bblean 0.6.0b2

BitBirch-Lean Python package

Overview

BitBIRCH-Lean is a high-throughput implementation of the BitBIRCH clustering algorithm designed for very large molecular libraries.

If you find this software useful please cite the following articles:

• BitBIRCH: efficient clustering of large molecular libraries: https://doi.org/10.1039/D5DD00030K
• BitBIRCH Clustering Refinement Strategies: https://doi.org/10.1021/acs.jcim.5c00627
• BitBIRCH-Lean: (preprint) https://www.biorxiv.org/content/10.1101/2025.10.22.684015v1

NOTE: BitBirch-Lean is currently beta software, expect minor breaking changes until we hit version 1.0

The documentation of the developer version is a work in progress. Please let us know if you find any issues.

⚠️ Important: The default threshold is 0.3 and the default fingerprint kind to ecfp4. We recommend setting threshold to 0.5-0.65 for rdkit fingerprints and 0.3-0.4 for ecfp4 or ecfp6 fingerprints (although you may need further tuning for your specific library / fingerprint set). For more information on tuning these parameters see the best practices and parameter tuning guides.

Installation

BitBIRCH-Lean requires Python 3.11 or higher, and can be installed in Windows, Linux or macOS via pip, which automatically includes C++ extensions:

pip install bblean
bb --help

We recommend installing bblean in a conda environment or a venv.

Memory usage and C++ extensions are most optimized for Linux / macOS. We support windows on a best-effort basis, some releases may not have Windows support.

From source

To build from source instead (editable mode):

git clone [e-mail]:mqcomplab/bblean,
cd bblean

conda env create --file ./environment.yaml
conda activate bblean

BITBIRCH_BUILD_CPP=1 pip install -e .

# If you want to build without the C++ extensions run this instead:
pip install -e .

bb --help

If the extensions install successfully, they will be automatically used each time BitBirch-Lean or its classes are used. No need to do anything else.

If you run into any issues when installing the extensions, please open a GitHub issue and tag it with C++.

CLI Quickstart

BitBIRCH-Lean provides a convenient CLI interface, bb. The CLI can be used to convert SMILES files into compact fingerprint arrays, and cluster them in parallel or serial mode with a single command, making it straightforward to triage collections with millions of molecules. The CLI prints a run banner with the parameters used, memory usage (when available), and elapsed timings so you can track each job at a glance.

The most important commands you need are:

• bb fps-from-smiles: Generate fingerprints from a *.smi file.
• bb run or bb multiround: Cluster the fingerprints
• bb plot-summary or bb plot-tsne: Analyze the clusters

An example usual workflow is as follows:

1. Generate fingerprints from SMILES: The repository ships with a ChEMBL sample that you can use right away for testing:

   bb fps-from-smiles examples/chembl-33-natural-products-sample.smi
   

   This writes a packed fingerprint array to the current working directory (use --out-dir <dir> for a different location). The naming convention is packed-fps-uint8-508e53ef.npy, where 508e53ef is a unique identifier (use --name <name> if you prefer a different name). The packed uint8 format is required for maximum memory-efficient, so keep the default --pack and --dtype values unless you have a very good reason to change them. You can optionally split over multiple files for parallel parallel processing with --num-parts <num>.

2. Cluster the fingerprints: To cluster in serial mode, point bb run at the generated array (or a directory with multiple *.npy files):

   bb run ./packed-fps-uint8-508e53ef.npy
   

   The outputs are stored in directory such as bb_run_outputs/504e40ef/, where 504e40ef is a unique identifier (use --out-dir <dir> for a different location). Additional flags can be set to control the BitBIRCH --branching, --threshold, and merge criterion. Optionally, cluster refinement can be performed with --refine-num 1. bb run --help for details.

   To cluster in parallel mode, use bb multiround ./file-or-dir instead. If pointed to a directory with multiple *.npy files, files will be clustered in parallel and sub-trees will be merged iteratively in intermediate rounds. For more information: bb multiround --help. Outputs are written by default to bb_multiround_outputs/<unique-id>/.

3. Visualize the results: You can plot a summary of the largest clusters with bb plot-summary <output-path> --top 20 (largest 20 clusters). Passing the optional --smiles <path-to-file.smi> argument additionally generates Murcko scaffold analysis. For a t-SNE visualization try bb plot-tsne <output-path> -- top 20. t-SNE plots use openTSNE as a backend, which is a parallel, extremely fast implementation. We recommend you consult the corresponding documentation for info on the available parameters. Still, expect t-SNE plots to be slow for very large datasets (more than 1M molecules).

Manually exploring clustering results

Every run directory contains a raw clusters.pkl file with the molecule indices for each cluster, plus metadata in *.json files that captures the exact settings and performance characteristics. A quick Python session is all you need to get started:

import pickle

clusters = pickle.load(open("bb_run_outputs/504e40ef/clusters.pkl", "rb"))
clusters[:2]
# [[321, 323, 326, 328, 337, ..., 9988, 9989],
#  [5914, 5915, 5916, 5917, 5918, ..., 9990, 9991, 9992, 9993]]

The indices refer to the position of each molecule in the order they were read from the fingerprint files, making it easy to link back to your original SMILES records.

Python Quickstart and Examples

For an example of how to use the main bblean classes and functions consult examples/bitbirch_quickstart.ipynb. The examples/dataset_splitting.ipynb notebook contains an adapted notebook by Pat Walters (Some Thoughts on Splitting Chemical Datasets). More examples will be added soon!

A quick summary:

import pickle

import matplotlib.pyplot as plt
import numpy as np

import bblean
import bblean.plotting as plotting
import bblean.analysis as analysis

# Create the fingerprints and pack them into a numpy array, starting from a *.smi file
smiles = bblean.load_smiles("./examples/chembl-33-natural-products-sample.smi")
fps = bblean.fps_from_smiles(smiles, pack=True, n_features=2048, kind="rdkit")

# Fit the figerprints (by default all bblean functions take *packed* fingerprints)
# A threhsold of 0.5-0.65 is good for rdkit fingerprints, a threshold of 0.3-0.4
# is better for ECFPs
tree = bblean.BitBirch(branching_factor=50, threshold=0.65, merge_criterion="diameter")
tree.fit(fps)

# Refine the tree (if needed)
tree.set_merge(merge_criterion="tolerance-diameter", tolerance=0.0)
tree.refine_inplace(fps)

# Visualize the results
clusters = tree.get_cluster_mol_ids()
ca = analysis.cluster_analysis(clusters, fps, smiles)
plotting.summary_plot(ca, title="ChEMBL Sample")
plt.show()

# Save the resulting clusters, metrics, and fps
with open("./clusters.pkl", "wb") as f:
    pickle.dump(clusters, f)
ca.dump_metrics("./metrics.csv")
np.save("./fps-packed-2048.npy", fps)

Public Python API and Documentation

By default all functions take packed fingerprints of dtype uint8. Many functions support an input_is_packed: bool flag, which you can toggle to False in case for some reason you want to pass unpacked fingerprints (not recommended).

• Functions and classes that end in an underscore are considered private (such as _private_function(...)) and should not be used, since they can be removed or modified without warning.
• All functions and classes that are in modules that end with an underscore are also considered private (such as bblean._private_module.private_function(...)) and should not be used, since they can be removed or modified without warning.
• All other functions and classes are part of the stable public API and can be used. However, expect minor breaking changes before we hit version 1.0

Contributing

If you find a bug in BitBIRCH-Lean or have an issue with the usage or documentation please open an issue in the GitHub issue tracker.

If you want to contribute to BitBIRCH-Lean with a bug fix, improving the documentation, with usability, maintainability, or performance, please open an issue with your idea/request (or directly open a PR from a fork if you prefer).

Currently we don't directly accept PRs with new features that have not been extensively validated, but if you have an idea to improve the BitBIRCH algorithm you may want to contact the Miranda-Quintana Lab, we are open to collaborations.

To contribute, first create a fork, then clone your fork (git clone [e-mail]:<user>/bblean. We recommend you install pre-commit (pre-commit install --hook-type pre-push), which will run some checks before you push to your branch. After you have finished work on your branch, open a pull request.