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

Drug discovery is 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. We see a massive failure rate even in the clinic. The whole process can cost upwards of something like three 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. That's just kind of not escaped from that. This 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? Clearly by non-binding opinions on spreadsheets. What is it 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 after. 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 Schorsenit, who leads foundational AI research there. We'll unpack what alpha fold 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 alpha fold, 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. 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. 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 3 billion. So 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. Well, would you say that the biggest misconception people have is when they share 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 design 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, that 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? That would be very interesting. Yeah. 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 clinic 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 and antibody? Maybe a peptide? And how am I going to develop that chemical matter from the 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 of it 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 thinking about what an AI drug design drug actually looks like. And maybe just to build on that quickly, on this misconception point. And I think you're getting this from back, you're right. And so I sometimes get this question of, oh, how can I help cure cancer with whatever I'm doing? But it's just so overwhelmingly many capabilities. If you really want to take a look at every piece of that, we think, OK, how would you do this with modern data driven AI approaches? And yeah, so kind of getting beyond this, oh, it feels so overwhelming. There's so many different things that can go wrong. And finding root-not problems, where if you can make meaningful progress, you can then see the impact across many different drug design programs. That's something that we're very much focusing on, especially in the early years of our small effects. OK, that makes a lot of sense. Let's see 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 the number of things that as a human drug designer, you kind of need to optimize for as you go through that optimization phase. So 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 politely, which is going to-- you can take as a tablet, which is going to survive the various acidic conditions in your stomachs, 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? So 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 a steady path of optimization through to a drug candidate. So as you probably saw in the ISO DDE, Technicor 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 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. And 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. And what-- 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're usually sort of like giving up and giving over to the machine a little bit. It sounds like-- Yeah, exactly. And that, that sort of is a level of trust that people have to build a little bit as humans. Yeah. Then we sell, even, in a lot of ways. Exactly. 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. And 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. So if this 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 that's really helpful. That's like kind of like confident scoring. Yeah, exactly. Exactly. And then you get into this world of, OK, 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 in creative ways. 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? A binding model that tells you, OK, 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? to be able to put together quite a sophisticated high-poll. 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 like text, images, or code, like what we're traditionally looking at with AMO. Yeah. 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? It's actually, it's a modeling artifact created based on the X-rays, right? 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 the kind of the processing of the data. Obviously, the cost of generating new ground truth data, knowing even what data you might want to be generating. So you spend a lot of time on that. The other part is like, if you think about this whole structure biology, atom model space, right? Think what you're trying to do. You're trying to place individual atoms. So it's the models are incredibly sensitive to every tiny detail. And if you think about this kind of from model perspective versus LMS, 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 and your LMS, you kind of get to reinspect, you 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. Oh, right. 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 some reverse engineer that you figured it out. Yeah. 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. But then if you look at the, let's say, old fold three, 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 infants 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 and pause and aspects of it. It's very fascinating to work with. Yeah. It's amazing to be able to. 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 and navigate this. Any big ones that were that are memorable. Lots, but I have to think about which ones I can speak about. 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 and something that you can actually tie to progress on the things that you care about on your drug design programs. So you can make a lot of progress. And that's maybe less. Now it's very easy to make progress on some even scientifically very interesting metrics. And benchmarks and you can feel kind of the curve is going up. But if you are not seeing the impact on your drug design programs. What are you doing? So like the equivalent of not seeing the productivity data show up in like GDP or whatever. Yeah, it's like, where is it? 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 by 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 can kind of can speak to that over the past few years, you actually see new things working every time. And that then kind of gives us the confidence to keep making these investments and these very expensive foundation models. Yeah, maybe a, maybe a cadet to that that we, it's incredible for us on the drug design side. Because obviously we're in the same building as this 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 in inter place with this new model. I'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, I can keep up. Great 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 science and, you know, making that work. Yeah. So I think it's also, you kind of 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, but 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 heart attack long term missions, right? Like you're almost you're seeing some people burnt out a bit by this M lab, lab rays and school uses what 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. Right. And even though you might spend a year of your life grinding out like a fraction of that percentage point on the foundation model, as you stack these up, we only have to solve that problem once, right? Like to be able to do certain things at experimental accuracy. And then you can bank that forever forever. Yeah, exactly. So that's just amazing, right? It is. That is cool. Well, what's the, 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 like the next hardest challenge related to that for drunk discovery? Like what else in our biology do we need to model in order to be able to predict reactions? Like is it, are we going bigger as in like we need to model how cells interact with each other? Or is there other mechanisms we need to go smaller? What's what's the right next challenge to solve there? Oh, that guy for Michael. Maybe we can start and you can add an. I think in a sense, it's, it's both you still need to make a lot of progress at the atom level modeling. And then you need to integrate a lot of other biological scale. And so maybe on the first one. There's proteins. There's kind of there's also for two moments. Can you kind of get a crystal structure of most proteins with experimental accuracy? Then there is the kind of awful three moments. 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? There's still a lot of work to be done there to get this to full experimental accuracy. And then now a day 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, yeah, cell level, organ level, toxicology, all of these things. 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, what I'm wondering? I think you tip it. You end up using slightly different modeling paradigms. I would say all of them are extremely challenging. It's probably the fair. Yeah, maybe back you want to add onto 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. And that kind of is how they enact their biological effect. Whereas now and 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 may be blocks the function of the protein that you're trying to inhibit. Or maybe these 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 so much possibility for the design of new new drugs, new modalities. And these are things we can start to model because we understand. and how proteins interact. Whether there's a pocket at the end of phase, 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 new world for us. - I'll bet you're learning new stuff every day right now, aren't you? - Yeah, right, yeah, 100%. - Maybe what's kind of interesting is, so this is what we kind of 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 satellites, effects that vacuum mentioned. So there's induced fit opening a cryptic pocket. So suddenly you're targeting something somewhere that you didn't necessarily expect, but the model of the SQ2. So like 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 all of reaching 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 if you are just very, very good at atom level modeling, then all of these therapeutic modalities, it's ultimately your modeling atoms interacting with each other, right? Rather than trying to be very specialized, I have my fine tune model for this target, this therapeutic modality. Right, you can do a lot more if you have a general model because you can't do it. Yeah. And that's my slightly counterintuitive against, but say some common narratives around, or there's just not enough data or you just, yeah, the models cannot generalize. So that's some of these 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 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 non-protinent actions. Well, 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. So 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 molecule. OK, so you know where you're targeting, you know which crevice, which pocket you're going after. Now you need to identify something at some starting point, some kind of toe holding so that you can then optimise that. So we have generative modelling capabilities, so denover models. They work on just the single amino acid sequence of the protein. And you can use those to generate a chemical matter for pockets denover. And that's something that we found an incredibly powerful way 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 other biotech and farmer, which is where you might screen compound libraries of millions against the protein target and you look for hits essentially. What we're doing is maybe synthesising a very small collection of bespoke molecules that have been designed denover using the models, scored using this kind of suite of inference models, binding affinity, admi predictions and others, get those made bespoke in a lab, test them. And we've been using that to find hits which is much faster, much cleaner. And then when you get a hit, you know, because it was predicted by the model, you're leaning into an area where the model is working well, enabling you then to push the project forward much faster. So this would be the medicinal chemist working within the system to run the generative models, to generate chemical matter and to score it in to select the best molecules for synthesis. Very cool. That's a really interesting process. I had no idea that it was, as, I mean, I guess I knew it was technical, but the idea that you're able to zero in and be like, here's the little niche problem. This is very important to do that. You have to be able to do that, right? Yeah. Is that the hardest part? Is just figuring out the hypothesis, then? I mean, obviously all of it's hard, but like, to really say like what is actually the exact mechanism or the exact area that we could target to do something meaningful? That's a kind of whole challenge in itself. We call that kind of target ID. And it's an incredibly difficult area to make meaningful impacts, because identifying biological targets is hard. You need to identify targets that are going to have a functional effect in your patient population that are not going to be associated with safety or other kind of risks. Is that target going to actually have an impact on the ethical? Is it going to have an efficacious impact on the disease? And sometimes this has to be a clinical experiment, so you then have to get all the way through to the clinic. And if you've gone all that way, maybe a decade in, you know, huge amount of money spent and then it fails for efficacy, that's a big loss. So there's big potential here, I think, for AI to make an impact. What is your pie in the sky hope to be able to turn a drug from hypothesis to out helping people in the real market? Pie in the sky dream would be that you can take a protein and in maybe one iteration, this is my dream. North star pie in the sky dream, that's what we can work towards, right? Whether we get there, I don't know if it's possible, but that can be what we work towards. And then is there any way that being able to do that speeds up the actual trial, another trials that you would have to do, like with human trials and all of that, like can you speed things up on that side as well? Or it like is the hope that, you know, you save as much time as possible on the front end, and then that just takes as long as it needs to take. Yeah, I suppose the fast part is just what are you actually sending to the trial, right? Well, the hope is that you're sending much higher quality molecules to the trial, and then ultimately that gives you much higher success chance. That itself might not shorten them that much. But then I think over time, there's now a lot of companies, a lot of creativity around of this kind of toxicity modeling on chip, basically trying to pull kind of more and more earlier into the modeling, so that by the time you get it to humans, you've already done kind of just more and more with models. So I think that's an emergency. It's exciting to watch. Where if you have an accurate representation of like an organ and you could accurately figure out how it's going to react, then you kind of know, okay, we don't have to do as many tests on actual humans because we have a pretty good idea of what it's got. Or even just beginning to save some of the small end-of-mill trials about already being like having that verified with the kind of organ on chip, so you're saying? Well, I think you maybe build on that by saying that even if you still have to do the same number of human trials, you'd hope your failure rate is decreased. So you've done all of this kind of modeling up front in your discovery phases and your pre-clinical phases. So by the time you get into the clinic, you've got really good confidence that your molecule isn't going to fail at least for a safety or tox reason. And if you can get the target right, then you don't fail for efficacy either and we start to see that clinical failure rate go down. And if you think about how high that clinical failure rate is at the moment, let's say 90%. We don't have to make too much of a dent into that to see really meaningful change. Yeah. I mean, maybe there's also like one aspect that just historically, before a lot of this modeling evolved, before we had a lot of these computational tools, modern experimental approaches, right? You just didn't really know even what exactly you were sending to the clinic, maybe, right? Or you just had to take the risk because you were out of money to optimize it further on some dimension. No, you didn't have it principally to optimize it further. I hadn't even considered how much money this could save in the drugs every process. That is not even a thing that had crossed my mind until right now. But gosh, it's got to be huge over, I mean, when you think about, you know, the time of expensive scientists, of expensive materials, lawyers and all of the many things that get involved in that process, I just, I can't imagine. Yeah, I mean, this is a process that takes over a decade and actually the a recent. publication, or fairly recent publication, but capitalized costs here. So accounting for failure rate up at six billion per new drug. Wow. Absolutely huge. And that makes it not commercially feasible to develop drugs for small patient populations, for example. Where diseases, things like that. Where diseases, whereas if you can bring that cost and time right down, then we should be able to have more of a no patient left behind mentality. Wow. I guess a big part of the drug cost is not in production. It's in the research and development it takes to get it to market, I assume. Yeah. So actually a huge amount of the cost falls in the clinical phases. There's also cost before that and getting model kills into the clinic as well. So do you think it's realistic in our time and our lifetimes for there to be something equivalent to a personalized drug workflow? Where at some point, well, these models get so good that you could make a custom drug for one individual. Let's start. Go for it. Well, I was actually going to say that I'm probably not the best person to make this judgment because if you'd have asked me 10 years ago, would a, would an AI model be able to predict with experimental accuracy, how a small molecule binds to a protein? I would have been like, no way. Progress in this field is just mind blowing. And I mean, I would say I'm going to be optimistic. I'm going to say, yes, I think there's a kind of question on the economics of it, right? Like you sometimes now read some, like interesting internet stories about scientists who develops their own kind of customized cancer vaccine, sorts of, some of these stories. But then at what point will you have that kind of broadly available at reasonable cost? That's another question. But I'm optimistic. Well, even if you can do it for a small groups of, you know, people who do have more rare diseases and be able to target that, I think you think that is just an accomplishment in and of itself. Because how many people go untreated with things? Yeah. I mean, you see this example, right? When was the first human genome sequence kind of two and a half decade ago? Right? That was a late 90s or something early 2000. Yeah. So that was a gigantic effort. It caused a gigantic amount of money. And now you can order, I think, for a few hundred dollars, maybe some of the full sequencing, but you can order something that basically gives you some information about certain genetic risk profiles that you have. Right? So there's just this incredible kind of over time, others of magnitude, cost reduction on the science side of these. Yeah. That's really neat. That's really neat. Something I wonder in this is when the model and your human scientists disagree. What happens? How do you, how do you, uh, it is, I bet it is. I'm just curious like how do you resolve that? So I mean, this is not something that's unique or new. This is, you know, this is something that we come across a lot. And so you are working day in and day out with these models, um, using them to form a hypothesis, uh, so that you can essentially only commit to experimental work, which is slow, lengthy, expensive when you have confidence in the molecule that you've designed, or the experiment that you, you've got a good hypothesis there. Um, and sometimes the predictions made by the models don't align with what you're expecting. So you kind of have a number of choices in front of you. Do you go with what the model predicts or do you try and do something else? And I guess our ethos here is, well, let the model guide you because it's going to give you information. It's either going to be correct and then it's pushed you forward in what you were trying to do. So dress your hypothesis or it's incorrect. And then it's told you something about the model itself that actually it's not to make a prediction in this space. Or, um, maybe you need more data or, you know, you go to the ML team and you say, Hey, this happened. And, and like, you know, what's the next steps here? And that becomes a really interesting conversation in itself. I take especially like in the earliest years of the company where we had much earlier versions of the model obviously. So a lot more was going wrong all the time. It's having that feedback loop of the human intuition and being able to correct tons of things. It's incredibly useful. Obviously, there's still things going wrong. I mean, all the time now, but it happens. Got to make progress. At least now we have some models where we are feeling fairly good and we've kind of seen them being used across a number of drug design problem. We're just building up that confidence where we can make stronger cases for actually, maybe you don't need that experimental data anymore. And ultimately, you have to think about the total economics of wanting to do drug design. Right. Like you need if you build a platform with foundation models, you need to be able to do kind of many programs. To take many shots on goal and parallel to be able to kind of amortize the investment in these foundation models for any individual program. Of course, you might always say I also want to have any data, any experimental data. Like the trials are going to be so much more expensive than any kind of amount of data that I spend on little extra experiments early on. But then on a platform basis, you need to get to the point where you stop doing that on some programs. Right. Let me show you. Actually, so back in the early days of ISO, we were using Alpha Fold 3. We would sometimes be using, so we're using the models to make a prediction of how a small molecule binds to a protein. And sometimes we'd notice visually kind of violations, structural violations there's maybe you were expecting an aromatic ring to be flat and maybe it's puckered. So, you know, going to the ML team and saying, like, why is this, like, why is this model of your being predicted in a binding pose that I know is actually physics, it doesn't obey the laws of physics. But that's like essential feedback because now you can see in the ISO DDE report that that doesn't happen with the latest models. So this kind of regular feedback is really important for us to like take those meaningful steps forward. That's interesting to add like a physics layer on top of it or something like, like, make sure you follow physics laws. How did you solve that? I think we'll not comment in the kind of full detail on that. But it's I think it's worth saying that you keep seeing deep learning, being able to do things that you maybe didn't expect to be able to do. And so, yeah, you ultimately find creative ways of solving these problems, even without having explicit, explicit physics. Like, for example, you see these video models now that can actually kind of model the physics of, yeah, basically the world. And they're trained on videos. They're not necessarily trained on physics, right? Right. It's like an emergent property. Yeah. One, I think one really exciting aspect to all of this is sometimes a model will make a prediction and will test it experimentally and it will give us this beautiful result. And then you have this converse question to the ML team of like, how did it know that? Like, how could it make that kind of general inference? And I think that's another thing that like, you know, makes us think as a team like, like, how did it do? How did it generalize over here? There's no data. How did it know that? And I think for us working here, it's like being like kids in a sweet shop. There's like much. There's so much exciting science that you can do when you combine these predictions using these models and then actually going out and verifying them experimentally. It's a really kind of beautiful collision of the two worlds. What, I guess, I guess in the early phases here in the years to come, what types of diseases or modalities do you think will be the most likely to benefit soon? Well, we're working at the moment we're focused on oncology and immunology. So I think for anyone, you know, listening, we all probably resonate with kind of the oncology area. We know how important that is. You know, unless you can find a cure, there's always going to be a need for treatments there. So we feel that we're doing really impactful science by working in the oncology area. And it's the same for immunology. There's a lot of our medical need there. So we're quite excited to be working in both of those areas. Well, 10 years from now, what would you say, like, yes, AI fundamentally changed medicine? What would make you say that? And what would make you say the impact was meaningful, but maybe more incremental than advertised. So in one hand, what, where would it, will it fundamentally change medicine in your production? And where will it be making progress, but not as much as we think? I would find it incredibly exciting and kind of see our hypotheses played out if we actually see that some of these targets, where we now see the models generalize into, so really like the most difficult and tractable targets, where you may be, there's nothing on the market, even though there is a kind of known this is biology. And it's just kind of, there are all these undruggable proteins. And if you actually see in 10 years that lots of kind of first and class or kind of step change, best and class. If that comes to market, or it's kind of about to come to market in 10 years, that for me would be kind of the real success of that step change hypothesis. And if it was more kind of marginal and you see a lot of kind of this first follower type, so you kind of it's it maybe works out on some level commercially, but it doesn't really revolutionize. I don't know how you see this as. Yeah, I feel the same way. I think where we have this real, I guess, unique position is, can we find small molecule or biological drugs for targets that have been labeled as undruggable? So targets that we know are validation disease, we know they cause disease, the field have been trying to drug them for many years, but we've not been able to make progress because they're just too challenging. And we start to use the technology we have to make progress there because I think that will that will be massively impactful. And it will open up new areas of biology, new patient populations in the clinic as well. Is there a common example of an undruggable? and drug-able protein that we might know of. - What would you know of? Have you heard of K-Ras? - No. - K-Ras, it was, - Tilted as undruggable for decades. And we have huge numbers of people across the field working on K-Ras. And then recently we've seen some beautiful progress in that field. And I mean, you may have seen some of the headlines recently about improved survival in pancreatic cancer because of K-Ras drugs. - Doubling it, right? - Doubling survival in pancreatic cancer. So this was a target that was labeled as undruggable. The amount of work across many research group in kind of experimental sciences, taken multi decades. And now we're finally seeing that paying off. And can we now do that for other targets that would have that same label? - Without decades. - Oh, fully. That's where I need to call. - That's amazing. - Rebecca, Michael, thank you so much for joining us. It's been a delight. It's been great to chat. - Thank you for having us. - Yeah. - What's the best way to keep up with you all and I see more of it? And what's coming down the pipeline? - Well, we have a block where we post updates like the technical report. Obviously we're hiring across the board, across many roles. I think this is a mission that many people find incredibly meaningful to come to. And I hope it really resonates with your listeners as well. We're hiring in Boston, we're hiring in London, we're hiring in Switzerland. Yeah. - That's amazing. I would say the work you all are doing is very much what many of us see as the real promise of AI in the future. Is the ability to extend life, improve life, fight new diseases. And I think what you're doing is a fantastic thing. And I'm grateful to see there are so many people focused on it. - We're excited to. We're excited. - Oh, right. Well, if you haven't yet, please reach up and hit the subscribe button right above you, or right below you. It'll be below you. And we'll have links to everything we talked about here today in the description below as well. Please also pop by the neuron.ai and sign up for our newsletter. So you can get the latest AI news every morning, right in your inbox. And on that note, that's all we have for today. Farewell for now, humans. (upbeat music) (upbeat music)