Anthropic Claude Vault Protein Structure Explained: First Bacterial Vault Found, Built From 96 Protein Copies
Illustrative image: the human vault particle, built from 78 copies of the major vault protein — the new bacterial vault mapped by Anthropic's team has 96 copies. Image: Wikimedia Commons / Lapenta et al. (bioRxiv 2025), CC BY.
📌 KEY HIGHLIGHTS
- Anthropic announced on 4 November 2026 the first 3D structure of a vault particle found in prokaryotic (bacterial) cells.
- The vault is built from 96 copies of a single bacteriophage protein, L05 — 48 copies per half — measuring about 54 × 45 nanometres and weighing roughly 9.5 megadaltons.
- Claude searched about 15 billion predicted proteins to find vault relatives, then helped design DNA, speed up purification and process cryo-EM data.
- Claude also made real mistakes: one protocol error cost the team two weeks, and an early map came out blended with its mirror image.
- The work is a preprint submitted to bioRxiv and has not been peer-reviewed; the particles have so far only been made in E. coli.
What happened?
On 4 November 2026, Anthropic's biology research lab published its account of a four-month project in which its AI model Claude helped solve a structure that has puzzled biologists for 40 years. Vault particles are large, hollow barrels of protein found in nearly all the cells in our body. They were discovered in 1986 by Leonard Rome and Nancy Kedersha, and their biological function is still unknown — despite being about three times the mass of a ribosome, making them among the largest known molecular complexes in complex cells.
Until now, vaults were known only from eukaryotes — organisms with complex cells, like mammals. Anthropic's team showed that related proteins in bacteria and the bacteriophages (viruses) that infect them can also assemble into complete vault-like particles, and mapped one of them in full 3D. The team calls it the first 3D structure of a vault particle found in prokaryotic cells.
The vault, in numbers: the new bacterial vault vs the human vault
The phage-derived vault, built from a protein the team calls L05, is shorter and wider than the human version — but holds about 25% more volume inside, because it is more spherical.
| Feature | New bacterial vault (L05) | Human vault |
|---|---|---|
| Protein copies | 96 (48 per half) | 78 (39 per half) |
| Size | ~54 × 45 nanometres | ~67 × 40 nanometres |
| Mass | ~9.5 megadaltons | ~13 megadaltons |
| Wall repeats per protein | 7 | 9 |
| Structure resolution | 3.2 ångströms (cryo-EM) | Previously solved |
| Enclosed volume | ~25% more than human vault | Baseline |
One detail stands out: the cap of the shell closes in three steps, like a tiered dome. In every group of three neighbouring protein copies, one ends at the first tier, one at the second, and one continues to the tip. The team also engineered the vault: duplicating two of the protein's seven wall segments gave it nine — the same number as the human vault — and produced closed particles with about 20% more enclosed volume. That suggests vaults could be resized on purpose, a step towards using them as nanocapsules that carry enzymes or drugs.
Structure of the vault complex from a rat liver cell, showing the barrel shape that gives vaults their name. Image: Wikimedia Commons.
How Claude helped — and where it went wrong
Claude was involved at every stage, and Anthropic is unusually honest about both the wins and the failures. Using Claude Science, the team searched a database of about 15 billion predicted proteins and found vault relatives in 29 major groups of bacteria and four groups of archaea — rare, but widely spread. They chose 44 candidate proteins to test; 15 of 24 produced soluble protein in E. coli, seven of those formed vault-like particles, and cryo-electron microscopy (cryo-EM) confirmed the vault shape in four of them — two from phages and two from bacteria.
The speed gains were real. The conventional way to purify vaults — spinning samples through increasingly dense sugar solution — takes about three days. Claude helped the team switch to size-exclusion chromatography, cutting purification to about 90 minutes. Claude then ran the open-source RELION software to process the cryo-EM data, producing a first full structural model in 34 hours — work the team estimates would normally have taken weeks. The whole project took under four months.
But Claude also caused setbacks the team discloses in full. A protocol it wrote included a growth-medium additive that blocked DNA from entering cells, delaying the project by two weeks. And an early cryo-EM run skipped a standard step, producing a flawed map blended with its mirror image — the researchers told Claude to stop improvising. Anthropic says it shared its lab notes with the teams that train Claude, and that newer models made fewer of these errors as the project went on.
Why it matters for World Tech readers
This is not a new product or a chatbot upgrade — it is a demonstration of AI doing real science, with the receipts. A small team with limited structural-biology experience solved a structure in under four months because Claude filled the gaps: searching billions of protein sequences, designing DNA, optimising lab methods and crunching microscope data. Anthropic says most of that workflow could be repeated by a typical lab using existing Claude products — only the cryo-EM analysis, which used hundreds of GPUs in parallel, needed more computing power than most universities have.
It also matters because the finding itself is genuinely new biology. A bacterial vault gives researchers a far simpler system to study than the human one: bacteria grow fast, and the particle can now be made, isolated in 90 minutes and resized at will. Each of the 96 protein copies also holds a pocket containing a small molecule the team has not yet identified — a computational search of 25,000 molecules points to coenzyme A or a related molecule, which could be an early hint about what vaults actually do. For India, where biotech and AI research are both growing fast, it is a concrete example of how AI agents are changing what a small lab can attempt. The same week, Baidu's ERNIE 5.5 Flash Base open model showed how quickly AI labs are now shipping big open models — the tools and the science are accelerating together.
Should you care?
✅ Yes, if you follow AI or biology — this is one of the clearest demonstrations yet of an AI model driving a real discovery end to end, with honest reporting of its mistakes. ⏳ But keep the caveats in mind: the result is a preprint, not peer-reviewed; the vaults have only been made in engineered E. coli, not in the bacteria and phages where the proteins naturally occur; and nobody yet knows what vaults do. The breakthrough is the method and the structure — not a cure or a product. For context on how fast AI capabilities are moving, see IBM's AI magic-state factories for fault-tolerant quantum computing and OpenAI's Images 2.6 GIF creation.
FAQs
What is a vault particle?
A vault particle is a large, hollow, barrel-shaped shell made of protein, found in most complex cells. It was discovered in 1986 and is about three times the mass of a ribosome, yet its biological function is still unknown after 40 years of study.
What did Anthropic's Claude actually discover?
Working with Anthropic's biology team, Claude helped find and map the first vault particle from bacterial life: a 96-copy shell built from a bacteriophage protein called L05, resolved by cryo-EM at 3.2 ångströms. Claude searched 15 billion predicted proteins, designed DNA constructs, sped up purification and processed the microscopy data.
Is the Anthropic vault finding peer-reviewed?
No — not yet. Anthropic has submitted a technical preprint to bioRxiv and plans to release the structures and additional data soon. Treat it as a credible but unreviewed result until independent replication and peer review land.
What could vault particles be used for?
Researchers have proposed using vaults as nanocapsules to deliver enzymes or drugs, since their hollow interior can carry cargo. Anthropic showed the shell can be resized by duplicating wall segments, but cargo delivery has not been demonstrated — the proven result is control over the shell's geometry.
Takeaway: Claude helped map a 96-copy bacterial vault in under four months — purification in 90 minutes, a first model in 34 hours — but the honest story includes its mistakes, an unreviewed preprint, and a 40-year-old mystery that is still unsolved.
Follow World Tech on TechByCountry for tomorrow's 6 PM update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Comments
Post a Comment