On November 4, 2026, Anthropic's biology research lab published an account of using Claude to investigate one of the odder puzzles in cell biology: vault particles, giant protein shells found in almost all of our cells, discovered 40 years ago, whose job is still unknown.
In under four months, a small team that, in its own words, had "limited experience in structural biology" used Claude to search for vault proteins across the tree of life, design and make dozens of candidates, and solve the first vault structure found in prokaryotes, the group of single-celled life that includes bacteria. They didn't solve the mystery of what vaults do. But how they got this far says a lot about where AI in science is heading.
This article covers what the team did, what Claude did, where it went wrong, and what could come next. In the future section, we label what Anthropic showed, what Anthropic says could follow, and our own analysis separately, so you can tell facts from forecasts.
What vault particles are
A vault is a large, hollow protein shell. Under the electron microscope, its ribbed shape reminded its discoverers, Dr. Leonard Rome and Dr. Nancy Kedersha, of the vaulted ceiling of a cathedral, which gave it its name in 1986.
- Structure: in mammals, the shell is 78 copies of a single protein, the major vault protein, arranged as two identical halves joined at the middle.
- Size: roughly three times the mass of a ribosome, making vaults among the largest known molecular complexes in complex cells.
- Function: unknown. Proposed roles include transport in and out of the cell nucleus, protection against toxic compounds such as cancer drugs, and innate immunity. None is confirmed, and mice without vaults survive.
- A research blind spot: vaults are missing from the favorite model organisms of biology, including baker's yeast, fruit flies, and nematode worms, which is one reason they've been hard to study.
What the team did, step by step
1. Searching the tree of life
The team started with Claude Science, Anthropic's research product, to search for relatives of the vault protein across a database of about 15 billion predicted proteins. They found that vault-protein relatives are rare but widespread: present in 29 major groups (phyla) of bacteria and four of archaea, but in less than 1% of species with a sequenced genome.
2. Designing and making candidates
Claude turned each chosen protein sequence into DNA suited to the host cell, added the control sequences that switch a gene on, and prepared the designs for ordering. The funnel looked like this:
| Step | Result |
|---|---|
| Candidates tested in a cell-free system | 44 |
| Best performers moved into E. coli | 24 |
| Produced soluble protein | 15 of 24 |
| Assembled into vault-like particles | 7 of 15 |
| Vault shape confirmed by cryo-EM | 4 of 7 (two from phages, two from bacteria) |
3. Purifying them: three days down to 90 minutes
This is the most human part of the story. Claude "first argued against" using a sizing column for purification, then recommended a specific column and ways to track the particles. It also proposed loading the contents of broken-open cells directly onto the column. The team was reluctant, since columns are expensive and raw cell contents can clog them, but it worked. A purification that conventionally takes three days took about 90 minutes. The team notes it's not the first group to isolate vaults this way.
4. Solving the structure
Working in an internal multi-agent setup with access to Anthropic's compute, Claude installed the open-source cryo-EM software RELION and ran each processing step, while the scientists advised on analysis decisions. A first structural model was ready 34 hours after processing began, which the team estimates would normally have taken weeks. Final maps were ready about a week after processing started. The cryo-EM step used hundreds of GPUs in parallel, the only part of the project that needed more computing than an academic institution might have.
What they found
The lead particle, called L05, comes from a protein found in a phage (a virus that infects bacteria):
| Mammalian vault | L05 | |
|---|---|---|
| Protein copies | 78 (39 per half) | 96 (48 per half) |
| Size | About 67 × 40 nm | About 54 × 45 nm |
| Wall repeats | 9 segments | 7 segments |
| Enclosed volume | Baseline | About 25% more |
The team also engineered the particle. Of 15 altered versions, some single changes near the cap stopped it from closing, producing open, rosette-like particles. Duplicating two wall segments (giving nine, as in mammals) increased the enclosed volume by about 20%, which hints that vaults may have changed size over evolution by copying pieces of their own wall.
One more tantalizing detail: each of the 96 copies has a pocket holding a small molecule, and Claude noticed its shape resembles coenzyme A, a molecule central to metabolism. That's unconfirmed. If it holds up, it would be an early hint at what vaults do.
Where Claude went wrong
Anthropic is candid about the mistakes, and they're as instructive as the wins:
- A two-week setback. A protocol Claude wrote called for a growth medium with an additive that interferes with getting DNA into cells.
- A mirror-image map. During cryo-EM processing, Claude skipped a standard step and produced a flawed map blended with its mirror image. The team's response has become a good motto for working with AI: "Please let's do normal things unless they don't work."
- It got better. The team shared its lab notes with the researchers who train Claude, and says Claude "got much better at these tasks as new models arrived."
Who did what
| Claude | People |
|---|---|
| Searched about 15 billion predicted proteins | Chose which proteins to test |
| Designed the DNA constructs | Did the hands-on lab work |
| Recommended the purification approach, and argued against one plan | Decided to try the approach, and checked the results |
| Set up RELION and ran the cryo-EM processing | Advised on analysis decisions and caught errors |
| Spotted the coenzyme A-like shape | Treated it as unconfirmed |
What the future holds
Here's how we'd separate the signal from the hype.
What Anthropic actually showed
- A small team, working outside its expertise, completed a full research cycle in under four months: search, design, make, purify, and solve a new structure.
- Some steps got dramatically faster: purification from days to 90 minutes, and a first model in 34 hours instead of an estimated few weeks.
- The AI made consequential mistakes that people had to catch.
What Anthropic says could come next
- Figuring out what vaults do. The team plans to continue, asking whether all vaults do the same thing or whether their function evolved.
- Open data. It says it will release the structures and more data soon, and it has submitted a preprint to bioRxiv. As of today, the structures aren't public yet.
- Biotechnology. Anthropic suggests vault-like particles "could be adapted for biotechnology," for example as nanocapsules for delivering enzymes. That's presented as a possibility, not a result.
Our analysis: what else could be possible
This section is our own reasoning, not Anthropic's claims.
- Small labs doing big-lab science. The most important line in the post may be about computing: apart from cryo-EM processing, nothing needed more than an academic institution might have. If AI handles the specialist software and the protocol design, a small lab or a lab at a teaching university could take on projects that used to need a dedicated structural biology group. Programs like Anthropic's AI for Science credits push in the same direction.
- Reviving "orphan" mysteries. Vaults sat half-understood for 40 years partly because they're missing from standard model organisms. AI that can search billions of proteins and test dozens of candidates quickly could reopen other neglected questions: proteins and structures nobody had the time or funding to chase.
- AI as a lab partner that pushes back. The purification breakthrough came from Claude arguing against a plan, not just following orders. Expect more research workflows built around AI proposing alternatives, with people deciding.
- Design-build-test loops that run in weeks. The wall-duplication experiment shows how quickly you can go from "why is it this size?" to testing it. Engineered protein containers for enzymes, sensors, or drug delivery are plausible long-term directions, but any medical use would still take years of testing and regulatory review.
- New kinds of jobs. The team lacked deep structural biology experience but knew enough to steer and check the work. That points to a growing role for scientists who combine domain knowledge with strong AI skills, much like the shift we describe in our OODA loop guide.
What hasn't changed
- Hands in the lab. Claude designed and processed; people grew, purified, and imaged.
- Verification. Both big mistakes would have shipped without expert review.
- Expensive steps. Hundreds of GPUs for cryo-EM processing is beyond most labs today.
- Peer review. The findings were released early, and the preprint hadn't gone live at the time of the announcement. Treat specific numbers as preliminary until the paper and data are out.
- Safety rules. Faster biology raises real dual-use concerns. Biology work with Claude is covered by Anthropic's usage policy, which restricts work that could contribute to weapons or other serious harms, and lab work still goes through institutional biosafety review.
What you can do now
- Researchers: see our guide to Anthropic's AI for Science program, which covers partner credits, direct applications, and the discounted Claude Team plan for scientists.
- Students and career changers: pair a science background with AI skills. Start with our AI learning path, and if you want to build tools, our review of Building with the Claude API.
- Everyone: use the same habits the team did. Let AI propose and do the routine work, check its output against standard practice, and keep people on the decisions.
Check your understanding
Answer each question from memory before opening it. Recalling, not rereading, is what makes learning stick (here's why).
1 What is a vault particle?
A large, hollow protein shell found in nearly all the cells in our body. In mammals it's made of 78 copies of one protein, arranged as two identical halves. Discovered in 1986, its function is still unknown.
2 What was new about the structure Anthropic's team solved?
It was the first vault structure found in prokaryotes. The lead particle, from a protein found in a phage (a virus that infects bacteria), has 96 copies of one protein instead of 78 and encloses about 25% more volume than a mammalian vault.
3 Name two things Claude did in the project, and two things people still did.
Claude searched a database of about 15 billion predicted proteins, designed the DNA constructs, recommended the purification approach, and ran the cryo-EM data processing. People chose which proteins to test, made the analysis decisions, and did the hands-on lab work.
4 What mistakes did Claude make, and what's the lesson?
It wrote a protocol with a growth-medium additive that blocks getting DNA into cells, costing two weeks, and it skipped a standard cryo-EM step, producing a flawed map mixed with its mirror image. The lesson: expert review and standard checks still matter, even when the AI is fast.
5 Which future uses does Anthropic itself mention, and how sure is it about them?
It suggests vault-like particles could be adapted for biotechnology, such as nanocapsules for delivering enzymes, and it plans to keep studying what vaults do. Both are presented as possibilities, not results.
6 Why does this study matter for small labs, even if they never study vaults?
Apart from the cryo-EM processing, which used hundreds of GPUs, Anthropic says nothing in the project needed more computing than an academic institution might have. It suggests AI can help small teams work outside their expertise, as long as they check the work.
Sources
- Anthropic: Investigating vault particles with Claude (November 4, 2026)
- Anthropic: preprint (PDF), submitted to bioRxiv
- Claude Science
- Kedersha and Rome, 1986: Isolation and characterization of a novel ribonucleoprotein particle, Journal of Cell Biology
Dr. Rome advised the project as a paid consultant, as disclosed in Anthropic's post.
← All AI & Workflows