Speaker A
HANNAH FRY: So, I mean, you literally say, make me a drug for X disease, off it goes, says, here's the molecule you need.
Exploring how AI revolutionizes drug design and disease treatment through protein folding and molecular modeling.
Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.
Generated from the transcript and can be wrong — check the timestamp.
Key Takeaways
- AI is revolutionizing drug discovery by accurately predicting protein structures and designing molecules.
- No disease is off the table for AI-driven drug design, with potential to tackle previously intractable conditions.
- Understanding disease biology at the molecular level is critical for developing effective treatments.
- AI will eventually transform clinical trials and drug development, speeding up the process and reducing costs.
- The integration of AI in science is becoming essential, akin to the role of mathematics in research.
What the video covers
- AI advancements like AlphaFold3 have transformed protein structure prediction, crucial for drug design.
- Isomorphic Labs, spun out of Google DeepMind, integrates AI and biology to accelerate drug discovery.
- AI models can now design molecules targeting specific proteins implicated in diseases rapidly and accurately.
- The complexity of diseases like cancer requires evolving drug designs to match biological changes.
- Understanding the biological drivers of diseases at the molecular level is essential for effective treatments.
- AI is expected to impact not only drug design but also clinical trials and drug development processes.
- The vast chemical space of potential drug molecules is being navigated efficiently with AI generative models.
- Explainability in AI models helps identify biases and improves understanding of disease mechanisms.
- AI-driven drug design aims to create safer, more effective drugs by predicting off-target effects early.
- The future of science and medicine will heavily rely on AI integration, making traditional methods obsolete.
Chapters
- 00:00Introduction to AI in Drug Design
- 01:44AlphaFold3 and Molecular Structure Prediction
- 03:26Expanding Disease Space with AI
- 05:24The Drug Design Process Explained
- 07:26Challenges in Understanding Disease Biology
- 09:13AI Models and Molecular Generation
- 11:36Testing AI-Designed Molecules
- 14:40Future of AI in Clinical Trials and Drug Development
- 18:26Conclusion: The Role of AI in Science and Medicine
Full Transcript — Download SRT & Markdown
Speaker A
MAX JADERBERG: Yeah, yeah.
Speaker A
HANNAH FRY: Do you think that's possible?
Speaker A
MAX JADERBERG: It's possible.
Speaker A
I think, it's possible. I think everything is pointing in that direction.
Speaker A
REBECCA HALL: Well, I've seen firsthand what can come out of this explosion of two fields coming together.
Speaker A
I think when you've got 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.
Speaker A
MAX JADERBERG: In five years' time, doing drug design without AI will be like doing any sort of science without maths.
Speaker A
And to be honest, I think the whole of science will be like this. It's like if you're not using AI, what are you doing?
Speaker A
HANNAH FRY: Welcome back to "Google DeepMind," the podcast. My name is Professor Hannah Fry.
Speaker A
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.
Speaker A
Now, everything comes down to proteins in the human body—how they fold, how they function.
Speaker A
But not very long ago, working out the structure for one single protein could take months, or even years.
Speaker A
And then, with the release of AlphaFold2, the algorithms developed at Google DeepMind, the entire field has been completely revolutionized.
Speaker A
More recently, AlphaFold3 can predict the structure of all of life's molecules with unprecedented accuracy, which turns out to be absolutely pivotal for drug design.
Speaker A
These developments have paved the way for a new company spun out of Google DeepMind.
Speaker A
It's called Isomorphic Labs, to represent the synergy between biology and AI. And joining me today are two of its most notable hires.
Speaker A
Rebecca Hall is Head of Medicinal Drug Design with years of experience in the process of discovering new drugs.
Speaker A
And Max Jaderberg 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.
Speaker A
Max, Rebecca, thank you so much for joining me.
Speaker A
REBECCA HALL: Thank you for having us.
Speaker A
MAX JADERBERG: Yes, a pleasure to be here.
Speaker A
HANNAH FRY: Well, it's a delight to have you because there's some big stuff happening.
Speaker A
I know a lot has been made of this claim that AI is going to be able to solve all diseases.
Speaker A
Is that realistic, Max?
Speaker A
MAX JADERBERG: This isn't going to happen overnight. Let's be clear.
Speaker A
But I think the exciting thing is that we can actually see there's, perhaps, a practical path towards that point.
Speaker A
And it's very different at this point in time than it has ever been, because we've got these AI machine-learning models that understand the biological world and the biochemical world in a completely different manner than what we've had before.
Speaker A
And so that's opening up tons of disease space that we didn't think was tractable before.
Speaker A
And that's just today. So as we start to develop these models further and further, it's really just the very beginning.
Speaker A
HANNAH FRY: Does that mean—I mean, every disease is on the table here, Becky.
Speaker A
REBECCA HALL: So I would say nothing is off the table at this point. And the journey for me here has been a big one.
Speaker A
I used to be much more conservative in this space. But having come to Isomorphic Labs, seeing 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 five, ten seconds, it's completely shifted my mindset. So now, I would put nothing off the table.
Speaker A
HANNAH FRY: And is it just in drug design, or is AI going to affect clinical trials, as well?
Speaker A
MAX JADERBERG: I think over time, yes, it will definitely affect clinical trials. We're focusing really heavily on the drug-design phase at the moment.
Speaker A
But you can imagine a world where, as you start to get better and better at drug design, actually, more and more of the bottleneck comes onto the clinical-development side of things.
Speaker A
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.
Speaker A
It's still 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, 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.
Speaker A
So that's not right now for us at Isomorphic, but it's very much in our future.
Speaker A
HANNAH FRY: What is the big idea then? What's the big ambition of Isomorphic?
Speaker A
MAX JADERBERG: It really is stepping towards that solvable-disease space.
Speaker A
And so this is really—before you go into clinical trial, before you even start testing out on people, you need something to test.
Speaker A
You need to create a molecule and a drug. And a drug, what is that?
Speaker A
It's something that goes in and modulates some function of the body, some function in a cell.
Speaker A
And so, really, the key first phase of Isomorphic Labs is, how can we create this AI drug-design engine that can take 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
Speaker A
and start modulating the function of these proteins, the function of cells, and then change the disease state for positive patients.
Speaker A
HANNAH FRY: The thing is, I mean, diseases have been cured in the past. We have come up with drug solutions that, effectively, make them a remnant of history.
Speaker A
Why are some diseases so much more difficult than others to solve or cure?
Speaker A
REBECCA HALL: So, for example, some types of cancers, you can develop a treatment to cure them because, for example, those cancers might be quite stable.
Speaker A
So there's maybe one mutation that drives that cancer. And then you treat that. You treat with that molecule that hits that particular mutation, and that can be curative in that disease.
Speaker A
But if you had a cancer which was continually evolving and 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.
Speaker A
So that would be then a much more difficult disease to solve.
Speaker A
HANNAH FRY: And, I guess, are there others where there's just a real gap in our understanding of what's going on in the body?
Speaker A
REBECCA HALL: Yeah, there's also that massive gap in understanding what's actually driving this disease.
Speaker A
Is it multiple things together? We don't have all the answers yet to biology. Biology is so, so complex.
Speaker A
One of the fundamental things we need to do is find out 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
Speaker A
an improvement in the symptoms, or a disease modification to that disease.
Speaker A
HANNAH FRY: Is that what you're trying to do then?
Speaker A
So actually, something like cancer, for instance, once you get down to the level of molecules and proteins, it's that something's gone wrong at that level, which then escalates into the scale of the human body.
Speaker A
REBECCA HALL: Absolutely. So you've got little—you can think of proteins as mini factories or engines inside your cell.
Speaker A
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,
Speaker A
or maybe it's mostly off. But now in a cancer cell, something's happened, it's always on.
Speaker A
It's going to just continually drive a signal. And that signal could be grow, proliferate.
Speaker A
And that will then drive the formation of a tumo
Speaker A
What makes this a well-suited problem for AI, Max? MAX JADERBERG: It's actually such a perfect application of AI and machine-learning, this whole playing LEGO with molecules.
Speaker A
We've seen over the last five, six, seven years the rise of models like AlphaFold.
Speaker A
We had AlphaFold1, AlphaFold2, understanding the structure of proteins. Before AlphaFold2, no one could really understand the structure without going into a lab and experimentally resolving these structures.
Speaker A
And that can take months, it can take years. Sometimes, it's not even possible for some proteins.
Speaker A
Obviously, AlphaFold2-- Nobel Prize winning breakthrough in chemistry. And now, we've taken that even further with things like AlphaFold3, where now we can understand the structure of proteins with small molecules.
Speaker A
And these small molecules are the little LEGO blocks that come in, and we use those drugs to inhibit the function of a protein, or change the function of a protein.
Speaker A
And the reason why this is such a good domain for machine-learning is that what we're trying to do, in essence, is predict the 3D coordinates of this biomolecular system.
Speaker A
And this fits really, really nicely into some of our classic, supervised learning, modeling domains.
Speaker A
It fits really, really nicely into our diffusion modeling frameworks that have been so, so successful for things like image generation, or video generation.
Speaker A
HANNAH FRY: And in terms of being suited for the supervised learning stuff that had already gone before, is that because there is some metric of success here, some way of being right, inverted commas.
Speaker A
MAX JADERBERG: There's a very clear metric of success here, which is super helpful for developing these models and for research.
Speaker A
And the reason is, over the last 50 years, people have been experimentally resolving these protein structures-- structure of these proteins with small molecules, with DNA, with RNA.
Speaker A
They've been doing it by hand in a lab, and then depositing the results into a big database called The Protein Data Bank, PDB.
Speaker A
And this gives a really rich source of information. It's a couple of 100,000 3D structures.
Speaker A
And each structure has thousands of atom coordinates, so there's a really high information density.
Speaker A
And that's a perfect scenario for supervised learning. Now, this isn't web-scale data, so it's not the scale of data that we might be used to for training large-language models.
Speaker A
But incredibly, we've worked out ways to design these neural-network architectures and these training regimes so that we can only train on a couple of 100,000 structures and excitingly get what we call generalization.
Speaker A
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.
Speaker A
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.
Speaker A
HANNAH FRY: But then, also, I guess, the possibilities of molecules that you could design-- I mean, it's-- well, a very big number, I imagine.
Speaker A
MAX JADERBERG: Yeah, that number is huge. People throw around numbers like 10 to the power of 60 is the possible number of drug-like molecules out there in the universe.
Speaker A
So it's a huge combinatorial problem. And that's also a really exciting spot for AI.
Speaker A
We can create great predictive models of how these molecules fit together, even how strongly they fit, or the properties of them.
Speaker A
But with the design space of 10 to the power of 60, we're reaching the level of atoms in the universe.
Speaker A
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.
Speaker A
So what do you do? OK, maybe you subsample that space and you search through some large number-- a million, 10 million, a billion, 10 billion.
Speaker A
HANNAH FRY: It's nothing, is it? MAX JADERBERG: Even though you're not-- yeah. HANNAH FRY: Yeah.
Speaker A
MAX JADERBERG: You're not even scratching the surface. And that's where 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
Speaker A
to explore across that whole space but without exhaustively searching the whole space. HANNAH FRY: OK, tell me about AlphaFold3 then.
Speaker A
In terms of designing drugs, what does it actually allow you to do? REBECCA HALL: So as a medicinal chemist, we always want to be able to visualize how our molecule binds to our protein, so how our LEGO block fits into the bigger
Speaker A
picture of LEGO blocks. HANNAH FRY: Thank you for going with me on this analogy, I appreciate it.
Speaker A
REBECCA HALL: The LEGO analogy. 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 which vector to explore.
Speaker A
You need to have some kind of target in mind. I'm going to explore that part of the pocket, or I'm going to that part of the protein structure.
Speaker A
That looks like a good place to go. And so for years, we've invested, as a scientific community, in ways to do that.
Speaker A
So X-ray crystallography is an experimental technique where you can actually go into a lab, and you can spend a lot of time crystallizing your protein.
Speaker A
You fire X-rays at it once it's bound to your small molecule. And then you can actually visualize, atom by atom, how your small molecule is binding to your protein.
Speaker A
Now we can do that with AlphaFold3, and the latest iterations of that model, into seconds.
Speaker A
And so something that might have taken me months when I was doing my PhD or in my early-stage research, I'm now just seeing on my screen all the time.
Speaker A
And so you can just iterate and iterate in silico until you get to something that actually looks really quite promising.
Speaker A
And then you take that into the lab. HANNAH FRY: So which way around does it work then?
Speaker A
Are you saying, OK, I think something like this, this, this, and this would work.
Speaker A
Let's try it out and see if it fits. Or is it the other way around?
Speaker A
Are these models telling you this is something that might fit? REBECCA HALL: So in the way that we've constructed our drug-design platform at ISO, you can do both.
Speaker A
So I can come in as an experienced medicinal chemist and I can say, I think this.
Speaker A
And I can test it then and there, a couple of minutes, and get that feedback.
Speaker A
But you can also take the opposite approach. I don't actually know what's going to work here so I'm going to apply the generative models we have.
Speaker A
I'm going to do some, what we call, virtual screening. So I'm going to take an area of chemical space that's commercially available.
Speaker A
I'm going to screen that against my protein. And I'm going to get the models to tell me what's best from that subset.
Speaker A
HANNAH FRY: Can I see what it looks like when you're actually designing something? REBECCA HALL: So the small molecule is fitting into this little groove in the protein, and it's forming interactions with the protein.
Speaker A
So the dotted lines you can see, those are interactions between that small molecule and the protein itself.
Speaker A
And as medicinal chemists, we want to optimize, or increase, the number of those interactions, because that's increasing the strength of that relationship between the small molecule and the protein.
Speaker A
HANNAH FRY: Let me describe what I've got, what's going on here. So you've got-- so the curly stuff is a protein.
Speaker A
REBECCA HALL: Yes, that's the protein, yeah. HANNAH FRY: And it's folded. So you've got the three-dimensional structure.
Speaker A
REBECCA HALL: Indeed. HANNAH FRY: And then over here, you've got-- I mean, this looks like the kind of thing you would do in GCSE chemistry, one of those kind of diagrams.
Speaker A
REBECCA HALL: Exactly. This is a small molecule. HANNAH FRY: This is sort of plugged into the protein?
Speaker A
REBECCA HALL: Yeah. HANNAH FRY: Wow. I mean, it really is like 3D jigsaws then.
Speaker A
I mean, that's-- REBECCA HALL: 3D, yes, exactly. HANNAH FRY: --what you're doing. REBECCA HALL: That's what we're doing, yeah.
Speaker A
HANNAH FRY: That crevice could be the thing that's making you feel pain, or the thing that's causing tumor growth, or whatever it might be.
Speaker A
REBECCA HALL: Yeah, causing your disease, yeah. HANNAH FRY: 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.
Speaker A
REBECCA HALL: Exactly. And I can quickly show you. So we have lots of different functionality on the platform that you can try.
Speaker A
And this is my favorite, which Max always tells me off about because I can use my expertise as a medicinal chemist.
Speaker A
And I can say, OK, I want to make some specific changes to this molecule.
Speaker A
HANNAH FRY: Oh, wow. REBECCA HALL: I can actually view what I'm doing in 3D.
Speaker A
So this is now going to fetch that structure prediction. And I can actually make modifications to this molecule.
Speaker A
And I can see the predicted structure in real-time. HANNAH FRY: And normally, this would have taken-- I mean, before AI-- a long time.
Speaker A
REBECCA HALL: I mean, if you're going to go into a lab and experimentally determine this, it could be anywhere from weeks to years.
Speaker A
HANNAH FRY: Why do you tell her off for this one? MAX JADERBERG: I guess you're referring to the fact that, over time, we want to do more and more from the model itself.
Speaker A
The really exciting frontier, from my perspective, is 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.
Speaker A
And then 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.
Speaker A
These are my design constraints. I want a molecule that does X, Y, Z and has these properties, and looks like this, and maybe interacts over there, and makes this sort of shape.
Speaker A
What can you come up with? And maybe set this running, go away, have a coffee, go home, come in the next morning and see what the agent has come up with.
Speaker A
HANNAH FRY: I guess with all of the projects that you've applied AI to generally in DeepMind, it has gone through that process of starting off with human expertise, and then slowly building in more knowledge and expertise within the model itself.
Speaker A
MAX JADERBERG: 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, and I've been working on them 10 years ago.
Speaker A
But they were rubbish. And they were spitting stuff out that looked like language. It kind of made sense.
Speaker A
But it also didn't make sense and you had to correct, and it clearly wasn't human.
Speaker A
And then, they got steadily better little by little. And suddenly, they just passed through this human-perceptible threshold where you can't really tell whether this is generated by a human or not.
Speaker A
And we're getting to the same point with our molecule-design models where, maybe five years ago in this field, you had generative models of molecules, and they'd spit stuff out.
Speaker A
But you'd give them to a chemist like Becky, and she would probably tear her hair out, like this is rubbish.
Speaker A
HANNAH FRY: Well, did you see those kinds of models? REBECCA HALL: Yes, I did, yeah, because I actually worked at an AI company prior to joining Isomorphic Lab, so I've seen that progression in that journey.
Speaker A
HANNAH FRY: And tell me, what kind of stuff did they spit out? REBECCA HALL: For a long time, it would just be nonsense because you're obviously giving the model an uphill function.
Speaker A
You want it to get to something that's going to bind really potently, for example.
Speaker A
But to do that, maybe it just makes the molecule massive. HANNAH FRY: Oh. REBECCA HALL: But then, well, it's not going to actually be absorbed through the intestine into the bloodstream because-- HANNAH FRY: Because you've got more to think about than just the molecule itself.
Speaker A
REBECCA HALL: Exactly, yeah. So there's a lot to piece together here. HANNAH FRY: Oh, that's interesting.
Speaker A
So AI, if we go back to our LEGO analogy, it was just like building a massive LEGO wall all around the protein.
Speaker A
REBECCA HALL: Yes, yeah. HANNAH FRY: I see. OK. REBECCA HALL: Essentially. HANNAH FRY: And now, this is-- have you seen that moment, that tip over with the language models that Max describes?
Speaker A
REBECCA HALL: Yeah, I think-- I've been so surprised by the quality of some of the molecules that come out of the generative AI.
Speaker A
And sometimes the molecules that come out, you think, oh, I would have-- why wouldn't I have come up with that?
Speaker A
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 do something else.
Speaker A
So-- HANNAH FRY: So it's working with you. REBECCA HALL: Yeah, you can work together.
Speaker A
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.
Speaker A
MAX JADERBERG: We've had some really fun moments where, for example, we've had our models submitting molecules blind, and then other people looking at them and seeing, OK, what are we going to send off for testing?
Speaker A
HANNAH FRY: Oh, really? MAX JADERBERG: And people look-- HANNAH FRY: Like a chewing test for molecules.
Speaker A
MAX JADERBERG: Exactly. And people looking at these molecules, 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 generative model instead.
Speaker A
HANNAH FRY: But OK, let me understand this, though, because using the analogy of large-language models, it makes sense there that you have these tokens, you break words down into little bytes, and then you can build up from there.
Speaker A
How do you do it in a way that makes sense chemically? I mean, you're not just taking atoms, are you?
Speaker A
MAX JADERBERG: We are, actually, just taking atoms. For bigger things like proteins, we chunk up into amino acids, so one token per amino acids.
Speaker A
So instead of characters of a sentence, letters of a sentence, we have amino acids of a protein.
Speaker A
And then for the small molecule, we chunk it up just into its individual atoms.
Speaker A
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.
Speaker A
And then we feed it through a structured model like AlphaFold3. And AlphaFold3 uses transformers.
Speaker A
But unlike in large-language models where transformers are used on one-dimensional sequences of characters, of letters, here, we use what we call a pair former, which operates on a two-dimensional interaction grid of all of these molecular elements.
Speaker A
So we can consider every single possible interaction that could occur between every amino asset, every part of the protein, and every atom of the small molecule, and everything in between.
Speaker A
And then this creates neural-network features, which condition a diffusion model. And diffusion models are generative models.
Speaker A
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.
Speaker A
HANNAH FRY: And it just so happens they work. MAX JADERBERG: And it just so happens this works phenomenally well.
Speaker A
You get these amazing structure predictions that-- when you go to the lab and experimentally resolve these structures-- and we do this on occasion.
Speaker A
Something amazing is predicted, like a completely new pocket or a new mechanism of action.
Speaker A
We go into the lab. We want to check that, are these models grounded at all in reality?
Speaker A
And what comes back is, yeah, this is within 1 angstrom, the tiniest unit of distance accurate, which is phenomenal.
Speaker A
HANNAH FRY: Yeah. But then through that training process, does it manage to extract a conceptual understanding of how chemistry works?
Speaker A
MAX JADERBERG: It's really hard to think about concepts in this atom space. But I do believe that there's some notion of reasoning in molecular and atomistic space that these models are doing because of the amount of generalization we're getting out of them.
Speaker A
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.
Speaker A
And so in our case, we give them all the molecules that might exist naturally, that people have worked out before, that people have designed before.
Speaker A
When your model gets better and better, you get things that look like they could have been designed before, which, then it starts to be imperceptible from a human design.
Speaker A
HANNAH FRY: What do you think? REBECCA HALL: You can certainly see that in the molecules that we get back.
Speaker A
They look like molecules that myself or someone else might have designed. You do sometimes get something crazy, though.
Speaker A
HANNAH FRY: Do you? REBECCA HALL: Yeah, we still do, but we have ways of filtering that out now.
Speaker A
HANNAH FRY: Is it like a hallucination, in a way? REBECCA HALL: Yeah. The model is really confident, but it's confidently wrong.
Speaker A
HANNAH FRY: Confidently wrong. OK. What does it look like when it hallucinates? REBECCA HALL: So we get back a number of structures that the models believe are good solutions for this problem, for this particular protein pocket.
Speaker A
And then our job is to go, OK, well, which of those molecules should we actually select to put into synthesis?
Speaker A
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, OK, of the solutions you've given me, where's the gold in that?
Speaker A
Where are the molecules that are actually going to push this project forward? And we'll put them into what we call chemical synthesis, where we make them and we test them.
Speaker A
But sometimes we get back results which, actually, the compound doesn't bind to the protein at all.
Speaker A
So here, the models, essentially, completely hallucinated a solution and so convincingly that actually, an expert looks at it and goes, yeah, that looks a really good solution.
Speaker A
The confidence metrics would suggest the same. So essentially, it is a hallucination. And I think we find it a really fascinating research question to say, OK, how do we find the really good stuff that the model is giving us?
Speaker A
HANNAH FRY: Beyond this 3D jigsaw or LEGOs-- we're mixing our metaphors. Is it just about structure?
Speaker A
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?
Speaker A
REBECCA HALL: There are so many other considerations. That's what makes this problem just incredibly complex.
Speaker A
So it's even beyond the shape. You can have something that maybe fits in the shape, but it's got to bind really strongly to that protein.
Speaker A
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.
Speaker A
The picture is a helpful guide, but you need to be able to predict that binding affinity separately.
Speaker A
And then all of those things together, that's just how your molecule binds to your protein.
Speaker A
You've also got to think about, is that molecule going to bind to any of the other 20,000 proteins in the body?
Speaker A
Because if the answer is yes, that could drive a side effect that you don't want.
Speaker A
That's going to drive you some toxicity. Is this molecule going to be stable? It's got to survive the really acidic conditions of the stomach.
Speaker A
It's got to survive going through the liver, which is going through a war zone.
Speaker A
The liver wants to do everything it can. It's like, this is a foreign molecule, I need to get rid of it.
Speaker A
So your molecule has got to be really robust. It's got to survive that journey.
Speaker A
It's got to be soluble. When you take a pill, that pills 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.
Speaker A
And a lot of these parameters are pulling against each other. So for something to be soluble and dissolve really well, it needs to be water loving.
Speaker A
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 water hating.
Speaker A
HANNAH FRY: Well, how do you possibly solve that if you need opposing characteristics? REBECCA HALL: So we've-- up until this point and still now, to some degree, drug discovery is a very iterative process.
Speaker A
So the human brain can only think about so many things at one time. So you're like, OK, I'm going to solve this binding-affinity problem a bit, and then, I'm going to start to think, OK, is my molecule soluble?
Speaker A
And then I'm going to start to bring in gradually these other properties. And honestly, it's like Whack-a-mole.
Speaker A
You play like this three-year game of Whack-a-mole where you're like, I fixed this problem.
Speaker A
Hooray. And then this other one pops up and you're like, OK, I'll fix that, but then the other one's gone bad, again.
Speaker A
So to be able to predict all these things in silico is still a really, really hard problem.
Speaker A
HANNAH FRY: Are you working on tools that will help with those elements, too? MAX JADERBERG: Yeah, absolutely.
Speaker A
So we think about, ISO, how do we do drug design end-to-end, which really means solving all of these very, very hard problems with cell permeability, solubility, toxicity liver clearance everything.
Speaker A
And none of this is solved. These are really hard problems to even model. So we spend a lot of effort, a lot of research, on creating new models to really understand this better.
Speaker A
And then as Becky was talking about, how do we then start to find these needles in a haystack molecules that are somehow just balancing the properties just right to be a perfect drug?
Speaker A
And it's really, really hard. And actually, there are a lot of analogies to maybe what I used to do at DeepMind in, for example, "Capture the Flag" or "StarCraft" is there's not just one agent, or strategy that solves "StarCraf"
Speaker A
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.
Speaker A
In the same way, we need to be searching this huge combinatorial-molecule space. So just like you might have tree search in a game of "Go," where at every move, you elucidate some other possible moves, and you start searching through that tree of possible
Speaker A
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.
Speaker A
So you start with a part of a molecule and you start to hypothesize, what are the different things I could add or take away from this molecule?
Speaker A
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.
Speaker A
HANNAH FRY: But how do you even know that the perfect molecule exists? Maybe there's just some crevices in the proteins that just are unfittable.
Speaker A
REBECCA HALL: We do have this concept of undruggable proteins where the crevice is really flat, and you can't really get anything to grip in there.
Speaker A
And those proteins might need different solutions. So actually, we have a whole emerging field which we call molecular glues.
Speaker A
And this is where you have two proteins that come together, and the pocket that's formed when they come together is actually a much more suitable pocket.
Speaker A
So now, you need to design a molecule that sits in the middle of them and glues them together.
Speaker A
So there's this whole explosion of all these different modalities now, which makes this an incredibly exciting field to work in.
Speaker A
MAX JADERBERG: From my perspective, the fact that we've actually found any drugs at all already, given how hard and complex the problem is-- and we've basically been doing a bit of human intuition and a lot of random screening and experimental testing.
Speaker A
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.e., there's probably lots of different solutions that could work,
Speaker A
but we've just got to find them. HANNAH FRY: Have you actually tried to make any of these molecule, or at the moment, do they just exist on the screen?
Speaker A
REBECCA HALL: Oh, no, we make a lot of molecules. We've got a huge experimental footprint.
Speaker A
Yeah. HANNAH FRY: And how-- do they turn out how you expect? I mean, what are the results like?
Speaker A
REBECCA HALL: We've had some incredible success in some of our projects where I'll find Max at his desk and I'll be like, have you seen this thing?
Speaker A
And we'll just both be really mind-blown about it. HANNAH FRY: That when you get the molecule, it actually works.
Speaker A
MAX JADERBERG: Yeah, exactly, exactly. And we have our own drug-design programs, so things that we've started from scratch ourselves.
Speaker A
We also work with pharma-company partners, people like Eli Lilly and Novartis. In these collaborations, you'll get specific targets to work on.
Speaker A
And these are ones that these companies have high conviction behind and probably a bunch of evidence behind.
Speaker A
Some of the collaborations we're in, we've been given very, very hard targets. These are things that people have worked on sometimes for over a decade and not made significant progress, to the point where you've got something on the market.
Speaker A
HANNAH FRY: Things like cancer and that sort of stuff. MAX JADERBERG: Whole host of therapeutic areas and disease areas.
Speaker A
And then Becky and team sit down, start designing with these models, and can start finding completely novel chemical matter for completely novel mechanisms that no one's really discovered before, which is mind-blowing for me-- HANNAH FRY: Yeah.
Speaker A
MAX JADERBERG: --as a computer scientist. REBECCA HALL: Yeah, it's mind-blowing for me. There's been some almost career-defining moments, where the AI will give you a hypothesis.
Speaker A
It will suggest something. And you think, I'm not convinced I would do that, but the model's telling me this thing.
Speaker A
And it's really quite convinced about this thing, so I should maybe just test this hypothesis.
Speaker A
And then, actually, it turns out that the model was right, and you were absolutely right to test it, and it's really pushed forward your project, or even that field.
Speaker A
So 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 and not going, oh, that doesn't fit with my worldview,
Speaker A
so I'm not going to test it. HANNAH FRY: But then you are also human, right?
Speaker A
REBECCA HALL: Yes. HANNAH FRY: So I do wonder whether if you see lots of hits with the model, as it were.
Speaker A
If the model is coming up with lots of good stuff in a row, do you start to maybe trusting it more than yourself?
Speaker A
REBECCA HALL: 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 cutoffs.
Speaker A
HANNAH FRY: What kind of cutoff? REBECCA HALL: For example, we have a model which we call binding probability, and it goes from 0 to 1.
Speaker A
So 1 is the model is convinced your molecule is definitely going to bind to your protein.
Speaker A
Zero, the model is telling you, this is definitely not going to bind. And when you can build a little bit of confidence over time that the model really does understand, OK, anything below 0.7, it's really got a very low probability
Speaker A
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 is quite likely to be right.
Speaker A
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 why doesn't it like it?
Speaker A
HANNAH FRY: But then what if it makes something that you don't understand, I mean, or that doesn't make sense?
Speaker A
I mean, does it need to explain itself? REBECCA HALL: I think at the moment that explainability is quite important for now, because the process is quite driven still by the human.
Speaker A
It's not end-to-end yet. We have to go in there, and we have to say, what comes next?
Speaker A
So if there's no explainability there, you don't know what would be next, right? And so that would be very difficult to work with.
Speaker A
I can imagine in a future state where, actually, the process is a bit more end-to-end, like in one step, the model.
Speaker A
Here's a drug. HANNAH FRY: Here's a drug. REBECCA HALL: Yeah. Then actually, maybe you don't need that explainability.
Speaker A
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 directionality, then that explainability is important.
Speaker A
And that's where I think, for me, the AlphaFold models really come in because, OK, the model is predicting this molecule is going to be good.
Speaker A
I can rationalize that with what I'm actually seeing. I know what I would do next.
Speaker A
HANNAH FRY: You have a slightly different view on explainability, don't you? REBECCA HALL: I do have a slightly different view on explainability, but I think you need explainability when your model sucks, basically.
Speaker A
And we don't have perfect models yet, so I think there's a good amount of room for explainability.
Speaker A
But I always hear the call for explainability and think, look, we need to make this model better.
Speaker A
And actually, the interesting thing about explainability is it can help you understand the pathologies that this model has, the biases that it has.
Speaker A
Where's that wrong, 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 do a final round of verification
Speaker A
in the lab at the end. HANNAH FRY: So, I mean, you literally say, make me a drug for X disease, off it goes, says here's the molecule you need.
Speaker A
MAX JADERBERG: Yeah, yeah. HANNAH FRY: Do you think that's possible? MAX JADERBERG: I think it's possible.
Speaker A
I think it's possible. I think everything is pointing in that direction. We're getting better and better.
Speaker A
We're already reducing the amount of experimental cycles you need, reducing the amount of lab time you need.
Speaker A
And, yeah, this is just the beginning. HANNAH FRY: Absolutely extraordinary. I mean, I suppose it does depend on knowing what protein you're targeting too right?
Speaker A
REBECCA HALL: Yes. HANNAH FRY: So still-- and the diseases that we don't have a full understanding are still going to be difficult.
Speaker A
REBECCA HALL: Yeah. And that kind of-- we call it target ID space, where you actually need to identify the protein that's causing your disease.
Speaker A
It's actually a really important part of drug discovery because if you're not hitting the right biological target from the start, you can design the best molecule in the world.
Speaker A
It's not going to do what you want it to do when you put it into a human.
Speaker A
So there's a lot to be done in that target ID space. And I think AI's got a big role to play there, as well.
Speaker A
MAX JADERBERG: It's one of the big frontiers of AI for biology is really understanding, what are those driving mechanisms of disease?
Speaker A
Can we start to understand how mutations in our DNA translate into changes of expression of RNA, and how that changes the type of proteins and expression levels of proteins, how those proteins interact with each other and build up into these signaling pathways,
Speaker A
and how changes in those signaling pathways change the disease state, as well? And, of course, if we can start to understand these bits, we can start to work out, where do we need to modulate this biological system?
Speaker A
But all of this is really, really hard. And there's some amazing breakthroughs happening in the field understanding DNA better, understanding this translation better.
Speaker A
Even through understanding how proteins interact, can we build up these interaction networks better? This is some of the really exciting frontier research that we're also doing at ISO.
Speaker A
HANNAH FRY: So you have a team working in that space, as well. MAX JADERBERG: Yeah, that's right.
Speaker A
We have a whole computational biology team, whole machine-learning modeling team focused in this space, yeah.
Speaker A
HANNAH FRY: But then what about personalized medicine? Because I guess each person is different in some ways.
Speaker A
MAX JADERBERG: I mean, this is the really exciting, potential future where we can understand much more about, for example, cancer individuals' mutations in their tumor, and through generative AI and design agents, be able to come up with molecules that work specifically
Speaker A
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?
Speaker A
But we're moving towards a place where that technology could be potentially there. HANNAH FRY: I mean, I'm thinking here about chemotherapy drugs, which come with really devastating side effects.
Speaker A
You think there's real hope on the horizon for that kind of thing? REBECCA HALL: Yeah, I mean, we think about chemotherapy drugs.
Speaker A
They're basically drugs that are nonspecific. So they're going into the body, and they're trying to halt that rapid cell proliferation.
Speaker A
But 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?
Speaker A
And that might have the same effect. But you're not just generally using something that's just very toxic to rapidly dividing cells.
Speaker A
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.
Speaker A
It's going to stop your hair follicles. You're going to lose your hair, whereas, we now know we can go in, we can target a very specific protein, the one that's actually causing the disease.
Speaker A
And if you inhibit that particular protein, that's not going to cause-- hopefully, if you get it right, it's not going to cause all these other side effects.
Speaker A
MAX JADERBERG: You can do some also really, really cool stuff with targeting particular cells.
Speaker A
So if you know that a particular cell type is expressing something on its surface, you can start programming things like antibodies to come in and find those particular receptors.
Speaker A
And so you're delivering your payloads directly to that particular cell type and not more broadly to the body.
Speaker A
HANNAH FRY: I just want to go back to the point that you made earlier, Becky, about, once you've got the drug design, then once you put it into the human, there's all of these other potential problems because, I mean,
Speaker A
there have been examples of this before where drugs have been made and looked like they were very good.
Speaker A
And then once you actually put it into a human, it causes some massive problems.
Speaker A
I think there was one which 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.
Speaker A
How do you mitigate against that, or can you not at this stage? REBECCA HALL: Well, one of the problems we have is that we often use animal models to then translate things into the clinic.
Speaker A
And animal models, they don't replicate human physiology very well at all, actually. So we know when we're working in the discovery and preclinical space, which is all of that space before you go into a human, we're working with different animal models, which
Speaker A
might model the disease we're interested in. And we're looking for molecules which have an effect in those animal models.
Speaker A
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.
Speaker A
But from that point until the market, there's a 90% failure rate. HANNAH FRY: Wow, 90%.
Speaker A
REBECCA HALL: So all that investment up to that point, which is huge. HANNAH FRY: What makes them so likely to fail?
Speaker A
REBECCA HALL: So molecules fail in the clinic for toxicity. They fail in the clinic for lack of efficacy.
Speaker A
And I think a lot of it comes back to the animal models we use just are not very good at replicating human physiology.
Speaker A
HANNAH FRY: Because the mouse is different to a human. REBECCA HALL: A mouse is different.
Speaker A
So we can cure a mouse disease, probably be quite good at that. HANNAH FRY: We've got loads of medication that works.
Speaker A
REBECCA HALL: Yeah. HANNAH FRY: Yeah. REBECCA HALL: But, yeah, that translation is a big part of science that we need to fix.
Speaker A
HANNAH FRY: Can AI help here, as well? I mean, if animal models are this stumbling block with such a low level of success, what can you do about it?
Speaker A
MAX JADERBERG: Well, this is where we can actually use some of the technology and models we've been developing and think about, OK, how can we understand toxicity better, understand the effect on human cells better, and see how that translates to organs?
Speaker A
If you think about some of these off-target effects, probably, there are many drugs that 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
Speaker A
and stopping the function of your heart. That's an off-target effect. HANNAH FRY: I mean, side effects, in general, are off-target effects, though, aren't they?
Speaker A
MAX JADERBERG: 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
Speaker A
in the human body, all 20,000 proteins? And you can start building up this fingerprint of interactions that this molecule, your drug molecule, is having across the body.
Speaker A
And so that can give you clues, maybe even concrete signal, into the toxicity or side effects of this molecule.
Speaker A
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.
Speaker A
So by the time you've gone through all of your molecule design and you get to the point where you're like, yeah, I want to go into humans, you've been thinking about these side effects in a very rational way
Speaker A
for a long time. And so, hopefully, your chances of actually hitting some of those radically reduces.
Speaker A
HANNAH FRY: You're taking your structure of LEGO bricks, or your jigsaw, and you're just trying it with every other possible combination that it might encounter in a human body.
Speaker A
MAX JADERBERG: Yeah. HANNAH FRY: That's amazing. MAX JADERBERG: We're going to make every possible LEGO combination.
Speaker A
HANNAH FRY: Well, if that's the design stage, then, Becky-- I mean, you also have to put this into clinical trials.
Speaker A
Just talk us through the process of clinical trials, if you could. REBECCA HALL: So the first time that your molecule ever goes into a human, that's a phase I clinical trial.
Speaker A
So it will be a small number of patients. Some of those might actually be healthy volunteers.
Speaker A
They don't necessarily have the disease that you're interested in. And what you're looking to see is, does your drug actually reach the level of exposure in the patient that would be needed to generate an effect?
Speaker A
And is the drug well tolerated, or do you suddenly start to see some side effects that you weren't anticipating?
Speaker A
If all is good, you'll proceed to a phase II clinical trial, which is now you're going into people who actually have the disease, and you're going into larger numbers.
Speaker A
You're really looking to answer the question, does your molecule actually have efficacy against the disease that you're interested in?
Speaker A
And this is where we do see that big failure rate. So 70% of molecules going into phase II don't actually pass through into phase III.
Speaker A
For those that do pass into phase III, that's where you're going into much bigger patient populations seeing if your drug is effective across that bigger population.
Speaker A
It's not just got to be safe, but it's got to be better than the standard of care.
Speaker A
There's got to be some-- 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.
Speaker A
HANNAH FRY: And this whole thing has a 90% failure rate, as you said. REBECCA HALL: Yeah.
Speaker A
HANNAH FRY: I mean, does that mean that there are people who work in this space who never, never succeed?
Speaker A
REBECCA HALL: Yeah. So I'm a medicinal chemist, and we often have this number where, actually, only 1 in 20 medicinal chemists will ever get a drug to market.
Speaker A
So 19 of us out of every 20 will never get a drug onto the market through our careers.
Speaker A
So, yeah, we are a profession where we're used to seeing significant failure. HANNAH FRY: You're comfortable with failure as a profession.
Speaker A
REBECCA HALL: We're comfortable with failure. We learn from it. HANNAH FRY: Extraordinary, extraordinary to imagine.
Speaker A
How long do you think it will be until the first AI-design drug is on the market?
Speaker A
Because all of these additional levels really take some time, don't they? REBECCA HALL: So those AI-designed drugs that are in the clinic now in clinical trials and the different levels of AI input into those current drugs, I would imagine that in the next five years
Speaker A
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
Speaker A
and really start to deliver molecules for patients? That, for me, will be when AI is having a really big impact.
Speaker A
HANNAH FRY: When you can start to say, here's the target, and then it pops out a drug at the end.
Speaker A
REBECCA HALL: Yeah, and you can put that straight into the clinic. HANNAH FRY: And be confident that it's not going to cause any damage to a person.
Speaker A
REBECCA HALL: And even a slightly improved level of confidence in where we are now would be quite impactful.
Speaker A
Yeah. MAX JADERBERG: 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.
Speaker A
We're just going to see more and more of that. In five years' time doing drug design without AI will be like doing any sort of science without maths.
Speaker A
And to be honest, I think the whole of science will be like this is, if you're not using AI, what are you doing?
Speaker A
There's just so much information to be gained there. So yeah, as Becky said, it's more like, how do we actually see that rapid increase in disease areas that we're able to tackle the targets, that we're able to unlock, ultimately, patients
Speaker A
that we're able to help. HANNAH FRY: Amazing. Thank you both. That was really interesting.
Speaker A
REBECCA HALL: Thank you for having us. It was so much fun. MAX JADERBERG: Yeah, it's been great to be here.
Speaker A
HANNAH FRY: I think I now realize that medicinal chemistry is one of the hardest jobs in the world.
Speaker A
It takes years to design a drug. Even if you get it to clinical trials, 90% of them fail.
Speaker A
And only one in 20 of your colleagues ever manages to see their medicine improving the lives of patients.
Speaker A
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
Speaker A
of possibilities, it's like someone has just turned on a floodlight. And, OK, of course, we are still very, very far away from a big AI button that's just going to solve all diseases.
Speaker A
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.
Speaker A
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.
Speaker A
And, of course, we have plenty more episodes on a whole range of topics to come, so do check those out.
Speaker A
See you next time. [MUSIC PLAYING]
Topics:AI drug designprotein foldingAlphaFold3Isomorphic LabsGoogle DeepMinddrug discoverymachine learningclinical trialsmolecular modelingcancer treatment
