molfeat - the hub for all your molecular featurizers

Docs | Homepage

---

Molfeat is a hub of molecular featurizers. It supports a wide variety of out-of-the-box molecular featurizers and can be easily extended to include your own custom featurizers.

• 🚀 Fast, with a simple and efficient API.
• 🔄 Unify pre-trained molecular embeddings and hand-crafted featurizers in a single package.
• ➕ Easily add your own featurizers through plugins.
• 📈 Benefit from increased performance through a trouble-free caching system.

Visit our website at https://molfeat.datamol.io.

Updates

Molfeat 1.x narrows the project to small-molecule featurization and a modern, maintainable stack. It removes the legacy DGL, DGLLife, Graphormer and protein adapters, keeps PyTorch Geometric as the graph backend, and adds maintained foundation-model integrations including CheMeleon and Mol-JEPA. Model loading is lazy, external checkpoint licences are explicit, and official representations are covered by integration tests.

HTTP model downloads use the required Datamol 0.13 series; they no longer require an unreleased dependency. SELFIES support is explicit, and InChI decoding now returns SMILES instead of silently failing.

See the complete changelog and the 1.x migration guide for the complete 1.0 transition.

Release maintainers: see the manual release guide.

Installation

Installing Molfeat

Add Molfeat to a uv-managed project:

uv add molfeat

Pip and conda-forge remain supported: python -m pip install molfeat or mamba install -c conda-forge molfeat.

The next major release requires Python 3.11 or newer and RDKit 2024.09 or newer. PyTorch 2.5 or newer is used on maintained platforms; macOS Intel uses the final available 2.2 wheel series with NumPy 1.26 and is tested on Python 3.12. On Mac Intel, the transformer extra selects Transformers 4.57 rather than 5. Use safetensors checkpoints there: current Transformers blocks pickle-based weight loading with PyTorch below 2.6. SAFE's notation core can share this environment, but its Transformers 5 model/training extras cannot.

Optional dependencies

Not all featurizers in the Molfeat core package are supported by default. Some featurizers require additional dependencies. If you try to use a featurizer that requires additional dependencies, Molfeat will raise an error and tell you which dependencies are missing and how to install them.

• To install Hugging Face Transformers support: python -m pip install "molfeat[transformer]".
• To convert SELFIES without Transformers: python -m pip install "molfeat[selfies]".
• To install PyTorch Geometric support: python -m pip install "molfeat[pyg]".
• To install FCD support: python -m pip install "molfeat[fcd]".
• To install HDF5 and Parquet cache support: python -m pip install "molfeat[cache]".
• To install S3 and Google Cloud model stores: python -m pip install "molfeat[cloud]".
• To install Mordred descriptors: python -m pip install "molfeat[mordred]".
• To install map4: see https://github.com/reymond-group/map4

python -m pip install "molfeat[all]" installs every maintained optional dependency compatible with the current interpreter. DGL, DGLLife and the legacy Graphormer adapter have been removed: their binary and build constraints conflict with the supported modern stack. PyTorch Geometric is the maintained graph backend. Protein featurizers have also been removed so Molfeat 1.x has a precise small-molecule scope. See the migration guide for details.

Compatibility

Installing Plugins

The functionality of Molfeat can be extended through plugins. The use of a plugin system ensures that the core package remains easy to install and as light as possible, while making it easy to extend its functionality with plug-and-play components. Additionally, it ensures that plugins can be developed independently from the core package, removing the bottleneck of a central party that reviews and approves new plugins. Consult the molfeat documentation for more details on how to create your own plugins.

However, this does imply that the installation of a plugin is plugin-dependent: please consult the relevant documentation to learn more.

API tour

import datamol as dm
from molfeat.calc import FPCalculator
from molfeat.trans import MoleculeTransformer
from molfeat.store.modelstore import ModelStore

# Load some dummy data
data = dm.data.freesolv().sample(100).smiles.values

# Featurize a single molecule
calc = FPCalculator("ecfp")
calc(data[0])

# Define a parallelized featurization pipeline
mol_transf = MoleculeTransformer(calc, n_jobs=-1)
mol_transf(data)

# Easily save and load featurizers
mol_transf.to_state_yaml_file("state_dict.yml")
mol_transf = MoleculeTransformer.from_state_yaml_file("state_dict.yml")
mol_transf(data)

# List all available featurizers
store = ModelStore()
store.available_models

# Find a featurizer and learn how to use it
model_card = store.search(name="ChemBERTa-77M-MLM")[0]
model_card.usage()

How to cite

Please cite Molfeat if you use it in your research: .

Contribute

See developers for a comprehensive guide on how to contribute to molfeat. molfeat is a community-led initiative and whether you're a first-time contributor or an open-source veteran, this project greatly benefits from your contributions. To learn more about the community and datamol.io ecosystem, please see community.

Maintainers

• @cwognum
• @maclandrol
• @hadim

License

Molfeat's code is licensed under Apache-2.0. See LICENSE. The CheMeleon adapter retains the Chemprop MIT notice. Distribution metadata records both as Apache-2.0 AND MIT; this is not a choice of licence for Molfeat's own code. Downloaded model weights and custom Hub code have their own terms; Mol-JEPA requires explicit acceptance of its non-commercial licence. See the model licensing policy.