AI tools can now take a request for a protein’s function in ordinary language and turn it into candidate designs. The approach, used by a system called Pinal, lowers the amount of structural detail a user must provide—but successful designs still depend on computational filtering and laboratory testing.
From a function to a design
Many protein-design systems work in two stages. First, a user supplies constraints, such as a requirement to bind another molecule or to arrange certain amino acids in a particular way. One tool proposes a protein shape; another assigns an amino-acid sequence that could fold into it.
That workflow asks users to know a great deal about protein chemistry. For a catalyst, they may need to specify which amino acids matter to the reaction and where those amino acids should sit in three-dimensional space. The information can be demanding to gather and to express in the format a design system expects.
Pinal aims to let users begin with a description of what they want a protein to do. Its developers drew on databases in which protein sequences are paired with labels describing their functions. The goal was to train a system to connect those descriptions with structures and sequences that could perform the requested task.
How Pinal translates language
Pinal still follows a two-stage design process, but it begins with text. Its T2struct module turns a natural-language prompt into a structural blueprint: a linear array of tokens describing spatial relationships between amino acids and assigning probabilities to those possibilities.
A second system, SAProt-T, uses both the original prompt and T2struct’s output to produce an amino-acid sequence. It reassesses the structural options globally, which the authors say can help avoid missing a less probable structure that could support a better functional sequence.
To train the system, the researchers used protein sequences from three databases alongside structural information predicted by AlphaFold. They also had a large language model turn protein-function tags into plain-language descriptions, using both forms of text in training. One of the databases contained 800 million keyword–protein pairs.
Predicted structures can be inaccurate, and that could pose a problem when a request concerns a relatively obscure protein. The article notes that structural data is available for many common proteins, which could help inform AlphaFold’s predictions. The researchers report that enough of the training data was accurate for the approach to work.
Lab tests put the designs to work
The researchers tested Pinal on four kinds of proteins: a fluorescent protein, two enzymes, and a protein that participates in a large multi-protein complex. In each case, the process involved generating many candidates, narrowing the list computationally, and then testing a smaller set in the lab.
For the fluorescent-protein test, Pinal generated 50,000 candidates requested to fluoresce like GFP. The researchers filtered out sequences that resembled known fluorescent proteins, while looking for features associated with GFP: a barrel-like structure and a key amino acid that can be chemically modified to enable fluorescence.
That screening left 29 high-priority candidates. All were made in E. coli, and one fluoresced. The result shows why a natural-language prompt should not be mistaken for a guarantee of a working protein: candidate designs still need to survive multiple rounds of selection and experimental checks.
The hit rate also needs context. The researchers excluded candidates that looked like known fluorescent proteins and did not test designs that lacked a GFP-like structure. That means the experiment focused on a selected group of alternatives, rather than every possible way to produce fluorescence.
The same idea reaches DNA
A separate announcement described Carbon X, an open-weight model focused on DNA sequences. It was introduced by a collaboration involving Hugging Face, the Zhongguancun Academy, and the University of Naples “Federico II.” The source describes it as similar in spirit to Pinal, but focused on DNA rather than proteins.
Together, the projects point toward a more accessible way to express a biological design goal: describe a desired function, then let a model propose candidates. The practical work remains substantial. A prompt can start the design process, but filtering, production, and lab tests determine whether a proposed sequence does what the user asked.