AI model designs proteins from plain-language prompts
Researchers at the University of Washington published ProteinChat, a 16-billion-parameter model trained on structure-function pairs that generates novel protein sequences from natural-language descriptions — validated in silico and in two wet-lab assays.
- Model produces sequences with predicted fold confidence above 0.85 on AlphaFold3 for 78% of prompts
- Two generated enzymes expressed correctly in E. coli in independent validation
- Preprint peer-reviewed and published in Nature; code and weights publicly available