MIT team builds model that predicts molecule traits from little data
Researchers at MIT have built CheMeleon, an open-source machine-learning model that predicts molecule properties when little data exists. Their paper in the Journal of Chemical Information and Modeling says it beats other methods at narrowing down candidate molecules.
The team, led by graduate student Akshat Shirish Zalte and Professor William Green, first trained the model on descriptors for about a million molecules. It then adjusts to each new property, such as solubility. Green says several pharmaceutical companies already use it, and nine of the top 20 entries in an OpenADMET contest used it. Those usage claims come from the research group itself.