DeepMind’s protein folding AI solved a 50-year-old biology challenge

In this year’s CASP, AlphaFold predicted the structure of dozens of proteins with a margin of error of just 1.6 angstroms, or 0.16 nanometers. This far exceeds all other computational methods and, for the first time, matches the accuracy of experimental techniques for mapping protein structure in the laboratory, such as cryoelectron microscopy, nuclear magnetic resonance Yes X-ray crystallography. These techniques are expensive and time consuming – they can cost hundreds of thousands of dollars and years of trial and error for each protein. AlphaFold can return to protein form in a matter of days.

But identifying the structure of a protein is very difficult. For most proteins, researchers have the amino acid sequence on the strip, but not the way they are folded. And there is usually an astronomical number of possible shapes for each sequence. Researchers have struggled with the problem since at least the 1970s, when Christian Anfinsen won the Nobel Prize for showing that sequences determine structure.

This breakthrough could help researchers design new drugs and understand diseases. In the long run, predicting protein structure will also help design synthetic proteins, such as enzymes that digest waste or produce biofuels. Researchers are also exploring ways to introduce synthetic proteins that will increase crop yields and make plants more nutritious.

Shocking advance

“This is a very substantial advance,” says Mohammed AlQuraishi, a systems biologist at Columbia University who has developed his own system for predicting protein structure. “This is something that I wasn’t expecting so quickly. It’s shocking, in a way.

“It’s a really big deal,” says David Baker, director of the Institute for Protein Design at the University of Washington and team leader behind Rosetta, a family of protein analysis tools. “It’s an incredible achievement, like what they did with Go.”

The launch of CASP in 1994 gave the field of protein structure prediction a boost. Every two years, the organizers publish around 100 amino acid sequences for proteins whose forms have been identified in the laboratory but are not yet made public. Then dozens of teams around the world compete against each other to find the right way to fold them using software. Many of the tools developed for CASP are already in use by medical researchers. But progress has been slow, with two decades of incremental advancements failing to produce a shortcut to painstaking lab work.

When DeepMind entered the competition in 2018 with its first release of AlphaFold, it gave CASP the shock it was looking for. It still couldn’t match the accuracy of a lab, but it left other computational techniques in the dust. The researchers took note: soon, many were adapting their own systems to function more like AlphaFold.

This year, more than half of the applications use some form of deep learning, says Moult. As a result, the overall accuracy was higher. Baker’s new system, called trRosetta, uses some of DeepMind’s ideas from 2018. But it arrived in a “very distant second,” he says.

In CASP, results are scored using what is called a Global Distance Test (GDT), which measures on a scale of 0 to 100 how close a predicted structure is to the actual shape of a protein. identified in laboratory experiments. The latest version of AlphaFold performed well for all challenge proteins. But it got a GDT score above 90 for about two-thirds of them. Their GDT for the hardest proteins was 25 points higher than that of the next best team, says John Jumper, who leads the AlphaFold team at DeepMind. In 2018, the advantage was around six points.