Discovering new molecules drives pharma, electronics and energy sectors, yet traditional discovery relies on trial‑and‑error because accurate property prediction is still missing. A MIT team reported a method that outperforms existing approaches when only a few data points are available, published in the September 14 2023 issue of the Journal of Chemical Information and Modeling. The model, called CheMeleon, is built on the open‑source ChemProp package and is pre‑trained on about one million molecules from PubChem using roughly 1,600 molecular descriptors. After pre‑training, fine‑tuning with a handful of target‑property samples enables prediction of common properties such as solubility and viscosity as well as specialized bioactivity or enzyme reactivity. In 58 benchmark cases CheMeleon ranked above other foundation models and expert‑driven models, dramatically cutting the number of compounds that must be synthesized and tested. Several pharmaceutical companies have already adopted it as the default modeling tool for early‑stage projects. The model is fully open‑source; users can download it via ChemProp and fine‑tune it for their own tasks.
Review: CheMeleon proves that a well‑designed foundation model can extract chemistry knowledge from limited data, offering a fast, cost‑effective route for molecular screening.