Estructura tridimensional de una proteína fluorescente amarilla, con su barril de láminas beta y el cromóforo al centro

· Translated from the Spanish original

AI and biotechnology

Pinal designs proteins from a sentence and two of its enzymes worked without optimization in the lab

The 16-billion-parameter model published in Nature produced a fluorescent protein, two enzymes and an H-protein that raised a product 1.7-fold.

What happened

  • On October 21, 2026, Nature published “Programmable design of functional proteins from natural language,” introducing Pinal.
  • Pinal is a 16-billion-parameter foundation model that proposes candidate proteins from a natural-language description of the desired function.
  • It was trained on 1.7 billion synthetically annotated protein–text pairs.
  • The authors applied it to four tasks: a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and a metabolic H-protein. In all four they observed the intended function, with catalytic activity in both enzymes.

Why it matters

  • Until now designing a protein required starting from a natural one or translating the goal into shape and sequence constraints. Here the entry point is a sentence.
  • The practical figure is the designed H-protein: without iterative experimental optimization, it raised product output 1.7-fold over the native E. coli protein in a CO₂ fixation pathway.
  • For a Chilean mining or environmental biotech lab, the PET hydrolase points to a concrete use: enzymes to degrade plastics. The paper shows it worked, not how much plastic it breaks down.

The number

1.7-fold. Increase in a CO₂ fixation pathway’s product with the H-protein designed by Pinal over the native E. coli one.

Context

The abstract notes that current design models have limits: they require natural starting proteins or reformulating the goal as geometric constraints. In computational evaluations, Pinal’s candidates combined predicted foldability, alignment with the functional description and sequence diversity.

What’s next

  • The authors present the result as a basis for prioritizing designs that can be tested in the lab.
  • No dates for an open version of the model or for industrial applications.

Bottom line

AI-driven protein design and discovery already has recent precedents: 950 Claude agents described an unreported enzyme system and AlphaFold added 2,800 viruses to its catalog. Pinal adds the reverse step, from text to protein.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes