Nature paper describes Pinal, a 16bn-parameter model designing proteins from text

Researchers reporting in Nature on 21 October 2026 have described Pinal, a protein design foundation model of 16 billion parameters that takes a written description of the function a user wants and returns candidate proteins to make and test¹. The team applied it to four separate design tasks and saw the intended function in all four when the designs were produced in the laboratory¹.

What the paper reports

Pinal is aimed at two constraints in current design models. Those models generally need an existing natural protein as a starting point for optimisation, or they need the functional goal restated by hand as geometric and sequence-level constraints before candidates can be generated¹. Pinal accepts the goal as natural language instead, which the authors put forward as a higher-level interface to design¹.

The model was trained on 1.7 billion protein-text pairs whose annotations were generated synthetically, and that training is what connects a stated functional intent to protein sequence and structure¹. In computational assessment, the candidates scored highly for predicted foldability, aligned with the functional descriptions they came from, and were diverse in sequence¹. The authors present those scores as a way of ranking which designs are worth testing experimentally, not as evidence of function¹.

Four targets were taken into the laboratory¹:

• a fluorescent protein
• a polyethylene terephthalate hydrolase
• an alcohol dehydrogenase
• a metabolic H-protein

Each one showed the function it had been asked for, and the two designed enzymes were catalytically active¹. The clearest quantitative result came from the H-protein. Put into a multi-enzyme pathway for CO2 fixation, the Pinal design increased product titre 1.7-fold against the corresponding native Escherichia coli H-protein, with no iterative experimental optimisation applied first¹.

What it means for the market

The commercial point here is about who can specify a design. Protein engineering groups in pharma, biotech and industrial biology currently depend on specialists to convert a functional requirement into the structural and sequence constraints a generative model will accept. The authors argue that natural-language input reduces that reliance on manually specified constraints¹. If that holds up outside the four cases reported, it changes the staffing and the cycle time of a discovery campaign more than it changes the chemistry, because the slow step moves from problem formulation to laboratory validation.

What the paper does not supply matters as much for buyers. None of the four targets is a therapeutic protein: the set is a reporter, two enzymes and a metabolic protein¹. Whether the same interface produces usable antibodies, binders or other drug modalities is not tested¹. The abstract also says nothing about how the model is made available, so licensing, access and any commercial route for drug developers are unknown on the evidence given.

Caveats

The experimental support is four design tasks, each with one intended function¹. The in silico measures of foldability and description alignment are predictions, and the authors describe them as a basis for prioritising candidates rather than as validation¹. The training annotations were synthetic rather than curated by hand¹. The 1.7-fold titre gain was measured for a single protein inside one engineered pathway, so it is not a general claim about yield¹.

The authors' own framing of the next step is programmable exploration of design space through language, with less dependence on hand-written structural or sequence constraints¹.