Anthropic’s biology research lab has achieved a significant milestone in structural biology by determining the first 3D structure of a vault particle found in prokaryotic cells. The team utilized Claude to search sequence literature, optimize isolation methods, and map the structure, reducing a process that traditionally takes weeks to just 34 hours for the initial model. This work, detailed in a preprint submitted to bioRxiv, demonstrates how large language models can accelerate fundamental biological discovery by handling complex data processing and experimental design tasks.

The Mystery of Vault Particles

Vault particles are large, hollow protein shells discovered in 1986 by Leonard Rome and Nancy Kedersha, yet their biological function remains elusive after four decades. While found in nearly all human cells, they are absent in common model organisms like yeast and fruit flies, complicating study. Mice lacking vaults survive normally, suggesting the particles may only be necessary under specific environmental conditions. Anthropic’s team hypothesized that studying prokaryotic relatives could reveal the vault’s origins and potential functions, given that bacterial versions of other human immune proteins have previously provided such insights.

Accelerating Structural Biology

The research team used Claude Science to search a database of 15 billion predicted proteins, identifying vault relatives in 29 bacterial phyla and four archaeal groups. Claude then assisted in designing DNA constructs and optimizing an isolation method using size-exclusion chromatography, which cut purification time from three days to approximately 90 minutes. For structural mapping, the team directed Claude through an internal multi-agent harness using RELION software. The AI set up the processing pipeline and ran the steps, producing a first full model in 34 hours and final maps in about a week, a timeline Anthropic estimates would normally take weeks or months.

Structural Findings and Engineering

The newly solved structure, derived from bacteriophage protein L05, consists of 96 copies of a single protein, significantly more than the 78 copies in human vaults. This prokaryotic vault is shorter and wider than its eukaryotic counterpart, measuring roughly 54 by 45 nanometers, but encloses 25% more volume due to its spherical shape. The team also engineered larger variants by duplicating repeating segments of the protein wall, proving that vault size can evolve through segment duplication without disrupting assembly. A notable unexplained feature is a small molecule pocket within each protein copy, which Claude suggested resembles coenzyme A, offering a potential clue to the particle’s function.

Key Takeaways

  • Claude reduced vault purification time from three days to 90 minutes using optimized chromatography.
  • The first prokaryotic vault structure was modeled in 34 hours, compared to weeks for traditional methods.
  • The phage-derived vault contains 96 protein copies, unlike the 78 found in mammals.
  • Anthropic shared lab notes with ML researchers to improve Claude’s experimental reasoning capabilities.

The Bottom Line

This project proves that LLMs can move beyond code generation to drive wet-lab scientific discovery, turning weeks of structural biology into days of automated analysis.

Limitations and Future Work

Despite the speed, the process was not flawless. Claude initially argued against the chromatography approach due to contamination risks and later made protocol errors, such as selecting a growth medium additive that interfered with DNA uptake, setting the project back by two weeks. The team also had to intervene during cryo-EM processing when the model skipped standard steps, requiring human guidance to ensure data integrity. Anthropic acknowledges that the function of these vaults remains unknown and invites the scientific community to contribute ideas for further testing, marking this as an early but pivotal step in AI-assisted biology.