Nine of the top 20 entries in a recent industry contest tasked with predicting the properties of previously unsynthesized molecules shared a common engine: CheMeleon. Developed by researchers at MIT and BASF, the foundation model is reshaping chemical research by predicting molecular behavior with a fraction of the experimental data traditionally required.
The findings, published in the September 14 issue of the Journal of Chemical Information and Modeling, outline how the team trained the AI on a million molecules from the U.S. National Institutes of Health’s PubChem database. By correlating molecular structures with roughly 1,600 expert-developed descriptors, the model can be fine-tuned for tasks like determining bioactivity, viscosity, or solubility using minimal additional data.
Bypassing Data-Heavy Bottlenecks
The development team—comprising MIT graduate student Akshat Shirish Zalte, Jackson Burns, Research Scientist Charlles Abreu, Professor William Green, and three researchers from BASF—designed CheMeleon to circumvent the reliance on massive, literature-based datasets. While standard deep-learning models demand exhaustive information, CheMeleon calculates specific descriptors, such as branchiness or weight, to interpret underlying chemical patterns.
The CheMeleon foundation model drastically reduces the number of molecules that need to be tested experimentally to accurately predict the properties of new molecules,
said William Green, the Hoyt Hottel Professor of Chemical Engineering at MIT. This is a huge advantage, since synthesizing and testing a large number of molecules is slow and expensive.
From Competition Floor to Corporate Policy
The model’s utility was stress-tested during an OpenADMET competition held earlier this year. Its performance caught the attention of the pharmaceutical industry long before the research paper cleared the peer-review process. Inquiries began flowing to the team while the manuscript was still under submission.
As word of the model’s contest performance spread, adoption accelerated. Green noted that at least one pharmaceutical firm has already integrated the tool into its internal workflow, establishing a mandate for its research staff: Anytime you're modeling a property, start with CheMeleon, that's our new company policy.
Benchmarking Against Classical Systems
To validate the model, researchers measured CheMeleon against both existing foundation models and classical, expert-designed systems across 58 test cases. Zalte confirmed the model consistently outperformed these alternatives.
Built using ChemProp—the open-source machine-learning package maintained by the Green group at MIT—CheMeleon is currently available for public download. The project was backed by the Machine Learning for Pharmaceutical Discovery and Synthesis consortium, with additional funding provided by MathWorks and the U.S. Department of Energy.