What happened
An MIT team developed CheMeleon, an open model that predicts molecular properties when little data is available. The study, with three researchers from the chemical company BASF, was published in the Journal of Chemical Information and Modeling. A preliminary version has circulated since last year and several pharmaceutical companies already use it.
Why it matters
Finding molecules for drugs, electronics or energy is still largely trial and error, because there is no way to predict exactly how a compound will behave. Machine learning helps decide what is worth synthesizing and testing. CheMeleon’s contribution is that it works better precisely where data are lacking, for example with a newly discovered cellular target for a possible drug.
The number
In a contest organized by OpenADMET to predict properties of new molecules, nine of the top twenty entries used CheMeleon.
Context
The model was built on ChemProp, an open-source package from the group of William Green, professor of chemical engineering at MIT, and anyone can download it through ChemProp and tune it to their needs. Green said that after the contest they began receiving inquiries from many places. The authors are student Akshat Shirish Zalte, recent PhD Jackson Burns, researcher Charlles Abreu and Green, together with three BASF scientists. According to the text, the method outperforms existing ones at narrowing the field of molecules to review when data are scarce. On related topics, Mamífero published An open LLM turns zeolite recipes into lab steps; MOSS-Image: open model for long text in images; Xiaomi opens MiMo-V2.6 and beats Opus 5 on two tests and Blood molecules from a 115-million-year-old dinosaur.
What’s next
What comes next depends on use: being open, other groups will be able to tune it to their own properties of interest and compare it with alternatives. Its performance will be judged on molecules synthesized and measured in the lab.
Bottom line
It is a filtering tool: it speeds up the choice of candidates, but verification still happens in the laboratory.
Sources
- Phys.org, summary of the MIT study published in the Journal of Chemical Information and Modeling (accessed November 9, 2026).
- Cover image: Belikov Maxim, CC BY 4.0, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
