Claude Mapped a 40-Year-Old Protein Vault Mystery

Anthropic’s biology team has mapped the first three-dimensional structure of a vault particle from bacterial life — and it did the work with Claude in the loop from the first search to the final reconstruction. The particle, built from 96 copies of a single phage-derived protein, is a hollow barrel about 54 nanometres long. It is also a mystery in the most literal sense: vaults were discovered in 1986, and after 40 years nobody knows what they do.

Quick answer: Anthropic used Claude to map the first 3D structure of a bacterial vault particle: a 96-copy protein shell, 54 nm long, whose job is still unknown after 40 years.

Vault particles are among the largest protein structures in our cells, and their function is still unknown. They are hollow, barrel-shaped assemblies first identified in 1986 by Leonard Rome and Nancy Kedersha. In humans and other eukaryotes they are made of 78 copies of a protein called the major vault protein. Until now, nobody had shown that anything like them exists in bacteria or the viruses that infect them — or what such a structure looks like in three dimensions.

How Claude searched 15 billion proteins to find the vaults

The discovery started with a search, not a microscope. Anthropic’s team used Claude Science to search a database of roughly 15 billion predicted protein sequences for distant relatives of the major vault protein. The search found candidates across 29 bacterial phyla and four archaeal phyla — but in fewer than 1% of species with sequenced genomes, which is why vaults in bacteria had stayed hidden for so long.

The funnel from search to structure was narrow, and Anthropic published the numbers:

Stage	What happened	Result
Protein search	Claude Science searched ~15 billion predicted proteins	Vault relatives in 29 bacterial + 4 archaeal phyla
Candidate testing	44 proteins selected for cell-free expression testing	15 of 24 candidates produced soluble protein
Microscopy screen	Soluble proteins examined under electron microscopy	7 formed vault-like particles
Cryo-EM confirmation	Cryo-electron microscopy of the 7 candidates	4 confirmed vault structures — 2 bacterial, 2 from phages

The team then chose one phage-derived protein, called L05, for detailed structural analysis, because it was among the best at forming complete, closed particles. All four confirmed structures were produced in laboratory-grown E. coli — a detail that matters, and one the team states plainly.

What the L05 vault actually looks like

The bacterial vault is bigger in copy number, shorter, wider — and roomier inside — than the human one. Cryo-electron microscopy resolved the L05 shell at a nominal resolution of 3.2 angstroms, and Claude helped process the cryo-EM data through the open-source RELION software, run inside an internal multi-agent harness.

Feature	Human (mammalian) vault	Phage-derived L05 vault
Protein copies	78 (39 per half)	96 (48 per half)
Wall segments per protein	9 repeating segments	7 repeating segments
Dimensions	~67 nm long × ~40 nm wide	~54 nm long × ~45 nm wide
Mass	Not the headline	~9.5 megadaltons
Enclosed volume	Baseline	~25% more than the human vault

The cap of the L05 shell closes in three steps: within each group of three neighbouring protein copies, one ends at the first tier, another at the second, and the third continues to the tip. The team also tested whether the structure could be redesigned — duplicating two wall segments gave L05 nine repeats and produced closed particles with roughly 20% more enclosed volume.

There is one more detail inside the shell. Each L05 protein copy contains a pocket holding an unidentified small molecule. A computational search of 25,000 molecules suggested coenzyme A, or a derivative of it, as a candidate — but neither the molecule’s identity nor any functional implication has been confirmed, and the article should not present it as settled.

The part Anthropic did not hide: where Claude went wrong

The most useful part of Anthropic’s report is the failure log. Claude contributed throughout — protein searches, DNA design, purification and microscopy analysis — and it also made two mistakes the team disclosed in full.

First, a growth protocol written by Claude contained an additive that blocked DNA uptake into the host cells, delaying the project by two weeks. Second, during cryo-EM processing, Claude skipped a standard step and produced a flawed structural map that blended the particle’s image with its mirror image. Human researchers caught the error.

The speed gains were real, and they are Anthropic’s own figures. An optimised purification step cut the cells-to-grid preparation from roughly three days to about 90 minutes, and the first complete 3D model emerged within 34 hours, though the final reconstruction took approximately a week. The point of the story is not that AI did the science alone. It is that a human team, using Claude as a collaborator, got to a structure faster — and caught the machine’s errors before they became the result.

What this proves — and what it does not prove

This is a real result with a clearly marked boundary. The findings appear in a technical report submitted to bioRxiv as a preprint, which means they have not been peer-reviewed or independently replicated as of 5 November 2026. Anthropic says the structures and additional data will follow. No new model or code was released with the work.

Claim	Status
Vault particles exist in bacterial and phage proteins	Demonstrated in E. coli, confirmed by cryo-EM
First 3D structure of a bacterial vault	Reported 4 November 2026; preprint, not yet peer-reviewed
L05 vault is 96 copies, 54 × 45 nm, ~9.5 MDa	Measured by cryo-EM at 3.2 Å resolution
The vault’s biological function	Still unknown — unchanged since 1986
The particle forms in its native bacteria or phages	Not tested — shown only in lab-grown E. coli
The pocket molecule is coenzyme A	Unconfirmed candidate from a computational search
Claude can run a biology project alone	Not shown — humans caught both of Claude’s errors

That boundary is what separates this story from the usual AI-hype cycle. Anthropic’s own benchmark context makes the same point from the other direction: the best frontier models still score only around 12% on multi-hour real-world agentic tasks in SemiAnalysis’s GDPx benchmark, as covered in our earlier piece on that benchmark. A model that helps map a protein in 34 hours and still needs a human to catch a mirrored map is a collaborator, not a replacement.

Our take: five questions to ask of any AI-assisted science claim

Treat every “AI solved X” headline as a claim with a checklist, not a verdict. This vault story passes most of the checks, which is exactly why it is worth covering — and the checks are the original element of this article. Before sharing the next AI-science breakthrough, ask:

Is there a preprint or paper, or only a post? Here there is a technical report submitted to bioRxiv. A social-media announcement alone is not evidence.
Was it peer-reviewed or independently replicated? Here, not yet — and the article says so in the headline area, not in a footnote.
Did the AI work alone, or did humans check it? Here the failure log shows both: two weeks lost to a bad protocol, one mirrored map caught by humans.
What was actually produced, and where? Here, particles in lab-grown E. coli — not yet shown forming in the bacteria or phages where the proteins naturally occur.
What is still unknown? Here, the function of vaults themselves. Forty years after discovery, that question is untouched — and the honest article leaves it open.

If a claim fails the first two questions, treat it as a rumour. If it passes them but hides the failure log, treat the speed numbers as marketing. The vault story is strong precisely because Anthropic published the errors alongside the result.

Frequently Asked Questions

What is a vault particle?

A vault particle is a hollow, barrel-shaped protein structure — among the largest protein assemblies found in cells. Vaults were discovered in 1986, and their biological function remains unknown. Human vaults are made of 78 copies of the major vault protein; the newly mapped bacterial vault uses 96 copies of a phage-derived protein called L05.

How did Claude help map the bacterial vault?

Claude searched roughly 15 billion predicted protein sequences to find vault relatives, helped design DNA and purification steps, and processed cryo-electron microscopy data through the RELION software. Anthropic reports the first complete 3D model emerged within 34 hours, and an optimised purification step cut preparation from about three days to 90 minutes.

Did Claude do the science on its own?

No. Claude made two disclosed mistakes: a growth-protocol additive that delayed the project by two weeks, and a skipped step in cryo-EM processing that produced a flawed mirrored map. Human researchers caught both errors. The work is a human-led project with Claude as a collaborator.

Has the finding been peer-reviewed?

Not yet. Anthropic submitted the technical report to bioRxiv as a preprint, and the structures and additional data are to follow. As of 5 November 2026, the result has not been independently replicated, and the particles have only been shown forming in laboratory-grown E. coli, not in their native bacteria or phages.

Why does it matter if nobody knows what vaults do?

Because it shows where AI-assisted science is genuinely useful right now: searching vast protein databases, speeding up lab workflows, and reconstructing structures — while humans still supply the judgement, the error-checking, and the open questions. The unknown function is not a flaw in the story; it is the next forty years of work.

Sources and methodology

Sources consulted and verified on 5 November 2026: Anthropic’s announcement of 4 November 2026 (via its @AnthropicAI post and technical report submitted to bioRxiv); Unite.AI, “Anthropic Uses Claude to Show Bacterial and Phage Proteins Form Vaults“; Interesting Engineering, “Anthropic maps world’s first 3D vault built from 96 protein copies“; The Neuron, “Claude Helps Investigate a 40-Year-Old Cell Mystery“; AlphaSignal, “Anthropic’s Claude Maps a Mysterious Cellular Vault in 90 Minutes“. All figures — 15 billion proteins searched, 44 candidates, 96 copies, 54 × 45 nm, ~9.5 megadaltons, 3.2 Å resolution, 34 hours, 90 minutes, two-week delay — are Anthropic’s own, reported by at least two independent outlets, and labelled as such where they rest on the company’s account alone. The preprint status and the E. coli-only limitation are stated from the same sources. This article was desk-researched; no hands-on lab work was performed, and no claim is made from personal testing. Verification date: 5 November 2026.

Arva Rangwala

Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.