MIT Unveils CheMeleon AI That Tops 58 Tests Using 1 Million Molecules

1 articles · Updated · MIT News · Nov 6

CheMeleon, an MIT-developed foundation model, predicts molecular properties accurately even when only a few task-specific data points are available, helping narrow which compounds should be synthesized and tested.
The system was pre-trained on descriptors for about 1 million PubChem molecules and roughly 1,600 chemical descriptors, then fine-tuned for targets ranging from solubility and viscosity to bioactivity and reactivity.
Across 58 test cases, the researchers said CheMeleon outperformed both other foundation models and classical expert-designed approaches, cutting the amount of slow, expensive experimental work needed.
Nine of the top 20 entries in an OpenADMET molecule-property contest used CheMeleon, and MIT said several pharmaceutical companies had already adopted the open-source model before the paper completed peer review.
The Sept. 14 Journal of Chemical Information and Modeling paper, co-authored with BASF researchers, points to broader use of AI to speed drug, electronics and energy materials discovery.