The Machine Learns Medicine: Daphne Koller’s Vision for AI in Biotech

0:00I think that the future is in a partnership between the human and the machine. I think for every technology that we've constructed in the past, people are like, oh my God, this is going to take away my job. And in fact, it did take away a lot of jobs. I mean, electricity certainly took a lot of jobs, right? I mean, the agricultural revolution took away a lot of jobs. this will certainly be the case here as well. But I think human creativity, human innovation is still something that is an important partner. Hi, everyone. Welcome to Leadership Next, the podcast about the people and trends that are shaping the future of business.

0:45I'm Diane Brady. And I'm Kristen Stoller. Kristen, this week we are speaking with a MacArthur Genius recipient. She's a genius. She's in our midst. I'm so excited. I'm sounding like a cynical Canadian here. I actually am very fascinated with her. I'm fascinated, first of all, with the concept of genius writ large. I have a friend who got one of those grants who's a jazz musician. Here's a woman who co-founded Coursera, now has In Citro, which is drug discovery. I mean, what does it mean to be a genius? I mean, founding two companies is a great start to being a genius in my book. That's a great start.

1:23And Stanford professor too. Yes, let's not forget that. Let's not forget the Stanford professor. Well, look, I think the most exciting use case for AI is in healthcare, drug discovery. And she's right in the heart of some of the most difficult diseases. She really is. They're focused on ALS. They're focused on fatty liver. Those are two really big ones that haven't seen a lot of innovation. No, and heartbreaking ones. I have, My old boss has ALS, and it's heartbreaking to see the decline people have. And I think with her in particular, she does seem to be somebody who follows her curiosity from previous interviews I've seen.

2:00So I'm curious to know what it is about this particular part of the AI ecosystem that drew her to it. She could have gone anywhere. I totally agree. And speaking of AI, I think she's catching us on a very good week, because we had such an AI-themed week last week. You were even interviewing my digital avatar. I was. I was interviewing your digital self to prep for an upcoming conference. I had an AI dinner on Wednesday that I went to. And then Thursday, you and I sat down with this fascinating CEO in the AI space. Sadly, we have to gatekeep his name a little bit. But remember how fascinating the debates we have?

2:35Yeah. Look, I disagree with him on a lot of things. Well, that's made it fun. Number one being the idea that AI will reduce drinking because it'll help us connect more socially. I think that's open to debate. Let's unpack that because he was saying that if there is, you know, chat GPT, you could have like a chat GPT and AI in a social setting, helping tell you like conversation starters, questions to ask, and no one needs to take a shot anymore. Look, some of us may be neurodivergent. I'm not making any projections here, but maybe that would help some people. I don't think that's why everybody drinks, A.

3:12A, and B, look, I think EQ and IQ are two different things. So let's just leave it at that. And I do think AI has a lot of potential. Clearly, this is an area that excites people. And I'm really curious to know what it is and when she hopes to be making some of these big advances. I'm also curious about Coursera. I mean, to me, it's a fascinating company that really did not fully realize the dream of online learning. Don't you think? I mean, we've basically proven the fact in the same way that AI cannot help you at a party necessarily to be determined. I don't think online learning helps you learn as well as being in a classroom.

3:57I'm with you there. I think the bigger problem for sure is in Citro because I feel like, you know, of course, education is an issue. But with drug discovery, it's such a field that has so much potential. It's such a costly field. Billions of dollars are invested in every year. It takes so long to bring a drug to the market, 10 to 15 years at best. And it's really successful. They have a 90 % failure rate in clinical trials, not very effective. Big pharma. Absolutely. Yeah. Look, she has a phrase that I've heard used in other interviews. Calculus is to physics as AI is to biology. We're going to have her unpack that.

4:34But I do think that we now have an infrastructure to improve drug discovery, no question. Can a startup disrupt big pharma or is it really additive to what they're doing to be determined? Well, let's find out with Daphne after the break. Great. We'll be right back. The best business leaders today know the value and importance of empowering those around them, personally and professionally. By encouraging and enabling others to grow, take risks, and fuel innovation, business leaders are not only driving greater engagement and performance, but also future-proofing their organization for years to come.

5:10I'm joined by Jason Garzadas, the CEO of Deloitte U.S., to talk more about this. Welcome, Jason.

5:16Daphne Koller:Well, thank you, Diane. Great to be here. Innovation is about empowering the people around you. And that's something that a lot of CEOs struggle with. How do they embed it into their leadership style? Well, I think there's all types of CEO leadership styles clearly and proven that there's maybe not one recipe for success. But it does require, I do believe, a commitment to inclusive leadership where all are expected and invited to contribute around innovation. I think there's also a collaboration and a collaborative culture that's a requirement. That's also not something that maybe comes as naturally and has to be cultivated and be intentional about.

5:51Daphne Koller:And then also, I think, giving leaders some autonomy to actually look at opportunities for innovation, look at opportunities for creative new ideas to bring forth. That requires a degree of trust and a degree of openness by CEOs in particular to allow for that within an organization. So, Jason, I want to, on a personal note, I'm talking to a CEO here. What are some of the most effective strategies you think for fostering open dialogue, collaboration? A lot of what you're talking about is the ingredients to innovation. Well, for me, it starts with being genuine and authentic as a leader, being clear that the single leader doesn't have all the answers to every question.

6:27Daphne Koller:And certainly in my case, it's inviting a very broad organization to participate in addressing the issues and challenges that we face. So I think that genuineness and that transparency and authentic leadership style is the key ingredient from my experience. Good advice. Thanks for joining us, Jason. Thank you, Diane. Daphne, in doing my research, I came across an interesting quote that you said, which was, calculus is to physics as AI is to biology. What did you mean by that? So disciplines become something that we can manage as humans when they become predictable, when we can have a formalism that allows us to predict what will happen in an experiment and have it be approximately correct.

7:10And calculus was that for Newtonian physics. Whether it's a ball rolling down a hill or a pendulum swinging, we were able to make predictions about what a system will do. And we've never had that for biology because biology is complicated and intertwined and we really can't predict. for almost any experiment, what will happen. And now, with new measurement technologies on the one hand, and the power of AI on the other, we can start to create models of biological systems where we can start to make predictions that are at least somewhat likely to be correct. And I think that is the path forward to many applications, including, of course, in human health.

7:48But I think it's going to be an incredibly important discipline in the 21st century. You talk in these amazing philosophical phrases. That wasn't the only quote that Diane and I came across that I thought was interesting. You have a lot of others. And you're a Stanford professor. Do you come up with all these yourselves? Do you have friends that help you? I think much of what all of us do is a distillation of bits and pieces of things that we hear and kind of put together in slightly new ways. I can't say that I invented these from whole cloth, but rather, you know, you develop a certain way of thinking about things when you get exposed to all these little bits and pieces and begin to understand how to think about the world.

8:33I do think, and I want to get back to, of course, AI and what you're doing now at InCitro, but I do think that having an eclectic background does allow for original thinking. And you are somebody who has had a MacArthur Genius Grant. That impresses me because I know two other people who've had it. One is a novelist, Edward P. Jones. Other is a jazz musician, Mary Halverson. And here you are, you know, to put you on the spot a bit. How do you define genius? You know, I always felt uncomfortable when that word was applied to myself because I'd always had a very aspirational definition of what genius means.

9:14And, you know, that was Albert Einstein or Leonardo da Vinci. It wasn't me. And when I did get that MacArthur Award, I actually felt very humbled and unworthy. And one might even say that, and I've even said that to the people who gave me the MacArthur Award, that much of my career journey following the MacArthur Award was an attempt to kind of pay back to prove myself as having deserved that. No offense, that sounds very female imposter syndrome. Well, maybe it is, but that's certainly how I felt. And I think the Coursera experience was definitely, I felt the need to go into something. Which you co-founded.

9:52Which I co-founded. for those who don't remember? Yes. And that was an attempt to do something that was sufficiently ambitious and world-changing, that it deserved that recognition. And so I don't know what genius is, but I can tell you that one of the things that I consider to be my superpower trying to avoid that female imposter syndrome is that ability to connect the dots across different disciplines and see connections that are oftentimes maybe obvious in retrospect, but weren't obvious at the time. And I find that that ability to transfer ideas from one discipline to another and create a synthesis that did not exist before is where a lot of really interesting ideas often emerge.

10:44And you can think about what we did in Coursera as the kind of bringing the web in a truly meaningful way to education. And you can think of what we do in Citro as the blend of biology and computer science. And in fact, it's even in the name. In Citro is the synthesis, the deep integration of in silico, which means in silicon in the computer, and in vitro, which means in the lab. And that is kind of the ethos of the company. Oh, I had no idea as an in vitro baby. I love that. Oh, that's lovely. Yes. So I'm wondering, Daphne, what was it that was the spark that made you kind of transition from academia into being a founder?

11:27And how did you feel? Were you scared? Because I would have been terrified. I was absolutely petrified. I was not only had I never founded a company, my career journey was at a company. I was an academic through and through. My parents were academics. I was convinced I would retire as an academic, be a professor emeritus like my father. And then, as I mentioned, that MacArthur thing happened and really helped put a point on, I wanted to make an impact. I wanted to do something that was directly changing the world as opposed to publishing papers and hoping that someone will read them and do something with that or, you know, graduating students that go off and do great things.

12:11And that's amazing. Especially Stanford. That's the hub. You also grew up in Israel, which is also a hub of biotech and entrepreneurship and tech. And you did college and high school simultaneously. I think you graduated 17 from university, correct? Yes. So talk a little bit about what intrigued you there, because computer science wasn't really as much of a discipline per se when you and I were growing up. No. And for women especially, I might add. I think that's really impressive. For sure. No, I got into computers actually when we were in sabbatical, when my father was here on sabbatical in the United States, and I was 12 years old.

12:52And I was fortunate to go to a high school that had a computer lab. And it's, you know, computer lab. Like Bill Gates. Well, yeah, kind of like Bill Gates a few years later. So we had somewhat better computers than he had access to when he was a preteen. But that ability to kind of tell the computer what to do and then it does stuff, I found fascinating. And it really fit in with my kind of wish to create models of complexity. And that theme has been with me throughout my career journey of taking complex sort of gnarly things and create models that lend clarity. And I hesitate to say that computers understand because computers are not conscious.

13:39They don't understand. So to me, a computer can be said to understand when it's making predictions that are reasonably accurate, which comes back to the comments that I started with. So creating models that are predictive of real world systems, to me, was just an incredibly fascinating endeavor. I think that's so important that they don't understand because right now I think a lot of people are using AI like it's a real human with emotions and feelings. And my friends tell me they use it for therapy and I'm like, I don't know if we should be doing that. Don't fall in love with your operating system.

14:12Yeah, exactly. How are you advising people? I mean, look, if it's useful for you as a human to speak out what you're feeling and have something that prompts you in an intelligent way to bring things out, I think that's great. And I think it's fine to use a computer as a therapist. But I think it's important to realize that there is no emotions on the other side. There's no sentience. There's no consciousness. It is effectively a really sophisticated mirror that allows you to work your way through, for example, complicated feelings. Thus making it particularly well suited, I think, to things like drug discovery, right?

14:51Tell me a little bit about the genesis of In-Sitro and what it is that you felt you could accomplish, especially in a world of big pharma where it takes years, decades in some cases, to get to the kinds of drugs that you're currently developing for diseases for which there's currently no cure. Yeah, so my sort of journey into drug discovery came out of my actually early academic work at Stanford where I'd been working at that intersection of biology and computer science and machine learning for quite a number of years prior to my detour at Coursera. And when my time at Coursera was over, I realized, because I'd actually been at Coursera when the machine learning revolution in 2012 started, in 2016 I lifted my head over the trenches and I was like, wow, machine learning is changing the world, but not having as big of an impact in life sciences.

15:47And I felt like I could potentially make a difference in that. And so I spent some time at Calico, which is a great company, and learned about drug discovery for the first time. And I was like, wait, that's how we make medicines? I mean, this is so, what's the word, serendipitous. Drug discovery is all about. I thought you were going to say antiquated. I was thinking the exact same thing. Fair enough. It's both. and this is not you know I don't want to this is not to say anything bad about Calico specifically the whole process is like a scientist has an idea and based on some intuition reading some paper whatever does some bespoke experiments to kind of try and prove that out and then you push that into the clinic and it fails over 90 % of the time and 90 % of the time is when you put that is from the time that you put this into the clinic.

16:43It's sort of, that's just pretty late in the process. Why is that? Because we are not able to predict, uh, upfront what the experiment. Your eureka moment could just be a flight of fancy. Yeah. I mean, you are making predictions based on experiments in often animal models that are just not human. You know how many mice we've cured of cancer? Many, many, many mice. If only mice were the target audience, we'd be in great shape. But we're not able to do experiments in humans until the very end of the process, as one should in the clinical trial. And we're not able to make predictions up front of what will or will not have a therapeutic impact, will make people better or not.

17:27And so what I wanted to do was to bring the power of machine learning on the one hand, And equally importantly, the power of modern day data collection from humans and different types of human derived systems bring that together to make as accurate a set of predictions as one can around whether a drug will have a benefit to a particular patient or not. And that requires having the right AI and having the right data to feed the AI. And those together are what is allowing us to create an engine that turns and makes more and more predictions about therapeutic hypotheses that we think are both novel and more likely to have success in the clinic.

18:10Tell us about this data because I'm curious where you're getting that, where it's coming from. What's that like? So data is the fuel for AI. I think everyone understands that now. The thing that has given us that chat GPT moment is the incredible amount of data that is available on the web in terms of text and images. What has given us alpha fold is the significant amount of data about proteins folding. And so in order to give us that next moment, that next chat GPT moment, we need the right data to feed the AI. And so at InCitro, we actually put together two different types of data that we think are synergistic and complementary.

18:43One type is data from humans, where we measure a human system and we understand how changes in human genetics, the DNA that makes you and me different, has impact on traits that are relevant to health and disease. And we're able to use measurement technologies like imaging of whole bodies or different organ systems, omics from blood and tissue, to really get a picture of human physiology and understand how it relates to genetics on the one hand and the clinical outcomes on the other. Which is great, but you can do experiments in humans, and you do need to experiment with your hypotheses. So the other half of what we do is we print massive amounts of data in our lab.

19:26And this, of course, is data in microcosm systems and cells. But those cells are allowing us to kind of make interventions and see what a particular change in a gene does in this case to cellular. Almost like scenario planning. Exactly. Cellular systems. And those together can triangulate us on hypotheses that are actually going to work. Let me pause a second on the human data on two fronts. One, I was very early, me and my family, I forced them to spit in test tubes and send it off to 23andMe. Me too. Still too scared to do that. And the promise, because the promise of how DNA collectively could unlock drug discovery excited me.

20:05I'm curious why those models, without making it too personal, have proven to not be as promising, it seems, as I don't, maybe it's just the business model of some of them. But then I wonder, secondly, about the trust issue. I trust spitting in a test tube. Kristen may not. But also there are a lot of populations that have been reluctant to share their data. Do you see any pain points around genetic data, privacy, trust, and just because it's so promising clearly. If everybody in the world contributed their information, how lucky we would be, and they're not. Yeah. So I think that is something that has been done poorly in some cases and very well in others.

20:52So let me point to a success story, which is something that we leverage a lot, which is resources like the UK Biobank, where I think it was done as well as one can possibly hope for. They got half a million people to agree to have their DNA measured, as well as a whole bunch of other what we call phenotypic measurements, which are various traits. and they took people and imaged them. They collected blood samples. They took all sorts of anthropomorphic measurements. And actually, it turns out that those together are way more powerful than either one in isolation. When you have both the genetics and the phenotypic data, this is where you really start getting insights about what genetics does to actual clinical outcomes.

21:35And they did it in a way where people were consented. They knew what they were doing, and they are excited to contribute to medical research. And I think if this was something that was done more broadly, I think we would be, to your point, in a much better situation. And it is possible to do it in a way that is privacy-preserving. You create research environments that are protected. The data are anonymized. You can't download the data. And all of those things are solvable engineering problems. What one needs is the recognition of just how valuable this is from a medical research perspective, not just medical, by the way.

22:15The UK Biobank gave us a tremendous amount of insights about non-medical situations like the impact of various other environmental factors, for example, on human health and other human outcomes. And you create that resource. It's just a gift that keeps on giving. And I wish other governments were, especially ones, to your point, that have access to more diverse populations, because the UK Biobank was very Eurocentric, would make that effort and create these resources that would allow us to address human health across different ethnic backgrounds. I think diversity was my next question because going along with the ethics of it, having a diverse data set really worries me, especially when it comes to health care.

22:57How are you at InCitro ensuring that all the data sets are diverse and inclusive? So I think it's important to distinguish between data that's used for scientific discovery and data that's used for the delivery of care. And we've seen that there is significant differences in terms of, you know, for instance, diagnostic tests that are developed for one population being less accurate than tests that are developed for another. When you're doing basic biology discovery and what you're coming to is the recognition that this gene is a significant causal driver of disease, of a particular disease, the prevalence of the disease might differ in different populations.

23:37but typically the underlying biology is consistent. And so that same driver is likely to be relevant across populations, even if the prevalence of the disease might be more relevant in some populations than others. So I think it's less of a concern when you're doing drug discovery than when you're doing the delivery of care. However, what we've also seen, which is why diversity is, I think, really important for all of us, is that as you get insights about new populations, new biologies emerge that might be kind of not just not there in another in another population. And so you might uncover even more new things by having access to more and more different diverse populations.

24:20So I wish we had that. Can you talk a little bit about how you chose ALS, fatty liver disease? What was it that made you target the diseases that are currently on your radar screen? So first and foremost, the lens that we begin by applying is how significant of a disease is this and to what extent is there what we call unmet need? Because there are certainly a lot of diseases that are reasonably well treated right now. And there are companies out there that make, you know, me too, me better drug for those diseases. And, you know, there is value in that. But that is not what we wanted to do. What we wanted to do was to uncover new intervention points that make a really significant difference.

25:06So ALS, to take a very concrete example, has absolutely nothing. There are four approved drugs. They extend lifespan by two to three months. The patients have a three to five year lifespan. I mean, it's worse than most cancers, to be honest. And patients, an excruciating death, they basically, they start losing, their motor function to the point that eventually they just are unable to breathe. And that's just a horrible way to die. And so we believe that in the work that we did in ALS, we have uncovered a novel master regulator, actually, we think maybe even more than one, that modulates a number of key proteins that we know are implicated in the pathophysiology of ALS.

25:58And so now it's early days, we haven't put this in a human. And unfortunately, ALS is one of those diseases where the animal models are particularly poor. And so our proof of success will only be when we put this in a human, which by the way, is true in general, but maybe even more true in ALS because of the lack of good animal models. But we are so excited about the experiments that we've done across multiple different ways of measuring ALS biology and finding that our targets modulate these different aspects of the disease. So we're really excited about that. And we keep looking for other examples of that.

26:36Now, the other element, of course, is one is unmet need. And the other is, do you have a unique differentiated ability to do something about this disease because there's a lot of diseases out there that have unmet need, but what makes us the best people to tackle that? And so what we look for are diseases that manifest in meaningful ways in the kind of high content data that we have access to on the one hand and have a sufficient genetic basis on the other so that we know that genetics is a sufficient driver of the disease so we can find those genotype-phenotype connections. I mean, if it's a disease that's driven by smoking, I mean...