Anthropic uses Claude to map first 3D structure of a prokaryotic vault particle

A biological mystery since 1986

Vault particles are large, hollow protein shells found in nearly all cells in the human body. They were first discovered in 1986, but scientists still do not know what they actually do. Because they are so large—roughly three times the mass of a ribosome—they are considered one of the largest known molecular complexes in eukaryotes.

Despite their prevalence, vaults are missing from many common laboratory organisms like baker’s yeast and fruit flies. In mice, animals bred without vault proteins still develop normally, which suggests these structures might only be necessary under specific environmental conditions. Researchers have proposed several theories, such as protecting cells from toxic drugs or transporting molecules, but none have been proven.

Anthropic's biology research lab decided to investigate these structures by looking at prokaryotes. Bacteria provide a simpler system for study because they grow quickly and can help reveal the microbial origins of human proteins. The team specifically targeted bacteriophages, which are viruses that hijack bacteria to produce their own proteins.

Mapping the phage vault

The team successfully uncovered the structure of a vault from a bacteriophage using cryo-electron microscopy. This technique involves flash-freezing particles and imaging them with an electron beam to build a 3D model. They identified a specific protein, called L05, which assembles into a shell made of 96 copies. This is different from the human vault, which is composed of 78 copies of a different protein.

The phage-derived vault has a unique shape that is shorter and wider than its mammalian counterpart. It measures roughly 54 nanometers long and 45 nanometers wide. Even though it is shorter, it holds roughly 25% more internal space because its shape is more rounded than the human version. The wall is built from seven repeating segments, whereas the human version uses nine.

One of the most interesting features discovered is how the caps of the vault close. The wall narrows toward the poles in three steps, resembling a layered dome. Within every set of three adjacent protein copies, one finishes at the first step, one at the second, and one reaches the very tip. This specific arrangement allows the shell to seal properly, a fact the team confirmed by making 15 altered versions of the protein.

Claude as a research partner

Anthropic integrated Claude into every step of the research process to speed up the discovery. The model was used to search a database of 15 billion predicted proteins to find vault relatives. It then turned those sequences into DNA designs tailored for host cells. This allowed the team to screen dozens of candidates in a fraction of the usual time.

Claude also helped optimize the physical isolation of the particles. The team used size-exclusion chromatography, where samples flow through beads to separate molecules by size. Claude initially argued against this method but eventually recommended a specific column and a way to track the proteins. This optimized method took about 90 minutes, whereas the conventional method for eukaryotic vaults takes three days.

For the final 3D modeling, Claude worked within an internal multi-agent harness to run specialized processing software. This allowed the team to produce a first full model in 34 hours, a task that usually takes weeks of manual work. The researchers noted that while Claude made some errors, such as proposing a growth medium that interfered with DNA uptake, it successfully compensated for the team's lack of experience in structural biology.

Biotech and future steps

The ability to quickly isolate and engineer these vaults suggests significant potential for biotechnology. Because they are large, hollow shells, they could be adapted into nanocapsules for delivering enzymes. Understanding how these proteins assemble by duplicating pieces of their walls also shows how they might have evolved over time. These findings provide a blueprint for how such complexes can be manipulated in a lab setting.

While the structure is now known, the exact biological function remains a mystery. Claude identified a pocket in the protein copies that holds a small molecule, which it noted resembles coenzyme A. If this is correct, it could provide a major hint at how the vault functions within a cell. The team is now preparing to release the structures and additional data to the wider scientific community.

The project also served as a test for how AI can be used for fundamental discovery. Anthropic shared their lab notes with the engineers who train Claude to show where the model struggled. This feedback loop is intended to help the model become more reliable at identifying pitfalls in biological experiments. The team plans to continue testing these vaults in different bacterial and phage environments to see how they behave in nature.