Anthropic's Claude maps mystery vault particle

Deep inside almost every cell in your body sits a bizarre, hollow protein barrel called a vault particle. It is roughly three times the size of a ribosome, yet ever since its discovery in 1986, biologists have had no idea what it actually does.

Studying vaults has long been a nightmare. Common lab organisms like fruit flies and roundworms do not make them, and mice bred without vaults seem to get along just fine. Now, Anthropic's biology lab has cracked the door open—not by staring harder at mammals, but by using Claude to track down and solve the 3D structure of a vault from prokaryotes for the first time.

From a 3-day chore to 90 minutes

Bacteria and the phages that infect them are far easier and faster to grow than mammalian cells. If vaults could be found and assembled there, researchers would finally have a nimble model to study.

Anthropic had Claude Science scour a database of 15 billion predicted proteins, spotting rare vault candidates across 29 bacterial and 4 archaeal phyla. From 44 selected candidates, Claude helped translate the protein sequences into tailored DNA constructs. When moved into E. coli, 15 produced soluble protein.

Then came the purification bottleneck. The classic eukaryotic method requires spinning samples through sucrose gradients for roughly three days. The team wanted to try size-exclusion chromatography instead. Claude actually argued against the move at first, warning that misfolded junk would come out in the same fraction. But once the scientists decided to push forward, Claude adapted, picked a specific column, and floated a wild idea: dump the raw ruptured cells directly onto the delicate column without prior purification.

It was a risky shortcut that could easily ruin expensive equipment, but it worked. The purification turnaround collapsed from three days to just 90 minutes. Cryo-electron microscopy quickly confirmed that four candidates formed intact vaults, led by a phage protein named L05.

34 hours to a 3D model

Turning noisy cryo-EM images into an atomic-scale 3D map usually requires weeks or months of tedious manual tuning. Anthropic hooked Claude up to an internal multi-agent harness with compute access, letting the AI set up and execute the standard processing software RELION.

With human researchers guiding high-level decisions, Claude produced a full structural model in 34 hours, wrapping up final maps within a week. The phage vault spans 96 copies of a single protein and weighs in at 9.5 megadaltons. It is shorter, wider, and holds about 25% more volume than the 78-copy human version.

Working with Claude, the team even engineered 15 variants. By duplicating two wall segments, they expanded the vault's internal volume by another 20%. Claude also spotted an unassigned density tucked inside a pocket on each protein, noting its shape closely resembles coenzyme A—a tantalizing clue toward its elusive function.

Leonard Rome, who co-discovered vaults 40 years ago and advised on the project, noted that figuring out their function remains enormously challenging, but Anthropic's broad data will inevitably bring more researchers into the fold.

Why it matters

To us, the real headline here is what this says about AI in the wet lab.

Until now, AI in biology mostly meant offline structural prediction like AlphaFold or summarizing papers. What Anthropic demonstrated is Claude acting as an active bench partner across an entire pipeline: debating protocol risks, optimizing messy sample prep, writing DNA designs, and driving specialized domain software via CLI.

Compressing a weeks-long modeling slog down to 34 hours changes the economics of discovery. When routine lab friction shrinks by an order of magnitude, scientists can afford to explore obscure biological mysteries that were previously deemed too slow or unrewarding to touch.

That said, nobody should assume AI is doing science on autopilot just yet. Anthropic admitted on social media that Claude made simple blunders along the way, requiring real-time corrections from researchers. But by pairing an LLM with execution harnesses and human oversight, Anthropic just showed how frontier models can deliver genuine scientific discoveries rather than just slick demos.