RSS
Research

Anthropic used Claude to map the first prokaryotic vault particle in 3D

Anthropic's biology lab directed Claude through finding, isolating and mapping the first 3D structure of a vault particle from prokaryotes: 96 copies of a phage protein, about 90 minutes from cells to microscope grid, a 34-hour first model, and the mistakes along the way.

a close up of an orange and blue substance
Photo by National Institute of Allergy and Infectious Diseases on Unsplash

Updated November 20, 2026: Expanded with the structure details, what Claude did and got wrong, and the now-posted preprint.

Anthropic's biology research lab has reported the first 3D structure of a vault particle found in prokaryotes, built from a protein encoded by a bacteriophage, in a preprint that has not been peer reviewed. Claude worked at every stage under the scientists' direction, from searching sequence databases to processing the microscope data. Anthropic's write-up went up on November 4 and the bioRxiv preprint on November 5. Anthropic says the project took under four months.

What a vault particle is

Vaults are large, hollow barrels of protein, discovered in 1986 by Leonard Rome and Nancy Kedersha. They sit in most cells of the human body and are roughly three times the mass of a ribosome, yet nobody has pinned down what they do. Proposed jobs include moving molecules between the nucleus and cytoplasm, shielding cells from toxic drugs and supporting innate immunity; none has been confirmed. Mice bred without a vault partner protein or the vault RNA develop normally, which suggests vaults may matter only under specific conditions.

Part of the reason the question has stayed open is that vaults are missing from the standard lab organisms: baker's yeast, fruit flies and nematode worms. Bioinformaticians had spotted distant relatives of the vault protein in bacteria and in the phages that infect them, but nobody had shown that those proteins build vaults. Bacteria are quick to grow, so a working prokaryotic vault gives researchers a simpler system to study.

What the team found

Searching about 15 billion predicted proteins, the team found vault-protein relatives in 29 bacterial phyla and four archaeal phyla, but in fewer than 1% of species with a sequenced genome. It tested 44 in a cell-free system, moved the best into E. coli, where 15 of 24 produced soluble protein, and found that seven of those 15 formed vault-like particles under the electron microscope. Cryo-EM confirmed the vault shape for four, two from phages and two from bacteria. (The preprint abstract counts eight homologs isolated as vault particles; the two documents count on different bases.)

The detailed structure is of a phage protein the team calls L05, picked because it was among the best at forming complete, closed particles. Its shell holds 96 copies of that one protein, 48 per half, against 78 in the mammalian vault, with a combined mass of about 9.5 megadaltons. It is shorter and wider than the mammalian version, roughly 54 by 45 nanometres against 67 by 40, and encloses about 25% more volume. The preprint calls it among the largest protein shells of defined size outside virus capsids.

Its wall has seven repeating segments instead of nine. Some single edits near the cap stopped the shell from closing and produced open, rosette-like particles. Adding two wall segments, for nine as in human vaults, gave a larger vault with about 20% more enclosed volume. The map also shows an unidentified small molecule in a pocket on every copy. Claude noted that its shape resembles coenzyme A, a common helper molecule cells use to shuttle chemical groups around; the molecule has not been identified, and the team treats this as an early hint, not an answer.

What Claude actually did

  • Ran the database search through Claude Science and, once the scientists picked 44 candidates, turned each protein sequence into a DNA design tailored to the host cell, ready to order.
  • Helped build a faster isolation method. The usual approach spins samples through a sugar gradient for about three days. Anthropic was not the first to isolate vaults on sizing columns, but its version, with Claude's suggestion to load raw cell contents straight onto the column, took about 90 minutes from cells to microscope grid.
  • Set up and ran the open-source cryo-EM software RELION inside an internal multi-agent harness, with the scientists advising on analysis decisions. A first full model was ready in 34 hours, against weeks the team expected; final maps took about a week.

Claude did not get everything right. It first argued against the chromatography approach that ended up working. One protocol it wrote called for a growth-medium additive that interferes with getting DNA into cells, which cost about two weeks. Early in the cryo-EM work it skipped a standard step and produced a flawed map blended with its mirror image, and the team told it to stop improvising. The scientists say mistakes were caught quickly in areas they knew well but cost time where they did not. They also report that Claude made fewer DNA-design errors as newer models arrived during the project.

Limits and disclosures

The vaults have only been made in E. coli, not in the bacteria and phages where the proteins occur naturally. The preprint offers the shells as a size-tunable scaffold for cargo; no delivery use has been shown. Anthropic says most of the work could be done by a typical lab with existing Claude products, except the cryo-EM analysis, which used hundreds of GPUs in parallel. All preprint authors are Anthropic employees and may hold Anthropic equity; one author declares an outside interest in Stylus Medicine; Anthropic has filed a provisional patent application; and Dr. Rome, who advised, was a paid consultant.

Our take

In our view, this is a useful agent case study because it lists the failures next to the speed-ups. For agent work where experts can review the decisions, the pattern looks worth copying: let the agent run the long, tool-heavy steps, keep experts checking the choices, and expect errors to cost most where the humans know least. Whatever vaults turn out to do, a 90-minute isolation and a 34-hour first model are concrete numbers other labs can test against.

Source: Investigating vault particles with Claude (Anthropic), Prokaryotic homologs of the major vault protein form engineerable vault nanocapsules (bioRxiv preprint)

This story first appeared in the claudefox daily briefing for November 6, 2026.