In the past few months, Baker’s team has worked with previously commented biologists trying to figure out what the proteins they were studying looked like. “There is a lot of great biological research that has been really accelerated,” he says. A public database of hundreds of thousands of ready-made protein forms should be an even greater accelerator.
“It looks amazingly impressive,” says Tom Ellis, a synthetic biologist at Imperial College London who studies the yeast genome, and is excited about the database trial. But he cautioned that most of the predicted forms have yet to be verified in the lab.
In the new version of AlphaFold, predictions come with a confidence score that the tool uses to indicate how close each predicted shape is to the real thing. Using this metric, DeepMind found that AlphaFold predicted the shapes of 36% of human proteins with accuracy down to the level of individual atoms. Hesabis says this is good enough for drug development.
Previously, after decades of work, the structures of only 17% of the proteins in the human body were determined in the laboratory. If AlphaFold’s predictions are as accurate as DeepMind says, the tool more than doubled that number in just a few weeks.
Even predictions that are not entirely accurate at the atomic level are still useful. For more than half of the proteins in the human body, AlphaFold predicted what shape should be good enough for researchers to learn the protein’s function. The remainder of AlphaFold’s current predictions are either incorrect, or they are for a third of the proteins in the human body that have no structure at all until they bind to others. “It’s flexible,” Hasbis says.
“The fact that it can be applied at this level of quality is impressive,” says Muhammad al-Quraish, a systems biologist at Columbia University who developed his own software for protein structure prediction. He also points out that having structures for most proteins in an organism would make it possible to study how these proteins function as a system, and not just in isolation. “That’s what I think is the most exciting,” he says.
DeepMind releases its tools and forecasts for free and won’t say if it has plans to make money from them in the future. However, he does not rule out the possibility. To create and operate the database, DeepMind is partnering with the European Molecular Biology Laboratory, an international research institution that already hosts a large database of protein information.
For now, Qureshi can’t wait to see what the researchers do with the new data. “It’s so amazing,” he says, “I don’t think any of us thought we’d be here so quickly. It’s mind-boggling.”