Anthropic — Investigating Vault Particles with Claudeauto-ingested

A result from Anthropic's in-house biology lab: the first 3D structure of a vault particle in prokaryotes, with Claude used at every step — literature and sequence search, DNA construct design, purification-method optimization, and cryo-EM processing. The lab-notebook detail matters more to this wiki than the biology: it is the wiki's first worked case of AI-accelerated research, including where Claude was wrong. Subscribed-feed capture (Anthropic Research), auto-fetched and auto-ingested 2026-11-08; awaiting human review.

The biology, briefly

Vaults are large hollow protein barrels, discovered in 1986 by Leonard Rome and Nancy Kedersha, found in nearly all cells in the human body, and of still-unknown function after 40 years. They are about three times the mass of a ribosome. Part of why they have stayed unexplained: they are missing from the model organisms geneticists rely on (yeast, fruit flies, nematodes), and mice bred without a vault partner protein or the vault RNA develop normally, so there is no defect to follow.

The team showed that vault proteins in prokaryotes and in the bacteriophages that infect them assemble into complete vault-like particles — related proteins had been found in those genomes before, but nobody had shown they form vaults. The structure solved is of a phage protein they call L05:

	Phage L05 vault	Mammalian vault
Copies of the protein	96 (48 per half)	78 (39 per half)
Mass	~9.5 megadaltons	—
Dimensions	~54 nm long × 45 wide	67 × 40
Enclosed volume	~25% more (more spherical)	—
Wall repeats per chain	7	9

The caps close in three tiers: in every group of three neighbouring copies, one stops at the first step, one at the second, one reaches the tip — identical chains taking different conformations. Single changes near the cap left open, rosette-like particles instead of sealed shells. Copying two wall segments (7 → 9, the mammalian number) built larger vaults, ~20% more volume — a demonstration of how vaults could have changed size over evolution by duplicating their own wall.

Why prokaryotes: bacteria grow fast and give a simpler system, and bacterial relatives of human proteins have previously revealed the microbial origins of human immune defenses (cGAS–STING, gasdermins, viperins). The difference here is that the human function is not already known.

Where Claude was used, step by step

1. Search. Claude Science searched ~15 billion predicted proteins for vault relatives. Result: rare but widely distributed — 29 bacterial phyla, 4 archaeal, in under 1% of sequenced species.
2. Construct design. The humans chose 44 proteins; Claude turned each sequence into host-tailored DNA, added control sequences, and prepared the designs for ordering. 15 of 24 moved into E. coli produced soluble protein.
3. Isolation. The conventional eukaryotic method (density-gradient spin) takes ~3 days. They tried size-exclusion chromatography instead. Claude recommended the column and the tracking/assembly checks, then proposed loading ruptured-cell contents straight onto the column with no pre-purification — unconventional and risky (clogging, contamination, expensive columns). It worked. ~90 minutes from cells to microscopy grid. 7 of 15 proteins had assembled into vault-like particles; cryo-EM confirmed 4 (two phage, two bacterial).
4. Structure. Claude set up and ran open-source RELION in an internal multi-agent harness with internal compute, while the humans "advised on the analysis decisions." First full model in 34 hours; final maps about a week later. The team estimates weeks or months of manual work conventionally.
5. Noticing. Claude flagged an unexplained small molecule in a pocket in each of the 96 copies and observed that its shape resembles coenzyme A — if confirmed, an early hint at function.

Whole project: under four months, by a team where "none of us had worked with a protein assembly this large" and with "limited experience in structural biology." They state that most of it is reproducible by a typical lab with existing Claude products; only the cryo-EM analysis (hundreds of GPUs in parallel) exceeded typical academic compute.

Where Claude was wrong (the most useful part)

• A protocol it wrote called for a growth medium containing an additive that interferes with getting DNA into cells. Cost: two weeks.
• In cryo-EM processing it skipped a standard step and produced a flawed map blended with its mirror image. The humans' instruction: "Please let's do normal things unless they don't work."
• It argued against the chromatography approach (misfolded clumps elute in the same void volume). The humans went ahead anyway and were right — then Claude made the method work.
• The load-bearing sentence: "In areas we knew well, we caught mistakes quickly, but they cost us time in areas we didn't."

Lab notes were fed back to the ML researchers who train Claude, informing training data; the team reports Claude improved at construct design and pitfall-spotting across the project as new models shipped. Verdict: "the benefits of working this way far outweighed the drawbacks."

Key claims (verbatim, for the record)

• "a vault purification that conventionally takes three days took us about 90 minutes, and a first structural model was ready 34 hours later, which we estimate would normally have taken us weeks."
• "Claude first argued against taking the chromatography approach… We went ahead anyway, and Claude helped make it work."
• "we told it to stop improvising."
• "Most of this work could be achieved by a typical lab with access to existing Claude products."
• Leonard Rome (paid consultant on the manuscript): "Figuring out the function of vaults is still enormously challenging… The broadness of Anthropic's finding[s] suggests that it is going to be a complicated function and an interesting ride."

Caveats the authors state: the vaults were made only in E. coli, not yet in the native bacteria and phages; the function remains unknown, and "our work only deepens the mystery." A preprint is submitted to bioRxiv.

Connections

• AI-Accelerated Science — the first case study behind the scenarios; and a challenge to how its model treats physical bench time.
• Verification Requires Domain Knowledge — the "areas we knew well" sentence is this page's main contribution to the wiki.
• AI Use Tracks AI Capability — cryo-EM processing is verifiable and went to Claude; the two failures were in steps where verification was hardest.
• The Human Sandwich — "under our direction," "we advised on the analysis decisions," "we told it to stop improvising" is the pattern in a wet lab.
• Agent Autonomy Spectrum — an internal multi-agent harness run semi-synchronously.
• The Bottleneck Shift — humans chose the question; Claude proposed an unconventional method and a candidate hypothesis, which nibbles at the ideation line.
• AI Commoditizes Expertise — a team without structural-biology expertise solved a structure; the cost of the missing expertise shows up as uncaught errors.
• Falsification Is Asymmetric — Claude's prior-based objection to chromatography was one hypothesis; one experiment settled it.
• Anthropic — the Life Sciences team and Claude Science.
• Method Reflection — T3 (who holds the knowledge).

Sources

• raw/2026-11-04 Anthropic — Investigating vault particles with Claude.md (full text).
• Preprint: bioRxiv 10.64898/2026.11.04.764334v1 (submitted; not read).