Regina Dugan, PhD, President and CEO, Wellcome Leap
Margo Georgiadis, Co-founder and CEO, Montai Health
Najat Khan, PhD, Chief Data Science Officer and Global Head, Strategy & Operations, Pharma R&D, Johnson & Johnson
Moderator: Dr. Lloyd Minor, Carl and Elizabeth Naumann Dean, Stanford University School of Medicine

Brainstorm Health 2023 - How Ai Is Advancing Healthcare R&D Today 

Conferences

Regina Dugan, Wellcome Leap
Margo Georgiadis, Montai Health
Najat Khan, Johnson & Johnson
Moderator: Dr. Lloyd Minor, Stanford University School of Medicine

How Ai Is Advancing Healthcare R&D Today 

The health care system is ailing and some say artificial intelligence and machine learning could lead to the cure. Its applications fit many aspects of the health care space from the ground level of research and development all the way to patient care. The expectation and anticipation of what A I and M L can do are enormous with the belief that A I can help solve problems that are beyond the limits of human abilities. So let's talk about how the technology can be used at scale to help build better systems and healthier lives. Dr Lloyd Minor is Dean of the Stanford University School of Medicine, which is brainstorm Health's academic partner. Dean Minor will lead a conversation with Regina Dugan phd CEO of welcome leap, a nonprofit that works to accelerate breakthroughs for global human health through bold initiative, innovative programs, Margot, Georgia co-founder and CEO of Monte Health, which pioneers leading edge technologies to preempt and treat chronic disease. And Neat Khan PhD, Chief Data Science Officer and S V P and Global Head of Strategy and Operations for R and D at Jansson Johnson and Johnson. Please welcome them all. Well, good afternoon, everyone. I'm delighted to be with three outstanding innovators and leaders. And we plan to include you our audience in this discussion as well. But maybe we can get started Regina with you. And could you describe Welcome Leap, your organization and how you're using A I today and the work that you and your colleagues are doing and we'll go down the line with, with everyone for that question. Happy to Lloyd. So, welcome Leap is an organization that is designed to increase the speed and number of breakthroughs in human health. And we are using A I and machine learning across a number of our programs and they're very diverse. The applications of now. It's my sense that one of the reasons the life sciences have been a little late to A I and machine learning is two fold, one some initial skepticism, but also lack of data. And if we see now the explosion of data in life sciences, that's part of what creates the opportunity for A I and machine learning. And that's across computer vision for teaching and training systems in surgery, for learning how young Children develop language and so on. Fantastic. Thank you Margo. So just you know, building on that theme of the explosion of data, Monti health is really scaffolding off that to create um breakthrough therapies, small molecule therapies to help us stem the chronic disease burden. And the reason I got into this is two reasons, right? We've talked about it almost the last two days, right, two billion people around the globe suffering with these diseases. About half of them are inflammatory based and most of them don't have treatments that enable early intervention or long term use. And the R and D engines really aren't keeping up. Right, 30% of the therapeutics that go into the clinic fail because of safety problems alone. And that's $15 billion a year going down the drain. In addition, if we look at a lot of the therapies that are available, we have some amazing things like biologics, but most of those are only available at later stages of disease. And people only about 5% of the people who have these diagnosed conditions can get access to those because of the safety issues as well as cost. So we really took on this challenge of how could we use A I to reimagine, how we could create safe small molecule therapeutics, do that efficiently, rapidly and repeatedly so that we could get solutions to as many patients as possible. And it's actually really personal for me. I have a son who's spent the last 20 years suffering with several really debilitating chronic diseases. And when you watch what happens in someone's life, when they wake up every day in pain or they wake up every day with severe health issues, it motivates you to say there's got to be a better way to tackle some of these things. And I'm excited about the possibilities in A I to enable us to decode the language of chemistry, which is what we're doing and then building it scale for the first time, the largest understanding of what we call answer molecules. These are human qualified molecules, these are molecules that have coexisted with humans for hundreds of thousands of years in our food herb supplements other places. And we're actually understanding and using this ai capability to survey that landscape as never before essentially industrialized scale. So that we can identify the best possible keys that exist within the products of nature that are human qualified, that we can actually use as potent therapeutics. And we've demonstrated we can do this in a wide range of application areas. And we're excited to be advancing therapeutics to the clinic. You know, I might add Margot that in one of our programs that's focused on RNA. What we know is that for an RNA based vaccine like the COVID vaccines, there are approximately 10 to the 6/100 power different ways to encode the instruction set. We can't computationally solve that. So we're using A I and machine learning to address that problem. That's another example in the drug development side. Yeah, a great example, for example, in any molecule, there's roughly a million permutations of structure and function in a single model, every edge of that molecule, which really makes the drug, some of it would say this is bio and other people would say how soluble it is permeable, how toxic it might be. And so without compute capabilities, we can't disentangle that language. Just like chat GBT learns the English language. We have to learn the language of chemistry and then use compute to enable us to solve these really hard problems. Yeah. Now, first of all, great to be here. So how we're using data science A I machine learning, you know, the core of what we do is how do you actually develop medicines, right, that are effective and safe. And if I start from the inception of how you figure out what's going wrong when someone has a disease is what's driving that disease, right? That's the first order problem that you're trying to solve. And if you look at so many diseases today, let's take Alzheimer's Parkinson's, it's named after the physician that figured out the syndrome that has nothing to do with the biology or the disregulation that's driving the disease. So that's step one, right? And a lot of the data sets that are being developed multi omics, let's just call it genetics transcript omics so forth, helps you in leveraging machine learning. We're finding out like depression. Actually, there's several subtypes of depression, it's not one disease. Um You know, look at Parkinson's, there's actually various drivers of Parkinson's and more and more that we are learning. Why is that important? Because if we can know what's driving the disease, then you have a much better chance of treating the right patient with the right medicine at the right time. So that's problem. One, the second area that we're using it quite a bit is around precision medicine, right? So, um you know, one of the aspects, let's say in oncology, even if you know what driver mutation is causing the disease, you look in the real world, most people are not getting sequenced. So, yes, you know what's causing the disease, but you're not finding the patients, you got to find the patients. And we heard other panelists say that. So we are using machine learning and A I on data that's collected routinely histopathology images E C G echo to be able to diagnose diseases earlier. So for instance, it just came out today, we just got breakthrough device designation for an A I algorithm on echoes that can pick up a amyloidosis, which is a rare disease years in advance, right? Um The other example is for instance, different mutations like F G F R E G F R, you can actually predict the mutation from the biopsy image. That makes a huge difference because when the patient gets a biopsy, you actually know what mutation they have. So then you can direct the right medicine at the right time. And then the third area, I mean there's many areas. The third big area that we're leveraging machine learning I would say is in creating better end points. The thing about it, you have a medicine, you understand what's driving the disease, then you start a trial. But if your end points are not specific enough, if it's just a composite, you know, pen and paper, you're not going to be able to understand the signal to noise you really want to see is someone benefiting from it the way you want to. So, you know, Margo, you talked a little bit about inflammatory diseases. Let's take and Crohn's IBD, only 30% of the patients actually respond. What that tells you is that we don't know how to stratify the patients mechanistically as well as we should. So we've developed an algorithm basically where instead of scoring how severe the disease is the way it's done today, which is if I have one flare in my large intestine versus you have seven, we get the same score, right. So it's not localized, it's not sensitive enough. So we have end points now where you're looking frame by frame thousands of frames as the endoscopy video is done. And you can actually get a full representation of the severity at such a granular level. So then you can figure out who's feeling better versus not. And then the last thing is clinical trials, we talked about diversity. We talked about, you know, being patient centric, we're using machine learning to say, OK, where are the patients in a de identified way? So it's all with the right privacy, et cetera where are my eligible patients. And how do we open up sites to where they are? Two benefits. One you go to where the patients are, you can recruit faster and get that medicine to patients faster. But the second is in a lot of cases, people of color, people with like tougher socioeconomic status are in areas of more remote areas. So then you can actually leverage this to get more diverse patients. And we have many examples where we have done that um doing both machine learning and using decentralized trials. So put it all together machine learning is being used at every single aspect. It's not an idea anymore across R and D. And that's how you capture the entire value, not just doing point solution but compounded impact and effect by improving the probability of success that all this hard work that we do is actually going to be a good medicine for the patient, right? Maybe one other quick question and then we'll open it up to the audience. But of course A I covers a variety of different disciplines and approaches, probably the newcomer and maybe the most transformative of all are the large language models and generative A I based on transformers. And I'm wondering how you're thinking about how you're going to leverage uh generative A I and approaches derived from generative A I in the work that you're doing in your companies and in your endeavors. Regina. Do you? Yeah. Well, I mean, I think, look, I think we've hit a real inflection point with, with a generative A I, I think we've been waiting a long time for it and it's early days still, I think the applications are opening up particularly in the health sciences and the other sciences. From our perspective, it will be one of the strategies that we use, mixed with other technological advances to get the kind of breakthroughs that we need. And I think, you know, Najat touched on the various measures that we're using and how A I and machine learning are playing in those areas. But there are also opportunities to intersect across those different measures. And we're seeing those opportunities emerge quite effectively. Depression is a very good one. So still after 20 plus years and $22 billion of investment, we have one out of three patients responding, it's just not good enough. And what we're seeing in one of the programs we're running multi channel psych is the opportunity to use A I and machine learning across the various measures from genetics all the way up to the digital phenotyping to get to better predictions for the medicines so that more people can be treated effectively on the first try. And, and what I think is generative A I is going to be a very important part of each of the lanes and in crossing the lanes as well. Yeah, I agree with that completely. Um I think we've had this conversation over the last day, which is we get the most accuracy when we start constraining the problem. So we see it in enterprise, right? If you go and ask chat GBT, a generic question, they're going to be not right, a lot of the time. But then if you break it down and you go and you go after a chat bot opportunity that enables you to get the return rate data at a scale where you can actually be precise. So I think that's what's exciting is to try to find those pieces. And that's really what we've done at Monti is we've set. If the most important thing we need to do is rapidly and efficiently find efficacious and safe therapies, we can strain the space. So we used a cousin of those large language models which is really scaled graph, keep learning models to understand this language of chemistry and enable us to make it searchable and understandable. In the same way you can do the English language, we just do it the language of chemistry and all those structure function relation ships that make up a molecule that becomes a drug. But then we attach and constrain the chemistry set to say the chemistry set has to be human qualified because we could use infinite types of chemistry and understand that. But then we have no understanding until we get much, much further in the drug development process. Is this actually a molecule that can coexist in the human body. And so by constraining, that gives us much faster feedback loops and a much greater understanding of data. So for example, there's like 2700 approved therapies today, but we've already created Anth grass of 100,000 of these much more diverse molecules that have coexisted with us. And we know from that last panel that food is medicine, we've just never understood it at the molecular level. And so what A I enables us to do is decon that chemistry and human biology connection so that we can rapidly and repeatedly for any known biology come up with powerful therapies. And in our case, we focus on the most powerful as you talked about pathways that truly underlie these diseases. And by using these safer molecules that have this history of human consumption that enables us to efficiently land in severe stages of disease from a development perspective, but then have the highest probably of success to be able to back those molecules up into earlier stages of disease with alternative dosing levels or formulations so that we have this chance to intervene earlier. When we have those great signals that we have the onset of these diseases, we have to think differently because otherwise, when we talk to our farm, you're not willing to take on those large scale clinical trials in those earlier stages of disease because the safety failure rates have been so high for so many of these therapies. And so we've got to find a way because we all want health care, not sick care. And if A I can enable us to take the serendipity of where drug discovery started, right, finding any fungals and any virals. And we can now industrialize that and make serendipity predictability. That could be transformational as we think about the future of developing safe efficacious medicines. And so that's why I'm excited about it to answer the question. I'm like, you know, and I think that's a great point. You raised Margo in terms of how we're using large language models in generative AI I'll maybe structure it in like three ways. One is really around, you know, understanding the language of chemistry. And I would say biology, why is that important? I remember when I was in grad school, it would take and I was doing in modeling, by the way, it would still take months to make a molecule, right? And you say, OK, I think this is a protein. This is what I want to design. Now, you can actually use these generative A I algorithms to say this is what the protein probably looks like. And I'm going to say what the molecule should be and back into it. The difference is you improve the starting point and if you improve the starting point, your probably a success of what would become a medicine is much higher. However, it's not important, it's not sufficient to just make it a tighter binding. You have to make sure it's safe, you have to like make sure it's not immunogenic. There's so many things you can do all of that in parallel. So that's one and I think that has, it's not prime time yet, but there's a lot of promise there. The other end of the spectrum is think about when you're actually starting a program, you're going to all these papers and trying to figure out who's done what we have that version internally, we call it met dot A I ask, but essentially like a bio GPT for where you can quickly form hypotheses, right? That searches all of the omics data, the essays we've done internal and external super super important. And then the last thing in the middle and we talked about depression. How is it that a patient who has depression today? How do we figure out if they're better or not? Again, you answer a questionnaire. How subjective is that quite subjective? Instead what we're trying to do and actually deploy in our trials is use speech, acoustic, right? Use speech, how you're typing all patient consented? And then we use large language models to transcribe that into a sentiment uh ranking. So is your response something that is more indicative of depression? Why are we doing that? You make it much more objective? Number one, number two, you personalize it to every single person because everyone's type of depression and how you, it is not the same. So again, all of these helps you bring together precision medicine. You know, the last thing I'll say, no matter what approach or technology or whatever you have, the most important thing is to answer a question that matters to make of medicine. Everything else is just an enabler. That's great. Well, this has been a wonderful discussion and we could go on for a long, long time. I think you could see why, you know, truly, thank you for your leadership. Each of you is doing such remarkable work in leading amazing organizations. Um And we have to do this again soon. Thank you. I'm sorry. We didn't get to questions, but this has been fascinating. Thank you very much.