Anthropic uses Claude to solve the first 3D structure of a prokaryotic “vault”: AI as a lab partner, flaws included

By Anthropic · November 4, 2026.

Anthropic's life sciences team has published the first three-dimensional structure of a “vault” particle found in prokaryotes (single-celled organisms such as bacteria). According to the company, the work is in a preprint submitted to bioRxiv, and the structures and additional data will be released later. Everything that follows comes from Anthropic's description, not from peer review.

A “vault” is a hollow protein capsule that is huge by cellular standards: roughly three times the mass of a ribosome. It was discovered in 1986 and is present in almost every human cell, but nobody yet knows what it is for. Mice without vaults survive normally, which suggests they may only be needed under specific conditions.

The central finding: related proteins from bacteria and bacteriophages (viruses that infect bacteria) do assemble into complete particles. According to Anthropic, this had not been demonstrated until now. The vault studied, called L05, uses 96 copies of a protein, compared with 78 in the human version. It measures about 54 × 45 nanometers and encloses around 25% more volume than the mammalian one. Using cryo-electron microscopy (cryo-EM, which freezes many copies and images them with an electron beam), they confirmed particle formation in four of seven candidates.

What interests us most is the method. Anthropic says Claude searched for relatives of the protein in a database of about 15 billion predicted proteins, designed the DNA for 44 candidates, optimized the isolation and processed the microscopy data with the RELION software. A purification that usually takes three days was cut to about 90 minutes, and a first structural model was ready in 34 hours, something that, according to the authors, would otherwise have taken weeks. The entire project, from screening to structure, took less than four months.

The piece is remarkably candid about the failures. A protocol written by Claude included an additive that prevented DNA from entering the cells and cost two weeks. In cryo-EM, the model skipped a standard step and produced a map mixed with its mirror image, until the researchers told it to stop improvising. They also report that Claude initially advised against size-exclusion chromatography, which ended up working, and proposed loading broken cells directly onto the column, an idea the humans were skeptical of and which proved valid.

**Our take.** Our view is that the value of this episode lies not in the vault but in the working pattern: a team with limited experience in structural biology tackled a new problem because the model covered what they didn't know. It is a real democratization of expert knowledge, and it fits our underlying view: the path to eradicating diseases runs through accelerating basic science, which today advances at the pace of weeks and months.

But it should be assessed with our feet on the ground. The authors acknowledge that in areas they knew well they caught errors quickly, while in those they didn't, the errors cost them time. That is the underlying risk: AI amplifies those who know how to verify it and can drag down those who don't. Nor is there yet any biological function or validation in the bacteria and phages where these vaults occur naturally; they have only been produced in E. coli.

There are also two contextual caveats. This is research by the very company that sells the model, and the outside consultant cited, Leonard Rome, was paid to review the manuscript, which Anthropic discloses. And although it states that most of the work can be replicated with existing Claude products, the cryo-EM analysis used hundreds of GPUs in parallel, beyond the reach of many academic labs.

More broadly, AI-assisted science is moving from “the model predicts” to “the model runs the hypothesis-experiment loop.” In the short term, that will reshuffle which skills matter in a lab: less procedural know-how, more judgment to spot when the agent is wrong. In the long term, if that loop truly speeds up, it is exactly the lever we need to understand cells, and with them, diseases. The finding that the function of vaults remains a mystery, and probably a complex one, is a good reminder that speed is no substitute for asking the right question.