Designing new proteins from scratch usually requires in-depth knowledge of their structures and chemistry, as well as familiarity with highly specialized software. But researchers have now developed an AI tool where users simply describe what they want a protein to do in plain language.
Since AlphaFold cracked protein folding in 2020, AI has revolutionized the field of structural biology. Since then, a wave of generative models have begun to work in the opposite direction, learning from vast collections of amino acid sequences, protein structures, and labels detailing their function to design proteins that do not exist in nature.
These tools remain difficult to use, however, as scientists have to guide them with formal technical specifications like structural constraints or instructions to mutate particular sites. That means trying to design something like a new plastic-degrading enzyme or a fluorescent imaging marker takes significant specialist expertise.
But in new research published inNature, scientists at Westlake University in China detail an AI system called Pinal that can generates protein sequences from natural-language prompts describing a desired protein function. And while the model outputs a lot of duds, lab tests showed that with enough tries, the system could reliably achieve a user’s goal.
“For decades, the field has largely required specialized expertise to translate desired functions into low-level structural or sequence constraints,” the authors write. “The importance of Pinal lies not only in its performance but also in reframing protein design as a problem that can be specified through functional language.”
Pinal works in two stages, neither of which asks the user for details about the molecular structure or sequence of amino acids. The first module takes the plain-language description of what the user wants and turns it into a structural blueprint. A second module then reads both that blueprint and the original text to decide the sequence of amino acids.
Crucially, the software doesn’t put all its eggs in one basket. The first module generates several candidate structures, all of which are fleshed out with their own amino acid sequences. Pinal then scores each of these combinations before settling on a final design. The researchers say this accounts for the fact that the best protein designs don’t always have the most obvious shapes.
Training the model took a vast amount of data. The researchers assembled roughly 135 million protein sequences and used large language models to write plain-language descriptions of their functions, generating about 1.7 billion protein-text pairs. They then used AlphaFold to predict the structures of the sequences.
The team used this corpus to train the two modules separately. The first learned to predict a structural blueprint from a text prompt, and the second learned to pick the right amino acid for each position in the sequence, given both the blueprint and the prompt.
To test the model, the team asked it to design four proteins: one that glows green, an enzyme that breaks down PET plastic, an enzyme that processes alcohols, and a shuttle protein that ferries chemicals between enzymes in a metabolic pathway.
The model was able to solve all four challenges, but its hit rate was a mixed bag.
When asked to produce a glowing protein, it created 50,000 sequences. These were filtered down to 29 for testing, and only one actually fluoresced. The system was better at producing the shuttle protein. Five of the 20 candidates worked, and the best one boosted a multi-enzyme carbon dioxide fixation pathway to 1.7 times the output of the natural version.
A different collaboration between Hugging Face and academics in China and Italy recently applied similar ideas to DNA sequences. The group released a 14-billion-parameter open model called Carbon-X that can answer questions about a gene sequence, predict what a mutation will do, or generate new DNA.
That model is based on Qwen3-14B, but was retrained on 300 billion tokens that mixed genomic sequences, biology text, and ordinary English. Getting it to learn DNA without forgetting English required a delicate balance though, the researchers write in a post, and forced them to do a long slow training run with a large model.
Across five benchmark tests, Carbon-X beat Opus 5.5 and Gemini 4 Argon and trailed GPT-6.1 Sol by less than a percent. And the model can run on a MacBook, which could be a major boon for researchers eager to avoid sharing sensitive or unpublished sequences with a commercial model provider.
Neither tool is perfect though. Pinal’s own developers concede that their screening heuristics “may restrict exploration to a limited region of the design space,” while Carbon-X struggles with long sequences and remains prone to hallucinations. Validating the sequences they produce also still requires a well-equipped lab and experienced scientists.
But these tools could nonetheless empower a far wider range of biologists to experiment with problems previously restricted to a handful of computational specialists.
