Anthropic Uses Claude to Show Bacterial and Phage Proteins Form Vaults

Anthropic reported the results on November 4, 2026, showing that prokaryotic and bacteriophage relatives of the major vault protein can assemble into vault particles when produced in E. coli. Claude helped the team map one bacteriophage-derived shell at a nominal resolution of 3.2 angstroms.

Vaults are large, hollow protein shells whose biological function remains unknown. Producing related particles in bacteria gives researchers a faster-growing system for investigating their assembly, contents, and origins. Anthropic describes the L05 result as the first 3D structure of a vault particle found in prokaryotic cells. The work establishes assembly and architecture while leaving biological function unresolved.

From Protein Search to Confirmed Vaults

Using Claude Science, the researchers searched a database of approximately 15 billion predicted proteins. They found major vault protein homologs in 29 bacterial phyla and four archaeal phyla, although these proteins appeared in less than 1% of species with sequenced genomes. The search therefore showed that the proteins are rare but broadly distributed.

The researchers selected 44 proteins for cell-free expression testing before moving leading candidates into E. coli. Fifteen of 24 candidates produced soluble protein, and seven of those 15 formed vault-like particles under electron microscopy. Cryo-electron microscopy confirmed the vault shape in four of them, two derived from phages and two from bacteria.

The phage-derived L05 particle was selected for detailed structural analysis because it was among the best at forming complete, closed particles. That selection moved the project beyond detecting barrel-shaped objects in microscope images to examining how individual protein chains fit together and where the shell could be modified.

Engineering the L05 Shell

L05 contains 96 copies of one protein, arranged as 48 copies in each half. Mammalian vaults contain 78 copies, with 39 in each half. Each L05 protein has seven repeating wall segments, compared with nine in human major vault protein. The resulting shell is shorter and wider than the human vault but encloses approximately 25% more volume.

Its cap closes in three steps. Within each group of three neighboring protein copies, one ends at the first tier, another at the second, and the third continues to the tip. The copies are identical along the wall but adopt different conformations as the cap narrows.

To investigate which contacts are necessary for assembly, the team made 15 altered versions of L05, mostly changing points where neighboring copies touch. Some single substitutions near the wall seam or cap produced open, rosette-like particles or aggregates instead of closed vaults, identifying regions required to seal the shell.

Duplicating two wall segments gave L05 nine repeats and produced closed particles with approximately 20% more enclosed volume. The authors interpret this as evidence that repeat duplication can adjust vault size without preventing assembly, offering a possible route to size-tunable nanocapsules. Cargo encapsulation and delivery are proposed engineering directions; the demonstrated result is control over shell geometry.

Claude’s Role and the Experimental Limits

The researchers selected the candidate proteins themselves. Claude converted the chosen protein sequences into DNA designs tailored to their expression hosts, added gene-control sequences, and prepared the designs for ordering.

Purification provided one of the clearest time comparisons. Anthropic says its optimized size-exclusion chromatography method reduced the journey from cells to a microscopy grid to about 90 minutes, compared with approximately three days for the conventional sugar-gradient method developed for eukaryotic vaults. In sizing columns, particles too large to enter the porous beads emerge first, while smaller molecules are delayed.

Claude initially argued against that approach because protein aggregates could emerge in the same excluded volume. It subsequently helped recommend a column, methods for tracking the protein and checking particle formation, and the direct loading of ruptured-cell contents without prior purification.

For structural processing, Claude operated inside an internal multi-agent harness with internal computing resources. It set up and ran the open-source RELION workflow while researchers guided the analytical decisions. Anthropic reports that a first full model was ready after 34 hours and final maps about a week after processing began, while estimating that an initial model would otherwise have taken weeks. The project took under four months.

Various Claude models and interfaces, including Claude Science and Claude Code, wrote the study’s code and performed its computational analyses, with the authors directing the work and accepting responsibility for the content. Anthropic says most steps could be performed by a typical laboratory using existing Claude products, although the parallel cryo-EM analysis used hundreds of GPUs beyond what many academic institutions may have available.

The workflow also produced consequential errors. One Claude-written protocol specified a growth-medium additive that interfered with introducing DNA into cells, delaying the project by two weeks. An early cryo-EM run skipped a standard step and generated a flawed map blended with its mirror image. Anthropic shared the laboratory notes with its model-training researchers and says subsequent Claude models appeared to make fewer DNA-design errors and identify experimental pitfalls more effectively.

Each L05 protein copy also contains a pocket with unresolved density from an unidentified small molecule. The technical report describes a computational search of 25,000 molecules that identified coenzyme A or a derivative as a candidate ligand. Neither the molecule’s identity nor any functional implication has been confirmed.

The particles have so far been produced only in E. coli, and the researchers have not tested whether they form in the bacteria and phages where the proteins naturally occur. Anthropic has submitted the technical report to bioRxiv and plans to release the structures and additional data later.

The immediate advance is experimental control over a family of protein shells that can now be produced in bacteria, mapped in detail, and resized through repeat duplication. The larger question remains biological: whether that simpler system and the unidentified ligand can help explain what vaults do, and whether the particles assemble in the organisms where their genes naturally occur.