Scientists have spent 40 years trying to figure out what certain mysterious particles do inside our cells. Now, Anthropic says Claude helped researchers uncover a new version of one of these particles, giving them another clue about how it works.
In a November 4 announcement, Anthropic revealed that its biology team used Claude to reconstruct the first 3D structure of a vault particle derived from a virus that infects bacteria.
Along the way, Claude reduced a three-day laboratory process to just 90 minutes. Pretty impressive, considering the researchers had limited experience with protein assemblies of this size.
But the AI also made mistakes, including one that delayed the experiment by two weeks.
So, how much scientific work can Claude handle before researchers need to step in?
The 40-year-old mystery Claude helped investigate
Back in 1986, scientists Nancy Kedersha and Leonard Rome discovered unusual protein particles that looked like tiny barrels.
They called them vault particles because their shape resembled the arched ceilings of cathedrals.
These hollow shells exist in many organisms, including humans. Scientists suspect they help move molecules around cells or protect against harmful substances, but nobody has established their exact biological function.
Previous researchers had already found related proteins in bacteria and viruses, although they couldn't confirm whether those proteins assembled into complete vault particles.
Anthropic's team wanted to find out.
With Claude's help, researchers identified promising proteins and produced them inside E. coli bacteria. They then used a powerful imaging technique called cryo-electron microscopy to reconstruct the particles in 3D.
One of the newly mapped particles contained 96 copies of a protein called L05, compared with 78 major vault protein copies in human vaults.
The scientists also experimented with its size. By duplicating part of the protein, they created a larger version with about 20% more internal space.
The results offer new clues about how these particles assemble and how their shapes might have changed over time.
Claude turned three days of lab work into 90 minutes
Here's where Claude's contribution gets interesting.
The researchers started by asking Claude Science to search a database of approximately 15 billion predicted proteins for relatives of known vault proteins. After the team selected promising candidates, Claude prepared the DNA instructions to produce them in the lab.
Then came a time-consuming problem: separating the vault particles from everything else inside the bacterial cells.
The conventional method takes about three days, so the researchers wanted to try a faster approach that separates particles based on their size. Claude initially advised against it because unwanted protein clumps could interfere with the results.
The scientists decided to test the method anyway.
Once they committed to the approach, Claude recommended equipment and suggested putting material from broken-open bacterial cells into the separation column. Anthropic says the revised procedure worked, cutting preparation time from roughly three days to 90 minutes.
And Claude's involvement didn't stop there.
It also helped process the microscopy images using specialized software. The team produced its first complete 3D model in 34 hours, but the final reconstruction took approximately a week.
Anthropic estimates that similar structural analysis would ordinarily require weeks of work. Those time savings sound promising, especially for researchers who might otherwise spend days preparing samples before they can even examine them.
When Claude got the science wrong
For all the time Claude reportedly saved, it also gave researchers instructions that caused some serious delays.
One laboratory protocol it generated included an unsuitable additive in the bacterial growth medium. The additive interfered with getting DNA into the cells, which Anthropic says set the project back by two weeks.
Claude made another mistake while processing the microscopy images. It skipped a standard step and produced an incorrect structural map that combined the particle's image with its mirror image.
The researchers caught the problem and told Claude to follow established procedures. Their instruction was amusingly straightforward: "Please let's do normal things unless they don't work."
These errors expose a problem researchers will need to address as AI takes on more scientific tasks. When scientists know a procedure well, they can usually recognize questionable recommendations. But what happens when they're relying on AI to help with experiments outside their expertise?
An incorrect instruction could go unnoticed until something fails, potentially costing days or weeks of work.
Anthropic says it shared the team's laboratory notes with the researchers who train Claude to improve its performance.
Could Claude help smaller labs make discoveries?
Anthropic says its team completed the entire project in less than four months, even though its researchers had limited previous experience in this area of structural biology. The company believes much of the work could be reproduced in a conventional laboratory with access to Claude.
There's a catch, though.
The microscopy analysis required hundreds of GPUs working simultaneously, which could put that part of the experiment beyond the computing budgets of many academic laboratories. The study also follows Anthropic's October experiment, where Claude designed proteins intended to attach to scorpion neurotoxins. Researchers tested 211 designs, and 17 met their binding criteria.
Together, the experiments suggest Claude could help scientists tackle research problems that would otherwise require additional specialist expertise.
But several questions remain about the vault discovery itself.
The researchers produced the particles inside laboratory-grown E. coli. They haven't confirmed whether the same particles naturally assemble in the bacteria or viruses where their proteins originate.
Claude also spotted an unidentified molecule inside the newly mapped particle that resembles coenzyme A, a substance involved in cellular metabolism. Scientists still need to determine what the molecule is.
Anthropic has submitted the findings as a preprint, and as of November 5, the research hasn't been reported as peer-reviewed or independently replicated.
The discovery could eventually help scientists understand what vault particles do and whether researchers can use similar shells to carry molecules for biotechnology applications.
For now, the experiment shows us Claude's capabilities in a research laboratory. It lets researchers investigate unfamiliar problems and speed up experiments, provided scientists can recognize when its advice goes wrong.
The next question is whether other laboratories can reproduce Anthropic's results and achieve similar time savings while catching Claude's mistakes before they derail an experiment.
