How an ‘accidental chemist’ honed his approach at UChicago on the way to a Nobel Prize

When future Nobel Prize winner John Jumper arrived at the University of Chicago for his Ph.D. in chemistry, there was a slight hitch.

“When I started, I knew no chemistry. None,” Jumper says. “I had to speedrun the whole thing. I managed to stay a week ahead of the undergrad gen-chem class I was TA’ing.”

Jumper, who was awarded the 2024 Nobel Prize in Chemistry, was an unusual student in other ways, too. His first love was physics. But after starting a Ph.D. in physics and finding little joy in it, he left and instead took a job at a company writing programs to model proteins with computation.

The problem hooked him. “It was so easy to see the connections; if we do this right, someone goes home from the hospital,” he says.

He decided to pursue a chemistry Ph.D. at the University of Chicago instead, working with Profs. Karl Freed, the Henry G. Gale Distinguished Service Professor Emeritus of the Department of Chemistry; and Tobin Sosnick, the William B. Graham Professor of Biochemistry and Molecular Biology.

His goal was to understand a question that had vexed scientists for more than half a century: how proteins fold.

“As an ‘accidental chemist,’ I wanted to go from understanding a very narrow slice of the problem, to getting a better sense of the problem as a whole,” Jumper says.

A knotty question

Every day, every hour, thousands of proteins are coming off the assembly line in your cells and twisting themselves into intricate pretzels. Your body makes hundreds of different kinds of proteins, each with their own jobs: transporting oxygen in your blood, making antibodies, enabling the synapses that power thoughts in your brain, and so on.

All of these specialized proteins are simply different combinations of the same 20 building blocks, known as amino acids. To do their particular jobs, they instantly and spontaneously fold up into different shapes. A protein knows how to do this instinctively. If you unfolded it, it would go back into the exact same shape once you let go.

With modern DNA sequencing, scientists could pinpoint exactly what amino acids a protein was made of. But there are more potential ways a protein could fold than there are stars in the sky, and no one could reliably use the genetic sequence to predict how that protein would actually then fold up.

“We’ve known the basic makeup of proteins since 1961,” Sosnick says, “and yet we’d been at this for 60 years, and we still couldn’t predict what shape a particular protein would take.”

Unfortunately, knowing the shape of a folded protein is very important. You could be seeing hints that a protein is involved in, say, Alzheimer’s disease, but without the structure, it’s hard to figure out where that protein binds to and what job it’s doing. Or you could discover that one particular protein is enabling a cancerous tumor’s growth, but without the structure, you can’t design a drug to block it. “Without knowing the structure of a protein, it’s very hard to understand a protein,” says Sosnick.

For years, the only way to get a good sense of the final structure of a folded protein was to make a large quantity of that protein, crystallize it, run those crystals through a high-powered beam, and interpret the way the light scattered off the crystals to recreate the 3D shape. This process worked well, but it takes time and money and large scientific facilities. Despite decades of advances in computation, no one had managed to make a program that could reliably take the genetic sequence and predict the outcome.

‘Even if I only held a pipette for three months, I benefitted’

At UChicago, Jumper began to study protein folding in Sosnick and Freed’s joint laboratory—an unusual combination of scientists who tackle problems from both theoretical and experimental perspectives, investigating questions at the intersection of biology, chemistry and physics.

“I think it was really an important part of my development to go from a physicist who happened to do biology, to someone who could think more deeply about the question,” Jumper says. “I’m a big believer in interdisciplinary science. You learn so much more by talking to a diverse group of experts to learn what they’re excited about and how they think about things.”

As part of the Ph.D. program, he rotated through different areas, including a brief stint in the laboratory of Prof. Steven Meredith, making beakers full of actual proteins. That was plenty long enough for him to know “the lab and I were better apart,” he laughs. “But it turned out to be a really valuable experience, to understand what experimentalists can do easily and what is hard; to know what problems they most need answered; to be familiar with the sources of uncertainty. Even if I only held a pipette for three months, I benefitted.”

For Jumper’s thesis, he built a code that uses machine learning to simulate the process of proteins folding up. (This is related to the protein folding question, but is a different research area than directly predicting protein structure from DNA.)

In 2017, he was awarded his Ph.D. and spent a few months as a postdoctoral researcher in Sosnick’s lab before taking a job with Google’s DeepMind.

There, he joined future co-Nobel Laureate Demis Hassabis and was quickly promoted to lead a team to attack the singular problem of predicting a protein’s final conformation based only on its genetic sequence.

Predictions and proteins

After several years of experimentation, Jumper, Hassabis and the DeepMind team unveiled AlphaFold—a program that, as the Nobel Prize committee would later write, “successfully utilized artificial intelligence to predict the structure of almost all known proteins.”

You can feed any genetic sequence into it—whether it be from a fruit fly or a cow or a human—and AlphaFold will spit out its prediction, and an estimate for how sure it is about that prediction.

It was an enormous leap forward in the field. “It hit the level where it transformed what people thought was possible,” says Sosnick.

AlphaFold is not 100% accurate—if they need it to be perfect, scientists usually still confirm the structure with crystallography—but it is tremendously useful. Thousands of scientific papers have already cited AlphaFold since it was released as an open-source program in 2021.

Scientists have used it to predict the structure of a protein in a new virus causing illness in humans, and hope it could speed up the response to future epidemics or pandemics. They have predicted possible new drugs. Having such a large number of proteins also allows scientists to look for overall patterns in evolution, such as scanning for useful properties, like bacteria that could break down pollutants, or understanding how viruses evolve.

AlphaFold draws from a database of more than 200,000 protein structures that have already been decoded and entered into the global Protein Data Bank. It uses deep learning, a type of AI, which takes insights from the bank of protein structures and applies them to predict what new structures might look like.

“It can be viewed as the first real scientific breakthrough of artificial intelligence,” wrote Johan Åqvist, member of the Nobel Committee for Chemistry.

For this, Jumper and his colleague were awarded half of the 2024 Nobel Prize in Chemistry. The other share went to David Baker of the University of Washington, for his work in a related but reverse problem: finding the genetic sequence that folds to a desired final protein.

More to offer

Meanwhile, the code Jumper wrote as part of his Ph.D. is still regularly in use. Sosnick and a team recently received a grant from the National Science Foundation to broaden the code to investigate different questions. “He’s the graduate student that keeps on giving,” says Sosnick with a smile.

Sosnick remembers Jumper as happy to engage in the particular UChicago tradition of friendly but vigorous discussion. “Everyone would happily argue through things to figure out what’s right and what’s wrong,” Sosnick says, whether in the lab or beer in hand at a backyard barbeque.

“He was very willing to spend his time contributing to others’ projects. He was a co-author on a lot of publications while he was here, which is a signal of how generous he was with his time and effort.”

Jumper and his fellow Nobel laureates will be awarded their Nobel Prizes at an official ceremony in Sweden on Dec. 10. Each newly minted Nobel laureate also will give a talk on their area of expertise.

As for the future, Jumper thinks machine learning and AI have more to offer science. “I think it’s going to start to change how we approach really complicated problems,” he says. “Nature is complex, and I think neural networks can handle complexity in surprisingly useful ways. I’m interested to see how far it can be pushed.”