Proteins are the small machines of life: they speed up chemical reactions, produce light, and can even break down plastic. Until now, asking an AI to design one required solid knowledge of molecular structure and geometry. A team led by Westlake University in China has described in the journal Nature a model called Pinal that starts from a plain sentence, such as “a protein that emits green light.”
Pinal works in two steps. A first module turns the description into a kind of structural blueprint; a second writes the amino-acid sequence that matches both the blueprint and the request. The tool prepares several blueprints and keeps the best pairing. To train it, the researchers matched 135 million proteins with descriptions of what they do, about 1.7 billion text-protein pairs.
The results are real but modest. For the glowing protein, 50,000 sequences were generated; 29 were tested, and only one actually glowed. For another protein, involved in carbon capture, 5 of 20 candidates worked, and the best made the process 1.7 times more effective than the natural protein. The authors acknowledge that controlling precise properties remains difficult.
There is something beautiful in watching researchers learn, step by step, the grammar of living things. These tools are advancing, but the road is still long and humble, made of trials and failures.