MIT and BASF's CheMeleon Predicts Molecular Behavior From a Handful of Data Points, and 9 of OpenADMET's Top 20 Entries Used It

A foundation model built by MIT and BASF can predict how a new molecule will behave, whether that means potency, toxicity or solubility, from a handful of measured data points, the kind of gap that normally stops a drug program before it starts. The work appeared in the September 14 issue of the Journal of Chemical Information and Modeling, and MIT's news office highlighted it this week. The model, CheMeleon, is open source and already in routine use inside pharmaceutical companies.

The bottleneck it attacks is data, not compute. Most deep-learning models for molecular property prediction need large labeled sets for each property, and those sets do not exist for a newly discovered cellular target or an unusual material. CheMeleon instead learns chemistry first: the team trained a network on about one million molecules drawn from PubChem, asking it to predict roughly 1,600 chemists' descriptors, quantities such as molecular weight, oxygen count, or how branched and bulky a structure is, from the structure alone. No literature mining was involved. "We do not look at the literature at all," said Akshat Shirish Zalte, the graduate student who led the work.

That pretraining produces a general sense of how shape relates to behavior. A user then fine-tunes the model on the property they care about, and it needs far fewer examples to reach useful accuracy. "Even if you fine-tune it with just a few data points for the task you care about, you can still get some good performance," said William Green, the Hoyt Hottel Professor of Chemical Engineering, because the model is not learning everything from scratch.

The team tested it on 58 cases and reports that CheMeleon beat both other foundation models and classical, expert-designed baselines. The stronger signal came from outside the lab. In an OpenADMET contest that asked entrants to predict properties of molecules that had never been synthesized, nine of the top 20 entries used CheMeleon.

Adoption moved faster than peer review. Green said inquiries from pharmaceutical companies arrived while the paper was still under review, and at least one drugmaker told staff that any property modeling should begin with CheMeleon. The authors are Zalte, recent graduate Jackson Burns, research scientist Charlles Abreu and Green, with Jochen Sieg, Christian Feldmann and Miriam Mathea of BASF.

The limits are real. The 58-case comparison is the researchers' own test set, accuracy depends on how closely a target property resembles what those descriptors capture, and every new target still needs some labeled data to fine-tune on. CheMeleon narrows the list of compounds worth synthesizing. It does not replace the lab, and the group says so.

For chemical and materials companies, the practical value is a shared starting point rather than a proprietary edge: the model ships through ChemProp, the open-source package Green's group maintains, so a small team can fine-tune it without building a data pipeline of its own. That is the pattern language models followed, a general pretrained base followed by narrow adaptation, arriving in a field where experiments rather than tokens are the expensive part.