Recursion Pharmaceuticals, Inc.

Good morning, everyone, and thank you for joining us. Before we begin, I'd like to remind everyone that today's discussion will include forward-looking statements. Next slide. Please refer to today's press release and our SEC findings filings for additional details. At Recursion, our mission is to decode biology to radically improve patient lives. and we do this by building transformational medicines with an AI-native product engine. Over the past year, we have reached an important inflection point. We are no longer just discussing the potential of our platform. We are demonstrating the ability of our AI-native product engine to generate differentiated programs and medicines. Just as a reminder, the engine you see on the left hand side is built as a continuous learning system. proprietary multimodal data created in our data factory powers frontier AI models. And these models then generate new hypotheses where every single prediction is tested experimentally. Each cycle strengthens both the engine and the products it creates. Ultimately, though, the measure of any engine is its output. So let's talk about that. First, our internal pipeline continues to mature. We now have five clinical stage programs, including REC 4881 and FAP, where we have generated some of the most promising clinical data in the company's history. Remember, in a disease with no approved therapies and a TAM of almost 10 billion. Second, we continue to make significant progress in our partnerships while learning from the best in the industry and also while validating our engine externally. Together with leading biopharma partners, we have generated more than 500 million in realized inflows while advancing differentiated programs with Sanofi and Roche Genentech. So today I'll share how we continue to strengthen our product engine and how we take these advances and are translating it into differentiated medicines, differentiated partnerships, and ultimately better outcomes for patients. So the question that naturally comes up, what makes our product engine different? There are many companies applying AI to drug discovery. We believe our advantage isn't AI alone. It's the combination of three capabilities that reinforce one another. First, we generate our own proprietary multimodal biological and molecular data at scale. This matters because AI can only learn well from high quality data. and much of the most valuable biology has never been measured systematically. Our 50 petabytes of data is designed specifically to train models, discover new biological relationships and improve over time as new algorithms emerge. Second, we connect these models directly to experimentation through a lab in the loop system, spanning biology, design and increasingly the clinic. Every prediction, as I mentioned before, is validated experimentally. Every result feeds back into those models. It is that recursive loop that helps us to move faster, improve our decision quality, and systematically build confidence in our programs. And third, and most importantly, we convert these capabilities into differentiated assets. That includes both our internal clinical programs, such as REC 4881 and FAP, REC 1245, RBM 39, and solid tumors, as well as our partnered programs with Sanofi, Roshan Genetics. So how are we doing? Let's look at the progress we've made over the year to date. As we look back over the first half or so of the year, I'm very pleased with the progress we're making across all three dimensions for business, our internal pipeline, our partnerships, and the continued advancement of our AI-native product engine. On the internal pipeline, we advanced REC 4881 with our initial FDA engagement, following encouraging phase two data and additional phase two data coming later this year that Vicky will talk about shortly. We have continued to build confidence in REC 1245 with early clinical safety and pharmacokinetic data. And we just received IMD clearance for REC 7735, positioning it to enter the clinic later this year. At the same time, our partnerships are also making progress. As you'll remember from earlier this year, we achieved another milestone with Sanofi, our fifth to date, on developing a novel lead series for a very challenging first-in-class oncology target. But I'd like to pause on a new milestone in particular that we're announcing today. Together with Roche Genentech, we are thrilled to announce that Genentech advanced the collaboration's first neuroscience target, a new unexplored target in neuroscience, into a joint early discovery program, providing early evidence that Recursion's platform can generate novel, biologically validated targets for drug discovery. To me, this represents much more than another partnership milestone. in an area where progress has been slow for decades. It provides early evidence that a fundamentally different approach, combining proprietary disease-relevant atlases, purpose-built foundation models, and that rigorous computational and experimental assays that we use to build confidence that these targets are actually causal. And of course, last but definitely not the least, the deep collaboration, scientific and technical, with the partner can uncover previously unexplored therapeutic targets. While it's still early, I believe this is an important proof point for both recursion and the broader field. It suggests that an AI-native engine can move beyond optimizing known biology to discovering new biology, compelling enough to advance into drug discovery with one of the world's leading neuroscience organizations. So that's just the left-hand side, but we have a lot more coming ahead. For REC 4881, we will present additional phase two data at the CGA IGC Conference, a premier medical congress for inherited GI disorders, our specific target audience for FAP. And we will also provide an update on our FDA interactions, as well as continue advancing what we believe could become a transformational therapy for patients with FAP. Remember, nothing approved to date, no approved therapies. For REC 1245, we are continuing our dose escalation and generating additional phase one data, and we'll have a more wholesome update later this year. With Sanofi, we expect the potential nomination of an oral INI development candidate, a very important milestone that would further validate our ability to design differentiated small molecules against challenging targets with the potential to impact multiple immune mediated diseases. and finally, we expect to initiate the phase one study for REC7735, further expanding our clinical oncology pipeline with another precision design program from our engine. Taking together these milestones reflect a company that is delivering ambitious proof points that matter while executing with focus and discipline. But equally important, we continue to strengthen the engine itself. Let me show you a few examples of how that innovation across biology, chemistry, and clinical development is making our engine faster and smarter. Let's start with biology. One of the biggest challenges in the industry is that much of human biology remains unexplored. We believe the answer isn't simply building larger AI models. It's generating proprietary disease-relevant data that these models can actually learn from. To do that, we have generated and aggregated more than 50 petabytes of multimodal biological data, creating what we believe is one of the largest proprietary datasets in the industry. And as that dataset grows, our models become better at discovering novel biology, and every new discovery further strengthens the engine. That learning then carries into design. Because our biology models generate higher confidence hypotheses, our chemistry platform focuses on designing better molecules more efficiently. There's much to share here, but one thing I'll mention is we are advancing candidates using roughly 330 compounds over approximately a year and a half. So going from target to candidate in a year and a half, compared with industry benchmarks for small molecules of roughly 2,500 compounds over four years. That's a meaningful improvement in both speed and capital efficiency. And finally, we extend that same philosophy into the clinic. The clinical development is where a lot of value is ultimately created and where also a lot of programs fail. By bringing AI into trial design, picking the right patients that can't enforce that enough, and Site Selection, we're already seeing improvements in enrollments, speed and patient matching, helping us to run smarter and more efficient studies. But one more important point, this isn't three different capabilities, it's one continuous learning system. Every experiment improves our data, better data improves our models, better models make better molecules, and then clinical data is fed back into the system to make the next generation of products even stronger. Perhaps the best example of the flywheel in action is what we have demonstrated with Roche Genentech and we're announcing today, where our biology engine discovered a previously unexplored and new neuroscience target. I'd like to spend a few minutes just to take you behind the scenes as to how we got there and why we believe this represents an important new approach to discovering medicines. Together with Roche Genentech, As we worked in this area to discover a new unexplored target from our AI driven map of biology, we focused on a few specific elements. Why does that matter? First, this wasn't about finding another target within a well studied biology. It was about uncovering previously unexplored biology and building enough evidence experimentally to advance it into drug discovery with one of the leading neuroscience organizations in the world. Second, we believe this validates something bigger than a single target. It provides early evidence that when you combine the right data, build the right models, do very rigorous computational and experimental validation, and pair that with the right complementary collaboration, you can actually systematically uncover novel biology. And we believe this is just the beginning. The underlying biological maps are reusable. This is a really important point, with the potential to generate many more therapeutic opportunities over time. Finally, across our collaboration with Roche Genentech, we've now achieved more than $216 million in upfront and milestone payments, with the opportunity for more than $300 million in additional development, commercialization, and sales milestones for each future small molecule All right, so let me show you how we built this engine. To understand why this milestone matters, the question is why neuroscience? It's worth stepping back and asking that question. Neuroscience remains one of the greatest unmet needs in medicine. More than 3 billion people worldwide are affected by neurological diseases. And yet, CNS drugs, as we know, continue to have amongst the lowest approval rates in industry. Neuroscience is particularly challenging because the biology is extraordinarily complex, difficult to model, and we have repeatedly returned to the same small set of well-understood targets with only incremental success. We believe meaningful progress will require discovering new biology, not just simply optimizing what is already known. And that's exactly what this collaboration was designed to do. So the next question comes, what does it actually take to discover a target that people will have confidence in? And before I go into the details, just a huge, huge thank you to Roche Genentech for this deep shoulder-to-shoulder collaboration. It's one of the few rare ones that I've seen where the teams are looking at the same data, talking about the same models, going through what validation needs to be done. So that joint collaboration was critical here. Everything starts with disease-relevant biology. We asked ourselves a simple question. Are we studying neurons in a context that actually reflects human disease? In our case, that meant creating iPSC-derived neuronal cells, both neuronal and microglial cells, at an unprecedented scale, more than a trillion neurons and hundreds of billions What this does is it creates a rich disease-relevant atlas that can be reused again and again to discover multiple future targets. We view this atlas as one of the most important long-term competitive advantages. But generating proprietary data, while important, isn't enough. The next challenge is making sense of it. Before asking the models to find something new, we grounded every analysis in causal biology that we understand today. So really grounding it in genetics. We introduced hundreds of disease-causing perturbations and anchored our searches around well-established drivers of neurological disease. That matters because it gives every subsequent prediction of biology from a causal target from the very beginning. Rather than searching blindly across the genome, we are searching from a foundation grounded in causal genetics and disease biology. Now, as that's established, AI can help us on our foundation models ask a much more interesting question. What is not seen? What can be unexplored biology that we don't know of today? This is where our foundation models come in. Instead of evaluating one hypothesis at a time, the models compare the biological signatures of more than 17,000 genes across tens of millions of data points. They build relationships across the entire genome and identify genes that consistently behave like known disease drivers, even if they've never been implicated in that disease before. That allows data and foundation models, not preconceived hypotheses, to compile a prioritized list of new novel potential targets. Now, AI can generate hypotheses, but medicines and programs require evidence. Together with Roche and Genentech, we predicted every target. We looked at every predicted target and then put that through a rigorous experimental validation cascade. We build confidence in layers. First, we establish that the target actually sits in the right biological pathway. Second, we show that changing the target actually can improve cellular function, for instance, neurons or microglia. And finally, very critical, we demonstrate that this target and modulating it can meaningfully affect disease-relevant biology using multiple orthogonal assays. These assays are very robust, but they also include other multi-omic data layers, such as proteomics, transcriptomics, et cetera. While no single experiment tells the story, what we do here is build a body of causal evidence before advancing the target. So putting it all together, our collaboration combines four capabilities. Generating disease-relevant biology at unprecedented scale, and it's challenging to do, to actually have a trillion iPSC-derived neuronal cells that are high quality, standardized, viable. It takes a lot of specialized protocols and know-how to do that. Second, we use foundation models to systematically explore that biology. Third, we navigate from well-understood disease mechanisms towards previously unexplored new biology. and finally, a very important step is validating all of these predictions experimentally before we advance it. So our first neuroscience target, as I mentioned before, has now advanced into a jointly developed small molecule discovery program supported by our design platform. And again, what excites us most is of course this target, but the fact that this kind of data is highly reusable. The potential to mine it over and over again for unexplored targets, and also that this wasn't the result of one algorithm or one experiment. It's the result of a new operating model for discovering medicines. Before I hand it over to Vicky, I would like to highlight as we move on to our internal programs, the pipeline. As you can see here, we have multiple programs in the clinic. We're constantly looking at the data to make data-driven decisions. For REC 4881 and FAP, where there's no approved therapies today, and REC 1245 targeting RBM, a novel first-in-class target, first-in-class degrader with limited clinical competition to date. Combined with additional internal and partner assets, we believe this creates a diversified portfolio with multiple opportunities to create value in the coming years. With that, I'm going to turn it to Vicky to walk you through the internal pipeline in more detail.

Thank you, Najat. I'll start off this morning by talking about our REC 4881 program in FAP. FAP is a rare disease that requires lifelong management. Patients with FAP develop hundreds to thousands of adenomatous polyps in their GI tract and require colectomy to reduce the risk of colorectal cancer. Following colectomy, polyps may continue to develop and grow both in the residual lower GI tract as well as in the duodenum in the upper GI tract. Patients require ongoing endoscopic surveillance, may require additional surgeries, and they continue to be at risk for GI cancers. With over 50,000 post-colectomy patients in the US and EU5, there are no approved systemic therapies to alter the course of disease. This represents an over $10 billion potential addressable market. REC4881 is an oral MEK12 inhibitor with a differentiated dual mechanism of action in FAP with the potential to inhibit both new polyp formation via crosstalk inhibition of the beta-catenin pathway, as well as to directly interrupt signaling of the MAP kinase pathway, which is a key signaling pathway in advanced disease. So again, blocking potentially both new polyp formation as well as the existing polyps within the GI tract. So with that, I'd like to take a minute to discuss the impact of this disease on patients through a story of a woman named Jenny who lives with FAP. Like approximately 70% of FAP patients, Jenny inherited the genetic mutation responsible for FAP from a parent, in her case, her mother. Seeing what her mother experienced had profound psychological impacts on Jenny, who knew from the young age of eight that she also carried this mutation. She has since had to endure multiple surgeries, which have led to chronic and life-altering complications, including frequent bowel movements, malabsorption and dehydration, chronic abdominal pain, and anxiety with medical PTSD from all of the surgeries and procedures. We have heard from both patients like Jenny, as well as their treating physicians, an interest in a pharmaceutical intervention that can prevent polyp growth and disease progression and ultimately lead to a reduction in the need for repeat surgical procedures. REC 4881 has shown promising clinical data in the ongoing phase two, two below study. Patients who had undergone colectomy for FAP receiving 4881 showed a median polyburden reduction of 43% after three months of treatment. That treatment effect was durable with sustained reductions after three months off treatment. Additionally, reductions in polyp burden were seen in both duodenal disease in the upper GI tract as well as the lower GI tract. The upper GI tract in particular is an area of high unmet need as approximately 90% of FAP patients will develop upper GI polyps. When removal of these upper GI polyps becomes necessary, the thin mucosal wall of the upper GI tract increases the likelihood of complications, including bleeding and perforation. REC4881 has a manageable safety profile with predominantly mild to moderate adverse events consistent with the safety profile of other MEK inhibitors. We continue to enroll patients on the Phase 2 Tupelo trial, including patients 18 years of age and older, as well as a dose optimization cohort. We are pleased to share that additional REC 4881 data will be presented during the Presidential Plenary Session at the CGA IGC Conference in November. As Najat mentioned earlier, this conference is focused specifically on inherited GI cancer syndromes with a target audience which includes physicians who treat FAP patients. We also look forward to providing an update on FDA discussions later this year. Now I'll move on to REC 7735. PI3 kinase is frequently mutated in several cancers and is a clinically validated therapeutic target. Lack of selectivity for the mutated form over the wild type is a key challenge for existing agents as inhibition of wild type PI3 kinase drives hyperglycemia. Increases in blood glucose are both a safety issue, which often limits dosing, and an efficacy issue, as the resulting hyperinsulinemia can reactivate signaling through the PI3 kinase pathway, undercutting the efficacy of less selective drugs. REC7735 is precision designed to be 100-fold selective, greater than 100-fold selective, for the H1047R mutation. which is the most frequent activating mutation in PI3 kinase. Recursion's AI native platform identified a previously unpublished binding site and delivered a development candidate in 10 months with no identified off-target liabilities. As hyperglycemia and the resultant hyperinsulinemia are driven by inhibition of wild type PI3K, The selectivity of 7735 is expected to result in an improved safety profile with respect to hyperglycemia and may allow expansion into patients such as diabetic and pre-diabetic patients who are unable to tolerate current PI3 kinase targeting options. An improved therapeutic index, as I have described, may allow us to expand treatable patient populations both within existing PI3 kinase alpha inhibitor indications, as well as in additional solid tumors in which PIK3CA mutations are prevalent, including potentially triple negative breast cancer, ovarian cancer, and endometrial cancer, just to name a few. Additionally, the improved therapeutic index may allow expansions into earlier stages of disease within oncology as well as non-oncology populations such as PI3 kinase-driven vascular anomalies. With the IND now cleared by FDA, we intend to initiate a Phase 1 Zinnia trial later this year. Dose escalation will begin in patients with PIK3CA H1047R mutant solid tumors. Once tolerability is confirmed at an active dose, we intend to expand into the hyperglycemia vulnerable patient cohort to confirm the improved tolerability in this patient population. Dose optimization of two active and tolerated doses will then be performed in ER-positive HER2-negative breast cancer patients. We may also expand into additional tumor types based on emerging data. We expect to share the first data from this dose escalation part of the trial in the first half of 2028. And with that, I'll turn it back over to Najat.

Thanks, Vicky. And shifting gears a bit, we often get asked about whether advances in frontier AI can reduce or increase recursion's competitive advantage. We believe we have a truly competitive, unique competitive edge. As reasoning models and agents continue to improve, next slide, They become dramatically more powerful when paired with proprietary data, automated labs, and real experimental feedback. That's exactly the system we've been building for years. Now, we are deploying agents across biology, chemistry, and clinical development across the engine and also alongside our scientists. In biology, here's some very quick examples. Our target discovery connector is helping scientists interrogate our proprietary biological maps in hours rather than weeks. These are the large maps that we just talked about earlier in our partnership with Roche Genentech, but also the internal maps that Recursion has built over years, accelerating the discovery of novel targets. In chemistry, our design agent reasons across structure, SAR, and experimental data to prioritize the next design hypothesis, critical inflection points in programs. This helps our scientists decide what to make next and compress design cycles from roughly four hours of structural analysis to about 30 minutes. And in clinical development, the agentic workflows are already improving patient enrollment, contributing to about 1.3 to 1.6 fold improvements over historical benchmarks. That's significant. These are still early examples. But I will have Chris Redoux, our Director of Structure-Based Technology, who's in this day in and day out, walk you through a real example in practice. Chris?

How we design our drugs matters as much as the drugs themselves. It's not about a single method. It's about an ecosystem. Tools compute data. And a UI that lifts productivity whilst capturing intent. Every decision, every step. Working on difficult-to-drug targets can feel like walking a tightrope through chemical space, and we are very deliberate about where we step. We minimize the number of compounds we make through deep exploration in silico. We have captured 97 billion predictions across 5.5 billion compound records, traceable to the design runs that made them and the problem the designer was trying to solve. This becomes the playbook for future agents Automation and plentiful compute means we are able to run calculations proactively for each project compound This ensures design agents have a rich context for interpreting experimental data Here, a chemist asks how to improve potency In seconds, the agents identify an insight from a compound the team has set aside due to solubility issues Then they explain why. The agent pulls in pre-computed physics-based calculations to show that this gain isn't a new interaction, it's conformational strain. That tells the team exactly how to redesign. Relationships no single scientist could hold, surfaced, explained, and turned into the next designs. That's how our teams move faster. Our approach to design has always been well suited to automation. Our inputs are far easier to record than inspiration at the bench. Several years of capturing our own drug design work has built up an immense catalog of design knowledge and agents are helping us to unlock it.

Thanks, Chris. What you just saw wasn't a chatbot answering a question. It was an AI agent reasoning across our proprietary experimental data, our in silico data, our historical project knowledge, and structural biology to surface insights that would otherwise require scientists long time, but then also non-obvious insights. That's because in drug discovery, the bottleneck is really just generating ideas, is finding the right idea quickly enough to keep the make, test, learn cycle moving. As these agents continue to improve alongside frontier models, we believe they will become an incredibly powerful multiplier of what we have already built. And finally, I'd like to highlight another aspect of our AI strategy. AI is advancing incredibly quickly, and no single model will remain state of art forever. Our strategy isn't to depend on any one model. is to build an AI-native product engine that can rapidly develop and adopt the best advances, whether they're developed at recursion or by the broader open-source community. NESO-1 is a great example. We developed and open-sourced this model. This is a binding-infinity model that delivers both to level accuracy with 10 to 20x faster inference, helping advance the field while enabling dramatically faster design cycles. But look, the real advantage is in the model itself. It's our operating system. It's our operating model. It's our ability to rapidly integrate these models into our proprietary data. That increases prediction performance, accelerates the make-test-learn loop, and allows us to evaluate many more compounds at a lower cost. And finally, great technology only creates value if you have the right people to translate it into medicine. We firmly believe that. and that's why we have strengthened our leadership team in two critical areas. First, Dr. Hoifang Poon joins us as Chief AI Officer. Hoifang is one of the world's leading AI researchers with more than 15 years at Microsoft Research, where he led pioneering work in biomedical foundation models and AI for healthcare. Importantly though, he's not just a researcher, he has repeatedly translated frontier AI into real world applications and deployed that at scale. At Recursion, he will unify our end-to-end AI strategy, bringing together frontier research and applied AI across biology, chemistry, and the clinic. Second, Dr. Donovan Chen joins us to head up drug design. Donovan has spent more than two decades solving some of the hardest problems in drug discovery, from small molecules and RNA targeted therapeutics to proximity-based medicines and peptide modalities. Across Parabolas, Arrakis, and Novartis, he repeatedly helped unlock targets that were previously considered difficult or even impossible to drug. That breadth across modalities and that depth and experience of translating computational design into medicines is exactly the kind of capability we need to continue building at Recursion. Together, Hoi Fong and Donovan strengthened the two engines that will continue to define our future, world-class AI and world-class scientific design. Now, I'm going to turn it over to Ben to give us a financial update.

Thank you, Najat. As I've said in the past, we want to continuously increase the impact of every dollar we spend. We are demonstrating this today by lowering our 2026 full-year cash operating expense guidance to $375 million. In total, our revised 2026 guidance represents a nearly 40% reduction from comparable 2024 pro forma expenses. Through disciplined, data-driven management, we have been able to continue lowering OPEX while still advancing our differentiated internal pipeline, achieving a series of partnership milestones, and maintaining a leadership position in AI-powered drug discovery. We have been able to increase our return on investment through multiple levers across the company. In our clinical pipeline, we use our ClinTech platform to drive more efficient enrollment and planning of our clinical trials, reducing the time and cost for each important data. Najat and Chris described some of the systems that we use to make our internal discovery both more efficient and more effective. We also focus our technologies on predicting and answering the hard questions first so that we can prioritize those programs with clear potential clinical and commercial differentiation as early as possible. Because we deliver outcomes that are truly novel and differentiated, like our Roche Genentech milestone today, our partnerships have achieved over 500 million in cash inflows, including more than a dozen successful discovery milestones. All of our partnerships are designed to be break-even or profitable on a direct cost basis from the start with substantial value growth as we achieve milestones. In our product engine, we are able to build, test, and integrate AI models on real projects using the scale of our internal pipeline and partnerships. We know not only if the model benchmarks well, but if it matters when it's applied to a drug program. This direct application allows us to determine early which technology investments are likely to have real-world impact. We apply the same disciplined management style to our corporate operations. We have been able to maintain G&A at a relatively low percentage of total cost, which helps us maximize the scientific ROI of every dollar we spend. We ended the quarter with approximately $557 million in cash and equivalents, which we believe provides us with an operating runway through early 2028. And with that, I'll turn it back over to Nanjot.

Thanks, Ben. I'll close by looking ahead. We built an AI-native product engine. Now the focus is expanding its impact while continuing to translate its capabilities into the right and repeatable proof points. So on our wholly owned portfolio, you should expect to see continued progress across multiple programs, additional phase two data for REC 4881 and a regulatory update before year end, continued advancement of REC 1245 with a more wholesome update later this year, the initiation of REC 7735 that Vicky just mentioned, and progress across the broader pipeline. We are on track across those multiple fronts. With our partners, we expect to build on this year's momentum. Following the advancement of the first previously unexplored neuroscience target with Genentech, we see the potential for additional programs to emerge from our maps. And with Sanofi, we expect the potential to continue the progression of AI design molecules towards development candidates and later stage milestones. We're entering an exciting period with multiple opportunities to demonstrate the power of our engine. With that, thank you again for the time today. And I'd be happy to take your questions. Great. So I'm just going to go through some of the questions. The first question coming from Alec from B of A and Sean from Morgan Stanley. Thank you. How does a collaboration with Roche Genentech form a template for how you can leverage your platform with other partners? Maybe two to three aspects that you think are transferable and provide proof points. Yeah, I mean, it's a great question. Thank you both. Big picture, the way we develop our novel data sets for creating novel maps, and then we take those novel targets and design compounds all the way into the clinic, that sort of lab in the loop is something we use for both our internal programs and for our partner programs. So that template is something that will only get better, faster as we go on. And we can, you know, in terms of new partners or current partners, we'll continue to scale that. As I mentioned before, you know, our differentiation really lies in three areas. One is that data factory. I mean, especially in biology, given so much of it is not known well. having access to great biology and data is incredibly important. And that takes years to build. I want to emphasize that, you know, understanding how to generate that data, validate that data, develop the models and also have a supercomputer, which we have in a hidden location in Salt Lake City, having that entire stack to make sense of the data back into the lab and validate it. I think that is something we are one of the very few companies that can do that. And we continue to drive momentum there. Next question. Can you provide, and this is a question from Sean from Morgan Stanley, Gil from Needham, and Brendan from Cowan. Can you provide an update on FDA engagement on REC 4881 and FAP, the registrational pathway, and the data coming up at CGA IGC? Maybe Vicky, you want to get us started?

Sure, I'd be happy to. Maybe I'll start with the upcoming data at CGA IGC. So we presented data from the Phase 2 Tupelo trial for the first time back in December of last year via a webinar. We do think it's really important to put these data in front of the physicians who treat patients with FAP. And so this will be an updated data set, again, presented in an oral presentation at a presidential plenary session at that meeting, which occurs in November, where you may see additional analyses that help contextualize the clinical relevance of the data, as well as potentially additional patients in that analysis as well. So we look forward to sharing those details with the FAP treating community later this year. With respect to the FDA engagements, as we've said, these are ongoing. I think it's important to remember there's very limited regulatory precedent in FAP. And so our engagement here really is around making sure that we de-risk the study design from a regulatory standpoint, including things like what is the appropriate primary endpoint to demonstrate clinical benefit. and I would say as somebody who worked at FDA many years ago, those discussions, those conversations have been productive and I think are helping us get to a better point in terms of the study design. So nothing out of the ordinary there. Again, this is just a rare disease with limited precedent and we continue to have a productive dialogue with FDA and look forward to sharing once we have sort of something more concrete to share. look forward to sharing more details on that later this year.

Thank you, Vicky. All right, I'll move on to the next question. Ben, this is for you. From Priyanka JPM and Gail Firmino. Can you provide more color on what operating efficiencies were done to reduce the OPEX guidance? Is there potential for further belt tightening on OPEX in second half of 2026?

Yeah, great question. And I think as you saw in the presentation, Najat covered We haven't changed any of our full year guidance on what outcomes we're trying to achieve over the course of the year. And I think that's really important to remember because this reduction in guidance is actually from doing the same amount or more with less. and so what we've really tried to focus on is how can we get to the most important answer first you heard some of the description of the technologies that Chris took us through that Najat took us through and that really makes a difference on how we can operate it how we can deliver those outcomes so I think we started the year and we had some ideas of where we could go what we've seen is they actually have impact we are actually getting to the answers faster and more cheaply. I think the numbers that everyone should use are the numbers that we give in guidance, which is the 375. That is our expectation of where we will be operating. But at our core, we are always looking for a better and faster way to do everything that we do. We are a technology company and we should be getting more and more efficient over time. So we will keep looking and update you as we know more.

Thanks, Ben. Yeah. And just to just to maybe reiterate that we always have a commitment in order to ensure that every dollar goes further with some of the improvements we're seeing in our engine. You saw some of the examples around the fact that we design 90%, we physically make 90% less compounds for the one that goes into the clinic. We take about a year and a half versus four years versus industry. Those are meaningful improvements in the velocity that we see in our engine and we ensure that that actually parlays into our spend. We mentioned earlier this year that we changed our budget to an outcomes-based budget so that every single aspect, like Alec and Sean, going back to your questions, when we do a partnership, we know exactly the fully loaded cost of building a map, of a program, and so forth. And that really helps us to ensure that those efficiencies are realized. The other thing I'll also say, we continue to focus on our G&A. and ensure that every single dollar is actually going to our programs and our partnerships. So we will continue to put pressure, that's our commitment, just like our commitment is to deliver on proof points from what can be really a value inflection point for the broader community in terms of programs and the use of AI to create value. Okay. With that, I'll go to the next question. A platform question from Alec from BFA and many others. Okay. with multiple tech companies entering drug development. And as generative AI becomes increasingly available, how does recursion differentiate itself today and in the future? And what do you believe remains recursion's durable competitive advantage competitors will find hardest to replicate over the next five years? Great question, Alec, and everyone else who has that. I think that's why you saw the second slide in the presentation was really around our durable mode and our differentiation. And that evolves over time. I think number one is the data factory. Look, you just said generative AI is becoming increasingly available. Maybe some would say even commoditized. Where does the differentiation come from? If 80, 90% of biology is unknown, it has to come from high quality data generation. Models depend on good quality data to be trained on. And you saw the example with Roche Genentech that we shared today, but also across the board, Starting with disease-relevant data sets also matters. That just doesn't exist. So in order to build that, like a trillion iPSC-derived neuronal cells, that's a cell manufacturing capacity that we have in our Salt Lake City labs. Over years, we have gone through the pain and suffering of what works and what doesn't work. So think about it as a really mature and increasingly validated capability. So that's one of the data factories. And that's not just for biology. You heard from Chris Raduch. 10 years of actually doing small molecule design, you know, millions to billions of virtual cells, virtual molecules that have been generated also gives us a lot of rich data, not just in areas that are known to the world like kinases, but actually other targets that are less known and not as available in the protein database, PDB, for instance, and others. So that's one big pillar. Second, I can't emphasize enough, is that lab in the loop, that operating model. because it's one thing to have great data, it's another thing to have great models, but really important, we need to validate these predictions. The only way we get this to be useful utility at the end of the day to make a drug is if you're validating it back into the lab and that feedback, good or bad, goes back into the models to make them better and smarter. We did the same thing with AI agents. The more you engage with them, the more you give them feedback, they get better. I think that integrated lab in the loop is It's hard to build for two reasons. It takes a lot of technical expertise, yes. It takes a lot of years of knowing what works, what doesn't work, yes. It takes tons of reps and with partners that are some of the best in the industry, we learn faster. But so much of it is also culture. It's culture. I've always mentioned the piece that we have bilingual scientists that understand, better understand, I would say, both science and tech that have appreciation of the challenges and opportunities with both that open-mindedness where an agent gives you a different hypothesis from what you started when you were in medicinal chemistry that's worked in that space for decades, that takes a different mindset. And I cannot emphasize that enough. And then the third piece is what are we actually making from the engine? You know, FAP, first in class oral for a disease where nothing's been approved. It's a standalone high value asset. RBM 39, first in class targeted, first in class degraded or built from this platform with limited competition. So what you'll see in our pipeline isn't incremental improvements, but any one or two drugs that can actually be a standalone differentiated asset in its own right. And we all know that takes time. So I think those are the three big areas that are not just an advantage for today, but continues. Because with every week, we're doing two million more experiments in our labs. The data mode grows. With every week, we actually have people churning through that lab and they're learning. That grows. and, as you can see, with every week, month, we're making progress in our pipeline. And that takes time, resilience, focus, and discipline. And that's what we're doing. Okay, one more question for Vicky. PI3K questions from Brendan at Cowan and Dennis at Jefferies. Looks like 7735 passed your internal criteria for go-no-go decision with a phase one to start for second half of 2026. Can you tell us a bit more about the go-no-go process what it is about the preclinical profile that gives you confidence that this is the right candidate, and also what the Recursion AI platform has told you about the best development path forward in terms of study design, patient selection, et cetera. And then there's another sub-question, but I'll start with that.

Sure. So first, maybe start off by saying we believe that there's room for improvement in the PI3 kinase space. So again, this is a very common mutation in certain malignancies, including hormone-receptive positive breast cancer, but also extending beyond breast cancer into other GYN malignancies, as well as head and neck cancer and colon cancer and others. So important target. still remaining unmet need in terms of maximizing the therapeutic index and ultimately the efficacy that patients see. So the go-no-go process really involved a rigorous evaluation and confidence building in our preclinical data set. The selectivity that allows us to hit the target hard without seeing, you know, additional toxicity. So again, both in terms of the efficacy that we're seeing in preclinical models that look similar to at least similar if not improved upon competitor profiles, the safety profile including the lack of hyperglycemia, but also of course our GLP tox studies. These all helped us build confidence that this was the right molecule to move forward with into clinical trials. Of course, ultimately, you know, after evaluating these data, we made the decision to go forward. We've submitted the IND and that IND is now cleared and we look forward again to initiating that study this year. In terms of the AI platform, you know, I think one of the key pieces from a clinical perspective is These patients are going to be selected based on the H1047R mutation, so a biomarker which will require a diagnostic. And one of the key areas where I think the platform is helping us is in terms of our ability to find these patients, look for the right geographies and sites in which to conduct our clinical trial, and help us accelerate the enrollment of this patient population.

Thank you, Vicky. Maybe just a couple of things to add. You know, we talked about this early on, which is for this compound specifically, it is over 100x in activity over a while, so it's a while time sparing. Why is that important? Important from a perspective of can we actually have the patient stay on, like increased dose intensity, dose duration, as Vicky mentioned. So really try to improve the outcomes from patient and the TI. But also, you know, even with grade one, two, and so on. And so, you know, we have seen elements that it can lead to, you know, reactivating the exact pathway, PI3K pathway that you're trying to suppress. And that has the potential for also, you know, compromising some of the efficacy that can be seen. So there are multiple elements to why, as we look at this compound, what we want to test in the clinic is, is it actually giving us the better safety profile? kind of give us the better efficacy profile that would improve the therapeutic index. The other thing I would just say from the AI platform as well, as Vicky mentioned, one is recruitment. We know this is a competitive area. We're starting in with solid tumors, as Vicky mentioned, but it gives us optionality given based on what we will see in the profile to either go in on or non-on indications as well. And that's also another area where the platform can help. We're really thinking about what are the right patient groups and indications that we might select that others haven't maybe explored to date. So a lot more work to come, but step one is to go into the clinic and ensure that we are seeing the elements the compound was really designed for. And recall, the compound was designed in 10 months, 242 compounds synthesized, 13 cycles, and the pocket was a previously unpublished pocket. So we're not going after the same areas, which is why you see almost 130X selectivity over wild type. Super precise, super precision-based, 1047 is one of the most frequent mutations you see in this space, one of the ones that's tied to disease causality and progression the most. So we're excited. But again, it's part of multiple different programs that we're looking at. And based on data, we'll make the right go-no-go decisions as well. One maybe just sub-question, when should we expect initial monotherapy data? I think Vicky had mentioned first half of 2028. So stay tuned. And with that, I'm not seeing any more questions on the screen. Thank you again so much for joining us today. Looking forward to the progress over the next set of weeks and months. And as always, talk to you soon.