LAST: Latent Space-Assisted Adaptive Sampling for Protein Trajectories

Area of Science:

• Computational biology
• Biophysics
• Structural biology

Background:

• Molecular dynamics (MD) simulations are crucial for studying protein dynamics but are often inefficient due to local energy minima.
• Inefficient sampling hinders the exploration of protein conformational space within practical simulation times.

Purpose of the Study:

• To introduce Latent Space-Assisted Adaptive Sampling (LAST), a novel deep learning-based adaptive sampling method.
• To accelerate the exploration of protein conformational space and improve sampling efficiency in molecular simulations.

Main Methods:

• Utilized variational autoencoders (VAEs) to learn a low-dimensional latent space representation of protein conformations.
• Implemented a cyclic approach involving VAE training, seed structure selection in the latent space, and MD simulations for conformational sampling.
• Validated LAST on two protein systems: *Escherichia coli* adenosine kinase (ADK) and Vivid (VVD).

Main Results:

• LAST successfully identified seed structures on the boundaries of conformational distributions for both ADK and VVD.
• LAST demonstrated significantly enhanced sampling efficiency, achieving large conformational changes in shorter simulation times compared to conventional MD (cMD) and structural dissimilarity sampling (SDS).
• In ADK simulations, LAST revealed two transition paths, while SDS found one and cMD found none. LAST was three times faster than cMD for VVD simulations.

Conclusions:

• LAST is a promising adaptive sampling tool, comparable to existing methods like SDS.
• The method effectively accelerates the exploration of protein conformational space, enabling more efficient molecular dynamics simulations.
• The LAST method is publicly available to support further research in protein dynamics and conformational sampling.