Anthropic's Claude Maps a Mysterious Cellular Vault in 90 Minutes
Anthropic's biology lab used Claude to map the first 3D structure of a vault particle from bacteriophages, cutting months of work into days.
- Anthropic's biology team used Claude to map the first 3D structure of a vault particle from a bacteriophage.
- Vaults were discovered in 1986 but their biological function remains unknown after 40 years of study.
- The phage vault contains 96 protein copies versus 78 in mammals, and is shorter, wider, and more spherical.
- Claude drove sequence search, DNA design, purification, and cryo-EM processing, cutting a 3-day purification to 90 minutes.
- A full project that normally takes a specialist team much longer was completed in under four months.
- Details in the preprint; structures and data to be released.
Anthropic researchers report the first three-dimensional structure of a vault-like particle assembled from a bacteriophage protein. Bacteriophages are viruses that infect bacteria. According to the team’s research report, cryo-electron microscopy, or cryo-EM, resolved a shell made from 96 copies of a protein called L05, with a combined molecular mass of about 9.5 million daltons.
Claude supported the project from sequence discovery through computational reconstruction. Researchers performed the physical experiments and supervised scientific decisions. The result gives vault biologists a simpler system to engineer and provides a detailed test of how an AI agent can coordinate a wet-lab project. The findings appear in a preprint and await peer review and independent replication.
A 40-year-old cellular mystery
Leonard Rome and Nancy Kedersha discovered vault particles in 1986. Their ribbed, hollow shape resembled a cathedral ceiling, which inspired the name. At roughly three times the mass of a ribosome, a vault is among the largest known molecular complexes in eukaryotes, the cells found in animals, plants, fungi, and other organisms with nuclei.
Researchers have proposed that vaults transport molecules around cells or help protect cells from toxins. Their function remains unresolved. Mice lacking vaults can survive, suggesting that any essential role may emerge only under particular environmental conditions or cellular stresses.
Phage proteins offer a simpler test bed
Before this study, structurally confirmed vaults came from eukaryotes. Sequence databases contained distant relatives of vault proteins in bacteria, archaea, and their viruses, but researchers had not shown that those proteins could assemble into vault-like shells. Anthropic’s team expressed candidate proteins and found that phage-derived L05 forms such particles under laboratory conditions.
Bacterial hosts grow faster and are easier to engineer than mammalian cells, making L05 a more tractable system for studying assembly and function. Its role during a natural phage infection remains unknown.
L05 builds a wider shell
| Feature | Phage-derived L05 vault | Mammalian vault |
|---|---|---|
| Protein copies | 96, with 48 per half | 78 |
| Dimensions | About 54 × 45 nanometers | About 67 × 40 nanometers |
| Overall form | Shorter, wider, and more spherical | Longer and narrower |
| Relative enclosed volume | About 1.25× | 1× |
To test how the shell assembles, the researchers engineered 15 L05 variants. Single amino-acid substitutions near the cap prevented closure and produced open, rosette-like particles. Duplicating two of the protein’s seven wall segments created vaults about 20% larger. The experiment identifies repeat duplication as a plausible resizing mechanism; the structure’s evolutionary history remains unresolved.
Ninety-six pockets hold a clue
Each L05 copy contains a pocket holding an unidentified small molecule whose shape resembles coenzyme A. Cells use coenzyme A to carry chemical groups between reactions. Chemical assays still need to establish the molecule’s identity, but confirmation would give researchers a specific lead for testing vault function.
From 15 billion sequences to a 3D map
- Sequence search: Claude Science searched a database of about 15 billion predicted proteins and identified vault relatives across 29 bacterial and four archaeal phyla.
- Construct design: Claude converted 44 candidate protein sequences into host-optimized DNA designs with the regulatory elements required for expression.
- Purification: Claude recommended size-exclusion chromatography, which separates molecules by size, along with assays for tracking the particles. It also proposed loading ruptured cell lysate directly onto the column, a shortcut the researchers initially resisted.
- Cryo-EM processing: Claude used the open-source RELION software through an internal multi-agent harness. RELION combines many two-dimensional particle images into a three-dimensional density map. Researchers advised the agent on analysis decisions.
Researchers carried out all physical manipulations and retained control over judgment calls. Claude proposed protocols, revised plans, and executed computational jobs with access to Anthropic’s internal computing resources.
A 90-minute purification
Anthropic reports that vault purification took about 90 minutes, compared with the team’s estimate of three days for a conventional workflow. A first structural model followed 34 hours later, a task the researchers estimated would usually require weeks. The team screened dozens of candidates, prepared samples, and solved the structure in less than four months.
The conventional timelines are team estimates. No head-to-head benchmark isolated Claude’s contribution from staffing, hardware, prior experience, or other workflow choices.
A bad medium and a mirrored map
Claude generated a growth protocol containing an additive that blocked DNA uptake into host cells, delaying the project by two weeks. During cryo-EM processing, it skipped a standard step and produced a flawed map blended with its mirror image. The researchers intervened and instructed the agent to follow the established workflow.
Anthropic says notes from both failures went to the machine-learning team and informed changes to the training data used for later models. The errors also show where domain supervision remained necessary during the project.
The reproducibility gap
Anthropic has posted a technical preprint submitted to bioRxiv and says the structures and additional data will follow. The current release contains no new model or code. The workflow combined existing Claude products with an internal multi-agent harness.
Anthropic says a typical laboratory with access to existing Claude products could perform most stages. Cryo-EM processing is the main exception because the team ran jobs across hundreds of GPUs in parallel. The pending structural data and internal harness currently limit full reproduction of the workflow.
What the result supports
L05 provides vault biologists with an engineerable system for testing how these particles assemble, change size, bind small molecules, and behave under different conditions. Its natural function remains the central open question.
The project documents an AI agent linking sequence mining, construct design, purification planning, and structural analysis within one research program. Similar performance across independent laboratories could reduce the amount of specialist time required for planning and computation, leaving experimental capacity, instrument access, and compute as major constraints.
Anthropic previously reported an enzyme-system study developed with Claude. Released artifacts, independent replication, and results from other kinds of experiments will determine how broadly this approach generalizes.
Comments
