Finding a new drug or material is mostly an exercise in elimination: chemists synthesize and test enormous numbers of molecules to find the few that behave the way they need. A machine-learning model from MIT called CheMeleon aims to shrink that pile by predicting molecular properties from very little experimental data.
According to MIT News (November 6, 2026), the work was reported in the Journal of Chemical Information and Modeling and comes from the group of chemical engineering professor William Green, with researchers from the chemical company BASF.
The problem: too little data
MIT notes that discovering molecules for specific uses drives progress in pharmaceuticals, electronics and energy, yet it remains largely trial and error because no one can perfectly predict how a compound will behave. AI has helped speed up early screening, but most deep-learning models need a lot of data. In practice, researchers often have only a handful of measurements for a newly discovered drug target, which is where conventional approaches struggle.
How CheMeleon is trained
CheMeleon is a pre-trained foundation model, inspired by the large language models behind chatbots. Instead of reading scientific literature, it learns from molecular descriptors: numbers that chemists have developed over decades, such as molecular weight or the number of oxygen atoms. The team calculated about 1,600 of these descriptors for roughly one million molecules drawn from PubChem, the free chemical database maintained by the U.S. National Institutes of Health.
The pre-training task is to predict those descriptors from molecular structure. In doing so, the model picks up general correlations between how a molecule is built and how it behaves. Graduate student Akshat Shirish Zalte explained that the team is not training it for one property but building a general model that understands some level of chemistry, so that fine-tuning on just a few data points still gives good performance, because it does not have to learn everything from scratch.
What the tests showed
The researchers compared CheMeleon against other foundation models and against classical models that rely on expert-designed principles. Across 58 test cases, MIT reports that CheMeleon came out on top. It can be fine-tuned for familiar properties such as solubility and viscosity, and for more specialized ones such as bioactivity or reactivity with a particular protein or enzyme. Green says the model drastically reduces how many molecules must be tested experimentally to predict the properties of new ones, which matters because synthesis and testing are slow and expensive.
Already in use before publication
The open-source model was built with ChemProp, a machine-learning package also maintained by the Green group, and anyone can download CheMeleon through ChemProp. Its influence spread even before peer review finished. In a contest organized by OpenADMET this year, which asked entrants to predict properties of molecules that had not yet been synthesized, nine of the top 20 entries used CheMeleon, according to Green. Several pharmaceutical companies now use it regularly, and Green says at least one told its staff to start with CheMeleon whenever modeling a property.
What it means
For drug and materials teams, the appeal is cost. Every molecule that does not need to be made and measured saves lab time, and a model that works with a few data points is especially useful for new targets where little is known. Because the code and weights are open, smaller labs and startups can adopt the same approach without building a foundation model themselves. As with any prediction tool, its outputs are guidance for deciding what to test, not a replacement for experiments.
Bottom line
CheMeleon shows that a model pre-trained on cheap, computable chemistry descriptors can punch above its weight when real data is scarce. Its quick adoption by competition entrants and pharma teams suggests low-data chemistry is a place where foundation models can pay off soon.