Recursion's Q2 2025 earnings call. My name is Chris Gibson, and I'm the Co-Founder and CEO of Recursion. I'm excited to share with you today some of the latest updates on our company as we drive forward to decode biology. We've been talking for the last nine months since the business combination with Exscientia about the Recursion OS 2.0. I want to start there today and tell you a little bit about the way that we're bringing together the incredible components from both Exscientia and Recursion and building new components of the OS in order to drive forward our mission. At Recursion, we base everything off of proprietary fit-for-purpose data, whether it's data we generate in-house or data that we pull from partners. We're not just generating data to help discover targets or to help translate programs or to help with clinical trials. We're building a true end-to-end capability from target discovery all the way through to clinical trial simulation. We're really, really excited about the way all of these pieces fit together and add to each other. Everything we do Recursion is based on iterative cycles of learning. Much of our work is based on iterative cycles of dry lab predictions and wet lab validations. I want to talk about a few of the pieces of the Recursion OS that we really, really leaned into in the last quarter. I'm going to start off with talking about Boltz-2. This was a really exciting partnership with both the MIT and NVIDIA, where we were able to help lead the field of protein folding and lead the field of protein ligand binding predictions with this work that we did with MIT. We were able to actually open source this project. To date, there have been almost 200,000 downloads and almost 50,000 unique users. What I think is most exciting, what's gotten the most traction about this work, is that we were able to actually make binding predictions that are approaching the level of efficiency and the level of efficacy of free energy perturbation calculations. We're able to do this with about a thousand-fold less compute. That is really, really, really exciting. It means that a lot of the sort of real bespoke work that was done with physics-based computing can actually be done in a screening format. While there's more work to do in this space by us and many others, we are very excited about the way this tool and tools like this are going to be able to drive the field forward. What's more, we've already built this technology into the Recursion OS and even improvements on this technology into the Recursion OS. Another area we've been talking about for the last year has been our ClinTech platform. This is something that we are now deploying against every single one of our programs at Recursion. There are multiple components to this. The first is our causal AI applied to human genomics. This is really exciting. We're taking patient data that we get from Helix and Tempus, combining that with our perturbation biology data and algorithms from Recursion to help to connect our platform to patients. This is enabling us to identify targets, to stratify patients, and even to do indication expansion. We've also started to design and simulate our clinical trials at Recursion using in-house software that we've been building. This is allowing us to potentially improve the optimal dose for 30% more patients. This is really, really exciting. We are now deploying this against our programs at Recursion. Third, we're now using our AI not just to identify patients, not just to design our clinical trials, but actually to recruit and execute. The operation side of our clinical trials is really, really important as well. With the new software that we've built and the partnerships we've built in this space, we now have the potential for 50% faster enrollment projections at high-quality sites. This means we can activate trials up to two months faster. This is the early days of our ClinTech platform. What I'm most excited about is that we're already deploying these tools against the programs in our pipeline. We'll be deploying these against new programs in our pipeline soon. Najat Khan is going to be able to tell you more about that in a few minutes. We continue to advance a pipeline of both internal programs in oncology and rare disease, as well as a suite of R&D collaborations and programs with our partners Roche, Sanofi, Bayer, and Merck KGaA. We're really, really excited about all of these programs today. What I think is most exciting isn't any one of the programs. It's the platform we're building and these leading indicators where we're demonstrating that we can bring medicines to the clinic faster and at lower cost. Ultimately, these leading indicators are things that we believe over time are going to continue to improve. We are going to be able to continue to raise a high bar of quality on our programs and drive them forward at real scale. To tell you more about the way we are building momentum, let me turn it over to our Chief R&D and Chief Commercial Officer, Najat Khan. Najat? Thanks, Chris. Great to be here today. Let's dive into this. Chris mentioned the suite of partnerships and partner discovery programs and internal programs that we are progressing. A couple of things to note on the internal side, you can see there are six or so programs that are going through really, really important inflection points, both across oncology and rare diseases. What I'll do today is double-click a bit more on a couple of our more late-stage or later-stage oncology programs: CDK7, monotherapy dose escalation, as well as the initiation of our expansion cohort/combination arm, and RBM39. We'll share a little bit more around the biomarker enriched, the solid tumors, the patient populations, et cetera, and how we leverage our platform insights in order to hone in on where we go. On the partnership front, I get this question a lot. I just wanted to step back for a second and share. Across our partnerships, there are two major areas of value creation. The first is really around what Chris mentioned in the beginning, proprietary fit-for-purpose data sets that we're co-developing with our partners. An example of this is, of course, the Phenomap, the first neuronal Phenomap, iPSC-derived with Roche Genentech. The other area of value creation is around partnered programs where we are designing using our AI modules on the chemistry side, very challenging first-in-class, best-in-class programs. Just recently, we achieved a fourth milestone in our Sanofi partnership. More to come on that. Going to the next slide, I'm going to take a second to do a quick snapshot on the overall programs that we have in our internal portfolio. Then I'll go a bit more into CDK7 and RBM39. Just as a quick reminder, CDK7, really important target. The focus really is leveraging our AI-powered design module in our Recursion OS 2.0 platform to optimize the therapeutic index. This is a target that has been tried by others before. That's the area of focus. We should have more monotherapy dose escalation data by the end of this year, and as I mentioned, combination initiated. RBM39, this is an example actually identified using our Phenomaps where we identified a new MOA with synthetic lethal targeting opportunities in genomically unstable cancers. More on that. First half of 2026, we anticipate some initial data from our monotherapy dose escalation. You heard a little bit about the MEK1/2 and our FAP program. I just want to highlight this is, again, another phenotypic insight where we actually derive the fact that there's a connection and an important relationship in an unbiased fashion between MEK1/2 and the relationship with MAP kinase pathway signaling pathway and APC and WN signaling pathway, which disease this is for FAP. We should expect more data beyond the initial cut we shared earlier this year, second half of 2025, end of this year. MALT1, this is another program where now we're using and leveraging our AI-powered chemistry design portion of the Recursion OS platform, again, to lower the liability that's associated with UGT1A1 inhibition that's also in monotherapy dose escalation. To round it out, we also have a couple of preclinical programs here that are going through important inflection points in the development candidate/IND enabling phase. A lot of these programs, and we talked about this before, were really focused on the earlier versions of the Recursion OS platform. As we iterate and learn and add more components to our Recursion OS platforms, we expect the next wave of programs to be even more high potential and potential to do it in a more efficient way. I wanted to take you a little bit under the hood of what's actually in the Recursion OS, especially the 2.0 platform following the integration with Exscientia. If you just look to the left-hand side, we first start with the AI-powered biological insights. This is where we are actually deriving novel targets. This is from multi-omic data, whether it be phenomics, transcriptomics, et cetera, connecting that early on with the patient. This is the ML-based patient connectivity data layer that's really important. Data sets such as from Tempus, Helix, and others, ensuring that we can actually take these biological insights and deconvolute the MOA and very early on do a screening approach around triaging what are some of the binding affinities early on. This is where approaches such as Boltz-2 that Chris mentioned earlier are already being incorporated into our workflow. In addition to that, we're also developing proprietary algorithms in-house. As soon as we put this on a slide, I have to say it gets outdated because there's so much rapid iteration and work that's happening. In the middle, AI-enabled precision design, this is where we're designing our molecules, really optimizing both for novel scaffolds. This is where we use generative AI approaches and also active learning in order to optimize drug-like properties. This also includes using QMMD approaches, which is a 3D protein and atomistic model. One important point here is the wet and dry lab integration that we have. This is where aspects around automated chemistry, automated biology, and automated ADMET become incredibly important. We can design out certain elements earlier, faster, to ensure that we have better molecules out of discovery. Last but certainly not the least, and one that's close to my heart, is ensuring that we do this also in clinical development. Chris touched on this in terms of some of the areas that we're building out. You'll see some of the examples we're using in our current programs already around causal inference on patient stratification and also smarter trials and faster recruitment. As I go through each of the programs, I will actually highlight which area of the Recursion OS 2.0 module and platform we are integrating and actually highlighted insights for our program. Let's start with RBM39. In this program, as I mentioned earlier, the focus was really around leveraging our maps of biology. Just as a reminder for everyone, starting on the left-hand side, we start with these really large maps of biology, whole genome CRISPR knockouts. Then we profile compounds that are proprietary to us in order to get a better understanding of the initial chemical substrates that might actually modulate the biological insight that we have identified. The example here is how we identified RBM39, which phenomimics CDK12. CDK12, and this is to the panel to your right-hand side, has been an attractive target in oncology, right, for its role in DDR modulation. Generally, it has suffered from challenges in selectivity because of how homologous CDK13 is. Leveraging our Phenomaps, we actually identified that RBM39 is similar, phenotypically at least, to CDK12 and not to CDK13. That was the first insight. The second insight was the fact that we were actually able to develop molecular glues and degraders for RBM39, which you'll see in a moment, that also phenotypically mimic CDK12. This was our first inkling that this could potentially, RBM39 inhibitors or degraders could potentially provide a druggable potential analog. I want to say something else that doesn't get talked about enough, which is if you look in the middle panel, we also look not just for CDK12 or CDK13, but we look more extensively across the map to see if there are well-established dependencies that are known of already biologically that are also being validated. An example here is a CDK12 and cyclin K similar phenomic readout. This is just a small detail in the entirety of the map that we look at. If you go to the next slide, this is another expansion of that same map. What we see here that's actually quite intriguing is in the center in the black box is what I was referring to in the earlier slide, which is the RBM39 and the degrader itself and some of the associations that we see with CDK12, CDK13, and so forth. You look broader and you also see associations mechanistically in DNA damage repair, epigenetic regulation, cell cycle control, and transcription. When you look at it from an MOA perspective, which I'll turn on next, actually intuitively makes sense. RBM39, if you go to the next slide, is focused, is important for splicing fidelity. Degradation of RBM39 leads to splicing defects. If you combine that with tumors that are already genomically unstable, whether it's because of DNA repair pathway vulnerabilities or transcriptional regulation, that can actually increase the amount of instability leading to potential apoptosis and cell death. I just wanted to share with you how an insight is then triangulated with understanding of mechanism of action. That's not enough. If we go to the next slide, in addition to that, we also looked at in vitro and in vivo work. Starting with, when we look at the broader patient population, just given the connectivity across the maps that I noted, for replication stress, tumors that suffer from epigenetic dysregulation, cell cycle alterations, or oncogenic drivers are relevant, as well as those tumors that have DDR effects. Both of those. That spans several solid tumors from colorectal, breast, et cetera, along with some pretty clinically actionable alterations that we will be studying and looking into more, such as MSI high, MEC amplification, et cetera. We wanted to look at the in silico understanding and triangulate that with in vitro and in vivo work. If you look at the in vitro cell lines, you clearly see that RBM39 degrader, so REC-1245 in this case, there is greater sensitivity in cell lines that have higher replication stress versus cell lines that don't have higher replication stress. This was a good early signal for us. If you go to the next slide, we see a similar trend hold in in vivo as well, where you see a reduction in tumor volume across different tumor types that actually have high replication stress signatures. This helps us to do two things. Number one, better understand the importance of RBM39 as a first-in-class target in solid tumors. Second, also give us a better sense in terms of which patient population, tumor segments, et cetera, might be relevant for us to target. If you go to the next slide, we went a step further than that. We also wanted to look at the totality of it. You have the Recursion OS inside, definitely the preclinical data that I mentioned, but also looking into mechanistic validation in the middle panel. We see two things here. First, the Dmax is approaching almost 100% in RBM39 degradation with quite potent D50 numbers as well. Rapid and potent RBM39 degradation. We wanted to go even a step further. If you go to the next slide, which is, if you go one slide before, please. OK, that's OK. If you go to the next slide, this has actually helped us inform what our dose escalation and our combination arm is going to be. For RBM39, monotherapy dose escalation, but in terms of the cancers that we're looking after or going after, it's endometrial, ovarian, et cetera, cancers with high genomic instability. We will also be focusing on some of these biomarker-enriched populations such as MSI high. First patient dose, patients are enrolling in this study. We should have early safety and PK data from this monotherapy trial in the first half of 2026. Now, we'll go to CDK7, which is our next program. Here, we actually leverage two components of our Recursion OS platform. First, focused on designing a molecule that can really optimize for the therapeutic index. Second, leveraging some of our ClinTech approaches in order to hone in on which patient population and which combination arm we will hone in on. Let's go to the next slide. OK. Just a quick reminder in terms of how the molecule was designed. A couple of things to note here. CDK7 has been an important target for some time as well. It is a master regulator, both cell cycle progression as well as transcription. One of the challenges that other compounds have seen so far is challenges with permeability, efflux, and not rapid absorption. We want to change that around. We use generative AI models to actually design new scaffolds. I think this part is really important, which is leveraging active learning and experimental ADMET data to quickly learn, iterate, and optimize the molecules to reduce the components that we wanted to design out, such as ensure that there's high permeability, rapid absorption, and low efflux. Similar to RBM39 degrader, which was done in a very short amount of time, 18 months from start to IND enabling, with about 200 compounds or so synthesized. In this case, you also see about 136 novel compounds synthesized and getting to candidate ID in less than 12 months. One of the components for designing high permeability, rapid absorption, and low efflux was to ensure that we would have sufficient exposures. You see that on the right-hand side panel, both 10 mg QD, 20 mg QD, clearing the IC80 line. When we actually look at versus some of the peers, it's an order of magnitude higher than the exposure that they're seeing. As of November/December 2024 data cutoff, the compound showed one confirmed PR in ovarian cancer, as well as multiple cases of stable disease. So far, with a favorable safety profile and no MTD reached. If you go to the next slide, what we have done since then is to really design which combination arm we will focus on. The one that we're going to focus on that we have announced today is second line plus platinum-resistant ovarian cancer. How did we get to that? First, we looked at preclinical data. Cell panels in vivo, you see in ovarian, both of them are sensitive to CDK17, and there are multiple panels that were done. In addition to that, as part of our ClinTech approach, we also use causal inference using some of this multi-omic and clinical data. This was very important to better understand the cause and effect factors here. What we see is that a higher expression of ovarian cancer based on this data is associated with lower overall or worse overall survival. This was based on about 32,000 patient records. This gave the totality of the evidence from preclinical and also some of what we see in our early clinical data so far, combined with some of this causal inference work, gave us more confidence in terms of the first indication that we would go after, where there's significant unmet need in second line plus platinum-resistant ovarian cancer. If you go to the next slide, site selection and activation is in progress right now for the combination arm. The standard of care includes single-agent chemotherapy, beva plus chemotherapy, and in some cases, PARP inhibitors. In addition to that, the monotherapy arm is ongoing, and we anticipate more data from that later on this year. If you go to the next slide. I'll also share a bit more about some of our partnered discovery programs. Next slide, please. Great. If you look at Sanofi as an example, I just mentioned that we have our fourth program milestone achieved in the last 18 months. I just want to take a moment to say that some of these programs, both in immunology and oncology, first-in-class, best-in-class, some of the milestones that we're going through include important milestones in discovery, lead series, development candidate, and so forth. We have several programs advancing to those milestones, including development candidate in the next 12 to 15 months. This effort leverages what you saw in the Recursion OS 2.0 platform, a lot of the AI-powered chemistry design module. In terms of Roche, five Phenomaps built to date. You saw an example for RBM39 how we use some of our Phenomaps. These are specific for the neuroscience and GI/ONC space. For the neuroscience one that we delivered last year, over a trillion iPSC-derived cells used, whole genome knockout, and also other perturbations in terms of overexpression. You're really getting a very holistic understanding of biological pathways. A lot of work in progress there in order to take those insights and translate them into novel programs. More to come on that. Also on the GI/ONC indication, over four maps already developed there and already one program that has been optioned and more work happening. I think one point to note here, it's a real pleasure and honor to partner with partners such as Roche, Sanofi, Bayer, and Merck KGaA, where we bring the best of our capabilities, you know, the Recursion OS, the Recursion, you know, Drug Hunter expertise, and the platform tech expertise, along with the deep biology expertise and chemistry expertise in Genentech, Sanofi, and others. When it comes to Bayer and Merck KGaA, similarly, the second area of value creation that I mentioned earlier, which is challenging targets, developing molecules for them using our chemistry platform, or actually highlighting and nominating novel or undruggable targets from our maps of biology. With the potential here, a lot of work ongoing for over $100 million in partnership milestones by the end of 2026. With that, I'm going to hand it over to Ben Taylor, our CFO and President of U.K., to tell us a little bit more about our financial update. Ben. Terrific. Thanks, Najat. We had a good quarter and ended with a strong cash balance, as we go to the next slide, showing $533 million in cash at the end of the quarter. That was based on not only managing our expenses. At the time of the merger, we made a commitment to our shareholders that we would not only drive a lot of the growth and the programs and the technology that Chris and Najat talked about, but also manage our expenses. You have seen us go from a pro forma burn in 2024 to an expected cash burn in 2026 that's 35% less. That's really our commitment as a management team to making sure that we're doing this as efficiently as possible. We had some great cash inflows over the quarter. In addition to the Sanofi milestone payment, we also had a $29 million R&D tax credit. This is a U.K. tax credit. We will continue to receive this in the future, although it will be smaller as the legislation around it has changed. Our guidance has not changed, and we continue to project over $100 million in partnership inflows by the end of 2026 and managing a burn below $390 million in 2026, so next year. All of that comes together with an expected cash runway through the fourth quarter of 2027. That cash burn number that I gave you does not include any partner inflows or other financing or inflows that would come in. With that, I will turn it back over to Chris. Thanks, Ben. I just want to end by talking a little bit about how we're looking ahead at the future of Recursion. It's been an incredible last nine months post the business combination with Exscientia. We really feel like we've pulled together the best elements of both companies' platforms into the Recursion OS 2.0, as both myself and Najat talked about earlier. Going forward, I think you're going to begin to see us, while maintaining an extraordinarily high bar for quality, bringing unique biological insights identified with our multimodal maps across many different cell types. You're going to see us bring new ideas, new targets, new chemistry. We're going to use our MOA and target deconvolution systems, tools like Boltz-2, our QMMD systems, and even CRISPR screens to help prosecute those exciting targets. We're going to continue to deploy this ClinTech platform to help translate the models and the programs that we're developing Recursion with real-world evidence into programs that can move towards the clinic. We are focused on differentiated, high-quality programs that are going to go where others can't. We're excited for the Recursion 2.0 platform to start to show you some of those programs that are really bringing together all of the elements from target discovery all the way through to ClinTech in the coming quarters and years. Over the next 18 months, we have a catalyst-packed calendar. The second half of this year is looking really exciting, with multiple readouts, including FAP and CDK7, as Najat spoke to earlier. In the first half of next year, we'll be talking about our RBM39 program with early safety and PK from the monotherapy trial. Rolling into the second half of next year, we're going to be looking at both MALT1 and initiating our ENPP1 program, which we were able to bring in recently from our JV with Rallybio. In addition to what you see here from our internal pipeline, we're going to be delivering across all of our partnerships with the potential for additional Phenomap options, the potential for new project initiations, and the potential for programs being optioned by our partners. Recursion is continuing to deliver across both our internal and partnered pipeline while also building the future drug discovery platform that we think is going to help to improve the probability of success, the time, the cost, and the potential of the medicines that we're advancing. With that, we're going to move over to the Q&A portion. I'm going to go to the first question, which comes from multiple parties, which is about our Boltz-2 project. The question is, is Boltz-2 the initiative with a major partner on foundational protein structure modeling that I mentioned at JPM earlier this year? The answer is yes. This is the partnership that we alluded to at JPMorgan. One of the questions here is, why open source versus keeping it internal? We believe that discovering and developing medicines is really, really challenging. Biology is really complex. Chemistry is really complex. There are places where we believe we have a very differentiated advantage, such as with our large-scale phenomics platform and our design platform. These are places where we're going to keep those tools internal. There are other places where we need to be on the forefront, but we believe there are many competitive partners or groups working in the space. In those areas, rather than try to keep something internal that others have available to them, we actually think it's best to help commoditize that particular technology. That's exactly what we're doing with Boltz-2. We're commoditizing our complement, making sure that everyone has access to the kinds of tools that many groups are using, and then keeping proprietary those tools that we think nobody else really has. The second question is, are you still building proprietary models? The answer is absolutely. We were leveraging the Boltz-2 models before they were public. We also have large-scale internal data sets. One could imagine that we could take the same kind of architectures, the same kinds of models that have been built in Boltz-2, and train them across much larger proprietary data sets to give us an internal advantage. The second question, I'm going to go to Najat. The question is from Dennis at Jefferies. Dennis asked for the CDK7 combo expansion cohort in ovarian cancer. What standard of care are you allowing in the trial? Remind us the level of efficacy they showed in terms of OR and PFS. Najat, I'll come to the part two after you answer the first one. Thanks, Chris. Thanks, Dennis, for the question. Great question. The standard of care, as I mentioned during the presentation, will be single-agent chemo plus beva as well. The last that I've seen for that combination, the median PFS was about 6.7 months. The median OS was about 14- 22 months. For us, for the combination, we definitely want to see meaningful improvement to the standard of care. This is a patient population with very significant unmet need. The team will look through in terms of what other points might be more critical as well. For instance, the proportion of patients that reach a certain scan by a certain period of time and so forth. A lot of conversations are ongoing there. We definitely want to see meaningful improvement from the standard of care for PFS. Thanks, Najat. You hit part two there. I'll move on to the next question, which is Brendan from Cowen and Alec from BofA ask, you mentioned the multi-omic profiling that's ongoing for REC-1245. That's our RBM39 program. Do you expect the data from this analysis will in part dictate which patients you enroll in future studies? What data from this analysis would you be able to leverage when targeting or enrolling future patients? Finally, can you point to the differentiation of RBM39 compared to other CDK7 targeting assets? Great. Lots of questions. Thank you, Brendan. Thank you, Alec. I'll start with the first couple of questions, which is the data from this analysis dictating the patients that we'll go into, and then also future patients. As I mentioned during the presentation today, really, instead of having, I think the beauty of the maps of biology, the Phenomaps, the multi-omic approach, and so forth is instead of having like a single screen in a certain area for a certain target, you saw the holistic nature of how you can see the target being important and interesting across different pathways. That was really important for us to understand that, for various forms of replication stress, which can be epigenetic, which can be other areas as well, and DNA repair vulnerabilities are both very important for RBM39 as a target. That was step one. That's actually what helped us for our monotherapy dose escalation to select patients in those areas, as you saw in our press release this morning. The other thing I'll also say is the monotherapy dose escalation is going to be important. We're going to see patients with certain biomarkers recruited and enrolled and so forth. We'll make more of the honing in of where we go based on data that we receive. This is a great way of actually using some of this data, not just for a novel target discovery, but something I've said before, also while you're in discovery to have a better hypothesis of which patients you might actually want to go forward with. 100,000+ patients, and the expansion is because it actually targets a broad, has a potential, I should say, to target a broad set of tumor types that are genomically unstable. Thanks, Najat. The other. Go ahead. I didn't answer the question. I just wanted to make sure I answered that. The point of differentiation in RBM39 and CDK7. Look, RBM39 is not a kinase, right? A lot of the kinases, for instance, as I mentioned, CDK12 has always been for a long time an important oncogenic target. The homology with CDK13 just makes it challenging to really get that selectivity that you're looking for. For us, it was born out of that inspiration of selectivity for a target that's important for DDR modulation, but went beyond much more when we looked at the broader map. Trust me, the map I even showed you today just for DDR pathways, it's a big, beautiful map. It's much broader than that. At some point, I'd love to be able to show you more and what we see there. Thanks, Najat. OK, next question. Brendan from Cowen and Sean from Morgan Stanley ask, for the upcoming FAP data, where do you see the threshold for success in that readout that would give you confidence in the path forward? Given the high unmet need in FAP, do you think the magnitude of polyproduction you've seen to date would support approval and uptake in this patient population if replicated in phase III? Thank you, Chris. And thank you, Brendan and Sean from Morgan Stanley for the question. Yes, look, for FAP, the standard of care, there is no therapeutic that's been approved for FAP. Let me just back up by saying that. Celecoxib and others are used off-label for polyp burden reduction, about 20%- 30% or so forth. We are definitely looking for a meaningful improvement in the polyp burden reduction, and some of the initial data has been promising. However, what's going to be really important for us is to look at the data later on this year, where we'll have a broader patient population. The second question around support for approval and uptake, you know, following the data that we see later this year, of course, it's going to be important to have conversations with regulators. Once we do, happy to follow up and share more in terms of what's going to take from an approval perspective. Thanks, Najat. Next, we have partnership questions coming from Gil at Needham and Sean at Morgan Stanley. Najat, I'm going to send the first one over to you, which is, for the $7 million milestone achieved under the Sanofi collaboration, one of the latest in many milestones we've earned from that collaboration, can you go into more detail as to what exactly was achieved to merit this milestone? Great. The programs that we have, and again, up to 15 programs as part of this partnership, I can't disclose exactly, of course, the target. I can say that this was a challenging target in the immunology space. What we do see is the milestone is focused on lead series, right, actually being able to successfully accomplish that. Next upcoming milestones would be development candidate. I think the point that's important to note is, look, these are all very, very challenging first-in-class, best-in-class targets. To design them is hard. It's not how you do it traditionally. The fact that we've been able to get four out of four so far, knock on wood, somewhere, I think is an important testament to how new approaches can help us and augment what we could do before. More to come over the next 12- 15 months. I think this is one of the interesting things about the tech biospace, Najat and Ben is, or I should say Gil and Sean, is that a lot of the companies in this space that are partnering with large pharma are working on some of the very hardest targets that were not amenable to more traditional approaches. Progress by us and others on these milestones is pretty exciting. Ben, I'm going to turn it over to you now. What visibility, if any, do you have on the potential $100 million in milestones by 2026? Are any assumed in the cash runway calculations? This comes from Gil at Needham and Sean at Morgan Stanley. Sure. Thanks, Gil and Sean. In a way, we have a lot of visibility in the sense that guidance was only based on the existing partnerships and existing programs in those partnerships. Of course, we don't have certainty that those milestones would be accomplished. What we do is we actually look at all of the programs that we know and we probability weight them. This is a probability-weighted number, not the full amount. If we were to take the absolute number, it would be higher than this. We don't include any potential new business development or additional expansion on programs that are not yet identified. Those are two areas where we could grow potential milestones in the future. This is our best estimate that we felt safe in, given the existing business. Thank you, Ben. Next, we're going to go to Dennis from Jefferies and Mani from Leerink, who are both asking questions about our cash runway and how we get to our guidance of Q4 2027 cash out. Sure, absolutely. A couple of important notes here. One, it's really important to always focus on the cash flows when we're thinking about cash runway. If you look at our P&L statement, our operating expenses or our net income actually include a lot of non-cash expenses in it. It's really important to go to that cash flow statement and look down at what is flowing through there. Secondly, all of our guidance that we gave, the $450 million this year, the $390 million next year, is cash-based operating expense and CapEx, not including any partner inflows or new business development or finance. What we do is we then look at what are all the scenarios that could take us forward and get us to 2027. Actually, what we found is there are many different ways that we get to the fourth quarter of 2027. What we felt comfortable with is even just looking at our existing partnerships, like I was just talking about with the milestones, we felt comfortable that operating in a smart way that we are right now and trying to be as efficient as possible with our expenses, trying to really execute on our existing partnerships, and following the same sort of strategy that we have on other cash inflows, including financing, we felt very comfortable we could get to the fourth quarter of 2027. We will continue to move forward. As time goes forward, we'll look to optimize as best we can around those different variables. Thanks, Ben. Final question here from John, who asks or says, we've seen companies like xAI making bold moves, such as investing heavily in compute with millions of chips to accelerate their vision. Can you share how RXRX is similarly tripling down? What ambitious or transformative initiatives are you planning that reflect your next level of thinking? John, thanks. Great question, I think, to end it. First, I'll just say, if you've looked at the State of AI report that Nathan Benaich puts out, you'll actually see Recursion is, I believe, one of the only biopharma companies that's actually listed as the top 20 private or public companies in the world, non-governmental companies in terms of the scale of our compute. Now, we're nowhere near Tesla, xAI, or any of those companies. We really are driving one of the most sophisticated, large-scale compute initiatives in the whole of biopharma. I think that speaks to the kind of ambition that we have for how technology is going to drive this field forward. In terms of other initiatives, I've spoken at prior events, including JPMorgan, about our belief in this field racing towards what we call a virtual cell. This is essentially a computational model of cellular biology that would allow you to predict what would happen to a cell, many different kinds of cells, if you acted on them in any way. You add a protein, you change the effect of a gene or the expression level of a gene, you add a small molecule or multiple small molecules. We believe that building a reliable and robust virtual cell is going to require not just having really good protein folding data, not just having really good atomistic and physics modeling, and not just having good patient data or pathway data. It's going to require having all of those different data layers and being on the frontier of all of those. I think Recursion today, through our partnerships with companies like Tempus and Helix, is really driving the patient layer through our own work at Recursion, building the pathway data with both genome scale knockout maps across more than a dozen different human cell types. As you heard today, with our Boltz modeling and some of our QMMD modeling, we're able to really work at the protein folding and the atomistic modeling layer. I think being able to operate across all those layers is going to be a real advantage as we race towards the virtual cell and deploy early versions of that internally. What's more, we have a team Recursion called the Frontier Research Group. The Frontier Research Group is a dedicated group of folks who are working at the very frontier in high-risk but high-reward areas. While this virtual cell is a part of the work that group is doing, some of the work you heard about today, including the causal AI modeling using Tempus data, actually started in this Frontier Research Group and now has gone into production across the Recursion OS 2.0. These are the bets we make in high-risk, high-reward areas that then get deployed, in some cases, just six or nine months later. I can't tell you about all the things we're doing in that group. I will say one of the areas we think is super interesting, we're watching very closely, is the use of agents to automate the way we discover things and to automate the way we might discover medicines. That's certainly an area that we're working to stay really close to as well. Lots of exciting work is happening Recursion and across the whole field. It feels like a very, very exciting area to watch for the next half decade or so. I want to thank everybody for joining us today. We really appreciate having you. Really appreciate the questions. We look forward to seeing you at the next earnings call or perhaps sometime before then. Thanks, everybody.