Okay. So, it's just such an honor to be here and have the opportunity to thank Judea for his profound influence that he's had on my career. And as I was deciding what to speak about, do I want to go back and talk about, you know, the work that we did on Bayesian networks or probabilistic relational models, which Tom so graciously described, I decided instead that I would talk about what we're doing today. And I went back to the standard presentation that we give when we talk about our work at Insitro, and I look at every place where we use the word causal. And I highlighted those in the presentation today, which I'm going to go through, and I think you will see just what a profound influence you're having not only on my historical career trajectory, but also on the work that we're doing right now to help bring better medicines to patients. So, you'll see there's a lot of yellow highlighted words with the word causal appearing, and I think that was a very different type of hopefully impact that you're having. Okay. So, I'm going to tell you about the work that we're doing on causal AI for human health.
And I'll point out that as an industry, the drug discovery industry is actually pretty good at making molecules when it knows what the molecules need to do. So, when we understand what the cause of the disease is, we actually have a pretty good success rate. So, if we live today in an industry whose industry-wide success rate is 10% — a new mechanism that goes into the clinic succeeds 10% of the time. If we have causal evidence that supports that target, let's say from human genetics, that probability goes up to 25%. And if we actually know the right causal gene, that probability goes up to 50%. So, the difference between success and failure in this industry is an understanding of causality. And the problem is that that is a rare commodity in this industry. And there's just not enough of those targets to be had, which has led to a very unfortunate phenomenon from this industry, which is that everybody goes after the very small number of targets where we have true evidence of causality, which means that 70-plus percent of the industry is now focused on basically making the 17th GLP-1. And furthermore, people's willingness to take the risk when they don't have this evidence of causality has gone down dramatically, which means that the number of novel mechanisms going to the clinic on an annual basis has gone down from 100 to 30, which is what is making it so painfully hard to get medicines for the very many diseases for which we have nothing.
And so, what we've devoted our work to for the past eight years is to get much better at identifying new causal targets. So, to do that, we've built a whole bunch of things that I won't tell you about very much today. A lot of this has to do with how do you interrogate causality at scale, which I'll talk about in a little bit in a few slides, from human genetics as well as from experimental systems. How do you build hardware to do that, and how do you build the right type of AI that integrates the incredible insights on causality with some of the newer techniques from deep learning. So, really trying to bridge across that chasm. So, we've built a platform that is attempting to scale human causal biology data across different data modalities, across disease states, across biological scales. Algorithmic advances that unlock precision traits, which we'll show you, and causal genes — I forgot to highlight that one — with human genetic support, and we'll tell you how we've done that.
Okay. So, the foundation for everything we've done is that it's very difficult to get at causality if your data is purely associational. And so, the focus has really been on how to get causal data from both humans and from the lab. And in humans, fortunately, nature has done causal experimentation for us. Each of us is an experiment of nature. Each of us has genetic makeup that by and large sits above the phenotypic traits that we display. And so, this is one of those few places where an association is generally causal. And I'll talk a little bit about some of the exceptions to that. And that has created an at-scale resource for getting to causal inferences on the human side. On the cell side, we have the ability to actually generate causal data. It's one of those few places where we can do causal experimentation at scale by actually intervening in a system and seeing what happens when you actually do a gene — like, have a do experiment in real life. And then, what are the consequences of that? So, you can see human genetics anchors all of it.
So, this is our genome-scale causal data factory where we intervene in a population of 100 million cells, perturbing one gene at a time, one in each cell, and measuring the phenotypic consequences of that intervention, where each cell gets its own intervention, and then we are able to measure the downstream consequences of a do experiment in real life at the scale of at this point 20 billion cells and counting. So, across different cellular systems.
So, moving on to the human data, I'm going to give you one example of a place where we've deployed this to really uncover causality at scale, and this is an application in metabolic disease. It's in a place where the causal basis of liver disease were very poorly understood because we have never been able to measure enough human individuals with liver disease to actually get at the underlying causality. And so we basically undertook the effort of expanding the set of individuals for whom we have the ability to infer the causal basis by one and a half to two orders of magnitude by leveraging data modalities that spoke to or are indirectly relevant to liver disease, and then using that to uncover the causal basis of the disease. So one example is instead of measuring liver fat via MRI, which is the gold standard, we leverage things like DEXA scans which are much more readily available. Many of you have had a DEXA scan if you're old enough. Takes about 5 minutes. Costs $50, as well as blood biomarkers from circulating blood. And here is an example that shows that you could actually using modern-day AI, this is kind of that integration, allows us to get to very high accuracy measurements of liver fat from these other modalities. So this is that illustration that you can leverage high content phenotypic data to get the high accuracy of these phenotypes. And then on top of that, now you can basically start to uncover causality in a whole new scale.
So I mentioned that prior to this work, we had a tiny handful of causal genes that were associated with liver fat, seven. Because we only had 25,000 individuals from which measurements were taken. As you start to scale out the ability to impute these phenotypic traits, the number of causal associations goes up and up and up to the point that we now have over 300 to choose from. And we found for the first time causal support for genes that we since then have validated as being successful, as well as multiple new causal factors that drive this disease that have allowed us to identify new therapeutic intervention.
As Judea points out many times, causality is very tricky and you need to be thoughtful in how you disentangle causal associations even when you think that one thing is upstream of the other, often times you can be wrong. And so we ended up leveraging a lot of the techniques from Judea's book in order to — you can see here's a causal diagram, which I'm sure Judea recognizes from his book — that allows us to detect whether having identified an association, say with a biomarker, is that biomarker upstream or downstream of liver fat and therefore is this association causal? And so we use the do calculus in our work in order to get the conviction around have we got the right causal direction.
Coming back to the synthesis between modern-day deep learning and Bayesian networks, we have found that you can leverage representation learning techniques on the phenotypes to disentangle precision phenotypes in human disease. This for example is taking the kind of measurements that you can get from an OCT scan of the eye and really take that and disentangle the distinct biologies that occur within that complex system and give them names. So those precision phenotypes which are much more intelligible, much more interpretable than just a whole bunch of pixel level measurements, we have found empower causal inference from genetics far more than human defined traits. And that too has been a driver for our ability to discover new causal associations.
So that all gives us a growing atlas which powers an AI model of human causal biology which we are now collecting. So this is a matrix that accumulates all of our growing understanding of the probability of when intervening at a gene — as in do gene, as in gene up or gene down — what does that do to the probability of these precision phenotypes that exist across different biological scales. Some of that comes from measuring what happens to those interventions at different types of cells for different types of biologies. Some of that exists in the human at the organ level and some of those are traits that exist in the organism level. So now you can have for a given intervention, for example activating gene 4 or repressing gene 4, what happens at all of these different layers of biology together? And so that first and foremost gives you triangulation of evidence across the causal role that a gene plays in different aspects that are potentially related to the disease. So, combining all of these things to create a holistic, multi-scale view of the causal role of a target in a set of related biology.
We're now accumulating millions, tens of millions causal associations between genes and traits that create these kind of holistic picture of human causal biology as it arises from human genetics and enables AI models of causal inference at human scale. So, specifically, one example that we've recently shown is can we identify from all of the things that have gone into clinical trials, which of them had causal evidence and which of them did not. And it turns out that when you focus in on genes for which you have significant causal support from this atlas that I showed you before, you're able to far better identify targets that are likely to actually have an impact in a human. And what you see here, it's a precision recall curve. This is an audience for which I probably don't need to define what a precision recall curve is, but you can see that you end up with a far better precision recall curve and hopefully end up, for example, sending into the clinic far fewer targets that are not causal for the disease and that are not going to be distinguishable from sugar water in terms of their impact in the disease. And that can be done simply by virtue of focusing on those that actually have a causal impact.
So, I'm going to very quickly — how much time do I have? I have perfect. Give you three examples of programs that actually emerged from this work. The first is the work on liver disease that I introduced a moment ago where we were able to get for the very first time a very large number of associations with significant causal impact on both liver fat and liver fibrosis. This program is going into the clinic for the first real trial, which is a human trial, at the end of this year, but you can see that in animal models of the disease where you're similarly able to do this intervention, there is a significant causal impact on both liver fat and liver fibrosis.
A second program, which is kind of cool, is a causal regulator of brown adipose tissue. Brown adipose tissue is brown fat, a metabolically active form of fat. The more of it you have has been associated with being healthier, but the causality has been hard to come by because you might have more brown fat because you exercise more and exercising more also makes you healthy, so the association has not been proven to be causal. What we were able to do is by measuring brown fat at scale for the very first time in 50,000 individuals, measuring that fat quantitatively, and you can see by the way another potentially causal association, which is although again hard to be sure, that you have more brown fat in cold months and less of it in warm months, which aligns with the biology. We were able to identify the genetics of brown fat and show for the very first time that brown fat — via Mendelian randomization — that is causal of multiple cardiometabolic traits. So, things like kidney function, liver function, cardiovascular outcomes, and so on. It is not merely an association. It is actually a causal fact because we were able to uncover instrumental variables, which are these genes that drive brown adipose tissue, and therefore show that brown fat is upstream of these cardiometabolic traits. And here's the first target that is showing a weight reduction in a rodent from the first of those targets.
And then finally, a program that came out of our ability to actually manipulate and create causality is our program in ALS, which is something we're very proud of. We created a precision phenotype for ALS — what happens to a motor neuron as it develops ALS and dies. You can see that in the videos on the left. Identified a causal pathway that sits upstream of that process and actually is a protective pathway that is buffering the pathology of ALS. And what you see here is a neurite outgrowth recovery of ALS neurons, where the neurons that are damaged in ALS patients do not recover. They do not grow back. And unlike in healthy humans, with the treatment, they start to grow back. And so, we are very excited about this program because any of you who have heard about ALS, it is probably one of the worst things you could possibly die from. It's a 3-5 year death sentence, a horrible, horrible death.
So, to wrap up, what we've tried to build, what we're working towards building, is a causal engine for human health that makes predictions about do calculus in a human. And that gives the opportunity to unlock a massive set of therapeutic interventions for the 75 plus percent of diseases for which there is currently absolutely no treatment. There's over 75% of diseases for which there is no FDA approved therapeutic, zero. And we are desperately lacking causal targets there. And then, of course, there are other things you can do beyond that. For example, predicting which trials you should not invest a billion dollars in, or at the end of the day — and Judea was mentioning that to me as we spoke before this session — which is not only predicting what a target will do in a human, but what will a target do in this human? And that, of course, is something that you can only do with the do calculus that Judea developed. So, thank you very much, Judea, for everything that you've done for us.
