Characterizing Metastable States with the Help of Machine Learning.

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Atomistic simulations generate vast datasets, making analysis and identification of metastable states increasingly difficult.
• Understanding the slowest dynamical modes is crucial for characterizing complex systems.

Purpose of the Study:

• To develop an efficient method for analyzing long atomistic simulation trajectories.
• To identify and characterize metastable states within complex molecular systems.
• To demonstrate the applicability to both unbiased and biased simulations.

Main Methods:

• Variational approach to conformation dynamics (VAC) to identify slowest dynamical modes.
• Machine learning techniques to discover physical descriptors of metastable states.
• Application to protein systems like chignolin and bovine pancreatic trypsin inhibitor.

Main Results:

• Successfully located and hierarchically organized metastable states.
• Discovered physical descriptors characterizing these states.
• Analysis performed in seconds for complex protein simulations.

Conclusions:

• The proposed method offers an efficient and versatile approach for analyzing complex molecular dynamics simulations.
• It facilitates the discovery and characterization of metastable states and their underlying dynamics.
• Applicable across different simulation types, enhancing its utility in biophysical studies.