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

Thank you. Good afternoon, everyone, and welcome to the second panel of the Credit Suisse Healthcare Innovator Series focused on the intersection of drug discovery, development and artificial intelligence. Thank you, everyone, for joining for this very exciting topic today. This is Jason English from the Healthcare Investment Banking team here at CS, where I focus on health care technology and pharma outsourcing solutions. And I'm joined by John Hoffman, another Managing Director from our equity capital markets team, where he leads our biopharma and health care technology effort. As you may have seen this morning, our panel this afternoon, will have a very similar format, a coffee shop type style where we'll try to keep it more conversational and hopefully, lively for everyone. But with that, let me make some brief introductions for our distinguished panelists that we have with us today. First, we have Anne Wojcicki, CEO and Co-founder of 23andMe. A little different than maybe what you see on her screen. She doesn't work for AbCellera. Anne is a pioneer in the direct-to-consumer DNA testing space and has now unprecedented access to genetic information, which is something we'll be sure to discuss today as well. Previously, Anne also spent a decade on Wall Street, investing in health care. Next, we have Carl Hansen, who is the founding CEO and Director of AbCellera. Previously, Carl was a Professor at the University of British Columbia. He's credited with over 100 patents and patent applications. And also is a Co-Founder of Precision NanoSystems. Next, we have Carolyn Magill, CEO and Director of Aetion. Carolyn was also formerly the CEO of Remedy Partners. She's held senior executive roles at Evolent and UnitedHealth, and she has been working closely here recently with the FDA and on a number of fronts during the COVID pandemic. So we'll make sure to discuss that a little bit later with Carolyn. Next, we have Vijay Pande. He's a founding general partner of the Bio Fund at Andreessen Horowitz. He also sits on a number of boards, including Apeel Sciences, BioAge, Devoted Health, Freenome, insitro, Nautilus Bio and Scribe Therapeutics and probably some others that I'm missing on that list. Next, we have Dr. John Hamer, Managing Partner at DCVC Bio. He's held a number of senior positions, including at Monsanto. He was the CEO of Paradigm Genetics. And founded and served as CEO of Arête Therapeutics. And was formerly a Professor at Purdue. Next, we have Daphne Koller, CEO Founder and Board member of insitro. She was the Rajeev Motwani Professor of Computer Science at Stanford University, where she served on the faculty for 18 years. Was Co-CEO founder of Coursera and was named one of Time Magazine's 100 Most Influential People. And finally, we have Ramy Farid, President, CEO and Board Member at Schrodinger where he has been since 2002. Ramy also serves on the Board of Nimbus Therapeutics and on the Scientific Advisory Board of Morphic Therapeutic. So I tried to keep that brief, but I'm sure it could have gone on much longer given all the accolades with our panelists today. But with the intros done, let's go ahead and dig in and get our discussion going on AI drug discovery and development in the intersection of those different categories. So with that, let me turn it over to John to get us started.

Thanks so much, Jason, and thank you to all the panelists. It's just such an incredible group that we're privileged to be joined with today. Maybe I'll start with a high-level question, then we can drill a little bit deeper. What I think is most exciting about the panel today is we're not talking about theoreticals anymore in terms of some of the various technologies that are being deployed. It's in the here. It's now. Major progress is being made. Therapeutics are being approved. Pipelines are being expanded. I would love to have Ramy and Carl just give a little bit of a perspective on what are the key lessons learned? What have we discovered from what's gotten us here and kind of where are we going from an AI perspective within drug development?

Absolutely, would y to start? Or Ramy, would you like to go first?

It's up to you, Carl, why don't you go ahead and start, and I'd be happy to...

Sure. I'll maybe launch this off then. John, I think the big lesson that we have learned, and I think that the industry has learned is the awesome power of new AI and machine learning techniques in being able to pull information out of data that otherwise we could not parse. The lesson that sort of is sitting behind that is that it requires that you have special capabilities in the generation and organization of that type of data. So one of the things that we recognized very early is that biology is a data-driven space, nowhere more so in the immune system, and we set out to build an engine that could do a much better job at generating new types of data that connect cells to DNA sequence to protein function. At the same time, we invested in the tools not just to generate that data, but to organize it and to maintain its integrity and the relationships. And on that basis, we now have the ability to apply our platform to pull out new insights from that data to explore it analytically. That, to me, is the theme that you're going to see across many sectors. And the companies that can control data generation and have the means to maintain the integrity, organize it and bring power to it are going to get accelerating advantages. One more lesson that we had, before I pass off to Ramy, is that, that is very difficult to do if you don't centralize everything. If you're trying to pull data from the rest of the world, you are stymied by the fact that everyone is collecting data differently, you don't know exactly what -- how it was collected, you can't make the connections that you need. And so a big investment in a centralized engine to do that, we think, is a powerful way to get ahead in this space.

Yes. And so one of the things that we've learned, we're in a little bit of a different space, and it actually matters. So where -- in the chemistry space, where really the goal is to predict the properties, very complex thermodynamic properties of molecules, things like affinity and solubility and permeability and selectivity. And these are really enormously complex problems, but there's another really important factor that needs to be appreciated, which is that chemical space, the number of molecules that you essentially can build from organic elements is essentially infinite. It's around estimated to be 10 to the 50 or 10 to the 60. So what does machine learning require? Machine learning requires a training set, and that training set has to be able to represent the thing that you're trying to model. So the thing you're trying to model is essentially infinite. This idea that you can somehow collect enough experimental data is a little bit different than in some of the other applications, right? So one of the things that we found, and now answering the question about lessons is, there needs to be a way of generating massive training sets if we're going to leverage the power of machine learning. And one of the exciting things that's happened over the last few years is the development of physics-based methods, first principles methods. These are not trained methods. These are first principles methods that can very rapidly and accurately produce the kinds of training sets that are required to build local machine learning models that can capture some aspect of chemistry. So I think that's a very exciting development that it's the combination of these physics-based methods to produce the training sets and to capture the complexities of these highly complex thermodynamic properties with the speed and, of course, and the performance. And what Carl mentioned about machine learning, it's the combination, I think, that's allowed us to do something that's pretty extraordinary, wouldn't be possible without with any of the one methods alone. And obviously, the data is really compelling, right? As you said, there are quite a number of programs now that have progressed into the clinic as a result of this combination.

Yes, that's amazing, and data is a great segue. We've had the good fortune of working with Anne on a project recently and just hearing the story about the diversity of data that your platform is able to capture in a really unique way is just extremely impressive and shows, I think, just the breadth of capabilities for novel approaches of technologies into the biopharma space. I would love for Anne and Daphne to spend a moment talking through just the different types of data that your platforms ingest, what you do with it and then how you measure proof points of the success of your platform extending into drug development, I think, would be very important for the audience to understand.

I can start. When we started the company, it was really with this idea that everything that we've said here is that you have to have a large data set, it has to be structured. Reconciling all these different data sets coming together is incredibly complicated. So we took the belief that consumers could actually self-report and that we could empower people to get access to their genetic information, learn about themselves, and that we could ask people to -- if they want to participate in research, and that we could collect a lot of self-reported data instead of a medical record and that it would be highly accurate. And I would say a good part of the first 5 years of the company was about proving out that the data we were collecting is accurate, and that we can actually use that for research studies, and now we can obviously use it for therapeutic discovery. So by going direct-to-consumer, you have incredible abilities to be on TV, to recruit, to go really broad. And so we now have over 10 million people, and we have found that people are really interested in participating in research. And one of the most important things for us was having longitudinal data is being able to keep going back to our customers. And a prime example of that was actually with COVID-19 where we could go back to our customers and in a matter of weeks, we could put out a survey and get over 1 million people to actually take this COVID survey. So it shows the power of having -- it's like a living, breathing cohort of people that can generate data that then we can go and analyze. In 2018, we did a large partnership with GSK on target identification and therapeutic development, and it's been amazing to actually see what we can actually find. We had always known that we could make novel discoveries, but having this collaboration now with GSK has really allowed us to actually move all of these various variants that we can see and move that through the process of actually understanding the biology and now translating those into development programs.

Can I just jump in?

Yes, Daphne. It would be great for you to layer on.

Yes. So I want to continue from where Anne left off and talk about the critical importance of having human data as a ground truth because we spend so much time in the pharmaceutical drug discovery and development effort in dealing with model systems that are just not very good models for human biology, and then we become surprised when something that worked in a mouse doesn't work in a human because humans are just not mice, and we always find that shocking, but it shouldn't be. So to my mind, the ability to collect large amounts of data from humans, which allow us to untangle the relationship between genetics, between various measurable properties of human biology that you can get from imaging and other high content modalities and ultimately, clinical phenotypes, is just an incredibly rich source for disentangling human biology that is truly relevant to humans. The other type of data that we've spent a tremendous amount creating an infrastructure for generating massive amounts of data is to parallel the human data with human-derived cellular models that also speak to human biology and the diversity of human genetics using iPS cells that provide us these induced pluripotent stem cells that capture human biology in a very natural way and provide us with an intervenable platform where you can test out some of the hypotheses that come out from the genetic analysis of human cohorts. And so in effect, we have a very high throughput data factory that generates massive amounts of cellular data that when juxtapose with a human data provides us kind of with the best of both worlds. And so we find that that is an incredibly powerful way to generate an understanding of human biology and then rapidly test it out in the lab where machine learning really allows you to bridge the gap between what you see at the cellular level, what you see in dense human clinical phenotypes and what you see in terms of clinical endpoints.

Yes. Daphne, I watched the presentation that you gave in advance of this panel talking about an EPOC. And it feels very much like at that moment in time in drug discovery and development, there's been distinct technologies that have helped us transcend, paradigm shift the way that therapeutics have been developed historically, and the idea of being able to model biology, yes, on a cellular level and really understand what's going on to be able to model a human system as opposed to a mouse or some other type of preclinical experiment is just a highly compelling way to think about where the future will take us. And Carolyn, I think it may be interesting for you to touch on how data ingestion and what you can do with it doesn't stop during clinical design and strategy and running an experiment. Real-world evidence is obviously a really important component of how therapeutics are developed. So if you could just spend a minute talking about the richness of information that you have coming from that channel and what it does for therapeutics development, that would be fantastic.

Yes, absolutely, happy do that. And real-world data is just data that' reflecting our everyday experiences with health care, and we capture that in some different ways. So where Aetion comes in is building on what Anne and Daphne were sharing in terms of appreciating what are the implications with all these data at our fingertips. Not just what you've mentioned, but also claims, electronic health record, patient-reported outcomes in different contexts, especially in rare disease, and how do we analyze this data to create a level of evidence that is credible enough to drive a better decision. And so our platform takes these data oftentimes in its raw form, and applies a consistent methodology to make clear the impact that a given drug as an example or clinical intervention has on a specific patient population, especially those who have not been represented in clinical trials or for whom data in the RCT space hasn't been collected to the same degree, and this can be used to inform our earliest stages of development for a drug. In fact, we have a recent example of this in the context of COVID-19, where we recently coauthored paper in science. It was in partnership with UCSF. And the first step was, okay, let's get our bench scientists look at the features of COVID-19, understand the virus at the molecular level, test how certain drug properties work to fight the disease, and then we use those insights on our platform and combine it with RWD to understand the severity of COVID-19 in patients taking drugs with those properties. And that way, we can test if they have a protective effect against COVID or not. And from that, we found, okay, there's 2 classes of drugs for further exploration, and they'll then inform for us what clinical trials to run.

That's great. Maybe shifting gears a bit to 2 panelists that are not in operating roles at the moment, but obviously see a number of really exciting companies given their roles at Andreessen and DCVC. We'd be really curious to understand from Vijay and John's perspective what you're seeing in the private company landscape as it relates to companies working with what historically may have been siloed data sets, and how they make them unsiloed, really, get the full utility out of those data sets and how far they can take them into transforming the industry from an FDA perspective, from a regulatory perspective and beyond, a payer perspective? Any type of context there would be great.

Yes. Maybe I'll jump in first. I think when you think about the challenges in drug design or in health care from a data perspective, it really boils down to the fact that data is very expensive to get and very slow to get. And what we're starting to see first is that people are using AI, machine learning, statistical learning, these new techniques for handling data to be bolted on to existing processes. So you're more or less doing drug design the way you were or health care the way you were, but this is maybe the most efficient use of your data to think of it from a statistical point of view. And that's good. I mean especially considering how precious this data is, that's a great first start. But I think the next generation is going to be thinking about how to generate data with AI first that this is going to be going into machine learning. This is going to be -- machine learning is not something bolted on, but it's actually we have a generation of people that are trained in machine learning that are now driving the experiments, are driving the studies. And I think that will be a much more efficient use of this very rare and expensive resource. And philosophically, it's the same thing in either drug design or health care. Animal model is expensive. A clinical trial is expensive. An X-ray is expensive. An MRI is expensive. All of that, I think, will be effected. And in terms of regulatory, I think the good news is that I think FDA is excited to partner, and they're data-driven as well. We just have to be able to demonstrate that the data is there in terms of RWE, real-world evidence, or trial, whichever, that there's statistical significances there.

John, can I jump in on that relative to FDA?

Yes.

Because we're very focused on establishing standards globally, not just with FDA, with EMA, with PMDA, so that they can have the confidence when they're seeing the insights, the evidence from these analyses, there is a level of consistency and credibility that can drive true decisions around safety, effectiveness and value. So we're seeing RWE increasingly used in the context of external control arms as an example. Safety is kind of the more traditional application, but increasingly for label expansions even in initial applications. But that can only happen in a large scale way if we agree to how we assess whether data are fit for purpose, whether the methodology we're applying are appropriate, are we adjusting for confounding variables, are we using transparent science? And this is where I think collaborations among some of the people on the call today would be really important because we can think about the ways in which we create standards and they take into account those source data so that FDA can actually make a decision about whether or not to put a drug in the market or whether a drug is safe for a given population. And those analyses have to be transparent, and they have to be replicable. And so that's where we can marry a platform like ours as an example, but together with machine learning and AI to validate and to demonstrate that, yes, this is something we can make a high-stakes decision on.

Can I also just jump quickly in on Vijay's point because I think he made a very important point and something that really drives our work at insitro, which is the importance of designing data sets that are fit for purpose. One of the 2-edged swords of modern day powerful machine learning is that while it's incredibly good at seeing subtle signal, it's also incredibly good at honing in on subtle artifacts. And if you do not design your data sets with quality, reproducibility and rigor in mind, you're going to over fit the things that have 0 relevance to the underlying biology. So creating, as we did, and I think Carl spoke to this as well, a data factory that produces really high-quality data at the right fidelity and the right scale is absolutely critical to driving truly valid biological insights. The other part is also that when you design data sets where the goal in mind is not some hypothesis valuing or testing some biological hypothesis, but really driving a machine learning model, the experimental designs that you end up with are radically different and much more value per dollar spend and really driving the right kind of machine learning than just taking an off-the-shelf data set and hoping for the best. So I just wanted to reinforce those points because they're really critical to the work that we're doing at insitro.

And I would add that this is a philosophical and cultural shift. This isn't a small tweak, and, I think, if you want to do it right. And so that's something we're -- I think we're going to have to see large change in organizations or new organizations built to do it from the ground up.

Maybe just to finish up, we've spent a lot of time talking about data specifically. But maybe John, over to you to finish up the conversation around data. More specifically, as we think about all these various forms of data that we talked about, be it molecule design, other DNA driven data, we talk about protein cell type of data, various other forms of data that we all have access to. In your mind, which of these do you think are going to be the most critical or are the most critical? Are there certain of these we should be more focused on than others? And are there certain of these that you think kind of in the near term are more critical to the utilization of AI. Sorry, John that was over to you.

Yes, I mean I think one of the real challenges we see is -- yes, can you hear me?

We can hear you.

I'm having a little bit of trouble with my Internet here today, but -- good. So one of the real challenges we have is with clinical trials with things like enrollment times, time line for the first trials and failures in those trials. And yet there's very large data sets around clinical trials and clinical data, and it's very high-quality data. And so one of the new concepts that we're excited about in investing is in the concept of digital twins. It is building a longitudinal computationally developed clinical record that can substitute for a placebo patient in a clinical trial. We all know that as clinical trials get enrolled, you have to enroll a placebo group, you're going to have all group of patients that aren't going to be able to get the drug. A digital twin can be substituted for that placebo group, either in part or in full, eventually, and therefore, be able to drive that, be able to substitute for that patient, allow faster enrollment, allow -- have that digital twin be a complete match to each patient in the drug arm. And those sorts of ideas are now before the FDA. And a couple of trials are being run in just that format. So this is different than a database of placebo patients. This has actually a wholly created clinical records based on a generative model of that particular disease. So pretty excited about how that can change the curve on clinical trials and speed of drug development.

Maybe following up on that and shifting kind of from data into drug discovery time lines, probabilities of success, the development time line itself. Ramy, I'll direct this one over to you. But maybe you can talk more specifically about the application of AI and software in those different categories and how those are kind of importantly, producing better drugs, but also producing drugs more efficiently and cost effectively.

Yes, absolutely. So first of all, to really understand what the challenge is. In drug discovery, in preclinical drug discovery where you're trying to find sort of perfect molecule, it's an enormously complex problem. It's a multiparameter optimization problem. And the reason it's so challenging is because all the properties that you have to design into a molecule are anti-correlated. So everybody knows this, as you start to design a molecule that's more potent, it becomes less soluble. As you try and make it more soluble, it becomes less permeable, and then it starts to bind a herd and then it starts binding the site, right? I mean it's a whack-a-mole problem and it's a really, really complex multiparameter optimization problem. So how do you solve that problem? Well, obviously, you have to explore an enormous amount of chemical space. It's not enough in a traditional drug discovery program that's not using machine learning or these physics-based methods, you're looking at a few thousand molecules a year. Maybe after 5 years, you looked at 5,000, you picked the best one. I mean what are the chances that that's the optimal molecule? Extremely low. And that's why there's so many failure rates. And that's why you have to go 5 years, 6 years, you hear about projects going as long as 10 years. So if you can somehow explore in a short period of time 100 billion molecules maybe, right, instead of a few thousand, then you have obviously a much higher chance of finding that molecule that perfectly balances all those properties. And the only way that's going to be done is obviously through machine learning, with, I think, training sets that are built using these first principles physics-based methods. That's the only way I think to solve that problem. And the exciting thing is it's working. There is a lot of data now to suggest that you can cut the time in half from target selection to development candidate. Obviously, way fewer molecules are being made in the lab because you're making most of them on the computer. And the most important thing, I think, and they kind of go hand in hand, is the probability of success has been significantly increased. Because when you explore enough molecules, you will eventually find a molecule that somehow balances all those properties. You can essentially eliminate chemistry risk by exploring enough of chemical space. And as I said earlier, it's a fortunate situation the chemical space is essentially infinite. So usually, when you have a situation like that, you will find a molecule. You just have to be able to explore enough of them with very accurate models.

If I could just add a couple of thoughts to that. I think it's a great launching point. So building on what Ramy just said, we take a bit of a different approach because we benefit from a natural immune system that has insane diversity that is generated through immunization or through infection. So we start from the position where there are potentially trillions of molecules, and we apply AI not to invent or engineer new molecules, but rather to find and characterize and predict which of the existing molecules that are ground truth in nature will most quickly and efficiently go through to be drugs. So we do that in a few ways. One is applying AI to the discovery process. So being able to process many tens of millions of images within hours in real time to find individual cells with the right properties, the genomic AI and then the prediction of their properties from biophysical measurements in terms of development. And a great example of that, I think, in the real world recently of that applying is responding to COVID-19 where we were able to get a blood sample. And in a case where speed really mattered, we're able to use machine learning algorithms trained over years with data sets to quickly screen through 5 million cells in 5 days to isolate 500 unique molecules to generate hundreds of thousands of data points and then to have the analytics that we don't get buried in the data. And that's an important point that probably hasn't been brought out. It's not quite AI, but without the right software tools, you are in a position where you're looking at thousands of spreadsheets with hundreds of data points that are completely opaque to any human observer. And you will spend months, if not years trying to decide where is the information in that data. So that's another very important thing that machine learning and just software competency is unlocking in this space where generating the data is a big part of the problem. But without the analytics, it's an empty victory.

And I would add, we're kind of in the beginning of what I see as a 10- or 20-year arc that other industries have gone through before. So if you think about finance, maybe 20 years ago, it would be seeming ridiculous that computers replace traders. And now 20 years later, it seems crazy to think that a human being could do what a computer could do, and we've had that flip. I think that's what we're starting to see. Whether we're talking about machine learning or physical simulation or any of it, we're seeing this shift to essentially in a more industrialized one, something where we're not having sort of this bespoke artisanal process of designing a drug or recovering a drug, but we're really rethinking the whole process, doing it in a systematic data-driven way and really taking people out of the part that computers are going to be just better.

And maybe...

And so maybe getting -- sorry, go ahead, Daphne.

Well, I just wanted to jump on to Vijay's point, agreeing with it, and saying that in order to really deliver on that value that Vijay just outlined, which is through this complete shift in perspective on what the role of people are versus what the role of computers are requires a different kind of company because the cultural inertia of companies that have traditionally done things in a certain way and coming and saying, "Wait a minute," there's a whole group of like 1,000 people that can be replaced with a few computers plus another maybe different group of people with different skill sets who can really work in harmony with the computers. That is not a cultural shift, it's a lot -- that's very easy for an incumbent to take on. And that cultural inertia really makes it challenging, I think, for companies to make that shift. And similarly, when you look at the tech space, because as Vijay pointed out, this is not a transformation that's unique to biopharma, it's happened in many, many industries already. In very rare cases, is at the case that the incumbent makes that digital transformation versus you don't have Blockbuster turning into Netflix or Walmart turning into Amazon. And so I think it's an interesting question of how one builds a truly digital company in the biopharma space.

Getting into the nitty-gritty on that, Daphne or any other analysts. Just curious, you talked about disruption within biopharma being -- following a playbook that other sectors have followed historically. When you actually think about the code and the software that's being utilized to do that, how much is transferable between industries? Are companies borrowing code open source in nature from other industries to apply them in a unique way. Vijay, I know that you've -- I think you've spent some time thinking about this.

Yes, I've spent some time thinking and actually doing a lot of this as well is that I think there's different types of codes. So the basic infrastructure like TensorFlow or whatever, that's very transferable. And then the next generation is that people look at the amazing things being done in a certain area like image analysis. And in biology, there are images, maybe cellular images or radiology or whatever, and maybe that can just be wholesale taken over. But in the end, is a DNA sequence like an image? Kind of there's some similarities but some not. Is a small molecule drug like an image? That's a lot harder. And so I think the first generation is bringing over the infrastructure, bringing over the algorithms. But what's, I think, really special for now is that we've got deep domain experience biologists that also are great machine learning AI engineers, and they're developing new algorithms, algorithms that weren't intended for images or other areas that other tech companies are interested in but areas that are biology first. And that will be the big, big transformation. There's only so much you can do to shoehorn an existing algorithm to this other data set. And what is new now is that we have biologists that are trained in the machine learning, and machine learning engineers now understand the biology.

I can't help myself, John, I just want to reinforce something that Daphne said that I think really resonates as true, which is that the one thing that these new companies have as an advantage is that you get to come to this space on the heels of 20 years of unbelievable advancements in measurement and computation. And these tools are there for you, and you can rebuild at the frontier a new platform that is completely modern. If you're in a large organization that has been doing this type of work for 20 years, you're looking at some costs, you're looking at organizational structures, you're looking at business models that make it very, very difficult, even if you are rational, to put the effort on that. And so I do think that it is the case that these new companies will be leading the charge, and that will be good for the entire sector because they will find ways to bring the molecules forward that the bigger companies then have the expertise to take forward, and it's going to happen through collaboration. So I couldn't agree more with that that there's a sociology business model culture barrier, but it's hard to really pin down, but it's a real thing.

And there's the temptation to view AI as like the next top thing like the genomics or structural biology or whatever in that you want to sort of put it in as this new technology. But I think the big difference here is that this is not just another useful technology, this is something that transforms everything. And that it transforms every aspect of the way you do things, you should be even thinking about the deals you do, the clinical trials at every stage. And that's the part that's really hard to shift. You could slot in structural or genomics into existing pipeline much more easily than you can completely transform the way you do everything.

Yes. I heard people in biopharma saying that machine learning is going to be like x-ray crystallography. You're going to slot it in here and it's going to accelerate this thing that we've already done anyway. And that is just such an incorrect perception of the power of this technology.

Very limited.

Yes. I think -- I was going to -- maybe, John, one question to go over...

Yes, I think one of the things you got to keep in mind is...

Yes, please. We can hear you, John.

No, I'm just going to comment. I mean if you look at the way big pharma traditionally does -- okay, yes, I think my Internet is pretty unstable. Can you hear me now?

We can hear you fine. Yes.

I think he can't.

We could, I think he said that.

No. I was just going to point out that historically, big pharma always wanted to acquire assets -- yes. Okay. I think I'm going to hold off here. I think my Internet really acting up for some reason.

To build on what I'm hearing John say potentially is that we found that pharma has been optimized for each step along the drug development life cycle. And instead, what we're really on the precipice of doing is creating more of an enterprise view. How do we drive systemic change that is built for purpose? Now I would counter some of the hypotheses here that some of the larger companies can't do it because we've certainly seen amongst some of our customers who are in the top 20 biopharma a level of commitment, at least with respect to what data are we accessing across the organization and how do we at least start looking at that in a consistent way? And if I'm creating patient cohorts, I'm defining conditions, I'm thinking about the data that are relevant in the early days of drug development, what of that is also relevant once the drug is on the market or as I'm preparing to talk to payers to demonstrate to them which population should have access to my medications and at what cost share, as an example. And so if we think about more how to optimize for the system, we're able to get a better sense of then how each part of the conveyor belt should start interfacing with one another. And one of the main ways to do that is obviously through data and technology and thinking much more holistically about what we turn to when and why, and very specifically, what decisions do they drive? Like, why do I care about this new innovative data source? Because if it's not driving a different decision, then I don't have time to prioritize it.

Yes. I think with any technology adoption cycle, there's friction when you try to push a technology into people that are incumbents and obviously, have to adopt new ways of discovering therapeutics. But Anne, I'd be curious from your perspective, I think there's some really strong tailwinds from a pull perspective on the consumer side. It's the health care just seems entirely transformed right now with consumers empowered to want to control their health, understand it and just be engaged in a more holistic way. One thing that we've talked about in the past is the concept of your database being re-contactable, and what that does for the company? Would love to hear some perspective on that and how you think about that dynamic?

Well, I think it's an interesting asset because it ties in with all these things, but the key potential, I think, going forward for drug discovery is really going to be about absolutely applying machine learning. Like, getting all the right tools applied, but the foundation is you need the data sets. And so part of what we thought here is, like, how do you really create these massive data sets, like, bring together massive numbers of people who want to contribute. And one thing I think that I had sort of found had been overlooked by the industry is that people actually want to participate They're just not really given this opportunity. And so when we give people an opportunity to be part of the discovery process, which frankly is really the solution for consumers, like, they want to find a drug or they want to find a prevention, they want to participate. So when I think about all the companies that are here and everything we're talking about in terms of data sets and one of the core aspects of data for me was finding the right patient population and being able to have longitudinal massive numbers of individuals who want to be ongoing and participating. So I think it will be interesting to see in the future, and again, I've enjoyed hearing everybody chat about this is, like, how then all of these various -- how all this actually starts to come together to really make an efficient drug discovery process. And that's where I can see -- like, my ultimate goal when I think about the dream is how do you actually really go from gene to function in a computational way. And I just, like, I was amazed at -- and I'm sure others here also had that sort of sense of awe, like, what, DeepMind did with protein folding is amazing. And so being able to see what those next steps, but when I think about the long-term goal here is I want to be able to go from sequence to function and be able to predict that quite accurately. And then that's going to have -- you're going to be able to tie in all other aspects of then how do you actually modulate to change that end function.

Anne made a really, I think, important point about DeepMind in AlphaFold, which I agree is an incredible landmark of achievement, not least of which because there was a clear benchmark of performance, and it was a true blinded experiment, and there were all these other comparison points and yet they were able to exceed that performance. What was, I think, the important aspect there, though, is that they actually did have enough data of that mapping of, in this case, sequence the structure so they could train them all. We said, "You know what, everything we know about physics and chemistry that's all very well, but let's let the machine learn for itself, and that really relied on so much data creation that had been done by structural biologists and chemists over time. And I think those are -- that's the kind of data that we need to think about generating in order to solve the problem that you laid out and which is how do we go from a gene to a function. We just haven't made that same data collection effort. It's a much harder problem because the data modalities are so much more complex and multi-variant is so much dependency on context, but it is the holy grail of how we really make drug discovery more efficient and higher probability of success.

And even when we say function, do we mean protein function or cellular function or it gets complicated quickly.

Maybe, Vijay, one question over to you, and we can ask this at a couple of people here as we finish off. We've talked a lot about where we're at today. I think we've talked a little bit about some of the vision of where we can take things. If you were to look in your crystal ball 5, 10 years out, kind of where do you see some of these capabilities evolving to?

Yes. No, I think we're at the beginning of an industrial revolution. I think if you look at how labs are done, like, lab biology is done, if you take a picture today and put in black and white, they'll look the same as a picture of 50 years ago. There might be slightly different boxes on the benches, but it's a very bespoke process, very slow. And I think we're seeing beginnings of fully industrializing this. And like any changes, this is not going to happen overnight. There'll be low-hanging fruit that people will be able to go after immediately. But I suspect over 10 or 20 years, we will get to the point where a few drugs are designed with AI to all drugs will be designed with AI. And that it's crazy to think computers are useful to. It's crazy to think human beings could load up 1 million data points in their brain and rock it and say, "This is the drug and not this one." And it will take time. And I think the people on this panel are some of the key people driving this. And I think the exciting thing is that this isn't just about drug design. This is all health care data. So this is whether you should get this X-ray or this MRI, how to interpret the MRI? Should you get the surgery, how do you make any decision-making? To make decision-making driven by data and data ideally about you specifically would be such a fundamental transition in health care that -- and it's all -- in some ways, it's all the pieces are there. There's just a lot to build.

Maybe Ramy over to you the same question.

Yes. I completely agree with that. I would just be careful about using AI as a sort of surrogate for just anything computational. It's -- AI is going to play a critical role, but of course, advances in structural biology, maybe DeepMind is the beginning of that, but there's a lot more to do. Those are AFold structures. So we still have a lot of work to do to really understand what those structures look like when the molecule is down. So I think -- again, I totally agree. I think in the next 10 years, building on the advances that have already made, again, in our deep understanding of the physics of these interactions, which is really valuable combined with AI is totally going to transform the industry. No question about it. I think it's just amazing how fast already it's happened. I mean a couple of years ago, some of the things -- just 2 years ago, the kinds of things, access to hardware, what Google and Amazon and Azure are doing, the availability, right, of compute resources, again, our deep understanding of the actual problems, the details of the problems at an animistic level are profound. And there's no question, this is where the industry is going. But as Daphne said, I mean pharma companies are -- can be very challenging to interact with, people are worried about their jobs. I think this is -- you've seen that in other industries, right? Other industries went through the same thing, people are really worried. There's a lot of apprehension, but we will get there for sure. No question about it.

And to Ramy's point, there's a lot of pretraining that's missing in these models. Models don't know anything about physics. They don't know anything about biology. They don't know anything about chemistry.

Exactly. Exactly.

And that's what can be gained from a physical model. But a physical model in a sense from a learning point of view is just another way to handle the data that we have. And so I think we need to incorporate that in a pretraining sense. And in essence, AI can subsume all of it. It's just handling all the data, whether the data is...

Yes, I see your point. Yes, I think that's right. Yes.

It's just how to handle all the learning that we have.

Yes, yes, yes.

I was just -- I do think that I agree completely with the long-term asymptotic way that this is going. So it's really hard to look out 10, 20, 30 years, 40 years and even be close to where the impact of this will be. I'm a little more tempered on the prospect that we will look 10 years out and see a lab without a lab. I think that is -- that's probably not where we're going. I expect that you're going to see added investment into data generation capabilities. You're going to see AI applied to be able to do that faster and more automated. And you're going to see a flywheel effect of the companies that control that, that can generate high-quality data that is asking the right questions and have married that with computation, and the culture to be able to do that are going to get way ahead. But the one thing I've learned over many years of doing tech dev is it always is much better than you think it's going to be, and it always takes way longer. It doesn't mean that you shouldn't persist. It's going to be a long-term effort.

I think it was Bill Gates that said that we always overestimate the effect of any new technology in a 2-year time frame and underestimated its impact on a 10-year time frame, except in drug discovery and development, you probably want to multiply that by 4.

My bar is much lower. I just think we won't -- in 10 years, won't be doing drug discovery by trial and error anymore.

Carolyn, maybe your perspective, kind of where you see things about 10 years out?

Yes. I think all of what people are saying resonates. And I think the most important thing is that we bring a level of transparency. So even as we iterate, even as we have ideas about new ways that we can design, that we can use our computations, the more that we can help people bridge where we are today to where they're going. And we run into this all the time in the real-world evidence space, and we were built specifically for the purpose of transforming data to evidence to drive regulatory decisions. So we really understand the value in that. But we also understand that we are erring as an industry on conservatism for good reasons because lives are at stake, because there are methods that are tried and true. So part of what we also need to do, and this is societal change, this is -- I mean this applies to all aspects of our lives, I think, is recognize where people are today and help to illuminate that path. And a lot of times, that is with trust, replicability, transparency in what we're doing, why we're doing it. And so the more that we can validate some of our innovations, even in the context of the way people speak and look at the world today, the more likely they are to be receptive to innovations as they come.

Yes. Awesome. Well, thank you, everyone, for joining us today. I think we've come up against our time. This was a fascinating and amazing discussion. I think, incredibly important work, obviously, that all of our panelists are doing. So thank you, again, for being a part of our panel today. Thank you for all the attendees for listening in and wish everyone a happy afternoon.

Thank you.

Thank you.

Thanks so much.

Take care.

Thank you.