Progress in deep Markov state modeling: Coarse graining and experimental data restraints.

Area of Science:

• Computational biology
• Biophysics
• Machine learning

Background:

• Deep learning frameworks offer powerful tools for analyzing complex systems like proteins.
• Incorporating physical constraints, such as time-reversibility, is essential for applying these methods to biophysical systems.

Purpose of the Study:

• To advance deep learning methods for analyzing long-timescale protein dynamics.
• To incorporate experimental observables to correct simulation data biases.
• To develop hierarchical models and attention mechanisms for detailed analysis and residue importance identification.

Main Methods:

• Utilized deep learning frameworks with physical constraints (time-reversibility).
• Integrated experimental observables into model estimation to compensate for simulation biases.
• Developed a novel neural network layer for hierarchical modeling.
• Implemented an attention mechanism to highlight important residues for state classification.

Main Results:

• Demonstrated successful application on an ultralong molecular dynamics simulation of the Villin headpiece miniprotein.
• Showcased the ability to compensate for simulation data biases using experimental observables.
• Validated the effectiveness of the hierarchical model and attention mechanism in analyzing protein dynamics and identifying key residues.

Conclusions:

• The enhanced deep learning methodology provides a robust framework for analyzing complex biophysical systems.
• The integration of physical constraints and experimental data improves the accuracy and applicability of molecular dynamics simulations.
• The developed hierarchical models and attention mechanisms offer novel insights into protein behavior and residue function.