Anthropic: Claude Helps Map the First Prokaryotic Vault Particle Structure; Purification in ~90 Minutes, First Model in ~34 Hours

Core Background & Industry Pain Points

Vault particles are barrel-shaped protein shells discovered in 1986 by Leonard Rome and Nancy Kedersha. They are among the largest molecular complexes in eukaryotic cells, yet their biological function remains unknown. Homologs of the major vault protein appear in prokaryotic and bacteriophage genomes, but until now it was unclear whether those proteins assemble into real vaults. On November 4, 2026, Anthropic's biology team published the first 3D structure of a prokaryotic vault and framed the project as a case study in using Claude to accelerate fundamental biology.

Architecture Highlights & Internals

Claude helped search roughly 15 billion predicted proteins, design DNA constructs, optimize isolation, and run cryo-EM processing through RELION inside an internal multi-agent harness. The chosen phage protein, L05, forms a closed shell of 96 chains (48 per half), about 9.5 megadaltons, roughly 54 by 45 nm, enclosing about 25% more volume than mammalian vaults. Adding two wall repeats (7 to 9) produced still larger vaults with about 20% more enclosed volume.

Authoritative Benchmarks & Measured Scores

No coding benchmarks were published. Vendor-reported lab timing: conventional sucrose-gradient purification takes about three days; the optimized size-exclusion workflow reached the microscope grid in about 90 minutes; a first full structural model was ready in 34 hours, versus an estimated weeks of manual work. Anthropic also documents failures: a medium additive delayed DNA delivery by about two weeks, and an early cryo-EM run skipped a standard step and produced a mirrored flawed map.

Developer Hands-on Guide

For builders, the transferable lesson is the research-agent workflow: search, construct design, protocol optimization, and structured data processing inside a multi-agent harness with human review on high-risk steps. Structures and additional data are expected soon; the preprint PDF is already linked from Anthropic's post. Teams building science agents can use the write-up as a checklist for where models may run autonomously and where humans must intervene.