Representation of Protein Dynamics Disentangled by Time-Structure-Based Prior

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Representation learning (RL) is crucial for extracting meaningful features from high-dimensional data, with applications in diverse fields.
• Molecular dynamics (MD) simulations generate complex, high-dimensional data essential for understanding biomolecular behavior.
• Current RL methods for MD often focus on capturing slow motions, but disentangling underlying physical factors remains a challenge.

Purpose of the Study:

• To develop novel RL methods tailored for analyzing molecular dynamics (MD) simulation data.
• To improve the disentanglement of underlying physical factors within MD data for better interpretation of conformational transitions.
• To enhance the construction of Markov state models (MSMs) from MD trajectories.

Main Methods:

• Proposed two novel representation learning (RL) methods incorporating a temporal constraint prior in the latent space.
• Applied these methods to analyze molecular dynamics (MD) simulation trajectories of alanine dipeptide and chignolin.
• Utilized total variation measure for quantitative evaluation of disentanglement and comparison with state-of-the-art techniques.

Main Results:

• The proposed RL methods successfully construct Markov state models (MSMs) with implied time scales comparable to existing state-of-the-art approaches.
• Quantitative evaluation confirmed that the methods effectively disentangle physically important coordinates, aiding in the interpretation of biomolecular dynamics.
• Demonstrated improved interpretability of folding/unfolding transitions for chignolin.

Conclusions:

• The developed RL methods offer a powerful approach for analyzing complex biomolecular dynamics.
• Temporal constraints in RL facilitate the disentanglement of key physical factors, leading to better insights into conformational changes.
• These methods provide superior representations for downstream tasks and enhance the interpretability of MD simulation data.