DeepMind AI Speeds Up the Time to Determine Proteins’ Structures
DeepMind AI Speeds Up the Time to Determine Proteins’ Structures

DeepMind AI Speeds Up the Time to Determine Proteins’ Structures

The technology solves proteins’ 3-D shapes in minutes, when traditional methods may take years.

Lisa Winter
Lisa Winter

Lisa joined The Scientist in 2017. As social media editor, some of her duties include creating content, managing interactions, and developing strategies for the brand’s social media presence. She also...

View full profile.


Learn about our editorial policies.

Dec 2, 2020

ABOVE: © ISTOCK.COM, SERGUNT

The shape of a protein has a direct bearing on its function and is a key component in drug discovery, and it can take years of experimentation to figure it out. On Monday, November 30, the Protein Structure Prediction Center at the University of California, Davis, announced that the DeepMind artificial intelligence lab and its AlphaFold program have accelerated the time to determine protein shape in a fraction of what it takes traditional methods to accomplish.

The AlphaFold program uses neural networks to perform deep learning, identifying patterns in sequences and structures of proteins found in a global database. As it learns over time, the program can identify the structure of a protein in minutes. Traditionally, researchers would use techniques such as X-ray crystallography or cryo-electron microscopy to visualize the protein. This is a time-consuming process that can take years to complete or even a person’s entire career.

DeepMind, based in London and owned by the parent company of Google, participated in the 14th biennial Critical Assessment of protein Structure Prediction (CASP) contest, hosted by the Protein Structure Prediction Center. Teams work out the structure of around 100 full or partial proteins, a few at a time, in a process that takes months to complete. 

Around half of the teams used deep learning this year and took notes from AlphaFold’s last CASP showing, Nature reports, but DeepMind was able to set itself apart by revamping its approach and considering the spatial constraints of protein folds and focusing on the protein’s sequence as a whole, not just the likely position of individual amino acids as it did in 2018.

To participate in the contest, labs compete to computationally determine the 3-D structures of proteins that the organizers have already verified experimentally, but have not published anywhere that the participants would be able to access. Scores are given based on accuracy, out of a possible 100. According to the organizers, 90 is the threshold for meeting experimental values.

Science reports that at the inaugural event in 1994, the average score was around 20 and had only grown to 40 by 2016. When DeepMind debuted in 2018, it dwarfed the competition at 60. For this year’s showing, it had a median score of 92.4 overall, and on the most challenging proteins, it scored a full 25 points higher than the competition at 87. 

“Most atoms [from DeepMind’s results] are within an atom diameter of where they are in the experimental structure,” John Moult, a professor at the University of Maryland who cofounded the contest, tells The New York Times. “And with those that aren’t, there are other possible explanations of the differences.”

See “A Primer: Artificial Intelligence Versus Neural Networks

“This will change medicine. It will change research. It will change bioengineering. It will change everything,” Andrei Lupas, an evolutionary biologist at the Max Planck Institute for Developmental Biology in Germany who helped judge the contest, tells Nature, adding that AlphaFold took only 30 minutes to produce the structure of a protein his lab had been trying to figure out for 10 years. 

“I always hoped I would live to see this day,” Moult tells the Times, with regard to a computer topping the 90-point mark. “But it wasn’t always obvious I was going to make it.”