Name:
Front Row - AI In Drug Discovery Series II - Part 2
Description:
Front Row - AI In Drug Discovery Series II - Part 2
Thumbnail URL:
https://cadmoremediastorage.blob.core.windows.net/855f98b4-6639-46f9-837a-ac1987e53b8e/videoscrubberimages/Scrubber_5.jpg?sv=2019-02-02&sr=c&sig=uoXgzNsiZuUoDHKw2kYekUhgnO4yGqm8aPQlWQDg%2FSQ%3D&st=2024-10-16T00%3A32%3A24Z&se=2024-10-16T04%3A37%3A24Z&sp=r
Duration:
T00H25M46S
Embed URL:
https://stream.cadmore.media/player/855f98b4-6639-46f9-837a-ac1987e53b8e
Content URL:
https://cadmoreoriginalmedia.blob.core.windows.net/855f98b4-6639-46f9-837a-ac1987e53b8e/AI In Drug Discovery Series II - Part 2.mp4?sv=2019-02-02&sr=c&sig=sq3iN9tOkwCa4p44XU3CTC5k0ToI%2BFPfbrWPcnw2NWo%3D&st=2024-10-16T00%3A32%3A24Z&se=2024-10-16T02%3A37%3A24Z&sp=r
Upload Date:
2023-10-19T00:00:00.0000000
Transcript:
Language: EN.
Segment:0 .
[MUSIC PLAYING]
MALORY EBRANCA: Artificial intelligence is one of the hottest fields in biopharmaceutical drug discovery and development today. It's been hot for a while, and the fact is that there are now signs that it's coming to fruition. And they're very encouraging signs, things that are making people, investors, patients, drug developers very interested. And so we are pleased to introduce you to a number of leaders in this field who are going to tell us about how they are taking steps and strides to make artificial intelligence one of the leading technologies in biopharmaceutical drug discovery and development today.
ANDY BECK: It is a really exciting time in artificial intelligence for drug discovery as well as development. And the reason it's been a really incredible past decade of advances in the field, a few different fields, some core algorithmic advances in areas like deep learning, increasing availability of very large, labeled data sets, as well as increasing availability of almost arbitrarily large amounts of compute infrastructure to both build and deploy models to help in many different areas of trying to better understand which drugs will most benefit which subsets of patients.
ANDY BECK: And Path AI has very much been a part of this world where we aim to advance drug development, in Path AIs case, with the use of modern approaches in artificial intelligence and machine learning. And in particular, what's so transformative about this recently announced partnership between Path AI and the Cleveland Clinic is a few things. One is it really brings together deep expertise in both areas that we think are going to be critical for really impacting patient care with artificial intelligence.
ANDY BECK: And that is industry-leading technology, platform, medical applications, and products that Path AI has been developing in the application of artificial intelligence to pathology with a leading provider of health care and a leading institution of experts who are really deeply knowledgeable about their disease areas, their patients, the available therapies. And truly bringing together these two parts, we think is so critical for really transforming research and medicine with artificial intelligence.
ANDY BECK: And then, even just more specifically, the Cleveland Clinic has been collecting a very large and valuable data set of patients that have been treated at that institution over many years. And we hope to partner with them to create new products that can bring even more value to the patients around the world, and at the Cleveland Clinic are being treated by certain diseases.
ANDY BECK: As I mentioned, as one of the key drivers of why artificial intelligence has advanced so much over the past decade, it's because people have been making better and better use of the data that's collected on patients, both in clinical trials, but also in clinical practice.
MALORY EBRANCA: Speaking of the data, what data did you start with?
ANDY BECK: Yeah, so Path AI, we're now in our sixth year. So we've collected one of the largest data sets of pathology images matched to expert pathologist annotations of data. And over the years, our sources of data have been a few. One is we've been partnering very closely with many of the leading drug development companies in the world. And that's one source of pathology data. We've also partnered with labs and other hospitals and diagnostics providers where we've also learned from and deployed our systems on pathology data.
ANDY BECK: And we've partnered with about 400 or so board-certified pathologists who help curate and annotate these data to enable us to build models to improve the ability of physicians to accurately diagnose disease and quantify features on disease.
MALORY EBRANCA: So how much does Cleveland Clinic bring to you? And what kind of a difference does that make?
ANDY BECK: Yeah, it really makes a huge difference. This is really the future of how technologies like the ones we're developing are going to really meaningfully impact patient care. It's by partnering with the leading institutions in the world for building the best products and then being able to, in the future, deploy those, not only for research, but also for patient care. So for us, partnerships like this are absolutely critical to allowing us to really support the future of medicine in this way.
MALORY EBRANCA: But in terms of numbers, what number of samples will they bring to you?
ANDY BECK: Yeah, so in terms of the size of the data and the partnership, it's a large partnership. It will be over a million whole slide images included that we will work together on as part of this partnership.
MALORY EBRANCA: And what will you do together on these slides?
ANDY BECK: Yeah, so we're going to be building different models to help us better understand different diseases, build tools that can help physicians diagnose diseases across a range of different therapeutic areas. Pretty much, Path AI works on everything that involves histopathology. So that includes cancer, it includes areas like inflammatory bowel disease, and also liver diseases, like nonalcoholic steatohepatitis.
MALORY EBRANCA: So why is NASH such a hot topic these days?
ANDY BECK: Yeah, so NASH is a growing medical problem. So it's a really serious disease. It's impacting an increasing number of patients. So I think that's one reason. The other is there's not yet any FDA approved medicines to treat NASH. So many biotechs and biopharma companies are investing heavily in developing these new medicines to treat a chronic condition that's serious. It can progress to liver failure and the need for liver transplant.
ANDY BECK: So to address this market with new therapies. So there is a lot of work on this. And I believe, if it is already, it's quickly becoming the most common reason people need liver transplants, which is a relatively new phenomenon.
MALORY EBRANCA: Well, I looked up and there were over 1,000 trials--
ANDY BECK: Yeah.
MALORY EBRANCA: --for NASH. So it seems like it's a pretty interesting topic for pharmaceutical companies, correct?
ANDY BECK: Absolutely, yeah, for sure. It's an unmet medical need that's chronic, that impacts a large patient population. And there are not yet approved therapies. So absolutely.
MALORY EBRANCA: Do we know enough about it yet to know which direction to point the research?
ANDY BECK: Yeah, it's a great question. So I think the answer is no. We need to know a lot more. We may have approved medicines before knowing all we need to know, but it reminds me in some ways of the earlier days in oncology where it was treated as one disease, more or less, at least within an organ system. And I'm sure it's a lot more complex than we know today, but I think we've got to take the first steps to get the first therapies approved.
ANDY BECK: Then we can even better understand patient subsets and why certain therapies work very well in some patients, work more poorly in others, and to begin to have a more precision medicine approach. Some pharma companies are already doing this in terms of their patient selection strategies. But we're definitely at early days here. I think we're far enough along that there are therapies where there's a biological mechanism that is positive, at least.
ANDY BECK: And there's promising early data. And there's been tremendous advances in the therapeutic arsenal for NASH. And you mentioned the 1,000 plus clinical trials. What Path AI is very much focused on is how do we make the pathology really super accurate, reproducible, standardized so that these medicines, if they're effective, we're enrolling the right patients and accurately assessing therapeutic response on tissue samples so that the most effective therapies hit their endpoints, and the unaffected therapies miss their endpoints as they should in terms of reflecting reality.
ANDY BECK: So it's been a big focus of ours to try to have pathology match the advances in therapy because you really need both to have these trials run effectively and to get these therapies approved for patients.
MALORY EBRANCA: So this is a bit unfair to ask you, but what does Cleveland Clinic get out of this?
ANDY BECK: So they want they want to continue being at the forefront of medicine and research and science. So they get to partner with a leading company in AI-powered pathology. And it will be a deep collaboration, so we're excited to build new products with them. And we hope that this close collaboration leads to the implementation of these algorithms we're building, both into the research, but ultimately also into their clinical care.
ANDY BECK: And to speak to this deep-seated collaboration, they also are becoming an equity holder in Path AI, which even further aligns the organizations.
MALORY EBRANCA: Now, we spoke earlier, and you mentioned that you had a number of biopharma collaborations. How many do you have?
ANDY BECK: We're working with about 18 of the top 20 biopharma companies. We're in about 20 prospective trials, and we've analyzed dozens of completed clinical trials over the past couple of years.
MALORY EBRANCA: So when you say analyzed, what type of work do you do with these companies?
ANDY BECK: Yeah, so a typical clinical trial, when it completes and the biopharma sponsors typically invested hundreds of millions of in this trial, whether it's a successful trial, in terms of hitting its primary endpoint or secondary endpoint, or a trial that failed to hit its endpoint, there's tremendous value to our biopharma partners in really understanding why their therapies are successful or why certain patients show resistance. So that's really the core question we're typically answering is to better understand patient subsets, to better understand why therapies are working for groups of patients or not working.
ANDY BECK: And the way we do this is, for each patient, once a trial is completed, we can have three pieces of information for each patient. The whole slide image, taken before the treatment was given. Then we know which treatment they received, which could be a treatment or a placebo or treatment A versus treatment B. Then we know their outcome. So did they show good outcome following the treatment or poor outcome?
ANDY BECK: And essentially, we can then use that data in a machine-learning framework to better understand how the disease pathology at the time of diagnosis may predict which patients are going to fail therapy or succeed with therapy. And then it's valuable to get that data for our biopharma partners to then help them design the next set of trials they may want to carry out, or even the next set of, say, combination therapy strategies or new therapeutic approaches by really having a deep understanding of mechanisms, of response and resistance based on this AI powered analysis of human pathology.
MALORY EBRANCA: So when we started, you talked about the three things that have driven AI, and those were algorithms, data, and computing power.
ANDY BECK: Yeah.
MALORY EBRANCA: Can we go through those?
ANDY BECK: Sure.
MALORY EBRANCA: How have the algorithms--
ANDY BECK: From the perspective of Path AI or more broadly?
MALORY EBRANCA: How have algorithms changed? How have they advanced?
ANDY BECK: Yeah, so one of the big changes over the past decade, particularly in areas-- computer vision, I think, has been the most positively impacted by this. So computer vision is just a fancy word for image analysis or training computers to really be able to measure precisely the contents of an image. So that's been the area that's most relevant to Path AI, and it's most significantly impacted by the advances in algorithms over the past decade.
ANDY BECK: Not to get too technical, but the biggest algorithmic advance has been the development and further, rapid improvement in this area of deep, convolutional, neural nets, which are essentially very complicated algorithms with millions of parameters. So we really don't have a super crisp understanding what's happening on the inside, but we really provide these algorithms with two things, an image and the right label on that image. And you train it on potentially hundreds of thousands or millions of these examples.
ANDY BECK: And then you provide it unlabeled examples, and you say, put a label on this. And what's different about deep learning versus a lot of prior algorithmic approaches is all of the featurization of the image, the features of it, and the rules about making the prediction are learned in a data-driven way. So it's almost the way you teach a child. You don't teach a child what a bike is by saying a bike has two wheels, and a bike has a steering wheel, and it's about this size.
ANDY BECK: You just say, that's a bike. That's a bike. That's a bike. That's a motorcycle. That's a scooter. And they figure it out. Deep learning is the same way, versus the old system, you actually did have to have these human encoded rules, and those just happen to work much worse.
ANDY BECK: So now it's a more data-driven approach versus an expert rules-based approach or expert features-based approach.
MALORY EBRANCA: And in terms of the data, we talked a little bit about that. But now, is there a ceiling or a point at which you need a certain amount of data to make-- because people have been doing this for a long time. But at what point did you achieve enough data that your algorithms were stronger?
ANDY BECK: Right, so the thing is in pathology, and particularly in drug development, you can keep making the tasks harder and harder so that you continue to learn from more data. For certain tasks, there will be, I think, a saturation in terms of the amount of data you need to be essentially perfect. For example, identifying certain individual cell types in certain cancer types with certain stains. We may, for example, already have enough or soon will already have enough data to build an algorithm that's extremely robust for that task.
ANDY BECK: But for other tasks, like predicting molecular alterations, predicting patient outcome, predicting outcome to different types of therapies, those more complicated, predictive tasks are still benefiting from increasingly large data set sizes. So it really depends on the difficulty, in some sense, of the task. And as the therapeutic world keeps advancing, we're going to keep having new challenging tasks.
ANDY BECK: For example, new therapy may come out, a new combination of therapies, and we're going to have to learn from different data to predict or to really be confident in what subset of patients will benefit best from these new therapies.
MALORY EBRANCA: What's the challenge with NASH?
ANDY BECK: So one of the big challenges with NASH in the field has been that we've been relying, really since the field started, on expert pathologists analyzing under a microscope the features that they see with their eyes and their patterns to both identify which patients should be included in trials and then to identify whether those features seem to have gotten better or worse or stayed the same following treatment. The reason that's very difficult to do by eye is because these are all these semiquantitative scoring of things like inflammation, amount of fat, amount of ballooning or diseased-looking liver cells, as well as amount of fibrosis.
ANDY BECK: So just like you might wake up a certain day and estimate a certain number of cars on the road or something, and you estimate tomorrow, you'll estimate something different. Because these estimates all have significant intra within observer variability as well as inter, between observer, variability, it's been a real challenge in the field because there's no area where pathology is more important and no area where it's more difficult and requires experts.
ANDY BECK: So that's been a huge problem. And that's why this has been a major focus of Path AI because we really can solve this problem by creating a computational system that is very accurate, extremely reproducible, and can be used once it's fully validated and qualified for both enrolling patients in trials and assessing therapeutic response. So it's a very difficult task for pathologists, even for diagnosis, but it's particularly problematic in clinical trials.
MALORY EBRANCA: What has the progress been?
ANDY BECK: Yeah, so our progress in this area has been very strong. We've been working for several years. We've published widely on the performance of our system. And now, we're closely working with both leading biopharma companies as well as regulatory agencies to incorporate this increasingly into future NASH trials. And we're in current NASH trials as well. And we expect, over the next couple of years, this really will become the new standard in the field.
ANDY BECK: So it's been really rapid and exciting progress. And we're super excited for the years ahead.
MALORY EBRANCA: Well, one of the things-- the three things that you mentioned that have been spurring the field is the compute ability. And how is that changing, and how is that making things different for AI?
ANDY BECK: Yeah, it's been huge. So back when I did my PhD in this area about a decade ago, at least for us, compute availability was a big deal. And there were high-performance computing systems, but they tended to be difficult to access. And only a few of the leading, say, research centers around the world had access. Now, anyone with a laptop can access Amazon Web Services or Google Cloud or Microsoft Azure and access, essentially, unlimited amount of compute.
ANDY BECK: So it's been a huge boon to the field, and it's only getting stronger as more things move to the cloud, as more of these tech companies continue to fight with each other investing heavily in the growth of cloud computing. And then companies like Nvidia and others keep innovating on things like GPUs. It's a huge tailwind for this field and completely different than even a decade ago.
MALORY EBRANCA: So when you first got into this, it was completely different, right?
ANDY BECK: Absolutely, couldn't of-- it was extremely different considering it was only 10 years ago. On all these fronts, data, algorithms, and compute.
MALORY EBRANCA: What made you think that this was an opportunity?
ANDY BECK: Honestly, when I got into it, it was less about these macro trends and more that it just seemed like an important problem and one that you shouldn't just be using your eyes for, one that computers should be helping us with. I would say I was a bit lucky and naive to know that-- or to think that there are all these other trends that would really advance the field so much. But to me it was-- if you just learn about pathology, it becomes clear, and you see the power of computation for many other tasks.
ANDY BECK: This should be one that should be able to be assisted. So that's why I got into it because of the importance of the problem and the belief that the problem, one day, could be solved with technology. And then, yeah, these other macro forces have really advanced that.
MALORY EBRANCA: How has the market responded?
ANDY BECK: Yeah, the markets responded increasingly receptive just in the past couple of years. We started the company about six years ago. We had an early focus, and still a focus today, on drug development. But as I think mentioned in the preview to this, there's just been such a recent uptake in interest in using AI across drug discovery, drug development, and even commercialization of new diagnostics. And we're very much seeing that from the market.
ANDY BECK: So we're now well situated to help deliver on all of this increasing demand for use of AI across the drug development spectrum. And really, the area we focus on is pathology in development, which is translational research, phase one, phase two, phase three studies, and even the co-development of AI-powered companion diagnostics.
MALORY EBRANCA: What was the biggest hurdle so far?
ANDY BECK: There's a few, yeah, the biggest. So one is just technology. It's truly a technically hard problem to do it at scale, at very high quality, and reproducibly. So we have we still continue to make huge investments, and just how do we ingest these very, very large gigapixel images. These aren't tiny little pictures. These are huge slides. And then they contain on them hundreds of thousands of different objects.
ANDY BECK: So how do we generate the data set to really build computational models that can identify every object in image, do it very fast, quantify features, and then use those to better predict, for example, patient outcomes or which patients will respond best to therapies. So I think the number one challenge is technical. And then the other is how do we develop this in such a way so it can really provide great value to our customers.
ANDY BECK: And we very much built the company around that for supporting-- drug development is our first major area of focus in the sense of being able to analyze large numbers of slides in bulk for translational research, having the clinical operations and regulatory and quality expertise to support prospective clinical trials, and then having the quality systems and medical device capabilities to support companion diagnostics.
ANDY BECK: So I'd say that's been really the big second challenge.
MALORY EBRANCA: You say that AI has crossed the hurdle. There was a point where everybody was saying, oh, AI, black box, hand waving, do you think it's crossed the barrier? And if so, why?
ANDY BECK: I think in certain areas, and in certain-- the one thing I always say about AI is it's not that meaningful of a word by itself. So it's like you can almost always replace it with the word math and try to say is this a meaningful question? Has math crossed the barrier? Well, sort of for certain things. But for certain things it hasn't. But to get a little more specific, I think for computer vision tasks, it clearly works really well, and in the right hands, can build really powerful tools to assist, to augment what people do and to do it even better in terms of interpreting what's inside images.
ANDY BECK: That, without a doubt, there's no doubt that that's the future. It's also largely the present. In terms of the actual drug discovery, I think we're still waiting on proof in human that it's more efficient, and you have fewer failures using AI. I think there's a lot of hope there. There's many companies working on it, but to go from idea or discovery to an approved medicine takes a decade or longer.
ANDY BECK: And people have been working on this and talking about this for a very long time. So I think we're still in the early stages of that. I think, clearly, AI, in some form, is going to be central to everything because it's essentially just improved applied math. So I think that's not controversial. But in terms of the difference between quote unquote "AI today and AI a decade ago," I think that's the more tractable, answerable question.
ANDY BECK: Because of course, AI is important, but it's like, is today really any different than 10 years ago, I think is the real question. I think in vision it is. I think in other areas of drug discovery, I believe it will be, but I think it still needs to be proven.
MALORY EBRANCA: Well, thank you so much, Andy, for your time and your insights. And congratulations for the progress that you've made. We're looking forward to watching you and your company and the more that you achieve. Do you want to say anything final in terms of--
ANDY BECK: Well, thank you. Thank you for this opportunity. And no, I think we covered it well in the interview. So thanks for the time today.
MALORY EBRANCA: OK, thank you. Take care. [MUSIC PLAYING]