Recursion Pharmaceuticals, Inc. ($RXRX)

Ben Taylor, CFO of Recursion, here with us. Thanks so much for joining us.

Maybe to start, if you want to may provide an overview of your platform, how exactly it integrates AI in your process? And what are the key points of differentiation that you have versus other AI-enabled biotech companies?

Sure. Great. And first of all, thanks for having us here, Tommie. This is always a terrific conference and great to be here. So taking a step back, if you think about what Recursion does as a whole is we combine different ways of predicting models, whether it's biology, chemistry or ClinTech. And then we'll look at a program and say, why might that fail? Could it fail because we don't understand the biology? Could it fail because we don't understand the chemistry or maybe it's the patient population or how to design a clinical trial. And then we try and take those 3 different pillars of the platform and say, can we create a better data-driven model that's going to be able to get us either to solve it in drug discovery or in clinical development and predict that failure point. And so what we've built now is really going beyond our own data generation, going beyond our modeling systems and really focusing on the proof points of that because we're not in a industry that creates models, we're in an industry that creates drugs. So right now, what we have are 5 programs in the clinic, all of which have data over the next 12 months. A couple more coming behind that. We've got a partnership business that is primarily focused on Roche and Sanofi, but we've brought in over $500 million from those partnerships, and those are really high-value, high NPV programs. And underlying all of that is an AI-based platform. And so this is where we're AI native in the sense of we don't have infrastructure that we're replacing to put AI in place. We actually look at every single problem, and we say, is there a better way that we can do it using data, using AI, using computational power to be able to solve it differently than it's been done before. And so the goal is to improve the probability of success. And because we're using technology, we do it generally faster and cheaper. So we've published a number of the statistics, but we've been able to really bring down development and discovery time lines and the cost of being able to do that as well.

Yes. And at this point, it's been over a year since your merger with Exscientia. Maybe at this time point, how you're thinking of the key priority for the company, both on the pipeline, BD and platform side?

Yes, yes. So Exscientia came in with the chemistry background to Recursion's core capabilities in biology. And what we've done since then is be able to combine them so we can identify novel targets and then we can design the drugs that are optimized for those targets. We green shielded a ClinTech business about 2 years ago. So Najat Khan, our CEO, when she came over from J&J, she was heading up the AI business as well as the portfolio review of all of J&J's programs. And she had created inside of J&J a ClinTech business that used a lot of real-world data. And so we brought that in and built it up from scratch. And it's actually had a massive impact across all of our programs. We've been able to -- even though it's a relatively new part of the platform, we've been able to speed up our clinical trials by 30% to 60%. That's across multiple different pieces as well as identify lots of new sites in different patient populations. So the platform has actually changed a great deal. Our capabilities have changed a great deal. I think underlying it, there's also a cultural piece. So maybe it's because of the chemistry and biology focus, but Exscientia had always been deeply analytical, really focused on getting to that exact solution. And we ran a business model that was very associated with that as well. Recursion was always more big picture, driving towards the vision and trying to be really transformational across the industry. And so those 2 have actually come together really nicely. So we've done a lot of work on bringing in the cost structure. It's funny because we took 35% out of the budgets since the merger. And that wasn't actually what I would call cost cutting. What we did is we implemented a program that measured impact of literally every dollar that we spend. It doesn't matter if it's pipeline or tech or G&A and said, what's having the high impact and anything that didn't meet the bar was eliminated. And so it's really more of impact modeling. And we continue to do that, and we continue to bring down efficiencies in that way while still trying to drive towards a pretty broad clinical pipeline and partnership business.

And we think you'd be quite active on the BD front. You brought in new technologies for the platform. You have pharma partnerships, you had an NVIDIA partnership. So maybe what are your thoughts on the overall strategy here? And where are the future directions for activity as we think external?

Yes. Well, and this is actually a really important point that I think a lot of people miss. Our original mandate was not just to change the probability of success. Basically, the investors that funded a lot of the original work said, I want to invest in biotech, but I don't want to invest in binary risk. And so I don't invest in biotech. So how can you create an actual ongoing business model. And so through good times and bad over the last many years, we've held to that and said, we need to be able to bring that risk diversification in. And that's why you see in our pipeline, we have best-in-class, we have first-in-class. We've got a real nice mix. But in our partnerships in our BD as well, where we look at it and say, how can we get a balance of different partners that all contribute. We run that partnership business actually at breakeven or to a profit at all the times. So they pay us in advance for our direct costs as we get to in-licensing or profit that flows down to us. So that's been a really nice backbone. The other part of BD, we haven't done yet, but we are set up to do is basically out-licensing of our own programs. And so if you look at our pipeline, the 5 clinical programs, the 2 in preclinical or in IND-enabling, it is not a part of our business plan to take all of those programs ahead ourselves. And so we want to look at the data. We want to see how they advance. I'm sure we will always have 1 or 2 at least that are our own wholly owned programs. But we're also very open to out-licensing and other aspects like that. We just want to get to a really valuable transition point. And so the way that we look at it is it's all option value. We can do portfolio management. We can do capital management because we have this diversified business.

Yes, that makes sense. And so maybe let's turn to the pipeline. If you'd like to give an overview of the data that we'll see within the next 12 to 18 months before we get into some more detail.

Absolutely. So we've got a number of data points, both towards the end of this year as well as the beginning of next. So this year, our most advanced program is 4881, and that's in an orphan disease called FAP. So we're currently in FDA discussions on the registrational trial design for that, and we'll give an update on the registrational trial design. We are also continuing to enroll a broader patient population than the original data cut had been on. And so we'll put out that data probably first half of next year. But those are 2 major points, especially on the FDA registrational trial design. Everybody is waiting for that on FAP. And so we'll update that as -- the pipeline a program called RBM39, which is a really interesting example from our biology platform where we looked at the biological mimic of CDK12, which is a known good oncology target, but almost impossible to drug cleanly. And we saw that RBM39 had a very similar biological fingerprint for it. And so we created a degrader molecule to be able to go after that. We've done the initial dose escalation, which we just talked about on the last earnings call. We saw a nice clean safety profile, dose-proportional PK/PD. We didn't put out a lot of data. I kind of think of it almost like a -- we did a Phase I futility analysis because novel target and a degrader system, just take a look early and make sure everything is going the way that you wanted to. It was. That has a lot of potential, and we'll probably talk about it later because it really targets genomically instable cancer mutations. And so that's a pretty broad number of indications. More data on that later on this year as we continue up into that escalation. And then looking across the rest of our pipeline, early next year, we've got MALT1 and CDK7. Both of those will have data readouts. MALT1 in monotherapy, CDK7 in combination therapy. And then second half of next year, we have an LSD1 program that we just started up as well.

Okay. Great. So on your lead program in FAP, you're discussing population and trial design with regulators. What are the potential scenarios here for what you could announce in the second half? And as you think of ERAP's path in their trial, what could be the potential similarities and differences?

Yes. So ERAP set a template of what we could do in this area. So this is something there's no approved therapies. And there really wasn't a lot known about the disease as well. And we've published some of the natural history data that we've been able to do through our ClinTech platform that's really educated not only ourselves, but also the industry on this is what the management cycle, the health cycle for a FAP patient looks like based on all of the data. And so what we want to do is take that to the regulators, take that to ourselves, be able to do analysis and say, this is what's important. This is what patient populations look like in different settings. These could be the higher risk aspects. Talk to them also about different potential endpoints. Now it would be fine if we went with the same endpoints that ERAP had used. That -- it's not a bad trial design. It's just what we want to do is look at other areas where there wasn't as much known coming into it. So excisions is a really important part. This is -- it's not as invasive as obviously a colectomy. But when you've got literally hundreds of polyps throughout your gastrointestinal system and they're doing excisions on it every 3 to 12 months to remove those polyps and then they keep growing back, there's major bleeding episodes, there's infections. There's all sorts of problems in being able to digest and absorb. And so the comorbidities that go along with just that, which is only out of a colonoscopy or other endoscopic treatment are actually very substantial. And that wasn't included, for example, in the ERAPA trial. And so we want to ask questions like that. The FDA has been -- we haven't noticed any differences in any of our programs from the changes that have been going on at the FDA, but we're also not going to get in front of it, and we want to have the discussions.

Yes, makes sense. And you spoke a bit on the ClinTech platform that you have. Maybe if you could just lay out how it was used in FAP and how it's different than if you had used a more traditional method.

Yes. So there were 3 ways that we used it. 2 of them on the natural history side. So one, we took -- there's a university in Amsterdam that had been doing for 20 years, they've been tracking FAP patients. So it's the longest running database of FAP patients. And so we were able to take that and convert that into something that looked at, okay, what's your average polyp burden growth? What are the patient demographics for this look like? How does progression look and be able to compute that into something that showed on average, you've got about a 60% increase in polyp burden per year per patient, which is just a massive amount of increase. No one had ever actually shown that polyps continue to grow, right? Like before that, you couldn't say they will spontaneously reduce or grow, right? And so we've pretty definitively shown which way that goes. In addition, -- we have a database of over 300 patient lives that include not only claims data, but also all of the physician notes and those other aspects. And so what we did is we looked at all of the physician notes and found all of those that were related to FAP patients, it was about 250,000. And we -- in the course of the week, because we have our own coders, we have our own supercomputer, we were able to create an LLM that allows us to query in the same way that you would with Claude or Gemini, how many surgeries are these patients having here? What are their likely comorbidities? What are they getting treated with and really understand the entire patient journey in a much more detailed way. And there is no standard of care manual for these patients. And so we were helping to define that based on what is actual clinical practice. So those 2 were completely novel and just play off our strength in data analytics and AI. What we also can do is basically run the simulated trials. So this is something that there is a classical history of being able to do clinical trial simulations. People do too little of it. They should do a lot more. But what we do with all of our trials is we'll take the patient population. We'll change the inclusion/exclusion criteria. We'll look at different combinations or comorbidities and see how it affects the outcomes of the trial by basically being able to run different statistical analyses around it. And so we can better target what does a good trial look like, what does our statistical package look like? How can we do a more informed data-driven trial design.

Interesting. And then if we think about the FAP market, how big could this opportunity be as you think of prevalence, what your label could potentially include and the proportion of patients who are seeking systemic therapy?

Yes. So U.S. and EU5, there's about 50,000 to 70,000 depending on source, patients that are identified with it. That is mostly the hereditary population. So most FAP patients, at least most known FAP patients have germline mutations that are leading to this. And that's why they are diagnosed usually in adolescents or their teenagers. And so what we look at is, one, how can we treat those patients. There's no commercial therapy out there right now. And so their only option is surgical ones. So they either get resections. They almost always have a colectomy by the time they're in their 20s. And then on average, have 10 different resections over the course of their lives as they go through. That doesn't include all of the colonoscopies where they're having the polyps removed on a 3- to 12-month basis. And again, these patients literally could have hundreds or thousands of polyps throughout their colon. So that's very severe. What we look at from the market is almost all of those patients could benefit from this because the polyps never stop. They're all precancersous. So literally, every single one of them is a precancerous. There's no benign polyps in FAP. And so they have to be removed at some point. And so being able to control and manage the polyp growth and burden, there's something called Spigelman staging, which basically dictates to doctors when they need to take it out, and it's a combination of polyp burden and dysplasia. We're happy to see 40% of patients even in just the 3 months that we treated them for, 40% had a downstaging of their Spigelman scores, which is really exciting. So what we want to do is get in there, look at this as really a chronic care market. Most, if not all, of those patients could benefit from a ongoing chronic therapy. There's probably also a pretty heavy underdiagnosis of the somatic mutations as well. So you can get this somatically. And currently, about 20% to 30% of identified patients have somatic mutations that have led to it. It's just the APC gene going off. The question mark becomes how many more patients of those are just being caught when they get colorectal cancer? Because if it goes undiagnosed, they will get colorectal cancer. And so there should be some portion of that, that could also be treated earlier.

So maybe this is where we segue to the oncology portfolio. As you said you reported data from RBM39. What is giving you confidence in the therapeutic index that you're showing? And what -- when we think of what you're going to report next, what will be the key learnings from that next data set? And over time, how are you thinking of potentially making this a combination opportunity?

Yes. So RBM39, really interesting program. So CDK12 is a very common transcriptionally driven issue with oncology and could be a great target. RBM39 as a transcriptional component is able to also have a similar effect. And so what we've looked at and we've shown preclinically across dozens of different models is that whether it's DDR, some of the cell cycle issues, basically any of the different types of mutations that have that genomic instability could be affected by this. You get that synthetically lethal mechanism of action or you at least weaken it down. If you look at how you could potentially use that, what you would want to do is be able to have that synthetically lethal component ongoing for a long period of time. What we got excited about in the data is the preclinical model said we'd need about a 70% or 80% IC70 or IC80 on a pretty sustained basis. So this is a protein that actually resynthesizes quite quickly. And within 8 to 12 hours, you can actually regain back to normal function. So what we needed to see was an extended period of being able to have that -- and so what we -- we haven't released all of the data because we wanted to gather more before we put it out, but the profile absolutely looks like QD dosing would be appropriate. We had no DLTs in the run-up so far. That's encouraging from 2 perspectives. One, obviously, you don't want to see the DLTs, but also this is a degrader. And degraders have historically had a lot of problems with off-target toxicity. And so not seeing that is very encouraging. And then we also saw a dose-dependent PK/PD profile. So we're seeing the systemic distribution, and we're seeing initial biomarkers on the PD that look like we're all heading in the right direction. So in the next couple of doses, we'll be getting into the predicted therapeutic window for -- from the preclinical models. And we'll report more on that second half of the year.

Okay. Great. And do you think this is a mono or a combo opportunity overall?

I mean anything in cancer is a combo opportunity, just how it's used clinically. From a mechanistic perspective, if you had a genomically instable cancer, absolutely, you should see some activity. So what we tried to do is it wasn't a biomarker targeted Phase I/II. It was a biomarker enriched. So what we went is we went after indications that have a higher prevalence of genomically instable cancers. And so we do want to get some people who have genomically stable cancers because you always need the negative in data as well, but we'll also be looking at a number of different biomarkers that could indicate where this is the best patient population.

Okay. And you have CDK7 data in first half of '27. What's the success for this data set look like to you?

Yes. CDK7, it's one of those big -- it's a high-risk, high-reward target. So it's a master regulator around a lot of cell cycle and transcription. So you're really talking about a fundamental biological mechanism -- we've seen the CDK 4/6s, which basically target one stage of the cell cycle. That's a -- I know it's north of $10 billion drug class now, but it doesn't have the durability that you'd like to see. And part of that is because you're targeting that single point. And so the cancers are very good at getting around a single point aspect. And so CDK7 could be a better mechanism because you're targeting basically 2 stages of the cell cycle as well as some of the transcriptional elements. And the real question there is, can you manage tox? So what we did is we designed something that had a very short half-life that was reversible that basically could get in quickly hit to a high level, anything that was a rapidly dividing cell and then get out of the body. So trying to avoid the heme tox is a big area that you'd see or the GI tox. Also, previous efforts had made things that were very much transporter substrate. And so that's going to be very counterproductive because you're going to have a lot of GI tox and you're also going to have trouble staying in the tumor microenvironment. So these are examples of how we could use the design platform to say, I only want to be in the system for 6 to 8 hours, and I want to go into a novel chemical space that's not going to have that substrate issue. A number of different profile properties that we were able to design around. In the Phase I monotherapy, we saw that the design -- we saw all of those things that we had designed for happened. Now the question is really, okay, in a combination environment because this is more cytostatic than cytotoxic, in a combination environment, do I see efficacy, too? And can I do that in a manageable toxicity window. If that data is good, massive potential for it, but we'll have to wait and see.

Okay. So for PI3K, a very active market. You have Astra, Lilly, Relay Therapeutic, others. What is the niche that you hope to carve out with your program?

Yes. Well, and this is a funny one because it is a niche, but it's a very big niche. So about half of the patients that are suitable for PI3K therapy are prediabetic or diabetic. And so that means that anything that's causing hyperglycemia is going to be counterindicated against it. So far, all of the PI3Ks that are more advanced or on the market are causing some level of hyperglycemia. And so that was the core thesis going into it. Can we create something that is ultra-potent and selective because the hyperglycemia is coming from a selectivity issue. So if you look at like the Relay or Scorpion compounds, they were about a 10x selective over wild type for the PI3K. And so you get a reduced hyperglycemia compared to some of the -- like a Piqray, for example. But what we wanted to do is say that hyperglycemia is a sign that you're not just focusing on the mutant. And so we took the most common mutant, which is 1047, and we created something that's about 120, 150x selective over wild type. And it's been incredibly clean. We've had great preclinical efficacy with no hyperglycemia signals. So in one hand, it's a niche to go after that. On the other hand, it's actually a really large patient population. We put out some of the numbers. It's about 11,000 patients in breast cancer and about 20,000 patients outside of breast cancer.

And you have a MALT1 program. There has been some historical challenges with tox here. How are you thinking about addressing that?

Which program...

Your MALT1.

MALT1, yes, yes. So MALT1, another -- it's another one of the design stories. If you look at the compounds that were with J&J and Schrodinger, what we saw is basically, there's a UGT1A1 inhibition, which is not actually a part of the MALT1 mechanism, but causes hyperbilirubinemia. The issue with that is you would almost always use a MALT1 in combination with a BTK or a BCL-2. Both of those have liver tox. And so if you combine liver tox with hyperbilirubinemia, then you get Hy's Law. And so you can't use those together, right? Or you're going to be triggering a lot of really serious toxicity. And so that was the design. Can we take that out? Because our thesis was that was not an on-target tox, that was an off-target tox. And so what we showed in our preclinical is we didn't have any UGT1A1 inhibition. We got into a very novel chemical space. We had great potency and selectivity. And so now that's in dose escalation for B-cell malignancies and could be some really interesting data. That's another one where you would also see it very early, right? Like you don't have to do a big trial to see are you hitting UGT1A1 or not? You'll get it in a handful of dose levels.

Okay. Is there anything else from the pipeline that we missed that you would highlight?

No. I mean we've always got different things coming along. I think getting back to that business model point, one of the real cores here is we wanted to -- and this was also post the merger. We wanted to take the best of the best from both companies. And so we actually killed a number of programs, both clinically and preclinically because we looked at it and everything had to pass a multiparameter analysis. We had to say, okay, good science. Hopefully, it has that. But is there a good clinical pathway? Does the data support it? Is there a good commercial opportunity? And then can we do it ourselves or not. And so everything went through that, and we as we look at those programs, current clinical or IND-enabling programs has the potential to be a blockbuster, has the potential to really make a broader impact. The ones that are more design stories, we've also already seen some biological validation like we're just talking about with MALT1, for example, or LSD1 is similar, where there has been clinical evidence of those mechanisms modifying disease, but they come with a lot of tox. And so can we take the tox out or trying to diversify that risk and say, let's go for some really novel biology like around FAP or RBM39. And so we're trying to just balance that all together, take the best of the best. And we definitely don't need them all to work, but hopefully, a lot of them will.

And then as you think of your cash runway, maybe you can walk us through some of the assumptions baked into that and how there are also considerations between how you recognize revenue from your partnerships versus look the OpEx from those?

Sure. Yes, always a point of confusion. So cash runway out to early 2028. And we've actually been able to extend that out a couple of times just by better cost management and efficiencies. And we're going to keep focusing on that and keep trying to drive it down, really focused on we're a technology company. We should be doing more with the dollar tomorrow than we are today. So early 2028, the only assumptions besides our operating costs that we have in there. So our operating costs include advancement of all 5 of those clinical programs plus one of our IND-enabling programs and we assume that those go out. So obviously, that will be directly correlated. On the partnership side, all that we've included is probability weighting of inflows from our existing partnerships. So there's no new business development. And we haven't gotten into how we do probability weighting, but what I'd say is we try -- even though we always hope to exceed industry benchmarks, we use industry benchmarks for our own internal modeling to try and take a more conservative view. And that's how we've modeled out the potential milestone payments from the partnership. And that's all. There's nothing else in there, trying to keep it clean.

And then just maybe one last one on the platform. As we look at the AI-enabled drug development space, we see a lot more activity on the preclinical discovery side versus the clinical side, and you've been doing a lot of work on the clinical side. Maybe if you could just speak to how this shift has happened and what the challenges are and kind of your unique position here.

Yes. Well, and that's -- it's actually really amazing how the story has transformed because we were absolutely -- both companies started up as basically a point solution. But what we found is you don't get to a 95% failure rate from a single point. You just have failures all across the spectrum. And so what we do now is we actually do a lot of applied development. How we ended up with our 3 pillars, the biology, the chemistry, the ClinTech is we looked at how programs are going to fail -- and what do we need to do to try and prevent that. And so that just naturally led us to this broader platform. Now the interesting part is each one of those individual technologies that comes together into those pillars has to be independently validated. The only way to do that is actually working on programs. And so one of our big advantages is for 13, 14 years now, we've been doing applied development. So most people don't realize this, but about 2/3 -- a little over 2/3 of our budget is applied. So it's directly going into pipeline programs or our partnerships. And so that's how we like to develop technology. We -- only about 10% to 15% of our budget is actually on what a lot of people would consider sort of the research, the basic research part. And the rest of it is really, I want to understand that this model works because I'm using it in a program and it's making a difference. And so you do that year after year with a lot of programs with big partners who say, yes, this makes sense or it doesn't. And you end up getting this very large base of validated platform capabilities. Now we can look back and say, what's the problem? Do I have a tool in my chest that works on that? And what's more so is we have loved all of the development going on in agents. So we've got partnerships with a number of the large tech companies looking at we can deploy our own agents or agents that we bring in from the outside very quickly because all of the fundamentals that an agent would want to use have already been built and validated inside of our system. Like an agent will never fix your broken system. You have to have a working system that an agent can then run on. And so that has been a force multiplier for us of bringing even more efficiency in than we had before.

Yes. Well, thank you. I'm looking forward to all the progress.

Thanks. I appreciate you having us again.