Charlotte Deane: AI, Drug Discovery, and the UK Advantage
Professor Charlotte Deane of the University of Oxford, one of the UK's leading computational biologists and a co-lead of the OpenBind consortium, on what it will take for AI to improve drug discovery. Recorded at the ISMB/ECCB conference in Liverpool, the conversation covers why AI needs purpose-built data, how the UK can play to its strengths in AI, and how AI will change the way science is done. Laila Rizvi and Stephen Horn are joined by scientist Luke Yates.
Transcript
AI is changing biology, but can it crack one of science's hardest problems: drug discovery? At the center of that challenge is Professor Charlotte Deane. As one of the leads of research consortium OpenBind, Charlotte is shaping how massive biological data sets are built, shared and trusted.
Welcome to Agents of Tech. Here in Liverpool, where over 2000 of the world's top computational biologists, AI researchers and data scientists have come together to decode the future of life science at the ISMB/ECCB conference. This is a flagship meeting for bioinformatics and systems biology. I'm Laila Rizvi and I'm here with my co-host Stephen Horn and Luke Yates, a friend of the podcast.
Hello.
So drug discovery, big question, big challenge. What do we think? And in the UK, UK AI.
I'm intrigued about what Charlotte's views are on drug discovery and why it's so difficult for AI to tackle it. I really want to get a sense of, of why it is so challenging. And, what they're trying to do to address that, particularly with this OpenBind initiative, this consortium, which sounds, you know, quite interesting because it implies that they're collecting and curating large data sets. And I think they'd probably be, you know, this is probably structural biology based to build a large repository that you can then use for training purposes. But I think it'd be really interesting to hear what she has to say about this.
So I'm interested in two things. I'm interested in data and data sets. So I don't think we talk about this enough with AI. We talk about the technology, but we don't talk about data and the and the importance of having large data sets that are well managed, that we can interrogate and train the different models on. So I'm intrigued about the UK and AI, right. Today we are sitting in Liverpool, if you go that way 4 or 5 thousand miles you hit the US. But we don't talk about the UK and AI very often, do we?
I completely agree and I think also, Stephen, your point on data, I actually think that's the primary role that humans will have as AI get better, that we will be far more fundamental in the data collection process and how we curate that data and how we build that data, because that is essentially the oil of AI, right? The UK, yeah. I don't think we talk enough about the fact that we're producing this unbelievable research, that DeepMind was a UK company, but now it's American. And what's happening there? So I'm very much looking forward to asking Charlotte about the current state of AI in the UK, because I feel like there's a lot of things being said, but what I'm seeing is very different because I don't hear the UK in the AI conversation or in the AI race. I mean, all I'm hearing is US-China. Luke, you're a prominent UK scientist.
That's quite generous.
So what do you think when you hear people talking about the US all the time? And what's your feeling on this?
Yeah, I think their funding models are vastly different to the UK. They can take larger risks. I think philanthropy is a very different entity in the US. I think philanthropic donations, large sums to private universities, you know, to drive innovation is something we really don't have here. And I guess my question probably to Charlotte is why now? Why is the timing right now for this OpenBind consortium and trying to use AI to tackle drug discovery? Given the fact that the UK has, has obviously contributed but isn't necessarily the front runner when it comes to thinking about, you know, global standing in sort of AI development.
This is a question more so for you, Stephen, as an entrepreneur. Do you think we have a cultural problem when it comes to an aversion to risk? And do you think there was a time where we weren't as risk averse, you know, when the UK was the manufacturing center of the world 100 years ago?
So, so, so that's a, that's a crazily good question. So, so sitting here 200 years ago, sitting exactly where we are now, 200 years ago, you would have been at the heart of the Industrial Revolution, where, where the whole world was changing and pivoting around the very spot that we are now in this place. But let us look at another place in the world at the moment, Singapore. Singapore, when Singapore was created they understood that they weren't big on natural resources, that their natural resources were their people. Now, Singapore is one of the most innovative countries in the world at the moment, and they focus on innovation. And as you say, they have a culture of innovation, in America they have a culture of innovation, in this country we have marvelous scientists like Luke but we do not have a culture of innovation.
When do you think was the point at which it, you know, started going like that? But then what was it? What was the shift? And that's what I constantly kind of rack my brains where, how do we have this incredible scientific research but we just stopped taking risk? As in, what caused it?
What do you aspire to be? I mean, what do you aspire to be? I mean, what does Luke aspire to be?
I mean, you know, a professor, with some type of medal to my name would be nice.
Professor, but you don't aspire to set up a startup company.
No. Charlotte's CV is interesting, isn't it? Because she's obviously, you know, done so well at Oxford, but she spent two years at UCLA. So be very interesting to see what her perspective on that cultural difference is.
And you can ask her about the difference in innovation.
We're now delighted to be joined by Professor Charlotte Deane. Thank you so much for being here, Charlotte. So I want to kick off by asking you a bit about your work and particularly the OpenBind Initiative. Could you tell us a little bit about what that's trying to achieve, what you're doing?
I work in the area of trying to use AI to improve drug discovery, is always my shorthand for trying to describe my research work. And OpenBind really springs from the idea that, just one of my favorite phrases, AI isn't magic. It doesn't just do stuff from nothing. It is a computational mathematical technique. And in order to make it good, you have to have data.
And the concept of OpenBind is to, in the same way we had the kind of revolution with AlphaFold because we had a database full of protein structures, we had things that we could use. They were there, you could build them. And, you know, amazing science could be done. If we want to do the same thing, to understand how small molecules bind to proteins in our body, and that's effectively what drugs are, we need that same kind of data. And so OpenBind is about collecting that data.
But it's more than just collecting it by accident. So the PDB has been around a really long time, I think 50 years. And it's kind of almost an accident that we have the data set we need. The concept of OpenBind is collect that data not by accident. Collect it with the people who do AI and machine learning involved so you can collect the right data, collect it rapidly, collect it to power the programs that will actually change the way you can do predictions of how molecules bind. That's a sort of, I hope, a quick summary of what it is.
Yeah, I guess moving on from that, I guess this is trying to rival maybe what exists in sort of pharmaceutical companies where they have latent databases of their own structures with their own drug molecules that aren't available to the public. Is this trying to rival that in some way?
I think it's different from that as well.
Okay.
You're always going to have pharmaceutical companies going to work on the things that interest them. And obviously they are, if you like, honor bound, but also legally bound to make money so they can hold on to that data because it's part of what they need to do, what they need to do, but they always concentrate on, if you like, one molecule.
Yeah.
To take it to the end point. So you have lots and lots of data around a certain point.
Right.
And that's really powerful in one way. But if you want to understand binding in a much more general sense, I don't actually want what's in the pharma. I mean, I'd love them to give it away, just to be honest, but what we really need is a much more general data set. You want coverage of the space, you want to try and identify what you don't know, where the models are failing and generate data there, which right now might not be useful for a particular molecule you're trying to take to market, but in the end makes a much better model. So it's about the diversity of the data and the types of data and what you collect.
Yeah, a more generalized sense of what makes something bind to a particular pocket on a particular protein, and what are the rules for that.
So we can predict them. Yeah. So thinking much, much more generally about the problem, and I guess this is one of the reasons why OpenBind can be set up like it is as a kind of, you know, it's a public consortium. We can try to make all the data open, but have companies involved in sponsoring and as part of that, because you can see it as a pre-competitive step.
Yes, yes.
Generate this data. It helps all of them. But for each of them they're working on a specific molecule where they will probably generate data that, though I would love to see, I will never see. That will help them be really good at the drugs they're most interested in, that they're trying to take to market in particular ways. But it's interesting if you pull out a little bit.
It's an interesting conundrum all round, isn't it, because you say, quite rightly, that AI is not a magic trick, not a magic box, and we need data and we need large sets of data. But then you were saying, the people who make money out of this, it's not necessarily in their interest to generate the data sets. How do the data sets get generated more generally if we need bigger data sets, but there isn't a financial return for the data set?
So I think this is that whole, there's always been this conundrum when you're working on problems that are close to the edge of what can be used commercially, but it's also what public science is for, you know. I'll use AlphaFold as an example again, because it's easy and because I've been thinking about it a lot recently. No one company was ever going to create the data bank of protein structures. Right? But governments around the world have funded researchers to get protein structures, and they've done that because we all want better medicines. We all want a better understanding of our biology. We all think that that's really important.
And also, if you put it to the other side, we all want companies that are successful, have economic growth, help make money for the country as well as deliver brilliant medicines and do everything else. So it's kind of part of the bargain. And actually, to be fair to companies, they are putting quite a lot of money into these very competitive data sets, but no one company is going to do it on their own. So it's that mixture, I think.
Right. But if you flip that on its head, you talk about governments. So what confidence have we got that governments around the world, as you put, are in this together?
You're always going to have differences between what any individual government is trying to do. But science has become and really is a worldwide endeavor. You talk to scientists. We collaborate with people all over the planet. We are all funded in different ways by different mechanisms. But what we're doing is sharing, just like being here with each other, what we're doing, how we're doing it, what makes us go forwards. So I never quite think, you know, it's dependent on a particular government of the day. It's dependent on actually, if we're honest, the human race really likes learning new stuff.
I mean, it does. It doesn't matter what that new stuff is. For me, it happens to be, I love thinking about computational models. I find protein structures fascinating. I'm very happy there. But I also like learning new stuff about, you know, how a computer works. But I like reading new books, and people like learning new stuff. And so there is this consensus of movement that keeps us doing this. And at different points in time, I suspect different parts of the world and different scientists will be the places where it's happening.
I want to go more on to kind of innovation in the UK now, we're obviously in Liverpool, and particularly in AI, and the UK is home to, you know, some of the brightest minds in the world, particularly in AI research, in computational biology. I mean, it was where DeepMind started, but it's not where DeepMind is now necessarily. I mean, it's now part of Google. So where do you think UK innovation in AI is heading, and what work do we have to do to be able to compete with countries like the US and China?
So the first thing to start with there is it's actually an amazing fact. If you think, we're not very good at thinking like this, is the UK, yeah, we're not actually that big.
Yeah, yeah, right.
We think we're massive. But look at us on the map, we're quite small.
Yeah, yeah.
We don't actually have that many people. We've got a lot, but we're not that huge. We are the third best country in the world for AI. Take your measures, take your metrics, think about whether that's in terms of your spin outs, whether in terms of your, you know, the people we are training. That's incredible. Right. Start there.
And then also think whether the word is compete. Right?
Right.
The important thing to think about here is, I don't really like this term, but it's the easiest way to think about it, is strategic advantage. What is the UK uniquely placed to be good at? Where do we stand out? And you could argue one of those areas would be something I would pick would be around computational biology. We have an incredibly strong kind of pharma sector, biotech sector, as well as amazing biotech research going on and sort of biology research, as well as the computational research across our university landscape as well.
So you think about where we are good and what we are good at, and then you think how that adds to like the world landscape. What is the places? Because if we're not going to win everything, and the scary thing is, nobody is. Yep. And so that's not the way to think about it. And then the other part of this is to make the right partnerships, to work with the correct, the other countries which are excelling at the parts that maybe we're not excelling at, or the countries that also excel at what we excel at, to go as fast as possible. So there's massive opportunities here.
So spinning that the other way round, then what do you think we need to do to keep that strategic advantage?
So I mean I have lots of opinions on this. I have to be careful as I have multiple hats.
We were down the pub last night talking about this, would you believe.
Yes, I think about this a lot. One of the biggest things we need to do is keep training people. One of the biggest lacks in all of this is enough trained people who really understand what they're doing. And that's not just let's train loads of AI scientists. That's one thing, train people to actually understand the fundamental techniques of AI. But it's more than that. It's training all the people who do everything else to be able to talk AI. Right?
I mean, one of the ways I think about it is, could you imagine a scientist now who did any scientific discipline just going, well, I really can't add up, I find numbers really confusing. I mean, they might, but everyone would look at them and go, you're going to have some issues.
Yeah, yeah.
And one of the ways to think about this is we are moving into a realm where people really need to understand AI. They don't need to be able to write it themselves, but to understand what it is, what it can do, how it works. And that is what allows you in a kind of AI for science world to make a really big difference.
So, I mean, based on that point, if you were just kind of a general scientist today or a young scientist, where would be the starting point to understanding and using AI in the same way you know your mental arithmetic, your maths, where, where would you start?
So you start at the same place you start with everything for learning. Go and read about it. Always just go and read stuff. We live in an amazing age where you can find lots of stuff to read. The second thing is go and immerse yourself in it somewhere. I was talking to someone actually over breakfast, and we were talking about one of the things that's really important to do in science is to go and sit somewhere that's a bit uncomfortable for you, go and sit through a talk you don't understand. Go and ask questions about that.
You can talk to people about it because for every discipline, there'll be different people who can help you to learn this. But there are people. So go and listen to them speak. Find out what they're saying about it, find out what they're referencing, and that's the place to begin. Because actually, you know, it's just about learning. You can learn it. It's not impossible. It's there. And then going alongside that, actually make yourself a bit uncomfortable. I do this all the time.
And then the second one is don't be afraid. Yeah, I think in some ways we've turned AI into a scary thing. We've kind of made into, oh, this is really complicated, so you need to be really good at, and it's actually, no, it's not that scary. It is just a computational mathematical technique. It's amazing. You know, it does incredible things. But the basics of understanding whether it's working or not are the same basics that you should have learned as to whether any experiment is working or not. How do you test it, how do you train it? What's happening when you do this? Do you believe those results? Do they look weird? Yep. And all of those are the same. So for me, a lot of it is just trust yourself, don't be afraid and go make yourself a bit uncomfortable to find out more.
So when AI sort of is brought into the sort of everyday lives of scientists, I guess you are, you're really kind of pointing out the fact that we'll still need to be present to intuit, because maybe AI won't be able to do that for us, to pick out, you know, to pull the threads of a particular problem or to know or feel our way around a problem and know that something doesn't smell right or look right to us.
I mean, I always have this phrase about what's happening with AI that people are not quite seeing it enough at the moment. AI is going to completely change the way we do science. Yeah, and I don't think people have quite grasped that, they're still sort of seeing it as a bolt-on.
Yes.
It's going to not just change the speed we can do it at, it's going to change the way we do it. And I think most people have got that. They're going, well, we'll do it on a computer and then we'll check it with fewer experiments, that kind of thing. It's going to change the questions you can ask. So it requires you to think about that.
I mean, some people don't like this, but I always just think of it as another toy. Somebody handed me an amazing toy. And if you hand a child a toy, generally speaking, I've noticed the first thing they do with it is not what it said on the box to do with that toy.
No. That's right. Yeah.
And that's how I want people to see this, it's going to completely change the way we ask questions. So it's not, you know, when people go, does it mean we don't need humans do science anymore? No. Just means we'll do different things. And that's really exciting. But I agree it's about, we're always going to have to think, well, is this actually answering the question I asked, is it doing what I think it's doing, in the same way that we've got machines that run tons of experiments now, we always look at the output and go, hm. Yeah. Because we're trying to push the edge of what's possible and every time you do that you should always be looking at it with a little bit of suspicion.
Is this in any way a generational issue, in so much, is it going to be easier for younger scientists to ask the questions that you're suggesting than people who have maybe followed a different path for a number of years?
I think it's a real mixture there. Right. Because there is also something very, very powerful about deeply understanding a subject. Yeah. It's very hard. You know, if you think of when you're a brand new PhD student or brand new, you know, starting a research project and you're very early on in your career, it's really hard to formulate really good questions. It's actually one of the hardest things to do, like a question that is important and answerable, maybe, and, you know, has boundaries on it that makes sense and is actually useful in some way, as opposed to just, I can write down loads of questions people can do, but they're actually not interesting in the big scheme of will they actually change something?
And deep knowledge of a subject tends to be something that really helps you to do that. But, and it's always true, and that's why it's so exciting to be in academics. One thing, I say this all the time, I love best about being an academic is working with students because they continuously tell me I'm an idiot and they should because they see the world brand new. And so yes, in some ways it's easier for them because of that. So I think there's both sides of that. And so that is where you find fun.
Well, based on that, in terms of deeper understanding leading to better questions, do you think AI could ask better questions if it has a deeper understanding and greater breadth of knowledge?
I think that's one of the ways that science is going to be changed. So at the moment, would I trust an AI model, you know, a large language model, to come up with a brilliant science question? No, because I'd almost certainly find that it had come up with it based on something somebody had already written on the internet somewhere. Do I imagine a future where using it as an, people already are, is one of the ways you can, like it's almost like interrogate what's already known, but also force yourself to think about, you know, it's another way of having someone to chat to. But this someone is different from a human. This someone is someone who can hold the entire knowledge of the internet in its head, but doesn't make links the way I would make links.
Yeah, yeah, yeah.
And that's good. For sure.
And on asking questions and how they're going to change, in the field of drug discovery, how have you seen questions change with the advent of AI?
I think the biggest changes, and we're only just starting to see them, I was talking to other people about this today, it wasn't that long ago that being a computational biologist or doing computational work in drug discovery, people sort of ignored you. They sort of paid attention in the sense that you're a computer person. That's nice. You did something extra. Yeah, but actually that's not true anymore. I think the field as a whole has done enough stuff that we've earnt our right to be serious.
Yeah.
But the big change, and I think it links back to OpenBind, is the concept that when you are designing how you do a process for drug discovery or thinking about how you do it, you don't go, here is my experimental pathway, how can computation help me? You go design a pathway where experiments and computation are working together so you maximize the benefits of both, because much of what we do around computation at the moment doesn't maximize its benefits, we don't use it at the right point in time. We don't think about it.
I would love somebody to actually, and it would fail, just to be clear, but to imagine, to do a drug discovery project which was entirely computation first. There's a really old saying from Henry Ford about if you asked the man on the street what they wanted, they'd want a faster horse, not a car. And this is that kind of step change. And that's the difference I'm starting to see, where now I can say, if you collect your data in this way, the computer will work better, on people. Instead of going, go, okay, we'll try it, you know? And if I say, if I make these predictions, you know, I am actually going to speed up your experiment, they believe me. And it's not always true, you know, there is a success rate here. Not everything works. Nothing is perfect, but nothing else is perfect either. So I think that's the change. The change is actually in people's minds that allows us to do this. And that's the bigger change and the longer term change.
So you're hoping for AI to almost design your experimental, the experimental process in some way, or to have a significant input in how you might curate and collect data that might benefit it. Could you conceive that AI might be a good place to then input into that process?
Oh, yeah. I mean, that's exactly how you want to think about it. The models are pretty good at telling you when they fail. Yeah. Not always. Sometimes they spectacularly fail and they're just lying to you. I hate doing that because it makes them sound like they're human. They're not human at all.
That's right.
They just make horrendous mistakes sometimes because they're a mathematical model. But given that, you can also use that as a way of working out what data you should have that would make them loads more effective. Yeah, and that's a really important piece of knowledge. And of course, if you can do that, then they become more effective and then they get better at telling you that piece of information and you get better at making your predictions. Yep. And so I think there's a massive part about how you link up much more closely your machine learning method, your AI method, your computation with your experiment, with an understanding of what you need in order to have a good statistical basis for this, so that what you design and do follows in that way.
Well, Charlotte, thanks ever so much indeed for joining us. We really appreciate it. We loved having you here. And for me, I'm going to take one takeaway, if you don't mind, which is there's still a job for us because good questions are important.
Fair enough.
Thank you.
Thank you so much for joining us.
So guys, Charlotte Deane, what did you make of that? Very interesting views on drug discovery and UK AI.
I know, I thought it was really interesting that the aim of OpenBind wasn't to discover drugs, it was to really discover the general schema, the general mechanisms for small molecule interactions that you could then leverage for drugs. But a more general idea, a more generalized model. And I was just quite taken aback by this kind of idea that it's a pre-competitive consortium, in a sense that both pharma and academics invest their time and effort, money into this. Everyone benefits. You get no winners and no losers in that sense. And I thought that was a really interesting dynamic.
It's really interesting, wasn't it? The role of government. The role of government is to help create the data sets, because there isn't sort of money, if you like, directly tied in the data sets, but without the data sets, there's no magic, as you said, in AI, right? So that is very interesting to see the role of government. I was particularly, I love the fact that she said governments cooperate on this, which if you watch the news every day, you might not think that's the case. But, you know, I find that very interesting.
But I think compared to governments, I think the global culture of science and the fact that scientists don't really care about governments, they still come together. They put their data together, they build data sets like the Protein Data Bank, which was over 50 years ago. And there was a lot of, you know, global tension over the past 50 years, but it still continued to go as it has over the years.
Yes.
And the idea that not a single company could recreate the Protein Data Bank, which is worth, by a lot of estimates, about $20 billion if you were to recreate it. And of course, no single company could generate that. So it kind of speaks to the strength of lots of researchers across the globe contributing to this dataset.
What did you think about her point about the way we do science is going to change?
I agree with her in lots of ways. I think asking those really important questions is a skill that scientists have to acquire, and putting AI into that mix in order to help, you know, refine or drill down onto an important question, and maybe even getting it to think about ways in which you answer the question, I thought were really interesting. And I can clearly see a use case for that as we move on.
Yeah. And I also think it was interesting how she was saying that the place to start is that you just have to do it. You just have to start talking to ChatGPT, whether it's gibberish, but you just have to start using it. Make yourself uncomfortable and it's not going to be clean cut. It's not going to be obvious, it's not going to be pretty. But that's the only way to kind of get ahead because it's coming whether we like it or not.
So what do you think? What's your verdict on her answer as to how the UK competes globally on AI? Were you sold on her answer to strategic advantage?
We weren't sold on it yesterday.
I think she sold it to me because I think that the only way to kind of get ahead in a David or Goliath situation, you will be incredibly demoralized if you say, well, there's no chance. And the thing was that the UK, during the Industrial Revolution, was in a power position, and there have been instances where just focusing in on one thing and doubling down on it has resulted in a company or organization leapfrogging. That is what OpenAI did compared to Google. And then they just doubled down with a small team, say, our goal is AGI, and that's all we're going to do. And they concentrated all their resources there. And similarly with Google, contrastingly, in fact, they had a lot of kind of business models to concentrate on, whether that search, whether that's DeepMind, whether that's Isomorphic. So I think really narrowing in on our talent, on, you know, people who want to do really innovative, long term visionary research, we should, we should concentrate on that.
Do you want to make $1 billion?
Who doesn't?
Sounds like hard work.
Well, thank you guys for joining me today, Stephen. Luke, it was great to have you.
My pleasure.
Thank you for joining us on Agents of Tech and for watching. What an amazing episode this was on drug discovery and AI in the UK. Please like and subscribe to see more.