How Claude helped solve the first bacterial vault particle structure

Anthropic says its biology research lab has produced the first 3D structure of a vault particle from prokaryotic life, the simple, usually single-celled organisms such as bacteria. Vaults are large hollow barrels of protein found in nearly every cell in the human body and across most of the eukaryotic tree, from slime molds to humans. Leonard Rome and Nancy Kedersha discovered them in 1986, and their function is still unknown. Proposed roles, including moving molecules between the nucleus and the cytoplasm, protecting cells from toxic compounds and supporting innate immunity, have not been confirmed, and mice lacking vaults still survive. Distant relatives of the vault protein had already turned up in the genomes of bacteria and of the bacteriophages that infect them, but Anthropic's research post says no one had shown that they assemble into vaults. It reports that they do. That matters because bacteria grow quickly, so a bacterial vault is a simpler system for asking where vaults came from and what, if anything, they hold.

The structure is of a phage protein Anthropic calls L05, which forms a shell of 96 copies where the mammalian vault uses 78. The barrel is shorter and wider than the mammalian one, about 54 by 45 nanometers against 67 by 40, yet encloses about 25% more volume because it is rounder, and Anthropic calls it among the largest protein shells identified in prokaryotes. The wall protein is built from seven repeating segments rather than the mammalian nine. When the team copied two segments to make nine, the protein still assembled, into particles with about 20% more enclosed volume, which Anthropic presents as a way vaults could have changed size over their evolution.

The post's other subject is how much of the work Claude did under the scientists' direction. Using Claude Science, it searched about 15 billion predicted proteins and found vault relatives that are rare but spread across 29 bacterial and four archaeal phyla. After the scientists chose 44 candidates, Claude designed the DNA to produce them. The team then isolated particles with size-exclusion chromatography, a column of porous beads that lets particles too large to enter the pores wash out first, instead of the conventional sugar-gradient spin. Claude first argued against the approach, because clumps of misfolded protein come out in the same early fraction, then helped make it work, including proposing that ruptured cells go straight onto the column. Anthropic reports that isolation took about 90 minutes instead of about three days. Of the candidates, seven formed vault-like particles under an electron microscope, and cryo-electron microscopy, which combines images of many frozen copies into one 3D model, confirmed the vault shape in four. Claude ran that processing in the open-source package RELION on hundreds of GPUs, and a first full model was ready in 34 hours, a step Anthropic estimates would normally have taken it weeks. The whole project took under four months.

Anthropic also lists Claude's mistakes. A protocol it wrote specified a growth medium with an additive that interferes with getting DNA into cells, which cost two weeks, and early in the cryo-EM processing it skipped a standard step and produced a flawed map blended with its mirror image, prompting the team to tell it: "Please let's do normal things unless they don't work." The scientists say they caught errors quickly in areas they knew well but lost time in areas they did not, and that newer models seemed to make fewer errors over the project. Most of the work, they say, is within reach of a typical lab using existing Claude products, though the cryo-EM step needed more computing than an academic institution might have.

The result does not explain what vaults are for. Claude noticed an unidentified small molecule in a pocket of each of the 96 copies whose shape resembles coenzyme A, a common helper molecule, and Anthropic calls that an early hint at a function if it holds up. The particles have so far been made only in E. coli, not in the bacteria and phages where the proteins occur naturally. Rome, a paid consultant on the work, said the breadth of the findings suggests a complicated function. All of these figures and the division of labor between Claude and the scientists come from Anthropic's own account; its preprint has been submitted to bioRxiv, and the structures and additional data have not yet been released. The same lab's earlier finding of a CRISPR-like enzyme system in phage DNA was likewise published before its function was known.