An AI Solves a Fifty-Year-Old Biology Problem: Predicting the Shape of a Protein

At a competition held every two years, DeepMind's AlphaFold predicted the three-dimensional shapes of proteins almost as accurately as laboratory experiments that take months. Scientists who run the contest say the problem is, in large part, solved.

Every two years since 1994, scientists have held a contest called CASP: they take proteins whose shapes have just been worked out in a laboratory but not yet published, and they challenge computer programs to predict those shapes from scratch. For twenty-five years the programs came up short. This morning the organizers announced that one of them didn't.

Proteins are the working parts of every living cell — the things that carry oxygen, digest food, fight infection. Each one is a chain of smaller molecules that folds up into a specific shape, and the shape determines what the protein does. Figuring out a shape in the lab can take a year of painstaking work with X-rays. Predicting it from the chain alone has been one of biology's great unsolved problems since the 1970s.

AlphaFold, built by DeepMind — the Google-owned lab behind the Go program — scored a median of 92 out of 100 on the competition's accuracy scale, where a score in the 90s is considered as good as an experiment. The best entry two years ago scored in the 60s. John Moult, who co-founded the contest, said this is the first time a serious scientific problem has been solved by AI.

Only a few thousand of the roughly 200 million known proteins have had their shapes determined. Knowing the shape is the starting point for understanding a disease or designing a drug. If the predictions hold up outside the contest, the research that used to wait a year for a structure may now wait an afternoon. DeepMind has not yet said when, or whether, it will make the program available to other scientists.