DeepMind Says It Cracked the Protein-Folding Problem

Every couple of years, structural biologists run a blind test called CASP - the Critical Assessment of Structure Prediction. Researchers submit predicted 3D shapes for proteins whose real structures have already been solved in a lab but not yet published, and the results get scored against reality. It’s been running since 1994, and for most of that time, “solving” it wasn’t really on the table. This year, DeepMind’s AlphaFold2 walked in and did something that’s being described as a genuine turning point.

The numbers, per the CASP14 organizers: AlphaFold2 hit a median score around 90 on the GDT scale (Global Distance Test, roughly a measure of how closely a predicted structure matches the real one, out of 100). That’s in the range experimental methods like X-ray crystallography or cryo-EM typically achieve. For a computational prediction to land that close to physically measuring the thing is what has people in the field using words like “landmark” and “historic.”

Why does this matter beyond the leaderboard? Protein folding has been an open grand challenge since the 1970s: you know a protein’s amino acid sequence, but the sequence alone doesn’t tell you how the chain folds into its final 3D shape, and shape is basically everything. A protein’s function - how it binds a drug molecule, how it interacts with other proteins, whether it misfolds into something that causes disease - comes down to structure. Labs have spent years and enormous grant money determining individual structures one at a time. If a model can predict structures computationally at near-experimental accuracy, that timeline collapses from months to hours.

The obvious applications people are already talking about: drug discovery (knowing a target protein’s shape is often step one for designing a molecule that binds it), understanding disease mechanisms tied to misfolded proteins, and enzyme design for things like plastic-degrading catalysts. DeepMind has said it wants to work with the scientific community on making this useful rather than just publishing a paper and moving on.

Worth tempering the hype a little, though. Some biologists in the field are already pointing out that “solved” doesn’t mean every protein, every time - performance reportedly varies by protein family and complexity, and there’s a difference between predicting a single static structure and modeling how proteins actually behave dynamically in a cell, interact with each other, or change shape when they bind something. There’s also the standard caveat with any splashy result announced via blog post and conference talk rather than full peer-reviewed publication: the details matter, and we don’t have all of them yet.

Still, take a step back. An AI system just did in a CASP cycle what an entire field spent five decades working toward. Whatever caveats apply, that’s not a small thing, and it says a lot about where AI-for-science is headed.