MDANCE: Molecular Dynamics Analysis with N-ary Clustering Ensembles

A transformative framework for analyzing molecular dynamics simulations through advanced clustering algorithms

Installation • The Problem • Our Solution • Key Features • Quick Start • Publications • Contributing

The Problem

Molecular Dynamics (MD) simulations generate terabytes of conformational data, but extracting meaningful biological insights remains challenging. Traditional clustering methods struggle with:

• Exponential complexity - MD datasets are massive.
• Poor initialization - leading to suboptimal clustering.
• Pathway ambiguity - difficulty identifying dominant biological pathways.
• Native structure prediction - accurately identifying biologically relevant states.
• Pairwise similarity limitations - traditional methods only compare pairs of objects, causing performance bottlenecks.
• Stochastic variability - lack of reproducibility across clustering runs.

Our Solution

MDANCE introduces a novel n-ary similarity framework that transforms how we analyze MD trajectories. Our algorithms provide:

• Linear scaling - from O(N²) to O(N) complexity.
• Deterministic results - reproducible science.
• Biological relevance - algorithms designed for structural biology.
• Unprecedented accuracy - validated against experimental structures.
• Extended similarity techniques - swift identification of high and low-density regions in linear time.

Key Features

🪄 NANI - Smart k-means Initialization

Breakthrough: Deterministic centroid initialization using n-ary comparisons to identify high-density regions and select diverse initial conformations.

Key Advantages:

• Solves the seed selection challenge in k-means clustering.
• Creates compact, well-separated clusters that accurately find metastable states.
• Provides consistent cluster populations across replicates.
• Dramatically reduces runtime: clusters 1.5 million HP35 frames in ~40 minutes.

🧩 HELM - Scalable Hierarchical Clustering

Breakthrough: Combines k-means efficiency with hierarchical flexibility using n-ary difference functions.

Performance:

• Retains k-means computational efficiency while enabling arbitrary partitions.
• Successfully analyzes simulations with over 1.5 million frames.
• Achieves in ~34 minutes what traditional HAC requires 29 hours for 1.5 million frames.
• Builds hierarchy without expensive pairwise distance matrices.

🌳 DIVINE - Deterministic Divisive Clustering

Breakthrough: Top-down hierarchical clustering framework that recursively splits clusters based on n-ary similarity principles.

Key Features:

• Completely avoids O(N²) pairwise distance matrices.
• Deterministic anchor initialization with NANI.
• Multiple cluster selection criteria including weighted variance metric.
• Single-pass design enables efficient resolution exploration.
• Matches or exceeds bisecting k-means quality with reduced runtime.

🌿 mdBIRCH - Online Clustering for MD Data

Innovation: Adapts BIRCH CF-tree to molecular dynamics data with RMSD-calibrated merge tests.

Key Capabilities:

• Online clustering that processes frames as they arrive.
• Merge test calibrated directly to RMSD for physical interpretability.
• Completely avoids pairwise distance matrices.
• Scales near-linearly with number of frames.
• Two practical protocols: RMSD-anchored runs and blind sweep analysis.
• Processes hundreds of thousands of frames on a single CPU core in seconds.

🔍 SHINE - Pathway Analysis

Transformative: Hierarchical clustering that identifies dominant biological pathways from enhanced sampling data.

Key Advantages:

• Streamlines analysis of pathway ensembles from multiple MD simulations.
• Integrates n-ary similarity with cheminformatics-inspired tools.
• Identifies most representative pathway within each pathway class.
• Provides insight into dominant biomolecular transformation mechanisms.
• Lower computational cost than Fréchet distance approaches.
• Successfully applied to alanine dipeptide and adenylate kinase systems.

🌈 PRISM - Pathway Representation & Clustering

Innovation: Pathway Representation via Intrinsic Structural Medoids enables efficient clustering of multiple molecular pathways.

Key Capabilities:

• Summarizes each pathway by a compact set of representative conformations (structural medoids).
• Clusters pathways using hierarchical agglomerative clustering based on weighted average Hausdorff distance.
• Three construction strategies for medoid selection: global sharing, per-pathway independent, or two-stage refinement.
• Enables efficient analysis of enhanced sampling and multiple pathway ensembles.
• Designed for multi-trajectory pathway comparison and classification.

🎯 eQual - O(N) Clustering

Innovation: Transforms O(N²) Radial Threshold Clustering into O(N) algorithm with novel seed selection and tie-breaking.

Key Features:

• Uses k-means++ for efficient seed selection.
• Employs extended similarity indices for deterministic results.
• Eliminates memory-intensive pairwise RMSD matrices.
• Produces compact and well-separated clusters matching RTC quality.

🏆 PRIME - Native Structure Prediction

Game Changer: Predicts native protein structures from simulation data with unprecedented accuracy. Scientific Validation: PRIME (Protein Retrieval via Integrative Molecular Ensembles) perfectly mapped all structural motifs in benchmark studies and consistently identified native structures within 2Å RMSD of experimental data.

Algorithm Comparison

Algorithm	Complexity	Type	Key Feature	Best Use Case
NANI	O(N)	Initialization	Deterministic centroids	k-means improvement
HELM	O(N)	Hybrid hierarchical	k-means + hierarchical fusion	Large-scale analysis
DIVINE	O(N)	Divisive hierarchical	Top-down splitting	Multi-resolution analysis
mdBIRCH	O(N)	Online clustering	Streaming data processing	Large-scale trajectories
SHINE	O(N)	Hierarchical	Pathway analysis	Enhanced sampling
PRISM	O(N)	Hierarchical	Structural medoid representation	Multi-pathway comparison
eQual	O(N)	Flat clustering	Linear RTC replacement	General purpose
PRIME	O(N)	Post-processing	Native structure prediction	Structure validation

Quick Start

Installation

pip install mdance

Basic Usage

import mdance
import numpy as np

# Load your MD trajectory data
data = np.load('trajectory.npy')

# Use NANI for optimal clustering initialization
from mdance.cluster.nani import KmeansNANI
nani = KmeansNANI(data, n_clusters=5, metric='MSD')
optimal_centroids = nani.initiate_kmeans()

# Cluster with standard *k*-means
from sklearn.cluster import KMeans
kmeans = KMeans(5, init=optimal_centroids[:5], n_init=1)
labels = kmeans.fit_predict(data)

Tutorials

• NANI Tutorial - Smart k-means initialization.
• HELM Tutorial - Scalable hierarchical clustering.
• DIVINE Tutorial - Deterministic divisive clustering.
• mdBIRCH Tutorial - Online clustering for streaming MD data.
• SHINE Tutorial - Pathway analysis.
• PRISM Tutorial - Pathway representation and clustering.
• eQual Tutorial - Linear-time clustering.
• PRIME Tutorial - Native structure retrieval.

Publications

Our methods are backed by peer-reviewed research:

• NANI: J. Chem. Theory Comput. 2024; J. Chem. Inf. Model. 2026
• HELM: J. Chem. Inf. Model. 2025
• DIVINE: J. Chem. Inf. Model. 2026
• mdBIRCH: J. Chem. Theory Comput. 2026
• SHINE: J. Chem. Inf. Model. 2025
• PRISM: Digital Discovery 2026
• eQual: J. Chem. Inf. Model. 2025
• PRIME: J. Chem. Theory Comput. 2024

If you use MDANCE, please cite the paper of each method you use. BibTeX entries for all papers are in CITATION.bib.

Impact

MDANCE is enabling researchers to:

• Accelerate drug discovery by rapidly identifying biologically relevant conformations.
• Understand disease mechanisms through precise pathway analysis.
• Validate computational models against experimental structures.
• Scale analyses to massive simulation datasets.

Contributing

We welcome collaborations and contributions! Whether you're a:

• Computational biologist with novel analysis needs.
• Method developer interested in extending our framework.
• Structural biologist with challenging datasets.

Get involved:

• Open an issue for bug reports or feature requests.
• Submit a pull request for improvements.
• Reach out to discuss research collaborations.

Funding

This research was supported by the National Institute of General Medical Sciences of the National Institutes of Health under award number R35GM150620.