Anthropic’s biology researchers have used Claude-assisted analysis to investigate a protein shell whose function remains a mystery. Their new work shows that relatives of the major vault protein from bacteria and bacteriophages can assemble into vault-like particles in a laboratory expression system. The advance is a clearer view of structure and assembly—not an explanation of what those particles naturally do.
A structural result, not a solved biological function
Unite.AI’s November 4 report describes the experimental progression from protein-family searches to microscopy. The team found related proteins across bacterial and archaeal groups, tested selected candidates and used cryo-electron microscopy to confirm several vault structures. One phage-derived particle, called L05, became the focus of detailed analysis.
The important distinction is between recognizing a related protein sequence and demonstrating an assembled particle. A computational resemblance can identify a candidate; the microscopy work addresses whether the chosen proteins actually form the expected architecture under the tested conditions.
That still leaves another boundary. The researchers produced the particles in E. coli. They have not shown that the same structures assemble in the original organisms where the corresponding proteins naturally occur. A laboratory system can make a question easier to investigate without answering every question about its natural setting.
For the AI story, this is a more concrete outcome than a generated biological hypothesis alone. There is an experimental structure to examine. But the presence of a structure should not be confused with an established physiological role or a demonstrated medical application.
The preprint reveals an adjustable shell
The team’s technical preprint describes L05 as a shell built from 96 protein chains, compared with 78 in mammalian vaults. Its enclosed volume is roughly 25% larger than the human counterpart. Adding wall repeats produced a version with approximately 20% more internal volume than the original L05 shell.
Those comparisons concern geometry, not demonstrated delivery performance. The authors discuss an engineerable scaffold for cargo encapsulation, but resizing a shell is not equivalent to showing that it can carry a particular payload safely and effectively to a destination.
The structural analysis also identifies unexplained density consistent with a small molecule inside a pocket. Coenzyme A or a related compound is a candidate interpretation, not a confirmed identity. The paper leaves both that identification and the biological function unresolved.
This uncertainty is scientifically consequential rather than an incidental disclaimer. If the identity of a bound molecule is still open, an explanation of the particle’s purpose cannot treat that candidate as established evidence. The new structure creates a testable lead; it does not complete the chain from shape to function.
Claude’s assistance included costly mistakes
Anthropic’s own account says Claude helped with database searches, computational work and microscopy processing while researchers directed the decisions. The company reports that a first structural model was ready in 34 hours and that the project took less than four months. Its comparison with a slower conventional workflow is the team’s estimate, not a randomized productivity trial.
The same account documents failures. A model-written protocol caused a two-week delay, and an early microscopy-processing run produced a flawed map blended with its mirror image. The team also used hundreds of GPUs for parallel structural analysis—resources that Anthropic acknowledges may exceed those available to many academic laboratories.
These details make the account more informative than a simple claim of autonomous discovery. Human judgment, specialist validation and available computing resources remain part of the result. The model’s involvement does not isolate its contribution from those other inputs.
Anthropic says the report has been submitted to bioRxiv and that structures and additional data are planned for release. Until wider scrutiny and follow-up experiments arrive, the defensible conclusion is specific: AI-assisted work helped characterize and modify a family of protein shells. Why those shells exist, and whether their engineering potential becomes a practical application, remain research questions.










