Anthropic uses Claude to map a 96-protein vault from a bacteria-infecting virus

Biologists have a simpler system for investigating one of the cell’s enduring mysteries: giant, hollow protein shells called vaults. On November 4, Anthropic said its researchers, working with Claude, resolved the first 3D vault structure from prokaryotic life. The particle contains 96 copies of a virus protein, but its biological purpose remains unknown.

The finding concerns a protein called L05 from a bacteriophage—a virus that infects bacteria. Researchers produced its vaults in E. coli bacteria and mapped their structure using cryo-electron microscopy, or cryo-EM. This method combines images of many flash-frozen particles into a single three-dimensional model.

Vaults were discovered in 1986 and occur in nearly all human cells. Proposed jobs include moving molecules, protecting cells from toxic compounds and supporting immune defenses, but none has been conclusively established. Related proteins were already known in bacteria and their viruses; Anthropic’s contribution was showing that some assemble into complete vault-like shells.

A shell researchers can resize

L05 forms a shell roughly 54 nanometers long and 45 wide, with a combined mass of about 9.5 megadaltons. It is shorter and wider than a mammalian vault, yet encloses about 25% more volume. Its 96 identical protein copies form two halves, each containing 48 copies.

The caps reveal how identical building blocks can close a curved shell. In each group of three neighboring protein copies, one stops at the cap’s first tier, another at its second, and the third reaches the tip. The copies match along the wall but take different shapes at the cap.

The team made 15 altered versions to investigate assembly. Some single changes near the cap prevented closure, leaving open, rosette-like particles. Duplicating two of the protein’s seven repeating wall segments instead produced larger vaults, increasing enclosed volume by approximately 20%. That offers an experimental way to study how shell size can change.

From billions of sequences to a microscopy grid

Claude’s role began before the structural work. The team used Claude Science to search about 15 billion predicted proteins for vault relatives. These appeared across 29 major bacterial groups and four groups of archaea, another branch of simple-celled life, but in fewer than 1% of species with sequenced genomes.

Researchers selected 44 proteins for initial testing. Claude converted their sequences into DNA designs tailored to the host cell and added the controls needed to switch the genes on. Of 24 candidates moved into living E. coli, 15 produced soluble protein. Seven formed vault-like particles; cryo-EM confirmed vault shapes for four.

Getting particles ready for imaging required a less conventional choice. The researchers proposed chromatography, which separates material by size through a column of porous beads. Claude initially objected because clumps of misfolded protein could emerge alongside vaults. After the team proceeded, it helped select a column and checks for particle formation.

Claude then suggested putting ruptured cell contents directly onto the column, skipping prior purification. Researchers worried about clogging or contaminating expensive equipment, but tested the idea successfully. Anthropic says the optimized process took about 90 minutes from cells to microscopy grid, versus about three days for the conventional method it described.

Fast processing, costly mistakes

For microscopy processing, Claude ran the open-source software RELION through an internal system coordinating multiple agents, while researchers guided analysis decisions. Anthropic says a first full model was ready in 34 hours, against its estimate of weeks normally. Final maps took about a week. This stage used hundreds of GPUs in parallel.

The workflow also exposed errors. A Claude-written growth protocol included an additive that interfered with getting DNA into cells, delaying the project by two weeks. Interesting Engineering also describes an inaccurate structural map after Claude skipped a standard processing step. Researchers had to supervise and correct the work.

One unresolved clue sits inside the structure: each protein copy holds a small molecule that has not been identified. Claude noted a resemblance to coenzyme A, a molecule cells use to carry chemical groups. That remains a possible lead, not an explanation. Anthropic says its findings were submitted to bioRxiv as a preprint, with structures and additional data to follow.