An MIT team's open model predicts molecule properties from a few examples
MIT and BASF researchers say CheMeleon, an open-source model described in a peer-reviewed journal, learns from a million molecules, then needs only a few examples for a new task.
A model called CheMeleon can be adapted to predict how a molecule behaves after seeing only a few data points for the task, according to the researchers who built it.
The work comes from MIT in Cambridge, Massachusetts, USA, with three researchers from BASF SE in Germany: Jochen Sieg, Christian Feldmann and Miriam Mathea. MIT's announcement says it is described in a peer-reviewed paper in the Journal of Chemical Information and Modeling, in the 14 September issue.
The team first trained the model on a million molecules from PubChem, a large, freely available chemistry database maintained by the U.S. National Institutes of Health. The model learned links between each molecule's structure and about 1,600 descriptors, which are numbers that describe a molecule. Chemists have developed them over decades.
MIT graduate student Akshat Shirish Zalte, one of the paper's authors, says the team does not use published chemistry literature for this step. In his words: "We just calculate a bunch of molecular descriptors like molecular weight or number of oxygens, among others."
The model is not trained for any one property. Instead, one trained model can be adjusted, or fine-tuned, with a small set of examples for one job. The researchers compare it to a chameleon changing colours: the same model adapts to predict potency, toxicity or solubility.
Professor William Green, the Hoyt Hottel Professor of Chemical Engineering at MIT, says this greatly reduces the number of molecules that must be tested in the lab, and that making and testing many molecules is slow and expensive.
MIT says there is not yet any way to perfectly predict how a chemical compound will behave, so methods like this give initial guidance on which compounds might be worth making and testing.
The comparison is the team's own. It tested CheMeleon across 58 test cases against other foundation models, which are general-purpose models trained on broad data, and against classical models designed by experts. Zalte says in MIT's article: "we see that our model comes out on top."
MIT says several pharmaceutical companies already use CheMeleon regularly. Green says nine of the top 20 entries in an OpenADMET contest, which asked entrants to predict properties of molecules not yet made, had used it.
He also says at least one pharmaceutical company has told its staff to start with CheMeleon whenever they model a property.
CheMeleon is open source. Anyone can download it through ChemProp, an open-source package also built by the Green group, and fine-tune it for their own purposes.
Sources
- How to predict molecule behavior with scant data 1106 · news.mit.edu