Sacha 0:00 And so, similarly, like, the analogy I like to use is if you're going into a burning house and you're trying to put out the fire, you're not going after the match at the start of the fire. Right? It's irrelevant at that point. You're going for the actual fire itself.
Brandon 0:14 Hi. You're listening to Power to the Patients hosted by Brandon Lee, the podcast for leaders and innovators in pharma and biotech. In each episode, Brandon, the cofounder of Power, cuts through the noise to bring real world insights on modern drug development, clinical operations, and the science that drives breakthrough therapies. Because at the end of the day, it's about patients, not just molecules and markets. If you enjoy the podcast, make sure to subscribe and leave a five star rating on Apple, Spotify, or wherever you get your podcasts. Let's get started.
Unknown 0:51 Welcome to Power to the Patients. I'm your host, Brandon Lee. And today, I'm joined by Sasha Skirmohorn, founder CEO of Babylon Bio. Sasha, welcome to the show. I'm so excited for this conversation.
Sacha 1:01 Thanks, Brandon. Happy to be here.
Unknown 1:02 Yeah. Maybe just as a precursor, what is Babylon Bio, and how did this come about?
Sacha 1:07 So Babylon's a small molecule therapies company, folks on Alzheimer's. This is a disease I have cared about for a very long period of time. I started doing research. I was 15, I guess, when I started. I'm now 29. So I've been caring about the CZs for a long time. I started off on the scientific side and then, um, actually coincided with the diagnosis of my grandmother. And so I think just the kind of parallel track of, like, being on the research side, being very invigorated by a lot of the stuff we were seeing in the lab, that incongruity with, unfortunately, what happened to my grandmother. This was, like, uh, very kind of, like, sobering, uh, discordance that got me very passionate about starting a company here.
Unknown 1:42 Look, Alzheimer's is a thoroughly researched field. I'm curious, what did you see that convinced you that those are there, though? There's something different to do?
Sacha 1:49 I think a lot of it just started with actually talking to a lot of the top KOLs who were very jealous of my time. Like, when I started Babylon three years ago, a lot of it was just, like, the first year. So I technically incorporated in October '22. And the first year was literally just me sending hundreds and hundreds of cold emails. I probably met with 500 plus researchers in the Alzheimer's space in the first year. And people were very generous with their time and also honest about the lack of knowledge that we had as, uh, collectively as a field. I felt pretty invigorated by that. And I'd say, like, all of the confidence just came from understanding that there were fundamental pillars of this disease we still hadn't understood. That for me, generally, has just been an invigorating concept. And I think the reality is, like, known as answers here, and we've learned a lot as a field. And I think there's this opportunity to rewrite that playbook. me from understanding that there were fundamental pillars of this disease we still hadn't understood. That for me, generally, has just been an invigorating concept. And I think the reality is, like, known as answers here, and we've learned a lot as a field. And I think there's this opportunity to rewrite that playbook.
Unknown 2:34 And kind of, like, digging in a little bit to the the Babylon story, I know that you've got a a slightly unique way of approaching it. Uh, I know you can't get into too much of the deals, but, like, at the highest level, like, what is the thesis for how about all lots of purchase?
Sacha 2:46 I think your North Star is to develop a treatment for Alzheimer's. You have to get creative around financing that and amortizing the risk of any individual program. I think it is very hard as a company for your entire existence to be predicated on a single asset. And so to whatever extent you can mitigate the risk of being beholden to that one asset is probably a good thing. And so the way we like to think about it is just have a system that enables you to take as many orthogonal shots on goal as possible, and know that any individual program just may not have a high probability of success. But, collectively, you can amortize the risk of those individual programs not working by having as many of those orthogonal shots as possible.
Unknown 3:25 I've talked to a lot of drug developers. Of course. I've also talked to a lot of investors, and it kinda feels like a pitch like this is potentially hard to swallow. Like, we're gonna fund an amorphous, like, TBD cash generating, like, plan that can then, therefore, like, fund other TBD, like, plans. Like, what's underneath that and gives people the confidence to do it?
Sacha 3:43 Well, I think a big part of it is honestly just, like, assembling the Avengers of drug hunting. A big part of my early work was also going to the best people on the planet who had invented some of the best drug hunters alive and convincing them to come out of retirement to join us full time. A lot of this was, like, John Makor, for instance. I had read about his work originally through Remedipine, which then became Nurtec o d t and was later bought for $11,600,000,000 by Pfizer. But speaking to John and some of the rest of the folks that I ended up pulling out of retirement and just realizing that a key element to this whole thing working is derisking the execution. You're taking big bets on the biology no matter what.
Unknown 4:16 How do you get people comfortable with it? Like, such a amorphous plan. ohn and some of the rest of the folks that I ended up pulling out of retirement and just realizing that a key element to this whole thing working is derisking the execution. You're taking big bets on the biology no matter what. How do you get people comfortable with it? Like, such a amorphous plan.
Sacha 4:20 I think a big part of answering that question is really just showing the execution. And I think the earliest days, candidly, like, we're incredibly hard to fundraise on a lot of people. You know, I went out basically saying, I'm gonna go try and cure this undruggable disease. And, by the way, it's just me. And, by the way, I'm in my bedroom, um, in New York and, uh, no PhD lineup. And, you know, it was it was very, very hard pitch. It took a long time. Um, thankfully, there were certain investors who took big bet on me early on as just it was a founder of that and saw the mission alignment and that I've cared about this disease almost my whole life, and they were comfortable taking that bet. And I think over time, it just became a story of execution, being able to prove that, like, hey. Some of the key risk factors here are execution and so on. And so then on the execution front, it became my job to pull in some best in class drug hunters to help work on this problem. So I think the only chance you actually have is by recruiting, you know, the kind of the a team to to work on this problem. And so that was a critical part of that story. And then as time goes on, obviously, the proof will be in the pudding. It's gonna be about getting results in the clinic, and we'll endeavor to to be
Unknown 5:24 able to show that as well. You've talked a little bit about assembling the Avengers of, um, of drug discovery. But if we back up, you incorporate as 2022. You're, what, like, 26 of the time. Um, and you're going out, like, to these legends saying, do you wanna try another, uh, another bet here? Like, what does that conversation look like? well. You've talked a little bit about assembling the Avengers of, um, of drug discovery. But if we back up, you incorporate as 2022. You're, what, like, 26 of the time. Um, and you're going out, like, to these legends saying, do you wanna try another, uh, another bet here? Like, what does that conversation look like?
Sacha 5:41 It's one of these things where the more time you spend with the best people on the planet, the more you just realize they wanna work on the hardest problems. I think there's a world in which, like, you may a priority think that, oh, well, this is, like, this is Sisyphean. Right? Like, look at the graveyard of previous attempts and, like, why should we win? There's no reason for it. And, you know, maybe you should dilute the ambitions of the company to try and recruit people. But I actually think if you're uncompromising on that, that actually turns out to be a good bat signal for, like, extremely talented people. Um, and so yeah. I mean, you know, I think another part of that was just lowering the activation energy to getting them involved. I just started by saying, hey. You know, let's just catch up. I I'm not looking for anything. I just wanna kind of, you know, get to know you. And so for pretty much everyone who ultimately came on full time, it started as just like, hey. Let's do catch ups, like, once every month or every few months, whatever. Just talk science. Then it became, hey. Why don't we just bring you on as a consultant? Do five hours a day. And then it became, you know, a conversation around, hey. I think it's it's time for you to, you know, fail retirement and come join us full time. And, thankfully, we've we've, uh, had a good track record there too.
Unknown 6:41 Yeah. The slippery slope out of retirement is, like, a a funny recruitment tactic.
Sacha 6:44 The best people on the planet just, like, literally, they're indefatigable and also just I don't know. They're just once you get bitten by the bug, there's no going back.
Unknown 6:51 They can't stop thinking about it.
Sacha 6:53 Absolutely. Exactly. That's why, like, you know, I think if you just lead with passion, lead with enthusiasm, these best in class people just they they feed off of that, and they reciprocate it. And, yeah, it's, you know, one of the great pleasures of what I'm doing is, like, I get to work with people like John, Bill, the rest of the team. It's just great honor.
Unknown 7:09 Totally. Let's talk about Alzheimer's a little bit here. An incredible mission. What has everybody gotten wrong in the way that they've thought about Alzheimer's, approached Alzheimer's at age? And I'm curious if there's a, uh, point of view that you're holding today that's maybe feels underrated to the world, but, uh, do you think is appropriately priced on on your end? re. An incredible mission. What has everybody gotten wrong in the way that they've thought about Alzheimer's, approached Alzheimer's at age? And I'm curious if there's a, uh, point of view that you're holding today that's maybe feels underrated to the world, but, uh, do you think is appropriately priced on on your end?
Sacha 7:26 I feel uncomfortable kind of saying that implicit in my characterizing what's wrong is that I know what's right. I can't say I I have that. But, certainly, I think one thing that we have not yet updated our priors on is this kind of idea that there is a let's call it discordance between the accumulation of some of the pathologies we think are important with the actual emergence of symptoms. And so a patient that is starting to forget where their keys are and, you know, starting to have this memory loss that, you know, is infamous with Alzheimer's, that person, on average, has had amyloid in the depositing in the brain for about twenty to thirty years prior. And so, similarly, like, the analogy I like to use is if you're going into a burning house and you're trying to put out the fire, you're you're not going after the match that started the fire. Right? It's irrelevant at that point. You're going for the actual fire itself. And so I think similarly, we've yet, as a field, collectively update our priors as we think about some of these pathologies that, yes, are probably causative, but being the cause, the zombie and is the best treatment vector. Right? And so this is, like, one thing that, I guess, if I just, like, survey the landscape or kind of go to the conferences and so on, that is a burgeoning kind of concept that we should actually consider the timeline of these pathologies. But still, I think it remains incredibly underappreciated.
Unknown 8:39 How much sooner before symptom onset do you think we should be looking?
Sacha 8:42 I think it's whatever the biomarkers enable. And I think this is actually, like, a super interesting inflection point we're passing as the field is, like, in a world where you can predict with a 100% certainty whether someone would convert to the clinical stage of Alzheimer's within one year, that is, like, incredibly tractable from a prevention standpoint. Right? Because if you're actually gonna run a prevention trial today for the most and biomarkers are starting to help with this, but good luck running a a twenty to thirty year trial. Right? I mean, this is never gonna happen in a million years. So I think the way that I think real unlock for the field is gonna be when you can, with an extremely high certainty, predict within a short period of time, something like twelve months, actually run that prevention trial. Because that's when, in theory, the pathology if you intercept the right pathology, you should be able to stop those patients from converting. But similarly, if you have surrogate biomarkers that folks believe and specifically that FDA believes will eventually represent the conversion to to MCI due to AD. That's a tractable trial that also people will believe. That's all you'll really need. pathology, you should be able to stop those patients from converting. But similarly, if you have surrogate biomarkers that folks believe and specifically that FDA believes will eventually represent the conversion to to MCI due to AD. That's a tractable trial that also people will believe. That's all you'll really need.
Unknown 9:41 Are there biomarkers that you're most excited about today, Adam?
Sacha 9:44 The big one, I mean, waves in recent years was p tau two seventeen. I think it's incredibly interesting. Some more recent stuff around eMTBR two forty three, and these are, like a lot of the tau related biomarkers when you synergize them with amyloid positivity, specifically a beta 42 of 40, these are the kinds of things that give you a high AUCRC in terms of predicting conversion to MCI. The problem is if you use them in isolation, you can do it with high certainty within, like, a five to seven year time horizon, but you need that to to shorten significantly. I'm very convinced the next three years will have big breakthroughs there, and we're already seeing kind of early precursors to this. But then there's a separate problem, which is if you're trying to predict the conversion to MCI in sporadic Alzheimer's patients like, part of the reason people have gone too familiar with Alzheimer's is that it's a high certainty of conversion, and so you know when they should develop it. And so any exceptions to that conversion become very interesting. But yeah. On the other hand, you know, with some of these biomarkers, you just really need to do weight matrix of different factors, figure out what the conversion is. And then beyond that, understand that there's still a prevalence problem because maybe only five percent of the patients that you're actually tracking will convert. Maybe you have a hundred percent certainty for that five percent. But if it's only five percent, again, there's, like, a sample size problem that you're dealing with. And so these aren't solved problems, but I'm actually, like, very I think the the trend lines in terms of biomarkers and so on where the field's going is, like, you know, trending incredibly favorably.
Unknown 11:06 This reminds me of like, we've done some multiple sclerosis work where what you're tracking is relapses. And currently, like, in MS, there's some pretty good DMTs that I I think the average patient does has a relapse every, like, eight years or something like this, which means in order to have, like, a stat sig, like, improvement in, like, lengthening the time between relapses, you're tracking for quite a long time, and you have to have quite a large sample size in order to have anything that looks, like, remotely statistically relevant. like, eight years or something like this, which means in order to have, like, a stat sig, like, improvement in, like, lengthening the time between relapses, you're tracking for quite a long time, and you have to have quite a large sample size in order to have anything that looks, like, remotely statistically relevant.
Sacha 11:30 This is the horrible and tragic reality of of Alzheimer's, but also on the other side, you know, serves as an interesting scientific benchmark is, like, the decline for Alzheimer's is inexorable. Right? So the placebo the room for placebo effect is relatively like, the ceiling is very low. And so at the limit, the probability of a false positive is essentially zero. If you're giving a drug, it's either going after two years. If the patient is stable, if multiple endpoints are showing stability multiple years out, that's the holy grail. And so I guess I would, like, also say scientifically, I'm, like, atypically comfortable with very small sample sizes in terms of if there is an effect on multiple endpoints that are suggesting stabilization. Even if it's a small subset of patients, I believe those data after a certain period of time. Because, again, you can't placebo your way out of having Alzheimer's. Maybe this remains also underappreciated. It's like, that some of these trajectories, they're not gonna be reliable, but we are starting to triangulate trajectories with biomarkers. I feel pretty convinced that, like anyway, there's probably, like, more work to be done there, but you can stratify patients accordingly.
Unknown 12:32 Your stated North Star, let's solve Alzheimer's, the kind of stated plan is we're gonna be a a drug development company. But it sounds like the the real bottleneck is having something predictive enough first, like, to go after before developing any drugs in the first place. So previewing a little bit into the plan here, or am I off base? tar, let's solve Alzheimer's, the kind of stated plan is we're gonna be a a drug development company. But it sounds like the the real bottleneck is having something predictive enough first, like, to go after before developing any drugs in the first place. So previewing a little bit into the plan here, or am I off base?
Sacha 12:48 Of course, the perfect world is one in which you can kinda predict prior to running the trial what's gonna work. I think that's remains a pipe dream. Um, but there are I'd be remiss not to comment on the fact that I do think AI is gonna be some non trivial percentage helpful there. A part of that is just on the ideation phase, like, we found internally some of the internal tools we've built. We have an internal kind of coscientist, let's just call it, or co pilot that has been incredibly helpful for the team as they chase the idea maze on a bunch of different, you know, ideas. It can definitely help collapse the kind of iteration loop for ideation. The example that we had was, uh, the fine tuned model with OpenAI. So we spent six months with the OpenAI research team developing a model to do basically only one thing, which was predict clinical trial outcomes and specifically predict whether the primary endpoint would be statistically significant. And turns out that you can stitch together these clues and pretty reliably predict clinical trial outcomes. Now does that mean it's better than the best CMOs in the world? No. I don't think so. I just think it means it kind of adds a scaling factor to their expertise and allows you to actually make predictions in a much more high throughput way and serves as a nice existence proof that, like, fine tuned AI can clearly do something. And so, you know, that's maybe another thing that I think is, like, potentially mispriced currently is is just this idea that AI can be fine tuned to do be hyper specialized and good at specific tasks. Maybe folks aren't appreciating that enough as well.
Unknown 14:13 This idea of predicting clinical trial outcomes is fascinating. Can you tell me a little bit more about, I guess, specifically, what you think there is predictive power for? can be fine tuned to do be hyper specialized and good at specific tasks. Maybe folks aren't appreciating that enough as well. This idea of predicting clinical trial outcomes is fascinating. Can you tell me a little bit more about, I guess, specifically, what you think there is predictive power for?
Sacha 14:20 I think where these models perform extremely well is in trying to determine basically, just like with all AI, anything that's within distribution. So if you're trying to determine the probability of something being best in class, that's where, at least, we're seeing, you know, incredible performance with these models. In terms of things that are, like, out of distribution, like, first in class medicines, what have you, the predictiveness falls off. And so that begs an interesting question of, like, you know, what can we actually do to enable these models to then start underwriting things it hasn't seen before? That's an unanswered question. I mean, I'm definitely of the the belief that human experts helping tweak these models is very good way of tackling that problem. But the concept is relatively straightforward. Right? If you could spend, uh, a hundred thousand hours and had perfect recall and you could go and read, you know, as many papers as you'd like within that time, I'm pretty sure if you with enough time and enough iterations, you'd be able to figure out what's predictive of whether a trial will hit the primary endpoint or not. And so this is kind of the basic concept there in terms of these models. Right? They're just stitching together clues from previous trials, from, uh, information out there on the Internet, as well as proprietary databases in some cases, and making predictions accordingly.
Unknown 15:30 Is this all just, like, MOA believability? Is this, like, also incorporating trial design execution? Like, what kinda goes into that? ther clues from previous trials, from, uh, information out there on the Internet, as well as proprietary databases in some cases, and making predictions accordingly. Is this all just, like, MOA believability? Is this, like, also incorporating trial design execution? Like, what kinda goes into that?
Sacha 15:38 So the beauty of it is it actually determines quantitatively how it should properly weight those different factors. Right? And so at the end of the day, like, um, there's a lot of cases. So the model we built with OpenAI spit out a justification with each trial prediction. And so when we make a prediction there, it's not just giving us a probability of success for any given trial. It's also giving us a full justification that can then be underwritten by our team of best in class drug hunters or whomever. That interpretability layer was very important from a design perspective because it allowed us to say, okay. Well, what is the model using to underwrite this probability? And individuals can come in and scrutinize that that justification accordingly. We see in the justifications, like, sometimes it would say that the biology is conceivably actually, that targeting this biology would actually work on paper. But when I actually do the calculations on the PKPD, we actually think we're gonna get insufficient exposure to actually get the therapeutic effect, and it'll fail for that reason. And so we'd see it would converge on the same decisions from completely different justifications, And that on its own is is an incredibly, uh, kind of scintillating prospect, at least for me.
Unknown 16:43 Fascinating. This feels like the kind of thing that, if true, is a no brainer. What do you think is, like, stopping everyone from just doing this right now?
Sacha 16:50 I mean, a huge part of it is proprietary data and also know how. And I think the bigger question is, like, the trend line is just clearly that these models are getting better. And at the bar, in terms of what it requires to make models even better, is going up and up and up. And so I think as we kind of, like, shift rightward on the distribution in terms of what expertise is required to improve these models from a eval's perspective and beyond, kind of begs the question of, like, what point will fine tuning no longer be enough, and at what point will there not be enough, uh, data to be able to fine tune these models further?
Unknown 17:19 What data did you have to run this, like, fine tuning out? Because if I'm guessing how you did it, and you can tell me if I'm guessing wrong, if I'm guessing how you did it, you went through clinicaltrials.gov. You scraped out, what, the eligibility criteria, and that's, like, kind of there, but not exactly, like, the eligibility criteria for the study. And then you looked at whether each of the, uh, the kind of studies passed failed because that's all public record. Is that is that more or less right? ugh clinicaltrials.gov. You scraped out, what, the eligibility criteria, and that's, like, kind of there, but not exactly, like, the eligibility criteria for the study. And then you looked at whether each of the, uh, the kind of studies passed failed because that's all public record. Is that is that more or less right?
Sacha 17:42 So it started with the licensing of a dataset from a third party that had spent years fastidiously tracking hundreds and hundreds and hundreds of trials and putting in the dimensionality of the data we haven't reported. But there was a lot of parameters that went into that, including the trials on, including a bunch of other things. And so all of that got integrated. Yeah. Since then, you know, it's been a story of just tracking tracking accordingly. Right? You track perspective trials. Once you collect those data, you kind of recursively feed those back into the models. And and, you know, hopefully, if everything goes right, you see a trend line that still remains positive, um, as it sees more and more data.
Unknown 18:16 And I'm dying to that. Like, did you see a difference between the easy publicly available way I just described and the proprietary dataset that you described? I'm curious about what the difference look like.
Sacha 18:25 I probably shouldn't go too much into those details. What I can say is, obviously, the ability of the models to make good predictions is directly proportional to the amount of data you put in. And so there's proprietary data around specific trials definitely help. And so for our internal programs, it has access to full proprietary data that we have. And for the most part, you're relatively constrained in terms of the predictions you can make. For instance, for the trials that we made with the fine tuned model, OpenAI, and the the data that we had announced, this was very specifically only in publicly traded companies. And part of the basis of that was that they actually had a lot of first of all, they were forced to report the results in most cases. And then secondly, a lot of those came with previously reported clinical data that were in previous investor presentations or what have you. It provided a repository of data that allowed the model to enable the model to basically pull from these different sources. And, again, like, if if the kind of success of the model is directly proportional to the amount of data it sees, okay, naturally, you know, you'd want to kind of give it as much data as possible, and that's where, you know, the comprised error data element also becomes a relevant factor.
Unknown 19:29 So then, like, uh, playing this forward, like like, every top 10, top 20 pharma company builds this for themselves, and then IQVIA builds it and sells it to all the biotechs. Is that is that the, like, the the natural end state of market? comprised error data element also becomes a relevant factor. So then, like, uh, playing this forward, like like, every top 10, top 20 pharma company builds this for themselves, and then IQVIA builds it and sells it to all the biotechs. Is that is that the, like, the the natural end state of market?
Sacha 19:40 Sure. I think, like, on a time line that I'm not sure I can predict, there will definitely be the commoditization of all these models. Right? And so I think an important question for a lot of the players in the space to to be asking is, you know, how can we kind of lean into whatever your edge is? The opportunity to have an edge, I think, is collapsing exponentially. And so so, you know, a bad word taking is obviously that you can take tail end expertise and use those to be able to fine tune these models. And if you have best in class people as your evals for the, you know, these models, then that's probably better than having an army of people who are, you know, good but not, you know, the best. But, yeah, at the end of the day, people are gonna have to start playing strategies because that window is definitely closing in my opinion. And, um, uh, it'll be really interesting to see how it all plays out.
Unknown 20:21 The funny thing I think about the collapse of your edge is that, a, it follows the iteration cycles of, like, an industry. Right? Because eventually, like, these things get competed away through iterations. Every industry, every market has, like, a natural physics of, like, how fast, like, the iterations go. And it just so happens that, like, this market, like
Sacha 20:38 The longest and most expensive you can possibly have. Yeah.
Unknown 20:42 You might have an edge for decades, like, if you're very hush-hush about it, which is why you're not telling me what you're doing.
Sacha 20:47 That's, like, one version of the story. I think the other is, like, to properly validate, especially on things that so so there's kind of two mechanisms or two ways to interpret that. One is the one you just mentioned. I think the other is that it takes a very long time to prove out whether your strategy works. Right? And so to whatever extent you can come up with clever ways to kind of near term validate your your method, which we tried to do, by the way, by just by basically just training the model, uh, to make predictions on trials that were being read out so perspectively. So we didn't need to only focus on our pipeline. We could focus on the totality of all trials that we that were getting read out. Alright.
Unknown 21:20 I'm gonna I'm gonna read you a list of pending phase threes, and you tell me what you think. Alright? And then we're gonna go
Sacha 21:25 You wouldn't be the only person asking me for those. I'll tell you that.
Unknown 21:28 Can't invest on, uh, in the market. I've got too much, uh, too much proprietary info. That's fascinating. So, like, this is mostly like an asset selection source of alpha. Is that right? then we're gonna go
Sacha 21:25 You wouldn't be the only person asking me for those. I'll tell you that.
Unknown 21:28 Can't invest on, uh, in the market. I've got too much, uh, too much proprietary info. That's fascinating. So, like, this is mostly like an asset selection source of alpha. Is that right?
Sacha 21:39 Yes. For sure. I mean, so we internally, it helps us rank you have to constrain the search base somehow. Right? Because the number of, like, licensable assets is just, like, exceeds any individual's, you know, ability to to diligence them. Then we're a super small team. And so one version of this is, like, well, you just scale up the number of employees and have, like, a thousand employees whose only job is finding good assets. The other is find more intelligent ways of constraining the search base. And so the way we do that is, like, we can make these predictions because inference is cheap, and we fine tune the model, and I negotiate with OpenAI that we would get the model at the end of that process. And so we use this model to make predictions, and those predictions can we can rank order assets that we're interested in based on a variety of parameters and then rank order by the probability of success for whatever the x clinical inflection point is. So capital and time, and this is, like, where a model is, like, in theory, infinitely scalable. If the COGS just converge on compute shit, inferences, like, really cheap. We can just run this a million times. It doesn't really cost us much, and and that will help us find assets that are actually more interesting than than others.
Unknown 22:40 Where are you most skeptical of the way that it it kinda runs?
Sacha 22:43 I think, again, like, the thing that I'm, like, maybe most skeptical or concerned about is what happens when presented with something that's out of distribution. If this is a first in class medicine, you know, how does the performance actually fare? And so I think the best thing you can do to kind of get to that point is to try and extract the heuristics of what how a best in class scientist or drug hunter would underwrite this. And to whatever extent you can extract those heuristics, try and bake those into the model. Right? Like, the way John on our team thinks about things is incredibly idiosyncratic. There is just a certain way of thinking that comes from being in the game for forty years and having consistent successes, like inventing for FDA approved drugs is just like that is a nontrivial feat. And so, like, that is definitionally out of distribution to begin with. And so having someone like that give feedback on these models and actually integrate their ways of thinking into these models, I think is the only way you can get tackle that problem of, like, how do we then make a prediction on something the model has never seen anything like it before? of distribution to begin with. And so having someone like that give feedback on these models and actually integrate their ways of thinking into these models, I think is the only way you can get tackle that problem of, like, how do we then make a prediction on something the model has never seen anything like it before?
Unknown 23:39 Is a day in the life for, like, a week in the life over at Babylon, AI gods come down and they they say, like, these things look great, and then and then you've got, like like, the legends kind of, like, looking at it going, this is probably wrong for these reasons, or this is, like, like, probably worth trying?
Sacha 23:53 I wouldn't say it's just, like, it's clean as that. Right? I mean, these are at the end of the day, all these things are tools in the armamentarium. Right? And so the way I see this is, like, to whatever extent you can expand your arsenal, you should do so. And so we employ different tools accordingly. Right? Like, certain situations require, um, a little more, you know, heavy lifting on the AI front than others. We kind of depend very heavily on specifically only the expertise of our best in class team because an AI would not be able to even underwrite that program. So, yeah, I'd say it's just like assembling an armamentarium and then deploying those different tools accordingly. It's, uh, a day a week in the life of Balan is some approximation of what you characterized for sure.
Unknown 24:31 I mean, that sounds like a ton of fun. I can see why people wanna do it. Even just kinda abstract this out to what we're seeing with, uh, with kinda software development and engineering, It's able to kind of take care of a lot of the grunt work that was nobody's favorite part of the jobs anyways, and they get to focus on a lot of the thinking and, like, the hard stuff, which, um, at least the best people in the world, uh, love to do.
Sacha 24:49 By the way, I think that's like there's the sexy use cases that everyone wants to be, like, the AI use case for for drug discovery unit development. I think there's the long tail minutiae that are less sexy, but actually, like, probably the most impactful on the day to day. And so, you know, I think that's another thing for people building at this nexus, like, to be thinking about is there's the kind of, like, headline grabbing things, like, you know, honestly, the fine tune involved to make clinical trial predictions. That is one of them. And then there's the day to day or, like, you know, regulatory submissions or you know, there's a litany of different things that we have to do in the kind of discovery and development process that, um, should be automated. LMS have enabled a lot of the the kind of automation of this for for the first time, but but maybe because they're not sexy enough, people aren't pursuing them. I don't know.
Unknown 25:31 I'd love to just riff with you a little bit. I feel like I hear a lot about AI and drug discovery. You're kinda choosing a a sense of, like, AI in, like, screening. f automation of this for for the first time, but but maybe because they're not sexy enough, people aren't pursuing them. I don't know. I'd love to just riff with you a little bit. I feel like I hear a lot about AI and drug discovery. You're kinda choosing a a sense of, like, AI in, like, screening.
Sacha 25:41 Yeah. AI for screening molecules, but not high throughput screening, for instance. I'm more skeptical there.
Unknown 25:46 Talk to me about, like, the value chain of, like, the discovery, like, uh, world and, like, where you think it's got a great leverage or where you think it's maybe overhead? 0:25:41] Sacha: Yeah. AI for screening molecules, but not high throughput screening, for instance. I'm more skeptical there. Talk to me about, like, the value chain of, like, the discovery, like, uh, world and, like, where you think it's got a great leverage or where you think it's maybe overhead?
Sacha 25:54 Well, I think that's one of the interesting things about where to apply AI is a lot of folks are focusing on the discovery stage, and there's clearly some level of enablement there. But I think my concern is that the super majority of spend in the entire process of getting a drug to market is in the later stages. And so if the costs are going up exponentially as a function of time and progress, it makes more sense, in my opinion, to be focusing on that side of the distribution than the earliest stage where the timelines are shorter relatively and the overall cost is much lower. For me, it makes more sense to be trying to deploy these systems. You talk about leverage. I think the the leverage is, like, also, you know, the leverage and impact is directly proportional to the cost of that phase. And so we should be focusing we should be kind of, like, focusing more on the tail end of that process instead of the earliest stages. And I think another part of that is also, like, what kind of data are available. Right? I mean, for for a lot of, like, small molecule chemistry side of things, the majority of information around medicinal chemistry comes from the stuff that doesn't work, not the stuff that does. And so if you're only training a model on all published literature, for instance, or, you know, kind of FDA approved drugs or whatever, that's the tip of the iceberg. And so to be sampling only from the right tail of the distribution, that feels like the wrong way to tackle that problem when, actually, the majority of the information is in the kind of the massive the long tail of other stuff that didn't work along the way. Basically, part of the reason that I think, you know, at least we've decided not to take a bet on the kind of, like, AI chemistry nexus is that you learn from the med chemist pretty early on that the super majority of information is coming from the stuff that that doesn't work, not the stuff that does. And so your these activity cliffs are basically you know, you find a molecule where you start to see some improvement on some vector, and you have to contextualize that with the long tail of other molecules that were designed or screened that didn't actually work. Yeah. My my fear is that, like, a lot of these AIs that are just trained on publicly available information or FDA approved drugs, they're overly sanguine by default because they're only taking stuff that was on the right side of that distribution. The majority of the information's in the 99% of stuff that didn't work, not the 1% that did, unless you own that part of the value chain where you're actually screening the molecules and you have decades of information there. CDD Vault is a kind of an interesting one, although all those data are proprietary, so they won't be able to capture it. But I do think a lot of these folks that are playing at that layer and have actually run the screens over the past several decades, if they have stored those data, that is incredibly valuable. But I think absent that, it's just, uh, it's an incredibly hard problem to solve and also probably doesn't generalize across targets. Right? You may have an incredibly potent inhibitor of one kinase, but it's not gonna generalize to others because chemical space is incredibly jagged and just very hard to do. valuable. But I think absent that, it's just, uh, it's an incredibly hard problem to solve and also probably doesn't generalize across targets. Right? You may have an incredibly potent inhibitor of one kinase, but it's not gonna generalize to others because chemical space is incredibly jagged and just very hard to do.
Unknown 28:30 Totally. You you know, in in the kind of, like, conversation before the recording, you're telling me a little bit about, like, thinking about giving the the best drug developers in the world, um, AI superpowers. What specifically are you talking about when you when you think of AI superpowers for them?
Sacha 28:44 I mean, I think part of it is honestly just giving them, like, for a lot of these, they're used to doing drug discovery in a way that is, like, they've been doing drug discovery in the same way, more or less, for about forty years. And so the way I think about this is, like, how do you lower the friction to using a technology that they've never aerated before into their process? And part of it is showing them the value in a very near term. And so I think where we're seeing a lot of wins on that front is, like, they are chasing a new idea, and an AI can be very good at surfacing the relevant literature. That's the lowest kind of tier of the totem pole, and we're seeing a lot of wins there where it's just pulling more relevant information than you find on Google or Google Scholar or whatever.
Unknown 29:24 Have you seen this AI prediction market for clinical trials that seems to be trending on Twitter? You must be looking at at that going, man, like, you've either got it all wrong or you've got it all right. Uh, what what's your take? information than you find on Google or Google Scholar or whatever. Have you seen this AI prediction market for clinical trials that seems to be trending on Twitter? You must be looking at at that going, man, like, you've either got it all wrong or you've got it all right. Uh, what what's your take?
Sacha 29:36 Here's what I can say. Predicting the trial outcome is not the same as predicting the price action of the stock after the readout. And so we've seen that, you know, time and time again, and I think especially as it relates to the statistics of the primary endpoint versus the effect size. So we've actually done these kind of post hoc analyses because people have asked us this question in the past as well. And the long story short is predicting the price action of of, like, a lot of these publicly traded biotechs is like, it's so uncorrelated with the actual outcome of the trial. And a lot of times, it has to do with the run up prior to the readout. Maybe over the past year, the stock's up, you know, 70% for whatever reason. And so the room for it to go thereafter is pretty low. Maybe sometimes you hit the primary endpoint, but the effect size was not good enough. Like, think about a lot of the the GLIP ones where you may actually have stat sig, uh, weight loss, but the actual effect size of the weight loss is less than whatever the best in class drug is at the time, and therefore, stock goes, you know, trades below cash. So, yes, it's definitely, like especially with the retail investors, it's picking up so much steam, and I'm actually, like, glad people are, you know, paying more attention to this. But but, yeah, the correlation is, like, a lot lower than you think between the primary endpoints or at least even the effect size, honestly, and, uh, the actual outcome on the price action.
Unknown 30:48 Yeah. There seems to be two parts of the story. The first one is, like like, how do you feel about the the tradeability of the information and, like, the encouragement of, like, more, like, the exuberant retail and and biotech, which is, like, admittedly, like, a tough space to be in, to be a retail investor in. Then the second one is the the potential ethical, like, implications on, like, actual trial operations and, like, the way that, uh, we try to, like, maintain neutrality. Like, it is the work we do. Like, I I know that sponsors would be, you know, squeamish about the idea that patients could go look up and study and have some outside in perspective on whether it's likely to work or not. And does that impact kind of expectancy bias and, like like, placebo response and, like, all these different, like, opponents? I can just imagine all the, like, beagle croft, uh, that's gonna come from this. patients could go look up and study and have some outside in perspective on whether it's likely to work or not. And does that impact kind of expectancy bias and, like like, placebo response and, like, all these different, like, opponents? I can just imagine all the, like, beagle croft, uh, that's gonna come from this.
Sacha 31:34 That is an incredibly good point. That's actually an incredibly interesting thread. I mean, I I I think, um, um, that's really interesting. I have not considered that implication in terms of expectancy bias and how for sure. And also, like, from a recruitment perspective. Right? Like, is it fair, like, in a world where you have a system that is, like, not, you know, 100% accurate, people's enrollment in trials is not, like, per se driven by their expectation around the success of that drug. I think most people enrolling in a trial see that there is a drug for whatever disease or indication they're, you know, suffering from, and that is sufficient for them to want to enroll. I think when you stack on this layer around some external prediction, the outcome of the trial, there are ethical implications for sure on that. And I think the only solution to that is information suppression, which I'm not in favor of. And so I think it it kind of it presents a really interesting dilemma that, like, probably is unavoidable candidly, but it is, um, yeah. I mean, like, imagine, especially in a world where you have a false negative, I think the implications are really severe. Like, if you were to have a cancer treatment that an AI model predicts is gonna fail, in a world where that was wrong and it was an actual false negative, that has profound implications in terms of its effect on society. So that's a really interesting point, and I'm, yeah, well received.
Unknown 32:49 This has been such a fun conversation. I, uh, really appreciate you coming on board, shooting the shit, uh, maybe sharing a little bit more than, uh, um, maybe typically, uh, share it online, but I've so enjoyed this conversation. So thank you.
Sacha 33:01 Likewise. Thank you so much. This was super fun.
Unknown 33:03 Yeah. Of course.
Brandon 33:06 That wraps up today's episode of Power to the Patients. If you found this composition valuable, make sure to follow or subscribe so you don't miss the next one. For more insights on drug development, clinical operations, and the future of pharma and biotech, connect with us on LinkedIn. Thanks for listening, and see you next time.