DeepMind Releases the Predicted Shapes of Nearly Every Human Protein, Free

Eight months after its AI cracked protein folding, the lab has published predicted structures for about 20,000 human proteins, along with those of 20 other organisms, in a public database anyone can search.

DeepMind and a European bioinformatics institute opened a website today that contains something biology has never had: the predicted three-dimensional shape of nearly every protein in the human body. About 20,000 human proteins are in it, covering 98.5 percent of the ones the body makes, plus 350,000 proteins from twenty other organisms used in research — mice, fruit flies, yeast, the malaria parasite, the tuberculosis bacterium. It is free, and it is searchable by anyone.

Last November, the lab's AlphaFold program won a long-running competition by predicting protein shapes about as well as laboratory experiments do. This is the follow-through. Before today, scientists had experimentally determined the shapes of roughly 17 percent of human proteins, after decades of work. The database does not make the experiments unnecessary — a prediction is a prediction — but each entry comes with a confidence score, and for about a third of the human proteins the program rates its own answer as highly reliable.

The shape of a protein is usually the first thing a researcher needs to know to understand what it does or how a drug might attach to it. Scientists who have had early access describe the effect as skipping a year of work at the start of a project. The lab says it will expand the database to cover more than 100 million proteins — nearly every one known to science — over the coming year.