Molfeat

Molfeat - Molecular Featurization Hub

Overview

Molfeat turns molecules (SMILES strings or RDKit/datamol Mol objects) into numerical representations for machine learning: fingerprints, descriptors, pharmacophores, shape descriptors, and pretrained neural embeddings, all behind one scikit-learn-compatible transformer interface with state serialization and caching.

Current baseline (verified 2026-08-16): molfeat 0.11.0 (May 2025) is still the latest PyPI and GitHub release; the repository has had no commits since. All examples in this skill were executed against 0.11.0 on Python 3.10 with datamol 0.12.5, RDKit 2026.03.5, numpy 2.2.6 and torch 2.13.0. Python 3.11+ is not installable (requires-python = ">=3.9,=1.1.1, from the store!** — the HuggingFace-artifact download bug above. Load from the HF Hub with HFModel.from_pretrained.

A fingerprint parameter had no effect — FPCalculator logs unknown parameters as an error and then drops them; construction still succeeds. FPCalculator("ecfp", n_bits=1024) returns 2048 dimensions (the key is fpSize). Check len(calc) after passing parameters.

ImportError: Cannot import map4 — MAP4 is external; install it from the reymond-group repository.

Package will not install — check the interpreter first: molfeat 0.11.0 is capped below Python 3.11 and pip/uv will refuse to resolve on 3.11+.

3D featurizers return errors or zeros — desc3D, usr, usrcat, electroshape, cats3D and pharm3D need conformers. Generate them first with dm.conformers.generate(mol, n_confs=1) and pass Mol objects, not SMILES.

Reproducibility — save to_state_yaml_file next to the model, and record molfeat.__version__; state files carry the writing version in _molfeat_version.

Composing with the rest of the bundle

• rdkit / datamol → before: standardise and desalt first. A featurizer embeds whatever

string it is given, so a salt or mixture produces a vector for the wrong species — and no error.

• chembl → before: curated measured bioactivity is what you want to featurize, not raw rows.
• pytdc → alongside: the scaffold and cold-start splits. Featurization quality is invisible under

a random split, which reports a fantasy R² regardless of the representation you chose.

• deepchem → after: model fitting, if you want the training loop rather than just the features.
• admet-prediction → instead: for standard ADMET endpoints, a ready-made model beats featurizing

and training from scratch unless you have your own measured data.

• chemical-space / generative-design → after: features are what a similarity or diversity

selection over an enumerated set is computed on.

Try ECFP first. Across most QSAR tasks a count-based Morgan fingerprint with a gradient-boosted model is within noise of a pretrained transformer embedding, at a fraction of the cost. Reach for pretrained embeddings when you have shown ECFP is the bottleneck, not before.

Additional Resources

• Official documentation: https://molfeat-docs.datamol.io/
• GitHub repository: https://github.com/datamol-io/molfeat
• PyPI package: https://pypi.org/project/molfeat/
• Tutorial: https://portal.valencelabs.com/datamol/post/types-of-featurizers-b1e8HHrbFMkbun6

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: K-Dense-AI
• Source: K-Dense-AI/drug-discovery-agent-skills
• License: MIT
• Homepage: www.k-dense.ai

Install and usage instructions live in the source repository linked above.