Good afternoon, everybody, and welcome to Barclays Global Healthcare Conference in Miami. Do email us or connect with us on Bloomberg for if there are any questions you need to address. My name is Peter Lawson. I'm one of the mid-cap biotech analysts at Barclays, covering predominantly oncology-related companies. Really happy to have up on stage with me management from Exscientia. We've got, Andrew Hopkins, CEO, founder, and Ben Taylor, CFO. Thanks so much today for the time. Good to see you. Good to see you. I guess AACR's around the corner, essentially, more, more interesting data kind of emerging. Great to kind of walk through some of those molecules. I guess the LSD1, the differentiation there and kind of how you design that molecule and how you think you could have a better LSD1. That'd be the first question. Absolutely. I'll have Ben to jump in as well on all these things. It's really great to see actually particularly the progress of our translational biology, Precision Medicine Platform. A continuous stream of new data coming out of our new sort of facilities in Vienna that we just opened. Exemplified by four, you know, I think fantastic abstracts at the AACR that just came out today. Really focusing on, you know, thinking about how we look at the design of our A2A trial and results of that which we announced last year. See that's moving forward. Importantly as well, how we're thinking about more work on the biomarker. Really excited by the Gene Signature, the adenosine signature we're doing there. Also more work on thinking about how we also apply our patient-centric platform to discovery of new targets. This is what's really interesting about our platform, how we can use deep learning, patient-based approaches, actually, to I think to build some of best translational models in the industry. How we can use those not just in the biomarker discovery as we get into the clinic, but also bring it right upstream. This is where sort of a multimodal approach now to incorporating not just single-cell genomics, but also single-cell transcriptomics, single-cell next-generation sequencing, and using machine learning to crunch all our data together then to really help us identify potentially new targets. You know, it's a really good example there. What was also exciting as well, starting to unveil our first data at a conference on our new LSD1 molecule, which we'll talk about in a minute. Ben, what are you excited for the AACR process about? Yeah. Well, getting to your question, Peter, on LSD-1, I think that's a really exciting new compound entering into the pipeline. What we were able to do there is a couple of different indications where LSD-1 is a really important potential target, including small cell lung cancer, as is well known. About 50% of small cell lung cancer patients develop brain metastases. We wanted to be able to develop something that was going to penetrate into the brain, but this is also a target that can lead to thrombocytopenia. We needed a very precise dosing profile, and most of the compounds that had been developed in this space had been non-reversible compounds. We needed something that was reversible and brain penetrant. The problem with brain penetrant, and no one had been able to crack this and certainly not in a reversible inhibitor, was there were three different design parameters that were all fighting with each other. There was the potency, there was hERG signal, and there was brain penetrance. You could solve for two of them, but not all three. This was an example where actually our active learning really shined, because most of you are familiar with our generative AI systems on how we explore the universe. Active learning is actually the AI that we use to try and understand which drugs are the best ones to make, which ones are going to help us explore that space more efficiently. In this case, what it actually did was led us down a really novel chemical direction that said, this isn't the most potent, it's not the most CNS penetrant, but it will give you the most information to make this compound, this series. That ended up being something that we then went on to optimize and became the compound that actually solved all three. That was a great example of how using our systems complementary different forms of AI to really solve a problem that hadn't been able to be solved previously. This has a lot of potential patient impact, but also, is a great sign of the technology. Absolutely. I'll give you an example of that, Peter, that hopefully more people can relate to as well. You might remember when AlphaGo was competing against Lee Sedol, you know. It's the first time, you know, a complicated game like Go was being played against man versus machine. Through the game, there was a move, Move 37, where AlphaGo threw out an really unexpected move, actually led to winning the game eventually and throwing Lee Sedol off his game potentially. This is what we found with the active learning approach. Really good example of thinking how there was a particular dogma, what people thought was needed for LSD-1 chemical design. That dogma actually took you well away from being CNS penetrant. We really wanted to have a first in class differentiated molecule, CNS penetrant. It actually led us to create a new chemotype that allowed us to identify that. It was, for us, it was like our Move 37. It was a really interesting example of how, you know, the use of these machine learning approaches potentially aid creativity. That was a really inspiring moment, you know, trying to reconcile what appeared to be before then, irreconcilable sort of properties in the molecules. What we got to now is a really exciting molecule, which is the first of its kind to be brain penetrant. As Ben said, the practice of a reversible molecule allows much more precise control of the pharmacokinetics in terms of dosing regimes that can then be applied out, control the side effects that Ben mentioned. That's what's really exciting, actually. We're starting to see now how the tech is helping us now create molecules which we think actually lead to, you know, solving very challenging target product profiles. Got you. We should know we kind of think about this as kind of pattern recognition or if you've got other molecules in that space and then you just tweak around it. It seems like you made a step change when you say. Absolutely. making it. Move 37. Go in beyond the where the data is pulling you to think about how then you combine that concept of generative design and effectively active learning is looking for the white space. It's looking for the gaps in the model and looking for what's an edge of domain of applicability of our prediction. That's what's really exciting. It's not just about regurgitating and extrapolating inside the data set, but actually really thinking about how you jump outside the data and allow you to explore new space. That was really important actually in the breakthrough to create this new TPP. Okay. In most of these settings, you need other examples in that space to kind of at least put a framework around that thinking process. That's where bringing in different methodologies can really help. You know, when we talk about MALT1 as well, it's also a really interesting molecule. One of the key things that actually led to that was actually sometimes to break out of certain particular areas. You also want to use new methodology, and this is the first time actually we've really been developing our platform system, not just to use AI approaches, but also have developed now a whole new physics-based approach as well. Importantly, it's integration of that actually. It's the integration of how you use physics-based methods and AI-based methods to help you solve these design problems and really move away into new areas of chemical space that really then try to solve other problems. When we talk about MALT1, there's some really interesting problem there we have to solve as well. Okay. Now, yeah, let's move on to MALT1 and kind of how. Sure. That's differentiated in the space. Yeah. What Move 37 was there? Yeah, no, it's actually, well, it's also philosophically how we think about a compound. When we start development, what we want to do is understand not just what we need to get to in potency and selectivity, but really understand the full profile of the drug. When we looked at MALT1, what we saw consistently across the different compounds in the space was an issue around UGT1A1 inhibition. If you're familiar with UGT1A1, this is something that helps break down bilirubin basically. By inhibiting it, you're actually elevating levels of bilirubin, which can obviously be an issue around liver toxicity, especially when you consider MALT1 is something that you're going to be dosing along with BTK inhibitors in a lot of situations. BTK inhibitors have their own liver toxicity issues, and we've actually seen a number of cases of Hy's law in the clinical trials that have come out on BTK inhibitors. The last thing you want to do with something that's going to have that sort of an issue is give it another drug that's potentially going to lead to a significant increase as well. On a standalone basis, it's an issue. In combination, it could be a very serious issue. We started from the beginning saying, this has to be potent, it has to be selective, it has to have a good dosing profile, but it also has to be clean from this toxicity issue. This was another example where we really just had to go down a different route to be able to find a solution. This is something that we commonly see where a lot of times when there is a lot of existing data on areas that people have explored, we actually have to move away from it because most of the time, if they have explored it, they've probably run to ground the potential for that chemical area. This was another good example of actually a different sort of technology combination. This is a allosteric binding site. What we were able to do was combine not only the generative AI and the active learning that we were talking about before, but really bring molecular dynamics into it. Molecular dynamics isn't something that we talk a lot about, but we actually have made a pretty large investment into it over the last several years, and have a large team running on it. In this case, we were able to combine our AI technologies with molecular dynamics to basically go down a new route and find a really nice clean profile for our MALT1 inhibitor. Absolutely. Super clean against, you know, whole kinase panels, you know, several panels. This particular, you know, UGT1A1 issue, we screened a lot of molecules, a lot of examples from patents against competitor agents. Very much was seen as against the whole class. That's why it's important for us actually to think of this as a key differentiation, particularly when we think about potentially how these compounds could be dosed and the likely combinations that they're going to be used in. We are really excited again by the differentiation we see in this particular molecule as well, particularly from a patient benefit point of view. Got you. Okay. I'd love to kind of scoot back to the existing pipeline. Just the questions we have around the. The adenosine antagonist, kind of what we should expect to see over the next six months to 12 months, around that phase I data. Yeah. We have got approval from the authorities now. We're starting to. Interestingly enough, we are aiming for a combination study, actually, with PD-1. The goal is to look for patients which have been refractory, get resistance, particularly in RCC and also lung cancer. What's interesting, this has been driven by two factors. One of them is like we got good phase 1 safety data, which allowed us then to make the argument going into combinations. Looking at the science. You might remember at the end of last year, we started to publish work on the new sort of gene signature of the Adenosine Burden Score, which we think actually is potentially the way to identify actually signals in the immune system that would identify who could be real responders in this case. What we found actually from a large analysis that this Adenosine Burden Signature was inversely correlated with one of the key sort of PD-1 signatures used called the Tumor Inflammation Signature. Actually, where people scored well in potentially the Tumor Inflammation Signature should be responsive to PD-1s at a low score for the adenosine signal. Those who had a high score for the adenosine signal will then show the opposite, low in PD-1. It's an interesting correlation. That data together allowed us then to build the argument that the real use of this, we believe now, is going to be in thinking about as a targeted combination for particular patients. Basically what we're trying to do is to think about how you can reignite, you know, the activity of PD-1s, potentially what could be a high adenosine environment. That's the sort of the hypothesis of the trial. The trial itself is actually split into. There's actually two things we're trying to validate and test in hypotheses. One of them actually is the molecule, learning about how it behaves in patients, how we expand the dose and safety and the efficacy, of course. In parallel, also to look and validate the biomarker. Of course, once that's validated, then we can bring that in either directly as part of the inclusion/exclusion criteria or as part of the next trial as well. Ben, there's also some interest in the ways our. Yeah been designed, isn't there? Yeah. This is something that you'll see in both A2A and CDK7, is how we can be more adaptive and continuous in our clinical trial design. For example, in our phase I, rather than doing a traditional 3+3, which basically you're just waiting to see when you have two out of the three patients having some sort of serious adverse event, and then that's helping to define your MTD, what you can actually do is look at the progression. Every patient enrolled is actually contributing data into understanding, is this a safe drug or not? Is this a safe dose or not? You're actually building the understanding of the drug as you go. This is actually a much more refined way of determining a statistical guideline for when a drug is safe or not. Then it automatically actually transitions then into a efficacy study. That'll be a adaptive designed efficacy study as well. In parallel to while we're running this confirmation on the biomarker, we are also going to be running a adaptive continuous trial design. Got you. The existing trials that you're running, are you modifying those or You're still thinking potentially having single agent activity with adenosine or? Yeah. Yeah. For adenosine from a clinical perspective, you're unlikely to use it in a monotherapy setting. This was actually a really interesting point that we had discussed with the regulators as well. Our first healthy human study looked at it in monotherapy and measured the PD and PK from that trial. We saw good outcomes from that that matched very much what our modeling platform had determined. What we're doing is we're actually going directly into the same sort of setting that you'd use it in a clinic, which is in combination. To make that signal even stronger, what we're doing is going in relapsed refractory immunotherapy patients, re-challenging them with the same immunotherapy, but now giving them the A2A instead. There is some potential to use A2A in monotherapy, but realistically, in the clinical setting, everything's going to be in combination. Got you. Okay. Do you think you need to demonstrate any single agent activity to kind of help validate? So- The screen? What I'll say is the regulators are allowing us to move forward with this trial design. It would be very powerful if we saw responses in patients who are relapsed refractory in the immunotherapy agents. Absolutely. I mean, it's a massive opportunity, right? Absolutely. Relapse refractory load. Okay. I think another important note there is we actually see, the biomarker being, from the immune system, not from the cancer mutations, which actually means you're talking about almost a comorbidity or a predisposition for these patients. It's something that they're carrying around in their genetic structure, in their immune system, that when their immune cells are introduced to this tumor microenvironment, they're resulting in the high adenosine levels. That means it's actually not about a cancer indication that you're talking about. You're actually looking for a patient genetic predisposition. Perfect. No, that's very clear. CDK7 kind of, I guess the You've got a differentiated molecule, kind of how did it originally come through to that, and then kind of when should we expect to see data? Is that early 2024 or late 2024? Probably. Firstly, to describe the molecule, I think it's worth going back. It's really excited again by exquisite design first, designed in less than 150 molecules in less than a year. It's designed specifically again to be a sort of a non-covalent inhibitor. That was important again for controlling neutropenia. We want to really try to minimize the immune effects potentially you get from CDK. That was a key part of the design process. Also then to have exquisite PK. Again, allowing us to control dose, as we talked about with LSD1 earlier. This idea then about controlling dose, understanding of... Really did a lot of work then to really understand sort of, you know, what's our IC50 on target, you know, for about six to eight hours. Effectively, you know, try to give yourself a bit of a drug holiday for the rest of the day, potentially. That we found to potentially be sufficient in the modeling to get the effects we were looking for. This is a potent and selective molecule, and that's some things which some other molecules in the clinic haven't quite achieved the right selectivity and potency required. Importantly, avoid some of the other key class problems we saw for these, which was efflux issues. Efflux, when you think about it, is really important when you think about, you know, the gut toxicity that some people have been observing. Not surprising necessarily when you think about the recirculation of a drug and all of the immune cells actually that are present in the gut as well. This is an important sort of design angles, trying to build in some safety windows like this. That's an important aspect to this. Second important aspect, as Ben was describing with the signatures we've been developing for 5-HT2A. Taking a similar approach to our precision medicine platform and again, we at the end of last year, publishing the first of the posters, showing potentially that we have a, I think a really exciting potential gene signature that really allowed us to identify sort of ex vivo responders and non-responders in our platform. Really drilled down then at a multi-modal, multi-omics approach. You know, looking at transcriptomics, single-cell sequencing and the single-cell phenomics. Allowed us then to really distinguish between when we then go back and do prospective studies in our ex vivo biobank samples to say, this actually potentially is a signature that distinguishes. It's about eight to 10 sort of genes in the signature. Now, interesting about that, as Ben described, trial design, very similar in sort of an adaptive learning approach. Also looking to test two hypotheses, looking to test out the biomarker hypothesis in parallel to the testing the safety and efficacy of the initial drug. One advantage we have here is that it's potentially a wider, broader range of indications we could test CDK7 for. Again, looking what we developed is a cancer-agnostic biomarker to allow us to be. Because of that, potentially, it's already with the authorities at the moment. We're hoping in the first half of this year we'd be, or at least this year, we getting to a first dose in patient, get approval, first half of the year. If that is the case, it's possible that even though it's starting slightly behind by about a quarter of than an A2A, but at the same time, potentially, recruitment could be a little bit faster. I don't want to give guidelines on the exact readout. I think sometime in 2024 probably. Because in many ways, the readout is determined by the statistical significance of the adaptive trial. It tells us when it's ready to read out when it's got the strong signal. The interesting thing is, I think both of these trials now, starting almost to be thought about running in parallel and the timing could be one or the other, depending on the signal. I'll throw a third one in there as well, which is PKC theta, which BMS had in-licensed and just started dosing. That's right. in February. That is, immunology drug, so it's going to be a healthy volunteer, type of first study. Run a more traditional phase on We'll see that, probably along similar timelines, to the initial readouts from, the other 2 as well. Yeah. Good. That's another really exciting target because it was great validation for the capabilities that we have in-house. That was one where 15 different companies disclosed projects trying to design a selective PKC theta inhibitor. Only two made it into the clinic. One was a pan-PKC, which had a lot of issues because of obviously hitting more than just PKC theta. You were sending off all sorts of immunology signals. Inflammatory signals. The other part was a drug that went in that was selective, but wasn't able to reach dosing because of potency or other issues. For us to be able to come in where, many different large pharma have tried to design around potentially a very large value, drug category, and having failed, but we came in and in a little over a year and, you know, 150 compounds were able to solve that problem, I think is a great statement on the capabilities of the company. Peter, it's been a great start to the year actually. First half of the year, we're looking at, you know, two trials in patients kicking off, getting approval, moving forward. We're looking at one of our licensed products with our partner, BMS, moving now, starting phase I. We're looking at the announcement now of our pipeline in precision oncology, doubling in size and expanding with the LSD1, MALT1 inhibitors. I think we've seen real momentum now in the clinical pipeline of Exscientia. Got you. Maybe, what we got, 20 seconds, but just the long-term vision here. I mean, w ould you eventually out-license most compounds or do we model at you as, you know, there's two or three compounds you're going to move forward yourself kind of? I'll give you the 10-second answer and then Andrew can add on if he wants. Okay. The good news is we don't have to decide now. We can continue to play these things forward. We have enough cash in the bank to be able to move forward all of the programs we've talked about without driving another financing behind it. The real question comes in, when do we have a transformative valuation point? We are fine either taking them forward ourselves or potentially partnering them, and that'll be determined on valuation and return on investment. Perfect. Into extra time, so thanks very much. Thanks so much. It's always a pleasure. Pleasure.