Deep learning to decompose macromolecules into independent Markovian domains

Area of Science:

• Computational Biology
• Molecular Dynamics
• Machine Learning

Background:

• Modeling large protein dynamics is challenged by the exponential growth of metastable states in global configuration models.
• Existing methods struggle to efficiently gather state probabilities and transition statistics for complex molecular systems.

Purpose of the Study:

• To develop a data-efficient method for modeling the dynamics of large molecular systems.
• To overcome limitations of global configuration models by decomposing systems into independent subdomains.

Main Methods:

• Combines independent Markov decomposition (IMD) with VAMPnets, a deep learning approach for Markov modeling.
• Establishes a training objective to assess subdomain decomposition accuracy for Markovian dynamics.
• Implements an end-to-end learning framework for simultaneous subdomain decomposition and Markov state model learning.

Main Results:

• The developed framework provides a data-efficient and interpretable summary of complex molecular dynamics.
• Successfully learned Markov state models for independent subdomains within larger molecular systems.
• Demonstrated a significant step towards learning Ising models from simulation data for molecular complexes.

Conclusions:

• The new method offers a promising approach to analyzing the dynamics of large molecular systems.
• While inter-subdomain coupling remains an open challenge, this work simplifies complex system analysis.
• This research advances the ability to model large molecular complexes using simulation data.