Welcome to the Good Medicine Podcast. We're very excited today with our episode to have Chris Gibson here with us. I guess I should call you Dr. Chris Gibson, but Chris is great. Chris is the co-founder of Recursion Pharmaceuticals, and Chris holds a PhD in bioengineering from the University of Utah. And Chris and I met a long time ago when you were a student, and I knew Chris was destined for big things when I first met him. And Recursion Pharmaceuticals has really become a real leader in the use of artificial intelligence to enable drug discovery. And so we're so excited to have you here today and to have a discussion about, you know, there's a lot of hype out there about what AI is and what it does. And we hope to dig in and talk a lot about, you know, what is AI and how do you use it and how does it enable drug discovery. But before we jump into that, I think it's always great to hear about you personally and your background. I know it's, I find it fascinating of how everyone ends up where they are. So, you know, as a child, did you dream about being a scientist, or, you know, was your mom and dad, you know, so there's all kinds of stories. So I'd love to hear, you know, what is it in your life that led you to want to be involved in science and specifically what got you excited about being involved on the therapeutic side of things?
Yeah. Well, first, thanks for having me on. It's really exciting to be here, and I'm excited for everything that's happening in Utah. So this is a lot of fun. Boy, I was one of those kids who took everything apart when I was a kid and then tried to put it back together. I always loved to understand how things work, and I was pretty into science at a young age, but I thought I was going to be an astronaut. And then in, I think, sixth or seventh grade, I passed the height limit. So at least I heard at the time it was like six feet or 6'1" or something. I passed that, and so I had this like existential crisis in middle school, and I had to figure out what to do with my life because, like, astronaut was off the table. And, you know, I always was into science, and when I went to, when I was thinking about what to do, I talked to my dad a lot. He had built a company starting two weeks before I was born, and he was at the interface of electrical engineering and business. And so he built a company that was based on some of the first parallel processors that came out of Intel. And so I got to see that happen as I grew up. And I was like, I kind of want to do something like that. But I thought that the next big wave was going to be in bioengineering instead of electrical engineering. And so that's what I went into in school. And a long serendipitous path of how I ended up in Utah. But I ended up in the lab of a guy named Dean Lee, who you knew. And it was a project in Dean's lab that led me to join you up at the Huntsman for some work together in drug discovery. And one thing led to another, and Recursion came out of that. So a long, long path, a lot of serendipity. But I think in some ways always destined to be in the interface of different fields. I'm not very good at any one thing. But I can do a few things reasonably well, so it works out.
Well, I can completely understand that, and I feel the same way. So that's exciting. So you're a graduate student. You're working in a really great lab with a great mentor. What was the beginnings of the thoughts? So you kind of had this background where you were exposed to your dad being an entrepreneur, you know, as being a trained scientist. We really don't get trained to think about starting a company as kind of a career path. I know I was the world's most reluctant entrepreneur. I, you know, was a faculty member and had a grant and all this stuff, and, you know, that's the world I knew. So what was it that kind of flipped the switch for you to say, hey, I could actually take this and turn this into something that could actually be a company that I could build around?
Yeah. I had actually in grad school gotten a little bit pulled away from the entrepreneurial side. And I was doing an MD-PhD, and I thought I was actually going to be a CT surgeon. Like in Dean's lab, one of the things we got to do was hang out with other people in the lab, and there were a few cardiothoracic surgeons there. And so I would go with the pager at two in the morning when they did a heart transplant, and I would get to basically just watch and sometimes help collect tissue from, you know, the hearts that came out of patients. And I thought, like, entrepreneurship is really cool, but I'm actually, this is my future. And then we were working on a project in Dean's lab on a rare genetic disease called cerebral cavernous malformation. It was one of many projects that I got to be a part of as an engineer in his lab, and ultimately the traditional approach that we took failed. And so we decided to try, like, a different approach and use some basic machine learning classifiers to look at images of human cells and see, you've got images of human tissue here. So we used a machine learning classifier and made some progress in a place where we hadn't made progress before, and that just felt like it was something we had to pursue more. And so in speaking with Dean, he arranged for me to go to Stanford for the summer and do sort of a mini entrepreneurial project there. And that just snowballed into what would become Recursion. And it felt like it was one of those things I was going to regret if I didn't try. It felt like the most likely outcome was failure. And so worst case is that I was going to lose a year before I went back to med school, but I was going to gain a lot of experience. And so we took the shot and I took a leave of absence from medical school in 2013. And like by 2016 or 2017, I wasn't back, and they made me formally drop out of the program. So I'm a med school dropout too, which plays very well in Silicon Valley.
I bet it does. You're raising money. Well, that's wonderful. So tell me about, so, you know, I've always been involved in starting companies and putting that pitch together of this is what's unique about us and our approach. So as kind of a pioneer, really a pioneer in the area of adapting machine learning and artificial intelligence to drug discovery, what was the original pitch that you came up with that got people interested, and how did that morph over time, and how did it resonate with the early investors that came in to help you get going?
Yeah, it was interesting because there were not a lot of investors at the interface of biology and technology at the time. So we were kind of like looking for biology investors or technology investors who could see across fields. What was super interesting is that in 2012 there was this thing called ImageNet that was released. Google did this really the first big sort of machine learning project on images, and it created a renaissance in computer vision. So can you essentially take an image and use a computer to extract really meaningful, rich representations? And so that was happening in 2012. Recursion started in 2013. And we had this evidence that maybe images in biology were important. And so the pitch was that in biology, structure suits function, and if we can generate a massive library of images of biology and we can get really good at doing computer vision, we could decode biology. And the business pitch was that we were going to repurpose medicines. So there's all these thousands of medicines that people have tried to take to the clinic or to the market that have failed for a variety of reasons, and we're going to find what they should have been designed for. We're not doing that much anymore. So the pivot was that it turned out most of these medicines were shelved for pretty good reasons. We found a few interesting indications, and we have clinical readouts coming for some of those in the coming quarters. But we found less opportunity at the interface of shelved assets and rare genetic diseases, which is where we were looking, for a wide variety of reasons. And so the way we started morphing in those early days was we started doing our own chemistry too. We found interesting targets or interesting pathways, and we figured we had to make the right chemical at that time. And so that's been a journey. The scope of what we have to build keeps getting bigger. And so, you know, we never thought we'd do our own clinical trials. Now we're doing a whole bunch of our own clinical trials. We never thought we'd have our own supercomputer. Now we have our own supercomputer. So it's become just a bigger and bigger endeavor. But the core idea that you can use computers to extract meaningful representations of biology from data, scaled data assets, that's held true. And it's been a lot of work, but we're here eleven and a half years later still doing that same basic thing.
That's incredible. And I think it's, you know, it's fascinating because it is one of the central tenets of biology is structure matches function. And that's not just at the molecular level, that's at the cellular level, which is really what I mean back when physicians used to like really take time with their patients. I mean, I read a lot of William Osler and some of the incredible learnings that physicians back in the day could take just from looking at little things, right? The little dots on your lips or your tongue or the curvature of your nail beds. These sorts of things tell us so much about biology. And I guess it's not that surprising to imagine that the same thing could hold at the level of cells.
Yeah. And that we found it to be true. Yeah, I mean, we've now knocked out every gene in the genome in many different cell types. We're not only doing that work for ourselves, but doing that work with companies like Roche, Genentech, and Sanofi and others. But it also turns out biology is super, super complex. And so like everything, I think we're making incredible progress, but it's a very, very dense, challenging, complicated map that we're trying to build.
Yeah. And that's, I think that's the biggest challenge because, you know, most scientists want to take the simplest thing. And, you know, so we find a single gene or a single protein that we think is changed and we put that in a test tube and study that in isolation. But the reality is biology is never one protein acting by itself.
I think that's a great way to do it too. I mean, it's, you know, in algebra you try to minimize the number of variables and that's how you solve the equation, right? And so like that is a very, very viable path and it's how much of medicine, there's sort of been two paths, like that path or kind of serendipity and luck, like you drop the wrong test tube in the wrong vial, something happens, right? That's kind of how most medicines came to be. And it felt like technology was getting to a point where, and the tools in biology were getting to a point, like CRISPR were coming out, and then also the computational side of things. I mean, if you were to do what we're doing today at Recursion fifteen years ago, you'd have to spend a billion dollars just on compute and data storage. So it just wasn't feasible. And so all these things were starting to come to fruition at the same time. And the question a lot of people ask, I think now many people think this, like technology-driven approach will be an important part of the way many medicines are discovered in the future, not all. And the question is, like, were we on one wave too early? Were we on the right wave? We're probably not too late. Like we're still one of the earliest ones in the space. So it's a question of too early or right at the right time. And now there's hundreds of companies and most large pharma that are putting real resources into this idea of leveraging technology to look at large, high-dimensional data to take a more complicated view of biology. And in many ways, you still need all those tools of sort of this reductionist approach. And putting those two things together, I think there's going to be hopefully an explosion of more medicines or better medicines in the coming decades.
That's what all of us want. I think that's exactly right. And I think that, you know, the idea that we can add on to kind of the traditional way of doing things with something that's, because the hard thing about biology is it's so complex, you can get lost in that complexity really easily. But the way that Recursion's taken that approach to say we're going to start with that, that if when we have a disease cell or a disease, you know, diseased tissue, it looks different, right? And that look is actually the structure of that cell actually tells you something about what's going on in the function. So I love that idea that that helps reduce the complexity of biology. You say this is one thing we can track and we can measure and we can say this is the result of a bunch of complex biology, but it's something that we can wrap our heads around, right? Because a lot of times, you know, you've looked at it, I have, you know, transcriptomics and RNA-seq data. I mean, it just blows your mind. Like it's like, I don't know how to make sense of this stuff. But I love the approach that you've taken where it's something that at least I can conceptualize to say all this complex biology results in this and I can measure this. And correct me if I'm wrong, but I think what your approach to drug discovery is really, can I correct that to make that dysfunction look normal again, right? So if I can take a cell that has taken on this abnormal function and abnormal structure, can I restore it to look like it's normal again and therefore I should have fixed all that biology that was messed up.
Yeah, that's exactly how we started and what we still do. What we've added to that is not just these images, but also transcriptomics and data from patients. All this rich, high-dimensional data, which machine learning and AI are actually pretty good, if you set it up right, at looking across really large data sets and finding these patterns. What it turns out happens is if I take an image of one cell and one disease state, you know, I can describe how that cell looks different. It turns out computers are way better at that than we are. They can see way more subtlety than we can as humans. But I don't really know what's happening. I know that the cell looks different. Maybe I broke a gene, it looks different, and now I add drugs and maybe one of them makes it look healthy. That seems good. But as you start doing this across millions of different experimental conditions, millions of different small molecules, millions of different genetic perturbations, you actually kind of build this lookup table. And this is the part that gets most exciting to me when I talk about mapping biology is now instead of saying on a disease-by-disease basis, I can ask if I break a gene and the cell looks different in some way, across all other ways I could break genes, do any cells look really similar? And that can actually start to tell you something about the proximity of these really complex pathways, like does gene A interact with gene B? And you can start taking this a step further and asking, can I predict essentially what all of the pathways in biology look like in various cell states, and can I then correspondingly place small molecules or large molecules in all of those different areas of biology? And that's kind of why this experiment of Recursion has grown so much with supercomputers and like massive automated labs is because we saw the opportunity to maybe find some really exciting medicines if we were doing this rescue the disease state, but moving into this mapping and navigation approach.
This is like a boil-the-ocean project.
Yeah, for sure. I mean, with Roche, Genentech, we had to make like over a trillion neural iPSC cells in order to build this neural map. We think we became the world's largest producer of neural iPSCs just to make that one map. Right? Like we called a supplier and they said it was going to be between seven hundred fifty million and a billion dollars to buy the cells from them, and our budget was like seven million dollars. So how do we do it? So we're like, okay, we'll do it ourselves. We'll grow it ourselves. And so like it's become this really big project, but the more you build these data sets, the more connections, not just we can see because it's too complicated, the computer can see. And then it gives us interesting associations that we can then take into these more traditional approaches and try and take medicines forward.
Incredible. BioHive is a nonprofit building the future of healthcare innovation right here in Utah. As one of the nation's fastest-growing life science industries, they bring together scientists, entrepreneurs, students, and companies to create breakthroughs that save lives and strengthen our economy. Whether you're in a lab or at a startup, BioHive is here to build, brand, and bring together a thriving life science community. And there's a place for you. Learn more, get involved, and see their latest happenings at biohive.com. So I guess that sets up perfectly with the next question I have, which is, you know, there's a lot out there and you were one of the first, and so you've seen this whole field kind of grow of how do you adapt these machine learning, and by the way, tell me what the difference between machine learning and artificial intelligence, because I think I know, but you probably know better than I do. But how do we adapt these approaches into biology and what is the potential? You know, you probably saw on 60 Minutes last week, you know, the CEO of DeepMind came out and said in ten years every disease will be cured. And you kind of hear that stuff and you and I probably sit back and say, I mean, that's a little bit of a bold statement, but I love the vision that says, you know, we're going to make a difference and I believe it. I think that we're at the beginning of a revolution that's just going to get bigger and bigger. But help us see through what the real promise is and what a little bit of the hype that's around, because obviously things get a little out of control sometimes, and I've heard it many times in my life that this is going to be the next thing, and it kind of doesn't ever live up to its potential.
So yeah, I mean, biology is pretty humbling. I've learned that. I've been guilty of some of these things. I mean, it's possible I was on a podcast roughly ten years ago where I said we'd have one hundred treatments in our pipeline. I think we've got like between us and our partners thirty or forty. So we're making progress, but definitely less than I expected then. And so, look, I think when you see the power of these technology tools, even back ten years ago when we saw the power, it makes it really hard to feel like there's not going to be an exponential shift in this field. And the question is where we are on that curve. And I think, you know, as bold and ambitious as the comments that Demis made were on that 60 Minutes episode, I do think we are somewhere in the range of seeing things that are that crazy different than they are today in some number of decades. Right? Whether it's ten years, I thought it would be kind of closer to now, and it's a little slower than expected. Demis thinks we're ten years away. I'm probably closer to Demis than I am to Derek Lowe, who wrote an interesting rebuttal to Demis' comments. You know, Derek's perspective, and he wrote about what I said ten years ago. And he's probably surprised to hear we're still around and kicking, but, you know, I think his comment was that ML and AI are not capable of generating any new knowledge. And I think that is completely false. I think they are now very capable of generating new knowledge. We see it all the time internally. I think our colleagues at Roche and Genentech see the new knowledge that is coming out of our work, Sanofi as well. So I think that's demonstrably false. Where we are on this curve is discovering and developing a medicine isn't one step. It's not finding a target or then finding a molecule. You've got to go through, you know, ADMET, translation, clinical trials, etc. And I think that's what's slowing us down. I think if you were to ask, pick one step in drug discovery, can we achieve what Demis or I thought was true? Yeah, I think we're on in ten years. Absolutely. The question is how do you stack that up with all of the hundred things that are required, as you know better than I, to actually make a medicine, right? You've done that. I've not yet done that. And so you've got to deploy these technologies across hundreds of steps and putting those all together and stitching them all together and making them play nice, and just the infrastructure required to build the data sets. Like that's a lot of work, and I think that's why it's taken longer than I expected. And biology is hard, but I think we're closer to Demis' vision than we are to Derek's.
Yeah, I would agree with that. I certainly agree with that. So I guess the question is, so developing all these tools and getting all the front end right, if the back end isn't ready for the push through, I mean, that's where you run into these obstacles. So you said it earlier that you didn't think that you were going to be running clinical trials and doing all these things, but now you are. So I guess the question is, are you finding yourself thinking about, well, how do I take the approaches that I've taken on the discovery side? Am I going to start making AI for clinical trials? Am I going to, you know, is that something that's on the horizon for you guys?
We are doing that now. Yeah, we're getting frustrated. I mean, you know, clinical trials are hard. There's not a ton of data, but there are some companies now who are starting to aggregate more data in the space. So a good example is we have a partnership with a company called Tempest who's taken a lot of data from oncology patients. And any one of those data sets by itself is maybe not that useful. But when you look across tens or hundreds of thousands of patients, you can start to see interesting patterns, like if I had a patient who had this particular form of breast cancer and they got this particular medicine, you know, what was the outcome for that patient. As you start looking across tens of thousands of patients, we start to see really interesting patterns. And when you then pair them up with this kind of perturbative biology platform that we've built and you put AI across all of those, we see the opportunity to do things like define the patient population better for your clinical trials. And then we also are starting to do clinical trial simulations with machine learning and AI, and that's helping us sort of be more adaptive in the way we drive. The challenge, of course, is each clinical trial is tens of millions of dollars. And so unlike an experiment in the lab that might be five or ten cents and you can try millions of different versions, clinical trials, you get one shot. And so it's a high-risk place to innovate. But we hired this woman, Nisha Khan, from J&J, who was their chief digital data science officer. She had been, I think, a visionary in applying these tools to not just clinical development but actually commercialization at J&J. And she's a trained chemist. She knows the front end as well, and she's been helping us build out the tools in that space. We're very early. And some of it is like not AI or machine learning at all. Some of it is like basic software engineering that's just automating how we email sites back and forth, so we don't have to have fifty people on our clinical operations team, right? But there's some low-hanging fruit, and very little has been built in the space, we were surprised to see. So this is the challenge. I mean, the infrastructure required to develop a medicine is massive, and if you want to deploy technology at many of those steps, it just takes time. But it's, you know, there's hundreds of companies now building, I mean, you're building companies like this as well.
