DeepMind’s protein-folding AI stuns with a solution to one of biology’s biggest challenges

At the start of a biennial contest to predict the structure of proteins, the expectations for Google’s artificial intelligence unit DeepMind couldn’t have been higher. Think Mike Tyson in the mid-1980s: Everyone was expecting a knockout.

The results of the contest, known as CASP, came out Monday — and DeepMind didn’t disappoint, stunning the field by essentially solving one of biology’s most enduring challenges: quickly and accurately predicting the 3D structure of a protein from its amino acid sequence. The discovery stands to accelerate drug discovery by giving scientists more precise information about how proteins function within cells, allowing them to better target those proteins to counteract the mechanisms underlying disease.

Even John Moult, a computational biologist at the University of Maryland who co-founded the contest in 1994, said DeepMind’s performance this year doesn’t leave much left to figure out.


“This is a big deal,” Moult told Nature. “In some sense the problem is solved.”

In the contest, teams are asked to submit predictions for proteins whose structures have already been solved by other means, such as X-ray crystallography, but not publicly disclosed. A team of independent scientists then compares the predictions to the previously determined structure.


DeepMind’s program, called AlphaFold, outperformed its competitors — mostly academic labs — by a wide margin, scoring about 90 out of 100 on protein structures deemed moderately difficult, while most others scored around 75. Nearly two-thirds of its predictions were comparable to structures previously determined in the lab.

“I’m impressed,” wrote Derek Lowe, a drug discovery columnist for Science Translational Medicine. “We’re not up to ‘guaranteed protein structure for whatever you put it’, but getting that level of structural accuracy on that many varied proteins is something that has just never been done before.”

The company’s breakthrough essentially means that it figured out how to use AI to deliver relatively quick answers to questions about protein structure and function that would take many months or years to solve using currently available methods.

One researcher who judged the contest, Andrei Lupas, a biologist at the Max Planck Institute for Developmental Biology in Germany, said DeepMind was able to predict the structure of a protein that had confounded his lab for more than a decade. “This will change medicine,” Lupas told Nature. “It will change research. It will change bioengineering. It will change everything.”

DeepMind’s AlphaFold research team said in a blog post that it intends to describe its methods in a peer-reviewed journal in the coming months. In the meantime, the team described its performance in CASP as one of the company’s “most significant advances to date,” noting that it could help advance the understanding of hundreds of millions of unmodeled proteins.

“For all of us working on computational and machine learning methods in science, systems like AlphaFold demonstrate the stunning potential for AI as a tool to aid fundamental discovery,” the team wrote.