A method developed at MIT, described in the Sept. 14 issue of the Journal of Chemical Information and Modeling, was reported to outperform existing approaches for narrowing the field of candidate molecules when few data exist on a property of interest, such as ligand affinities for a newly discovered cellular target. The paper was by MIT graduate student Akshat Shirish Zalte, recent graduate Jackson Burns PhD '26, Research Scientist Charlles Abreu, Professor William Green, and three BASF researchers.

The open-source model, called CheMeleon, was pre-trained on detailed data for about a million molecules from the PubChem database and roughly 1,600 descriptors. It can then be fine-tuned for specific properties using few data points. In 58 test cases, the team reported, their model came out on top against other foundation models and classical expert-designed models. Green said the model reduces the number of molecules that must be tested experimentally.

An earlier version of the paper had been online since last year, and the team said CheMeleon was already in use at several pharmaceutical companies. Nine of the top 20 entries in an OpenADMET contest, Green said, used CheMeleon. The work was supported by the Machine Learning for Pharmaceutical Discovery and Synthesis consortium, with fellowship support from the U.S. Department of Energy and MathWorks.