State predictive information bottleneck.

Area of Science:

• Computational chemistry
• Statistical mechanics
• Machine learning

Background:

• Analyzing high-dimensional molecular dynamics data requires identifying low-dimensional manifolds, often termed reaction coordinates (RCs), to capture slow dynamics and distinguish metastable states.
• Existing machine learning methods for learning these manifolds are often criticized for lacking physical interpretability and connection to traditional chemical physics concepts.

Purpose of the Study:

• To develop a deep learning approach that learns interpretable reaction coordinates (RCs) from molecular simulation data.
• To bridge the gap between data-driven machine learning and physically meaningful interpretations in molecular dynamics analysis.

Main Methods:

• A deep learning-based state predictive information bottleneck approach was developed to learn the RC from high-dimensional molecular simulation trajectories.
• The method analytically and numerically demonstrates the connection between the learned RC and the committor, a key concept in chemical physics for identifying transition states.

Main Results:

• The learned RC accurately identifies transition states and is demonstrably linked to the committor.
• A crucial hyperparameter, the time delay, provides adjustable control over the coarse-graining level for metastable state classification.
• Comparisons on benchmark systems validated the effectiveness and control offered by the method.

Conclusions:

• This work presents a significant advancement in applying deep learning to molecular simulations by providing a physically interpretable framework.
• The developed method offers a systematic way to learn reaction coordinates and control the granularity of metastable state analysis.