Google’s DeepMind researchers among recipients of Nobel prize for chemistry
IT IS A recurring joke among chemists that the Nobel prize for chemistry is, more often than not, awarded for developments in biology. Recent examples include awards for the gene-editing tool CRISPR, in 2020; directed evolution of enzymes and antibodies, in 2018; and DNA repair mechanisms, in 2015. Some may view this year’s prize, which awarded work designing and predicting the structure of proteins, as a continuation of that trend. But the main message lies elsewhere: some of the best brains in chemistry do not only make molecules, they make computer models too.
One half of the award went to David Baker, a biochemist at the University of Washington, for his work on designing new proteins using computers. The other half was shared between John Jumper and Demis Hassabis from DeepMind, Google’s artificial-intelligence (AI) company, for their development of AI models capable of predicting three-dimensional protein structure, a long-standing grand challenge in biochemistry.
Dr Baker has long been considered a favourite to win. The choice of Drs Hassabis and Jumper, though, came as something of a surprise. But this has been AI’s year. The day before the chemistry announcement, the prize for physics was awarded for the development of the neural networks that underpin artificial intelligence models such as those DeepMind pioneered (a subject, some argued, that hardly counts as physics at all).
Proteins are the main chemical building blocks of life. They are made up of smaller molecules called amino acids, arranged in long chains which fold in highly complex and specific ways. A protein’s final folded form determines its biological function. In other words, to understand proteins—and, by extension, biology—one must understand their structure.
Dr Baker achieved such understanding through doing. In a landmark paper from 2003, he succeeded in designing a completely new protein. Using a computer program he named Rosetta, he found an amino-acid sequence capable of folding in ways not seen in nature. Once the sequence was recreated in the lab and the protein formed, he determined its final structure using a technique called X-ray crystallography: it was a close match to what he had set out to make. Rosetta, now called Rosetta Commons, has subsequently become a software package used by every protein chemist, and computational protein design has assisted in everything from vaccine development to the detection of toxic chemicals.
Going the other way, and predicting a protein’s structure from its amino-acid sequence, is a problem that took even longer to crack. Given the near-limitless number of configurations into which a protein can fold—by some estimates, as many as 10300 for a single complex protein—even computers had limited success. DeepMind’s AI-based AlphaFold 1 and 2 models, made public in 2018 and 2020 respectively, were the first to even get close. AlphaFold 2 now has a database of more than 200m protein structure predictions, with a prediction accuracy approaching 90%.
Though Drs Jumper and Hassabis have featured on various contender lists this year, many wondered if it was too soon for AlphaFold to be recognised. Yet it has already had real impact: DeepMind says that some 2m scientists already use it in their research. AlphaFold 3, released in May, goes beyond proteins to predict the structure of a host of other biomolecules, such as DNA, as well as small molecules that might function as drugs. It can also predict how different molecules with different structures fit together, such as how a virus’s spike protein might interact with antibodies and sugars found in the body.
By choosing, for the first time, to honour work performed with an AI model, the committee has opened the door for more such prizes in the future. That is just as well; AI has been seeping into all areas of science for some time now, as Dr Baker illustrated when he was phoned up during the committee’s press conference. He said that AlphaFold has inspired him to make generative AI models that can design new proteins. “Our new AI methods are much more powerful,” he said, sounding happy and a little tired. If recent history is anything to go by, researchers will be repeating that line for years to come. ■