Minimum-Excess-Work Guidance: Score-Based Sampling with Experimental Data or Sparse Restraints.

Area of Science:

• Computational Chemistry
• Statistical Mechanics
• Machine Learning

Background:

• Deep generative models like Boltzmann generators (BGs) and emulators (BEs) are crucial for molecular simulations.
• Refining these models with sparse experimental data is challenging due to a lack of standardized methods.

Purpose of the Study:

• To develop a principled method for refining pretrained generative models using sparse external information.
• To leverage thermodynamic principles to guide generative models towards experimental data.

Main Methods:

• Proposed a regularization technique inspired by thermodynamic work to guide pretrained probability flow models.
• Developed Path Guidance for rare transition state sampling and Observable Guidance for aligning with experimental data.
• Applied the framework to coarse-grained Boltzmann emulators for protein systems.

Main Results:

• Demonstrated improved sampling of transition configurations and correction of systematic biases using experimental data.
• Showcased the framework's versatility across different model protein systems.
• Provided theoretical bounds on distributional differences between guided and unguided models.

Conclusions:

• The proposed method offers a physics-inspired, efficient alternative to standard fine-tuning for data-scarce domains.
• Successfully bridges thermodynamic principles with deep generative architectures for molecular simulations.
• Achieved enhanced sample efficiency and bias reduction, applicable to molecular simulations and beyond.