EP 83: Daphne Koller, Founder & CEO of Insitro - Integrating machine learning and biology at scal...

Thank you, Patrick. Glad to be here.

So I started my career, actually if you go all the way back to my undergraduate days, as a mathematician, and initially was very much enamored by the conceptual beauty of mathematical frameworks for things that are challenging to model in the world. And so my work followed the thread of let's create these really elegant frameworks, and my thesis falls into that category. And at this point, if I look back, I mean, there's some really deep mathematical theorems there and some really interesting insights. But a turning point for me actually came at the end of my PhD when I went to Berkeley to do my postdoc, and my postdoc mentor, St. Russell, took me to lunch and asked me, 'So if you were to take your PhD thesis and I were to give you three exceptionally talented undergraduates to help you with the implementation, which part of it would you actually implement?' And I was dumbstruck because not only had no one ever asked me that question, the answer to that as I reflected on it was none, because it was just conceptual and it wasn't really good for anything. And I think that put me on a journey to really start to think much more deeply about how I make sure that the work that I do has an impact in the world. And that journey took me on a progression of going from the more theoretical to the more and more applied at every step, starting from work going from that thesis to working on machine learning, which at least in principle was more applied, then to applied machine learning, not just methods development but actually applying them to vision, robotics, natural language processing. Then moving from there to something that I felt was even more applied, which is how does one take machine learning techniques and apply them to biomedical data, which is a transition that took place around the late 90s, early 2000s as data from patients and from biological data were starting to come available. And then from there, I became more and more applied in how I applied machine learning to biomedical data. And ultimately, that actually was what precipitated my departure from Stanford in going out to found Coursera, because it was something that was not an expected part of my career journey, but it was an incredible opportunity to have a lot of impact in a very short amount of time, and I couldn't pass that up. And so in some sense, there was this journey that just kept going.

So actually, if I could go a little bit back to the reason to leave Coursera and go to Calico, because that was in some ways the beginning of that journey. That was because I'd actually been at Coursera for just about five years from the time that we started incubating it at Stanford as a company, as opposed to as an internal Stanford project. And so I looked at where Coursera was at the time, and it's a wonderful company and I'm incredibly proud of what we have accomplished with building that, but it was primarily a company where the impact was derived from having great content provided by the great universities, or at this point the broader set of organizations, and there wasn't a very significant role for technology and certainly not for science. And so that was to me an opportunity to ask myself where could I personally have the greatest amount of incremental impact. And I felt that Coursera without me would continue on a great trajectory with great people. But what I saw around me, and this was in 2016, four years after the machine learning revolution, and seeing that machine learning had been transforming the world around us but not really having that much of an impact in the life sciences. And when I asked myself why that is, it's because there's not a lot of people who speak both languages. By and large, machine learning people don't know very much about biology at all, and most biologists are certainly not trained in machine learning. And that's a chasm that makes it very hard to have a meaningful, deep application of machine learning methods to life science data, with a handful of people who can really make that connection. And so I felt that the impact that I personally could have would be much larger if I went and built something that probably would not get built if I or someone like me didn't do that, versus just continuing on the trajectory that we were on with Coursera, which would likely continue even if I left. And so that was at the point when I didn't quite know how that impact would translate. I had been out of both machine learning and science for about five years, so I didn't even have an idea of what I wanted to do. And I was fortunate enough to connect with Art Levinson, who was the CEO of Calico, previously the CEO of Genentech, and so that was an opportunity to work with a truly great leader, and I was really inspired by that. And so I went and spent 18 months at Calico and realized that ultimately I was not a biologist, I was still a computer scientist, a machine learning scientist. And so what I wanted to do was to build a platform that would help solve biological problems in a very different way, and specifically transform the way, hopefully, in which we design and develop medicines. And it didn't make sense to build that within a company that focused on a specific biology, the biology of aging. And so that led me to depart after about 18 months and form Insitro, which is four years old this week.

So I think the idea behind the platform that we're building is that our understanding today of human disease is incredibly out of date. The taxonomy that we use to classify diseases is based on very coarse-grained clinical manifestations, often times filtered through the subjective lens of a patient and the clinician. And we categorize diseases based on this coarse-grained similarity of these clinical symptoms that has very little to do, in many cases, with the underlying biology of the disease. This is a realization that has become very clear as we look at oncology, which is the area in which we've had enough molecular data from enough patient samples to realize that breast cancer is not a thing, it is multiple things. And at this point, we have different subtypes of breast cancer, each of which has a very different treatment, and it's what allowed us to move away from chemotherapy, which is truly the lowest common denominator of trying to address cancer and generally ineffective except in the very earliest stages and has horrible consequences, to something that is much more of a targeted therapy that treats the underlying causes of the disease with different treatments for different patients. We've not really had that level of clarity in the broader set of diseases, and that is partially because the data that we've been able to acquire from patients up until quite recently has been quite limited. In cancer, by sort of nature of the disease, you end up biopsying samples and extracting examples from humans, and at the same time, cancer cells because they're so robust are also very easy to culture in a lab, and so one can do experiments on those cancer cells and intervene and assess different treatments. It's a lot harder to do either of those in the context of most other diseases. That situation, however, has changed in the last few years with a growing amount of both high-content data that are objective and acquired from human patients, whether it's in some cases less invasive biopsies, in other cases it's other types of measurements like MRIs of different kinds, PET scans, and so on. So we finally have the ability to look at the biology of actual patients and potentially identify patterns there. And on the other side, there's been an incredible set of developments in terms of life science tools that enable us to interrogate the biology of these diseases in the lab. And that includes the development of cell-based systems that are much more similar to human natural cells, using this concept of induced pluripotent stem cells, which allow us to take any cell from any one of us and transform it to this stem cell status and then create neurons that have our genetics or liver cells that have our genetics. So we can actually start to look at what those cells look like with different genetic burdens of disease, and we can further modify those cells using technologies like CRISPR in order to introduce disease-causing variants and see what that does. And finally, importantly, you can also measure those systems in again very high-content, very detailed ways using microscopy, using single-cell transcriptomics, and various other approaches that really allow us to gain a very deep and holistic view of what disease looks like at the cellular level. So with that flood of data, what we have available to us is now an opportunity to really redefine the taxonomy of disease, and then for the appropriate subtypes, because we know what disease looks like at the biological level, understand what are some things that one can do to protect against or revert disease. And so I think this opens the door to exploring a range of different diseases that maybe share some characteristics, like the existence of genetic drivers. So it's not something that is derived from a lifetime of wear and tear and is really hard to model in a lab. So you need to have certain characteristics in order for the technology stack that I'm talking about to be applicable, but with that set of tools, you can interrogate a whole range of diseases. So what we have elected to do at Insitro is really industrialize that process, so create a technology stack that allows us to unlock not every disease, it's not going to be universal, but unlock enough of them that we can have a much more reliable and less artisanal and stochastic process for uncovering new treatments against those diseases. So we're still trying to figure out which are the right diseases for this type of intervention. The first four years of the company were in many ways an exploration to figure out which tools are the ones that are most valuable, most robust, most applicable, and where they are applicable. Not sure if NASH in fact is the right place to start. We started there at the beginning of our journey. I don't know that that is the right place where we would want to first deploy this technology, but we're now getting to a much greater level of clarity on where we can really deploy this and hopefully start to really reap the fruits of the platform that we've built.

So one thing that we've learned is first, we've learned that causality is really important, which I think was maybe fairly obvious in retrospect, although you'd be surprised just how many drug programs have been launched based on purely observational correlates that have no causal threads anywhere there. So this is maybe obvious. As an example, maybe controversially, there is some genetics that has supported those, but those exist in a very, very small segment of patients and have been extrapolated much more broadly than I think is justified. And maybe that has led to some of the incredible number of failures that have occurred in all comers patient trials. By the way, one important lesson learned here is the importance of selecting the right patient population for whatever intervention. And I think a lot of the mistakes, a lot of the failures that have come about across multiple drug trials for complex diseases have been that we have in fact tried to deploy interventions against a patient population that is much broader than it should be, and effectively trying to cure what are multiple diseases with one drug, which is the thing that really very rarely works. But sorry, I digress. So we learned that causality is critical. We learned about the incredible amount of information that exists in high-content data, in subtle patterns that people just can't see. That has turned out to be something that we've repeatedly shown across multiple data modalities, both cellular and human, that there is way more information in histopathology than even experienced pathologists, for example, can discern. Same is true for cellular images, and so on. So that's one thing that we've learned. We've learned about, as one considers the tools and protocols that we use in the lab, that if you want to apply machine learning to high-content data, the reproducibility and error parameters of the tools that you're applying are absolutely critical, because the more advanced the machine learning that you apply, the easier it is for the machine learning to get confused by things that are artifacts and batch effects and latch onto them and make incredibly accurate predictions that have absolutely nothing to do with the underlying biology. And so that has actually driven the choice in many cases in our lab of methods that are really where the selection is very much driven by how reproducible they are, how robust they are, how much can we protect against artifacts and batch effects. That also has driven a very big focus on automation, because one of the biggest sources of batch effects is when you have different humans doing the same task, or even the same human doing the task on two different days when they woke up on the wrong side of the bed or whatever. And then similarly, on the human cohort data, which is the other big data source that we use in our discovery work, really focusing on data sets that are large and well-curated and well-collected. And in that respect, the work that we did with the UK Biobank and many others, I mean to me the UK Biobank is an incredible resource that has driven so much insight across so many different aspects of human biology, human disease, and even just epidemiology, because it is so well-collected and well-curated. So the availability of such resources, I think, is absolutely critical to scientific progress.

So I think the UK is probably the most advanced of anyone, with maybe a couple of exceptions. I mean, Finland, Iceland also have some incredible resources in this respect. Some other countries are starting to deploy resources towards that. And I think it was just a combination of having the National Health Service on the one hand, which of course allows the system to track people throughout their lifetime, and so if people agree to participate in research studies, there is a wealth of data available on them that can be brought in with their consent, of course. The other is just a number of people within the UK data ecosystem who had remarkable foresight about the value of creating these very well-curated, very carefully collected large population cohorts, and equally the willingness to create that as a resource not to a small tiny subset of researchers but really to anyone who is able to do good science and is willing to comply, of course, with the constraints of the cohort, which is that insight should be introduced back into the cohort so that they benefit everybody, and the protection of patient or individual privacy. And that I think was just an incredible amount of foresight, and I really hats off to the people who put that in place. When I look at similar efforts elsewhere, I find that they often lag behind on one or the other of these dimensions in a way that I think is unfortunate. First, they often times don't make the same investment in really collecting very detailed, dense phenotypes. A lot of the efforts that I've seen are just electronic health records, maybe with some amount of genetics, but not the very detailed phenotyping that exists in the UK Biobank. Many of them don't do the longitudinal follow-up that the NHS provides, and many of them don't provide the almost universal access to qualified researchers that the UK Biobank has done. And I think that's a real shame, because one of the things that has really happened in the last, I would say, 20 to 30 years in science is the democratization of science. It's the fact that people all over the world, by having access to computers and technology and so on, can come up with new ideas and new forms of analyses. You don't have to be at a top academic center in order to have a really great idea of how to extract value from these resources. And if you shut those people out, then I think it just limits our ability to provide benefit to patients.

I agree it was a dilemma and not an easy decision. There is a lot that, if you go to one of the big pharmaceutical companies for example, there's a lot there. Certainly they have a lot of money. They also have, in some cases, a lot of relevant people already in place, and they claim, although sometimes with much greater confidence than is deserved, to have a lot of data. And while it's true that they do have a lot of data, the data is oftentimes fragmented, not harmonized, not harmonizable because it was collected by different research programs with very different approaches. And coming back to my earlier comments about batch effects and artifacts, it's often times difficult and perhaps impossible to actually create a single unified data resource from that that is actually valuable. So why did I not choose to do that? I think because fundamentally, I think that if you're looking to do something that is completely different, it's difficult to do it from within a company that has a very significant investment in doing things in the old way, and groups of people for whom it's in some sense their careers are at stake, their livelihoods at stake, to completely change the paradigm. And you are basically fighting that cultural headwind all the time. And I think that one of the biggest things to really struggle against is what I call cultural inertia. It's just the unwillingness of an organization to transform the way that it does things. And specifically, what I think is really challenging in a lot of big pharmas is to take an organization that has been led, where the research has really been led almost entirely by life scientists, and actually give the data scientists, the machine learning scientists, an equal seat at the table versus having them be kind of a secondary service organization. Because people don't think of data scientists like that. It's always been the way that the life scientists are calling the shots. And I think that the parity, the equality between those two groups of people, is absolutely central to this new paradigm of how we are going to do things in a very different way. And maybe to take an analogy from the broader world out there, if you look at the tech-enabled giants that have emerged in other parts of the economy, whether it's in media, whether it's in commerce, whether it's in transportation, the really successful companies have almost never emerged as transforming the incumbents. You know, Amazon did not emerge from Walmart, Google did not emerge from the Yellow Pages. And so I think it's useful sometimes to just take a blank sheet of paper and say, how should I build this in the right way?