#102, Decoding Human Biology with Single-Cell Insights, with Lindsay Edwards from Relation

So actually, if we start for people who aren't familiar, if you start with the incredible cost of drug discovery, so it's between $2 and $3 billion, it moves around, but let's just say a lot of money to discover a new medicine. And that's not because there's just terrible waste. It's because we fail a lot of the time. Welcome to the AWS Health Innovation Podcast, where you can learn from entrepreneurs and investors who are driving progress in health care and life science around the globe. Welcome back to the show, everyone. I'm Alex Merwin from AWS. Today's episode will be hosted by my colleague, Yin He. Yin welcomes Lindsay Edwards to the pod, the CTO and president of platform at Relation Therapeutics, a biotech company that's leveraging single-cell, multi-omics, machine learning, and clinical insights to develop transformational medicines. Lindsay will share his insights on Relation's unique approach to drug discovery, his unconventional career path from musician to drug hunter, and the company's focus on using single-cell data to find better drug targets. But enough from me. Let's get started. Hello, and welcome to the AWS Health Innovation Podcast. I'm excited to guest host this session today. My name is Yin, and I lead our work with our early stage health care and life sciences startups and investors at AWS. I'm trained. My background is a wet lab scientist. I have a PhD in molecular bio and genomics, and prior to AWS, I was an early employee and operator at a venture-backed startup building cloud labs and lab automation to help pharma and biotechs accelerate drug discovery. Today, I'm so thrilled to be joined by Lindsay Edwards, CTO of Relation Therapeutics. All right, let's go ahead and dive in. Welcome, Lindsay. So excited to have you. Thank you. Great to be here. Wonderful. To get us started, we always ask about the company. So tell us a little bit about Relation, what you all do, and how is your approach to drug discovery differentiated and unique? Sure. Thank you, Yin. I guess I should start by saying that there are a lot of companies out there right now who would fall on the intersection between machine learning and drug discovery, but that leaves a lot of room for uniqueness. So we are leveraging two really exciting technologies, single-cell, as well as machine learning. We're not just a tech company, though. So one of the other things that we have here is really deep expertise in drug discovery. A few of the senior leaders here, myself included, have worked in Big Pharma and have worked along every bit of the drug discovery value chain from target ID all the way through to post-approvals. I think all of us have successful medicines on our CV, certainly the CEO, David, and the CSOA did do. But then I guess the final thing, our superpower, is about integration. We're not these disciplines bolted together from the very birth of relation. We've really tried to make it a company where these disciplines come together. And I guess we'll get into this a bit more in the podcast, but I'm increasingly convinced that the reason this is important is because these are not machine learning problems or biology problems or data science problems in isolation. We're trying to tackle problems that live on the interface between those things. And so we've worked really hard at the company to build the physical infrastructure. So I'm sat here in relation HQ, but literally just 30 yards that way is our beautiful kind of state of the art wet labs where we can do CRISPR and handle human tissue and all that kind of stuff. And just the other side of a window are all the machine learning folk and the technical staff as well as clinical capability and the culture and the people and everything that you need to tackle these really important problems. Yeah, absolutely. So would you say you all develop your own therapeutics, provide a platform or software? Where do you sit on that spectrum? Absolutely. So we are an end to end drug discovery company. So we'll develop our own therapeutics. We'll also form partnerships, not a huge number, but a few to help us expand our pipeline. And yeah, the aim is to have relation medicines in the clinic. Yeah. And I'm sure with your background, we'll dive into that next coming from pharma, right? Really thinking about where you can form strategic partnerships, right? So it's not about the number or really the quality and then how do you tie into some of the strategic goals of the company? All right. With that, yeah. Tell me a little bit about your background and how your journey, right, to relation. How did that happen? Yes. I have a very non-standard career trajectory. We love you. These are the best. That's why we pick folks to feature on the podcast because if you think about, yeah, like founding teams or founders, it's never, hey, I was just a scientist or I was just a computer or software engineer, right? Like everybody follows these very non-traditional trajectories. And as you said, right now we're operating in a field that requires extreme cross-functional disciplines and skills, right, to put all of the pieces together. Yeah. Please continue. Yeah. So I started out as a musician. I was a pop musician here in the UK and I guess moderately successful. So we have hit records here. Okay, maybe we can change, yeah, the song for this session to one of your songs. All right. We're going to have to check it out. That would be great. I get royalty still. Yeah, every now and again, yeah, the YouTube videos pop up. But yeah, so the high point of that was we had a single with Emma Bunton from the Spice Girls as the singer and that was a big hit over here in the UK. So it feels like a lifetime ago, but weirdly enough, coming to a startup after being in a big corporate felt a bit like coming home because you're back to that feeling of not quite self-employed, but you do everything. So it did feel very natural. But at some point, I think it just had a good run, probably 10 years of being fairly successful. But I think I realized at some point that I actually wasn't really enjoying it in spite of the fact that it was fun for a lot of the time. And I went to university and I studied initially studied exercise physiology. Then I got a scholarship at Oxford to do physiological biochemistry in a group that was very interdisciplinary. So we did magnetic resonance and a lot of human stuff. So my PhD was around fat metabolism in humans, but it was also because the group was a magnetic resonance group. There was computation and physicist and engineers and people building things. And so it was a real melting pot of different disciplines. And I really enjoyed it. And I fell in love with experimental science and I fell in love with computational science. I went off on a standard academic career track. I've always been a nerd as well. Like I started learning how to write code in basic on a Commodore Pet, which ages me a bit. So it was nice to be able to bring those things together. And so as an academic, I was doing systems biology. So we did a lot of metabolomics, a lot of human tissue, and then computational modeling of kind of different systems. So enzyme kinetics, models of whole cells, machine learning. I like to say when it was weird and not cool to be doing machine learning and then came back to the UK. I had a small group at King's College London just over the river here. And then in 2014, I got recruited to GSK and did a number of different jobs at GSK. And it's a great company to be entrepreneurial within the company. So I started functions that didn't exist before. I started one of the early data science groups at GSK and started doing a lot of machine learning on microarray data a bit at the time. And then just grew within the corporate structure. So built various different functions. My last job at GSK was VP of machine learning engineering. I went to AstraZeneca for a couple of years. We worked somewhat tangentially on the COVID vaccine, which is very exciting, and built a very good machine learning research team there. But started to get this feeling that some of the things that I wanted to be able to do were going to be difficult to do within that large corporate environment. And I'm always quick to say, that's not because there's anything wrong with any of those companies. They're amazing companies that do amazing work. But large corporations are generally not designed for doing like rapid innovation or to be able to have the kind of agility that you have at a company like Relation. Yeah. So I took the decision, this is my third year at Relation. So two and a half years or so ago to leave corporate life and come and join a startup. And I've been loving it ever since. Curious. How did you stumble upon relation? Was it a relationship that you had or working tangentially in the space you knew about it? Yeah. So the machine learning community is actually still quite small. And particularly, if you go to the major conferences, so ICML in Europe, particularly, you get to know, everyone gets to know each other, at least to say hi to. And so one of the founders of Relation, who's a guy called Charlie Roberts, who's an NHS physician who now lives in California and set up a company called Freenome. So Charlie and I met at ICML in Long Beach in, I think, 2019 over a terrible taco and just got on really well talking about machine learning and applications and healthcare and stuff. And so we stayed in touch. And so I knew that Relation existed. I think I'd tracked their progress. And so when I was looking for the next move, I was looking for a company where it was early enough that I would feel as though I could really have some influence and say in how the company grew and developed, but also probably passed the two people in a garage stage. And also with Relation, some of the real basics had been laid down phenomenally well by the founders in terms of the vision and the idea of integration and the skeleton and framework of the company was fantastic. So it was a really good springboard. Yeah. No. It's usually somehow relationships, right, especially if you're joining a company in its earlier stages. So that's really great. I think you have an interesting perspective coming from more of the machine learning perspective and then also having done multiple pharma. I want to, from your perspective, what do you think are the top two challenges right now in drug discovery and development and then how does Relation's technology fit in? So it's drug targets, drug targets, drug targets in no particular order. I'll take one. Okay. It's fascinating because it's the, so it's unarguably the single most important thing that needs solving. If you look at the statistics of why we fail, so actually if we start for people who aren't familiar, if you start with the incredible cost of drug discovery, so it's between two and three billion dollars, it moves around, but let's just say a lot of money to discover a new medicine. And that's not because there's just terrible waste. It's because we fail a lot of the time. So drug discovery is this long process where you start trying to identify a target, you understand the sort of, you start by identifying a disease and then you try and understand a subgroup of patients or patients with that disease, you try and understand the biology in those patients versus some other group and then you try and develop a drug target and then you develop a drug and blah, blah, blah. And eventually you go into human studies and we have these phases and again, I know you're going to be familiar, but it's a people listening who aren't familiar with the process. You have phase one studies, which are typically when you very first time you give the medicine to humans and they're often healthy and you do a dose ranging study to see whether it's safe. And then the second round, the phase two is with the very first time that these drugs see patients, usually in a smaller study so you can look to see whether there's any kind of signal for it actually working. And then you do these big phase three studies, which are typically of the order of tens to hundreds of millions of dollars and those are the ones where, you know, and they could be huge like heart failure studies of these massive things. And the issue is that you don't really know whether it's going to work or not until phase two. So you do all of this work and then you might be unlucky and fail in phase one because it's not because it's got the adverse event profile, but typically you fail in phase two. And by that time you spent a lot of money, but you fail 95% of the time. So of that sort of two to three billion cost, a lot of it is down to the 95 failures for every five successful medicines that you get out the door. And when you break that down, if you look at why did we fail? The reasons typically are it doesn't work. So you've done all of this work, you've given a medicine to patients and it does nothing or it's toxic in some way. When toxic isn't the right word, but there are adverse events associated with it. And then you have this notion of therapeutic window. So it's what's the amount of this new medicine that I can give to somebody after it starts having an effect and before it starts showing like adverse effects. And it's really important because these two things are connected with each other. Whatever the medicine is, take enough of it, it's going to have an adverse effect. And the issue is that whatever you've decided to drug, you just can't get to a therapeutic dose without it causing side effects that are unacceptable. And that's why we fail like 60 to 70% of the time. So if you look at the overall cost, most of it is down to this like we've probably targeted the wrong thing. Some of it is not that it could be to do with the actual molecule itself, but the majority is this thing. And that pretty much boils down to the drug target, which is the protein in the human body that we've designed a molecule to go and target. And so getting the right drug target is the key to unlocking this. But the problem with that is that you don't really know whether you've got the right drug target until you've done all this work and you give it to humans in phase two. So it presents this extraordinary challenge. Yeah, exactly. And that's why people have tended to shy away from working on it. And to answer the second part of your question, the reason what relation does, that's pretty much early in the company, that's been 100% our focus. How do we become better at finding better drug targets? And the way that we tackle that in the main is to use two sources of information that we know to be useful. So the first one is human genetic evidence. So there are some excellent scientific papers going back five or 10 years now, showing basically that one of the few pieces of information that really increases, like more than doubles, your chances of success from a target perspective, is whether or not you've got human genetic evidence supporting that target, and that's a major one. And then more recently, people have started to look at single cell data and look to see whether data from single cell experiments can also boost that. And those are the two core information sources that we work with at the moment with relation. And the last thing I would add to that is there's a lot of effort over the years curing mice, and we are extremely focused here on human data. So we have a clinical study in our primary indication, which is osteoporosis, where we get bone samples from human patients with osteoporosis. And it's messy because it's humans, but it's human data. And so you have that confidence that what I'm looking at is the real biology in a real human patient rather than trying to cure all mouse diseases. Yeah, that's so important. It's like we have this very extensive animal work phase in this entire cycle, but rarely is it predictive when it goes into humans. And so I'm definitely keeping an eye on when that starts changing. There's actual pressure to change how things are done, because I think right now, animal studies, people still want to see them if you're doing published research and things like that. It's a really important data set that you have to submit, so it'll be interesting to keep an eye on the space and see how that evolves. You mentioned a little bit about the first indication in focusing on osteoporosis as the first therapeutic target, an area, presumably because you all have some data that you're finding the right targets in this space. Can you tell us a little bit about how you decided to focus on this area and some of the progress you've made? Absolutely. So osteoporosis is one of those things that half of us will end up with. I'm not sure of the exact statistic. I think it's that 50% of everyone over the age of 50 will end up with osteoporosis. Just very quickly, what osteoporosis is basically a weakening of the bones. And I guess traditionally it was thought of as being a disease primarily maybe of postmenopausal women and because of probably some elements of structural sexism, it hasn't been looked at as closely as it should have been. We do have some treatments for osteoporosis, but they all have issues. They're either horrible to take or going back to the earlier conversation about adverse events. The adverse events are not great. And the other feature of some of the existing medicines is that there's like a reversal. So if you stop taking the medicine, then you can end up just going back to where you were before. So you get a couple of years grace or three years grace. So there is actually a huge unmet need in osteoporosis. So that would be the first thing. And that's one of the kind of less well-recognized things about drug discovery is that the moral imperative and the kind of financial imperative align pretty well, like the bigger the population and the greater the suffering that you can address, the better it is for everybody, for the company and for the patients. And we put the patients at the heart of everything that we do, like our company Tagline, you'll see it on the website is the patient is waiting. And again, it's so important to have that human element to it. I have a close family member who has osteoporosis and so it helps bring that personal motivation to these kinds of this kind of work, which is so valuable compared to... Yeah, at the end of the day, that's why we're all doing this, right? Yeah, yeah, exactly. Rather than optimizing web click-throughs or something, this feels like it's an important work. Another thing that I think that's fascinating about this is in drug discovery, we talk about this idea of good sort of translational fidelity. You mentioned it just then in terms of like animal models, not translating into humans. But actually, this is something that applies all the way through the drug discovery life cycle, if you like. One way of thinking about drug discovery is this series of interlocking steps in terms of the biology where each proceeding step or each step as you go along, the biological models get less and less easy to work with and more and more like clinically relevant. So you start out with single cells, often immortalized single cells that are easy to like gene edit and stuff, but are capturing who knows how much of the biology and use... Then you have multicellular systems and then you, as you said, you might move into pre-clinical models until you get to humans. So it's this kind of... Each of these is a step as you go along and this idea of good translatability is that you try and have each of those models as representative as you can get it of the patient that you're trying to study. And one of the challenges, particularly in early drug discovery is like, what cell do I culture in the lab in order to understand the biology that's driving this disease? So if someone's got asthma, for example, that's something I used to work on, there's 65 different cell types in the lung. So if you're looking at the genetics of people with asthma, when you want to test what... You have to figure out which of these 65 cell types plus immune cells. It's a really tough task. And osteoporosis has this... One of the things that's nice about it is that the cells involved in making bone and turning bone over, it's a much smaller number. So you have these cells called osteoblasts, which make bone, osteoclasts, which break bone down. You have these cells called osteocytes, which kind of orchestrate everything. But if you want to study a model of bone being made, it has to be the osteoblasts. So it's good from a translation perspective. And there's loads of other reasons. So there's lots of good genetic evidence as well, because there's large groups of people with osteoporosis. You can study them quite easily. There's a large number of patients in UK Biobank with osteoporosis. Also diagnosis codes are often extremely noisy. So if you go into health records and you're trying to look for evidence that associated with the disease, you'll have people diagnosed with a disease that don't have the disease. You'll have people with the disease that aren't diagnosed with the disease. Whereas with osteoporosis, you have this relatively simple measurement, bone mineral density, which is like a measurement of bone quality. And you can use that in all of your modeling and all of your downstream stuff. Anyway, all of these things are what we would typically talk about in pharma circles as being good translatability. You've got good translatable cell models. You've got a nice, we call it, an endpoint in bone mineral density. So one of the things I've learned at Relation is all of these things make machine learning easier. So if you're trying to, say, predict drug targets in asthma using something like a knowledge graph and you want to test whether those are real targets or not, which cell type do you do it in? You've got like 70 to choose from. It's really difficult. Whereas at Relation, what we're able to do is take human genetic evidence from patients with osteoporosis. We can look at which genetic variants are predictive of having high or low bone mineral density. So we're not relying on diagnosis codes. We can take those variants and predict which genes are associated with them using one of our models. And then we can look in osteoblasts because we're really interested in drugs that drive more bone growth, so anabolic drugs. And so we look in osteoblasts and you get a straight answer. It's like this gene is affecting bone deposition in this human cell type. So it's much cleaner. And from a machine learning perspective, I used to joke early, there's, I think it's an Abraham Lincoln quote about only fighting one war at a time. Like you've got a model that's modeling one thing. And you don't want to be worrying about 50 other drug discovery related things. You want as clean an answer to your model outputs as possible. Yeah. So that's the reason we went after OP in the end. I know it makes a lot of sense in terms of how do you quickly accelerate the process with well understood biology with biomarkers that are well characterized that you can pull without a physician or clinician making some sort of diagnosis on top of that and depending on that step. Yeah, that's interesting. I feel like now what you hear is everybody wants to throw machine learning at these extremely complex problems and expecting that through it they're just going to get clarity and answers. So I think it's having you come from the other perspective. Hey, let's deploy machine learning, right? But on clear concise problems where we have very well understood biology and have a focus area. Ian, actually, just to comment on that quickly, it's one of the things you see a lot is like modern machine learning models are extremely complex. And because they're extremely complex, it's difficult sometimes to figure out what they're doing and also very easy to fool yourself that they're working way better than they actually are. And if you take a very complicated data set and you take a very complicated model and you fuse those two together, you can almost certainly make it look as though it's doing something. And if you really want to believe it is, then, you know, it's like that sort of built-in human bias. Whereas what we found is that it's much better, it's a little slower, but it's much better to feel your way into the problem and spend a lot of time kicking the model to see if it's really doing like we'd spend a favorite of time trying to prove ourselves wrong. What are the ten ways in which this model could be fooling me? But it's really important because if the model actually, if you've kidded yourself that the model is better than it actually is, it'll come unraveled eventually. Well, that's the thing. I mean, you will fail at some point, right, sooner or later, you go into humans or animals or whatever it is, you're going to find out. Yeah, sooner is better. Definitely. Yeah, yeah, ideally. Fail faster when you're talking about $3 billion process. Yep. Yeah. Okay. I'm going to take yours a little bit and talk a little bit more about some of the internal tools and I would almost say like lab infrastructure that you've built, right, specifically about deploying lab in the loop approach. Can you talk a little bit about what this means to you and how has that enabled you at relation to do what you all do? For sure. So for us, lab in the loop is our way of summing up the level of integration that we strive for here at relation, and it's this deep integration between the data generating processes, whether it's from human data or from the assays in the lab and the machine learning. And where the loop comes in is it's not just about data feeds, the model. It's about the models driving the experiments that we do that I think is really important. So to give you a concrete example, one of our core models we call Rosalind internally is a model that allows us to link natural human variation to the genes that are being affected. And this is an interesting problem because if you look at genome wide, we say GWAS but it stands for genome wide association studies. Like I said before, you take a bunch of patients with osteoporosis and a bunch of patients without osteoporosis and you look to see how they're genetically different. And what you find is a lot of the things that are different between the genomes of people with and without a disease are not in the bits of the genome that make the proteins. We call it non-coding DNA. And it's not necessarily clear how the changes in non-coding DNA are driving the disease. So we have this model that's trained on loads of DNA and loads of other kind of cellular data that helps us make those predictions. But key then is having made those model predictions, they feed back into the lab and that drives our experimentation in terms of using CRISPR to do gene editing to help us understand the role of those genes that the model has predicted in osteoblasts, like we talked about these cells that drive kind of bone deposition. And then that allows us to inform again the models that we make and how we train our models. So on the kind of day to day level, it's about this deep integration between the models and the data generation itself. And I think there's a bit more to it as well though, which is we're still very, not us here at Relation, but as a kind of, as a set of communities, we're still very siloed. We've just got back from ICML, which is one of those big, one of the big two machine learning research conferences. And there's certainly more pharma and biology presence at those conferences than there was say five years ago, but it's still mainly the machine learning community. And to just refer back to the point I made early on when you asked me about like why relation and why go to relation and why create a company like this. The integration needs to happen at this kind of level above the technical integration. And so for example, that project that I just described, there are, there's a team of people who sit together and can talk to each other about this stuff and think about how we design the experiments and think about how we test the models and think about what data we generate and how the model informs what data we generate. And those conversations, that's like an evolving thing and it takes time and space to breathe for the biologists to get really involved in the machine learning and for the machine learning guys to really understand the biology and for those conversations to happen. And although lab and the loop to us is about the kind of integration between the wet lab and dry lab at its simplest level, there is also this element to it, which is about the culture of the company and allowing people to spend significant time together thinking about that very specific problem. How do I use a large neural model to predict the effect of changes in human DNA on the downstream biology and how do I test that in the lab? Yeah, as long as there's still tons of humans in the loop and in touch points at the end of the day, we still have to figure out the challenges around communications and collaboration and these, again, as you mentioned, are more cultural aspects of a company. I think it's massively underestimated. It's one of those things where you talk about collaboration and it's a bit like osteoporosis. Everyone thinks, oh, we solved that because we've talked about it for such a long time. But we actually, I think we're in the foothills of how we design companies that work on these things. And one really surprising thing I noticed is a lot of scientists, particularly very talented scientists, have been successful because they take a problem and they go away into a corner and they think really hard and then they come back and they're like, "Okay, I think I figured." They're good for that independence, right? Yeah. Exactly. And yet that's almost like a failure mode for working on these interdisciplinary things because no one person has everything in their head to solve the problem. So if people start going off on their own and thinking about stuff, you watch the two halves of the conversation drift apart. So you have to do the work of pulling people back again and saying, "Let's go again." And trying to understand and the team becomes the unit of operation rather than the individual. Absolutely. Yeah. It's one of those things, again, that looks easy or people are like, "Oh, we know how to do collaboration." Wow. Yeah. In some ways, you do that. Yeah. And this for this... Like, where are these people? Let me talk to them. Yeah. Yeah, exactly. Have you done it? Yeah. I've seen this firsthand and this always fascinates me because, again, for these types of companies to work, you have to get in its most basic form, scientists and engineers to work together and to speak the same language. A lot of the times the engineers are developing internal tools for the scientists, but you be shocked at how often they would be coding at their computer, building these things, never stepping in a wet lab to truly understand what the workflow looks like, what are the actual needs of these folks, what does a UI/UX design process have to look like for them to... Because they're an internal customer at the end of the day. It's not just, "Hey, this is a task I have to do," and at the end of the day, we're focusing on external customers. You have to focus on your internal customers as well and it's the same process. You have to come up with requirements and build towards them. Yeah. We got to the point where we had to almost enforce and say, "Hey, these teams, as you mentioned, as the unit of collaboration," to actually go and spend time with them, to shadow them, to understand if you're building a tool for a scientist, you have to understand what their challenges are and perhaps they're something new. Because again, you can't always rely on the scientists to tell you what should be built for them in terms of tooling. They don't fully understand of the capability. We really had to almost force this and that led to better outcomes, but I'm always super intrigued by this because I think moving forward, at its minimum, these two teams have to almost operate like one very seamlessly. Yeah. Absolutely. I think as we're winding down here, I want to ask, you've had a pretty entrepreneurial journey. What advice would you give to folks that are listening in and thinking about starting their own venture here? In terms of career trajectories, I've been very lucky now to have done two things that I love. My music career was something that I enjoyed, at least to begin with, and something that was, I was naturally good at it, which makes it a bit easier. But I think more than anything, it was fun. My career now, and particularly with machine learning, it just really spoke to me. The first time I started training models and building computational models, I think what really intrigued me was this idea that you could build these little models of reality in a computer and then experiment on these little models of reality. There's something to me magical about the fact that we have models here where you can change the DNA sequence and the model will have a stab at telling you what's going on. To me, that's just inherently a cool thing. But I guess the reason I'm saying this is because if you have that, then going to work, I mean, it's the oldest adage in the book. If you enjoy what you do, you'll never work a day on your life, but it turns out that's true. What it doesn't account for is the flip side of it, which is having to get up and work harder the job that you hate is really soul destroying. We all have to do it at some point, but it's so much better if you do something that you love and you feel passionate about. That would be the first thing. The second thing I would say actually is that there's so much more possibility in the world now for the generations that are coming through with capabilities in these areas that I really would encourage people to go for it. There are investors out there that want to invest. There's all sorts of organizations, like is it entrepreneurs first, these kind of collectives? I know AWS is very supportive of young startups, NVIDIA is very supportive of young startups. So there's lots and lots of possibility. And I would say as well, I don't think this is a particularly London centric podcast, but London is a phenomenal place right now for everything that's going on. Like I look out the window here, I can see the Welcome Trust. The Crick Institute is just there. I can look across and see UCL. We've got Google DeepMind just up the road. So there's this kind of square mile called the Knowledge Quarter, which is just this amazing confluence of healthcare and academic institutions and startups and biotech. It's really exciting. So I guess the other thing I would say is, yeah, just go for it. If you have an idea and give it a go. One quick thing actually, one of the things I love about computer science, I think this is a Zuckerberg one, but where he was talking about software engineering as a discipline versus other engineering disciplines. And if you're a mechanical engineer and you want to build a bridge, you need a lot of money and people and support and stuff. If you're a software engineer and you have a good idea, you can build it. You can scale it or you can push it out to potentially millions of customers without ever going out. It's effectively something that you can design and build on your own in your front room. And I've always found that as well, like an incredibly attractive idea that I can have this idea and I can build it and scale it just me. And yeah, so I don't know if that made any sense at all, but do something that you love. Don't be scared to give it a go. And if the things that you love happen to be software based, you're in luck because you can build them and scale them at home. And obviously, again, like AWS has had a huge role in making that possible because the whole concept of Cloud Compute has made it possible for people to access huge amounts of computation again without going out. Yeah, absolutely. Just go for it. You have to have conviction. You have to have passion. You have to believe that what you're building needs to be built and that you are the right person to do it. There's nothing to lose. There's only things to gain. At the end of the day, even if you "fail" in terms of the business taking off or going out of business, think of everything you've learned right along that experience and it will only set you up to be better at your next venture. We talk to serial entrepreneurs day in and day out and they are the ones who actually we have more conviction in because they've done it once. It's almost better if they failed because they know what they're getting themselves into and actually they're going for it again. I think there's a lot to be said for that. All right. Very last question. I want to know, because you shared your music career, what from your music career have you taken away that impacts your life on a day-to-day basis in this role? Like a skill or some sort of ritual or anything like that? You know, there's actually a surprisingly large number of things. I'm comfortable in front of people. I have to say teaching really helped with that. So I was a university lecturer and I taught medical students. They're a tough crowd. I'm comfortable on a stage with quite a hostile audience, which is turned out to be really useful. I think as well, feel is incredibly important. We live like, I got into science, I think in part because I was attracted by the data-driven aspect of it. But in fact, a fair bit of what we do as scientists is about feel and creativity and inspiration and insight. And in that respect, it's really not very different whether you're sat in front of a keyboard in the studio trying to come up with an idea or whether you're sitting looking at a problem in the business or whether you're looking at some data or whether you're looking at a machine learning model, trying to figure out what's going on. In the end, these are all creative processes. And creative processes require a certain amount of gut. You look at something and you think, yeah, that feels about right, I'll give it a go. And I think one thing that happens as you progress in your career as a scientist is your gut gets better because you've seen more examples. So whatever part of your brain it is that does that subconscious pattern recognition that makes you go, huh, this looks like, even it's like subconsciously, this looks like a thing I've seen before, that gets better because it's seen more examples. But then you also get, I've got better at going, I'm going to trust that instinct and push on it for a bit. And because there's lots and lots of situations you encounter where the data are pretty inconclusive and you're not going to get better clarity and you still have to make a decision. And yeah, I would say that, I don't know if tool is the right word, that kind of internal spirit level that allows you to go inside and go, does this feel like the right thing to do? Or does this feel like the wrong thing to do? That was the same, whether I was trying to decide which high hat sound to use as when I'm figuring out what path might be a good path to take in a particular project. I would say that's continued to be useful, but also the medical students, wow. We'll have to dive into that in a part two, maybe. For sure. Yeah, yeah. No, I love that learning. We actually have this leadership principle called leaders are right a lot. And it's what does that mean? And so if you read about it, it's about folks that, as you said, have experiences and anecdotal data, right? Not just the traditional data that we think of, but through their lived experiences, they've been able to pattern match, right? And even in the case of uncertainty, say you have 60 or 70% of the data, but you have to make a decision or able to move forward with that. Yeah, I really like that. All right, Lindsay. Thank you so much. It was an absolute pleasure. Thank you, Yin. It's been a pleasure for me as well. Thanks for joining us today for the AWS Health Innovation Podcast. If you want to get in touch with AWS, please check out our show notes where you can find a link. If you enjoy the podcast, the best way to support us is to share it with your colleagues and friends. We also really appreciate your reviews and ratings wherever you listen to podcasts. We love hearing feedback from our listeners, so please don't hesitate to get in touch. Again, you'll find all the details in our show notes. See you next week. [BLANK_AUDIO]