On November 5, Anthropic announced that its biology team, using Claude, had for the first time resolved the three-dimensional structure of a prokaryotic Vault. This vesicle-like complex was discovered in 1986, and for forty years no one could explain what it actually does in living cells; human cells contain nearly 10,000 of these same structures per cell.

image.png

The breakthrough came from a needle-in-a-haystack search: the team pored over 15 billion predicted sequences for homologs, locked onto candidates from 29 bacterial phyla and 4 archaeal phyla, and selected 44 for laboratory validation. Seven assembled into vault-like particles in E. coli, and electron microscopy confirmed four complete structures. Among them, phage L05 reached 3.2 Å resolution—96 identical protein chains, each with 7 wall repeats (humans have 9); the particle measured 54.5 × 47.5 nm, was roughly spherical, and had an internal volume 25% larger than that of human vaults. The team then inserted extra wall repeats, and the variant’s volume increased by another ~20%, proving that the number of wall repeats can tune size.

Claude cut purification from 3 days to 90 minutes—and also caused problems

Claude wore many hats: it was involved in protein search, DNA design, purification optimization, and electron microscopy data processing. Traditional purification took 3 days, while the new approach cut it to 90 minutes; the first complete structural model was generated in 34 hours, whereas the old method would have taken weeks.

AI also showed its fallibility: an additive it suggested once interfered with DNA delivery, costing the research two wasted weeks; during electron microscopy processing, it skipped standard steps and produced inaccurate images, which were corrected only through intensive supervision. Although the structure has been resolved, the function of vaults remains a mystery. Data for L05 and its variants have been deposited in the Protein Data Bank, providing a size-tunable scaffold for cargo encapsulation—human-AI collaboration speeds things up, but also serves as a reminder that models are far from being able to handle things on their own.