DRUMBEAT: Temporally resolved interpretable machine learning model for characterizing state transitions in protein...

Area of Science:

• Biophysics
• Computational Biology
• Machine Learning

Background:

• Protein conformational transitions are crucial for function but challenging to study mechanistically due to complex dynamics.
• Existing network models like Bayesian networks identify residue interactions but lack temporal resolution for dynamic events.

Purpose of the Study:

• Introduce DRUMBEAT (Dynamically Resolved Universal Model for Bayesian network Tracking), a novel machine learning approach.
• Generate interpretable, time-resolved maps of cooperative events in molecular dynamics (MD) trajectories.
• Dissect the order and timing of conformational changes during protein folding.

Main Methods:

• DRUMBEAT combines a universal graph topology with sliding-window rescoring for MD data analysis.
• Applied to Fip35 WW domain folding trajectories to identify folding pathways and critical residues.
• Robustness validated across multiple sampling replicates.

Main Results:

• Successfully recovered major folding pathways and known critical residues for the Fip35 WW domain.
• Uncovered previously unknown protein features important for conformational transitions.
• Provided precise temporal sequences of residue contact closures during individual folding events.

Conclusions:

• DRUMBEAT offers a scalable and interpretable machine learning framework for analyzing protein folding dynamics.
• The method provides new mechanistic insights into the sequence and timing of conformational changes.
• Establishes DRUMBEAT as a generalizable tool for studying biomolecular dynamics.