Anthropic's biology lab has published the first 3D vault particle structure from a prokaryote: a 96-copy shell from a bacteriophage, produced in E. coli, set against the 78-copy human vault. For anyone who wants to experiment, the bacterial vault is the better model, because it is simpler and quicker to make and can be altered at will. For anyone interested in medicine, the human vault still matters more, and this result does not tell us what either one does (Anthropic research post, preprint; published 4 November 2026).
Last verified: 5 November 2026.
TL;DR
- The new vault particle structure comes from a phage protein named L05, built from 96 copies (48 per half), against 78 copies in humans (Anthropic).
- Claude turned electron microscopy images into a first structural model in 34 hours, a job that can otherwise take weeks (Science).
- Isolation dropped from about three days to about 90 minutes using size-exclusion chromatography (Anthropic).
- Vault function is still unknown, and the result deepens that mystery rather than solving it.
- Claude made errors that cost the team roughly two weeks, and humans had to rein it in during cryo-EM processing.
- The work is a preprint and has not yet been peer reviewed.
What is a vault particle?
A vault is a large, hollow, barrel-shaped protein shell found across most of the eukaryotic tree, from slime moulds to humans. Leonard Rome and Nancy Kedersha discovered them in 1986, and the ribbed shell reminded them of a cathedral's vaulted ceiling. Roughly 40 years on, nobody knows what they are for (Anthropic; Gadgets Now).
The mammalian vault has 78 copies of the major vault protein, 39 per half, and is about 67 nanometres long by 40 wide. Proposed roles include shuttling molecules between nucleus and cytoplasm, shielding cells from toxic compounds such as cancer drugs, and supporting innate immunity. None is confirmed. Mice bred without vault components still develop normally, and vaults are absent from baker's yeast, fruit flies and nematode worms, which removed the usual routes to studying them.
How does the bacterial vault compare with the human one?
The two shells share a design but differ in proportions. The prokaryotic particle is shorter, wider and more spherical, so it encloses more volume despite being shorter.
| Property | Human (mammalian) vault | Bacterial/phage vault (L05) |
|---|---|---|
| Protein copies | 78 (39 per half) | 96 (48 per half) |
| Mass | About 3x a ribosome | About 9.5 megadaltons |
| Size | About 67 x 40 nm | About 54 x 45 nm |
| Enclosed volume | Baseline | About 25% more |
| Wall segments per chain | 9 | 7 |
| 3D structure | Solved earlier by cryo-EM | First prokaryotic structure, 4 Nov 2026 |
Source for all figures: Anthropic. The end caps are flatter in L05 and close in three steps, like a tiered dome. In each group of three neighbouring copies, one chain stops at step one, one at step two and one reaches the tip, so three identical chains adopt different shapes.
How did the team find and make the vault?
Related vault-protein sequences had been reported in prokaryote and phage genomes before (2023 bioRxiv abstract), but nobody had shown they assemble into vaults. According to Anthropic, the steps were:
- Search about 15 billion predicted proteins with Claude Science (Anthropic). Vault relatives turned out to be rare but widespread: 29 bacterial phyla and four archaeal groups, yet under 1% of species with a sequenced genome (Anthropic).
- Choose 44 candidates and test them in a cell-free system. Claude converted each sequence into host-tailored DNA with control sequences.
- Move the best into living E. coli. Fifteen of 24 produced soluble protein.
- Isolate particles by size-exclusion chromatography, about 90 minutes from cells to microscopy grid (Anthropic).
- Image under the electron microscope. Seven of the 15 proteins formed vault-like particles, and cryo-EM confirmed the shape for four (two phage, two bacterial).
- Run cryo-EM processing in RELION, directed by Claude in a multi-agent harness. The first full model took 34 hours and the final maps about a week (Anthropic).
- Test 15 altered versions of the protein.
The chromatography step is worth a closer look. Claude initially argued against it, since clumps of misfolded protein elute in the same void volume. The team went ahead, and Claude then suggested a specific column and proposed loading ruptured-cell contents straight on with no prior purification. That was unconventional, columns are expensive and raw E. coli lysate risks clogging them, but it worked.
What did the mutation experiments show?
Some single changes near the cap stopped the shell closing, leaving open, rosette-like particles instead of sealed vaults. That identifies the cap contacts as the region that seals the shell. Duplicating two wall segments, taking the count from seven to nine as in human vaults, produced particles with about 20% more enclosed volume (Anthropic). This suggests how vaults may have changed size through duplication during evolution, though it is a lab inference, not an observed history.
Claude also flagged an unexplained feature: a pocket partway up each of the 96 copies holds a small molecule, visible in the map but unidentified. Its shape resembles coenzyme A. That is an early hint at function, nothing more.
Where did Claude go wrong, and who deserves the credit?
The failure list is the most useful part for anyone planning similar work. A protocol Claude wrote called for a growth-medium additive that interfered with DNA delivery, setting the project back roughly two weeks. Early in cryo-EM processing it skipped a standard step and produced a flawed map blended with its own mirror image. The team told it to stop improvising. The whole project took under four months (Anthropic).
On credit, Science's report is blunt: it is hard to tell exactly how much belongs to Claude and how much to Anthropic's financial resources, given the billions of dollars its investors have put in (Science). Anthropic says most of the work could be repeated by a typical lab with existing Claude products. Only the cryo-EM analysis, which used hundreds of GPUs in parallel, exceeded what an academic institution usually has. Independent coverage describes Claude as a research assistant in a longer experimental chain, not an autonomous discoverer (Fathom News).
What has not been shown yet?
No one has shown these particles form in the bacteria or phages where they occur naturally, only in E. coli constructs. Function remains open. Rome, now an advisor to the lab, said working out vault function is still enormously challenging and that the breadth of the finding points to a complicated answer (Anthropic). The preprint was submitted to bioRxiv, and the PDF is hosted on Anthropic's CDN.
Engineered vaults are already being explored as nanocapsules for delivering enzymes (PNAS, 2005), so a malleable bacterial version has practical appeal. Bacterial relatives of human proteins have been informative before, with cGAS-STING, gasdermins and viperins revealing the microbial origins of human immune defences.
Where does this fit with other AI-in-science work?
This is the lab's second major shared finding, after a CRISPR-like enzyme system reported in September 2026 (Anthropic). For wider context, see how Claude Science and multi-agent orchestrators are organised, how AI-designed viruses from Evo 2 raise different questions, how cheminformatics agents compare, and where agentic AI in health care currently stands.
FAQ
Q: Who solved the first prokaryotic vault particle structure?
A: Anthropic's biology lab, working with Claude, reported it on 4 November 2026 as a preprint. The structure comes from a phage protein called L05 built from 96 copies.
Q: How long did Claude take to process the cryo-EM data?
A: Claude produced a first full structural model in 34 hours and final maps about a week after processing began (Anthropic). Manual work with specialist software can take weeks or months (Science).
Q: What is the difference between a human vault and the bacterial vault?
A: The human vault has 78 protein copies and measures about 67 x 40 nm (Anthropic). The bacterial L05 vault has 96 copies, measures about 54 x 45 nm and encloses roughly 25% more volume (Anthropic).
Q: Does this result explain what vaults do?
A: No. Function is still unknown after about 40 years, and Science describes the result as deepening the mystery.
Q: Could another lab reproduce this work?
A: Anthropic says most of it could be reproduced with existing Claude products. The cryo-EM analysis, which used hundreds of GPUs in parallel, is the part beyond a typical academic budget.
Q: Is the vault research peer reviewed?
A: Not yet. It is a preprint submitted to bioRxiv, so treat the conclusions as provisional until review is complete.
This article was researched and drafted with AI assistance and checked against the primary sources linked above.
Top comments (0)