Chemprop
Open-source message-passing neural network for predicting molecular properties — widely used for ADME/Tox prediction, activity modeling, and property-guided screening in drug discovery and materials chemistry.
What it does
Chemprop is an open-source Python library implementing directed message-passing neural networks (D-MPNNs) for molecular property prediction. You provide a set of molecules (as SMILES strings) with measured properties, train a model on that data, and the model predicts those properties for new, unmeasured molecules. It learns directly from the molecular graph — atoms as nodes, bonds as edges — without requiring hand-crafted molecular fingerprints.
Originally developed at MIT (Yang et al., 2019, Journal of Chemical Information and Modeling), Chemprop has become one of the most widely adopted open-source tools for molecular property prediction, particularly in pharmaceutical and agricultural chemistry contexts where ADME/Tox profiles (absorption, distribution, metabolism, excretion, and toxicity) need to be predicted for large compound sets.
Best for
Predicting continuous (e.g., binding affinity, solubility, logP) or categorical (e.g., active/inactive, toxic/non-toxic) properties of molecules from a training set of experimental measurements. Particularly well-suited to: ADME/Tox modeling, selectivity profiling, potency prediction for lead optimization, and property-guided screening of virtual compound libraries.
Pricing
Free and open-source (MIT license). Available on GitHub and via pip install chemprop.
Strengths
- Learns molecular representations from the graph directly — typically outperforms classical fingerprint-based models (random forests, SVMs) on benchmark property prediction tasks
- Handles multi-task learning natively: train one model to predict multiple properties simultaneously, which often improves per-task performance via transfer
- Uncertainty estimation built in: predict confidence intervals alongside predictions, useful for identifying molecules where the model is extrapolating outside training data
- Well-documented, actively maintained, and widely used — substantial community of users means issues and edge cases are documented
- Works with data sizes from hundreds to tens of thousands of molecules
Limitations
- Requires labeled training data for your specific property — if you have no experimental measurements, there’s nothing to train on. Transfer learning from pre-trained models helps but doesn’t eliminate this requirement.
- Performance degrades for molecules structurally unlike anything in the training set (out-of-distribution generalization) — common in early-stage drug discovery when exploring novel scaffolds
- Training time scales with dataset size; very large datasets (>100,000 molecules) may need GPU resources for reasonable training times
- A well-trained model is only as good as the quality and coverage of its training data — measurement noise, assay variability, and selection bias in training sets all limit performance
How it compares
| vs. | Key difference |
|---|---|
| Molecular fingerprint + random forest | Chemprop’s graph-based representation typically outperforms fingerprint + RF on benchmark tasks, especially for activity prediction; RF is faster to train and easier to interpret |
| IBM RXN | IBM RXN predicts reaction outcomes and plans synthetic routes; Chemprop predicts molecular properties — different tasks, often used together in a discovery pipeline |
| AlphaFold | AlphaFold predicts protein structure; Chemprop predicts small-molecule properties — different domains, complementary in structure-based drug design workflows |
| Commercial ADME/Tox platforms (StarDrop, ADMET Predictor) | Commercial tools offer broader property coverage and polished interfaces; Chemprop is open-source, customizable, and trainable on your own proprietary data |
Related content
- Field Guide: Chemistry
- Glossary: Molecular Fingerprint, SMILES Notation, Virtual Screening