No, AI Isn’t About to “Solve All Disease”

Can AI really "solve all disease"?

AlphaFold and AI-driven drug discovery are remarkable achievements, but are we aiming at the right problem?

In this episode, I argue that many of today's most devastating diseases are not isolated molecular failures but emergent system-level phenomena. Using the analogy of traffic jams, we explore why precision medicine puts AI on the wrong path, and why the future of AI in healthcare requires learning to steer biological systems, not its individual parts.

Become a Member at dekyon.io/seanmcclure/science-in-perspective

If you enjoyed this episode and want to go deeper, come find me over at dekyon.io. With a premium membership you get the full Science in Perspective experience: interactive visualizations of the key concepts, a study space to learn the fundamentals behind each episode, and an "Ask an Episode" feature where you can dialogue with an AI about the details and start connecting ideas across the whole show.

There's also notes, transcripts, deep-dive articles, and a community of folks thinking through this stuff alongside you. It's where the episode stops being something you just listened to and becomes something you truly understand. So head to dekyon.io, click on the science in perspective app, and become a member today.

Transcript

So Google DeepMind co-founder and CEO Demis Hassabis recently posted on X. He said, "I've always believed the number one application of AI should be to improve human health. That work started with AlphaFold," the tool that they used, "And now at Isomorphic Labs, with the mission to reimagine drug discovery and one day solve all disease, we are turbocharging that goal with two point one billion dollars in new funding." And, uh, you can actually go take a look at the website for Isomorphic Labs and his company, and you can see in the top left really big letters that says, "Solve all disease." So I mean, this sounds pretty good. Who wouldn't think that's a great thing, right? I mean, if you ask people what they're most afraid of, right? Disease is gonna almost always be near the top of that list, right? Cancer, dementia, those kind of diagnoses that change everything. And if someone comes along and says, "Look, AI, with as, as exciting and as powerful that we know it is right now, these last couple years especially," someone comes along that's really at the forefront of that and says, "We're going to cure all of it, every disease within ten years," let's say. Uh, yeah, that sounds good. Who wouldn't want it, uh, who wouldn't want that, and who wouldn't be excited about it, right? Um, so ultimately, this is... You can kinda think of this as AI being used to supercharge how we currently fight disease, right? I mean, because AI is going to come in, it's gonna take the data that we've been collecting for decades and decades now, and it's going to find correlations at a scale that humans can't do alone. Um, it's basically going to supercharge what we do now to fight disease at, at a, at a greater speed, at a greater scale, and what could be wrong with that? And again, we're seeing so much success in AI currently. We've got, you know, the, these Erdos problems in mathematics that are now getting solved. Um, anybody who uses tools like ChatGPT and Claude sees that they're essentially passing the Turing test for all intents and purposes, and it's just being applied across the board. Yes, it, it has its, its issues, but these seem to get resolved almost on a daily basis, and the amount, uh, of, of problems left gets smaller and smaller. So in any case, this is an extremely powerful technology, and so applying it to fighting disease sounds like something everyone would agree with. What could be wrong with that? Okay, well, but turbocharging something only makes sense when the road you are on is the correct one, right? If you take a Ferrari off the fl-- uh, the flat road, then stepping on the gas pedal isn't going to do anything productive, right? The Ferrari is an amazing piece of engineering, but arguably the real innovation there is the flat road itself, that you can actually make a path almost perfectly flat for an extremely long distance. Everything, all the infrastructure that goes into making that possible is the only reason you have something like a Ferrari. The road is critical, right? If you're gonna have something that's super fast and super powerful, it has to be on the road where the speed and the power can be taken advantage of, or more to the point, isn't going to make something worse, let's say. Now you can say, okay, but w- why am I even talking about that, right? Like, why, why would something like AlphaFold, the tool that, that Demis alludes to, why would that be on the wrong road, right? Why would just AI in general being applied to cure AI-- uh, sorry, to cure disease, why would that be problematic? It seems to make sense, right? I mean, if you look at what AlphaFold is doing, which is essentially this protein structure prediction, that is absolutely a worthwhile target when it comes to fighting disease, it would seem, 'cause proteins play a role in disease, a major one. Uh, so getting control over designing them seems like a pretty good direction. I mean, after all, let's keep hammering on that, proteins matter In the area of disease and human health writ large, true, and most drugs actually target proteins, almost all of them. That's also true. Disease usually involves some protein doing the wrong thing, also true. Okay? So AlphaFold has cracked a real fifty-year problem in the life sciences, and that is true. Okay, so all that reasoning is pretty much airtight, except that we make a massive leap when we interpret those facts as proteins play a role in disease, so controlling protein controls disease. And you might say, "Well, you know, maybe people don't really mean it that strong." Well, no, they do, 'cause this is really the overarching message, and especially if you say we're going to cure all disease, right? It's kinda like how people, you know, they kind of know, uh, that, that correlation does not imply causation, and yet people always talk about it as though they d- as though it does. And so a study comes out, there's an association there, but it gets promoted and talked about and even become policy at times because we're basically assuming it is actually a causal thing. So the, the, the interpretation of the facts, so controlling proteins control disease, right? It's smuggling in a severe assumption that a disease is made of its proteins the way a wall is made of bricks. If you fix the bricks, fix the wall, okay? And there's this, this kind of underlying reductionism that I bring up a lot, right? This principle that the best way to understand the whole is to understand the individual parts that make it up. So if proteins are playing a major role in disease, why not go back to the proteins themselves, have some control over it, be able to predict structure, design maybe better ones or augment? Wouldn't that make sense that that's going to make a major difference in disease? I mean, yeah, the logic seems to hold, okay? But there's that assumption again. Is that connection really that tight, right? Just because you involve a thing doesn't mean control over a thing gives you the output, okay? In fact, most disease isn't really a wall at all, is using that brick wall analogy. It's more like a traffic jam, okay? And here's the thing about traffic jams, there is no root car, okay? So you can build the most perfect car ever designed. It can be the fastest, smartest, it ha- you know, can have the best brakes. It's like AlphaFold for automobiles, right? And you can put a million of them on the road, right? And you still get jams because the jam isn't a property of any car. It lives in the interactions between them. It forms, it moves, it even travels backward against the flow of traffic. The traffic jam travels, it can, in an opposite direction than the flow of traffic itself. It's its own thing. It's its own entity, right? It has, uh, a life, if you will, that no single car has. So perfecting the car doesn't touch it. That's emergence, right? And that's what most disease is. So AlphaFold gives you superhuman mastery of the car But the disease is the jam. Okay, so there's something off here. There's a famous experiment, Sugiyama two thousand and eight, where you put a couple dozen cars on a circular track. You tell everyone to hold a steady speed, right? No obstacles, no accident. There's no first car. A jam forms anyway spontaneously. One driver taps the brake a hair too hard, the car behind overcorrects, and, uh, and, and that little tiny ripple ends up amplifying backward down the line until cars far behind are essentially at a dead stop, right? It's like a stop-and-go wave that travels backward through the traffic. Like I said, it, it, it's kind of like a wave of water, I guess you could say. There's just, there's some fluid dynamics there, right? But there's no car to remove Okay, here's the point. The fluctuation that started it is long gone. Traffic jam's not gone, it's worse than ever. But whatever started it, whatever kicked it off is gone, right? The jam is now a self-sustaining thing living in the relationships or existing in the relationships between the cars, not in any car, okay? Now, hold on. You might say, "Okay, hold on, hold on. Let, let's back up a little bit here." I mean, is that the best analogy? There's two kinds of jams, right? If you wanna keep pushing on the, the accident, uh, on the car jams. Okay, there's like an accident jam. That would be like a real obstacle, right? A stalled car, uh, like a discrete cause. So if you remo- i- if you remove it, then the jam clears. Okay, so that's like the brick wall disease, right? It's a pathogen, a single broken gene. This is typically how we hear about disease, right? And we're good at-- And, and those do exist, absolutely, and we're good at those. We can cure those, or at least with a great deal of success, right? But then there's the phantom jam, right? There's no trigger. It emerges from the collective. There's nothing to remove to fix it, and that's your emergent disease, the one we're stuck on, right? And the whole mistake of modern medicine is actually treating phantom jams like accident jams, right? It's, it's like sending the tow truck out to find a wreck that was never there, okay? And you can push, you can keep coming back. Well, hold on, hold on. There are lots of accident jam diseases, those, those single kind of root cause ones, right? There's lots out there. True. True. In fact, you can go to the kind of official documentation that gets released, and it, it's typically quoted, let's say, between seven and ten thousand. You could say kinda single source diseases like that. So there you go, right? Isn't this worthwhile? But those are vanishingly rare. There's lots of them, but it could be one hundred to a thousand people maybe globally that get them, okay? Still a tragedy, obviously. But that's not what we're really talking about or that the healthcare industry is talking about or scientists are talking about when they say the word. It's definitely not what the layman is considering when you hear about disease, right? The diseases filling the hospitals and the obituaries are things like heart disease, stroke, cancer as essentially a class, type 2 diabetes, Alzheimer's, COPD, right? Every one of these are complex, multifactorial, age-driven. They are phantom jams top to bottom. Okay. So let's say we go back to 1900. Let's really be clear about this, right? What killed people around 1900? Tuberculosis, pneumonia, infections, diarrheal disease, right? Every one of those is an accident jam. It's like a pathogen, a discrete cause. It's a car to remove. So we remove the cars, antibiotics, vaccines, right? And we crush them, a world historic victory. And what moved in to take their place at the top of the death charts, right? What moved in to take the-- Well, it's very different things. Heart disease, cancer, stroke, diabetes, dementia, the phantom jams. Okay, so the story that we hear today appearing in all the, well, media and, and like, uh, comments made by Demis or kinda like the, you know, the science utopian nonfiction books, right? The books by people like Harari, Pinker, Kurzweil, very, very optimistic science is gonna solve all the problems, right? That, that story that you hear comes from history. It's not false, right? If you look historically, you're right. We have had large success in curing diseases, but it doesn't relate too strongly to what we are experiencing today, right? Also, a-a-and again, to the point, what I mean by that is, is, is you look at the cancer, you look at the dementia, COPD, Alzheimer's, you look at these diseases, they're not accident jams, right? They're phantom jams. They're emergent. They're multi, multi, multi-causal. There is no one car you can remove to clear the jam. By the way, also a huge share of that infectious disease victory that we keep talking about as though that's the, that's the world we live in today. That wasn't really-- A lot of that wasn't about magic bullet drugs at all. It was actually clean water, sanitation, nutrition, right? That's more systems level stuff. So even the accident jams were beaten partly by changing how the whole system relates, not just by pulling out cars that were problematic. So keep that in mind because it's gonna bring us to what I think is the proper way forward later on. But here's, you know, here's the thing. Medical progress now, it, it didn't slow because we got lazy. It slowed because we ran out of cars to remove. We're fighting different diseases now, right? We're in this kind of, uh, epidemiological transition you could say, yet we're acting like we're not. When we say things like solve all disease, use tools like AlphaFold, supercharge what we've been doing all along, just throw more data and scale it more, it's like being on the wrong road though. AI is powerful. Disease should be cured. We do historically have success with it. But bringing all those together and then interpreting that as, so let's just go ahead and apply AI to what we've been doing all along, I hope you're starting to see now there's something problematic with that. Because again, the diseases that are coming in at the top of the charts aren't like the ones in 1900. They, they're not going to be beaten using the same kind of success story where you can just go after a root cause or even a single gene and, and try to identi- So when you, when you look at tools like AlphaFold that have this precision The problem is precision isn't the road to be on. It's not a precision problem, right? Think about how medication actually works, 'cause a lot of people, I think, don't actually get this, right? People assume if you take a designed drug and it targets the problem, right, um, that that's how it works, right? Like you, you t- pharmaceutical companies, they put bill- you know, millions, billions of dollars, let's say, they design a drug, and then it targets the problem. But that's not really true, actually. What happens is the body is shifted into a different equilibrium that just happens to not have the symptom you didn't want, okay? So it still works, right? Great. This is great. Okay, but it's not, it's not a targeted thing. It's not really a precision thing. Precise in that it might shift the equilibrium specifically so that the symptom you don't want goes away, okay? But that's also why the side effects are so often quite numerous, right? The shift into a new equilibrium is not a costless thing. It's a risky thing because the drug trades one whole body state for another, and you don't get to keep only the parts you wanted, right? You take the entire new state, side effects and all. Moving the system is never free. It could still be worth it. It could still be a wonderful innovation. It could still absolutely improve the quality of your life. I'm not arguing against any of that. But let's just understand what it is. It is not a precision target thing. It's a take this, shift your body into an equilibrium that's not natural, although you might think the disease is not particularly natural either, so again, cost-benefit. You shift it, but it's not just the symptom that's going away. You now have to live in this new state which your body is not really naturally supposed to be in, and there was a cost to doing that. So, so Demis's pitch assumes, to use that as an example, that, uh, side effects are a targeting problem, right? If you make the molecule precise enough, then maybe those side effects will disappear. But if side effects are the cost of moving a coupled system, they aren't a precision failure you can design away. They're intrinsic to perturbing a network, right? The side effect-free designer drug is, is really the reductionist fantasy in its purest form. It's just quite simply not how things work, okay? Uh, let's think about, you know, you hear about the Human Genome Project, and you hear about genetics all the time, and it kind of-- really solid science, fascinating stuff. It always kind of gets portrayed as though now it's a precision thing. Like, now we can go after specific genes, and we can turn things on and off, and some of that is true, right? But think about the fact that we read the entire human genome, and then we sequenced millions of people, and then we went hunting for the gene behind top-of-the-chart diseases: heart disease, diabetes, schizophrenia. Well, we didn't find a gene. We found thousands, each nudging risk by a hair, all wired together. And the harder we looked, the more cause spread out across the whole network instead of concentrating anywhere, right? Pritchard's group named it the, uh, the omnigenic model in, in 2017, I think it was. Th-that's not a failure to find the root cause. It's proof that there isn't one. The cause is the web, right? It's a holistic thing. So we went looking for the broken brick, right? And what we found was that there is no brick. The cause was spread so thin across so many genes all talking to each other that no single one of them is the disease, right? The disease is the conversation between them. Let's take Alzheimer's as another example. The whole field bet that, uh, amyloid plaque was like the broken brick, right? And then we finally built the drugs that clear the plaque almost completely, right? Lecanemab, uh, uh, don- donanemab, drugs like this, they work. The plaque comes out, and the disease barely moves. Modest slowing at best. Real side effects. No cure. We hit the molecular target about as cleanly as anyone could hope to, and the patient still has Alzheimer's. The target wasn't the disease, okay? So what do we do about this then? Right? Because I'm, I'm argue-- you know, I'm, I'm looking at something, you know, Demis Hassabis' post on X, and it sounds very great, and I'm a believer in AI technology. I talk about it a lot. I'm involved in it. And, and I also think it probably is a good tool to apply toward the problem of, of, of disease, okay? But this notion that if you look at his approach, many people's approaches, it's just supercharging what we already do. We, we have this notion that there's like single causes to things, and it's a precision, and if you can just get more precise, if you can just design a better molecular, uh, you know, target or, or, or design a better molecule itself so that it targets the problem, it's just the wrong paradigm. That's not, that's not the situation we're in with these diseases. It's not 1900. It's not tuberculosis. It's not single gene problem. It's not single pathogen problem. This is extremely complex, emergent, multi, multi, multi-causal problems that don't have a target, right? So what do we do? I mean, how, how do, how would you fix the phantom jams, so to speak, right? The car jams. Well, not by finding a car. You fix them by changing how all the cars relate, right? You coordinate the spacing, you smooth the speeds, and, and if you do that, let's say enough, then arguably the wave never forms, right? So the poi-- and, and that's not to be too optimistic about that approach, but the whole argument here is that that is a systems level intervention, right? Remember when I said, you know, a huge part of even the 1900 diseases was sanitation, clean water, and better nutrition, right? It was actually systems level stuff, not precision stuff, right? System level intervention aimed at the interactions, not the parts, right? Which is exactly what curing an emergent disease would actually take, and exactly what AlphaFold working down at the level of the single perfect car can't hand you, okay? So let's go back to, to Demis's, uh, comment and his AlphaFold tool, which by the way, he won the Nobel Prize for, right? This work. Uh, AlphaFold predicts a static structure, right? What one piece looks like. But what we need is to steer a complex system toward a winning state, finding a strategy that no human wrote, right? This is the systems mindset, right? DeepMind invented the best steer-a-complex-system-toward-a-goal engine in history, and then when they turned to the disease, they reached for the structure prediction tool instead of the system steering one, right? So, so Google DeepMind and companies like this, they, it, it, it's-- they don't just use AI. In fact, most of AI has not been for precision at all, right? AI is, in some sense, the antithesis of precision. The way the learning actually happens is very broad. It's very, uh, you know, billions of parameters. It's actually a much more holistic thing. That's why you can't peel back the layers and go to the root cause of the input that produced the output, 'cause there's no such thing. It's actually what AI is, right? So what we need are tools that can predict how the whole network reacts to a perturbation, like a change, like if you bump it and it starts to wobble. Get good at that. It's a more of a global kind of thing, right? Discover what shifts the system to health with, let's say, the fewest downstream consequences. It's not gonna be a perfect thing. Not gonna be, not gonna be totally deterministic, right? But that's the kind of epistemological and, and quite, you know, scientific, biological, life science approach that we should be doing. That's more like coordinating the traffic, not trying to perfect a car, okay? So the only version of solve all disease that's even coherent is really the one that's built around a systems target, not a molecular one, right? What we need is to build a tool that can make decisions against a model of the whole system's dynamics. Now you can say, "Well, okay, but I mean, wouldn't that be too difficult? Like, wouldn't the model have to be like a replica of the body? Like almost a perfect replica, simulating every molecule, every reaction, right? The whole internal machine. I mean, that would be hopeless. Well, yeah, that would be hopeless, right? But that's not what steering an emergent system takes. Okay, again, you're falling for the trap of reductionism when you think you need to simulate all the things, uh, at, at a low level, right? Emergent systems collapse actually onto a low-dimensional description, right? Sometimes I like to say complexity is simpler than simplicity. When you get too simplistic about things, you think all the little individual pieces add up to make the whole. That's not what emergence is, right? Right, the output is actually greater than the sum of its parts. But more to the point, the levers that you pull are actually not that numerous in complex systems, 'cause there's only so many things you can pull in their system level, right? Um, so, so it's a low-dimensional description that actually controls complex systems, right? The substrate can be astronomically complex, while the governing variables are actually quite flu-- uh, few, right? You don't track ten to the twenty-three air molecules to do thermodynamics. You track temperature and pressure, to use a simple example, right? Or rather, a complex example. You don't model a driver's brain, right, to, to dissolve a traffic jam, right? A traffic engineer would move a couple of macro levers, ramp the metering, speed harmonization, and the flow would start to reorganize itself. It's kind of like letting the system do what it does best, and you only do top-level stuff. Again, that is absolutely how today's AI works, right? We don't design the systems the way people think. They're not programmed. They're not rules-based. You're not instructing it what to do. You do top-level scaffolding things, a few meta-level levers, and then you let the system run millions of iterations, and it self-organizes into the thing that you want, right? The, the, the driver's internals run themselves. You operate on the handful of variables that steer the collective, okay? Again, this, this is how deep le-- we're talking about deep-- using AI, deep learning. This is how deep learning actually works. Nobody hand sets the millions of weights, right? You specify the architecture, the objective, the data, the meta-level, right, stuff, the levers, and the internals self-organize over millions of iterations. We control the constraints and the target. The system finds its own way there, right? We control the constraints and the target, and then the system finds its own way there, okay? Similar to, to market dynamics, right? No, no specific person is setting the price. Where does the price come from? It's kind of an ill-posed question. It's not really coming from anywhere. It's not a root cause. It's an emergent thing that materializes, that precipitates out of a massive level of collective action, right? So the, the kind of meta-level model isn't a fantasy waiting on some breakthrough and simulating all of biology. It's a learning problem of exactly the type our modern tools already solve, quite frankly. It's just aimed at the emergent coordinates of the body instead of the static shape of a protein. So AlphaFold as a tool, it learned, you could say, the manifold of protein structure, kind of like the thing you were supposed to learn to get protein structure. Call that the manifold. Yeah, the same kind of machine pointed one floor up could learn the manifold of the body states, and that's the map you would actually steer on. I don't know if it's the AlphaFold tool that's gonna do this, but I'm trying to show you the difference. You-- It's not the technology that's actually wrong. You're just on the wrong road. You've got to go one level up to the meta. That's the map you would actually steer on. So they're not missing the capability so much, right? I argue that they're simply aiming it one floor too low. Okay? All right. So that's it for this episode. And, uh, thanks everyone for listening. Of those of you who are watching, again, as I like to remind people, if you enjoy this episode and you wanna go deeper, come find me over at dekion.io, D-E-K-Y-O-N dot I-O. Uh, with a premium membership, you get the full Science in Perspective experience, interactive visualizations of the key concepts, a study space to learn the fundamentals behind each episode, and an ask-an-episode feature where you can dialogue with an AI about the details and start connecting ideas across the whole show, okay? Because the, the, the Dekion platform is aware of-- has all the transcripts, has all the understanding, has all the concepts that I talk about, and it can make connections between them. It's a really good learning tool. Of course, there's also notes and transcripts, deep dive articles, and a community of folks thinking through this stuff alongside you, right? With the public forum. So it's where the episode stops being something you just listen to and becomes something that you truly understand. So head on over to dekion.io, click on the Science in Perspective app, and become a member today. All right, everyone. Thanks again for listening, and for those of you who are watching, until the next one. Take care.