What did Claude actually do in Anthropic's vault particle project?

In Anthropic's vault project, Claude searched about 15 billion predicted proteins for vault relatives, designed DNA for each candidate protein, suggested loading broken-open cells straight onto a chromatography column, and ran the RELION cryo-EM software in a multi-agent harness. Human scientists ran the lab work, and Dr. Leonard Rome, who co-discovered vaults in 1986, advised as a paid consultant.

What mistakes did Claude make in the vault research?

Anthropic lists two Claude errors in the vault research. One protocol Claude wrote used a growth medium additive that interferes with getting DNA into cells, which cost about two weeks. Early in cryo-EM processing, Claude skipped a standard step and produced a flawed map, and the team told it to stop improvising.

How is the bacterial L05 vault different from the mammalian vault?

The L05 vault from a bacteriophage has 96 copies of one protein (48 per half), while mammalian vaults have 78 copies of the major vault protein. L05 measures about 54 by 45 nanometres against 67 by 40 for the mammalian shell, encloses about 25% more volume, and has seven repeating wall segments instead of nine.

Are the vault structures and data available to download?

Not yet. Anthropic published a PDF technical report and submitted it to bioRxiv, and says it will update the link once the preprint is live. The authors say they will be releasing the vault structures and additional data, and they invite ideas about what vaults do by email.

Anthropic · 2026-11-04 · major

Vault particles with Claude — Anthropic solves a bacterial vault structure

In short

Anthropic's biology lab used Claude to find and solve the first 3D structure of a vault particle from a prokaryote: a 96-copy, 9.5-megadalton protein shell from a bacteriophage. The project took under four months.

Anthropic · 2026-11-04 · 3 sources checked

Anthropic research post artwork for Investigating vault particles with Claude

Claude helped a small Anthropic lab team find, build and image a giant protein shell no one had seen in bacteria.

Quick facts

Maker	Anthropic Life Sciences team
Result	First 3D structure of a prokaryotic vault (L05)
Size	96 protein copies, about 9.5 megadaltons
Search space	About 15 billion predicted proteins
Project time	Under four months
Outputs	Technical report; structures and data to follow

What is it?

Vaults are large, hollow protein shells found in many cells; their job is still unknown, nearly 40 years after their discovery in 1986. Anthropic's biology team used Claude to search for vault relatives outside animals and solved the first 3D structure of one from a prokaryote, a bacteriophage protein called L05. The L05 vault is built from 96 copies of one protein and is shorter and wider than the mammalian version, with about 25% more room inside.

How does it work?

Through Claude Science, Claude searched a database of about 15 billion predicted proteins and found vault relatives in 29 bacterial and four archaeal phyla. It turned each protein sequence into DNA for E. coli, proposed a faster purification method that cut a three-day step to about 90 minutes, and set up and ran the open-source cryo-EM software RELION inside an internal multi-agent harness. Seven of 15 tested proteins formed vault-like particles, and cryo-EM confirmed four.

Why does it matter?

The post shows an AI model doing hands-on parts of a structural biology project, not only reading papers, while staying honest about its errors. Claude wrote one protocol with an additive that blocked DNA uptake, costing about two weeks, and it skipped a standard cryo-EM step, which produced a flawed map. The engineered results, such as larger vaults made by duplicating two wall segments, give researchers new shells to study for packaging and delivery.

Who is it for?

structural biologists, protein engineers, AI-for-science researchers

Frequently asked questions

What did Claude actually do in Anthropic's vault particle project?: In Anthropic's vault project, Claude searched about 15 billion predicted proteins for vault relatives, designed DNA for each candidate protein, suggested loading broken-open cells straight onto a chromatography column, and ran the RELION cryo-EM software in a multi-agent harness. Human scientists ran the lab work, and Dr. Leonard Rome, who co-discovered vaults in 1986, advised as a paid consultant.
What mistakes did Claude make in the vault research?: Anthropic lists two Claude errors in the vault research. One protocol Claude wrote used a growth medium additive that interferes with getting DNA into cells, which cost about two weeks. Early in cryo-EM processing, Claude skipped a standard step and produced a flawed map, and the team told it to stop improvising.
How is the bacterial L05 vault different from the mammalian vault?: The L05 vault from a bacteriophage has 96 copies of one protein (48 per half), while mammalian vaults have 78 copies of the major vault protein. L05 measures about 54 by 45 nanometres against 67 by 40 for the mammalian shell, encloses about 25% more volume, and has seven repeating wall segments instead of nine.
Are the vault structures and data available to download?: Not yet. Anthropic published a PDF technical report and submitted it to bioRxiv, and says it will update the link once the preprint is live. The authors say they will be releasing the vault structures and additional data, and they invite ideas about what vaults do by email.

Try it

https://www-cdn.anthropic.com/76e55fbe85697afbaae16b839498dc00797dec30.pdf