Anthropic uses Claude to crack the first three-dimensional structure of a prokaryotic vault

Today, Anthropic once again crossed into biology, using Claude to launch a counterattack against a mysterious life structure that has puzzled humanity for 40 years!

In less than 4 months, the AI drew the first-ever three-dimensional structure of the "Vault" inside a prokaryotic organism.

From gene hunting and optimizing purification to processing cryo-EM data and manual expansion and modification, Claude was present almost the entire time.

What is even more striking is that its problem-solving speed can only be described as a "dimensional strike."

A purification experiment that originally took 3 days was compressed to 90 minutes; a three-dimensional structure that previously took weeks or even months to compute was on the table in just 34 hours!

A "ghost cathedral" has trapped the biology world for 40 years

In 1986, Leonard Rome's team at the University of California, Los Angeles, isolated a mysterious particle from rat liver.

Under the electron microscope, its internal arcs closely resembled cathedral vaults, hence the name "Vault."

Its mass is 3 times that of a ribosome, making it one of the largest protein complexes in cells. The mammalian vault is formed by 78 protein chains (39 on top and 39 on the bottom) closing together in half, enclosing a huge hollow container.

But what is the point of every human cell building tens of thousands of such hollow barrels?

Biologists' most commonly used organisms—yeast, fruit flies, and nematodes—strangely do not have it at all; and after knocking out the vault gene in mice, the mice still develop normally.

This exquisite "protein barrel" thus became a mystery hanging over the biology world for 40 years.

Hunting through 15 billion proteins, family size surges 36-fold

Since no breakthrough could be found in complex animals, the Anthropic team decided to look for answers in the simpler world of bacteria.

They brought Claude Science into the laboratory and launched a carpet-search across about 15 billion protein sequences.

The results showed that this type of protein is spread across 29 bacterial phyla and 4 archaeal phyla, and even the phages (viruses) that infect bacteria harbor large numbers of similar genes.

This search directly expanded the known vault protein family from 41 to 1,507, a 36-fold surge in size!

Moreover, they are quite rare in the bacterial world, appearing in less than 1% of species, and are often located right next to antiviral defense genes, making them very likely special equipment for bacterial antiviral defense.

Finding clues is only the first step; next they had to actually produce them.

The researchers selected 44 candidate proteins and had Claude handle everything from synthetic DNA design and expression switch configuration to ordering plans.

After transfer into E. coli, 15 candidate sequences successfully produced soluble proteins. But how to cleanly extract these particles, which are hundreds of times larger than ordinary proteins, immediately became a stumbling block.

First "playing devil's advocate," then a ruthless move: purification cut from 3 days to 90 minutes

The traditional vault isolation method requires placing samples in a sucrose solution and ultracentrifuging them; the entire process takes a full 3 days and is simply impossible for batch screening.

To speed things up, the human researchers planned to switch to "size-exclusion chromatography" (SEC).

The principle is simple: let the sample flow through a column packed with porous beads; small molecules get stuck in the pores and move slowly, while the huge vault particles cannot enter and can only take a straight path, rushing out first.

A funny scene occurred: at first, Claude actually voted against it.

Its reasoning was very professional: misfolded protein junk is also large and would rush out at the same time as the vaults, making them impossible to distinguish.

The team thought electron microscopy rechecking could serve as a fallback and decided to push ahead. After being overruled, Claude immediately switched into helper mode and recommended chromatography columns and tracking detection schemes suitable for ultra-large particles.

The real ruthless work was still to come.

Next, Claude in turn came up with an unusual move: skip the upfront purification and pour the crude lysate from crushed E. coli directly into the chromatography column!

The researchers tried it half-doubtfully, and the result was that the huge vaults rushed out of the column first, leaving tens of thousands of contaminating proteins far behind.

The purification process that originally took 3 days was compressed in one stroke to an astonishing 90 minutes!

After this batch of proteins passed initial electron microscopy screening, 8 had grown clear vault particles—the first time humans had clearly seen complete vaults outside eukaryotes.

Claude takes over EM reconstruction, solving 96 chains in 34 hours

After cryo-EM review, 4 of the candidate particles had the most regular morphology (2 from phages and 2 from bacteria).

The team selected the most completely assembled phage L05 vault and set out to tackle a high-resolution 3D atomic structure.

Previously, piecing hundreds of thousands of electron microscopy images back into a three-dimensional structure would take an expert weeks or even months.

But within its internal multi-agent framework, Claude installed the open-source software RELION itself and ran it in parallel on hundreds of GPUs, with humans only making decisions at key checkpoints.

In just 34 hours, the first complete three-dimensional model was on the table!

A week later, the refined map came out, with a resolution of 3.2 Å, so clear that even how amino acid side chains hold hands could be seen.

The resulting structure is quite stunning: the phage L05 vault is formed by 96 identical protein chains closing together in half top and bottom, and its internal capacity is a full 25% larger than that of the human vault, making it one of the largest known fixed-length protein shells besides viral capsids.

A mysterious molecule found on the inner wall, and it can be expanded by 50% like Lego

However, this also raises a geometric puzzle: 48 identical protein chains on one side converge toward the top, getting narrower and narrower—how do they avoid colliding?

High-resolution electron microscopy images revealed its exquisite division of labor.

Every three adjacent protein chains go their separate ways at the top: the first folds inward, the second folds outward to make room, and only the third continues toward the center, fastening into a perfectly fitted ring overhead.

Not only that, but when examining the electron microscopy density map, Claude also keenly spotted a detail that the naked eye could easily miss—inside the inner pocket of each protein chain, there was a mysterious substance lodged that did not belong to the protein itself!

Claude compared 25,000 kinds of intracellular small molecules one by one, and ultimately locked onto the prime suspect: coenzyme A (Coenzyme A), a core molecule of cellular metabolism.

This also handed key clues to the unsolved case: this large empty barrel is not an empty decoration; it is very likely a "molecular container" used to lock up key cofactors and nucleic acids when cells encounter viral attacks or metabolic crises.

After figuring out the assembly rules, the team, with Claude's cooperation, simply set about disassembling and expanding this nanoscale container.

By changing just one key amino acid at the waist or top, the originally sealed capsule instantly could no longer close, and under the electron microscope it flattened out into blossoming "protein lotus pedestals."

Conversely, by inserting two extra repeat modules into the middle of the barrel wall (L05+2), the lengthened protein chains could still close together perfectly.

After expansion, the internal space is 20% larger than that of the original phage vault and a full 50% larger than the natural human vault!

The pharmaceutical world has long wanted to use vaults to encapsulate drugs and deliver mRNA vaccines, but human vaults are extremely difficult to produce in E. coli, and their size cannot be changed.

Now this prokaryotic chassis, which can be rapidly mass-produced in bacteria and can also be resized on demand, happens to solve this problem.

The "Moore's Law" moment for scientific discovery has arrived

Surprisingly, this research was completed across disciplines by a team of "novices" in structural biology.

Apart from the internal multi-agent cluster used for 3D electron microscopy reconstruction, all the remaining gene mining, DNA design, purification protocols, and small-molecule screening were accomplished entirely with Claude's standard tools, which ordinary people can use at hand.

At the end of the paper, Claude was solemnly named—

All code in this study was written by Claude, and all computational analyses were performed by Claude. The authors were responsible for guiding the analyses and take responsibility for the content.

After reading the paper, Professor Leonard Rome, who personally discovered the vault back then, could hardly contain his excitement and highly praised this breakthrough for opening an entirely new path to unraveling the mystery of the vault's function.

For a problem that has trapped the academic world for 40 years, Claude not only found the blueprint in the bacterial world, but also helped humans disassemble and expand its structure.

The "Moore's Law" moment for basic scientific discovery has thunderously arrived.

References:

This article is from the WeChat public account "Xinzhiyuan" (ID: AI_era), author: ASI Revelation; editors: Moses, Taozi. Published by 36Kr with authorization.