Claude helped solve a bacterial vault structure in 34 hours, Anthropic says

Anthropic’s biology research team says it has solved the first 3D structure of a vault particle from a prokaryote, with Claude handling much of the work. In its announcement, the company says a first structural model took 34 hours.

A phage vault built from 96 copies of one protein

Vaults are large, hollow protein barrels found in cells across most of the eukaryotic tree, from slime molds to humans. Leonard Rome and Nancy Kedersha discovered them in 1986, and their function is still unknown. Anthropic’s team looked at a vault-like protein from a bacteriophage, a virus that infects bacteria, which it calls L05, and produced it in E. coli.

Using cryo-electron microscopy, the team found that the shell consists of 96 copies of that single protein, 48 per half, for a combined mass of about 9.5 megadaltons. Mammalian vaults use 78 copies. The phage vault measures roughly 54 by 45 nanometers, against 67 by 40 in mammals. Anthropic says it is shorter but encloses about 25% more volume because it is more spherical. Its wall has seven repeating segments rather than nine, and its caps are flatter.

The caps close in three steps, like a tiered dome. Within each trio of adjacent chains, the three finish at different heights: one stops on the lowest tier, another on the middle tier, and the third reaches the apex. Along the wall the chains are identical, yet in the cap they fold into three distinct shapes.

The team also made 15 altered versions of the protein. Some single changes near the cap produced open, rosette-like particles instead of sealed vaults. Duplicating two wall segments, which gives nine as in the human vault, produced larger particles with about 20% more enclosed volume.

Where Claude cut the time

The workflow began with Claude Science searching a database of about 15 billion predicted proteins. Vault relatives turned out to be rare but widely spread: 29 bacterial phyla and four archaeal ones, yet under 1% of sequenced species. The team picked 44 candidates, and Claude converted each sequence into DNA suited to the host cell. Of 24 proteins moved into E. coli, 15 yielded soluble protein, and seven of those formed vault-like particles under the microscope. Cryo-EM confirmed the shape for four, two from phages and two from bacteria.

The conventional purification spins samples through a sugar gradient and takes about three days. Anthropic says it switched to size-exclusion chromatography instead. Claude first argued against this, because misfolded clumps elute in the same fraction as vaults, but then helped make it work. It also suggested loading ruptured cells straight onto the column. The result, according to Anthropic, is about 90 minutes from cells to microscopy grid. Claude ran the open-source RELION software for the cryo-EM processing, and final maps arrived about a week after processing began.

According to Anthropic, a typical lab with access to current Claude products could carry out most of this work. The exception was the cryo-EM processing, which ran on hundreds of GPUs at once and needs more computing than many universities can offer.

The mistakes are part of the report

Anthropic is unusually frank about errors. A protocol written by Claude called for a growth medium with an additive that interferes with getting DNA into cells, costing about two weeks. Early in cryo-EM processing it skipped a standard step and produced a map blended with its mirror image. The team’s instruction was: “Please let’s do normal things unless they don’t work”. The company says it shared its lab notes with the researchers who train Claude, and that the model made fewer construct design errors as newer versions arrived. It also says the team caught mistakes quickly in areas it knew well, but lost time in areas it did not.

One open item is a pocket in each of the 96 copies that holds a small molecule the team has not identified. Claude noted that its shape resembles coenzyme A. Anthropic treats that only as a possible early hint about function.

A simpler system for an old puzzle

The figures give some context for why a bacterial vault matters. According to Wikipedia’s entry on the organelle, the human vault weighs about 13 megadaltons, and most human cells contain around 10,000 of them, with some immune cells holding up to 100,000. The same entry says vaults are overexpressed in many multidrug-resistant cancer cells, though a causal link is unproven, and that mice lacking vault proteins show no major changes. That combination of abundance and apparent dispensability is what has kept the function a mystery. Bacteria grow quickly, so a vault that can be made in E. coli gives researchers a faster system for testing ideas.

Anthropic points to precedent: bacterial versions of cGAS-STING signaling, gasdermins and viperins revealed the microbial origins of human immune defenses. For vaults, the human function is unknown, so the bacterial relatives may help reveal it.

What the report does not settle

The function of vaults remains unknown, and the work only deepens the question. Rome, whom Anthropic brought on as an adviser, said: “Figuring out the function of vaults is still enormously challenging.” Anthropic discloses that he was a paid consultant who gave feedback on the manuscript.

The claims come from the company’s own account and a preprint that it says it submitted to bioRxiv but that was not yet live at publication. It has not been independently verified or peer reviewed. The team has so far made the vaults only in E. coli and has not tested whether they form in the bacteria and phages where they occur naturally. Time savings such as the weeks-long estimate for the first model are Anthropic’s own estimates. The company says it will release the structures and additional data soon, and it invites researchers with ideas or test systems to write to the authors.