MDSCAN: RMSD-based HDBSCAN clustering of long molecular dynamics

Area of Science:

• Computational Biology and Chemistry
• Data Science and Machine Learning

Background:

• Clustering is a key unsupervised learning technique for grouping similar data points.
• Geometrical clustering of molecular dynamics (MD) trajectories reveals system conformational behavior.
• Existing methods struggle with long MD trajectories due to quadratic time/memory complexity, especially with the RMSD metric.

Purpose of the Study:

• To introduce MDSCAN, a novel software for memory-efficient RMSD-based clustering of long MD trajectories.
• To provide a user-friendly tool for non-programmers to analyze complex molecular dynamics data.
• To overcome the computational limitations of traditional clustering algorithms for large-scale trajectory analysis.

Main Methods:

• MDSCAN is inspired by HDBSCAN, a robust density-based clustering algorithm.
• It employs a vantage-point tree encoding for trajectories to reduce time complexity.
• A dual-heap approach is utilized to construct a quasi-minimum spanning tree, minimizing memory usage.

Main Results:

• MDSCAN successfully processed a 1-million-frame trajectory using the RMSD metric in approximately 21 hours with <8 GB RAM.
• This represents a significant improvement over accelerated HDBSCAN*, which would require >32 TB RAM for a similar task.
• The software demonstrates high efficiency and scalability for analyzing extensive molecular dynamics simulations.

Conclusions:

• MDSCAN offers a practical and computationally efficient solution for clustering long MD trajectories.
• It democratizes the analysis of large molecular dynamics datasets for researchers without extensive programming expertise.
• The software's memory efficiency and speed make it suitable for tackling previously intractable simulation data.