Anthropic says Claude drove discovery of first prokaryotic vault structure

Summary

• Anthropic reports its biology lab used Claude to find and solve a bacteriophage vault particle.
• Vault particles were discovered in 1986 and their biological function remains unknown.
• The run compressed lab stages normally taking weeks into days, Anthropic said.

The latest

A 3D structure of a vault particle from a prokaryote — the first reported — came out of Anthropic's internal biology lab using Claude, the company's science team said on November 4, 2026. The particle, called L05, was traced to a bacteriophage, and the result was posted as a preliminary paper submitted to bioRxiv. Anthropic did not name which Claude model ran the work.

Details

• The search: Under human direction, Claude searched a database of roughly 15 billion predicted proteins, converted candidate sequences into host-optimized DNA, added control sequences and prepared the designs for ordering, according to Anthropic. The company described a Claude Science tool for the database search and an internal agent system for cryo-electron microscopy work.
• The numbers: L05 is built from 96 copies of a single protein, 48 per half, at about 9.5 megadaltons, against 78 copies and 39 per half in mammalian vaults, Anthropic said. It measures about 54 nanometers by 45 versus 67 by 40, giving roughly 25% more internal volume, with 7 wall segments instead of 9.
• Engineering it: The cap closes over 3 layers. Of 15 modified versions, single changes near the cap blocked closure and produced open, rose-shaped particles, while duplicating two segments to reach 9 raised volume by about 20%, per the company's account.
• The purification fix: Claude initially objected to size-exclusion chromatography, then recommended a specific column, tracking methods and loading lysed cell contents directly. Purification fell to about 90 minutes from roughly 3 days, Anthropic said.
• The timeline: A first structural model arrived in 34 hours and final maps about a week later, against a typical estimate of weeks, with hundreds of GPUs running in parallel. Less than 4 months passed from screening to the first structure.
• The wider screen: Relatives were identified across 29 bacterial phyla and 4 archaeal phyla, present in under 1% of sequenced species. Of 44 candidate proteins tested in a cell-free system, 15 of 24 produced soluble protein in E. coli, 7 of 15 formed vault-like particles and microscopy confirmed 4 — two from phages, two from bacteria.
• The errors: A protocol that included an additive to the growth medium impeded DNA uptake and cost two weeks. An early processing run skipped a standard step and yielded a map mixed with its mirror image, after which the team told Claude to stop improvising.
• The people: Anthropic named Januka Athukoralage, Peter Yoon, Nicholas Perry and Matthew Durrant alongside its life sciences team. Leonard Rohm, a paid consultant who commented on the manuscript and co-discovered vaults in 1986 with Nancy Kedersha, said determining vault function remains extremely difficult.
• The limits: The particles were produced only in E. coli. Anthropic said their assembly in the native bacteria and phages was not tested, and the biological function is unknown.

Background

Vault particles were first described in 1986 by Leonard Rohm and Nancy Kedersha. Four decades later, no established biological role has been assigned to them, which is why a new structure matters as a starting point rather than an answer.

Between the lines

The claim here is about pace, not biology. Anthropic is describing an agent that moved from database search through DNA design, purification strategy and processing to model building — a sequence of lab stages, not a single task. The two documented failures, the growth-medium additive and the mirror-image map, show the compression came with human correction at each stage.

What's next

Peer review of the bioRxiv preprint, and any test of whether the particles assemble in their native bacteria and phages rather than in E. coli alone.

Source: Anthropic