Anthropic Says Claude Helped Solve First 3D Structure of a Bacterial-Type Vault Particle

Vaults are large, hollow, barrel-shaped particles that Leonard Rome and Nancy Kedersha first identified in 1986. Until now they were known mainly from animals and other eukaryotes, and their biological function is still unexplained. For Anthropic, the bigger point is that an AI system took part in most of the laboratory and computational work, from searching protein databases to processing the cryo-electron microscopy data. That supports the company's case that its models can speed up experimental science, not only software work.

The team says the project took less than four months, from screening candidates to a solved structure.

From Database Search to Shell

The researchers used Claude to search a database of about 15 billion predicted proteins. They found relatives of the major vault protein in 29 bacterial phyla and four archaeal phyla, although these appear in fewer than 1 per cent of species with sequenced genomes. They tested 44 candidates in a cell-free system. Of the 24 they then moved into E. coli, 15 produced soluble protein and seven formed vault-like particles under the electron microscope. Cryo-EM confirmed a vault shape for four of them: two from phages and two from bacteria.

The best-characterised particle, from a phage protein the team calls L05, is built from 96 copies of a single protein, 48 in each half. Mammalian vaults have 78 copies. According to Anthropic, the L05 shell weighs about 9.5 megadaltons, measures roughly 54 by 45 nanometres, and holds about 25 per cent more volume than the mammalian version. It was resolved at a nominal 3.2 angstroms. When the researchers duplicated two of the protein's seven wall repeats, the particles still closed and held about 20 per cent more volume. The authors suggest this could lead to nanocapsules of adjustable size, though cargo delivery remains an idea, not a result.

Where the Model Helped and Where It Failed

Anthropic says the researchers picked the candidate proteins. Claude turned the sequences into DNA designs for ordering, helped build a size-exclusion purification method that cut the time from cells to a microscope grid from about three days to about 90 minutes, and ran the open-source RELION cryo-EM pipeline inside an internal multi-agent setup. A first full structural model was ready 34 hours after processing began. Claude had initially argued against the purification approach that ended up working.

The report also lists the model's mistakes. A protocol Claude wrote included a growth-medium additive that interfered with getting DNA into cells, which cost two weeks. An early processing run skipped a standard step and produced a map blended with its mirror image, and humans had to catch it. The company also says the parallel cryo-EM analysis used hundreds of GPUs, more than a typical lab has.

The findings are limited. The particles have only been produced in E. coli, and the team has "yet to test whether they form in the bacteria and phages where they're naturally found". Each L05 copy has a pocket holding an unidentified molecule. A computational screen of 25,000 compounds pointed to coenzyme A or a derivative, but this has not been confirmed. Rome, who was paid to consult and gave feedback on the manuscript, said: "Figuring out the function of vaults is still enormously challenging."

Anthropic says it will release the structures and more data later. The bioRxiv preprint will be the first chance for outside structural biologists to examine the work in detail.