Anthropic's lab used Claude to solve the first prokaryotic vault structure, a 96-copy phage shell, and lists where Claude went wrong

Anthropic's in-house biology lab has published what it says is the first 3D structure of a vault particle from a prokaryotic cell, solved with Claude doing most of the hands-on computational work. The write-up, "Investigating vault particles with Claude", appeared on November 4. The preprint has been submitted to bioRxiv but isn't live yet, and the structures and data are promised "soon", so for now the only source is Anthropic's own description.

Vaults are one of the odder open problems in cell biology. They are huge, hollow protein barrels, about three times the mass of a ribosome, found in nearly every cell in the human body. Leonard Rome and Nancy Kedersha discovered them in 1986 and named them after the ribbed ceilings of cathedrals. Forty years later nobody knows what they do: mice bred without them develop normally, and they are missing from yeast, fruit flies and nematode worms, the organisms where geneticists usually look first. Rome advised Anthropic on the project.

What they found

Distant relatives of the vault protein had already been spotted in bacterial, archaeal and phage genomes, but nobody had shown they assemble into vaults. Anthropic's team used Claude Science to search about 15 billion predicted proteins. The relatives turned up in 29 bacterial and 4 archaeal phyla, but in less than 1% of species with a sequenced genome.

From there the funnel was: 44 candidates tested in a cell-free system, 15 of 24 producing soluble protein in E. coli, 7 of those 15 forming vault-like particles under the electron microscope, and 4 confirmed by cryo-EM, two from phages and two from bacteria. The detailed structure is of a phage protein the team calls L05.

The L05 vault is built from 96 copies of one protein (48 per half, against 39 per half in mammals), about 9.5 megadaltons in total. It is shorter and wider than the mammalian vault, roughly 54 by 45 nanometers against 67 by 40, and encloses about 25% more volume. Its caps close like a stepped dome: in every group of three neighboring chains, one stops at the first tier, one at the second and one reaches the tip, so identical chains take three different shapes. Of 15 engineered mutants, some single changes near the cap stopped the shell from closing and produced open "rosette" particles. Duplicating two of the seven repeating wall segments, to match the nine in human vaults, produced larger vaults with about 20% more volume.

There is also a loose end. Each of the 96 chains holds a small molecule in a pocket that the team couldn't identify; Claude noted that its shape resembles coenzyme A. If that holds up, it would be the first hint at what these particles do.

What Claude did, and got wrong

The part Anthropic wants readers to notice is the workflow. Claude turned protein choices into DNA constructs ready to order, recommended a size-exclusion column and a way to track the particles, and processed the cryo-EM data by setting up and running the open-source RELION package in an internal multi-agent harness. A first full model was ready 34 hours after processing began and final maps about a week later. The team says none of its members had worked on a protein assembly this large, and the whole project took under four months.

The best detail is the purification. Claude first argued against size-exclusion chromatography, because clumps of misfolded protein come out in the same fraction as vaults. The team tried it anyway, and Claude then proposed loading raw lysate straight onto an expensive column, which the humans were reluctant to do. It worked: about 90 minutes from cells to microscopy grid, against roughly three days for the standard sugar-gradient method. Anthropic notes it isn't the first group to purify vaults on sizing columns.

The mistakes are listed too. A growth medium Claude specified contained an additive that interferes with getting DNA into cells, which cost two weeks. Early in the cryo-EM work it skipped a standard step and produced a map blended with its own mirror image, until the team told it: "Please let's do normal things unless they don't work." Anthropic says errors were caught quickly where the team had expertise and cost time where it didn't, which is the honest version of the "AI scientist" pitch. Most of the work, it says, could be done with ordinary Claude products; only the cryo-EM step used hundreds of GPUs in parallel.

It is Anthropic's second life-sciences progress note in about ten days, after a report on drug-development work with Claude, and it lands as OpenAI keeps its own strongest biology model, GPT-6 Rosalind, behind application-based access. The vault function is still unknown. As Rome put it: "The broadness of Anthropic's finding[s] suggests that it is going to be a complicated function and an interesting ride."