On November 4, 2026, Anthropic’s Life Sciences Team reported the first 3D structure of a vault particle from a prokaryotic cell, built from a bacteriophage protein called L05. Vaults, large hollow protein shells co-discovered in mammals in 1986 by Leonard Rome (an adviser on the project), had not been structurally characterized outside eukaryotes. A preprint was submitted to bioRxiv.
The prokaryotic vault has 96 copies of one protein (48 per half), about 9.5 megadaltons, versus 78 copies in mammalian vaults. It is about 54 by 45 nm and encloses about 25 percent more volume. Its wall has seven repeating segments instead of nine; duplicating two segments enlarged it by about 20 percent in volume, and some single changes near the cap produced open, rosette-like particles.
Claude searched a database of about 15 billion predicted proteins and found relatives in 29 bacterial and four archaeal phyla, in under 1 percent of sequenced genomes. Of 44 candidates tested cell-free, 24 moved into E. coli, 15 gave soluble protein, seven formed vault-like particles and four were confirmed by cryo-EM. The team replaced a three-day density-gradient purification with roughly 90-minute size-exclusion chromatography, and Claude ran the cryo-EM processing in RELION, with final maps about a week after processing began.
Why it matters: it is another case of a model driving most steps of a wet-lab discovery - search, candidate selection, protocol design and structure processing - producing a new structural result rather than a rediscovery.
What it does not show: the vault’s biological function is still unknown, the particles were made only in E. coli, and a bound small molecule resembling coenzyme A is unidentified. Claude also made errors that cost time, including a growth-medium suggestion that blocked DNA uptake for about two weeks and a skipped cryo-EM step that produced a flawed map, and the processing needed hundreds of GPUs.