From data to noise to data for mixing physics across temperatures with generative artificial intelligence

Area of Science:

• Computational Chemistry and Physics
• Biomolecular Simulations
• Artificial Intelligence in Science

Background:

• Predicting molecular behavior at unsimulated temperatures is crucial for understanding chemical and physical processes.
• Traditional methods often struggle with accurate sampling of complex biomolecular energy landscapes across diverse temperature ranges.

Purpose of the Study:

• To develop a new computational framework for predicting molecular properties at arbitrary temperatures.
• To enhance the sampling efficiency and accuracy of biomolecular simulations using generative AI.

Main Methods:

• Integration of statistical mechanics with generative artificial intelligence, specifically denoising diffusion probabilistic models.
• Application of replica exchange molecular dynamics combined with AI for superior energy landscape sampling.
• Treating temperature as a fluctuating variable to sample joint probability distributions in configuration and temperature space.

Main Results:

• Demonstrated superior sampling of biomolecular energy landscapes for peptides and RNA at unsimulated temperatures.
• Successfully identified previously unseen transition and metastable states in all-atom water simulations.
• Bypassed the need for extensive simulations across multiple temperatures and enabled easy identification of unphysical states.

Conclusions:

• The developed AI-driven framework effectively bridges the gap between simulated and target temperatures for molecular systems.
• This approach offers a powerful, generalizable method for extracting information across varying control parameters in simulations and experiments.
• The method holds significant potential for accelerating discovery in computational chemistry, physics, and related fields.