Anthropic used Claude to solve the first structure of a bacterial-world 'vault' particle. What the AI did, and where it slipped

Anthropic says its life-sciences team used Claude to determine the first 3D structure of a prokaryotic vault particle, a large hollow protein shell, in this case encoded by a bacteriophage (a virus that infects bacteria). The company published the work on November 4 with a technical report it says it has submitted to bioRxiv.

Why it matters: vaults have puzzled cell biologists since Leonard Rome and Nancy Kedersha found them in 1986. They turn up across the tree of life, yet nobody has pinned down what they do. Anthropic says this is the first vault structure from a prokaryote (bacteria and archaea, plus the phages that infect them). Just as interesting for the tech audience is how the work was done: Anthropic describes Claude doing much of the search, protocol design and cryo-EM processing, with people steering. It also lists where the AI went wrong.

What they found

The particle studied in detail comes from a phage protein the team calls L05, produced in E. coli. According to Anthropic:

Despite being shorter, L05 encloses about 25% more volume because it is rounder, Anthropic says. The team also made 15 altered versions. Some single changes near the cap stopped the shell from closing, giving open, rosette-like particles. Adding two wall segments (for nine in total, like the mammalian version) produced a bigger vault with about 20% more enclosed volume. Anthropic suggests vaults may have changed size during evolution in the same way.

Each of the 96 chains holds an unidentified small molecule in a pocket. Its shape resembles coenzyme A, which the team flags as a possible early hint about function, not a finding.

What Claude did

Anthropic lays out the pipeline step by step:

1. Search. Using Claude Science, Claude scanned about 15 billion predicted proteins for vault relatives. It found them in 29 bacterial and four archaeal phyla, but in fewer than 1% of species with a sequenced genome.
2. Design. Researchers picked 44 candidates. Claude turned the sequences into DNA tailored for the host and prepared the orders.
3. Expression. Of 24 candidates moved into E. coli, 15 gave soluble protein and seven assembled into vault-like particles. Cryo-EM confirmed the vault shape for four: two from phages and two from bacteria.
4. Purification. Claude first argued against size-exclusion chromatography. When the team went ahead, it proposed loading lysed cells straight onto the column. Anthropic says that cut a roughly three-day purification to about 90 minutes.
5. Structure. Claude set up and ran the open-source cryo-EM package RELION inside an internal multi-agent harness, with humans advising on choices. A first model was ready 34 hours after processing began, and final maps about a week later. The team estimates the usual approach would take weeks.

The whole project took under four months, according to the company. Interesting Engineering's independent write-up reports the same core numbers.

Where it slipped

Anthropic does not present this as a flawless run:

• A protocol Claude wrote called for a growth medium with an additive that interferes with getting DNA into cells. That cost about two weeks.
• Early in cryo-EM processing, Claude skipped a standard step and produced a flawed map blended with its mirror image. The researchers' instruction afterwards: "Please let's do normal things unless they don't work."
• The analysis used hundreds of GPUs in parallel, more than many academic labs can get.

Caveats

• This is a company report, not yet peer reviewed. The authors are Anthropic staff. Leonard Rome advised as a paid consultant, Anthropic discloses.
• The vaults were made in E. coli only. The team has not shown they form in the native phages or bacteria.
• No function yet. By Anthropic's own account, the result deepens the vault mystery rather than solving it.
• The blog post does not give a numerical resolution for the maps, and says the structures and more data are still to be released.

What to watch

The bioRxiv preprint and the deposited structures are the real test: outside structural biologists can then check the maps and the claimed speed-ups. For labs weighing AI agents, the useful takeaway is the split of work. Claude did the search and the routine processing fast, but the two errors were caught by people who knew what "normal" should look like.