Curing All Disease with AI with Max Jaderberg

0:00So AI was not satisfied just whooping our ass in chess and in jeopardy and everything else where it looks like brains mattered. It's now taken over our physiology. Well, no, you pointed it in a good direction, aimed it at a good place, and we're getting somewhere. To solve our diseases. Yeah. So now it's going to cure us of all disease before it makes us its slaves. Because we need healthy slaves. All that and more coming up on StarTalk. Welcome to StarTalk, your place in the universe where science and pop culture collide. StarTalk begins right now. This is StarTalk Special Edition. Neil deGrasse Tyson, your personal astrophysicist.

0:48Special Edition means we've got Gary O 'Reilly in the house. Gary. Hi, Neil. Former soccer pro. Apparently. Yeah, and soccer announcer. Yes, definitely. And you still do that, don't you? I do. Chuck Nice, baby. Hey, announcing that I know nothing about soccer. You're in my club then. Announcing that you are American. American dog on it. Real football. Violence. So we're talking about AI today. Yeah. That's a favorite topic. We revisit that often. Only the future of the entire world. AI as it matters in biology. Oh, wow. Now that's a big deal. I know. Uh-huh. I know. Because people thinking about, you know, composing your term paper.

1:30Right. Or winning a chess. Right. But it's got a whole frontier ready to be explored. Yeah. And so tell me what you and your producers cooked up today. Okay, so we've been on the case to get these guys involved for some time, but they are so busy. So here we go. I'll say it. I'm made of proteins. Yes. You're made of proteins from strings of amino acids that fold into shapes that put all together form us. But there's a fundamental problem in biology that has implications for all of medicine. How do these proteins fold up? Oh. For this solution, we looked at AI and a Google DeepMind tool called AlphaFold.

2:12The second iteration of AlphaFold 2 won the Nobel Prize in chemistry last year for answering this very question. Who knew AI was smart, huh? We should have won all the Nobel Prizes. Yeah, just give them to them now. Just pack them all up. Pack them up. Now, the Isomorphic Labs, together with Google DeepMind, developed and released AlphaFold 3. Yes, we're on the third iteration. And that was last year and applied these new AI models for drug discovery. Oh, that's great. All right, so think this through. Could our next generations of treatments be computer generated? Oh, yeah. Oh, by the way, Neil, let's introduce our guest.

2:48I will. We've got Max Yoderbergh. Did I pronounce that correctly? Yeah, you got it right. Let's hear you say it. Let's hear you say it. Me say it. Yeah. Max Yoderbergh. Oh, he got it perfectly. That's what I got. He copied you. He was practicing. He was practicing. So you studied AI at Oxford? That's right. That's right. Now I hear that's a community college? Oxford Community College. That's exactly right. In Oxford, England. Yes. So I specialize in deep learning algorithms. I got your little bio here for understanding images. Nice. That was a big advance when a search can go into an image. I thought I'd died and gone to computer heaven when that started.

3:30Yeah. I mean, this was 10, 15 years ago, back before AI was cool. Right. where you talk about AI and it's something from a sci-fi book, but understanding images and videos was the big thing at that point in time. We couldn't actually do that very well. I searched my 9 ,000 images on my computer for the word telescope, and it found telescope written in Chinese on a photo taken at an angle in one of my images when I was visiting China. This is it. During my PhD, we took all of the BBC's back catalog and we ran my algorithm across it and created a search engine so you could pull up footage from decades ago that had this text or these objects.

4:15That's seriously some archive. So here's what I'm interested in. When you do that, do you tie? So when the AI is looking at an image, it's not seeing the image the way we do. We're not even seeing a whole image. Our brains, we're really just intuiting an image when we see it as our brain. Exactly, exactly. That's how we do it. But the AI actually - It's like a holistic processing. It's just a holistic processing. AI actually sees the image, and what it's seeing is pixels. That's right. And really all it's doing is just, oh, this pixel, this pixel, this pixel, in this arrangement, that's this image.

4:46So do you tie that to language, and that's how we search, or is the search just the AI knows the actual image itself? This was like the big breakthrough back then when I was doing my PhD, and this is what deep learning as well is all about. You can imagine if you have this image full of pixels, how do you actually code up how to read text from there? How do you tie it to the language of the text? It's unimaginably hard to code that up by hand. So instead, what you do is you put these, what they're called neural networks. They look at all of the pixels of the image and you give it lots and lots of examples of images that have somewhere it's got the text in it and you tell the neural network what the text is.

5:29And the neural network, through lots and lots of training, starts to work out its internal algorithm to extract the information from these pixels, piece it all together, and spit out the actual text or spit out what the objects are. Wow. So you're currently chief AI officer at Isomorphic Labs. This is a biology place. That's right. Do you have any biology in your background? Formal biology? No. No, okay. So they want you for your AI. That's right, that's right. So I was at a place called DeepMind beforehand. Google. Google DeepMind. Yes. Exactly. I was there for a long time. I absolutely love this core AI technology called deep learning.

6:05That's what I've been developing my whole career so far. At DeepMind, we were working on some crazy stuff, learning to play chess and Go and beating top professionals at games like StarCraft. Back then, it was about... All of it was a big deal. Yeah, and because the world didn't know what AI was, so we were trying to prove that this was even a thing. It seems crazy now, but back then, it was just proving that this was actually a real thing. But at the core, I love this technology. I want to see it have profound impact on our world. And I was thinking these things. That's where it begins. Yep. That's where the Terminator starts.

6:38It's always the innocent dreamer who says, this can change the world for such good. And it's in my closet now. Would you like to see? And then it's always like some evil businessman who's just like, with my weather machine, I will one day rule the world. Apart from that, carry on. The good thing is there's some pretty strictly good applications of AI that we can drive. Demis Asabis started Isomorphic Labs, spinning that out of Google DeepMind to really think, how can we apply AI to actually completely solve all disease? Okay, so it has genetic links back to its origin story. Exactly. It was in DeepMind.

7:18I feel better about that now. Yeah. Okay. You happy? I'm happier now. Yes. So I moved over as part of that founding team to head up AI in this space. And it's been about three and a half years now. It's been a crazy journey, but it's fascinating. It's so much fun. So you've got this AI expertise, and AlphaFold spins off this biological application of it. First, tell me the word isomorphic. What does that mean in biology? Isomorphic is this technical term, which is a one-to-one mapping of space, right? And the reason we're called isomorphic labs is really that we believe that biology is really, really complicated.

7:53In the world of physics, we can write down equations for physics with maths. And maths is that perfect description language for physics, but you can't really just write down equations in maths for biology, for the cell, it's just too complicated. — Biology is the most complex expression of chemistry that we know. — Oh, that's great! — There's just so many moving parts. — Did you just make that up? — Yeah. — So we're looking for a Rosetta Stone here, for the language of biology. — Exactly. So what could be that perfect description language for biology? We believe AI and machine learning is that.

8:23So that there could exist an isomorphism, a mapping between the biological world and the world of AI machine learning. Hence the name. Gotcha, gotcha, gotcha. All right, so tell us about protein folding. Because when we learn about chemistry, we learn about chemical reactions. And we're not really taught that the shape of the molecule should have anything to do with anything. It's just what is the chemical symbol. and when you write down the chemical equations, there's no shape in there. There's just what elements and molecules comprise it. And those equations don't really ever represent the three-dimensional nature.

9:05Exactly. You don't even know if it has handedness. Yes. Right? So take us from there. We think about proteins. Proteins are these fundamental building blocks of life. Yeah. They're inside of everyone. They make up everything we have, basically. And they're made up of what's called a sequence of amino acids. Each amino acid is a molecule. There's about 20 different amino acids, and you put them together in a long string. Ever or just in life? In life, you can have non-natural amino acids as well that you can make as well. You can make them. You can make them and actually use those for drugs sometimes.

9:37You string these amino acids together, and that becomes a protein, but they don't exist as these strings. they fold up spontaneously in the cell to create these 3D shapes. And why that's important is that these proteins, they're basically molecular machines. They don't just exist by themselves. They actually create these little pieces of machinery. They interact with other proteins. They interact with other biomolecules like DNA and RNA. That interaction is a shape-fitting. Exactly, exactly. So these proteins... It's a puzzle. It's a 3D puzzle. It's a 3D puzzle. Exactly. It's a 3D jigsaw puzzle.

10:12And it's not static. Which is way harder than a 2D jigsaw puzzle. And these are not static things. It's not just static puzzle pieces coming together. They change shapes. So something comes in contact, and that opens up something else on the other side of the protein, which changes the machine, and on and on it goes. And that's what I was going to ask you. What speed is this folding taking place? Is it continuous? Once it folds, that's it. But you've just told me, no, it just keeps moving through the whole thing. Yeah. These are really, really complex dynamical systems. Yeah. You know, composed of thousands, millions, trillions of atoms within our cells, unfolding over the course of, you know, microseconds and beyond.

10:47And this dynacism that you're talking about. What word is that? Dynacism? That's a word? Isn't that a word? Dynacism? I think I just made it up. Mism. Oh, dynamism. No, dynamism. Dynamism. Thank you. Thank you. I'm correcting grammar. This is a first. This is. Dynamism. Dynamism. That's a new kind of dinosaur. Dynamism. Not dinosaur. Dinosaur. Not dinosaur. Not dinosaur. But the dynamism that you're talking about within the cell. When you look at each one of us, since each one of us is so different, even though there's a general execution and blueprint, we all come out so different. Is that part of the process that you are looking at and mapping?

11:31I would say to a jellyfish, we all look identical. True. Okay? They're not saying, oh, is your skin color slightly different or you're slightly taller? You're describing functions at a cellular level. Is your job to understand that or is your job to figure out extra ways to fold proteins that maybe biology has yet to even figure out that can then solve problems that we encounter that the natural universe has not? Ooh, that was a good question. You're happy with yourself. I'm so happy with you, sir. I'm so happy with us today. Yeah. This is really interesting. We have these little molecular machines, these proteins, and we care about that 3D structure and how they work for two reasons.

12:17One, we want to understand how our cells work, because if something goes wrong with that, which is the case for disease, then we want to understand, okay, where do we actually need to go in and start fixing that? And how we can stop it from actually going wrong in the first place. Exactly, exactly. So that's one thing. And then when we think about, okay, how can we go and fix that? What we're actually saying when we're doing drug design, we're saying, can we create another molecule that will come into the cell and actually start modulating these molecular machines? This drug molecule is going to actually attach to this protein over here, and that's going to cause this protein to change shape, for example, and so it won't operate how it normally does.

12:56And so we stop that protein working, or we make it work better. These are the sort of things we do in drug use. It reminds me of messenger RNA vaccines that we developed for COVID. Yeah, you know, there's so many different types of molecular mechanisms that we take advantage for the drug design. Wow.

13:23I'm Ali Khan Hemraj, and I support StarTalk on Patreon. on. This is StarTalk with Neil deGrasse Tyson.

13:38Are the folding proteins generally following a set pattern in the way that they do fold and you're able to map them and when they misfold that's when you're able to flag that up or have I just reinvented something or talked rubbish? That would be cool. No, no, you're onto something. That would be great. So, you know, like the, I mean, the amazing thing is that we can actually, turns out, predict how these proteins fold. So they are modeling that process. Yeah, we're modeling that with deep learning, with neural networks. That's what AlphaFold, you know, and all its generations are all about. And that means that we can actually just take in a sequence of amino acids, knowing nothing about this protein before, and then get out the 3D structure.

14:20And normally this would take people months, if not years, to work out this 3D structure. So how is it that alpha fold knows how a large molecule wants to fold? It's got to know that in some way. It's learned this from a few hundred thousand examples. So chemists, biochemists over the last 50 years, they've been working out these protein structures by hand. They've been literally synthesizing protein, crystallizing it, then shooting x-rays at this to look at the electron scattering. And from that, you can resolve the protein structure. It's a pretty hard process. But people have been doing that.

15:02That's your way to photograph what the shape of the molecule is. That's your way to photograph it in reality. It's basically an electron microscope at that level. Yeah, similar. It's like electron scattering. Yeah, exactly. And so people have been doing that for the last 50 years and depositing these structures. and now we've taken all of that data and trained a neural network to go just from the input of what is this molecule description to try and predict all of that data. And the amazing thing is, and this is really remarkable, is that you can then train this on the last 50 years of data. That's a couple of hundred thousand protein and biomolecular systems.

15:40But you can apply it seemingly to everything we know about in the protein universe, in the proteome. Well, it has the proteome. Oh, we like that. Oh, proteome. Proteome. Yes. Proteome. Oh, well, I like that. So how accurate is AlphaFold and AlphaFold with on the third iteration with its predictions? Because AI has been around a little while, as you've already said, and you're not the only AI tool that's out there, but how accurate is this particular tool? Yes. So AlphaFold 2 was that big jump where we started to get experimental level accuracy for just proteins. And that's what won the Nobel Prize in chemistry.

16:18Oh, so you balance it off against empirical experimental... Yeah, the benchmark is doing the real lab work itself. Right. So alpha-fold 2 reached that level. Now alpha-fold 3 expands from just proteins to incorporate other biomolecular types. So proteins with other proteins, proteins with DNA, with RNA, with what's called small molecules, which are... They go to the neighborhood, they start mixing all that up. Or maybe not the neighborhood. Maybe that neighborhood gets a little bit of an upgrade. No, that's when you make the superhuman. So everybody thinks it's going to be the Terminator. It's not going to be the Terminator.

16:53It's going to be the superhuman. And then they're going to be like us. And they're going to look down on us and go, you know, why do we need you guys? And that's it. So anyway, you're able to predict these. And have you actually taken any of the modeled predictions and made the proteins? Yeah. Oh, yeah. Or tell us where you expect these to lead to new and innovative drugs. Because otherwise, it's just a puzzle exercise. Yeah. It's a great Lego set. Lego set. We want the guests to enjoy this. Yeah. It's like, oh, my God. How much that Lego set costs? Only$10 billion.

17:36Sorry. Go ahead. Yeah. Yeah. So if you take a particular disease, and we think that we can actually solve this disease by modulating a particular protein, the question is how we do that. So we design a drug molecule, and we want it to fit to this protein in a certain way. And so this is where traditionally, you would have to actually either just guess or go into the lab and crystallize each one of those combinations and then photograph them and see if it worked. But now you can model it and the AI can do a thousand of those in like a minute. Isn't this what you call the target proteins? Yes. Wow.

18:17So if you know you've got a certain target protein, do you not then run that against a list of drugs and think, yeah, this one, drug A works better with this or maybe it's drug D or whichever letter of the alphabet you're on. And now we become the sort of detective. And has this alpha fold three produced how many clues and how many answers? or are we still grappling with? Yeah, instead of trying to figure out the drug, the AI actually figures out the drug for you. Drug discovery. Well, exactly. If you let the guy speak, we might hear the two of you. We're figuring out this whole industry ourselves, Gary.

18:52Yes. Yeah, this is exactly where it's going. So we can start actually rationally designing these drugs. Traditionally, you would take a million random molecules and you would just throw them at these proteins and see what sticks. And that's how so many drugs have been created historically. You go back further and you're sifting through mud to find these sort of molecules. Is this why there's been such a low percentage of success rates with the sort of drugs for whatever the problem is? That's part of it because we don't necessarily understand how these molecules are working. But with something like alpha-fold 3, you can put the molecule, put the target protein into the system, into the neural network, and you get out the 3D structure, and as a chemist, you can start to understand, okay, how is this small molecule drug modulating this protein?

19:47Now, still the problem is, well, how do you find that small molecule in the first place that's going to be good for this protein? You know, it's estimated there's like 10 to the power of 60 possible drug-like molecules out there. That's, you know, 10 with 60 zeros. Yeah, people know what 10 to the 60 is. Yeah, yeah, yeah. Don't be rude. Even if you had the perfect alpha fold, you'd have to run that across 10 to the power 60 molecules, which is just computationally impossible. It's unfeasible. Until quantum computing. And so then what we need is something that we call a generative model, or an agent, which is able to actually search through that space, understand that entire molecular space, and come up with molecule designs for you.

20:35Oh, because the 10 of the 60 is if you just did it randomly. Right, right. Exactly. Right, right. If you don't do it randomly, then you can— But randomly is the state-of-the-art method. That's how people do it. I mean, it's how people currently do it. It's how people currently do it. Oh, cool. Well, he called it state-of-the-art. You're the state-of-the-art. Right, thank you. Let's use the word properly here. So what if the protein turns left when you've mapped it to turn right? is that when we have issues that even alpha fold 3 has a problem with exactly these are not perfect models at the end of the day they're very very accurate but they will make some mistakes so you still do currently need to go into the lab occasionally but the amount of lab work you have to do is so much less right and you and often you can find the area of molecular space where these these models work you know really really well and we then go out to the lab later later down the line and we crystallize these things and we see, yeah, this is perfect mapping of what the model predicted.

21:35So back to an earlier point. In the old days, like last month, the pharmaceutical companies, big pharma, would spend millions, maybe not quite a billion, hundreds of millions of dollars developing a drug. We think that holding aside what might be abuses of pricing, the fact that there's some truth to this first pill cost$50 million. The second pill cost$0.10 because they had to research to get the formula for that first pill. If you have narrowed the search space, then the cost of developing that first pill can be manifold smaller. It costs, on average,$3 billion to create a new drug. Wow. That's on average.

22:27Yeah. So, yeah. So, I was low. When I said$100 million, you were low-balling. I was low-balling it. Okay. So, this is a massive opportunity to, like, completely change just, like, the cost, the speed. The business model. The efficacy. And the business model as we do that. But is it proprietary? So, here's my real—because here's where you would revolutionize. So, if I come up with it and I'm company A, right, it's mine. And I get to determine everything. If you're an AI company and you're just doing this, okay, so that you can sell it, then it's yours, which one will make prices lower for the consumer?

23:01Our goal is to really redefine this way you do drug design so it becomes so much cheaper. We have so much more abundance of potential drugs and chemical matter that it really does change the business model and it changes the economics of the space. So you can actually revolutionize the cost of making drugs. Yeah, that's where we're going. That's where we're going. All right. I'm satisfied. Is one of the next steps with AlphaFold, whichever it's through, or maybe the next iteration or so, going to investigate why and what drives the misfolding of a protein so as you can kind of get ahead of even the story of that happening?

23:40Wow. So actually, the misfolding of a protein is another thing. That's what causes some types of disease where you'll have a genetic mutation, a mutation in your DNA, which will change a particular amino acid in that protein. And so it doesn't fold the normal way it should fold. And so it doesn't function as it normally should as a molecular machine. And so things like alpha fold can help us understand what are those mutations that cause misfolding and they're called missense mutations. And then, you know, these could be potential drug targets. So we could think about molecules that could mitigate against that.

24:17If I understand correctly, if you look at the PDR, is this thick and... What's PDR? The physician's desk reference. Thank you. Is this thick. So it's the size of an old-style Manhattan phone book. Okay, it's very thick, multiple inches across. And it's chock full of existing medicines available to the doctor to prescribe. Is it true that 100 % of those medicines are interacting with the patient chemically rather than through protein folding? so that if that's the case, does that mean that where proteins misfold, we can't combat it with any kind of folding algorithm. We just prescribe chemistry for your body to handle the impact of that.

25:06Is that, did I say any of that right? I think that is the majority of drugs. They are chemicals that we take. We take them as pills. The chemicals are not gonna fix the fold. It's gonna treat the symptoms of the misfolding that happened. We're not changing the mutations of the proteins. That could be something like gene therapy. But these are chemicals that come in and will attach themselves to these proteins and somehow mitigate something like a misfold or it'll change an interface, change how these molecular machines work. That's wild. Is there a particular disease isomorphic labs are focusing upon right now or is this a more of a broad spectrum?

25:44Let's go for proteins and cherry pick out certain things or will we be really looking at one particular? The technology we're creating is really, really general. We want to be able to apply this drug design engine on any protein, any target, any disease area that comes our way. Now, saying that, you know, practically as a company, you want to focus on a particular area. We're focusing at the moment on, you know, a lot on cancer and a lot on immunology. Of course. Two biggies. Two biggies. And the two that probably lend themselves best to what you're trying to accomplish, actually. Okay, the question everyone's going to want me to ask right now that's watching this and listening to this is, how are you getting on?