# Claude Helps Researchers Find Vault-Like Protein Structures in Bacteria and Viruses > An Anthropic team used Claude to identify candidate vault proteins, then tested them in the lab. The results suggest the mysterious structures may have deeper evolutionary roots than previously thought. > > OpenCheese AI · 2026.11.09 · Web page: https://opencheese.dev/blog/claude-helps-researchers-find-vault-like-protein-structures-in-bacteria-and-viru Vaults are found by the thousands in plant and animal cells, yet nearly four decades after their discovery, biologists still do not know what they do. Now, an Anthropic team using Claude has found evidence that proteins from bacteria, archaea and viruses can also assemble into vault-like structures, raising the possibility that vaults have a much older evolutionary history. ## A cellular structure with an unknown job Vaults are hollow shells made from copies of a protein called major vault protein, or MVP. In humans, a vault's outer shell contains 78 copies of MVP, with other proteins and RNA molecules inside. Researchers have proposed several possible functions since vaults were discovered in 1986. One idea is that they carry substances between parts of a cell. None of the proposed functions has been confirmed, and disabling the MVP gene in mice does not cause obvious health problems. Earlier research had identified similar genetic sequences in simpler organisms. What scientists had not confirmed was whether the proteins encoded by those sequences could assemble into structures resembling vaults. ## From protein database to lab bench Anthropic's team used Claude to analyze a database of 15 billion proteins predicted from publicly available genetic sequences. The model identified potential MVP counterparts across 29 major groups of bacteria, four groups of archaea and bacteriophages, the viruses that infect bacteria. Claude then helped the researchers select 44 genes encoding suspected vault proteins for testing. The team synthesized the genes and introduced them into Escherichia coli. Of those 44 genes, 15 produced soluble proteins that could be purified, and seven produced vault-like structures. Electron microscopy confirmed their shapes. The distinction matters. Finding a promising sequence is one step, showing that its protein can form a vault-like shell is another. The experiments demonstrate that the selected proteins can assemble when produced in E. coli. They do not establish what the structures do in their source organisms. Evolutionary biochemist Edward Marcotte said the candidate proteins were consistent with earlier research, while the experimental demonstration of their assembly was the more substantial result. ## A different shape from a bacteriophage protein The team took a closer look at L05, a protein from bacteriophages. When copies of L05 assembled in E. coli cells, they formed shells containing 96 protein molecules, compared with 78 in human vaults. The L05 shell was shorter and roomier, with an egg-like shape rather than the lemon-like form of a human vault. Researchers also engineered L05 variants that formed larger structures. Such shells could be of interest for delivering drugs or other substances in the future. For now, that is a prospective application, not a demonstrated treatment. ## What Claude contributed, and what it did not Anthropic says the project was also a test of AI's role in scientific research. The company reports that Claude helped the team reduce vault purification from three days to 90 minutes. It also says the model built a structural model from electron microscopy images in 34 hours, a task that would otherwise have taken several weeks. The model did not get everything right. Some of its suggestions led to unproductive laboratory experiments. Indiana University bioinformatician Yuzhen Ye said the project's success depended at least as much on conventional laboratory work as it did on AI. The division of labor is clear in the result. Claude helped find and select candidates, but the evidence for their assembly came from making the proteins and examining the resulting structures. ## A question about independence Computational biologist Mario Rodriguez Mestre of the University of Copenhagen said he had been developing a program, alongside his dissertation work, to model the assembly of similar proteins into bacterial and viral vaults. He also used Claude and regarded the coincidence as suspicious. Anthropic denies that his work influenced its analyses. The company says its biologists cannot access users' conversations with Claude. ## An older origin, still an open question The findings widen the range of sources from which proteins can be made to form vault-like shells, and they give researchers a way to study those structures using bacteria. They also support the possibility that vaults arose early in the evolution of life, although that history remains to be established. As for their purpose, the new shells have given researchers more to study. They have not made the old mystery any less mysterious. ## Related posts - 2026.10.16 · [AI-designed enzymes keep a protease's cutting core and rebuild everything else](https://opencheese.dev/blog/ai-designed-enzymes-keep-protease-catalytic-core-rebuild-the-rest.md): Researchers used AI protein design to keep the catalytic core of two protease types and replace the rest of the protein. The enzymes can be tuned to cut chosen targets, including proteins tied to neurological diseases. - 2026.10.25 · [ChatGPT Ads expands to 26 new countries, with over 100,000 advertisers already on the platform](https://opencheese.dev/blog/chatgpt-ads-expands-26-new-countries-100000-advertisers.md): OpenAI says ChatGPT Ads is growing to 26 more countries and that more than 100,000 advertisers have used it. The pitch: reach people while they explore, compare and decide what to buy. - 2026.10.19 · [Claude Mythos 5 designed quantum circuits that need far fewer gates than published ones](https://opencheese.dev/blog/claude-mythos-5-quantum-circuits-fewer-gates.md): Anthropic says Claude Mythos 5 optimised eight quantum routines in an automated loop, cutting gate counts up to 17-fold in some cases. The authors stress that this is circuit design, not a working quantum computer.