Westlake's Pinal Designs Proteins From Text

A team led by Westlake University in Hangzhou says it has built an artificial-intelligence model that turns plain-language descriptions into proteins which then performed their described jobs in the laboratory, according to a paper published in Nature on 21 October.

The model, named Pinal, has 16 billion parameters and was trained on 1.7 billion protein-text pairs whose text annotations were generated synthetically, the authors write. A user types what a protein should do, and the system proposes an amino-acid sequence intended to match. The first authors are Fengyuan Dai and Shiyang You, with Fajie Yuan as the last author; the work lists further affiliations including Shanghai Zelixir Biotech.

The group has released the model weights on Hugging Face and the inference and evaluation code on GitHub under an Apache 2.0 licence, alongside a public web server. It has withheld the full training corpus, citing biosecurity.

Four proteins that did their jobs

The clearest part of the claim is the wet-lab test. The authors prompted Pinal for four different proteins and say each showed its “intended function”. They designed a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and a metabolic H-protein. For both enzymes, the paper reports catalytic activity; for the fluorescent protein, the expected glow.

The most striking single number concerns the H-protein, which the team inserted into a multi-enzyme pathway that fixes carbon dioxide. There, the designed protein increased product titre 1.7-fold relative to the native E. coli H-protein, and the authors stress this came “Without iterative experimental optimization” — in other words, from the model’s output rather than rounds of trial-and-error tuning.

Those are four examples, not a success rate, and the proteins are research demonstrations rather than medicines or products. The benefit worth weighing is practical: designing proteins to order could speed up enzymes for industrial chemistry and carbon capture, the kind of beneficial use the field has long promised. The paper does not claim the proteins are ready for any application, and a 1.7-fold gain in one engineered pathway is a laboratory result, not a deployed process. Independent replication, and tests on functions where failure is easy to measure, will matter before anyone treats “describe it and get a working protein” as routine. What the result does show is that a text-to-protein model can clear the bar of making something that works at all, which earlier natural-language systems struggled to do.

Weights are open; the training data is not

The team has chosen a deliberately partial release, and the reasoning is spelled out in the paper. It makes inference code, evaluation code, the model weights and the web server public, a scope the authors compare with releases such as AlphaFold, ESM3 and ESMFold. What it does not publish is the large training corpus, because “The release of large-scale training resources for generative protein design requires careful consideration of potential dual-use biosecurity risks.”

The authors note that the structural data behind Pinal came from the AlphaFold database, which contains no viral proteomes, so “Pinal was not trained or intended for viral protein design.” Even so, they argue, publishing the full training resources “could lower the barrier for adapting similar systems to sensitive biological datasets”, and so those resources were “provided only under confidential access for editorial and peer review assessment”.

This is the balance the field keeps returning to. Open weights let other laboratories check and build on the work, which is how beneficial tools spread and how claims get tested; withholding the corpus is a defence against misuse that the authors judged worth the cost to full reproducibility. Reasonable researchers disagree on where that line sits. The relevant point for readers is that Pinal’s designers made the call explicitly rather than by default, and documented both what they released and what they held back. TSN has reported adjacent efforts, from open protein-design models to the Baker lab’s AI-designed enzymes and Anthropic’s work on toxin binders, where the same question of openness against safety recurs.