DRUMBEAT temporally resolved interpretable machine learning model for characterizing state transitions in protein dynamics.

Conformational transitions are central to protein function, yet their mechanistic analysis remains challenging due to the multi-dimensionality and timescales underlying the molecular motions. While interpretable network models such as Bayesian networks have advanced the identification of key residue interactions in molecular dynamics (MD) data, they lack temporal resolution and cannot capture event sequences during transitions. Here, we introduce Dynamically Resolved Universal Model for BayEsiAn network Tracking (DRUMBEAT), a machine learning approach that combines a universal graph topology with sliding-window rescoring to generate interpretable, time-resolved maps of cooperative events in MD trajectories. Applying DRUMBEAT to the benchmark Fip35 WW domain folding trajectories from DE Shaw Research Group, we recover both major folding pathways and critical residues highlighted by experiment. DRUMBEAT provides new insight by (1) uncovering unknown protein features important for transition, and (2) dissecting the order and timing of conformational changes, revealing the precise sequence of residue contact closures during folding. Robustness analysis demonstrates that both the universal graph and time-resolved results are highly consistent across replicates. These findings establish DRUMBEAT as a scalable and interpretable machine learning framework for dissecting the dynamics of protein folding and other conformational transitions, offering a tool for the mechanistic study of biomolecular dynamics.