On this page
Anthropic’s biology research lab published a Science-team research post, *Investigating vault particles with Claude* (https://www.anthropic.com/research/investigating-vault-particles-with-claude). The page marks publishedOn=2026-11-04, with article:modified_time 2026-11-04T19:51:49.000Z (2026-11-05 03:51:49 Asia/Shanghai). The claim at the center is structural, not a product launch: with Claude assisting under human direction, the team reports the first 3D structure of a vault particle built from prokaryotic / bacteriophage-related protein. Vaults—large hollow protein barrels first described in 1986—still have no confirmed biological function. DigIntoAI has not reproduced the experiments, has not run cryo-EM, and has not peer-reviewed the preprint; everything below follows Anthropic’s research post, public posts, and CDN PDF as self-reported.
The same day, @AnthropicAI posted a short thread (2026-11-04T20:11:28.000Z = 2026-11-05 04:11:28 Asia/Shanghai; https://x.com/AnthropicAI/status/2118073665296228863). The root post says the biology team worked with Claude on the first 3D structure of a vault from bacterial life; vaults have been mysterious since 1986 and were previously known mainly from complex animal and plant cells; this round shows that similar proteins from bacteria and the viruses that infect them can also assemble into vaults. A follow-up in the same second (https://x.com/AnthropicAI/status/2118073667858887068) links the preprint PDF.
Related DigIntoAI notes: Anthropic: Early Progress on Accelerating Drug Development with Claude, Anthropic Used Claude to Formally Verify 123 Open-Source Libraries, OpenAI Says Internal Models Produced Math, CS and Physics Results for Academic Shepherds, and UK Opens World's First AI Biosecurity Lab.
What a vault is
A vault is a large, hollow, ribbed protein shell. In mammals it is built from 78 copies of the major vault protein (two identical halves). Researchers have proposed transport, toxin protection, or innate-immunity roles, but none are conclusively confirmed. The Anthropic post is explicit: the lab has not determined vault function. Vaults are not described as “solved,” and no drug, nanocapsule product, or clinical use is claimed as live. Biotechnology suggestions in the post (for example, engineerable nanocapsules) stay forward-looking, not productized.
The post is not a new Claude model ID, not a pricing update, and not an API feature switch. It is an official research hard signal: structure discovery + how Claude was woven into lab steps.
Structure facts (self-reported)
Using cryo-EM, the team reports a vault built from bacteriophage protein L05:
| Piece | Self-reported figure |
|---|---|
| Copies | 96 total (48 per half) vs 78 in mammals |
| Mass | About 9.5 megadaltons |
| Size | About 54 × 45 nm (shorter and wider than the mammalian ~67 × 40 nm) |
| Volume | About 25% more enclosed volume despite being shorter (more spherical) |
| Wall | Seven repeating segments end-to-end (mammals: nine); flatter end caps |
The post frames the particle as among the larger protein shells identified in prokaryotes. Related vault-like proteins had been spotted in prokaryote and phage genomes before; the claimed novelty here is showing they can assemble into complete vault-like particles, with a detailed 3D model for L05.
The lab also describes engineering experiments (altered contact points; duplicating wall segments to enlarge volume ~20%). Those are structural assembly tests—not efficacy or clinical results.
Where Claude helped (assisted—not unsupervised)
Under team direction, Claude is described as helping across the pipeline:
- Search — relatives of the vault protein across a database of about 15 billion predicted proteins (Claude Science named on the page).
- DNA design — turning chosen sequences into host-tailored constructs and order-ready designs (the team chose which proteins to pursue).
- Isolation — moving purification from a conventional multi-day sucrose gradient (~three days) toward size-exclusion chromatography finishing in about 90 minutes from cells to microscopy grid (the post notes Claude first argued against that chromatography path, then helped make it work).
- Structure — directing an internal multi-agent setup through open-source RELION cryo-EM processing; a first full model in about 34 hours (team estimate: weeks in a conventional path), with final maps about a week after processing began.
The post itself notes these boundaries: particles were made in *E. coli* so far; the team has not yet shown they form in the natural bacterial / phage hosts; function remains open; structures and extra data are described as coming soon.
The same post self-reports mistakes—as potential and limits. One protocol called for a growth-medium additive that interferes with getting DNA into cells and cost about two weeks. Early cryo-EM processing skipped a standard step and produced a map mixed with its mirror image, after which the team told the model to stop improvising. Lab notes went back to Anthropic’s ML researchers. DigIntoAI has not audited those incidents beyond the public text.
Preprint status
Anthropic hosts a preprint PDF on its CDN (https://www-cdn.anthropic.com/76e55fbe85697afbaae16b839498dc00797dec30.pdf)—reachable as HTTP 200, application/pdf, about 11.3 MB, Last-Modified 2026-11-04T19:27:51Z (2026-11-05 03:27:51 Asia/Shanghai). The research page says the preprint was submitted to bioRxiv and that the public bioRxiv link will be updated when it goes live. No bioRxiv DOI is listed yet; the CDN PDF is a preprint, not a peer-reviewed journal article.
Advisor Leonard Rome is quoted saying function remains enormously challenging and that others will work on the problem.
DigIntoAI status
DigIntoAI has not reproduced L05 expression, purification, or particle assembly; has not collected cryo-EM or reprocessed RELION; has not independently checked the 9.5 MDa / 54×45 nm / 96-copy figures against deposited coordinates; and has not peer-reviewed the ~11.3 MB PDF. Research posts and preprint links can change; primary sources remain the Anthropic research page, the @AnthropicAI thread, and the CDN PDF.
Quick facts
- What. First reported 3D structure of a prokaryotic / phage-related vault particle (L05), Claude-assisted.
- Who. Anthropic Science / biology research lab; Claude under human direction.
- When. Research page publishedOn 2026-11-04; modified 03:51:49 Asia/Shanghai next calendar morning; @AnthropicAI thread 04:11:28 Asia/Shanghai.
- Structure (self-report). 96 copies; ~9.5 MDa; ~54×45 nm; cryo-EM; ~25% more volume than the mammalian vault despite being shorter.
- Claude steps (self-report). ~15B predicted-protein search; DNA design; purification ~3 days → ~90 min; first full model ~34 h.
- Open. Biological function still unknown; made in *E. coli* so far; natural-host assembly not yet shown.
- Not. New model, pricing, API switch, drug approval, or “function solved.”
- Preprint. CDN PDF available; bioRxiv entry pending.
- Related. Drug-development Claude progress, formal verification, OpenAI academic shepherds, UK AI biosecurity lab.
In short
Anthropic’s Science team says that, with Claude’s help, it has mapped the first 3D structure of a vault particle from prokaryotic / bacteriophage protein (L05; 96 copies; cryo-EM). Vault function remains unknown, work so far is in *E. coli*, and the post is a research hard signal—not a new Claude model or price change. A CDN preprint PDF is downloadable; the bioRxiv listing is still pending. DigIntoAI hasn’t reproduced the lab work or re-checked the structure maps—details above follow the research post, public posts, and preprint as published.