AI system solves 50-year-old protein folding problem in hours

An abstract illustration of pink, purple and green ribbons based on protein structures.
(Image credit: Shutterstock)

An artificial intelligence company that gained fame for designing computer systems that could beat humans at games has now made a huge advancement in biological science.

The company, DeepMind, which is owned by the same parent company as Google, has created an AI system that can rapidly and accurately predict how proteins fold to get their 3D shapes, a surprisingly complex problem that has plagued researchers for decades, according to The New York Times.

Figuring out a protein's structure can require years or even decades of laborious experimentation, and current computer simulations of protein folding fall short on accuracy. But DeepMind's system, known as AlphaFold, required only a few hours to accurately predict a protein's structure, the Times reported.

Related: Why does artificial intelligence scare us so much? 

Proteins are large molecules that are essential for life. They are made up of a string of chemical compounds known as amino acids. These "strings" fold in intricate ways to create unique structures that determine what the protein can do. (For example, the "spike" protein on the new coronavirus allows the virus to bind to and invade human cells.)

Nearly 50 years ago, scientists hypothesized that you could predict a protein's structure knowing just its sequence of amino acids. But solving this "protein folding problem" has proved enormously challenging because there are a mind-boggling number of ways in which the same protein could theoretically fold to take on a 3D structure, according to a statement from DeepMind.

Twenty-five years ago, scientists created an international competition to compare various methods of predicting protein structure — something of a "protein olympics," known as CASP, which stands for Critical Assessment of Protein Structure Prediction, according to The Guardian.

In this year's challenge, AlphaFold's performance was head and shoulders above its competitors'. It achieved a level of accuracy that researchers were not expecting to see for years. 

"This computational work represents a stunning advance on the protein-folding problem, a 50-year-old grand challenge in biology," Venki Ramakrishnan, president of the Royal Society in the United Kingdom, who was not involved with the work, said in a statement. "It has occurred decades before many people in the field would have predicted. It will be exciting to see the many ways in which it will fundamentally change biological research."

For the competition, teams are given the amino acid sequences of about 100 proteins, the structures of which are known but have not been published, according to Nature News. The predictions are given a score from zero to 100, with 90 considered on par with the accuracy of experimental methods.

AlphaFold trained itself to recognize the relationship between the amino acid sequence and protein structure using existing databases. Then, it used a neural network —  a computer algorithm modeled on the way the human brain processes information — to iteratively improve its prediction of the unpublished protein structures.

Overall, AlphaFold had a median score of 92.5. That's up from a score of less than 60 that the system achieved in its first CASP competition in 2018.

The system isn't perfect — in particular, AlphaFold did not perform well in modeling groups of proteins that interact with each other, Nature News reported.

But the advance is a game-changer.

"I think it's fair to say this will be very disruptive to the protein-structure-prediction field. I suspect many will leave the field as the core problem has arguably been solved," Mohammed AlQuraishi, a computational biologist at Columbia University told Nature News. "It's a breakthrough of the first order, certainly one of the most significant scientific results of my lifetime."

DeepMind previously made headlines when it created an AI program, known as AlphaGo, that beat humans at the ancient game of Go.

Researchers hope AlphaFold can have many real-world applications. For example, it could help identify the structures of proteins involved in certain diseases and accelerate drug development.

DeepMind is currently working on a peer-reviewed paper on its work on AlphaFold, the Times reported.

Originally published on Live Science.  

Rachael Rettner

Rachael is a Live Science contributor, and was a former channel editor and senior writer for Live Science between 2010 and 2022. She has a master's degree in journalism from New York University's Science, Health and Environmental Reporting Program. She also holds a B.S. in molecular biology and an M.S. in biology from the University of California, San Diego. Her work has appeared in Scienceline, The Washington Post and Scientific American.