Methods:

• Clustering of experimental Hsp90 NTD structures into native states.
• Unbiased molecular dynamics (MD) simulations to generate state-specific datasets.
• Training autoencoders with varying architectures and bottleneck dimensions (k=1-10).
• Adaptive biasing force Langevin dynamics simulations using learned CVs.

Main Results:

• Autoencoder performance is not significantly improved by an extra hidden layer.
• A two-dimensional (2D) bottleneck provides sufficient information for state representation.
• An optimal bottleneck dimension of five was identified.
• Selecting a 2D CV from a 5D latent space outperformed directly learning a 2D CV.
• Transitions between Hsp90 native states were observed using biased free energy dynamics.

Conclusions:

• Autoencoder-learned CVs are effective for characterizing Hsp90 NTD dynamics.
• A 2D CV derived from a higher-dimensional latent space offers superior performance for observing state transitions.
• This approach advances the study of molecular chaperone dynamics and conformational landscapes.