Pinal generates proteins from text prompts, with lab success in four classes

A written description of a protein’s job can now guide an AI system toward designs that work in laboratory experiments. In an October 21, 2026 Nature paper, researchers introduced Pinal and reported intended function across four protein classes. The advance is a simpler starting point for design—not an escape from expert screening and laboratory testing.

The distinction is important. Protein-design systems often ask scientists to translate a desired function into detailed rules about molecular shape or amino-acid placement. Pinal instead accepts a description of the desired function and generates candidate sequences. Its authors present natural language as a higher-level interface that reduces reliance on manually specified structural and sequence constraints.

A written request becomes a molecular blueprint

Pinal is a 16-billion-parameter model trained on 1.7 billion synthetically annotated protein–text pairs. Those examples connect protein sequences with descriptions of what they do. As John Timmer explained in Ars Technica on October 27, the researchers used a language model to turn functional tags into plain-language descriptions, then trained with both the tags and the descriptions.

Generation happens in two stages. A component called T2struct converts the request into a structural blueprint. Rather than drawing a complete three-dimensional shape, it produces tokens representing local spatial relationships between amino acids, along with probabilities. A second component, SAProt-T, combines that information with the original request to generate the amino-acid sequence—the chain that makes up the protein.

The second stage can reconsider the first stage’s possibilities. The developers say this helps avoid choosing a likely structure while overlooking a less likely one that could support a better functional sequence. Training also drew on AlphaFold-predicted structures, rather than relying entirely on experimentally measured structures.

Working molecules, selected from a much larger pool

The experimental tasks covered a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and a metabolic H-protein. The authors report intended function in each case, including catalytic activity for both enzymes. One result went beyond detecting activity: a designed H-protein produced 1.7 times the product concentration of the native E. coli H-protein in a multi-enzyme pathway that fixes carbon dioxide. That comparison required no iterative experimental optimization.

The fluorescent experiment reveals how deliberately the candidates were selected. Researchers rejected sequences resembling known fluorescent proteins, while favoring a barrel-like structure and a key amino acid associated with fluorescence. They were seeking sequence novelty while retaining selected features of green fluorescent protein. Candidates lacking the preferred structure were not tested, so the experiment does not measure how every generated sequence would perform.

Public tools, but not the full training recipe

The paper provides several routes for using and examining Pinal. Its code availability statement lists public resources, while drawing a boundary around reproducing the complete training process:

• Trained model weights are available on Hugging Face.
• Public code includes inference, evaluation and model-use instructions; Pinal’s code carries an Apache 2.0 license.
• A web server offers another way to access the system.

The authors are withholding the full pretraining corpus and complete training pipeline from public release, citing potential biological misuse. They warn that releasing those resources could make it easier to adapt similar systems to sensitive datasets. They also say the AlphaFold structural collection used for training excludes viral proteomes, and that Pinal was neither trained nor intended for viral protein design. The release therefore opens model use and evaluation without opening the entire training recipe.