Not reviewed. This note reports what the post itself says. Unlike the other notes in this collection, its claims have not been checked against the underlying papers, reports or data. The work concerns biological design that could be misused, and in-depth analysis of it triggers the safety safeguards of the AI system used to prepare these notes, so it was left at the level of the post's own summary.
What the post reports
Vault particles are large, hollow protein shells, named for a resemblance to a cathedral's vaulted ceiling, discovered in 1986 and present across most of the eukaryotic tree. Forty years on, the post says, nobody knows what they do. Several functions have been proposed and none confirmed; mice without vaults survive, which leaves no obvious defect to follow. The post's puzzle is why evolution would keep something so large (roughly three times the mass of a ribosome) if it did nothing.
Anthropic's biology research lab reports the first three-dimensional structure of a vault from a prokaryotic system, specifically one encoded by a bacteriophage. Relatives of the vault protein had been spotted in such genomes before, but the post says nobody had shown they assemble into vaults. A search across a database of about 15 billion predicted proteins found the relatives rare but widely spread: present in 29 bacterial phyla and four archaeal ones, but in under 1% of species with a sequenced genome. Of candidates taken forward, seven formed vault-like particles under an electron microscope and four were confirmed as vaults by cryo-electron microscopy. The one solved in detail, called L05, is built from 96 copies of a single protein against 78 in the mammalian version, with a combined mass of about 9.5 megadaltons; it is shorter and wider than the mammalian vault yet encloses about 25% more volume because it is more spherical. The post describes its end caps closing in three tiers, with identical copies taking different shapes at the pole. Altered versions showed that some single changes near the cap leave the shell open rather than sealed, and adding wall repeats produced larger vaults, which the post reads as a plausible route by which vaults could have changed size over evolutionary history.
The post also presents this as a demonstration of working method: Claude searched the sequence databases, designed the genetic constructs, helped devise a faster isolation route (about 90 minutes where the conventional approach takes roughly three days), and ran the structure-determination software under the team's direction, with a first model ready 34 hours later against the weeks they estimate it would otherwise have taken. The whole project ran in under four months. One unexplained observation is noted: a small molecule sits in a pocket in each protein copy, clearly visible but unidentified, which Claude suggested resembles coenzyme A, an early hint, the post says, and nothing more.
Who did the work and who checked it
The post is credited to Januka Athukoralage, Peter Yoon, Nicholas Perry, Matthew Durrant and Anthropic's Life Sciences Team. The detail is in a preprint submitted to bioRxiv, so not peer reviewed at publication. Leonard Rome, who co-discovered vaults in 1986, was brought on to advise and is quoted; the post discloses that he was a paid consultant who gave feedback on the manuscript. The team invites outside researchers with ideas about vault function to contact them, and says structures and additional data will be released.
Caveats the post itself gives
The function of vaults remains undetermined: the post says its findings only deepen the mystery, and quotes Rome to the same effect. The particles have so far been made only in a laboratory bacterium, not in the organisms where these proteins naturally occur, and the post states this explicitly. The unidentified molecule in the structure is flagged as unexplained. The post is also candid about where Claude cost them time: a protocol it wrote contained an error that set the project back two weeks, and in the structural processing it skipped a standard step and produced a flawed map, prompting the instruction "Please let's do normal things unless they don't work." Mistakes were caught quickly in areas the team knew well and more slowly in areas they did not. One step used hundreds of GPUs in parallel, beyond what the post expects a typical academic lab to have.
How it connects
- Teaching AI to build scientific models is listed beside it as related work from the same program.
- Introducing our Science Blog describes the Science blog this lab publishes through.
References
- "Investigating vault particles with Claude." Anthropic Science blog, 4 November 2026. https://www.anthropic.com/research/investigating-vault-particles-with-claude record