Anthropic says Claude helped solve the first structure of a bacterial vault, but the data are not yet public
What happened
On 4 November 2026, Anthropic published a research post titled “Investigating vault particles with Claude.” It reports what the company calls the first 3D structure of a vault particle found in prokaryotic cells, that is, in bacteria or the viruses that infect them. The contributors listed are Januka Athukoralage, Peter Yoon, Nicholas Perry, Matthew Durrant and Anthropic’s Life Sciences Team. Leonard Rome, who co-discovered vaults in 1986, advised the project; the post discloses that he was a paid consultant who gave feedback on the manuscript.
Vaults are large, hollow protein shells found in most human cells. Their function is still unknown. The post describes the team using Claude to search a database of about 15 billion predicted proteins for relatives of the vault protein, finding hits in 29 bacterial and 4 archaeal phyla. The team picked candidates, and Claude converted each sequence into DNA designed for expression in E. coli. The team chose 44 candidate proteins; in the expression tests the post reports, 15 of 24 produced soluble protein and 7 formed vault-like particles; four were confirmed by cryo-electron microscopycryo-electron microscopy A method that images flash-frozen molecules with an electron beam and combines many pictures into a 3D model..
The headline particle is built from 96 copies of a single protein from a bacteriophagebacteriophage A virus that infects bacteria., which the team calls L05. Anthropic says it weighs about 9.5 megadaltons, measures about 54 by 45 nanometres and encloses about a quarter more volume than a mammalian vault. Each copy holds an unidentified small molecule; Claude noted it resembles coenzyme A, but the team says it has not identified it.
Anthropic also reports process details. Claude suggested loading cell contents straight onto a size-exclusion columnsize-exclusion chromatography A way of separating molecules by size, by passing a mixture through a column packed with porous beads., which the post says cut purification from about three days to about 90 minutes. Working in an internal multi-agent setup, Claude installed and ran the open-source cryo-EM software RELION, while the scientists advised on analysis decisions. A first full model took 34 hours. The whole project took under four months.
What is claimed vs. what is verified
Nearly everything above is Anthropic’s own account. The company says a preprintpreprint A research paper shared publicly before, or without, formal peer review by a journal. has been submitted to bioRxiv and that it will release the structures and more data soon. At the time of the post, no Protein Data Bank or Electron Microscopy Data Bank accession codes were given, so outside researchers cannot yet inspect the maps or models. The post does not name the Claude model version used, and it does not state a resolution figure; one secondary site reports 3.2 ångströms, but that number does not appear in the post itself.
Independent coverage so far has repeated Anthropic’s account rather than tested it. Interesting Engineering’s report, published the same day, matches the post’s figures and quotes no outside scientists; it also describes the result as the world’s first bacterial vault, slightly broader than the post’s wording, and does not mention that Dr. Rome was paid.
The post does report its own failures: a protocol additive Claude recommended set the project back two weeks, and an early processing run produced a map blended with its mirror image. Claude also initially argued against the chromatography approach that later worked.
Why it matters
If the structure holds up, it is evidence that vault-forming proteins exist outside complex cells, which bears on how these shells evolved. It is also one of the more detailed public descriptions of a frontier AI model doing hands-on parts of an experimental biology project, from database search to image processing. The time figures, if typical, would matter for how labs plan structural work.
Open questions and other views
The authors note that the particles have only been made in E. coli and that no one has yet checked whether they form in their natural bacterial and phage hosts. Their function remains open; Dr. Rome is quoted saying that working out what vaults do “is still enormously challenging.”
Supporters of AI in science would point to the speed and to a fully disclosed list of the model’s mistakes as a sign of a workable division of labour. Sceptics would note that the scientists chose the candidates and directed each step, so it is hard to separate Claude’s contribution from expert guidance, and that the post comes from a company with a commercial interest in showing its model is useful for research. Structural biologists will be able to judge the work once maps and models are deposited. Until then, the result is a well-described claim rather than an independently checked one.
Sources
- Anthropic, “Investigating vault particles with Claude,” research post, 4 November 2026. https://www.anthropic.com/research/investigating-vault-particles-with-claude
- Aamir Khollam, “Anthropic maps world’s first 3D vault built from 96 protein copies,” Interesting Engineering, 4 November 2026. https://interestingengineering.com/ai-robotics/anthropic-first-3d-bacterial-vault-map