Future Proofing Clinical Development with Recursion

Speaker 1:
Welcome to TD Cowen Insights, a space that brings leading thinkers together to share insights and ideas shaping the world around us. Join us as we converse with the top minds who are influencing our global sectors.
Brendan Smith:
Okay. Welcome back to another episode of Machine Medicine: AI & Healthcare, TD Cowen's podcast series where we bring you the latest and most important takeaways from the state of AI in the healthcare sector today. I'm your host and TD Cowen healthcare analyst, Brendan Smith. And today I'm joined by Recursions, SVP of Clinical Development and Data Science, Sid Jain. Sid, it's great to have you. Welcome.
Sid Jain:
Thank you, Brendan. Looking forward to this.
Brendan Smith:
Yes. So for anyone new to our podcast series, Machine Medicine aims to break down the use of artificial intelligence across healthcare into bite-size digestible points one at a time, highlight the biggest misconceptions, and then ultimately re-contextualize each piece back into the bigger picture. And today, Sid and I are looking to dive into what it means to build out a truly proprietary ClinTech platform and how Recursion is more or less future-proofing their clinical development portfolio with AI. So I think with that, Sid, let's just dive right in. I guess we, as I'm sure you do, continuously get questions from investors, academics, corporates, really about how to maximize success in one of the most make or break stages of drug development, which ultimately is clinical trials.
I know Recursion sits at this intersection between an AI powered technology company and clinical stage biopharmaceutical franchise. And over the past few years, I know you all have invested a ton of time and energy into really expanding AI beyond this drug design and discovery phase into really every stage of your clinical development network. But maybe before we dive into the specifics of your platform, I guess high level, what would you say for people listening in are some of the greatest challenges that drug developers face specifically within this clinical development process?
Sid Jain:
Yeah, it's actually a pretty good place to start because the instinct in our field is to assume that the hard part is inventing the molecule. And trust me, discovery is hard, designing the right molecule is hard. But over time, especially with the work that companies like Recursion are doing in terms of advancing how we design molecules, how we make those molecules better, the tension still remains on the development side. That's where the biggest bottleneck is. If you think about all the advances that are happening over the last 10 years or so, the approval for the new molecules has more or less remained the same, so about 40 to 50 novel drugs a year, and that hasn't really scaled with the pipeline at all. So while the discovery is getting better, we're still not able to get more drugs to the patients faster. So development remains that bottleneck. So some of the stats are sobering.
1 in 10 drugs that enters a human trial makes its way to the patients through a regulatory approval. Think about that. The majority of the cost in R&D is in development and we don't even fail cheaply. And a lot of this failure is self-inflicted. So nearly 60% of the trial protocols require an amendment and about a third of those are completely avoidable. Roughly 1 in 5 sites activate but never enroll a single patient. And at the end of the day, only a low single digit percentage of eligible patients ever make it into a trial at all. So the point is, we are going to continue to make discovery better, faster, cheaper, but if we can't raise a probability of success and eventually tackle the development bottleneck, we're not going to be able to get these drugs to patients any faster.
Brendan Smith:
Yeah, that's great. And I think that really helps level set a handful of the challenges and there's always kinds of top to bottom and left to right processes that need fine-tuning across the board. But I guess maybe drilling into Recursions platform specifically, you all launched your next gen ClinTech platform last year, thousand foot view, I guess through that lens that you just laid out for us, could you maybe walk us through the platform as it stands today and ultimately what are some of the most important considerations when building out a new AI/ML capabilities specifically within clinical development?
Sid Jain:
So the problems we're trying to tackle are the ones I just described. One is, how do you increase that rate of 1 out of 10 drugs making their way into the patient's hands and how do we do that cheaper and faster? So the way to do that is... Our ClinTech platform is centered around these key questions. The first question is, who is the patient that's most likely going to benefit from the molecule? If we get that part right and get the trial designed right, you improve the probability of success significantly. You make the trial cheaper and faster because you then have to enroll less patients. And then at the same time, you want to accelerate the operations and the actual process to bring that molecule, once it's in trial, to the patients. So how do we enroll faster? How do you automate a number of these workflows that takes weeks and months into days and hours?
I mean, just recently we've done some work with a dose prediction and the work that it takes to read out. Instead of taking several days, it now takes hours. So think about that compounding across multiple processes. And then in the end, the evidence generation piece. So how do you... Especially in the rare diseases and early stage portfolio, there's very little literature at times. How do you contextualize the evidence that you're generating with the natural history data? And there there's tremendous potential to use AI/ML versus some of the historical ways through manual chart reviews the evidence was generated. And in fact, some of the times that was too cost prohibitive to even take on as an undertaking.
So we're applying this question across the board from, as we look at the preclinical data to translate the animal models and in vitro models, how do we embed the real world clinical data into the same embedding and the foundational models to predict better patient response? So who are the super responders, non-responders, and based on those, make our indication selection and patient selection choices, all the way to recruitment and evidence generation.
Brendan Smith:
Yeah, I guess what would you say have actually been some of maybe the more notable improvements in your platform that you've already seen today? And I guess maybe through that same lens, something we're often asked is, how are you actually able to measure success in those instances, I guess in a more kind of discreet, concrete way?
Sid Jain:
Yeah, that's the hard part. I mean, we just talked about development being a 7 to 10-year process and we've started on this journey about a year ago. But at the same time, there are leading indicators which I can talk about. First thing is, even the process it takes to get a study up and running. Instead of relying on the manual processes and historical ways of doing your feasibility and site selection to having access to vast amounts of real world data, we're talking about over 300 million covered lives, linked with all the operational data. Our teams, our clinical operations teams, our clinical development teams have access to all that data at their fingertips. We can run these feasibility analysis in matters of hours versus days and months. Sometimes we had to wait to work with a CRO to get the results back and have the contract, et cetera.
So along the way, for several of our programs, we're seeing 30 to 60% improvement in recruitment compared to the benchmark rates for the same indication for the same phase studies. And you can imagine, that's leading to, in some cases, studies that were delayed are now either coming back on schedule or ahead of schedule for read up. That has a big impact. And then we're using real world data as we have regulatory conversations to be able to contextualize the evidence for our own study. And so a combination of those things is giving us more confidence. And then the ability to continue to use the technology across the life cycle from, like I said, the protocol design to authoring all the way to submission, we're continuing to innovate in all these areas.
Brendan Smith:
Yeah, that's great. I think it helps really set the stage for just how many opportunities there are to move the needle on a lot of this using AI/ML across the board here. But I guess the flip side of this is also something we spend a lot of time thinking about, and as folks try to understand where there are opportunities to make a difference. At the same time, I would be curious to see after all this work you guys have all done, are there any areas actually where some of these bottlenecks are actually largely insulated from AI? And I guess how do you go about navigating those areas when you're making some of these pipeline decisions?
Sid Jain:
Yeah, I mean, great question. I mean, the first is just the biological time. In clinical development, you dose a patient and you wait. No model can shorten a survival curve. You still have to dose the patient, wait for the appropriate amount of time to pass, in many case 12 months, 24 months before you can have your endpoint readout. You can't shorten that endpoint from a 12 to 24 months into a weekend. Can you make better prediction? Can you predict efficacy better? And this is where the point around patient selection, can you identify the right patient population that is likely to respond? That could have a significant impact, but you still have to do the readout of the trial. Where it does help though is if you can pinpoint the patient population that is going to respond versus not respond, the ability to enroll the right patient reduces the sample size, the number of patients you need to enroll, and therefore helps with the duration of the trial, the cost of the trial, a number of different things.
And ultimately, that's what we all want. We want to bring patients into the trial that are most likely to benefit so they're not suffering unnecessary consequences. So this is where we can de-risk some of this approach, but there's certain areas where the time it's going to take is the time it's going to take, but we can make progress in those areas too.
Brendan Smith:
Yeah, that's great. I think, look, I know you obviously all have your own internal therapeutics pipeline and you partner with a number of really heavy hitters across the industry, and you've kind of laid out a number of things that you guys are all watching for and opportunities to, again, move the needle on some of this. But I guess what are some of the more notable efficiencies or even benchmarks, maybe kind of zooming out, a little bit more broadly that you've observed either as a result of your ClinTech platform or just improvements from other folks in the industry who are all of you collectively leading the charge in this space? Is there anything that folks can watch on a broader scale and turn to see how things have actually already evolved and how they might continue to move forward?
Sid Jain:
Yeah, so the first benchmark I would talk about is still our own. If you look at REC-1245, going from biology to a first in class development candidate in about 18 months, roughly twice the typical pace. I mean, that to me is progress. I mean, we have the molecule in clinic, we have great hopes for it, but that's the type of progress where you're shortening one end of the spectrum. Operationally, near term, there is tremendous opportunity and we're seeing this in several places. There's measurable ROI across the whole industry. The opportunity for on trial execution side, like I talked about, site selection, even projections. Because by using AI/ML, when you make much more informed projection for how long your trial is going to take to recruit and number of patients it's going to take, it has significant downstream impact we're starting to see for CMC, for all the planning.
Imagine you make a poor projection for, not imagine, I've seen that in several studies in the past lives, in some cases enrolling up to 15,000 patients, where if you make the wrong projection, you could either be out of supply or you can have too much supply. And if you're out of supply, then that causes a pause, delays the trial. And I'm starting to see changes there across the industry where we have much better projections as we bring real-time data and more AI/ML into generating hypotheses and projections. And even the FDA's own push towards this continuous real-time trial review could be a general unlock for every sponsor. I know there's a lot of work there still remaining, but we're very hopeful.
Brendan Smith:
Yeah, that's great. I know that real-time clinical trial is something you and I just caught up with a couple other folks involved in that recently at our conference in Dana Point, California a couple of weeks ago. And I guess, excuse me, more on this conversation about your pipeline, other folks' pipeline, and I think really just how investors in particular look at the size and breadth and pace of pipeline progress. And this is something that I think a lot of people are really interested in, but from where you have come over the last few years to today and maybe forecasting a little bit moving forward, do you think as drug developers start to see more and more of these time and cost savings, whether on discovery, early stage or through the clinic, would you expect that some of those productivity improvements ultimately lead to more programs coming through the pipeline or redistribution of capital back into innovation? How should we think about that push and pull?
Sid Jain:
The cop-out answer is both, but that really is true, at least in my opinion. To me, by those gains, if we free up capital, we have an opportunity to take more shots in goal. And as we make these progress in the areas of discovery and better design molecules, do we increase the probability of success of that molecule? And we would want to take more shots at goal. And so by reducing the cost from the system of bringing efficiencies into the system, especially in clinical development, that gives us an opportunity to run more studies and give ourselves a better shot. I mean, this is a probabilistic game. We know that 100% of the drugs that go into the clinic are not going to work. Right now, it's 10%.
So if we can, with better molecules and much more efficient processes, take more shots in goal that have a higher likelihood of success, that's what we would want to do. And that's how we would deploy the capital. And at the end of the day, we're using AI/ML for better ranking and prioritization, but we still all know we have to run these programs for the clinic.
Brendan Smith:
Yeah. And I think that really gets at this idea of what is an appropriately sized pipeline for each company. I think this is also something that investors are constantly coming to us about, is trying to understand the size of a pipeline today versus even 5, 10 years ago and the cost of actually developing a lot of these assets is pretty consistently in flux now and it is continuing to evolve.
Sid Jain:
Yeah, just final point of that, we would have to bring the cost of that pipeline development down in order to have more shots. The point I just made is, if the cost structure remains the same, even with the quality of the molecules increasing, you cannot have more shots at goal.
Brendan Smith:
Right. Yeah, because presumably if you've got even a great drug design, AI/ML discovery engine, and you're really cranking out much faster and much cheaper than you have historically, but the actual cost to develop each of those in the clinic is the same, ultimately you're going to limit yourself on how much you're able to really maximize the ROI and all that early stage discovery. So I think it's feeling more and more like this is an invaluable opportunity to really get the most out of some of what's already been invested in that side of things too.
So I guess maybe from where you sit, and I have a guess of a few ideas of how you might answer this, and what you've seen over the last couple of years, what do you feel now in 2026, what does the investment community likely either underappreciate or just misunderstand altogether about AI within this ecosystem that you think is really essential for anyone listening in to understand?
Sid Jain:
Yeah, I mean, there are a few things. First is, people are waiting for a single breakthrough, like an alpha full moment for efficacy. I don't think it's going to arrive that way. Progress is going to be still sequential, and I do feel discovery improves first, you design better molecules, less toxicity, and over time clinical efficacy. That's going to be structurally last. And the companies that are going to win are going to be positioned for that sequence and not for a miracle.
The other point I would make is, what's most important is, the durable mode is not going to be access to the best public models. These models are getting extraordinary. I mean, we're seeing that in real time, but I don't think they're going to be the differentiators. The edge is going to be the proprietary data, the proprietary workflows, and how we can integrate that with these models and other data sets that are available and create that learning loop. We talk about the lab in the loop ourselves, for every program, every readout. Those who can do that, that's the competitive vote.
And eventually, really these models, the better models won't get full credit until we have proven success across many prospective readouts. Even if we do develop a model that can predict efficacy in vitro and silico. Regulators are not going to drop the requirement for a human trial anytime soon. So there is that kind of valuation overhang where we're all waiting for the proof points, and I do think they're going to come. It's going to take time, but we have to keep at it.
Brendan Smith:
Yeah, I think this is something we hear quite a bit as a theme. It's less of a panacea for all of drug developments woes of which there are many, many, many across the entire spectrum and more, if you think of it as an efficiency tool, how would you think of any past efficiency tool over the past 50 years and what is going to come out of that as well? And if we view a lot of this through that lens, then I think we can start to better understand what the realistic path forward here is.
I think with that, I want to thank you for hopping on and talking us through what really is kind of the cutting edge of this marriage between software and healthcare technology innovation. I'm sure we'll have plenty more to discuss over the weeks and months ahead as it pertains to all of this, but really Sid, thank you so much for joining. Thank you everyone for listening.
Sid Jain:
Thanks, Brendan.
Speaker 1:
Thanks for joining us. Stay tuned for the next episode of TD Cowen Insights.