Stanford professor and digital biology pioneer Daphne Koller joins Joel Hellermark for a thought-provoking conversation spanning her early leap into academia, the evolution of AI and machine learning, and their growing impact on biology and medicine. Daphne reflects on the challenges and opportunities at the intersection of data quality, interdisciplinary teamwork, and scientific discovery, offering her vision for how AI can transform drug development and personalized care. With practical insights drawn from decades at the frontier of technology and science, she traces the path toward a future where biology is decoded at unprecedented scaleāunlocking new treatments, deeper understanding, and healthier lives.
Watch the full episode above or read the transcript of their conversation below.
So I, found myself quite bored with school from a relatively early age. And so when I was 13, I convinced my dad that I wanted to start university. Was very reluctant to let me do that. But I said I was really bored in high school. And so he went and connected with some faculty members. And I started to study university when I was 13.5, and ended up doing a lot of my university work in parallel with high school, which basically meant I was carrying two jobs same time, if you will.
But but that was and I was also kind of like the kid at the university, so a little bit like everybody's mascot. but it was a lot of fun. And so I graduated high school at 16 and then college at 17, and I started studying computer science as honestly, it was becoming a discipline. Most universities at that time didn't even have a computer science department.
They had maybe an applied math department, or maybe it was part of engineering. and so I was kind of relatively early into the field of computer science as a whole. I've always kind of found myself being early into the field of whatever it is, early into computer science, early into AI, early into machine learning, and then early into, you know, machine learning applied to biology, what we now call digital biology.
And so what was your intuition there? Why did you find the computer science? So, so interesting?
Because it was so, malleable and yet versatile, so malleable in the sense that you could tell the computer to do stuff. And if you told you what you need to do it in the right way, then it would just go and do stuff that felt very empowering, especially to a young kid. So, you know, I this was very early in the days of accessible computers.
I mean, this was as computers were moving away from, like, these very inaccessible mainframes that only institutions could have. And the first personal computers had come out a few years earlier. And, and so I was able to start programing some of those early personal computers. And it was just incredible to have that ability of power. And I think even at the time, I was able to glimpse just the incredible versatility of that technology and the fact that you could use it for so many different use cases.
And that was very exciting.
And then you ended up at Stanford? Yes. Well, what, what were the early days there like.
So at Stanford? I got into, field of basically reasoning under uncertainty. I'd done my master's degree in reasoning in multi-agent systems, where intrinsically there is uncertainty about what the other player is going to do. And that's a very fun. And I worked on Game Theory and ended up even during my PhD, doing some work, for example, on, multi-agent games like poker and such.
But over time came to the realization that we still hadn't really nailed the problem of acting under uncertainty. Even we don't have the complexity of other agents, just the world itself was out to get us. And so how do we interact with the world? And so moved away a little bit from multi-agent reasoning to just reasoning under uncertainty, which at the time really wasn't I because I didn't embrace uncertainty at the time.
It was in fact anathema to AI to even think about numbers and probabilities. And so on. Everyone was doing logical reasoning and, you know, deterministic systems. And so we were not part of the community. and so when people ask me, when did you start doing AI, what I often answer is, when did I start doing what I was doing?
and so eventually what we were doing, which was initially or what I was doing, which was uncertainty in AI and then realizing that in order to appropriately make decisions under uncertainty, you first needed to have a model of the world. And building by hand, the model of the world was excruciatingly painful and frankly, never going to be successful.
So you needed to learn a model of the world, which is how I got into what is now machine learning. And and it really was the the field of AI first embraced machine learning and then really pretty much became machine learning. So I became machine learning over time.
And when did you start feeling that it could have a very big effect on biology?
So initially I would say, you know, I can't claim to have had particularly perceptive vision at the time. I started to get into AI in biology in, around like the late 90s, 1998, 99, relatively soon after I joined, came back to Stanford as a faculty member, which was in 95 and initially it was because the data sets that were available in core machine learning at the time were, frankly, rather boring and, aspirational.
So how inspired could you get about spam filtering or 20 newsgroups? I mean, it's just like doesn't really excite you. And so at the time I looked around and this was one of the first data sets in biology were coming about, actually the first one that I worked on, interestingly, when you think about it from today's perspective, was an epidemiology data set for tuberculosis or transmission of tuberculosis from one patient to the other, which of course, in the post-Covid era seems like a very important problem.
Back then, the data sets were rather small, but they were still interesting. And then the other and then starting to look around. There were other data sets, like for example, inferring gene regulatory networks from the activity levels of 20,000 genes. This was the day when we were first starting to measure activity levels of gene across the entire genome.
And it was like, wow, this is kind of cool. It's much more interesting than 20 newsgroups. So I started to work on a purely from technical interest at the start, and then became interested in biology and medicine in their own right, and so subsequently had like this weird bifurcated existence as a Stanford faculty member where half my lab continued to core machine learning, published that, you know, NeurIPS and icMl and the other half to basically biology research published in, you know, science and Nature genetics and things like that.
And a lot of my Stanford colleagues, even as recently as a few years ago, was like, so why did you start doing the whole gee, after Coursera? It's like, no, that happened in the 90s. and a lot of my biology colleagues had no idea I was in the computer science department. And so it was kind of like this weird bifurcation.
1st May argue that this is the most exciting time in the history of biology, because at last we might be able to sort of, crack it. Tell us more about, like, why is this such an exciting time and why could this be sort of the equivalent to, you know, what we did with physics and chemistry and so on?
No, I think that's you're asking exactly the right questions. And one thing that I forget where I heard, but I think to me, really kind of captures it is that, calculus is to physics what AI is biology. So it's creates a scientific found, a mathematical foundation that allows you to start making predictions about what is going to happen in the natural world.
And calculus allows you to do that in the physics realm. But it's of course, it's nowhere an adequate for what the complexity of a biological system and AI machine learning on biological data is that predictive underlay or that predictive foundation that I think is going to allow us to, over time, make better and better predictions about what will happen when I intervene in a biological systems such as giving medicine to a patient or trying to create an organism that breaks down plastics or whatever it is that we're trying to solve, which right now we have the ability to make that kind of, you know, prediction in a reliable way.
So the reason why now's the time is really a convergence of two forces. There's obviously the incredible advancements that have happened in AI and machine learning over the last. What is it, ten, 12 years, where we are finally able to build models, take massive amounts of data and make predictions by pulling together threads that no, right now, no human can really understand how the machine is able to do that.
But but the fact of the matter is, it's able to do that. But the other half of that, equation is of course, the data and data in biology has been notoriously lacking in terms of scale, resolution, quality and so on. But that too is changing. So we have now, the ability to via the tools from bioengineering, from, cell biology, generate massive amounts of data at increasing depth, increasing resolution, increasing scale on different types of biological systems.
And on the other hand, we're now starting to be able to measure again at increasing scale data about humans. with electronic medical records and better imaging modalities and so on and so forth. And so that data layer is obviously the critical piece that's necessary in order to drive a new level of understanding in biology. And so those two are converging at the same time.
And one I, one I think really important observation is that people often ask, well, you know, these AI companies and biology have been around, some of them not us, but some of them for 15 years. And where's all that progress? And I like to think about the exponential curve that we've been on in AI in general, where 15 years ago people would have told you about I mean, where's all that progress?
I mean, you keep talking about AI since 1950s and where's the progress? And the point is, on an exponential curve in the early stages, you don't realize you're on an exponential curve because it's very, very flat. It looks really, really fun. And all of a sudden you hit the inflection points like, oh my God, what's happening? And that's where we are today in AI.
For the last, I would say 2 to 3 years in biology, we're probably ten years behind that because the data that has that is a necessary driver for that exponential transformation hasn't been around. But it's coming. It's is already here and more and more is coming. So that exponential curve is merging with the exponential curve of AI and really getting us to the point where I think we'll be able, as you said, to crack, some of the problems biology.
Which leads us to, to in situ and one of the primary focuses for you initially as, as well as it's been on the quality of, of the data sets that, that you, that you create, what were some of the missing pieces in the existing data sets that you were sitting, and how are you planning on addressing that within zero?
So when I started to work in computational biology, back when I was at Stanford, I was like, oh, cool. All these people are generating data and read their papers in there is a supplementary materials and they, you know, not all of them, but some of them publish their data and like, let's let's just go and analyze it and then you start looking at what's there and you realize that, first of all, in many cases, data aren't made available.
It's very poorly annotated. It's, you know, there's a whole bunch of missing stuff. And then when, many of the assumptions about how the experiment was conducted are not even stated. and so and then worse than that, when you start to bring in together what are oftentimes not very large data sets from different sources, even when they look like they're doing the same thing, they are not.
So the experimental design is different, the assays are different, and biological systems are very sensitive to a lot of different conditions. And so you end up with different mishmash of different pieces that are just completely incompatible. And machine learning is really good at latching on to subtle effects of like specifics of one data set or another and getting totally misled by that.
And so I realized that if you really wanted to interrogate complex biological systems, data quality was much, much more important, than in many other applications. And so we the part of the vision behind the CTO is to really build a hybrid company that both generates massive amounts of high quality biological data and then also interrogates it using, AI and machine learning.
When you think about the name Citra, which is probably maybe not apparent to everybody, it's the integration of in silico, which means in the computer and in vitro, which means in the lab, and really merging them together into a single integrated whole. And that's been the vision behind the company, both scientifically as well as culturally, which is really important.
And hopefully we'll get to that later. and that's what we've really built here is that.
And how can you create the bridge between the lower quality data and the high quality data that that you produce? I've seen some work you've done in sort of using embeddings there as the as the bridge. But but how do you assure that you can both use this high volume, together with the quality one secret here.
So, I mean, first of all, what we create here is very high volume because the technologies that have emerged in the last few years allow you to create, for example, data at the single cell level in very large format with extensive automation. And so you are able to create fairly large data sets in a lab such as ours.
Now, in terms of the bringing data from that, that is internal, together with data from the outside, we're very selective. There is a handful of data sets externally that we feel are both large enough and high quality enough that they're worth making an investment. And I think for us, it's been maybe something like on the order of 7 or 8 large data sets that we've brought in, we've made the effort to clean, harmonize, curate each of them so that they are really as good, as good as you can get.
There is a very long tail of data sets that are, you know, just not worth the effort to bring in. They're either not sufficiently high quality or they're not big enough, or they're just not fit for purpose for machine learning. They weren't designed with machine learning in mind.
And one thing you, you spend a lot of time on as well is, is trying to remove the, the biases from, from the data. So there was, you know, the work you did where you had the pathologists that had sort of embedded their knowledge. And when you trade an image and a model on on that it overlooked some of the data.
And you taught it to like, overlook some of, some of the data. because that's what the pathologists historically had had done. Tell us about that importance of sort of removing the human biases from, from the data.
So this is not a data quality issue because I think the, you know, the pathologist scores in that data set were actually quite high quality scores. it was a single pathologist reading all the images. And I think it was a very qualified pathologist. It's just people are taught to look at data in a very certain way, and by omission, they're not taught to look at it in other ways.
And we initially in that project to which you're learning, started out by training the model to anchor on the pathologist scores. And it did actually a pretty good job of predicting the pathologist scores. And we got a fair bit of insight from that. But when we removed that anchor and said, basically just, you know, find patterns in the data, don't think about what the pathologist would say about this image.
Just, just think about predominant patterns. It found a completely different set of things that are potentially equally valuable as what the pathologist looked at. So this is not a quality issues much as the science has created a set of lampposts. And if you let the machine learning anchor on those lamppost, then it'll find stuff under that lamp post versus other dark regions that might be equally valuable.
And do you think this is how we can then, you know, enable the models to supersede the human knowledge to a large extent?
If I you know, I completely think that a lot of the interesting discoveries, a lot of the important, sort of insights that that the machine can use to become predictive will arise from things that people don't know to work for. And then, of course, there is that challenge, which I think is a real challenge of, okay, on the one hand, you want to let the machine be far ranging and creative and not be over anchored to what a person looks at.
But then how do you gain conviction that what it's looking at is real and not an artifact? Because of course, machines are very good at latching onto artifacts, as we've discussed. And so one of the things that we've developed here as a range of different methods is how do you gain conviction around what the machine is looking for?
If it doesn't line up with what the person knows to look for, how do you dress issues like explainable ability? How do you address replicability across diverse cohorts that you're not like? Overfitting to the specifics of a single cohort, how do you, credential is the word that we like to use findings relative to orthogonal data sources that were never presented to the algorithm.
All of these are ways that we use to gain conviction around the insights. When a person can't look at the results and say, yep, that's right.
And do you think you'll ever sort of, you know, a decade from now when we have this very large sort of foundational models, does it make sense for you to fine tune on top of that? So you get sort of the general capabilities of of those models, especially as the the transfer learnings you're seeing now are, are quite intriguing.
Like you train a model to understand math and then it gets also better at entity recognition or you train it on language. It also gets better at code than if you train your code. It also gets better at language. so it seems to be incredibly capable of drawing this analogies on a very sort of high dimensional, view.
do you think it will also be able to do that for, for, for medicine?
I think the answer is yes. Whether at what level you get to sort of really trying to tie together things that are completely disparate at different levels of biological scale. I think that's a question. We're not there yet, but certainly the fine tuning of models, towards specific tasks, which is where we also started on the AI side, is absolutely critical.
So you start by training. I mean, I'll get use imaging as an example because it's so, you know, so it's a place where machine learning is so powerful and is also easy to understand. We started out by training models that start by taking models that were trained on, you know, standard web images, cats, dogs, airplanes, cars and so on, and then using them on images of cells.
And that actually worked surprisingly okay. And we started out there. And then the question is, okay, can you do better than that? Then we started to take some of those models and put in images of cells to kind of fine tune and refine the models. That turned out to work better. But even better than that was to try and understand what are some of the specific characteristics inductive bias, as we call it, in the field of cell images.
And how do we think about training models specifically aim for example, with rotational invariance, which of course you don't turn a car on its head that often, but cells have no there's no like up and down. And so how do you, how do you put those inductive biases into the models or to up the performance even further?
So there is a certain level of kind of as you said, foundation model as a starting point. But then the fine tuning, the refinement and ultimately the optimizations of a specific domain turn out to be really, really important. Now, at what point are we going to get models that were trained on cell data to help make models that are trained on whatever human MRI better?
I don't know, that seems a little bit on the more distant horizon.
But if if we are to learn something from from labs is how much skill matters and and of course, it's a slightly easier problem because there's with more text and then there is, images.
Of cells or even more so images of MRI is. Yes.
But it does make sense that that ultimately you can have this multimodal models and leverage scale to, to, to a large extent for sure.
And we already leveraging multimodal models, we already collect data across multiple modalities, both on the human side and on the cellular side. And we're gaining an incredible, impact and value from that, not only because of scale, but also because the multiple divergent perspectives that you get on a single biological system, where you measure it in different ways, often reveals very different biological phenomena.
And by bringing them together, you really end up with a more than the sum of the parts. so I think over time, yes, we will start to build in more and more multiple modalities on a single level of biology, and then start to sort of expand up and down, if you will, in two different biological scales. One of the favorite examples that we have on our side for this is we had, a lot of work going on in, in cell imaging on the one hand, like imaging of cells.
And then we had a lot of work on MRI. And then in the middle we have work on histopathology like, you know, microscopy images of biopsy samples. And then we started to create bridges. So models that span MRI and histopathology models that relate histopathology to cellular data by doing what we call spatial biology, which is measuring on a tissue slice, not only the histopathology but also individual expression of of individual proteins.
So those are all creating bridges between different biological scales. And the insights that you get from that are just mind blowing.
And how does this ultimately translate to more human interpretable, discoveries as well? Have you found things that the models have discovered that you subsequent were able to understand in a, in a lower dimensional way?
Well, I mean, in some ways, everything that we do ends up having to be reduced to something that I don't know if you'd call understanding, but you it has to be actionable. So a lot of what we end up needing to reduce to is, okay, I need to tell you the exact targets that you would need to modulate in order to create a drug, for particular indication.
That is an interpretable fact. Now, why intervening in this actually has the effect. You can start to create kind of like diagrams in your mind of how that happens. And certainly it's possible. And we've done that to have the AI inform those diagrams like oh okay, I'll give you an example. one of the really think, important advancements that we're proud of and that have been very impactful to our work is the ability to take like a histopathology image, which is one of those biopsy samples that you put under the microscope and collected from every cancer patient pretty much anywhere in the world, and look at it and actually impute from it the expression
level, the activity levels of genes across the genome. So you can look at a sample which was stained with like two colors. And basically imagine you have thousands of different stains on it that say this gene is active, this gene is not active. This gene is active in this part of the tumor. It's not active in this other part of the tumor.
All by using machine learning to kind of create these, synthetic layers, if you will, on top of the images that is intrinsically more interpretable to a person than the image itself, because they can start to then say, well, I kind of know what this gene does. And if I see a whole group of genes that are together in what we call a pathway, then that says, oh, this pathways is like a proliferative pathways pathway.
So maybe this is like a hyper proliferative tumor. You can start to create interpretable stories from that in ways that you'd never be able to do. If you just looked at the image and say, oh, there's a whole bunch of cells and they're stuck together or something, which is just very, that's very descriptive. It doesn't give you any kind of mechanistic understanding.
So that understanding is coming out of the work that we do. And it is definitely a key step towards an intervention that modulates a biology and actually, hopefully helps lead us to a, therapeutic intervention.
And in your term, you have the machine learning engineers and you have the biologists, and hopefully you get them to talk to each other. But they come from two quite different, contexts. How do you get them to work effectively together?
You know, one of the things that I'm the proudest of and how we've built in Citro, is that we are able to take people not only from, you know, suddenly biologists and engineers. There's machine learning scientists, there's automation engineers, there's drug discovery experts. I mean, all of these need to talk to each other. How do you take people from these extremely different backgrounds and create, like a common language?
And there is a number of tools that we've used towards that. One is are you we have very, very sort of, well laid out cultural tenets on how we engage with each other and how we collaborate. And one of my favorite of our values is, that we engage with each other openly, which means being open to naive ideas and naive questions and IV Virgin perspectives, constructively, which means that we all are asking these questions not to be the smartest person in the room, but in order to make the outcome better and with respect for what the other person can bring to the table, and we really enforce that and we hire to
that. That's one thing a lot of the people, just by nature, we attract people who come here because they want to learn about disciplines outside their own. So there is a built in incentive for people to to have that type of dialog. And we did hire a certain group of individuals that are kind of like bridges in the sense that they speak both languages.
Those people are, unfortunately, are quite rare. but but they exist and they can help kind of create almost like a translation layer. And then over time we find that there's actually natural convergence. That is, there's people on each side who just want to learn more about the others who have vowels, who learn to program and do some data science.
And we have people who we hired from, like whatever tech companies who never seen an image of a cell, who can now talk to you about mitochondria and lysosomes and the soma, and you know, which stain is like, they're not going to be biologists, but they can talk biology pretty well. And I think that's really incredible that we're able to bring those, create a larger group of, of bilingual, colleagues.
And do you think we'll ever look back at biologist Assembly? Maybe we look back at linguists and the importance of linguists for a will. Will we ultimately abstract ourselves away? And ultimately it would just, machine learning problem?
I don't think so. I mean, look, the biology, a linguist I don't want to speak to whether we look how we look at linguists. But when you look at the, the discipline of linguistics, it is an interpretation. It is an analysis layer on top of something that emerged organically, which is language in biology. That layer is also how do we, you know, the biology is not there on the surface.
You need to measure it. You need to figure out how to measure it. You need to design experimental procedures, assays, protocols, devices, reagents, Crispr. You need to discover things like Crispr. Those things are not something that you just like. They don't come from nowhere. They come from biologists who really think deeply about how we measure biology and how we surface up the kind of biological data that would be grist for the mill, for machine learning and AI models.
And I would say so that's one place where you're not going to replace biologists anytime soon. I would say that even on the other side, it's it's going to be a while before an AI system just says, here's what we need to do in order to create a medicine. And it's just like, turns out to be true.
I mean, I think there's going to be a lot of interplay between the human and the machine and making sure the right questions are asked, making sure that we, that that we avoid, that we kind of bring together the different pieces into a conclusion now. So I think there is a fair bit of runway here for biologists across multiple different areas.
I'm sure your team would be very excited to hear that.
You know, I think to me, it it and I'm going to editorialize a little bit here. There is a a certain kind of, tech hubris that, manifests in many cases where tech people come into new domain without really understanding the challenges and the complexities. And like, we're I, I, I think the silver bullet is going to solve everything.
It's very powerful tool for sure. But that hubris is often, I think, a real barrier to understanding and to effective deployment of the technology in in a way that maximizes impact. And I think the the respect piece coming into a discipline and having respect for the challenges of discipline and the insights and the, the, kind of us and the technologies that that the discipline has developed.
Now, do you need to come in and blindly accept everything that has happened in that discipline as gospel truth set? This is how we do things now? Of course not. That's the whole point of transforming and rethinking and disrupting how things happen. But that doesn't mean throwing away everything that's there make sense.
So unfortunately, when when we look at medicine, it's been on this sort of inverted Moore's law, the Moore's Law, and as you, as you like to I did.
Not I did not coin that phrase. but yes.
and so we're effectively saying that the cost of the drugs now to if you account for all of the, the failures along the way, they hit 2.5 billion. And it's been on an exponential decay for, for decades now. What do you think are the core reasons for that?
You know, entire papers have been written about that. I think that, one of the big factors is just that we don't understand biology well enough to be thoughtful about how we intervene in it. And so there has been this incredible, you know, failure rate that has just plagued the industry, you know, for, for quite a while.
And, and so of drugs that go into clinical development, which, by the way, is already a pretty late stage because there's years of pre-clinical stuff that happens before the only about 10% at most are successful. That's a, you know, that's a 90% failure rate. No other industry has a 90% failure rate. I think that contributes hugely to the cost.
And so what we're trying to do at insightful is really make better predictions about what is likely to be successful and what is not, so that we can avoid some of those failure rates. Now, I do think that once you start being able to make better predictions, it also helps address some of the other elements of what contributes to the cost.
So for example, clinical development that clinical trial piece is very, very costly. You're doing experiments with people. Every time that you touch a person it becomes complicated and expensive. And there's multiple human touch points and it's by nature less scalable. however, if you are able to, for example, to make better predictions about which in individuals which patients are more likely to respond to your drug, you can hugely accelerate clinical development because you can recruit only the right patients and not the other.
So your trial becomes smaller so you touch fewer people, the effect sizes become larger, and if the effect sizes are larger, you might be able to have the trial be shorter because you're going to see statistical significance emerge faster in the in the timeline. So there's certain things that are incompressible, right? Is Alzheimer's disease is is, you know, slowly progressing disease.
And you have to wait a certain amount of time to see a gap opening between people who have cognitive decline and people who don't. That's just that's how long it takes. But if you're able to identify people who are more likely to respond well to the drug, the gap becomes between the two curves, becomes larger. Between the control curve, the possible curves in that, and then the case curve.
And so you're able to potentially halt the experiments earlier and declare hopefully victory. So I think there's places where making better predictions are going to reduce not only the are not only going to reduce the failure rate, but are also going to reduce the costs associated with certain pieces.
And on on what scale do you think will be able to reduce the cost as a as a function of this?
Look, drug discovery in development is a multi-year journey. This is not you sit down in the basement with, your, you know, little writing, little up, up in the cloud and recruiting customers or users. And, you know, within three months you have, whatever, a blockbuster app. That's not the way it works.
It's a much simpler life. Yes.
Yeah. And sometimes I wonder, why did you do that? And I remember that there's patients who are waiting for us to help them, you know, help them with illnesses.
And then you kind of did that before as well with, with Coursera. So you've had
I did have that. And that was very, very fast moving. This is much slower. you know, I think to really see the effect sort of dramatically shift the needle will take probably well over a decade just because of the, rate at which clinical development has to happen. But I think we're already starting to see some of the pieces becoming faster and cheaper.
and frankly, we're already seeing some pre I, you know, capabilities shifting the needle. So for example, it's been well-established now across multiple studies that, drugs whose target has human genetic support, support from, you know, human association with a change in a gene in a human. And and the outcome, as you can ascertain, in clinical cohorts, they're about 2 to 3 times more likely to be successful in the clinic.
2 to 3 x is a very big deal, right. that type of analysis is not AI, but it's a precursor to AI uses statistical methods. It uses data in a rigorous way to identify a pattern in the data that suggests that this gene is playing a significant causal effect in this clinical outcome. So you can think of it as like a precursor to the kind of methods that we're now employing in a way that digs deeper into the data in maybe more sophisticated forms of analysis.
And hopefully we'll find more such insights in in a more reproducible way. But it's kind of like following that trajectory in that 2 to 3 x. That's that's a very big deal. And it has started to shift the needle on your Ohm's Law.
That's that's massive. What are some ideas that you'd like to execute that are currently limited either by compute or death? but if you had sort of the compute and this time you could run, sort of any experiment, what would you do?
Yeah. So I would say that we at least are not currently limited by compute. I mean the kinds of data skills that exist for us in biology and medicine, even when they're very large, like images of cells and things which are quite, quite large, they're nowhere at the scale of what you, you know, of the large elements that are trained on web scale data.
So it's not compute that's rate limiting. It's data of the straight limiting for us right now, not just for us, for the community as a whole. so, if I had my, you know, ideal data set that doesn't exist today that I would love to have. so there is a data resource out there, which is one of my favorites.
It's called the UK Biobank. It's a data resources started about 15 years ago. They recruited 500,000 people and they measured them every, which way that was available time. And they also saved biologists biological samples so that they could run more assays on it downstream. And now they've been tracking these people for 15 years. And if they were like kind of on average 50, 55 when they're recruited, they're now about, you know, hitting 70.
So many of them are getting sicker. Unfortunately, that happens when you age. And so you're now able to say, what did we see in these people 16 years ago that may or may not have driven, you know, disease state today. and really start to put in also multimodal or multilayer analysis from the different data modalities. And they're constantly enriching and expanding that data set.
So if I had like an ideal data set let's take that. Not for 500,000 people, mostly of European descent, but a much larger number of people, also measured across multiple different clinical data modalities and tracked for a longitudinal outcome. Because ultimately, what we really care about is the trajectory over time. so that we can really start to understand what disease looks like before develops, what does it look like after it's manifest, and what are some of the potential causal factors could might drive differentiated outcome.
And what's limiting us from from getting that data now?
Oh, boy. I think a number of different things. Part of it is that no one has there's very few people have had the will to even contemplate a project of the scale of the UK Biobank. And really, hats off to them for, having taken that on 15 years ago was very, very forward thinking. partly it's I think.
The a very somewhat counter, intuitive norms around health care privacy, make it impossible for, for an individual, even when they want to make their data available for biomedical research. There's no way for them to do that. There's no way for you to, you know, when you go to the whatever the Department of Motor Vehicles, DMV, as we call it here in the US, you can check a button saying, I want to become an organ donor.
so this is, you know, something happens to me. My organs go to a good purpose. There's no way for you to check a button that says, I want my data to be used for biomedical research. It's just it's not there. And so if if it were there and you could have, like, a consented cohort of people who volunteered, opted in to having their data recorded, and you could come in and say, oh, I want to supplement that data by running these assays on whatever blood samples that they provided.
Or I want to, you know, re-invite them in to have a second, whatever imaging study ten years after they joined or whatever, you would have an incredibly rich cohort, because when I talk to patients and when others of my colleagues talked to patients, many of them, although they would love to be able to help drive forward medical research, even if it's not going to be helpful to them just because of timelines.
It might be helpful to patients like them. And we don't we don't accommodate for that in any way. And that's really sad.
And so who do we need to convince to add this, this box?
Well, I mean, unfortunately it's not just a box even for organ donations. You, you know, there is an organ donation network and the database. And so there's a bunch of stuff that needs to happen, infrastructure on the back end. So someone needs to build this. And it has to be I think at the government level, it's not going to happen just by, you know, here and there.
And so that's something that maybe some forward thinking government will eventually doing. There are actually governments, unfortunately not here in the United States. But, that are starting to think about how they can create a nationwide data network. Nations do that are going to be so way ahead of the curve as we sort of start to turn biology into digital biology.
Arguably, this is the most important machine learning problem that the folks could be working on, at the moment. So we want as many of the best machine learning engineers as possible to be pursuing this, this problem. How do you convince the machine learning engineers that this is the most important problem they should be focusing on?
So, you know, machine learning scientists and engineers come in different flavors, and there's some that just want to be in there for something that's quick, like a quick success. but I think there is an increasing number of machine learning scientists who are driven by a combination of I want to something that really makes a difference, that really changes the lives of individual people.
And I would argue that the next shopping app might not be that, but here you have the opportunity to help people who are grievously ill potentially become better. And there's very few of us who have been so lucky that no one in the neither us nor people in our closest sanity have been touched by grievous disease, especially now in the post-Covid era.
And that makes it very personal to people that this is something that, that will really make a difference to the world. I think what's holding a lot of them back is the impression that if you're doing this, you're not going to be doing, you know, I it's at the frontier of your field. You're going to end up doing kind of like mundane, kind of like simple, whatever linear regression, random forest type stuff you're not really going to develop in in your discipline, in your profession.
And what I thought when I talk to machine learning scientists who are of the aspirational type, who don't just want to build a quick app, is mostly to explain to them just how interesting and challenging and intellectually stimulating the problems that we encounter in this field are. That is, you don't have to just work on the big labs in order to do really interesting, you know, technically challenging things that we offer an equally exciting and difficult set of technical problems to work on that, you know, off the shelf models, while often a reasonable starting point, are rarely the end state of the most the best thing that you can build.
And the other thing I also tell them, and again, it attracts a certain group of people, not others, is there's people who just love to learn new things, and just if they feel like it expands their minds. And one of the things that is true here in insider, because of a range of disciplines that we bring together and the fact that those disciplines are themselves incredibly fast moving, like every few weeks there's like a new development on bio engineering or such that allows you to measure biology in a whole different way.
So they will be in an environment where they're constantly challenged to learn stuff that is very much outside of their scope, very much outside of their comfort zone. And so they will just always have something new to get excited about. And then maybe one last thing, which I love because, it really speaks to, you know, the fun of being here.
You can think of the kind of work that, the kind of, approaches that I have taken in, has taken in the last decade is creating a set of increasingly sophisticated and interesting building blocks that you can start to put together in different ways. You can, you know, you have the the convolution layer with the attention layer.
And, you know, you can start to think about how all these things and you build you can program in these abstractions. Biology is on the similar journey where you can say, well, I'm going to edit the cells using Crispr, and then I'm going to put in this type of in this type of microscope, and I'm going to stain the cells in this way.
So you similarly have a set of primitives that you can start to create more and more complex experimental designs here in seed for not only do we have both sets of primitives, we actually interleave them, because a lot of what you see in our lives is actually they are is deeply embedded into the experimental protocols. It's not just the end consumer of the data that comes out of the protocols.
The machine doesn't work if you don't have an AI microscope, an AI driven microscope, and so you end up programing in two languages simultaneously, which is something that really gets people excited about, you know, a new kind of technical challenge.
And it's also a very difficult problem where you have so many aspects you need to get right. As the CEO, how do you find your personal motivation in this? When things get tough.
You know, building companies is definitely not for the faint of heart. there's always very hard moments, whether it's on the science, on the fundraising, on the people, on the business side. I mean, there's always like, something that you have to think hard about, but when you think about the ways in which you could spend your time, I mean, sure, you could go and spend, you know, be a beach bum somewhere and spend time doing pretty much nothing.
But if you want to do something that that's really that is challenging, doing something that is challenging and in service of actually making the world a better place for individual people, where you can point to people, as we were able to do in and say, look, this person, their life would not have been the same if we hadn't given them the gift of Coursera.
I'm definitely one of those.
Witnesses to hear that. But now I'm going to basically say, but you know, you'd probably have been a perfectly you had a fine life without that. And I'm really glad that we were able to bring you to the, you know, to a whole new level of what your ambitions and aspirations were. But in what we're doing here in CTO, a lot of the people that we're looking to help will likely die unless there is a treatment that we that is that someone can provide to them.
And that, I think is the biggest gift that you can give someone is the gift of extra years of healthy life. One of the diseases that we're working on is ALS, and that's a disease that kills you in 3 to 5 years from, like gnosis, it usually strikes people and relatively relatively young, maybe not relative to you, but, you know, your 40s, or early 50s, devastating, devastating disease.
And, you know, we've been in touch with a number of patients. One is actually going to speak to us in our upcoming, very, very well known patient advocate coming to speak to us about, what it's like to live with ALS. it's you look at that and you say, God, if I can do something to help people like that, probably not him, because unfortunate.
Just the timelines are such that not likely that we'll be able to get something in time. But, you know, people like that, you think about, like, if I can, you know, when I get to the end of my career and look back and I can say, look, these are the people who are alive today because of what I was able to or at least lived 3 to 5 years longer.
That's just like, there's nothing quite like that.
And where do you think we can be decades from now when we solve this intersection of machine learning and and biology? We found treatments to, to a lot of diseases. What will that world look like?
So I think right now we don't understand most diseases, we define diseases in a very kind of coarse grained, symptomatic way, like dementia or even Alzheimer's is not one disease. There's different biologies, different types of patients. And I think as a consequence, the ability that we have to provide meaningful treatments to people is is very limited. So if you think back to the early days of of cancer treatments in oncology, we didn't understand cancer.
It was just like a proliferative disease. Cells divide. And so we're going to hit it over the head with a very blunt hammer, which is soft cells from dividing. And then give us chemotherapy. chemotherapy is a brutal treatment that only marginally helps, for, for a lot of people, it's a very high cost. certainly very rarely cure, we no longer think about cancer that way.
We now understand that there's different biological processes that give rise to different types of cancers. And the consequence we have much finer grained intervention tools that are really dramatically impactful for that subset of patients. I think, in, you know, and hopefully in that rosy future that we're talking about, we will be able to define diseases in a way the corresponds to the underlying biology, not of the symptomology, and therefore create treatments that have a much higher effect size will be able to recognize that the diseases that someone is on path to a disease long before the disease manifests in the same clinical way that we have today, because we'll have bio analytes from people,
or maybe even data from noninvasive monitoring like your whatever, your watch and such, and be able to say, oh, someone is likely to need this type of care and so intervene early on. This is a lot easier to treat. And with that scalpel like intervention. And I think we'll be able to provide a lot more people, you know, the opportunity of having a much healthier, longer life.
Now, will you get rid of disease altogether? I don't think that's necessarily feasible, but, just think about how many healthy life years you can give to people.