Acceleration of Molecular Simulations by Parametric Time-Lagged tSNE Metadynamics

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Molecular simulations are computationally expensive, limiting their application.
• Enhanced sampling methods, like metadynamics, are crucial for accelerating simulations.
• Metadynamics requires careful selection of collective variables (CVs) to define bias potentials.

Purpose of the Study:

• To utilize unsupervised machine learning for designing collective variables (CVs).
• To accelerate molecular dynamics simulations using metadynamics with ML-designed CVs.
• To test the efficacy of parametric time-lagged t-distributed stochastic neighbor embedding (ptltSNE) for CV discovery.

Main Methods:

• Employed parametric time-lagged t-distributed stochastic neighbor embedding (ptltSNE), an unsupervised machine learning technique, to identify and design collective variables (CVs).
• Applied metadynamics simulations, incorporating the designed CVs, to study protein folding dynamics.
• Utilized both standard metadynamics with an alpha-RMSD CV and parallel tempering metadynamics to explore different conformational landscapes.

Main Results:

• The ptltSNE method successfully generated a conformational map and distinguished between fast and slow conformational changes from a Trp-cage trajectory.
• Metadynamics with an alpha-RMSD CV resulted in one observed protein folding event.
• Parallel tempering metadynamics significantly enhanced sampling, yielding 10 folding events in a shorter simulation time with multiple replicas.

Conclusions:

• Unsupervised machine learning, specifically ptltSNE, provides a powerful approach for designing effective collective variables (CVs) in molecular simulations.
• The developed methodology accelerates enhanced sampling techniques like metadynamics, enabling the study of complex biological processes such as protein folding.
• Combining ML-driven CVs with advanced sampling methods like parallel tempering offers a promising strategy for overcoming computational limitations in molecular dynamics.