Why AI Cannot Cure Disease Yet | The Next Endeavour 2026

So it's going great and it is a journey because the world that we live in right now is the world where AI bits meet real-world atoms. And I think people often underappreciate just how much complexity exists at that interface and the fact that when you move out of the realm of the world where all of this data is being generated by the minds of eight billion people, images, text, and so on on the web, code that you can generate and test immediately and so on, and you have to exist in a world where data generation and verification that what you're doing is actually working, you're working at the pace of the physical world. You're no longer working at the speed of light. And so I think there is a, I have a great admiration to all of the companies that are building world models where they are actually modeling the physical world with all of its complexity. That's the world where we live in except we have an even one layer extra of complexity which is the world that we're intervening is the world of the human body. So the company that I'm building right now is a company that does AI for making medicines and the fundamental technology at its root is a causal AI model of human biology. So you can make a prediction of what will happen if you perform an intervention in the real world. The real world being the real-world body of individual patients like all of us. What will happen in an experiment that has never been performed. So you don't get to perform the experiment. The experiment is at the end of the journey when it's called a clinical trial and you actually put a drug in a person and you get to see what the experiment has done. How do you make that prediction about that counterfactual world that has never happened with data that isn't actually doing that experiment? And so what we've had to do in the last eight years, and that's why it took eight years, is to collect a unique-of-its-kind data set that combines observational data from millions of people with a causal underpinning in human genetics. And combine that with causal data in experimental systems that we can generate and perturb at the scales of tens of billions, soon to be hundreds of billions of cells. And bring that together to the point that we can make accurate predictions in what we call the virtual human which is a first-of-its-kind we believe AI model for human causal biology making causal predictions in the real world of human beings. And so making data takes a long time, making data at the quality and reliability that an AI model doesn't get hung up on artifacts and batch effects and all of the many things that can make an experimental system go wrong. And then taking that data, collecting it, curating it and training that AI model is what we've been doing for years. Not to mention making molecules against that and doing all the pre-clinical studies that regulators legitimately require before they will let you put a brand new molecule into a human being. And so that's what we've been up to and would be excited to tell you about some of our first programs.

That would be nice actually if someone were to pay me to write. No. I write in my own free time. So I think that there is a lot of enthusiasm and excitement about the role that AI can play in the physical world and specifically the role that AI can play in human health and I think that is a not unreasonable thing and we all would love for AI to do something that helps make the world better as opposed to, you know, all the predictions of doom and gloom. And I think there is, and that has caused a lot of predictions that are aspirationally correct but on timelines that I think are unrealistic. So when you say something like AI is going to cure all disease in 5 to 10 years, I like to think back on Eric Topol's response. Eric Topol is an incredible physician scientist and one of the smartest people I know and he says if you list the number of diseases that we have quote cured outside of infectious diseases and underneath he has an empty box. We have not yet cured a disease other than infectious diseases and a few pediatric cancers. It's hard to cure anything. You can slow disease, you can treat disease. So saying we'll cure all disease in 5 to 10 years is, you know, I think it's a bit optimistic. I think directionally that's what we'd love to do. But I think we also need to understand what is feasible and what it will take to get there. And so one of the main messages that I make in that article is we are now better positioned due to the incredible advancements of technologies like AlphaFold which originated in, you know, in the company that owns this campus to make better molecules, far better molecules, far faster, far easier than we've been able to make. Unfortunately in most diseases we have no idea what to make molecules towards. We do not understand the biology. It is not a chemistry problem. It is a biology problem. And the amount of data that we've been able to acquire for understanding human biology is a diminutive fraction of what we need. And the more so as you get towards the diseases that are much more intrinsically human like anything in neurodegeneration, you can't collect that from animal models. You can't collect that from cells. You need to collect that from humans. And that becomes a much more complicated problem. So I just wanted to create a sense of reality while also maintaining a sense of the incredible importance and aspiration of this problem that we're trying to solve.

It's an incredible number. It's a depressing number and also one that to your point, people were maybe less risk-off than they are right now. I think that there has...

Risk-off means...

It is, it's a term of art. No, it's a term of art. It's an investor term of art especially in the biotech industry where it refers to people being unwilling to take biological risks. So 10 years ago I think there was maybe more openness to exploring novel targets and as the statistics of our industry which remain a dismal, dismal 10% success rate for a novel mechanism that goes into the clinic, people are like why do I want to take that risk? I'd much rather go after a target that, you know, my fellow pharma bro has de-risked for me and I'm going to be second to the clinic and I'm just or third or fourth or 17th. And by the way, there's 37 targets that have over 50 programs against them, which is a shockingly high number. I mean, you can make the argument that maybe it wasn't the first statin that was the best. Maybe it was a sixth or seventh statin that was best, but what are you going to get from the 30th? I mean, not very much. Certainly patients are not going to get very much but we have this incredible what's called target crowding, another term of art, of people not wanting to take biological risk, not wanting to explore novel biology. Now let me add another stat that you didn't cite but I think is equally important: the fraction of diseases that have any FDA approved therapeutic, depending on how you define the denominator, any FDA approved therapeutic ranges from shy of 10% to 22%. That means the vast majority do not have even a single FDA approved therapeutic. That subset that do, a lot of these drugs don't do very much. One of the programs that we're really excited about at Insitro that we're working on is ALS. There are four approved drugs for ALS. How many of you are familiar with ALS? Yeah, a lot of you. Ice Bucket Challenge. Three to five year death sentence. Horrible, horrible death. The drugs that are out there extend lifespan by two to three months. That's all they do. They're palliative at best. So coming up, and so this is in a subset, that small subset that actually has a quote FDA approved therapeutic, just doesn't do very much. So that is where we should be devoting our attention and coming back to my previous statement, it is not a chemistry question. The reason why we haven't made more drugs for ALS isn't because there is a target that we know is going to work and just nobody's been able to drug it. It's because we have no idea what to drug. And so that is the problem that I think presents an incredible opportunity for uncovering new biology that will give us all of, you know, potentially the opportunity to go after disease after disease after disease and come up with drugs that actually work that are actually, another term of art, disease modifying.

They talk to each other all the time and it's actually an incredible source of our innovation how that cross-disciplinary fertilization happens. So to your point about the right data, most of the data that we use to train AI models right now is observational. They tell you what is the state of the world. It's patterns. It's correlations. Correlations are great and if you're trying to produce a manuscript or an image that's awesome but if you're asking what would happen if we did X, that is a causal question, that is a question that involves taking an action in the world and if that...

I'm saying we need AI to be causal if we're going to make predictions about taking action.

I think it was mainstream for a while. If you talk to any statistician or economist or epidemiologist or clinician, they will tell you that correlation and causation are very, very different things. And the place where people often go wrong is because they interpret correlation as causation and they do something where two things happen to be correlated because of some shared confounder or something but acting on one has no consequence on the other. And when you're making a medicine to put in a person and undergoing what is a multi-year often billion-plus dollar clinical trial, you better have your causality right. You better have some conviction that when you take an action that has never been taken in the world before, which is intervening in a particular way in a human being, that that has a causal impact on that person. And so that is a really critical gap in the data that we have available to us. And what we've done in our labs, to your point, in South San Francisco, is we have a causal data factory. We take billions and billions of cells from different cell lineages. Some of them are neurons. Some of them are liver cells, cardiac cells, immune cells. We intervene in them by knocking down a particular gene, amplifying a particular gene and then seeing what happens. And when I say seeing, I mean that literally because microscopy is a significant part of our technology stack. And it's amazing what AI combined with advanced microscopy techniques can see in a human cell, especially when it's a living human cell because we have live cell microscopy. We also look at what happens to other transcripts using sequencing type technologies. And we're collecting the largest causal database of human biology that has ever existed when you look at cells. And then we're combining that with data, with causal data that you get from humans. And you ask how do you get the causal data from humans when we can't do an experiment in humans correctly? So until a clinical...

I could but I won't. Okay. So we created a set of cells that replicated ALS biology by creating sort of an ALS in a cell. We put our microscopy stack on that and identified a never-before-seen disease axis for ALS that showed what happens when a motor neuron develops ALS and gradually dies. And then we deployed our cell-based screening with a causality to identify targets and when you knock them down, made the unhealthy cells revert back to a healthy state. And then we took a whole bunch of different traits that we know exist in human neurons and occur when those neurons die and we showed that those same targets revert those. So specifically in the picture, in the spirit of a picture is worth a thousand words, damaged neurites, those little neurites that emerge from a neuron, when they're damaged in an ALS patient, do not grow back in the same way that they do in a healthy individual. And when you put our drug into those same neurons, they actually do grow back. And that is the fundamental functional deficit that occurs in ALS patients. So in addition to a whole bunch of molecular reversion that we were able to do, we are also seeing a functional reversion. So we are so excited about this program because we believe it could be a truly disease-modifying intervention for the 97% of ALS patients that are sporadic ALS that have what's called TDP43 pathology. And that will be, if I do nothing else in my entire life, that will have been worth it.

Yes. Until you got to 23andMe. I'm showing there's a risk, right? No, there is a risk, but I was trying to be funny. No, I didn't think...