Claude helped map mysterious cell particle in bacteria

Anthropic's biology team has determined the first 3D structure of a so-called vault particle from bacteria. Vaults are small, barrel-shaped particles inside cells that were previously known from mammals. According to Anthropic, Claude played a central role in the work.

The work has been submitted as a preprint to bioRxiv (as of November 4, 2026). That means it has been made public but has not been documented as reviewed by other researchers (peer review).

Searching among 15 billion proteins

Claude searched databases of about 15 billion proteins for relatives of vault proteins. The search found them in 29 bacterial lineages and four lineages of archaea, which are another group of single-celled organisms. However, the proteins appeared in fewer than 1% of species with sequenced genomes.

The bacterial vault consists of 96 copies of the protein, compared with 78 in mammalian vaults. It weighs about 9.5 megadaltons and is 54 nm long and 45 nm wide. Despite its shorter length, it has about 25% greater volume than mammalian vaults.

Of 15 candidate proteins that yielded soluble protein, 7 formed vault-like particles under an electron microscope. Four structures were confirmed with cryo-EM, a method in which samples are frozen and imaged in an electron microscope. The team also made 15 altered protein versions to investigate how the particles assemble.

Claude's role

Claude helped with literature and protein searches, design of DNA constructs, optimization of isolation (including the choice of chromatography columns), tracking of vault proteins and processing of microscopy data. Claude also steered the structure determination itself in the RELION software.

According to Anthropic, purifying the particles took about 90 minutes, compared with about three days using the conventional method. The first structural model was ready after about 34 hours, compared with typically weeks. The figures are Anthropic's own and have not been independently verified. The entire project took about four months, so the short times apply to individual steps.

Mistakes along the way

Claude also made mistakes. It first argued against chromatography, which turned out to be wrong. It wrote a protocol with an incompatible growth medium, which caused a delay of about two weeks. In data processing, it skipped standard steps, so the maps of the structure were flawed, and at times it improvised inappropriately. According to Anthropic, the mistakes were costlier in areas where the team itself was less familiar.

What vaults are actually used for is still unclear. Leonard Rome, who was a paid consultant and adviser on the project, says that the function of vaults is still very difficult to determine.