# Ep79: Daphne Koller on Machine Learning for Drug Discovery

- **Show:** The Long Run with Luke Timmerman
- **Published:** 26 Oct 2020
- **Source:** https://soundcloud.com/longrunpodcast/ep79-daphne-koller
- **Transcript:** https://podcasts.opennash.com/shows/the-long-run/ep79-daphne-koller-on-machine-learning-for-drug-discovery-c26266/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Daphne Koller argues that drug discovery can improve if a company builds its data and its machine learning together from the start. insitro makes human-cell experiments and pairs them with human genetics, so algorithms learn from data designed for them. She says this matters more than a clever algorithm run on poor data, and that even modest gains in predicting clinical outcomes would be valuable.

Koller's path runs from math prodigy to machine learning professor, Coursera co-founder and Calico, steadily favouring work that is useful over work that is merely elegant. The tension is between that ambition and biology's complexity. She rejects an all-knowing predictive model, aiming instead for better target choices and patient subgroups. She says that needs a culture where biologists and computer scientists are equal partners.

## Key insights

1. **Data quality beats algorithm sophistication** Koller says a good algorithm on a poor data set loses to an okay algorithm on an excellent one. So insitro builds an engine to produce data designed for machine learning, instead of reusing data collected for other purposes.
2. **Genetics plus stem-cell experiments train predictions** Human genetics shows what natural variation does to people, and lab-grown human cells let the company test its own interventions. Machine learning links the cell measurements to clinical outcomes, which answers the problem that human intervention data usually arrives only in late-stage trials.
3. **Koller chose a startup over big pharma** She cites cultural inertia: big drug companies are science-first, with data scientists in a siloed group. She wanted biology and data science to share decisions from day one, which she judged easier to build from scratch than to retrofit.
4. **Better target selection means patient subgroups, not blockbusters** Koller expects future drugs to work for genetically defined subsets, as in precision oncology. She says PCSK9 started with rare variants yet helps a broader group, and she expects disease categories in neuroscience to be redrawn within 10 to 15 years.
5. **Her career shifted from elegant to useful** A Berkeley postdoc mentor asked what three undergraduates could build from her PhD work, and she had no answer. That pushed her toward applied machine learning, then biology, then companies, so her work could reach patients and learners.

## Questions this episode answers

### What is the difference between machine learning and artificial intelligence?

Koller says artificial intelligence is the task of recreating human-like intelligence, while machine learning is a set of methods that learn complex tasks from data. Many AI problems are best solved with machine learning, but not all, and she says insitro's tasks are often ones humans cannot do well.

### What does insitro actually do to find new drugs?

It combines human genetics with experiments on lab-grown human cells made from stem cells, then uses machine learning to link what happens in the cells to patient outcomes. The aim is to predict which targets and patient groups are most likely to respond to a drug.

### Why did Daphne Koller start a new company instead of working inside a big drug company?

She says big pharmas were built as science-first companies, with data scientists in a small separate group. She wanted scientists and machine learning experts to decide together from the start, and she thought that was easier to build fresh.

### Does Koller think AI will solve drug discovery completely?

No. She says an all-knowing predictive model is unrealistic and keeps her ambitions reasonable. Because drug success rates are so low, she argues even a modest improvement would greatly help the industry's productivity.

### Which diseases is insitro focusing on, and why?

Liver disease and neuroscience. The liver work came from a partnership with Gilead on NASH, a fatty liver disease. She chose neuroscience because good cell-making methods exist there, the diseases are strongly genetic, and patients' unmet need is large.

## Facts

- **Daphne Koller** [0:02:50]: "I think artificial intelligence is the task of recreating intelligence as we know it in people, whereas machine learning is methodologies that learn to perform complex tasks by learning from data."
- **Daphne Koller** [0:04:18]: "I finished high school at 16 and college at 17 and had always been interested in disentangling complexity."
- **Daphne Koller** [0:05:19]: "My father was a professor, actually a professor of biology as it happens, which is one of the reasons why I steered as far away from biology as I could as a child"
- **Daphne Koller** [0:08:55]: "I was a military intelligence analyst. So it was actually kind of fun because I got to, it was my first encounter with Bayesian reasoning."
- **Daphne Koller** [0:14:53]: "AI has gone through peaks and troughs multiple times in its history, generally caused by the fact that certain groups of people made hyperbolic promises about what artificial intelligence would be able to accomplish"
- **Daphne Koller** [0:17:08]: "But no, it really sort of took off as a discipline within computer science and engineering in the early 90s."
- **Daphne Koller** [0:18:49]: "if I gave you a group of like three smart undergrads to work with you and build a system that would be useful based on the work that you did, what would you have them do?"
- **Daphne Koller** [0:24:05]: "how do you use gene expression data to uncover gene regulatory networks. And that was, I think, work that was quite influential. The paper is still very highly cited."
- **Daphne Koller** [0:25:48]: "my sabbatical in a biology lab at UCSF with Jonathan Weissman, who's one of the leaders in the functional genomics CRISPR field, incredible scientist."
- **Daphne Koller** [0:28:02]: "had worked on some of the earliest machine learning-based analyses of cancer biopsy images under the microscope."
- **Daphne Koller** [0:32:18]: "from the very beginning of the journey, which was in the fall of 2011, which is when we launched the first MOOCs while still at Stanford, up until the time that I left, it was almost exactly five years."
- **Daphne Koller** [0:31:24]: "So I decided and I ended up resigning what was a chaired, tenured faculty position at Stanford, because I felt like I really needed to see that Coursera journey through."
- **Daphne Koller** [0:34:24]: "it was a very memorable meeting because he said, look, I can answer your question, but let me tell you about Calico."
- **Daphne Koller** [0:35:14]: "I wanted to build a platform and it didn't seem to make sense to build it there"
- **Daphne Koller** [0:38:47]: "they really started in early 18 as I was kind of starting to transition out of Calico."
- **Daphne Koller** [0:37:53]: "So the first investor was Jim Tannenbaum. And then very soon thereafter, Krishna Yeshwant from GV and Bob Nelson from Arch."
- **Daphne Koller** [0:38:07]: "So we ended up with five major investors, actually, in that initial group, with Third Rock being the fifth."
- **Luke Timmerman** [0:00:41]: "as it raised a $143 million Series B financing in May 2020, led by Andreessen Horowitz, and which included earlier investors like ArchVenture Partners, Foresight Capital, GV, and Third Rock Ventures."
- **Daphne Koller** [0:39:15]: "if you ended up giving me the tradeoff between the best machine learning algorithm on a crappy data set versus an amazing data set with kind of like an okay machine learning algorithm, the second will win every day."
- **Daphne Koller** [0:42:32]: "the data scientists in most of these companies are usually this kind of smallish siloed group that is sitting in some corner somewhere."
- **Daphne Koller** [0:47:12]: "They are 70 % life scientists, but they know how to program in Python and analyze their own data."
- **Daphne Koller** [0:52:38]: "where we can take cells from you or me, convert them into pluripotent status. stem cells and then differentiate them into a Daphne neuron or a Daphne hepatocyte or any one of the set of lineages that the community has developed differentiation protocols for."
- **Daphne Koller** [0:55:25]: "It came out of the partnership that we have with Gilead on NASH, non-alcoholic steatohepatitis."
- **Daphne Koller** [0:56:03]: "the heritability is you know 30 40 50 70 percent for instance in autism so it's very high but also very complex"
- **Daphne Koller** [1:00:24]: "a BRCA1 breast cancer is very different from a HER2 breast cancer and are in fact treated with completely different treatments right now."
- **Daphne Koller** [1:01:42]: "A great example of that is the poster child here, which is PCSK9, where it was discovered in a rare familial variant in hypercholesterolemia on the one hand, familial hypercholesterolemia, familial hypocholesterolemia on the other, sort of showing this beautiful dose response."
- **Daphne Koller** [1:02:25]: "I think over the next 10 to 15 years, we'll actually end up with a very different disease taxonomy than we have today, especially in the neuroscience area."

## Transcript

**[0:00:01] Luke Timmerman:** Welcome to The Long Run. This is a podcast for biotech adventurers. I'm your host, Luke Timmerman. Today's guest is Daphne Koller. Daphne is the CEO of South San Francisco-based Insitro. The company is seeking to develop a new platform for drug discovery that leans on a combination of wet labs and machine learning algorithms to spot new biological targets for drug discovery. Artificial intelligence and machine learning has been stirring imaginations in biopharma and generating more than a little hype for a few years.

**[0:00:41] Luke Timmerman:** Smart people who know how complex drug discovery is are pining for something to make it less risky, less time-consuming, and less expensive. insitro represents one of the prominent wagers in this AI for drug discovery category as it raised a $143 million Series B financing in May 2020, led by Andreessen Horowitz, and which included earlier investors like ArchVenture Partners, Foresight Capital, GV, and Third Rock Ventures. Daphne comes to this challenge with a fresh set of eyes.

**[0:01:14] Luke Timmerman:** She was a math prodigy as a kid. She became a tenured professor of computer science at Stanford University. Then she became a successful entrepreneur, co-founding Coursera, the online learning platform. But when it came time for her to find challenging new problems to solve intellectually and important problems to solve for humanity, she decided that tackling disease was the place to be for her next act. We talked about Daphne's journey and some of the key aspects of her strategy to make insitro succeed in this new frontier.

**[0:01:48] Luke Timmerman:** Now, quickly before we get started, do you enjoy the Long Run Podcast? Maybe you'd like to raise awareness of your company, your organization, or your services with a high-powered crowd of entrepreneurs and venture investors who listen to this show. Tell me about your company and why it could be a good fit as an advertiser on the long run. Luke at TimmermanReport.com Now, please join me and Daphne Koller on the long run. Daphne Koller, welcome to the long run.

**[0:02:22] Daphne Koller:** Great to be here, Luke.

**[0:02:23] Luke Timmerman:** So, Daphne, I'm going to throw a surprise question at you. It's kind of a nerdy one about terms. Machine learning, artificial intelligence. People throw these terms around a lot interchangeably in different settings. You describe what you're doing at Insitro as machine learning for drug discovery quite often. How do you think of that term and why do you use that term as opposed to just AI for drug discovery?

**[0:02:50] Daphne Koller:** I think it's a great question, and my view on this is not necessarily representative of that of everyone. Some people will tell you that machine learning is a subset of artificial intelligence, and I actually view it quite differently. I think artificial intelligence is the task of recreating intelligence as we know it in people, whereas machine learning is methodologies that learn to perform complex tasks by learning from data. Now, it turns out that the right solution for many AI problems is by employing machine learning, but there are AI problems that don't involve machine learning.

**[0:03:29] Daphne Koller:** And maybe more relevant to what we're doing is that the problems that we're solving are not generally ones that people actually solve well. So we're creating, in some sense, a different type of intelligence, not what humans do, but in many cases, tasks that humans cannot do.

**[0:03:47] Luke Timmerman:** Okay, okay. I'm going to come back to this. That's a good working definition for starters. Okay, so as you know, with this show, I like to start out with some of the personal background, like how you came to this place, where you could begin to think about these hard problems with machine learning for drug discovery. So I know you were born in Jerusalem, grew up in Israel. Can you tell me a little bit about your experience growing up there?

**[0:04:18] Daphne Koller:** Absolutely. So I had always found school to be a somewhat less than scintillating experience, at least the schools that I got to go to. And so I ended up at a relatively early age, time sharing between sometime at school and sometime in college so that I ended up finishing almost contemporaneously. So I finished high school at 16 and college at 17 and had always been interested in disentangling complexity.

**[0:04:54] Luke Timmerman:** Well, Daphne, let me back up just a second before we get there. Growing up, I mean, you're something of a child prodigy. And math was kind of your first love, right?

**[0:05:05] Daphne Koller:** Yes, because it was so conceptually elegant.

**[0:05:10] Luke Timmerman:** Uh-huh. Uh-huh. How did you, what did your parents do for a living and how did they nurture this, you know, budding intellect?

**[0:05:19] Daphne Koller:** So my parents were both sort of academics. My father was a professor, actually a professor of biology as it happens, which is one of the reasons why I steered as far away from biology as I could as a child, because kids always want to do something different from what their parents did. My mom was an English teacher and they, I think, didn't quite know what to make of me. So when I asked for math books as gifts as a child, then my father went out and bought them.

**[0:05:46] Daphne Koller:** But it wasn't like they pushed me, which is what a lot of people think, that somehow you ended up doing this kind of weird trajectory that I had because the parents pushed you. That was definitely, from my perspective, a pull rather than a push. and I was actually the one who came to my father when I was 13 and said hey dad I want to go to college in parallel with high school and he said no and I said no really please and so he helped facilitate an introduction but everything else was just stuff that you know I just drove that process.

**[0:06:16] Daphne Koller:** Why was he saying no? Well I mean you don't usually want to send your 13 year old to college, especially when in Israel, a lot of the, most of the college students are actually post-military service. So your average college student in Israel is in their mid twenties.

**[0:06:34] Luke Timmerman:** So I was way younger than everybody else. A real social, I mean, misfit in a way. I mean,

**[0:06:44] Daphne Koller:** well, I, you know, I, I don't want to say misfit because that suggests sort of some kind of sort of a dysfunctional relationship. And I think it was kind of like I was the kid and people kind of cherished me as almost like a mascot, which is, you know, it was fine. And I had a good time with that. And I was, you know, I was actually a teaching assistant. In fact, the teaching fellow when I was 17, teaching a class of mostly people in their late twenties in an advanced database class.

**[0:07:17] Daphne Koller:** And that was fine. And everyone treated me with respect. So I don't

**[0:07:20] Luke Timmerman:** think it was dysfunctional. So I didn't mean like misfit in that sort of sense. It's just like you, you, you're, you know, not with your peer group and that's social, that socialization part is, you know, so important to schooling, especially at that age.

**[0:07:38] Daphne Koller:** Yeah. So there were some people who Israel has this program where some 18 year olds actually get to attend college pre-military service. And I mostly hung around with those people. And so if I

**[0:07:53] Luke Timmerman:** was 16 and they were 18, it wasn't that big of a gap. Okay. So you enjoy math, you're seeking out these challenges. Did you encounter a teacher who like really helped inspire you or lead you down

**[0:08:12] Daphne Koller:** that path? I had a lot of really great teachers, but I think the more inspiring interactions that I had actually came later on when I was at Stanford as my doing my PhD in terms of people who really inspired my path, my PhD and postdoc specifically. Okay. Okay. We'll get there. So now

**[0:08:36] Luke Timmerman:** Now you graduate from college at, was it 18, you said?

**[0:08:42] Daphne Koller:** My bachelor's degree was at 17 and my master's at 18.

**[0:08:46] Luke Timmerman:** Okay. And this is at Hebrew University? And then did you go serve in the, perform your Israeli military service?

**[0:08:55] Daphne Koller:** I did. I was a military intelligence analyst. So it was actually kind of fun because I got to, it was my first encounter with Bayesian reasoning. How do you take lots of little pieces of evidence and put them together to create a bigger picture that makes sense? And it was actually a really stimulating intellectual exercise, very different from anything that I've had done before, but helped significantly reinforce my love of untangling complexity.

**[0:09:24] Luke Timmerman:** Interesting. So did you, you said intelligence, did you work on like signals, kind of like what we think of at the, like the NSA here?

**[0:09:33] Daphne Koller:** I shouldn't probably talk about my military service, but I will just say that I was an intel analyst in the sense you got lots of different pieces of information from different sources and kind of said, okay, this is one tiny little shred of evidence without context. It really doesn't make sense. And here's another one. And here's a third. How do you put them together and discard the stuff that doesn't fit and create a picture that then provides actionable insights to people

**[0:10:04] Luke Timmerman:** who need to make decisions. Okay. Okay. So this would have been something like what, the late 80s? That was in the mid 80s. Okay. Okay. And how long did you serve there?

**[0:10:15] Daphne Koller:** Three years. So I was an intel officer. So as an officer, you end up serving longer than the stipulated two years. I ended up serving closer to three. Was there any other, like,

**[0:10:26] Luke Timmerman:** important aspect to shaping your career, doing that kind of, you know, that service work, where you knew, like, eventually you're going to go to, you know, get a PhD or pursue some kind of career other than this. But, I mean, we talk about this in the U.S., right? This idea of, should there be compulsory service? And there's some arguments for it and against it. But

**[0:10:48] Daphne Koller:** what did it mean to you? Well, for one thing, it was a very maturing experience because the Israeli military, because it is not just a volunteer army, everyone gets conscripted and has to do the service. First of all, you get to encounter people from all walks of life. And for those of us like myself who were privileged to have a more sheltered upbringing with a school that was relatively homogeneous from a cultural perspective, it was really interesting to just have that exposure.

**[0:11:24] Daphne Koller:** The other piece is that because everyone joins the service approximately when they're 18, a lot of young people end up having relatively quickly roles of significant responsibility. And while I wasn't like, you know, at the top of the heap by any stretch, I did get to kind of have a little, you know, at the age of whatever, 19 or 20, this little unit that I was in charge of. And so that's not something that most college students in the US, for instance, get to do.

**[0:11:56] Daphne Koller:** And so I found it to be incredibly interesting, both in terms of just doing something that was entirely non-academic, unlike the work I'd done in school and university, and also just in terms of the exposure to a different mindset, both the leadership and the diversity of the people I worked with.

**[0:12:14] Luke Timmerman:** Interesting. So I saw something you had written down that you came to the U.S. July 4th of 1989, a rather memorable date. How did that come about?

**[0:12:27] Daphne Koller:** Just by chance. I mean, I was actually with my then boyfriend, ultimately husband, and we were supposed to come over together. We had this beautifully planned vacation that we were going to take in Europe after we both finished the service. It was supposed to be three months and then I was going to come and do my PhD. And he had a job lined up and about two weeks before that was supposed to happen, his future boss said, actually, we need you, called and said, we need you here now.

**[0:12:57] Daphne Koller:** So basically we both finished the service and we got on a plane

**[0:13:00] Luke Timmerman:** the next day and it happened to be July 4th. And this was to go to Berkeley? Stanford. Oh, Stanford. Okay. So it was my postdoc. Okay. Okay. So what were, what were you there to study at

**[0:13:13] Daphne Koller:** Stanford? Computer science. Straight away? PhD in computer science. Okay. Okay. My undergrad was in math and computer science. So that was the natural continuation. And what kinds of problems

**[0:13:27] Luke Timmerman:** were you interested in then in, you know, late 89 or the early 90s? So I started out, and I think I

**[0:13:36] Daphne Koller:** mentioned this earlier, being really enamored by the elegance of mathematical solutions. And so at the time, I was trying to think about how should people make decisions and how do we create decision-making environments that allow them to make better decisions? So I studied and worked on tools first from game theory related to multi-agent systems and then also in single-agent decision-making using tools from probability theory and decision theory. And it really was at that point for me, in a way that I feel a little bit almost embarrassed to think about right now, just how elegant a conceptual framework I could put together with sound, strong mathematical principles.

**[0:14:23] Daphne Koller:** And so I ended up doing some work on both of these fronts, both decision-making and multi-agent systems and the decision-making and single-agent systems in a way that when I now look at it, it's very elegant, perhaps, but it's not really good for much.

**[0:14:44] Luke Timmerman:** What was the discussion around the Bay Area, the field, around artificial intelligence in those days?

**[0:14:53] Daphne Koller:** This was right around the time of one of the earlier AI winters. So AI has gone through peaks and troughs multiple times in its history, generally caused by the fact that certain groups of people made hyperbolic promises about what artificial intelligence would be able to accomplish in three years, five years. and those promises at that point didn't really materialize. And the late 80s, which is when I started my PhD, was just coming off the tail end of one of the boom periods, which was the mid-80s, and then getting into one of the trough periods.

**[0:15:37] Daphne Koller:** And the truth of the matter is I didn't perceive what I was doing at the time as artificial intelligence. Artificial intelligence was at that point in many ways a somewhat squishy discipline. It was kind of like we're going to put rules upon rules and then some kind of behavior is going to emerge. And those systems were oftentimes quite brittle, very difficult to engineer. Unexpected behaviors would just emerge in ways that were totally hard to correct for.

**[0:16:07] Daphne Koller:** So what I was doing was really more mathematical modeling. And so people will often ask me, how did you get into artificial intelligence? And my response is, I didn't move into artificial intelligence. Artificial intelligence actually moved to encompass what I was doing.

**[0:16:24] Luke Timmerman:** It sounds like you were comfortable with the predictive ability of what you were doing.

**[0:16:30] Daphne Koller:** So at the time, this was right around, you know, the tail end of my PhD was in the early 90s, 92-ish or so. And at that point, I was really starting to get into machine learning because of the realization that the systems that one needed to build in order to make intelligent decisions were really hard to engineer well. And the only way that one could really circumvent what was then called the knowledge engineering gap or challenge was to not have to engineer the system, but rather have it learn for itself from data what the model ought to be.

**[0:17:08] Daphne Koller:** And so that's when I started to really get into machine learning as a discipline. So I've been working in that space, I would say, since the early 90s. I know a lot of people now think that machine learning is a relatively recent phenomenon, and it's something that is just recently emerged. But no, it really sort of took off as a discipline within computer science and engineering in the early 90s. I mean, statistics had existed long before then, but it was very different.

**[0:17:44] Daphne Koller:** And the two fields kind of merged. Statistics became more predictive and machine learning took off within computer science. And I think the two together, plus ideas from neuroscience and optimization, those are really, it's the confluence of those that gave us what we think of as modern day machine learning.

**[0:18:04] Luke Timmerman:** Interesting. So you said you moved on to Berkeley for your postdoc. Is that when you started looking around biology, you know, that field that your dad had done and you had no interest in as a kid?

**[0:18:18] Daphne Koller:** No, that actually happened later. But I will say one of the things that happened after I returned to Stanford as a faculty member, and I can talk about that in a minute. But one of the things that really did happen to me at Berkeley, which to me was a sort of pivotal, almost kind of transformative event, was in the very first conversation that I had with my then postdoc mentor was when I first came and he took me to lunch and said, so you did all this really great work in your PhD and it's very elegant.

**[0:18:49] Daphne Koller:** So if I gave you a group of like three smart undergrads to work with you and build a system that would be useful based on the work that you did, what would you have them do? And I was totally dumbstruck because no one had actually ever asked me that question. And it bothered me that I didn't have an answer. So in some ways, that was this inflection point in my thinking of if I do something, it shouldn't just be pretty.

**[0:19:20] Daphne Koller:** useful. So that's when I started to work on, well, first of all, machine learning more broadly, because I felt like that was just a generally useful bag of techniques, which I think history has proven that is in fact the case. But also over time, and that started at Berkeley and continued in my faculty role at Stanford, working on applications of machine learning that would both be useful in themselves and also drive the development of new machine learning techniques, which I could have some assurance were actually useful because they arose in the context of a more applied problem that I was working on.

**[0:19:59] Daphne Koller:** So I gradually became more and more interested in that interplay between foundational methods and applications. It's really interesting that he

**[0:20:08] Luke Timmerman:** prodded you to think about what are the problems you really want to try to solve? And this is This is often where, I mean, kind of what allows the great scientists to stand out is selecting those problems that are important. I mean, intellectually interesting, but important as well.

**[0:20:30] Daphne Koller:** Yeah, and I think I wish that more people actually spend time thinking about the nature of the problems that they work on and what value they provide to society. And I think that that manifests on both sides of the spectrum. There are people who are doing like really obscure theory and don't spend time thinking about, is this going to benefit humanity in any way? And then on the other hand, you have really equally smart people working on problems that while applied are, shall we say, not as aspirational as one would hope.

**[0:21:11] Daphne Koller:** And how many people do we need applying machine learning to optimizing, you know, the stock market or selling ads online? I mean, maybe some of that, but you really would like to see more people spending their mental energies applying machine learning to problems of social good. And so to me, that was actually, again, the sort of natural next step of my evolution in trying to become more applied is not only wanting to do something that was useful, but actually useful for good.

**[0:21:45] Daphne Koller:** And that was part of the transition into biology, which happened towards the late 90s when I was a relatively junior faculty member at Stanford.

**[0:21:56] Luke Timmerman:** So you're computer science junior faculty at Stanford. And how did you get to know what those people are doing in the wet labs?

**[0:22:06] Daphne Koller:** So initially, I was mostly looking for cool data sets. and the data sets that we had to work with as computer scientists at that point were rather lame. I mean, they were not as challenging technically, but more importantly, when you said, if I solve this problem, will anybody care? And the answer is I can publish a paper that said the accuracy went up from 83 to 87%, but so what? So I went and started talking to people on the biology side, my first problem, funnily enough, that I worked on was with someone who subsequently, who was at an epidemiologist working on tuberculosis, subsequently went to the Gates Foundation to work on TB there.

**[0:22:52] Daphne Koller:** And so it's actually a problem that's very relevant to our COVID problems today. And that was my first foray. And then over time, I expanded from epidemiology and problems that were more at the clinical level down to biology and then back up again to work again on more clinically relevant data. And so I kind of did almost a full circle from clinical data, or in that case, epidemiological data, down to very in-depth biology using some of the earlier microarray data sets, some of the earlier genome sequencing data sets, and so on.

**[0:23:33] Daphne Koller:** And then over time, back to working on data that was from clinical samples.

**[0:23:41] Luke Timmerman:** Okay. So you mentioned that this was the era of microarrays and gene sequencing. I mean, the human genome project was up and going, but they didn't have, you know, that wasn't done yet. So we weren't in the genomics age. What was really lacking there in biology from your view? Was it just not enough data or that the data types weren't properly linked or both or what?

**[0:24:05] Daphne Koller:** I mean, initially, certainly there were some interesting data sets there. And I think we did some work that I'm still quite proud of on, for instance, how do you use gene expression data to uncover gene regulatory networks. And that was, I think, work that was quite influential. The paper is still very highly cited. And so I think one could make do even with the data sets that were out there and do some interesting, certainly development of machine learning techniques, but also extract interesting biological insights.

**[0:24:45] Daphne Koller:** But I think what I realized over time was that as I developed my own interest in biology and said, here's a problem that I think if I were able to solve it, it would be super cool and I think quite valuable. Look around and the data sets just weren't there. And so you would go to a fellow biologist and say, hey, if we generated this data set together, or think about the cool stuff that we can do.

**[0:25:19] Daphne Koller:** And they would say, oh yeah, that could be quite interesting. Very cool. I don't really have a postdoc that's able to work on this right now. So why don't we write a grant proposal together to the NIH? And by the time you write the proposal, and then you rewrite it when you get your first round of reviews, and then you hire the postdoc, and then you onboard the postdoc, three years have passed. And by that point, you don't really care.

**[0:25:45] Luke Timmerman:** You're alluding to some of the challenges of academia.

**[0:25:48] Daphne Koller:** Exactly. And so that was rather frustrating not to have any ability to sort of say, look, this is a very important problem and I would love to just be able to generate data that's fit to purpose. So to me, that was part of the shift in thinking that ultimately led me to Insitro. But I will say the truth is that the tools to do what I would have wanted to do were just not there. So I did, as an example, my post, not my postdoc, sorry, my sabbatical in a biology lab at UCSF with Jonathan Weissman, who's one of the leaders in the functional genomics CRISPR field, incredible scientist.

**[0:26:31] Daphne Koller:** and when I did my sabbatical there, we were working with model systems like yeast and trying to do things like looking at the effect of one gene knockout versus two gene knockouts. And even in yeast, which is a much more tractable organism, these experiments were incredibly slow and painful to do because the biological toolkit just wasn't there. We didn't have CRISPR. We We didn't have all of the ways in which one could take individual cells and measure their transcriptional profiles.

**[0:27:11] Daphne Koller:** So I think my ambitions, if you will, the science just wasn't ready for them at the time.

**[0:27:19] Luke Timmerman:** But a tool-driven revolution was beginning to happen. Like the ability to generate massive amounts of data of the types that you are mentioning. I mean, that's step one. Once you can generate all of this different kinds of data, well, now you can sink your teeth into and analyze it and think about how it might go together and perhaps put machine learning to work. But we're getting a little bit ahead of ourselves now. So you had an entrepreneurial experience in here with Coursera, online learning, famous, right?

**[0:27:53] Luke Timmerman:** Co-founder and co-CEO of this thing. Now, did you actually leave Stanford full-time to do this, or was it a part-time thing?

**[0:28:02] Daphne Koller:** So it was a full-time thing, but it was a temporary thing. So I hadn't intended to leave Stanford. I had this mindset that I would retire as an academic. But even as an academic, I was feeling this increasing level of frustration internally of not being able to make the kind of impact on the world directly as I would have wanted to have. So I actually tried. I mentioned earlier that I'd had this clinical experience sort of that I'd circled back to and had worked on some of the earliest machine learning-based analyses of cancer biopsy images under the microscope.

**[0:28:44] Daphne Koller:** And that was, again, a piece of work that even though the machine learning is now drastically out of date. It was a very impactful piece of work and one that I thought could benefit patients. And I tried really hard from within academia to get some adoption of that approach into the standard of care, into hospitals, to get maybe device manufacturers that scanned tumor samples to use the technology. And it was very difficult to get someone to adopt a piece of paper.

**[0:29:16] Daphne Koller:** You needed to actually build a product for them. And that doesn't align with what an academic is supposed to do or their students or trainees are supposed to do. And so that was a point where I was like, if I'm going to make an impact, I can't really do it from within where I sit. And so that was right around the time that a project that died project for me, it wasn't really my research focus, but I'd always had this passion project on education, working in parallel with my research.

**[0:29:48] Daphne Koller:** And that just kind of ended up erupting as the launch of the first Stanford massive open online courses, the so-called MOOCs. And when that happened, I suddenly found myself looking at this thing that we'd inadvertently created, where each of those courses had 100 ,000 people taking limited time. And some of those were people from truly disadvantaged backgrounds, really poor countries, really low socioeconomic status who wanted a leg up by having access to a Stanford quality education.

**[0:30:25] Daphne Koller:** And I felt like it would be really remiss of me to just say, oh, that was nice. We did that for a while. And I'm going to go back and write some more papers and assume that someone would take that on and build an enterprise around this that would continue benefiting so many people. And so that was the time when I said, I'm not going to go back to the lab. I'm going to take what was supposed to be to your leave of absence from Stanford, going to form a company around this and really try and bring this to people.

**[0:30:57] Daphne Koller:** And everything around Coursera was with the premise that at the end of two years, I would leave it in someone else's hands and come back to Stanford and continue to be an academic. But when that two-year leave of absence came to an end, Stanford has a pretty rigid leave of absence policy. They said, hey, you've got to decide. Are you coming back? And I said, I can't come back right now. I need at least another year to get this to the point where it's much more mature and stable.

**[0:31:24] Daphne Koller:** And they said, well, you've got to decide. So I decided and I ended up resigning what was a chaired, tenured faculty position at Stanford, because I felt like I really needed to see that Coursera journey through.

**[0:31:36] Luke Timmerman:** Which is something not a lot of your peers would do or often do. People think, boy, Daphne, this is kind of nutty.

**[0:31:46] Daphne Koller:** Yeah, my mom thought I was totally nuts because my mom, you know, she was married to my dad who was an academic. And for her being a faculty member at the world's top computer science department or whatever top handful, you know, there's a tie among the first of the top few, like giving that up. I mean, it was the pinnacle of one's aspirations.

**[0:32:10] Luke Timmerman:** Yeah. Yeah. So you decide to stay in both feet in there at Coursera. And how long did you stay?

**[0:32:18] Daphne Koller:** So from the very beginning of the journey, which was in the fall of 2011, which is when we launched the first MOOCs while still at Stanford, up until the time that I left, it was almost exactly five years.

**[0:32:28] Luke Timmerman:** Okay. Okay. So what was the next move? Was that Calico?

**[0:32:34] Daphne Koller:** Yes. And that was when I, towards kind of mid-2016, Coursera was already in a good trajectory. And, you know, it's a wonderful, it still is a wonderful company. But fundamentally, it's a content company. It helps faculty create courses and delivers them to learners worldwide. why there's not a lot of hugely deep technology in there and certainly not a lot of machine learning. And I looked around me and the machine learning that I'd done at Stanford had made so much progress.

**[0:33:16] Daphne Koller:** I mean, huge step function in ways that I would never have imagined back in 2011, when I left Stanford, how far it would come in such a short period. And I was wanting to be part of that. And so.

**[0:33:31] Luke Timmerman:** Get back into the action. And your work was done there, really, at Coursera. Other people could take it from there. Exactly. Okay. So you get to talking with, I mean, this is a pretty first rate crew, Art Levinson and Hal Barron, David Botstein and others. What was the draw there?

**[0:33:51] Daphne Koller:** so I wasn't actually thinking of calico as a place to go I had encountered art in a number of other occasions just here and there partly you know his son was in the computer science department at Stanford and I was on his PhD thesis committee I'd also encountered him in a couple other occasions. So I reached out to Art thinking that he was someone who knew the world of, you know, at least drug discovery, biotech, and so on.

**[0:34:24] Daphne Koller:** And at the same time, because of his involvement at both Google and Apple, he understood the tech world. So I thought I would just ask him for advice on what was most interesting to pursue in this space. And it was a very memorable meeting because he said, look, I can answer your question, but let me tell you about Calico. And so he did. And I don't know how much you've had an opportunity to interact with art and then subsequently with how, but they're quite, shall we say, compelling people.

**[0:34:51] Daphne Koller:** And I found it hard to pass up the opportunity to spend some time working with them.

**[0:34:57] Luke Timmerman:** Okay. So you go there for a little while and about two years, right?

**[0:35:03] Daphne Koller:** Yeah. So it was actually 18 months.

**[0:35:05] Luke Timmerman:** Okay. Okay. So, and then comes along this opportunity for insitro. Was that the sequence?

**[0:35:14] Daphne Koller:** It was a combination of that plus, I think, the internal realization that fundamentally I was really an engineer, machine learning scientist. And what I really wanted to do was to engineer this broad-based platform for doing drug discovery differently. and that didn't make sense to build in within a company that is focused on a very specific biology and and really is driving forward an understanding of aging and I didn't personally want to become an aging scientist I mean it's not who I am it would be like a you know pretty late career change I wanted to build a platform and it didn't seem to make sense to build it there

**[0:36:03] Luke Timmerman:** If you like listening to the Long Run Podcast, you'll love reading Timmerman Report. This is where you'll gain a deep contextual understanding of biotech from my writing and the writing of many talented contributing writers. This is the kind of stuff that will help you get ahead of the curve. It's a bargain at $169 a year for an individual to subscribe. Discounts are available for groups with multiple readers. Go to TimmermanReport.com slash subscribe to show your support today.

**[0:36:33] Luke Timmerman:** And are you a fan of the Long Run Podcast? Trying to raise awareness of your company, your organization, or your services with a high-powered crowd of entrepreneurs and venture capitalists who listen to this show? Ask me about advertising opportunities. For more information, Luke at TimmermanReport.com.

**[0:36:55] Daphne Koller:** so how did this uh new company come to be so you know it's um i was introduced initially to one of my future investors who had been looking in fact for quite a while for a machine learning company to back and had done a lot of due diligence on a number of companies, didn't really find any of them to be what he wanted to make an investment in. And he reached out to me. I think he had heard somehow that I was thinking about next steps.

**[0:37:39] Daphne Koller:** And then I ended up meeting a few of the other of the investors that ended up being part of the initial consortium that founded insitro. I mean, Vijay Pandey, who was -

**[0:37:51] Luke Timmerman:** Who was this first investor?

**[0:37:53] Daphne Koller:** So the first investor was Jim Tannenbaum. And then very soon thereafter, Krishna Yeshwant from GV and Bob Nelson from Arch.

**[0:38:04] Luke Timmerman:** Jim's from Foresight Capital. Yes. Yep.

**[0:38:07] Daphne Koller:** And then Vijay had been a Stanford colleague and I had a lot of respect, have a lot of respect for his understanding as a scientist who bridges those two worlds. So he was a very natural person that I wanted to add to that consortium. So we ended up with five major investors, actually, in that initial group, with Third Rock being the fifth.

**[0:38:33] Luke Timmerman:** Now, Vijay was at A16Z, Andrews and Horowitz at the time. Okay, so these conversations, yes. So these conversations were occurring about when, like late 17, early 18?

**[0:38:47] Daphne Koller:** Yeah, they really started in early 18 as I was kind of starting to transition out of Calico.

**[0:38:55] Luke Timmerman:** Okay. Now, you mentioned that a whole lot of things had happened while you were away at Coursera, even in years previous. What else had happened by this point, 2018, that got you excited that, okay, this is the time to start a company around machine learning for drug discovery?

**[0:39:15] Daphne Koller:** So one part, of course, was just the incredible emergence of machine learning technologies that were capable of solving problems that were inconceivable for machine learning to be able to solve when I started to work in the field. But the other was addressing what was, to me, the other half, which is the truth is that machine learning is only as good as the data that you feed it. And, in fact, if you ended up giving me the tradeoff between the best machine learning algorithm on a crappy data set versus an amazing data set with kind of like an okay machine learning algorithm, the second will win every day.

**[0:39:59] Daphne Koller:** Now, the best of both worlds, of course, is a great machine learning algorithm on a great data set, but I really not wanted to go back to doing machine learning on data sets that are not that great. So, the observations that I had both before and during my time at Calico about the technologies that had emerged on the side of bioengineering, cell biology, and so on. And the idea of what would happen if you took all these little, not little, because each of them is hugely important on its own, but took all of those pieces, put them together, and built a sort of perfect engine for data production.

**[0:40:42] Daphne Koller:** What would that look like? And then applied machine learning as a way of extracting maximum information from data sets that were designed to purpose, to have maximal benefit from a machine learning algorithm.

**[0:40:57] Luke Timmerman:** This is an important point. I mean, you allude to the garbage in garbage out problem. And, and, you know, you're also by this point, you've looked around the pharmaceutical industry and seen, you know, a lot of big legacy organizations, they've got, it's a highly regulated industry, there's certain ways of doing things, data sets that, you know, don't necessarily go together very well. Um, so did you, did you consider like whether the alternative of, you know, trying to do something machine learning for drug discovery related in a big company with a lot of resources and presumably a lot of people to help you, um, or like start with a clean sheet of paper and and feed the data in

**[0:41:40] Luke Timmerman:** that, that you think needs to go into the right algorithm.

**[0:41:44] Daphne Koller:** I absolutely did. And that's a very good question. As you say, there are many advantages, at least in principle, in doing this kind of work within Big Pharma. There's already data sets that have been generated. There's already people there, lots of money in principle. And so it was a natural direction to consider. And I think the thing that ultimately decided me against it is what I think of as cultural inertia. That the companies, when they've been around for long enough, develop a certain way of doing things, a certain hierarchy of who gets to make the decisions and how those decisions are made and who gets to have a say.

**[0:42:32] Daphne Koller:** and none of these companies were built as data-first companies. They were built as science-first companies. And so the data scientists in most of these companies are usually this kind of smallish siloed group that is sitting in some corner somewhere. And when the scientists have finished conducting whatever experiment they're done conducting, they ship it over to the bioinformatics team to produce a set of charts and analyses that then the scientists take back and figure out what science to pursue.

**[0:43:06] Daphne Koller:** And that is really not the culture that I wanted to set up, nor, by the way, was its mirror image, which you often see in companies that are born from the tech world, where it's a bunch of engineers who get to decide on the strategic direction and call the shots. And they make the decisions and you have, whatever, a few biologists sitting there providing advice on, yeah, this is something that would be useful for doctors or biologists or whatever.

**[0:43:32] Daphne Koller:** Building a company where the two groups, and it's actually more than two, but for simplicity, let's just think of it as two groups, really walk into something as equal partners. And by that, I don't just mean solving problems together, but figuring out from day one, what are the problems that ought to be solved, that is a very different mindset. And I didn't think it would be easier to take a 100 ,000 person biopharma company and shift it to that form of decision making.

**[0:44:04] Luke Timmerman:** You know, just a very minor tangent here, Daphne. But I came up in the newspaper industry in the late 1990s. And so I was a young reporter when newspapers with the internet age came along and everybody realized that they needed to make this transition to online. and legacy organizations that have been around 100 years, you know, had all kinds of structure and culture and everything built up around print. You know, you have a whole bunch of like afternoon deadlines and, you know, a whole set of copy editors that would review copy before the desk.

**[0:44:34] Luke Timmerman:** And I mean, it was a hard transition for many of these companies to like start creating content in a way that was more suitable, the way that people wanted to consume it on their computers and their mobile devices, et cetera, and publishing systems and everything needed to be re-engineered. And so there were advantages for some of those first crop of web 2.0, I guess, online only journalism startups. They didn't have all that legacy baggage, if you will.

**[0:45:03] Luke Timmerman:** They could start from the first principles and say, well, what is the best way to set up an organization to serve readers the way they want the news now? And I think there may be an analogy here It's what you're trying to do.

**[0:45:13] Daphne Koller:** I think absolutely. And it's, you know, most of the newspapers, as you know, did not actually survive this transition. And conversely, there is a whole crop of newcomers who are very influential now and didn't exist back in the days of early newspapers. Now, some, a small handful, I think, of traditional print newspapers actually made the transition successfully. But as we said, most didn't. And when I think of some of the big tech-enabled companies out there that are dominating the world, the Amazons and Googles and such, Amazon wasn't Walmart with a website.

**[0:45:55] Daphne Koller:** Google wasn't the yellow pages and Netflix didn't come out of Blockbuster or Hollywood Studio.

**[0:46:04] Luke Timmerman:** So they borrowed some important concepts from those predecessors and but discarded a lot of the stuff that wasn't necessary for the next phase.

**[0:46:13] Daphne Koller:** Exactly. And I think that is understanding what is what's good to keep, but what needs to be engineered from the ground up.

**[0:46:22] Luke Timmerman:** Okay. Okay. Sorry to interrupt, but you did mention culture here. So the idea that you're going to start a company, you're going to be the CEO and founder, you get to set that culture. You get to pick the people and set them up in an organization that, as you see fit. And you need both computational people and wet lab biology expertise. And I love this comment you made when we did this event for the Petrie Bio Group a month or so ago.

**[0:46:52] Luke Timmerman:** You talked about recruiting damp scientists who are neither wet nor dry, but both. And I know that people who are listening and they love that comment. Can you expand on that a little bit? It's kind of like a fun way of thinking about cross-disciplinary coexisting.

**[0:47:12] Daphne Koller:** No, it is one of, I think, my favorite concepts. And a lot of the people that we've been lucky enough to recruit to in Citro are exactly that. They are 70 % life scientists, but they know how to program in Python and analyze their own data. or they are 70 % computer scientists, but they spend some time in a wet lab. And so they understand the complexities of getting a biological system to actually do what you want it to do and all of the issues of dealing with living cells and such.

**[0:47:45] Daphne Koller:** And so there is a good group of that at insitro. Now, we're not 100 % that. We have some folks who are purely life scientists, but they just love the idea of being at a place where there's this immediate use case for their data in feeding a machine learning algorithm. A lot of them grew up in a wet lab where they produced massive amounts of data and they realized they were hitting a wall in terms of extracting only a tiny fraction of the value of their data using simple analytical tools.

**[0:48:19] Daphne Koller:** And then on the other side, we have some people who are just core computer scientists. Some of them don't have any life science experience, but they really want to be earlier discussion at a place where what they're working on isn't selling ads online, but something that does provide a much greater societal benefit. And so we're getting a lot of people who say, I want to spend my life doing something worthwhile. Because the thing about machine learning people, of course, is they can work on anything, right?

**[0:48:51] Daphne Koller:** They can go, I mean, any tech company, in fact, any company at all today is hiring data scientists and machine learning people. The ones that we get are ones that really want to make a difference.

**[0:49:03] Luke Timmerman:** So they're coming to work here to put together a platform, a new kind of platform for drug discovery. And it's an amalgam of in silico and in vitro. That's the inspiration for the name in situ. Can you talk a little bit about that platform? What are the key components? What are

**[0:49:21] Daphne Koller:** you building? I'm going to actually start with the higher level aspiration, which I realize is long-term and grandiose and will take many years, which is the drug discovery journey from being that you're going to work on a particular disease or indication to the point that you actually have a drug that works in people, that's a many-year journey. And there are many steps along the way that we make decisions that are currently many times driven by intuition or animal models that we don't really believe in, but we don't have anything better, or historical heuristics.

**[0:50:03] Daphne Koller:** And the question is, what if some of those decisions could be guided by predictive models that are kind of like a compass that tell us of the many forks on the road, which one is more likely to be successful. And those four includes things like which target do I pursue and in which patient population and of the multiple molecules that I can potentially develop, which are the ones that are going to job of binding to and modulating this target and what's the right biomarker for clinical efficacy so that when I go into clinical trials, I know which of my patients are doing better and which ones aren't.

**[0:50:42] Daphne Koller:** All of these are opportunities for machine learning. Over time, we'd like to do multiple of these, and that's going to take a long time and a lot of resources. But that's the aspiration is to really build a biopharma company from the ground up. Now, that being said, we're a startup. We need to focus and can't do all of these things at once. And so the thing that we started to build our very first platform, if you will, is something that tries to address what to me is the fundamental problem of drug discovery, which is if I make an intervention of this type in a human being, what clinical outcome is that likely to have?

**[0:51:25] Daphne Koller:** And that is fundamentally at the core of drug discovery. And the problem, of course, is that you don't get to collect interventional data like that from humans until the very end of the drug discovery process when you actually do a clinical trial. So how do you collect training data that informs a model on what an intervention will do in a human? And that's where we've actually put together two very complementary data modalities, each of which I think is really powerful, but putting them together is even more powerful.

**[0:51:58] Daphne Koller:** One is human genetics, where we make, our nature has made interventions in a human, in each one of us. Each one of us is an experiment of nature, and we can see, based on the set of perturbations that comprise our genome, what do we look like? What is our clinical outcome? So that's one type of data modality. The other one, which we think is very complementary, is because we ultimately need to be able to make interventions in a system ourselves so that we can do experiments, and we can't do the experiment in a human, we don't want to do the experiment in an animal model that doesn't usually translate, is we're using human-derived model systems that are based on

**[0:52:38] Daphne Koller:** each of our genetics, where we can take cells from you or me, convert them into pluripotent status. stem cells and then differentiate them into a Daphne neuron or a Daphne hepatocyte or any one of the set of lineages that the community has developed differentiation protocols for. And then I can ask myself, by intervening in that cellular system, in that Daphne neuron, with either a genetic or chemical intervention, what effect does that have at the cellular level?

**[0:53:12] Daphne Koller:** And if I can then use machine learning to interpret what I see in many, many measurements of those cells and align them with what I see about my clinical outcome, that's where the machine learning allows me to kind of put those two data sets together and hopefully make more reliable predictions about what interventions that I get to test in the lab, like a new drug or a new target.

**[0:53:37] Luke Timmerman:** So you start with human samples. and iPSCs. Now, you also are able to collect a lot of the genetic information that you alluded to. Is it exome in nature?

**[0:53:58] Daphne Koller:** It varies. Sometimes it's exome. Sometimes you start with genotyping. I mean, obviously, the closer you get to sequencing, the more information you have about variants that are more rare and therefore in many cases have a larger effect size on the phenotype. So you're often underpowered to see those, but when you see them, they're very, very revealing. So you kind of end up putting together genotyping and sequencing to get as much information as you can.

**[0:54:28] Luke Timmerman:** And you're able to do CRISPR knockouts at a pretty high speed across. You can ask and answer lots of questions.

**[0:54:38] Daphne Koller:** Absolutely. And that's where there is, again, these beautiful technologies like doing pooled CRISPR screens and then doing high content phenotyping of the cells in a pool. So you can say, oh, I've made perturbations in thousands of cells with thousands of different perturbations and measure, for instance, a genome-wide transcriptional profile and really understand what it did to the cells, each of them.

**[0:55:05] Luke Timmerman:** You mentioned hepatocytes and neurons. I don't think that was an accident because I think you've said your first two therapeutic areas are liver and CNS diseases. That's right. What is it about those cell types or those diseases that makes that a reasonable place for you to start with building your platform?

**[0:55:25] Daphne Koller:** So liver was in some ways serendipitous. It came out of the partnership that we have with Gilead on NASH, non-alcoholic steatohepatitis. we uh partly it was selected because we really liked gilead as a partner the data that they brought into the collaboration the quality of the scientists that we got to interact with and uh it partly because it did actually align with the scientific platform that we were looking to build so we said no that that looks like a good match neuroscience was one that was driven by us and was a consequence of several observations.

**[0:56:03] Daphne Koller:** First of all, for whatever reason, some of the best differentiation protocols out there are actually in the CNS space, in the neuroscience space. And so it's a very, for whatever reason, tractable system to work with from an experimental perspective. The other one, the second reason is that even though we understand a lot of the underlying pathways in the in cns diseases they are incredibly genetically driven so for instance for a lot of the neurodevelopmental neuropsychiatric neuro degeneration the heritability is you know 30 40 50 70 percent for instance in autism so it's very high but also very complex and therefore an opportunity for machine learning to provide insights that currently people don't have.

**[0:56:53] Daphne Koller:** And then maybe the last and most important one is the extent of the unmet need. The ability that we currently have to provide meaningful treatment to almost any neurological disorder, with a very limited number of exceptions, is exceedingly low. And these diseases are hugely a burden on the individuals who are afflicted, on society. And so we felt like there was an opportunity for a very different suite of methods that hadn't really been deployed in this context to make, we hope, a really big difference to these patients who don't currently have other options.

**[0:57:38] Luke Timmerman:** Now, this might be a hard question to answer. It might be too early. But I mean, knowing the tools that you have, and that the data that you're able to collect and analyze, I mean, it's not perfect. Like you don't have everything. You don't have the whole cell modeled or the whole organism. I mean, the complexity is just unbelievable. It isn't like, you know, Boeing where you can, you know, put your flight simulator and you can draw all these parameters of a new plane and, you know, predict on a computer program whether that thing will fly with a certain amount of cargo and passengers, et cetera, et cetera, right?

**[0:58:14] Luke Timmerman:** So, given what you have, where do you think this might, machine learning for drug discovery might be useful first? Might it be in identifying like rare genetic subpopulations of CNS patients who are likely to respond to something? Or is it maybe you can get an early warning sign on, you know, where adverse events might crop up? because even just like nibbling on the edges a little bit of the drug discovery problem would be a big difference because it's just it's such a i mean so much time and money is wasted right now because of this inability to create the you know the omniscient predictive model well so i'm not

**[0:58:55] Daphne Koller:** sorry let me rephrase that i'm sure we're not going to get the omniscient predictive model That is a very, very aspirational goal, and I try and keep my ambitions while large to be at least somewhat reasonable. But as you said, our current failure rates are so high, or rather our success rate is so low that I think even a modest improvement, relatively speaking, can make a huge difference to the productivity of the industry and to being able to address diseases of unmet need.

**[0:59:32] Daphne Koller:** So, as I mentioned earlier, I think there's opportunities all throughout what's typically called the, you know, Pharma R &D value chain, and many of those actually will make a difference. the ones that the one that we've decided to focus on is really identifying those novel targets or at least novel within an indication and and and many of those as you said will in fact be within the context of a specific patient population to me that is a huge and in some ways overdue transformation when you think about polygenic germline disease when we think about the progression that's happened in oncology and specifically precision oncology in the last decade or so, we now realize that breast cancer is not one disease.

**[1:00:24] Daphne Koller:** It's, you know, a BRCA1 breast cancer is very different from a HER2 breast cancer and are in fact treated with completely different treatments right now. And we haven't really done that in most other non-cancer indications. We still think of Alzheimer's as a disease, even though we recognize that there is an incredible heterogeneity. Autism is called a spectrum for a reason. And the same is true for diabetes and cardiovascular disease and many others. So there may be some number of these blockbuster, all-comer drugs still out there that if they are there, that's great.

**[1:01:06] Daphne Koller:** We should obviously look for those. But I think a lot of the future is going to be in the same place that we see it in precision oncology, which is to identify a subset of patients that are well tackled with a particular target that works for that patient population. Now, does it have to be a rare patient population? Not necessarily. I think that in some cases, we see that there is a signature, the sort of end of the thread in a genetically defined rare patient population.

**[1:01:42] Daphne Koller:** But in fact, the treatable population is nevertheless considerably larger. A great example of that is the poster child here, which is PCSK9, where it was discovered in a rare familial variant in hypercholesterolemia on the one hand, familial hypercholesterolemia, familial hypocholesterolemia on the other, sort of showing this beautiful dose response. But it turns out that the PCSK9 drugs are actually valuable even for patients that don't have a genetic variant of hypercholesterolemia in the same way that the familials do.

**[1:02:25] Daphne Koller:** So I think there is a possibility to identify using genetics, those rare variants, and then expand from there into a broader population. It's still not going to be, in most cases, the universe of all patients that present with what we currently call a certain disease, partially because, frankly, I think a lot of our disease taxonomy is just not right. It's based on very coarse-grained clinical manifestations, many of which are very subjective. And I think over the next 10 to 15 years, we'll actually end up with a very different disease taxonomy than we have today, especially in the neuroscience area.

**[1:03:03] Daphne Koller:** Um, but, um, so we're not going to find, I think the Alzheimer's drug, but, but maybe there's going to be a patient population that isn't 0.01%, but something that is reasonable that really sees dramatic effect sizes. I mean, that's really the hope that's, I think what we all hope.

**[1:03:23] Luke Timmerman:** It's really interesting. And, and in this case, the biology is being redefined, um, down there at the molecular level, and that will guide therapeutic development, like you say, with the PCSK9 elegant example. Okay, so Daphne, we're almost out of time. You've raised a lot of money. I think you did $140 million Series B in the spring. So congratulations. You got some decent runway here. You got good, solid investors, and they seem to be patient people. So can you give me an example or something in the last, I don't know, the last six months on how you're doing where you got up in the morning and said, holy cow, today was a good day.

**[1:04:06] Luke Timmerman:** And here's why.

**[1:04:09] Daphne Koller:** You know, it actually happens on a pretty regular basis, because what we see on a day-to-day basis is that magic that happens when you take people from very different disciplines, and you put them together in a single working group and you say, figure out what you can do together that neither of you would ever have envisioned as even being possible working alone. And so we see the synthesis of sort of high throughput data generation and machine learning kind of coming together in so many different ways, some of which range from just core infrastructure.

**[1:04:51] Daphne Koller:** Like, can you get a machine learning system to automatically culture pluripotent stem cells and differentiate them? Because you know what? An expert biologist can look at well and say these cells are going in the right direction and these ones aren't if you're a really good developmental biologist. But a machine learning algorithm can do it even better. And so we're embedding that in just even the day-to-day operations of what we do. And then when you get to the higher level of, okay, now let's understand disease phenotypes.

**[1:05:28] Daphne Koller:** So moving beyond just the mechanics of how we build the company or how we build our assays, but really understanding human disease, we have seen patterns in, for instance, diseased versus normal liver cells in both human samples as well as in in vitro cell cultures that an expert biologist just could not see until the machine pointed it out to them. And that was the beginning of the understanding of this is the biology that we now need to modify in order to revert the cells from the unhealthy to the healthy state.

**[1:06:12] Daphne Koller:** And when we did it once, it was like, oh, my God. And then when we did it again, it was like, wow. And now it just keeps happening. And that, to me, is the promise of that synergy between the creation of the right data, asking the right questions, and letting the machine learning add the data using the best of the best machine learning methods.

**[1:06:36] Luke Timmerman:** So you get the people on board, you get the platform up and running, and pretty soon it starts pumping out more and more of the data that you really need.

**[1:06:47] Daphne Koller:** And then the other thing I will say, just coming back to some of the earlier threads that we were talking about before, the culture that we've been able to build, where there isn't a single project, I think, in pretty much the entire company, where it's just one group of people with like background working together. everything spans a range of disciplines and, and the fact that people are viewing themselves as just like part of a team, even when they come from these so different backgrounds, that is, that is magic.

**[1:07:18] Daphne Koller:** And I've not seen that happening elsewhere very much.

**[1:07:21] Luke Timmerman:** Well, and we're out of time, but last thing I wanted to say is like, you know, we talked about this earlier at the event we did that you create a culture of mutual respect across for people across disciplines. So like there, There really aren't dumb questions. People can challenge assumptions and be unafraid. And actually, this is where you consciously cultivate this to try to create that magic.

**[1:07:45] Daphne Koller:** That's exactly right. And what we find is that oftentimes the biggest insights come when someone says, so wait, wait, why can't we do it this way? And the normal inclination will be like, well, that's a dumb idea. We don't really have, we don't do it that way. But when you build that culture of, you know what? Huh, you're right. Why don't we do it that way? Let's just suspend disbelief for a moment and try and think about what would happen if we tried to do it that way.

**[1:08:13] Daphne Koller:** And oftentimes, not every time, but often enough, it's like, you know what? Something like that could actually work. And obviously, the domain expert takes the idea and reshapes it and actually makes it doable. But the idea often comes from that orthogonal way of thinking. It happens in all directions. So it's not just from, you know, the machine learning people to the life scientists or vice versa. And I think that's where the magic is going to happen.

**[1:08:39] Daphne Koller:** And I've seen it across other areas where the magic really happens at the boundary between disciplines, because that's where you get these like things that would never have occurred within the little silos that we build for ourselves in academia and in industry.

**[1:08:56] Luke Timmerman:** Daphne Koller, I wish you all the best as you go through the inevitable ups and downs with machine learning for drug discovery on the long journey. Thank you for joining me today on the podcast.

**[1:09:08] Daphne Koller:** Thank you so much, Luke. It was a pleasure.

**[1:09:12] Luke Timmerman:** Thanks for listening to The Long Run, a production of Timmerman Report. Pedro Rosado of Headstepper Media was the sound editor. Music is from D.A. Wallach. See you next episode. Thank you.
