A large group of researchers working in China has developed a protein design system called Pinal that accepts ordinary written descriptions instead of technical specifications, according to Ars Technica. The tool takes a plain-language prompt and works toward a protein that performs the requested function. In a separate announcement, Hugging Face said it now hosts a tool that applies a comparable approach at the DNA level.
Ars Technica describes the development as a kind of "vibe coding" arriving in protein design, with an important caveat: the approach is most useful to anyone with access to a laboratory, since many candidate designs will fail and need to be sorted from the ones that work.
Why it matters
Existing protein design software is powerful but demanding. Ars Technica reports that the leading tools typically operate in two stages. A user first defines constraints, such as the need to bind a particular molecule or to place certain amino acids in specific positions. One AI model then proposes a protein shape that satisfies those constraints, and a second tool fills that shape with an amino acid sequence that should fold into the intended configuration.
That workflow assumes real expertise. Ars Technica notes that designing a catalyst requires knowing which amino acids drive the reaction and where they must sit in three-dimensional space. Expressing that knowledge in a form the software accepts takes a sophisticated grasp of protein chemistry.
The gap Pinal targets is one of vocabulary. Ars Technica points out that two specialists would not need such precision with each other: saying "serine protease" conveys a great deal about the reaction mechanism to a trained listener. Current AI tools, by contrast, handle 3D arrangements of amino acids well but understand little about the functions those arrangements produce. Pinal is built to close that distance.
How the system turns text into a sequence
Pinal keeps a two-stage architecture but changes what the user supplies at the start. According to Ars Technica, the first stage is handled by a module called T2struct, which converts a natural language prompt into a structural blueprint. T2struct does not generate a 3D shape directly. Instead it produces a linear series of tokens describing how each amino acid relates spatially to its nearest neighbors, along with a probability for each.
That output passes to a second component, SAProt-T, which combines the original text with the T2struct result to build the amino acid sequence for the final design. The probabilities from the first stage carry weight here because SAProt-T re-evaluates the design globally. Ars Technica reports that the authors say this guards against discarding a slightly less likely structure that could support a much better functional sequence.
Training relied on scale. Ars Technica reports that for every set of function tags attached to a protein, the researchers fed those tags to a large language model and had it write a plain-language description, then used both the tags and the description in training. Sequences came from three databases and were paired with structural information predicted by AlphaFold. One of those databases alone held 800 million keyword–protein pairs, according to Ars Technica.
The reliance on AlphaFold carries a known limitation. Ars Technica notes that AlphaFold's predictions are not fully accurate, which could matter when the target is an obscure protein. For widely studied proteins, however, experimental structural data exists and would presumably inform the predictions. Ars Technica reports that enough of the structural data proved sound for training to succeed.
Four test cases, and a very large candidate list
The team evaluated Pinal on four dissimilar targets: a fluorescent protein, two enzymes, and a protein that participates in a large multi-protein complex. Ars Technica reports that the procedure followed a similar pattern across cases, and uses the fluorescent protein to illustrate it.
The classic reference point for fluorescence is GFP, the green fluorescent protein, whose discovery earned a Nobel Prize. Pinal was asked to generate 50,000 proteins capable of fluorescing like GFP, and the researchers then narrowed that pool using a series of filters. Ars Technica's account of the filtering process is cut off in the available reporting, so the full criteria and the final success rates are not detailed here.
What Ars Technica does make clear is the overall picture: if a working protein is the measure, the hit rate may be modest, but the system still delivers results. The practical consequence is that Pinal looks less like a replacement for expert-designed proteins and more like a way to generate large batches of candidates that a lab can then screen.
The Hugging Face announcement adds a second front. Ars Technica reports that the platform is now hosting a tool that does similar work on DNA rather than proteins, extending natural-language design prompts from the protein level down to the genetic sequence level.
Watch for independent benchmarking of Pinal's success rates and for how the Hugging Face DNA tool is adopted, since both will indicate whether conversational prompts become a practical front end for biological design or remain a laboratory curiosity.
