I really appreciate it. Good morning, everyone. My name is Akash Jawari. I head our farm and biotech research efforts here at Jefferies. This is our New York City Healthcare Conference, which I always enjoy, and it's great to see the attendance being as strong as ever this year. The topic of today's discussion is something which I feel like we have heard way too much. It gets brought up in graduation speeches and it gets loudly booed. But I think it's a topic where in health care, it's funny. No one really has an idea of how AI is going to impact drug development, how it's going to impact the health care industry, despite health care being the biggest employer, one of the biggest sources of revenue and employment in the United States. And there always seems to be this kind of disconnect when you hear, let's say, people who are just heavily in the tech industry who are convinced that AI will transform drug development. And then you'll hear, you know, people that are more on the traditional side express skepticism. And the three panelists we have here, I think, are really on the cutting edge, who probably more than as much as anyone else in the world understands where AI is helping in terms of how they run their companies, how they design their businesses, and how they actually develop better drugs to patients, which is ultimately the most important thing. So joining us today from Alto is Amit Etkin, founder and CEO, from Iambic, Tom Miller, and then from Idea, Yudiro Hata. Thrilled to have the three of you on. I think maybe just to kind of start off, because, you know, it's funny, we've all chatted before this panel, too. And I can tell there is levels of exuberance when it comes to how AI will impact health care. And I think all three of you kind of strike a very pragmatic tone, which is optimistic, but also very pragmatic. So let's just start when you hear, you know, Dario or Jensen Hong talk about how all diseases are going to get cured. And, you know, a lot of and biology becoming a computational problem. Right. And I think that's kind of the fundamental idea. You know, how skeptical or optimistic are you with that type of framework? that curing diseases is an issue of computation and that the way we think about biology is really going to fundamentally transform as AI gets introduced? Any of you can start this one.

I would answer yes. There's a lot of questions in there. Yes to everything. But also, you know, I think, look, we have to realize that we have new tools and with that comes new capabilities but also the same problems persist the reason we're here and we haven't solved all diseases yet is because it's hard you know from within the industry I find it useful to break down and be curious what my colleagues here say the different silos in which AI is useful it's useful in many ways but in different ways and in different sources of benefits so we don't do as much on the discovery side that's certainly where a lot of computation could help we're a clinical stage biotech company we're in psychiatry and so inherently we have some challenges some are shared with all clinical stage biotech companies some are more specific for psychiatry shared with all clinical stage biotechs is that it's really expensive to run trials and you can't run that many trials which means if you have more molecules you actually have not necessarily solved your core problem which is you still have to run the trials and you still don't know necessarily if the thing is going to work because the predictive value of something that's early for something that's late is still pretty bad that takes iteration to improve not just computation but the approach that we've taken is one of a precision psychiatry lens. So we take a broad diagnosis like depression, we collect a variety of biomarker data, and then we use what used to be called AI, that I think probably more aptly would be called machine learning, to find biomarkers that segment the population that we could replicate that we had not necessarily anticipated. Incredibly useful, incredibly illustrative of what can be done in that space. But if I had 10 times as many molecules i would have a huge headache because i wouldn't necessarily know how to prosecute those in a different way it also helps tremendously with how we run our trials we're doing a ton and we can talk about this more as we go but we can we're doing a ton with llms in the context of running trials i actually like throw a hot take out there i think we are at the point now where a person with the right experience could probably run using consultants and contract organizations and so forth can probably run an entire drug development program as a single individual all that's great but the problems are harder and the tools will never be able and I don't think there's any tool in the history of our field that's been able to truly bend the curve when it's come to the difficulties of simply not knowing enough about human biology and human disease I'll think about on that.

It's a great question. Iambic works extensively with NVIDIA in terms of the development and scaling of our models. So we're familiar with Jensen's framing of the field. A lot of that relates to other aspects of ways in which AI can impact drug discovery as well. You mentioned the impact that AI can have in terms of clinical design and patient selection. There's the challenge challenge of actually making the right drug for a target of interest, which can be an incredibly big challenge. There's the challenge of identifying which targets to go after. So you've got targets, you've got to make the drug, and then you have to develop clinically, and all of those present different challenges. Iambic works in the category of making best-in-class and first-in-class drugs, novel chemical matter, for really challenging and important targets around which we have a lot of conviction. And we use the combination of large proprietary transformer-based models to holistically design molecules for success to clinic and through clinic, and the execution of that in combination with high-throughput experimentation. So agent-driven experimentation informed by proprietary models, and we've shown that that can execute the clinic repeatedly and successfully at a pace that is roughly three times faster than the industry wants. Understood.

Yeah, so I think the short answer is no. I think very unlikely, and I'm not sure that's the objective either to cure all diseases with AI. I think from our perspective, it's really about how do you gain efficiency in what we do and you know for us at idea that's it's really around direct discovery now can we accelerate the time from an actual hit ligand to a molecule in the clinic by X percent and you know we think that's where really you're gonna see the gains ultimately in terms of curing all diseases I think that's a very complex question I think ultimately and you know in particular you have with disease or we have adaptation so you know that target is continually moving, the therapies you're creating is always going to be trailing adaptation. So I think that's a wonderful kind of big-term, longer-term objective, but I think highly unlikely.

Understood. And no, I think one of the points that we've certainly seen in our conversation, it's funny, I was, you know, hosting Apogee two nights ago for dinner, and there was a discussion of like, hey, the FDA has said that you may only need to run one clinical trial, not two clinical trials. And there is also this idea that maybe you don't have to run, you know, a phase three trial in the future. And it's funny when you talk to the companies, especially whether it's on the Smith cap or even the pharma side, they're like, well, we still want to run our trial. We want to prove that we have a best in class drug. And so actually the upfront cost of running a clinical trial is de minimis versus the decision that would change whether you have a best in class molecule or not. And ultimately clinical data is going to be what's going to prove it. So, you know, I think what the three of you have said is spot on in the sense that ultimately there's an efficiency advantage. But the real question is, can we actually deliver kind of better molecules? And I know all three of you are working on that. I wanted to start on the neuropsych side, because I think it's actually incredibly important. You know, drug overdoses, suicides are the reason why, you know, the U.S. lifespan lags behind other developed countries. If you remove those. We're in line with other European nations, but we dismiss that. I mean, that's a tragedy. And you think about how neuropsychiatry is treated and patient selection, there's comorbidities, there's real human problems, right? So when you think about how AI can actually help identify patients who would benefit from your medications, and how do you use that to actually implement a strategy to get these drugs ultimately to the, you know, depressed or schizophrenic individuals who suffer from these disorders, how is AI going to help with that?

So we have a unique problem in psychiatry, as you're highlighting, which is that we have phenotypically characterized our patients. Let's start with depression. Depression may be, you know, One in five, one in ten, depending on how you want to sample the population has had depression. It's a set of symptoms that tell you about something quite broad. And within that, you know, there's things like eating too much, eating too little, sleeping too much, sleeping too little, which seemed like specificity, but I can tell you that any individual item on any depression scale correlates better with things that seemingly have nothing to do with that function, if you think of it as a measure of physiology. And more to do with just the general mood. I feel bad, therefore, X, Y, Z. So we've had this impenetrable structure that has led to a hard time modeling in animals, a hard time discovering drugs, a hard time knowing how to develop it, really high failure rates, non-predictive early trials. trials, and even when you get to phase three, it's sort of a coin toss in the historical way we've done trials. So that's a fundamental issue of knowledge. There is almost no way without tools like AI, and I'm painting AI very broadly here, that the problem will just solve itself. We need something to be able to look at data and reveal patterns that we cannot see. data that are too complex, ways to describe data that have layers of complexity that we couldn't even imagine. For example, we collect biomarker data, so things like EEG, brainwave recordings, cognitive tests, and the like. And EEG data is very rich, and there's been various ways to characterize it, but there's various ways that we haven't even thought of to characterize it. And across all of those ways to characterize it, we don't know which signal is the most important for any one patient, any one subgroup, any one drug. That's really where discovery out of the data, machine learning, AI, and different ways to formulate what that construct means can allow us to discover patterns that we'd not anticipated. But an example for a drug we're developing right now turned out to actually be pretty cool in linking back to mechanism. Let me give you an example. Alta 300 is a drug, an antidepressant actually approved in Europe and Australia as an antidepressant in an all-comer population, has a different mechanism of action. We're developing it here as an adjunctive treatment in depression. And we're trying to find what that biomarker-defined responsive group is. So we use machine learning, defined a pattern out of the EEG, prospectively replicated it in another sample, and are now developing the drug using that as a patient selection approach. There's no differences in symptoms between people with and without the biomarker, but then when you look at the pattern, what the EEG tells us from that biomarker, it turns out it's simple enough for us to understand and bring back to the lab, bring back to experimental manipulations in humans and in animals to show that actually, as it turns out, what we discovered was a biomarker that reflects essentially the opposite of what the drug does to the brain. So if you perturb the brain pharmacologically opposite of the drug, you induce the biomarker. So we actually discovered something and learned something and actually maybe even learned a strategy for the next thing in an entirely agnostic way that's open to discovery, which would not have happened were it not for an AI slash machine learning approach. and that's the the sort of this getting through an impenetrable wall of our own definition of disease will need that kind of third party if you will perspective that we otherwise don't have understood and actually uh i wanted to get your perspective from tom and uger um obviously both of your uh companies are working primarily on the treatment of cancer um but it is interesting we

don't you know I think also is kind of a rare example of a you know AI pro AI approach when it comes to treating your psychiatry disorders I'm curious why haven't we seen maybe the same advent of AI drug development in some of these neuropsychiatry disorders versus you know oncology is there is there a biological issue there or is it more about target selection and there just doesn't seem to be a lot of great targets right now in your psych There actually I think are targets, there's actually plenty of targets.

If you look at the landscape actually it does seem like we keep reinventing the same, more molecules against the same targets because they keep coming up in animal experiments and in post-mortem and genetic studies and so forth. What I think what we're waiting for, what we're hoping to provide at ALTO is kind of a turning point, a change in imagination for what's possible, because the field is very much operated kind of like oncology did 15 to 20, well, probably more like 20, 25 years ago, prior to the IO revolution, which is needing convincing that something that seems obvious, that you should use biology to segment your populations, actually works. If you remember, oncology at the time actually did not, had not proved that, even though there was Herceptin, you know, Gifitinib and so forth. It didn't really spark a revolution in terms of a precision approach until IO came. That's kind of, we're waiting for our IO moment. And then I think people will flip. It's not because – it's sort of like a traditional way of thinking that until you have something that's proven that's better, it's at least the thing that you know.

Yeah. I can actually offer a pretty pragmatic defense of oncology as a choice for AI companies. You know, I think everybody is excited to see where are the AI discovered drugs. And one of the reasons that Iambic focused on oncology as a therapeutic area initially is because, yes, there's tremendous unmet need, and yes, there's established targets, but, you know, you can do the phase one in patients with disease. So Iambic got our first drug to clinic in less than two years. And within 18 months after getting into clinic, we were able to be at ESMO showing differentiated clinical activity and safety of our drug. But not just safety. The point is that we can also show that clinical activity at that early proof point. And yes, our early assets are focused on oncology, but through the expansion of our pipeline and through our partnerships, we're working in neuroscience, we're working in GI, we're working in immunology, we're working in metabolic disease. So we're working in a number of different therapeutic areas, expanding from the initial proof point.

I mean, the other observation I'll make from that is just hard endpoints sometimes, you know, like a mad risk score. I've been staring at that and made some bad calls in my career because we overinterpret a phase two, you know, separation and things kind of revert to the mean. So I think that that is probably also playing a role here. You know, another kind of question I'd love to ask the teams, when you think about running an AI drug discovery business model, right? And we've seen companies like Schrodinger, Recursion, and these are really pioneers in the space, and I think they deserve a ton of credit. it it's not been traditionally a very successful business model and of itself and ultimately it seems like the value is is making these drugs but i'm going to give you kind of a counter consensus question why is that why should we give up on that model and that concept because the way i think about it i mean we have sam altman saying hey you know if you if we make a drug for you we'll get five percent of of an ownership stake and in the back of my mind i'm like i don't know why i would trust them to develop a drug versus let's say the three of you where you know you understand the biology you understand regulatory interactions you have proprietary data sets academic partnerships right so if the power is ultimately the data and the quality of it i mean it lies really with the people that can melt both together so the idea that you could you know create a multi-billion dollar company by just going to pharma business and saying look we will create you great medicines you give us timelines and we can get it done do you think that is actually a sustainable business model in the future i could we think a lot about that um you frame it um in a really nice way um iambic does a lot of we have our internally owned pipeline um as well as uh extensive partnership activity we've announced partnerships even in just the last

year, a $1.7 billion deal with Takeda, large neuroscience deal with Lundback, a large deal with Revolution Medicines and oncology. And you're absolutely right that people can most easily assign value to drugs as opposed to assigning value to software, right? So structuring deals around the endpoint delivery of drugs and having a track record of actually creating best-in-class and first-class drugs is really important for those. But those partnerships do offer a way to have a promising business model. You can use those partnerships to expand your data mode. You can use those partnerships to have significant income and use that to drive your internal pipeline. Over 50% of iambics clinical costs are defrayed through our partnership income. So we're running it, you know, half the cost for a clinical problem.

Understood. I'm curious.

Yeah, look, you know, we're I would say we're probably a bit of a later stage company. You know, you have nine programs in the clinic. Our first drug, we're getting hopefully ready to launch commercially here fairly soon. It will be a revenue generating company with a sales force on the ground. So, I mean, I think the way we think about value creation as it relates to AI, ML, drug discovery, which is our entire focus, and where we think ultimately for the industry we're perhaps the most important area for AI, we think a good portion will be in the drug discovery area in particular. Our view is the value is going to get created between that intersection between AI drug discovery and people that can really bring the novel cancer biology to the table. And, you know, we've seen a lot of AI drug discovery in the last couple of years that are, you know, largely sort of me too compounds that have been trailing. Right. We can give hundreds and hundreds of examples of that. And I think when we had caught up over the phone before this panel, I gave you the example, and I won't specify who this group is, but there was a publication that came out using AI against one of the novel targets that we've been one of the pioneers on with the structure that we generated through patents that were published. And it was bound to the pocket, but it was inverted, right? And so that's, I think, a good example, you know, where you have to just be really thoughtful and careful of kind of how you're applying it. And ultimately, that's sort of how at least our view of the most optimal business model of value creation is really to combine novel first-in-class targets. We have, you know, wet labs that were doing the basic research, really explore the fundamental biology of these pathways, and then go after targets that no one's ever drugged before. You know, whether that's in the area of transcription factors, HILA cases, other very difficult target classes that through normal means, you know, as we mentioned. In terms of typical calculations, if you did not use AI, it could take you over a decade to figure out, you know, can we accelerate that computationally to within weeks or months?

You know, that's that's so important. I'd love if you could give even maybe just like a brief case study because you'll hear this very often. like, well, AI can make a known target more selective and, you know, make a better hammer to hit a known nail. But I've sensed from, you know, both talking to you and Tom, that's not been your experience at all. In fact, it's doing what is practically impossible already. So I'd love for both of you, illustrate maybe, and, you know, in a way that doesn't ruin your patents, how AI has actually done stuff that a traditional drug development platform simply couldn't provide.

Yeah, I mean, I could give a specific example. You know, we have a program in the clinic now that, frankly, you know, it would have probably had taken us a couple more years to get a molecule into the clinic. It was extraordinarily valuable. AI really accelerated that time to DC nomination. And sure enough, you know, we're now in the clinic with a major pharma company, and it's just us and them side by side. We're sure many other companies have been trying to solve this. And that's where I think early adoption is going to be critical. However, as we had talked about earlier, it's very clear that, at least in our experience, AI is actually not very good at optimizing certain parameters. And so it's really that knowledge and know-how that you have to have, because if you sort of apply it broadly, it's actually not very effective. In many ways, it's going to take you down the wrong path, and you could end up losing time. So I think it's really going to be about how you, you know, getting into the depth of understanding that question, which is where is it effective? Where is it not effective? And that answer actually changes depending on the target classes that you're working on. And that's where I think for companies that adopt early, build those internal proprietary databases early. Things like your machine learning process will get more sophisticated over time. And that will be ultimately, I think, the company's competitive advantage. So the last part I will mention is lead optimization in general. I think with AI powering, it's pretty straightforward. I don't think that's really the issue, at least for us right now. You know, we look at it as where is the biggest mathematical problem or challenge? And at least we think that biggest bottleneck today, especially with novel first-in-class very difficult to drug targets, is actually the hit finding. And it's just a pure mathematical problem, right? Which is, it's the, you know, that first moment where you're trying to find a ligand to a target, that computational challenge is the greatest. That's the example I mentioned. You could wait 10 years to really find a real viable hit. So there's actually more even back end, you know, software, whatever you want to call it, that we think needs to be created. even with, you know, Schrodinger or other platforms, we actually don't think have the capacity to do this in an efficient fashion. That's what we've been working on. We've been doing it sort of quietly behind the scenes. We've been forming academic partnerships to enable us to do this. But we realize that we have to actually build this ourselves because it does not exist currently.

And if we can untap that, then we would change this whole area of hit finding um in an extraordinary way and you know that that would become and we don't we wouldn't plan to partner that for us we would essentially fuel that into our internal pipeline and that would be the advantage ultimately and hopefully the value to our shareholders understood you know tom i remember the conversation uh we had recently i think one of the unique perspectives you have is you know again like if you think about ai as a machine you put an input and then you put an output and it's this whole idea of can you see the dynamic system in play and can we get to a point with ai where you can back solve from the output you know um how to discover novel targets right and and that's that's really maybe a generative ai approach right can you talk about um you know how iamic is really working on not just doing lead optimization but also working on discovering truly novel systems that we probably couldn't find just looking at a proprietary data set.

Yeah, I'm happy to do that. I think that it's tempting to look at drug discovery and to say, you know, this stage is hard or that stage is hard or that stage is hard. But the reality is there are probabilities of failure all the way from hit finding through lead op, through preclinical preparation, through phase one, you know, each one of those presents different associated challenges. And what makes drug discovery generally so hard is you have to get married to the molecule you're going to be carrying forward really early in the program, and then you have to kind of put all those resources into it once you've hitched your wagon to it to ultimately try to get that into clinic. Iambic has very purposefully built multimodal transformer models that allow us to predict both on the earliest aspects of drug discovery, like kit finding, as well as selectivity, as well as in vivo PK, as well as manufacturability, as well as human PK properties and talks. So we are able to design from the earliest stage with enriched probability of navigating those subsequent bottlenecks and we do that in a way that that allows us to get rapid feedback with experimental data to then drive forward those optimization cycles so the you know you ask about ways in which that's really been proven to work and in ways in which that leads to things that people simply haven't been able to find before. You know, our first program, we have more novel targets, of course, but our first program is HER2. You couldn't think of a more established target. People have been working on that for 25 years, making TKIs that are all dose-limited by liver toxin, by rash, and that aren't brain penetrant. That is the state of the field. And Iambic found a novel way using our platform to bind the HER2 enzyme. We have the first ever HER2 TKI that binds in the inactive confirmation that has led to thousand-fold selectivity versus off targets that leads to avoidance, as we've shown now in clinic, of the dose-limiting rash and liver tox and leads to extremely high brain penetrance and broad activity. So through, you know, this sort of approach, you can design for success not only into clinic, but through clinic as well.

Understood. I want to hit on an angle that I think doesn't come up enough, but I always get fascinated with IP. And protecting your technology is incredibly important, especially as you can look at a patent and reverse engineer compounds relatively easily. But one of the things I've noticed talking to companies that adopt AI-centric approaches is they'll jokingly tell me, like, I can tell you the target. You still wouldn't be able to do what we're doing because it would take you a year and a half to develop the preclinical models and even then getting into clinic so there's a real advantage to using ai when it comes to making sure that whether it's in china but even in the united states where look there there are me too drugs everywhere how does ai protect or at least preserve that kind of two three year advantage when you discover a novel target or you know a novel biomarker uh to prevent companies from quickly adopting your technology so so i actually don't think it's the ai part itself it's the data yeah yeah and probably we all have data that we would hold particularly dear for our particular uses that when you then

apply ai and so forth allow you insights that data usually is very hard to collect so in our world collecting data on people just there's never enough of it but especially longitudinal data with biomarkers in psychiatry turns out to be one of the things that is the most lacking that is the first area that we invested in at the beginning of Alto is collecting that data even on standard of care treatments to just understand what happens to people and their brains over time and that kind of data you never put in a patent right you're putting a final end use that's narrow with respect to how you're going to take the product forward but the data that's generative for more ideas that's really where the value is and then as computation and tools improve you can take new looks at the data and discover new things but that part i think ultimately you know i mean there's this sort this data is you know the new oil or whatever uh the people have said over the years understanding what data you need to collect kind of looking around the corner for what data you might need given what you've just done or what you think you could do that is really where the game is at uh in my mind understood i'm curious if you guys have a perspective maybe on the again because that's you know what we've seen with a lot of targeted oncology drugs is you know you can have competition very quickly so maintaining that kind of first mover advantage

is critical here how is using ai just drug discovery made it harder for that to actually occur but do you think it hasn't yeah i can i you know i think we see um you know competition overseas we see ai in general is just the rising tide of competitiveness in terms of drug discovery that everybody has to compete in terms of. And then, you know, you can use that as just the baseline standard for what is good enough. And, you know, finding stuff that are sufficiently novel and that are sufficiently protectable is still going to be the game. And you can use AI to slice and dice chemical space and to do that as thoughtfully as possible. I really like the point about the fact that, you know, certain data goes into a patent, but the expanding data mode of an organization is a sustaining thing. And with multimodal AI models that can learn on one kind of data and get better on a different kind of predictive endpoint, you get complementarity across different programs through doing that. So there's a flywheel benefit of doing lots of programs, acquiring that data mode, and ultimately having predictive capabilities that are differentiated.

I think this is an important topic. I'm actually more worried about the negatives related to AI and intellectual property. And we've seen several situations where patents have surfaced fairly quickly, likely AI driven. And if you look at a lot of those chemical structures, you know, 99 percent of them are essentially junk, but they're, you know, canvassing the universe. Yeah. So I think it does bring up a key issue for the sector, because I don't think that's actually good for the industry. Right. Because all of a sudden it's flooded with all this IP that you have to navigate, you know, these fast followers. So I do think what's going to become even more central for all of the companies on this panel, as well as big pharma and other companies, is, you know, what is your timing of disclosure around targets and initial chemistry? I think you just have to be a lot more mindful, a lot more cautious. We all want to publish in Nature and Cell and, you know, American Chemical Society. But this is very real, right? And I think AI is going to likely, you know, just magnify this even more over the next couple of years. I think this disclosure question is going to be critical.

I couldn't agree more. And not an AI question, but I'm curious. And this kind of comes in the context of the House Appropriations Committee and what's going on with China. The sense I get talking to some of our pharma companies is, look, we probably want to continue to partner with China. It's incredibly important for drug development. I think that's fair. But there are things that I think to preserve innovation in the United States. And one of those topics is the timing of patent data disclosure. So not an AI question, but I'm curious, is your view that maybe in the next few years, we could actually see the USPTO adopt, like, we'll grant you a patent, but you don't necessarily have to disclose your chemical structure for a couple years? Is that a possibility here?

It's hard to tell the future on that. One thing I want to say is that, you know, there's a lot of kind of versions of the same question in discovery and drug development. You're often dealing with incomplete noisy data, either in finding your drug or in making a decision about what to patent or in making a decision about how to, you know, to develop from a regulatory perspective. And one of the benefits of AI, it is simply the right framework given a situation to optimize for success. So I do think that it's, you know, it can't change the uncertainties. It can't change the way in which the rules change, you know, due to external regulation factors or competitive factors. But it can give you, subject to the known uncertainties, the best strategy to optimize your success. And I think we see that theme both on the development side and in the discovery side throughout.

Yeah, maybe just a directly answered question there. I mean, I think on the PTO side, I mean, there probably is things that could be done to give that first mover, you know, the true innovator that came in. Because as you know, even around initial patent filings, you usually have a certain period of time to add to that patent filing. That's another area perhaps, you know, that could be explored. Could you extend that to give a little more window of time to bolster that IP before there's sort of a floodgate that occurs? Second, I would say, is also just our peer-reviewed settings where we publish. A lot of these major journals or even medical conferences, you can't even submit if you don't show the structure. Are those things that should be reconsidered? Right. And, you know, we're all for obviously full disclosure of information. But if it is becoming damaging, you know, companies like us are put in a hard situation because, you know, yeah, of course we want to present, you know, the triple meeting or whatever ACR. But those requirements are significant, especially around disclosure, around chemical structures. And that's probably something worth considering. Obviously, we're not going to resolve any of that on this panel, but I think to this point, because I think it's a very real risk for sponsors and innovators.

Yeah, so I'm not going to necessarily touch that topic on how to deal with competition from other places with other models. But related to that is the question of whether you can even get a patent. This whole question of non-obviousness for patent, right, it's always been limited by the degree to which, you know, a human being could basically guess the answer, right? But if you systematize this information, because, you know, there is the innovator, right, that creates some new molecule for some new target. But so much of biopharma has been the follower drugs that might be better and actually may even capture the market more effectively because they have fewer side effects, better on-target activity, what have you. All of that depends on there being patents that can be gotten. And at what point, you know, I don't know that we have as a society and legal system really wrestled with this question of when is there non-obviousness and when do you lose that because of AI? How much do you actually have to show to lose that? That's a really good point.

I mean, the entire concept is opposed to ordinary skilled interest in the art. It's a very different person now. It's a really interesting point. So maybe thinking about how a lot of our discussion has been about ai for drug discovery i definitely want to circle back to it but there's also this idea it's it's interesting you know albert borla um you know had our um fireside chat yesterday and he made a really interesting point he's like look everyone wants to look for the best drug when it gets in ai and he's like look when i think about pfizer and he's like i look at large cap pharma there's people who are ai skeptics and there are people who are all in he's like i'm all in and his point is i need to go to the people who've been doing this for 20 years on the finance team and say look you've done a great job but you have to start adopting ai because you know we need more productivity out of you and he has to convince thousands of people in a global organization to change how they're doing their day-to-day tasks i can tell you i don't use a lot of ai i'm probably the annoying person in jeffries who doesn't you know adopt enough but you know his whole point was the edge is going to be how much you are able to change an organization from the ground up. But you guys have all a very unique opportunity, which is, you know, you're starting these companies and you, I mean, in your case, you've developed multiple drugs, but the point is like, you could be truly AI native, right? And what does that mean for a SmidCat biotech to be AI native? Forget just drug discovery, but when it thinks, when you think about operations team, how you run your organization, how you seek external capital, what is AI native mean versus a traditional SMITCAT biotech drug development approach?

Yeah, so I'll take a development lens because again, that's where we focus is really phase one and beyond in humans. And go back to a comment that I'd made before, which is that I actually think that the era of a single person biotech is upon us if you have the right experience that you can do a ton by yourself that creates far more nimble probably more cost-effective organizations than what we've had before you you you'll have specialists that you pull in consultants contract organizations those exist anyway whether you're staffed up internally or not it's a really exciting future actually Because now you liberate people who are creative, who are skilled, who are motivated, and you give them superpowers because now they're able to execute on things at a much faster clip, doing things like writing a protocol. How complicated is it for AI to write a protocol? Not really. They take the same form no matter what you're doing. and you as the person with the knowledge get to imbue the content and what you want without having to spend a ton of time and a ton of money on the generation process and that goes for anything, regulatory and overseeing trials and any number of areas that's really exciting actually because now you liberate ideas in a way it does create challenges in terms of our capital allocation but maybe that's sort of like more individual minds get unlocked and then we have to think about how we stage funding for more ideas, more innovation, but a more competitive process and then getting the bigger rounds of funding for ultimate late-stage trials.

I think the question also connects on how to kind of win the hearts and minds of the people who are doing that work within an existing organization like Miser or how to build an organization that is AI native. And I honestly think, and you see that through enterprise solutions, like what that conversation relates to, you see it through drug discovery, AI solutions as well. It has to be a pull mechanism. You can't just tell people, you know, okay, now you're going to have to do it with AI, right? It's like, you know, now you have to do your normal job, but now you've got to do this weird way that makes everything less convenient, No one's ever going to want to do that, and they're going to revert to the norm, and you can't build a company around that. You have to build it around the fact that you know you're doing this job and you want to do it as well as possible, and these are your pain points. And this is a tool that you can use that answers your questions and makes those pain points a lot easier, whatever it be. So I think really just focusing on helping industry experts become superpowered by removing their pain points is what we should broadly achieve to do to get adoption.

Yeah, look, I think here with AI and sort of organizational adoption, you know, who are the early movers or not, I think, frankly, at this point, if you're not an early adopter or integrating AI through your organization, you're going to be at a significant disadvantage. And I think it's really now to the point where it's become table stakes. And if you're not utilizing it for chemistry, drug discovery, frankly, we are. And it doesn't matter how many resources you have. If you're a farmer, we're probably going to beat you to that end. And, you know, now we're using it real time and we're seeing it now. So, as you know, we just got our tour for our lead program and we just filed pre-submission one. And I even know from a regulatory perspective, we're starting to implement AI for the first time for our regulatory documents, including the NDAs. That's big. First time as a company we're doing that. In fact, that was pitched by our regulatory team. And when you look at just the level of efficiency we would gain in terms of time, capital, dollars versus our deploying actual people, because, you know, as was noted earlier, a lot of this is fairly standardized, right? And so then you just sort of have someone that's essentially just optimizing the final 20% versus doing the full 100. And the amount of efficiency gains is going to be quite extraordinary. And actually, when I saw this side by side, it's an obvious decision, right? And so I think you're going to see this also, I think, in our sector in the regulatory front. And that's actually great for small companies. So it just allows us to be more efficient, use capital more efficiently. And then we can use those dollars on more high-value areas that can perhaps start new programs, right, invest in other clinical trials.

Right. That makes sense. Now, you know, one of the topics I think we've heard, and we'll go back to the drug discovery side. I remember in London we had, you know, Tim from Argenix, Ugar, and then John from who was the founder of, one of the founders of L-Nylam. And they echoed, I think, a sentiment that's been said often by pharma, which is this is a data problem and a compute problem, but it's actually a data problem first, which is the information. You know, we don't understand enough about human biology or there's this national self-selection of, well, I'll present the data that's good, but I'm not going to present the failures. And, you know, I know Dave said, like, we think maybe the NIH should be really working on thinking about a consolidated data bank so that we can really, you know, solve this data and quality of data problem in a holistic way. And it hasn't been done. I'm sure you guys are very aware of that sentiment. How true has that been? You know, where is that sentiment correct?

And where is that sentiment actually not correct? where you actually have enough data right now to really optimize drugs quite well versus other indications or other disease states where it's it's it's lacking still maybe tom yeah chemical and biological space is just extraordinarily vast so you might have a great data set but nobody feels like they've got it all covered and they've got all the data they want right i think one of the catchphrases that you see across the entire ai industry now is something called super intelligence which sounds very grand but basically it's sort of the recognition that the big models be they the big models are basically trained on all the publicly available data at this point right so being able to have a factory of additional data creation that can feed those models is an important part of actually building the next generation of models that are going to be better because everyone else has the same data and you can't get that differentiation. So, you know, we focus on that through combining our predictions with high-throughput chemistry and biology experimentation and creating data per program at roughly 50 times the pace of pharma companies. So that expands our data mode as we're doing those programs in that way.

I'm curious, any other? I mean, from our lens is just way too little data, way, way, way too little data. But I think where people are playing with some ideas is what can you learn from data that are large but are not necessarily suited for your purpose to create models that you could then apply to smaller data that are better suited for your purpose. That feels like kind of a next frontier in bridging between different kinds of data, different scales, because there's some areas where you'll just never have enough data. I mean, for us sampling human beings, you're limited by how many people you can interact with, how many observations you can make on them. And people are not going to just like sit there and just be data machines, you know, in the way that a cell culture, you know, setup might be replicable at scale. I do want to go back, though, to the other discussion a moment ago, which was actually an interesting observation we've made internally about people and how they start looking at AI, which is that we have actually added a performance goal for each person. So we run our own clinical trials. We have clinical operations, goes down to like CRAs and all the sort of day-to-day staff of running clinical trials, not necessarily people you'd think of as all kind of AI native folks, but we added a performance goal. And not uniformly, but very largely, across people, it started to get them to think about what other things they could solve and new opportunities. And across the organization, in areas that are of nothing to do even with drug development, drug discovery, just the running of an organization, people just started looking for it because they found that it solves their problems, to your point, right? Right. So there is a culture that you can create that when whether you're now then talking about data or how to do a trial or or how to run analyses, it's, you know, you open up so much more of your organization to flexible thinking that leads to unanticipated exciting outcomes. So that's part of the fun of being in this dynamic time.

That makes sense. Maybe just quick on the data part. Obviously, at the end, more data equals more accuracy. If you think about the question around data availability data, particularly around data integration, and I know this is topics that were probably talked about 10, 15 years ago, specifically around things like electronic medical records. Could there be a future in which data could get integrated a specific city, state, nation across countries to basically inform things like clinical decision making? Could there be a future where just broad global data integration could occur and hospitals and physicians could all access that data equally to make more informed decisions? Hopefully, that's where the future is going, and you can imagine where AI would be extraordinarily powerful, and that would be a true global impact and change, right? And this is not where we're necessarily competing with each other because we're trying to find a drug first, but, you know, purely just – and I know that conversation, you know, had started many, many moons ago. We're obviously not even close to there yet, but, you know, hopefully that will be something we'll see in the future.

That makes sense. Now, as we get to the tail end of the discussion, I wanted to hit more on the regulatory side and also this idea that we can really shorten drug development times. The Ray Linum step seems to be also the FDA, but also just, as I mentioned earlier, the natural incentive to run phase three trials, because in the end of the day, a humanist, a doctor is still going to be prescribing a drug and patients are going to be making a personal decision. They want to make sure they're taking the best medication. So, A, when it comes to the FDA, they talked about, you know, adopting Bayesian models. They've talked about, you know, real-time clinical trials. But it, you know, that can be a surface-level discussion. I'm curious, in terms of, you know, discussions you've had with regulatory agencies, has there been a change, let's say, under this administration, to adopt more forward-thinking ways of drug development? And if so, what specific examples should we be paying attention to?

I'll be surprised if any of us comment on our own specific discussion. But I do think that we have seen activity from the FDA in terms of being open as to using AI for assessing toxicity and guiding dose selection. I think that is definitely a positive sign. I think the use of AI to identify connections between already approved drugs and maybe underserved patient populations or rare diseases is also another area where we've already seen just material activity happening. I think that that will increase, and I suspect that the degree of adventurousness of the FDA will be related to the degree of unmet need in terms of those patient populations.

Yeah, look, I think probably where from that piece, regulatory perspective, just sort of kind of taking the question as it comes here, but, you know, perhaps is in the area of real world evidence and real world data. and would there be, because at the end of the day, right, as you know, it's always about what's in that control arm, but ultimately if you could rely on that control arm being actual real-world data and apply some single arm design and you rely on that real-world data as your control arm, right, this is again back to the data integration piece, because now as a sponsor, you actually have to access that real-world data, going to sites, centers, you know, cooperative groups, whatever that is. But if there would be more open to that, that would be significant, right? Because now your trials are more streamlined. You're relying on data that's external, that's out there. And especially when you think about orphan type indications, you know, that would be a significant, I think, game changer for the industry and would hopefully be in line with this, you know, forward-looking perspective. Understood.

Yeah. I mean, a lot of the concepts that apply to use of AI from a regulatory perspective are actually already there and I'll highlight just one and it's just a matter of understanding how you apply principles to a set of tools and data and and hold by the way any data set that you submit to a regulator to the same level of rigor right it's not like when you use AI that means the rigor automatically changes in some way there's an enrichment guideline that was published in 2019 that is critical to understanding how to develop drugs for subpopulations. Obviously, it's something we think about a lot, but has no psychiatry in there. And it has no information on how we discover biomarkers. But it's clear enough that the concepts are laid out so that as we discover a biomarker, whether by AI actual intelligence or AI artificial intelligence, if you can lay out the data in a clear enough way, here's a subpopulation, here's how it reproduces, here's what I understand about its biology and why the drug might work, you follow the same principles from a regulatory sense. So you're upskilling your ways to get there, but the principles are still the same. Maybe you're getting there faster because you can, right? But at least our experience of FDA has been a lot of openness, just wanting the same level of data being shown, right? You You just can't assume that there's magic through some new tool, right? It's still the same bar in terms of evidence, to your point, about what needs to be shown to a doctor and a patient.

Maybe, lastly, I want to think about, and again, when you think about being AI native and we think beyond just drug discovery, a lot of it ends up being you're launching a drug and you're identifying patients. And that can apply to A, recruiting clinical trials, but B, you know, once you're on the market, making sure the right patient's actually getting your drug. You know, I think a lot of people, and maybe I'm just speaking for myself, when I think about, you know, pharma and, you know, direct-to-consumer advertising, I think of like the MyPillow commercials, and you'll get then like a ton of pharmaceutical commercials, and you'll be like, okay, well, that's how you've targeted patients historically, and it never seemed to be that effective, at least to me. But, you know, in my conversations with SmidCap companies, but also with pharma companies, there's been an incredible improvement in terms of identifying patients who really should be on your drug. And, you know, technology, whether that's AI or not, has has improved the specificity of that.

Can you talk about, you know, again, because you're taking this from scratch, you can run an organization from scratch as well. how is ai really going to help increase the productivity of your sales force and identifying the right patient should who should be on your drug i i think it's going to be huge moving forward i mean i think the sort of old sales force model i mean whether it's pharmaceuticals or other sectors i think is evolving extremely quickly right and even we didn't talk about that topic but even around clinical trial enrollment right companies like us are now even utilizing ai to even think about site selection right and for efficiency in the same way if you're now a salesperson you can implement ai about how do you more efficiently target the right medical centers so every individual is is maximized so i i think it's going to be significant and we're already seeing that and i think that's only going to accelerate moving forward so i think that's going to be great for everybody you know as i think it was mentioned sort of this n equals one individual can do so much more and i think there's probably no greater place that that could also be applied uh as it relates to a sales force i think what's particularly exciting is the tools that are being developed now are actually the tools that are being developed now i think in drug discovery are actually going to be useful for that task because if you know what we're doing right now is saying for this patient what is the optimal drug you can flip that question around is so

given an optimal given a drug what is the optimal set of patients so the exact tools that are being developed can can i think begin to address that question and erase the boundaries between drug discovery and diagnostics and prognostics and patient selection and even commercialization into one unified framework which should be the aspiration i'm really curious for your answer too because again you think about neuropsychiatry you know this is a human problem these are people that often don't have access to the healthcare system they don't have you know they often lose touch with their family they might have a cell phone right um they might still interface with technology in ways so when you think about identifying patients who would benefit from

your drug using you know next-gen technology including ai is is there an angle there that you're seeing yeah so so actually even just kind of current gen technology with llms yeah and tools to engage people. So half of people who have depression actually don't get treatment. So there you go, right? There's a huge population that could benefit, that should benefit, but because of things like access, stigma, not really understanding the disease and what treatments are available to them, they don't engage in treatment or they engage and drop out. Things to keep them in treatment, things to, I mean, nobody I don't think has done this yet. Correct me if I'm wrong, but as part of an advertising approach have an educational chatbot that helps engage people and answer questions that they may not even have somebody they can pose that question to that will open up awareness and an understanding for patients of what their their options are and what they're even experiencing a lot of that is like just basic person engagement right separate of the pharma angle just actually expanding care and awareness that lm tools do in scalable ways that we just don't have the staffing for understood uh last question and i hate to make it on the reductive side but when you think about um again efficiencies when it comes to the sales force

could we be in a future where you don't hire the same amount of sales reps as you traditionally would or um you know you have ai agents that are out and identifying patients and getting them on on drug and then you'll have maybe a supervisor monitoring them i mean how does that change you know your operating structure too is there is there benefits there you see i mean that's been the trend right anyway that you hire fewer and fewer people because i mean for a remote yeah you know telehealth system that we've increasingly become post-pandemic i don't know how you access the doctors in the same way that makes sense we're out of time but this is a wonderful conversation and i really do appreciate it thanks so much for uh you guys and you guys for uh attending and the thoughtful answers.

Really appreciate it.