Who Gets Bench Time When Anyone Can Design A Protein?
By Morgan Kohler, Executive Editor

A researcher who cannot specify a protein backbone can now describe the function they want in a sentence and get a working enzyme back. Four times out of four, in the cases that have been tested.
Pinal, as described in Nature on Oct. 21, is a 16-billion-parameter foundation model that designs candidate proteins from natural-language descriptions of function. It was trained on roughly 1.7 billion protein-text pairs, assembled by taking approximately 135 million protein sequences and using large language models to write plain-language descriptions of what each one does.
It runs in two stages. A prompt is first translated into structural information, and a protein language model called SaProt then designs sequences conditioned on both the user’s stated intent and that generated structure.
The researchers tested four design tasks: a fluorescent protein, a polyethylene terephthalate (PET) hydrolase, an alcohol dehydrogenase, and a metabolic H-protein. In each case, the intended function was observed experimentally. On the H-protein, five of 20 candidates worked, and the best of them raised the output of a multienzyme carbon dioxide fixation pathway to 1.7 times that of the natural version.
The coverage that followed in early November emphasized the four-for-four record. The authors themselves are considerably more measured, and what they concede is the more useful part of the paper.
What This Means For Drug Discovery
Let’s start with what wasn’t tested. None of the four validated tasks is a therapeutic protein and there is no binder, antibody, nanobody, and target engagement anywhere in the experimental record.
The assays differ in kind, not just difficulty. Fluorescence and catalysis can be read in bulk, cheaply, with unambiguous readouts. Target binding, proteome-wide selectivity, and immunogenicity cannot, and the paper does not attempt them. The validated envelope is four well-characterized enzyme and reporter functions with abundant training data and convenient functional readouts. This is a reasonable place to demonstrate a method and a poor basis for inferring what it will do against a drug target.
The authors are candid about the constraints. They concede that their screening heuristics may restrict exploration to a limited region of the design space. Models of this class can hallucinate unreliable or repetitive sequences, and the exact phrasing of a prompt affects the structure that comes out the other side.
That last limitation deserves more attention. A structural specification is a controlled input, but a sentence is not. If wording shifts the generated structure, a design process built on prompts has an uncontrolled variable sitting at its front end, which is a reproducibility problem before it is a performance problem.
It is also worth noting how much of the screening remained computational. Five of 20 is the wet-lab confirmation rate on the task the model handled best. Across published work in this area, computational success rates and wet-lab confirmation rates routinely differ by an order of magnitude, and any team evaluating a tool like this should insist on the latter.
Pinal is a statement about who can run a design campaign rather than about what a design campaign can currently produce for a drug program. If a biologist, a medicinal chemist, and a translational scientist can each generate protein candidates from a text box, protein sciences and assay development become the rate limiter and the gatekeeper simultaneously. And someone has to decide whose prompts get bench time. That is a resourcing and governance question.
The nearest-term value sits in tool proteins rather than therapeutics: engineered enzymes for assay reagents, biocatalysts for synthetic routes, reporter constructs. Those are real cost centers, the assays are cheap, and a pilot there is defensible now. For therapeutic binder design, the responsible position is that this is upstream capability rather than a current option. The thing to wait for is a published campaign against a real target with real selectivity data.
Interfaces, though, tend to change organizations well before they change science. That is the part worth planning around.
The Discovery Question
If anyone in your organization could generate a protein candidate tomorrow, who decides which ones reach the bench?
Related Drug Discovery Online Editorials
- From Receptor Signaling to Therapeutic Design: Rethinking Antibody Discovery
- The 2026 Pharma R&D AI Sourcing Map: Partnerships, Concentration, Inflection
- How Peptides Are Expanding The Possibilities Of Drug Discovery
About The Editor
Morgan Kohler is the executive editor of guest expert articles for Drug Discovery Online, working with industry experts to develop insightful and informative pieces about drug discovery and preclinical drug development science, regulatory strategy, trends, analyses, and commentary.
Morgan holds a master’s degree in professional writing and web content development and has worked in the publishing industry since 2018. Before working for Drug Discovery Online, she guided and edited industry-expert-bylined content on topics related to radiology and economics.