Drug and materials discovery runs into the same wall again and again: for the property you care about most, there is hardly any measured data. Deep-learning models usually need more data than traditional methods, not less. CheMeleon, an open model from MIT and the chemical company BASF, is built for that gap, and it was in industrial use before its paper cleared peer review.
The idea is to pre-train on something cheap and plentiful. Instead of mining the chemistry literature, the team took about a million molecules from PubChem, the free database run by the US National Institutes of Health, and computed roughly 1,600 standard descriptors for each, values chemists have defined over decades such as molecular weight, the number of oxygen atoms or how branched a molecule is. The model learns to predict those descriptors from structure, which teaches it a general sense of chemistry. It can then be fine-tuned with a handful of measurements for a specific target, from solubility and viscosity to binding to a particular protein, without learning everything from scratch.
The results reported so far are strong. Across 58 test cases, the authors say CheMeleon came out on top against both other foundation models and classical expert-designed models. In a large OpenADMET contest earlier this year to predict properties of molecules not yet synthesised, nine of the top 20 entries used it. According to William Green, the MIT professor who leads the group, at least one pharmaceutical company then told its staff to start every property model with CheMeleon. The paper is now published in the Journal of Chemical Information and Modeling, and the model can be downloaded and fine-tuned through ChemProp, the group's open-source package.
The approach is a neat inversion of the usual foundation-model story. Language models such as the ones behind ChatGPT and Claude learn from vast amounts of human text; CheMeleon learns from numbers a computer can calculate for free, which sidesteps the scarcity of experimental data rather than waiting for it to be solved.
Some caution is due. The evidence here comes through a university press release, which quotes the authors' summary of the benchmark rather than giving error figures. Descriptors capture what chemists already know how to compute, so a property driven by something they miss may still need real experiments. And a contest win shows usefulness for screening which molecules to make next; it does not replace testing the ones that matter.