Pod Insights a16z Podcast Digital Biology with insitro's Daphne Koller About this episode Daphne Koller is the founder and CEO of insitro, a company using AI and machine learning to engineer drug discovery. In this conversation with a16z's Vijay Pande, she discusses how an LLM for cells could revolutionize drug discovery and how to bridge the technological and cultural divide between building with atoms and building with bits.
[0:00] Why life sciences?
[3:23] AI in the life sciences
[7:02] LLM for cells
[11:37] Engineering disease and drug discovery
[13:33] Bits vs. atoms
[17:36] The opportunity ahead
This conversation is part of our AI Revolution series, recorded August 2023 at a live event in San Francisco. The series features some of the most impactful builders in the field of AI discussing and debating where we are, where we’re going, and the big open questions in AI. Find more content from our AI Revolution series on www.a16z.com/AIRevolution.
Transcript 0:00 Speaker 0 We built a language model for biology. So all of us are like, now everyone's an expert in language models. You have to explain this to people. It's like, oh, language of biology, no one knew what I was talking about. But now it's like, I'm just saying, look, it's just like GPT, but for cells.
0:41 Speaker 1 Daphne is like the OG's OG in AI. She was a pioneer at Stanford in uh different areas of AI, especially in uh PGMs. She left Stanford to found uh co-found Coursera with Andrewing. Uh, and actually is now the founder uh CEO of In Citro, a tech biocompany using AI to uh develop drugs in life sciences. So Daphne, you know, given all the things you could be doing, like why life sciences?
1:11 Speaker 0 It's one of the really hard and really important problems. And there is very few things that are as challenging, as exciting as um intervening in a safe and effective way in human health. And so it's just a thing that absolutely needs to be done if we are going to use AI for good, which I think is one of the things that I think I at least really strive to do. Um, the second part of the answer is why now? And what I what brought me back to this field back in um 2016, post-Coursera, was the realization that we can now finally for the first time measure biology at scale, both at the cellular level, sometimes the subcellular level and at the organism level via ways of quantitating human biology.
1:59 Speaker 0 And that gives us for the very first time the ability to deploy machine learning in ways where it is truly meaningful to do that because the data sets are large enough for really interesting machine learning methods to be um to be deployed. And that's something I spent a lot of time thinking about to your point, Vijay, because at this point, AI people can do pretty much anything.
2:25 Speaker 1 Yep.
2:26 Speaker 0 And I am a big believer in leverage. That is places where you can have a disproportionately large impact. And because of the fact that I had spent a large part of my Stanford career working in these two spaces simultaneously, core machine learning on the one hand and machine learning in service of biomedical data on the other, I actually have the ability to sort of bridge bridge the chasm between these two very disparate disciplines. And when I was leaving Coursera in 2016, and I looked around me and I saw even at that time, which of course is uh tiny compared to where we are today, um, that machine learning was changing the world, it wasn't having much of an impact in the life sciences. And I believe one of the main reasons for that is that because there's so very few people who actually have the language of both disciplines and are able to bring them together.
3:17 Speaker 0 So I felt like I could have impact in AI across many things, but sure I could have disproportionate impact.
3:23 Speaker 1 Well, you know, you spoke about the why now. Um, what's your take on AI for life sciences? What's the why now there? And is there what's different now than you know, even what we could do, even just let's say five years ago.
3:35 Speaker 0 So I think it comes back to this ability to collect, but even more than collect and generate data at scale. So one of the things that we have at in Citro that is truly unique is we have a data factory. We have put together um the tools that um have been developed by people who are taking pluripotent stem cells, which are cells from you or me or anyone in this audience and turning them into this pluripotent status, which can make a Daphne neuron in a dish or Daphne hepatocyte and vision is going to be different than the VJ neuron and the VJ hepaticide because we have different um different genetics, and that's going to manifest in how these cells um behave and how these cells look and different measurements. Uh, we can engineer those to introduce a disease-causing mutation and ask what is that disease-causing mutation do to a DAFT neurone versus what does it do to a VJ neuron and what is this mutation do versus that mutation?
4:26 Speaker 0 And so we're able to kind of do um data generation on spec. And that is a truly unique capability, which frankly is not that easy to do even in other areas where AI is being is being deployed. You don't get to make your own data in many cases, but here we do, and that creates both really important discovery opportunities for life sciences, but also really cool and interesting machine learning problems because you could start doing active learning, you could start the sort of experimental design, and it's actually like a really exciting technical discipline at this point.
4:58 Speaker 1 Well, maybe you could dive a little deeper and give an example. So, like uh your your paper on the Posh uh approach, I think just that came out on archive. Could you double-click on that, tell people what what you did there? Especially like why is AI and life sciences a big deal? What could you hope to get?
5:12 Speaker 0 So, well, first of all, let me tell you a little bit about that platform, um, which is called Posh or pooled optical screening in humans. You take a bunch of cells, they can be cancerous cells, but in our cases they're mostly those induced polypose cells with you know neurons or whatever, and you put them in a with a pool of CRISPR guides that edit them, and each cell gets a different guide. So now you have a bunch of cells, each with a genetically diverse mutation. And now they're all sitting there in a pool, you can measure them with a microscope, you can measure them as they move around and do their thing, you can basically fix them, and you sequence the bark that came with the guide. So now you can say, oh, this cell that got this guide behaved this way, and this other cell behave that way.
5:57 Speaker 0 And I can tell you that one of the really challenging things about cells is because they're live. If you put different cells in different wells, then they each have a slightly different environment and you get subtle differences, and it's really hard to reconcile. When they're all a pool, you eliminate all of those artifacts, and all of a sudden you have the ability to measure a genome-wide CRISPR screen, basically. So 20,000 genes in the genome, all modifying the same cellular background in the same dish with a different genetic intervention. And you're measuring that in a genome-wide scale in like 10 or 12 plates in two weeks. Now imagine doing that rinse repeat and doing genome-wide scale on this genetic background or in this cell type. And so you can really start to decipher the genotype phenotype connection and the effect in which individual genetics makes a difference on cellular phenotypes, which we then translate to what we believe they will have in terms of clinical impact.
6:54 Speaker 0 And that is the beginning of an understanding of what it is that we want to modify in order to have meaningful therapeutic interventions.
7:02 Speaker 1 Well, yeah, but no, the biology part is really critical because now you get the data, and we all know how important that is. But I think one of the things that's really I found intriguing is the creation of a latent space for human biology, uh, and especially being able to tell the difference between disease and non-disease or even different disease phenotype. So, what does that uh how does that come about? And like especially, you know, how is AI driving it?
7:24 Speaker 0 So actually, I'm gonna go back one step further because he said, of course, one of the things we need to do is get the data, and I should have mentioned that it's impossible to run this instrument without AI being built into it because you can't even segment the cells. You can't call the barcodes. I mean, all of it is an AI-enabled architecture. Every part of our technology stat is intrinsically AI enabled. But then, to your point, Vijay, now you have a whole bunch of cellular images, and what do you do with them? And so the first thing we do is we built this latent space. We built a language model for biology. So all of us are like, now everyone's an expert in language models. Used to have to explain this to people. It's like, oh, language of biology, no one knew what I was talking about. But now it's like, I'm just saying, look, it's just like GPT, but for cells, and people think they understand. So um,
8:09 Speaker 1 maybe we can double-click on that from today. Sure. Yes.
8:11 Speaker 0 So we have the language of you know, cells and what cells look like, or the transcriptional or gene expression profiles of cells, and you measure hundreds of millions of cells in different states, and now with a much more limited amount of data, because we have this latent space, then just like the large language models for natural language, with a small amount of data, you can start asking, okay, how does disease move you, like a disease-causing gene from one place to the other? How does a treatment move you hopefully back in the opposite direction from the disease state back to the healthy state? And that's super powerful, and it's uh, and it's the gift that keeps on giving. Kind of like other language models, you have this, it keeps getting better the more data you feed it, and over time you end up with a better and better like competitive mode, if you will, of how does understanding the core foundations of biology um help you understand disease and health better.
9:07 Speaker 0 And let me just say, this is not just for cellular data. The other source of data that we use is clinical data. So we do the same thing with histopathology. There's so much more in histopathology than your pathologist typically looks at in MRI data, your radiologist doesn't see more than like a small percentage of what's there in your radiology images, but also not just imaging. There's also other modalities where there's an equal amount of information left on the table. And over time, we're learning these languages of different biological modalities and the ability to translate between them.
9:41 Speaker 1 I think these this concept of foundation model for biology is particularly exciting because um, you know, 10 years ago, you could have ML that was predictive. You just needed maybe a hundred actives.
9:51 Speaker 0 Yeah.
9:51 Speaker 1 And the problem is like if you have a hundred cases, examples of a drug that works, you don't need to design a drug. And so these low shot zero shots. Approaches that come from a foundation model are really night and day. So how far does this go? I mean, the big problem in biology is that biology is hard.
10:08 Speaker 0 Biology is really
10:09 Speaker 1 hard. Well, biology is very hard. And you know, uh
10:11 Speaker 0 sometimes I ask myself, like, really, why am I doing this? I could go write an app or like a SAT age of
10:16 Speaker 1 do a FART app or something like that. Yeah.
10:19 Speaker 0 It would be so much easier, wouldn't it? Yeah,
10:21 Speaker 1 it would be a lot easier. But but like, so why are you doing it? What is the big win? How does this like where does this go? Let's say in by the end of the decade, like what could you hope to do that we can't do, couldn't do before.
10:33 Speaker 0 We want to do it in a different way. We want to come up with a very almost like systematic recipe for how do you go from a decision that I want to work on ALS or want to work on fatty liver disease through a sequence of steps towards something that results in a meaningful intervention in the right patient population. The hope is that by the end of the of this decade, we will have built this process, we will have run through it a number of times. We will have delivered some medicines to patients in our first trunche of indications, but then we will have learned enough from that so that we can now say, okay, and here's how we're going to do it here and here and here. And because it's not only machine learning that moves forward over time, it's also the biological tools that we're relying on. I mean, used to be that there wasn't any CRISPR.
11:19 Speaker 0 There was just SIRNA, actually, there wasn't even that. And then there's like CRISPR-base editing, and now there's CRISPR prime that replaces entire regions of the genome. So the tools that we're building on also get better and better over time, which unlocks more and more diseases that we could tackle in meaningful.
11:37 Speaker 1 Well, let's step back for a second, because I think this may not be clear for everyone, like why biology is so hard. And one of the biggest reasons is that, you know, if if if we can do tons of experiments on mice, you know, I have to feel like it's a great time to be a rich mouse. You can be cured of any disease, right? Like all these diseases can be cured in mice. But you know, it's obviously unethical to experiment on people. And that's why uh that's one of the big reasons why trials fail, right? Because when you go into a clinical trial, you spend all this money to get there, you're spending all this money, hundreds of millions of dollars in the trial, and turns out mice are different than people, and it fails.
12:11 Speaker 0 Yeah.
12:12 Speaker 1 So, like, how can AI help that?
12:14 Speaker 0 So, first of all, and and this notion of, you know, we can cure lots of mice, um, is something that really drove our discovery strategy at in Citro, which is all of our work is done in human and human-derived systems, so that incorporates at least some subset of uh of human cells working together. So that's one piece. And the nice thing about it is that it allows you to intervene in those systems and ask the what if questions, the counterfactual, like what if I had this person's um biology, but in a world where this gene was doing was inactive versus active or the other way around. So that's great, but obviously you want to cure people, not cells or even organoids. And so the other source of data that we bring in is data from people from clinical records.
13:01 Speaker 0 Um, and what we end up doing is kind of bridging between the two using machine learning. So using machine learning on the cellular data, using machine learning on the human data, bridging between those in representation space, but also in genetic space, because genetics is kind of like this thread that ties the two together. And uh, without machine learning, uh without AI, the space would be so complex and so high dimensional that you couldn't even make sense of it far less bridge between those two different worlds. Yeah, well,
13:33 Speaker 1 that makes sense. Um curious to change um uh gears a bit and talk a bit about company building. Yeah, one of the interesting things that you've done is that you've brought together people that are biology experts with people that are ML AI experts. And and how do you build that culture? What does that look like, especially since they're from fairly different you know, parts of the universe?
13:53 Speaker 0 How do you build that is really hard because you take your average, you know, machine learning scientists and your average life scientist, even if they're very well-intentioned, you put them into the room together, they might as well be talking pie and swahili to each other. The languages are different, the ways in which you think are totally different. So, how do you create a shared, um, a shared language, a shared vision? Um, and so there's a few tool tricks or I or approaches that we use. First of all, we hire some number of people, you can't get enough of them, unfortunately, who are in the middle, who are able to be translators and talk to both sides and kind of bring them together. Um, and then I think the other really important part is that you create a culture of and you hire very rigorously to that culture of people who are genuinely interested in engaging with the you know, the other side.
14:45 Speaker 0 And we have a list of company values, and the final value, which is one that I hold particularly dear, is it's last, not because it's least important, it's because they're ordered from what we do to how we do it, um, which is that we engage with each other openly constructively and with respect. And each of those words matters, engage means that we're not siloed. All of our work is done in like cross-functional project teams. Openly means an openness to asking really naive questions when you don't understand, and to accepting really naive suggestions from somebody else, because sometimes the best ideas come from an orthogonal mindset.
15:22 Speaker 1 It's something that I experienced even as a kid. So um my parents are scientists, and my mom warned me, no matter what I do, even though this is what I was doing as a kid, because of programming as a kid, I should not get into programming or computer science. No one's ever going to make money by selling software. Uh yeah, I I think maybe then, you know, especially for this audience here that are coming from uh the AI side, especially as AI gets into areas that are sort of not just the world of bits, but in the world of atoms. Yeah. Yeah. Any advice for how to bridge those gaps?
15:54 Speaker 0 Uh I mean, first of all, having uh deep respect for atoms um is really important. Yeah, I mean, some
16:00 Speaker 1 of my closest friends are atoms. I
16:03 Speaker 0 think we're all atoms, but I think having an appreciation for the complexity of atoms and the facts that especially, especially when your atoms are part of live systems, they behave in unexpected, unpredictable, idiosyncratic ways that sometimes cause a lot of pain. And I I can tell you that when you do biological experiments, one of the strongest signals when you apply machine learning to it is what was the technician who actually did the experiments? You could read that very clearly off the cells, because they behave a little bit differently. They pipette a little bit differently, they treat the cells a little bit differently. It's amazing how hard that is to clean that up, which is one of the reasons why we spend so much of our time building robots because they do the same thing over and over again.
16:53 Speaker 0 So I think having a lot of respect for atoms, but also I would say an appreciation for the fact that the next frontier of uh what a of the impact that AI can have is when AI starts to touch the physical world. And we've all seen just how much harder that is. We've all seen how hard it is, astonishingly to build a self-driving car compared to building a chat bot, right? I mean, we've made so much progress on building chatbot, and self-driving cars are still blocking fire trucks in San Francisco. Um having an appreciation for that complexity, but also an appreciation for the magnitude of the impact if you can actually nail it.
17:33 Speaker 1 Where does this all go? Like, you know, you're talking about life sciences in terms of healthcare and drug design, but there's a lot more to biology than just uh than just drugs, right? But where do you think this confluence between AI and life sciences goes from here?
17:48 Speaker 0 So I actually think that there is this incredible opportunity at this point at this intersection between the two fields. And I will kind of think I think about it from a little bit of a historical perspective of think back on the history of science. And at certain times in our history, there have been eras where a particular scientific discipline has made incredible amounts of progress in a relatively short amount of time because there was kind of like a click where we started to uh see the world in different way, or there was a tool that wasn't available before. So if you think back to you know, the late 1800s, that was chemistry, where we suddenly realized, you know, we couldn't really turn lead into gold. And there was this thing called the periodic table, and there were electrons, and it really um shifted chemistry.
18:38 Speaker 0 And then in the early 1900s, obviously that discipline was physics and the connection between uh between energy and matter and between space and time completely shifted our understanding of the universe. In 1960s, that discipline was computing, where we get these machines that perform calculations that up until that point only a human was able to perform. And then in 1990s, there was this interesting bifurcation. Uh, on the one side, there was data science that ultimately, you know, that that grew on computers, but also had elements of neuroscience and optimization and statistics, and ultimately gave us modern day machine learning and AI. And then the other side was what I think of as quantitative biology, which was the first time where we actually started to measure biology in a scale that was more than like trap three genes across an experiment that took five years, and that was the first uh microarray data and the first human genome and so on and so forth.
19:37 Speaker 0 And I think this this time that we're living in is the time when those last two disciplines are actually going to merge. And they're giving us um an era of what I think of as digital biology, which is the ability to measure biology at unprecedented satellite and scale, interprets the unbelievable masses of data, different biological sources. Scales and different systems using the tools of machine learning and data science. And then bring that back to engineer biology using tools like CRISPR, genome editing, and so on, so that we can make biology do things that it would otherwise not want to do.
20:14 Speaker 1 Like what?
20:16 Speaker 0 So I think there's obviously, as we said, applications in human health, but I think there's applications in agriculture. I don't think we need to tell anybody anymore, although there's still some people who might need to hear it about the impact of global warming or climate change on our on our world and the fact that we need to have crops that are much more resistant to drought and severe weather. And to
20:38 Speaker 1 feed 10 billion people.
21:08 Speaker 1 Yeah, yeah.
21:10 Speaker 0 So there's that. I think there's, you know, uh biomaterials and so on. There's so many opportunities at this intersection that I would encourage any of you in this audience who are looking for something truly aspirational and exciting to do. I think this convergence is a moment in time for us to make a really big difference in the world that we live in using tools that exist today that did not exist even five years ago.
21:38 Speaker 1 Yeah, with that, I think that's the opportunity at hand. Um we'll wrap up there. Let's thank Daphne one more time.
Previous Engineering Rapidly Reusable Rockets Next AI Revolution: Top Lessons from OpenAI, Anthropic, CharacterAI, & More Page updated August 25, 2026 · transcript and summary are machine-generated