Recommended Articles
Tech & AI

Anthropic's Claude Helps Solve a 40-Year Biology Mystery With the First Prokaryotic Vault Structure

By Angeled · November 5, 2026 · Tech & AI
Anthropic's Claude Helps Solve a 40-Year Biology Mystery With the First Prokaryotic Vault Structure
Anthropic's Claude Helps Solve a 40-Year Biology Mystery With the First Prokaryotic Vault Structure

Anthropic reported on November 4, 2026, that its biology research team, working with Claude, had mapped the first three-dimensional structure of a vault particle from bacterial life. The work resolves a mystery that has stood since vault particles were discovered in 1986, and it is a striking demonstration of what an AI assistant can — and cannot — do inside a real laboratory.

Vault particles are barrel-shaped objects among the largest protein structures found in cells. They were first identified in 1986 by scientists Nancy Kedersha and Leonard Rome, and they have been studied mainly in eukaryotes, including mammals, ever since. Nobody knew whether equivalent structures existed in bacteria and the viruses that infect them, or what they were for. Anthropic's team set out to find out, using Claude throughout the investigation, from identifying promising proteins to reconstructing their molecular architecture.

The Search Through 15 Billion Proteins

The project began with a computational search on a scale that would have been impractical a few years ago. Using Claude Science, the researchers screened a database of approximately 15 billion predicted proteins, looking for distant relatives of the major vault protein — the building block of vault particles. They found candidate homologs in 29 bacterial phyla and four archaeal phyla, a sign that vault-like proteins are far more widespread than anyone had assumed.

From those candidates, the team selected 44 proteins for laboratory testing. Fifteen of 24 candidates taken into cell-free expression produced soluble protein, and seven of those 15 formed vault-like particles under electron microscopy. Cryo-electron microscopy, a technique that flash-freezes samples and combines thousands of particle images into a three-dimensional reconstruction, confirmed the vault shape in four of them: two derived from bacteria and two from bacteriophages, the viruses that infect bacteria.

The 96-Protein Shell

The team selected a phage-derived protein called L05 for detailed structural analysis, because it was among the best at forming complete, closed particles. The result is the headline of the paper: inside E. coli cells engineered to produce the protein, L05 self-assembles into a giant hollow barrel containing 96 copies of itself, arranged as 48 copies in each half, with 48-fold rotational symmetry around its equatorial midline. The particle measures approximately 54 nanometres long and 45 nanometres wide, with a combined molecular mass of roughly 9.5 megadaltons.

That is bigger than the mammalian vault, which contains 78 copies. By duplicating part of the protein, the researchers also created a larger version with substantially more internal space, showing the structure can be engineered, not just observed. It was resolved at a nominal 3.2 angstroms, fine enough to see individual protein chains. Each L05 protein copy also contains a pocket with unresolved density from an unidentified small molecule; a computational search of 25,000 molecules identified coenzyme A or a derivative as a candidate, but neither the molecule's identity nor any functional implication has been confirmed.

How Claude Helped — and Where It Went Wrong

Claude contributed at every stage: protein sequence searches, DNA design, purification, and analysis of the microscopy images. One contribution was dramatic. Anthropic says an optimized size-exclusion chromatography method cut the journey from cells to a microscopy grid to about 90 minutes, compared with approximately three days for the conventional sugar-gradient method developed for eukaryotic vaults. The first complete structural model was assembled in 34 hours, and the final reconstruction took about a week.

Anthropic is unusually candid about the failures. One Claude-written protocol specified a growth-medium additive that interfered with introducing DNA into cells, delaying the project by two weeks. An early cryo-EM run skipped a standard processing step and produced a flawed map blended with its mirror image. The team caught both errors, but the lesson is plain: an AI assistant in the lab still needs expert supervision, and its mistakes can cost weeks.

The findings are also not yet settled science. Anthropic has submitted the technical report to the bioRxiv preprint server, but the work has not been peer-reviewed or independently replicated. The particles have so far been produced only in engineered E. coli, and the researchers have not tested whether they form in the bacteria and phages where the proteins naturally occur. Anthropic says it will release the structures and additional data, with atomic coordinates to follow.

Why It Matters for Founders and Investors

For AI-for-science, this is a proof point with numbers attached: a search across 15 billion proteins, a three-day process cut to 90 minutes, a first structure in 34 hours. It suggests AI assistants are moving from answering questions about biology to doing biology — designing experiments, processing data, and accelerating the slow, manual steps that dominate lab work. Startups building AI tools for drug discovery, protein engineering, and lab automation should take note: the market is shifting from software that analyses results to software that drives the experiment.

The caveats matter just as much. The errors Anthropic disclosed are a reminder that AI in science is not a replacement for expertise, and that the winners will build verification into the workflow. For investors, the signal is that the most valuable AI-for-science companies may be those that treat the model as a junior collaborator with a senior scientist watching — and that publish their failures alongside their breakthroughs.

Angeled.com connects founders with investors.

Sources: Anthropic technical report and announcement, November 4, 2026; Unite.AI, November 4, 2026; Interesting Engineering, November 4, 2026; The Neuron, November 5, 2026.

Read Next