Mysterious cellular capsules studied with the help of Claude

Researchers at Anthropic characterized a protein capsule of viral origin and modified its size with Claude. Its natural function remains unknown.

A bacteriophage protein, a virus that infects bacteria, assembles into a hollow capsule when produced in E. coli. This is one of the results reported by Anthropic’s biology lab, which used Claude to carry out the computational research and support the experiments. The team reconstructed the structure of this capsule, then obtained a larger version by modifying the protein that makes it up.

The work presented by the company focuses on “vaults,” large protein shells known for forty years in organisms such as humans. Related proteins had already been identified in bacteria and their viruses. The new study shows that they can form assemblies of this kind in the laboratory. Their presence in this form in their native organisms remains to be verified. The results are the subject of a preprint that has not yet been peer reviewed. A microbial version of a capsule known since 1986

The name of these particles comes from their appearance: their ribbed walls evoke the vaults of a cathedral. Discovered in 1986, they are made, in mammals, of 78 copies of the same main protein, arranged in two halves that join together. Their biological role remains poorly understood, despite hypotheses related to molecular transport or cell protection.

The particle studied in detail, called L05, comes from a bacteriophage protein. Produced in E. coli, it brings together 96 copies of this protein, or 48 per half. Electron cryomicroscopy made it possible to reconstruct its three-dimensional organization from images of many frozen particles. Shorter and wider than the human vault, it encloses about 25% more volume.

The difference also lies in the construction of its wall. Each chain contains seven repeated segments, compared with nine in the human version. These elements align to form the sides of the shell, while the ends of the chains take on different shapes in order to close its poles. The study thus describes how identical copies of a protein contribute to a closed structure. Computational analyses conducted under the direction of the researchers

The research began with the exploration of a database containing about fifteen billion predicted protein sequences. Using Claude Science, the team searched for relatives of the vault protein before selecting candidates to test. The assistant then helped prepare the DNA constructs needed for their production. The researchers carried out the experiments and selected the proteins for further study.

Isolating the particles was another stage of collaboration. The team chose a size-based separation, although Claude initially raised reservations about this approach. The assistant then helped optimize it. According to the published account, the preparation from the cells to the microscopy grid took about 90 minutes, compared with roughly three days for the conventional protocol to which the authors compare it.

Image processing involved several Claude interfaces and an internal system of multiple working agents. Under the team’s supervision, the AI installed and used the RELION software for the analyses. An initial structural model was available in 34 hours; the final maps required about a week of processing. The account specifies that this phase benefited from hundreds of GPUs used in parallel.

The reported gains were accompanied by documented errors. A culture medium choice recommended by Claude delayed the manipulations by two weeks. During image processing, an omitted standard step also produced an incorrect map, mixed with its mirror image. The researchers had to intervene to correct the analysis. The preprint indicates that Claude wrote all the code and performed the computational analyses, under the direction of the authors, who take responsibility for them. A shell whose size can be modified

The team then tested changes to the protein to understand what keeps the capsule closed. Certain modifications near its end prevent closure: the images then show open, rosette-shaped assemblies. These experiments identify regions that are important for building the shell.

Adding two repeated segments in the wall, by contrast, produces a larger closed capsule. Its internal volume increases by about 20% compared with L05, which is approximately 50% more than that of a human vault. The authors see this as a path toward making nanocapsules with adjustable size, capable of housing molecules or enzymes. A biological transport application remains to be developed and tested.

The structure also retains one mystery: a pocket located in the wall contains a small molecule that the researchers were unable to identify. The analyses suggest a resemblance to coenzyme A, involved in many cellular reactions, or to one of its derivatives. This attribution remains hypothetical. Determining what the capsule actually retains could help explain its function.