Latest / ThinkData Podcast / S4 | E6 | AI in Drug Discovery with Josh Haimson - Inductive Bio

Transcript

S4 | E6 | AI in Drug Discovery with Josh Haimson - Inductive Bio · ThinkData Podcast

• 0:03
  Alex Hutchings: Welcome to the Think Data podcast brought to you in partnership with Mydataworks. If you want to stay up to date with the latest breakthroughs and trends in the world of data and artificial intelligence, and if you're curious about some of the strategies that companies and founders use to launch data and AI products, then you're in the right place. Our aim is to bring together a diverse lineup of fantastic guests from the founders, through to accomplished leaders and product owners at some of the most fascinating data and AI companies worldwide. They will each offer you their own unique insight into what it takes to launch and scale a great data business. Thanks for tuning in, and I hope you enjoy the episode. Welcome to the Think Data podcast, and today it's great to welcome Josh Hameson to the show. Josh is the co-founder and CEO of Inductive Bio. They are a cutting-edge biotech startup that's using advanced AI to run virtual chemistry labs that help scientists predict how drug molecules behave in the body. And as well as, their solution actually helps accelerate the discovery of better therapeutics faster and more efficiently than ever before. And obviously someone that's coming from... this space. You work for a well-known company, obviously on the product side, pretty early on before launching Inductive, what, three, four years ago now. So I'm really keen to, I suppose, dig on the backstory because going from one early stage company to then kind of co-finding or co-founding your own company is a whole different kettle of fish. So yeah, over to you, Josh, what brought you to Inductive and first and foremost to this space within biotech?
• 1:51
  Josh Hameson: Yeah, absolutely. So before Inductive, I was at Flatiron Health. which was in the real-world evidence kind of clinical research space. So we were essentially trying to understand large-scale amounts of data to figure out which treatments were working and which were not for cancer patients. That company was an exciting ride and an exciting startup journey. You know, when I joined, we were around 100 people. When I left, we were around 1,500 people. So it scaled kind of very rapidly. And... Towards the end, we got acquired by Roche, which, of course, is one of the largest drug discovery companies in the world. And at Flatiron, we were really focused on kind of clinical stage research. So once you have a drug, how do you know if it's actually working for patients? But I found myself running a large AI group inside of a large drug discovery company. And there was so much hype at the time around AI and drug discovery, kind of set out on this journey to figure out what if that, if anything, is real and kind of ready to move the needle. before you get to the clinic. So I did a ton of research initially kind of within the four walls of Roche, Genentech, and really started to land on this area of molecular optimization as an area that was really ripe for kind of the current state of the art to make a really big impact. And decided didn't want to do that within the four walls of a large company. I really wanted to do that as a platform for the industry as a whole. And so kind of left the mothership and started Inductive, you know, a few years ago.
• 3:17
  Alex Hutchings: How was that journey? I know you obviously said you went on that journey, which obviously you weren't spending your own capital with. Obviously, when you're at Flatiron, you're this hyper growth business, 1,500 people at the point of leaving before acquisition. How did you manage that kind of early stage as a founder? Because a lot of people I have on this podcast, it's the excitement of becoming a founder. Sometimes it doesn't always marry up with the reality. But how was it for you?
• 3:44
  Josh Hameson: Yeah, I mean, to be clear, we started the company. after we had already left. So that journey was really kind of the research journey of saying, what anything can I do here to actually make a difference in this drug discovery process? I don't think it would have been possible, at least for me and the way kind of I work to have then also done kind of nights and weekends getting the company off the ground. And so that was really more about learning about the space and trying to understand where the opportunity was. And then we kind of left and kind of got started for real. It was, It was definitely a shock going from... At that point, we were part of a massive company to just me and my co-founder with a whiteboard trying to figure out how we actually get started. But it's been a fun ride.
• 4:26
  Alex Hutchings: Yeah, sounds it. And for people in layman's terms who are listening to this who may not be from this world, in practice, what does the solution do, the platform do, and what part does kind of artificial intelligence play within this?
• 4:41
  Josh Hameson: The way we think about what we're doing is... is this idea of a virtual lab, which you mentioned at the top. And really, if you think about how science is done today, it happens in a physical wet lab where you have scientists who are looking at data, trying to come up with scientific hypotheses, and then running physical experiments, making molecules, putting those molecules in cells, measuring the outputs. to see if those hypotheses work. And they do that iteratively. So they come up with an idea, they test it, they learn that it's partially right, partially wrong, and they do that over and over again. The process of drug discovery teams typically make thousands of molecules over the course of four or five years before finding that one needle in a haystack that really kind of meets all of the properties that could actually have a good shot of working in humans. So what we're doing at the virtual lab is really mirroring that wet lab process. where we have AI chemistry assistants. So these are essentially AI scientists that have learned a lot of the best practices of medicinal chemists and drug discovery scientists who do this work and can start to come up with hypotheses about interesting ideas to test. And we pair them with AI models of some of the key experiments that are happening in the wet lab. So that way they can run that design-make-test cycle virtually just by kind of scaling up with compute. And there's no replacement of you what's actually happening in the wet lab. You always need to make the molecule and run the experiment. But what this really allows you to do is scale up the scientific process by just adding more compute. And so one scientist, rather than coming up with a few ideas and testing them over the course of a few weeks, they can be paired with this virtual lab that is running 24-7 overnight while they're sleeping. It's coming up with hundreds of thousands of potential hypotheses that are interesting and then testing them in these virtual models of biology and then kind of From that process, they surface some of those best ideas to the human scientists that actually run in the lab. And so that really dramatically increases the amount of hypotheses they can explore. And in doing so, leads to much higher quality molecules and ideas that come much faster.
• 6:46
  Alex Hutchings: Interesting. And on that faster bit, from my understanding of the interviews I've had on here, the sheer cost of these, say, big pharma and drug development companies to take a drug to market is... we're talking exponentially, multiple, multiple millions, hundreds of millions in some cases. In terms of your business model then for them, is the benefit to them a shorter lead time of getting those drugs to market or just a more effective and more due diligent process to ensure that the drugs they're developing are fit for market? Does that make sense? Which way does it work?
• 7:20
  Josh Hameson: Interestingly, it's a bit of both because the reality of drug discovery is that Teams kind of set out on a timeline and say, we're going to give ourselves two years to discover the best possible molecule to hit this particular target. And assuming the data is supportive at the end of that two-year effort, that's the drug that is then going to go forward into these clinical trials, which are much more time consuming and much more expensive. But kind of the likelihood of success of those trials is determined by that kind of two-year effort to do some of that later stage work. And then. six, seven, eight years of work before that to figure out what is the right chemical space to operate in and what is the right target to drug. And so what we see with our partners is we can take that, we call it five, six year process of finding chemical matter and optimizing it into a drug to go into clinical trials and cut that down into one or two years. And we've done that kind of multiple times now. And then they essentially have a choice based on what the science is telling them, what the market is telling them to. continue to optimize and improve that drug over the next year to get the best possible shot on goal on the clinic or to kind of get to the clinic first and that that's you know a whole component of of the drug discovery strategy interesting and i know we're in this kind of ai world where obviously everyone's on the kind of the
• 8:41
  Alex Hutchings: hype train where obviously ai just seems to be that perceived as that silver bullet to solve all problems but you within drug development especially the size of organizations that you're probably partnering with what's been that big shift that you've started to see in terms of their mindset and thought process around computational models specifically?
• 9:00
  Josh Hameson: Yeah, I think we are, you know, sort of at the shift. And it's been interesting to watch in my time starting this company, because it has that shift has really happened and played out during the life of the company, where when we were getting started, there were kind of a few innovative teams that were starting to leverage these tools. But I think the default attitude of the industry was that they're not ready for prime time yet. Never. there. That shift has really fundamentally changed. I would say now, if anything, we see a little bit too much hype where I think at the same time, drug discovery is insanely difficult and insanely hard. And there are kind of a lot of very hype-y AI efforts that I think promise a little bit too much. You know, there is no silver bullet here. At the end of the day, AI is still a very powerful tool that helps teams do their job more effectively, but it's not going to uniquely solve drug discovery. And so what we the shift we've seen is really going from insane amounts of skepticism. This will never work to, OK, we know AI is going to be useful. How do we figure out what of that is real and what of that is just kind of hype and noise?
• 10:05
  Alex Hutchings: Yeah, you're right. I'm fortunate enough to meet kind of great founders on here. And I think in the last three years, well, three and a half years that we've had this thing, the sheer pace at which companies are launching, scaling, getting investment, products coming to market is just accelerating exponentially. You're right, drug discovery in this kind of biotech space is still, there's still some trepidation, right? There's still this nervousness around fully backing AI in this. So there's kind of human in the loop still. And it's an interesting inflection point, I think.
• 10:35
  Josh Hameson: Yeah, and I think it's important to kind of ground ourselves in what drug discovery is, which is an effort in biology at the end of the day. And part of why there are limitations for where AI can be useful here and cannot be useful is the fact that at the end of the day, we still have an incomplete map of biology. If you ask any drug discovery scientist, do we actually know at a really deep level how life works and how the body works, they'll tell you no. for the last 20, 30 years. We've been really dramatically increasing the amount of biology we can measure. So we sequenced the genome. We started sequencing the transcriptome, the proteome, the microbiome. But every time we've kind of expanded the space of what we can measure, we've learned how much complexity there is to life that was really just beyond our vantage point a year ago. And so there's always this gap between what is possible to measure in biology and what is actually happening at the end of the day. And because of that gap, if you think about what AI is. It's really still just statistical pattern recognition. And so if you don't, if you're not actually measuring the totality of what's going on and there's lots of unknown unknowns out there, there's fundamental limitations to what these tools can and can't do. And that's really guided our approach of where do you apply these tools and how do you really turn them into tools that humans can use to make better scientific decisions, but not a kind of magic black box that outpops the silver bullet drug.
• 11:56
  Alex Hutchings: Yeah. Really, between the lines, it's almost like we need to slow down a bit and go, well, hang on a minute. Let's not forget the core principles that sit behind drug discovery. And let's not just invest all of our time and reliance and trust into effectively a model.
• 12:12
  Josh Hameson: Exactly. And, you know, we have a saying that we quote, I think it's attributed to George Box a lot at the company, which is all models are wrong, some are useful.
• 12:22
  Alex Hutchings: Yeah.
• 12:22
  Josh Hameson: At the end of the day, it's still a model of a very complicated, unknown system, but that model can be very useful for drug discovery teams. And, you know, where we've focused our efforts are really on this molecular optimization phase, which is a four or five year, you know, tens of millions of dollars effort that has a tremendous impact on the likelihood of success of a drug. And in that specific area of drug discovery, you have many shots on goal. You're making thousands of molecules and... iteratively running experiments and so you can feed that data into a model and it can learn and it can update and it can accelerate that process over time and cut it from you know three or four thousand molecule effort over five years to a 500 or 600 molecule effort over a year but there's a hundred other areas of drug discovery that we just don't have the data to to build a model that is useful yet yeah and so i say no models are not worth there
• 13:20
  Alex Hutchings: wait if there's no data to back it up, right? So they're kind of in this position where you've got to still go with the data and the facts and the trends before you can start building, relying on models.
• 13:29
  Josh Hameson: Exactly.
• 13:30
  Alex Hutchings: Yeah.
• 13:31
  Josh Hameson: I think equally important to the data is also validation. I think that's an area we focus a lot on and historically the field has maybe under-invested in. It's really easy. These are statistical pattern matching algorithms. It's really easy to trade correlation for causation and trick yourself into thinking you have a model that is actually useful. But in practice, it's just memorize something that's in the literature.
• 13:57
  Alex Hutchings: Yeah.
• 13:57
  Josh Hameson: It doesn't actually generalize. And so a big part of what we focused on is how do we build the right data sets to really rigorously validate our models as well as kind of benchmark them publicly. So there have been a couple of recent kind of public benchmarking efforts in ADMET optimization, which are the set of models that we focus on. from groups like Open AdMet and Polaris, where we've come in first place in these independent blinded competitions, hundreds of competitors, but really importantly is validating it in these kind of really rigorous ways that represent how drug discovery is actually done on blinded data sets that have never been discovered before. Because otherwise you can really fool yourself into thinking that you have a great model that will design the perfect drug, but it's actually just memorized the last drug that went to market.
• 14:41
  Alex Hutchings: Yeah, yeah, really interesting. And I'm keen to touch on Josh, the founder here. So obviously you're supremely knowledgeable and you're passionate about the space and the problem statement, but putting your kind of founder hat on or co-founder hat on, obviously following kind of investment, there's a lot of organizations that... They either do two things. They go hard on the product development side, invest a lot of capital into that, or they also look at whether it's their go-to-market motion, their product motion, their marketing motion. How did you prioritize what to spend capital and also equally your time following investment?
• 15:20
  Josh Hameson: Yeah, I think we sort of let the market guide us there. I think key guiding point for us early on was just talking to the market and understanding, you know, first, where can this actually be useful? And then once we understood that, what would it actually take to believe that it is useful? As I mentioned, you know, early on when we were getting started, the default mode of the industry was deep skepticism that these tools would ever work. And so that sort of led us to focus a lot of our efforts on proving the science and proving the product and generating the data packages to kind of convince folks that this isn't hype that. that this can actually add value. Of course, it's not a silver bullet, but really showing people that we can go from early drug discovery to drug development candidate in a year instead of five years. And we did that with a lot of kind of early collaborative partnerships with some of the leading kind of biotech companies. And so focused a lot of the early days of Inductive on just proving that the science works and generating the evidence that it can really create value for the industry. And then once that evidence was in place. We then started focusing on both scaling that product, scaling the suite of capabilities we could offer and scaling up the go to market motion around that.
• 16:35
  Alex Hutchings: On the partnership side, I get this from a lot of folks. They'll see the best form. And I think you mentioned it earlier, the best form of feedback. You're steered by the market feedback. So in terms of what they're saying, what is needed, what's not needed, and you obviously can pivot accordingly. But is there a moment in time? over the last few years, which stands out as that kind of light bulb moment, which made you think, ah, this is the path we should be following. And equally, was there anything which was said, which you kind of just ended up going with your gut anyway?
• 17:04
  Josh Hameson: Yeah, there's definitely, I think, a couple of moments early on where we're talking to some veteran scientists who've been around the block. They had a ton of skepticism around AI and kind of computational modeling. They've been promised the world for 30 or 40 years from computational tools, and it really hasn't delivered. And so their mode kind of coming in was, great, another kind of overhyped computational startup. I guess I'll give them 30 minutes of my time. And I think some of the light bulb moments that really clicked for us was honing in on what aspects of the problem did we actually think AI could solve and talking to them about it in ways that turned their skepticism into, I wouldn't say. believers at first, since they still wanted to see the data, but believing at least there was a possibility that this could be real and could create value for them. And converting those kind of first two early, highly skeptical kind of scientific folks into potential champions for us, that really kind of validated to some extent that one, that the space we were focused on molecular optimization really is valuable if you can solve it. And direction we were taking around building extremely high quality data sets to build the best models and validate them rigorously is something that could actually work for bringing along some of the more kind of skeptical scientists.
• 18:26
  Alex Hutchings: And how is that journey coming along now? Do you think you've converted a few of the skeptics?
• 18:31
  Josh Hameson: Yeah. You know, we've brought on, you know, dozens of biopharma partners, helped run many drug programs, brought multiple molecules to the clinic, and have now also started expanding out the kind of suite of capabilities that we offer, including kind of integrating both the virtual labs and the AI models we're building with a lot of experimental workflows in the wet lab and getting kind of the best of both worlds. And so product development is always a journey of kind of finding the wedge, figuring out how to validate that we can create something useful there, and then expanding into kind of the next version of that. And so now we're facing skepticism on the next set of things that we're developing and really trying to navigate into that useful. kind of valuable products based with them.
• 19:17
  Alex Hutchings: Exciting. I'll go back to something you said earlier where in the last, almost that point of launch, the last few years, this kind of acceleration of obviously AI, we touched on whether it's hype or not, but in terms of genuine, the value that's being added, the speed at reiteration and product releases, AI is a huge component of that. Over the next three to five years, What do you think is going to be a big opportunity for AI in drug discovery specifically? Equally to that, what do you think is the biggest, not threat, or the biggest thing we should be mindful of?
• 19:52
  Josh Hameson: Yeah, I think one thing to be clear about is that in drug discovery, three to five years is actually quite short. Not long,
• 19:59
  Alex Hutchings: yeah.
• 20:00
  Josh Hameson: So the typical timeline of... going from the start of a drug program to a drug approved by the FDA on the market is 10 to 15 years. And you can kind of break that down into three different bottlenecks. This piece is basic science and a lot of biology research, which is really trying to understand the body, understand biology, understand disease pathways, and come up with a really good hypothesis that you can validate that says, if we drug this particular thing, if we drug EGFR, that might kill cancer cells. And you lead to an effective drug for cancer. So that's five to 10 years of basic science usually being done often in academic labs, as well as some industrial labs. Then there's another five to seven years of kind of designing the drug and the chemistry work that we focus our company on, which is we have a target. Now we need to come up with the best possible molecule to hit that target and make sure that not only does it hit that target, but it goes through this crazy journey through the human body and it can navigate through, you know, Through your liver, it can permeate across the intestinal walls into the blood. It can dissolve into the bloodstream, distribute throughout the body, do that all without creating any toxicity issues. So all of that really complicated multi-parameter optimization work to get to a high quality molecule. And then you have five to 10 years of clinical testing and running clinical trials to say, is this safe and efficacious in humans? I'm not convinced that last bottleneck around clinical testing is really going to change. too much. I think AI can help on the margins, but the reality is you need to give a drug to a human for a certain amount of time and wait and see if it works. And AI doesn't really change the laws of physics. I think the first area is really compelling of basic science. You know, these tools are very helpful pattern matching algorithms and they can surface interesting patterns that can then become hypotheses, but you then still have to go test those hypotheses in the clinic. And so Really where we see the three to five year time horizon is dramatically changing is that middle bucket, which is I have a target. I want to come up with the best possible molecule to test that target in humans. There I think we can collapse what is today a five or six year process down to one to two years. And at the end of that, come out with much higher quality molecules that are much more likely to work in the clinic.
• 22:19
  Alex Hutchings: It's interesting. Yes. I can see the excitement. I think you touched on it really. saliently earlier in terms of just kind of not getting too carried away, but it's evident the value and benefit that AI can bring as long as it's used, on your point just now, in the right areas. Because as you say, it's not going to be able to change the fundamental laws of physics, right? In a human body, they can't accelerate that, but actually getting something to that point could be improved.
• 22:45
  Josh Hameson: Yeah. And increasing the odds that when you do bring that drug to the clinic, it has a much higher likelihood of being safe, has a much higher likelihood of hitting the target in the way you want it to. and can really rigorously test those interesting hypotheses that are coming out of all the advances around AI at the earlier stages as well.
• 23:01
  Alex Hutchings: Yeah, makes sense. And what final question for you in terms of what inductive bio can people can start to see or hear from you guys over the coming months and kind of next three, six months? What can they keep an eye out for?
• 23:13
  Josh Hameson: Yeah, so we started the company focused on virtual labs really focused on this concept of admit. So absorption, distribution, metabolism excretion, you could kind of think of it as. The entire journey that a drug molecule needs to take throughout the body, there's a lot of complicated optimization that goes into making a drug work there. That was a lot of our seed round when we raised our Series A earlier last year. A lot of the focus was both expanding that, but then also expanding the suite of capabilities really in two directions. The first is going after some of the very hard and challenging toxicity problems that come up in drug discovery, often only get realized in kind of later stage clinical trials. drug-induced liver injury, cardiotoxicity, for example, about 40% of safety-related drug failures come from one of those two issues. And they usually get caught in phase two, phase three, very expensive, very slow clinical trials. So we've actually announced towards the end of last year an initiative funded by ARPA-H in partnership with Amgen and a bunch of academic pioneers to really go develop, we almost think of them as digital organs that can simulate how will a drug actually work in the liver or in the heart. that to go kind of increase the likelihood that a drug will actually be safe in humans. That's one big effort that we are continuing to push down and seeing a lot of progress on. And then the second, as I mentioned earlier, is really integrating the virtual lab with the wet lab. At the end of the day, you'll need to go run an experiment, test the hypothesis in the physical world. And so we've partnered with some of the leading folks here, folks like IncoBioworks, folks like Enamine, and we're building a kind of a list of best-in-class partners. to really effectively combine what is possible in the wet lab with what is possible in the virtual lab and get sort of an end-to-end feedback loop that helps our partners get to higher quality drugs much, much faster.
• 25:02
  Alex Hutchings: Well, keep me busy. There's no surprise you've had this rapid kind of adoption excitement around inductive, you know, three and a bit years, investments, some really good partners. So, no, I wish you guys and the team a huge amount of success with this. And I look forward to kind of sharing this because I think it's a space which isn't necessarily a normalist would listen to and go, oh, people look at drugs, people look at, and it's just there. But they actually don't know what happens in, let's say, the 10, 15 years beforehand. and you know You're playing such a big, important part of that process. It's, yeah, it's been great to have you on, Josh. And thanks. Thanks so much.
• 25:39
  Josh Hameson: Yeah, thanks for having me. It's a fun conversation. And I think, you know, the more you dig into the world of drug discovery, the more fascinating and honestly inspiring it is of the fact you get to go to a pharmacy, pick up a drug and immediately treats your disease. You don't see the decades of. really really rigorous work the hundreds of millions billions of dollars that went into actually getting that drug into your hands and confirming that it works so um you know excited to to share a little bit more about that with uh your audience and kind of the broader world here since it's you know a phenomenal space to be in yeah i love it thanks so much josh it's been uh super interesting thanks again likewise thanks alex thank you