Claude Just Cracked a 40-Year Biological Mystery in Days

Biologists discovered massive hollow barrel-shaped structures inside cells back in 1986, promptly naming them vault particles and then spending nearly forty years scratching their heads over what they actually do. These mysterious capsules were suspected to handle intracellular transport or immune defenses, but proving it required mapping their exact architecture—a task that stubbornly resisted traditional methods. A research team decided to see if Anthropic's flagship AI could speed things up by tackling the structure of a prokaryotic vault-like particle from scratch.

The digital assistant started by scanning a database of roughly 15 billion protein sequences to track down evolutionary relatives, helping scientists cherry-pick 44 variants for lab testing. Instead of manually designing DNA constructs, the researchers used the model to prep the genetic blueprints for bacterial production, completely bypassing the usual bureaucratic nightmare of trial and error. When it came time to harvest the material, standard three-day isolation protocols were compressed into swift 90-minute sessions that produced clean samples ready for microscopy.

For the heavy-duty structural analysis, the workflow relied on cryo-electron microscopy alongside specialized data processing software called RELION, which the AI operated directly to chew through experimental readouts. The resulting macromolecule turned out to be a hefty 9.5-megadalton beast built from 96 identical protein copies arranged cleanly as 48 per half-shell. Researchers also engineered 15 mutant variants, intentionally breaking the closing mechanism on some versions to map out the exact molecular seams required for full assembly.

The entire empirical loop wrapped up in under four months, yielding the first-ever high-resolution 3D model of this specific protein family in a mere 34 hours of active compute time instead of the customary weeks. Naturally, the silicon co-pilot wasn't flawless, managing to throw a two-week wrench into the gears with a bogus laboratory protocol recommendation and completely skipping a data preprocessing step that initially spat out a bizarrely mirrored structure. Speed always demands a tax, and in this case, the toll booth operator occasionally hands out fake tickets.

When an algorithm starts doing the heavy lifting in wet labs, the boundary between computer science and molecular biology dissolves into a blur of automated pipettes and frantic prompt engineering.