MIT and BASF's open-source CheMeleon model predicts molecular properties from small datasets

Researchers at MIT and the German chemical company BASF have described CheMeleon, an open-source model for predicting the properties of molecules when only a small amount of experimental data is available for a given task. MIT publicized the work on November 6. The underlying study appeared in the Journal of Chemical Information and Modeling, online on August 7 and in the September 14 issue. A preprint had been on arXiv since June 2025.

How the model works

Machine learning models for chemistry often struggle because measured data, such as toxicity or solubility results, are expensive to produce and usually come in small sets. CheMeleon takes a different route to pretraining. Rather than learning from noisy experimental measurements or from quantum-chemistry simulations, it was trained on roughly one million molecules from the PubChem database to predict about 1,600 molecular descriptors, numerical features that software computes directly from a structure. According to the project's code repository, the descriptors came from the Mordred package. Computing them took about 500 CPU hours, and pretraining took about 1,000 GPU hours.

The resulting model has about 10 million parameters and is built on directed message-passing neural networks. Users then fine-tune it on their own small dataset. In the paper, the authors tested it on 58 datasets from the Polaris and MoleculeACE benchmarks. On Polaris tasks they report a 75% win rate, against 68% for Random Forest, 36% for fastprop and 32% for Chemprop. On MoleculeACE assays they report 97%, against 50% for Random Forest.

Adoption and availability

William Green, a chemical engineering professor at MIT who leads the group, said nine of the top 20 entries in an OpenADMET contest held earlier this year used CheMeleon to predict properties of molecules that had not yet been synthesized. The figure comes from Green in MIT's announcement; OpenADMET's own results were not checked for this article. The team also includes MIT graduate student Akshat Shirish Zalte and BASF scientists Jochen Sieg, Christian Feldmann and Miriam Mathea.

The code is released under the MIT license, and the weights are archived on Zenodo. Chemprop, the software maintained by Green's group, can load the model with a single command-line flag. Chemprop's documentation says training takes longer per epoch because the model is larger. The developers present CheMeleon as a way to narrow down candidate molecules, not to replace experiments. Promising compounds still need to be made and tested in the lab.