scikit-mol

Scikit-Learn classes for molecular vectorization using RDKit

The intended usage is to be able to add molecular vectorization directly into scikit-learn pipelines, so that the final model directly predict on RDKit molecules or SMILES strings

As example with the needed scikit-learn and -mol imports and RDKit mol objects in the mol_list_train and _test lists:

pipe = Pipeline([('mol_transformer', MorganFingerprintTransformer()), ('Regressor', Ridge())])
pipe.fit(mol_list_train, y_train)
pipe.score(mol_list_test, y_test)
pipe.predict([Chem.MolFromSmiles('c1ccccc1C(=O)C')])

>>> array([4.93858815])

The scikit-learn compatibility should also make it easier to include the fingerprinting step in hyperparameter tuning with scikit-learns utilities

The first draft for the project was created at the RDKIT UGM 2022 hackathon 2022-October-14

Implemented

• Descriptors
  • MolecularDescriptorTransformer

• Fingerprints
  • MorganFingerprintTransformer
  • MACCSKeysFingerprintTransformer
  • RDKitFingerprintTransformer
  • AtomPairFingerprintTransformer
  • TopologicalTorsionFingerprintTransformer
  • MHFingerprintTransformer
  • SECFingerprintTransformer
  • AvalonFingerprintTransformer

• Conversions
  • SmilesToMol

• Standardizer
  • Standardizer

- safeinference - SafeInferenceWrapper - set_safe_inference_mode

• Utilities
  • CheckSmilesSanitazion

Installation

Users can install latest tagged release from pip

pip install scikit-mol

or from conda-forge

conda install -c conda-forge scikit-mol

The conda forge package should get updated shortly after a new tagged release on pypi.

Bleeding edge

pip install git+https://github.com:EBjerrum/scikit-mol.git

Documentation

There are a collection of notebooks in the notebooks directory which demonstrates some different aspects and use cases

• Basic Usage and fingerprint transformers

• Descriptor transformer

• Pipelining with Scikit-Learn classes

• Molecular standardization

• Sanitizing SMILES input

• Integrated hyperparameter tuning of Scikit-Learn estimator and Scikit-Mol transformer

• Using parallel execution to speed up descriptor and fingerprint calculations

• Using skopt for hyperparameter tuning

• Testing different fingerprints as part of the hyperparameter optimization

• Using pandas output for easy feature importance analysis and combine pre-exisitng values with new computations

• Working with pipelines and estimators in safe inference mode for handling prediction on batches with invalid smiles or molecules

  We also put a software note on ChemRxiv. https://doi.org/10.26434/chemrxiv-2023-fzqwd

Roadmap and Contributing

Help wanted! Are you a PhD student that want a "side-quest" to procrastinate your thesis writing or are you simply interested in computational chemistry, cheminformatics or simply with an interest in QSAR modelling, Python Programming open-source software? Do you want to learn more about machine learning with Scikit-Learn? Or do you use scikit-mol for your current work and would like to pay a little back to the project and see it improved as well? With a little bit of help, this project can be improved much faster! Reach to me (Esben), for a discussion about how we can proceed.

Currently we are working on fixing some deprecation warnings, its not the most exciting work, but it's important to maintain a little. Later on we need to go over the scikit-learn compatibility and update to some of their newer features on their estimator classes. We're also brewing on some feature enhancements and tests, such as new fingerprints and a more versatile standardizer.

There are more information about how to contribute to the project in CONTRIBUTION.md

BUGS

Probably still, please check issues at GitHub and report there

Contributers:

• Esben Jannik Bjerrum @ebjerrum, [email protected]
• Carmen Esposito @cespos
• Son Ha, [email protected]
• Oh-hyeon Choung, [email protected]
• Andreas Poehlmann, @ap--
• Ya Chen, @anya-chen
• Rafał Bachorz @rafalbachorz
• Adrien Chaton @adrienchaton
• @VincentAlexanderScholz
• @RiesBen
• @enricogandini
• @mikemhenry
• @c-feldmann