Claude Cuts Vault Protein Particle Isolation from 3 Days to 90 Minutes

Scientists have studied vault particles, large hollow protein capsules inside cells, for forty years, and still don't know what they are for. Anthropic says Claude helped break the deadlock: the agent analyzed a database of 15 billion protein sequences and helped prepare experiments, and as a result particle isolation in the lab fell from 3 days to 90 minutes. Imagine a project build that used to take three days and now takes an hour and a half. That is roughly the gain the biologists got, in a field where every attempt costs time and expensive reagents. For anyone building AI agents, this is a rare case that offers both strong numbers and an honest list of mistakes.

What are these particles, and why is there debate about them?

Vault particles were discovered in 1986. They look like hollow capsules and are thought to take part in intracellular transport and cell protection. Their exact function is still unknown.

In the new work, researchers used Claude to reconstruct, for the first time, the three-dimensional structure of a vault-like particle found in prokaryotes, organisms without a cell nucleus. This did not reveal the particle's role: the authors state plainly that its biological function remains a mystery. But knowing its shape and composition is a first step, and without it one can only guess at its purpose.

What exactly did Claude do?

The tasks were varied. Claude combed through 15 billion protein sequences, helped design DNA constructs, optimized lab procedures, and processed cryo-electron microscopy data, the images from which a 3D model of a protein is assembled. It worked inside a multi-agent system, ran the specialized RELION software, and processed data on a compute cluster.

The results in numbers:

• particle isolation was sped up from 3 days to 90 minutes;
• the first 3D model was produced in 34 hours instead of several weeks;
• the structure consists of 96 copies of the protein, with a mass of about 9.5 megadaltons (on the order of 9.5 million times the mass of a hydrogen atom);
• the entire research cycle took less than 4 months.

Importantly, the agent did not just answer questions in a chat window. It was built into the workflow: it searched a huge database, proposed protocols, ran professional software, and analyzed the results. This is the kind of setup that product agents are trying to reproduce today, except that here the environment is more expensive and less forgiving of errors.

Where the agent went wrong and what follows from it

The most useful part of the report is the section on failures. One bad recommendation from Claude delayed the experiments by two weeks. A skipped processing step led to an incorrect 3D map. In a lab, such errors can't be rolled back like a commit in a repository: they eat up weeks and materials.

The practical takeaway is simple. If an agent speeds up routine work several-fold but sometimes silently skips a step, the value of the system also depends on how quickly a human notices the miss. A good-looking final model does not prove that every step was completed, so checkpoints are needed after expensive and irreversible steps, not only at the end. Logging the agent's actions also helps: the log shows which step was skipped.

The second takeaway concerns economics. The main gain here comes from compressing long chains of routine steps, not from one-off clever answers. Give the agent real tools and compute, but keep a human in the role of reviewer: Anthropic stresses that AI agents carry out a significant share of complex scientific work under researchers' supervision.

There are caveats too. This is a report by Anthropic about its own model, and it contains no independent evaluation. Comparisons such as "instead of several weeks" refer to one specific project, and it would be rash to apply them to any lab. Still, the example gives a realistic picture: an agent can speed up science many times over, but it does not replace the scientist. It turns the scientist into an editor and a reviewer.