Hi, everyone. Before we begin, I'd like to draw your attention to the disclaimer statement, the notification on the screen and also on the printed materials. Good afternoon. My name is Joanna Shields. I'm CEO of BenevolentAI, and I'm delighted to be with you here today. For many of you, it's the first time you've had the opportunity to hear from the company and to meet our leadership team. We hope this is going to be a long and you know successful relationship with all of you, so we look forward to getting to know you. I've been leading BenevolentAI for the past four and a half years, and during my tenure, we've been heads down. We've been building a technology platform that is AI-enabled and has resulted in a substantial pipeline. Before joining Benevolent, though, I've spent many years in the technology industry. Over three decades, I built and scaled technology products, platforms, and operations globally for companies that you'll recognize, like Google, Facebook, AOL, and Bebo. But just before Benevolent, I also spent five and a half years in the U.K. government as the Digital Economy Minister for the U.K., or the Digital Economy Advisor to Number Ten Downing Street and the Minister for Internet Safety and Security. I was responsible for cybersecurity in the U.K. The reason I tell you about my background is the context from which I decided to join BenevolentAI. I have seen technology disrupt markets over the course of my career, but I've never seen the potential of BenevolentAI and what we're trying to accomplish. This endeavor is by far the most impactful and important thing I've ever had the opportunity to work on, and really excited to share it with you today. BenevolentAI has evolved into a leading clinical stage AI drug discovery company with a unique and differentiated approach that focuses on harnessing advanced AI and machine learning to decipher complex disease biology and discover novel drugs. Today, we will share with you how our work tackles some of the most significant challenges facing the research and development of the pharma industry and how this approach and the work that we are doing can be transformational for millions of patients suffering from diseases without adequate treatment. The Benevolent drug discovery platform is commercially and scientifically validated. It has generated all of the in-house pipeline of 13 named drug programs and 10 exploratory stage programs. Our most advanced asset, which we'll share with you today, is for atopic dermatitis. It's in phase II clinical development, and we also have a novel asset for ulcerative colitis in IND-enabling studies. We will talk more in depth about these programs later. Commercially, our platform continues to deliver in ongoing collaboration with AstraZeneca. As of now, three novel targets identified by our platform have been validated and selected for entry into the AstraZeneca portfolio, which is for us an important revenue driver, but it's also external validation of our scientific and technology leadership. We have demonstrated the predictive power of our platform recently. It was early in the pandemic when we identified and published a hypothesis that an existing rheumatoid arthritis drug could be used to treat COVID-19. Our findings were later validated in multiple randomized clinical trials, then and ultimately resulted in approval by the FDA for treatment of hospitalized patients. Now, identifying this existing drug so quickly and accurately predicting its potential to address some of the implications of COVID-19 was really an illustration of how the potential of AI can enhance the understanding of disease biology and accelerate discoveries. We believe that this enhanced understanding of disease biology that our platform provides is what enables scientists to discover new medicines with a higher probability of clinical success, and we're excited to share our progress with you today. First, I'd like to start with why we do what we do. In a world where so much has been reimagined by technology, we should be asking, why are there still thousands of diseases with no effective treatment? Under the traditional pharma model, it takes approximately 10 years and costs an average of $ 2.6 billion to bring a drug to market. The discovery and drug development process currently has a 96% attrition rate, and 30%-50% of top-selling drugs don't work effectively for the patients. This approach is not sustainable. A key reason for these failures is a lack of understanding of the underlying biology. Today, given the exponential growth in the availability of biomedical data and research, we have an unprecedented opportunity to fundamentally rethink drug discovery. To give you a sense of the scale of information that researchers have recently deposited 4.5 PB of data to the US National Cancer Institute's Genomic Data Commons. Now, this is just one of thousands of open data initiatives. It won't be long before we're talking about exabytes of data being generated, which is one with 18 zeros or 1 billion GB to give you some context. Now, when you add to the mix the tens of millions of new publications, patents, clinical trial data and other databases, the challenge to make all this data accessible and useful for scientific inquiry is enormous. Now, we as humans are limited by the amount of information we can absorb and process. So how can we make sense of all these diverse and rich data resources and enable scientists to put together the clues and relevant connections to make life-changing discoveries? This presents the perfect AI and machine learning opportunity and is precisely the problem we set out to solve, and we are making great progress. At Benevolent, we see enormous opportunities hidden in data. We believe even a single data point has the potential to unlock the mysteries of disease. We harness the power of this vast and growing corpus of biomedical data and make it accessible and useful for scientific inquiry, uncovering novel connections that the human mind alone might not find. We are unique in our scope and ambition as we deliberately choose not to restrain our platform to any one data type, therapeutic area, modality or experimental methodology, all of which have limitations and biases. Instead, we take a multimodal approach with a diverse and scalable data foundation consisting of over 85 data sources and tens of millions of peer-reviewed papers to create a comprehensive view of disease biology. Our transformative drug discovery platform consists of products and tools that allow scientists to reason across this vast and growing corpus of biomedical data, ask and answer complex biomedical questions, run in silico experiments in real-time, and advance hypotheses. We model human disease for exploration and inquiry in unique ways due to the scale and diversity of data in our flexible architecture and our platform. Our platform has the potential to work in any disease or drug modality, from small molecules to biologics to cell therapies. Our discovery engine is prolific. It's designed for scale and consistently to generate novel disease targets to support both our internal programs and our external collaborators. We follow a philosophy at BenevolentAI of building technology in the service of science, meaning the needs and requirements of our drug discovery scientists are central in developing our platform. Our team of over 350 work together in cross-functional teams with experts in biology, chemistry, informatics, drug discovery, working side by side with AI and machine learning and data scientists. Now, they work together to leverage our discovery platform and spark new ideas and creativity. We are also supported by a board of industry luminaries. Our chair, Dr. François Nader, is a veteran Biopharma Executive who also serves on the Board of Moderna and is Chairman of Acceleron. We have new non-executive directors, including Dr. Olivier Brandicourt, the former CEO of Sanofi, Jean Raby, former CEO of Natixis, Sir Nigel Shadbolt, Principal at Jesus College at Oxford and Founder of the Open Data Institute, and Stanford University ethics expert, Dr. Susan Liautaud. With EUR 300 million invested in our platform to date and capital raised from our listing on Euronext Amsterdam in April, we are in a solid financial position to progress the development portfolio towards the clinic and also to increase investment in our technology. I'm delighted to be joined here today by our executive leadership team, and I will hand it over now to our COO and Co-Founder, Dr. Ivan Griffin, who will give you an overview of the market and our business model. Oh, okay. Thanks, Joanna. I'm Ivan Griffin, COO, and one of the Founders of Benevolent. My background is I was a neuroscience researcher at Oxford, spent several years as an investor in life science businesses before doing a small stint in government, setting up something called Genomics England, which is a U.K.-wide DNA sequencing initiative linking genomic data with NHS medical records for people with cancer and rare diseases prior to starting Benevolent back in 2014. I'm gonna give a quick market overview of this emerging AI drug discovery sector for those that are new to the space before diving into, in a bit more detail, our approach and business model. The use of AI in drug discovery is growing rapidly, and to quantify this, here's a chart from RBC looking at the amount of investment that's gone into the space in the last seven years, totaling about $13 billion, which covers, you know, equity, collaborations, and M&A. In terms of what's being done with this investment, on the right you see a chart by Boston Consulting Group looking at the internal pipelines of AI drug discovery companies. Here you can see large growth in the number of assets in preclinical and increasingly going into the clinic. At a technical level, there's three sort of broader trends that are enabling this to happen. Firstly, advances in machine learning algorithms. Secondly, increasing availability of biomedical data that can be queried by these algorithms. Finally, an increase in computational power, meaning that less time is required and also more complex calculations can be done. The next question, therefore, is: Why is all this money actually going into the space, and what's the benefit going to be? We see there's two main areas where AI approaches can add value in pharma R&D. Firstly, direct R&D cost savings and secondly, and in the longer term, much more importantly, increasing the probability of success. Taking each in turn, industry data shows it takes an average of five and a half years and $33 million to go from the start of chemistry to the clinic. Internally, we're tracking more than 50% cheaper and at least two years faster than this. One key thing to note is this data starting at chemistry ignores the many years that goes on in pharma R&D in basic biology and target discovery that no one has been able to quantify 'cause it's so large. We're tracking one year overall from start to finish in terms of delivering targets into our portfolio. However, the much larger impact of AI will lie in improving probability of success, and this is because the reason you read it takes $2.6 billion to develop a drug is not because that's the cost of the one that gets to market, it's because you pay for all the failures along the way, and you often fail very late and also very expensively into the process. Phase II is the stage which has the highest clinical attrition, where approximately only 30% of programs succeed and half of phase II and III trials are due to lack of efficacy. Our belief is the most critical element is to get the right target at the outset, as without that, you're doomed to failure no matter how good a molecule you develop. Look at this financially, this is an illustrative model which shows what a 25% improvement in probability of success at each stage in the clinic can do to the economics of R&D. Put simply, if you improve each clinical stage by this amount, you more than double the chance of a drug that enters the clinic of getting to market, meaning rather than needing nine drugs, you need four, and more than triples the NPV of the program. Now, of course, as context for how realistic this actual 25% boost is, it's worth noting that already phase II trials that use patient preselection biomarkers are 50% more likely to succeed. Also as an added note, as part of our business combination with Odyssey, a third-party consultancy did a lot of interviews with industry executives and peer companies, and the broad consensus was the potential improvements of up to 45%. Now we can just look at which part of the drug discovery process that various companies in this space are active in. This slide shows some work done by Oliver Wyman on the sector, and although a simplified view is actually quite a helpful place to start, it looks, starting from the left, firstly, which area of the process and which problem are they trying to solve. You have companies that are focused on target ID, where it's identifying targets at the start of the process, where you have a lot of complexity due to poor understanding of disease biology, and the other sort of broad group of companies which are more focused in the chemistry. Once you know the target, you're looking to optimize or design a molecule, and for that, where you have a lot of computational complexity. The second dimension relates to the overall drug discovery approach, whether you take a sort of hypothesis-driven experimental method, where you specify a specific hypothesis you're going to test, the alternative being sort of non-hypothesis phenotypic screening, where you test many compounds and you see what hits, and then later on try and work out what's driving the effect. I should add this Oliver Wyman analysis really looks at where the company started, and it's possible, and many companies are active at multiple points in the value chain. Lastly, just to comment on pharma companies and what they're doing in this space, and here we see quite a bit of a mix. Some companies are just starting to explore how they're gonna use AI and data in R&D. Others are developing quite a lot of internal capability, and then others are doing collaborations with players such as us to see what works. Overall, we feel this is an emerging sector with large sort of long-term structural tailwinds behind it, so there's plenty of room for different approaches, and it's not a zero-sum game. As you can see, we're very much focused on the biology and identifying targets as a key first step. Let me now come on to talk about our approach and why we've gone about things the way we have. We start from the place that the idea that the best place to start understanding disease biology is to ingest, organize, make available as much biomedical data as possible. This includes data from the literature, patents, genetics, chemistry, electronic health records, proprietary data generated by ourselves and our collaborators, and also what we also have is AI methods to bring this data together and then infer new knowledge and new insights that are only available to us. One really important point which Daniel will come onto is get well is made a conscious decision to integrate as many independent and orthogonal data sources as possible on the basis that with all these independent sources of data, we can get a higher confidence of what the true biological signal is among the often contradictory evidence that exists. We then use our predictive algorithms to predict targets for whatever disease we're focused on, and then we've built a series of software tools to help our scientists explore those targets, make decisions in a data-driven manner around each step of the process, explore which ones to progress, look, consider things such as safety, opportunity to differentiate, et cetera. We send our targets to our Cambridge lab for experimental testing, or in some cases, use academic centers where they have specific assay systems we want to use. We'll be hearing about that later today in our ulcerative colitis program. Finally, this experimental data is fed back into our graphs, so we learn over time and improve. I now want to touch in a bit more detail on why we actually chose this approach and the benefits we think it can bring. Down the left-hand side of this slide are what we feel are some of the key underlying challenges behind poor R&D productivity in drug development. First, and most importantly, is the fact that understanding biology seems to be the rate-limiting step, accounting for approximately half of late-stage failures. Digging into this a bit deeper, one factor is scientists tend to be only focused in a single therapy area, so your job is to identify targets for pain, for example. They're unlikely to look across the scientific literature into other diseases where there may actually be commonalities and across mechanisms. Another data factor is that each type of data only gives you part of the picture. Genetics, for example, is very useful, but it doesn't actually tell you what proteins get expressed and then what the ultimate phenotype is. You need the proteomics and clinical data as well. In terms of our platform, we consciously focused on understanding biology at the start as that's where we felt there was the biggest ROI, and we bring together biomedical data across all diseases and all data types to overcome these silos and biases. As a consequence of this means we can actually work in any disease area and also any modality because fundamentally we're trying to work out what the right target is in the first place, and then the next step is whether we should prosecute it with a small molecule, antibody, siRNA, et cetera. This also enables our scientists to find novel targets, either by linking things that already exist but no one's actually linked, or also inferring things that should be known but haven't been discovered yet. We've got examples of both in our portfolio, and this leads you to both first-in-class and best-in-class programs. Another challenge in current R&D, especially as I've said around brute force high-throughput screening or phenotypic screening approaches, is that you can rapidly generate hits, which is great, but then you have a long period of target deconvolution to actually understand what's driving those effects. In this case, again, as I've said, we take a very hypothesis-driven approach, so we can understand the effects in the lab and move forward quickly. Another element that gives us speed is our platform and software tools, which Danny will show later, which bring all the information scientists need to make decisions together in one place. Finally, fundamentally, we see target discovery still very much based on a sort of small number of disease experts, each with their own biases and favorite targets, reading papers and going to conferences, rather than being tackled in a holistic and systematic fashion. By developing a software approach across disease areas, we can enable scale, learning, and improvement from each time we look for targets. Of course, coming back to the original point, the whole goal of doing all of this is to increase the probability of success of the programs we pursue. To convert these potential benefits into numbers, firstly, on the probability of success or how likely are targets to succeed. Obviously, we haven't run 20 clinical trials, but we can say we have a 23% conversion rate in disease-relevant assays of targets our platform predicts, which is high. We can also point to our Nature paper from 2020 showing our platform is highly likely to identify subsequent clinical successes. As shown in the illustrative POS model earlier, this can lead you to an almost twice improvement in getting to market. In terms of cost, internally, we track from going from target to candidate in two-two and a half years, compared to an industry benchmark of four and a half years. Also, as I mentioned before, no one has quantified the many years pharma research units spend looking for targets, whereas we have been shown that can be done in a year. Finally, in terms of cost to go from target to IND filing, internally, we're looking at $15 million, greater than 50% savings versus industry. Overall, we feel this shows our approach can dramatically increase R&D productivity. Finally, looking at this compared to others, L.E.K., a life science consultancy, did an analysis of the number of IND filings per year by major pharma and biotech, compared to where we are currently trending and where we want to go. This chart shows that we are potentially able to match the R&D productivity of much larger companies with a vastly higher market cap at a fraction of the cost. Because as described earlier, our platform is scalable and flexible, so the only incremental cost is on the development side, not the many years generating targets or pursuing projects that ultimately fail. Now let me come on to our business model in a bit more detail. As you've seen, our technology platform can generate high-quality drug targets, and we have strong in-house capabilities in taking these through discovery and into development. This slide covers at a high level our business model for exploiting this. In terms of our in-house pipeline, we plan to keep our drug programs to at least the start of phase I studies, and we will then take a balanced approach in deciding which drugs to take through into late-stage clinical development and which to out-license. I'll come to how we make those decisions shortly. In addition to our own programs, we will also target a small number of additional platform collaborations similar to our AstraZeneca deal. These collaborations provide non-dilutive funding, further validate our platform, and importantly, we can leverage the data generated to enhance our platform to the benefit of our in-house programs. We also have non-commercial collaborations where we put our platform to good use for wider societal benefit. Our COVID-19 work, which identifying baricitinib, now FDA-approved, is an example of this, as well as our partnership with the DNDi, looking at repurposing approved drugs for dengue fever, a major healthcare burden in the developing world with no cure. Given our flexible model and ability to work in any disease area, the key question is how do we decide which diseases to work on in the first place, and then how far do we take those programs ourselves? On the left of this slide are the key factors underpinning how we make these choices, where overall, we are looking for the overlap of all these factors for diseases we will pursue. The first is scientific, and what I mean by that is we will focus on complex multifactorial diseases where uncovering the root cause is the key issue, which is often the case in, say, neurological or immunological disorders. This is opposed to monogenic disorders, where the target needing to be modulated is well known, so it doesn't play to our strengths. Secondly is competitive and commercial, and by this we mean the level of unmet need and also whether disease is already fairly well served by heavily genericized products. For example, ALS obviously has a huge unmet need with no cure, whereas insomnia would not be something we'd pursue. Thirdly, we look at factors relating to clinical development. This is in terms of the cost of trials and also time to see clinical proof of concept, which can be highly variable across diseases and depends on whether you can measure biomarkers of response, among other things. The final factor is relating to data in our platform, whether we have or could incorporate some data sources that could give us a specific advantage here. There's certain ultra-rare diseases where hardly any data exists, so those we wouldn't go after. Historically, we wanted to demonstrate our technology works in a wide variety of disease areas, hence the broad pipeline we have today. However, as we have grown our pipeline and now start moving more programs into the clinic, it makes sense to narrow down and focus on a few key therapy areas. Going forward, we're gonna be coalescing around three main areas: immunology, neurology, and oncology, where there are several diseases that satisfy our criteria well. Within these, we'll look to either take things into early clinical development and out-license or take some into later stage development and potentially beyond. This will be done in a staged manner, especially when considering any future investments in late stage or commercial infrastructure. That is a few years ahead of us in terms of needing those capabilities. We also have a small number of existing programs outside these therapy areas, and in these cases, we will look to take them to a point where we can find a licensing partner, but not beyond IND filing. In terms of therapy or disease areas we don't want to work in, this is where we'd look to collaborate with pharma companies from the start, like we have with AZ, as this overcomes many of the factors preventing us from doing it ourselves. Heart failure with AstraZeneca is a great example of this, complex disease, large unmet need, but not something we would do in-house given the massive clinical development timelines. Also, in this category are non-commercial collaborations, such as what I outlined earlier with the DNDi. I want to touch briefly on some key validation and evidence that our approach is working across three separate categories. Firstly, our internal pipeline, most advanced assets in phase two, which is all developed using our technology as opposed to being acquired or in-licensed like other companies often do. Secondly, strategic validation by delivery on our target partnership with AstraZeneca, which now covers four of their key disease areas. Finally, regulatory validation, that our platform is the only one of many people who tried to identify a treatment for COVID-19 that was proven out in the clinic and received full regulatory approval. This slide provides a snapshot of our drug portfolio from target identification all the way through to the clinical development. All of these programs have been generated and developed in-house, or in the case of CKD and IPF, in partnership with AstraZeneca. Our most advanced asset is in phase II for atopic dermatitis, after which we were looking to find a partner. Following behind is one of our ulcerative colitis program, where PDE10 is a target with zero linkage in the literature to ulcerative colitis. Our portfolio comprises a broad range of disease, and we have examples of both best-in-class and first-in-class programs, but Anne can cover this in more detail later. We have a very successful collaboration with AstraZeneca, initially focused on identifying targets in chronic kidney disease and idiopathic pulmonary fibrosis, but expanded at the end of 2021 into heart failure and lupus. This collaboration involves combining our technology and AI approaches with AstraZeneca's internal data and scientific expertise. Specifically, we've created a separate cloud instance of our platform in a secure environment, which AZ can then deposit their internal data for us to integrate with our knowledge graph, run our AI models. The targets that come out then go to AZ for experimental validation and progression. In terms of milestones achieved to date on this partnership, we've already delivered three novel targets into the AZ portfolio across CKD and IPF. Both diseases we've never worked in before and experimentally validated to sufficient quality to enter the AstraZeneca portfolio. On the back of this success, we expand the collaboration into heart failure and lupus at the end of 2021, and as part of that, AZ made an investment into our PIPE financing, so they're now a shareholder as well as a collaborator. Financially, these agreements are structured as typical platform collaboration deals, upfront payment, FTEs to cover our cost, and then milestones and royalties. We retain upside as the programs move through clinical development. Lastly, our COVID-19 work is a great example of the power of our platform and what can be done by combining public data and our AI approaches. In a bit more detail, if you go back to the start of the pandemic, given the urgency of the situation, we looked for already approved drugs that could be used to treat patients. We looked for something that would be anti-inflammatory to stop the cytokine storm and also antiviral to slow disease progression. We identified baricitinib, JAK inhibitor owned by Eli Lilly, and approved to treat rheumatoid arthritis as the top-ranked drug. Baricitinib is a well-known anti-inflammatory, but the reason we focused on it versus all the other JAK inhibitors or other anti-inflammatories was because of an off-target effect our platform discovered that suggested it could also have an antiviral effect. This was something that not even Eli Lilly, the owner of the drug, were aware of. This was some work we completed over a weekend using only the data in our knowledge graph. It was only so fast because we'd spent over five years building the platform in the first place. We published this in The Lancet in February 2020. Since then, the utility of baricitinib by its dual anti-inflammatory, antiviral approach was validated experimentally and in three large placebo-controlled randomized controlled trials run by NIAID, Eli Lilly, and also in the RECOVERY trial in the UK. It's been shown to be the most effective therapy for reducing mortality in hospitalized COVID-19 patients. The Lilly trial showed baricitinib delivered 38% reduction in mortality, rising to 46% for those on supplemental oxygen. Baricitinib received emergency use authorization from the FDA in November 2020 and full approval earlier this year. This work led Eli Lilly to invest in Benevolent in 2020, and it's reported that there have been more than 80 different artificial intelligence attempts at repurposing drugs for COVID-19, and this is the only one that's worked. That concludes my section. Now to help you understand our platform better, we asked our Director of Engineering, Olly Oechsle, to record a short animated version of the onboarding overview we present to all new employees, which helps to connect the key parts of our platform from data processing ingestion through to novel target identification and ultimately the clinic. Hi, I'm Olly Oechsle, Director of Software Engineering at BenevolentAI. I'm here to give an informal tour of our system, which is something that we do for all of our new starters at the company. We'll touch on various points already highlighted by Danny and visualize how it works from one end to another, from data through to drugs. On the way, I will trample through some scientific jargon to make it more real. I recommend you wait for Anne, our Chief Scientific Officer, and Professor Tom MacDonald to do full justice to the science later. BenevolentAI harnesses the power of biomedical data to evolve the way drugs are discovered. We do this by following a process that takes us right the way from data through to drugs. In more concrete terms, from relevant biological information on the left side of the picture, to the understanding of disease biology at a cellular level and through to drugs in the clinic. We're gonna take a tour of that process starting from the beginning. It all starts with data. Having the right data, covering the right topics is crucial for any algorithm expected to create useful insights. We invest in the most comprehensive view of biomedical data we can. It's designed to power the insights of AI models and to assist our team of scientists in making the right calls. It starts from around 85 different data sources covering fundamental concepts, like 23,000 known diseases, all 20,000 or so genes in the human genome, and complemented with compounds, drugs, processes, cell types, tissues, and many more. Crucially, the millions of connections that describe the ways that each of these concepts interacts inside a human system. The central store for these millions of concepts and connections is our knowledge graph. A single graph for all diseases gives our technology the widest possible view and our scientists access to everything in one place, and it underpins our ability to work across different areas at once. Far, so abstract. To see those graph connections in action, let's look at an example based on some information from our scientists who tell us, one, the messenger molecule cGMP has various effects in the body, including in inflammation. Number two, inflammation of the colon is associated with the disease ulcerative colitis. These facts can be represented together in the graph, including nodes for the messenger molecule cGMP, the inflammation process, ulcerative colitis itself, and colonic tissue. These nodes are then connected by relationships showing that cGMP modulates inflammation, in turn associated with ulcerative colitis, which occurs in the colon. This is now data that machines and humans can begin to reason over. Millions of facts, like the previous illustration, are described in the scientific literature, which often goes beyond the information available in curated datasets. The 30 million or so articles in the biomedical domain are processed using BenevolentAI's natural language processing architecture, which converts text into connections. These connections are scored and aggregated by their semantic meaning before they flow into the graph. Altogether, this builds the graph's strong basis in fundamental science. We still want more. The ideal mix must also weave in insights from patients, for example, transcriptomics, genetics, or other experimental insights. In other words, data derived from real people suffering from a disease. Indeed, studies show that drug programs grounded in patient or genetic data have a higher chance of success. Over the last few years, we've increasingly exploited genetics and genomics data to add new information to the graph, including patient groups, comparisons between diseases and controls, and differential expression of genes, all of which have grown the graph significantly. Let's turn to how the graph is used by our scientists. BenevolentAI works on disease areas that have high unmet need, suitable data, and commercial opportunity. As an example of this, let's return to the inflammatory disease, ulcerative colitis, a chronic lifelong disease affecting the digestive tract. In 2018, BenevolentAI scientists began research on ulcerative colitis, so we're gonna continue the tour with the process, of UC in mind. Our scientists start with access to a range of purpose-built tools within the BAI platform, which are built on top of our graph. Our users and collaborators can interrogate the graph for key concepts like ulcerative colitis, its related proteins, mechanisms or treatments. They can visualize these proteins in a network to get a better understanding of the processes and pathways involved. Using this, scientists can prepare their biological questions they'd like to pose to our machines. These questions ask, in different ways, for novel targets associated with a disease such as UC. But why do we take such an interest in novel targets? Most clinical trials fail because of a lack of efficacy. This can be because the drug did not target the optimal point of intervention that would have reversed or corrected the process leading to disease. By contrast, BenevolentAI disease programs always start with a comprehensive overview of novel or optimal targets, as proposed by a suite of AI models, which we call a fleet. This is something that our machines are uniquely patient at doing for us. One AI model in the fleet, called Graph Embedding, learns directly from the graph and is able to propose targets by ranking the entire genome from most likely to least likely involved. Another model, Enterprise, makes predictions based on English language queries. A scientist types a question, and Enterprise answers with suggested targets. The value of AI models is that they leave no stone unturned. As we've just seen, they can evaluate the entire genome in one go. By contrast, human scientists bring practical knowledge and expertise in creating realistic hypotheses that can grow into successful drug programs. The two come together in Benevolent Platform's triage tool, where scientists evaluate the predictions suggested by the AI model fleet. The triage tool uses the graph to provide further context around each prediction, covering progressability, toxicity, and other key aspects, allowing scientists to make informed decisions about what should proceed into assay experiments. For ulcerative colitis, one of the predicted targets was PDE10, and this stood out because the protein PDE10 was not previously associated with ulcerative colitis in the scientific literature. It was previously most associated with its activities in the brain and diseases like schizophrenia. The science behind all of this is gonna be described better shortly, but in the meantime, here's another example of how scientific examples are represented in the graph. Our scientists looked into the PDE10 prediction using our tools to form the following hypothesis based on the information that we have in our graph. Let's go back to our previous graph representation containing cGMP, inflammation, and ulcerative colitis and add on some additional information, which is PDE10 is upregulated in the colon of ulcerative colitis patients and is also highly expressed there. These information come from our precision medicine team. PDE10 itself breaks down cyclic GMP, which we all know is already associated with inflammation. Together, these facts place PDE10 under some suspicion. It's active in UC and the tissue where UC occurs, and it breaks down the very molecule whose absence affects inflammation, a key component of UC itself. Hence, our scientists believe that lowering PDE10 levels using a PDE10 inhibitor drug would treat ulcerative colitis. The next step in the BAI process is to test such a hypothesis in living cells, either at our labs in Cambridge or with a suitable collaborator. In the case of UC, experiments were conducted with our expert collaborator, Professor Tom MacDonald, whose team confirmed positive readouts from assays validating the PDE10 hypothesis back in 2019. With a validated hypothesis, the next step is to perform further checks before embarking on a drug program. Benevolent started a drug development program back in 2019. BenevolentAI's drug development team then works to achieve a balance between various factors specific to the disease in question, including potency, selectivity, dosage, and toxicity. This requires many iterations of compound design, chemical synthesis, testing in vitro, and DMPK studies, all of which can be accomplished at our Cambridge facility, allowing us to pursue our most promising hypothesis through to drug development, lead optimization, and then onward to the clinic. As a result, Benevolent's chemistry team identified BEN-8744, our lead compound for the treatment of ulcerative colitis. Visualize how we work as a company. Now I'll talk more about the detailed technology that underpins everything that we do here. To introduce myself, I am Daniel. I'm the Chief Technology Officer here at BenevolentAI. I have a background in biomedical computation at Stanford, and a PhD researching neuroscience and machine learning at ETH in Switzerland. I've worked in tech consulting in Silicon Valley before this, and I've been at BenevolentAI for about five years. I've also built some of the TargetID tools that Olly referred to and I'll be showing you more about today. Excited to show you this. All right, let's return to the slide that Ivan introduced earlier, and building on what you've just seen from Olly. What I wanted to do is to talk in this more detailed way about the technology we have developed for drug discovery. In the first part, I'll talk about the left-hand side, the data foundations for a platform that support all of our systems. In the second part, I'll talk more about the tools that and user interfaces that we use to help augment our expert scientists and make them better at what they do. We believe, again, that the key to unlocking complex disease biology is to integrate and make available as much biomedical information as possible. We want comprehensive data foundations across a wide range of data sources and data types, as you can see here on the left. Now, we harmonize these sources to make them available through our tools and to support our expert scientists in developing new drugs. The goal is to holistically represent biomedical knowledge. Cumulatively, of course, these data foundations span a breadth of information that is beyond any individual expert. Now, by leveraging these consistent data foundations within our tools, we can build this positive feedback loop that you see on the right, one that helps us improve the platform every time we pursue a disease program. Predictive algorithms, software that helps support triaging targets and assessing the progressibility of targets, and then ultimately, the experimental validation of the hypotheses from the system, are fed back into our underlying data. By instrumenting how these decisions are made and what information is useful where, building metrics on how the data is used in the platform, we can optimize and improve this process over time. Since we have a hypothesis-driven approach where there's reasoning behind each prediction, we can avoid many of the false positives, as well as side-step complex issues with target deconvolution that you run into with a phenotypic approach. The output is the portfolio programs you see on the right, and Anne will be talking about more in a moment. First, let's begin talking in a bit more detail about our data foundations. Let's look at the diversity of the data types that we bring in. We've built comprehensive representations of biomedical data here at BenevolentAI that are therapeutic area agnostic, that is, not just one single disease area, and drug modality agnostic, that is, not just small molecule or antibody. As all data have their own unique missingness, bias, and noise, we can achieve fundamentally better representations of biomedical data by gathering and integrating all the available information. If we have a fact, for example, that's surfaced from the experimental data on the upper left, but it's also additionally supported by human genetics and fundamental scientific research, we can be that much more confident about it as a fact to reason from. That, of course, is a much more comprehensive and correct approach than if we were stuck in a single data modality like imaging. We do heavily leverage the scientific literature, but that's actually just one of the many data types that we use. You can see more of the examples on the slide here, experimental data, omics, biological systems, so forth. Each of these data types requires their own special way of processing. What we do is we ensure that we have data pipelines that are paired to each of these data modalities. You know, for the scientific literature, you need an NLP processing pipeline. You can see an example segment of our NLP pipeline here, where we are running named entity recognition in orange, identifying entities of interest and relationships in blue, that describe the relationships between things that we care about. Our data is proprietary, public, and in-house. We have a mixture of all kinds. The goal is to bring it together and harmonize it. These domain-specific data processing pipelines that we built support that processing integration and harmonization of each of our data modalities, and that facilitates our ability to bring together a variety of the information that you need in order to make a call about a target, as you're building it into your portfolio. Let's explore these data foundations and numbers. At the top left, you can see some of the data modality types that we have in our knowledge graph. We have a variety of data here, ontologies and dictionaries, genetics and clinical data, and experimental data. We have very large amounts of literature, scientific literature, as well as structured databases. Those might be, for example, chemistry databases that we pull into the system, as well as omics-derived databases. For example, correlated modules from gene expression signatures. We aggregate this data from over 85 different data sources that covers everything from the major scientific publishers to emerging databases relevant to biomedicine, and it represents critical biomedical entities, 33 in total, as you see on this slide. Molecules, clinical traits, biological pathways, tissues, and so forth. It supports our nuanced and specific reasoning. We've done a lot actually with this data foundation, right? You can see also how much the system is growing over time. Our in-house proprietary representations have continued to grow at pace over the past year, which yields a formidable competitive moat of unique data representations. We now house around 409 million relationships in our knowledge graph, which is up almost 250% since June 2021. Over 187 million of those relationships are proprietary to us. That's 46%. Yeah. Now note that we have more proprietary data now than we had total data in June last year. In order to develop that proprietary data if competitors want to get a head start on us, not only would you need to develop all of our in-house algorithms and processing techniques, but you would also need access to that huge volume of data, which we've spent years figuring out how to build and integrate correctly. Moreover, as a company, our platform benefits as the quantity and quality of biomedical data that's out there continues to expand and compound and provides us structural tailwinds as a company. Just to give you a little bit more flavor of the detail that goes behind, you know, ingesting all of this data, I'll walk through two example data processing pipelines that we have internally at a high level. At the top, our scientific papers are made available in a variety of ways, via APIs from scientific journals, via flat files curated and licensed to us from databases of scientific literature. These inputs are automatically downloaded and ingested so we can ensure that we reliably bring up-to-date, fresh scientific literature into the system. We then normalize these documents, convert them to a consistent format, remove extra metadata, the standard types of processing that you need to get it ready to run in a standardized NLP processing pipeline, where we actually wanna extract information from those documents. The NLP begins with named entity recognition, or NER, you see here, where we identify the key concepts and texts, and then we run it through the relation extraction methods, that include both rule-based and ML-based methods. Ultimately, the output of our NLP processing pipeline produces a robust set of edges, the relationships that feed into our data foundations and support the reasoning and the assessments that our scientists need. That's the literature pipeline. Parallel to this is our genetics precision medicine processing pipeline seen here at the bottom. Now, using genetic information is critical, with evidence showing that drugs that have genetic support can have double the likelihood of success in clinical trials. We bring in summary statistics as well as data corresponding to particular genetic cohorts, as well as annotation features. We use this data then in our internal GWAS pipeline, a Spark-enabled biobank scale genetics processing pipeline that helps us link traits to various locations efficiently and at scale. Of course, when you do this, you will find the majority of implicated loci are outside the protein coding genome, and so we need a method to map a variant to a gene. We have an ML approach to doing this using the annotation features, eQTLs, chromatin interaction, locus distance, so forth. That maps our variants to specific genes. Now, the critical part about all of this, as you can see on the slide, is that we have to ensure that we have the same harmonized representation across all the systems. Whether a gene is found via the genetics pipeline or it's mentioned in scientific literature, we're grounding it to the same concept that allows us to reason effectively over it. Of course, we have a similar precision medicine set up for our whole exome sequences, and our whole genome sequences, and ultimately outputting into that same shared representation of biology. My key takeaway here really is just like this is just two of our pipelines that we have. There's also bulk transcriptomics, EHR, and so forth. The key takeaway is that we have an automated, well-engineered set of processing pipelines that are paired to all the data types that we bring in, with the outputs converging onto a consistent data foundation that gives us comprehensive coverage of biomedical knowledge and naturally benefits from the growing quantity and diversity of biomedical information that becomes available. All right, now we'll move on from the data foundations and begin discussing in more detail the second core, the tech platform, the user interface tools that empower our scientists. I'm really excited to show these to you today. But first, I think it actually helps to understand how they fit into the drug discovery process. We'll do a quick primer on the earliest steps of drug discovery, and then we can step into how we use these tools internally. We chose to develop our tool suite to improve the earliest stages due to the impact and leverage of those first few steps, right? If you choose the wrong target at the beginning, it doesn't matter how great your chemistry is or how great your clinical trial capabilities are, it will fail. Whereas we focus on using AI, data, and tools as early as possible so that we can develop only the most promising hypotheses and maximize our overall success rate. Let's begin at the target ID stages of hypothesis generation and hypothesis validation, and then we'll walk through those boxes that you see here. I'll just be kind of reading out the text under there, if you wanna follow along. You can see the tools that are specifically supported, you can see the stages specifically supported by tools in green. Beginning on the left in the light blue, we begin with selecting a disease of interest for our tools. We make a defined start, stating how we're gonna pursue a given disease. We'll choose that by analyzing unmet need, data availability, tech opportunity, and so forth, the factors that Ivan talked about earlier in his section. We begin to scope potential assays so that we can test those targets effectively. That helps us stick to our rapid timelines as a company, where we can produce, test, and move on. We work to define the tech approach, iteratively formulating a hypothesis for the tools to predict against and developing and refining the context used to define the disease state. We apply our algorithms that rank the genome and then generate the predictions to triage, and then during the biological triage step, our scientists go through each one in turn, analyzing the evidence, reviewing it, and examining the targets to see if there are any red flags for that target, working to assess its biological relevance and plausibility to the disease. Following that, we have the target progressability step, or TPA, as it's shown here on this slide, in which we assess the progressability potential of the target. That's typically a much deeper dive than the initial quick triage step. TPA can occur before, during, or after the targets are sent to assay, and that's why it's shown here as being before the pink assay step and sometimes after the pink assay step. There might be times when, for example, you want to only test. You only have space to test a few of the most promising targets. There, you wanna make sure that the targets that you put into assay are going to succeed, and so you wanna run that TPA step first. Alternatively, you might just want to use data as feedback into the platform or for future disease iterations, and so you can afford to do the TPA step after you've gathered data from your initial assays. Now, finally, in the last few steps, we prepare for portfolio entry. In these steps, the evidence is combined with a validation package that allows us to consider whether it's ready. After being taken into the portfolio, the program is guided through the standard stages of discovery, hit ID and expansion, lead op candidate seeking, and eventually into development and clinical trials. Now I'll talk about how our tech supports those steps in green. Zooming into those green steps, which are there on the left for reference if you want, we'll go into a bit more detail. We define that tech approach. This is what Olly referred to earlier in his presentation in the video, about using our in-house tools and algorithms that allows us to explore the data and define the input to our predictive models. You can see an example here. Can we treat ALS by reversing autophagy impairment in microglia by reducing oxidative stress? Each of those underlined terms is grounded to a concept in our data foundations. It's enriched with connections, links, and context, and it allows us to create predictions based on those grounded terms. We provide tools to our scientists to explore and define this tech approach. Often, our scientists will have some way in their minds of thinking about the disease and their unique angle of attack, and we provide ways for them to translate that from their mind in suggesting new mechanisms they may not have thought of and translating into something the platform can use. In the next step, you can see on the right is our target prediction algorithm. It's a bit of the secret sauce that combines our multiple ML approaches and leverages data from our previous disease programs, so that we can identify the best, most promising therapeutic hypotheses. You can see a few of the example models that we have here. Olly pointed out a couple in the presentation, but just to reiterate, basically, we have graph models that help us identify missing knowledge graph links. We have transcriptomics in the models. We have other data modality-specific models, like predicting directly from genetics. We're really excited actually about the introduction of that last one, the large language model approach. Think of being able to train on all of the volume of scientific literature that we have access to and do really nuanced, specific inferences on top of that natural language. It's really cool. Now each of these methods produces their own set of targets. It all has to get aggregated together. The way that we do that successfully is by using our previous results. We have metrics for all of this. We track usage, and we know what the signature of a good target looks like. What we can do is we can look over all these different scores and metadata about the targets and use that to create a prioritized list of targets that we wanna look at. That takes us to that third step. Now, how do we actually assess those targets? Our job as technologists is to bring together all the diverse data that a scientist needs in order to make a call. You know, extracted sentences from underlying scientific papers, models predicting safety, previous explorations and clinical trials, that kind of data. That's that third step, triaging and assessing progressibility. Now, without our analysis tools, in a pharma company, it takes substantial effort and time to root around and find all that needed information, whereas we can present it efficiently in one view, helping this process to occur quickly and correctly. Now, ultimately, all of these targets are then validated experimentally so that we can assess the target's potential. Now let's actually look at the tools. We provide a lot of backgrounds to help you understand and contextualize what you see here, but these are, you know, expert-driven tools, and I think what's really exciting is the ways that they can contribute to their own scientist workflows. Here on the left-hand side, what you can see is a scientist working to define the tech approach for a specific disease. Our algorithms recommend important related concepts in the context of a disease. For example, here, visualizing the overlap of a disease and some suggestive mechanisms. In this case, it's ALS and cellular response to oxidative stress and astrocyte activation. The relationships of these associated terms are derived from the harmonized data of the diverse sources that we bring together. This saves our scientists time, but it also sparks new insights or reveals novel associations, as data from other parts of biology are recommended and surfaced. You know, one of the things that's really important about this as well is that we augment that with the data metrics that, a scientist needs to understand, say, the relevance, the novelty, and the specificity of some of these associated terms. Helps them understand how they wanna think about the question. Is this a new angle of approach? How relevant is it? Now, that set feeds into another visualization, which you can see here on the right, in which the scientists explore the biological relationships themselves. The platform presents these relationships between concepts selected in that previous step, uncovering these communities that you see here as shading. Those are tightly connected biological interactions. As the scientists explore and traverse these interacting communities, exploring the underlying data and the relationships between them, they begin to understand the purpose of these biological clusters and enrich pathways to uncover functionality and ultimately the causal biology driving them. The purpose here is to allow our expert scientists to use this interfaced to build out the nuances of the tech approach for the disease. Perhaps selecting a mechanistic context that was recommended by the tools, helping them assess their initial way of thinking about the disease and how they wanted to pursue it. They will result in a set of something like that tech approach that I showed earlier and was introduced by Olly, using terms that are well-represented and intriguing, and those would form the input to the predictive models. The algorithms I mentioned, you also saw in Olly's video, create the predictions using the entirety of our knowledge graph and the data foundations as data input, and that draws insights across disease silos and data types that might be constraining the way a scientist would think about the disease. The output lists for triaging, as you see here, is a list as you see here for triaging. This is a demo board. Each of the predicted targets, MAPKAPK2, LDHA, it's all represented by a summary tile. In our drug programs, a scientist would systematically step through and triage each of these predicted drug targets to determine what we would actually test. Now, our scientists select criteria to filter and sort these predictions based on a pre-specified strategy. You know, shown here in the criteria filters like biological rationale or existing evidence to understand how well known it is or safety, and each has been color-coded to match the filters that they're interested in. Each disease has unique needs, right, in terms of biological evidence versus novelty, safety, and so forth. In a disease like glioblastoma, a life-threatening brain cancer, the tolerance for side effects will be different, and quite different than that of a much less fatal disease like atopic dermatitis. Setting criteria helps rank the hypotheses that most closely match what a scientist is actually looking for. You can see that color coding here for the ones that actually match what the scientist is interested in. Now, this is just the summary view because by clicking through on each of these tiles, we want to be able to support any decision that the scientist would then make on it. So as the scientist reads through the supporting card for each target, going through the literature evidence and other supporting information, they begin to form an idea of whether or not that target could form a plausible hypothesis for the disease. Eventually, they'll make a decision, which is also captured by the platform, that records whether the target is one they would take forward or not. So in that process, we actually captured valuable annotations that support that reasoning. Having used this same platform in all of our disease programs over years, we have captured tens of thousands of detailed annotations from expert drug discovery scientists, the ones we have in-house, and allows us to train and refine our machine learning models, that helps us bring together those target hypotheses that we talked about earlier in that predictions stage. Now, the whole system was co-designed between our tech teams and our drug discovery scientists. It's a really important part of our company. The goal is to provide them with all the information they need in one place while they assess these therapeutic targets. The platform populates the information, metadata, insights that they need to make a decision, all in one carefully optimized view that we've iterated on over time. Now, pulling this information together manually outside of the platform we have built is challenging and costly considering the sheer breadth of information that we've surfaced and summarized. All right. Stepping ahead to that target progressability assessment. Once we've done a quick triage to initially assess a target, we want to support our scientists in performing a deep dive on our targets to identify which ones they want to invest in as the targets that are gonna be the most likely to succeed. This is our first time talking about this tool publicly, which is kind of exciting for me, and it represents the culmination of one of our key objectives this year in developing the technology to improve how we do this target progressability assessment. We now use this within our own internal disease programs. The TPA tool that you see here automatically surfaces insights about the suitability of a target to enter the portfolio from a chemistry and competitive landscape perspective. Within this tool, our users are able to assess the opportunity to differentiate from existing assets, clinical assets, the drug ability of a protein, for example, on its 3-D structure, the selectivity potential using prediction of likely off targets, and the patent landscape of the chemistry around the target. You can think about all the different kinds of data that needs to be pulled together to enable that to happen. The tool replaces manual processes again that are used to gather decision-making data, improving the efficiency and the consistency by which our scientists are able to make calls about targets, for our portfolio. In the view you see here, it's, we're looking at one example of visualization where we're exploring the possible, target selectivity by analyzing similarity in a detected drug pocket compared to other similar targets. We have a rich set of visualizations, which you can see on the lower right, and that allows our scientists to navigate the space and identify other interesting, areas to pursue. Similarly to the previous tool, stepping through each of these targets into the relevant category provides a wealth of supporting information, allowing our scientists to understand and correctly make decisions about possible targets for the portfolio. All right, I think at this point it's worth bringing it all the way back. Like, why have we made all these decisions of designing the platform the way it has, and what is our usage over many years yield to us at the end? Ultimately, these tools were designed to leverage our unique technology approach, produce the benefits that you see here on this slide. I'll walk across the top row. By focusing on removing data silos and building on a breadth of biomedical information, we have produced a platform that can pursue any therapeutic area. It's a disease-agnostic platform. With a focus on biology and target ID, our tools are not tied to any particular data type. They can support chemistry and antibody approaches, and for a modality-agnostic platform that avoids the limitations of companies that, say, are constrained to work in only one space, like chemistry. By integrating wide data types and large volumes of information, we can spark creativity in scientists and surface novel targets they haven't considered before. In fact, our predictive algorithms rank the whole genome for all of our disease programs. You know, as further proof of these top three benefits, note the breadth within our business collaboration with AstraZeneca. Chronic kidney disease, idiopathic pulmonary fibrosis, heart failure, lupus. These are quite disparate diseases and demonstrate the breadth of applicability that we can achieve with the platform. Despite little biological overlap between, say, renal diseases and respiratory diseases, our platform has found targets for our partner. Moreover, despite the substantial research already done in this space, perhaps even unpublished, our platform has connected evidence to find novel biology that AZ is pursuing, having now nominated three targets into their portfolio from our collaboration. Working along the bottom of benefits on this slide, it's worth saying again that we've built these tech products for our scientists to understand how to support them, what they do better. Summarizing data from a huge volume of underlying information, they're able to efficiently make data-driven decisions about programs that can be supported by all the tools we have built. This accelerates discovery. By taking a tech approach with foundations designed for volume, metrics that analyze our platform, and well-engineered systems and workflows, we have developed a scalable and repeatable approach which we apply again and again to different diseases. This process just compounds and becomes more valuable over time. Ultimately, we do all of this to solve the hardest and most important problem in our industry, increasing the probability of success of our drug assets. Now I'll hand over to Anne to walk through our drug discovery and pipeline review. Good afternoon. My name's Anne Phelan, and I'm the CSO at BenevolentAI. In terms of my background, I have a degree and PhD in genetics, but more significantly, I spent about 16 years working at Pfizer in the U.K. During that time, I had the opportunity to work across a range of therapeutic areas, leading teams from early target discovery through the drug discovery process and through into clinical development. I've been with BenevolentAI for almost four years now, and I'm responsible for all things drug discovey. As Danny said, and he's just described systems. The next step in the target ID process is to test these hypotheses experimentally. To enable this home to over seven state-of-the-art iPSC, CRISPR, and sequencing capabilities to further support the target validation workflows on our wider drug discovery portfolio. Access either technologies or complex human patient-derived cell-based systems that may not otherwise be readily available to us. Oh. Is that better? Yes. Yeah. All of the data generated are captured in a centralized data repository and can be fed back into our systems to further enrich our data foundations. Ultimately, once a target has gone through our rigorous hypothesis validation workflow, fulfills our requirements around safety, opportunity to differentiate, and other commercial elements, we will make the decision whether or not to bring that target into our portfolio. As Ivan mentioned, this slide provides a snapshot of our platform-generated portfolio, comprising both first-in-class and best-in-class approaches across a broad range of therapeutic areas. We believe that the diversity of our portfolio emphasizes, and is in fact testament to, the data-driven, disease-agnostic applicability of our approach, which enables us to identify therapeutic targets irrespective of the disease of interest, which, coupled with our molecular design capabilities, has enabled us to rapidly build a portfolio of targets with a substantial projected market value. I should point out that our target ID platform is also modality-agnostic, and we're in the process now of validating potential antibody and siRNA therapeutic targets for portfolio entry in the coming months. With over 10 early discovery programs currently undergoing target validation, we'll continue to rapidly scale and build our portfolio through 2022 and beyond. The most advanced asset in our pipeline is a topical pan-Trk inhibitor ointment, BEN-2293, now in a Phase I-B, II-A clinical trial. BEN-2293 has been developed to both provide rapid itch resolution and relieve inflammation in patients with atopic dermatitis, which we believe will be a key differentiator for this asset. Atopic dermatitis, as you may know, is the most prevalent chronic inflammatory skin condition, with around 70% of patients presenting in the mild-to-moderate disease category, which is our intended target market, where we're looking to provide important steroid-sparing therapy. We also know that the prevalence of this autoimmune disease is increasing, leading to a forecasted market value expected to exceed $14 billion by 2028, the window of anticipated launch for BEN-2293. To give you a little bit more detail into the mode of action of BEN-2293 and why we believe it will provide therapeutic benefit, 2293 is a potent pan-Trk inhibitor, able to simultaneously block signaling through the TrkA, TrkB, and TrkC receptors, which are associated with inflammatory, pruritic, and pain signaling pathways. In terms of the mode of action, inhibition of the TrkA receptor is expected to block nerve sensitization and reduce pruritic signaling in primary sensory neurons, thereby suppressing the itch reflex. We believe that the concomitant inhibition of TrkB and C receptors will reduce the Th1 and Th2-mediated dermal inflammation, treating the inflammatory component of AD. By addressing the two most troublesome symptoms of the disease simultaneously, we aim to develop a disease-modifying therapy targeting mild to moderate AD patients in the first instance, with the option to pursue severe patients who require additional add-on treatment beyond their standard of care therapy. There's a strong scientific rationale behind the additive benefit of blocking the three Trk receptors. It's been shown that TrkA receptor levels in the skin increase dramatically in response to inflammatory stimuli, as do levels of NGF, which together lead to enhanced sensitization of primary afferents, resulting in chronic itch in already broken and inflamed skin. TrkB, the cellular receptor for BDNF, shows elevated expression in the skin-resident eosinophils of AD patients, leading to an increase in the chemotactic index and a pro-inflammatory response. Signaling through the TrkC receptor via NT3 is known to potentiate T cell-mediated inflammatory responses, further exacerbating the inflammatory microenvironment within the skin. Taken together, there is a sound logic to the simultaneous inhibition of all three Trk receptors in order to tackle the key elements of the disease. In terms of the pharmacology data package we've generated in support of this asset, we've been able to show that by targeting the three Trk receptors, we're able to dose-dependently inhibit release of pro-inflammatory cytokines, such as TNF alpha and IL-4, from stimulated human PBMCs, and you can see these on the red and blue dot response curves at the top of this slide. These results are indicative of a robust anti-inflammatory effect. We've also demonstrated that BEN-2293 is able to inhibit release of the itch-provoking neuropeptide CGRP from explanted rodent DRGs, as shown in the lower left figure, clearly illustrating the antipruritic potential of this asset. Finally, we've been able to demonstrate that topical application of a pan-Trk inhibitor significantly reduced most ear inflammation in a PMA model, performing comparably well to a topical steroid ointment. With this data set supportive of therapeutic efficacy against the two major symptoms of the disease, together with the excellent in vivo safety and tolerability, CTA enabling package, we've been able to embark on an adaptive phase I-B, II-A clinical trial directly into patients. The first clinical study for BEN-2293, which successfully completed at the end of 2021, was a phase I-B, first-in-patient dose escalation safety and tolerability study. For this double-blind study, each cohort comprised eight mild to moderate AD patients, six of whom received active drug, two received placebo. As this is a topical agent, we were able to escalate the dose by increasing the dose strength of the ointment from a 0.25% ointment to the higher strength 1% ointment. Increase the duration of dosing from seven days - 14 days, increase the body surface area to which we applied the ointment from 10% - 30%, and increase the frequency of drug application from once per day to twice per day. The study culminated in twice per day dosing of our 1% ointment over 30% of body surface area for 14 days. Patients were monitored for all standard safety and tolerability endpoints, and although it was only a small sample size, we did also assess a number of efficacy endpoints in this study. Now, this slide illustrates a snippet of the efficacy data generated in this pilot phase 1-B study. I can't emphasize enough that this is a phase I study. It was not powered to meaningfully assess efficacy with only six patients dosed with active in any given cohort. I also recognize the fact that an efficacy study would ordinarily run for 28 days, not the maximum of 14 days we have in this safety and tolerability study. That said, this is one example of the type of data we generated. What we're looking at here on this figure, on the X-axis is the number of days of dosing out to a maximum of 14 days. On the Y-axis, we have a mean change from baseline of the percentage body surface area affected in treated areas. If we focus on the 14-day dosing curves, you can see the placebo group in orange, where there's a small early effect, but from about day 11 onwards, the score then essentially returned to baseline. The green line illustrates the patients dosed with 1% ointment on 30% of their body surface area once per day. Finally, the purple group, where the dosing regimen was escalated to twice per day dosing. Considering a change from baseline of score of two would be regarded as clinically meaningful, these data can only be really interpreted as a trend effect, but there does appear to be a clear window of effect between the placebo and the BEN- 2293 dose groups, which is certainly encouraging at this stage. After a full safety review of the data from the phase I-B study, we completed comprehensive statistical modeling to inform the final study design for the phase II-A efficacy study that initiated earlier this year. This placebo-controlled, double-blinded study is designed with a seven-day washout from any existing medication, followed by a three-day run-in to enable exclusion of any placebo responders before randomization into the study designed to assess improvements in itch, AD rating scales, and key inflammatory efficacy outcome markers. This is in patients treated with 1% BEN-2293 twice a day for 28 days. The study is on track to complete recruitment by fourth quarter of this year, with results anticipated in the first quarter of 2023. With results in hand, it is our intention to out-license this as a phase II-B-ready asset for onward clinical development. BEN-2293 is being developed to address key unmet need in the treatment of atopic dermatitis, as described. We believe with rapid onset of itch resolution and a significant anti-inflammatory effect, this asset could displace ineffective or poorly tolerated second-line treatments for chronic use in mild to moderate adults and pediatrics, with the potential to treat severe AD patients as an adjunct therapy option. As I said at the outset, pan-Trk is the most advanced clinical asset in our portfolio. Coming along quickly behind it is our second most advanced program, which is a PDE10 inhibitor for the treatment of ulcerative colitis. To provide an introduction to ulcerative colitis as a disease and describe some of these groundbreaking research in this arena, I would like to introduce Professor Tom MacDonald. Which one makes it go forward? Yeah, this one here. Oh, hi. Thank you very much. I'm an immunologist, so I worked around the corner here at Barts Hospital for, like, forever, and I'm now banished to Whitechapel for my sins to work in the Blizard Institute. I discovered that TNF was a bad molecule in inflammatory bowel disease, and I discovered Mongersen as well. About 10 years ago, I got fed up publishing papers and sort of writing grants. I decided to try and put myself in the space where to try and help pharma to develop, well, get over the valley of death, essentially, by using human tissue instead of animal models, because animal models for inflammatory bowel disease are terrible. I've been working with GSK, essentially developing, helping them since the mid-noughties. Despite all this, I've never helped them discover a drug. Which I'm very happy about actually. We set up this collaboration in 2018 because I think that what we try to do is put us, as I said, in the position where we can sample gut biopsies and resected samples and put them in a little dish, and actually they don't know they're in the dish. They still think they're in the body, and still start make cytokines, and the signaling molecules they make are still the same as if they're in the body. It's actually quite a useful sort of ex vivo system to explore actually human biology and essentially what drugs might work in patients, because you're dealing with patients' samples. Ulcerative colitis is a not very nice disease that starts in the rectum and moves up towards your small intestine, and it can be cured actually by colectomy. But people like to have their colon actually in their body. It affects a lot of people. There's a big unmet need here, actually. These are severe patients, but the moderate-to-severe patients. You know, there's quite a lot of therapies out there now, cheap therapies like steroids, but the problem is people relapse. There's no long-standing effect. The chronic inflammation leads to people getting increased colorectal carcinoma, so it's not a pleasant disease at all. I mean, unpleasant disease. About half the patients don't respond to anti-TNFs and things, and a lot of patients are now on vedolizumab, Entyvio, the anti-α4β7 integrin, but that doesn't really work either. Lots of patients are on it because it's safe. A few years ago, we decided to take samples from patients 'cause the gut is one of the tissues that's endoscopy happens a lot, and we've got ethics approval to take 60 biopsies for experiments. We can take these biopsies, and we can culture them in a little dish and serum-free media, and they continue to think they're in the body. They make cytokines. The signaling pathways are still there. We can come in with, actually, with, say, JAK inhibitors and things and show, actually. If we add prednisolone to the samples, which is a steroid, everything goes down. It's actually. What works in mild disease works in these samples as well. Interestingly enough, we don't get resected bowel anymore because of the success of the therapy. People are keeping their colons inside them now and their small bowels, which I think is quite good. Here we go. This is just the sort of stuff we do. Actually, spent a lot of time working with VHsquared, who actually made llama antibodies. They're little 12-kDa molecules. This is an experiment here that we did. In which we took this only available anti-TNF. We added infliximab to the biopsies, actually. What happens actually is that the amount of IL-6 goes down, which is really good. The amount of IL-8 goes down. The amount of TNF-α obviously goes down. This is actually the antibody that we tested, V565 as an antibody. Was a to mab. This is, we're still working with them. We made a few and looked at a few other antibodies. This is done by Scott Crowe, and we had a really good collaboration with Scott. This is something else we do. The one sitting there is, you know, a selective JAK inhibitor. Yeah, right. Yes. Like it's not. Yeah. That's a lot. Never read what drug companies say on the label. That's never true actually. We can do it. The tons of proteins. We can tell by mass spec, we don't do it by these things there. What else do we do? This is something else we RIP1 and RIP2 inhibitor programs that we spent a lot of money on. They evolved the PROTAC, which is you put an E3 ubiquitin ligase to the drug. Actually, not only do you block the effect of the active site of the drug, but you degrade the drug as well by sending it to the proteasome. This is some experiments we did actually just with different cytokines, and you can see measuring at the top here TNF, IL-4, etc. That was done at GSK, and we looked at the Crohn's disease biopsies and ulcerative colitis biopsies. This is an amazing thing. The protein, the RIP1, RIP2 protein disappeared completely within 12 hours, and all the cytokine levels went down. On the left-hand side is IL-1β to IL-6. You can see they're all going down at 50 nmol. Amazing. Absolutely amazing. I'm quite happy in this space. I think it's good helping people. It gives some information. It's great working with BenevolentAI because they're just. They're in London, so I don't have to travel very far. I can get there in the Hammersmith and City Line, which is a great advantage. I don't have to go to Cambridge or any place like that. Okay. I'm finished. Thanks. Thanks, Tom. You can see why we enjoy working on ulcerative colitis. One of the reasons we selected UC as a disease area of choice, for the reasons Ivan articulated earlier, is a complex multifactorial disease with high unmet medical need where we felt the application of an AI-augmented approach could be very fruitful. This slide outlines the workflow we embarked on that resulted in the discovery of PDE10 as a potential target associations linking PDE10 with ulcerative colitis. It was entirely novel. The team prioritized a short list of targets for experimental validation using the ex vivo assay described by Prof. MacDonald. We were able to demonstrate a robust anti-inflammatory effect with a tool PDE10 inhibitor in this system. Using the available metadata, we were able to build the rationale for why a PDE10 inhibitor had the potential to be a UC therapy. Essentially, why the hypothesis generated by the models was correct. The target itself, PDE10, functions to reduce intracellular levels of the signaling molecule cyclic GMP. We know that low levels of cyclic GMP are directly associated with increased inflammation, and you can measure this through elevated levels of markers such as TNF-α. So logic says that by inhibiting PDE10 and restoring cyclic GMP levels, you should have a direct anti-inflammatory and disease-modifying benefit in the colon. Data further supporting our confidence in this target comes from transcriptomics, where we demonstrated an upregulation of PDE10 in ulcerative colitis patients relative to healthy donor tissue. This was mirrored by a reduction in guanylate cyclase, the enzyme that synthesizes cyclic GMP. This results in a net reduction in cyclic GMP levels, supporting a direct role for PDE10 in ulcerative colitis disease. Finally, when considering the basal expression profile of PDE10, shown by the blue band in the lower right figure, it's clear that PDE10 has low levels of peripheral tissue expression, reducing the potential liability of targeted inhibition of this enzyme. To further validate the hypothesis, we tested a representative PDE10 inhibitor in the surgical colonic biopsies taken from UC patients that are refractory to standard of care therapies, which is the precise patient group we would initially target in the clinic. These studies were conducted by Prof. MacDonald's lab, making a side-by-side comparison with the JAK inhibitor ruxolitinib and the steroid prednisolone. As you can see in the bar graphs, the PDE10 inhibitor, BEN-3218, was able to demonstrate a significant reduction in the inflammatory cytokine markers of chronic inflammation, IL-6 and IL-8, relative to the controls. This was an extremely exciting moment for the project as we believe it was the first-ever demonstration of a PDE10 inhibitor showing efficacy in a UC system. In terms of the underlying mechanism, there's a direct anti-inflammatory effect associated with elevated cyclic GMP and cyclic AMP levels, which we believe can be achieved via PDE10 inhibition. Increased cyclic GMP levels reduce TNF-α mediated cytokine release with concomitant reduction in tissue-resident macrophage activation, which in turn will result in reduced intestinal inflammation. Elevation of cyclic GMP has also been shown to improve tight junction assembly and fluid mucus homeostasis, leading to improved barrier integrity, raising the potential for a disease-modifying benefit from this approach. Overall, we believe this constitutes a compelling data package, which is why we chose to bring this program into our portfolio. On this slide, you can see the rate of progress for this program and its current status. From validation of this entirely novel target for the treatment of UC in 2019, the team have been able to generate a potent, selective, and peripherally restricted candidate molecule within two years of the project initiation. The asset, BEN-8744, has now transitioned through regulatory tox studies with a clinical trial application anticipated in December of this year. I should also mention our ongoing precision medicine activities, through which we are in the process of using patient-derived molecular descriptors to identify the optimal patient cohorts. This information will be used to inform our clinical trial design, but also to identify efficacy biomarkers in order to further increase our probability of clinical success. With this data package, we believe our PDE10 inhibitor will provide an efficacious disease-modifying treatment. In terms of its position in the market, BEN-8744 is a best-in-class, peripherally restricted, potent, and selective drug designed to target the moderate to severe patient population, meeting the unmet need left by existing therapies, such as those patients that are refractory to anti-TNFs or other biologics, improved safety and tolerability profile compared to the competitors, and a precision medicine approach to target key responder patient cohorts, avoiding the safety risks associated with ineffective therapies. We're aware that this is a competitive market and believe that we can differentiate from the competition not only by equaling or improving on efficacy, but also by providing an improved safety and tolerability profile compared to the frank immunosuppressants recently approved. Those PDE10 inhibitors that have been assessed in phase II studies for schizophrenia were regarded as safe and well-tolerated. The major dose-limiting side effect published is a centrally mediated sedation. Our peripherally restricted agents should be devoid of this side effect liability, and consequently, we anticipate it to be safe and well-tolerated in humans. In the interest of time, I've only touched on our two most advanced assets, but clearly we have a robust pipeline of programs progressing through the drug discovery process, reflecting the fact that we have created an AI-augmented model for sustained portfolio delivery across a range of therapy areas. Some examples of our recent progress through the first half of this year are listed out in the box. We recently nominated a candidate for our GBM program. A second ALS asset has transitioned through into candidate seeking. We transitioned two more oncology programs through into lead optimization and brought a new project into our portfolio for the potential treatment of Parkinson's disease. However, in order to maximize the value of each of our in-house programs, in addition to their primary disease, where we have our primary focus, we systematically evaluate each target to assess alternative disease indications or additional opportunities for that asset. Using our precision medicine approaches, we're able to consider indication mapping or perhaps phrased differently, drug or target repurposing opportunities across the whole portfolio. When we think about each target, we assess opportunities for therapeutic benefit, perhaps in an entirely different disease or on a more granular level in the identification of a specific patient cohort or responder group within a selected disease. At this stage, we also evaluate biomarkers or trial design outcomes options. This is a capability that we currently apply to our in-house portfolio, but has obvious applicability in a wider repurposing context and could be readily applied to approved drugs or drugs currently in clinical development. To wrap up my section, as you've heard from the team, we have a number of key inflection points on the horizon from a drug discovery and development perspective. Our atopic dermatitis program is anticipated to read out results first part of 2023. Our ulcerative colitis asset will initiate its phase I study early 2023, followed shortly after by a phase II-A study. We have our glioblastoma multiforme candidate that is now going through IND enabling studies with a view to submitting the clinical trial application next year. Moving forward into a phase I study, we continue to progress our pipeline, moving projects, as you've just heard, through each of the standard drug discovery transition stage gates. We expect to add between four and six named drug programs to our portfolio year on year, culminating in the progression of one-two CTA/IND stage drug candidates every year from 2024 onwards. With that, I'll hand off to Nick to discuss our interim results. Can I put that down now? Oh, that's good. Thank you all, and thank you, Anne. Good afternoon, everyone. My name is Nick Keher. I'm the CFO of BenevolentAI, and I, having joined in March of this year, will highlight the first half performance of the company from both an operational and financial perspective. Turning to my first slide, I'm pleased to report the first half operational highlights, again reiterate the validation points of the platform you have heard in depth today. Across our wholly owned internal pipeline, we have seen continued progress with BEN-2293 entering a phase II-A clinical study in the first half of 2022. Recruitment is on track, and we expect to have results from that phase II-A study in Q1 2023. Preclinical work on BEN-8744 for UC is also on track, and a CTA filing is expected in the Q4 of this year, with the phase I to start in 2023. Lastly, from a clinical and preclinical perspective, as Anne has highlighted, we have IND-enabling studies for BEN-28010 in GBM, while BEN-9160 is also on the cusp of entering IND-enabling studies for ALS. Why is that important? With all this progress, we are setting up a steady stream of news flow from a clinical pipeline standpoint that will continue into 2025 and beyond. Across our early-stage pipeline assets, we have seen two programs transition to lead optimization, with another expected in the second half. We have two new named assets enter the portfolio, and we expect two-three more to enter in the second half. This represents solid progress across our pipeline and our platform and a build-out of deep long-term value within the company. With our partners, we have seen continued success, in particular with the AstraZeneca partnership, where a third novel target was selected to enter in the first half, and we expect more news to follow in the second half of this year. This provides strategic validation to the platform, but it also highlights the disease-agnostic capability that we have from our platform. We've also seen progress on our not-for-profit work with the DNDi, which also provided novel targets for review in dengue fever. Elsewhere, as discussed previously, you've seen the application of our technology was validated yet again, this time selecting baricitinib for full approval by the FDA for COVID-19. Not forgetting what underpins this company, we continue to invest in the BenevolentAI platform itself to strengthen our leading position in the AI drug discovery field. We continue to grow and enrich our data foundations with new and generated insights, primarily due to the increase in patient-level data and enhanced natural language processing recall. As you've seen from Danny's' slide, we've seen an increase of around 250% in relationship volume in the knowledge graph from June 2021 to August 2022, 46% of which is proprietary to BenevolentAI and hence creating a meaningful barrier to entry for anybody looking to enter this market. The knowledge graph continues to evolve, and a new version was launched during the period to enhance recall of facts from scientific literature, supporting more nuance and specificity in our biological relationships. Beyond our data representation, our AI tools for scientists and predictive algorithms have improved at pace. Extensions and enhancements to our existing suite of tools have now completed that allow for novel targets best prosecuted by alternative modalities. As a result, we are aiming to bring a monoclonal antibody into the pipeline in 2023, which would provide more validation to the platform's capabilities and further differentiate us from our peers. On April 22nd, the company closed the business combination and listed on Euronext Amsterdam on the 25th, raising EUR 225 million of gross proceeds. As part of this, we gained two highly valuable and experienced board members in Olivier Brandicourt, the former CEO of Sanofi, and Jean Raby, the former CEO of Natixis Investment Managers. This is already adding to a leading industry board that we have at BenevolentAI. I have subsequently added Dr. Susan Liautaud as well. Lastly, we've been adding depth and capability throughout the organization to help deliver as a listed company, but also to progress the pipeline and maintain our leadership in the AI field. Turning now to the financial results themselves, and the performance of the company, we performed in line with what internal expectations, with revenue growth of EUR 3.1 million - EUR 4.8 million, thanks to the extension of the AZ collaboration, with further upside potential to come from milestones related to new target selection, asset development, and ultimately revenue-based royalties upon commercialization. In terms of spend, we saw drug R&D rise to fund the phase II-A study for BEN-2293, as previously outlined, and progress the development of other molecules within the pipeline. Headcount has increased across the business, but with both our product and tech business as well as our central functions, we expect there to be minimal increase from here as we focus on expanding our clinical capability. Net, this led to a normalized operating loss of EUR 55.3 million, which is before we consider any R&D tax credit, which is captured below the line. Given the fact we have listed through a business combination, which has a material impact on the face of our reported figures, particularly in the first half, we wanted to spend two minutes to outline the walk from a reported to a normalized position to help the reader understand our go-forward cost base. Importantly, many of these exceptional costs are non-cash and are either now finished or, in the case of share-based payments, are set to reduce materially. As a result, while the overall cost base has increased some GBP 9 million, this does largely reflect the phase II-A study for BEN-2293 and the incremental costs of our listing status. Turning now to the cash flow and the key takeaways here are threefold. First of all, the clean read from EBIT to operating cash with little CapEx or capitalized development spend in our cash flow. Secondly, that we have an R&D tax credit of around GBP 12 million that we expect to receive in the second half. Thirdly, we finished the end June 2022 with a cash position of GBP 265 million, which provides us with sufficient capital to deliver on our key objectives, which leads nicely to my final slide. With the GBP 165 million we have of cash at the end of June 2022, we see sufficient cash runway to Q4 2024 before we consider any incremental cash from out licensing assets or collaboration deals. This strengthened cash position will also help us deliver near and medium-term value inflection points, which we have categorized into the five bullets on the right, and it informs our capital allocation process. Firstly, we will ensure the completion of the phase II-A study for BEN-2293 in atopic dermatitis before aiming to out-license this asset to a large partner who has capabilities in the dermatology space. Next, we will fund the phase I study of BEN-8744 in UC and commence the phase II study shortly after in 2024. Alongside this, we will aim to initiate two other phase I studies by 2025, with both BEN-28010 in GBM and BEN-9160 in ALS front runners to enter the clinic in this timeline. Again, to underpin the value of the platform, we will focus investment on our continued leadership in the AI drug discovery field to enhance and maintain the BenevolentAI platform. Finally, we will focus our resources on internal capabilities to support both our listing status and to support further collaborations which we'll sign in due course. I'd now like to hand back to our CEO, Joanna Shields, to provide our closing remarks. Thank you, Nick. Is the mic working? Okay, thank you. I think what you've seen today is the impact of artificial intelligence and machine learning on research and development in the pharma industry is clearly at an inflection point. Within our peer group of newly listed companies, BenevolentAI stands out with a scientifically and technologically validated and differentiated approach that is already producing a rich portfolio of drug programs and consistently delivering in our collaboration with AstraZeneca. All of this is supported by our intellectual property strategy, which includes strong patent protection on our drug pipeline and copyright and robust protection of trade secrets of our technology platform. Today, we have shown you what we believe to be is a category-defining business, underpinned by a world-class team, as you've heard from my colleagues today. Thank you, Professor Tom. Much appreciated. Also a very distinguished board, which is very supportive and helpful to us as we transition now into a listed company. We have shared the output of our prolific discovery platform that is enabling our scientists, and we hope soon many scientists around the world, to better understand complex multifactorial diseases from the outset to help them discover novel drugs at scale. We believe that all this has the potential to accelerate discovery and increase the probability of success in the clinic. The funds secured in our recent business combination will support the clinical development of our own in-house pipeline. These assets will drive long-term value. We also plan to out-license assets over the next one to three years. We will look to increase the size of our pipeline with a healthy balance of first-in-class and best-in-class assets, with 1-2 CTA or IND stage drug candidates every year going forward. Our strengthened financial position also solidified our plans to develop further and scale our AI-enabled drug discovery platform. We look to sign new collaboration agreements with pharma companies to leverage our disease-agnostic capabilities into therapeutic indications outside of our focus areas to generate incremental revenue. Finally, what gives BenevolentAI a significant edge is the number of years we've invested in building our drug discovery platform, and we plan to maintain this leading position through increased investment in our technology capabilities while building our metrics to exemplify the differentiation and impact of our approach. I would like to close by pointing to why this all matters. Billions are spent each year supporting a drug development system that leaves patients without effective treatments. This must change. This sign is in the lobby of our London headquarters, and the message is illuminated in neon, and it reads, "Because it matters." This represents the sentiment that drives everybody at BenevolentAI, where we work to push the boundaries of artificial intelligence to unlock the power of decades of data and research, to understand the underlying cause of disease, and to develop treatments for the millions of patients who need them. We do it to transform the way drugs are discovered and developed, and we do it because it matters. With that closes our formal presentation, but I'm actually delighted to open up the floor and also our online participants the opportunity to ask questions of myself or our leadership team and Professor Tom MacDonald. Thank you. Okay. Marvelous. Okay, I think we're ready to go. Do you have a first question? Hi. Maybe a question to Daniel regarding the technology. How do you assess the performance of your platform? Yeah, great question. So the question was how do we assess the performance of our platform? I think if I understand what you're getting at, ultimately, the platform is assessed by the eventual success of our drug candidates, right? That is many years down the road. We've developed a system of internal proxy metrics that allow us to assess the value of our predictive algorithms. You can imagine various ways that we do this, thinking about looking through the scientific literature, looking at predicted performance on assays, looking at performance on our own internal programs where we've hidden data that allows us to assess that. We also have a family of metrics that help us understand how we measure our data and our volume, as well as the data quality that we're bringing into the system. Overall, the goal here is what correct proxy metrics can we develop that actually correlate really meaningfully with eventual clinical success. We've developed and iterated on those over time. Danny, maybe it's an interesting opportunity to mention the Helix partnership with Stanford University. Absolutely, yeah. There are a number of, you know, challenging opportunities for research in the area. We have a partnership with Professor Russ Altman, who leads the Helix group at Stanford University. Together, our teams are actually collaborating on a really interesting research project around discovering and adjudicating contradictions in literature. As you might imagine, if you extract facts at scale from the full scientific literature, you're gonna uncover facts actually that disagree with each other. How do you actually resolve that contradiction? Actually, this is one of the areas where our broad recall of scientific literature really helps us because it turns out that nine times out of 10, the difference is actually the context of the sentence. Maybe it's a disease state versus a healthy state, or it's a particular tissue versus another one. You know, by putting together these systems where we've measured these metrics, we can identify some of the contradictions, and then we develop research approaches that help us resolve that. We can measure the performance improvement over time. Thank you. Next question. Right. I was wondering if you could expand a little bit on the sort of concept of drug repurposing. Because I think most people understand this area or think of it as that you're defining novel targets and then sort of going out into the chemical universe and finding some molecule that interacts with them, and that's actually how you're saving time. Now, you talked about it sort of saving two years on the drug. Throughout the presentation, you didn't imply that many of the programs or any of the programs even were repurposing attempts. Okay. Although you talked about baricitinib, which is obviously a repurposing attempt. I'm just a bit, a little bit confused, but, you know, what is really going on here? How many of the 20 or so, you know, or 15 or whatever there is? Yeah. You know, are repurposed molecules that you've taken in and licensed in or, maybe or whatever, however you've sourced them, you know. How many are de novo, you know, drug designed attempts? I'm glad you asked the question. I'm happy we have the opportunity to clear that up. We are not a repurposing company. None of the drugs in our current portfolio, the named, the main drugs in the portfolio or the experimental, the early stage stuff are repurposed assets. Everything has been generated through the process that we've demonstrated here today, the use of the platform to discover novel biology and then design and develop the right molecule to target that, you know, that condition. I don't know if you wanted to add anything, Anne. Yeah. I'd be happy to, actually. Yeah, as Joanna said, none of our programs are repurposing programs. They're all wholly owned internally. We have composition of matter, IP on our lead assets, so we take them through the drug discovery process ourselves with our medicinal chemists in-house. The repurposing element of this from our internal portfolio perspective, if you think about PDE10, we have found an alternative use for that as a target in ulcerative colitis. Once we've invested in a considerable amount of effort actually in creating all the assays, getting the chemistry, getting the composition of matter IP, we are then taking our own asset and thinking, is there more we can do with this asset than just ulcerative colitis? Can we use our models as we've shown that we found ulcerative colitis? Can we go through that same process and find another additional disease that we could consider for any of the assets for which we have the IP? This is it's an opportunity to sort of evergreen our own portfolio rather than necessarily considering it as, you know, the opportunity to repurpose other people's drugs. Obviously because we have this pipeline, we do have that capacity to, as we've shown with baricitinib, but that is not the main focus of our work by any means. Yeah. Just explaining the decision we made to look at approved drugs that could be used for COVID-19 was just, you know, seeing the escalating global health emergency. We felt like we had this capability that Anne's describing to look for additional indications, and we thought, "Let's use that workflow, see what we come up with." We found some really strong evidence to support our hypothesis, which is why we, you know, we signed the Wellcome Trust pledge. We went out and we published this with all the caveats that we thought that this drug had this off-target effect, and we believe that to be true. Over the course of our relationship with Eli Lilly, we were able to prove that hypothesis, you know, first of all in silico and then ultimately in the large scale clinical trials. It was more or less, you know, our response to giving our technology to humanity at that stage when it was a really difficult time and we thought vaccines were going to be a longer, much longer, time for delivery. Oh, interesting. Okay. If I might follow up with another one. I was just wondering on 2293 on atopic dermatitis, if you able to compare that in any way, or were you able to compare that with the sort of relatively new molecules that have been approved? Obviously, there's JAK and the TYK2, if I'm right, there in that space. Obviously, the reason that you've gone with a topical versus an oral, you know, is that your differentiation is, you know, for example, two pills. Do you want me to? I can take that. Yes, go ahead. Yeah. We're looking for some of the agents that have been approved in recent times suffer from application site irritancy, and the JAKs, for example, still carry a black box warning. Even the topical JAKs carry that very significant safety warning. We have the opportunity to differentiate on the basis of safety and tolerability, things which is we anticipate no application site irritancy with our asset. One of the reasons to go for a topical is because the pharmacokinetics of our drug are such that it's absorbed into the skin. Where it binds to the sensory neurons that are in the lower levels of the skin. And then once it gets into the systemic circulation, it's very rapidly cleared. We get essentially zero systemic exposure. This gets us away from any of the side effect potential profile of any of the oral pan-Trk or Trk inhibitors that have gone before us. There's a safety element to being topical. It's also for something like atopic dermatitis, there is a therapeutic benefit of topical application to the affected body surface area, as opposed to taking a systemic drug when you may only have 10% or 15% of your body surface area affected with disease. Hi. Thanks. If your platforms are clearly quite far reaching in terms of identifying potential disease targets, with what frequency are you identifying targets that might traditionally be considered undruggable? When that happens, and presumably that happens quite frequently depending on the disease area, what are your capabilities in terms of developing modalities, perhaps the emerging kind of modalities like targeted protein degradation and others that can address such targets? Yeah, we have the capacity, as you suggested, to find entirely novel targets for the treatment of potential diseases. When we're looking at our target validation workflows, we have multiple ways of validating in our experimental setups. Where there's a good quality small molecule inhibitor available that we believe is potent and selective, we can use that. But we also use a lot of CRISPR and RNAi knockdown sort of technologies where there isn't a good off-the-shelf drug available. Once we get one of these first-in-class previously undrugged, that's not to say undruggable. I think undruggable is becoming a kind of redundant concept. Where we've got targets that are previously undrugged or a first-in-class, we have the capacity in-house to take PROTACs, for example, or a degradation approach, and we have. We're in the process now of setting up strategic relationships with monoclonal antibody providers, and we're in discussions with groups who specialize in siRNA therapeutic delivery. Actually, albeit internally we have to date focused on small molecules and chemistry, through a network of collaborations, we can and will prosecute any modality, whatever we regard as the best modality to prosecute what we believe to be the best target for that disease. I would just add a little bit to what Anne was saying. Danny presented our target progressibility tool, and I think it's really important that we've moved that multifactorial analysis of what could potentially lead to failures down the line into the actual target assessment phase. We look at that alongside of the target workflow that we use. I think it's really important that these technology developments continue to enhance the ability and so you can identify and say, "That would have failed, therefore we're moving in a different direction." You can modify your hypothesis, you can change the criteria, you can look at different properties and, you know, you can do all this in a, you know, instantly with all those factors considered early stage of target identification. I dunno if you wanna add anything, Danny. Okay. Hi, John Priester at JP Morgan here. I had a couple of questions. First one, what's your capacity for additional platform collaborations over the next, say, two -thre years? Maybe just your kind of thinking around your preferences between collaborations, in-house development and out-licensing, and how you see that developing. Do you wanna take it, Ivan? Yeah. I think on the first one, John, I don't think in terms of internal capacity, if we were to sign, say, another collaboration, maybe there'd be a small amount of additional headcount, but fundamentally our platform's very scalable, so I don't think there'd be a capacity constraint internally on the amount of collaborations. In terms of, I think the second part of your question was around focus on internal pipeline versus collaboration, and I think for us we see the long-term value of this company is always in obtaining assets and taking them through clinical development. However, in the shorter term, the collaborations bring benefits in terms of non-dilutive capital and validation. I think we look, as we've guided previously over the next couple of years, to do maybe one-two, but this isn't a sort of volume business because the challenge there is, you know, you close off other therapy areas you may want to be in yourself. Maybe focusing on the lead asset BEN-2293, can you speak at all about potential interest you've had around that? I noticed that it is referred to as a phase II-A study. In terms of potential progressibility for a partner, is there a likelihood there for necessity for a phase II-B study, or is this likely something that could go straight to phase III? Ivan, did you wanna take that? I'd say I don't wanna give a running commentary, but as you'd expect, we're always talking to potential partners for our assets as they move through the development. I'd say we're regularly talking with interested companies in the space. I think it's not for us to say what the next trial will be. I think for us, we think of it as we'll generate a phase II-B-ready data package. It depends. That's very much on the choice of the partner. Anne, did you wanna add anything? No, I think in our current trial, we will have dosed 90 patients, 45 on active, 45 on placebo. If a partner deems that to be, you know, an adequate cohort size, you know, they could move into a phase III. I think the likelihood is they would want to broaden it out into a wider phase II-B trial. That's entirely at their discretion. Can I have another one? Go ahead. Right. Okay. Well, it's a bit variation on John's question. I just... If you look at the workflow from target identification, I was sort of wondering if, you know, you do that, but, you know, what proportion of times do you find an answer that you can take forward? I mean, is it, you know... I don't know, could it be, like, one in five, or is it every time you get, like, as you illustrate, four options to then consider what's the most promising? Or is it... Do Do you find that you just can't find something sometimes? Well, I'll start by saying the iterative nature of a hypothesis-driven approach is that you're constantly refining those questions and asking and answering and solving problems as you do. The richness of information available to the scientist through these complex multifactorial questions that they're asking gives them the indication on how to move forward. You can always find something, I would say. You know, we haven't explored a disease area that we didn't find something of interest, you know. In the triage tool, there's like lots of ranked targets and, you know, there's always something of interest because there's such an unexplored space. I don't know, Anne, if you wanna talk about how we map diseases to mechanisms and how some of that mechanistic understanding sort of opens up new spaces as well for us. Yeah, absolutely. Direct answer to your question is, we have never sought to find a target for a disease and failed to find something that, you know, is interesting and of adequate quality to bring into the portfolio. It's actually the opposite problem because we have really, really rigorous criteria for what we bring into the portfolio. We're constantly looking to narrow the number of interesting targets down because for any given disease, we only want to bring a small number of projects into our portfolio because we don't want to dominate the portfolio with any particular disease. We are constantly looking for reasons to invalidate targets when it comes to bringing them through into the portfolio. In terms of John's reference to the mechanistic mapping, we think about disease and disease targets, not really in terms of a disease classification or a diagnosis. It's a mechanistic dysregulation at the cellular level. Because we are approaching the target identification process through considering human biology and human mechanistic dysregulation, we are able to reason over not just the literature or the omics data sets precisely generated in response to ulcerative colitis. We can look at the totality of biomedical literature and all of the data at a mechanistic level and draw parallels across biology. We may find a target that had been you know validated in schizophrenia, and we find an applicability for it in ulcerative colitis. Because we can reason so holistically across mechanistic biology, we have never failed to find a target of interest that passes all of our very rigorous criteria for portfolio entry. Hi. Thank you. Zoe Karamanoli from RBC. Identifying the target is obviously a very important part of drug discovery and development, but also, selectivity and optimization of the target are also equally important. Just wondering now that you also mentioned that you're gonna try and bring in targets that are antibody, how good is the platform, how it compares, and whether there is any opportunity for collaboration given that, you know, one can't be best at everything? Thank you. Yeah. I would say yes. Fundamentally start on the biology. I mean, we didn't talk about it, but we do have tools in chemistry and small molecule optimization as well. To the specific question around the antibodies, yes, we can't be wonderful at everything, regrettably. I think we would. That's the kind of thing we'd look to do in partnership with a specialized company that is their core business. Now, whether that be more of a fee for service or a strategic relationship, these are the things we're exploring. You know, in the first instance, we have to have a target that it makes sense that that's the right therapeutic approach based on all the criteria. Hi. Thanks again. Does the platform have the ability to identify potential combinations of targets to, you know, go after hitting those even before you've worked out whether one is viable? I mean, appreciating that you're gonna need to test those in safety studies alone first, but most companies developing combination plays are obviously taking something that's already approved or something where the efficacy is already established and adding something else that makes sense from a kind of. It's a known target in the disease type already. Are you able to kind of synthesize these ideas from scratch? Do you wanna give an answer from the user perspective, basically? Yes. As a straightforward capacity to identify two novel targets or two previously or drugs with the targets that don't have like a launched product drug, that's not something we can do very straightforwardly. We can do it through the sort of synergy of two lists, if you like, and we can look at the interplay between them. We don't immediately just pull through a combination suggestion. That said, when you're talking about a drug on top of standard of care, which is oftentimes the place you're gonna be, particularly in the cancer arena where people will not come off the standard of care therapy, you're looking at an add-on. That is something that we try to model through, the likely horsepower and benefit of our target in the context of the patients already being on standard of care. Did you wanna talk about the tech angle? No, I think that's a really good summary. I think that through the combination of multiple lists, we can approach it. We've also done it through the kinda data exploration view that I showed earlier. You can look at, say, a community of tightly coupled biology and think about, okay, what would it be like if we hit these different communities, and what would the effect be? I think, you know, the opportunity space is still just so rich and even single targets, and the challenges are so great that it makes sense to focus there to start with. Do we have any online viewers? Oh, we have another one. Great. No problem. Thank you. Just building on the earlier questions on 2293, if we play this out post-data, you talk about obviously the plan is to license the product out. In the scenario you are either unable to find a partner to license or unable to agree terms, what would plan B look like? Mindful of the investment already made, is there any scope you would take it further forward yourselves or is there an alternate scenario? Well, I think that's the unfortunate scenario. There is potential, I'd say. I'd be perfectly honest today, atopic dermatitis is not a disease we want to take things a long way, but I think we probably go through the process of what data are we missing. Is there specific analysis we're missing? Then see if we could do that. I mean, that's one of the reasons we spent so much time talking to the potential partners as we go along, so we don't end up in that situation. I wouldn't say it's a core plan is to take it forward. Yeah, of course. We have all day. All right. Okay. right. Okay. Right. It's me again. Looking at BEN-8744, and sort of noting your earlier answer to my question about the repurposing. Obviously, you could have tried to or been able to license in a compound that's probably got phase II data, and that would go, you know, be a lot further ahead there. You wanted to design your own peripherally restricted compound, is what I think you've done here, if I'm understanding correctly. I mean, are you concerned that there'd be CNS side effects with that approach? Is that what determined that strategy? I'm going to yield to Anne on that. Yes. Yes. That's exactly the strategy. There are existing clinical-grade compounds that we used in our early validation work because at that stage, we didn't have our own proprietary compounds. We used those centrally penetrant compounds in the first instance as part of the validation package. Yes, the reason that those drugs have not succeeded in clinical trial is because of the CNS-mediated side effects of predominantly sedation. They were in trials for schizophrenia, and there's profound sedative effects. We have gone to great lengths to ensure that the drugs that we generated are peripherally restricted, do not get into the CNS and will be devoid of those side effects. We couldn't take an off-the-shelf agent and just repurpose it. How is it peripherally restricted? I mean, is that by the chemistry or is i t just doesn't cross the blood-brain barrier at all? Yes. No, it's by design. It is. Right. It effluxed out of the brain, so you get no measurable levels in the brain at all. Okay. Thank you. Do we have any online participants that would like to ask a question? Please press star one to ask a question. No, no. You're okay. I was like, I don't see a star one anywhere. Okay. It's not on the computer. We have no questions, so I'll hand back to the room. Mm-hmm. Okay. Last chance, everybody. Before we can have a coffee together. Any other questions? We're happy to take any at all. Oh, you. Can you hear me? Can you hear me now? Okay. All right. Well, thank you very much for joining us today. As I said, it's been a real privilege to be able to share our journey with you and our story and how the company has evolved, and we're immensely proud of what we're doing. Really feel like we're onto something exciting here, and we're delighted to have your interest and happy to engage with you know, anytime, all of us, to answer any further questions and to keep you updated on what's happening. With that, I think we're going to move to a coffee break. Okay. Thank you so much.