Chalcedon

Fast, memory-efficient Butina clustering and train/validation/test splitting for chemical datasets. Use this package to minimize data leakage when splitting chemical data to improve the evaluation and generalizability of your models.

Installation

uv pip install chalcedon

Quick start

Recommended

For the recommended case, run directly from SMILES. Chalcedon computes Morgan fingerprints (radius 2, 2048 bits) internally and clusters in float32:

import chalcedon

smiles = [
    "CCO",
    "c1ccccc1",
    # ...your dataset
]

splits = chalcedon.butina_split(
    smiles,
    fractions={"train": 0.8, "val": 0.1, "test": 0.1},
    cutoff=0.65,
    dtype="float32" # or np.float32
)

train_smiles = splits["train"]
val_smiles = splits["val"]
test_smiles = splits["test"]

Using custom descriptors

We recommend dtype="float64" for non-binary descriptors, where dot-product magnitudes can exceed float32's exact range.

import chalcedon

descriptors = my_descriptor_generator(molecules)  # numpy.ndarray of shape (n, d)

cluster_ids = chalcedon.butina_cluster(descriptors, cutoff=0.65, dtype="float64")
splits = chalcedon.greedy_cluster_split(
    cluster_ids,
    fractions={"train": 0.8, "val": 0.1, "test": 0.1},
)

train_indices = splits["train"]  # numpy.ndarray of indices into `descriptors`

pairwise_tanimoto(fingerprints) is also exposed if you want just the similarity matrix.

Benchmarks

Chalcedon can quickly create Butina clusters of large chemical datasets on consumer hardware with near linear memory scaling.

See benchmarks/report.md for a detailed analysis of algorithm performance and benchmarks/ to reproduce results.

Citation

If you use Chalcedon in your research, please cite:

@software{chalcedon,
  title = {Chalcedon: Clustering and dataset splitting for chemical data.},
  year = {2026},
  url = {https://github.com/rowansci/chalcedon}
}

Acknowledgements

• RDKit for cheminformatics infrastructure and the CrystalFF torsion library (Riniker & Landrum, J. Chem. Inf. Model. 56, 2016)
• GEOM dataset for the benchmark SMILES (Axelrod & Gomez-Bombarelli, Sci Data 9, 185, 2022)

This package was created with Cookiecutter and the jevandezande/uv-cookiecutter project template.