Anthropic Says Claude Helped Map First Prokaryotic Vault Particle Structure

At a glance

The publish date is November 4, 2026, from the Anthropic research page and @AnthropicAI. The claim is the first 3D structure of a vault from prokaryotic cells, from the research post. The particle is phage protein L05: 96 copies (48 per half) and about 9.5 megadaltons, from the research post and the preprint PDF. Size is about 54 by 45 nanometers, shorter and wider than mammalian vaults at about 67 by 40 nanometers, from the research post. Enclosed volume is about 25% more than mammalian vaults, from the research post. Resolution is a nominal 3.2 Å cryo-EM map, from the preprint PDF.

Anthropic says Claude’s role covered a database search, DNA constructs, an isolation method, and RELION processing, from the research post. Speed claims are isolation in about 90 minutes versus about three days, and a first model in 34 hours against an estimate of weeks, from the research post. The preprint was submitted to bioRxiv, and the PDF is hosted on Anthropic’s CDN, from the research post and the PDF.

What Anthropic claims

Vaults are large, hollow, barrel-shaped protein shells discovered in 1986 by Leonard Rome and Nancy Kedersha. Their biological function is still unknown. Mammalian vaults are built from 78 copies of the major vault protein. Anthropic says related proteins were already known in prokaryote and phage genomes, but—to its knowledge—no one had shown they assemble into vaults.

The detailed structure is of L05, a bacteriophage protein produced in E. coli. Cryo-EM shows 96 identical chains (48 per half), a mass of about 9.5 megadaltons, seven wall repeats instead of nine, flatter caps, and a more spherical, egg-like shape that encloses about 25% more volume despite being shorter. The preprint titles the work Prokaryotic homologs of the major vault protein form engineerable vault nanocapsules and reports a nominal resolution of 3.2 Å.

Anthropic also says seven of fifteen expressed candidates formed vault-like particles under electron microscopy; cryo-EM confirmed four (two phage, two bacterial). Cap-sealing mutants produced open rosettes, and duplicating two wall repeats (to nine, matching human MVP) enlarged enclosed volume by about 20%.

How Claude was used—and where it failed

Under human direction, Claude searched a database of about 15 billion predicted proteins, designed DNA constructs for 44 candidates, helped optimize size-exclusion chromatography so purification fell from roughly three days to about 90 minutes, and drove open-source RELION processing in an internal multi-agent harness. Anthropic says a first full model was ready in 34 hours, with final maps about a week later—time savings it estimates against weeks of conventional work.

The same post documents failures. A Claude-written protocol used a growth-medium additive that interfered with DNA delivery and cost about two weeks. Early cryo-EM processing skipped a standard step and produced a flawed mirror-blended map until researchers intervened. Anthropic says lab notes were shared with model trainers and that later Claude versions made fewer construct-design errors. Leonard Rome, brought on as an adviser, is quoted on the still-unsolved function problem; Anthropic discloses he was a paid consultant who gave manuscript feedback.

Limits worth stating precisely

The X announcement summarized “bacterial life.” The research note is more precise: the solved shell is a phage-derived vault that would be built inside a bacterium, and related bacterial proteins also assembled in the screen. Particles have so far been made only in E. coli; Anthropic has not tested whether they form in the native bacteria and phages. Function remains unknown. The bioRxiv link was not live when the post said it would be updated; the technical PDF is available from Anthropic’s CDN. Speed comparisons are Anthropic’s own estimates.

Bottom line

Anthropic’s November 4 note is a first-party claim that Claude-assisted wet-lab and cryo-EM work produced the first solved prokaryotic vault structure—an engineerable L05 nanocapsule—while openly logging model mistakes and leaving biological function unsolved. Treat assembly, dimensions, and timing as Anthropic (and preprint) claims until independent groups reproduce them.