Nick Myerberg, Partner & Head of AI at Braidwell on AI's Next Era in Drug Discovery

0:24Nick Myerberg: Nick Myerberg. Wonderful to have you on today, Nick. Thanks, Rahul. Excited to be here. So Nick, to kick us off, we're going to cover a wide range of topics today at the intersection of AI and biotech. But for context, I'd love if you could talk to us about the arc of your career and what got you to where you are today. Sure. And maybe we can start more recently. So I've been at Braidwell since almost the very beginning. I think it may have been employee number six or eight. And Braidwell's a life sciences investment firm. focused on trying to figure out how to build and back the companies that are going to transform human health.

1:02I was just talking to Alex Carnell, one of our co-founders and our CIO, and he put it this way. I really liked it. He said, over the coming decades, we want to figure out a way to impact a billion human lives. And so we want to do that through finding the companies that are making the medicines or devices and diagnostics that are going to improve those lives, or in the cases they don't exist, building those companies ourselves. Practically, we do three things. We invest in public and private companies through our flagship strategy at Bridewell Partners Fund. We make loans to companies in our space.

1:30And then, as I mentioned, through our Bridewell Labs subsidiary, we build new companies where we find massive market needs where somebody isn't playing in that space yet. I came to Bridewell largely because I thought the opportunity to work in biology was really exciting. And I followed a bunch of people that I built strong relationships with over the years that I'd spent at a previous company at Bridgewater Associates, a large global macro hedge fund. Our other co-founder, Brian Kreider, was one of the best people I'd worked with. I had other people that I'd spent time with who had followed him there.

2:02And I thought, great problem, great group of people, an opportunity to scratch an entrepreneurial itch that I'd had. And so I jumped at the opportunity. I'd spent my entire career before in various ML practitioner roles and machine learning research and engineering roles, largely focused on figuring out how to extract noise from messy data, trying to figure out how to make hard decisions under uncertainty. And so I had this background really professionally as somebody who had built software, built models and pointed them to business problems. I came at it from a little bit of a, maybe an unusual place where I joined Bridgewater, which was really my first job out of grad school after spending time being trained as historian and philosopher of science.

2:47And this is kind of like a common story for me across my career and across my education, which is that I love living at the intersection of two things. I love saying, you know, I'm trained as a philosopher, but I work on ML research, or I spent my career doing machine learning and software engineering. And now I'm thinking about biology. Those are always really fun places. And I find them to be exciting places where you can put very different ideas in conversation with one another. The fact that at Braidwell, we're trying to improve human health just gives a North Star. That's really exciting.

3:16Nick Myerberg: And Nick, talk to us a little bit about your responsibilities at Bridgewater. That's first part of the question. And then perhaps what surprised you when making the switch from that role to obviously a very different sector and what carried over well? At Bridgewater, I was a technologist. I spent time building software. I spent time building statistical models, doing machine learning. One of the exciting parts about being at Bridgewater at the time, by the way, was I think I had just started when the famous attention is all you need paper had come out that had inaugurated this many now generations of models based on the transformer architecture that undergirds your chat GPT and your Claude.

3:54And my job was largely to find ways to understand these models, whether they're these super sophisticated transformers, statistical learning, or more sort of prosaic rule-based models and figure out how to use them to push the business forward. So really, I had this hybrid role of technologists, machine learning researcher, and somebody thinking about business. When I came to Braidwell, I, of course, leveraged a lot of these skills. My job at Braidwell really is split in two. On one hand, I think about how we build systems internally to invest better, to take our team of phenomenal biologists and biostatisticians and all these other specialist disciplines and couple them with software, couple them in particular with agents these days to go and solve questions faster and more creatively, more laterally, and in ways that are more systematic.

4:45So that's a piece of continuity. The domain though is very different. In a lot of ways, my specialization has been in being a generalist, at least before coming to Braidwell. My job was to get really smart on some piece of the business, go about three inches deep, learn enough to build the model, and then move on. Biology is really interesting, and I'm sure we'll talk about this more, in that it is deeply recalcitrant simplification. You have to go deep, you have to understand it deeply, and you have to bring people with you, because the depth of it requires lifetimes of study and of specialization.

5:17And so that was, if not a surprise, was a pleasant difference. It's a place where I've been fortunate enough to sit with really smart people have spent careers in biology, not in computation, and try to bring those perspectives together with a computational perspective and with a market perspective to understand how we can turn that into an investment thesis or into the thesis for a company that should exist.

5:40Nick Myerberg: And at Bridgewater, I understand you ran an internal skunk works and worked on applying the firm's systematic investment approach to kind of how the firm actually managed itself. Are you running a similar playbook now at Braidwell? And love if you could talk about that as in turning the fund itself into a machine. And what does that actually look like day to day? Yeah, you know, I think that the thing that's neat about coming out of a culture as strong and as pervasive as Bridgewater's is that when you come and create a new firm, you get to make a decision on what you bring with you and what you don't.

6:12And I think everything we do is, of course, influenced by the things we learned at Bridgewater. The biggest thing that we did from the very early days when we're thinking about building our investment process was to think of it like a machine and think about it systematically. A, because in the short to medium term, you get better outcomes by understanding your process, understanding where you're great and where you're bad, how to improve the places where you underperform. You only get there by collecting data, by writing down the outcomes, by understanding how you can modify that design to get you better results.

6:42The benefit of that is that along the way, you generate data sets of real-life investment decisions based on real-life evidence where you have an outcome. And we've been measuring this and collecting this data for five years at this point, such that when you start thinking about how to integrate that into a modern agentic system for investments, you avoid this cold start problem that I think a lot of people run into. We have this special collection of insights that allows us to now start integrating AI throughout our investment process. These days, we're in an early beta of a system we're very excited about with some promising early results.

7:21But we're only able to do that because we started with a systematic approach to how to think about building an investment organization.

7:28Nick Myerberg: Great, Nick. So now shifting gears a bit, you've been working in AI for quite some time. You're now obviously working at the intersection of AI and bio. and where we are right now. For listeners who've only really followed the headlines, I'd love if you could walk us through the actual history here of, let's say, the first wave of AI-native biotechs, AlphaFolds, the enormous funding that followed. And from where you sit now, what did that first generation actually promise? And then I'd obviously love to talk about what's different this time around. Yeah, it's a great question. And I think it's really easy to look at AI for biology and say that it began in recent memory.

8:10And I don't think it did. I think biology really spent more than a century creating representations and doing calculations that made a difference. You can go back to the early 1900s and look at people like Darcy Thompson, Fisher, Wright, Haldane, Lotka and Volterra doing modeling for ecology and see the early seeds of people trying to produce this computational revolution in, at that time, kinds of biology that were very far from the cell. Even more recently, if you look to the 60s through the 80s, we had this massive focus in computational biology and drug discovery in particular. You had quantitative structure activity relationships, you had structure-based drug design.

8:51People have been trying for quite some time to think about how they can take a process that is, I think, arguably one of the hardest things that human beings do and bring some mathematical tools to the fight. The thing that was different in those days versus the deep learning era, let's say through today, was that in those periods of time, people, they've relied on these human chosen representations. The scientists had to sit down and say, for the problem that I'm trying to solve, how do I craft something that the machine can understand? And that would be molecular descriptors, force fields, similarity assumptions.

9:28It's easy to look back at that and say, oh, well, that was silly. Why didn't they build a big deep learning model? But in a lot of ways, that was an epistemic virtue, not a weakness. The idea was that you had real assumptions that were encoded in these models, and they were real scientific knowledge. And I'm not sure exactly how to pinpoint this pivot. Let's say in the early 2010s, this deep learning revolution did something really important to that representational labor. All of a sudden, you started having these machines that could start developing their own pictures of the biological processes we were trying to understand.

10:03And this was the context in which you saw this first generation, really, of AI for bio companies, call them Emerge. They were combining the previous general physical models with representation learning, with deep learning. They were trying to fuse those things to push towards better understanding of ligand design, better understanding of how to do classic structure-based drug design tasks, or doing things like large-scale phenotyping through microscopy across many different ways to enhance how high-throughput screening already worked. We saw a number of companies form and show really strong results in some cases in phase one, and then falter as those drugs went through the pipelines.

10:45We can talk a little bit about why that's true. But I think suffice it to say, there was a massive amount of progress made on just understanding how to measure things at scale, how to build models that are suitable for specific biological questions, and then how to build organizations that brought together computer scientists, biologists, people from the business side, and think about how to build an AI native drug discovery organization. In more recent years, this move towards generative modeling really just changes one important thing. It shifts from predicting supplied objects, say, you know, hey, what do we think the PK and PD of this particular molecule is, to creating objects that satisfy certain kinds of criteria.

11:27There's nothing too unusual, at least at a conceptual level, of what's happening there. I think the biggest change and the biggest point of evolution is this move towards systems that are, I hate to use the word agentic because I think it's overloaded and it's a little bit of a marketing term, but systems that seek to be more of a co-pilot, seek to own more of the experimental loop that have integrations with the physical world that push us towards higher order questions like which hypothesis shall we pursue? What experiment shall we run? And what objective are we optimizing towards? It's a quick version of the story.

12:01It's a partial one. But as I mentioned, I think there's this impulse to tell this Whig history in which we say these early stages of computational drug discovery were groping towards deep learning. Factually, these methods solved real problems, and they built many of the tasks, data sets, and representations that modern AI inherits. Their limitation wasn't just inferior modeling, but I think it was something that's a little bit more conceptual, and I think it's also true today. It's that a lot of the most consequential scientific search problems sat outside of their frame. When we think about how these models work at Braidwell, when we are having these discussions, we talk about them as playing a role in their search towards better medicines.

12:44And as you open up this frame, you collect different kinds of data. You can tell different kinds of stories about that search problem and push towards better kinds of solutions. Maybe a way to frame it is I've been calling this thing the given trap. And I think it's an interesting concept. Much of drug discovery starts with a given. We say, given a target, find a binder, or given a molecule, predict affinity, given an assay, optimize this readout. But in a lot of ways, this given hides this decisive uncertainty, the target, the mechanism, the modality, the objective. In a lot of ways, and I think this was true from the early days, the danger is solving this conditional problem beautifully without ever interrogating the condition.

13:29No system, whether we're talking about the most contemporary AI systems that try to transform scientific discovery, we look further out, eliminates these givens. But in an interesting way, I think they just add them to the frame of things that they're able to optimize over. Maybe the point here is that each generation we've talked about moving from representing things manually to the machines learning the representations to being able to generate based on those things and then adding more of the physical world to the story just moves more choices from given to searchable. And I think in doing so just expands the amount of science that we can seriously intent.

14:04And that is the most exciting part of where we are today. But it's also, I'm sure, was exciting for everybody watching it happen along the way from the 60s forward. A badly chosen given can just destroy a decade of work. A well-chosen given, which in a lot of ways could be this tacit understanding of what works and what doesn't, is something that really can push you towards interesting new science and can unlock cures for a lot more people. Maybe a great case here we can talk about it is AlphaFold because I think that's a case where we have actually the success is relative to a really good and crisp given.

14:38We can go there if you want.

14:43If you're an HR or hiring manager in biotech, you know all too well that the pool of experts seeking full-time employment is shrinking. Filling key full-time positions can be a long, drawn-out ordeal. It can slow the pace of execution and growth. Throw away the old hiring playbook. Now you can build a biotech dream team in a fraction of that time. Find out how. Visit Chlora.com. Chlora, talent optimized.

15:15So we've talked about the lineage here, And I think the big sort of step change for a lot of people as you're looking at the field, and maybe it's the reason people think there's this rupture between the past and the present, was the announcement of the results from AlphaFold. DeepMind goes after this, of course. You know, they won the Nobel. It was the year before last. Dennis Sassabas and John Jumper are rewarded for this thing. It was undoubtedly a model breakthrough. But it was also simultaneously the harvest of this scientific institution that was built over generations of scientists. you'll hear a lot of our peers say maybe we're going in that we're going to go solve all of human disease in 10 years and i think a lot of that may come from looking at this case and bracketing out the given a little bit and saying well this is a general purpose because actually the alpha actually solves something very specific they solve protein structure prediction it was just extraordinary architecture engineering compute and scientific insight that just came together to solve this thing.

16:10But it was also decades of experimentally determined structures. So like the RCSB protein database was the transformative thing here. But alongside that sequence databases, shared representations and curation and deposition norms that were just true for structural biologists who had spent roughly a half century and order of magnitude, say$50 billion, building out hundreds of thousands of experimentally resolved structures. it was almost at the end that there's this great insight that you could approach developing these models on top of it. And in particular, that the AlphaFold approach was going to be a winner.

16:51Maybe one way to frame it is that AlphaFold was not just trained by DeepMind, but also by these generations of structural biologists going after this thing. In a lot of ways, it was like the perfect problem for them to go after. Luca Neff and Michael Bronstein, who are at Proxima, one of our portfolio companies, They've made this great argument recently that protein structure is actually this unusually coherent scientific object. It's comparatively stable. You can represent it really nicely. It's geometrically constrained. You have a lot of favorable things that are working. There are a lot of facts that are working in your favor that make it susceptible to modeling.

17:26So maybe look at it this way. You have a great unlock from a modeling perspective. You have 50 years of data collection, and then you have an intrinsic fact about the task that makes it tractable. So bringing those things together, this unlock is phenomenal. But it also puts you in a position to ask, how do you do the next one? And how do you iterate forward from here? Especially if our goal is to widen the aperture and to include more things in our experimental system. It's also true that biological data are hard. You look in particular at functional readouts. These things are dependent on donor and cell state, reagent lots.

18:05They have these massive batch effects. You look at this data and you have this morose feeling that you might have a model that's appearing to learn disease biology, but it's actually learning which laboratory or instrument produced the sample. The way out here is to expand the aperture and include more in the story to generate clean data, to understand the machines that produce the data, to continue making the picture bigger pushes you towards the ability to not solve the point problem, to solve what's the structure of this thing, which is critical, but also to solve the bigger question, which is given another structure, this thing, given that I found a molecule that binds it, what's this going to do in a human, which is fundamentally what we're trying to do.

18:48Maybe like one way that I've liked to think about it, and this has sort of been a North Star for me a little bit, is we talk about scientific bandwidth a lot. How do we increase the scientific bandwidth of the world? How do we increase a company scientific bandwidth? How do we increase the scientific bandwidth of a discovery program that we're running internally? One way we think about it is how do we think about how do we make every experiment teach us more? Well, we measure more and we build models that can combine all those things into a coherent picture. And we reframe it less as a point prediction and reflect it.

19:20And we frame it more as the search through hypothesis space towards something that matters.

19:25Nick Myerberg: And Nick, to go a little bit deeper in one area, which is looking at the past, right? So over 100 billion had gone into this space at the intersection of AI and biology over the last 15 years or so and didn't result in a single drug getting approved by FDA or EMA. And I know in your neuro IPS piece, you had mentioned that only now is the field developing benchmarks real enough to survive contact with real biology? Is that the postmortem? Is that we couldn't tell a good model from a bad one? And so capital flowed towards, you know, kind of whoever's demo was the most powerful? I'm not sure it's causal, but I do think it's diagnostic.

20:06In a lot of ways, in these previous generations, we were at a place where we had seen great success, as we've discussed, from the AlphaFold context. We'd seen early signs of life from a lot of these other approaches, but that ultimately these signs of life were largely related to the most atomic kinds of tasks that these models have to do. Maybe one caveat here as well is that, you know, as we're discussing, we've had a long history of computational success in drug discovery and that we have a name for CAD or computer-aided drug design scientists should tell us that we have found places where they are useful.

20:42I think this first generation of companies were folks who were using the best data they had, the best models that they had, reaching though for relatively atomic tasks like predicting binding affinity or trying to do phenotypic readouts that even if successful don't tell you whether a drug is going to work in a human or not. I'm sort of fond of saying that drugs aren't made out of binding affinity. And I think it's quite true here is that through that process, people developed co-folding models, for example, that were able to achieve pico-molar binding affinity on challenging targets, but ultimately didn't have the right frame to understand how that would push them towards a drug that's more likely to succeed in phase two and onward.

21:27For these companies that you mentioned, by the way, we did see enhanced success rates in phase one. I think that's worth a celebration, but that's the first step of a long story. On the investor side, I think the challenge is less that we were watching people chase bad benchmarks and more that understanding a company in this space takes a lot more than underwriting just the computation or underwriting just the biology or some third option. It's rather that you have to be able to answer a range of questions that are this inextricable combination of computation and biology. We spend a lot of time thinking about the right process and the right kind of team and the right kind of orientation towards these questions to push us towards answering these properly.

22:12I'm not sure that as an industry, we've landed on the right way to do it. I can tell you how we think about it. And I think that if we'd asked these same questions in the previous generation, it probably would have landed approximately the same position, which is that it really does take this rare constellation of capabilities to transform drug discovery meaningfully rather than adding AI to it. We ask five questions. First, we want to know if the model is genuinely world-class, not good for biology, but this is a group of people really doing fundamental AI research. Funnily enough, when I go to the AI conferences, when I go to NeurIPS or ICLEAR or ICML, some of the most forward-leaning fundamental AI research, especially in these things called diffusion models, which are also what are generating the images you ask Claude or Chet for, are coming from people who are focusing on biological questions because these diffusion models sit at the heart of structure prediction and co-folding models.

23:09The second question we ask is, does this team have strong therapeutic instinct is what we call, but it's really just biological judgment. Are they picking the right target, the right disease, the right modality? Are they picking those things, not just in a vacuum, but picking them in a context where they can say, this is the drug to go after because it biologically makes sense. And also it's tractable with the AI that we can build. The next two are sort of prosaic, but they're important. Like, Can the business pay for itself? And can it raise enough capital to outlast the science? Drug discovery is expensive itself.

23:42Add to that the fact that you need to have a team of AI researchers working on a big cluster of GPUs, generating a lot of data. It ends up being expensive. You want to make sure that they're in a position to pay for itself, certainly to pay for itself through time, but also to raise the money they need to get these step changes. And then finally, and I think this is the most important one and might be the one that's hardest to fake is, do they genuinely have a compoundable loop? that allows them to incrementally run more informative experiments and drive towards a more informative answer on a particular asset.

24:13If I look backwards, I look at the companies that have come and gone in the previous generations. I think they fall short on a couple of these items. They fall short, not in desire to change human lives and make better medicines. Without a doubt, these folks were committed and focused, super smart. The place where I see a lot of them falling short is really on, have you built a system that compounds through time? Have you found a way to generate differentiated proprietary data sets that a competitor can't scrape off the internet or get off the public web, but rather actively drive you towards a therapeutic answer?

24:51In a lot of ways, I don't think it was possible with some of these models previously. And I think a lot of ways, the companies that we're looking at today have made that a first-class element of how they approach these problems.

25:02Nick Myerberg: And Nick, now in many ways, we're still in the early innings of this next generation of AI and bio. What are some of the companies that have particularly excited you and the team at Braidwell where you've actually invested and how is their approach innovative relative to some of the companies in the first wave of AI that we talked about? We've encountered a lot of smart people in this space who hit the mark on those five criteria and do it in very different ways. And I can speak about a couple of them. The first I'd mentioned is a company called Proxima. I mentioned them in the context of the AlphaFold content, which is they're a company working on proximity modulation, working on glues, working on ProTax and RipTax and a wide range of proximity modulators.

25:46And what makes them special, makes them different, I think is in particular their ability to think really hard about how to generate data at scale that drives them towards good answers there. Glues are hard and in general have, up until quite recently, largely been found through serendipity. And that's in some ways surprising because much of biology doesn't act like an inhibitor. It acts like a glue. It acts like a proximity modulator. And so finding a company like Proxima that has brought together a truly exciting, differentiated AI model architecture with very differentiated ability to generate interesting data that drives that model forward and couples that with a lot of experience from the co-founders, from Zach Carpenter, Luke and Neff, actually working with pharma and understanding what matters to them to push a drug forward, puts them in a very interesting position to both do a lot of fundamental science, both in the AI and biology side, and turn that into things that partners actually want to buy because they find that the modality interesting.

26:51And also they see the glimmers of the possibility that perhaps this approach might unlock proximity modulation broadly. So find them. That's a terrific team, a terrific company. We're very excited to partner with them. We've several, perhaps about a year ago, we were really excited to co-lead Lila Sciences Series A. For those that don't know, Lila Sciences is a company with the ambition to build scientific superintelligence. broadly construed, both in the life sciences and material sciences. And they achieve that by taking very seriously this point about building a loop that we were talking about.

27:26So finding ways to develop models that can make predictions, can see what the models say, and then pretty rapidly and in a very cost-effective way, validate those predictions out in the real world in autonomously driven laboratories. Maybe to go back to your question about where things fell down in previous generations, a lot of it is how do you as quickly as possible make your predictions confront nature and confront reality? Lila has built that in as a first class citizen and lots of people are following suit because it really in a lot of ways is the way in which you want these models to work.

28:00You want to be able to survey the universe of possibilities. You want machines that can search. Computers are phenomenal at search. And then you want to bring them to bear as quickly as possible against the question you're actually trying to solve. So they're able to run discovery campaigns in this interplay between the biological world and the computer world. And the results that we've seen have been phenomenal. The last one I'll mention is a company that is focused on antibody design, which as people have seen, is a relatively popular space now for lots of reasons. It's popular because the datasets that exist are quite good.

28:39The publicly available datasets are quite good. It's popular because in a lot of ways, evolution has made protein-protein interfaces a little bit easier to solve than protein-small molecule interfaces. And finally, because over the last 50 years, biologics have played a massive role in our ability to attack human disease. is this company that I'm speaking of brings together a powerful model that is able to design and rank antibody candidates with a lab in the loop that is able to go and again, can make these designs confront the real world. It's in doing so in their ability to do that in a cost-effective and rapid way that they're able to push their model beyond, can I design a binder that has strong affinity to, can I design a binder that has a strong affinity and then also has all of these translational properties that matter when we talk about unlocking phase two and on for AI bio companies.

29:35All three of these companies are folks that have really brought together the phenomenally strong AI, differentiated data generation capabilities, and the therapeutic instinct that's required to understand how you turn that computational and lab operational expertise into something that can actually drive us towards better medicines for more people.

29:56Nick Myerberg: And so Lila builds robot factories to generate, my understanding, its own experimental data. Proxima built a technology to generate protein interaction data that doesn't exist, right? Is the actual thesis that the bottleneck at the intersection of AI and bio moved from, let's say, algorithms to proprietary data, and that as models continue to evolve at the pace that they're evolving, that models are going to become somewhat more commoditized while data isn't? Yeah, I think it's pleasantly true. The one refinement I'd make is there certainly is a bottleneck in terms of the ability to get the data that you need to solve these problems, the most relevant data.

30:37The bottlenecks also, how do you create that data and run it back into your model in such a manner that the next piece of data that you generate is more informative than the last? Having a takeoff in your model's ability to perform depends, of course, on human beings doing research on both ends of the equation, but also depends on your ability to turn those predictions into results that drive the model towards better answers across programs, across modalities, that there's transfer across everything. And so with that refinement, I totally agree. I think that we've gotten to a point where, by and large, the value that was latent in publicly available data sets has been exhausted, at least in certain domains, and that it didn't get us particularly far.

31:22Now, there are really smart people I respect enormously in the space who would argue that there's a lot more to do out of public data. But from what I see, I think our failures to go and solve some of these problems has been that in spite of our ability to continue building phenomenal models and exploiting this data, we haven't been able to actually unlock the next step. I think that model architecture, to your point, is likely already is and will continue to be more and more commoditized. Of course, the value isn't so much in the model architecture. If you look across the frontier labs in the LLM space, that models by and large look the same with important differences, but the structure is the same.

31:56The difference for our biology companies is going to be, can those weights and those models integrate enough knowledge about the world and can climb that ladder of biological complexity from does it bind to is it manufacturable to will it actually work in a human being in the lab such that they're able to continue pushing towards owning more and more of the discovery process in silico, making the process faster, making it more repeatable and making it less idiosyncratic.

32:24Nick Myerberg: And, you know, if you think about the arc at the intersection of AI and biotech over the next 10 years, and I'm picking intentionally somewhat of a short-term horizon, given the pace of model development, as we just spoke about right now, break it down for me. Like, what are you excited about in zero to three years, three to six years, six to nine years, As we think about, let's say, like AI native discovery platforms, there's a lot of AI enabled workflows within existing biotechs. And then lastly, is it infrastructure and tooling companies that support the ecosystem? How do you think about where we are right now?

32:59Nick Myerberg: What excites you now? And what excites you perhaps in year 10? I think the thing that I'm most excited about, well, we'll start here, which is I'm excited about truly AI design drugs hitting the clinic, being shown to be productive, useful medicines that can go into human arms and ultimately can cure disease. And I think we're at a place where for particular domains, let's say in particular in the antibody design space, where this is happening so quickly. And I think that proof point is going to be so important, both to founders and prospective founders who are going to go chase after the next category of problems, important to capital allocators.

33:36As folks think about how they want to run investment firms, they're going to have to take very seriously the fact that discovery is changing in front of them. And then finally, of course, as I mentioned, it's going to be exciting for patients who are going to benefit materially as these drugs get approved, because these companies are able to go after hard targets. They're able to do so in ways that reduce the time, reduce the risk, and there will be knock-on effects for those folks. When I think about what I'm excited about from a transformation perspective, pharma and biotech, is this pivot from point solutions to infrastructural changes to how discovery works.

34:18We have, through this entire process, and I think with notable exceptions from the companies that we've spoken about, we've been in this point solution era for a long time, where we ask these models to predict a property or rank molecules, generate a candidate, it, summarize literature, whatever the case may be, they're very task oriented. And the coordination of these tasks is still loose and doesn't necessarily have this end to end gloss to it. We've started to see evidence that these models are going to certainly be able to do those things, but also start influencing the kinds of questions that we ask, that drug hunters ask, the kinds of experiments that get selected, which, you know, how to synthesize evidence, how to remember failures, how to allocate scientific resources, I see a lot of possibility even over the next three years to start imagining what discovery operating system looks like for these organizations.

35:10Maybe one way to frame it is just relative to what electrification looked like. The full value of electricity appeared when factories got redesigned around distributed power, not when they first applied electricity to a motor and saw it spin. When I think about what the goal should be in short order, it's not giving every scientist a co-pilot. It's redesigning discovery around cheap cognition, rapid hypothesis generation, programmable experimentation, and thinking about what this challenging task looks like in a very different ecosystem. I want to see, and I think we're right on the corner, I think I want to see a world in which we're not talking about AI as something that accelerates existing scientific workflows or generates many more candidates, but pushes us towards changing that workflow, generating better candidates.

35:57They're more likely to survive. I think that's eminently possible in the next three years. I think you've started to see pharma take that seriously. The Noetic GSK deal from about six months ago, I think, was the first indication of this, sort of frontier lab style licensing, where GSK, from my perspective, I have no deep insight here, but it seems to me that GSK is thinking, we want to strike deals with a company like Noetic, where we can integrate this throughout our process and ensure that we're able to actually reaccommodate how we think about discovery and how we think about clinical development around what this capability is.

36:32That to me is deeply exciting. But I think in a lot of ways, it depends on something non-technical. Again, we've talked a lot about building a better model or generating more data or building a bigger loop. I think a lot of it has to do with how we think about integrating computationalists and biologists in under one roof as collaborators. I think it's a really hard problem to solve because these are two groups of people who have grown up in disciplines with very different norms and very different approaches, very different desires, kinds of problems they work on. It's very easy to, and we've seen this historically in these companies, see worlds in which you have a, you know, a computational subordinated to biological focus or vice versa, rather than these teams where you have these joint, you have folks who can hold both pictures in their mind at the same time.

37:20You have folks who can think deeply about what the actual operative scientific question is, what distinctions matter biologically, how to pull these things together and push them into something that matters for the result for a particular asset. I think that there's a little bit of a circularity to it in that we're going to need these folks to be front and center in leadership at pharma, front and center in leadership of these organizations who are building these new kinds of tools. And we'll need them also to be the users of these new kinds of tools. If we want to take seriously this possibility that we're going to go and reinvent how discovery happens.

37:47Nick Myerberg: And Nick, if we fast forward that, let's say 10 to 20 years again, what does a biotech look like then if let's say the first draft of science will happen within models rather than in plates. And then in addition, so, you know, we think about, let's say if discovery is 10 X cheaper and gets 10 times faster, the historical moat within biotech, right, which is, let's say, capital time scarcity and number of shots on goal required. As that starts to evolve and perhaps erode, what shape does the economic structure of the industry take from your perspective? Yeah, it's a phenomenal question. And it sort of leads me just to reflect on what it looks like inside the walls at Brayville right now.

38:31We're experimenting with trying to understand what these companies might look like and doing so through prosecuting our own discovery campaigns with AI partners and with our own internally developed AI tools. a because we think the time is right now that the thing you're asking about is actually true we think it's entirely possible to start prosecuting novel ideas in important therapeutic areas using these tools and so it you know as i walk around outside and i see art scientists i see alex or cio i see you know them just like their mind spinning a little bit on what is just different about the process versus what was true maybe three or four years ago, which is time to a plausible hit is much shorter.

39:16The ability or one's ability to de-risk it from a translational perspective has never been better. And your ability to essentially come to the conversation with a description of what you'd like in the world and test whether or not that's possible has never been faster and cheaper. And so we're running discovery programs that are taking a fraction of the time, costing us a fraction of the money, and ultimately we believe are leading towards better medicines. So to more directly answer your question, what's it look like when everyone can do this? Well, from one perspective, as we're able to unlock more biology, the market gets bigger.

39:51From our perspective, we think about the world of indications that matter. We want people working on all of them. We want biotech companies to find it easy to sort of spin up a zero FTE shot on goal to see whether they can attack a particular problem and see if they have some insight that is going to unlock the possibility for a new kind of medicine. The moat just ends up being, to your point, ends up being different. It ends up being certainly the same classic questions around exclusivity once you have a drug in the market. But when you get to early stage, a lot of it's going to be questions of understanding what kinds of medicines to go after, clearing space around what you choose, and being able to prosecute them incredibly quickly.

40:28So you can be able to market in a way that's more akin to building software or building consumer product than it is akin to having these deep theoretical and conceptual biological modes around the things that you're pursuing. And so net net, I think it's going to be phenomenal for the industry. It just is going to require, as we're talking about, some changes in how people orient themselves towards building programs, thinking about the role of computation and biology and thinking about how to push them forward quickly.

40:56Nick Myerberg: In all of this, you mentioned, I think, Lila, that the tech multiplies scientists rather than replaces them. Where does the human scientist end up in all of this? And what advice perhaps to folks that are listening right now, would you provide, you know, let's say that 26, 28 year old bent scientist in this new age of AI biotech that we're in? ultimately there's a human element to how you decide what to build what you build and how you build i think we increasingly see the models being phenomenal in cooperation with human beings that have steered these models towards the right kinds of questions when we're talking about the history i think one of the themes there is that through time even from the early days the ceiling of biological complexity that the models can accommodate just gets higher and higher and so i think that human beings who can understand where the models are performant where they're bad how to think as a biologist in a computer science context and vice versa are the people who end up being the most impactful in the space.

41:55Ultimately, and I think that's a bit of a hackneyed point because people have talked about it in the context of software and everything else that's being automated by AI. Taste matters. And it's a different kind of taste, I think, in drug discovery than it is in building products, building a web app or something. It's taste for the biology that you want to go after. I think one of the most important decisions these companies make, even in the AI space, before they think of building a model is what part of biology am I participating in? And then what kind of patient population should I go and serve?

42:27And these are the first one models will get quite good at, I'm sure, and they'll get increasingly better. And folks will operate a higher level of abstraction. But the choice of trying to figure out how to serve what kind of population, what kind of product most directly impacts the patient population you seek to help ends up being really a product decision that a human being is a part of. And we think that's going to continue being true. Co-creation, I think, is what the future looks like.

42:53Nick Myerberg: And to that point of taste matters, Nick, in the broader tech space, we've seen high growth startups like Scale focus on AI evals and reinforcement learning to power model development. How do you think about AI evals in biotech companies? Critically important. It is the idea of trying to improve a model without being able to measure it is, you know, it's almost nonsense, but building these evals is really hard. There are folks out in the world working on these kinds of problems who bring with them just massive histories of understanding what kinds of noise can be introduced in assays and then using that to build better data sets, build better evals.

43:35One I think of in particular is Leash Bio. I'm sorry, say founder, you know, it's this Ian Quigley and Andrew Blevins, the co-founders there who have spent massive amounts of energy thinking about how do you generate the kinds of data that you need to, of course, train models, but data that's out of distribution, hard and interesting to actually determine what these models are capable of doing in the small molecule case. You're going to have to see a lot of people at these companies develop this kind of expertise. It's so easy to convince yourself that these models are learning things when they're not.

44:07It's so easy to build an eval that doesn't actually relate to the ultimate success or failure of a medicine. And so it takes a lot of thinking. It takes a lot of effort. When we look at the best evals in our space, and we love to benchmark, it's one of the levers we have as we think about whether these companies are likely to deliver a model that works or doesn't. You just have to be incredibly selective about how you interpret the outputs. It'd be so easy to say, well, this model A is the most capable on this very constrained question, so therefore it's a good model. When in fact, the thing we're optimizing for is miles away from does it have particularly tight binding affinity.

44:46When I talk to folks, founders in our space, when I talk to our portfolio companies, they all take very seriously this question of generating the right data to evaluate whether model is any good. The next step past that is how do we figure out the right organizations out in the world who can generate high quality evals for the broader world, who can generate evals that are available to academics, who can develop evals that are available to brand new companies. So they can actually benchmark in a way that matters and benchmark against these more translational import kinds of questions. Yeah, great.

45:19Nick Myerberg: Well, Nick, thank you for educating us on this new frontier that we're all operating in. Two questions before we let you run. First is, what's one piece of advice you wish you could provide your younger self, knowing what you now know? It's a great question. I think I'd encourage my younger self to reflect on the fact that being a generalist is a phenomenal thing. Learning how to learn is one of the most important things you can do. And that means not just focusing in on biology or computer science, or if the case might be, but reading broadly. because I find myself in places where the decisions I make often are driven by history or driven by a conceptual framing that came from reading some philosophy.

46:01It matters, especially today. It matters to have the broadest possible diet of ideas if you want to try to dive into these important and interesting questions.

46:10Nick Myerberg: Wonderful advice. And what's a book that you've read that had a meaningful impact on you can be professional, personal, et cetera, that you'd recommend? Yeah, there's a phenomenal book by a historian named Andrew Pickering called The Cybernetic Brain. And it's a retelling of the history of cybernetics and how that led to artificial intelligence that departs a little bit from the traditional story of, well, they had to go and build machines for war. But rather it takes you to understand how the early thinkers in the space were trying to actually really, they were trying to solve scientific problems.

46:45And so I appreciate hearing about the long history of people trying to solve the hardest scientific problems using computation, even when the computers weren't that good.

46:55Nick Myerberg: Wonderful. Well, Nick, thank you again for joining us today. It was a pleasure having you on. Look forward to continuing to follow all the exciting work that you and the rest of the Braidwell team are pursuing. My pleasure. Thanks so much for having me.

47:11Nick Myerberg: Thank you for listening to this episode of Biotech 2050. This episode is hosted by me, Rahul Chaturvedi, and Alok Thayi. If you enjoyed this episode of Biotech 2050, please subscribe to our podcast and leave us a review. Also follow us on Twitter and Instagram at biotech2050pod. Again, that's biotech2050pod. Until next time.