Nobel Prize lecture Demis Hassabis Nobel Prize in Chemistry 2024

So our next speaker is Demis Hassabis. He was born in 1976 in London, United Kingdom, and he received his PhD in 2009 from University College London. And he is presently the CEO of Google DeepMind, also based in London, United Kingdom. So please, Demis, we're looking forward to your Nobel lecture, and please join me in welcoming Dr. Hassabis. Well, first of all, I want to start by thanking the Royal Swedish Academy of Sciences and the Nobel Committee for the honor of a lifetime. It's so wonderful to see so many family and close friends and colleagues here today. It's been an amazing few days. I'm going to talk about AlphaFold, of course, and the impact it's had so far, but I'm also going to touch on the critical work we did in the lead up to developing AlphaFold and where I think the future of AI for science is going more broadly. So my journey to AI actually started, surprisingly, with games, and specifically chess. I started playing chess when I was four and played very competitively, captaining many of the England junior chess teams growing up. That's a picture of me on the right playing for board one for the England under-11 team when I was nine. And when you play chess seriously at such a young age, it's a very formative experience. And for me, it really got me thinking about thinking itself. I was sort of fascinated, intrigued by how does our mind come up with these ideas, these plans, these strategies, and how can that be improved? And as part of the training camps that we went to, we got actually access to early chess computers like this one here, the Fidelity Chess Challenger, and some of you remember, these early chess computers, you actually had to press the squares on the board physically, the LEDs on the board, to move the pieces around. And of course, we were meant to use these machines to improve our chess skills, but I remember actually being fascinated by the fact that someone had actually managed to program this inanimate lump of plastic to play chess well against me. So I was intrigued about this, and I started reading some books about it and ended up coding my first AI program on my trusty Amiga 500 that you'll see in the middle there when I was 11, but to play the classic game of Othello. And this program, my first AI program, it managed to beat my kid brother. Admittedly, he was only five at the time, but I was still amazed that you could make something that could do something sort of independently of you. And that got me hooked on computers and AI for the rest of my life. So fast forwarding then more than 20 years later, and games are still central to my approach to AI and actually at the core of our early work at DeepMind. We started DeepMind in 2010 as an industrial research lab with a mission to build artificial general intelligence. And the vision behind it was to create a kind of modern day Bell Labs. Initially, we used games as the sort of perfect proving ground for AI because it's very easy to generate a lot of data on your computers. You can have the systems play against themselves, and it's easy to specify a clear objective or clear metric to hill climb against. So in most games, you can specify to maximize the score or to just win the game. So it's very clear if you're making progress with your algorithmic ideas. Now, of course, the pinnacle for games AI, which has had a long history, actually from the beginning of AI. So pioneers like Alan Turing and Claude Shannon, they all tried their hands at making and programming chess computers. But the pinnacle for games AI is being Go, the game of Go, the ancient game of Go. It's probably the most complex game we've ever devised. One illustration of how complex Go is is that there are 10 to the power 170 possible positions, which is more than there are atoms in the universe. So there's no way you could write a program that could brute force a solution to playing Go well. So our first big breakthrough really was in, major breakthrough was in 2016 with our AlphaGo system that managed to beat the world champion, the South Korean grandmaster Lee Sedol 4-1 in a famous match in Seoul. But it didn't just win the match. Importantly, it actually came up with new creative strategies and ideas that had never been seen before, even though we've played Go for thousands of years. And the most famous one of these new strategies was actually move 37 that I'm showing here in the bottom right. And in game two of this challenge match, and it's outlined as the black stone outlined in red. And this move had sort of never been seen before in Go, in top level Go play. And it turned out to be the perfect move to decide this game two match 100 moves later, as if AlphaGo had put it there presciently ahead of time. So how did we do this? Well, the AlphaGo system and its successors, AlphaZero, more general successors, mastered the game of Go and also all other two-player games through a process of what we call self-learning. So that's where these systems, instead of being programmed directly with solutions, they learn strategies and solutions for themselves. In this case, by playing many hundreds of thousands, in fact, millions of games against itself and incrementally learning from its mistakes and improving its strategies. And through this process of self-learning, we can build up a useful neural network model of Go. So AlphaGo and this 2016 result was a bit of a watershed moment, I think, in modern AI, as it was the first kind of big proof point that these sorts of systems, these kind of learning systems, were able to do something pretty impressive and thought to be intractable. So with this neural network systems, this neural network model, we can actually efficiently guide the search process so that we don't have to search all of the possibilities in a game like Go or in any kind of complex space. So in this cartoon diagram here, I'm showing a tree of possibilities in a Go game. And if you imagine each node in the tree is a particular Go position, and you have to make a decision about which move to take, and that takes you down this tree. So there's this massive set of intractable set of possibilities from any particular Go position. And what this neural network model allows you to do, AlphaGo to do, was to only analyze and look at and consider a tiny fraction of all the possible paths. So here kind of illustrated in the paths in blue. And it does this by basically figuring out what are the most likely moves, what are the most promising moves to look at in any particular position, and also which side, black or white, is likely to win from the current position, and the probability, the estimated probability of winning from that position. And then once the thinking time has run out, say a minute or two, then it outputs the best line it's found so far, the most optimal line it's found so far, out of the ones it's considered, out of those blue lines. And here in this case, it's the line colored in purple. So now that we've sort of cracked the pinnacle of games AI, we sort of knew that we had the techniques and methods to now transport that and tackle important real-world problems, including big challenges in science. So what then makes for a suitable problem for these kind of AI methods to tackle? Well, we look for kind of three criteria that make a problem suitable. Number one, finding a path through a massively combinatorial search space. Can the problem be couched in these terms? Secondly, can we specify a clear objective function or metric to optimize or hill climb against? So in games, this is pretty easy. It's winning the game or maximizing the score. And then thirdly, are there a lot of data available to learn the neural network model? And, ideally and, but it could be or, an accurate and efficient simulator to generate more synthetic data in distribution. So it turns out, actually, when you look at problems in this way, there are a lot of real-world problems that actually fit this profile or can be made to fit this profile, including many problems in science. And for me, the number one thing I always wanted to do with these kinds of systems was protein folding. It was always top of my list ever since I first came across it as this fascinating problem as an undergraduate at Cambridge. So as you heard earlier, proteins are the building blocks of life. Almost all biological processes in every living thing depend on proteins for their operation, from the twitching of your muscle fibers to the firing of your neurons. So in essence, you can think of proteins as these kind of exquisite nanoscale bio-machines. And they're, of course, extraordinarily beautiful as well. So proteins, as we heard earlier, are specified by their amino acid sequence, which you can see an example of one here on the left. And these sequences fold up into complex 3D structures. For example, this one here folds into this beautiful structure on the right. So you can think of it like a string of beads that kind of scrunch up into a ball. And knowing the 3D structure of a protein tells you a lot about its function, and is, of course, critical for things like understanding disease and accelerating drug discovery. So the question then is, can the 3D structure of a protein be directly predicted solely from its one-dimensional amino acid sequence? And in his Nobel lecture in 1972, Christian Enfinsen famously conjectured that this should theoretically be possible. And he advanced his thermodynamic hypothesis, the idea that a protein would uniquely take the shape that minimizes the free energy in the system. This then became known as the protein folding problem, and kicked off a kind of 50-year quest to find a computational solution to this grand challenge. So why is this problem so hard? Well, of course, the normal way that you would go about finding the structure of a protein is doing it experimentally. And this is very painstaking and difficult work. And it can often take years, months, and even years of work to just determine the structure of even a single protein. And Leventhal actually encapsulated this difficulty very well in what became known as Leventhal's paradox. And he estimated that for a typical-sized protein, there are maybe up to 10 to the power 300 possible confirmations that that protein can take, a truly astronomical number. And so, obviously, enumerating all of these possibilities would take longer than the age of the universe. And yet, somehow, in nature, these proteins might fold spontaneously, sometimes in a matter of milliseconds. So that gives us hope, really, that there must be a solution to this, because it must mean that there's some topology in the kind of energy landscape that guides the folding process very efficiently in nature. And perhaps we can recover what that process is. Another key aspect of why we took on this challenge was that there was readily available data to learn from, and also, very importantly for AI development, very clear and great benchmarks to try and measure your progress against. So after decades of experimental work, 170,000, roughly, structures had been determined experimentally through lots and lots of painstaking work for tens of thousands of experimentalists and collated into the protein databank, the PDB, which is an incredible and valuable resource that we used as the starting point to train AlphaFold. And then secondly, there was the CASP competition, which was considered to be the gold standard benchmark for structure prediction, and was run every two years, ever since 1994 by Professor John Malt and his colleagues, and to test the best computational systems. And the beauty of this competition was that it's a blind assessment competition. So the proteins that you're trying to predict computationally have just been, the structures have just been discovered experimentally, but have not yet been published. So no one external from those groups knows what those structures are. So it's a true test of whether these computational systems are capable or not. We were always told that the critical threshold for a computational system like this to be of use, to practical use, to experimentalists, was that you had to get the accuracy down to within the width of an atom. So on average, less than one Angstrom error, which is incredibly a high degree of accuracy. So with AlphaFold 2, we managed to achieve this atomic accuracy. We started off with AlphaFold 1 back in 2016, and entered it for the first time in CASP 13, the 13th edition of CASP in 2018. And we really advanced the field here. And if you look at this bar chart of progress, this is basically the winning score for the top team, each edition of CASP as a kind of distance measure of how accurate the prediction is compared to the ground truth. And you can see for about a decade prior, there hadn't been much progress in increasing the accuracy of these predictions. So with AlphaFold 1, it topped the leaderboard of the CASP 13 competition and improved the predictions by a large amount. And it was really the first time that a machine learning system was introduced that had machine learning as the core component of the system. But we still hadn't reached atomic accuracy, which is sort of represented by the 90 GDT score line here on the chart. And we were able to do that by taking the learnings from AlphaFold 1, then re-architecting a brand new system with AlphaFold 2, and entering that into CASP 14. And that reached the atomic accuracy needed and led the organizers of the competition to declare the problem essentially solved. So this is the diagram of the innovative architecture of AlphaFold 2. I'm gonna leave John to cover some of the technical details of this in his talk. But the high level thing that takeaway is that actually there was no silver bullet to solving this problem. AlphaFold 2 actually had to incorporate over a dozen different innovations into quite a complex, what we call a hybrid system. And one of the key things of that was to build in evolutionary and physical constraints into the architecture, along with the learning components of the system, and sort of combining those two elements together. And it was key that the fact of this working is that the AlphaFold team was a multidisciplinary team composed of expert biologists and chemists, as well as machine learning and engineering. And it was a full end-to-end system that started with the amino acid sequence and directly output the 3D structure prediction. And it used this for the recycling stage that allowed it to iteratively refine the prediction over many iterative steps. So you can see that happening here with this complex protein that was actually in the CASP 14 competition. You see on the left-hand side, the ground truth in green and the prediction in blue very much overlapping. And then on the right-hand side, you can see how AlphaFold 2 sort of iterates over 192 steps in this case to the final predicted structure, each time improving its accuracy. So then with AlphaFold, once we had the system, of course, we wanted to make the maximum impact possible with it. And AlphaFold is not only very accurate, but it's also very fast. And we sort of quickly realized that it was fast enough to actually practically be able to fold all proteins known to science. So roughly over 200 million proteins that we know the sequences for, and we would love to know the structures for. And so over the course of the next year, on our computers, we basically folded all the 200 million proteins. We open-sourced AlphaFold, and we built the AlphaFold database with our fantastic colleagues at EMBL-EBI, the European Bioinformatics Institute. And we provided them free and unrestricted access for everyone to use to all these predictive structures. Of course, we were cognizant about safety and ethics. So we consulted with over 30 biosecurity and bioethics experts before we put the database out to ensure that the benefits far outweighed any risks. And the impact so far has been beyond what we could have imagined. With over 2 million researchers from around the world make use of AlphaFold in its predictions, over 30,000 citations, and it's very much become a kind of standard tool in the biology toolkit. Just a small sampling of some of the progress that AlphaFold has helped to accelerate across actually a huge range of problems. And some of my favorites, tackling plastic pollution with designing new enzymes, helping with neglected diseases for underfunded diseases that affect a lot of the poorer parts of the world, helping with fundamental structural biology, things like determining the structure of the nuclear pore complex. And even just recently, just a couple of weeks ago, discovering new mechanisms in reproduction. And I think very much we're just at the beginning of the impact that programs like AlphaFold will make. Of course, we've been continuing on with developing AlphaFold and just earlier this year, we released the newest version of AlphaFold, AlphaFold 3. And this is also a big advance because with AlphaFold 2, you can think of it as essentially solving the static picture of a protein and what the structure of it looks like. But of course, we know that biology is incredibly dynamic and actually all the interesting things happen in biology when different aspects of biology interact with each other. And AlphaFold 3 is our first step towards modeling the interactions and the dynamics. And it's actually able to model pairwise interactions between proteins and other proteins, but also proteins and RNA, proteins and DNA, and proteins and ligands. So this is a big step towards what one would need to use AlphaFold for things like drug discovery. So I just wanna end with the last section on zooming out a bit and talking a little bit about the implications of this kind of work and some of our colleagues on the bigger picture of AI for scientific endeavor. So we take a step back and look at the essence of what our system's been doing, both AlphaGo, AlphaFold, and some of the other systems that we've built. And really, we can describe them as finding the optimal solution in this enormous combinatorial search space. And we do that by learning a model of that environment, either from data or from simulation, and then using that model to guide a search process according to some kind of objective function that you're trying to optimize. And it turns out that this is a very general solution and that many problems can fit this approach. So I showed you earlier this tree diagram of finding the best Go move in a Go game, but one could easily switch out those nodes for game Go positions for actually chemical compound designs. And now think about the tree of search that you're doing is finding the best molecule or the best drug compound in through chemical space. And I think you could very much use the same kind of techniques that I've outlined here for this kind of problem. So that means if we may be entering perhaps this new era of what I like to call digital biology, I've always thought of biology at its most fundamental level can be thought of as a information processing system, albeit a phenomenally complex and emergent one. And just, I think it's such a complex system, it's gonna be hard to boil down the workings of biology to a few mathematical equations. So I think maths has been an incredible tool to describe sort of description language for physical phenomena for physics. And I think in the same way, AI may be potentially the perfect description language for biology. And we hope that AlphaFold is a kind of proof point that could, when we look back maybe in 10 or 20 years time, maybe help usher in this new era of digital biology. And we're trying to contribute to that where we recently spun out a couple of years ago, Isomorphic Labs, a new company to build on AlphaFold and to reimagine the drug discovery process from first principles with AI. And perhaps we can reduce down drug discovery, which is incredibly difficult and time-consuming and expensive process down from years to perhaps months or maybe even one day, weeks. And we sometimes think of this kind of accelerated process that I've just shown you with AlphaFold and the protein structures and perhaps one day will happen with drug discovery as sort of doing science at digital speed. And I've had a longstanding dream actually to think maybe one day be able to simulate a whole virtual cell, not just a protein or a couple of proteins interacting with each other, but an entire cell. And then those predictions would be useful for experimentalists. So perhaps I'll end then, given this is a Nobel lecture, I thought I would end with a couple of new, perhaps slightly provocative ideas in the spirit of Christian Anfinsen and his 1972 lecture. So ever since doing AlphaGo actually, I've been thinking a lot about what are the limits of classical systems, classical computing systems. And I think there's a big debate going on at the moment in computing circles about quantum computers versus classical systems. And I think classical Turing machines, basically the underpinnings of modern computers today, I think can do a lot more than we probably previously thought. And how can they do that? Well, they do that by perhaps doing this massive amount of pre-compute ahead of time and use that to develop a good model, a good model of the environment, good model of the problem that you're trying to solve. And then you can use this model to efficiently explore solutions based in polynomial time, what's called polynomial time in complexity theory, so an efficient amount of time. So a sort of loosely proposed conjecture that I'm thinking about is that maybe any pattern or structure that can be generated or found in nature can be efficiently discovered and modeled by classical learning algorithm. That doesn't mean everything, all quantum systems, because there'll be lots of things that don't occur in nature that have no pattern or no underlying structure to learn. So for example, factorizing large numbers or abstract problems like that. But I think systems in nature like proteins and perhaps materials will potentially have structure that can be learned by these kinds of processes that I've outlined today. And if it turns out that classical systems then therefore can model some types of quantum systems, I think that could have some quite big implications for areas like complexity theory, including P equals NP, and maybe even some aspects of fundamental physics like information theory. So with that, I just wanted to talk about the final piece, which is that I've talked a lot about AI here for biology and life sciences, but actually we're doing a lot of work at DeepMind across the sciences, medicine, climate, mathematics. And here's just a little selection of some of the work we've done in the last few years, applying some of the AI techniques I've talked about today to medical areas like image analysis and diagnosis of medical images, genetic sequences, and spotting missense variations, whether they're benign or pathogenic, controlling the fusion, the plasma in a fusion reactor, coming up with faster algorithms like matrix multiplication, state-of-the-art weather prediction systems, and the discovery of thousands of new materials that we've never seen before, all using AI as a fundamental tool underlying the sciences. So I think with this, we may be entering a new golden era of discovery helped by these new types of powerful AI tools. Now, of course, I've worked my whole life on AI because I think it has the incredible potential to help with humanity's greatest challenges, but also AI is a dual-purpose technology and it must be built responsibly and safely and be used for the benefit of everyone. And in order to do that, I think it's critical for scientists and technologists at the lead of this field to engage with a wide range of stakeholders from government to academia and civil society to incorporate all the views of how these systems should be deployed and best used. And I think a technology as transformative as AGI will be, I think this will be akin to something like the invention of fire or electricity requires exceptional care and foresight to steer us through the next stages of the development of this incredibly powerful technology. But if we're able to steward this technology safely through, then I think AGI could end up being the ultimate general purpose tool to help us understand the universe around us and our place in it. So I just want to end, of course, by thanking the amazing and incredible AlphaFold team, almost all of whom are here today, and the AlphaGo and AlphaZero teams, our wider colleagues at DeepMind and Google for all of their incredible support and infrastructure that supports all of this work that you've seen today. Our wonderful collaborators at EMBL-EBI, the CASP community, and the PDB and experimental biology community. And of course, lastly, and most importantly, I want to thank my wonderful family, closest friends and colleagues, all of whom are here today, for their love, support, and encouragement without which none of this would have been possible. Thank you.