Chelombus

Billion-scale molecular clustering and visualization on commodity hardware.

Chelombus enables interactive exploration of ultra-large chemical datasets (up to billions of molecules) using Product Quantization and nested TMAPs. Process the entire Enamine REAL database (9.6B molecules) on a single workstation.

Live Demo: https://chelombus.gdb.tools

Overview

Chelombus implements the "Nested TMAP" framework for visualizing billion-sized molecular datasets:

SMILES → MQN Fingerprints → PQ Encoding → PQk-means Clustering → Nested TMAPs

Key Features:

• Scalability: Stream billions of molecules without loading everything into memory
• Efficiency: Compress 42-dimensional MQN vectors to 6-byte PQ codes (28x compression)
• GPU acceleration: Optional CUDA support for PQ encoding and cluster assignment (~25x speedup)
• Visualization: Navigate from global overview to individual molecules in two clicks
• Accessibility: Runs on commodity hardware (tested: AMD Ryzen 7, 64GB RAM)

Installation

From PyPI (recommended)

pip install chelombus

From Source

git clone https://github.com/afloresep/chelombus.git
cd chelombus
pip install -e .

Platform Notes

Apple Silicon (M1/M2/M3): The pqkmeans library is not currently supported on Apple Silicon Macs. My plan is to rewrite pqkmeans with Silicon and GPU support but that's for a future release... For now, clustering functionality requires an x86_64 system.

GPU Acceleration

Every step in the pipeline supports optional GPU acceleration via the device parameter: encoder training, PQ encoding, cluster training, and label assignment. When a CUDA GPU is available, device='auto' (the default) uses the GPU transparently; otherwise it falls back to CPU.

Requirements: torch and triton (both installed with pip install torch).

encoder = PQEncoder(k=256, m=6, iterations=20)
encoder.fit(training_fps, device='auto')          # GPU KMeans fitting
pq_codes = encoder.transform(fingerprints)        # GPU batch assignment

clusterer = PQKMeans(encoder, k=100000)
clusterer.fit(pq_codes)                           # GPU Triton assign + CPU centroid update
labels = clusterer.predict(pq_codes)              # GPU Triton kernel

# Or force a specific device
labels_cpu = clusterer.predict(pq_codes, device='cpu')
labels_gpu = clusterer.predict(pq_codes, device='gpu')

GPU benchmarks on 1B Enamine REAL molecules (RTX 4070 Ti SUPER 16GB, K=100,000):

Cluster training with the default tol=1e-3 converges earlier (changed centers dropped from 11.8% → 0.2% over 5 iterations), reducing training time further. Extrapolated to 10B molecules: ~1.2 days.

To reproduce (requires 1B MQN fingerprints as .npy chunks):

python scripts/benchmark_1B_pipeline.py --chunks /path/to/chunks --n 1000000000

The GPU implementation uses a custom Triton kernel for cluster assignment that tiles over centers with an online argmin, never materializing the N x K distance matrix. The kernel supports any number of subvectors M via compile-time unrolling (tl.static_range). VRAM usage is ~(M+4) bytes/point, so even an 8 GB GPU can process hundreds of millions of points per batch.

Quick Start

from chelombus import DataStreamer, FingerprintCalculator, PQEncoder, PQKMeans

# 1. Stream SMILES in chunks
streamer = DataStreamer(path='molecules.smi', chunksize=100000)

# 2. Calculate MQN fingerprints
fp_calc = FingerprintCalculator()
for smiles_chunk in streamer.parse_input():
    fingerprints = fp_calc.FingerprintFromSmiles(smiles_chunk, fp='mqn')
    # Save fingerprints...

# 3. Train PQ encoder on sample
encoder = PQEncoder(k=256, m=6, iterations=20)
encoder.fit(training_fingerprints)

# 4. Transform all fingerprints to PQ codes
pq_codes = encoder.transform(fingerprints)

# 5. Cluster with PQk-means
clusterer = PQKMeans(encoder, k=100000)
labels = clusterer.fit_predict(pq_codes)

Project Structure

chelombus/
├── chelombus/
│   ├── encoder/          # Product Quantization encoder
│   ├── clustering/       # PQk-means wrapper
│   ├── streamer/         # Memory-efficient data streaming
│   └── utils/            # Fingerprints, visualization, helpers
├── scripts/              # Pipeline scripts
├── examples/             # Tutorial notebooks
└── tests/                # Unit tests

Choosing k (Number of Clusters)

The scripts/select_k.py script sweeps over k values on a subsample to help pick the right number of clusters. It supports checkpointing, if interrupted, rerun the same command and it resumes from where it left off.

python scripts/select_k.py \
    --pq-codes data/pq_codes.npy \
    --encoder models/encoder.joblib \
    --n-subsample 10000000 \
    --k-values 10000 25000 50000 100000 200000 \
    --iterations 10 \
    --output results/k_selection.csv \
    --plot results/k_selection.png

Results on 100M Enamine REAL molecules (RTX 4070 Ti SUPER 16GB, AMD Ryzen 7 64GB RAM):

Guidelines:

• k = 50,000 is a good default — under 20% empty clusters, median size ~1,900, and the avg distance improvement starts plateauing beyond this point.
• k = 100,000 if you need tighter clusters and can tolerate ~27% empty clusters.
• Beyond 200K, over a third of clusters are empty — diminishing returns.

Documentation

• Full docs: https://chelombus.gdb.tools
• Tutorial: See examples/tutorial.ipynb for a hands-on introduction
• Large-scale example: See examples/enamine_1B_clustering.ipynb
• API Reference: See docs/api.md or the hosted docs

Testing

# Run all tests
pytest tests/

# Run specific test file
pytest tests/test_encoder.py -v

Citation

If you use Chelombus in your research, please cite:

@article{chelombus2025,
  title={Nested TMAPs to visualize Billions of Molecules},
  author={Flores Sepulveda, Alejandro and Reymond, Jean-Louis},
  journal={},
  year={2025}
}

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch: git checkout -b feature/my-feature
3. Write tests for new functionality
4. Submit a pull request

License

MIT License. See LICENSE for details.

Acknowledgments

• PQk-means by Matsui et al.
• TMAP by Probst & Reymond
• RDKit for cheminformatics functionality
• Swiss National Science Foundation (grant no. 200020_178998)