Our Terms & Conditions | Our Privacy Policy
John Jumper currently serves as director at Google DeepMind.
John Jumper (S.M. ’12, Ph.D. ’17) was awarded the Nobel Prize in Chemistry in 2024 for his contributions to the development of AlphaFold, an AI model that revolutionized protein structure prediction. He spoke to the Maroon about the journey that led to this historic achievement, the challenges faced, and his broader vision for the future of AI in biology and beyond.
The following transcript has been edited for clarity and brevity.
Chicago Maroon: What first inspired the development of AlphaFold?
John Jumper: For a long time, people wanted to do this because the protein sequence contains all the information about the protein structure. For me, personally, I kind of came to it indirectly through by chance ending up at this company studying how proteins move and getting very interested in these questions. After leaving that company and coming to UChicago as a grad student, I became interested in how we can use AI on this problem. Ultimately, it’s really about how we understand biological systems to make people healthy. How do we understand these proteins that are very important and their structure? Then protein folding, or protein structure prediction, is one of the problems we want to solve, but there are many. It’s a way in.
CM: Would you say that the ability to apply computational skills to biology is one of the main things that an education at UChicago provided you?
JJ: Yes, you could say that. My undergraduate was physics and math. I had come from some time at this company, D.E. Shaw Research, which was doing custom computer chips for studying protein simulation. What I really came to UChicago for was actually a better understanding of wetland biology. I wanted to do a Ph.D., but I also wanted to understand what are the questions people actually really want to answer with these systems. I came to it from the computational side and was looking for the scientific side.
I kind of learned computation from my own job. I think for multiple years UChicago was the place I learned about biology and biophysics. For me, UChicago was about learning what experimentalists want to do with these systems. How can I design and work together with the experiment to understand the problems? What problems are worth understanding? I gained a lot in the lab, understanding protein structure and protein predictions’ role in the larger experiments of understanding biological systems.
CM: What would you say was the most difficult problem that you had to overcome during the development of AlphaFold？
JJ: There are many hard problems, but for the really fundamental ones, it is that we have relatively little data by AI and machine learning standards. It’s a hard problem. We came up with a bunch of ideas. There’s about 200,000 known protein structures. Each structure takes a year of a Ph.D. student’s time to get. Therefore, you have a very, very fixed data set. The data set is growing very slowly, about 14,000 new structures per year. It’s very fixed for an AI system. AI systems aren’t like chatbots that get to learn from the entire internet. You have a small amount of data to use, and you have to use it really well.
Then you have to figure out how you put your understanding of biology or geometry or physics into this neural network such that it learns more from each data point, because you don’t really have enough data. You could go crazy and have fun. We tried writing a simulation with this very limited data, but that didn’t produce a predictive model. We had to find this kind of halfway between, in my view, simulation and pure machine learning, and that turned out to be more effective than either extreme.
CM: You are also the first scientist to be awarded a Nobel Prize for AI technology. What first sparked your interest in AI?
JJ: What really actually sparked my interest was when I was working at D.E. Shaw Research. I was working on some data analysis problems related to science. There was this undergraduate that had come in, and he knew a lot more about statistics than I did, and he had some really cool tools of high dimensional statistics. I hadn’t really learned modern statistics as I only learned the statistics of linear regression. But then the modern statistics can tell you a lot about really big, high dimensional systems that you can work with things with millions of variables and thousands of observations.
I started to fall in love with these tools of modern statistics. I started to read everything that I could in statistics and then machine learning. It was really David MacKay’s book about information theory and early AI that made me realize these ideas were really powerful. At the same time, I moved from having these custom computer chips at my job to having whatever computer resources were available to a first-year graduate student. I now had access to these incredibly powerful computing systems that were 1,000 times faster than anything I had access to before.
I was trying to build back with algorithms and AI what I could no longer do with custom computer chips because I wasn’t [at my previous job] anymore. That led me to my Ph.D., trying to figure out how all these ideas in AI relate to protein simulation and motion. Then that kind of naturally led to my work in DeepMind about how we are going to use AI to produce an effective system on protein knowledge.
CM: What other areas are you planning to explore with AI, either in or outside biology？
JJ: I think there’s a tremendous amount. There are many problems, and we have no shortage of them. We made a later version of AlphaFold that solves more problems and predicts how proteins bind to DNA. Then there’s the question of how we do drug development with these systems. How do we understand the cell? How when we think about all the things that go on in the cell and even the most perfect alpha form only describes a part of this? How do we use all this data? How do we learn from it? How would we form perspectives on it? I think there’s quite a lot here, but there’s really a huge number of different problems available. The question will really become which has good enough data sources which are kind of favorable to be solved. And we’ll find out.
CM: What do you think is the future of AI? There is a lot of talk about how it could potentially overtake humanity. Do you think that could happen?
JJ: I think the question is really one of time and time scale. It does seem likely that we will make generally intelligent AI systems at some point. But is that 10 years away or one hundred or a thousand? I don’t think we really know for sure.
I think maybe one distinction to draw is that chatbots are doing the most fundamentally human things that other humans can do. We’re really amazed that we can teach machines to do these things that previously we didn’t know how to program a computer to do. Then there’s systems like AlphaFold that are doing something that humans can’t. There’s no human that is good at protein structure prediction that they can do experimentally. But as a pure prediction, no one is good at that. So I think AI for science work is going to be about how we do things that humans can’t. Certainly, we’ll see lots more of those systems, maybe not a totally completely general model, but a lot of really important problems that this turns out to be the right tool to solve.
CM: Do you have any advice for aspiring scientists or UChicago students in general?
JJ: One really big advantage of being at a university like UChicago that not enough people take advantage of is that you have world experts in all sorts of things. A lot of people, especially Ph.D. students, will go to their group meetings and seminars for their department. That will be their Ph.D., and they will learn exactly the things that they are expected to learn. But they don’t go to the seminar for other departments, and they don’t go talk to the world’s leading computer scientists. They don’t talk to the world’s leading economists or anything else. They’re missing out on there. There’s a lot that comes from immersing yourself in different disciplines and doing different things. You learn a lot more by joining in with those experts. No one will kick you out. You just sit in the area, and you won’t understand anything the first week before you will. And they don’t even take attendance; you can just stroll in.
My piece of advice is: go find interesting experts and things you don’t know when just sitting in their seminars. Talk to them and then you can get a much broader and more interesting education than from pouring ever more time into exactly this narrow area of the same molecular biology that you’re expected to be learning. You should learn that too but take advantage of being at a university.
Images are for reference only.Images and contents gathered automatic from google or 3rd party sources.All rights on the images and contents are with their legal original owners.
Aggregated From –
Comments are closed.