—

None of these enzymes are real. Not one has ever been expressed, crystallised, assayed, or observed. There is no organism they came from, because those organisms do not exist either.

How they are generated. Everything you see is produced in your browser, procedurally, from a single number. A seeded pseudo-random sequence assembles a plausible backbone out of helices, strands and loops. The same sequence names the enzyme in a binomial-adjacent style, invents an Enzyme Commission number, fabricates a source organism, assigns a substrate, writes a one-sentence function, and predicts a mass. The rotating figure is a cartoon of that backbone. The downloadable .pdb file contains the very same coordinates, formatted to look like a deposited structure. It will open in real structural-biology software. It describes nothing.

What that says about structure prediction. Generative structure prediction has turned protein folds into an aesthetic commodity — ribbon diagrams as wallpaper, model outputs as desktop backgrounds. After a few minutes of scrolling gorgeous, plausible enzymes that will never exist, the point arrives on its own: a great deal of generated biology is now visually indistinguishable from catalogued biology, and a database of predicted structures is a catalogue of confident guesses, not a catalogue of things that are known to be true. A predicted structure is a hypothesis wearing the clothes of a measurement.

Where we drew the line. This piece stays firmly on the enzyme side — metabolic enzymes, structural proteins, luciferases, extremophile oddities. No toxins, no virulence factors, nothing pathogen-shaped, and no sequence-level design guidance of any kind. The beauty here is in catalysis and folding, not in threat, and that is both the responsible line and the better-looking one.