By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg

Drug discovery and development has long relied on expensive, time-consuming empirical trial and error.

Max JaderbergguestElad Gilhost

Nov 12, 202639mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

Transcript

40 min read · 8,221 words

0:00 – 0:45
Trailer
1. MJMax Jaderberg
  The large part of the industry makes medicines is this very empirically driven process. Where things really fail is in the clinic because we never could see the problems until we get to the point where we put these drugs in patients and we say, "Oh, no," like, "There's toxicity that we didn't account for." And so the really big opportunity is to have much more of that understanding before you go into patients.
2. EGElad Gil
  Mm-hmm.
3. MJMax Jaderberg
  How can we bring all of that drug discovery work onto a computer? So how can we get out of the real world? I'd be really disappointed if we fast-forward even five years' time, like the predictive understanding of disease biology has not step changed in like a massive way.
4. EGElad Gil
  Do you have a prediction on when you think, for example, 80% of human cancers will be treatable? [upbeat music]
0:45 – 1:24
Max Jaderberg Introduction
1. EGElad Gil
  Today on No Priors, we're joined by Max Jaderberg. Max is the president and part of the founding team of Isomorphic Labs. Isomorphic is backed by Alphabet and a variety of other, uh, investors, and its goal is to solve all disease, which is very exciting as a topic. We're gonna talk a little bit about the role of AI in the future of biopharma, human health, longevity. And, uh, Max, of course, has a background in machine learning. He has a PhD in the topic working on machine vision. Eventually sold the company to Google DeepMind, rose up the ranks there, and now is president of Isomorphic. So very excited to have this conversation. Max, thank you so much for joining us today on No Priors.
2. MJMax Jaderberg
  No, it's such a pleasure to be here.
3. EGElad Gil
  Yeah.
4. MJMax Jaderberg
  Thank you.
5. EGElad Gil
  Isomorphic,
1:24 – 5:23
Isomorphic Origins
1. EGElad Gil
  I think, is, is one of the more fascinating companies right now that's happening in technology in general. Um, it-- Alphabet is a major shareholder and sort of sponsor. Um, Demis, uh, who pioneered a lot of, uh, models at DeepMind and other places, which I believe you sold the company to before, um, ha- you know, really helped with the inception and setup and now running of the company. And you folks are tackling one of the biggest problems that's left for humanity to tackle, which is human health. Could you tell me a bit about how you came to, uh, work on this company, and what was the motive force to starting it?
2. MJMax Jaderberg
  I was part of the f- you know, the founding team of Isomorphic Labs. Um, I was at DeepMind for a long time beforehand. You know, personally, I was working on, uh, early days of generative models, language models, deep reinforcement learning, all of the big, you know, game playing challenge problems. Those were my jam, like AlphaGo, AlphaStar. Um, and it got to the point where, you know, we were, we were going after these like, you know, quote unquote, "impossible problems" at that point in time. You know, it's just probably 2018, '19. That sort of-
3. EGElad Gil
  It's like gameplay and can AI beat video games and things like that.
4. MJMax Jaderberg
  Exactly. You know, at that point in time, you know, we were still trying to prove to the world that AI was even a thing, right?
5. EGElad Gil
  Mm-hmm. Mm-hmm.
6. MJMax Jaderberg
  Um, and so we were going after these impossible problems, and I could see, you know, firsthand like how to-- like just systematically how to solve these.
7. EGElad Gil
  Mm-hmm.
8. MJMax Jaderberg
  And it got to the point for me personally where it's like, okay, sure, we're not at AGI yet, but we clearly have this amazing, you know, um, set of technologies, you know, in deep learning and reinforcement learning and, you know, what are the problems we really need to be focusing this stuff on? And I was, you know, I was thinking all of this. Of course, Demis was thinking all of this. At the same time, we're seeing the AlphaFold results, you know, the first embers of that starting to emerge and really starting to see for us this was like the first proof of concept that you could get a deep learning model that generalizes in, you know, a meaningful space in biology. And so that was really the proof of concept that gave us conviction that actually, you know, now is the right time to go after this, you know, the big mission of, you know, changing the way that we think about human health-
9. EGElad Gil
  Mm-hmm
10. MJMax Jaderberg
  ... you know, going after solving all disease. Um, and that's gonna take like a really substantial effort, and that warrants its own company, its own culture, you know, specific people of different disciplines coming together for this and, and the ability to run really fast on that. And so that was the genesis of if- Isomorphic Labs in, in about 2021.
11. EGElad Gil
  Oh, amazing.
12. MJMax Jaderberg
  Yeah.
13. EGElad Gil
  Your background really started off more on the deep learning side and AI modeling side, and I believe you sold a company to DeepMind in, was it 2014?
14. MJMax Jaderberg
  Yeah, that's right.
15. EGElad Gil
  So quite early. I mean, you were at this stuff before people realized how important it was, or many people, uh, realized how important it was. Tell-- Could you tell me a little bit more about sort of the journey you took, how you got there?
16. MJMax Jaderberg
  Yeah, of course.
17. EGElad Gil
  What got you interested in it to begin with?
18. MJMax Jaderberg
  Yeah, I mean, I, I, I, I come from an engineering background. Um, ended up almost by accident doing a PhD-
19. EGElad Gil
  Mm-hmm
20. MJMax Jaderberg
  ... um, in computer vision, and this was actually just at the turning point. I think I started my PhD, um, pre AlexNet.
21. EGElad Gil
  Oh, 'cause AlexNet was 2012, was it?
22. MJMax Jaderberg
  Yeah.
23. EGElad Gil
  Yeah.
24. MJMax Jaderberg
  Exactly. And then I saw that emerge. I was like, "Wow, this is awesome." I remember, um, speaking to Yann LeCun about convolutional neural, neural net-- conv- convnets, um, and I was like, "We need to translate this into the core computer vision space that we, we're working on in Oxford." Um, and yes, ended up developing all of these methods. We ended up having the, uh, ImageNet winning network of, of 2014, best object recognition, text detection, all of this sort of stuff. Um, and I had a comp-- like built a company in this space, uh, Computer Vision as a Service, and trying to raise money. And at that point in time, you know, De- uh, Demis and DeepMind had just sold to Google. So trying to raise money very-- and very quickly ended up just speaking to Demis Hassabis, um, asking for angel investment. And he was like, "Well, why don't you just come to DeepMind to solve AGI?"
25. EGElad Gil
  Yeah.
26. MJMax Jaderberg
  I was like, "What's AGI?"
27. EGElad Gil
  [laughs]
28. MJMax Jaderberg
  Um, but so began the journey into DeepMind-
29. EGElad Gil
  Uh-huh
30. MJMax Jaderberg
  ... and, and everything there.
5:23 – 11:49
Drug Discovery Stagnation
1. EGElad Gil
  Amazing. And your focus now is basically a vertically integrated company that's doing everything from models to the biology, and the idea is to find drugs that can help people with dif- different aspects of human health. Um, and so i- if you look at the pharma industry, one of the things that I think is really striking, like I've, I've graphed this out in the past, where if you look at the number of new drug approvals per year, it's roughly flat.
2. MJMax Jaderberg
  Yeah.
3. EGElad Gil
  But that's despite an explosion in knowledge. So all these scientific papers have sort of exponented in terms of the number of papers. You know, the number of papers have gone up ten, 20X during that period. Um, the number of new modalities, the type of new drugs that you can do has gone up. So there's antibody-based drugs, there's small molecules, there's proteins, there's peptides, there's all these new ways you can treat drugs. The price has gone way up, right? It went from a hundred, a few hundred million dollars per drug to, um, one point five billion. It went from Eight or nine years to 15 years. So every metric has gotten worse, while the number of things have gotten flat, the knowledge base has exploded, the number of mo- types of molecules that you can do has gone up. W- why is that? Why has it been so static despite all that progress in other ways?
4. MJMax Jaderberg
  Yeah. It's crazy to see just how it keeps getting harder and harder to make new medicines more and more expensive, longer and longer. Um, and, you know, I, I, I think there's probably a number of reasons behind this. Like, the s- the thing that I really, um, think quite a lot about is that, you know, uh, like the way that we actually make medicines in general in the industry and as, as a species, you know, forever has been a very, very empirical process, right? At the very beginning, it was, you know... And, and, and, and this was, you know, a long time ago, like 100, 100 years ago, you know, the medicines that we came up with, they came about by chance. We just, like, happened to stumble upon something that people realized had some, you know, physiological effect, and people adopted it. Things got a bit more sophisticated. It's a bit more empirical trial and error.
5. EGElad Gil
  Yeah, a lot of medicines we just guessed, right? The, like aspirin, I think for a long time we had no idea how it worked.
6. MJMax Jaderberg
  Exactly.
7. EGElad Gil
  Yeah.
8. MJMax Jaderberg
  Aspirin, we didn't know how it worked. Paracetamol, we didn't know how it worked. Um, and the way we've been making medicines and still, like, lo- how the large part of the industry makes medicines is this very empirically driven process. Empirical as in you need to, um, you know, you have a hypothesis on, you know, how a disease might work. You have a hypothesis on what a good molecule design could be, but you have to make it and test it in the real world and do this very iteratively.
9. EGElad Gil
  Mm-hmm.
10. MJMax Jaderberg
  That's slow. It's really, really difficult to get signal out of that. And the space of biology is really, really big and complex. These are like big com- you know, the cell is a big, complex, chaotic system.
11. EGElad Gil
  Mm-hmm.
12. MJMax Jaderberg
  Super high-dimensional. And then molecule design, whether it's small molecules, antibodies, peptides, you know, they're really, again, really high-dimensional objects, combinatorially large design problems. And the only tools we've got available, you know, to us, uh, you know, is, is basically trial and error screening processes.
13. EGElad Gil
  Mm-hmm.
14. MJMax Jaderberg
  And so that limits, to me, the complexity of solution that you can find in, in a finite period of time or finite amount of money that you spend on that empirical work.
15. EGElad Gil
  Mm-hmm.
16. MJMax Jaderberg
  Uh, and so I think, you know, what you've seen over time is like the low-hanging fruit of medicines that can be found-
17. EGElad Gil
  Mm-hmm
18. MJMax Jaderberg
  ... with our empirical process of designing medicines are being found.
19. EGElad Gil
  Mm-hmm.
20. MJMax Jaderberg
  We're finding the low-hanging fruit, but then it's getting harder and harder to sort of ratchet up the complexity and, and find things that are out of reach from this empirical process.
21. EGElad Gil
  That's interesting. But don't you think-- So if I look at any other industry, often what happens is the age of the industry impacts the rate of change in it.
22. MJMax Jaderberg
  Hmm.
23. EGElad Gil
  In other words, the older the industry, the slower it moves.
24. MJMax Jaderberg
  Hmm.
25. EGElad Gil
  Things become more bureaucratic. There's more regulatory capture, et cetera. I think the last time there was a biopharma company that was worth over $50 billion, uh, besides Moderna, and I think Moderna was almost an accident in COVID, was the, uh, 1980s, I think, uh, with Regeneron.
26. MJMax Jaderberg
  Yeah.
27. EGElad Gil
  And so it's been, um, almost 40 years since we've had a new tens of billions of dollar biotech built.
28. MJMax Jaderberg
  Yeah. Regeneron, Vertex, these, you know-
29. EGElad Gil
  Amgen was that era, too.
30. MJMax Jaderberg
  Yeah.
11:49 – 14:35
Hiring at Isomorphic
1. MJMax Jaderberg
  with AI.
2. EGElad Gil
  One of the things that's really fascinating about, um, Isomorphic is the degree to which you have a very diverse base of people all working together towards a common cause, often who come from backgrounds that are very disparate. Can you tell me more about what sort of people you hire, what are their profiles, who you're looking for currently?
3. MJMax Jaderberg
  Yeah. I mean, it's an interesting space because, uh, the, the, the sort of mashup of, you know, frontier AI and drug discovery, chemistry, biology, hasn't really existed for very long. So you can't kind of go out into the world and say, "Hey, I wanna hire the world experts at AI for drug discovery," because- They basically don't exist.
4. EGElad Gil
  Yeah.
5. MJMax Jaderberg
  And if they do exist, they're not using our models. They're using different techniques. So we think about going after literally just the best people in core AI and deep learning-
6. EGElad Gil
  Mm
7. MJMax Jaderberg
  ... who maybe don't even have a background in life sciences or biology or chemistry, but really wanna see their work on deep learning translate into amazing impact for, for humanity. So we hire just core people in AI. We hire core people in biology and chemistry, maybe who haven't had much exposure to AI.
8. EGElad Gil
  Mm.
9. MJMax Jaderberg
  And then, you know, we work hard internally to build that culture that everyone talks and learns from each other. And then internally, we get these sort of cross-disciplinary experts-
10. EGElad Gil
  Mm
11. MJMax Jaderberg
  ... expert scientists emerging, which has been really exciting to see.
12. EGElad Gil
  How do you think about the ratio of those things? So, um, I started a company a long time ago, uh, called Color, and I was CEO for the first couple years. My co-founder took over a decade ago now. So this was a long, [laughs] you know, ways back, and we were basically using ML and data science as applied to genomics, and we were building these genomic pipelines. And what we found is you could have, like, five ML people and then one research scientist from the traditional field, and that was kinda the ratio-ish-
13. MJMax Jaderberg
  Mm
14. EGElad Gil
  ... that we adopted. What is it like now? Or how do you think about how many people of each discipline do you optimally have in a pod working on a problem?
15. MJMax Jaderberg
  Yeah, for some of these core AI, you know, superpower projects, the c- the AI modeling projects, um, I think that's actually not too dissimilar. You know, you, you have your core modelers like engineers, uh, AI scientists, and then you just need some key domain experts, people who really understand the application, the nuances of the data.
16. EGElad Gil
  Mm-hmm.
17. MJMax Jaderberg
  Like, what do the metrics mean? Um, make sure that we're not, you know, chasing smoke metrics, and these are real things that are gonna translate into actually making medicines. Um, so yeah, those sort of ratios, one in five, like that, th- I think those resonate quite a lot for what our teams look like.
18. EGElad Gil
  And are those specific roles that you're actively hiring for currently or just across the board, you know?
19. MJMax Jaderberg
  I mean, we want the best AI researchers, um, coming into ISO and helping us push on this mission. I think there's, there's literally so much work to do on the-
20. EGElad Gil
  Mm-hmm
21. MJMax Jaderberg
  ... on the, like, core simulation models like AlphaFold+, ++, but also on the core reasoning models, like taking all of the frontier models and, and open source models and, and adapting them to our space.
22. EGElad Gil
  Mm-hmm.
23. MJMax Jaderberg
  Um, so that's, I think, a really important area for us to be hiring on.
14:35 – 19:31
Bringing Drug Discovery Onto Computers
1. EGElad Gil
  What are the components that you tackle? Or what are the main areas that you all are focused on in terms of how you approach new drugs or, you know, you-- Do you have your own labs? Do you run ro- robots that do pipetting? Like what-
2. MJMax Jaderberg
  Yeah
3. EGElad Gil
  ... what is the approach you're taking?
4. MJMax Jaderberg
  Yeah. So, um, you know, if you, if you go into our like, uh, sort of our core AI team and, uh, like our North Star and the thing that we're really trying to achieve is how can we bring all of that drug discovery work, all of that drug design work, understanding of, you know, disease biology onto a computer?
5. EGElad Gil
  Mm.
6. MJMax Jaderberg
  So how can we get out of the real world?
7. EGElad Gil
  Mm-hmm.
8. MJMax Jaderberg
  Um, you know, I, I know a lot of people are, uh, building ways to be, like, interfacing with the real world really, really efficiently, so you can do lots and lots of real world itera- iteration.
9. EGElad Gil
  Mm.
10. MJMax Jaderberg
  That's great, and it's complementary. But fundamentally, the, the big time bottleneck, the big cost bottleneck is synthesizing real atoms [laughs]
11. EGElad Gil
  Mm-hmm
12. MJMax Jaderberg
  ... and then testing them out in real assays and, you know, uh, experiments. And, you know, just like we've seen, I think, pretty much every other area of engineering move into a computer, uh, computer-aided design scenario, and that's just changed the way that, you know, changed what we can design, changed what we can engineer, um, we wanna do the same thing in this space.
13. EGElad Gil
  Mm-hmm.
14. MJMax Jaderberg
  And so really we-- like, when we think about the core AI drug design engine development side, this is really focused on, well, how can we achieve experimental accuracy-
15. EGElad Gil
  Mm-hmm
16. MJMax Jaderberg
  ... for our simulation models, our world models? If you can get a predictive model, you know, let's use AlphaFold as an example. AlphaFold looks at structure prediction, s- the pr- the, the structure of proteins and how they interact with each other. Normally, that takes, you know, months or years to do X-ray crystallography and, and, and get a structure out. If you can do that with a simulation model like AlphaFold, then you don't have to go into the real world and do that. You save months and months of time. And there's many, many-- You know, drug discovery is a, is a combination of many, many different really hard scientific problems and engineering problems. And if you can bring the-- you know, create predictive models instead of the real world's, then, you know, you're saving tons and tons of time. So internally, we have, you know, a dotted line on all of our leaderboards that is experimental accuracy.
17. EGElad Gil
  Mm.
18. MJMax Jaderberg
  And we-- You know, the goal of all of these, like, superpowers that we're trying to create is to get to a point where it's experimental accuracy. At that point, you don't need to go into the real world.
19. EGElad Gil
  Mm-hmm.
20. MJMax Jaderberg
  You can do so much more design and iteration on a computer, so that when you do go into the real world, when you make your molecules, when you test out your scientific hypothesis-
21. EGElad Gil
  Mm-hmm
22. MJMax Jaderberg
  ... the probability of success is much, much higher than it would have been traditionally. And so you don't have to do as many turns of the real world, and things actually become quicker and problems fall easier.
23. EGElad Gil
  Mm. Well, so when I look at drug development, um, I think it's about a fifteen-year process now.
24. MJMax Jaderberg
  Mm.
25. EGElad Gil
  It costs one point five billion dollars. My sense is the small molecule and the in vitro work, the tissue culture and mous- mouse work is basically, say, two to three years and then it's, say, fifty million dollars or that order of magnitude. And then the other one point four five billion and twelve years-ish, uh, is the clinical trial and, you know, the different phases and safety. And, um, how much of that do you think you can actually encapsulate in a computer?
26. MJMax Jaderberg
  You know, you have the discovery phase where you go from idea to you think you've got a good, uh, like, a good molecule that you could put in patients.
27. EGElad Gil
  Mm-hmm.
28. MJMax Jaderberg
  Well, before that point, you have to do preclinical, you know, animal studies to make sure that this is safe and working as intended. That takes another year. Then you go into clinical studies. You have phase one, phase two, phase three.
29. EGElad Gil
  Mm.
30. MJMax Jaderberg
  Ratchets up. And so, you know, when you look at the timeline and when you look at the failure rates-
19:31 – 20:30
Isomorphic’s Datasets
1. EGElad Gil
  do you get the data to model those things? Like, where does that dataset come from?
2. MJMax Jaderberg
  You know, we have this, like, ridiculously comprehensive, you know, da-data strategy. There's data all over the place. But I think fundamentally what we've learnt over the last, you know, five years is there's a lot of opportunity to generate this data yourself.
3. EGElad Gil
  Mm.
4. MJMax Jaderberg
  Um, there's of course, like, big repositories of data. Sometimes that's really good for deep learning. Um, oftentimes, these repositories of data, you know, whether they're public or in pharma or, you know, um, in companies, they've been the byproduct of many years of passive, you know, data collection from doing, you know, the activity of, of making medicine. So they can be very, very biased-
5. EGElad Gil
  Mm-hmm
6. MJMax Jaderberg
  ... which often doesn't really help you if you're trying to make a general solution to drug discovery.
7. EGElad Gil
  Mm.
8. MJMax Jaderberg
  If you really care about generalization out of these models, there's maybe a different distribution of data you need.
9. EGElad Gil
  Mm.
10. MJMax Jaderberg
  So we go out and generate this ourselves. You know, we've, we-- and we've seen some really, really amazing, uh, data scaling laws as a result.
20:30 – 22:50
Rethinking Endpoints
1. EGElad Gil
  How do you measure whether, uh, some of these things will have these side effects? Because ultimately, there's the old saying-- I used to be a-- I have a PhD in biology and history-
2. MJMax Jaderberg
  Yeah
3. EGElad Gil
  ... in a variety of fields, cancer, neuroscience, et cetera. And on the cancer side, we cured cancer in mice hundreds of times, and those molecules wouldn't translate into people. And so a lot of the data you actually needed, you had to generate in human clinical trials.
4. MJMax Jaderberg
  Mm.
5. EGElad Gil
  How do you go about generating that data that actually informs it? Because if you're using animal models, often the data that comes out of that is incorrect relative to human biology.
6. MJMax Jaderberg
  Well, exactly. It sounds like you experienced a lot of this firsthand.
7. EGElad Gil
  It's very painful [laughs] .
8. MJMax Jaderberg
  It's very painful. And you know, if you think about, um, you know, toxicity, like sources of toxic-- like the gold standard that, you know, you need to get through for regulators is you need to prove that your molecules aren't toxic on rats and other animals. Um, but still, even when things are approved-
9. EGElad Gil
  Doesn't matter necessarily. Yeah
10. MJMax Jaderberg
  ... and they've been approved on rats-
11. EGElad Gil
  Yeah
12. MJMax Jaderberg
  ... they end up being, you know, often toxic in humans.
13. EGElad Gil
  Mm-hmm.
14. MJMax Jaderberg
  And, well, why is that? Well, okay, a rat isn't actually a human. It's not actually very predictive. I think, you know, choose your study, but it's about thirty percent predictive of, of, of human clinical tox, like in a rat. Um, and you know, if, if one of my, like, AI scientists or engineers came up to me and said, "Hey, we've got this amazing model. It's, it's, it's-"
15. EGElad Gil
  Yeah
16. MJMax Jaderberg
  ... "thirty percent predictive." You're like, "Okay."
17. EGElad Gil
  Yeah, yeah.
18. MJMax Jaderberg
  "Can maybe do a bit better than that."
19. EGElad Gil
  Yeah.
20. MJMax Jaderberg
  Um, and so and I think there's a lot of opportunity to actually rethink how we, you know-- like what are those predictive endpoints for human clinical tox? Like forget animals.
21. EGElad Gil
  Yeah.
22. MJMax Jaderberg
  You know, they've been a proxy for the last, you know, fifty, seventy years.
23. EGElad Gil
  Uh-huh.
24. MJMax Jaderberg
  Um, but we can maybe do better. We can do better by, um, you know, bringing together in vitro endpoints. So things you could measure quite easily in the lab together with, you know, AI predictive models. Also together with this-- there's some amazing new, um, experimental technologies coming out in the space of organoids, organs on a chip, like synthetic human tissues-
25. EGElad Gil
  Yeah
26. MJMax Jaderberg
  ... which sort of approximates and create these like little microcosm environments of human organs.
27. EGElad Gil
  Mm-hmm.
28. MJMax Jaderberg
  And so you can imagine there's probably an opportunity to bring all of this stuff together-
29. EGElad Gil
  Mm-hmm
30. MJMax Jaderberg
  ... and then directly try and model human clinical tox. Now, there's not much data on that side-
22:50 – 23:57
External vs. Internal Dependences
1. EGElad Gil
  of approaches or techniques, do you tend to partner for them? Do you bring them in-house? Like, I'm a little bit curious as a company how you operate relative to-- 'cause there's so much that could be done.
2. MJMax Jaderberg
  Yeah.
3. EGElad Gil
  And a little bit curious how much you wanna just emphasize the computer science side and then, you know, biology, often there are these things called CROs or these contract research organizations that people use. Unfortunately, the quality tends to be pretty mixed in terms of the output of these types of organ-- I'm just a little bit curious, like how much you depend on external parties versus internal.
4. MJMax Jaderberg
  Yeah. So like making our drug design engine, all of the AI models, we do that all in-house. Um, and you know, we see a big opportunity to generating proprietary data, um, and owning that ourselves. But, you know, biology and, you know, experimental platforms like this is really complex and long-tailed, and there's some amazing people working on some fantastic platforms here. So yeah, we work with the whole world, whether that's, you know, stuff we do ourselves, contract research organizations, CROs, um, other innovative companies-
5. EGElad Gil
  Mm-hmm
6. MJMax Jaderberg
  ... um, to actually help us create the data that is, uh, you know, powering our models.
7. EGElad Gil
  Yeah.
8. MJMax Jaderberg
  Uh, so yeah.
9. EGElad Gil
  That's cool.
23:57 – 26:22
Automating Biologic Labs
1. EGElad Gil
  How much, uh-- So Anthropic recently announced that they're doing a wet lab in San Francisco. And, you know, reading between the lines without any special knowledge, just reading external papers, it kinda looks like there's some potentially robotic system or closed loop system or something else they're trying to build. And in general, if you look at AI models, as you know better than me, um, you want closed loop systems to be able to iterate quickly.
2. MJMax Jaderberg
  Mm.
3. EGElad Gil
  In other words, if you can get rapid feedback, like if you're writing code and it can instantly get tested, then you have a fast way to test if the model's working well or not, right? And so in biology, one of the issues is you set up an experiment, you wait three weeks, then you get the results. You have to interpret those. There's a big loop, and so that slows down progress and the ability to iterate on data. And building these r- these robotic platforms or other things may e-expedite that. Um, how do you approach that problem, or do you think that's part of the solution that you all will be pursuing in terms of-
4. MJMax Jaderberg
  Yeah, I mean, it's, it's been really exciting to see, like, how much better these, you know, automated labs, how, how much better these robotic solutions for experimental labs have been getting. And, you know, we've be- we've been working with, um, these sort of automated labs for a long time now. Everyone should, and we should keep driving down that, like, iterate, iteration time with the real world. And fundamentally, I wanna get out of the real world, right?
5. EGElad Gil
  Mm.
6. MJMax Jaderberg
  And, and, you know, can we actually... Because there's gonna be just like you're gonna hit the finite laws of physics at some point. Like, how long does it take to synthesize a molecule?
7. EGElad Gil
  Mm.
8. MJMax Jaderberg
  At some point, you have to just wait for the reaction to happen, and that's how long the reaction takes. How long does it take to culture a cell?
9. EGElad Gil
  Mm.
10. MJMax Jaderberg
  At some point in time, you just have to wait those days for the cell to get to the right state before you can test things on that. And so yes, we can, you know, I think do a lot of work, and we should, and we are a-a-a- as a as a species as well, is on, on, on, you know, improving automation and improving the way we do things.
11. EGElad Gil
  But you want to bring it in silico, it sounds. You want to bring it in a computer.
12. MJMax Jaderberg
  But I, but I think we should also be driving really hard at bringing that in silico-
13. EGElad Gil
  Yeah
14. MJMax Jaderberg
  ... on a computer so that we can reduce the amount of real world time so that we can do much more predictive work on a computer, so that when we do go into the real world, yeah, it looks like we go into real world to validate-
15. EGElad Gil
  Yeah
16. MJMax Jaderberg
  ... not necessarily to learn and iterate.
17. EGElad Gil
  Mm.
18. MJMax Jaderberg
  We still need to go into real world to validate because we're coming up with novel, you know, scientific hypotheses, and we want to test them out on people at the end of the day. So you have to make sure that they work and they're safe. So you have to go into the real world to validate. But the question is, how much can you reduce the amount of time you go into the real world to experiment and to iterate?
19. EGElad Gil
  How
26:22 – 27:19
Internal Development vs. Partnerships
1. EGElad Gil
  much of your approach is focused on partnering with existing pharma companies versus developing your own drugs?
2. MJMax Jaderberg
  We make the AI drug design engine in-house, and then we, we actually, you know, don't sell the model. We like-- We, we don't license out the model. We don't sell tokens in that way. We use it internally to, to make medicines.
3. EGElad Gil
  Mm-hmm.
4. MJMax Jaderberg
  And then we do that in two ways. We have programs, drug discovery programs-
5. EGElad Gil
  Mm-hmm
6. MJMax Jaderberg
  ... that we fully own and fund ourselves and work on ourselves, our internal pipeline. And then we have programs that we work on with, you know, amazing pharma companies, biotech, um, where we partner on these, on these projects. And, um, those are really exciting as well because often those companies, uh, and those teams, they've got, you know, amazing ideas and amazing expertise in specific disease areas. And so they come to us with like-
7. EGElad Gil
  Mm-hmm
8. MJMax Jaderberg
  ... these are the hardest problems, maybe problems that they've been working on for sometimes ten, twenty years-
9. EGElad Gil
  Wow
10. MJMax Jaderberg
  ... not making much progress. And then together, using our drug design engine, we can make a huge amount of progress.
27:19 – 29:21
Areas of Human Health Focus
1. EGElad Gil
  What are the indication areas or areas of human health that you all are focused on?
2. MJMax Jaderberg
  Internally, we're focused on m-- you know, much more on oncology, immunology, and inflammation.
3. EGElad Gil
  Mm-hmm.
4. MJMax Jaderberg
  Um, but the cool thing is, you know, the drug design engine or the models, they're general. They're disease area agnostic.
5. EGElad Gil
  Yeah.
6. MJMax Jaderberg
  They're target class agnostic. They're modality agnostic.
7. EGElad Gil
  Yeah.
8. MJMax Jaderberg
  So they work across multiple modalities.
9. EGElad Gil
  But you're starting with, like, cancer and is it autoimmune disease or something else? Or...
10. MJMax Jaderberg
  Yes. So, you know, um, immunology, inflammation, they're often quite related. Um, so it's that sort of space. And when we go after these things internally, whether it's oncology, you know, cancer, inflammation, y-you know, I think there's a big opportunity to go after some of these root node problems. The, the sort of things that could affect, you know, positively if you have a solution, could affect, you know, lots and lots of different patients for different downstream indications.
11. EGElad Gil
  Mm. Mm-hmm.
12. MJMax Jaderberg
  Um, so that's where-
13. EGElad Gil
  What would be an example of that?
14. MJMax Jaderberg
  Um, I mean, if you, if, if you think about cancer-
15. EGElad Gil
  Mm-hmm
16. MJMax Jaderberg
  ... like this is a really, really complex and long tail disease, and it shows its itself up in, in patients in many, many different ways.
17. EGElad Gil
  Mm-hmm.
18. MJMax Jaderberg
  But we also know there are some-- like maybe there are some core driving mechanisms that are common to many different things.
19. EGElad Gil
  Sure.
20. MJMax Jaderberg
  Or, um, you know, we saw this amazing example earlier this year, um, you know, Firaxstan, like K RAS inhibitor-
21. EGElad Gil
  Mm
22. MJMax Jaderberg
  ... standing ovation at a big canv- cancer conference, like drastically improving, um, you know, the, the, the, the lives of, of, of sort of like late stage pancreatic cancer patients from Revolution Medicines. And, you know, again, great example of like going after some sort of root nodes.
23. EGElad Gil
  I see. Yeah, RAS has been popping up in the cancer literature for like thirty years.
24. MJMax Jaderberg
  Exactly.
25. EGElad Gil
  Or twenty years. Yeah.
26. MJMax Jaderberg
  But these, these problems are devilishly hard-
27. EGElad Gil
  Mm
28. MJMax Jaderberg
  ... from a like drug discovery and design-
29. EGElad Gil
  Yeah
30. MJMax Jaderberg
  ... point of view. Like just-
29:21 – 32:17
Multidrug Approaches
1. EGElad Gil
  It's actually an interesting point because, um, for a lot of different drugs, uh, for a lot of different areas of health, this also was HIV, right? You end up with a cocktail of drugs that you give eventually, and that's actually what allows you to suppress AIDS or HIV. In cancer, similarly, a lot of drugs which were developed as, as individual drugs versus cocktails.
2. MJMax Jaderberg
  Mm.
3. EGElad Gil
  And to your point, often within a tumor, you have a spectrum of mutations that you need to suppress simultaneously if you really wanna cure the thing, 'cause otherwise it'll sort of-
4. MJMax Jaderberg
  Yeah
5. EGElad Gil
  ... evolve out under, given the sort of substrata of mutations that exist in a distribution across the cells. Are you thinking about those more complex approaches of cocktails or multidrug approaches or-
6. MJMax Jaderberg
  Yeah. So, so why aren't we giving more cocktails-
7. EGElad Gil
  Yeah
8. MJMax Jaderberg
  ... drugs? Well, the problem is like often these molecules, they are-- they work, but they have some toxicity.
9. EGElad Gil
  Mm-hmm.
10. MJMax Jaderberg
  Or they exhibit some effect on the liver, such that if you're taking multiple of these molecules, these medicines at the same time, they have these drug-drug interactions that can be really harmful for patients.
11. EGElad Gil
  Mm.
12. MJMax Jaderberg
  And so it's actually because we're not great at engineering exquisite molecules and medicines, that means that we can't combine them together in a safe way for patients. And so we're maybe limiting like the, the surface of applicability.
13. EGElad Gil
  Yeah. I guess you only have so many patients too, and so many people who can enroll a clinical trial. And so the combinatorics of it, my sense is also kinda adds up in terms of the clinical trial side.
14. MJMax Jaderberg
  Especially when there's a lot of, um, like risk and uncertainty about how these things will combine. And so that's where we see an opportunity. Yeah. If you can actually Much more accurately engineer these molecules to be safe, to not have these effects on the liver, to not, uh, interact with each other in bad ways, then you can feasibly come up with actually, I would really hope some quite complex, you know, combinations-
15. EGElad Gil
  Mm-hmm
16. MJMax Jaderberg
  ... that can modulate not just a single target at a time, but a whole vector of targets that span these, like, very complex biological pathways at the same time and so have a better effect on disease.
17. EGElad Gil
  There's a very good YouTube video of, uh, this guy, Janssen, who started Janssen Pharmaceuticals. It was acquired eventually by J&J, and he was one of the big pioneers of drug development in pharma, and I think he was one of the, um... They had one of the most productive pharma companies in terms of actual drugs launched. And in it, he- his view that he expresses pretty strongly is that the main constraint on drug development is an overemphasis on safety.
18. MJMax Jaderberg
  Mm.
19. EGElad Gil
  And that regulators kind of curtail progress, and he pointed out how during World War II, Winston Churchill tried to find a cure for soldiers of some venereal disease, uh, in the field. And so they, they rediscovered, um, antibiotics, because penicillin was discovered in the '20s by Fleming and then kinda put on a shelf and forgotten. And then in a nine-month period, they rediscovered and, um, commercialized, uh, penicillin from scratch specifically to treat this. And it was because we removed all regulatory obstacles. So that was an example that Janssen pointed to. One could argue the COVID vaccine was similar-
20. MJMax Jaderberg
  Exactly
21. EGElad Gil
  ... in terms of being able to rush it. Um,
32:17 – 34:26
Risk Appetite Variation
1. EGElad Gil
  how much do you think there's too much of an emphasis on safety? Like, Janssen's point was it's about risk-reward.
2. MJMax Jaderberg
  Mm.
3. EGElad Gil
  You know, you need to make trade-offs and you can't just monolithically say it's only safety. You should also say, well, it's about benefit, and in some cases risk should be taken as well. Like, how do you think about that?
4. MJMax Jaderberg
  I think it's, like, centrally important to just always have patient safety in mind, to be honest. Um, and but then you need to understand, like, what are the risks and the appetites of patients, and that can be different for different indications.
5. EGElad Gil
  Mm-hmm.
6. MJMax Jaderberg
  Uh, you know, COVID's a good example where the risk appetite for patients-
7. EGElad Gil
  Mm-hmm
8. MJMax Jaderberg
  ... because it was so pervasive and so dangerous just changed dramatically.
9. EGElad Gil
  Mm-hmm.
10. MJMax Jaderberg
  And that changed, I think, the demand and the force-
11. EGElad Gil
  Mm-hmm
12. MJMax Jaderberg
  ... for doing things slightly differently. You even see this across different disease areas. If patients have, you know, late-stage cancer-
13. EGElad Gil
  Mm-hmm
14. MJMax Jaderberg
  ... and not many options in front of them, there's a very different risk appetite-
15. EGElad Gil
  Yeah
16. MJMax Jaderberg
  ... compared to if you're trying to treat a chronic low-level, you know, uh, inflammation, um, uh, indication. And so I think it's actually really, really important that, like, we keep focusing on patient safety-
17. EGElad Gil
  Mm-hmm
18. MJMax Jaderberg
  ... but that we can modulate that, you know, depending on what the patient need is. So ultimately, you know, ultimately, we want patients who are living healthier lives, and they're receiving that treatment as quickly as possible.
19. EGElad Gil
  Mm-hmm.
20. MJMax Jaderberg
  And, you know, we, we, we speak a lot to regulators here, um, you know, FDA here, but also all over the world, and I think the one thing in common is, like, you know, people really want to get patients the best drugs and the best solutions possible.
21. EGElad Gil
  Mm-hmm.
22. MJMax Jaderberg
  Um, and I do think that over time we will see a difference in our, you know, probab- you know, the success rates and the safety profiles of molecules coming out when they're designed with much more understanding about how they work-
23. EGElad Gil
  Mm-hmm
24. MJMax Jaderberg
  ... with, you know, we're being able to search the design space to engineer something, you know, much better than you could do from an empirical- empirically driven process. And I hope to prove out over time that this is very stat- sta- statistically s- uh, different.
25. EGElad Gil
  Mm-hmm.
26. MJMax Jaderberg
  And that will change the way that we think, I think, about, about regulation and, and getting drugs to patients.
34:26 – 35:42
Disease Treatment Outlook
1. EGElad Gil
  When do you think you'll have your first clinical trial?
2. MJMax Jaderberg
  I would love to give you a, like a, a hard date. This is-
3. EGElad Gil
  Yeah
4. MJMax Jaderberg
  ... like, really hard frontier science.
5. EGElad Gil
  Yeah.
6. MJMax Jaderberg
  Um, we've got some really exciting, um, medicines in the works.
7. EGElad Gil
  Mm-hmm.
8. MJMax Jaderberg
  Um, and th- these are going really well.
9. EGElad Gil
  Mm-hmm.
10. MJMax Jaderberg
  So we're pushing really hard to get these to patients as quickly as possible. We've been building up our clinical development team, so we're getting ready for this and, uh-
11. EGElad Gil
  Do you have a-- I know these are always impossible questions.
12. MJMax Jaderberg
  Yeah.
13. EGElad Gil
  Do you have a prediction on when you think, for example, 80% of human cancers will be treatable or, you know, some other metric of why this is so important for the world?
14. MJMax Jaderberg
  I'd be really disappointed if we fast-forward, you know, even five years' time, if we don't have the base tools like the, like the predictive understanding of disease biology has not step change in, like, a massive way.
15. EGElad Gil
  Mm-hmm.
16. MJMax Jaderberg
  Um, our ability to design molecules-
17. EGElad Gil
  Mm-hmm
18. MJMax Jaderberg
  ... that are much more precise, much less toxic-
19. EGElad Gil
  Mm-hmm
20. MJMax Jaderberg
  ... has not step change in such a massive way that, you know, still drug, drug development and discovery and development takes years.
21. EGElad Gil
  Mm-hmm.
22. MJMax Jaderberg
  And that'll reduce over time, but it takes years now. Um, but we should have the tools in place to really start be- like, start cracking these large, like, large portions of disease-
23. EGElad Gil
  Yeah
24. MJMax Jaderberg
  ... in a different way.
25. EGElad Gil
  Um,
35:42 – 39:16
Human Longevity and Drug Development
1. EGElad Gil
  if you remove, uh, if, if you removed cancer from society and heart disease, I think the stats suggest that the average person would live, I can't remember if it's five or seven years longer. And so the actual impact to longevity for two of the biggest killers, um, is on some-- in some sense, quite large, and another point of view, it's kinda small. Uh, there's a number of companies that have tried to focus on aging or longevity as a more general process, which seems to clearly be genetically programmed, right? If you knock out certain genes in C. elegans or Drosophila or in mice, they'll live 20% to 30% longer. Or in the case of C-, uh, C. elegans, the, the little worms, they'll live three times longer, right, if you knock out certain genes and certain pathways. I actually worked on those pathways as part of my PhD.
2. MJMax Jaderberg
  Amazing.
3. EGElad Gil
  Uh, it was really fun. Um, how do you think about that in the context of human lifespan? Like, should we just hit a point eventually where people live three times longer, five times lo- like, what-- how do you think about longevity in the context of drug development and, you know, all these tools that will now exist?
4. MJMax Jaderberg
  Yeah. No, it, it, it's interesting. Like, there, there are some of these major diseases that afflict so many people-
5. EGElad Gil
  Mm-hmm
6. MJMax Jaderberg
  ... that if we could start to make impact on those, that, that would have a concrete impact on, on how long people are living. I think even outside of, um, you know- Um, things like cancer and heart disease. You just have the, um, sort of chronic afflictions-
7. EGElad Gil
  Mm-hmm
8. MJMax Jaderberg
  ... uh, of aging that start to add up.
9. EGElad Gil
  And the root cause is literally aging.
10. MJMax Jaderberg
  I- yeah, I-
11. EGElad Gil
  In other words, like the root cause is, um, the fact that you're allowing these systems to degrade over time.
12. MJMax Jaderberg
  Grade. And then-
13. EGElad Gil
  Stem cell compartments-
14. MJMax Jaderberg
  But yeah, and they, and they-
15. EGElad Gil
  ... et cetera
16. MJMax Jaderberg
  ... combine in a way that means that people's ... You know, even if people live to a lovely long age of-
17. EGElad Gil
  Mm-hmm
18. MJMax Jaderberg
  ... 80 or 90, like they ... Maybe the last like 10 years of their life-
19. EGElad Gil
  Mm-hmm
20. MJMax Jaderberg
  ... or even 20 years of their life isn't not, isn't necessarily as great as it could be.
21. EGElad Gil
  Yeah, yeah.
22. MJMax Jaderberg
  Because they've, you know, because of, um, you know, your bones start to degrade, your muscles start to degrade. Like you have a fall and suddenly you can't, you're not as mobile as you want to be.
23. EGElad Gil
  Yeah.
24. MJMax Jaderberg
  That has impacts on your mental health and suddenly-
25. EGElad Gil
  Mm-hmm
26. MJMax Jaderberg
  ... your quality of life has dropped dramatically. You still might live-
27. EGElad Gil
  Yeah. I-
28. MJMax Jaderberg
  ... another 10 years
29. EGElad Gil
  ... in the, in the aging community, they talk about that, about as lifespan versus healthspan.
30. MJMax Jaderberg
  Exactly. So we think a lot about that.