Anthropic used Claude to determine the structure of a bacterial virus protein assembled into a vault, a barrel-shaped particle of unknown function; the preprint reports that both viral and bacterial proteins form these barrels
Anthropic used Claude to determine the structure of a bacterial virus protein assembled into a vault, a barrel-shaped particle of unknown function; the preprint reports that both viral and bacterial proteins form these barrels
Anthropic, the developer of Claude AI, published an account of the work on November 4, followed the next day by a bioRxiv preprint (a paper that has not yet been peer-reviewed). The authors introduced genes from bacteria and bacteriophages (viruses that infect bacteria) into Escherichia coli, and the cells assembled vaults. According to the preprint, Claude wrote all the code and performed all the computational analyses under the authors’ direction.
A vault is a hollow barrel made from 78 copies of a single protein in mammals; a human cell contains about 10 000 of these barrels. Nancy Kedersha and Leonard Rome discovered vaults in 1986, but Cell rejected their paper without peer review, adding that the journal would reconsider it once a function had been identified. Forty years later, their function remains elusive: mice lacking the major vault protein live as long as normal mice.
Anthropic turned to bacteria: they grow quickly, and bacterial relatives of human proteins have often helped reveal what those proteins do. Relatives of the vault protein had previously been found in bacteria and phages, but according to the authors, no one had shown that they assemble into vaults. In a database of 15 billion proteins, where the laboratory had already found the ART enzyme system, Claude identified such relatives in 29 bacterial phyla and prepared DNA sequences for 44 selected proteins.
Isolating vaults by centrifugation usually takes about three days. The authors tried to speed up the process using a column from which larger particles emerge first. Claude objected that clumps of misfolded protein would emerge alongside the vaults. They tried it, and Claude suggested loading the contents of disrupted cells onto the column without further purification. This risked clogging an expensive column, but it worked: going from cells to a sample ready for microscopy took 90 minutes. Electron microscopy showed that at least seven proteins had assembled into vaults; the others formed rings, open “rosettes,” and clumps. Cryo-electron microscopy (imaging frozen particles to reconstruct a 3D model) confirmed the vault shape for four proteins: two from phages and two from bacteria.
For structure determination, they chose the phage protein L05, one of the proteins that most readily assembled into closed vaults. The L05 vault consists of 96 protein copies rather than 78: it is shorter than the human vault but wider, with a quarter more internal volume. Each copy contains seven repeated segments rather than nine. In some naturally occurring proteins, adjacent segments are very similar, suggesting that one arose through duplication of the other. The authors made the same change to L05: they duplicated two segments, bringing the total to nine, as in the human protein, and increased the vault’s internal volume by approximately 20%. They suggest that vaults may have grown this way during evolution.
A pocket in each protein copy contains a signal from an unidentified small molecule. A computational search pointed to coenzyme A (a carrier of chemical groups) or one of its derivatives. This may offer a clue to the vault’s function.
The first 3D model was obtained after 34 hours of processing, and the entire project took less than four months for a team with limited experience in structural biology. Claude’s errors cost time: an additive in the protocol that interfered with DNA entering cells caused a two-week delay, while a missed standard step produced a structure mixed with its mirror image. At that point, the authors stopped Claude, saying, “Please, let’s stick to standard methods while they are working.”
Rome, an adviser on the project, asks:
“Will we find that they all do the same thing? Or will we see how vault function has evolved? Perhaps they do the same thing as human vaults, which would also be very interesting.”