2026-04-22-agentic-ai-as-the-oncologist

Date: 2026-04-22 Primary topic: agentic-ai Cross-referenced from: agentic-ai, bioinfo-tools Word count: ~12,693

Auto-generated captions

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

Let's get started. Good morning, everyone. Welcome to this session at AACR 26. Today's session is Agentic AI as the Oncologist: Clinical Decision Support and Human AI Collaboration. My name is Renato Umeton. I'm the Vice President of Data Sciences and Chief of AI for St. Jude Children's Research Hospital. And today we're going to explore this session with two giants, one from academia, one from industry. The reason why in picking the speakers, we decided to structure it this way is because personally, if you have known me, if you have worked with me, I am very opportunistic as of where the solution comes from, right? I'm married to the outcome. I'm married to seeing a patient outcome improvement, patient experience improving, clinician experience improving. Improving our healthcare systems, right? Optimizing cost. And AI is something that is now ubiquitous among all of us, our lives, and in healthcare is particularly impactful. Our first speaker, I'm very pleased to introduce Dr. Dorin Comaniciu, Senior Vice President and Chief Expert for Healthcare AI at Siemens Healthineers. His work in computational imaging and machine learning has translated into a broad portfolio of clinical products across diagnostic imaging, image-guided therapy, precision medicine. He's an elected member of the National Academy of Engineering, National Academy of Medicine, the Romanian Academy, and a fellow of IEEE, ACM, and MICCAI Society. He holds the IEEE Longuet-Higgins Prize for fundamental contribution to computer vision and a honorary doctor from Friedrich-Alexander University. With over 550 granted patents, 350 peer reviews manuscript, and bringing the perspective of a leading biomedical company that many of us trust already in our hospitals, especially within radiology and beyond, I look forward to what he'll share with us. And thank you very much for joining us, Dr. Comaniciu. The floor is yours. Thank you so much for the kind introduction, and especially thank you so much for coming here to listen to us in this beautiful San Diego morning. So, today I will speak about agentic AI in radiation oncology from imaging to therapy. I'm coming obviously from Siemens Healthineers, and all the disclosures are that I'm employed by Siemens Healthineers. So a couple of words about our company. We have in the world something like 300-- I'm sorry, 700,000 equipment in the world. So covering everything from imaging, CT, MR, X-ray, nuclear medicine, and including also laboratory diagnostics and also moving to therapy like interventional oncology, but also radiation therapy. A couple of years ago, we acquired Varian, and today we're number one in radiation oncology. So we see a lot of our focus is on non-communicable diseases such as neurodegenerative, cardiovascular, stroke, and the focus of today, cancer. Now, these devices that we have in the world have something like 300,000 patient touch points per hour. So think about during my talk, about 20 minutes, more than 100,000 patient touch points. This gives us a lot of purpose in what we do. And each time, say, colleagues are joining us in the R&D, it's very inspiring that actually after so much work in the school, PhDs and so on and so forth, ultimately you have an opportunity to give something back, right? So now, the base of what we do is very much related to healthcare AI, and I'm a strong believer that healthcare AI is perhaps one of the few technologies that has the potential to make healthcare more efficient, more effective, but also to expand care to many people in need. So you can see here a simple picture of healthcare AI, the way we see it from our portfolio. So our focus is to make imaging faster and better, what we call patient to image, and to help interpretation of images in what we call image to report and plan. And then how to use that information in order to plan and deliver the therapy. Now, in the space of healthcare AI, we've been working for quite a bit of time. We have today more than 100 AI-powered medical devices approved in the US. We're also innovation leader in this space in AI for MedTech with quite a number of patents and intellectual property. I'm responsible for the healthcare AI centers of Siemens Healthineers. So you see here we have-- I'm coming from Princeton, New Jersey, and we have a presence in Germany, Eastern Europe, India, China, and we're currently building in Canada, in Alberta.Now, what differentiates us is what we call a global healthcare AI factory. For more than actually 10 years, we're building, improving the factory. We have supercomputers. You can see here we have more than one exaflop computational power. We run more than 1,800 deep learning trainings per day, and everything is sitting on a very large database. And the algorithms that are coming from there are maturing day by day and being translated into our businesses and following regulatory approvals. I want to show you one example. So, actually, four years ago, we published a first paper showing a foundation model from 100 million medical images. What you see in front of you, this is learning from half a billion medical images. And pretty much you need a lot of compute power, a lot of data to build these models. Now, if you really think how much time would be to look at these images 10 seconds at a time, you need something like 150 years. So these models constitute the base for what we do in terms of developing the products. And today, for example, we're training with the next generation with more than one billion. So now, why are we all here? We're here to enhance and improve the patient experience. And especially my talk today will be focused on efficiency. We all know that every 40 weeks of delay in cancer treatment can increase mortality by 10%. And the reality is that everything in oncology needs more efficiency, needs more effectiveness, and again, as I said, to expand care. Now, I would like to go back and think about how we can formalize everything that's happening today in treatment, in therapy. Ultimately, you see everything is about controlling the evolution of a physical biological system, right? So now with this, I'm going back to some old theories, actually, starting with a control theory in the 1940s, 1950s. Ultimately, to be able to manage disease, everything reduces to controlling the evolution of a biological system. Now, the challenge is that such systems are not observable. The state is partially observable due to the complexity. So, there is this POMDP, partially observable Markov decision process that is coming from many years ago, and in the sense that the agent cannot directly observe the underlying physical state. Yeah? So now, you see, in trying to manage the evolution of such system, there are a couple of well-defined steps in the sense that you have the perception, and then from the measurements that you get from the system, you have a belief model in building what we call patient state or the system state. What the clinician, what we believe is the current state of the system. And then you have a disease progression model, and based on that, there is a policy defined and what we call treatment planning, and then going to plan and control. Right? So now making more specific, when you put the patient in the center, then the patient, again, is only partially observable. The patient has a very high complexity at a molecular, cellular, organ, functional level, and everything is so dynamic, right? So now, again, we have the topic of sensing, we have the patient state estimation, the disease progression model, the therapy planning and therapy. And I believe it's always very important to keep everything in a context. We're dealing ultimately with five learnable models in the terms of how do I do sensing, how do I do the estimation of the state, what is the disease progression model, what is the therapy planning, and how the therapy is delivered. So now let's take it one by one. Right? So the first topic is that how do we do sensing? As I said, right here we provide a lot of imaging, laboratory information. And actually, I want to show you one example on how we improve sensing. And that example is coming from MR, where two years and a half ago we introduced DeepResolve. DeepResolve is about a technology that makes MR much faster by reconstructing from the k-space into the image with the help of AI. Yeah? So now, an MR is not 10 or 20%, but can be, you see, 73% faster. So we have many organizations that are improving their flow in MR by increasing the capability with DeepResolve to scan more patients more efficiently. And you see in the world up to now, in two and a half years, we have more than 44 million scans done with this technology with AI-based reconstruction. Now, the next step is how do you interpret images? And here the next generation comes, is not only looking at specific, say, radiological findings, but looking into how do we generate a full report. And there are new technologies emerging, large databases in which the AI technology is capable to generate full report and also us being able to integrate it in the radiologist workflow.The next topic in the space of radiation oncology is once the decision is made about radiation therapy, is how do you generate a plan? And the first topic is to do the simulation, and then very important here is to have the auto-contouring. Auto-contouring means that you need to focus on the organs at risk, you need to focus on the lesion, and you see here this is a regulatory approved system that we have. It covers more than 200 organs at risk, plus clinical target volumes, and includes lymph nodes, stations in pelvic, head and neck, and breast cancer. Yeah. And we have recently released also the contouring for brain metastasis from MR, together with multiple other organs at risk that correspond to the brain. And this is also fascinating because such technology would allow a better and better treatment of metastasis. And the next step is radiation therapy planning. Now, radiation therapy planning is today, if we look at the head and neck scenario, you see it's a quite involved process coming after the simulation and going into creating the prescription, initializing the plan optimization, and then going to physicians review, the refinement, and then final evaluation. So this actually can take quite a bit of time. So the question is that where AI can change everything. So now we have technology in the development where actually the entire optimization, starting from the simulation to the deliverable plan, goes in under a second on a standard hardware. So this is a really fascinating evolvement, some order of magnitude faster of what's happening. So if you really think today what would be the potential evolution of radiation therapy planning is that the interactive plan optimization powered by AI end to end or what we call AI RT, in which you press the button and you see the plan, and if you want to do any iteration, you do an iteration, and that would be a joint session of physician and planner. And after that, of course, there are many, many other topics where AI is relevant, relating into how do you deliver therapy, such as the motion estimation, multiple elements here, looking at the brachy, looking at the external beam, interventional oncology. And once the therapy is being delivered, the question is how do you handle the outcome management? And ultimately, it's an entire loop that goes again and again, looking into, again, sensing how the therapy has made and what needs to be done, what is the patient's data again and again. So therefore, I believe it's very important to understand that this loop is not one shot, but it's happening again and again. The patient state is constantly changing. And for that, also our digital oncology system, Oria Core, we try to have everything that helps the information integration. But again, we try to introduce a lot of efficiencies in processing this entire information. As you can see in my very short lecture here, I address a substantial complexity of the topic that is happening only in radiation oncology. But now another element is that how do we make everything more efficient? And for that, let me talk just a couple of words about the operational twin for department of hospitals. So now, this relates to the capability of agents to achieve user-specified KPIs, key performance indicators, under dynamic conditions. What I mean dynamic conditions for a department can be is that patient delays, workforce unavailability, machine down, cybersecurity, whatever you can imagine that can change the typical workflow. So now, the new technology, again, think everything in terms of having the capability of a supercomputer to train. So the new technology is that you're capturing the data from the department. Based on that data, you do exascale simulations, billions of simulations. And with that data, use that data to train the agents that learn how to deal with various exceptions, with various variabilities, with various dynamic conditions while focusing to deliver the quality and the time. So we have now agents that are having the runtime and also making strategic changes on how everything should run. So now, one specific topic is that we integrate such technology, advanced scheduling technology into Oria. So where AI evaluates patient needs, machine availability, and scheduling constraints, and then generate optimized appointment and recommendations in one click. So you can see here the AI scheduling process. You have the patient demand, machine capacity, operational constraints, clinical priority, machine availability, and computer scheduling rules, and then pretty much the AI is delivering recommendations, and you have quite a number of benefits.But now think about that. Again, having such scheduling, and this is only one example, involves quite a bit of complexity coming from the environment. Each environment is very different. So now when we think of in terms of the autonomous agents, is the capability to capture the environment complexity with the overall large-scale optimization framework. And this is where the autonomous agents are coming into place, and the story is, again, providing the proper synergy between gen AI, the capability of understanding very high complexity with the capability of high-end simulations and high-end optimization. So now think about where things are going in the future. We see the roadmap into scaling the operational twin, not only at the department but for the entire health system. So now you see at the base, you see the topic of staff and patient and machine scheduling for various departments. But then when you go to the high level, you're talking about service coordination and balancing. Meaning, what is the journey of this patient, and how do I make sure that the journey of each and every patient is optimized across various departments? Going even higher level is about capacity management, and ultimately, multiple CEOs are interested in what are my operational strategies. Can I model at a larger scale my operation strategies? So now let's go back and draw some conclusions from here. So from a technology standpoint, there are three very important elements. You have the gen AI that has a very important promise, and already there are so many examples out there in which the knowledge of the world comes back to our fingertips. So that's a very important thing. But certainly there are two more elements here. One is large-scale optimization. How do I develop agents that are benefiting from the knowledge from the world, but also they are optimizing in a dynamic fashion certain tasks, like scheduling, patient journey, and so forth. And then the other part, which is equally important, what is the role of simulations? So I had there the disease evolution model. And today a lot of disease evolution model is mostly part of the guidelines. But still now, think about how do I model the tumor complexity? How do I model the tumor environment? And so on and so forth. So we have quite a bit of simulations out there. Now, those simulations were not scalable up to now, and the synergy between the large-scale simulations and everything that's coming from gen AI, this will be yet another paradigm that is evolving. So think about these three triangle that would drive us forward. This triangle that would drive us forward. Number one is everything that's happening in gen AI. Number two, large-scale simulations, and number three is that large-scale optimizations. And with that, thank you so much for listening. Thank you. Thank you so much for the illuminating presentation, Dr. Doran. I'll open it now to the floor for questions. We do have time for a few questions. There are mics in the room. Don't be shy. Hi, Saren Rothberg, stage four cancer survivor, 33 years. And I'm just curious, are you using patients and patient advocates to develop these systems? I'm a AI. I wrote the first AI study in cancer survivorship in 2019. Just curious. Yes. The short answer is yes. Especially related to activities in radiation oncology. We have quite a number of contacts and advising, and actually you see among our colleagues, we have patient advocates. I would say that it's very important because let me try to reiterate. That's the reason we're here, to make a difference for the patient. It's also true that we're trying to bring more, as I said, efficiency, effectiveness, but also the capability to address more patients to cancer. All statistics are showing that cancer will only grow. So having patients very close is very important. Yeah. It was really impressive and very clear. Thank you. Thank you. Good morning. Steve Schneibel, Cancer Research Catalyst. Thank you for the excellent talk. Very interesting. Obviously, there was a lot of work being done with AI. That's something of an understatement. And as these tools are emerging and ultimately being adapted between healthcare systems, there is going to be a certain amount of competition. So, while recognizing that you are here on behalf of Siemens, I'm just curious what your perspective might be for how healthcare systems can think about which tools being developed right now are best, and how different tools can talk to each other, and to make sure that as different AI modalities emerge for healthcare, they're not being siloed off from each other. Thank you. Yeah. Thank you for the question. So no, I can give you a bit of a perspective. I'm in this field for more than 20 years, 25 years, something like that. So there is a clear evolution. And I'm coming from a generation that was doing machine learning 25 years ago, and when we tried to do it in healthcare, people were asking question, "Why the heck are we spending time with that?" So that's my humble beginning in this field. But now it's the opposite. We see a lot of promise in AI. And it's always useful to see the evolution from tools to systems. I like to use the word systems because everything that's happening today in AI is beyond a tool. And if you read, I just finished reading this, "A Giant Leap" where the author is speaking about the future of AI systems, where actually we have to treat such systems as really powerful systems that would provide scientific advice, that would provide guidance, and so on and so forth. And we should start thinking of having such systems evaluated with the same rigorousness that we evaluate today clinicians. Meaning that to have a clear curriculum to test these systems, to test everything, to reassess their performance, and so on and so forth. So again, it's a very, very, very interesting evolution. And again, coming back to this triangle, the capability of gen AI to help us handling complexity, plus the capability of us doing a lot of very powerful simulations. So that means that you can scale simulations, and you can scale also the optimization of the systems. To me, the future is fascinating. But to me, again, for a patient, it's very simple. What is the contribution for me? What is the contribution for me? What exactly do we need to do as a community to make the patient's life better? Thank you very much. I have another question. We have another few minutes. Your group worked on AI RT, and that impressive software generates a plan in under a second, let's say, for prostate. I was wondering, this compresses, importantly, the turnaround time. It could also compress the experience for the physician, right? And so in the context of AI in the loop, I feel there is a good question there to be asked. How do we make sure that these new systems, this agentic workflow speed up without eroding useful, meaningful human supervision? And what does meaningful look in practice to you? Yeah. So thank you for the question. I would see this topic a bit from the other side, in the sense that the capabilities of evolving AI to get answers, to get solutions much faster than before, it would eliminate, I would say, what we call a friction that we see today in everything in the IT systems of taking time to get answers, taking time to get optimization, taking time to not do everything in real time. Because at the moment where everything is run in real time, that you can imagine a lot of workflows being changed and being much, much more efficient than today. Today, that's the story, right? If the system doesn't give me an answer immediately, then due to the time pressure, I need to switch my work to a different problem and come back today, tomorrow, and so on and so forth. So this is very, very, very interesting paradigm, eliminating friction from delays in the answer of the system. Yeah. Thank you very much. There's one more question. Just quickly. So thank you for the talk. It's incredibly fascinating. I would look at this and think then, how do I apply this to an academic hospital? Have you done test cases or taken these systems and applied them in hospitals? And I have a second part of the question which is, AI is really rapidly evolving. You're watching it change daily. And rapid evolution and clinical systems are not normally words that come together in the same sentence. How do you see the rapid evolution to be incorporated into clinical processes that are normally very controlled and very slow to change? Yeah. Thank you for the question. So almost everything that you have seen here, we run with hospitals. We run with academic hospitals, and very much depends on the nature of the problem. We run the high end with academic, but also we work quite a bit with underserved hospitals or hospitals serving various underserved communities. So it's a range of problem. And to me, the second topic is equally relevant in the current environment. I guess that we all experience, number one, is that we are overloaded by documentsCreated by gen AI, by messages and so on, so forth. But also you see a lot of software nowadays, and while there is so much software out there, still the pipeline to get topics regulatory approved is still very slow. So we hope that a lot of transformation will happen in the translation aspect. While of course, keeping the focus on the safety and effectiveness for the patient. Excellent. Thank you again, Dr. Komoniyi. Our next speaker is Dr. Taehyong Hwang, Founding Director from the Center of Molecular Artificial Intelligence at Vanderbilt University Medical Center. The first center of its kind in a clinical setting, dedicated to leveraging molecular AI to directly improve patient care. He also serves as Director of AI Research in Surgical Sciences, and now a chair and professor at VUMC. His work integrates AI with molecular biology and advances cancer research, transplant science, precision medicine, developing technology such as spatial 3D and 4D molecular profiling to uncover molecular mechanisms, identify novel biomarkers, and designing target therapeutics. Bringing the perspective of a research lab with such a deep knowledge is actually how we can turn the molecular AI into translatable tools. I really look forward to what he'll teach us today. Dr. Hwang, the floor is yours. Thank you. Thank you. Thank you very much. I have one question for everybody. Have you ever wondered how a drug works in human body? You might know the mechanisms, but we never watched that. So today, we'll watch together. So this is the real tumor. This is HER2-positive cell. What you are looking at is real tumor expressing HER2. The NR2 is in a pink color. That's the drug that are very successful at treating patient with HER2-positive cell. Now, what we can see is that when it administered the drug, after 24 hours, whether those drug actually delivered to the tumor, whether they bind to the target tumor cell, whether they are really killing it, and then whether they internalize how this payload release, and then do you see the response? That's something we never have seen before, but today, that's what we're going to watch together. So this is not a prediction. This is not a simulation. This is real tumor, real drug that we treat today for patient. So my name is Taehyong Hwang. I'm an induct director of the AI research at Vanderbilt. Today, what I want to talk about is that it's not about simulations, it's not about research, even it's not about development. The platform that we use daily, the platform that we want to use to treat our patient, our loved ones. So the talk is about the 4D oncologists and then how AI help us to understand and watch the drug works. So we understand that the world that we live is in 3D. Our cell is 3D. Our body is 3D. When drug enter to human body, now it becomes 4D because we have time domain. And then when drug penetrate, we always need to calculate how much dose, how long we need to give this drug, and then whether how long does it last. So every single momentum that matters, and then we need to think about 4D domain. But when you treat patient, when we use our drug, as long as we give the drug to the patient, we give the drug with the hope. With the hope that that drug can reach out to the tumor, and then it's going to bind, and then it's going to kill the tumor cell. But again, those decisions should have been made within 4D dimension, not the 2D. The today's standard of care, everything we simplify. We cut the tumor, we make a slide as 2D. We use the information to prescribe treatment, even though the way we should treat our patient is to 4D. So one of the example what we're going to show that is that as part of NCI-Cancer Moonshot Project, initiated by NCI with my colleague Lingha Wang. So we are heavily studying gastric cancer. So the consequences is simple. If you don't use 4D, this is what we see. Patients with gastric cancer, 800,000 deaths per year. The challenge is that once it's metastasized, their median survival is less than 11 months, and then majority of patient do not have targeted therapy. Then how are we going to change this? We trying to build the model, not just model, with the platform. Cancer is very complex. One single model with large parameters cannot solve everybody's tumor problem or cancer problem. So we want to use agentic AI, the agent with AI, with intelligence, bring them together, orchestrate them, and then integrate all this information together, make one most important decisionWhich drug, A drug, will work for my patient? So the first agent is developed with LG AI Research. The second one is how to watch it. The third agent is we want to see their behavior with tumor. And then the last one is integrators. The first, the agent is, again, developed by LG AI Research, is very critical. What we're trying to do here is not about research, like I said. It's not about development. We want to implement it as of today. We want to deliver it to the patient. All this research and development, there is one thing we need to do, which academic cannot do. Regulatory implementation, make it as the product so that our patient can use it. So not just for LG AI Research's amazing model, but again, all these logistics we want to build together. So also what we are doing is that we are not building it in a separate platform. We want to embed it in clinical setting, in standard care. That's what we do. So we don't interrupt any standard care mechanisms or asking any additional material. But once we have those standard care material, which is H&E as a diagnostic, that's how we start our journey. So this is LG Axon AI model. I don't think we have the voice, but let me introduce that. So once we have H&E from surgical paths or routine pathology lab, then agentic AI co-developed by LG AI, we look at all these NCCN guideline, ESMO guideline. Because what we care about is that if there are standard care, we can actually maximize our algorithm to provide better treatment. That's something we want to know. And then once we have certain markers, the markers even we don't have from the NCCN guideline or ESMO guideline, our agentic AI don't stop there. We want to look for the additional potential targets. Potential target to use from off-label potential target, we believe that AI can guide better treatment response. We also do not just simply looking at that. We also bring the large language model developed by LG AI Research. We're looking at all these regulatory process. We're looking at all potential clinical trial based on inclusion, exclusion criteria, so that that guide us that what we should prioritize even to staining, even though it's not the NCCN guideline or ESMO guideline. That give us the time that are most precious for our patient. So let's say that the agentic AI we develop with LG AI provide a list of things we should do, including the standard care. So for example, there is potentially HER2, then we're going to do that IHC staining because that's the standard care. But let's say that maybe there are FGFR2 which is emerging treatment modality. Still in the clinical trial, there are potentially a lot of the expression or expressing cancer cell we might be useful to try. Then our agentic AI can prioritize that. Even though this is not standard care, please generate IHC staining for FGFR2b so that if our drug doesn't work from standard care, maybe we can rapidly enroll this patient to the clinical trial with FGFR2 inhibitors or ADC. So now second one is that before we give the drug, we have the chance to look at that, which drug is really going to be effective for our patients. During the biopsy, we can do standard thing like IHC staining, other thing, but again, we can actually use that, culture them to make patient-derived organoid or cell line. Now even before the standard care like HER2 ADC, we can actually treat the drug with our patient tumor, your tumor, and then we can watch it whether ADC can really kill the tumor cell from our patient body. When you're looking at this, this is how ADC works. When you give the drug, it bind to surface, and then ADC internalize, and then lysosome comes in, and it's wrapping up, and it cleaves linker. And then payload release and the apoptosis happening. We can measure how much payload we need to give. We can measure that. Which ADCs might be the better ADC to provide to our patient. On the right side, what you are looking at is the case that the ADC Trop2 didn't work. The reason why Trop2 ADC didn't work is not because of ADC for Trop2 is bad. It's because it's patient tumor. Those patient tumor cell doesn't express lysosome, which means that we cannot cleave this linker, and then we cannot release payload. So we can tell you that based on all this experiment, which we can do less than 30 days, we can plan all these multiple drug, existing drug, FDA approved or investigate drug so that we can provide this information with the matrices, with the quantification to our oncologists. The next one is that how it's going to behave in the future. All the standard care what we are looking at is that the first thing is staging. Oh, this is early cancer stage one. I'm sorry, this is stage four because it already metastasized. But when you give the new adjuvant therapy or adjuvant treatment, do you know that how this tumor going to behave even though you treat the patient? Using very special device with microfluidic device with aligned extracellular matrix. Now using the tumor, we can actually let see that how this tumor going to behaveWhat you are seeing right now is the tumor from breast cancer. And then what you are seeing is that it has a lot of metastatic potential. What does that mean? Even though we treat patient with the standard care, we should be very cautious. We should very aggressively observe and monitor the patient tumor development, because what we are seeing is that this patient, this tumor, unfortunately, likely invade to distant organs. So that we can plan ahead. We don't wait to six months, and then we don't want to say, "I'm sorry." We want to say that "We have plan for you, and now we know how to behave, and then we going to do best thing for your tumor to prevent this is happening." And then what we do is that we accumulate all this information, millions of billions of zillions of information from all different parameters and the picture of the videos and then all this information. And then we're going to put them together. That's our orchestra. We're going to use more than millions of billions of the time point to build this model from, again, actual patient tumor. And then we're going to decide that should we give this drug, even though we see how it works in their patient tumor or organoid or cell lines, to really show that this will really benefit the patient. So that's something we build as called Terrain. This is work with Lingha Wang from MD Anderson. This is real tumor. This is stomach cancer. We resect two regions, and then we generate data, which is spatial multimodal data using singular genomics G4X platform. In each side, this is the two different regions, and because we believe there are two different tumor microenvironment. And we also generate the protein data so we can look at the vascular structures. Why? When you simulate or predict how drug works, drug actually give through IV, which means that it circulates through the blood. I mean, the vascular structures. So we need to understand the structure, how we simulate that. And then also we need to look at all the standard guideline. Based on their profile of what kind of tumor cell, what's the composition? Are there a lot of stromas? Are there a lot of exhausted T-cell? Are there a lot of other things we should care about? Now, within seven days, we can generate all these data right before we prescribe or decide their treatment. So now using all this information, again, the information we already learn from the patient who didn't respond to treatment. Using this information, we can accumulate a lot of the decision process. What you're looking at is that we can profile and prioritize all existing 56 FDA approved ADC drug, and then how it's going to work. And then now you can watch it, how this drug going to penetrate and deliver and kill the tumors. And then what's the potential toxicities? What's the potential resistance mechanisms you're going to develop? So once you can look at that, then again, all this information comes to agent, the agent who can actually interact with our oncologist and then our medical experts. That's something we never done before because cancer is biology. It's very complex. There are millions of zillion information. When you go to the real tumor board, we never talk about their tumor biology. We just talk about their staging, there is mutations, and that's it, even though we can collect a lot of information. So what you're looking at is that this is the real time we simulated, and now you can see that based on each cycle, when time goes by, how much drug can penetrate and kill that based on the data from real patient tumor, we about to provide the treatment. We have a lot of options, and we can always simulate that because, again, this is real-time decision. Once you have all this information, what you are looking at is that in this actually the figures and then different drug, different sequential treatment. Now we have some answers. Maybe it's not the perfect answers, but again, it's better than guessing. That's something we can provide. So once we have all this information, what we're going to do that is that, like I said, we're going to discuss that, the discussion with agent AI. So all this information already went through. So we did all this simulation in a backbone. And when you meet with tumor board, this is something what we want to do. We want to discuss what's the recommended based on NCCN guideline, what's what we simulated, and then now with the real patient and their oncologist going to decide that which one would be the best treatment. Now what you are seeing is our agent giving some of the opinion. The human expert says that this particular treatment, two of them agree. One says we want to be cautious. Another one says that maybe we should give a go. And then our agent also provide information, information with the molecular data, with billions of zillions of parameters. And then human decide that, because at the end of day, it's the decision we make with our patient and with the oncologist. So you already saw that what we did. And then like I said, we don't want to just say that our agent or AI can do everything because it's not true. But one thing I can tell you that is that we can try our best to simulate and predict and observeSo that when you have a lot of answers, or when you don't have any answers, that is something we can help with you. So next two slide, I'm going to show you that in the real case scenario. This is a patient with HER2 palliative, and then there are also ineligibility because this patient can get the immunotherapy. Now you wonder that, should I give to HER2 ADC or should I give it to anti-PD-1 immunotherapy? Usually, we give the anti-PD-1 first. But based on our agent AI with real data we generate, we found that there are a lot of stroma, which means that drug might not penetrate. Drug never reach the tumor. Then what are we going to do? That's what we are doing. From the patient tumor, we drive their cell line. We simulate that. Simulate means it's not simulation like the virtual. Use their cell line co-culture with their fibroblasts, and then we treat them real time. And then we can see that how this tumor behave or it respond to this drug that we are about to prescribe. If it works, that's the hope. That's better than the guessing. And that's something we want to utilize. What we are looking at is that patient with HER2 positive, HER2 negative, that's what we're mimicking. And we can show you the ADC with by standard impact. We can actually still have some of the benefit of the clinical decision. After that, again, what we do? We discuss with AI all this information with real human expert. And then we're going to say that actually, even though immunotherapy with good drug, maybe ADC can open the door for stroma barriers. And then anti-PD-L1 would be the better choice. That is something we can use AI we've never done before, using real experiment. Another case is much harder cases. This patient doesn't have any markers. HER2 negative. There is the lack of the immune cell expressing PD-1, PD-L1. There is no colon 8.2. There is no biomarker. Still, we can use her tumor to try to see that whether some of the drug might work. So that's what we are doing on the left side than right side, that we're showing the Trop2 ADC is very effective at killing her tumor cell. After this, we can also do surgical resections, and then based on behavior with tumor we saw from the scalp, from the data we see, we can prepare what's the next. What if this treatment doesn't work? That's the biggest fear of the patient. Then we can say that we already have plan. We already try all these drugs, so we know what's the next. So that's something we want to give to our patient, the hope. So this is what we showed today, and then I just want to close this. Until now, like I said, we gave the drug to the patient with the hope. The hope that this drug reach to your tumor. The drug will kill your tumor cell. And then we wait. Many of cases, it didn't work like what we hoped. But as of today, we can watch it together with our patient, and then we can decide together with our patient, our doctors, and our agent. So thank you very much. Outstanding presentation. Thank you so much. Please come to the mics for questions. Hi. Marcus Munari, B Pharma. The question I have is probably twofold. One is you started out with saying we don't consider the time dimension so far in treatment, but we do PK sampling, we do biomarker. So the first question is that component. Mm-hmm. So how do you see the big difference there? And the second one is how does it scale? Mm-hmm. What are the statistics behind it? So have you compared this system versus whatever's used now state-of-the-art? Mm-hmm. Do they fall apart? So good question. Actually, it comes to just one question. So there are drug screenings. You can measure the efficacy, toxicity, and PK study. Like animal model, cell line. What's the difference? The difference is that if you're just thinking of ADC as biologics, how ADC works, it bind, first of all. It should bind on target, off target, right? We need to measure that. And then we need to look at internalization. Internalization does not equal to drug works because we need to look at whether linker is cleavage. Because linker cleavage doesn't mean that it's going to work because what's the payload? So the payload release doesn't mean that it's going to work, because if it leakish, it's going to actually make a lot of toxicity. So all this information I just talked to you, how we actually get the result. You get the one plate paper saying that this is cell viability with one hour, two hour, 24 hour. What I can tell you that is that ADC, all this internalization, lysosome, it doesn't work like what we expected. When internalized, sometimes lysosome doesn't come in, which means that even though drug internalize, even the drug works, because of those tumor, it doesn't really actually express lysosome. You cannot cleavage the linker. Which means that it's not because this drug fail because your drug is not efficient. It's because just how this tumor behavior. So this is something we're trying to provide. Not just replacing what you do, PK study, but we want to show them that this is the information you can add. And then the most important question, when you give the drug through the mouse model or human, how do you know how much drug actually delivered? You never know. You cannot guess. Even you cannot get the information. But based on what we are showing here, we can see it. We can count it, how many drug, how many doses, how much of them, and in real time. So I didn't show that real time intro... focal meeting, but that's something I hope to show next year. But that's something we can actually provide. So I think there are a lot of discussion we can have, but we can have it after maybe this presentation. On the statistics bits, so did you- So statistics, same. So all this data we have, like AI, all this, we have locations we can show that based on 2D level and the static level, on the real-time level. So how many cell, how many drug, how many mouse we need to do? I can tell you that whatever you do, we can do equally, and then we can add more dimension, and we can also add more the statistics. But we can discuss that later. Thank you for the excellent presentation. I was just wondering, given the promise that this has to be used in the clinic, right now seemingly, off the shelf, it looks like the software is ready. Are you considering any kind of measurement of this platform's outcomes for clinicians who may use it and clinicians who are not using it, something like that to measure the difference that it might make? That's great question. So, I just want to make clear, it's not something we use right now to treat patients. So this is something we are developing together with our collaborators and industry partners like LG. So to do that, of course, we need to go through all this regulatory process. Maybe we need to do some clinical trial or some other thing. So yes, because again, patient safety is the best thing. So and that's something we differentiate with others. We don't interrupt standard care. We don't say it or claim it, you need to change the standard care. So like I said, even patients that standard care, we can prepare all the experiment and result, and then we know the measure of patient don't respond to treatment. Then what's the next? When you don't have any NCCN guideline. When you don't have any ESMO guideline, patient who fail first line, second line, third line. Because then the last step is that either you get the chemo, which you don't know about what we're going to get, or you go to actually, unfortunately, hospice setting. But what if there is some drug we can give you the answer, the answer based on your tumor? That's something we're trying to do. So that's the opportunity. Still we can give the hope for the patient who didn't really respond to standard care. Excellent. Thank you very much. There is another question. Go ahead. First thing is your slides are incredible. Did you make them? Yes, I made them. You use AI or you just design? Slide scale, yes. Cloud design. Did you? But every single data and then the video is our real experiment. There is no AI. Okay. That was- Just a style. But I guide that. Yeah. It's so fun. Okay, that was number one. Number two is, does the patient get to see- Yes ... this creativity of, and how the science is changing using AI, or is this just behind the curtain for the medical team? No. Everything involved with patient. That's why we make the visualization. Visualization that even non-human expert can understand. The reason is that this is real case scenario. So right now, patient use AI. So when patient come with the news of the cancer, they have pile of papers, and then they tell their doctors, "I need to get this A drug, and then B drug, and C drug." And the medical oncologist says that, "Oh, this is not guideline." "I don't care about guideline, I'm going to pay for it. Give me this drug and this drug because I read that." Because as you know, in clinic, it doesn't work like that way, right? Then sometimes we need to explain why, as the experts, as you're a doctor, why this wouldn't be best option. Mm-hmm. We can show is that, for example, you want to get the immunotherapy, and the doctor says that you are not eligible. You don't understand why. We can take your tumor from biopsy during standard care, right? And then we can show, quantify that how healthy your immune cell are. Maybe on post-treatment, your immune cell are not healthy anymore. Then there is also answer that we can actually boost the immune cell later, right? That's something we can also consider. But this process, showing your real tumor, and then you wonder that after treatment, they said that, "Oh, thank you very much. You finish your treatment. Please come back six months later." Then what we're going to think about? Every day, every morning, every time you're going to think about, "Am I getting the best treatment or this is really working?" That's the question, the fear we have. We want to eliminate it. When you come here, I will tell you, "Where is the drug?" I will show is that they are working really hard for you. That's the information we want to provide. If it doesn't work, we also want to inform you and our physician so that we can intervene earlier than later when it's too late. So are you licensing this out to other hospitals, or is it proprietary just to Vanderbilt? No. So we're going to make the commercial product. That's something we want to do because, again, without commercialization, we cannot use it in clinical setting, which means that we cannot use for patient. That's the path. So we designed this, again, it's not because of research. It's because I want to use it. Because I have my family members touched by cancer. When cancer becomes personal, you cannot wait. You need to do it. And this is something I do, and I want to use it for my family member, too. So thank you for being here. I really appreciate your time. Excellent. Thank you very much. Thank you. Fantastic presentations. Thank you so much. I thought, with such a deep and wide set of presentations, how could I contribute a little bit to this conversation? And so here I wanted to bring to you the perspective of the academic medical center. AI that works across research, clinical care, and education. So, before I start, I wanted to thank everybody at St. Jude who has made this possible. All our leaders, our patients, their families, the Office of Data Science, Dr. Madan Babu, who architected most of what you'll see, and our partners in IES, the CEO, the grad school, and legal. What I want to show you today is a story of acceleration. Why? Because St. Jude's mission is to advance cures and means of prevention for pediatric catastrophic diseases through research and treatment. If you can ask one question to the next one who talks to you about AI and clinical AI, let it be this one. Ask them to measure time to patient. There is a deafening level of hype around AI. Ask them the point question. Time to patient. I'm going to show you now a few examples throughout what we do at St. Jude along these five dimensions and how we are getting better at time to patient, how we are reaching more patients, and what are the AI systems that, for instance, we are exploring and with whom. The very first thing that you'll realize you need when you want to traverse that path is you need some sort of compass. You need something that sets the North Star. And that, for us, is really the product that was put through the vision for the data science initiative created by our CEO and our scientific director. They wanted for St. Jude to become a global leader in applying data sciences to biological discovery. What does it mean to do that? Well, after much consultation and introspection and looking at what everybody else was doing and where the industry and the academia was going, they realized that for St. Jude to deliver on that mission, they had to create a seamless ecosystem where experimental scientists, clinical scientists, data scientists would work together toward generating new knowledge that advances our understanding of healthy and disease states for preventing and curing pediatric catastrophic diseases. That is why they put together the Office of Data Science. The Office of Data Science has in scope, really, three types of topics. Pan-institutional initiatives. Think about data governance across the organization, enterprise-wide. How do you tackle that? I promise you need an office for that. Departmental level initiatives. What is one exploration within one department? How can we supercharge it, bringing data scientists there? And then forward-looking initiatives. Quantum computing. We know it's part of the research reality 20 years from today, but is it also 10? Five? What about two? What should we do today to bring closer that future to us, and that means to our patients? Center for Protein Design, Frontier Intelligence Labs, and edge computing. For that, St. Jude has allocated about $200 million over the next five years. We're going to hire 115 new positions, and that includes also funding for enhanced infrastructure. So I'll start from the grad school and give you some examples there and walk our way through global impact. What are we doing in the grad school? One of the very first things I'm excited to try next year with the next cohort is the concept that Cogniti brings, which is an AI clone of the teacher. That is an AI created by the teacher, created by the scientist, who can then be available to the researcher, to the students 24/7. The second one is a Socratic bot, something that doesn't give you the answer but teaches you to ask better questions. The second one that I'll show very briefly is some work that we're doing in foundational research, and for that, I wanted to share that we gave our workforce access through Cloud Research, Cloud Code, BioMCP, BioOmni, Cadence, you name it. If you've seen something flashy at AACR or beyond, I promise you, we are piloting it within the right guardrails with data that is meaningful and de-identified and all of that. What does that mean? It means that, for instance, to compile a report about answering the question, "Catch me up on cell-free DNA landscape across academia and industry," it takes now 20 minutes. It used to take two weeks just for such a scoping review. And now in 20 minutes, we got something like that, that it's tailored to me, Renato, because it knows my background, knows my CV and all of that. Another example I wanted to bring from this aspect is what we're building with Google and DeepMind. We introduced Co-Scientist at Kids just in October of last year, and what we observed is fascinating. Our research teams went from, "How do I make sense of this data?" That they had been looking into for about two years, all the way to, we have now the CRISPR screening to validate the mechanic C hypothesis. Used to take two years.Thanks to Coscientist and the close collaboration with Google, we looked at that in months. We are now validating in the lab new hypotheses that were generated thanks to these extremely powerful AI, which is really a systems of AI plus ELO ranking, and please read the paper. It's fascinating. It is something that we are currently exploring. What you'll realize pretty quickly after you start implementing some of these very powerful systems is that you need to train your entire workforce, and that doesn't scale. This AI is extremely powerful. Just explaining Coscientist to a single scientist, it is a lot of work. And so for that, we're building a community of practice. That is something that spans the entire organization, from students to researchers, department chairs, global, everybody. That is something that, for instance, we are seeing materialize now in 160 people that show up to our community of practice events that has multiple special interest groups. I'll mention very briefly that the one called AI Coding Agents has already 90 internal participants. And these people are hosting meetings. What is your frustration? How do you this? How do you install that? It is incredibly powerful to harness that internal community. We're also organizing biohackathons for them. 160 participants showed up last year, 23 internal departments, 55 total organization working on 22 projects that were identified by St. Jude. The next thing you realize is that you cannot just build these bottom up. You need also to inform all the scientific and data leaders. And for that, the Office of Data Science is working with every single department chair and data leader in the organization. We are bringing them together, whether they are from basic science or clinical sciences, all the way through. Just in February, we hosted the Data Intensive Projects Retreat, the first of a series, and we have 79 participants across EVPs, SVPs, department chairs, data delegates from 22 departments. We are creating this community. The next thing you realize, and I'm so glad that it was shown in the previous presenter's slides, is that you cannot do this alone. Especially after ChatGPT, it is evident that you need access to a level of compute, a level of data that sometimes you don't have access to internally. But which companies and where is this expertise? Well, if you look at the training compute that is required for specific advancements, and I'll point here at AlphaFold 2, because this was the 2024 Nobel Prize winner. Very few hospitals will have access to this computing power, and this is log scale. That means that you're going to need thousands of years to reach, let's say, the computing power that was used to train a GPT-5 level model. So how do we tap that? Partnerships. Networks on networks. ML Commons is an international nonprofit that has the mission of benchmarking AI and ML across industry, and that allows us to access the organizations that you see and the hyperscalers that you see on the screen. The second one that we're tapping is called the Trillion Parameter Consortium. Over there, we have multiple universities, multiple national labs in the US, amazing industry partners. And if you look at the research institutes now, I'm so proud that we have also St. Jude. What does it mean for our researchers? It means when you look at the top 500, which is the list where you have all the most powerful computers on Earth, nine out of the top 10 computers participate in TPC. And from my perspective, when we have to train over 115 data scientists, even having access to this network as a workforce development opportunity, it's really a game changer. Next one, I want to show you some examples from what we're doing in clinical. What we're bringing there is, of course, through a clinical AI and analytics council. We're looking at specific use cases, evaluating one by one and asking the tough questions. Are we really improving patient outcomes? Are we improving their experience? Are we improving the provider experience? Are we making improvements also in healthcare access? Are we optimizing costs and all that? The other thing I'm excited about and that we'll start doing next is doing something similar to what the FDA does. They bucket an intervention into a risk category, and that means we will be able to organize certain AI types, clinical AI, into given boxes for which we know the risk, for which we don't know the benefit, and we want to evaluate it and assess it. But at least bucketing the risk, this is important because this allows us to bring more innovation faster to our patients. Lastly, I want to mention what we're doing with Global. We heard about the guidelines and how they are transformative. Over there, I'll share that we are building the AREA Guide, which is an adaptive guideline for all the 450-member organization that St. Jude Global works with. And in that context, we're looking into developing an AI agent that would gather the evidence, present that evidence to the clinical committees, and then tailor that down, tailor that up, depending on what's available in that specific hospital. SoIf you look at the partners that are working with us, this is 450 of them across the globe. Here now we are measuring not just time to patient, but it's time to patient in Kenya, time to patient in South America, in that hospital. And so this is really the story that I wanted to convey to you and a summary of how we are really measuring time to patient. And if I can leave you with one message, let it be this. Next time you see an AI presentation, ask the question, time to patient. That means, how long did it take from, "Hey, I published this molecule in 'Nature'" to, "Hey, I injected the first child." Right? "I've run the first clinical trial, and I've run the first clinical trial in a phase III, and I've reached a underserved population." With that, thank you, everybody, and I'll be happy to take any questions. Hi. I'm Anne Marie. I used to work at Dana-Farber with Nancy Lynn, Melissa, and Greg. So I was just curious, so how do you decide which initiatives should involve your team in AI at St. Jude and maybe even previous work too? So that's an excellent question, and it's a prioritization one, right? We have only a limited number of experts, although we are training our workforce. But AI expertise is still niche. So we're doing two things. One, we are training the entire workforce, at least on the AI basics, right? And next, we'll look at the data basics and even answering the question inside any organization, who do I work with to explore question X? I promise it's going to be a fishing expedition in any academic medical center. So we're working towards building better navigation from that, and that's why the Office of Data Science is important. The second one, what to pick? This is really what we try to answer with these five questions that are ranked in this order. Right? If you are able to show patient outcome improvement, your project will be prioritized. Similarly, if you are able to show that you're actually improving healthcare access, it will be prioritized. Let's say you have a new molecule, something that it's easier to export, doesn't need refrigeration and all of that. So this is a ranked order that we are developing. But what we have front of mind is always impact to our patient. That is the number one question that we ask. Thank you. Sure. So I also have question. So this is amazing. So one of the key challenges as institutional decision maker, right? So every single model upgrade, sometimes within weeks, within a day, right? And then as they're implementing this version, sometimes we are not allowed to use the better version, even if you know that it is much, much better, like 10 times. Yeah. So how St. Jude's actually place the policy? Again, potential uncertainties were like... I don't know how I can actually describe. But as big organization, as the first, one of the largest institution in implementing all this data science, how are you guys going to actually plan and deal with that? That's an excellent question. We build, we test, we validate on GPT 5.4. Tuesday, 5.5 is released, right? How do we adapt for that? That's a fantastic question. We are very fortunate to work also with regulatory colleagues. They pointed us to what the FDA does, specifically for AI as a medical device, and they suggest using PCCP, which is basically a plan where you tell the FDA what you're going to take as a reference cohort, what you are going to take as a benchmark, and then they ask you to write down a plan of what is acceptable within what parameters. Now, if GPT 5.5 or 5.7, you run the benchmark, you analyze it against the data, and you are within parameters, you're in the clear. You don't have to go back to the FDA because you have not changed the risk profile. You might have improved, actually, on the efficacy profile. And so this is the stance that we're trying to take internally, and that's why, for instance, that risk grouping that I showed in the clinical area makes a lot of sense because it allows us speed without really slowing down the progress. It allows us to catch the new risk. Maybe one comment from me. So great observation on the PCCP. Actually, we have quite a number of products now approved under the PCCP, and that allows us to evolve the products to the next feature or to the next capability. But again, according to the constraints provided by the plan. Yep. So this is what we have in mind, but if anyone has better plans, come tell us. The FDA, we were at the meeting with them. They're also exploring new ways because they know how fast AI is developing. Go ahead. So I just did an article with Forbes, and I was pro-AI in healthcare for women, and the article was negative. Like very suspicious, very negative, very antiquated. And so I was trying to figure out how do we get people off that mentality? And I was wondering ifMaybe we put together a patient bill of rights of sorts. So because you are on the other side, supposedly, what would you suggest goes in that? Because it shouldn't be limited what you're doing, but maybe if I can present that, we're looking at the protections of the patient, what would each of you suggest should be in that, that wouldn't be so constrictive, but I can offer? That's an excellent question. Thank you so much, and thank you for the very important role that you bring. In general, one thing that I mention to everybody developing clinical AI is, if you don't have the patient at the table, or if you ask the patient what they think about the product when you're talking to the FDA, you're at least two years too late. So that is of key importance. The question that you ask is really something that we look at answering towards education. Education, because there is a lot of fear about AI. There was a lot of fear about computers. There were probably a lot of fears about cars when we were using horses, right? And education, having a license, having a risk that everybody can understand and can relate to is the key that allowed us these big leaps. And so that's why, for instance, we teach it in the grad school throughout. That's something that the new students will have at their availability, and they have to learn how to properly use so that those antiquated articles maybe are not written in the future. How can we stop it today? I would say maybe your organization can write a comment piece on that article. Patient voices are, as my colleague stated, they are the reason why we are here and why we're not in another organization or we're not in another line of business. So from my perspective, I would say the patient voice should be heard, especially by those who claim that they're speaking on behalf of the patient and sometimes they're just projecting their own opinion or fear. It was definitely-- It was as bold as, AI is meant to trick and own your material, and you're not getting any bene-- I said, "Take me out of the article unless you write my perspective." Yeah. And so they wrote my perspective. But I'm really trying to figure out, what can I show a journalist or somebody that's really negative, say, "But here's what they're agreeing not to do." Mm-hmm. They're not selling our data. They're using security precautions. They're not putting in our things into open systems. This is closed system. I'm just trying to get something to offer- Yeah ... so that they quiet down and just really get excited about what we're all trying to do here using artificial intelligence. Yeah. If I were to guess what happened is that the journalist that wrote the piece wanted attention, wanted the piece to be sensationalistic, and wanted to contrast certain opinions, and they exaggerated the views that were presented. I think when you have clearly stated, as you said, that this is the risk box, right? That we're giving this new tool, right? Is not something that people want to argue with because they're already thinking about the benefit. So I would say education is something that you should always point and you could point to these people, especially in the context of fears. As it's recognized by everybody, we are not risk-free at all, all the time, right? What we have in our hospitals is not a risk-free business at all, right? Chemotherapy is a sort of poison, right? And so are all the interventions. So this idea that the risk has to be zero, even when you talk to a single doctor, it is not zero, right? Because as my colleagues also anticipated, you are going to have multiple opinions, right? And as a patient, you'll wrestle with the question, am I getting the right one? Am I getting the wrong one, right? And that is something that we are still figuring out as a society. But if I can give you one, it would be really education for the masses. I feel AI should be taught even in elementary school. We're teaching computers already. It's not that big of a delta to say, "Hey, this is generative AI. This is non-generative AI. This is the math that goes in." The math is already taught. It's just that we have to wrestle a little bit as a society with this new big experiment. Do you have a patient education program on AI? We do, and that is something that we are developing closely with the colleagues at SCEO. We have this program called Together, which is geared towards the patient. How do we use AI to bridge that gap? It's an open question, but I promise you that the patients will be at the table. Because, again, we're doing it for them, with them, not to them. Thank you. Yeah. Hello. Thanks for the excellent talk. My name is Anna. I'm working in data science as well. So first of all, it's really amazing how you bring data science to the clinics in practice. And I was wondering if you could tell us a bit more about your perspective, where AI currently makes the biggest differences for the patients.So is it diagnostic tools? Is it drug development? Or is it rather right now more processing information and make it accessible? So maybe you could give your perspective on that, how it is now and how you see it in five years, maybe. Thank you. That's an excellent question. I would say there is a review that we are compiling with 30-plus colleagues to answer exactly that question. If you leave me your email, I'll make sure to send you the PDF when it's out, or I can show you the draft we have, but we are answering exactly that question, right? At every step of the way, what is the impact that we're making at that level? To me, there is also another component. Right now, if you look at this picture, right? It feels a little bit like we're retrofitting. We are putting AI here, we're putting AI here, we're putting AI here, here, and here. Right? In the hope that it accelerates every step. The other part of the answer to me will be AI that will allow us to go all the way through faster in a safer way. I haven't seen it yet, but this is the time compression. Is this still going to take us $2 billion and 10 years, 15 years to take a drug to the market? Or we can compress that entire thing end to end. That's the other thing that I look forward to seeing. I hope it answers your question. Yeah, thanks. That's an interesting perspective, and also looking forward to read that report. Yeah. Thanks. What's your time to pi-- I'm just kidding. You said to ask that. I really enjoyed seeing the org chart and seeing that your entire institution is all in on making sure that employees are upskilled. Obviously, the big focus here is improving patient outcomes and improving the speed of research. But there's also a tremendous amount of opportunity for these same systems to be integrated across large, necessarily bureaucratic institutions like St. Jude. You see thousands, if not many more, of patients, and there's a lot of paperwork that goes along with that. So, as part of this initiative, is St. Jude looking to implement AI on the administrative side as well? Absolutely, 100%. One thing that I did not show is that we have a mirror council in admins and in operations. Exactly as we have the clinical AI one for AI and analytics, we have one over there, and these are the colleagues that are leading that effort together with others, let's say, in finance, in HR, right? All the key administrative components. I would say the other thing that could be extremely interesting from what you asked is bringing a platform that means a common shared platform, let's say, across our 450 partners' hospitals. If you ask me what would be something really impactful, that would be part of the solution. Bring a software platform, and not just that, because that is going to be the easy part. Bring also the processes to operate that platform and allow you as a researcher, as a patient, as a clinician, to traverse this faster, right? Okay, I published this molecule. Who do I talk to to have the clinical trial? How do I design that, right? I'm a biologist. I know my biology. I don't know how to run a clinical trial. How do I find that, right? So the processes. And importantly, people. People will be the key part, and this is really the reason why I'm at St. Jude, the amazing people that I found there. I hope it answers. Thank you. All right. That's a wrap, everybody. Thank you so much. Thank you to our speakers. Thank you for being here.