Divide and Cluster: The DIVINE Framework for Deterministic Top-Down Analysis of Molecular Dynamics Trajectories

Area of Science:

• Computational chemistry
• Biophysics
• Data analysis

Background:

• Molecular dynamics (MD) simulations generate large datasets requiring robust analysis.
• Traditional clustering methods for MD trajectories often struggle with scalability and reproducibility.
• Existing methods may require significant computational resources (e.g., O(N^2) pairwise distances) and can be stochastic.

Purpose of the Study:

• To introduce DIVIsive N-ary Ensembles (DIVINE), a deterministic clustering framework for MD trajectories.
• To develop a scalable and reproducible method for analyzing complex molecular dynamics data.
• To offer an interpretable and efficient alternative to existing MD clustering techniques.

Main Methods:

• DIVINE employs a deterministic, top-down hierarchical clustering approach.
• It utilizes n-ary similarity principles to recursively split clusters, avoiding large distance matrices.
• Supports multiple cluster selection criteria (e.g., weighted variance) and deterministic initialization (NANI).

Main Results:

• DIVINE achieved comparable or superior clustering quality to bisecting k-means on a 305 μs villin headpiece trajectory.
• Demonstrated reduced runtime and eliminated stochastic variability compared to conventional methods.
• The single-pass design allows efficient exploration of various clustering resolutions.

Conclusions:

• DIVINE offers a scalable, interpretable, and deterministic solution for MD trajectory clustering.
• It presents a practical and robust alternative to current standard methods.
• The framework is available as part of the open-source MDANCE package.