Can AI Really Design New Drugs? Google DeepMind Spin-out Isomorphic Labs Explains

0:00Drug discovery is a messy problem. So it can take upwards of a decade to get a drug, all the way from the initial stages of a project, all the way through onto the market. And we see a massive failure rate, even in the clinic. The whole process can cost upwards of something like$3 billion. It's not a single problem. It's not even just a few problems. It's so, so many different capabilities that you actually need to develop. So it's kind of a Manhattan Project-style effort. If you get the single atom wrong that's actually responsible for a key interaction, there's just kind of no escape from that.

0:35This is my dream. One iteration you can get to a drug candidate. So one design cycle. Maybe you make a couple of hundred molecules in that design cycle, but in there you've managed to make such accurate inference across all of your models that you've got a molecule that's suitable as a candidate. Welcome, humans, to the Neuron AI Explained. I'm your host, Corey Knowles. Today, armed with curiosity, caffeine, and a legally non-binding opinion about spreadsheets is our own Grant Harvey. How are you today, Grant?

1:05Corey Noles:Legally non-binding opinions on spreadsheets. What is a legally binding opinion on spreadsheets? I really don't know. I really don't have an answer. I'm not sure what one would be, but I'll bet it's out there and someone will tell us in the comments. How are you, Corey? I'm doing fantastic. Doing good, man. Doing good. Really excited about this call today. Who are we talking to? Yeah, today we are talking about one of the most consequential questions in AI. Can foundation models move from predicting biological structures to actually helping create drugs? So our guests are Becky Paul, who leads medicinal drug design at Isomorphic Labs, and Michael Schwarzschmidt, who leads foundational AI research there.

1:42Corey Noles:We'll unpack what AlphaFold made possible, why drug discovery is still so hard, what Isomorphic Labs drug design engine is trying to do, and where the line is between real scientific acceleration of drug discovery and the AI hype cycle. Becky, Michael, welcome to the Neuron. We're so excited to have you. Thank you. It's great to be here. Great to be here. Awesome. Well, I guess to get started, let's say for listeners who know AlphaFold, but not the messy reality of drug discovery, what problem is isomorphic labs trying to solve? So drug discovery is, as you describe it really well, It's a messy problem.

2:18So it can take upwards of a decade to get a drug all the way from the initial stages of a project all the way through onto the market. We see a massive failure rate even in the clinic. 90 % of compounds that enter the clinic don't come out the other side. So a huge failure rate and the whole process can cost upwards of something like three billion. So a huge expense to get new drugs to patients. So we really need something disruptive to kind of change the way that we do drug discovery.

2:48Corey Noles:What would you say that the biggest misconception people have is when they hear AI design drugs? Like, what do they get wrong about that? So at Isomorphic Labs, we are developing a number of foundational models. So Michael will tell you a lot about this as well. And you saw a bit of this in our ISO DDE report. and we're using those to help us drive an AI-first approach to drug design. There's a lot to unpack within a sort of an AI-designed drug. What's the contribution of AI to that drug? Is it wholly designed? How much human in the loop? How are the models used? So maybe we can kind of walk through the process like piece by piece and we can talk about what that might actually mean and what actually is the process of discovering a drug all the way from beginning to end.

3:35That would be very interesting. Yeah. Yeah. Right at the beginning of drug discovery, you need to form some kind of biological hypothesis. So what biological target do I need to modulate and in which patient population to actually impact patients in clinical in the clinic? That's kind of the first phase. And that's already something that's really quite difficult to do. The next piece is, OK, I know which biological target or collection of biological targets that I need to modulate. Actually, how am I going to do that? Which kind of chemical modality do I need? Do I need a small molecule, an antibody, maybe a peptide?

4:12And how am I going to develop that chemical matter to modulate this target in the right way? And then the final piece is in the clinic. How do I make sure that the right molecules are getting into the right patients in the right clinical trial, demonstrating efficacy so that that drug and safety, so that that drug can get all the way onto the market? And we think at isomorphic labs that AI is poised to make an impact across this whole process. But there's lots of detail within each of these phases. I'm just sort of thinking about, you know, what an AI drug, design drug actually looks like. And maybe just to build on that quickly on this misconception point.

4:51And I think you're getting this from Becky, right? It's not a single problem. It's not even just a few problems. It's so, so many different capabilities that you actually need to develop. And so I sometimes get this question of, oh, how can I help cure cancer with whatever I'm doing, right? But there's just so overwhelmingly many capabilities. So it's kind of a Manhattan Project-style effort. If you really want to take a look at every kind of piece of that, rethink, okay, how would you do this with modern data-driven AI approaches? And yeah, so kind of getting beyond this, oh, it feels so overwhelming.

5:29there are so many different things that can go wrong and finding root node problems where if you can make meaningful progress, you can then kind of see the impact across many different drug design programs. So that's something that we are very much focusing on, especially in the early years of isomorphic labs. Okay, that makes a lot of sense. Let's say from the medical chemistry side, what does AI need to get right before a human drug designer is going to feel comfortable with this and be excited about what it has to bring to the table. So a number of things that as a human drug designer, you kind of need to optimize for as you go through that optimization phase.

6:08So you're going to start from something that, maybe you've got something that binds very weakly to a protein, but you need to take that all the way through from that to something which binds really potently, which is going to, you can take as a tablet, which is going to survive the very acidic conditions in your stomach. It's going to be soluble. all of these kinds of problems, you're going to want to understand how that molecule actually binds to the protein, which pocket is that molecule binding to, which interactions are important, and how might I build on those to make this molecule bind more potently to this protein.

6:43So a huge number of models, as Michael says, that need to kind of come together for you to be able to make meaningful predictions about a molecule and to drive it forward towards sort of a steady path of optimization through to a drug candidate. So as you probably saw in the ISO-DDE technical report, one of those really foundational models is a structure prediction. So being able to predict how a molecule binds to a protein enables us in sort of real time, by real time, I mean, you know, maybe five seconds, maybe a minute, to be able to visualize how a molecule binds to a protein. And that's something that before we had these capabilities, kind of co-folding capabilities, we'd need to actually define experimentally.

7:27And that sometimes takes someone's entire PhD to be able to get that information experimentally. So that would be an X-ray crystal structure. And that would allow you to visualize how the molecule binds to the protein. And what we see at isomorphic labs is that these structure prediction models, in some instances, have been so predictive that when we've actually invested in obtaining a crystal structure, it's been identical to the original co-fold so then you sort of start to think is this worth the investment at this point or actually should we just really begin to trust this model and build on it seeing if we can push forward this project using the information that it's giving us

8:05Corey Noles:and so so basically what you're saying is at this point you're sort of like it's really good i think we do need to just trust it and that will save us a lot of time here you sort of like giving up and giving over to the machine a little bit, it sounds like. Yeah, exactly. And that sort of is a level of trust that people have to build a little bit as humans. Themselves, even, in a lot of ways. Exactly. And because once you can trust what you're seeing, then that allows you to use that as a hypothesis and a solid foundation, and then you can build on it. And you can use it to answer your next scientific question.

8:43And maybe you don't need that warm, fuzzy feeling of the experimental validation so much anymore because you've seen this model able to make very accurate predictions many times. You've validated them. It's making prospective predictions. And so therefore you can develop that trust in the model. And interestingly, maybe on that one, the models themselves also have very well calibrated error predictions, right? So if the structure model tells you that it feels very highly confident, it most likely is. And if it tells you, I just don't know what this is, it is most likely going to be garbage. So yeah.

9:18That's really helpful. That's like kind of like confidence scoring. Yeah, exactly. Exactly. And then you get into this world of, okay, you have the confidence scores, but then you also have, let's say your binding model and you begin to integrate a lot of these different signals and creative ways, right? You might have a generative model that creates a new molecule and then you have a better structure model with a better confidence model that tells you how confident am I in this generation.

9:42Corey Noles:The binding model that tells you okay, how tightly does this bind? Do these models agree? What about the prior version of that model? What about scaling the inference on that model? And if you run many samples, so you kind of get to be able to put together quite sophisticated hypotheses. It's not really, you call the model, you look at the output, but yeah, you generate quite complex ideas with it. That makes sense. Michael, from like the AI side, What makes working with biology, chemistry so different from training models on like text, images or code, like what we're traditionally looking at with AI?

10:21Yeah. So first, the data is incredibly messy because the raw ground truth data are these ultimately experimental artifacts. Right. But even the data that you're looking at from, let's say, an x-ray cosmography, right, is actually it's a modeling artifact created based on x-rays. So even what you might see as the ground truth data for a structure model itself often has flaws and then models have to learn to deal with that. So on the one hand, there is the actual just kind of the processing of the data. Then obviously the cost of generating new ground truth data, knowing even what data you might want to be generating.

11:00So you spend a lot of time on that. The other part is like if you're thinking about this whole structure biology at a model space, right? Think what you're trying to do. you're trying to place individual atoms. So the models are incredibly sensitive to every tiny detail. And if you think about this kind of from a model perspective versus LLMs, first you have obviously tons of data and you have very large models. And if you have tons of data, even if a lot of data is garbage or noisy and you have a ton of capacity in the model, then you ultimately, you just kind of learn to smooth over that. And obviously, if you slightly rearrange the grammar somewhere and you get to kind of, nowadays, you have your reasoning trace in your LM, so you kind of get to re-inspect.

11:44You can ultimately correct a lot. But if you get the single atom wrong that's actually responsible for a key interaction, there's just kind of no escape from that. So I think you need to develop an incredible level of patience with the details of the evaluation and data. And maybe like an interesting point to think about on that is in the LM world, you see this incredible diffusion of capabilities very quickly, right? Like one lab releases something and then like a month later, another lab has kind of reverse engineered it and figured it out. So it's kind of moving very, very quickly, which is kind of great as the users of these models and kind of very exciting.

12:26But then if you look at the, let's say, Alt Fold 3, that took almost kind of two years to like fully, fully reproduce or nearly reproduce kind of open source. Even though you have a really long paper, you have open source code and inference code, you have all the algorithms listed. It's just like so many details that even if you have all of that, you can still struggle to reproduce important aspects of it. So it's very fascinating to work with. Yeah, it's amazing to be honest. I bet you've learned a lot of lessons along the way in the process, too, as far as like, I can see that where there had to be some amount of trial and error in here as you're learning and finding your way through to navigate this.

13:05Any big ones that are memorable? Lots, but I have to think about which ones I can speak about. Oh, that's fair. I think for us, if you think about this whole timeline of drug design and what Becky laid out, right? So you're going to make decisions today that the ultimate kind of drug outcome you're going to see really, really far downstream. But you still have to ground your modeling process in something that you can actually tie to progress on the things that you care about on your drug design programs, right? So you can make a lot of progress. That's maybe the lesson. It's very easy to make progress on some even scientifically very interesting metrics and benchmarks.

13:46And you can feel the curve is going up. But if you are not seeing the impact on your drug design programs, what are you doing?

13:55Corey Noles:It's kind of like the equivalent of not seeing the productivity data show up in GDP or whatever, right? Everyone's like, where is it? Yes. I think for us, this is just really, really important that we have the drug designers in-house. And we are trying to sell a software where we say, oh, look at how amazing we are. The benchmarks kind of buy this model. We ultimately have to make it work on the programs. And what we are seeing excitingly is that some of these core foundational model metrics, as we release better models, and maybe Becky kind of can speak to that over the past few years, you actually see new things working every time.

14:34And that then kind of gives us the confidence to keep making these investments in these very expensive foundation models. Yeah, maybe I could add to that. It's incredible for us on the drug design side. Because obviously we're in the same building as this like world-class ML team. And they're like, you know, fighting for every performance improvement on all of these models. And then we get this new model released. And let's say suddenly it just unlocks this kind of new kind of project, for example, where maybe the previous model could kind of make some predictions, but it was getting some element wrong and suddenly it's all fallen into place with this new model.

15:10I've never worked somewhere where the pace of innovation is so fast. It's really, it's very exciting. The AI space has a way of being that way. Yeah. Can you keep up?

15:21Corey Noles:Grant knows that very well. Well, especially to Michael's point with all of the complicated details that you all are working with, Like to try and keep a mental model of all of it in your head. I find that that's probably the hardest part beyond the actual like science and, you know, making that work. Yeah. So I think it's also you kind of you have to commit to that, right? Like even to like really get productive. So we actually on the AI side, we hire mostly deep learning journalists, right? Like we typically we don't hire a lot of people with, let's say, a specific biology or chemistry background because obviously it's welcome.

15:56but just having these general kind of deep learning tools that are then kind of applicable to a lot of future problems that we have to solve. But then you have to actually commit to learn for a very long time to begin picking up these details. But I think people are really nowadays like more excited to commit to these like hard tech long-term missions, right? Like you're almost seeing some people burned out a bit by this LM lab race and school releases a lot faster. And then kind of to think about this fact that there are these problems, if we don't solve them in 50 years, 100 years, drugs will still take a decade to market.

16:34Right. And even though you might spend a year of your life grinding out like a fraction of that percentage point on the formation model, as you stack these up, we only have to solve that problem once, right? Like as you met here, to be able to do certain things at experimental accuracy. And then you can bank that forever. Yeah, exactly. So that's just amazing, right? It is. That is cool.

16:58Corey Noles:Well, so I've always wondered this, and you might be surprised to hear this, but I am not a biologist. Shocking, I know. But I've always wondered, so there's protein folding, but what is the next hardest challenge related to that for drug discovery? What else in our biology do we need to model in order to be able to predict reactions? Are we going bigger, as in we need to model how cells interact with each other? or is there other mechanisms we need to go smaller? What's the right next challenge to solve there? Oh, go for it, Michael. Maybe I can start and you can add in. I think in a sense, it's both.

17:36You still need to make a lot of progress at the atom level modeling, and then you need to integrate a lot of other kind of biological scale. And so on the first one, there's proteins. There's kind of this alpha-fault-2 moment. can you kind of get a crystal structure of most proteins with experimental accuracy. Then there is the kind of alpha-fold 3 moment. Okay, for drug design, you don't care just about this protein. You care about can you co-fold other molecules onto that, let's say small molecules. And then other atom-level problems, such as the kind of binding affinity, how tightly do two molecules bind to each other to ultimately be able to improve the potency?

18:20There's still a lot of work to be done there to get this to full experimental accuracy. And then nowadays, there's a lot of exciting therapeutic modalities that look at multiple molecules involved. So even at this atom level space, you see still quite a lot happening. And then there is the other kind of levels of abstraction where you think about cell level, organ level, toxicology, all of these things.

18:45Corey Noles:Does that get harder and harder and harder as you try to model that because of all of the complexities and ways that it can? Does the complexity scale exponentially, I guess, is what I'm wondering. I think you end up using slightly different modeling paradigms. I would say all of them are extremely challenging. It's probably the... Yeah. Maybe Becky, you want to add on to the different scales. Yeah, I mean, I think what you said about the emerging sort of therapeutic modalities is really interesting because traditionally we might have thought of drugs as small molecules that bind to maybe an enzyme, block the function of that enzyme.

19:22And that kind of is how they enact their biological effect. Whereas now we're in this really exciting paradigm of like all these new modalities, you know, people will talk about molecular glues. These are small molecules which maybe glue two proteins together and that maybe blocks the function of the protein that you're trying to inhibit. Or maybe you glue two molecules together, but one of them actually degrades the other and actually makes it disappear from the cell entirely. So this is like a whole new frontier for us that opens up so much possibility for the design of new drugs, new modalities.

19:57and these are things we can start to model because we understand how proteins interact whether there's a pocket at the interface is that pocket something that's induced by a ligand or can we predict that happening can we predict how something might be degraded or blocked so it's kind of opening up this whole whole new world for us i'll bet you're learning new stuff every day right now aren't you yeah right yeah 100 yeah maybe what's kind of interesting is so this is what we have released in this technical report where it's a lot about this step change progress in generalization. There are a lot of these very, very small scale subtle effects that Becky mentioned.

20:38So there's induced fit, opening a cryptic pocket. So suddenly you're targeting something somewhere that you didn't necessarily expect, but the model tells you to. So very subtle changes or kind of conformational changes where people might have thought you need kind of a whole different class of modeling or it's just kind of completely still out of reach another big change and you see that emerging basically and then the other one is that we are very much focused on this general foundational modeling paradigm and so we see playing out this generalization not just on small molecules but let's say also on antibodies and so So if you are just very, very good at atom level modeling, then all of these therapeutic modalities, it's ultimately you're, yeah, you're modeling atoms interacting with each other, right?

21:28Rather than trying to be like very specialized, I have my fine-tuned model for this target, this therapeutic modality.

21:35Corey Noles:Right. You can do a lot more if you have a general model because you can, yeah. Yeah. And that's maybe slightly counterintrusive against, I would say, some common narratives around, or there's just like not enough data or you just, yeah, the models cannot generalize. So that's some of these kind of benchmarks we're looking at in the technical report where they had kind of pointed out that maybe prior generations of co-folding models were actually really, really struggling with generalization. So there was this idea in the community, okay, maybe they haven't actually learned all that much, and they're mainly just kind of memorizing known protein interactions.

22:11Wow. Becky, I have a question. How does a medicinal chemist interact with these systems on a day-to-day basis? Like, is the model proposing ideas, ranking ideas you bring to it? What does that look like? So the first thing you're going to do on a project is use the structure models to examine your protein, find out where there are predicted pockets. Maybe if there's known chemical matter, you can use the co-folding models to predict where those bind so that you can understand where maybe you want to target. In the very early stages of a drug discovery project, you have to have some kind of biological hypothesis that if I modulate this biological target in the context of this patient population, I'm going to have an effect on this disease.

22:55So I've got to inhibit or activate this protein or this collection of proteins, and that's going to have some positive benefit for this collection of patients. Then once you have that understanding, you're then going to look at that protein and be like, OK, how am I going to inhibit this protein? I don't just need to bind to it. I actually need to functionally inhibit it. I need to stop it doing the thing in the cell that it's doing too much of to cause this disease. so you need to understand a little bit about how that protein works how it interacts with the kind of cellular machinery around it which interfaces you might need to actually bind to to stop that functional effect that's going to sort of give you an indication of the pocket you need to go after and we call them pockets because proteins are they're not smooth surfaces they've got loads of crevices on them and so you're going to be targeting one of those crevices with your with your molecule um okay so you know where you're targeting you know which crevice which pocket you're going after now you need to identify some something that some starting point some kind of toehold in so that you can then optimize that um so we have um generative modeling capabilities so de novo models they work on just the single amino acid sequence of the protein and you can use those to generate chemical matter for pockets de novo and that's something that we found an incredibly powerful way to to replace traditional methods that you might use to find small molecule start points like high throughput screening is something that will be very familiar to people who work in in other biotech and pharma which is where you might screen compound libraries of millions against a protein target and you look for hits essentially what we're doing is maybe synthesizing a very small collection of bespoke molecules that have been designed de novo using the models scored using this kind of suite of inference models binding affinity admi predictions and others, get those made bespokely in a lab, test them.