Anthropic reported on November 4, 2026 that its biology team used Claude throughout a project investigating vault particles, large protein shells whose biological purpose remains unknown. The team says it mapped the first 3D structure of a vault-like particle from a prokaryotic source. For readers interested in AI research tools, the useful question is what software did in the workflow and where human scientists still had to catch mistakes. This is the team's own account of a research project, not evidence that an AI system independently solved biology.
What did Claude contribute to the vault study?
Anthropic says researchers directed Claude to search protein databases, help design DNA constructs, refine a method for isolating particles and process cryo-electron microscopy data into a structural model. The team worked with a protein from a bacteriophage, a virus that infects bacteria, and reports a particle built from 96 copies of that protein. The human vault structure discussed in the report has 78 copies. Those counts describe the studied structures; they do not reveal what vaults do inside living cells.
The team reports that a purification step that would conventionally take about three days took roughly 90 minutes in its workflow, with an initial structural model ready 34 hours later. These comparisons are Anthropic's estimates for its own project. They are not a controlled measure of how much faster Claude would make an ordinary lab or a general-purpose research app.
Where did the AI-assisted workflow fail?
The account is unusually specific about errors. A protocol written by Claude included a growth-medium additive that interfered with introducing DNA into cells, setting the work back two weeks. During microscopy processing, Claude skipped a standard step and produced a flawed map combined with its mirror image. The scientists spotted and corrected these problems. The example shows why a plausible instruction or output needs domain checks before it becomes an experiment or a published result.
Anthropic says its team completed the work in under four months, but the project also relied on people with laboratory expertise and substantial computing for part of the cryo-EM analysis. The team's statement that most other steps could be done with existing Claude products is a claim about its experience, not a demonstration that a new researcher can reproduce the complete study with a chatbot alone.
What has the study established, and what remains open?
The researchers say they made the vault-like particles in E. coli for study. They have not yet tested whether these particles form naturally in the bacteria and phages where related proteins occur. Nor have they identified the biological function of vaults. Anthropic links a technical preprint and says it plans to release structures and further data. Other researchers will need those materials to inspect and build on the finding.
For makers of scientific software, the story is about workflow design as much as a model's capabilities: a useful assistant can connect search, experimental planning and data analysis, while recording enough steps for specialists to verify and correct it. VIBPS is a social network for software, including tools people build for research, but it has no role in Anthropic's study and has not tested Claude on this task.