science & techMIT News

How to predict the behavior of molecules from scant data

the gist

Researchers at MIT built CheMeleon, a machine-learning model that is first pre-trained on about a million molecules to learn basic chemistry. It then adapts to predict specific properties such as potency, toxicity and solubility with only limited data. The approach aims to make early drug and chemical design faster when experimental data is scarce. The model is a single base system that is fine-tuned for each new task. The work targets specialists in computational chemistry.

so what

It may reduce the lab testing needed to screen new molecules. Its direct impact is mainly within drug discovery research.

Reprinted with permission of MIT News

Image: MIT News