Confronting pitfalls of AI-augmented molecular dynamics using statistical physics

Area of Science:

• Computational chemistry and biophysics.
• Application of artificial intelligence in scientific research.

Background:

• Artificial intelligence (AI) enhances molecular simulations by identifying slow modes to accelerate computations.
• Molecular simulations often face limited data, posing a risk of AI models converging to incorrect solutions (spurious regimes).
• Incorrectly identified reaction coordinates (RCs) in AI-driven simulations can lead to significant deviations from ground truth.

Purpose of the Study:

• To develop a novel, automated algorithm to address the challenge of spurious AI solutions in molecular simulations.
• To ensure the reliability and accuracy of AI-driven reaction coordinate identification, especially with limited data.
• To enable more robust and trustworthy application of AI in complex molecular simulations.

Main Methods:

• Developed an automated algorithm based on statistical mechanics principles.
• Utilized a maximum caliber-based framework to learn timescale separation from limited data.
• Focused on maximizing the timescale separation between slow and fast processes for reliable AI solutions.

Main Results:

• Demonstrated the algorithm's applicability on three benchmark problems: peptide conformational dynamics, protein-ligand unbinding, and protein G folding/unfolding.
• The novel algorithm effectively identifies reliable reaction coordinates even with limited simulation data.
• The method ensures AI-driven simulations avoid spurious regimes and maintain accuracy.

Conclusions:

• The developed algorithm provides a trustworthy method for AI-driven molecular simulations.
• This approach enhances the reliability of AI in characterizing complex molecular systems.
• Facilitates increased and robust use of AI in molecular simulations, overcoming data limitations.