Worth knowing
How to predict the behavior of molecules from scant data
It shows how pre-training foundation models on easily calculated synthetic descriptors can solve the data-scarcity bottleneck in scientific AI applications like drug and materials discovery.
Reported by MIT News ·
What happened
Researchers from MIT and BASF developed CheMeleon, an open-source machine learning foundation model pre-trained on 1,600 calculated descriptors of one million molecules from PubChem. The model outperforms existing methods when fine-tuned on scant data to predict novel chemical properties, and is already being adopted by pharmaceutical companies for drug discovery.
Who it’s about
- MIT
- CheMeleon
- ChemProp
- BASF
- William Green