Accelerating scientific discovery with AI

so our next speaker is uh Demis hassabis he was born 1976 in London United Kingdom and he received his PhD in 2009 from University College London and he is presently the CEO of Google Deep Mind also based in London United Kingdom so please Demis we're looking forward to your Nobel lecture and please join me in wel coming Dr hassis well first of all I want to start by thanking the Royal Swedish Academy of Sciences and the no Nobel committee for the honor of a lifetime uh it's so wonderful to see so many um family and

close friends and colleagues here today it's been an amazing days I'm going to talk about Alpha fold 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 leadup to developing Alpha fold and where I think the future of AI for science is going more broadly so my journey uh 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 gr up that's a picture of

me on the right um playing for board one for the England under 11 team uh 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 um and how can that be improved and as part of the training camps uh that we went to uh we got actually access to early chess computers uh like this one here

the Fidelity chess Challenger and uh some of you remember these early chess computers you actually had to press the squares on the board physically uh 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 um but I remember actually being fascinated by the fact that someone had actually managed to program this an inan lump of plastic to play chess well against me so I was intrigued about this and I started reading um some books about it and ended up coding my

first AI program uh on my trusty igga 500 that you'll see in the middle there when I was 11 but to play um the classic game of aell and this program my first AI program it managed to beat my kid brother um admittedly he was only five at the time but uh I was still amazed that I you know someone could you could make something that could do something sort of independently of you and uh that got me hooked um on computers and AI for the rest of my life so fast forwarding then 20 more

than 20 years later and games are still Central to my approach to Ai and actually the at the core of our early work at Deep might we started Deep Mind in 2010 as an industrial research lab with a mission to build artificial general int intelligence and the vision behind it was to create a kind of modern day Bell Labs initially we use games as the sort of perfect Proving Ground for AI because it's very easy to generate a lot of data on your computers um you can have the systems play against themselves um and it's

easy to specify a clear objective or clear metric to Hill Climb against um so in most games you can specify to maximize the score or to uh 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 um which has had a long history actually from the beginning of AI um so Pioneers like Alan churing and Claud Shannon um they all tried their hands at making and and programming uh chess computers um but the Pinnacle for games AI is being Go the

game of Go the ancient game of Go um it's probably the most complex game um 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 play and go well so our first big breakthrough really was in um major breakthrough was in 2016 with our alphao system um that managed to beat the world champion uh the South Korean Grandmaster Lisa do

4-1 in a famous match in Soul but it didn't just win the match importantly it actually came up with new creative strategies and ideas um that had never been seen before even though um we've played go for thousands of years and um the most famous one of these new strategies was actually move 37 that I'm showing here uh in the bottom right um and in game two of this uh this challenge match uh and it's outlined as the black stone outlined in red and uh this move uh had sort of never been seen before uh

in go in top level go play uh and it turned out to be the perfect move to decide this game two match 100 moves later as if ala for go had put it there presly uh ahead of time so how did we do this well the alphao system um uh and its successors Alpha zero more General successors mastered the game of Go and also all other two two-player games through a process of what we call self-learning so that's where these systems instead of being programmed with 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 and through this process of self-learning we can build up a useful neural network model of go so alphao 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 uh intractable so with this new network

systems this new network model uh we can actually efficiently guide uh the search process so that we don't have to search all of the um uh possibilities um 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 new network model allows you to do alpha go to do was to only um uh uh analyze and look at and consider a a fraction a tiny fraction of all the possible um uh all the possible uh uh paths so here kind of Illustrated in the paths in blue and uh it does this by um uh basically figuring out what are the most likely moves what are the most promising moves to look at in any particular position and also uh 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 um the best line it's found so far the most optimal line it's line so it's found so far out of the ones it's considered out of those blue lines and here in this case is the line uh uh colored in purple so now that we've sort of you know cracked the the Pinnacle of games AI we were sort of knew that we had the

techniques and methods to now transport that and Tackle important real world problems uh including big challenges in science so what then makes for a suitable problem uh for these kind of AI methods to tackle well we look for kind of three criteria uh that make a problem suitable number one um finding a path through a massively combinatorial search base 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 um or

maximizing the score and then thirdly uh are there a lot of data available to learn uh the neural network model and ideally and but it could be all an accurate and efficient simulator to generate more synthetic data in distribution so it turns out actually when you look at at the at the at problems in this way there are a lot of um real world problems that actually fit this profile or can be made to fit this profile um including many problems in science and um for me uh 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 a 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 extraordin 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 uh uh 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 um 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 Enon 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 uh well of course the normal way that you

would go finding the structure of protein is doing it experimentally and this is very painstaking um 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 levental actually encapsulated this difficulty very well in what became known as lentiles Paradox and um he he estimated that for a typical size protein there are may be up to 10 to the power 300 possible confirmations that that protein can take a truly astronomical number so obviously enumerating all of these possibilities would take longer

than the age of the universe and yet somehow in nature these proteins um 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 um there was a readily available data to learn from and also very importantly for uh um AI development a

very clear and great benchmarks to uh to try and uh measure your progress against so after Decades of experimental work um 170,000 roughly structures had been determined experimentally um through lots and lots of painstaking work for T of thousands of experimentalists and collated into the protein Data Bank the pdb which is incredible and valuable resource that we used as the starting point to train alphafold and then secondly there was the Casp competition um which was considered to be the gold standard Benchmark for uh structure prediction and uh was run every two years ever since since

1994 by professor John malt and his colleagues and to test the best computational systems uh and the beauty of this competition was that it's a blind compet assessment competition so uh the the proteins that you're trying to predict computationally um have just been the structures have just been discovered experimentally but have not yet been published so no one external from those groups uh knows what those structures are so it's a true test of uh whether these computational systems are are capable or not and we're always told that the critical threshold for a computational system like

this to of use to um 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 eror error which is incredibly uh High degree of accuracy so with Alpha fold 2 uh we managed to achieve this Atomic accuracy um we started off uh uh with uh with Alpha fold 1 um back in 2016 entered it for the first time in Casp 13 uh the 13th edition of Casp in 2018 and we really Advanced the field here and if you look

at the this bar chart of progress um this is uh basically the the the winning score uh for the top team each each edition of Casp um on uh as a kind of distance measure of of how accurate um uh the the the prediction is compared to the ground truth and you can see for about a decade uh prior there hadn't been much progress um in increasing uh the accuracy of these predictions so with Alpha fold one um uh it it topped the leaderboard of the C 13 competition uh and improved the predictions by a

a large amount and it was really the first time that um 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 on the chart um and we were able to do that by taking the learnings from alpha fold one then re architecting a brand new system with Alpha fold 2 um and entering that into C 14 and that reached the atomic accuracy needed and led the organizers of the competition

to declare the problem essentially solved so this is uh uh the diagram of the Innovative architecture of Alpha 2 um I'm going to leave John to cover some of the the the 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 Alpha fold 2 actually had to incorporate um over a dozen different Innovations uh into quite a complex what we call a hybrid system and one of the key things of that was to build in uh uh uh evolutionary

and physical constraints uh 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 alpha fold team was a multi-disciplinary team composed of expert biologists and chemists as well as machine learning and engineering and it was a full end to-end system um that uh started with the amino acid sequence and directly output uh the 3D structure prediction uh and it used this s of recycling stage that allowed it to iteratively refine the prediction uh over

many iterative steps so you can see that happening here with uh this complex prot that was actually in the C 14 competition you see on the left hand side uh the the ground truth in green and the prediction in Blue uh very much overlapping and then on the right hand side you can see um how alpha f 2 sort of iterates over 192 steps in this case uh to the final uh predicted structure each time improving uh uh uh its accuracy so then with Alpha 4 once we had the system of course we wanted to

make the max impact possible with it um and Alpha fold 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 uh over 200 million proteins that have um we know the sequences for um and we would love to know the structures for and so over the course of the next year um we we on our on our on our computers we basically folded um all the 200 100 million proteins we open

sourced Alpha fold uh and we built the alpha fold database with our fantastic colleagues at emble ebi the European bioinformatics Institute and we provided them free and unrestricted access for uh everyone to use to all these predicted structures of course we were cognizant about safety and ethics so we consulted with over 30 biocurity and bio ethics experts before we put the data base out to ensure that the benefits far outweighed any risks and the impact so far has been beyond uh what we could have imagined uh with over 2 million researchers from around the world

have made use of alpha fold in his predictions over 30,000 citations and it's very much become a kind of standard tool in the biology uh uh tool kit just a small sampling of some of the progress that Alpha folders helped to accelerate across actually a huge range uh of problems um and some of my favorites you know tackling plastic pollution with designing new enzymes um helping with neglected diseases for um underfunded diseases that affect a lot of the poorer parts of the world um helping with fundamental structural biology things like um determining the structure of

the nuclear Poe complex and even in just recently just a couple of weeks ago discovering new mechanisms uh in in reproduction and I think very much we're just at the beginning of uh uh the impact that uh programs like Alpha fold will make of course um we've been continuing on with um um with uh developing Alpha fold and just earlier this year we released the newest version of alpha fold Alpha fold 3 and uh this is also a big Advance because with Alpha fold 2 you can think of it as essentially uh uh solving the

static picture of a protein uh 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 Alpha 43 is our first step towards model in the interactions in the Dynamics uh and it's actually able to model uh pairwise interactions between proteins and other proteins but also proteins and RNA proteins and DNA and proteins and ligans so this is a big step towards what one would need to use Alpha fold for

things like drug Discovery so I just want to end with the last section on um zooming out a bit and uh talking a little bit about the implications of this kind of work uh and and uh some of our colleagues on the bigger picture of AI uh for scientific Endeavor so if we take a step back and um look at the essence of what are our systems been doing both alphao Alpha fold and some of the other systems uh that we've built and really we can describe them as um finding the optimal solution in this

enormous combinatorial search base 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 uh in a go game but one could easily switch out those nodes for of 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 use very much use the same kind of techniques um that I've outlined here for this kind of problem so that means you know 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 orbe it a phenomenally complex an emergent one and just um

I I think it's such a complex system um it's going to be hard to boil down the workings of biology to a few mathematical equations so um I think you know 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 Alpha fold is a kind of proof point that could when we look back maybe in 10 or 20 years time maybe helped Usher in this new era of digital

biology and we're trying to contribute to that uh where we recently spun out a couple of years ago isomorphic Labs a new company to build on Alpha fold and to reimagine the drug Discovery process from first principles with AI and perhaps we can reduce down drug discovery which is incredibly difficult on timec consuming expensive process down from years to perhaps uh months or maybe even one day weeks and we sometimes think of this this kind of accelerated process um that I've just shown you with Alpha fold 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 long-standing dream actually to think uh maybe one day be able to simulate a whole virtual cell uh not just a protein or a couple of proteins interacting with each other uh but an entire cell um and then those predictions would be useful for experimenting this 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 uh in the spirit of Christian an Fen and and his 1972

lecture so ever since doing alphao actually um I've been thinking a lot about what are the limits of classical systems uh classical Computing systems um and you know I think there's a there's a big debate going on at the moment in Computing uh circles about quantum computers versus classical systems and I think classical touring machines basically uh the the the 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 precompute 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 solution Space in polom Time what's called polinomial 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 a classical learning algorithm that doesn't mean everything um 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 uh or abstract problems like that but I think systems in nature like proteins and perhaps um materials uh will potentially have structure that can be learned uh 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 quanton systems I think that could have some quite big implications for areas like complexity Theory including p m p and maybe

even some aspects of fundamental physics uh 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 uh at Deep Mind Across The Sciences medicine climate mathematics and here's just um a little selection of some of the work we've do we've done in the last few years um applying the some of the I techniques I've talked about today to uh medical areas like uh image analysis and diagnosis of medical

images uh genetic sequences and spotting missense variations whether they're benign or or pathogenic controlling um the fusion uh the the plasma in a fusion reactor um coming up with faster algorithms like matrix multiplication um state of the art weather prediction systems and the discovery of thousands of new materials that um uh we've never seen before all using AI as a fundamental tool underlying uh The Sciences so I think with this you know we may be entering a new golden era of Discovery help uh 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 in the in at the uh at the lead of this field to engage with a wide range of stakeholders from government to Academia and Civil Society uh 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 um you know 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 uh technology but if we're able to Steward uh uh this technology safely through then I think um 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 Alpha Team almost all of whom are here today um and the alphao and Alpha zero teams our wider colleagues at Deep mine and Google for all of their uh incredible support and infrastructure that supports all of this work that you've seen today our wonderful collaborators at emble ebi the cast Community uh 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 um uh for their love support and encouragement without who

without which none of this would have been possible thank you