Transcription
So, I mean, you literally say, "Make me a drug for X disease." Off it goes. Says, "Here's the molecule you need." Yeah. Yeah. Do you think that's possible? It's possible. I think it's possible. I think everything is pointing in that direction.
What I've seen firsthand is what can come out of this like explosion of two fields coming together, I think when you've got like experts in two fields and they come together and they're really curious, deeply curious about the other field and they want to apply their thing to your field, that's when you get this kind of magic.
In 5 years time, doing drug design without AI will be like doing any sort of science without maths, you know, and and join us. I think the whole of science will be like this is like if you're not using AI, what what are you doing?
Welcome back to Google DeepMind the podcast. My name is Professor Hannah Fry. Now, you'll already know that Demis Hassabis and John Jumper won the Nobel Prize in 2024 for their work applying artificial intelligence to protein folding. Now, everything comes down to proteins in the human body, how they fold, how they function. But not very long ago, working out the structure for one single protein could take months or even years. And then with the release of AlphaFold 2, the algorithms developed at Google DeepMind, the entire field has been completely revolutionized. More recently, AlphaFold 3 can predict the structure of all of life's molecules with unprecedented accuracy, which turns out to be absolutely pivotal for drug design.
These developments have paved the way for a new company spun out of Google DeepMind. It's called Isomorphic Labs to represent the synergy between biology and AI. And joining me today are two of its most notable hires. Rebecca Paul is head of medicinal drug design with years of experience in the process of discovering new drugs. And Max Yardberg is its chief AI officer. Anyone who's been following this podcast will remember Max from his earlier days at DeepMind, teaching agents to play Capture the Flag and Starcraft. and today he thinks that agents will be instrumental in the future of drug discovery.
Max, Rebecca, thank you so much for joining me.
Thank you for having us.
Yes, a pleasure to be here.
Well, it's a delight to have you because there's some some big stuff happening. Um, I know a lot has been made of this this this claim that AI is going to be able to solve all diseases. Is that realistic, Max?
This isn't going to happen overnight. Let's be clear. I think the exciting thing is that we can actually see there's perhaps a practical path towards that point and it's very different at this point in time than it has ever been because we've got these AI and machine learning models that understand the biological world and the biochemical world in a completely different manner than what we've had before. And so that's opening up you know tons of disease space that we didn't think was tractable before. Uh and that's just today. So as we start to develop these models further and further, it's really just the very beginning.
Does that mean every I mean every disease is on the table here Becky?
So I would say nothing is off the table at this point and you know the journey for me here is you know been a big one. I used to be much more conservative in this space but having come to Isomorphic Labs seeing you know the kind of models that we have things that I thought in the past we would never be able to predict and now being able to do it every day within you know 5 10 seconds um completely shifted my mindset so now I would put nothing off the table.
And and is it just in drug design or is AI going to affect clinical trials as well?
I think over time yes it will definitely uh affect clinical trials we're focusing really heavily on the drug design phase at the moment Um but you can imagine a world where as you start to get better and better at drug design actually you know more and more of the bottleneck comes onto the clinical uh development side of things and so we really need to rethink how we do that I don't think we've really changed the way we do clinical development for a long long time uh it's still yeah very slow there's of course a lot of regulation for good reason but as we understand more and more about how these molecules work you know the true mechanisms of disease how these molecules also interact with the rest of the body and affect everything like toxicity inside of us. We can start to rethink even how we do those clinical trial designs, how we first go into people and start measuring the efficacy of these molecules. So that's not right now for us at Isomorphic but it's very much in our future.
What is the big idea then? What's the big ambition of Isomorphic?
It really is stepping towards that sol space. And so this is really before you go into clinical trial, before you even start testing out on people, you need something to test, right? You need to create a molecule and a drug in a in a drug. What is that? It's something that goes in and modulates some function of the body, some function in a cell. And so really, you know, the the the key first phase of Isomorphic labs is how can we create this AI drug design engine that can take you pretty much any disease, any protein target that's implicated in that disease and work out how to create a molecule that will go in and start modulating the function of these proteins, the function of cells and then change the disease state for you know positive of patients.
The thing is I mean diseases have been cured in the past right we have come up with like drug solutions that effectively make them you know a remnant of history. Why are some diseases so much more difficult than others to to solve or cure?
So um you know for example some types of cancers you you can develop a treatment to cure them because for example those cancers might be quite stable. So there's maybe one mutation that drives that cancer and then you know you treat that you you you treat with that molecule that hits that particular mutation and that can be curative in that disease. But if you had a cancer which was continually evolving and um you know the cells were evolving to overcome the drug that you're treating with then you need to be continually evolving the drug that you're treating them with. So that that would be then a much more difficult disease to solve.
And I guess are there others where we where there's just like a real gap in our understanding of what's going on in the body.
Yeah. There's also that, you know, massive gap in understanding, you know, what what's actually driving this disease. Is it multiple things together? We we don't have all the answers yet to biology. Biology is so so complex. Um, one of the the fun the fundamental things we need to do is find out, you know, what what biology is actually driving disease before we can then develop a small molecule or some kind of chemistry that then will modulate the right biology to actually make an improvement in the the symptoms or a disease modification to that disease.
Is that what you're trying to do then?
So, so actually something like cancer for instance once you get down to like the level of molecules and proteins it's that something's gone wrong at that level which then you know escalates in the in to the scale of the human body.
Absolutely. So you've got little you can think of proteins as like mini uh factories or engines inside your cell. Um and those they have a function they do something. So if you have a mutation or something that changes about that protein, that means for example, it's always switched on where in a normal cell it might be going on and off or maybe it's mostly off. But now in a cancer cell, something's happened. It's always on. It's going to just continually drive a signal and that signal could be grow, proliferate, and that will then drive the formation of a tumor. And so your target then you know that there's something going on with that protein. So you're targeting then that protein with a molecule of some medicine or drug.
Exactly. Okay. So you want to if you think of that that little factory working away, you want to design like the perfect shaped wrench that you can throw into that little set of gears to stop it working so that then it's going to stop driving that signal.
Does it come down to shape?
Very important. You can think of proteins as having pockets or grooves or crevices and you really need to design a molecule that perfectly fits inside that groove to block it.
To block it. Oh, really? It's literally like that.
Yes. And is that true of of of all drugs that that already exist? I mean like paracetamol for example, is is that what it's doing?
Yes. So paracetamol is a small molecule binds to its protein targets, blocks the function. Um and in in the case of paracetamol or painkillers, it's preventing you from from experiencing pain. So um yes.
So in some ways then you're sort of playing Lego at the at the molecular level.
Exactly. Yes. And we love it.
Why why would that be a good situation for AI? what what makes this a a wells suited problem for AI Max?
It's actually such a perfect application of AI and machine learning. Um this whole you know playing Lego with molecules we've seen over the last you know five six seven years the rise of models like AlphaFold. We had AlphaFold one AlphaFold 2 understanding the structure of proteins. You know before AlphaFold 2 no one could really understand the structure without going into a lab and experimentally resolving these structures. And that can take months it can take years. Sometimes it's not even possible for some proteins. Obviously, AlphaFold 2 Nobel prizewinning breakthrough in chemistry. And now we've taken that even further with things like AlphaFold 3 where now we can understand the structure of proteins with small molecules. And these small molecules are the little Lego blocks that come in and we use as drugs to inhibit the function of a protein or change the function of a protein. And the reason why this is such a good uh domain for machine learning is that what we're trying to do in essence is is predict the 3D coordinates of this uh biomolecular system. And this fits really really nicely into some of our classic supervised learning modeling domains. It fits really really nicely into our diffusion modeling um frameworks that have been so so successful for things like image generation or video generation and in terms of being suited for the supervised learning stuff that had already gone before.
Is that because there is some some metric of success here like some way of of of being right in vertic?
There's a very clear metric of success here which is super helpful um for developing these models and for research and the reason is over the last 50 years people have been experimentally resolving these protein structures structure these proteins with small molecules with DNA with RNA they've been doing it by hand in a lab and then depositing the results into a big database called the protein databank PDB and this gives you know a really rich source of information it's you know a couple of hundred thousand 3D structure structures and each structure has thousands of atom coordinates. So there's really high information density and that's a perfect scenario for supervised learning. Now this isn't webcale data. So it's not the scale of data that we might be used to for training large language models but incredibly we've worked out ways to design these neural network architectures and the these training regimes so that we can only train on you know a couple of hundred thousand structures and excitingly get what we call generalization. You know, we see that these models can generalize. They can be applied to completely new proteins, completely new molecules that people have never seen before in history. And of course, that's essential if you're doing drug design. Drug design is about creating completely new molecules that we've never seen before in nature even to actually modulate these functions.
But then also, I guess like then the the possibilities of molecules that you could design. I mean, it's sort of well, a very big number, I imagine.
Yeah, that that that number is huge. um you know people people throw out around numbers like uh 10 to the power of 60 is the the possible number of drug-like molecules out there in the universe. So it's it's a huge combinatorial problem and uh that's also a really exciting spot for AI. You know, we can create great predictive models of you how these molecules fit together, even how strongly they fit or the properties of them. But with a design space of 10 to the power of 60, you know, we're reaching the level of, you know, atoms in in the universe. So even if you had the perfect predictive models of how this fits together, you wouldn't be able to exhaustively search that massive space. So what do you do? Okay, maybe you subsample that space and you search through some large number, a million, 10 million, a billion, 10 billion. Even though you're not Yeah, you're not scratching even scratching the surface. And that's where you know we can then fall to new types of models. Things like generative models, search methods, agents which instead of exhaustively searching the full molecular space, we can really smartly start to explore across that whole space but without exhaustively searching the whole space.
Becky, tell me about AlphaFold 3 then in terms of designing drugs. What does it what does it actually allow you to do?
So, as a medicinal chemist, we we always want to be able to visualize how our molecule binds to our protein. So, how our Lego block fits into the bigger picture of Lego blocks. Thank you for going with me on this analogy.
Lego analogy.
Um, the reason we need to have that visualization is because when you're optimizing a small molecule binding to a protein, you need to know uh which vector to explore. You need to have kind of some kind of target in mind. I'm going to explore that part of the pocket or I'm going to you know that part of the the protein structure like that looks that looks like a good place to go. Um and so for years we've invested as as a kind of scientific community in ways to do that. So X-ray crystallography is an experimental technique where you can actually go into a lab and you can um spend a lot of time crystallizing your protein. you fire X-rays at it once it's bound to your small molecule um and then you can actually visualize um atom by atom how your small molecule is binding to your protein. Now we can do that with AlphaFold 3 and sort of the latest iterations of that model you know in just seconds. And so something that you know might have taken me months when I was you know doing my PhD or in my early stage research I'm now just seeing on my screen all the time.
And so you can just iterate and iterate in silico until you get to something that actually looks really quite promising. Then you take that into the lab.
So which way around does it work then? Are you saying okay I think something like this this this and this would work. Let's try it out and see if it fits or is is it the other way round? Is is are the are these models telling you this is something that might fit?
So in the way that we've constructed our drug design platform at ISO, you can do both. So I can come in as an experienced medicinal chemist and I can say I think this and I can test it then and there couple of minutes and get that that feedback. But you can also take the opposite approach. Okay, I don't actually know what's going to work here. So I'm going to apply the generative models we have. I'm going to um do some what we call virtual screening. So I'm going to take an area of chemical space uh that's commercially available. I'm going to screen that against my protein um and I'm going to get the models to tell me what what what's best from that subset.
Can I see what it looks like when you're actually designing something?
So, the small molecule is fitting into this little groove in the protein um and uh it's forming interactions with the protein. So, the dotted lines you can see, those are interactions between that small molecule and the protein itself. And as medicinal chemists, we want to optimize or increase the number of those interactions because that's increasing the strength of that kind of relationship between the small molecule and the protein.
Let me see, let me describe what I've got, what's going on here. So, so you've got the the sort of curly stuff is a protein.
Yes, that's the protein.
Yeah. And it's it's it's folded. So, you've got the kind of threedimensional structure.
Indeed.
And then over here you've got a I mean, this looks like the kind of thing you would do in GCSE chemistry by those kind of diagrams.
Exactly. This is a small molecule. This is sort of plugged into the the the protein.
Yeah.
Wow. I mean, it really is like 3D jigsaws then. I mean, that's exactly what you're doing.
That's what we're doing.
Yeah. That crevice could be the thing that's making you feel pain or the thing that's, you know, causing tumor growth or whatever it might be.
Yeah. Causing your disease.
Yeah. Amazing. So then you're trying different versions of this molecule to see if you can get the best possible fit in that little crevice.
Exactly. And I can I can quickly show you. So we have lots of different functionality on the on the platform that you can try. Um this is my favorite which Max always tells me off about because I can use my expertise as a medicinal chemist and I can say, "Okay, I want to make want to make some specific changes to this molecule." I can actually view what I'm doing in 3D. So this is now going to fetch that structure prediction and I can actually make modifications to this molecule and I can see the predicted structure in real time. And normally this would have taken I mean before AI a long time. I mean if you're going to go into a lab and experimentally determine this, it could be anywhere from weeks to years.
Why do you tell her off for this one?
I guess you're referring to the fact that over time we want to do more and more from the model itself. The really exciting frontier from my perspective is uh you know that there's going to be lots of scenarios where Becky will want to go in and make those changes by hand and test out very specific hypotheses that she has on why this molecule works and how we can make it better. And then you know what we also should be doing is asking our generative models and our agents to say hey this is how I'm thinking about the problem. These are my design constraints. So I want a molecule that does XY Z and has these sort of properties and kind of looks like this and maybe interacts over there and makes this sort of shape. What can you come up with and you know maybe set this running go away have a coffee go home come come in the next morning and see what the agent has come up with.
I guess with all of the projects that you've applied AI to sort of generally in DeepMind, it has gone through that that that that process, right, of starting off with human expertise and then kind of slowly building in more knowledge and expertise within the model itself.
Yeah. And I think there are a lot of analogies to that moment that we had with large language models where we've had large language models for a long time. I've been working on them, you know, 10 years ago, but they were kind of rubbish and they were spitting stuff out that looked like language. It kind of made sense, but it also didn't make sense. You had to correct and it clearly wasn't human. And then they got steadily better little by little and suddenly they just passed through this, you know, human perceptible threshold where you can't really tell whether this is generated by a human or not. And we're getting to, you know, to the same point with our molecule design models where, you know, maybe 5 years ago in this field, you you you had generative models of molecules and they'd spit stuff out, but you'd give them to a chemist like Becky and and she would probably like tear her hair out like this is this is rubbish.
Well, did you see those kinds of models?
Yes, I did.
Yeah, cuz I actually worked at an AI company prior to joining Isomorphic Labs. So, I've seen like that that progression and that journey. and and tell me what what kind of stuff did they spit out?
Um, you know, for a long time it would just be kind of nonsense. Um, because you you're obviously giving the model an uphill function. You want it to to to get to like something that's going to bind
Really potently, for example. But to do that, maybe it just like makes the molecule massive. Oh. But then you know, well, it's not well then it's not going to actually be absorbed through the intestine into the bloodstream, because because you've got more to think about than just the molecule itself.
Yeah. So there's a lot to piece together here. Oh, that's interesting. So the AI, if we go back to our Lego analogy, was just like building a massive Lego wall all around the protein.
Yes. Yeah, I see. Okay. Essentially, and now this is—Have you seen that moment, that tip over, like like with the language models that Max describes?
Yeah. I think I—I've been so surprised by the quality of some of the molecules that come out of the generative AI. And sometimes the molecules that come out, you think, "Oh, I would have—why wouldn't I have come up with that?" Like that's really—that's really amazing. And of course, you don't have to go and make that exact molecule, but you could then use that as inspiration to like do something else. So, um, so it's working with you.
Yeah, you can work together. I mean, I'm guessing at some point in the future, it will be so good that you'll be like, "Oh, there's nothing I would change." We—we—we've had some really fun moments where, um, you know, for example, we've had our models submitting molecules blind and then other people looking at them and seeing, okay, what are we going to send off for testing? And oh, really? Like test for molecules.
Exactly. And people looking at these molecules—like very, very experienced medicinal chemists—saying, "Wow, there's a lot of experience behind the design of this molecule," and actually not knowing that this was designed by a general instead. So, but okay, let—let me understand this though, because using the the analogy of large language models, it sort of makes sense there that you—you have these tokens, you know, you kind of break words down into little—little bites, um, and then you can kind of build up from there. How do you do it in a way that makes sense chemically? I mean, you're not just taking atoms, are you?
We are actually just taking atoms. Um, for bigger things like proteins, we chunk up into amino acids. So, one token per amino acid. So, instead of, you know, characters of—of a sentence, letters of a sentence, we have amino acids of a protein. And then for the small molecule, we chunk it up just into its individual atoms. And so, we have a sequence of amino acids and a sequence of atoms. And we put them together. And that's one big sequence. And then we feed it through a structure model like AlphaFold3. And AlphaFold3 uses, um, you know, transformers, but unlike in large language models where transformers are used on one-dimensional sequences of—of characters—of letters, here we use what we call a pair former, which operates on a—on a two-dimensional interaction grid of all of these molecular elements. So we can consider every single possible interaction that could occur between every amino acid, every part of the protein and every atom of the small molecule and everything in between. And then this creates, um, you know, neural network features which condition a diffusion model. And diffusion models are generative models. We probably know them from these amazing image generative models, video generative models. And instead of generating the pixels of an image, instead our diffusion models are generating the 3D atom coordinates of this whole biomolecular system. And it just so happens they work. And it just so happens this works phenomenally well. Um, you get these amazing structure predictions that, you know, when you go to the lab and experimentally resolve these structures—and you know we do this on occasion—we, you know, something amazing is predicted, like a completely new pocket or a new mechanism of action. We go into the lab, we want to check that, you know, are these models grounded at all in reality, and what comes back is like, "Yeah, this is like within one angstrom," like, you know, the tiniest unit of—of—of distance accurate, which is phenomenal.
Yeah. But then through that training process, does it sort of manage to extract a kind of conceptual understanding of—of—of how chemistry works?
It's really hard to think about concepts in this atom space, but I do believe that there's some notion of, um, reasoning in molecular and atomistic space that these models are doing because of the amount of generalization we're getting out of them. And if you think about what these generative models are trained to do, they're trained to fit to the data distribution that you give them. And so in our case, we give them, you know, all the molecules that might exist naturally, that people have worked out before, that people have designed before. When your model gets better and better, you get things that look like they could have been designed before, which, you know, then is—starts to be imperceptible from a human design.
What do you think?
You can certainly see that in the molecules that we get back. They look like molecules that myself or someone else might have designed. You do sometimes get something crazy though.
So do you?
Yeah, we—we—we still do, but we kind of have ways of filtering that out now.
Is it sort of like a hallucination in a way?
Yeah, the model's like really confident that it's—kind of confidently wrong.
Confidently wrong. Okay. What does it look like if it—when it hallucinates?
So, you know, we get back, you know, a number of structures that the models believe are, you know, good solutions for this problem, for this particular protein pocket. And then our job is to go, okay, well, which of those molecules should we actually select to put into synthesis? And so this is never a model on its own in a silo. This is a model working really closely with an expert to say, okay, of the—of the—of the solutions you've given me, where's the gold in that? Where's—where's—where are the—the molecules that are actually going to, you know, push this project forward, and we'll put them into what we call chemical synthesis, where we make them and we test them. But sometimes we get back results which are, you know, actually the compound doesn't bind to the protein at all. So here the models essentially completely hallucinated a solution, um, so convincingly that actually an expert looks at it and goes, "Yeah, that looks like a really good solution," you know, confidence metrics would suggest the same, um, so it's essentially—it is a hallucination, and I think it—we find it a really fascinating research question to say, okay, how do we find the really good stuff that the model's giving us beyond this 3D jigsaw—or Lego, as we're, you know, mixing our metaphors—is it just about structure, is it just about finding something that will plug a particular hole, or are there other considerations that you have to have as well when it comes to drug design?
There are so many other considerations. That's what makes this problem just incredibly complex. So, it's even beyond the shape, right? You can have something that maybe fits in the shape, but it's got to bind really strongly to that—to that protein. And that prediction of binding affinity, as we call it, is actually different. You can't really gauge that from just looking at a picture. The picture is a helpful guide. But you need to be able to predict that binding affinity kind of separately. And then all of those things together, that's just how your molecule binds to your protein. You've also got to think about—is that molecule going to bind to any of the other 20,000 proteins in the body? Cuz if the answer is yes, that could drive a side effect that you don't want. That's going to drive you some toxicity. Um, you know, is this molecule going to be stable? It's got to survive like the really acidic conditions of the stomach. It's got to survive going through the liver, which is like going through a war zone. Um, you know, the liver wants to do everything it can. It's like, "This is a foreign molecule. I need to like get rid of it." So, your molecule's got to be really robust. It's got to survive that journey. It's got to be soluble. When you take a pill, that pill's got to dissolve in your stomach, and it's got to stay dissolved all the way through your intestine, because otherwise, it's not going to absorb into your body. And a lot of these parameters are pulling against each other. So, for something to be soluble and dissolve really well, um, it needs to be water loving. But for it to bind to the protein and to gain affinity in that kind of pocket, that Lego connection, it actually needs to be kind of water hating. So how do you possibly solve that if—if you need opposing characteristics?
This—So we've up until this point, and still now to some degree, drug discovery is a very iterative process. So you know that the human brain can only think about so many things at one time. Okay, I'm going to solve—I'm going to solve this binding affinity problem a bit, and then I'm going to start to think, okay, is my molecule soluble, and then I'm going to start to bring in gradually these other properties. And it's honestly, it's like whack-a-mole. You play like this like three-ear game of whack-a-mole where you're like, okay, fix this—this problem. Hooray. And then—and then this other one pops up, and you're like, okay, I'll fix that. But then the other one's gone bad again. So to be able to predict all these things in silico is still a really, really hard problem.
Are you working on tools that will help with those elements, too?
Yeah, absolutely. So you know we think about at Isomorphic Labs, how do we do drug design end to end, you know, which really means solving all of these very, very hard problems with, um, you know, cell permeability, solubility, toxicity, liver clearance, everything, um, and none of this is solved—these are really hard problems to even model, so we spend a lot of effort, lot of research on creating new models to—to really understand this better, and then as—as Becky was talking about, how do we then start to find these needles in a haystack—molecules that are, you know, somehow just balancing the properties just right to be a perfect drug—and it's really, really hard, and actually there are a lot of analogies to maybe, um, what I used to do at DeepMind in, for example, Capture the Flag or—or StarCraft, is you know there's not just one agent or strategy that solves StarCraft or a particular game like Go—you have to completely start mixing up these strategies and working out exploits for each individual strategy and basically searching this huge combinatorial strategy space, and in the same way we need to be searching this huge combinatorial molecule space. So just like you might have, you know, tree search in a game of Go where at every move you elucidate, you know, some other possible moves and you start searching through that tree of possible strategies, going deeper and deeper, and just like you can do that for moves in a game of Go, you can imagine doing a similar thing for designing a molecule. So you start with a part of a molecule and you—you know, start to hypothesize what are the different things I could add or take away from this molecule, and you get to a whole tree of possible futures that you can then score and work out a value associated with that to create that perfect molecule for this very specific indication.
But how do you even know that the perfect molecule exists? Like maybe there's just some crevices in—in the proteins that just are unfittable.
We do have this concept of like undruggable proteins, um, where the crevice is like really flat, you can't really get anything to grip in there. Um, and those proteins might need different solutions. So actually we have an—a whole emerging field which we call molecular glues, and this is where, um, you know, you have two proteins that come together, and the pocket that's formed when they come together is actually a much more suitable pocket. So now you need to design a molecule that sits, uh, in the middle of them and glues them together. So there's like this whole explosion of all these different modalities now, which makes this an incredibly exciting field to work in.
From my perspective, the fact that we've actually found any drugs at all already, given how hard and complex the problem is, you know, and—and we've basically been doing, you know, a bit of human intuition and a lot of random, you know, screening, experimental testing, and we've managed to find molecules even though the design space is huge. Actually, that gives me a lot of hope because that means that there's probably a lot of redundancy in chemical space. I—I—there's like probably lots of different solutions that could work, but we've just got to find them.
Have you actually tried to make any of these molecules, or do—at the moment—do they just exist on the screen?
Oh, no. We make a lot of molecules. We've got a huge, um, experimental footprint.
Yeah. And how do they turn out—how you expect? I mean, what are the results like?
We've had like some, you know, incredible success in some of our projects, um, where, you know, I—I—I'll find Max at his desk and I'll be like, "Have you seen this thing?" And you'll just both be like really kind of mind-blown about it—that when you get the molecule, it actually works.
Yeah. Exactly. Exactly. And we have our own drug design programs. So things that we've started from scratch ourselves. We also work with, um, pharma company partners, people like Eli Lilly and Novartis. In these collaborations, you'll get specific targets to work on, um, and you know, these are ones that these companies have high conviction behind and probably a bunch of evidence behind, um, you know, some of the collaborations we're in—we've been given very, very hard targets, you know, these—these are things that people have worked on sometimes for, you know, over a decade, and, uh, you know, not made significant progress to the point where you've got something on the market—things like cancer and that sort of stuff—whole host of—of—of therapeutic areas, um, and disease areas—and then, you know, Becky and team sit down, start designing with these models and, you know, can start finding completely novel chemical matter for completely novel, you know, mechanisms that—that no one's really discovered before, which is mind-blowing for me as a computer scientist.
Yeah, that's mind-blowing for me. There's been some like, um, almost like career-defining moments, like where you—the AI will—or the—will give you a hypothesis, right? Right? It will—it will suggest something, and you think, like, "I'm not convinced I would do that," but the model's telling me this thing, and it's really quite convinced about this thing, so I should maybe just test this hypothesis, and then actually it turns out that the model was right, um, and you were absolutely right to test it, and it's really pushed forward your—your project or even like that kind of field. So, um, I think for me it's not about how we trust the models, it's about how we are open to testing the hypotheses that they put in front of us, um, and not sort of going, "Oh, that—that doesn't fit with my worldview, so I'm not going to test it," but then you are also human, right? So I do wonder whether if you see, you know, lots of—of hits with the model, as it were, if the model is coming up with lots of good stuff in a row, do you start sort of maybe trusting it more than yourself?
We actually put a lot of trust in the models. We actually use, for example, some of the models we have—we use them as quite strict cut-offs.
What kind of cut-off?
Like, for example, we have a model which we call binding probability, and it goes from zero to one, so one is, um, you know, the model is convinced your—your molecule is definitely going to bind to your protein, zero—the model is telling you this is definitely not going to bind, and you know you can build a little bit of confidence over time that the model really does understand, okay, anything below 0.7, it's really got a very low probability of success, so we just—define that as a cutoff and be like, "We're not going to put anything in the lab that's got a probability of less than this," because actually the model's quite likely to be right—it's probably not going to be any good. And so you do—even though—and that's quite hard because as a chemist, you design something and you think that was a really clever idea that I just came up with, and like, why doesn't it—like it? But then what if it makes something that you don't understand? I mean, or—or—or that doesn't make sense? I mean, do—does it need to explain itself?
I think at the moment that explainability is quite important for now because the process is kind of quite driven still by the human—it's not end to end yet. We have to go in there, and we have to—we have to sort of say what—what comes next. So if there's no explainability there, you—you don't know what would be next, right? And so that—that would be very difficult to work with. I can imagine like in a future state where actually the—the process is a bit more end to end—like in one step the—the model here's a drug—yeah—then actually that—you maybe you don't need that explainability, but when you've got to go in there as a human and you've got to iterate and you've got to do a bit more of that, um, directionality, then that explainability is important, and that's where I think for me the AlphaFold models really come in because, uh, okay, the—the models predicting this molecule is going to be good. I can rationalize that with what I'm actually seeing. I know what I would do next.
You have a slightly different view on explainability, don't you?
I do have a slightly different view on explainability, but I think you need explainability when your model sucks basically. And, um, you know, we don't have perfect models yet. So, I think, you know, there's—there's a good amount of room for explainability, but I always like hear the call for explainability and think, look, we need—we need to make this model better. And actually the interesting thing about explainability is it can help you understand the pathologies that this model has, the biases that it has, where is it wrong, you know, given the science that we know about. And so we can start patching that and make it better and better and get to this point where, yeah, actually we can just do end-to-end design purely in silico and maybe to do a final round of verification in the lab at the end.
So, I mean, you literally say, "Make me a drug for X disease." Off it goes. Says, "Here's the molecule you need."
Yeah. Yeah.
Do you think that's possible?
I think it's possible. I think it's possible. I think everything is pointing in that direction. We're getting better and better. Uh, we're already reducing the amount of, you know, experimental cycles you need, um, reducing the amount of lab time you need, um, and yeah, this is just the beginning.
Absolutely extraordinary. I mean, I suppose it does—it does depend on knowing what protein you're targeting too, right?
Yes. So still—and the diseases that we don't have a full understanding of are still going to be difficult. Yeah. And that—that kind of—we call it target ID space—where you actually need to identify the protein that's causing your disease—it's actually like a really important part of drug discovery because if you're not hitting the right biological target from the start, um, you can design the best molecule in the world—it's not going to do what you want it to do when you put it into—into a human. So there's a lot to be done in that target ID space, and I think AI has got a big role to play there as well. It's one of the big frontiers of, you know, AI for biology is really understanding, um, you know, what are those driving mechanisms of disease. Can we, you know, start to understand, um, how mutations in our DNA translate into changes of expression of RNA and how that changes, you know, the type of proteins and expression levels of proteins, how those proteins interact with each other and build up into these signaling pathways and how changes in those signaling pathways, you know, change the, um, disease state as well. Um, and of course if we can start to understand these bits, you know, we can start to work out where do we need to modulate this biological system, but all of this is really, really hard, and there's—there's some amazing breakthroughs happening in the field—understanding DNA better, understanding this translation better, um, even through understanding how proteins interact—can we build up these, um, interaction networks better—this is some of the really exciting frontier research that we're also doing.
At ISO, so you have a team working in that space as well. Yeah, that's right. We have a whole computational biology team, a whole machine learning modeling team focused in this space.
Yeah. But then what about personalized medicine? Because I guess each person is different in some ways. I mean, you know, this is the really exciting, you know, potential future where we can understand much more about, for example, cancer: individuals' mutations in their tumor. Um, and you know, through generative AI and design agents, be able to come up with molecules that work specifically for these sort of mutations. Now there's a whole question of how do we actually operationalize that and get these drugs to patients and approve this framework, but um, you know, we're moving towards a place where that technology is, you know, could be potentially there.
I mean, I'm thinking here about chemotherapy drugs, which come with really devastating side effects. You think there's real hope on the horizon for that kind of thing?
Yeah, I think when we think about chemotherapy drugs, they're basically drugs that are like non-specific. So, they're they're going into the body and they're kind of trying to halt that um rapid cell proliferation. Um, but what we have what we have now is an ability to think about actually what's the specific target, the protein target that we want to inhibit. We want to stop its function. Um, and that might have the same effect, but you're not just generally using something like just very toxic to rapidly dividing cells. Yes, it's going to stop your tumor cells dividing, but it's also going to stop the cells that line your stomach and your intestine. Going to make you feel nauseous and sick. It's going to stop your hair follicles. You're going to lose your hair. Um, whereas we now know we can go in, we can target a very specific protein, the one that's actually causing the disease. And if you inhibit that particular protein, that's not going to cause—hopefully, if you get it right—that's not going to cause all these other side effects. You can do some also really, really cool stuff of targeting particular cells. So if you know that like a particular cell type is expressing, you know, something on its surface, you can start programming um things like antibodies to come in and find those particular receptors. Um, and so you're delivering your um payloads directly to that particular cell type and not more broadly to the body.
I just want to go back to the point that you made earlier, Becky, about uh once you've got the drug design, then once you put it into the human, there's there's all of these other potential problems because I mean, there have been examples of this before um where drugs have been made and looked like they were very good, and then once you actually put it into a human, it causes some massive problem. I think there was one which uh people were very excited about the impact it was going to have on pain, but it turned out that protein also was quite crucial to making sure your heart kept beating. How do you mitigate against that, or can you not at this stage?
Well, one of the problems we have is that we often use animal models to then translate things into the clinic. And animal models, they don't replicate human physiology very well at all, actually. So, we know when we're working in the kind of discovery and preclinical space, which is all of that space before you go into a human. We're we're working with um different animal models which might model the disease we're interested in, and we're looking for molecules which have an effect in those animal models, and we have to show that they're not toxic in those animal models, and then we use that bank of evidence to go to the drug regulatory bodies and say, right, we're ready to go into a human. But from that point until the market, there's a 90% failure rate.
Wow, 90%. So all that investment up to that point, which is huge. What makes them so likely to fail?
So molecules fail in the clinic for toxicity. They fail in the clinic for lack of efficacy. And I think a lot of it comes back to, you know, the animal models we use just are not very good at replicating human physiology. We—a mouse is different to a human. A mouse is different. So we can cure mouse disease. Probably be quite good at that. We've got loads of medication that work. Yeah. But yeah, that translation is a big part of science that we we we need to fix.
Can AI help here as well? I mean, if if animal models are this sort of stumbling block with such a low level of success, what can you do about it?
Well, this is this is where we can actually, you know, use some of the technology and models we've been developing and and think about, okay, how can we understand toxicity better? Uh, understand the effect on human cells better and see how that translates to organs. If you think about, you know, some of these off-target effects, probably, um, there are many drugs that, you know, you go into the clinic and you're hitting your target of interest and it's curing your pain, but then it's hitting another target that's in another protein in your heart and and stopping the function of your heart. That's that's an off-target effect.
Yeah.
Yeah. Exactly. But you can imagine that if we've been building models that understand really well how this molecule interacts with your target of interest, you could also ask the question, well, how does this molecule interact with every other target in the human body, every, you know, all 20,000 proteins? And you can start building up this, you know, fingerprint of interactions that this molecule, your drug molecule, is having across the body. And so that can give you clues, maybe even concrete signals into the toxicity or side effects of this molecule. And the nice thing is we can get that signal not when you're going into humans, but actually at the very, very beginning of the design process. So by the time you've you've gone through all of your molecule design and you get to the point where like, yeah, I want to go into humans, you've been thinking about these side effects in a very rational way for a long time. And so you hopefully your chances of of actually hitting some of those radically reduces. You're taking your your structure of of Lego bricks or your jigsaw and you're destroying it with every other possible combination that it might encounter in a human body.
Yeah, we're going to make every possible Lego combination.
Well, if that's the design stage then, Becky, I mean, you always have to put this into clinical trials. Just just talk us through the process of clinical trials if you could.
So, the first time that your molecule ever goes into a human, that's a phase one clinical trial. So, it will be a small number of patients. Some of those might actually be um healthy volunteers; they don't necessarily have the disease that you're interested in. Um, and what you're looking to see is um, does your drug actually reach the level of exposure in the patient that would be needed to generate an effect and is the drug well tolerated or do you suddenly start to see some side effects that you weren't anticipating? If all is good, you'll proceed to a phase 2 clinical trial, which is now you're going into people who actually have the disease um, and you're going into larger numbers. You're really looking to answer the question, does your molecule actually have efficacy against the disease that you're interested in? Um, and this is where we do see that that big failure rate. So 70% of molecules going into phase 2 um don't actually pass through into phase three. For those that do pass into phase three, that's where you're going into much bigger patient populations, seeing if your drug is effective across that bigger population. You you've—it's not just got to be safe, but it's got to be better than the standard of care. There's got to be some, you know, for doctors to actually prescribe this to their patients, they've got to say this drug is better or this drug is safer than what I currently use. And this whole thing has like a 90% failure rate as you said.
Yeah. I mean, does that mean that there are people who work in this space who never never succeed?
Yeah. So um I'm a medicinal chemist, and um we we often have this number where actually only one in 20 medicinal chemists will ever get a drug to market. So 19 of us out of every 20 will never get a drug onto the market through our careers. So yeah, we are a we are a profession where we're used to seeing significant failure.
You're comfortable with failure.
Comfortable with failure. We learn from it.
Extraordinary.
Extraordinary to imagine. How long do you think it will be until the first AI-designed drug is on the market? Because all of these additional levels really take some time, don't they?
So, there's there's AI-designed drugs that are in the clinic now in clinical trials. Um, and there are different levels of AI input into those current drugs. Um, I I would imagine that in the next five five years or so, we're going to see an approval of one of those medicines. But for me, the big thing is going to be when can AI start to really fill out this pipeline and start to get drugs into the clinic really quickly and really start to deliver molecules for patients. That for me will be, you know, when when AI is having a really big impact—when you can start to say here's the target and then it sort of pops out a drug at the end. Yeah. And you can put that straight into uh into the clinic and be confident that it's not going to cause any damage to a person. And even a a slightly improved level of confidence in where we are now would be quite impactful.
M, yeah, because as Becky said, there's already molecules in the clinic that have been touched by AI, that have been enabled by AI in some way. We're just going to see more and more of that, you know, in in 5 years' time. Doing drug design without AI will be like doing any sort of science without maths, you know, and and join us. I think the whole of science will be like this is like if you're not using AI, what what are you doing? Right? There's just so much information to be to be gained there um so yeah, as Becky said, it's more like how do we actually see that rapid increase in in sort of disease areas that we're able to tackle or targets that we're able to unlock um ultimately like patients that we're able to help.
Amazing. Thank you both. That was really interesting.
Thank you for having us.
It was so much fun.
Yeah, it's been great to be here. I think I now realize that medicinal chemistry is one of the hardest jobs in the world. It takes years to design a drug. Even if you get it to clinical trials, 90% of them fail, and only one in 20 of your colleagues ever manages to see their medicine improving the lives of patients. But strangely, that is precisely what I think is so exciting about this space. Because if everything we've done up until now has effectively been like working in the dark, slowly laboriously navigating the most infinitesimally small areas of the vast landscape of possibilities, it's like someone has just turned on a floodlight. And okay, of course, we are still very, very far away from a big AI button that's just going to solve all diseases. But there is so much headroom here for improvement, so much scope to move the dial, and simultaneously so much opportunity to directly impact the lives of all of us.
You have been listening to Google DeepMind, the podcast with me, Professor Hannah Fry. If you enjoyed this episode, then do subscribe to our YouTube channel or leave a review on your favorite podcast platform. And of course, we have plenty more episodes on a whole range of topics to come. So, do check those out. See you next time.