NederlandsAI model designs working proteins from a sentence in plain language
Researchers in China have built an AI model that designs proteins from a description in plain language. In four trials, the result actually did what was asked of it: give off light, break down plastic, convert alcohol and help capture CO2.
Proteins do almost all the work in a living cell. They break down substances, build others up, give off light or speed up chemical reactions. Anyone wanting a protein that does something very specific has so far had to follow one of two difficult paths: start from an existing natural protein and refine it step by step, or first translate the goal into precise requirements about shape and building blocks. Both demand years of experience.
Researchers from, among others, Westlake University in Hangzhou, China, presented a model in the journal Nature on 21 October 2026 that takes over that translation step. It is called Pinal. One describes in plain language what the protein should do, and the model delivers candidates. In four very different assignments, those candidates in the lab actually did what was asked.
From sentence to molecule
Pinal works in two steps. First it devises a shape to match the description: the three-dimensional figure the protein should adopt. It then searches for the chain of amino acids that folds itself into that shape.
The model has 16 billion adjustable parameters, the size of a substantial language model. It was trained on 1.7 billion combinations of a protein and a corresponding description, largely assembled automatically from public protein databases.
For each assignment, Pinal generates thousands of candidates. The researchers first filter these on the computer: they check whether the design is likely to fold up neatly and whether it indeed matches the description. Only the best few go into the lab.
Four assignments, four results
The first test was a light-emitting protein, the kind biologists use to make cells visible. Of 29 designs tested, one gave off light, with a sequence of building blocks that differs strongly from the natural example taken from the jellyfish.
The second was an enzyme that can break down PET, the plastic used in soft drink bottles and clothing fibres. From more than 2,400 promising designs, the researchers tested the 94 best-scoring; one of them actually attacked the plastic.
The third assignment was an enzyme that converts alcohol. Of eight designs tested, five turned out to work. And the fourth was an H-protein: a component of a chain of enzymes that together convert CO2 into a usable substance. The best design yielded 1.7 times as much end product as the natural H-protein from the gut bacterium E. coli — without any rounds of refinement being needed.
Open, but not entirely
Pinal has been released under an open licence. The code is on GitHub, the trained model on Hugging Face, and there is a website where it can be tried out without any installation.
The training data and the complete training method, however, the researchers are keeping to themselves. Their reasoning: anyone who has those assets in hand could build a similar system on sensitive biological data. Nature's reviewers were given confidential access to them. The model itself was not trained on viruses, because the database of protein shapes used contains no viral proteins. Large models such as AlphaFold and ESM3 are released in a similar way.
What does this mean for the road to the singularity?
The interesting thing here is not that a computer designs a protein — that has been happening for some time. It is the control knob. Until now, anyone who wanted to build something in living nature first had to learn the language of the specialist: folding patterns, binding sites, scoring thresholds. Now a sentence is enough.
That is the same shift that programming underwent when it became possible to tell a computer what was wanted rather than how to do it. And it ties in with what is being seen elsewhere: AI models that carry out a complete design cycle themselves instead of an isolated computational chore.
At the same time, the lab has the final say. Of thousands of designs, only a handful make it to the finish line, and someone still has to build and measure them. The real gain lies in the fact that the idea stage is no longer a bottleneck: the question shifts from "can we design this?" to "what do we actually want?". The fact that the researchers deliberately do not release their training data shows that this question has a downside too — and that the field is now thinking about that itself.
What is the technological singularity? Read the explanation →
