Anthropic reported on 4 November that its biology team had used Claude to help determine a three-dimensional structure of a vault particle made from a bacteriophage protein, Unite.AI reported. The team describes it as the first such structure for a vault particle found in prokaryotic cells, a group that includes bacteria. At a nominal resolution of 3.2 angstroms, the map gives researchers a way to examine how this particular shell fits together and closes, although the biological purpose of vaults remains unknown.
Key points
- Anthropic reports that the bacteriophage-derived particle contains 96 copies of one protein.
- The team searched approximately 15 billion predicted proteins before testing candidates in the laboratory.
- Claude helped with DNA designs, purification and microscopy processing, while researchers directed the analysis.
- The research has not undergone peer review.
A search across 15 billion predicted proteins
Vaults are hollow, barrel-shaped assemblies of protein first identified by Leonard Rome and Nancy Kedersha in 1986, Interesting Engineering reported. Mammalian vaults contain 78 copies of a major vault protein. Their function remains uncertain, despite proposals involving transport within cells, drug resistance and immune defence. A related particle that can be produced in bacteria would give researchers a simpler experimental system in which to study how the shell assembles.
Using Claude Science, Anthropic’s researchers searched a database of approximately 15 billion predicted proteins for relatives of that major vault protein. They identified candidates in 29 bacterial phyla and four archaeal phyla, though the proteins appeared in less than 1% of species with sequenced genomes. A match in a database was only a starting point: a protein also had to be produced and shown to form a particle.
The researchers selected 44 proteins for initial tests before moving candidates into E. coli. Of 24 candidates tested there, 15 yielded soluble protein, and seven of those formed vault-like particles under an electron microscope. Cryo-electron microscopy — which images samples kept at very low temperatures — confirmed the vault shape in four cases: two proteins came from bacteria and two from bacteriophages, the viruses that infect them. The team chose one phage-derived protein, L05, for a detailed structural study because it formed complete, closed particles particularly well.
L05 closes with 96 protein copies
Anthropic’s L05 structure contains 96 copies of the same protein, with 48 in each half. Each copy contributes to the wall, but near the top identical copies take different shapes as the shell narrows and closes. The team tested that arrangement by making 15 altered versions of L05. Some changes at points where neighbouring proteins meet left the particles open or produced aggregates rather than closed shells.
Each L05 protein has seven repeating wall segments. Duplicating two of them produced closed particles with approximately 20% more enclosed volume, according to the research. This is a change to the shell’s geometry: putting cargo inside it and using it for delivery remain proposed engineering directions.
Packing a container could become a different problem if its walls could be extended while its lid still closed. The altered shell would offer more room for possible cargo, but the reported experiment tested its shape and enclosed volume rather than whether it could carry anything.
Claude and the 34-hour structural model
The researchers chose the protein candidates, while Claude helped turn their sequences into DNA designs for production in the selected hosts. It also contributed to a purification procedure based on size-exclusion chromatography. In this method, a sample passes through beads with small pores: smaller molecules spend longer inside the beads, while large particles emerge earlier. Anthropic says its procedure took about 90 minutes from cells to a microscopy grid, compared with approximately three days for a conventional sugar-gradient method used for vaults from more complex cells.
Claude initially argued against the column approach because unwanted clumps of protein could emerge alongside the particles. It later helped choose a column and checks for whether vault-like particles had formed. That sequence matters to the account of its contribution: the reported workflow involved researchers making the experimental choices, rather than accepting an uninterrupted set of model instructions.
For the microscopy data, Claude worked through an internal system that could run the open-source RELION processing workflow, while researchers guided analytical decisions. Anthropic reports that a first full structural model was ready after 34 hours of processing and final maps about a week after processing began. Its estimate for obtaining an initial model by conventional methods was weeks. The project took under four months. The reported workflow also included a flawed protocol and a corrupted microscopy map that required correction.
Anthropic has previously reported a different kind of laboratory result involving Claude: 17 scorpion-toxin binders designed with its help passed lab tests. In the vault project, the model’s role extended across candidate search, preparation and image processing, making the researchers’ supervision and the recorded errors part of the practical account of how that work was done.
Anthropic said it had submitted a preprint to bioRxiv and planned to release the structures and additional datasets, Benzinga reported on 6 November. The research has not undergone peer review. Rome, who co-discovered vault particles, advised Anthropic on the project as a paid consultant and gave feedback on the manuscript.
