Anthropic's in-house biology lab has published the first 3D structure of a vault particle from a prokaryotic cell, and the more useful news for most readers is how the work was done. Vaults are large, hollow protein barrels found across most of the tree of life, from slime moulds to humans, and nobody has confirmed what they are for. The team says Claude was built into every step, from searching sequence databases to processing microscope data, and that this compressed weeks of work into days.
The particle comes from a protein made by a bacteriophage, a virus that infects bacteria. Using cryo-electron microscopy, the team found the shell is assembled from 96 copies of a single protein, 48 in each half, where the human vault uses 78. It weighs about 9.5 megadaltons, which Anthropic says puts it among the largest protein shells identified in prokaryotes. Its wall is built from seven repeating segments rather than the nine in mammalian vaults, and its caps close in three tiers, with neighbouring copies of the same protein taking different shapes at the top.
The team then changed the protein in 15 ways, mostly where neighbouring copies touch. Some single changes near the cap stopped the shell from closing and left open, rosette-like particles. Adding two extra wall segments, matching the human count of nine, produced larger vaults with about 20% more internal volume. That suggests one way vaults could have changed size over evolutionary time, and why researchers interested in protein nanocapsules for delivering enzymes may pay attention.
The claim about speed is specific. Anthropic says purifying vaults usually takes three days and took about 90 minutes here, and that the first structural model was ready 34 hours later, against an estimated few weeks by hand. Claude, working under the scientists' direction, searched protein databases, designed DNA constructs, tuned the isolation method and processed the microscopy data. This follows the same lab's earlier report of an enzyme system found with Claude, and it is the clearest public example yet of Claude used as a working tool in wet-lab biology rather than as a chatbot.
The limits matter. The findings are in a preprint submitted to bioRxiv and have not been peer reviewed, the structures and extra data are promised but not yet released, and the speed figures are Anthropic's own estimates. The central question, what vaults actually do, is still open; Dr Rome, a vault researcher who advised the team, calls working out their function enormously challenging. Readers weighing Claude for research work should treat this as a strong case study from its maker, not an independent benchmark.