Finding a molecule with a useful property often means making and testing candidates through trial and error. An MIT-developed model called CheMeleon aims to make that search more efficient by helping researchers predict how molecules may behave when data for a specific task are scarce.
Learning broadly before tackling a specific property
CheMeleon is trained first on molecular descriptors drawn from a million molecules in PubChem, a freely available chemical information database maintained by the U.S. National Institutes of Health. The descriptors include features such as molecular weight and the number of oxygen atoms, along with many others.
The model learns relationships between a molecule’s structure and a catalog of about 1,600 descriptors developed by chemists over decades. It is not trained at this stage to predict one particular outcome. Instead, that initial training gives it a general grounding in chemistry that researchers can adapt to a more focused question.
That distinction matters when only a few examples are available for the property a team wants to predict. Rather than learn everything from scratch, CheMeleon can be fine-tuned using those limited data points. The researchers say it can be applied to familiar properties such as solubility and viscosity, as well as more specialized questions involving bioactivity or reactivity with a protein or enzyme.
A different route from expert-built models
Traditional expert-driven models rely on specialists to specify which molecular features may be related to a property. For example, researchers might use knowledge about molecular weight, oxygen content, or how branched or bulky a molecule is when modeling solubility.
CheMeleon’s pre-training takes another route. The team uses calculated descriptors rather than drawing on chemical journal literature to teach the model its initial task. The resulting foundation model is then adapted for a particular prediction problem. Its starting point is broad, while the fine-tuning focuses on the property under study.
The approach was built using ChemProp, an open-source machine-learning package developed and maintained by MIT’s Green group. CheMeleon is available through ChemProp, where users can download it and fine-tune it for their own purposes.
Testing performance when data are limited
The researchers compared CheMeleon with other foundation models and with classical models based on expert-designed principles. Across 58 test cases, MIT graduate student Akshat Shirish Zalte said the model came out on top.
Its adoption has also extended beyond the study. In a contest organized by OpenADMET to predict properties of molecules that had not previously been synthesized, nine of the top 20 entries used CheMeleon. The result prompted inquiries from pharmaceutical companies, according to William Green, the Hoyt Hottel Professor of Chemical Engineering.
Green described one company’s internal guidance as: “Anytime you’re modeling a property, start with CheMeleon, that’s our new company policy”. The model’s use was spreading while the paper was still going through peer review.
Why screening can matter
Predicting a molecule’s properties does not remove the need for experiments. It can help researchers decide which candidates are most worth synthesizing and testing, reducing the number that need experimental evaluation. That could save time and expense, since making and testing many molecules is slow and costly.
The broader challenge remains: no method can perfectly predict all the many properties and responses a chemical compound may exhibit. CheMeleon is presented as a way to improve early screening, particularly when little data are available for the specific property of interest.
The findings were reported in the Sept. 14 issue of the Journal of Chemical Information and Modeling. The paper’s authors include MIT researchers and Jochen Sieg, Christian Feldmann, and Miriam Mathea of BASF SE in Germany. The work was supported by the Machine Learning for Pharmaceutical Discovery and Synthesis consortium, with fellowship support from U.S. Department of Energy and MathWorks.