Schrödinger, Inc. (SDGR) Earnings Call Transcript & Summary

Welcome, everyone, and thank you for joining us for our first ever Byte-ology Conference on The Convergence of Biotechnology and Technology. Our first panel today is leveraging AI, machine learning and computational biology for drug development, and I'm joined by my colleague, Chris Shibutani, who covers the biopharma space. On this panel, we're pleased to have Exscientia, where we have Andrew Hopkins, CEO; Recursion Pharmaceuticals, Chris Gibson, CEO and Co-Founder; Relay Therapeutics, Sanjiv Patel; and Schrödinger, Karen Akinsanya, Head of Discovery R&D and Chief Biomedical Scientist. With that, thank you, everyone, for joining us. And to start here, just for investors that may be unfamiliar, could you briefly introduce yourselves and give an overview of your company and platform, including what you would single out as a key advantage or differentiating factors? And Andrew, why don't we start with you?

Thank you. Absolutely brilliant. Pleasure to be here today, especially amongst my speak colleagues. We actually don't think ourselves as competitors. The real competition is against disease here. So Exscientia, we're an AI-driven pharma tech. We use AI to improve every aspect of new drug creation. And we are committed to true precision medicine. We want to make discovering the best drug for every individual patient a real possibility in the fastest, most effective manner. And to do this, we think one of the key advantages Exscientia has been building is a unique AI-first end-to-end process. And actually, we start off with the patient. We start off with the clinic. And it's about how then do we build the best translational models by taking sort of patient tissue samples, the data associated with our patient and drive then those models to drive all aspects of drug discovery, drug creation and patient selection. A good example of how AI is driving back, just a few weeks back, we published the first ever functional precision oncology platform to successfully guide treatment selection using an AI process. And that was a prospective interventional study, really show that AI-driven drug selection could improve potentially where just a physician-driven approach. And we want to use then that sort of clinically-validated approach to translation then back into our sort of way of driving AI-driven design. And the key other thing with that is that when you're thinking about drug design, when you're thinking about AI-driven approaches, we really think of drug design as precision engineering at the molecular scale. And what we want to do then is that we use creative AI to generate the best molecules possible. And that allows us then to rapidly translate from scientific concepts into precision design drug candidates and then take those into the clinic. Ultimately, we see our philosophy as not drug discovery as a screening problem, but we see it as a learning problem and all our technologies are set up, how do we then try to learn as fast as possible.

Great. Chris Gibson?

Sure. Well, thanks, Salveen and thanks, Chris, for hosting. And my name is Chris Gibson, Co-Founder and CEO of Recursion. We are pretty convinced that the challenge facing us right now is one of understanding how to modulate biology the right way and I think the proof is in the clinic where 90% of drugs that go into the clinic don't make it out the other side to affect patients' lives for the better. And so, what we've been focused on for the past 8 years is building maps of biology. So, we use what you see over my shoulder, one of our large automated facilities where robots do giant experiments every week. So we do about 2 million experiments every week in a wide variety of human cell types and complex co-cultures, and we generate omics data from all of those experiments. We've actually just celebrated this week our 100 millionth experiment as a company in that laboratory. And so, that generates a huge amount of data. We're generating now close to -- we just looked at the statistic. Every week, you need nearly 1,000 of the newest model of iPhones to hold all the data that Recursion generates in a single week. And all of that data gives us incredible insights into relationships across biology between genes and targets and pathways and molecules. And so, we use sophisticated computational systems, including convolution neural networks and AI, which we train on our own supercomputer to allow us to identify those exciting patterns. And what we like to say is we want to go where no Nature paper has been before. We've really built this system to help us identify targets that are not in the literature anywhere. And so, that's what we've been working on the last 8 years. And what I think really differentiates us is what I'll call integration. So, we have not only one of the most sophisticated, highest throughput wet labs on earth, but we also have an incredible team of data scientists and software engineers working on one of the fastest supercomputers in the world. And I think bringing those 2 worlds together, creates a lot of excitement for us, 4 programs headed into Phase II trials in the next just about 2 quarters and about 40-plus programs following behind. And as you know, Salveen, we just announced yesterday a major partnership with Roche/Genentech to go after the whole of neuroscience and also have been partnered with Bayer to go after fibrosis. So, really creating both our own internal pipeline and these partnerships to go after these big intractable areas of biology, which is only possible because of this integration between the wet and the dry lab and because we're building across the whole stack, not just the early discovery, but all the way through animal models and now into the clinic.

Thanks, Chris. Sanjiv?

Thank you to Goldman for the invitation, and also it's a great pleasure to be with the other panelists today. Relay Therapeutics is a company that was highlighted in 2016, was one of the first companies that sat right at this intersection of leading-edge computational and experimental techniques. We are very focused in our approach. We're pragmatic. We know biology is complicated. And so, we focus on validated targets, so genetically validated or clinically validated. So we're not mapping novel biology. We're just looking at targets that we know that if we could find a better therapeutic for, we know that we should be able to have an impact on patients. And with that, we use the combination of experimental techniques, structural biology, biochemistry, biophysics to better understand the proteins that we are trying to drug and then combine that with a range of computational tools to first of all visualize the proteins and understand their dynamic confirmations and come up with motion-based hypotheses for how to drug them, identifying novel starting points and pockets and then use the computational and experimental approaches to rapidly optimize those chemical starting points of small molecules into drug like molecules. And so 5 years later, now we have 250 people. We have taken 2 precision oncology programs right the way through this process and into the clinic. And most importantly, in October, we showed for our FGFR2 program, the first clinical validation of our approach and impact on patients. And so if you come to the question on what is unique, I think for us, it is the experience that we have across multiple programs now. And the ability to combine very deeply, bottom-up a company that was created using leading edge experimental and computational techniques. We believe that, that really is the core of Relay Therapeutics, the ability to integrate.

Karen?

Thank you very much for having me on the panel and happy to talk about Schrödinger, which is a 30-year-old company. We have a platform that really spans the gamut from structure-based drug designs, protein structure refinement, all the way through precision design of molecules using atomistic physics-based models. The company essentially has software that's used by the entire industry, all of big pharma and a very large and increasing portion of biotech uses the software to design molecules with a high degree of accuracy. That means that we're using these atomistic physics based models, that are designed specifically for each program and essentially are able to solve the multi-parameter optimization problem. And that means, all of the parameters that you're trying to optimize to get a high-quality molecule that is appropriate for testing therapeutic hypotheses. I think the differentiation is really the platform and how broadly it's used by the industry and how well validated, I would say, that it is over quite a long period of time. Over the last 30 years, of course, the platform has been developed. The company is about 600 or more people, and a good portion of that, over 200 PhDs are working on the existing platform, but also taking the platform to new heights to solve an even broader range of problems that face drug discoverers. In terms of the history of the company over the last 10 years, we have worked with a large number of companies both pharma and biotech to really use these methods at enormous scale. That means that 6 molecules now are in clinical development, actually, one of which is approved and another 6 are in clinical development between Phase I and Phase II, and there are 6 other programs that are in IND-enabling studies from those collaborations. So, the impact of the platform has already been established and is broad. I would say the next piece in the story of Schrödinger is that over the last 3 years, we have developed our own internal pipeline. We've been able to identify targets that we think are of great interest and develop molecules for those, which -- the first of which will be in the clinic in the first half of next year. We've assembled a team of 100 or so people from pharma and biotech, who have deep expertise in drug discovery and development. And we believe that with the advent now of the broad availability of structures, structure-based drug design is really about to take off in a very broad way. And all of the tools that we've been building over the last 30 years are very well poised to benefit from the broader availability of structures.

Karen, maybe to one of the points that you raised, and we could go backwards through all the teams on this question, but how can you tell when a discovered or designed therapeutic is optimal, which signals do you look for and which benchmarks do you consider through the process?

Yes. I mean, I think it all starts with the target product profile. Of course, once you've established that a target is biologically interesting, the question is how do you prove that biology in the human setting. And that means, you have to have a pretty good understanding of the characteristics you want in that molecule. How long should it last in the blood? What affinity should it have for the target? What are your target -- what's the target product profile? Is it supposed to be brain penetrant? All of these sorts of questions, obviously, are posed at the very beginning of a program. And so the question is, how do we accomplish the goal of designing in those characteristics of a compound, before it goes into clinical testing. Historically, of course, this has been achieved by empirical methods. It can take anywhere from 3 to 10 or even longer years in traditional approaches to come up with a molecule that has sufficiently well balanced properties to go into the clinic and sufficiently test the hypothesis. That is the challenge in drug discovery and translation. Of course, there's a huge biology challenge in translation, but I think having the right molecule to adequately test the hypothesis, that has all the right properties, is a key component of that. And so the question is, have we been able to demonstrate that being able to design molecules with these balanced properties are appropriate for testing in the clinic. And as I mentioned, there are many molecules that have gone in from our platform. Any one of those properties can completely derail a program. So, if you don't reach the right exposures in humans, you're not really going to be able to sufficiently probe that target. And so, I think increasingly, we're able to assess some of that preclinically. And I think Andrew talked about our ability to figure out do we have the right patient, who has the right target at the right state of disease that we can actually design this target product profile for. So, I think there's a level of understanding of the problem statement in the human and then designing a molecule that essentially can be well positioned to address that question.

I just want to answer what Karen saying actually and fully agree with her that the key starting point is defining the objectives. How does your biological knowledge define a target for a profile, it's more than discovering a drug target. Actually, if we think about a drug from a patient's point of view, we need to think how they're going to be exposed to that, how they're going to react to it, what's the dosing profile is going to be important. This is the difference between sort of a nearly run molecule or potentially in a blockbuster. So ideas then are focusing and understanding the importance of a target product profile and then converting that then into the objectives with the multi-parameter optimization is key. For us, we use an entire AI-driven approach when driving that for multi-parameter optimization and we believe it is essential. Just wanted to sort of underline what Karen was saying and the importance of that. And also, the other way we think about that is that when we think about all the different types of assays we want to use and measure all those different aspects of a target product profile, we think that's always important to be what we call data agnostic. It's not just about a crystal structure, and it's not just about high content. You need to use whole wide range of different types of assays, where it's structural, where it's high content, where it's pharmacology ion channels, et cetera. All of these producing different data types, which require then sort of different modeling techniques. And then the third thing, just the emphasis of what Karen is saying is that fundamentally to think about the question you asked, is that relevance of our translational-able assets, the relevance of how then we go into a test in the lab and have confidence of the results we get in a lab are going to lead results in the clinic is the fundamental crux, I believe, of changing economics of the industry. And that's why for us, focusing on patient-centric assays, focusing on understanding the heterogeneity that we have of patient response is vital, and that's when we can build assays when we look at individual patient responses and look then across a wide variety of patients. We then start to understand that certain drugs, certain drug targets would hit then just a subset of that disease with its ovarian cancer or melanoma. And it allows us then to understand then how we're going to target our drug for variety of patient population and that's key. What we find about -- the beautiful thing about bringing these assays from that clinical experience then into drug discovery is that you can then think about patient stratification at the discovery level, even before you identify the drug candidate and knowing which patient population will go after and how then we can think about targeting the right biomarkers to identify the patient population. I think that is essential, actually, to this question of how -- even when we are -- precision engineering drug molecules and AI, making sure that they are relevant for a patient as possible.

I mean, maybe just to add to what you said, Andrew. I think you're absolutely right around -- the target is critical. The question is, the one that we all wrestle with, what's the optimal molecule. And so for us, we try to make it as simple as possible and reduce as many of the unknowns. So we work on validated targets, where there is a well-trodden pathway around the assays and the translatability of potential in vitro, in vivo models into the clinic. And so, that's how we try and reduce the unknowns. If we believe that we can make in silico predictions and test them in vitro, then if your predictions are correct, then you hope that they would work in vivo and then so and so. And I think we showed the first validation of that with our FGFR2 program that there is a straight-line between these. I think for predictions, because I think the focus is, can you use computational predictions to make this whole process more efficient. I think we're right at the beginning of understanding whether and how efficient this is going to be. And I think we'll get more efficient over time. And so just in our own experience over the last 5 years, we started out being able to predict potency and then get more confident at that. Now we started to predict selectivity, got better at that. Now we're trying to predict drug like properties, we'll get better at that. But all of this is really in the early innings. And so, we've probably -- we'll be back here in a few years' time and say we're much better at it than we are today.

I agree with everything that's been stated. And Sanjiv, I think you raised an important point there, which is one of the huge fundamental benefits of machine learning and AI is going to be to improve the efficiency of designing, developing molecules for the biology that we understand and that's been demonstrated. And I think that -- I'm very excited about all of that tech, and that's some of the stuff that we're working on here. I think over the coming decades, the additional promise of machine learning and AI at the intersection of biology is to push the bounds of the biology that we understand and to identify targets that maybe are not validated. And I think that's where Recursion is really focused for the last 8 years. We actually are huge fans of the work that all these other companies are doing, and we could imagine even partnering with them in some way, shape or form at certain points because they've built really fantastic technology. Where we really focus is on trying to identify a novel relationship between some particular target and some disease and that's our favorite thing. When we use these maps of biology we've built and we query them against some disease and the target comes up and you do the Google search between the 2 and there's no results, that's the exciting part for us because that's, I think, how we're going to ultimately really shift the impact for patients. You look at the field of Alzheimer's, for example, where right or wrong, the industry has been focused on the same target for the last 20 years, more or less. Where are the companies that are going to really push the bounds of what those targets might be. And I think in our partnership we just announced, that's a huge focus. It's going to be going after these massive and tractable diseases and to go after novel targets, things that nobody else is working on, that aren't being published right now. And certainly, that comes with risk, but at the same time, I think, for patients, there's the potential for breakthroughs there. And that's really where we've focused. And then beyond that initial map building exercise to understand where these new targets are, how do we expand that, using some of the great technology that our peers have made or building some of our own to make it even faster to advance those medicines to patients.

And the commentary here is a great segue to our next question and let's try to get some perspective from all of you. You are all very much pioneers. And yet you also bring to bear here, a lot of credibility from your industry backgrounds, Andrew from Pfizer. I think. Certainly, Chris and Karen, I think you're both Merck [indiscernible], hopefully you're good [indiscernible]. And just trying to understand where we're at with the ecosystem. And as Sanjiv had mentioned, we are early days. And we think about the industry here in 2021 and looking over the next few years and really the next decade, we're kind of left to question why is it that every company does not seem to be incorporating some elements of these artificial intelligence-based methodologies to enhance their approaches to R&D. I realize it's a bit of a spectrum, but maybe what do you think are kind of -- what's holding them back? What are the inertia points that are making things challenging? Maybe I'll turn to you Sanjiv. You really seem to introduce this kind of where we are.

I mean I think -- there's 2 sides to that, which is, look, the industry has been successful without it to date. All the wonderful medications that we have around us have been largely developed without these tools that are now emerging. And so I think that's the first thing. That's just inertia. Like why would we change? And as part of that, there's been lots of false storms. So if you look back 15, 20 years, there was a dawn of genomics and look, all this data is going to change and revolutionize everything around us. And it didn't play out in the way that people thought it would play out. So I think that's the first reason. I think the second reason is on the other side is it's hard. I mean, I think you will see it. We've all lived the battle scars around the panel here. The culture of building a company that's bottom-up using computation is tough to do in a world where people in our industry are not used to doing it. And then I think, as Chris points out, there's a huge amount of data and experience curve that is involved in what we do. Each one of us, the more we do, the better we get. And building those kind of data factories that each one of us has to different extents and different elements, just takes time. And like we've seen it over the 5 years, there's a significant experience curve, both in terms of people, approaches, culture, the way that the companies are kind of engineered. And then I think the final thing is talent. There are very few people that are -- have been built into this world because traditional pharma companies have not created these kinds of people. And so, I think there'll be a huge talent war and the shortage, which will restrict people really creating these types of companies.

Yes. Chris Gibson, you're freshest off of this sort of interface, having just announced a collaboration. I'd be curious to know that relevant to the comment that Sanjiv made about the people, was there something about the level of folks that you were interfacing with on your new partner side that you think was really compelling? Was it kind of like the next-generation books that were more looking to get the hands to come to agreement here? Maybe a little bit of a sense for what -- how you were able to recently overcome some of those inertial factors?

Yes. Well, I mean, you look at who's leading gRED now and Aviv Regev, who is really a computational biologist. And I think that is a strong signal to the industry of the direction things are moving. And as my Co-Founder, Dean Li, who's at Merck said, if we can't convince Aviv, that it's going to be hard to convince the rest of the industry. And of course, what we found and Aviv was somebody who appreciated not only the biology and what we were building, but I think appreciated what we had built from a culture perspective. And I'd echo Sanjiv's remarks. The reason a lot of the companies have not been able to or chose not to build as deeply as these 4 companies at the intersection of technology and biology in some cases, is that the language between these worlds is extraordinarily different. Software engineers and data scientists speak in very, very different ways than biologists and chemists. And I think traditional companies in this space struggle even with the political tension between the chemists and the biologists. And so, imagine adding on top of that data scientists and software engineers who bring their bayesian statistics into every conversation. And all of a sudden, you create, I think, potentially a really challenging environment, unless you can, from the very beginning, build a culture that embraces these things and embraces the differences between these groups and harnesses them in a way where a software engineer and a data scientist have the same say over a program as a biologist and the chemist. And that takes a lot of retraining, right? The chemists from certain companies and the biologists from others are used to kind of having the decision-making authority. And when you build a company where the data scientists can kill a program, that just takes a tremendous amount of intentionality from a culture perspective. And I think the tech is hard, but all of our colleagues and the larger, more traditional companies are sophisticated and talented. But hiring, recruiting and building culture at the intersection of this space is, it is the most difficult thing that any of us do. And I think it is the essential key ingredient for us to be successful.

Karen, at Schrödinger, you've actually had a very active business, a decade of experience of trying to essentially partner with companies on the software part of the business. As Joel tries to perpetually wrestle with these partnerships, do you think things are getting better? Are they improving? And what's the pace of that sort of getting the buy in, so to speak?

It's a great question. And I think actually, it's a matter of scale. I think that companies -- and I was at Merck 4 years ago, I think companies are experimenting to some degree with this new paradigm. And they're doing this at a scale where I don't believe they're seeing the full impact. And so, companies that choose to collaborate with Schrödinger and use the technology, at least for the properties that we've spoken about, selectivity, potency, permeability, solubility, the things that matter in making drugs, when they use it at full scale, it's pretty obvious. This is transformational. I think the way companies are set up right now, and I think the other panelists have touched on this is that the teams are constructed very differently. Computational scientists are either not on the team or they are outnumbered in significant fashion by traditional either by chemists or traditional biologists. I believe that tide is shifting. And -- but as you mentioned, with people like Aviv joining in the senior leadership positions, Dean, that tide is changing because I think what you're seeing and what we are doing at Schrödinger, we have a very equal distribution of computationists, physicists, mathematicians, biologists, chemists, translational and clinical people. And when you put that team together and you have access to the technology at its fuller scale, which is happening broadly in our collaborations, pretty obvious, you can move very fast. Our first IND comes about 1.5 years, if not 2 years after we conceived of that project. That's transformational. And if this technology is used at scale, by the way, on novel targets as well as those that we've already been working on as an industry for a decade or decades, I believe that the momentum is only going to get faster. And by collaborating with BMS, who both have access to the technology or using it broadly, but are also seeing what we do and how we do this at enormous scale, they're beginning to see that knowledge transfer. And I think those are going to be the moment that ignite a true transformation and a true paradigm shift to more of these technologies. And I do just want to add one other comment. So, I think we've talked about well-validated targets and what Chris is doing in finding brand-new targets. Having the ability to come up with a toll molecule that has the right properties to phenocopy of the genetics or phenocopy of the biology that we see in an assay system like Chris has set up that are actually very close to what you will need to then go into the clinic. Being able to do that at scale, I think, is extremely exciting, and I think pharma is beginning to realize that too, because we need to test more hypotheses. The reason why drug development's so difficult is because we're either all working on the same thing, and we're not really edging into that new frontier of new biology. And if it takes 10 years to get a molecule, and it takes $1 million to do a high throughput screen, the number of hypotheses you're going to test is going to be small. So I think we're at the moment where a lot of that's going to change.

Indeed. And I think amidst all this enthusiasm and pioneering leaning in, I think, Regev, you said something that was very important and that this is difficult. So Andrew, I'll turn to you for the question of are there modalities or disease areas that you think companies should actually be wary of applying AI-based approaches too? Some cautionary thoughts there? Andrew, you're on mute. I apologize.

When we look across it now, we look at what's possible technology. We fundamentally believe that all modalities can benefit from the use of AI in their design and optimization. In fact, I fundamentally believe that all drugs should be designed by AI, probably within the next decade as well because what we're seeing now is that the productivity gains are so comprehensive that actually makes a very compelling economic case to do so. But when we think about where we should go, when we think about our own pipeline, our own portfolio, one of the driving factors first now is thinking about how we got the right sort of understanding of the disease, can we build using sort of deep learning approaches to multi-omics, et cetera, about how we can understand how we're going to translate. How can we go from the clinic to the lab and understand that disease and some heterogeneity or patient response and how then can those models help us drive their new molecules from the lab back into the clinic, we have, hopefully, a much higher chance of success than traditionally has been the case. So, when we think about where we should be wary about applying the AI to, the way we think about it is like, are there areas where we can actually build now and understand our biology and use our new technology to help us build these translation assays to guide it. That's what actually drives us and allows us to think about where to face our bets and where we think we're going to really shift the curve in the understanding of how we're going to change the economics of the industry and understanding disease, understanding how we can model that then in the lab. And from our position, we believe it's by understanding the patients better, allows us then to think about where should be the areas actually we should go after and where we can make a big difference.

Sanjiv, would love to get your thoughts on this realm as where as well. Where might there be some warning signs or areas where you think it may be less optimal to apply these approaches?

I agree with Andrew. I think, ultimately, the approach is it's applicable broadly. And so, it just comes back to the pathway to get there. And for our own approach, we're just trying to be pragmatic. It's complicated, and there's lots of unknowns here. And so early days, which is where we're at, let's just try and keep this as simple as possible. And so, we focus on small molecules, like many of the companies on the panel because they're tried and tested. There's plenty of data out there. The biggest data sets all sit in small molecules because they've been around for many, many decades now. And so, I think that's probably the best place to be pragmatic and hone our approach and get better and build experience and then slowly expand out from there. But there's nothing inherently to say that it cannot be used across any modality. And I think it's just a pathway to get there.

And Karen, you've been tasked with sort of quarter backing of the Schrödinger effort on your own proprietary pipeline, which puts you in position to make some technical decisions of over how you would prioritize this same question and framing it from the perspective of how you chose your plan of attack with the -- with your own internal portfolio. Are there areas that you thought maybe not in that dimension or maybe not that realm?

Yes. I think it's a good question. I mean before I sort of dive in on that, I will say that the idea that AI, which let's face it, is a sort of umbrella term for anything that has computation in it, that word has been thrown around a lot, right? And I think the idea that AI is somehow going to solve this multi-hundred year problem that we've had in a sort of black boxy fashion, I think we have to be careful there. And to your last question about the adoption, I think the more we sort of use these general terms without digging in on the validation and the actual ability to accurately predict experiment, I think, we have -- there's a danger of running into trouble. And when we also talk about data and how data is going to change things, we don't have all the data we need. So I will argue, and that's why I'm excited about what, for example, Recursion is doing in neuroscience. We don't have that many approved trials. So our ability to learn what we've done wrong in the past is challenged by the fact we haven't been very successful. And so the idea that data in general is going to solve this problem as opposed to fundamentally trying to understand the biology, the pathways, where we are in the cascade from health to disease. These questions, I think, can be -- we can dig into those and really try and understand the fundamentals, but I think it's a danger in saying, we're going to solve with AI what hundreds and maybe thousands of scientists haven't been able to do previously. We need to do the work. We need to do the work and dig in. And so, what I would say about the choices we've made, we're on a journey. It's been a long journey for Schrödinger. First, the platform and now, how we apply it. Obviously, the methods, the physics-based methods work regardless of modality. We're doing a BMS degraded effort now. That's a whole new space. It's very exciting. We think these methods will be able to help inform that on top of what we're doing with them on small molecules, biologics are amenable. What I would say is that for us, we wanted to focus on where we thought we could have the greatest impact on solving some design challenges. That's where we started. Where I think we'll end up is contributing to this broader understanding of structure function, and understanding of, yes, the target, but how do you drug that target? And how do you do that at scale? How do you use these methods to create molecules that allow us to marry our understanding of the good-looking molecule with whether that biology is translatable. Those 2 things have to come together, I believe. And so, the journey we're on is moving towards more and more of those insights that will tell us whether this is a great target, both to drug, but ultimately to take into the clinic. And that's a very wide space that's ahead of us. And I think we figured out how to do some of this at scale.

Yes. I have to say that I'm so struck that there is an aura of mutual respect amongst you. Salveen and I co-host many panels and gatherings and sometimes you have to expect a fight to break out. Instead, this feels like a conclave of the cool kids, and you all have so much sort of mutual respect and energy for each other, which is really wonderful. And Chris Gibson, I'll just close to ask you on this topic. I'll close it, Karen, through you some kudos because if we -- the question I was asking was, where should be people be wary of and people in general are wary of CNS. The degree of difficulty is in that -- the need is tremendous. And if you're going in that direction. So then, Chris Gibson, tell us, would you say that there might be a realm that anybody should be wary of in terms of this as application?

Yes. Well, a lot depends on how you're applying technology. And I think the reason you hear the comradery among these companies is that there are hundreds of steps and many different ways to apply technology to improve the process. And as Andrew said at the beginning, there's a lot of disease and suffering in the world, and none of us are close to eliminating all of it. Maybe in the next 15 or 20 years, we'll be more competitive because we'll be honing in on the last few diseases that need to be treated. That would be a great dream and a wish, I think, for all of us. But what I would say is I think of what Recursion is building like a map, right? And we've knocked out every gene in the genome, 4 or 5 different ways, in multiple human cell types. We've profiled nearly 1 million small molecules at multiple different doses and multiple cell types and that creates a lot of data. We actually have hundreds of billions of relationships that we have identified based on all of the data that we've collected. And the challenge there, it's like if you open Google Earth, you can zoom in all the way down to the level of individual houses and knowing where to look is really challenging. So where we like to go in biology is places where we have an anchor. We'd like to kind of know what city are we looking at. So we like things like genetic diseases and something that we found in the space of neuroscience, and we'll be working with our colleagues at Roche and Genentech to explore is that there are rare genetic causes of many different elements of neurologic disease. Those can be kind of the initial coordinants that can allow us to take these hundreds of billions of relationships and to zoom in on the thousands of relationships that might be most important, using those kind of anchor points of biology. And so as Recursion builds our maps, that's what we like. We like finding an anchor and then looking for all the novel biology that exists around it and then trying to -- as Andrew and Sanjiv mentioned, making sure we've got really translational models to turn those hypotheses into rigorous tests of what we're exploring.

Thanks, Chris. Maybe to turn over to the question of platform validation, and Sanjiv, we could start here with you. You've seen this early proof-of-concept in FGFR2. And how do you think that, that helps you understand the forward here? Was that really true clinical validation that then it derisks your platform from the standpoint of translation for every other program you're going after, given for you, the biology that in terms of the targets have been validated.

I mean, I think it goes a long way to kind of drawing that line between hypothesis that we create using our motion-based understanding of proteins to the in vitro modeling and in vivo and then in the clinic. And so, the ultimate goal is always creating a therapeutic. And I think that obligation, all of the panelists here feel, right, which is we are playing in a space which is new and emerging, it will be judged in time only on one thing, which is how many new medications are delivered to patients. And so, we can talk a lot about all of the other things, but that is the single biggest factor. And so I think the data that we showed a few months ago really goes a long way to showing that. But I do think we're in the early innings. And as I keep coming back to, I think it's a long way from hitting a button and finding a medication to come out. It's not a panacea, as Karen has said. There's lots of data that we don't have. And I think we'll see lots of emerging companies and approaches. And the panelists here are great examples. We have a lot of respect for all of them. Chris discovering novel biology is a fantastic effort, like we would love to drug that novel biology over time. And so, I think you'll see lots of companies emerge. Our belief is that the computation must be combined with real experimental feedback loops. And we will see tools and approaches that will not work. And so, I think we all have to be open to that. There are plenty of things that we try internally that we love. That didn't work. We'll move on and we'll see what else works. And I think the final thing that we'll see over time is the dots being joined. And so, we're all focused in different areas. It's only just because it's so hard. And so I think we'll see novel biology linked with our ability to drug, linked with approaches in the clinic to create almost a full stack approach. And so, we're right at the early genesis of all of this.

Chris, maybe you can touch on your approach here with regard to platform validation, where in the early 4 programs you've mentioned, you've in-licensed them and then you're optimizing them with your platform, but then you have your own internal programs and then now with Roche/Genentech going to go after very hard disease areas. So, how do you look at validation and then optimization to apply in parallel since you're taking a couple of different approaches?

Yes. Thanks, Salveen. I think if you go back maybe 30 or 40 years, there was a pretty big shift in the industry to being very sort of molecularly targeted. And that has paid some dividends, and I think we've learned a tremendous amount. But one of the challenges there is that the tools that we've had at our disposal of scientists have been pretty reductionist, right? And I think one piece in common among all the companies here in different ways is that we're synthesizing much larger data sets that act more like a system. So in the case of a relay, for example, in the past, people have looked at sort of the static proteins when they're doing their work and the work that relay is doing is allowing them to look at all of the inherent motion that exists in the real-world in these proteins to find better drugs. It's a much more complex problem than looking at a single structure. And I think at Recursion, it's the same thing, and it's how we think about validation. Rather than taking the bias of, I believe, that target X is the target I want in the disease. Did I modulate target X and is there some low dimensional unimodal readout that confirms my hypothesis, which comes with a tremendous amount of confirmation bias. And despite incredible scientists in the industry, we've seen the power of confirmation bias across hundreds of programs that have failed in the clinic. I think what we like at Recursion is to ask really, really broad, what I call, high dimensional questions. If we're interested in a piece of biology, we don't just want to know does it move one downstream marker that we think is associated with the disease because the reality is we probably don't understand most of the biology. The question is, if you model that disease, do hundreds or thousands of things change? And can you find a medicine that can make those hundreds or thousands of things go back to the way they are in a healthy sample? That, I think, is really, really exciting. And it's difficult to explain all of it, but finding the validation steps in high dimensional biologics and systems, that's what we think is key. And some of those molecules we in-licensed were not expected to be useful in the diseases in which we've explored them. The literature doesn't really support the hypothesis, but we found that they had the potential to be used in these new ways because they were rescuing these really complex, highly subtle signatures of disease. And so it made a lot of sense to take the great work that other chemists had done and bring that in. And I think that's what we'll be doing with Roche and what we're doing with Bayer and others is to look for these very, very complex signatures of disease and ask whether we can rescue those. And I think that it's much more difficult to have confirmation bias creep in or other kinds of bias when you hold yourself to these more kind of systems-level standards. And that's how we think about it. But ultimately, Sanjiv's right. All of us will be judged by the number of programs in the clinic that benefit patients at the end of the day. The field is new. So all of us are at the beginning of that journey, but that's where all of us are headed, and that's the ultimate validation for every company here.

Karen, it's been interesting to note that apart from being partnered with biotech and pharma, you are also partnered with tech with Navidea. And so, how do you apply all these learnings? And when do you think you'll have a sense of how the platform is playing out clinically?

Well, I think from the collaborations, which have been in motion for around 10 years, we already see some of those discovery projects have now transitioned not just into safety, tolerability, dose escalation, but beyond that, to the question of proof-of-concept for the target or for the target product profile that we were trying to solve for. A great example of that would be the small molecule integrin inhibitor that Morphic have been working on. For a long time, we didn't know how to make small molecules for the integrins because we didn't understand structure and function sufficiently. When you have that data and you have those insights, you are able to solve for some of those challenges that have existed for decades sometimes in the industry and achieve that target product profile and take it to the clinic and see whether you are right. Now, that journey is ongoing, obviously, for a number of programs. There's about 30 programs that we've worked on collaboratively and that number keeps growing. And as they make progress in the clinic, I would say that, that adds to the validation that not only can you design molecules with the balanced properties that are required to fully test the hypothesis in the clinic, that as Chris and Sanjiv just pointed to, that these molecules matter to patients, physicians and obviously, those interested in broad application of therapeutics. From a technology standpoint, I do just want to point out that while these methods are validated for most of the endpoints that we think matter in the [indiscernible] of a development candidate and that's shown by the breadth that this platform is used by the industry, we also recognize and the continued investment in the platform is because there are a lot of things that we have not yet solved, right? And so, when it comes to acne and toxicology, these remain some of the biggest challenges from a prediction standpoint. And that's why we are doubling down on the platform build. Obviously, once you start doing drug discovery, there's a potential to kind of only focus on executing on those specific programs. At Schrödinger, we've chosen to do both. Let's move more programs into the clinic, but test more high therapeutic hypotheses with good quality molecules that could eventually become the drugs. And let's keep trying to solve what I think at this point is not yet well validated, which is some of those later stage discovery endpoints like acne and tox.

Andrew, how do you think about these issues of platform validation?

So when you want to change a multitrillion dollar industry, which, as Sanjiv said earlier, sometimes things are fine as it is. But we believe they're not. We're going to transform them. We fundamentally believe that 1,000 diseases we need to tackle, and we need to bring the latest science to patients quicker. That validation has been central to us. I mean when we started nearly 10 years ago, on the back of work we first published in Nature, our first algorithms, what we saw then is that, that's great for an academic. But actually to change the industry, the onus is on us to validate it. That's why for us, it is so important now to really show that AI-generated molecules can optimize far quicker, far more efficient than conventional approaches. Now, we've got the first free AI design, true AI design molecules now in the clinic, that really helps and validates that. And then how we can show our AI can help us translate ology better. Now we've published our first EXALT-1 clinical trial, showing that an AI-driven approach can improve outcomes in cancer for the first time that actually can help us select which medicines a patient which has failed all other therapies should take. This then gives us a real validation, we believe, from both how we think about AI being applied to our [ royalty ] platform, how AI is being applied to our design platform. And where we see it right now is that, based on these sort of findings, based on these real AI-first for the biotech industry now, how do we scale that now? How do we then bring this forward? Because I think all the companies here today, we focus on industry transformation. That's what drives us. That's why I said at the beginning, we're not competitors. Disease is the thing we're all competing against. And that's why it's been so important as well when one tends to change and transform such an important industry, that's why I think this time to get to the point of having these validation proof points now. And I believe, actually, the technology platform is now showing that and now is exactly the time to then take those findings, take that success and scale it up now because there are such a huge number of patients now that are waiting. We have to be impatient for the patient.

Great. Well, thank you so much on behalf of Chris and I. We really appreciate your time today.

Thank you.

Thank you.

Thank you.

Thank you.

Thank you everyone, for joining us. Fantastic conversation.