How to predict the behavior of molecules from scant data

MIT researchers have built a machine-learning model, CheMeleon, that predicts how molecules will behave when very little experimental data is available, MIT News reports. The aim is to cut down the slow and costly trial and error involved in finding new drugs and materials.

CheMeleon is a foundation model for chemistry. It was first trained on about a million molecules from PubChem, the public database run by the U.S. National Institutes of Health, by having it predict roughly 1,600 descriptors that chemists have developed over decades and that are easy to compute, such as molecular weight. It can then be fine-tuned with only a few data points to predict a specific property, from solubility to how a molecule reacts with a particular protein. The team says it came out ahead of other foundation models and of classical expert-designed models across 58 test cases.

The work, by graduate student Akshat Shirish Zalte, Jackson Burns, Charlles Abreu, Professor William Green and three researchers at the chemical company BASF, appeared in the Journal of Chemical Information and Modeling. The model is open source and available through MIT's ChemProp package.

It was in use before the paper had been reviewed. Green says nine of the top 20 entries in a property-prediction contest organized by OpenADMET used CheMeleon, and that at least one pharmaceutical company told its staff to start with it whenever they model a property.