Unbiasing Enhanced Sampling on a High-Dimensional Free Energy Surface with a Deep Generative Model.

Area of Science:

• Computational Chemistry
• Statistical Mechanics
• Machine Learning

Background:

• Enhanced sampling methods using collective variables (CVs) are crucial for studying molecular conformations.
• High-dimensional free energy surfaces pose challenges for traditional density estimation in unbiasing simulations.
• Temperature-accelerated molecular dynamics (TAMD) can incorporate multiple CVs but requires accurate probability distribution modeling.

Purpose of the Study:

• To develop a novel unbiasing method for enhanced sampling simulations.
• To leverage score-based diffusion models for accurate density estimation in high-dimensional spaces.
• To enable the generation of unbiased conformational ensembles for complex systems.

Main Methods:

• Proposed an unbiasing method utilizing score-based diffusion models, a type of deep generative learning.
• Applied the method to multiple Temperature-Accelerated Molecular Dynamics (TAMD) simulations.
• Evaluated the performance against traditional unbiasing techniques.

Main Results:

• The score-based diffusion model unbiasing approach significantly outperformed traditional methods.
• The method successfully generated accurate unbiased conformational ensembles.
• Demonstrated that TAMD can effectively utilize CVs for improved sampling efficiency.

Conclusions:

• Score-based diffusion models provide a powerful solution for unbiasing enhanced sampling simulations.
• This approach enables accurate evaluation of ensemble averages for chemical features.
• Facilitates the study of complex molecular systems by generating reliable conformational data.