Certara, Inc. (CERT) Earnings Call Transcript & Summary

Hello. My name is David Lebowitz. I have with me Neena Bitritto-Garg from the Citi Biotech team. We're happy to welcome you to the second session on the Citi Panel Day. I guess to start out this panel is probably a little bit different than a lot of the other panels, as it's focusing actually more on the technology behind development than specifically on the product of development. And what this panel is going to go after is talk about artificial intelligence; machine learning; computational medicine, which is something that's become increasingly prominent in recent years, considering that way back when, when drugs first started, it was really about people standing over bench tops and mixing test tubes together. We've progressed a long way since then. And with that, I'd like to have each of our panelists to introduce themselves and tell us briefly about their company and where they fit in this spectrum. I guess we can start with Karen from Schrödinger.

Hello, and thank you so much for including us in this discussion. I'm Karen Akinsanya. I'm the President of R&D at Schrödinger. Schrödinger is a company that's been around for about 30 years. We produce software that's used by the whole industry. It's atomistic, physics-based modeling to allow one to predict the properties of compounds. We have deployed that technology, obviously, with the software side, but we also have a large number of collaborations with both biotech and pharma, deploying this technology and latest breakthroughs to drug discovery projects. So there are 7 programs now in the clinic from those collaborations and a large number coming up behind them, as well as our wholly-owned internal programs. Delighted to participate in this discussion today.

Thank you, and we'll jump over to Piet from Certara.

Thanks for having me. I'm Piet van der Graaf. I'm Senior Vice President at Certara, and I lead our Quantitative Systems Pharmacology group. So Certara is a global leader in biosimulation, and we provide integrated end-to-end platforms that are used by more than 2,000 biopharmaceutical companies across the world, including most of the top 50 companies. Our products are also used by many regulatory agencies all over the world. And I'll probably finish the intro by saying that in the last 8 years, 90% of all FDA new drug approvals were supported by Certara's services and software. Thanks.

Got it. And Matthew from Roivant.

Thank you, and thank you all for having us on the panel. So Roivant is -- first and foremost, actually, we are a biopharma company of our own. So we develop -- discover, develop and commercialize our own medicines. We have sort of 2 ways in which we do that kind of differently than others. One is this Vant model, where we develop drugs at a family of independent, nimble, entrepreneurial biotech companies that are sort of subsidiaries or part of our organization, will be different to capture some of the biotech individual spirit with some of the benefits of scale that come from being a bigger pharma company. And then the other one, as a part of this discussion, is from really the beginning throughout the spectrum from discovery into development, we've used computational tools to facilitate. We think faster and lower risk outcomes. And that ranges from, I'd say, more boring, less sophisticated tools in development that allow us to see site level clinical trial data faster, understand what's going on in our trials from an operational perspective, understand our patient population better, to some of the more sophisticated tools. And others on this panel has talked about like molecular dynamics and machine learning. And we combine both of those things in our drug discovery efforts, principally focused on discovering and really designing engineering new small molecules, especially of the form of heterobifunctionals or bifunctional small molecules like targeted protein degraders.

Got it. And Ben from Immuneering.

Great. Well, first, I want to thank you, David and Neena, for the opportunity to be on the panel, and thank everyone for listening. Just start out by showing some love for my GC by reminding you all that we will be making forward-looking statements today. So please see our public disclaimers for more information. Now that that's out of the way, I'm Ben Zeskind, I'm the Co-Founder and CEO of Immuneering. Started the company 14 years ago, really with the premise that every cancer patient deserves a durable complete response. Every cancer patient deserves a durable complete response. And what we started out doing was to really study that small percentage of patients who [indiscernible] these durable CRs at the time, mostly checkpoint inhibitors, using bioinformatics to really understand the mechanisms that were happening differently in those patients from others. And over time, we built up a computational platform based on some of those insights about 4 years ago, started applying that to create our own pipeline of wholly-owned drug programs. And today, our lead product candidate is on track to file our IND in the third quarter of this year. So it's an exciting time for the company.

Got it. Now I hear words like artificial intelligence, AI, machine learning, computational medicine, bioinformatics. And aside from the fact that when I say those words, I know people think I'm intelligent. But at the same time, what does it all mean? How do these words actually play into the world of drug development? And where in the process, from the very, very beginning of drug discovery to long past approval and the days in the post-marketing days, can these technologies be employed? Anyone, feel free to chime in.

Well, I think it's important to perhaps distinguish the sort of terms being used since you've introduced some of those terms. I mean I would say the word artificial intelligence is almost like saying the Internet. It is an umbrella term that describes a segment, but the specifics obviously matter a lot. We prefer to use the words like machine learning that speak about the actual capabilities. Obviously, over the last few decades, we've had more access to significant sort of escalation of data in the biomedical space. And that certainly has allowed the deployment of machine learning methods to derive insights from that data. I mean I think it's really true that it's beyond the capacity of humans now to even read all the papers that are published in your domain of expertise. And so machine learning in the form of text recognition has an impact. The ability to analyze images, I think, is really one of the most exciting opportunity, whether it's high-content cell screens or pathological images. Machine learning has come a long way, and that biomedical data is now very accessible and analyzable, if you like, using machine learning. We also have to recognize, of course, that machine learning has lots of limitations and just putting a whole lot of data into these machine learning algorithms doesn't tell you necessarily everything you need to know about how to design or develop a medicine. And so in our view, combining accurate physics-based methods with machine learning for scale works extremely well for designing small molecule drugs. And of course, this physics-based method is agnostic to data. It's actually target-specific, so it could be a biologic or a small molecule. It really doesn't matter as long as you have a protein structure. So I believe this space is very important to our industry. I think we just have to be careful about the applications and the interpretation of the data. I mean I often rename AI as augmented intelligence, augmented insights that we, as biopharmaceutical drug developers, have access to this now. It's wonderful from that perspective, but we have to be careful.

Piet, what are your thoughts? Go ahead.

Sorry, I didn't mean to interrupt. Piet, you can go ahead.

All right. Well, thanks. So to me, the various terms that have already been mentioned, AI, machine learning, data science, biosimulation, are all part of a bigger picture, which we and all of us in the industry typically call model-informed drug development. Now it's important to realize that model-informed drug development, or MIDD, has been around for quite some time, several decades probably. And I think that it's kind of evolution, it's kind of a hockey stick, where you have early companies getting into that space very soon and early in the 1980s probably already. Then a second wave where some of the methodologies for example, pharmacometrics, became fairly mainstream actually, both in the industry as well as regulatory agency. And now I think we're in the next phase, which I think is driven both by things like AI and data science and machine learning, but also by the development of mechanistic platforms, where we really try to develop mathematical computational frameworks that are based on deep understanding of biology and pathology and pharmacology. And I think that is really, as far as I can see, a very important kind of next phase that has already started, but I think we will really kind of become the next main item in model-informed drug developments. Now I think the -- obviously, these various things, hopefully, will start to kind of be integrated and be used in a more synergistic manner. In terms of its application, I think you said, certainly, we are somewhat unique that we apply biosimulation all the way from very early drug discovery before you may even if design your compounds, all the way to late-stage clinical development as well as even post marketing. So we kind of span really the whole R&D cycle with our biosimulation platforms. Thanks. Matt?

Yes. Thank you. Yes. I guess -- I was going to say 2 things. One is I'm delighted to be on this panel. But I think at least one of the things that we're trying to bring about is an era where -- for example, we don't have panels these days of like people who use mass specs in drug discovery, right? Everybody uses mass specs in drug discovery. It's just part of how drugs are discovered. We don't have panels for like who uses reagents in drug discovery or pipettes or test tubes. I think what we're all building here are foundational tools, and there's many of them, right? Part of what's complicated about this is it's not like we're building just the next mass spec or the next test tube. Rather, we're building an entire generation of new tools that I think if we're successful, will be as instrumental to the discovery of new medicines as any of those tools are. And in fact, in some ways, make it easier than those tools do because the cycle times in a computer or a function of how many cores you can scale up or how big a cluster you can build versus the cycle times in a mass spec or in some of these other instruments, which are set by sort of biophysical limits. So I think it's about creating this next-generation tools. For me, and I'll say, I'll focus -- I agree with everything that both Piet and Karen have said. I think there's applicability across the entire spectrum. Maybe to just pick like a piece of what Roivant is focused on, which is using AI and molecular dynamics both, specifically for what I would call sort of biophysical chemistry problems, right, thinking about the interaction between a small molecule and a protein you might want in a drug, which is just one of many, many, many we talked about, many, many, many problems. For me, what we are able to do with the tools that we are building is to take something that fundamentally historically used to be a kind of luck of the draw, spaghetti against the wall exercise literally, right? You would take warehouses full of compounds and you would throw the big pharma companies literally millions or hundreds of millions of compounds against the target because we actually didn't know which ones of those hundreds of millions are priori, were more likely to bind to the protein and which ones were less likely. And now we've got tools that allow us to not only answer that question, but to engineer specific solutions to chemistry problems, to take a physical problem that we know like this protein, when it sort of folded open like this has 1 biological effect. And when it's folded close like this, has a different biological effect and to say, well, we're going to create a [indiscernible], but we stick inside the protein to hold it open because we want that biological effect. It used to be, the way that works, is you got a protein, you kind of understood what it did and then you throw a million things against it and try it ultimately into cellular assays. And you found out after the fact that, hey, we were holding the protein open and it worked. Now we can start with the problem of understanding that we need to hold the protein open and design something that fixes that problem. And to me, when you say it that way, it just feels like an obvious gap. The reason it hasn't been done universally yet is because it's really hard, and everyone acknowledges that, right? Actually, like -- that sounds easy, but understanding -- even just understanding what it means for protein to look like this versus like this is a really hard problem. But I feel like computational tools have now reached a level where we can start to use that knowledge in various forms, whether it's sort of machine learning and training on real data about how proteins interact with each other or whether it's molecular dynamics and teaching computers the physics of what an open protein versus a closed protein looks like. We can start building tools that do those things for us so that we can engineer solutions to those problems.

Ben, Immuneering users bioinformatics. Can you tell us where your efforts fit into this space? And you're also exclusively on the drug development end, although you started on the services end. Elaborate on Immuneering approach.

Sure. Yes. So our platform really starts with the biology. And specifically, how can we come up with counterintuitive insights into the biology. And I think that's important because for all the progress the industry has made, there's still millions of people dying of cancer every year. And if we keep doing the intuitive thing, we're going to keep getting the same result. So the question for us was really how can we do something that's counterintuitive, but deeply rooted in data. And one of the observations we made over our first decade of working on a number of very effective medicines and really looking at what was happening in those patients that had strong response because what we wanted to do was to say, okay, if we can better understand what's happening in those few patients that have these good responses, we can make it happen in more people, right? We can achieve broader activity. And so one of the things we saw was that frequently the transcriptomics, the gene expression signatures, help to explain what was happening in those patients. And more specifically, we saw the genes that were being upregulated in disease were being downregulated by effective therapies and vice versa. So again, it sounds simple as a concept, to Matt's point, just like it sounds simple to fit the molecule into the right spot. But sort of the math, the algorithms to implement that, to really comprehensively assess how well a particular drug or perturbation is counteracting disease associated transcriptomic changes, that was hard. That took us a while to build, and that ultimately resulted in our patented Disease Cancelling Technology. And so what we did, again, looking for broad activity, we said, let's start with the MAP kinase pathway. This is RAS/RAF/MEK/ERK. It's activated in more than half of all solid tumors. So this is driving a lot of bad cancers. And there's kind of a first generation of MEK inhibitors that have really been limited to RAF-mutant disease, and they're quite toxic. And so we said it, can we understand why? How do we make a MEK inhibitor that has broader activity and better tolerability? And so when we started using our platform, our Disease Cancelling Technology, to look at the biology and specifically the transcriptomics, the gene expression profiles induced by these MEK inhibitors -- this first-generation MEK inhibitors at various time points. What we saw was really interesting. The early timelines made a lot of sense. So at 3 hours and at 6 hours, a first-generation MEK inhibitor does a great job of counteracting disease-associated gene expression changes. But by 24 hours, it's the complete opposite. By 24 hours, a first-generation MEK inhibitor actually induces gene expression patterns that amplify disease-associated changes. And we said, well, that's not good. But the second thing we said is, well, actually that explains a lot in terms of the challenges that these first-generation MEK inhibitors have faced. And that -- those insights, in turn, really led us to do a deep dive into the dynamics of the pathway, the signaling, the timing. And that ultimately led us to design our lead product candidate, IMM-1-104, which we developed in-house from scratch to have this deep cyclic inhibition, where essentially we block a feedback loop called CRAF-bypass. We have a very short half-life. And the combination of those things enables us to hit the tumor hard and release, and basically deprive the tumor of the sustained high level of signaling that it needs to grow and divide, while giving the healthy cells enough of MAP kinase pathway signaling to be less affected. And that ultimately is what's driven the -- what we've seen, which is very strong to our growth inhibition across 5 different animal models with good tolerability. So I think that's an example of how using computation and specifically bioinformatics can really lead to these counterintuitive insights, right? Like you would think intuitively, this pathway is bad, it's on in cancer, just shut it off, right? Just turn it off 24/7. But what the counterintuitive insight from the computation and the bioinformatics was no. By 24 hours, you don't want it off, you want it back on. And so I think that's one example. Later in the panel, we can get into other ways of how we apply computation, but that's an example of using bioinformatics early on to give a counterintuitive insight.

Neena?

Well -- sorry.

Sorry. Go ahead.

Well, I was about to say that's very interesting insights, right, that comes from the biological -- the application of computation in the biological space. And I think what I'm certainly excited about with computation is the idea that we can generate these hypotheses at an accelerated rate. And then at an accelerated rate, as Matt was alluding to, a design with some precision, the molecule that will address that hypothesis and actually go and test this in the clinic. I think that's an exciting iterative loop of computation sort of impacting biology, computation impacting chemistry and probe design or molecule design so that we can test more hypotheses and get to the right answer.

Totally agreed, Karen. And I think that it's -- the chemistry -- the chaining of sort of biological insights from computation to the kind of insights from chemistry to develop better molecules faster as you described and Matt described it. And then I think as -- there's a lot in between from when you have that initial hit to sort of getting to an IND. And I think the kind of work that Piet does also really helps to accelerate that. So it's a complicated process with a lot of steps, but I think there's -- we're seeing, just across these 4 companies, there's room for a computation to contribute at each step.

Yes. Great. So Matt, I did want to just ask you kind of a similar question, right? So within the Roivant portfolio, you've got a number of companies that are working on software-based approaches. You've got VantAI, you've got Roivant Discovery. Maybe if you could talk a little bit about what each is working on and how they kind of fit together ultimately?

Yes, sure. And I think that's actually -- it dovetails perfectly on the comments that Karen and Ben were just making in the sense that, look, medicine will say and certainly drug development really is just like a Matryoshka doll of incredibly complicated problems. We're like we've got really complicated chemistry problems that benefit from computational approaches and really complicated problems to intracellular protein signaling and then really complicated problems of cell biology and then really complicated problems in tissue and organ biology, and then really -- sort of all the way up and then like human health systems data and sort of -- and every one of those levels, the complicated problems that need to be addressed. So we, in particular, I'd say VantAI and Roivant Discovery via a combination of principally machine learning approaches at VantAI and principally, they're not entirely molecular dynamics or physics simulation approaches at Roivant Discovery. Those engines are focused. Now what I would call again, biophysical chemistry problems, right, thinking about how small molecules can be designed to solve problems. And I think it's actually interesting and informative, even in the broader context to think about the difference between a machine learning approach, such as the one VantAI might be taking in those situations, or some molecular dynamic approach. Let's just take protein degradations. In protein degradation, you're designing a small molecule that's designed to bring 2 proteins together. In this case, a protein that you want to downregulate and an E3 ligase that will tag it so the body's natural recycling system will throw it out. And one of the difficult questions there is what kind of -- you're really designing like a tiny little -- like a chain or something that's kind of like a tow truck. It's going to bring these things together, what is the correct place for those proteins to sort of sit next to each other in order to have the effect you want? And on the one hand, from sort of a biophysical molecular dynamics perspective, sort of the Roivant Discovery approach, you can go and you can take a biophysical data about these actual proteins and do structure-based drug discovery. You kind of think people have been doing it for a long time, but on increasingly complicated systems involving multiple proteins where you can start to simulate the interaction between them. Then on the other hand, you take a machine learning approach. Well, there, rather than sort of starting with the structure of the complex or the structure of the proteins, you say, our body brings proteins together all the time, what can we learn from the way our body brings proteins together. What can we learn from all of those other protein interfaces that exist all over the place that might inform us on how these 2 proteins might come together. So even within sort of the smallest piece of the drug discovery question, there are sort of learning-based approaches based on real-world data, and there's what we call ontological approaches based on teaching the computer the rules and then trying to figure out what's going to happen. Then I'd say if we go one step out from there, we've got things like Datavant, which is a company that we originally built, maybe it's not one step out, maybe this is actually 50 steps out. But 50 steps out, Datavant was built to answer questions about patients in the U.S. health care system. So we started at the very smallest level, at the interaction between 1 small molecule and 1 protein. Now we're saying, okay, let's think about a patient's journey between a hospital and a doctor's office and Quest Diagnostics. Well, as it turns out, if you want to study, I don't know, myasthenia gravis, for example. Like the diagnosis of myasthenia gravis in the U.S. exists, let's say, in electronic medical record system. The fact that the patient is receiving IVIG exists in a claims database and the fact that the patient has certain lab values exists in a Quest Diagnostic database. So what Datavant exists to do is to allow us to link the data between those different data sources on a de-identified basis, so that we can answer biological medical questions about those patients. And where the rubber hits the road, I guess, if you really want to think about it this way, is, okay, so we're studying an antibody, actually something that is not particularly relevant to the discovery work that we do, but we're starting an antibody, FcRn. And we sort of stumbled upon new biology. It turns out if you activate the FcRn pathway in a certain way, it drives up LDL cholesterol. We can talk about the biology of it some other time, but we didn't expect it. And so now all of a sudden, we've got these patients who are on our drug, who are receiving a significant therapeutic benefit but are seeing elevated LDL levels. Well, first, you might want to answer the question, does that matter? How many myasthenia gravis patients have elevated LDL? How many of them have hyperlipidemia? Really hard question to answer unless you studied it specifically because the U.S. health care data ecosystem makes them possible. Datavant can help answer those questions. And then you can say, okay, fine, so now we know, we know whether it's an issue in the disease population and which parts of the population is an issue. Now you might say, okay, now we need to go back to the drawing board and say, can we reengineer [indiscernible] a small molecule can prepare this drug with another drug that affects the way that you're impacting [indiscernible] the albumin-binding domain FcRn so that you can go and engineer around that problem, or is there some other biological solution, or is there some other sort of patient health solution. And I think this is just like a perfect example of how if you stack these tools together, you can answer questions at one level, but then give you sort of engineering insights for better design at another level. And that's really why we've got all of these different pieces working together, is you need the tools at each level of the process in order to sort of generate the cyclical benefit.

That's an interesting question. It actually brings me to my next question for Piet. You talked about -- I read an article you wrote on probability of success in development. And you certainly operate in a lot of different parts of that process. Could you speak to where AI could contribute in the process? And then how does it affect this concept of probability of success? And one of the interesting things I found in the article is -- and certainly juxtaposing over what investors think about probability of success is probability of success isn't necessarily measured in terms of does it get approved or not, which often it is. It's a much more complicated question because approved drug is not necessarily a useful drug. And so if you could go through, number one, talk about how Certara plays across the process and this concept of probability of success, and then certainly other panels can chime in as well.

Absolutely. Thanks, yes. So yes, the editorial you mentioned is one that came out a few weeks ago that I wrote for Clinical Pharmacology & Therapeutics. And that was based on the fact that Novartis published a paper, which I think is a really important paper on a framework to quantify probability of success when you transition from Phase II to Phase III, which is obviously the most -- the point where failure is most possible, right? So I think that's why this was such an important paper. And there was a commentary by 2 leaders, tech leaders from Pfizer and also FDA on that paper. So basically, what they say is you can't just define probability of success basically on your gut feeling, right? Because there's all kinds of biases and typically, people have a somewhat optimistic view of the probability of success. I mean they, for example, use the term anchoring as a cognitive bias where people say, well, a hint of efficacy in Phase II typically leads people to believe that, that will be amplified in the Phase III study, whereas more often than not, that may actually be the opposite. So anyway, they provided this really interesting Bayesian framework to really kind of quantify. They also pointed out that, obviously, success is multifactorial, right? It includes not just technical success, but also medical success and regulatory success as well as commercial success. So it becomes really multifactorial. So what could be success for the scientists may be failure for his or her commercial colleague, right? Now in the commentary, I then go back and say, well, obviously, failure between Phase II and III is incredibly expensive at a program level. At the portfolio level, arguably Phase II failure is a more significant issue because still many compounds fail when they are tested for the very first time in Phase II. And obviously, you need to work your way back all the way from early discovery when you start a project and have to define what's the probability of success long term. Now this just becomes quite complicated. Now also, very recently, Pfizer published a series of 2 very important papers where they described what I would call the transformation of their portfolio. So Pfizer said, look, 10 years ago, they were operating below the industry average when it came to Phase II success, which is fairly low in the industry, about 30% or so. Pfizer was operating at 10%, 10 years ago. And they now published data for the most recent cohort. And now they are leading the industry and have a 50% success in Phase II, which is a complete transformation. So they said that was really kind of due to largely speaking, 3 things. One is much more focused on biology, much deeper understanding of science and it really raises sharp focus on a limited number of therapeutic areas rather than kind of a gunshot on everything. Second, the broadening of modalities from small molecules or even conventional antibodies to novel modalities. And then the third one speaks directly to this panel, I think, was actual decision-making based on quantitative metrics and the use of modeling and simulation, data science to really inform decision-making in an objective and quantitative manner. And they -- Pfizer broadened the whole concept of the Novartis framework and used a thing like 3 pillars and signs of clinical -- early clinical activity, et cetera. So I think biosimulation really fits into this picture, where you say we can actually use these platforms and technologies to informed decision-making all the way from very early discovery to make a decision, this is a good target to go after. And if so, should we go after a small molecule or an antibody or something else. Subsequently, selecting candidates, predicting first-in-human dose, and then from there, actual prediction of clinical efficacy and safety in patients, but also in specific patient populations. So this, I think, is where biosimulation fits in really kind of well into this whole paradigm of quantitative decision-making using simulation to inform R&D from start to finish.

It's interesting you used the term MIDD at the beginning, model-informed drug development. I think the comments that Ben made earlier about the real-world success for patients of well-known mechanisms and our ability to kind of take all of the data, the discovery data, the clinical data and perhaps even the real-world data and build the model for what actually is the optimal profile for that medicine. I think in the next 3, 5, perhaps 10 years, a lot of those initial products that went out where perhaps we didn't understand the right sort of framework for how to interrogate that mechanism in the clinic, all of that data now informing next-generation drugs, I think, is a really exciting opportunity. And of course, once we understand that better, in the case of MEK, I think it was that Ben referred to, we can come up with more effective versions as the drugs that have been developed over the past decade or so and actually get to those durable responses, meaningful responses that we're really all looking for, for patients. And so this model-informed, I think there's this sort of loop where we bring all of that data, as Matt was also saying, all the way out to patient outcomes to sort of relearn what is the profile we need for a drug. And I think that's exciting in terms of the products and the targets that we've all been focused on. But I think it's also going to teach us something about the next generation of new targets and what are the generalizable learnings that we can now apply to more novel targets because let's face it, the industry is working on a lot of the same targets and programs. And I think those learnings will help us with that next generation. I think, we're really excited about this at Schrödinger, because right now, we're working on sort of a lot of best-in-class programs, where the learnings from the clinic have actually informed on how best do you go and drive this with a better target product profile using the simulation results, but also then deploying these physics-based methods to really design with great accuracy the profile that you're looking for. But then think about all the targets we've not touched yet. There are thousands of proteins that we haven't yet even approached, whether it's by small molecule or other approaches. And I think that the ability to leverage all of these learnings for our next-generation targets here at Schrödinger, working on first-in-class and novel targets where we understand a lot more of the biology, I think, is an exciting next step.

It's a great point, Karen. And I think it's -- to some extent, it varies from disease area to disease area, right? So I mean with oncology, for instance, right? I mean cancer doesn't work by magic, right? It's not some paranormal force. I think decades of research have really given us a pretty clear understanding of the biology of cancer. And so there, the challenges are more around how do you kill cancer cells while sparing healthy cell, right? How do you thread that needle to get therapeutic index when cancer is so sneaky, if you will, about taking pathways that healthy cells use and sort of co-opting them? So I think there's a lot of room for innovation. But then you look at a disease like Alzheimer's, right? You go to the neuroscience area, and we have a small effort in Alzheimer's to early-stage programs. And there, it's just completely wide open. I think the understanding is much earlier in terms of the biology. And I think there's a huge opportunity for novel targets. And our programs are based on actually having -- not just nominal targets, novel -- redefining the disease, right? So we actually -- we said what if this disease that we call Alzheimer's is actually several different diseases at the molecular level that just sort of all result in the same set of symptoms? And so we analyzed gene expression profiles from hundreds of Alzheimer's patients and found these very distinctly different sets of biology. And then we were able to use those transcriptomic profiles to identify completely novel targets that hadn't been previously appreciated to your point. So they're going towards novel -- completely novel targets and do some validation work on those. So I think every -- and I'm sure in other disease areas, it's at a different stage. So I think it's interesting that each disease is really at a different stage in the research. But I think all of them can benefit from novel chemistry, like what you've described, Karen, all of them can benefit from the kind of modeling and simulation that Piet talked about, and Matt, very interesting around the population analysis kind of after the fact. So there's -- it is really exciting just how all these technologies can help.

One thing that often gets lost in this discussion from my perspective is I think there's an air of like -- and mostly, I find this just come from people approaching the question from different directions, but there's like an air of magic, like permeates discussions around the use of computers and drug discovery. And the truth is that like, look, computers are good at, roughly speaking, 2 things, right? They're good at making predictions based on really complicated sets of rules that they understand well, meaning if we know things that are true about the universe, we understand physics, you can program physics into a computer and then use it to make and it's difficult depending on how much physics, but like predictions based on what's going to happen, or computers are really good at extrapolating -- synthesizing data and extrapolating around it, right? And so those are the 2 things, right? I don't -- I think you find almost every problem that anybody is solving with a computer and drug discovery or drug development is doing one of those 2 things. Either we have rules about the universe that we believe we understand and they could be rules of physics that we understand reasonably well, but there's just way too much of it to do a good job of without a lot of work, or they can be rules of biology, where we think we understand them, but like we don't really know, but you can program them anyway, and you can see what happens if you apply them or you're taking data from the world, right, patient-level data or tumor-level data or physiochemical data or whatever it is, and you're pulling that data in and then trying to extrapolate from it. I think listening to Ben's comments at the beginning of this call about, I can't sort of -- I can't speak to exactly how the platform works, but like you can imagine, computers might be really good at reading hundreds and hundreds and hundreds of papers on PubMed and trying to draw ontological conclusions from those papers that aren't obvious to a human reader because there's just so many pieces to that, right? Like paper is about x tumor type, like a huge number of them mentioned target y, paper is about y tumor type mentioned targets y and z. Maybe Target z is actually relevant to the first cancer even though it's not mentioned in a bunch of papers about the first cancer. And so you'd never get it by reading papers about the first cancer. You'd never get it by reading papers about the second cancer, but a computer that's read both papers might draw an inference that a human might not draw. And the answer is that inference is there for the taking, only because people have done the science in Paper A and in Paper B. If they haven't done the science in Paper A and Paper B, there's no inference to be drawn, there's no extrapolation to be drawn. These models are only as good as the input data or is the rules. But the truth is, at this point, the thing that's really exciting, at least to me about the current moment, is unknowingly, we've been generating data for decades without the tools to fully analyze it or at least we were generating data for decades, and now the quality of the tools has increased, not like overnight, but over a 20-year period to a point where we're starting to realize that we can answer questions about old data. And that to me, when you think about the sort of Cambrian explosion, when you think about the hockey stick, the reason that I feel like we are close to an inflection point on this stuff is because we have 100 years of insights to pull through and just now are able to start pulling all of that data into one place and extrapolating from it. But still fundamentally -- look, I don't think anybody on this call would disagree with me. If you've got a single cellular assay and you generated a bunch of data from it and then you plug that data into simulation and you simulate it, it's not better than the cellular assay, it's the same as the cellular assay. It's just a computer recapitulating a distribution of data from a cellular assay. So there has to be something novel about the connection of the data or something novel about the rules in order to generate an insight.

Yes...

Sorry, go ahead.

Well, I was just going to point to the fact that there are domains in which computation, predictions have immense power to accelerate our efforts. When we think about binding selectivity versus off targets, when we think about solubility, permeability, all very important characteristics of molecules that are required to come up with a development candidate. We think we have a pretty good handle on that. A lot of that is physics-based and we're really excited about what we're seeing in real life projects. But I think as Piet has pointed out, there are -- this isn't magic, right? There are other aspects of drug discovery and certainly development that we still need to master. And so having a significant component of ongoing development of predictive tools to look at some of the more complicated aspect in the integrated system like PBPK modeling and sort of predicting toxicology outcomes, predicting ADME outcomes, that's an area of exciting opportunity for us because having come up with the molecule -- well, having first come up with the biological insight, coming up with the molecule that has generally good properties, we also need to kind of cross that hurdle into humans. And I think we have great tools already. And Certara and Piet's team are working on that, but I think there's even more we can do. So that's exciting, I think, for the future.

Okay. Well, I actually had a couple of questions, but I think you guys have already addressed. So one was going to be just around what do you think the limitations of these approaches are? And I know Matt did touch on a few of them. And then on the flip side, how do we kind of understand what the market opportunity or the total kind of addressable market is for these sorts of approaches?

Yes. I mean I can speak a bit to the limitations because I think the right deal of saying in computer science is garbage in, garbage out, right? And I think you just have to always be very careful with the data that you're feeding into these algorithms and these methods, right? So there have been studies where, what, 2/3 of scientific literature is not reproducible and people go to kind of repeat the experiments rigorously. So I think that leads to a question of how do you build on prior scientific knowledge, while making sure that the input is reproducible and robust? And this is where having spent a decade doing bioinformatics services for pharma, we have a lot of scars, a lot of lessons we kind of learned, learned the hard way by working with large pharma. Biostatisticians have a unique personality type. I love them. They're extremely rigorous, and they don't take -- they don't flex on anything, and it was a great crucible for us because the computation that we built from the beginning had to pass this kind of pharma-grade statistical muster. And then there's certainly a large pharma group of sort of, I'd call them traditionalists, who say I don't believe any of this newfangled computational stuff. Making a prediction that I can then go and validate, to Matt's point, in a cellular assay or in another assay in the lab. And again, for us, that was great because it -- as we iterated on our computation, the things that made predictions that came through in the lab, we kept and refined and everything else we got rid of. So I think the limitations around the data quality and just having this, really intense focus on reproducibility, quality control, avoiding batch effects, things that aren't glamorous but are actually really important. Becky Kusko, in our leadership team, she's on the -- an FDA-led consortium, She's on the Board of it called the MAQC, which is focused on data reproducibility, best practices. And we always say it's one of the least glamorous things, but probably one of the more important things that we do. So data quality, I think, is key.

Yes. I would agree with that. And I think a lot of these computational tools are obviously being generated outside of pharma by groups like ours. And one of the limitations, I guess, is that question of the validity and the impact of these methods. Schrödinger has invested obviously over 30 years in demonstrating the power of these platforms, the platform that we have. And having that now adopted by every pharma at one level or another, I think, is same process you just described then. There's a lot of skepticism at first, but then people try it and they see that actually it's having an impact on their projects and now broadly deploying it across the industry. I think one of the limitations and to your question about TAM is that when you sort of dip your toe in the water and you try these things, the ability to then disrupt essentially the way you've been doing some things previously requires a pretty significant shift in mindset in the way you're organized, in the way teams work. And I think it's going to take a little bit of time for essentially, what we've all been talking about here today to totally disrupt the way in which medicines are discovered and developed. I think we're definitely far into that effort. I think all of us in some way or another are interacting with pharma or they're using the solutions, despite the fact that all pharmas are using Schrödinger and Biotech, actually has thousands of customers using the software. Internally, we use the software at many x log scale, sometimes fold the way pharma uses it. And so the TAM, I think, is pretty large when you think about the biological computation all the way through computation in the clinical arena. I think with full deployment of all of these methods, TAM, I would say, is pretty significant. We're only really at the beginning of this, in my opinion.

So I think, Ben, and Karen already mentioned, well, major limitation or maybe barrier is the fact that some people maybe on this call may think this all sounds a bit like magic and pie in the sky. Well, it's not. Let me just cite a recent survey from the International Consortium for Innovation and Quality and Pharmaceutical Development or IQ Consortium, which is pretty much all the U.S. pharma companies. So they published a survey on the impact of MIDD in some 20 or so recent case studies. So the conclusion was that per program, time savings were up to 2 years because of application of MIDD associated with $30 million to $70 million in development cost per program. And the reasons and mechanisms for these also and time savings were speeding up the time to go, no-go smaller clinical studies, but perhaps most compellingly, the replacements of clinical studies by biosimulation. So our Simcyp PBPK platform, which has been developed over some 20 years now, I think, and it's pretty much used by most kind of people in the industry. So our Simcyp platform has now been [indiscernible] in more than 275 drug labels, where basically, the conclusion is that the potential for drug-drug interaction of this particular medicine is low, not because we did the clinical study, but because we did the in silico virtual trial using the Simcyp simulator. So that's 275 studies. Now running these clinical studies, as you know, is incredibly costly. So obviously, the idea that you can kind of replace that with biosimulation may sound like futuristic. It's not. It's happening all the time. And it will only increase, I think.

I don't mean to pick on the question. I don't love the use of the word TAM, to be totally honest. And the reason I don't love the use of the word TAM is in fact to what I said at the beginning. No one asked what the TAM of the lab bench is, right? A lab bench is a tool that's used in drug discovery. It's an instrumental tool. No one does drug discovery without it. No one even asked what the TAM of the mass spec is. And I guess like the -- reason I worry a little bit about that language is because I think that language feeds into some of what Ben was talking about around the sort of -- and what Karen was talking about around the inherent skepticism that you get because, like you come in and you say we're going to introduce this new method or this new technology or this new idea into drug discovery. And the discussion at some level starts with the limits. It starts with the skepticism. It starts with like convince me to change what I'm doing, tell me that the market is big enough, tell me it's going to solve enough problems for me. And I guess like I approach it the other way. I think like when you hire a really great chemoproteomics person to work on your team, they don't come in the door and immediately face a barrage of questions around how they're going to add value to drug discovery and whether they're going to add value to drug discovery and like prove yourself, right? They're like, here's a hard problem, like work on this problem and help us solve this problem, and then we'll move on to the next hard problem. And people and tools that help you solve hard problems get used in more hard problems and become instrumental, and people and tools that help you solve less hard problems become less instrumental. And the cool thing about computers is they're not one tool, they are many tools. And so the thing we need to figure out isn't what the TAM of a computer is. The thing we need to figure out is which tools are going to help us solve more hard problems and which are going to help us solve less hard problems. And then we need to build more tools that are going to help us solve more hard problems. So that's like a broad sort of annoying, whiny philosophical answer, but I feel like some of the discussions that compare -- even like compare computational drug discovery to RNAi or something like that. It's a new modality. Computational drug discovery is not a new modality. It's a series of tools that could help with any modality. And some of those tools will help more with many modalities and some of those tools will help less with many modalities. But I feel like the sooner we stop thinking of it as a modality in and of itself and start thinking of it as a collection of new equipment and ideas to solve problems, I think the easier it will be to integrate into people's discovery processes. Sorry, that's a little bit of an off-target rant, but it is something that I think is a barrier that we sometimes have to cross.

Thank you all very much. We have actually went up to the end of the time. I know that, that question on TAM is always one that I'd love to ask even though I know I'm never going to get an answer. We just go at a dose of perspective. I think that's from the investor point of view is they don't know -- have perspective on what it means. They know that the pharma industry spends $250 billion a year in R&D. And clearly, AI can't do that at all and computational medicine cannot do that all. But where is it, $1 billion of that $250 million? Or is it $10 billion? Is it $20 billion? That's where they're trying to get. I get it's difficult. But -- and I'm sure you're going to keep getting asked that question again and again and again, probably almost every client you speak with from our end. So with that in mind...

Yes, also, it's the medicines that come out that will drive the value from these methods, I think.

Yes. Absolutely. No question. It's there. So we actually got to wind up, and thank you very much again, and I look forward to chatting with you all soon.

Thank you.

David, thank you.

Thanks, everyone. Thank you.