Live on Bloomberg TV

00:00Max, I want to start with you. When we think about A.I. and drug discovery and the models that you and Iso labs are building for the layperson, what does that actually look like on a day to day basis for you and the team? So what we're doing at Isomorphic Labs, we're building this drug design engine. So you can think of this as a machine which is able to, you know, come up with new molecule designs for different disease areas, different indications, and even for different modalities, not just a single modality. And, you know, underlying that drug design engine are a whole suite of A.I. models. These are these are, you know, not just a single model like Alpha Fold, but actually we're going to need, you know, maybe half a dozen alpha fold like breakthroughs that come together into this drug design engine. So these are predictive capabilities, generative AI capabilities, things, for example, like for predict the structure of proteins. Proteins are the fundamental building block blocks of life. They form these molecular machines. We need to understand what these proteins look like physically in 3D and also how these proteins interact with other biomolecules things like DNA, RNA, small molecules, small molecules. The class of molecules that often drugs that you take as pills. So if you can start working out and predicting, okay, what shape is this protein going to take when it interacts with DNA? And if I design this small molecule drug, how is that going to come in and affect the shape of the system? Because ultimately, if you can start changing the structure, 3D structure of the system, you start modulating the function, you change the function of this molecular machine that changes some functions within the cell, and ultimately that could change the phenotype or the, you know, the state of your disease. So so these sort of components like alpha for these structure models, models that can understand how strongly these molecules bind together binding affinity models. And then of course, drug design is not just about making a molecule that will modulate your molecular machine in a particular way, but it also has to be a good drug. It's got to be safe for people. It's got to, you know, be absorbed into the gut. If you're taking it as a pill, it's got to get to the right part of the body, maybe even got to penetrate the cell wall to get into the cell itself. So there's a whole spectrum of models that come together into this drug design engine. There's so many so many different variables that you touch on there. So, Rebecca, given your biomedical background, how how do the systems that Max and his team are building change drug discovery from your perspective work versus the legacy systems? What are you and the team able to do with these kind of models, this kind of technology that you wouldn't be able to do with legacy systems? Well, if you think about the way that we tackle drug design more traditionally, it's very iterative. So you maybe make a slight tweak to your molecule, your drug molecule. Then you got to go in a lab and you've actually got to make it. That might take a couple of weeks, maybe even a couple of months. You ain't got to test it in a system which is approximating, you know, your biological system that you're interested in and you get the data back, then you have to iterate again. It's very, very stepwise. And this is why it takes so long and it takes so much time. So what we're doing is lifting that all out of the real world by putting it in this virtual world so that you doing all of this on a computer. So you're designing a molecule, you're testing it through these models, which, you know, thanks to the incredible ML group we have or in some cases approaching experimental accuracy. So you do all this design on the computer and then maybe multiple iterations on a computer, and you take the best things and you take those into the lab. So you're skipping steps and you're managing you're getting all the way to something that looks much better for half the time. And that's where we get the shrinking at the time frame that we're that we're aspiring exactly to achieve. And Max, you touched on Alpha Fold, and many in the industry had assumed that it was almost an impossible problem to solve, which is the protein folding problem. You and the team managed to do that with Alpha Fold two and then Alpha three, which gives scientists and medical experts the ability to look at how those molecules interact. You talked about the need for other models. What is what what comes beyond alpha fold? Three What are the breakthroughs do you need or other models? Do you need beyond half of all three to get the kind of breakthroughs that you and Rebecca aiming for? Yeah, you know, I think we have much better understanding of this biomolecular world today than we had, you know, even four or five years ago on a computer with models like Alpha two of three. But we're not the full way yet. You know, there's still a huge amount of accuracy to be gained on these models. And, you know, we're seeing, you know, big step changes in accuracy internally with our sort of next versions and iterations of these types of models. And then as I touched on before, it's not just about understanding the structure of these components. That's really just one part of the problem. You need to understand how strongly. These molecules come together. You need to understand how these molecules interact with different parts of the body. You know, a common cause of clinical failure is from toxicity, and a big source of toxicity can be from, for example, how these molecules not just interact with the target protein that you're trying to design for, but interaction with all the other proteins in your body. And so you can imagine that we need to understand not just how these molecules interact with what we're designing for, but with everything else in the body. That's a big task. And then there's this whole question of actually how do we design these molecules? Because molecular space is massive. You know, something like ten to the power of 60, that's that's a huge number. So you can't brute force such that you can't just like ask these models, hey, how does this molecule look? How does that molecule look? People try to do that. You can maybe feed, you know, 100,000, a million, a billion molecules through these systems and start reading the answers. But that barely scratches the surface of this massive molecular space as a here. There's a huge opportunity for generative models, search processes and agents to be, you know, instead of having to exhaustively search molecular space, molecule by molecule, have generative models that come up with designs that can effectively search the entire space and change that down to just ultimately a handful of molecules that then, as Rebecca said, we can take into the lab and test. And Rebecca, on that point, what is the biggest challenge in getting and you were talking about this the in silico molecules that the the models may suggest and making sure that the in silico versions, the computer versions essentially work in the real world in that in the lab conditions. What are the biggest challenges around that part? I mean we can think of the models as enriching what we're doing so they might not get it perfectly right every single time. And we have to be aware of that. We have to have a strong hypothesis for why we're making a particular molecular change. And then once we have that hypothesis and we've got some conviction, we take it into the lab, we make it and we see and sometimes we're absolutely blown away by what we get back. You know, we've we've generated something that's novel, it's potent, it's got the right kind of properties, solubility and, and everything else that it needs. And sometimes things don't work as expected. And we would imagine that over time, as the models get better and better and our processes get better and better, we should start to see this increased success. And what is that progress looking like right now? I mean, do you have days where you kind of stand back out? Wow. That this this progress has been phenomenal? Or are there more days where you've got your head in the hands thinking it's not working out the way I hoped? There have been some days where things are just phenomenal and I might be go find Max and just say like, can you believe this? You know, it's one of our structure models predicting the pose of a molecule binding to a protein always perfectly. And you get an x ray structure and it confirms, you know, to just this incredible accuracy. And it's something that's never been seen before. Just absolutely mind blowing. And do you do you and your team now, Rebecca, have a list of molecules that you've identified that could be that could be candidates? Are we at that stage yet with some of our more advanced programs? We are moving through pre-clinical development and then we have a whole load of other programs coming through that are much earlier with lots of opportunity associated with those as well. So yeah, and Max, on the question of data, because medical data, biometric data, it can be or biological data, it can be complicated, it can be messy, it can be siloed, it can be biased. How do you ensure that you're getting the data that you need into your models? Yeah, You know, machine learning the field for machine learning is data. And you know, there's the old saying garbage in, garbage out. So the data has to be good and it's got to be, you know, the right distribution to learn from. We have a very, very comprehensive data strategy where we will look at as many sources of data as possible. There's a huge amount of publicly available data that works really well. There's a lot of historical data that's also being collected. But as you touched on, it can be highly biased because it's not being collected for machine learning. It's being collected as a byproduct of maybe historical drug design campaigns. So I still believe there's a huge opportunity to actually be creating the data that's needed for these machine learning models. And we do a large amount of work in that respect, as well as actually generating our own wet lab data for the sole purpose of training. And you're confident in the quality of your data at this point? Yes. You know, I think that's where we invest a lot in, you know, understanding what data is out there, working out how to, you know, ingest and clean and really work out where the signal is from that data. These are sort of big areas of research for us. And that's where you can see, you know, huge benefits. And both you and Demis have talked about the generalizable technology that you and Iso Labs. So labs are building, which is different to some of your competitors. Just explain why that is important, why that's the route that you've taken. Yeah. This. This this notion of generalizability, that's the key for what we're building, because ultimately we don't want to do just a single drug design program and then we're done. No, we want to be building this drug design engine, this machine that we can apply again and again on any target in any disease area, even against, you know, any different modality. So it's really important. What does generalizability mean that we have these models that are trained on a certain, you know, set of data, but then they can be applied on something completely novel that can be applied to a protein, a target that's never been worked on before. They can discover chemical matter that's never been seen in humanity before. And, you know, in practice, we're seeing that bear out day to day. We have benchmarks internally that measure, you know, how generalizable our models are. But, you know, the proof is in the pudding. Can you actually use these these models on a completely novel drug design campaign? Is it fair to say that that generalizable model is more ambitious and more challenging? The outcomes can be that much more significant, but the building of that model versus a more targeted one is more ambitious, is more challenging. Yes, I think that's that's the case. We you know, I said we're focused a lot on, you know, new molecules for new targets, you know, not just changing existing, you know, published drugs by small amounts. So we want to be creating novel chemical matter, you know, against problems that people have been working on for a huge amount of time but made no traction. So it's really important that we have these these generalizable models that can work for that. But they are harder to build. It's much easier to collect a small amount of data around the very restricted problems that you know about, create a small model around that that can be used just in that little restricted problem set but doesn't can't be applied to the next program or to the next program for us that won't that won't cut it, that we won't get get this sort of reusable engine. And so from day one, we've been building these big foundation models, things like Alpha Fold three, which do strongly generalize to completely unseen systems. And again, we're seeing that across the board of our model space. And Rebecca, why does it make sense for the internal pipeline to focus on immunology and oncology? So so cancers, what was the rationale behind that? What does that mean in terms of what they what the achievements could be? And it makes sense to focus on those two disease areas. And clinical trials are much more tractable and we can run them in a shorter time frame. They're diseases with a lot of impact as well. I mean, we all know someone affected by cancer and immunology is also important as well. Affects a huge proportion of the world people who have inflammatory diseases. So they made sense as places for us to start really good preclinical models as well with good translate ability into the clinic. So you can make a prediction about something working based on models that you apply before you go into humans and things should translate better. What is your outlook? I know this is a difficult question because there's all sorts of variables, but given what you see right now and what you're working on, what would be your best assessment as to when we may be able to cure cancer? And I think cure cancer is a massive statement because actually cancer is a collection of a huge number of diseases. But I think we can see this as a stepwise process towards taking people with different cancers and giving them a hugely better outlook where actually cancer becomes more of a chronic disease where you maybe take medicine for life, but you have a normal lifespan. We already see this in some cancers where medical treatments can, you know, give you a normal lifespan. And I think for me, that would be incredible, right? Cancer is no longer something you need to worry about. It's just something that you need to make sure you get the right treatment and you maybe have to take treatment, but you have a normal lifespan. And that proposition is is not is not too far off. We're talking years rather than decades potentially. I think it's really difficult to put a time stamp on that. I mean, there are some cancers that are incredibly complex and especially as cancers become more and more advanced and become more difficult to treat. So but I think we can we can start to think about how we might approach this problem. Now, Max, give us a sense of the power of the models that you're using right now. Can that can they analyze an entire family of proteins, for example? I'm just getting a sense of what what they're able to do in the here now. Yeah. So these models, they can be applied to not just a single type of protein or a single protein, but we can apply this across what we call the whole human proteome, the whole universe of proteins that are within us. And that means the other thing is because these models, you know, run really quickly, the neural network models say, you know, we can run thousands and thousands of cores of these models in parallel, which means we can think about analyzing not just the protein in isolation. Yeah, the whole families of proteins, whole universes of proteins. And this is why this whole world of in silico drug design is such a departure from the experimental world, because in the experimental world, you know, to study a protein, you have to go and synthesize your protein, You have to get it shipped to you then have to do the physical experiments. You can't be analyzing, you know, thousands of proteins in parallel against the same thing. So it's a really changes sort of experimental design based on these models. Is is this you're an expert an expert in many things, but I bet you an expert in reinforcement learning. Is that still the right approach to try and refine this technology? Reinforcement learning is is one of these fundamental tools in the machine learning and AI toolbox. And it's part of the spectrum of tools that we use in Ice morphic labs where reinforcement comes into play. Reinforcement learning comes into play is where, for example, we have, you know, very powerful generative models. We have these, you know, really powerful scoring systems that can tell you how good or bad a molecule is. And then you want to work out a process of iterating on that molecule and making it better and better towards, you know, the sort of design criteria that Becky and team will come up with. So that's where reinforcement learning can come in, can help train these agents that are able to explore molecular space towards these design criteria for for drugs. On the partnerships, Becky, with Novartis and Eli Lilly. Can you give us some detail in terms of some of the milestones that you and the team have been able to achieve in those in those partnerships? So these are incredibly collaborative partnerships, and I've been working as part of one of them. We work really closely with Novartis, and I think Novartis came out against it initially and said these are really challenging targets. The targets we've been working on for many years in some cases, and some of them are sort of a graveyard of pharma effort and we've managed to make some really good progress on some of them. Some of them represent what we might call dark chemical, that dark protein matter where there's never been a molecule identified to bind to these proteins. And we've managed to identify some of the first chemical matter binding to these proteins. So they've seen some incredible stepping stones from the Isomorphic platform already. Okay. Yeah. Now that's huge. And if you were to look back, if you look out ten years, if you project out ten years, how does this technology how do you think the balance of power shifts around big pharma versus biotech and some of the startups and some of the tech players? And I think that's maybe quite difficult to predict. But we see what we see from Novartis is this really forward thinking mentality where they they have, you know, really deep scientific expertise internally, you know, incredible. But they also be forward looking and they want to partner with with biotechs who are really on the cutting edge of pioneering these new deep learning models so that they can ensure that using the latest capabilities in drug design really try to push that frontier forward. And Max, Demis' forecast when I spoke to him, that's in 2 to 3 years time, potentially we could be at the point where we're reducing the AI drug discovery part of the process down to maybe four years, possibly two months, possibly to a shorter timescale. What are the technical hurdles that you and the team need to overcome to get there? Yeah, absolutely. You know, this is the goal and we're really starting to see some incredible progress towards that. The key technical hurdles around this is really being able to, you know, predict more and more from about the outcome of experiments in Silico on a computer. And so the closer we get our models to that experimental level accuracy, you can imagine, the less you have to rely on going out into the lab, synthesizing your molecules, actually doing this manual work. That's a really for us, a big goal for a lot of our modelling capabilities is, you know, this experimental level accuracy for some of these we've actually, you know, achieved these and starting to achieve these especially in different pockets of of the chemical universe. So that's one of the real key things for us. And Rebecca, I'm going to ask you to forecast again, but in ten years time, given what you're saying, given what you and Max are working on, what does our world of medicine look like for people interacting with medicine, interacting their doctors, having to take therapeutics, how how different is it going to look in ten years? I think it's difficult to put a time frame on it, but I can see this really exciting future where you don't have to worry. You've got symptoms, you might go to your GP or maybe you are actually interacting with, you know, some sort of A.I. tool that can help to diagnose you at that point. And that would hopefully be a medicine to treat you, to treat your disease. That's obviously way, way into the future and there's lots to solve before then. Isomorphic labs, you know, will want to collaborate with others to kind of achieve this big this big vision that we have to try to solve all disease. It's a massive ambition, but I think it's a really exciting future where we can just really make a big impact on human health and disease. That's what gets me out of bed in the morning, really thinking about how can I impact patients lives with them, with the work that we do and solving all disease guys that does that happen whilst you're both still at Isomorphic Labs? Does that is that feasible to think that that you could meet that goal potentially in both of your lifetimes? I mean, you know, this is a big ambition, right? There's no denying that. It's also not going to happen overnight and it's also not going to happen alone. We're going to have to galvanise, you know, not just ourselves, but I think the whole industry, many industries together to start stepping towards this. You know, what is true is that there is this fundamental new technology which is quite different. It's generalizable, it can be reused across different spaces and that's something it's very rare to find more broadly in biotech, and we haven't really seen that happen often in the past. And so we really have an opportunity now to actually step towards, you know, this holy grail of solving all disease. Yeah. Rebecca And yeah, just to echo all this, it's an incredibly exciting opportunity. You know, we were talking before that I've worked in this space for a while and now to be able to use some of the tech with deep learning coming out of Max's team to actually design molecules in this virtual world actually make those and see the incredible jumps that we can make in condensing this part of this process down with the aim of getting molecules to to patients faster. And it's really it's a really exciting time. And I've been blown away by some of the technology and I hope that we we get to see some big, big changes in the time that we're both at Isomorphic labs.

No schedule data available.

BTV Channel FinderWatch BTV in your area

Isomorphic Labs: Turning Cancer Into a Chronic Disease

More From Bloomberg Technology
All episodes and clips

Bloomberg Markets
"Bloomberg Markets" is focused on bringing you the most important global business and breaking markets news and information as it happens.
More episodes and clips