Explaining reaction coordinates of alanine dipeptide isomerization obtained from deep neural networks using...

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Artificial Intelligence

Background:

• Identifying reaction coordinates is crucial for understanding molecular transitions.
• Deep learning models using artificial neural networks (ANNs) are increasingly used for this purpose.
• Explaining the contribution of input variables in complex ANNs remains a challenge.

Purpose of the Study:

• To apply Explainable Artificial Intelligence (XAI) methods to interpret deep learning models for reaction coordinate prediction.
• To identify the key collective variables that contribute to reaction coordinates in molecular systems.
• To develop an AI-aided framework for explaining reaction coordinates in complex molecular systems.

Main Methods:

• Utilized Explainable Artificial Intelligence (XAI) techniques, specifically Local Interpretable Model-agnostic Explanation (LIME) and Shapley Additive exPlanations (SHAP).
• Applied these methods to analyze nonlinear regressions with deep learning for the committor of alanine dipeptide isomerization in vacuum.
• Evaluated the contribution of collective variables to the predicted reaction coordinates.

Main Results:

• XAI methods (LIME and SHAP) successfully determined the contribution of each collective variable to the predicted reaction coordinates.
• The identified important features align with previously reported dihedral angles from committor test analysis.
• Demonstrated the effectiveness of XAI in interpreting complex deep learning models for molecular systems.

Conclusions:

• XAI provides a powerful tool for explaining reaction coordinates derived from deep learning.
• The AI-aided framework enhances the interpretability of molecular dynamics simulations, especially with increasing degrees of freedom.
• This approach offers significant advantages for understanding complex molecular behavior and reaction pathways.