Anthropic said on November 4 that its biology research team, working with Claude, has mapped the first three-dimensional structure of a vault particle from prokaryotic life, a hollow protein shell whose purpose has puzzled scientists since its discovery in 1986. The particle, built from 96 copies of a single protein from a virus that infects bacteria, was produced in E. coli and resolved with cryo-electron microscopy, and the team says the project took under four months from screening to solved structure.
A barrel biology could not explain
Vault particles are large, hollow protein shells found in nearly every cell in the human body, discovered in 1986 by Leonard Rome and Nancy Kedersha and named for the way their ribbed shell recalled a cathedral ceiling. They take their name from that shape, and they have been found in cells across the tree of life.
At roughly three times the mass of a ribosome, they rank among the largest known molecular complexes in complex cells, yet their function remains unknown after four decades of study. That absence of a visible defect is why vaults have resisted explanation, even as their size suggested they must matter.
Researchers have proposed that vaults ferry molecules between the nucleus and cytoplasm, shield cells from toxic compounds such as cancer drugs, or support innate immunity, but none of those ideas has been conclusively confirmed. Mice bred without vault components still develop normally, leaving no obvious defect to chase, which is one reason the mystery has lasted so long.
How Claude helped at every step
The team used Claude Science to search a database of about 15 billion predicted proteins for relatives of the major vault protein, finding them rare but widely spread across 29 bacterial phyla and four archaeal phyla. The proteins were rare, appearing in less than 1 percent of species with sequenced genomes, but their spread suggested an ancient origin.
From 44 candidates tested in a cell-free system, the best moved into living E. coli, where 15 of 24 produced soluble protein and seven formed vault-like particles under the microscope. Cryo-electron microscopy then confirmed the vault shape in four of them, two from phages and two from bacteria.
Claude turned the chosen protein sequences into DNA designs tailored to their host cells and helped rebuild the isolation step: the conventional sugar-gradient method takes about three days, while the team’s optimized size-exclusion chromatography route took about 90 minutes from cells to microscopy grid.
Working in an internal multi-agent harness, Claude ran the open-source RELION processing workflow under researcher direction, producing a first full model in 34 hours and final maps about a week after processing began, against an estimate of weeks for a first model by hand.
The shell itself
The mapped particle comes from a bacteriophage protein the team calls L05, resolved at a nominal 3.2 angstroms. It holds 96 copies of one protein, 48 per half, against 78 copies in mammalian vaults, with a combined mass of about 9.5 megadaltons.
It measures roughly 54 nanometers long and 45 wide, against 67 by 40 in mammals. Shorter and wider, it still encloses about 25 percent more volume, with seven repeating wall segments per chain instead of nine in the human protein.
Its cap closes in three steps: in every group of three neighboring copies, one ends at the first tier, another at the second, and the third continues to the tip. The copies are identical along the wall but take different shapes as the cap narrows.
The team also made 15 altered versions of L05, finding that single changes near the cap could leave open, rosette-like particles instead of sealed shells. Duplicating two wall segments gave the protein nine repeats and produced closed particles with about 20 percent more enclosed volume.
Our read
The strongest part of this story is not the speed alone but the disclosure that comes with it. Anthropic reports that a Claude-written protocol called for a growth-medium additive that blocked DNA uptake and cost the project two weeks, and that an early microscopy run skipped a standard step and produced a flawed map blended with its mirror image.
The researchers directed the work, caught the errors, and fed the lab notes back to the teams that train Claude. Anthropic says later Claude models made fewer DNA-design errors over the course of the project and grew better at spotting experimental pitfalls before they cost time.
That makes the vault a useful test case for AI in science: a real, checkable structure, produced by a documented workflow, with its failures printed alongside its results. It is not yet a drug, a delivery vehicle, or an explanation of what vaults do, and Anthropic is careful to say so.
The comparison that matters is with other labs putting models inside biology, such as OpenAI’s gated biology model, where the claim must survive replication rather than headlines. A structure other teams can rebuild and check is worth more than a speed record alone.
Why It Matters
For working labs, a bacterial vault system is a practical prize: bacteria grow fast, the particles can be made and isolated quickly, and their size can apparently be tuned by duplicating wall segments. That gives researchers a simpler handle on a structure that was previously locked inside slow, complex cells, and a possible route toward engineered nanocapsules.
Anthropic has already been pushing Claude into laboratories, including credits for 20 science groups announced on October 30, and this vault result is the programme’s most concrete output so far. It shows the science push producing basic research, not only faster office work.
The limits are stated plainly and should stay attached to the result. The vault’s function is still unknown, each protein copy holds a pocket containing an unidentified small molecule that only resembles coenzyme A, and the particles have so far been made only in E. coli rather than in the bacteria and phages where the genes naturally occur.
The work is a preprint submitted to bioRxiv, not yet peer-reviewed, and the microscopy analysis used hundreds of GPUs in parallel, beyond what many academic labs can spare. Other labs are moving the same way, with Google’s Antigravity agents entering drug discovery, so independent replication will decide how much weight this first structure carries.
Background
Vaults were found across most of the eukaryotic tree, from slime molds to humans, which convinced researchers they must do something useful, yet they are missing from the organisms biologists rely on most, including baker’s yeast, fruit flies, and nematode worms. A simpler bacterial relative, now mapped in detail, is the first system that lets scientists study their assembly directly.
Source: Anthropic’s vault research and Unite.AI’s full report.
0 Comments