Skip to content

2026-04-18-ai-in-biomarker-discovery-and-drug-development

Date: 2026-04-18 Primary topic: agentic-ai Cross-referenced from: agentic-ai, bioinfo-tools, virtual-cells Word count: ~12,941

Auto-generated captions

This transcript is Vimeo auto-generated English. Expect misheard names and technical terms. Use the ripgrep/search approach for exact quotes.

Good afternoon. A very warm welcome to our session here. I'm not Jakob Kata. I wish I were, but I'm not. Jakob could not make it today with some of the logistic challenges coming from Germany to here. So we are very fortunate, however, that he was able to send his recorded talk. So we'll get there in a second. I'm George Visilo. I'm the Chief AI for Science Innovation at AstraZeneca, and I'm going to pass here to my impromptu co-chair. Hi. Very nice to meet all of you. T. Rubihani from Recursion Pharmaceuticals, VP of Translational Strategy. So it's an immense pleasure to have all of you here during lunchtime today because we are witnessing a new era of translational research, clinical development across academia and biopharm. It is happening. It's a reality, and I think here we will have the opportunity of discussing how the research is evolving today, how the future looks like, and what it is that we need to do to get to the future that we're all talking about. Because AI is no longer the future, it's happening today. So without any further ado, let's move to the first talk, that's by Professor Jakob Kata. He is the Chair of Clinical AI at University of Dresden in Germany, and he also has a laboratory at University of Heidelberg in Germany as well. He has been a leader in terms of the development of AI-driven biomarkers in oncology, and he has also developed a number of frontier solutions in the agentic space, actually helping us reimagine tumor boards from first principles through the lens of AI, not with AI. It's extremely exciting. So, I'm going to ask our technical team here to play the video now. Let's hope it works. Fingers crossed. Hello and welcome. My name is Jakob Kata. I'm Professor of Clinical Artificial Intelligence in Dresden and Heidelberg in Germany. My flight was canceled. I'm late to the conference, so I have to pre-record this. Very sorry, but I will be there for the Sunday sessions and Monday sessions. Please get in touch. And now I want to share my slides and share with you my talk about multimodal and agentic AI in biomarker discovery. Here are my disclosures. Let's look at how the field of AI developed in the last 20 years. It all started with what we now call classical machine learning methods, where you train machine learning models to, for example, work with tabular data or with pre-extracted features from images to solve tasks such as prediction, prognostication, classification, et cetera. But, of course, as we all know, these methods were not sufficient for the complexity of data we face in cancer research, specifically image data, for example, histopathology images. So the field was completely changed in the 2010s when deep learning came along, and we were able to train deep learning models on large image datasets or other complex datasets for tasks such as classification, object segmentation, many others. In the 2020s, in the last couple of years, the field has been changed again by the advent of foundation models. These are big models that you train once with self-supervised learning on a large heterogeneous image cohort or dataset, and then you can use them and fine-tune them or apply them to additional downstream tasks. And typically, if you do this two-step approach, then you don't need as much data for the downstream tasks, and you can solve them much better. Another type of foundation models is, of course, large language models. These are just big foundation models that have been trained on natural language, so they can be used in natural language processing, but they have also obtained a generalist world knowledge and of course, are the basis for all the chatbots, ChatGPT, and everything that everybody's using. I think the most interesting part of these large language models is that they have enabled the technology of AI agents, where you take a large language model and you connect it to other tools so that it can solve complex tasks. Here, I'm going to show you a few examples how all of this comes together, how we can use deep learning to develop multimodal or unimodal biomarkers, and how we can then use AI agents to develop them much faster. First of all, let's speak about the biomarkers. And I want to just share this terminology that we have developed in our professional society, ESMO, European Society of Medical Oncology last year, and this has been published as a guideline. We've looked at the entire field of AI in cancer research and oncology, looked at the publications, looked at the products that are out there. And there's three classes of AI-based biomarkers, where AI is being used to derive properties of tumors that help us make better treatment decisions. Class A is using AI for biomarker quantification, and this is very simple because here what you're doing is you're just recapitulating the human workflow. Okay? So imagine a PD-L1 stained immunohistochemistryslide in lung cancer. Normally, a human would read this and would score the PD-L1 staining intensity. Now you can train AI systems to do the same thing, obviously with potentially much better reproducibility and also faster turnaround, et cetera. And many of these Class A biomarkers are already on the market. Class B would be the use of AI to indirectly measure existing biomarkers. For example, using pathology slides to make inferences about genetic alterations. There's now a lot of research available and also commercially approved products whereby you can predict microsatellite instability in colorectal cancer from pathology slides, EGFR mutations in lung cancer, gene fusions, many other mutations, tumor mutational burden, and many of these can be inferred just from the pathology images. This is a Class B biomarker. There's also other works that have shown that you can predict IDH mutational status from brain imaging in glioma patients. So there's a lot of work here. These biomarkers are typically not perfect, but good enough to be useful as pre-screening tests. Class C biomarkers are entirely new biomarkers where we train, for example, on radiology images, pathology images, genomic data, one of these, or all of these, or any combinations of these to predict treatment outcomes directly. For example, you can train an AI model to predict resist response based on these baseline images, or you can train an AI model to predict survival and/or adjuvant treatment efficacy, et cetera. You can predict clinical endpoints directly, and this would be a Class C biomarker. If you're interested, if you're working in this field, this has been published last year in "Annual of Oncology" and we think that this could be a scaffold to talk about these AI biomarkers. Now, what does it have to do with agents? I've already mentioned initially that large language models, of course, everybody knows large language models that can be used as chatbots. They can also be enriched with additional context. For example, medical guidelines or your personal documents that the chatbot can access. But if you generalize this more, what you get is an agent. Agents are large language models that can interact with any other software or any other tool to perform additional actions that they cannot perform natively. For example, to interact with a web browser or a command line interface. And then you can build them also in a way that they iterate and perform multi-step problem-solving. Let me introduce some terminology here. The AI model itself, the large language model, has weights. It's just a bunch of weights with the architecture, and this is the actual model. It has been trained once and then you use it. You can place this in harness, which would allow for example, computer use, browser use, terminal use, et cetera, and then you can package this harness in apps that people use. There's just some examples here. This is of course, already a bit outdated, but I think it's important to get this terminology right. And how does this apply to drug development, biomarker development? Well, it applies because a lot of what we do in biomarker development, drug development is just a human using a computer, right? In a very simplified way, it's a human researcher looking at experimental data, making a hypothesis, writing code, running code, debugging code, making presentations, et cetera. All of these steps can be performed in principle by an AI agent. Now, probably today's agents are not ready to do this completely autonomously, right? But again, what we are seeing right now is that the LLM-based AI agents that are available right now already automate part of this pipeline. And ultimately, I think it's safe to say that anything a human does with a computer can ultimately be automated. Let me show you some examples. We've built these agents in our research group mainly for clinical applications. So these are very relevant, of course also for biomarker development and biomarker application. One example is this work where we built a tumor board agent. You have a large language model. You give it access to a patient case, all the radiology, all the pathology images, all the genomic data that the patient has, access to a database of clinicaltrials.gov, et cetera, and then the agent does stepwise reasoning and uses tools to analyze all these data and comes up with a recommendation. We validated this obviously against human experts and found a very good performance. We've also developed autonomous agents, and this is unpublished, should come out soon, where we have agents that can look at patient cases and then not just piece all the diagnostic information together, but also take actions in that they, for example, can request additional imaging, request procedures, request laboratory tests, and then you can basically, we again benchmark this against human doctors, so you can basically automate the entire treatment journey in this scenario.So how does this relate not just to biomarker application, but also to the development of biomarkers? Well, as I said, anything a human developer does with a computer can ultimately be done with an agent. Right now, there's different components of this that can be done with agents. Many people are using this. In my lab everybody's using this. In many other labs, in industry, everybody's using agents, specifically coding agents, software development agents such as Claude Code or Codex are, of course, very powerful. And they are generalist agents, general purpose agents that are also very good at bioinformatics tasks and research tasks and biomarker development tasks. And in fact, many people are using them for this purpose. This is just an example, asking Claude Code to do a QC on genetic data. It's probably not perfect, but it can be done. Of course, I have no conflict of interest here, but there is many companies out there that are developing this, basically domain specific agents. And now the question is, and I think that's an unresolved question, what is really the secret sauce? What makes a drug development agent, what makes a biomarker development agent unique, right? Can you just use Claude Code or do you need more specific scaffolding? I think the answer is evolving, right? For as many cases, generalist agents are very good. For other cases, you might need specific agents. Another example where I don't have any conflict of interest, right? One of the trailblazers in this field, one of the pioneers was, of course, Jure Leskovec and his group's paper, Bi omni, and their agentic applications in biomedicine. Some of these you can reproduce nowadays with Claude Code. Some of these you cannot because they require specific tooling that is embedded. So the field is evolving super fast. What I think is one of the most important scientific questions is how to evaluate agents. I really like and highly recommend this Pokemon evaluation of agents where you can basically take large language models and give them a harness that lets them play Pokemon Red in this case. And then you can see how these agents solve the problems, how they basically succeed in solving tasks iteratively, how they advance through the game, through the different milestones in the game. This example is of course highly outdated, but you can see that each of these lines is one large language model in an agent harness playing Pokemon. And you can see that there's differences in how they advance through the game. And this is a benchmark. This is a classical agent benchmark where you test the capabilities of these agents for multi-step problem solving. And I think these are the types of benchmarks that we need. We cannot just like previously in AI, have one system that produces one output. We compare it with the gold standard, and then we have our performance. No, we have to look at the procedures that the tool calls benchmarking is now-- Benchmarking of agents is now becoming very complex, and we still need to develop the conceptual frameworks for this. And I think if we can formulate biomedical research, biomarker development as we can formulate games such as Pokemon Red, then we will make advancements in providing good benchmarks for agents. I want to close with one warning slide. Agents are of course-- any AI systems and agents in particular are of course susceptible to many pitfalls. There's of course hallucinations, many other potential problems. We are very interested in attacks on agents, attacks on AI systems, for example, prompt injection attacks whereby a user that interacts with an agent can hide prompts in the input data, even in images, even like invisible texts and images. And of course, there are many efforts ongoing to harden large language models and other AI systems and agentic systems against these attacks. But there is no universal cure here, and it's still possible deliberately or accidentally to perform prompt injections. For example, when pathologists write with a pen on a slide, PT3, then a lazy agent will typically just look at the text and not the image and become confused, right? We've done these two studies. Dr. Fersch who led this study on the right-hand side, he's also going to give a talk here in another session. So if you're interested, please reach out to him. And with that, I want to say thank you for your attention. I hope that this was interesting. Please reach out. Here's my email address. Again, I will arrive a bit later and there's a session tomorrow. There's a plenary session on Monday at 8:00 a.m. Thank you very much. Thank you very much to the other speakers and see you there. One of the benefits of technology. And actually, before we move to the questions here in the room- Hello and welcome. Just because I praise technology, we have a glitch. Anyhow, so there was a question online already that what are The AUCs, the sensitivity and specificity values associated with Class 3 biomarkers. And I had the opportunity of actually asking Jacob as he flies here. So his answer was that this is an excellent question and not an easy one. That there's no universal cutoff. Actually, in this context of the Class 3 biomarkers or Class C here, what's imperative is that there is clinical utility and that validation in adequate data sets is performed. One important point, in particular, when we move to these biomarkers for which there is no gold standard that we can perform directly an analytical validation, is that we may not need full explainability of the model. What we need is validation in multiple data sets that are fit for purpose. Imagine the levels of evidence that we require for clinical utility being also applied to this framework of biomarkers driven by AI. Just because we are using really cutting-edge technology does not give us poetic license to bypass the levels of evidence that are required and sufficient for us to deliver something in the clinic that is transformational. Do we have any questions in the room? Tiru and I will try to fill them in the absence of Jacob. One, two, three. So thank you so much. So we move to the next talk ahead of schedule. Can you believe it? Efficiency. Let's do it. So I have an immense pleasure here to invite Tirup Behanie, my impromptu co-chair today. She's the VP for Translational Strategy at Recursion, and it's a company that actually is doing drug discovery and clinical development through AI first principles. Tirup, it's such a pleasure to have you with us today. I look forward to your talk. Thank you very much. Good afternoon, everyone. Thank you for being here, and thank you to the ACR organizers for inviting Recursion to be a part of this very important discussion. As my co-chair has said, it's very topical for our industry, and I'm excited to be a part of it. Today what I was hoping to do is to walk through a couple of examples of our clinical stage programs that really highlight how Recursion is using AI in practice to inform everyday decisions across the R&D continuum. So from biology through to chemistry through to the clinic. So first, a quick introduction on Recursion Pharmaceuticals. We are a novel medicines company. We are aiming to produce better molecules, more differentiated molecules, for more meaningful outcomes for patients. And we're using AI, as I mentioned, across the R&D continuum to help us get there. We have proprietary multimodal data at industrial scale. We've designed purpose-built models with the ability to answer key biological, chemical, and clinical questions. And importantly, we've built bilingual teams, so both scientists that are fluent in AI, as well as AI engineers that are fluent in science. And together, these teams can create biological insights that we can translate then into clinical strategies and eventually execution, and hopefully we can show meaningful outcomes for patients. But the real differentiation for Recursion is how it all comes together. Because we believe that we're one of the few AI-native companies that has end-to-end operations, so really applying AI, again, across the R&D continuum. And we believe that's because it's important not only to improve the decisions that we make in one area of R&D, but really across the entire value chain. So here I'm introducing our Recursion platform, the Recursion OS. And it really is sort of broken up into three distinct pieces. In the first piece, the left-hand side, you'll see that we can go from biology to insight. Here's where we can use our platform to really identify causal disease mechanisms. And what this does is it allows us to identify novel targets and the associated diseases for those targets. Then in the middle, we have our precision design part of our platform. So here's where we can really optimize our chemical synthesis and produce molecules that have differentiated properties, optimized properties, and those that will really help us achieve the therapeutic outcomes and the therapeutic indices that we're really going for. And then on the right-hand side, you can see we go from molecule to patient. So this is our clin tech platform. Here's where we can identify the right patient using causal modeling. We use RWD analyses to help contextualize our trial data, and we also are able to run our trials more efficiently by using AI-informed enrollment. The goal, of course, across this entire platform is to be able to not only improve our chance of success, but also reduce the time and the money that's typically used for traditional R&D processes. So what this has led to is we're very proud of our wholly owned pipeline at Recursion. We have five clinical stage assets and a couple of programs coming behind those clinical stage assets. And what really differentiates each one of these programs is that they've been informed by a specific insight or specific set of insights that have come from our platform. We are very encouraged by our clinical programs and the momentum that we have and again, the number of programs that we have coming behind those.We believe that they have first-in-class and best-in-class potential. And again, in the next couple of slides, I'm going to show you examples of three different programs and how we've applied AI in practice across these examples. So one thing to note is that we don't apply our platform uniformly across all of our programs. Instead, we use specific parts of our platform to really hone in on key challenges and answer key problems within each program. So on the left-hand side of this slide, you're seeing our platform again, just in a slightly different format. You've got the biology to insight piece, the precision design piece, and the clin tech piece. And each one of these components, you can see how they're being used across our programs on the top. So for today, I'm going to focus on three pieces of our pipeline, the REC-1245 program, which is an RBM39 degrader, the REC-617 program, which is a CDK7 inhibitor, and the REC-4881 molecule, which is a MEK1/2 inhibitor for the potential treatment of a rare orphan precancerous disease called FAP. So first, let's get into RBM39 and the REC-1245 program. As I mentioned, we don't use every part of our platform for every single program. But what's great about REC-1245 is this program's a great example of how we've used our phenomaps and biological insights to really identify a novel therapeutic target. So back in 2021, using genome scale phenomapping, which are basically our maps of biology, what we uncovered is a really important relationship. We found that RBM39 was actually a functional analog of CDK12, and this is an already well-established oncology target. So what we were then able to do is profile a number of compounds in our phenomap to show functional biology and phenocopying of the CDK12 protein. We did this rapidly, and we were able to optimize degraders and land on our lead candidate with only 200 molecules synthesized, and it went from target identification to IND-enabling studies in just under 18 months. So very proud of how fast we were able to do this. Again, all of this in an effort to improve our clinical success, but also reducing the time and money that traditional R&D processes take. On the next slide, we've expanded the map that I showed you on the previous slide. In the middle is the smaller portion of the map that I showed on the previous slide, showing the RBM39 and CDK12 functional relationship. But if you zoom out, what you see is that there's also a relationship with DNA damage repair, epigenetic regulation, transcription, cell cycle control, and of course, all of this is very consistent with RBM39 biology. So this really validates our phenomap to be able to uncover the actual biology that's taking place. So very quickly, mechanistically, RBM39 is a protein that regulates splicing fidelity. So when you actually degrade this protein, what you end up getting is widespread splicing defects. You can imagine then that tumors that have global genomic instability, high replication stress, or have DNA damage repair defects would be quite susceptible to RBM39 degradation. And so that's what we went forward with, and the hypothesis we tried to confirm here. So one evidence of genomic instability are tumors that have high MSI high status, so microsatellite instability. And what we're showing here is that in a model of ovarian cancer that has MSI high status, you can see dose-dependent degradation of the RBM39 protein. And this actually translates in vivo, and you can see monotherapy activity of REC-1245 in the model that has microsatellite instability. We took this one step further and confirmed this in a number of other models. On the left-hand side is another ovarian model that has MSI high status. This model also happens to have DDR defects, again, going along with the biology we showed before. And then when we profiled not only additional MSI high models but compared these to MSS models, so those with more stable genomes, you could see that there was a trend that those models that had higher global genomic instability were more sensitive to REC-1245 than those that had more stable genomes. All of this led to our clinical study design and our eligible population. We have an ongoing Phase 1/2 monotherapy study called the DAHLIA study. And what we've done with this study design is allowed patients that have tumor types that naturally have high levels of genomic instability, as well as biomarker-enriched populations of patients with tumors that have high MSI status. We're continuing dose escalation right now, and we will be hoping to share a preliminary safety and PK update in the first half of this year. But what we really wanted to highlight is that RBM39 and the REC-1245 program is a great example of how we could take a novel biological insight and actually turn it into a drug program, and that led us to a translational hypothesis and eventually our patient selection and clinical study design. Next, we have the REC-617 program. This is a CDK7 inhibitor. For this program, we leveraged two parts of the platform, the precision design piece of the platform, where we really optimized this molecule to have properties that we were looking for to increase the therapeutic index.Now, it took us only 136 molecules to come up with our lead candidate, and we did this in under 12 months. Again, really highlighting the speed and efficiency with which we can do this type of work. Today, I'm going to focus more so on the right-hand side and how we got to the decision to prioritize ovarian cancer as our indication in our first in-human trial. So again, for this program, there are a number of biological insights, the generative AI and active learning that we did to get to a lead candidate and use our precision design capabilities. We have our in vivo data that I'll show you on the next slide, but really, again, I want to focus on some of the clinical data and the causal modeling we did to identify ovarian cancer as a sensitive population. So here on the left, you're seeing an in vitro ovarian cell line that has sensitivity to REC-617. This is dose-dependent cell growth inhibition, and this translates into the in vivo model where you see all three doses of REC-617 as a monotherapy produce not only significant tumor growth inhibition, but actually tumor regressions in all three doses. Based on some causal inference and multi-omic and clinical data, what we were really able to show is the relationship between CDK7 and a vulnerability in ovarian cancer. So on the right-hand side, what you're looking at are thousands of patient records that we mined and analyzed to show that patients with high CDK7 in their tumors had lower overall survival and poor prognosis versus those patients that had low CDK7 expression. And again, this helped us establish more of a causal relationship for CDK7 in ovarian cancer versus just a correlative one. And I bring this up because a number of other CDK7 programs are actually going into a different indication. And so we followed our data, followed our AI-enabled insights to really prioritize ovarian cancer as a differentiated indication for this molecule. Importantly, in our dose escalation, we saw that we had a PR in a patient that had ovarian cancer. Again, this really just confirms our hypothesis that ovarian cancer patients would be sensitive to a CDK7 inhibitor. Our Phase 1/2 trial is ongoing. It's called ELUCIDATE. We've reached MTD. We're exploring alternative dosing schedules, and we've started our combination cohorts as well. We look forward to sharing an update on all of this in 2027. The final program I'll turn to is the REC-4881 program or MEK1/2 for the potential treatment of familial adenomatous polyposis, or FAP. So this program actually came out of an insight from one of the earliest versions of our platform. And what the insight actually showed us is that we could take a known mechanism inhibiting MEK1/2 and actually repurpose it for a completely new indication in which this mechanism has never been studied before clinically. Currently, we have an ongoing trial, and we're using our clin tech part of our platform to really use our real-world evidence analyses to contextualize this data, which is super important because this is a rare disease. So a little bit of background here on FAP. This is a disease that's caused by the loss of function in a gene called APC. And when you lose APC function, patients with FAP will develop hundreds to thousands of polyps in their colon and rectum. And if they're left untreated, nearly 100% of these patients are at risk of developing colorectal cancer. About 50,000 patients in the US and EU5 have FAP. And unfortunately, these patients undergo a lifetime of endoscopic surveillance, frequent excisional interventions, and life-altering surgeries, and all of this leads to very poor quality of life, morbidity, and poor long-term outcomes for these patients. So really highlighting the unmet need for this group of patients. So within the rare disease space, natural history data is very important in order to be able to contextualize the data that you're getting from clinical trial results. Unfortunately, for FAP, there's little known about the disease outside of clinical trials. And so our clin tech team aimed to fill this gap by building a-- I'm sorry, I missed a slide. Let me first tell you about how we discovered 4881. So REC-4881 came out of a phenotypic discovery map, essentially where we knocked out APC, and we were able to show that after profiling thousands of compounds, that 4881 was able to produce a rescue effect in these cells. So it converted a disease state cell into a healthy state cell. Now we'll move to the natural history part. So for rare disease, again, in order to contextualize your clinical data, natural history data is very, very important because there is just not a lot of ability to often do controlled trials and understand what your trial results mean. Unfortunately for FAP, a lot of quantified data doesn't exist outside of clinical trials. And so this is where our clin tech team aimed to fill that gap. So they created a custom LLM analysis of 1,000 US FAP patients using over 250,000 US physician notes in order to be able to understand and contextualize the natural progression of polyp burdenHere you can see that we did an additional follow-up with the Amsterdam University Medical Center and used their largest and longest running FAP registry to additionally contextualize our results coming from our clinical trial. This registry actually has over 20 years of follow-up for about 200 patients. And so what this allowed us to do was really understand at scale for the first time how polyp burden progresses in these patients. And you'll notice that 87% of untreated FAP patients had annual polyp burden increase shown in this waterfall plot here, and an average of 60% polyp burden increase annually. I want you to keep this graph in mind as I move to the next slide because it's really going to show you the difference between an untreated population and our trial results coming out of our phase one, two Tipelo trial. So here you see the increase in polyp burden. However, when we look at our trial data, so these are FAP patients treated with 12 weeks of 4881 therapy. And you can see after just 12 weeks of therapy, a reduction in polyp size, not an increase, compared to the natural history data. 75% of the patients responded, and the overall median polyp burden reduction was 43%. Importantly, the safety profile was consistent with MEK1/2 inhibitors and this class of compounds, and the majority of the adverse events were grade one/two. What's really interesting is what's coming up on the next slide. So not only did we see this polyp reduction at 12 weeks, but we saw that there's actually durable polyp reduction response in these patients. So this is a graph showing the polyp burden after being off therapy for 12 weeks, and you can see that 82% of the patients remained responders and that the polyp burden reduction remained at 53%. So this is very exciting for the Recursion team. This is the first time we've shown clinical validation of our platform insight and really showing the full arc of how you can go from a biological insight all the way to clinical validation. And I think what's very, very interesting about this program, again, was the ability for us to take our map and actually come up with a novel indication for a MEK1/2 inhibitor that has never been uncovered before. So hopefully I've shown you today the different ways in which Recursion and the way our industry can use AI across the R&D continuum. We believe that it's really important to make sure that we're answering key questions all along the way from the biology standpoint, making sure that we're identifying novel targets. From a chemistry standpoint, using AI to really precision design medicines that are differentiated and optimized, and then in the clinical space to really use AI for selecting the right patients as well as designing your trials in a way that will hopefully make them be as successful as possible. We have a couple of additional things that we're trying to do. Recursion is always innovating. One thing that we're doing is building something called Cell Neighbor. So both of these things on this slide are aimed at really trying to close the translational gap that exists. So for Cell Neighbor, what we do is we actually map cell lines to patient tumor biology. And what that allows us to do by putting these both in the same transcriptional space, in practice, it allows us to use models that are more translatable to patient biology in the clinic, but it also allows us to avoid models that do not translate into the clinic. Again, a very important piece of R&D in cancer research. We're also building TxFM. This is a state-of-the-art transcriptomics foundation model. The goal here is to really bridge the gap between lab perturbations and patient biology. We're very excited about this. It's early days, but we believe that it has widespread applications that will be able to use TxFM across target identification, mechanistic understanding of our molecules, and as well as patient selection. I'll just highlight that our Cell Neighbor will be shown at AACR on Monday. There's a poster session from 9:00 to 12:00, so I highly encourage you to check that out, and the TxFM foundational model will be shown at a conference next week. In closing, I just want to again reiterate our true belief that applying AI across the R&D continuum is critically important to make sure that we're driving informed decisions at every step of the process in order to increase our chance of success and to reduce time and cost. And thank you again. Thank you so much, Jeroo, for the fantastic presentation and for showing us how you can deploy AI solutions in a way that guides your, let's say, drug discovery and clinical development in such a transformational way. I have one quick question before we open here too for questions that we have already by the microphone. So for CDK7, you made a really important point that others are pursuing different indications, and you went into that particular indication because of the insights that you are deriving through your AI approaches and the new models that you have built. Could you just elaborate a bit more as to how did you get yourself comfortable enough to challenge the current status quo and to venture into a completely different area? Yeah.I think it was really a data-informed decision. So we had the in vitro and in vivo preclinical proof of concept. But I think really having thousands of patient data sets from Tempus that shows that causal relationship really drove that decision at the end of the day and helped us prioritize. It also helped that we saw a partial response in our ovarian cancer patient. We allowed advanced solid tumors in our dose escalation study. But it was the ovarian cancer patient that had the partial response. So that further validated and really made us feel confident in that decision. So just to test my understanding, it was both the mechanistic explainability together with the additional reasons you believe based on validation. That's right. It's a combination of the two. Yeah. Microphone starting here on the left. Please identify yourself and then ask the question. Thank you. Hello. Oh, excellent. It works. Dr. Behani, thank you so very much. That is a wonderful presentation. I have a question on the Recursion race. It's clear that the link through across the entire value chain is substantial. My question is around the benchmark. How many molecules and mechanistic analyses actually link through to IND, and what's the baseline? That's a really good question. You mean across our pipeline? That's correct. How many go from left to right all the way through our platform? Mm-hmm. Again, we don't use our platform uniformly across programs, and you may know that Recursion also had a business combination with a company called Exscientia that also brought a couple of molecules into our pipeline. They had their own precision design platform that then got incorporated into our overall platform. So again, not all of the molecules go from left to right all the way through. But the important piece, I think, is making sure that we're using the right pieces of our platform to answer the specific challenges of each program that we have. Thank you so much. So here on the microphone on the right. Please identify yourself and ask the question. Thank you. Thank you. I'm Mila. I'm a biomarker scientist. It's really exciting to say you already use the foundation model for the drug development. I also have question about CDK7 because CDK7, it's in ovarian cancer, right? It's a enzyme. So my question is, are you using CDK7 expression for patient selection strategy or not? I'm also wondering about the response difference between, for example, CDK7 high expressers versus CDK7 low expressers. Yeah. As of now, we're not selecting patients based on CDK7 expression, but it's something we're continuously evaluating to understand baseline profiles of all of our patients to really understand what our patterns of response are going to be. Okay, thank you. Just to test my understanding, how are you incorporating that then into your clinical development program? So just continued analysis, interrogation of our clinical patient data to see where the patterns of response will be. Is that what you mean? Yes. And what would trigger then a decision to go with a patient selection strategy that is based on some type of phenomics? Yeah. It would have to be a large enough data set that really gives us confidence. So what we're constantly doing at Recursion is mixing our preclinical data, our clinical data, and our computational abilities to continuously interrogate and ask these types of questions as incoming data is coming in from the clinic. Right now nothing to share there, but that's a continued analysis that's happening at Recursion. Thank you so much. Microphone here on the left. Thank you. Hi, this is Marta Milo, AstraZeneca, Oncology Data Science and AI. Thank you very much for this inspiring talk. I'm intrigued about the causality and how you use causality and AI in these particular case studies that you showed us. Could you elaborate more how you use causality, how you evaluate it, and how you evaluate the uncertainty around causality? Mm-hmm. And how you use the data to model it. So I'll use the CDK7 example here. So what we did to make sure that it was truly causal in nature that the CDK7 high expressers had overall worse prognosis, is that all of the other variables in the analysis were actually fixed so that we could really get gene-specific changes and causality based on each gene that's changed. That's about as much as I can say. I have my computational biology leader here. I'm happy to answer more questions after the talk. Thank you. Thank you. Microphone here on the right. Hello, thank you for your presentation. I am Min Jung from Berkeley. I'm from Witgen, which is developing virtual single-cell Gen AI. So thank you for the presentation. It was very good to see all these kind of assets that get into clinical trials. Because maximizing AI utility in the drug discovery is kind of very difficult and was taking a long time. So my first question is, how are you evaluate the assets that Recursion Therapeutics have done compared? How fast do you get from the conventional methodology using AI? And the second question is, so clinical data is especially very difficult to acquire for a pharmaceutical company. So how Recursion Therapeutics are acquiring clinical data to get into clinical trials, and are you using also the kind of data in the preclinical trial in order to bridge the gap between in vitro and in vivo?Okay. So your first question- Was how faster are you getting? How fast? Okay. Is there a benchmark? Thank you. It's been a long couple of days. How fast? Well, hopefully, we showed you that with the CDK7 example, we only had to synthesize around 136 molecules in under 12 months in order to land on our lead candidate. This is again, because we're using closed loop, AI guided design for our molecules that allows us to rapidly optimize our chemical series. And then the same thing for RBM39. We only needed about 200 molecules. And we went from target identification to IND-enabling studies in about 18 months. So again, both examples of just how fast we can go if we're using AI guided information. Kira, just one question so that others who are not as familiar with this. What are the industry benchmarks for the 136 and 200? That is a good question. I think that's probably beyond my scope. I've been with Recursion for just a couple of months now, but happy to entertain those questions again. I have the full team here with me. So let's continue that. And the access to the data, the second question, how do you have access to clinical data? That's a real challenge for pharma. Yeah. So, the CDK7 example was using Tempus data. So we had a relationship with Tempus to evaluate thousands of patient data sets in order to really come up with that causal relationship. And then you saw for the 4881 trial, we really utilized that Amsterdam UMC registry again, because they have the longest and largest registry for FAP patients that have such long follow-up. But in general, your question is a good one that where can we fill in the gap with preclinical data? I think that's where, again, I just reiterate that we have data coming from multiple sources, right? We have preclinical data. Even for our clinical stage compounds, we're continuously generating preclinical data to try and better understand our molecules. We're getting clinical data from our ongoing trials, and then we have these data sets like Tempus' data set or other registries that allows us to put everything together to really have a large enough data set to make some clear associations and causal relationships. Thank you for this. Thank you. Microphone here on the left. Hi. I'm Catherine from University of Toronto. My work involves CDK13 and splicing, so I'm really curious about the RBM39 project at Recursion. I'm curious about what's the advantage of designing or looking for genes such as RBM39, who mimics the loss of function of CDK12 versus improving the existing current degraders of CDK12 and 13. And kind of to follow up on that, did you observe any compensatory effect of CDK12 and 13 when you inhibit RBM39? Thank you. Yeah, sure. So one thing I didn't talk about in the talk is that when we saw this initial relationship between RBM39 and CDK12, what was key was that at that time, CDK12 inhibitors often were coming with this liability that they would also hit CDK13. And I think there was a belief that the inhibition of CDK13, along with CDK12, was really causing toxicity issues and causing really narrow therapeutic windows. So that's why we were so excited when we saw this relationship, that there was a way to have functional analog of CDK12 by targeting RBM39, and that way you don't have to inhibit CDK13 at the same time. And the hope is that you'll have an improved therapeutic window. As far as your compensatory mechanisms question, of course, for all of our programs, we're continuously trying to understand where we have the greatest benefit versus not, or greatest efficacy, and constantly mining the data to understand if there are compensatory mechanisms that are popping up that would reduce the effect of our drug. Can't share any data on that right now, but that's something we're always doing. Thank you. Yeah, the microphone on the right. Hi. Thanks for a great talk. Victor Chubuko from Freedom. I'm curious if the MEK inhibitor continues to be efficacious in FAP. Do you see a path forward in sporadic CRC, and why or why not? Potentially, certainly that pathway is relevant in CRC. I think it's a pretty crowded space. And I think that with the FAP program, we have the real opportunity to tackle an unmet need there. But certainly, the fact that these polyps will eventually turn into colorectal cancer, it does give you the hypothesis that this could work in sporadic CRC as well. But not much more to share there. Thank you. Here on the left. Hi. Thanks. Excellent talk. Some of the questions I had already answered, but I have one more. My name is Itai Jacob from AstraZeneca. I think your brilliant operating system from biology to insight, from insight to molecule, and from molecule to patient is amazing. Wondering, have you been leveraging that to actually train and build models? For example, the model that you have, the transcriptomics data foundation model, for example, and other models that you've probably been building. So have you been leveraging that too as part of the training strategy and model architectures that you've been developing? Yes. Again, I'm not an AI person. I know I said that we're all bilingual, but I've only been here for a couple of months, so I'm going to need a few more months to really become bilingual in AI. I'm much more of a translational medicine subject matter expert. But what I can say is that we're constantly trying to improve each pieces of this platform. And like I said, the translational gap, we're really trying to close that by using these transcriptomics and foundational models that we're building. So yes, Recursion in general, at a high level, is constantly trying to innovate in order to do things better, to do things faster. Hopefully that answered the question, but again, the whole team is here with me, so please feel free to come find us after the talk. It would be great to touch base on that more. Sure. I think that this is the frontier, and you guys are at the frontier, so I would like to learn more about how you're actually leveraging all this end-to-end to build those models. Yeah. That would be great. Thank you. Thank you so much, Ita. In the interest of time, perhaps you can ask the question to Tiru after the session concludes. So I apologize, but we need to get to the final talk. Thank you so much, Tiru, for the excellent talk and for fielding so many questions. Thank you. Okay. Now I have the distinct pleasure of introducing George Reis-Filho. He's the Chief AI and Science Innovation at AstraZeneca, and his talk is AI Transformation of Pharmaceutical R&D. Thank you so much. It's such a distinct pleasure to be here with you today and to see so many familiar faces and so many colleagues, and actually so many frontier AI people in the audience as well. I feel really humbled that all of you are with us here today. So what I'd like to do is to discuss with you how we are seeking to apply, really intentionally, AI to transform every aspect of what we do in R&D. And we are a big pharmaceutical company. So how do we do it? So these are my conflicts of interest from my distant past, but they remain current anyhow. So let's get started here. Where this has all begun. I think from our perspective, I would submit here that critical developments in oncology research and also oncology clinical development actually mirrored technological developments over the last 30 years. I think it's self-evident. If you stop for a second and you think about it, there is a clear correlation, and it's causal. It's not only a correlation. Actually, we are moving, and we have moved from the idea of requiring extremely well-curated, extremely harmonized tabular data sets, to actually using raw multimodal data. Can you believe that in the space of what? 20 years. And if you recall the big data thesis of 2017, 2018, we have moved such a long way. It's quite incredible. Second, as Jacob alluded to in his talk, we are moving from that idea of handcrafted annotations, fully supervised models, to actually self-supervision, extracting features, and utilizing the raw data first principles. And actually, we are able to derive many more insights by doing that. Not only that, we are going from self-supervised learning in a single modality, the unimodal foundation models, to multimodal foundation models. Not that we're quite there yet. We have not yet solved the problem of the semantic layer across data modalities. It's a problem for us to solve together, I would submit. In parallel with that, we have seen the birth and the growth and the transformation brought about by agentic AI. And no longer can we code tools. Actually, we can perform relatively complex tasks in a semi or completely autonomous manner now. And we have moved into the space of coding agents now. And as of last week, we've seen the development of automated alignment researchers. Just think about that for a second. Automated alignment researchers. So models can train models and get better performance without a human. It's quite incredible. And we are moving even further here, that of going from agency to delegated awareness for the agents. Again, let's just stop here for a second. Let's think here. Delegated awareness. So the agent understand the system that we are actually confronting and is able to reason what are the best solutions, the most possible solutions for the problems we have. So it is something that's happening. It's not science fiction. It's happening today. Can we again agree, all of us, that AI is no longer the future, it's today? Just for a second, can we just share that reality? So if you accept that, then you understand that what our premise was back then in 2025, when actually we reasoned that our core thesis was this. That AI is going to change every aspect of what we do in a big pharmaceutical company, from accelerating drug discovery, to transforming clinical development, to optimizing clinical operations, to delivering journeys for patients that are more seamless, more impactful, and more meaningful.So with that key thesis, our premise was that frontier AI models would outperform us in reasoning and cognition tasks by the end of 2028. That was our premise, that we'll get there before the end of 2028. And why would we come to our leadership to articulate that premise? For three reasons, yeah. First, even last year, even more now, if we take these AI models offline and do an IQ test without any data leakage, the IQ sits comfortably above 145. It's a reality. Second, if you look at the way these models are evolving in relation to the state-of-the-art benchmarks, they're just advancing so quickly. First, they were increasing one order of magnitude every eight months. They are now doubling every four months, and it's quite incredible. Some models have reached a level of reason and cognition that the developers are uncomfortable in releasing it in the wild. And third, that the length of tasks that can be autonomously performed by agents keeps advancing almost now on a weekly basis to the point that coding, that was an incredibly important skill that we need to have, and it was really hard. You needed specialized knowledge. Actually, it's available to all of us now. It's not solved everything, and that's not what I'm saying, but it's available to us, and we can iterate, we can build things, first principles ourselves. Even if you don't know how to code with Julia to optimize the hardware that you have, you can do it now with some of the solutions available. So with that in mind, what I'd like to focus on today is that one of our key strategies relied upon this particular principle, that by applying frontier AI models to our key differentiator, our data, we could transform clinical development in oncology. That was one of the key elements of our strategy. And what do I mean by that? It's a lot of big words, yeah? But what do I mean by that? AstraZeneca has intentionally developed a uniquely diverse portfolio of therapeutic agents. Five different classes of agents that you can see here on the right-hand side of this color wheel. On the right-hand side, you can see five different classes of agents that target the cancer cells directly. ADCs, radioconjugates, molecules that target DNA damage response, treatments that target disease drivers such as EGFR mutations or ER signaling, and also epigenomic alterations that locking of the cells into a specifically state through epigenetics. And we also have three different classes of therapeutic agents that leverage the immune system to attack the cancer cells. We have bispecifics in terms of the traditional IO approaches. We have synthetic immunity in the form of immune engagers and cell therapy for not only hematological diseases, but most importantly for solid tumors as well. So this diversity enabled us to accrue translational data over the last 15 years. And what we're doing here is then we're using our data to reinforce, fine-tune, and augment frontier foundation models that we source into our system. And why do we use that strategy? Why are we not developing the foundation model ourselves? Because in this space, in clinical development, because the performance of our foundation model depends upon the data diversity and volume. So the diversity across data modalities and the diversity within a data modality. For anyone to attain state-of-the-art performance, very large volumes of data with the right level of diversity are essential. So we have decided to form partnerships to bring these frontier domain-specific foundation models. We created a partnership with Tempus and Pathos on one hand to develop the largest ever multimodal oncology foundation model for one reason. Because they had the best, let's say, multimodal clinical oncology data set available at that time, with EHR data, de-identified clinical notes, lab tests, pathology images, pathology reports, radiology reports. So we could use that to pre-train a model and then reinforce, fine-tune, and augment with our data to derive new insights. In the pathology space, we have also intentionally benchmarked several models and created a partnership with Model AI that led to the acquisition of the company earlier this year. But the reason to bring these models is that the insights that we generate are to address four main tasks that we have as a pharma company. First, to reimagine patient selection through a multimodal biomarker space. But here we are not trying to do this, the most complex biomarker that you'll find, a small population that respond. No, we are developing this on the basis of maximum parsimony.So what is the smallest set of parameters that would be required and sufficient for us to identify the largest population that would objectively benefit from one of our therapeutic agents alone or in combination? Second, we are using this approach to de-risk the transition from our Phase 1, Phase 2 studies to Phase 3 studies. Why? Because these models are able to generalize from smaller sample sizes. So the idea here is, okay, we don't want to be sitting with 40 patients' worth of data after Phase 1 and render that critical decision to move to a Phase 3, having that level of uncertainty. If we can bring the insights from these models, we can make better decisions. That's why I was asking the question about the pursuit in that domain here. The third, how can we improve the probability of success of our Phase 3 studies? And here, it's using another property of the foundation models as well. That's to generalize from sparser data. When we have a Phase 3 study, we don't have the depth of translational data that we have from a Phase 1 or Phase 2. So can we use those models to help us actually render better decisions there as well? And all of this, to be honest, is just to address our fourth point. We'd like to develop the next generation of combination treatments that will be transformational for patients. I think it's a shared goal that all of us have to move the response rates to a much higher level through the combination of therapies, and I think we'll get there together. At the end of the day, the ultimate goal is to transform patients' lives and actually to advance our bold ambition in oncology. That's to eliminate cancer as a cause of death. I think it's a shared goal that all of us have here. We realize that that approach with the multi-modal foundation models coming from a clinical perspective is one of the scales that we need to be aware of. We also understand now that for us to get to the point that I'm articulating here, we'll need to have the different scales for the development of these domain-specific models, from an atomic level to the cellular level. And you hear a lot about virtual cells, and I'd like us to define exactly what you mean when we say a virtual cell being developed, and then from the patient level. What we are learning as well as these are developed is that if we take the embeddings from one scale and actually refine and fine-tune them with the others from another scale, we get even better insights. So it's likely going to be an ensemble of models that will allow us to get to the point where we need to be. Yes, to derive those insights that actually will be transformational for patients at the end of the day. But we needed to start somewhere. And I have to say that for us at AstraZeneca, AI is really not the future, it's our reality. We're doing this. We actually have been pursuing this for over a decade. And one of the spaces where AI has had already an impact is in computational pathology. That idea that we can apply artificial intelligence to computer vision of pathology images actually started over 10 years ago. First with deep learning, convolutional neural networks applied to pathology images. It was really exciting, but now we are getting to the point where we're seeing the maturity of those solutions. A great example is our quantitative continuous scoring system. That's basically a fully supervised method where we use handcrafted annotations to develop a model that can quantify the expression of targets at a subcellular level, membrane, cytoplasm, and nucleus, in a really quantitative way at the single-cell level. We actually applied this first to the development of our antibody-drug conjugates. The first was our antibody-drug conjugate that targets Trop2. In that space, immunochemical assay, semi-quantification by pathologists, people like me looking down a microscope, actually, we could not get to a predictive biomarker. With the use of the quantitative continuous scoring system, we have now demonstrated in the context of the tropionlongo one and tropionlongo two that actually you can quantify the expression of Trop2 on the cell surface in the cytoplasm, derive what we call the normalized membrane ratio. And that normalized membrane ratio is a surrogate for the internalization of Trop2 as a target. And that is a surrogate marker for the response to our Trop2 target in ADC. And actually, this is a great example of one of the types of biomarkers that Jacob Cutter articulated in his talk, the type C, yeah, where we don't have an equivalent by humans, and we are developing something new here, something that cannot be quantified trivially with our approaches. It's really hard then to develop and validate this. You need validation in multiple datasets. And actually, we're seeing the impact of this biomarker in the next few months, few years with some of the prospective clinical trials we're running at AstraZeneca in the space of non-small cell, non-squamous, non-AGE lung cancers. The reality, however, is that for us to develop the solution that received breakthrough device designationBy the FDA. The first ever to receive that designation in that context of therapeutic prediction. Actually, for us to develop that model, it took us... I will let you guess. Imagine. How many hours of pathologists handcrafting annotations to train the model were required? Do you have an idea? So I'll tell you, 40,000 hours of pathologists to develop one model for one indication, one tumor type. It's hard to scale, isn't it? So we asked a question last year. What if we brought a pathology foundation model, highly capable, to actually reduce or even replace the need for handcrafted annotations for the development of the model? Not the biomarker, the model to quantify a specific target. And what we have witnessed is that you can go from 40,000 to four and a half, even to zero. You just do the benchmarks and the validation extensively. It's quite incredible, isn't it? And it's extremely exciting because not only for us, it's for the whole community. This is a principle that if we have the most capable models, we can start pressure testing, as long as we have very clear evaluations and benchmarks. Those are absolutely essential. And I'm going to reiterate again, evaluations and benchmarks for everything that we do, and then extensive validation of the biomarker once the biomarker is developed. Where we're trying to get to now is to integrate not only pathology, but multimodal data to help inform patient selection strategies. And for us to do that, we need to be able to connect our data with the models to develop insights. And for that, we are developing a whole agentic framework to do just that. That circuitry that's required and sufficient to connect our data with the models in a way that we derive the embeddings from these models. We can actually use them to derive the new hypothesis and test them, and even generate new data to test the hypothesis. But for us to get there, I'm afraid to say that it's going to be really, really hard to do this as an island, and I would not suggest you to try this at home. Actually, rather, what we need to do is to build an ecosystem of partnerships whereby we can actually bring together domain experts on one hand with people with extremely deep technical expertise, the right orchestration on top, and then at the bottom, the right infrastructure. And when I say infrastructure here, it's your data infrastructure, your compute, all of the platforms that are required and sufficient to do that. If one can bring all of those components together, then we allow ourselves to ask the killer questions, the questions that actually really matter for our patients. So we have been building this ecosystem with hyperscalers on one hand, with the frontier AI companies, with specialized AI companies. And we have several declared partnerships with Tempus, with Pathos, with Modelo, with Immune AI, with Okin. This is all happening. And we also have partnerships with academics. And I have to say that I've joined AstraZeneca, and I have an incredible number of meetings with academics who are actually developing the next generation of solutions, seeing from around the corner, and we need to be there because the pace of technological development is unprecedented. And finally, that actually we started last year in September, we've brought into these ecosystem partnerships patient advocacy groups. Because developing the AI solutions without having the voice of the patients is really, really hard. We need to understand what their needs are. Not that they will constrain everything that we're doing, but having that information help us define what the path should look like. Another point that I would like to discuss with you is that we also approach AI from a three-level perspective. As I mentioned to you, in clinical development, if you think about the first layer, the model layer, the data of incredible volumes in adequate diversity with the edge cases being adequately represented is essential to develop the most capable models. And I subscribe to the notion that we first need the most capable models, then we can distill them down to find, let's say, the smallest model possible that would provide the answers to us, but we need to develop the most capable to begin with. So we are actually there developing these models through partnerships. Where we actually are actively engaged internally is in developing the solutions, the second layer, utilizing those models to address specific questions. For that, we are also building a really complete stack of evaluations and benchmarks, task-specific and end-to-end, to test how performant these models are. With those solutions, then we have to develop the third layer, that's the applications. What will touch the patients? What will actually get to the point of informing decisions? And we build this in-house using our datasets, but we actually partner then with diagnostic companies, tech bio companies, biotech companies, to get this to the patients. We're not a diagnostic company. Our primary purpose here is to develop the transformational treatments to our patients.So we are focused on our primary purpose, and we do this through partnerships. And I'm certain that by bringing all of this together, this ecosystem of partnerships, and understanding the roles that we'll be playing, we can get to the next wave of innovation. One final point, one final thought that I'd like to share with you, that I've come to realize through experience and exposure, first and foremost. I had not thought about it when I first started, that AI in a pharma company of the size that we are is not a technological challenge, it's a challenge of people, process, and culture. And we need to help and enable our team members to upskill and to transform themselves. And we have to reimagine processes, first principles, to deliver the solutions that actually will be transformational for our patients and will allow us to take many steps forward towards our bold ambition in oncology. That is, to eliminate cancer as a cause of death. With that, I'd like to thank you so much for your attention, and I'd happily take any questions you might have. Thank you. Thank you very much. Kiro, I'm afraid that we are the only thing between them and lunch. So let's just be aware of that. Sounds good. While people are making their way to the microphone, I'll give you a question that came up online. Have you used BCMA and CD33 also as an ADC AI-induced biomarker? I'm not in a position of commenting on specific solutions here today, but we are investigating all of the different potential strategies for patient selection, not only in solid tumors, but in heme as well. So watch this space. Sounds good. We'll take a question from the right-hand side. Hi, George. Thank you for the very inspiring talk. I embrace your enthusiasm, okay, but here is a question. You mentioned a lot-- Sorry, Andrea Sotteriva, Computational Biology, Human Technopole Milan. You know me. You mentioned a lot foundation models, multimodal foundation models. Now, in language and images, they work very well because their information is very dense, okay? And we have three, four orders of magnitude more data, okay, for free from the internet. Now in biology, information is sparse and data is scarce. And I understand you propose, as you said, to build up this data infrastructure. But realistically, if you take off your AstraZeneca hat and you keep your research hat, as a scientist, as humanity, how do we generate enough of this multimodal data in biomedicine to really build models that are predictive? Andrea, that's such a great question, and thank you so much for asking. I think here, if you think about petabytes in biology and clinical development, a lot of it lies in the imaging space. A lot of petabytes of data from radiology, from pathology images. What we've learnt over the last few years is that if apply self-supervision to pathology images and radiology images, if we solve feature extraction really well, we can derive new insights. Really new insights. And we are putting this to test, and actually, there is a delta between just doing this with traditional approaches, and if you use the embeddings to derive the new insights. So this is one modality. Then when you combine pathology images with pathology reports into a vision language model, you derive even more insights. And actually, you can use then those models to help you understand the mechanistic basis of some of the predictions made. So it's building that stack. We are also seeing in parallel many companies developing extremely sophisticated strategies to develop the data sets required for the training of these models. So we are getting to that point where the data is going to be sufficient for us to develop the first generation. What is really important here is that it's not been solved yet. Two problems have not been solved yet. The development of the semantic layer across data modalities in a truly multimodal way, and the second is that the utilization of time as a first dimension in the multimodal embeddings. These are two problems that we hope to help solve over the next year. Does that make sense? Yeah. We can talk a bit more later, maybe, but thank you very much. Thank you. We'll take a question from my left. Hi, George. Thank you for the amazing talk. My name is Arshad Ahmed. I'm with OneCell AI, and I'm a little biased here, so just bear with me one second. A lot of the discussion centered around, we hear Tempus a lot, in terms of the data and basically the foundation of the data. If that's not solid, then everything else falls apart. Right? So I see three big limitations, and I would love to hear your thoughts on those. One is these are done in archival samples, right? These are not fresh biopsies. They're done in bulk. These are not done in single-cell level. And these are not done in longitudinal samples, right? So these are done at the baseline, but it's hard to get biopsies longitudinally unless you do a ctDNA, but it's not single cells.What's your view on the next generation of the Tempest database that'll drive the next generation of foundation models, which will be fresh, longitudinal, and single-cell level data, which also have perturbations in it? Is anyone working on that, or is that a white space? Is there even a need for it? What's your view on that? Thank you so much for the question. I had that slide with the atomic, cellular, and patient level. The patient level is what we could do today. And actually, one important point is that those longitudinal samples are like gold dust, because if you can include them in the pre-training, it becomes really more powerful. Second, we need to bring the time dimension to that, and one can start deriving that dimension through the de-identified clinical notes. Because there are so many time points, we may not have the samples, but we can start understanding chronology and temporality. Now, your final question, do we need to generate the single-cell data with perturbations, let's say, where we understand exactly which parameters we changed, and quantify before and after? Absolutely. And I believe that many in this ecosystem are directly involved in the generation of those data sets. And they will actually bring the next generation of the models. We'll take another question from online. What is your perspective on the black box nature of these AI tools and their use in a regulatory-heavy environment? That's such a great question, actually. I'm not sure if that word black box is applicable anymore. Because actually, the level of system introspection that these models can perform today is quite incredible. The regulators are actually evolving with us. Before, mechanistic and explainability was a condition that was sine qua non, was mandatory. And as you could see in the ESMO guidelines, the EBHI, and also in discussions with several regulatory agencies, the main requirement now is multiple validation of the solutions derived. If we have explainability, even better. But what's essential is that we demonstrate the performance of the particular AI solution in terms of clinical utility, very clearly, with adequately powered data sets, and done in a way that's scientifically rigorous. That's what I would submit. But the partnership with the regulatory agencies now is super exciting because they're prepared to iterate with us and to think about the next wave of innovation. I think we can have time for just one more question. We'll go to the right of me. Hi. Thank you so much for the great talk. My name is John Wu Kong from UC San Diego. I had a really similar question from the question actually asked by the online audience, which is model interpretability. And my two questions, one question is how much validation would be needed? For example, how many cohorts or how many number of samples in addition to the training data set and the validation and benchmarking data set, can we be sure that the model is we think is going to be generalizable? Generalizability is the question. Absolutely. And the second question is, once you go beyond unimodal foundation models and building a multimodal foundation models, I would think that the interpretation will be much more harder because you now have to investigate not only the unimodal model features, but you have to really understand how these multimodal features interact with each other, which raises another layer of complexity in terms of interpreting the model. So I'm wondering, what's your thought on that? So let's begin with the second question first. Mm-hmm. Let's imagine a narrowly fusion model where most of the additional information derived is in the form of nonlinear correlations in the latent space. How do you explain this, yeah? And that's why the regulators are saying you need extensive validation, because that we may not be able to semantically translate into words. But for us to get there, because then we are getting into the territory of open-endedness, aren't we? We don't know the answer to the question. We're trying to derive a new question because our traditional train of thought cannot get us there. Is that correct? Yeah. If so, then we need evaluations and benchmarks for all the sub-tasks and tasks as we progress. And then we need validation in multiple data sets. The magic number, I do not know. It's going to be context-dependent and solution-dependent, and we'll need interactions with the regulators so that they can articulate what would be required and sufficient for us to satisfy the principle of safety for the patients, and that the AI is actually performant enough and robust enough to be shipped clinically. Got it. Thank you so much. I think we're all set, and I think we're going to have to move out of this room. So we have to vacate the room. First, I'd like to thank you so much for your attention. I'd like to thank you so much for asking the provocative questions and for pressure testing us, and for sitting through this session that despite the logistic disruptions that we faced, Lufthansa did not help. We actually got the message across here, and I would like to thank you again for spending this Saturday with us. Thank you.