Anna Gressel: Yeah, it's a great question. So protein folding, it is the science of how little tiny proteins fold into all different kinds of shapes. And you might think like, why does a shape matter? Like what is the importance of a shape? But a shape actually really is what gives a protein its function. So if you think about it, its shape means function. And so what we're really talking about when we talk about protein folding is what proteins can actually do in practice. And so many people think about this as like a lock and key. If you have the right key, meaning the right shape, you can open the right lock. You know, proteins interact with each other, so they sometimes both have to have the right shape to create the right reaction or, if you want to think about it, sequence of events. So we're talking about functions that create important events in our body. And that might be like causing a disease, or it might be like curing a disease. So you can think about this as having a lot of important effects in practice. It really is kind of the backbone of biology in many senses.
Katherine Forrest: Okay, so first of all, why you're a lawyer and not a doctor, I know not, but that is for you to ask yourself in the dark of night, those kinds of deep questions. But what I'd like to understand is what is, first of all, how did you get into it? Because that would be, I think, an interesting story. And two, what does protein folding have to do with AI?
Anna Gressel: We'll get to the second part of that. I mean, I think we'll talk all about that, but I'll just tell you a little bit about the first part of it. It is, kind of does, answer your question a bit of like, why did I not do this? And why am I doing what I do now? But for folks who know me, like most of my good friends know that when I was growing up, I really wanted to be a neuroscientist. Like my dream was actually to be a research scientist uncovers all of these hidden insights about the brain and particularly about cells. And I loved molecular biology. And by the time I got to college, the real issue that fascinated me was this important question about how viruses are able to actually get into our brain cells. All right, that sounds like it's a weird question, but think about the kind of COVID day we live in. Like there are effects when viruses get into various systems, and I was interested in the brain. So I was interested in what happens when things like AIDS or other viruses get into our brain cells and kind of wreak havoc on our brain cells.
So you can think about this from a molecular biology perspective, almost like a Trojan horse. The virus is a Trojan horse and it needs to get into a cell, that's Troy. Like it needs to get into the place it wants to be. And so actually, what it needs to do is unlock the door to Troy and open it and go inside. You know, those actions, like the door itself and the key, are proteins. And proteins make that entire thing happen. So the question I was really interested in is like, how does the virus know how to unlock? Like how does it make the key to unlock the door? And can we do anything to change that lock so the virus can no longer get inside? It can no longer use a key in that lock. But it's important, too, for the normal people of Troy to go back in and out of the door even as we're locking the virus out. So it's actually quite a complicated problem. And I wrote my thesis in college about how to change the shape of the lock so it no longer accept the key from the Trojan horse, but it would let all the normal people from Troy use their normal keys to go in and out of Troy and kind of keep the brain, the building, functioning, right? These were the kinds of problems that fascinated me. Katherine, I can see the look on your face.
Katherine Forrest: Okay, all right, all right. No, yeah, too bad the audience can't see the look on my face. You know, we should do one of those podcasts where we actually videotape ourselves as we're talking,
Anna Gressel: This should have been in it.
Katherine Forrest: And people would see all kinds of things. But anyway, you were doing all of this work before AI was an available tool, right?
Anna Gressel: Yeah, I mean at the time you actually had to use what was called x-ray crystallography to take photos of what we're calling the lock and the key then to actually understand what shapes they were making. And again, remember the shapes dictate the function. So like. how the lock and key work depend on knowing what that shape is. But my god, that was a long process. It was a super, super, super, super manual process, and it would take people a billion hours. And so there were hypothetical ways of thinking about this problem and testing around it, but there was not really a direct way in. And so it just wasn't that feasible to mathematically model. And I mean, I will tell you, Katherine, that some of that is also why I was like, I'm not going to spend 30 years of my life probably solving this one problem. Maybe I should go out in the world and not be a research scientist. But if you fast forward, there have been so many advances from AI that have completely revolutionized the way that we deal with these kinds of problems. And the major one, I think the one we should spend some time talking about today, is AlphaFold, which is a project that originally came out of DeepMind, which is now part of Google, was originally kind of its own lab. And that was an amazing AI model that could simulate shapes of proteins in the real world and tell us about their function. And there's a lot to unpack about how AlphaFold works. But suffice to say, it really unlocked this whole new terrain of understanding protein folding and all of the scientific discoveries that kind of scaffold on top of that.
Katherine Forrest: Okay. Now, I have actually heard of AlphaFold. Do you know why?
Anna Gressel: I probably do, but tell me, Katherine.
Katherine Forrest: Okay. Because I find that I will actually watch who wins the Nobel Prize every year when the results come out. And AlphaFold, DeepMind's AI system for predicting these protein structures, the individuals or a couple of the individuals who were involved, Demis Hassabis and John Jumper from DeepMind, they won the Nobel Prize for that.
Anna Gressel: It was such an amazing and, I think, appropriate recognition of the value of AlphaFold. And it, you know, just to also benchmark, this is only one of the ways that people were using AI in that kind of early AI era to unlock scientific discovery. We could have whole conversations on other things. But other issues that I was interested in in the time and that kind of motivated me to get into AI were people who were using images, like image recognition models, to map all of the connections between all of the neurons in your brain almost to create like a fingerprint of your brain and your personal identity in your brain. It's a field called connectomics. So there's a lot that AI was doing to kind of advance and revolutionize science in so many different areas that kind of drove me back into this, notwithstanding that I've of course had gone to law school and become a technology lawyer.
Katherine Forrest: Right, connectomics, wow. That's actually a word I hadn't even heard about. I mean, never heard that word before. Like, not once.
Anna Gressel: I digress, Katherine. You would actually love connectomics because it's all about the sense of self and identity and how that might be rooted in your biology. So you and I can have a separate conversation on this.
Katherine Forrest: You're going to delay my book. You're going to delay me finishing my book because I'm now going to go down a rabbit hole on a whole other area of research. But let's talk about AlphaFold in some more detail because it's going to, I think, wind up being a microcosm for the story that this episode is really about. And the Deep Mind team, well, actually the Nobel Prize committee, wrote that Demis Hassabis and John Jumper of the DeepMind team had developed an AI model to solve what they called “a 50-year-old problem predicting proteins’ complex structures.” And the prize committee there, when they made that statement, was referring to something called this protein folding problem that you have been obsessed with. So tell us a little bit more about what all that means.
Anna Gressel: I talked before about why it's so hard, or why it had been so hard, to actually take photos and model protein folding. But let's talk a bit more about why it actually matters. Proteins are really the engines of life, and they're the basic tools that organisms use to keep ourselves, you know, human machines, running. And I would actually say every kind of life, even plants, are based on proteins. And proteins, in turn, have as their building blocks, a set of 20 amino acids. And those 20 amino acids combine together in different sequences to form hundreds of millions of different proteins.
Katherine Forrest: Now that's my, sort of, biology classes from long, long ago are sort of coming back to me. And I know that a sequence of amino acids can tell you about a protein.
Anna Gressel: Exactly. It tells you about what the protein is made of, but it doesn't really tell you about how the protein functions and why. I mean, you might know something else about that from other science that's happened, but you don't really know what the shape of it is, and you therefore don't know, you don't fully know what the function is. So to understand the shape, you have to kind of understand how all of those amino acids come together and twist and turn and form a cohesive whole. I'm trying to think of a good visual example of this, but it's almost like if you think of a bunch of paper clips, you could drop them on your desk and they would look a bunch of different ways. You could make some 3D shape of it, you can make some flat shape, and proteins can kind of create all different shapes, even though they have all these similar building blocks. So the basic problem here, and this is what we're calling the protein folding problem, is going from a sequence of amino acids to the real structure of the protein was one of the big challenges in biology for decades. All we could really do previously was figure out shapes through experimental observation and photographing, et cetera. But learning to go from an understanding of the properties of the amino acids to theoretically deduce the protein shape was very challenging. And people had been working on it. There were other models. But doing it well really presented a challenge for decades and decades and decades, notwithstanding that people are kind of consistently plugging away at this issue.
Katherine Forrest: Okay, so here's my takeaway so far. Proteins have a shape. The shape is really important.
Anna Gressel: True.
Katherine Forrest: And AlphaFold predicts shapes.
Anna Gressel: Exactly.
Katherine Forrest: All right. Now, and you've given me these numbers, but before AlphaFold existed, less than 10% of the proteins in the human body had validated shapes or structures. We didn't really know what they were, and only 17% of the amino acids in the human body were mapped to a protein with a known structure. So that's a pretty low amount. But then AlphaFold comes out, and there's a total explosion because that AI tool is able to double and then triple and quadruple, and we suddenly end up with, I think, a couple hundred million predicted structures. Am I right about that so far?
Anna Gressel: Yeah, definitely. And there was a big kind of publication in 2022 about that huge number and looking at that trend line, that growth, and what it meant for one of biology's biggest challenges. I mean, it's really incredible. It was solving a problem that had existed for decades, kind of beyond anyone's wildest expectations, I think. I mean, you know, someone could say, “no, no, I expected this.” But I think it was it was pretty astounding. And I don't think I'm ever going to not be amazed that that happened, but we should talk a little bit about the practical impact of this. So how it went from the lab into people's hands and like, what does this really mean?
Katherine Forrest: Right, I think it's the what does this really mean that people are probably walking around with their coffee cups saying “this is such an unusual episode, what does it really mean?” So let me just sort of lay a little bit of groundwork to give you that ramp that you're going to jump off of and tell us what it really means. But we know that AlphaFold grew to be used by actually millions of researchers, and it offered the ability for these researchers to predict protein structure. And now virtually everywhere in biology, from pure protein research to drug development to the classroom, researchers are able to rapidly generate and refine, I guess, hypotheses about protein structures. So what does that do for us? What does this AI tool called AlphaFold actually do for us?
Anna Gressel: One of the things it does, I mean, just one of many, is it really speeds up drug development. And that's a super exciting area for us and for some of our clients, just for the world. The reason that this is so game changing is that understanding how proteins interact with their environment is really key to designing drugs to, for example, combat a disease evolution in the human body. If you know what the function is, if you know what it does in the body, you can identify a promising drug candidate much faster. So it really speeds up the time to identify potential drug targets and get those candidates into clinical trials.
And it gets even more impressive. DeepMind released its AlphaFold 3 model in 2024, just last year, which was another significant jump forward because it can predict structure, not only structure, but also interactions between proteins and other biological moleculesThat ability to model the interactions themselves is really critical. All of these proteins are interacting with other proteins. And so being able to understand how they relate to each other, it's like understanding the lock and the key together. And that is, it's a true game changer and just so phenomenally impressive that they were able to do that.
Katherine Forrest: Right, and so if we understand the interaction, it gives us a much richer or deeper picture about how a drug candidate might actually work in practice.
Anna Gressel: Exactly. And there's so much happening in this space. We'll come back to it in other episodes. But we can talk more about some of the work that companies like Isomorphic Labs are doing and using agents to explore what they call the chemical landscape of the world and come up with potential molecules that might be promising drug candidates. Or other companies like Pasteur Labs are doing really cool work to simulate physical properties in the world. So there's so much happening in the AI and scientific space that's worth us exploring and spending time on. So I don't know, to our listeners, let us know if you like this kind of stuff, and we'll just keep going. We enjoy it, clearly.
Katherine Forrest: I want to do also, let me just put a little plug in, is I want to talk about some of those little nanobots that run around inside the body and fix things.
Anna Gressel: Let's do it. I love that.
Katherine Forrest: All right, can we do that?
Anna Gressel: Yeah.
Katherine Forrest: As part of our AI biology little side show? Okay, all right. Well, that's all we have time for today. I'm Katherine Forrest.
Anna Gressel: And I'm Anna Gressel. Like and subscribe if you're enjoying the podcast.
