Pinal: when natural language becomes a blueprint for biology

A decade ago, designing a protein from scratch was like trying to build a car engine with the manual in Japanese. It was possible, but it required years of specialized training, complex calculations, and a significant dose of luck. Today, an artificial intelligence model called Pinal, developed by researchers at Westlake University in Hangzhou, China, allows anyone to describe in plain words what they want a protein to do and receive, in return, a functional amino acid sequence tested and validated in a laboratory.

The results were published in Nature on October 21, 2026 and represent not just a technical leap, but a paradigm shift in how scientists and even non-specialists can think about molecular design.

The problem no one could solve

Existing protein design models face a fundamental limitation: they require the researcher to already have a natural protein as a starting point for optimization, or to manually reformulate functional objectives into geometric and sequence-level constraints to guide candidate generation. It is a process that demands expertise in bioinformatics, crystallography, and computational modeling, barriers that have kept most molecular biologists on the outside.

The authors describe the approach as "a programmable instruction set for biology," one capable of democratizing protein design and shifting the paradigm from incrementalism to programmable exploration.

How Pinal works

The model has 16 billion parameters and was trained on 1.7 billion synthetic protein-text pairs, a massive collection linking functional descriptions in natural language to protein sequences and structures. The central idea is elegant in its simplicity: instead of asking the model to generate an amino acid sequence directly from text, Pinal operates in two stages.

First, it generates a candidate protein structure from the language instruction. Then, it designs an amino acid sequence conditioned on both the structure and the original description. The authors justify this two-step approach with a solid argument: protein structure is more conserved than sequence, making it a more stable intermediate representation to generate from a textual description before committing to a specific chain of amino acids.

It is the same reasoning that led language models to first learn latent representations before generating text, structure precedes surface.

Four proteins, four distinct tasks

To test the model, the team chose four very different tasks, spanning from fluorescent proteins to enzymes with real catalytic activity.

Fluorescent protein -- the model generated candidates that emitted green fluorescence, similar to the original Green Fluorescent Protein (GFP), but with an entirely de novo sequence, not derived from any natural organism.

PETase -- an enzyme capable of degrading polyethylene terephthalate (PET, the plastic used in bottles). Of the top 94 candidates selected for gene synthesis, one (pinalPETase36) showed catalytic activity, with progressive accumulation of plastic degradation products over seven days. Considering that the chemical recycling of PET is one of the biggest environmental challenges in the plastics industry, an enzyme that degrades it efficiently could be transformative.

Alcohol dehydrogenase -- another enzyme with catalytic activity, involved in redox reactions that have applications in industrial biocatalysis and fine chemical production.

Metabolic H-protein -- this was perhaps the most surprising. The protein designed by Pinal not only worked as expected, but outperformed its natural counterpart: in a multi-enzyme carbon dioxide fixation pathway, the model-generated protein produced 1.7 times more product than the native Escherichia coli H-protein. This is not just "it works", it is "it works better than nature".

All four were expressed, purified, and measured in the laboratory, without the iterative experimental optimization that traditional de novo protein design has typically required.

Why this matters

The value of Pinal lies not in any individual protein, but in the workflow it enables. A natural language prompt, run through a single model, produces a structure and sequence that someone then expresses, purifies, and measures, and it works. Without the multiple rounds of optimization that de novo protein design has traditionally demanded.

For students and researchers in computational biology and synthetic biology, this means the barrier to entry for protein design drops dramatically. A biochemist studying bacterial metabolism could describe what they need in words and receive a functional candidate, without needing to learn AlphaFold, Rosetta, or other modeling tools.

Nature itself called it "a new chapter in programmable biology." And there is a reason: if natural language becomes the interface for protein candidate generation, the entire field opens up to anyone who can write a prompt.

What comes next

The authors are careful to emphasize that Pinal does not replace experimental validation, it opens a more accessible path to generate testable candidates. The model's value lies in opening a new and more accessible way to generate testable candidates, rather than replacing experimental validation.

But what Pinal achieves is a milestone: the first foundation model trained specifically to translate functional descriptions into entirely new proteins, with confirmed experimental validation across four distinct tasks. If the coming years bring models that multiply precision and diversity, the combined effect could accelerate the discovery of industrial enzymes, protein therapeutics, and biodegradable materials at a pace that today seems science fiction.

The question is no longer "when" will language-programmable biology impact medicine and industry, how fast will we need regulations for what happens when anyone can ask an AI to design a protein.

Sources: Nature, EduFabTech, BioEngineer

✓ Independent sources cross-checked and verified before publishing