Anthropic says Claude helped map the first 3D structure of a vault particle from bacterial life

Anthropic said on November 4 that its biology research team, working with Claude, has determined the first three-dimensional structure of a vault particle from prokaryotic life. Vaults are large, barrel-shaped protein shells that cell biologists have observed since 1986, when Leonard Rome and Nancy Kedersha first described them, yet their function has never been confirmed. Science's news section also reported on the work, framing it as research that deepens the mystery around these particles rather than solving it.

What the team found

Until now, vaults had been characterised structurally only in eukaryotes such as mammals, where the shell is made of 78 copies of a major vault protein. Anthropic's team looked for distant relatives in bacteria and in bacteriophages, the viruses that infect bacteria. Of 15 candidate proteins tested in the lab, seven assembled into vault-like particles and four structures were confirmed by cryo-electron microscopy, two bacterial and two phage-derived.

The most detailed result concerns L05, a phage protein that forms a particle of 96 copies, 48 per half. According to the reported figures, it measures about 54 by 45 nanometres and weighs roughly 9.5 megadaltons. Although shorter than the mammalian vault, it encloses about a quarter more volume. By duplicating two segments of its wall, the researchers produced an even larger shell.

What Claude did, and where it went wrong

Anthropic says Claude searched about 15 billion predicted protein sequences, finding vault relatives across 29 bacterial and four archaeal groups. It designed DNA sequences to express candidates in E. coli, proposed a chromatography-based purification method that cut a roughly three-day step to about 90 minutes, and ran cryo-EM processing software. Interesting Engineering reports that the first complete structural model took 34 hours, against an estimate of several weeks with conventional analysis. Human scientists carried out the lab work, and Rome advised the project as a paid consultant.

The company also listed mistakes. A protocol written by Claude included a growth-medium additive that blocked DNA uptake, costing about two weeks, and Claude skipped a standard processing step, producing a flawed map. Researchers caught both problems and tightened supervision.

Open questions

The structure does not answer what vaults are for. Anthropic has submitted a technical report to bioRxiv, but at publication the preprint had not appeared, the work had not been peer reviewed, and the structures and supporting data had not yet been released.