Drug discovery
Knowing a target protein's shape is the starting gun for designing a drug that fits it. Researchers are compressing early-stage discovery from years toward months.
A protein's job is decided by the exact shape it folds into — and for fifty years, working out that shape was one of biology's hardest problems. Then DeepMind's AlphaFold predicted the structure of nearly every protein known to science in about a year, and gave the whole library away for free. This is the story of the problem, the breakthrough, and the hockey-stick moment when a fifty-year climb went vertical. Every figure and quote is real; sources are at the bottom.
Your body runs on proteins — they digest food, fight infection, carry oxygen, fire neurons. Each one starts life as a flat chain of chemical beads called amino acids (there are 20 kinds). In a fraction of a second, that chain folds itself into a precise three-dimensional shape, and that shape is what the protein does. Get the shape wrong and the machine jams — misfolded proteins are implicated in Alzheimer's, Parkinson's, and cystic fibrosis. The hard problem: given only the flat chain, what shape does it fold into?
The sequence is easy to read. The shape is not. Modern machines can read the amino-acid sequence of a protein quickly and cheaply. But the sequence is just a list of beads — it doesn't tell you how the string folds up in space, and the folded shape is the only thing that matters for how the protein behaves.
The beads pull on each other in thousands of tiny ways at once — some repel water, some attract each other, some carry charge — and the chain settles into the one shape that balances every force. Predicting that final shape from the sequence alone is the protein folding problem.
In 1969 the molecular biologist Cyrus Levinthal pointed out something strange. A protein could fold into an astronomical number of possible shapes — yet real proteins find the right one in milliseconds. If a protein tried every possibility one at a time, it would take longer than the universe has existed. This is Levinthal's paradox. Drag the length below and watch the search space explode.
Nature solves this instantly; computers could not. For decades, the only reliable way to learn a protein's shape was to stop calculating and start measuring it in a lab — atom by painstaking atom.
Before AI, biologists determined a protein's shape experimentally — growing crystals of it and firing X-rays through them, or freezing it for an electron microscope. The work is exacting and slow. A single structure could take months to years and cost hundreds of thousands of dollars; determining one protein was often enough to fill a researcher's entire doctorate.
Put those two numbers together and the scale of the problem is clear: at the pace of hands-on experiments, mapping every known protein wasn't a matter of years or decades. It was effectively never.
DeepMind trained a neural network on the ~170,000 structures biologists had painstakingly solved, teaching it to predict a folded shape directly from an amino-acid sequence. At the field's blind test in 2020, it didn't just compete — it effectively solved the 50-year problem, then scaled from a handful of proteins to hundreds of millions.
Here is the whole story in one line. Every protein structure humanity had ever determined by experiment, accumulated over five decades, versus what AlphaFold added in two years. On a normal (linear) axis, half a century of Nobel-grade laboratory work barely lifts off the floor. Switch to a logarithmic axis to see that history — and the near-vertical jump that dwarfs it.
Doing all of that experimentally would have taken the field an almost unimaginable amount of time. DeepMind's own framing, and its founder's, put a number on it — and then they gave the entire result away.
"It's kind of like a billion years of PhD time done in one year."
DeepMind estimates the database has saved researchers hundreds of millions of years of collective effort. But the number isn't the point — the point is what all those saved years unlock. Here is what people are already using it for.
Knowing a target protein's shape is the starting gun for designing a drug that fits it. Researchers are compressing early-stage discovery from years toward months.
Structures of proteins in malaria, tuberculosis, and antibiotic-resistant bacteria are helping scientists target diseases that kill millions each year.
Teams have used AlphaFold structures to engineer enzymes that break down single-use plastic, a route toward faster recycling.
Because it's free, labs in lower-income countries — often working on diseases the market ignores — get the same starting point as the best-funded institutions.
AlphaFold is the clearest case yet of a pattern worth understanding: a problem declared intractable for half a century, then reframed and dissolved by machine learning — and the value multiplied by giving the result away rather than hoarding it.
The protein folding problem was a fifty-year grand challenge until, suddenly, it wasn't. It's worth asking which of your own "unsolvable" bottlenecks are next.
DeepMind's influence grew by releasing AlphaFold free. Open assets can compound into more value — and goodwill — than locking them away.
AlphaFold didn't simulate physics faster; it learned patterns from 170,000 hard-won examples. The lesson generalises to most real AI wins.
The hard part isn't believing AI can do remarkable things — AlphaFold settles that. It's identifying which problem in your organization is a protein-folding problem: enormous, expensive, and quietly ready to fall. That's the conversation we have with leadership teams before they commit budget.