MDSCAN

MDSCAN performs memory-efficient clustering of long molecular dynamics (MD) trajectories using HDBSCAN with vantage-point tree encodings and a dual-heap quasi-minimum spanning tree strategy to enable scalable RMSD-based conformational analysis.

Key Features:

• Memory Efficiency: Significantly reduces RAM requirements compared to traditional HDBSCAN implementations, enabling processing of long trajectories without prohibitive memory demands.
• Vantage-Point Tree Encoding: Encodes trajectories as a specialized class of vantage-point tree to decrease time complexity for large datasets.
• Dual-Heap Quasi-Minimum Spanning Tree: Employs a dual-heap strategy to construct a quasi-minimum spanning tree, reducing memory usage and improving performance.
• Clustering Algorithm: Leverages the hierarchical density-based spatial clustering of applications with noise (HDBSCAN) algorithm to identify clusters and noise in MD trajectory data.
• Distance Metric: Uses the root-mean-square deviation (RMSD) metric for trajectory frame comparison and clustering.
• Benchmark Performance: Processed a trajectory of 1,000,000 frames using RMSD in approximately 21 hours while consuming less than 8 GB of RAM, compared to an accelerated HDBSCAN* implementation that would require over 32 TB of RAM for a similar task.

Scientific Applications:

• Protein Folding Analysis: Identification and characterization of conformational states and folding pathways from long MD trajectories.
• Ligand Binding and Dynamics: Analysis of ligand binding poses and conformational rearrangements in drug–target simulations.
• Conformational Landscape Mapping: Exploration of biomolecular conformational ensembles and dynamic processes across large-scale MD datasets.

Methodology:

Encodes MD trajectories as vantage-point trees, applies hierarchical density-based clustering via HDBSCAN using the RMSD metric, and constructs a quasi-minimum spanning tree through a dual-heap strategy.