DeepMind Just Gave Away 350,000 Protein Structures for Free

This week DeepMind and the European Bioinformatics Institute flipped the switch on something that’s genuinely hard to overstate: a public, searchable database of roughly 350,000 protein structures predicted by AlphaFold, covering about 44% of the human proteome. Anyone — a grad student, a biotech startup, a curious hobbyist with a browser — can now pull up predicted 3D structures for proteins that, for decades, would have required months of crystallography or cryo-EM work to resolve, if anyone bothered to resolve them at all.

Protein structure has always been the bottleneck. You can sequence a genome in a day now, but knowing the sequence of amino acids tells you surprisingly little about what a protein actually does, because function follows shape. A protein’s job — binding a drug, catalyzing a reaction, signaling to a neighboring cell — comes from how that chain folds into a 3D shape. Figuring out that shape experimentally is slow, expensive, and often doesn’t work at all for trickier proteins. That gap between “we know the sequence” and “we know the structure” has been one of biology’s oldest and most stubborn problems.

AlphaFold’s splash at CASP14 last December already suggested DeepMind had basically cracked this at a level competitive with experimental methods. What’s new this week isn’t a bigger claim about accuracy — it’s scale and access. Instead of a research paper with a handful of showcase structures, there’s now a browsable database anyone can query, with EMBL-EBI hosting it so it’s positioned as a durable public resource rather than a one-off release tied to a single lab’s servers.

The timing is also interesting. This drop comes just days after the Baker Lab published their own structure-prediction system, RoseTTAFold, in Science. Two independent teams converging on deep-learning approaches to the same problem in the same summer is a pretty strong signal that this isn’t a fluke or a one-company party trick — protein structure prediction as a solved-ish problem looks like it’s here to stay, and now there’s competition pushing the field forward instead of one company holding all the cards.

What actually happens with 350,000 free structures is the fun part to speculate about. Structural biologists get a head start on proteins nobody’s had the bandwidth to crystallize. Drug discovery teams get more candidate binding pockets to poke at computationally before touching a wet lab. Basic researchers studying disease mechanisms get a shortcut past a step that used to eat entire PhD theses. None of this replaces experimental verification — a predicted structure is a hypothesis, not a fact, and I’d expect plenty of edge cases where AlphaFold’s confidence scores are more honest about uncertainty than people give them credit for. But going from “structure known” being the exception to being closer to the default, for a huge chunk of the human proteome, is the kind of infrastructure shift that quietly reshapes a field over the next few years rather than making headlines every week.

Worth watching whether DeepMind extends coverage beyond humans — model organisms used in research (mouse, yeast, E. coli) would be an obvious next target, and there’s been talk of eventually covering most of UniProt. If that happens, this stops being a nice dataset and starts being the default reference layer for molecular biology.