Recursion Pharmaceuticals, Inc. ($RXRX)

Thanks for attending the 2026 Bank of America Healthcare Conference. My name is Alec Stranahan. I cover Recursion and SmidBiotech at Bank of America. Pleased to be joined today by Ben Taylor, Chief Financial Officer of Recursion. Thanks for being here, Ben.

Happy to be here. Thanks for the invite. It's always good to come out every year.

Yes. Yes. Great. Well, maybe just to start, there's a lot going on at Recursion. This is AI-enabled drug discovery for those who are uninitiated and Recursion is really Trailblazer. They've been doing it for many, many years. I guess, Ben, to start, you've got sort of real value drivers at the company. You've got your internal pipeline. You've got your pharma partnerships, and we've also got the Recursion OS platform itself. I guess, where do you see the biggest rerating opportunity for the company? Is it from the pipeline validation? Is it from advancing and expanding your partnerships? Or is it really from the platform and sort of what you're driving from that.

Well, at the heart of it, we're a therapeutics company. And so we really focus on making sure that we're advancing the pipeline and getting the value drivers from that. I mean if you look at one of the fundamental tenets of the business is to create a more risk-diversified biotech model. And so that has, over the last decade-plus involved a lot of platform and a lot of the partnerships as well. But in the end, it's all for the goal of actually developing therapeutics that are going to reach patients. And so right now, what we have is 5 different clinical programs that all have clinical data in the next 12 to 18 months that's really impactful for showing proof of concept and/or we've got 488 that's setting its registrational trial. And so those are really the core value drivers, not only from an investor perspective, but also just for the core mission of advancing medicines towards the patients. The other components, it's funny because for us, it's a continuous stream. We don't really view the platform separately from the pipeline and the partnerships. We're actually just doing the same work that we would do on our internal programs, what we're getting paid in [indiscernible] it and a nice NPV on it. So all of it really flows back to does this drug have differentiation? And is it likely to be a good product for patients.

Great. And maybe for those less familiar with the company, maybe you can just talk about sort of the full stack capabilities that you have. And how that's evolved over time, where you're sort of at today? And how do you sort of see yourself settling into the rapidly evolving AI landscape.

Yes. If you take a step back and you think about, all right, how can you make a differentiated model in biotech? There's 2 components that you have to do. One is improve the probability of success. If you've got a 90-plus percent failure rate in the industry, you're always guessing. And so that's the part that really focuses on building a better data-driven model to it. If you think about how we continue to advance along and bring up the other pipeline programs, what we've been able to do is show that we can do that across a number of different facets on where that probability of success can come down. So if you think about a program at whole you need to think about the failure of a clinical trial based on the statistics. It could be chemistry. It could be biology. It could be patient selection, it could be trial design. And so all of our systems are actually saying, can it create a better predictive model to be able to solve that reason. [Audio Gap] That patient better. I want to be able to explore the biology better. I want to go through a completely novel chemical space. And so there's not a single point of technology that really defines us. we just have the largest integrated tool set to be able to solve those problems that cause failure for clinical trials. So I think that is a massive differentiator for us, the fact that it's in created within a unified system so that it works together, where most of the companies in the space right now are really developing point solutions for a single area. The other aspect is the data. I mean, because we've been doing applied work for so long, we've created a massive data set over 50 petabytes of internal proprietary data that allows us to build better models. So if you're just using the public data or even if you're just using data that has been generated outside of the ML context, the fidelity of it is pretty low. And so it doesn't actually drive a lot predictability. What you really need to do is create something that supports a better data analysis. We just had a paper published that showed a much smaller data set of highly annotated data actually creates far more predictive models than a massive data set of poorly annotated data.

Interesting. And I wanted to talk about something that you got asked at our breakfast this morning, and I thought your answer was pretty interesting. Just thinking in the near term, the industry is evolving in terms of where innovation is being rewarded, right? Typically, that was through pharma partnerships like what you have, a lot of the pharmaceutical companies are going to China now as well. I guess on the AI drug development landscape with like alpha fold kind of established now. Where is the next sort of white space for development in the next couple of years?

Yes. Well, it's just -- it's absolutely massive. So if you think about all of the pharma industry, all of the biotech industry for as long as they have both been running. And all of the approved medicines and all of the drugs that are currently in clinical trials, you only cover a little over 10% of the genome. And so that means almost 90% of biology is sort of in the dark to us from a therapeutic perspective. And the chemical space is actually in a similar state, if not even worse. And so first of all, there's a lot of white space that we need to explore. And that's part of why we need new techniques be able to evaluate it. But I really think we're going to continue to see evolution is one, over the next couple of years, well, where we are right now, people understand that patient selection is important, right? If you get the right patients into the trial, you can increase probability of success. And obviously, we've got a lot on that in Quintech. What I think people are just starting to understand is the chemistry space, which is really potency and even selectivity are only a small piece of the puzzle because realistically, how is it absorbed? How is it metabolized? What are the other interactions that you're dealing with? And so we, in legacy Exentiahad predicted multiple clinical trials that failed before there was any data on the clinical trial, just looking at the chemistry. And so I think people are just starting to understand, okay, a potent compound that works in mine, which by the way, a preclinical model that's basically designed just to focus on potency, does not actually taking to picture the entire patient environment and all of the biology. And so understanding that side from chemistry, I think, is just starting to come up. And then the biology side of it. People still think are very linearly. ABC1 connects to XYZ2, and that's how biology works, which is versus not. It's incredibly kind of multivariate problem. And so I think we are just starting to get the tools to be able to explore that better and be able to get into areas that are far off that what I was talking about earlier was really exciting Virtual cell is obviously a big word, a lot often used. But the publication that we just had on Nature Biotech, basically, what it showed is for a number of different cell lines, we were able to experimentally predict what would happen in that cell without having a B at all part of the training data. And so we're actually getting to the point where you can use multimodal biology. So cellular phenomics, which was recursion was originally founded on, combined with transcriptomics, combined with other different aspects that you can bring in. And all of a sudden, that signal that cloudy in any single medium becomes much more clear and you can make predictions on actual underlying biology even without having original knowledge. Still a long way to go there, but we're just tip of the iceberg on it.

Yes. And I mean, the way I think about alpha fold is you had a good data set, and there's not that many. I mean, there's a lot of different ways that proteins can fold and interact, but it's not as complicated as like neutrons and electrons in like small molecules. So like the multimodality of that made it an easier first step actually, which was a little bit counterintuitive. Is the right way to be thinking about it that you're just going up in orders of complexity going from proteins to small molecules and then to a virtual cell.

Well, it's really interesting. I think there is going to be fits and starts across all the categories, but you're absolutely right in a way, like the alpha fold, which, by the way, completely amazing and fantastic discovery, but it was based on 30 years of imitated protein data that existed in the public space, so that there was a much larger base to be able to build from. And then being able to put that together, you're basically looking at physics principles and how that combines together and it becomes a computational statistical problem. With chemistry, you're talking about -- so just to put it into context, like if you take all of the antibodies, for example, that are known and out there, you're talking about 10 to the 15th about. If you take all the potential medicinal chemistries, you're talking about something around 10 to the 60th, so effectively infinite from what we're thinking about. The biology, obviously, you're interacting between all of those different pieces. But I think what's really important is helpful itself didn't solve all proteins. It may be far better at estimating and predicting what then you can prove out experimentally later on in proteins that are not well known, but the prediction capability is far higher in areas where there's a lot known about proteins. And the same is true in chemistry, like generative systems work better around areas where there has been a lot of data created. And so what you'll see is there's going to be aspects of biology, let's call it, the most complex because you're now combining all of those different [indiscernible] or whatever. And you're going to have amazing breakthroughs in that but they're going to be around areas that you can build off of, just like we're continuing to advance in proteins and continuing to advance the chemistry.

Great. That's really interesting. But yes, maybe we can move from the theoretical now to the practical because this is...

You got me going.

Yes. I think it's super interesting. But in terms of what moves stocks, including Recursion potentially the tangible aspects of what's coming out of the platform and for you guys, that would be your clinical and preclinical program. So maybe we can just go down the list because you've got a bunch of interesting programs in our pipeline. Maybe starting with 4881 in FAP. For those less familiar, maybe just talk a little bit about sort of what this disease is. There's really no approved therapies here, although there's maybe some context that you can put 4881 into that makes sense from a biologic perspective, and -- so what's the clinical course for these patients? And what are you trying to solve for the treatment?

Yes, really, really tough disease states. So this is usually identified hereditary though, there probably are a lot of somatic mutations that cause us as well. But 50,000-plus, and that's primarily the hereditary population, who throughout their entire life as long as they live, they're going to continue to get hundreds or even thousands of malignant polyps throughout their GI tract. And so this usually starts up in adolescent 13s. Patients who usually have a colectomy before their 30, and then they'll continue to get further and further resections as time goes on. They're likely going in and having excisions of all those polyps can be on a several month basis. So really, really difficult disease state, very, very high comorbidities and completely chronic. So what we've been able to show with the data is within 3 months of treatment, the pallet burden was reduced by about 50%. And then we were able to actually also take patients off of drug and maintain that pallet benefit, which is really, really important for a chronic disease drug. Those aspects basically gave us confidence that we would be able to move forward with a registrational pathway. And so we're currently in FDA discussions on starting the pivotal trial, and we'll give another update on that later on this year. But huge unmet need, there are no approved therapies for it. And we're really excited because that was a discovery that came out of our Bio AI platform. It was a novel connection that we were able to look and say, okay, how can we potentially reverse this [indiscernible] mutational effect with a drug and then convert that drug into now patient proof of concept.

Yes. And I guess as we're thinking about the translation of [indiscernible] to functional outcomes for patients, I know you've got a pretty substantial kind of natural history set that really helps you understand sort of the patients with this disease. I guess -- how should we kind of bridge that gap between the 50% polyp burden to functional endpoints, like, I don't know, CRC prevention in others? And what's the most important for the FDA, I guess.

Absolutely. Well, so every single one of those polyps is precancer. And so basically, if left untreated, all of these patients will progress into colorectal cancer. And so that's obviously one of the end points. But there's a number of different end points that you can look at with the FDA. There are things like there's a Spiegelman score, which is sort of a composite of polyp burden and dysplasia that leads clinical assessment and treatment there are excisions because when -- without FAP goes in for colonoscopy and gets polyporremoved. It's painful, it's uncomfortable. When you're going in every 3 or 6 or 9 months and having 100 polyps removed you're talking about massive excisional burden, bleeding infections, quality of life changes. And so you can look at those pieces. You can look at resections and serious surgeries these polyps actually can go up into the duodenum and you can't do like a colonoscopy to be able to remove those policies, actually need a different endoscopic procedure, which is a far more serious procedure. So there's a lot of different true clinical endpoints. If you talk to the clinicians, they view poly burden as being incredibly important because it basically determines how they stage patients. The other thing that we're able to do, and this is sort of an example of how we have the infrastructure because we have the supercomputer because we have AI scientists because we have access to 300 million patients of real-world data, literally in the course of a week, we created an LLM that looked across those 300,000 patients, found 250,000 patient records associated with FAP created something that we could query and say, okay, what is the standard of care in the setting, how many surgeries are these patients getting? The answer is 10, by the way, across their lifetime of serious surgeries, not counting all of the colonocopies, and that gives us a lot of insight that never existed or that we can also take to the FDA and talk to them about the pathway.

Yes. I guess how central is the real-world data package to your FDA discussions? And does this reduce the likelihood of a randomized control trial.

Yes. Well, I won't get in front of the FDA discussions, certainly. But what I'd say is it's definitely informative, definitely appreciated. I think even if it doesn't -- so the FDA loves data and loves transparency. Anyone who thinks differently hasn't worked with them, they love it. But even if it doesn't make a difference with our FDA pathway, it makes a big difference with how we design that trial, which patients we go after. We can do things like take a patient population and change our inclusion, exclusion criteria and say, okay, first of all, how does this change my patient population, commercial opportunity across the U.S. and Europe. And then we can say, where are the sites with patients that have this sort of demographic and then we can build in there. That's why we've seen a 30% to 60% improvement in our enrollment rate across our clinical trials where we're using our clin tech. So it makes a big difference.

Yes. And that should pay dividends across the pipeline when you're finding new studies. I guess on RBM 39, this is your molecule that can, I guess, [indiscernible] to shutting down CDK12, we're expecting clinical data in the first half, I believe, what's sort of the first in human data that you need to show to give you conviction to advance? And I guess, how quickly could you make this go on the [indiscernible].

Yes, absolutely. So we RBM 39, just to take a step back, was a program that we founded with our biologics platform is basically CDK12 has been a very interesting oncology target for a long time. It's transcriptionally related. And so there's obviously a lot of opportunity with high mutational burden in cancers, but incredibly difficult to target because the pockets of CDK12 and 13 look the same, you get a ton of side effects with CDK13. What we were able to do is say, how can we identify the same biology and another target, and that was a novel connection. And then what sort of drug could then reverse that disease impact. And both of those were discovered phenotypically on our platform, and then we optimize the molecule. So what we have been enrolling in is looking at patient populations that generally are going to be higher in high like MSI high populations, there are different areas that are going to be potentially more susceptible to this type of drug. And we wanted to do like an early look, almost like a futility analysis. And saying, okay, this is a completely novel target that no one's discovered. This is a decorator, which has its own benefits and potential concerns and we're doing early dose escalation and trying to be really efficient with all of our money. Sure, we'll talk about that at some point in to. And so what we wanted to do is take a look, and this was -- the first look we did was last week, on our earnings call. So basically, we had -- we showed that we had PK/PD that was dose proportional. We had the profile that we wanted to see, which was basically this is a protein target that you'd need to knock down probably 70% plus on a pretty constant rate throughout the day, and we were seeing that exactly how we had designed it to happen, which is great. And we're just getting close to what would be the predicted therapeutic doses. This is monotherapy, but there's potential to see signal coming out of monotherapy because if it was cancer type that exposed to synthetically lethal mechanisms or other things like that, we could potentially see a signal. So that's where we're going next. The next [indiscernible] more thorough data set second half of the year and it could be really, really exciting. That's actually a very large addressable patient population has a different sort of splicing and stability disorder.

And I guess on that point, we'll be looking to see which tumors you select. I guess, is there a decision tree that you make either around which indications to pursue like unmet need vis-a-vis market opportunity? And then how do you make the decision to mono or combo.

As you will respect, we are incredibly data-driven. So we do a lot of real-world profiling and statistical analysis around that. We'll also do that for our commercial analysis, and we can do both of those aspects in-house. And then we're going to be looking scientifically at the data. We've got series based on what we see coming out of the Bio platform and what we believe to be true, and we're going to be looking in the data and say what holds up and what doesn't.

Okay. And maybe shifting over to 7735, which you announced, I think, earlier this year. This is your mutant selective PI3K inhibitor, obviously a pretty hot space right now with [indiscernible] and others, relay, et cetera. I guess here, I guess you have the potential to address some of the safety shortcomings with the hyperglycemia and dermatitis that's been seen. Is that really the leg you stand on from a competitive perspective? Or is there something else you expect to gain from the molecule? And I guess, what is sort of the IND-enabling data you need to show to persist into the clinic?

Yes. Well, I mean, there are other things than the hyperglycemia, but that's a really big one because this is always the hidden secret and oncology drug as you see the results come up from a clinical trial. You should always go back and look at the inclusion exclusion criteria. So most patients right now, and we did this on a real-world analysis as well. I mean, if you've got diabetes or prediabetes, which is actually in a lot of the cancer indications, around half of the patients. You're not going to get this drug or you're not going to get a PI3K that are currently out there or you are likely to come off at very quickly. So the real world times of being on, for example, Piqray is only a couple of months, and that's because of side effect profile. Even if you don't have diabetes or prediabetes, Grade 3 hyperglycemia is a very serious issue. And so that is a really big important part. I think what also interesting is by having -- so we're seeing 100, 150x selectivity over wild type. Now what that means is an ability, we believe, to drive deeper into the dosing of it for the mutants. And so I think there's 2 levels of the question. One is that orphan population that is not eligible for these therapies right now because of some sort of glycemia issue. But the other aspect is, can we go deeper on dose to drive higher response rates in this population before you start to get to MTDs. And so we've seen a lot of drugs come up in the space. Our original thesis on this continues to hold. As far as we can tell, we are about an order of magnitude more selective than the drugs that are currently advancing crew development. And so we like this drug.

So we'll stay tuned for the data from that program in the second half. I guess I want to ask about another arm of our business, which is partnership. The 2 main ones that we hear talked about are Sanofi and Roche. And maybe we can roll this into sort of the capital allocation burn question, which is -- how could milestones potentially bolster your cash position over the next 12 to 18 months? How are these partnerships going? And how do you sort of see the allocation between your internal development servicing your partnerships and good stewards to capital.

Yes. And I'm going to back it out into the capital allocation question as the core part is there because I think -- as I said earlier, we are a therapeutics company, and with the long-term value is going to come out of the drugs that we developed. And so if you think about how we invest in platform and what we do with our partnerships, it's got to be towards driving that bigger goal. And so first of all, on sort of the expense management, what we did post the merger was take a look at everything across the company and actually put a analytical framework to measure impact and basically said, anything that we can't clearly see high impact from or high probability from, cut it. That literally took out 35% of the budget over the last year. And so we continue on with that framework now expense guidance for '26, even with the 5 clinical, 2 preclinical, the big partnerships with Sanofi and Roche. It's still -- so it's less than $390 million. We're focused on getting to that less number because we are, at our heart, a tech and efficiency company. We should always be getting more impact for less cost. We're going to continue doing that. We're also going to continue to be really disciplined in how we make decisions. So right now, most people don't realize this, about 2/3 of our cost, almost 70% is going directly into pipeline programs and our partnerships. So there's not some big 30% of the budget going into platform development or something like that. That's not how we do it. It's applied platform and pipeline development. And so what we like to be able to do is say, okay, is the pipeline advancing, is the partnership advancing? And if it's not, I'm going to cut all of the associates spend with that. If it is, great, because it's going to be generating value on its own. And so from the partnership side, we've brought in over $500 million from those partnerships. Just in the last about 2 years, we've hit 7 partner milestones, 5 with Sanofi, 2 with Roche that continue to bring in more money. And over time, that business, right now, we run it basically to the breakeven to a mild profit because our partners prepay us for our expenses as we get to opt-ins on Sanofi as we get into advanced states on Roche, actually, that becomes all profit that rolls in on a go-forward basis because we have no more operational obligation.

Okay. Great. Well, there's obviously more we could talk about, always a lot going on at Recursion, but I think for time, we'll leave it there. So thank you, Ben for the great discussion, and thanks, everyone, for your interest in Recursion. Thank you.