ABOVE: David Baker (left), Demis Hassabis (center), and John Jumper (right) won the 2024 Chemistry Nobel Prize for developing tools to design proteins and predict protein structures. Ill. Niklas Elmehed © Nobel Prize Outreach
The 2024 Nobel Prize in Chemistry goes to researchers who cracked the code for proteins’ structures, the Royal Swedish Academy of Sciences announced today (Oct 9). David Baker, a biochemist at the University of Washington, won one half of the prize for his work on computational protein design. Demis Hassabis and John Jumper, computer scientists at Google DeepMind, shared the second half of the prize for their work on protein structure prediction.
“In order to understand how proteins work, you need to know what they look like,” said Johan Åqvist, a member of the 2024 Nobel Committee for Chemistry. “And that’s what this year’s laureates have done.”
Proteins are the building blocks of life. They are made up of strings of amino acid molecules which fold themselves into complex three-dimensional shapes. These structures consist of thousands of atoms, whose position in space relative to each other determines the protein’s function, ranging from aiding biochemical reactions to defending against pathogens.
“To understand how life works, we first need to understand the shape of proteins,” said Heiner Linke, chair of the 2024 Nobel Committee for Chemistry.
In the 20th century, scientists were focused on trying to understand how proteins take certain forms. Nobel prize-winning research by Christian Anfinsen in 1961 laid the foundation for these studies when he showed that the amino acid sequence governs the three-dimensional structure of a protein. This meant that biologists should be able to predict a protein’s structure based on its amino acid sequence alone, without having to rely on tedious methods like X-ray crystallography. However, this protein prediction problem challenged scientists for decades.
Then, around the turn of the century, David Baker made a breakthrough while exploring how proteins fold. Using insights from his experiments, he developed a computer program called Rosetta that would predict protein structures from amino acid sequences. Soon, Baker and his team realized that they could reverse engineer the problem. In an interview with The Scientist earlier this year, Baker said, “It wasn't too long after our first successes in structure prediction that we started thinking, well, maybe instead of predicting what structure a sequence would fold up to, we could use these methods to make a completely new structure and then find out what sequence could fold to it.”
Using Rosetta, Baker’s team designed a completely new protein in 2003 and synthesized it in the lab.1 It served as a proof-of-concept that computational protein design was feasible.
Despite the success of the Rosetta software, the accuracy of protein structure prediction remained low—a protein of even a few amino acids can take a wide range of structures. For many years, even the most successful computer programs could predict the three-dimensional structure of an amino acid sequence with only 40 percent accuracy. Then, artificial intelligence (AI) models entered the scene.
Prior to cofounding DeepMind in 2010, Hassabis trained as a neuroscientist studying neural networks for AI. For his newly founded company, he applied this expertise to develop AI models for popular board games, including Go. However, he wanted to apply his skills towards something that could have a greater impact on humanity. Hassabis and his team developed AlphaFold, an AI model that predicted protein structures with an accuracy of 60 percent. Although a major improvement, it was still far from the 90 percent goal he set out to achieve.
In 2017, Jumper, a newly-minted PhD, joined DeepMind with hopes to use his experience of protein simulation and theoretical physics to improve AlphaFold’s accuracy further. Jumper and Hassabis co-led the charge on training the deep learning-based model on the sequences and structures of around 100,000 proteins, to more accurately predict protein structure.2 The new model also borrowed from concurrent advances in the field of AI: transformers, or neural networks that could more flexibly search for patterns hidden in large amount of data. They iterated their model to improve its accuracy, and by 2021 AlphaFold2 could predict protein structures with almost 90 percent accuracy.3
Hassabis and Jumper have used their model to predict the structures of most of the proteins that researchers have identified. Since the launch of these tools, researchers around the world have been using them to develop new proteins for vaccines and targeted therapies, and to determine the structures of proteins that decompose plastic or lead to antibiotic resistance.
“The impact of this year’s prizes in protein design and protein structure predictions is truly huge,” said Åqvist.
Work done by Baker, Hassabis and Jumper has not only solved a decades-long protein folding problem, but it also helped usher in a new era of AI-based modeling to understand and design the fundamental molecules that enable all life.
- Kuhlman B, et al. Design of a novel globular protein fold with atomic-level accuracy. Science. 2003;302(5649):1364-1368.
- Senior AW, et al. Improved protein structure prediction using potentials from deep learning. Nature. 2020;577(7792):706-710.
- Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589.