Ω · Oneironexus · Blog
The Map Got Bigger
Two rooms where AI drew what no one had seen: the shapes of proteins and the catalogue of stable crystals. With the numbers the maps came with.
14 October 2026 · Aaron Michael Weis
Imagine a museum with two new rooms. Both opened in the last five years. Both hold maps of places that exist but had never been drawn.
Walk in with me.
Room one: the shape of life
A protein is a long chain of building blocks that folds into a shape, and the shape is the job. Hemoglobin's shape carries oxygen. An antibody's shape grabs a virus. For seventy years, finding a protein's shape meant growing crystals, firing X-rays at them, and reading the pattern — months or years for one protein.
In 2020 the field held its regular blind test. Teams were given protein sequences no one had solved and asked to predict the shapes. The answers were then compared with the real structures, measured in the lab and kept secret until judging.
AlphaFold's predictions landed within about one ångström of the real backbone, at the median — a 0.96 with the interval running 0.85 to 1.16. One ångström is about the width of a hydrogen atom. The next best method in the same test landed at about 2.8.
That is the kind of result that gets written up once, carefully, in Nature, and then changes what a biology lab does on a Monday morning. A shape that used to take a year to measure is now a starting point you can read before lunch.
Room two: the catalogue of stable crystals
Every solid material you have ever touched — a battery electrode, a solar cell, a chip, a magnet — is a crystal: atoms arranged in a repeating pattern. The pattern decides what the material can do. Humanity's catalogue of stable crystals, the ones that hold together, had grown slowly for a century, one careful discovery at a time, to somewhere in the tens of thousands.
In 2023 a graph network called GNoME, trained on first-principles physics calculations, drew the rest of the room. It proposed 2.2 million crystal structures. Of those, 381,000 are newly stable by the same physics test the field already uses — enough to bring the total of known stable crystals to 421,000, almost ten times more than before.
And here is the detail I keep coming back to. The authors checked their predictions against the world's experimental database of crystal structures, the ones real labs had grown and measured on their own, with no connection to the paper. 736 of GNoME's predicted structures were already there — found independently by hand, in laboratories, by people who did not know a computer had drawn the same shape. The map and the territory agree where they overlap.
What a bigger map is for
A map does not build the road. A predicted protein still gets confirmed in the lab; a predicted crystal still gets synthesized, tested and, if it works, manufactured. The map's gift is different: it tells you where to look first.
A battery chemist with 381,000 candidate structures starts the week with a shortlist instead of a guess. A drug designer with a protein's shape starts with the pocket where a molecule fits. A student opens the AlphaFold database and sees, in minutes, a structure that would have been a whole doctoral thesis a decade ago.
Both maps are open. The protein structures are in a free public database. The crystal structures are published with the paper. Anyone can walk into either room.
The habit the rooms share
Both results arrived with the number that tests them. AlphaFold came with the blind-test score and its confidence interval. GNoME came with the 736 independent matches. That is how a map earns trust: it shows you where it has already been checked against the ground.
At Oneironexus we keep the same habit on our own work: a number travels with where it came from and when. A map people can check is a map people can build on.
The map got bigger. The territory is still there to walk. Now we know where to start.
Where the numbers come from
Protein structure prediction at CASP14: Jumper J et al., Nature 2021;596:583–589, DOI 10.1038/s41586-021-03819-2 — median backbone r.m.s.d.95 0.96 Å (95% CI 0.85–1.16) vs 2.8 Å (2.7–4.0) for the next best entrant. Crystal discovery: Merchant A et al., Nature 2023;624:80–85, DOI 10.1038/s41586-023-06735-9 — 2.2 million structures, 381,000 newly stable on the computed convex hull (total known stable 421,000), 736 matched to structures independently recorded in the Inorganic Crystal Structure Database; the authors are at Google DeepMind, and the stability figure is computational, not synthesis. CC BY 4.0 — attribute as Aaron Michael Weis, Oneiron Holdings LLC, oneironexus.ai.