At this time, it is my pleasure to turn the program over to your host, Alkesh Shah. Please go ahead.
Thanks, Robbie. Good morning, everyone, and welcome to the B of A AI Conference. So I'm Alkesh Shah, and I lead our software equity research team, and I'm spending a huge amount of time on AI. So today and tomorrow, we've gathered pioneering leaders in artificial intelligence and how companies are building to meet the future needs of AI, and also to explore how transformative this technology is and how it's going to shape the world. So AI isn't just about algorithms or data. It's about redefining industries, augmenting human potential with copilots, and ultimately having autopilots that may help us with some of our greatest challenges. So over the next few days, you'll hear our B of A research team discuss with leaders across semis, healthcare, finance, security, data centers, and also native AI-native vertical companies.
Whether you're real deep in the weeds with AI or just curious about what's likely ahead for the most important theme in tech, this conference is designed to spark new ideas, equip you with the opportunity to hear from industry leaders, and follow up on any of the topics with our team. Thanks for joining, and now I'll pass it to Mike, Alec, and Andrew to start our healthcare AI panel discussion.
Great. Thanks, Alkesh. Yeah, my name is Mike Ryskin, and I'm on the life science tools and diagnostics team here at Bank of America, and I want to thank you all again for joining the session. The topic and the goal of the session is to sort of discuss how AI and technology is disrupting drug discovery specifically. I'm joined by a few of my colleagues here at B of A, and we're joined by three panelists. We'll go through directions in a second, but just as a quick overview of the format, it's going to be a fireside chat for the majority of the session. You have an opportunity to submit questions via the Veracast portal. Feel free to do that, and we'll throw those in as they come in.
Also, don't hesitate to message us on Bloomberg Chat or email if you want, if that's easier for you. With that, Alec, I'll pass it on to you to do a quick introduction.
Yeah, perfect. Thanks, Mike. So my name is Alec Stranahan. I'm one of the senior biotech analysts here at B of A. I cover about 30 or so mid-cap biotech stocks, one of which is Recursion. I also covered Exscientia, which they are now one and the same, so happy to have Ben Taylor, who's a new joiner, from Exscientia at Recursion on the panel. Andrew?
Yeah, thanks, Alec, and thank you all for being here. My name is Andrew Moss. I'm on the U.S. software team where I cover Certara and enterprise software companies more broadly, and also spend a lot of time on AI from a thematic perspective. So I think maybe for our panelists, William, maybe let's start with you, and we can go through and some quick intros.
Well, thank you, Andrew. Appreciate it. I'm William F. Feehery. I'm the CEO of Certara. So Certara is focused on what we call biosimulation, but what our company is about is around modeling the disease states and the interaction of the drug with the body. So we're interested in things like pharmacometrics and the kinetics of drugs and how drugs get to the active spots in the body and what concentrations there are. And in particular, we're interested in what happens in the clinical phases, not entirely, but that's where a lot of money gets spent in pharma. And we're interested in sort of the differences in populations.
So we're not just interested in modeling how a drug works in one body, but we're interested in the variations that exist across many human populations, whether they're people with comorbidities or genetic differences or it could be differences in food intake or in other drugs they're taking. And we build up a position of quite a number of products that basically inform what we like to say is we like to inform your clinical trial before you actually do it. So the question is, whether you should do it, clinical trials, absolutely you should, but you should take into account all that's known about the science of the drug and what's happening with that disease state before you make that very expensive step in both human risk and in cost in moving into a human trial. Right. Ramy, maybe to you next.
Sure. Yeah. Hi, my name is Ramy Farid, CEO at Schrödinger. At a high level, the platform that we've developed helps scientists design molecules. And the way we do that is essentially by mimicking experiment. So what that means is that on a computer, you can now synthesize a massive number of molecules on a really large scale, and then essentially computationally assay them. And it's the same thing that William just said, the same basic idea, but here it's in discovery. So if you can do that on a really large scale, synthesize a huge number of molecules and test them, predict their properties before you go and make them in the lab, you can imagine the impact that that can have on a drug discovery program. First of all, you can explore way more molecules, of course, than you can explore experimentally by doing it on a computer.
And what that has actually resulted in is a pretty profound impact on the progression of drug discovery programs. It's most definitely, of course, reduced the time and the cost to getting to a development candidate, but probably more importantly, it's pretty significantly increased the probability of success. Of course, the more molecules you explore, the more likely you are to find that perfect molecule that has all the properties that are required for it to be a drug or a material. We're working in both spaces. And then maybe most excitingly, I think, is the most exciting thing about this is it's also, and if you think about it, this is sort of logical, also resulting in much higher quality molecules with better properties, which of course has a pretty big impact on the probability of success of the molecule, for example, in the clinic.
Now, the whole industry is using these technologies. I think we're going to get into that today, so we license our software to the pharma biotech industry and also materials science companies globally, but as we'll talk about later, we're also using that technology in collaborative programs with pharma companies, biotech companies, some of which we've co-founded, and also we have our own proprietary programs, three of which are in the clinic.
Great, and that probably leaves me left over. All right, so Ben Taylor, CFO and President of Recursion UK, formerly at Exscientia. And just to give a quick summary of what we do, it's really, even though the two companies were separate and Recursion focused more on the biology side and Exscientia more on the chemistry side, actually the business model and the vision was very similar, which was create a really tight learning loop between what you can virtually model out, like what Ramy was just talking about, in trying to predict properties in better clarity, but then also combine that very closely with the experimental side to try and validate it and create a closed loop process for that so you can learn very quickly. Because one of the problems in designing molecules or finding targets is really there are so many possibilities out there.
What you need to be able to do is bring together all of that information and learn into a sparse data environment quickly. So what all of our technologies are really focused on is thinking about why a clinical trial will fail, what will happen to a drug as it is progressing, and then trying to create a better predictive system that you can use in discovery to identify the target, understand the target, and design the molecule with better clarity, and then trying to verify that in discovery. Because one of the problems with the current system is so much of the work that we do in the development phase is currently experimental. And ideally, what you'd like to move it to is you're doing your exploration in discovery and you're doing your confirmation in development. And so there's lots of different pieces and technologies.
Some of it's AI, some of it's automation, some of it's traditional experimentation at scale, but they're all on that same goal of how do you improve probability of success by predicting points of failure.
Yes, that's terrific. Thanks, Ben. And I think all three of you in your intros, you left some good hook for us to jump in on. But maybe since some in the audience may be new to the drug development process, maybe we can just level set by saying, what are the challenges or the bottlenecks or the shortcomings of traditional drug discovery? And how do you think leveraging AI on your different platforms can overcome these? Maybe starting with Ben, and then we can go to William and Ramy after that.
Sure. Yeah. I mean, I think there's two different aspects. One is really understanding data, and the other is multi-parameter optimization, so on the understanding data, and this is where the newer technologies that we're all using make so much of a difference. Traditionally, a lot of the experimentation was done in singular siloed, more or less biased processes, and so when you do that, you have to go step by step, and you can't take a broader view of the complexity of biology, and so that's a really important part that a lot of the machine learning has brought in the ability to process all of this complex data that we had to simplify before, so that's a major change. The other part is the multi-parameter optimization, so traditionally, if you look at discovery, it's very sequential.
Solve one problem first, hand off to another, try and solve that problem, hand off to another. What ends up happening is as you optimize each one of those pieces in sequence, earlier parts of the process can get thrown off or you're not able to get total optimization. By doing this in a computational environment, you're able to look at many parameters in parallel. That's why, like if we look at drug candidates, being able to design things that are much more balanced and have as good of ADMET profiles as they do potency is an important design part or being able to identify biological correlations.
So I think that's just been a fundamental change that has led to not only better quality output, but also a lot of efficiency on the time and cost because it's also a much better process in taking that through.
Okay. Thanks. Let me jump in. So our view of the answer to that question has to do a little bit with changing how pharma thinks about things. So if you think about the development of drugs, it ought to be a process of collecting lots of data from disparate sources and synthesizing that to make decisions. What are the decisions? For example, what some of my colleagues here are talking about in terms of what's the molecule that you want to take forward. But it's more than that too. At every point, you have a choice to, do I want to continue with this drug development or not? What's the next trial that should be done to give me the most information that I do not currently have? And so what we find a lot in pharma is that because of the history of pharma, there's two problems.
One is it's a statistically oriented science, so people are oriented to go do a trial, do some with, do some without, have a placebo, get some statistics, and then you prove your case, and that's fine, but it's really expensive, and the other problem is that a lot of data that's out there that's just very difficult to access or understand, so the way we think about things is what you should do when you're developing a drug is apply more of an engineering mindset, so you build a model using, at any point, all of the data you have available on that drug. It might come from AI. It might come from experiments. It might come from clinical trials on other drugs that are similar.
We can find all kinds of data that indicate how a drug might actually perform when it really gets out into a human population. And you develop a model and use that model to say, okay, what do I know and what do I not know? And of the things that I do not know, what experiments am I going to run for the capital I have to advance that at each point? And it sounds logical. It's not really the way most drugs are really developed today.
A lot of the reason why drugs are so expensive is because when you really get into the really expensive parts, at least per molecule around late-stage clinical trials or any human clinical trial, a lot of them really weren't thought through to the extent that they could have been if you'd really developed a model that took into account everything you know about the disease and what's really happening. We see a really big opportunity there. Obviously, we can talk about this unfortunately. I think AI is a piece of this. It's not the whole thing because there's a lot of understanding that we have from literally hundreds of billions of USD of scientific research that's out there that you can take into account.
And there's a big opportunity in the pharmaceutical industry today as they think about how do you become more cost-efficient to basically leverage that. And we're seeing that as we've been seeing more and more of that every year. And that's kind of what we're leveraging as we develop the technology we have here at Certara.
Yeah, so let me.
Good.
Yeah. Is it okay? Let me try and yeah. So.
Yeah. No, please.
Yep. Yep. So let's talk about the design of molecules. So I think you've heard this before, that chemical space is infinitely complex. It's estimated that the number of ways you can combine organic elements into a drug-like molecule is 10 to the 80, right? That's the number of atoms in the universe. So we can call that infinitely complex. So the challenge we have in trying to describe chemistry is what is the training set going to look like? And this has been, and I think William was touching on this, has been a huge challenge for the field is what are we going to train these machine learning models on? And so, of course, up until pretty recently, we didn't have many choices. It was experiment, right? We'd make a molecule in the lab, and then we test it experimentally.
And now we have one data point for our training set. And we keep doing that for a little while until we built up a training set that maybe can describe the chemistry and the properties of that class of molecules. So on a typical drug discovery project, you may make a few thousand molecules to get to a development candidate. So there you go. Now you've got 1,000 molecules out of the 10 to the 80 to describe chemistry and chemical space. Obviously, that's going to work eventually on that one project. But here's the key. Have you developed a global model that describes all of chemistry and all of physics with those few thousand molecules? And of course, the answer is no. And that's been the challenge of machine learning. So where are these training sets going to come from?
I think we've talked a little bit about this. So this shouldn't come as a surprise. But of course, they can come from simulation. They can actually come from physics, from first principles methods. That's another pretty effective way of generating massive training sets on a scale, of course, that's many orders of magnitude larger than just experiment. In fact, at the moment, given the speed of computers and the speed of the algorithms, in one day, it's possible to produce as much data as it would take 10 years to produce experimentally. One day equivalent to 10 years. So you think about now, that's pretty exciting. Now we have a way of really taking advantage of machine learning and AI because we can generate large enough training sets. That's been the challenge for so long.
But it required, of course, developing these first principles methods, these physics-based methods, which, of course, are not machine. I mean, obviously, this would be pretty circular if it wasn't, right? These are completely different approaches. They're using fundamental principles of chemistry and physics and actually simulating experiment. Now, of course, those are computationally pretty expensive. And that's why you absolutely need to have machine learning. Without machine learning, you wouldn't be able to amplify the power of physics, the accuracy of physics, and really scale it up to the level that you need to scale it up, which we've said already is given that multi-parameter optimization problem that Ben described, it turns out to really find that perfect molecule that balances, again, as Ben said, potency and ADME properties. That's a one in a hundred billion molecule.
So you have to find some way to analyze, to test hundreds of billions of molecules. And it turns out now the advances in machine learning plus advances in the ability to build the training sets for machine learning now are allowing us to do something pretty extraordinary. And it is having a really big impact on drug discovery projects across the industry to the extent that essentially the whole industry is using these technologies now.
That's helpful, Ramy. I want to follow up on a point I think a couple of you guys touched on. Ben and William, you kind of approached sort of the strategy or the mindset. And obviously, the technology is a part of it. My interpretation of those answers are, it seems like there's two impediments to even broader adoption. One is having the technology, having the know-how. And obviously, a lot of that has evolved in the last five, 10 years. What all of you, all three of your companies are doing today would have been very difficult five years ago, almost impossible 10, 20 years ago in terms of the technological improvements and the capabilities. So that's making leaps and bounds.
But the other problem is sort of the culture and the mindset and the approach of pharma where they've been doing drug discovery the same way for the last 50 years. High-throughput screening today is high-throughput screening for 50 years ago. Maybe you're doing more high-throughput, but it's still sort of the same approach. And it's just sort of hitting a hammer with a nail. And that culture and that mindset is taking longer to change and to sort of convert. I mean, would you agree with that? And sort of anything you can comment on in terms of have you seen any of the companies, any of the traditional biopharma start to realize that maybe the older way, the traditional way of doing drug discovery is not the right way? Ramy, maybe I want to.
Yeah. Thanks. I think there's something so interesting that I think will shed some light on this, so let's describe first what is traditional drug discovery. What does that mean? What were chemists doing before, so here's what chemists were doing. They were kind of doing machine learning in their head. What they were doing is looking at a collection of molecules, pictures of them generally. Eventually got to a little bit more sophisticated looking at a 3D picture, but essentially an image of the molecule and looking for patterns, and then they would apply those patterns. They'd say, "Okay, I added a methyl group here in this project. Maybe if I add it over here, it'll have the same kind of effect it did on this other project," right? Pattern recognition, and that turned out, of course, not to be very effective.
We kind of know the impact of that. That's why we have 10% success rate to getting a development candidate and 1% to get a drug to market is because that wasn't working. But here's the thing that's really what we discovered recently, and I think really shed some light on what you just said, Mike, is that it turns out that because of the complexity of the physics that underlies these properties, the correlation between the structure and the property is highly unintuitive. So now think about what that means. So you tell a chemist, you run a calculation, and it tells the chemist what molecule to make. But more often than not, that prediction is highly unintuitive.
So the chemist is looking at this and saying, "I've been doing it my way for a long time, and I have this intuition I thought I built up in my head, and the calculation's telling me something completely different." And you're a chemist that has to go spend two months in the lab making that molecule. Are you going to do it? If the answer is no, if you're not used to the prediction being reliable and you're used to using your own intuition, the answer is, it turns out, no. They're really reluctant to go in the lab and put all that work into an unintuitive result from a calculation. Now, the good news is what you said is correct. The technology is still relatively new. You're right. A few years ago, we couldn't do this. So here's what happens. This is what's necessary.
It's a little crazy, but this is what happens. And we've seen this over and over again. We'll tell the chemist, "Okay, fine. Make the molecule you thought was going to work, but just humor us. Make this other molecule." Sounds kind of crazy, right, that you have to do this. This is what's necessary. Make this other molecule that we know you think isn't going to work, but the calculation's telling you. And we've actually done some statistics on this. Around 80% of the time, that's a big number, the calculation is correct, and the prediction is not. So the prediction from the human is not. And you can imagine now the next step. That's it. We're done, right? They've been convinced, and they start to put aside their decades-long intuition that they've been relying on. But you can imagine the barrier to doing that.
That's the reason why the adoption is kind of slow. It's breaking with the sort of human intuition, and you see this, by the way, in a lot of other fields that have been transformed by CAD, by computer-aided design. Airplane design, movie making, right? The cartoonists were totally skeptical that you can ever make a software program make a human look human and look like they're walking normally. You just had to do it once. That was Toy Story, by the way, right? That proved it to the industry, but so you need that. You need these successes. You need to convince humans that the calculations are working even though it doesn't look like they are initially.
Yeah. And if I could build on what Ramy's saying, I think you're right. And it's even broader than in discovery. So everybody in pharma has been basically applying human pattern recognition to every piece of this. There's toxicologists that do some tests, and they say, "Okay, what do those tests mean?" You get into clinical trial design. You kind of look at, "Well, how are other drugs designed?" Or we look at other things like, "Does my drug interact with other drugs out there? If I'm going to take this dosing to pregnant women or to babies, how am I going to do that?" So I think that's sort of that model around, "Hey, we're kind of using human-powered pattern recognition across the whole field," is really what's going on. And I think there's two things that are so you have a conservative, highly regulated industry.
Things don't really happen overnight. But there are two countervailing forces. One is I think there's a widespread recognition across the pharma industry that something has to change, right? You've got an industry which is spending $350 billion on R&D. And if you look at FDA approvals in a year, 40-50, maybe. So productivity of a global industry is not exactly impressive. You're failing more than 90% of the time, depending on when you want to pick up your time. So there's a lot of smart people in this industry, and there's a lot of thought about, "Okay, well, individually, it's very hard to fix that. But system-wise, there's got to be something." And then the second piece of it is actually on the side of the regulators.
So what we found is the regulators really don't like it when you just dump a bunch of clinical trial data on their desk and say, "Look, there's a 95% probability that this is statistically significant." They would like to know why, right? Because that means there's a 5% probability that this is an anomalous result, and you're going to worry about that. And by the way, it might change depending on who you give the drug to. And so they are very receptive to, "Here's a model on why this drug works. Here's a bunch of experiments that prove my model is correct." And based on this model, you can assume certain things about who should get this drug, how safe it is, what the dosing should be in different populations. And they have been encouraging that, right?
I think that's kind of a positive thing as we think about this going forward as far as adoption.
Yeah, and maybe just to add to both the points, and Mike brought up a really good point because it actually shaped our entire business model. If you think about it, we do do partnerships, Sanofi, Roche, others. However, we actually do all of the execution in-house, and that was for a very specific reason because originally it wasn't that way. We actually did design as a service where we would just do point solutions basically and then hand it off, and what we found was it was so hard for the pharma partners to basically organize around the process because when we're doing design, for example, we're doing chemistry, we're doing toxicity, we're doing patient enrichment strategies, all as one component. It's all brought into the same design cycle when we're going through it. They're not separate parts of the process.
When we do our generic systems leading into the modeling, leading into the active learning, in that modeling system, we'll have our experimental assays, we'll have our physics-based assays, we'll have our different toxicity predictions. We'll have the whole. We actually run about 2,000-3,000 modeling simulations bringing in all of the different components that you'll want in that design profile later into the design cycle itself on every design cycle.
And so that was really hard to get into a large pharma system where they're used to saying, "I'm going to do my hit idea over here, and I'm going to hand off to someone else to do this part of it and hand off to the tox team to look at this." And so we ended up integrating it all into a single point where we're basically agreeing on a target with the partner and then delivering the molecule at the end precisely for that reason. And it's actually really interesting. The way that we tried to build it up was actually a plug-and-play system for this exact reason. So we actually don't use the same models for any particular project. Some of them overlap, but a lot of them will be brought in just for a particular project. Sometimes we don't know the structure at all.
I mean, Recursion's background is in phenotypic target identification. And sometimes we're working where we don't know what we're actually targeting. But that's okay as long as you have an experimental system to go through. That's another thing that's really hard for the pharma to get their arms around. How can you not have a clear pathway validation? Now, once you get to the demonstration of in vivo efficacy on the preclinical side, they can buy into it. But if you try and convince them at the very beginning, "This is going to be something that'll work," you may spend years literally validating a target before you ever even start design versus we want to be able to move through the phenotypic discovery into the design, into the validation in vivo.
There's a lot of those cultural contexts that come into it that we needed to just change the way that we work and we partner with outside parties.
Thanks. William, let's talk about Certara specifically for a moment. So can you help us understand how to view the company? Is Certara a healthcare services company, a software company, or somewhere in between?
Yeah. Thanks for the question. Look, we obviously have both software and services. I would say we're a software-centered company because that's really the development and the hard science. We're creating software products that lets us scale across the industry. But we're also practical in that we're in an industry where this is pretty complicated stuff. The number of really high-end experts that can fully use all of our software is not infinite in the industry. Obviously, we're trying to train more people and more coming out. But if we want to change the global pharma industry, we need to think a little bit bigger.
And so we also have a services arm, which is kind of, think about it, in terms of if you don't want to, or you can't buy our software, or you want to scale faster, we can go in and help. So we kind of think about it as we want a center of expertise, and we can't just have a bunch of software developers in a room. We need a bunch of active drug developers here to do it. So we derive a lot of benefit from having the consultants in a room. I mean, some of our cutting-edge projects lead to our next features in our software, for example. And then there's literally hundreds and hundreds of biotechs out there that probably aren't going to feel good about starting a modeling department from scratch if they're running fast.
And so accessing that market means that you need some degree of expertise that you can go in and help. So I guess the short answer to your question is both because that's just the nature of the industry and the nature of what modeling really is. It's not like, "Here's a piece of software. Press a button. It spits out an answer." You still need expertise and human judgment to know how to use these things profitably.
That's great. So Certara's core offering really revolves around its AI-powered biosimulation platform. So I guess first, can you just explain at a high level what is biosimulation, and then how does it drive efficiencies for your customers? And any specific examples would be great.
Yeah, sure. So kind of the software that's the most unique to us is focused on what's called physiologically-based pharmacokinetics. So what that is, is basically we're predicting from when a drug enters the body, however it enters, orally or through an injection or however we're going to do this, where does it go, what are the concentrations in the various organs of the body, and then how is it eliminated or metabolized, right? We're also interested in modeling once the drug is there, we know how much, let's say you're working on a cardiac drug. If it was an injected drug, you're interested in how much the drug gets to the heart. And you might be interested in how much of the drug gets to other parts of the body, maybe where you didn't want it. You got to think about toxic effects.
So we can model that not only for, as I was saying earlier, for one human, but it varies depending on lots of things: age, weight, genetic predisposition. A lot of current drugs, the FDA is very interested in how it will vary depending on what you ate recently or didn't eat recently. And some of the questions we're trying to answer are around dosing. So when you go to first in humans, what should be the dose? Are there subpopulations where the dose is different? Are there populations where you don't want to do clinical trials where you still have to come up with a dose?
A classic one would be, "I have a bunch of adult clinical data, and I need to figure out how to dose children of different ages." Or, "I had to figure out how to dose babies or neonates." Those are really difficult trials to do. They might not even be, in some cases, ethical to do some of those trials. But it doesn't mean you don't need some informed answer. Coming back to the AI side, we developed most of these models. We could call it AI, but the definition of AI, let's be honest, has kind of changed depending on what people are talking about. But we developed a lot of these models mechanistically, right? So we've been mining lots of data and lots of scientific literature looking for the underlying systems biology.
How does the biology of the cell and of the human body really work? What do we know? And the answer is we know a lot, right? And it doesn't all get taken into account in drug development, but we know a lot. We've been building that up over the years. And now AI has come along, and I guess two flavors came along. So one is what happened, let's call it, prior to the large language model revolution of a couple of years ago. And that was kind of the idea around, well, if we feed it a lot of data, we can do pattern recognition. We can sophisticated statistics. We can pull out relationships between variables that we wouldn't have otherwise known based on the scientific literature. And yes, we have used a lot of that to inform our models. Not always.
We prefer to actually have an experiment and have an underlying understanding, but we take data where we can get it, and then more recently, I think, and maybe we've talked more about this, but a couple of years ago, everybody woke up to what they can do when they process language, text, data, unstructured data, and that's opened up a whole world of things that either I wouldn't say we couldn't touch it, but it was just so expensive to process that kind of data, so just for example, we're developing tools that can look through an entire body of literature and create knowledge graphs of how all of the different parameters work together, what the equations might be, what the parameters are, and then kind of list out for a human to go through and say, "Okay, does this make sense or not? How do I document it?
If I'm going to go to the FDA and explain this, how would I explain it?" That would take us. That's probably reducing our cost. It is kind of stopped by a factor of 10 right now. I think everybody will probably agree with me, but that technology has only really been practical for, what, two years now, so this is still kind of early stage in where we're going, so I'll pause there and give some of the other guys a chance.
Yeah. I wanted to.
Yeah. Thanks, William.
Yeah. I was going to jump in quick because it's funny. I feel like AI is too highfalutin of a word, and machine learning is too basic. The reality is somewhere in between, right? Because all it really is, is statistical models that are designed to reflect logic and be able to analyze things on a far larger level than a human can. And so it's kind of like I always think back when computers were coming up or think of Excel, right? No one would think to do math on a calculator if it's complex anymore. They would open up Excel, and they would do that. Think of AI and machine learning, whatever you want to call it. It's statistics-based computation as that next generation moving up and beyond the levels of computation that were available before.
And so all of those different things, and there's lots of different ways that you can do the learning. There's lots of different ways that you can process logic, but they're all based on that same fundamental principle. And so what it's really opened up is this ability to change the way that we think about both big data and sparse data problems. I think one of the biggest misconceptions that a lot of people have is that AI is really focused on big data problems and that correlation analysis. And obviously, LLMs get incredibly high-profile attention. But in reality, I think a lot of the more practical uses of it are actually in sparse data environments where you're trying to quickly identify how to get down to the right answer when you've got that infinite problem like Ramy was talking about before.
That's where you can use active learning. You can use other different methods that incorporate in the negative data to be able to exclude parts of it and focus in on the areas that are more important. And so that's been really, that's why I brought up at the beginning the tie to experimental data. If you put those two together where you are statistically analyzing where you're most likely to have the right answers, and then you tie that to something that can quickly test that experimentally, you can actually narrow down your focus very quickly. It would be called exploration and exploitation in the tech world. And so what you want to be able to do is quickly narrow down in that exploration phase and then optimize in that exploitation phase. And that's where actually you have to get really, really specific.
When we're doing a project, probably about a third of the modeling systems that we'll use are specifically designed for that project. There are some things that are more general that you can use from project to project. Sometimes you can use the public data, though usually the public data is really poor. Most of the time, what we find is for the really important questions, you need to design a new modeling system, a new assay, a new way of testing it, and really develop that proprietary data along with it because what we design are such specific outcomes. The other point that I really want to bring up because there's so much hype around this space is we're actually seeing drugs come through. That's the important part. We're not here to advance tech. We're here to advance medicines.
I know Ramy has drugs in the clinic as well. We've got several drugs in the clinic. We've had a lot of drugs that have in the lab been able to show that they are quantifiably better from a chemical sense than ones that have traditionally been made. But now we're actually seeing that data come through in the clinical trials as well, where we're solving side effect profiles that have prevented targets from being drugged before. We've seen it in multiple different clinical trials where the predictions that we had made back in discovery and design now actually are showing through and starting to make a difference with the outcomes that we're getting in the clinical trials.
This is really the exciting time as we start to transition from what has been theory, what has been promise, into actually seeing how it might change outcomes and the probability of success in the industry.
I have one minute left. Is it okay? Just real quick.
Yeah, yeah, sure.
Did you have something?
I was actually going to pivot to you, anyway, Ramy, but go ahead. What were you going to say?
Yeah. I just want to say something that everybody should take as a pretty encouraging sign. I think you're hopefully hearing that from us. But it's important for everybody to understand, which is what Ben is alluding to, is that actually the whole industry is using AI and machine learning and physics-based methods to develop the training sets. It seems like it's not, because for some reason, what's happened is there are a few companies that are just being labeled as AI companies. And I think a lot of people think those are the companies that are using AI and everybody else isn't. There are a huge number. I would actually say it's the whole industry now. They're using AI, using machine learning, using physics-based methods, and developing. The time is now. It's actually happened.
And even though it's still, I think there's still a much bigger opportunity in front of us because the technology is just coming online. We're still dealing with that skepticism from chemists. But we should feel very excited about what's possible over the next few years as more and more chemists embrace these technologies because they now all have access to it.
I kind of want to, Ramy, where you I think you were kind of talking about this just now, and it's something that both William and Ben touched on earlier, sort of the different ways you can leverage the technology. I want to ask you about Schrödinger's business model and commercial models. I know that one of the things you guys have is you have a software business that's sort of very distinct, vertical, at least on the P&L, it looks distinct. I think you can talk about how it's integrated. Then there's the you've got collaborations, partnerships with pharma and biotech, and then you have your internal pipelines. You've got multiple different approaches of using your underlying technology and platform. What do you see as the benefits of that? How does that help you continue to evolve and develop and push the barriers on what you're doing?
Thanks for the question. I think it's really absolutely key to the success of the company. And here's the reason. So when a software company develops these technologies in a vacuum, if you will, and relying solely on feedback from customers and validation from customers, it's extremely challenging because of the nature of what we're talking about here. This is disruptive technology. It's completely changing the way drug discovery is done. And if you don't break the chicken and egg problem, which is if companies don't have access to the technology at the appropriate scale, they're never going to show impact. And if they can't show impact, they're not going to be convinced to buy the technology. So how do you break that cycle? And you can understand, we talked about all that skepticism. How are you going to break that?
I think the most effective way to do it is to simply use the software yourself. And that's what we did. We had to use the technology ourselves, advance those programs, first in collaboration with companies like Nimbus and Morphic, which were obviously highly successful, and many other ones, and then eventually with our own pipeline. And that generated the validation. That convinced the chemists to do that experiment we talked about earlier, to actually make a molecule that is not matching with their intuition. And so the validation that we got from the collaborations and the proprietary program has been really transformative. And the other thing is you can imagine we have tens of thousands of users of our software over decades.
Think about all the learnings from all of those interactions, all the things that customers do to beat on the software and try different things, and all of that knowledge has gone into the platform. So I think you have to have both to be able to advance the technology in the way that we have.
Thanks, Ramy. Yep. That's great.
Yeah. And maybe one for you, Ben, specific to Recursion. Obviously, the deal that just closed is probably the biggest news item of the year for you guys. How do you look at the combined entity as maybe stronger than some of its parts? And where do you sort of see consolidation for the industry going more broadly, either from the bottom up or from the top down, from your perspective over the next few years?
Yeah. Great question. And a couple of different pieces that go into that. I think we liked it from a technology platform perspective because of that biology versus chemistry and the balance of bringing them together. So a lot of Exscientia had been very efficient on the chemistry and probably spent more money and took more time on the biology. And I think Recursion was sort of the opposite. And so by bringing those two together, we do hope to get a lot of not only efficiency in process and cost, but also efficiency in getting to the best possible outcome because we just have a really nice experience combination. I think another important part, though, is for the industry, which is scale is likely to matter here because what we've seen is a couple of different things.
One, that combination of the virtual with the experimental we think is just so critical. And that takes a lot of investment. That takes a lot of time to put that all together. And if you don't, what we have now is a much broader platform that is integrated together that both of us spent about 12 years building independently and now to bring together. And all of the proprietary data and experimentation and modeling systems that come with that is really fantastic. But also scale on the pipeline and partnership side. I mean, I am a huge believer in risk diversification. And the traditional biotech model really terrifies me. I've lived it. I was operating in the biotech environment before I came over here. And it's really hard to have good data-driven management decision-making and program advancement when you've only got one program.
And so what we have the benefit of is multiple different programs, both internally and with partners, that we can both learn and validate our systems on, but also that we can use to make the best business decisions. And so the entire company doesn't rest on one drug doing well. It's actually a system that we put together. And so those different pieces all come together. Obviously, we were both some of the first companies in the space using these technologies to put it together. But we feel really comfortable that we've got the right set of assets to be successful. And as Ramy was talking about earlier, the industry is going to need to change.
The results that all three of us are producing here are showing people that they have to change the way that they work because the new methods are better. And so having a leadership position in that, I think, is going to be important.
Great. So we've got about eight minutes left and maybe time for one more question for all of you to answer. So let's take out the crystal ball. So how do you all expect the biopharma industry to evolve over, say, the next three to five and five to ten years? William, let's start with you.
Yeah. Look, I think that there's a number of trends that maybe inform the answer. So if you look long-term, look back 10 years, 20 years, look at how much money is spent in pharma and how many drugs come out, you will rapidly come to the conclusion that despite the advancements in all kinds of science, that the industry hasn't gotten more efficient at all. So in fact, you look at peak revenue per drug and the amount of R&D taken to create an active drug, it's been sort of slowly going the wrong way for a long time. And I don't think in the industry people would really debate that. Everybody sees it happening. Now, there's reasons for it, maybe increasing regulation, those types of things. But there's some negative things happening too. So you've got the IRA that hit the industry.
So that's effectively a form of price controls on the really big drugs. And then outside the U.S., there's even more controls on that and some political backlash on sort of the business model here, which is, hey, we don't really need to hit very often as long as we can just charge a whole lot of money. So I think if you look forward, I think there's obviously a tremendous amount of interest in AI in the industry. I think, honestly, it's difficult to escape the assumption that some of that is probably hype, but that's okay. I mean, AI is still new. The industry is putting some money in to figure out what's real and what's not real. Not all of it will work. Some of it will. There are significant scientific advances from all three of the companies on this panel.
They're all saying we're not the only people doing what we're doing in the industry by any means. So this is kind of spreading through. I would say generally, regulators exist to kind of slow things down, but you have generally a fairly sophisticated regulatory group in this industry that actually is paying attention to science and paying attention to the new ways. They don't allow everything, but there's a healthy discussion going on. So I would predict that there's going to be a change. Now, a drug takes 10 years to be developed, so you wonder. So one problem with this industry is like, okay, well, what do I look at to know if it's really happening? I mean, that's a problem.
But if you look over the next 10 years, I think you're going to see basically a change in that, in basically the number of drugs that gets developed for the number of dollars. You don't need to change everything in the industry to have a big change in the outcome. I mean, if you could just change the probabilities of each phase by, let's fail like 5% less or something like that, it would have a massive impact in terms of the drugs that can come out. And then I think the other thing, and maybe some of th e other guys will talk or comment on this, but I suspect you'll see a difference in the types of drugs that come out too, right?
So if you can develop more targeted drugs, it's more likely you'll have higher quality drugs targeted to more subpopulations because if you kind of think about it, if the overall statistics of the industry change and everything doesn't have to be a huge blockbuster, and you can target smaller things and you can have a reliable pipeline coming out because we know more about the science and we can reliably push things out, then you'll get a better outcome in terms of drugs that are targeted more for subpopulations and not for the really wide populations where you kind of tolerate what works really well for some people and doesn't work well for other people. So that's my prediction. I think you'll have more and you'll have a quality difference come out over the next decade.
Thank you.
Yes, Ramy, maybe William.
Yep. I think it's really simple in the short term. What we're going to see is very significant scale-up of the use of computation over the next few years. We're just seeing the signs of it now. I would go so far as to say I think we're getting pretty close to this. We've heard this actually from NVIDIA that the use of their GPUs, which of course are in all the three clouds, Azure, AWS, and GCP, is dominated by the pharma industry. That's kind of amazing if you think about crypto and all the other stuff going on and GPUs. That's a really great indication of what's going to come in the next few years, significant scale-up of the use of computation, physics, and machine learning, and that's really exciting because it's going to have a really big impact. We're already seeing it.
We're already seeing in isolated cases, but I think it's going to be industry-wide. And I think the cloud providers are prepared for this. That is, the compute power is not going to be the bottleneck. I think in the longer term, there's something sort of interesting. One of the challenges that we're facing, actually, that's a barrier to full at-scale usage of these technologies is people to run it. There's such a high demand for computational chemists. There's a lot of work that we're contributing to this, but a lot of people are trying to figure out how to retrain chemists, obviously, right? This is what always happens. The experimentalists need to transition to using computation. That happens in other fields that are transformed by CAD, by computer-aided design. But I don't think it's going to be able to catch up.
I don't think we're going to have enough people to run it. So one of the exciting things, and I think it's worth talking about at an AI panel, is the idea of AI agents actually being able to help humans run these and make really complex decisions. At the moment, this multi-parameter optimization problem you described, that's completely humans making decisions about which parameters do I care about and how do I compute them? And how do I run these calculations to get and how do I do things at this massive scale that we're talking about? I think it's interesting. In the longer term, you asked, I think, 5 to 10 years, that second half, the really fantasy-type timeline. I think keep an eye out for agents, which we're seeing in other AI agents, which we're seeing in other fields playing a pretty transformative role. It's very complex.
I think it's going to take some time, but since we're fantasizing about things in 10 years, I think it's easy to make predictions like that.
Thanks. I don't know if it'll take 10 years to see agents, but I guess we will see it. I think we have to wait for that.
I don't think it will either. I think it's just not in the next three to five years, I think. But I agree with you. I don't think it's 10 years. I think it's sooner than that in this field. I agree.
Yeah. Just quick note at the end. I mean, the industry has to change, right? We have so much money and such smart people, and our outcomes are terrible. I mean, we're failing more than 95% of the time. You can't make a good decision either as an investor or a manager in a 95% failure environment. And that's because of all of these design and discovery steps. Because once you go into development, you're locked in. It all starts before that, right? You're just discovering the properties that already exist. And so I think everyone has now accepted that AI is a way to make it better, that automation and experimentation combined with advanced computation are a way to make it better. I mean, five years ago, most large pharma didn't admit that they had an in-house AI program going on.
Now, every single one of them talks about it, or almost everyone talks about it. And that's just because it is actually getting to better results. Now, there's a big question in how quickly people can change process, what they can integrate in, and how they apply it. I also liked what we were talking a bit about the development aspect. We actually are spending a lot of our time focused on patient enrichment, looking at different ways to optimize clinical trials, both from a design perspective but also a patient selection perspective. There are so many opportunities to actually use the advanced modeling techniques to actually create a more efficient process overall. So you're going to see it change because it has to. We have such a poor starting line to go from, but I don't know the timing of how it all flips over, but it will.
Great. With that, we're actually just over time. So we're going to have to wrap it there. I think we could have kept this going a little bit longer. Thanks, everyone, for joining us on the Veritas. I want to thank our panelists. That was a great conversation, really, really insightful. Ben, William, Ramy, really appreciate your time. Hope you enjoyed it.
Thank you.
Thanks.