David Baker | Using AI for Science to Solve Humanity's Biggest Problems

‍Beatrice Erkers: I am very excited to be here today with David Baker. David is wearing his cool goggles. If you're not looking at the video recording of this, you're missing out. Thank you so much for joining us, David.

David: Thanks.

Beatrice Erkers: So yeah, David, you're the head of the Baker Lab and you've done a lot of work over the years on protein design. We're very excited at the Foresight Institute because we know that you spent so much time working on protein design. I think it was a decade or two decades ago you were awarded the Feynman Prize from us, and then last year you were awarded the Nobel Prize for the same work on protein design, so we like to take a little bit of credit perhaps. But no, in all seriousness, I think that it’s just been really, really cool to see and follow the work that you've done on this. And also, especially with the recent advancements and the work that DeepMind has done with AlphaFold, it's just been really exciting to see this field start to really bloom, and also you receiving the recognition for that. At Foresight, we've also had a few people from your lab who have been fellows at the Foresight Institute. So overall, I should say we're just really excited to start chatting and diving into the work that you're doing today.

So why don't we just start with you sharing? I heard, for example, that you actually started out in social science and you weren't even planning to be a scientist, and now you're this Nobel-awarded scientist. So how did this all happen? What's your background story, David?

David: Well, when I started college, I really didn't know what I wanted to do. I was interested in social studies and then philosophy. I think I took almost every introductory course that was offered. Somewhere through my junior year, I got a little fed up with the language games that seemed to be a lot of the subject of modern philosophy. And then in my senior year, I took a developmental biology class and I learned about how rapidly biological research was advancing. It was just quite a contrast; rather than studying these somewhat arcane arguments from a hundred years before, just to see things that were discovered really in the last couple of years was just completely different.

And so that inspired me to go to graduate school. I hadn't really done any advanced research before, but it seemed really neat. I actually took a year off before starting graduate school. I visited schools across the country and then traveled in Asia for a bit. Then when I started graduate school, I was really ready to try and do something meaningful.

I found that when I applied to graduate school, this class I had taken was on developmental biology. And so I said I was interested in developmental biology and how the brain works; neurobiology and developmental biology were my interests. But when I got to graduate school, I learned that studying those areas, first of all, it was very slow. And second of all, you had to cut open animals, which I really didn't want to do. And so then I ended up working in graduate school on how proteins get organized inside cells. Then that was interesting. And what I really liked is working with other people to solve hard, unsolved problems. I just found it was really fun. So I was studying how cells get organized as a graduate student. And then during that time, I got very interested in how biological self-organization arises.

So then I decided to make a pretty big switch. I hadn't done any structural biology. I'd actually never touched a computer except maybe a couple of times. Mainly, I had ridiculed anyone in the lab who did anything with a computer as it being a total waste of time. And I could be pretty harsh in those days. So then I decided to switch and try and learn some structural biology for my postdoc. Actually, I was only planning for a year to learn some structural biology. And I thought, well, maybe I'd come back and apply that knowledge to the cell biology problems I was working on. But then on my first day as a postdoc, there was a computer terminal on my desk. And I asked my new postdoctoral advisor what that was for. He said, well, it's for computing. And so I had to learn what that was. And then as I started working on protein folding, I really got fascinated by the problem and started learning how to program.

Then when I came to the University of Washington to start my own group, I decided to really double down on trying to understand how proteins fold using a combination of experimental work initially. And then as we developed a picture from our experiments about how proteins fold, we started developing computer models for modeling protein folding, and those grew into methods for protein structure prediction. That was the first version of Rosetta, the software which we developed for many years. And then after we had gotten reasonably good at going from amino acid sequences to protein structures—nowhere near as close to modern methods like AlphaFold—but then at that point, we realized we could go backwards: from structure to sequence, rather than from sequence to structure. And Ryan Coleman came to my group really with the idea of doing that. And that was when we designed the first proteins. And that's probably when I first interacted with Foresight. So that was indeed 20 years ago or a little bit more now. So that was the very roundabout way that I got into protein design. I had no idea what protein design was. I barely knew what proteins were when I was in college.

Just one more sentence: that was in 2003 that we found we could make brand new proteins. And then the challenge was, could we make proteins that do interesting and useful things? And that's really what we've been working on since.

Beatrice Erkers: Yeah, I think that would be a really interesting point to just dive into. Like you said, you're actually working on protein design. And many people may not actually know what that is. So, and what it enables, because that's the really exciting part. So maybe you could say a little bit about what is protein design and what potential applications could we actually get from this in the world that could make the world a better place?

David: Yeah, well, in our bodies and in all living things, proteins are the miniature machines that carry out all the important functions. And so the DNA in our genomes, what it does is it specifies what proteins we have in our bodies. And so if you look in nature, there are all these amazing things that different types of organisms do, like plants absorb sunlight and use that energy to make molecules. We eat plants and we have proteins that break down those molecules. All the currents flowing through our brains as you're listening to me or I'm talking are mediated by proteins. So proteins basically do everything in biology and they evolved over millions or billions of years to solve the problems that were relevant during evolution, like being able to run away from predators or being able to really efficiently turn sunlight into making chemical bonds.

And so a lot of the study of biology has been figuring out for every biological process what proteins are involved and then understanding what the structures of those proteins are and how they carry out that function. So the idea of protein design is that the space of possible proteins is far, far larger, astronomically larger than the very small set of proteins that have been explored in evolution. And also we face a whole new set of problems that weren't relevant during evolution. We're spreading plastic around the world and we'd like to break it down, but there wasn't plastic during evolution. We're heating up the planet, so we want ways of taking greenhouse gases out of the atmosphere. We live longer, so there are diseases like Alzheimer's disease, which are really important. So there are all these new problems that we face, but there aren't really proteins to fix them because all the proteins that we knew about until recently were the proteins that came about through evolution and natural selection. So maybe if we waited for a really long time, another million years, new proteins would evolve that would solve these problems. But we don't really want to wait millions of years for these problems to be solved.

So the idea of protein design is that we've learned now how to make brand new proteins that have new functions, like they can bind to cancer cells and shut them down, or they can mimic pathogens to be really potent vaccines, or they can break down toxic molecules in the environment. And so what we're doing at the Institute for Protein Design is developing computational methods for creating new proteins. And then we make synthetic genes that encode these proteins and we put them into bacteria, and the bacteria make the proteins. And then we can extract the proteins and see whether they actually have the function that we designed. And if they do, then we try and get them out in the world where they can actually solve the problem.

Beatrice Erkers: Yeah, I think I watched your TED talk on this and there's a great data visualization of the amount of proteins that we have and the amount of possible proteins. So that was very vivid to see, or very effective, I would say. So highly recommend that.
I heard you say that molecular machines could be the next industrial revolution. Could you also maybe talk a little bit about that because that's something that's dear to us at Foresight.

David: Yeah. I remember hearing Eric Drexler speak about this at the meeting I went to 20 years ago for the Feynman prize and talking to Eric a little bit about it. And that has now come back to me because now we're actually close to being able to do this with proteins. So it's really very exciting, and I can sort of take you through the steps. One of the first things we tried to do, or learn to do, was design proteins that bind to other proteins, for example, to viral proteins and block them. And so we've gotten very good at designing proteins that bind to some target. The next thing which we tried to do was design proteins which could have multiple states, which could switch between those states. And the third thing that we have worked on is, as I mentioned, proteins that can break chemical bonds, like catalysts.

So a basic machine has to take some kind of fuel, and that fuel should power some cyclical process where it can change the surrounding environment. And so by combining those three capabilities—binding, conformational change, and catalysis—we can actually start making machines.

Now, some of the problems we're interested in making machines for are in medicine. For example, inside our cells, there's very elaborate quality control machinery. So if anything goes wrong, it gets fixed quickly. And that's how cells, like bacterial cells, can keep dividing and everything, and the cells in our bodies are in good shape. But the problem is outside of our cells, there really isn't any system like that which uses energy to fix things that are broken. And so one of the things that we're very interested in now, which I think was in Eric's talk 20 years ago, is designing nanomachines that can circulate and clean up stuff that shouldn't be there, repair broken tissue, and so forth. And because we can design these new catalytic sites, we can use fuel molecules that nature never thought of. For example, it could be something that's in your diet, like there could be some type of food or food additive that actually when you eat it, it goes into your circulation through your body. And that's what happens when you eat triglycerides and sugar and things; they sort of go into circulation. And this now becomes the fuel to power those machines.

Other examples are machines that could, for example, help untangle the amyloid fibrils that have been associated with neurodegenerative disease. Outside of medicine, there are a huge number of possibilities. You can imagine machines that help pattern atoms in specific ways, so allow fabrication of materials in a bottom-up approach. So there's really a huge number of possibilities. We're trying to work out prototypes for these first machines; we have versions that are powered by a chemical fuel and other versions which are powered by light. But the neat thing about them is that since they're made out of proteins, even though they're very sophisticated molecules and we'll be able to do hopefully these amazing things, they're still just encoded in genes like anything else. So they would be very cheap to produce and completely biodegradable. So it's not like the nanomachines are made out of metal and eventually degrade in your body; these are proteins that would eventually just be returned back to amino acids.

Beatrice Erkers: That's really nice to hear that Eric's ideas are being picked up on like this. What do you think we need right now in order to maybe help unlock this? What can we do to speed this up?

David: Yes, well, it's a good question. I think there's still some basic methods development that has to be done. We're trying to figure out how, like if you have a fuel molecule, it's kind of like in a combustion engine where you have a fuel, the gasoline that's being burned, that's the energy, and then that has to be turned into motion that say powers the car. And so you can have a fuel and you can burn it or hydrolyze it, would be the technical term. But it's not—you have to figure out how you connect that to the actual motion to make it really efficient, which is really unclear right now. So no one's ever designed a molecular machine like that from first principles.

We know there's a lot of descriptive work about the natural machines in our bodies, for example, the ones that power motion of muscle and the ones that move things around in cells and the ones that do the quality control processes. But they're very, very complicated. And so there's still some controversy about what the basic mechanisms are. So the first phase is we need to understand how to efficiently convert the energy of consumption of the fuel into driving the process that we want to drive. And that will involve machines that are simpler than the ones I described. And that is sort of a basic research problem.

Well, as everyone knows, funding for research right now is a little bit in jeopardy. That's part of it. Finding government or others who are interested in funding this type of research is going to be important for it to really move forward. And then once we understand the basic principles, there's the question of, for actually achieving the goals in medicine for the tissue repair option, what are the best targets to go for and how can you... what fuel to use is one thing that we think about a lot. What should be the fuel molecule that you eat to power the motor? It probably shouldn't be something that's in all of food because then you'd never be able to control the machine when you want it on or off. And then, like I said, what exactly it's going to do in the body. And then you have to be able to test it. So there's a number of different parts to this.

Beatrice Erkers: Yeah, I hope that Foresight can help coordinate some efforts around this.
Yeah, this leads me, your comments lead me into some other questions I had, which is just like, you've led this very productive lab now for decades. Do you have any thoughts on what it takes to do good science? How can we do it in the best way?

David: Yeah, I think it's a really exciting area. Yeah, I have lots of thoughts on that, of course. I think people are tired of me talking about this, but my metaphor for the lab is what I call a "communal brain" where everyone is talking to everybody else all the time. And I think if you have a group of really smart, really motivated people who are just talking and bouncing ideas off of each other all the time, you can do really amazing things. We extend that to people from the outside who come visit all the time. We have collaborations with people all over the world. And so we try and extend this communal brain and make it as powerful as possible. So I think that's one thing.

I think another thing is that it's very important to be working on unsolved problems. Often, I think there's a desire to do things that you know will work, but it won't really push the boundaries of knowledge so much. But at the same time, you have to choose unsolved problems which are kind of in the "adjacent possible," where they are solvable within a few years of a really talented graduate student working on them. So I think this machines problem is not one—20 years ago, when Eric talked, I really didn't see how to do that, but now is the right time for that. So different problems become accessible at different stages. Sometimes things can move more rapidly than you think. Those are some of the core principles.

Beatrice Erkers: Yeah, I heard that when you went to Stockholm for your Nobel Prize, there were like 200 graduates from your lab or something.

David: Yeah. Yeah. It was so fun. Yeah. It was really amazing. It was just a nonstop party celebration. And then, you know, it was also very important for me. After I got done with my Nobel Prize lecture, when the lights came on, I saw that a quarter of the audience were people that I'd worked with in the past, which was just amazing. So it was just kind of like, we just had party after party. It was really fun, but it made it very clear to me that, well, I'm really excited about all the research I'm describing. I mean, I'm excited about the stuff in the future, not so much excited about the stuff that we've already done, but I think clearly the impact for me, the big impact is in all the people that I've trained and mentored. So that's really now, like my lab is really almost all graduate students and postdocs. I feel like that's a really important part of this for me.

Beatrice Erkers: Yeah, that's really nice. I think it must be also very nice for people who have joined your lab and seen this way of working, and they can take that on when they go to the next lab or something. Speaking of the system we have for science, do you think that there are things that the academic system could be doing better to select for? Well, obviously there are things maybe right now that could be done better in terms of creating security and these sorts of things. Do you think there's... I think at Foresight we're often talking about wanting the "crazy scientists" a little bit more, we want to give them a space. Is there anything that we could do better to select for long-term actual impact in terms of scientific progress?

David: Yeah, it's a good question. I mean, I will say just coming back to the current situation, it is really a critical time because, you know, my colleagues or the graduate students and postdocs in my group, and with all this uncertainty about science and science funding, it's... we're really in danger of losing a generation of scientists in the United States. People are leaving the country or choosing other careers. So we're really at a critical time. You know, I think that is a problem for entities like Foresight and for philanthropists to really think about how can they support the work of graduate students and postdocs and enable them to continue with their careers through this very critical time. I think one of the metaphors the government has is that if you shake things up and you create some chaos, in a business, in a company, that can be good. But I think for science, it's just really bad because you have these people that are at a really sensitive career stage. That's the first thing.

As far as how to recognize research that is going to be really impactful in the long term, I mean, I think things like you had the Feynman Prize. I mean, that was 20 years ago before we'd really done anything. So I think you're doing a good job in recognizing things that have potential. So I think that's a very important function and that's probably a way in which entities like Foresight can really have a big impact. It's kind of as a society-wide thing. I think it's a bit of an investment strategy. I mean, you invest some in things which are kind of lunatic fringe, which like when I talked about de novo design back then, it really seemed crazy. Or maybe, we had designed one protein, but the idea that you could actually ultimately make machines or make drugs that way seemed totally crazy. So I think supporting work like that is good, but there's also, you know, more traditional science where you're sort of expanding the bounds of knowledge in a discipline. That's also important. So I think it takes a range of different things. But I think models where there's more integrated collaboration, I think, can be better than some things where you have many small groups that aren't talking to each other; that's, I think, less effective at solving really hard problems.

Beatrice Erkers: Yeah, I think that another thing just in relation to this whole like meta science discussion is the whole AI for science thing that's happening a bit now and where you've obviously played a really crucial role. I heard someone say that you and Demis Hassabis sort of represent two different approaches to innovation where you focused on this domain for a really long time and just thrown different tools at the same sort of challenge while he developed this one tool and is applying it across many domains. Do you think that there are particular strengths or limitations to those approaches? Would you recommend choosing one sort of area and focusing on that or what would you recommend?

David: Yeah, I think they're both, they both obviously can be very, very successful. I think what DeepMind has done has been really, really impressive in solving problem after problem. I think it also depends on what the problem is. And I think even if we just look in the realm of proteins, there's protein structure prediction and protein design. They seem on the surface of it to be very, very similar because in the case of protein structure prediction, you're going from the amino acid sequence to the three-dimensional structure. And in the case of protein design, you're just going backwards from the three-dimensional structure and function back to an amino acid sequence.

But I think what we've seen is that these models are, the two models you described, have very different, are differently matched to these two problems. So, in the case of protein structure prediction, it was a problem which it turned out that the very large number of protein structures which have been solved and deposited in the protein structure data bank actually contained enough information to solve the problem. And that is what DeepMind demonstrated with AlphaFold: that you could get remarkably accurate predictions of protein structures just by training on the data that existed and building architectures, networks that were really well suited, were sort of really tuned for that. And so in that case, the protein problem is really straightforward, right? You have a large number of protein structures and you know their amino acid sequences and you have to train the model to predict what the structure is from the amino acid sequence. So in a sense, it's a very well-posed ML problem because you have a large amount of data, a really large amount of data, and basically the question is how do you best apply modern AI methods to solve this problem? And DeepMind, being really at the cutting edge of AI methods, could put together solutions to those problems that were really, really effective.

Now, protein design, even though it's the inverse problem, is much more open-ended because it's not that you want to design the native protein structures, because we found really early on that that's kind of a ridiculously easy problem. If you have the structure, you can pretty much work out the sequence. What you want to do is design new proteins, and you want to design new proteins that do new and interesting things. But what should those things be? It's a very open-ended problem. Like I was talking about earlier, there are all these different problems you could solve with protein design. And then it's not purely a computational problem because let's say you want to design a protein to cure cancer. You do your design calculation, you have an amino acid sequence, but what do you do then? There's no way of knowing whether that amino acid sequence cures cancer. Whereas in the protein structure prediction case, you could just compare your predicted structure to the actual structure to see how well you're doing.

And so this is where having the domain expertise, like for us, when we started developing AI methods for protein design, very much inspired by DeepMind's success with structure prediction, we thought, well, if they could do that well in structure prediction with AI methods, we should be able to develop AI methods for protein design. We already knew what the important problems were. We knew how to make and experimentally test proteins. And people are sometimes surprised that from not really doing any AI at all, in a couple of years, we were able to develop methods for AI-based protein design that people are using all around the world that are clearly better than what we had developed before with traditional scientific models. But I think it's because we knew what the problems were and we knew how to make the proteins. And DeepMind has been working on protein design for a very long time, AI basically since the time that they started AlphaFold, but it's been very hard because they don't have that broader connectivity. I think it's a very long way to say that I think the generalist and the kind of domain specialist approaches really depend on the nature of the problem.

Beatrice Erkers: Yeah, that makes sense. So in this whole AI for science, in the little bubble that I exist in, that's the most exciting term you can use right now, AI for science. What do you think the promises are, and what do you think is maybe overhyped, if anything?

David: Yeah, well, I think that the example of protein structure prediction and protein design, there might be a little bit of a tendency to over-extrapolate from it. And because what I said earlier was that the really key thing is that the protein structure database—the first structures were solved in the 1960s. And since then, many, many tens of thousands of graduate students and postdocs spent their careers solving the structures of proteins. Like each protein structure was a PhD. Then another large group of scientists took on the effort of curating a database of all this that's very, very precise and accurate. And the global investment in this effort was probably in the tens of billions of dollars. So just a huge public, concerted effort over 50 or 60 years to generate the data. So it's kind of like the fossil fuel; we burned fast because there was all this stuff that was done before.

But as we move up the biological complexity hierarchy, we don't really have datasets of that category. And so what I see now happening is there are many different groups developing their internal datasets, and there's a new one every day. It's not clear that any entity, even one that's very deeply funded, can collect enough data. And it's not even clear what data to collect. So I think it's a big experiment. And we'll see. I mean, to some extent, the problems of protein structure prediction—that's a clear biological mapping from sequence to structure. The other sort of advance that's been made in AI is things like ChatGPT, where you train it on a very boring task, just to predict the next word or the next letter, the mass token recovery. And then suddenly it gets this emergent property of being able to answer very complicated questions, this apparently near-human level reasoning. And so I think that's driven some work or quite a bit of money to be spent on just collecting tons of data and training models to do things that are not necessarily interesting themselves in the hope that more exciting emergent properties will come about. And that may or may not be the case. In the case of text, there's just so much of it. I would say that I'm sure AI will have an impact as we go up the biological complexity ladder, but I think the extrapolation from what's been done with protein structure prediction and protein design, people may be over-extrapolating.

Beatrice Erkers: Yeah, that's, I think that's a really interesting take and yeah, good data and good benchmarks is what I heard you need to be able to do good AI for science currently. I would also be curious to hear just like, how does it even work in practice? Like, do you have machine learning engineers at your lab? And are the scientists at their computers and then they test in the lab, or how does it actually work?

David: Yeah, yeah. Well, like I said, people were surprised that we were able to be, from not doing any ML at all, to be able to in just a couple of years be developing cutting-edge ML software from scratch. And I think the reason why is I just get incredibly smart, brilliant, talented people coming to my group all the time who want to do the next big thing. And so when it became clear that the next big thing was going to be AI for protein design, we learned it. And people came in who hadn't done much ML before but were interested in learning. I think it's like anything; mastering the last 30 years of a field is hard, but with something like ML, where the methods are moving really fast—when we started, it was convolutional neural networks, and now they're ancient history and transformers and attention are the thing. So you only, it's actually very easy to learn. You don't have to worry about this whole ancient history, and the tools are pretty advanced. So we could very rapidly start putting together really interesting new models.

And so since that time, there have been some people in my group who've really explicitly been interested in new methods development. And then there are people who are interested in applying the new methods to longstanding challenges in protein design like catalysis. And those people are not only playing around with the software, but then they're also designing proteins. And right next to you, if you visit my lab, what you'd see is it's a big room and on one side people are sitting at desks doing computer stuff and on the other side people are pipetting and taking synthetic genes encoding the latest designed proteins and testing them. And so as our AI methods got better, we were starting to be able to design more proteins better, more quickly. And so we developed methods for very rapidly taking a set of new designed proteins that came out of the computer and very rapidly making and testing the proteins in the lab. So some super talented postdocs developed a method they called "cowboy biochemistry," which let us make proteins and test them from very inexpensive DNA fragments encoding the proteins. And we could get the proteins back in like three days after we got the DNA fragments. So we developed really quick ways of testing the proteins. And so we could go back and forth between the computational methods development and the experimental feedback very quickly. And I think that's really a key thing that we've continued to really optimize. So that's very much the way my group works today, where there are a number of really smart people just trying to develop the latest methods. And then there are others who are really excited about taking those methods and testing them out in the lab. So it's very closely integrated.

Beatrice Erkers: Yeah, that's great that you can have such short feedback loops. That's amazing. I think on that also, you know, this communal brain of science, how have you... I'm asking because I know you do, I think, chocolate Wednesdays or something like that. What are the things that you have done?

David: Yeah, well, lots of free food and drink is one of the principles. It's gotten a little bit harder recently because we're under this austerity regime imposed by the university with a changing financial climate. So we have lower value treats than we used to, but there's still, there's still fresh fruit Thursdays, chocolate Wednesdays, and then Mondays and Fridays, we have happy hours with drinks and chips and munchies, and Tuesdays is bagels. So, and then, so there's lots of gatherings all the time. We also have two meetings a week. I don't have a very long attention span, so they're one hour and three people talk about their research for 20 minutes each. So six people present their research every week, and after that we have happy hours where people brainstorm and talk about things. And then it's just the density is super high. I think we probably set some kind of record. And so people are just always bumping into each other and talking over coffee or lunch or as they bump into each other in the hallway.

Beatrice Erkers: That seems like, so food seems like the number one sort of technology that we have for the social connection. Yeah. I also, when I was reading up on your work, as I understand it, you've been a big part of sort of helping to build up this biotech ecosystem in Seattle. And I was a bit curious because right now it's easy to argue or say that it's quite centralized where we have a lot of the scientific and technological innovation. Like either it's in the Bay Area in the US or Boston maybe. And even if you look globally, the Bay Area is still for technological innovation, it's just the hub and it's hard to compete with. What do you think, why was it important to you to maybe, if it was, to have more startups in Seattle? And do you think that we need to do something about this centralization or is it better to have this? Is it better to have the Bay as the capital? And then, yeah.

David: Yeah. Food helps. Food and drink help, yeah. Yeah, well, I think different locations have different specialties and, you know, Silicon Valley definitely was very good for emerging computer technology. I would say probably, at the time of the Feynman Prize and my visit to Foresight Institute, I was not thinking about companies at all. But as we were able to design proteins that solve problems and looked like they could be useful, increasingly the people coming to my group, who I said are super highly motivated and smart and ambitious, would sometimes feel like, "I designed this new protein. Now I want that is supposed to cure cancer. Now I actually want to see whether it can." To do that, you have to start a company. There's kind of a limited amount you can do in an academic environment.

So we developed this translational program, and the way it works is if you're a graduate student or postdoc, after you write your paper and publish it in Nature describing your amazing new protein, you can then transition to the translational investigator program. And then we were able to raise some funding from philanthropy so that people could carry on developing the technology to the point where you can start a company and then start talking to VCs about spinning out a company. And we've actually spun out 21 companies now, which is I think kind of a record. It's really just, I'm not really very deeply involved, but there's this amazing number, amazing people come in.

And then it's actually sort of a funny story. When we first started doing this, I talked to the VC firms, which were large firms at the time, like Flagship and Third Rock. And we never got past like the second call because they always wanted to be in Boston. I said, well, it has to be in Seattle. Now it's completely changed. When we talk to VC firms about starting companies, they know it's going to be in Seattle. It's not even a question. And I think that's good because, so every place has its own advantage. Seattle is very strong in AI now, and obviously the tech industry is really building up here. So there's a lot, and you know, we've got some really fantastic medical research organizations like the UW and Fred Hutch and Benaroya Institute and Seattle Children's and stuff. So there's a lot of talent. I grew up in Seattle actually. That was one of my goals was to make biotech a new big thing in Seattle, make Seattle a leader in that. I think it's really happening. I mean, the whole mentality around company formation in Seattle has completely changed and we're sort of filling the Seattle landscape with new biotech companies, which is very exciting.

Beatrice Erkers: Yeah, that is very exciting. So this is the Existential Hope podcast. So I do want to ask you some vision-setting questions, I would say. If you think about all the possible applications of protein design, which do you think could have the largest positive impact on the world over the next few decades?

David: Well, I think we sort of have to look in different areas. So increasingly, I and the people in my group are interested in issues relating to sustainability. So let's first talk about bioremediation, which is basically taking compounds like plastics or forever chemicals that humans have put into the environment and breaking them down. One of the things that proteins in biology are really good at is breaking chemical bonds. So I think there's a huge opportunity for designing proteins that break down pollutants and toxins and plastics and so forth.

Then we have really major climate challenges. So we need to remove greenhouse gases from the atmosphere. We need to have better ways of capturing carbon. And again, these are things that proteins in nature figured out how to do, but sort of in the context of biological systems and specific environmental niches. So no proteins evolved to do this in smokestacks, right? And so the things that we now as humans want to optimize for in terms of climate are different from the things that organisms evolved for. So we're now, for example, interested in things we can do, for example, for carbon capture in the oceans, which was never something that was really optimized in biology.

And then in health, there are huge challenges. So my colleague Neil King actually made the first de novo design medicine, a vaccine, and he's working to make universal flu vaccines and in general vaccines that could protect against as-yet-unknown viruses. And then on the therapeutic design, as I mentioned, we're working on better treatments for cancer, for autoimmune disease, and therapeutic approaches that can be developed quickly to deal with future pandemics or biological warfare agents. And then neurodegenerative disease I mentioned earlier is a really big and important one.

In technology, I think there's huge opportunity for integrating proteins with inorganic materials. I've been fascinated by how in our bodies, bones and outside our bodies, shells are made out of proteins interacting with inorganic minerals. And so we're starting to be able to make new types of materials that integrate proteins with inorganic stuff. And I think that could be very impactful as well as figuring out how to combine biology and proteins with electronics. Again, that was never encountered during evolution. So I think all these things could have very, very great impact, but it's in different areas.

Beatrice Erkers: Yeah, I think if we get half of this, we'll be in a much, much better place. If we sort of envision that it's 2045 and protein design has achieved most of what you hope, how do you think the world looks differently? Like if you paint the picture for us.

David: Well, yeah, it's a good question. I mean, a lot of the suffering from disease... of course, life is finite. So we may cure some diseases only to have others or aging to become an increasing challenge. I know there's a lot of optimism now about combating aging. We'll see how that goes. That's a very complicated biological problem compared to the ones that I mentioned.

With climate, I think it's going to be, you know, it's going to take many types of approaches and the collective will to deploy things. I think one of the problems for the solutions I described is, and I think it's true for most of the solutions to climate-related problems, you know, at what stage do you decide to pull the trigger and really apply them earth-wide, which is a big commitment.

And then the technologies, I think they'll allow us to do sorts of things that we really can't even envision today from making new types of materials, green chemistry to replace... I didn't mention that earlier, but proteins are really good at doing catalysis. A lot of the catalysis that gets done industrially now is very energy intensive and produces, involves a lot of toxic compounds and solvents. So replacing that will be really good.

But I'm quite confident in 2045 that there will be a new set of problems for the next set of people to tackle then. You know, probably a lot of the people I'm training now, I hope will be tackling problems in 2045 and 2065. So I guess I'm an optimist in that sense. You know, I think there's a popular thought now that AI is going to control our lives within a few years. And I think I'm... I think that humans will stay in charge of AI for a considerable time to come.

Beatrice Erkers: Yeah, there's a lot of hot takes on this topic. I know you also recently, your lab launched a podcast, the Baker Lab podcast. Do you want to tell us a bit about that?

David: Yes. Yeah, I do. I think that's another interesting story. So I've worked with really amazing people throughout my career and they come from all over the world and have all kinds of backgrounds. Originally, my plan was to write a book describing some of these ideas about scientific innovation and the communal brain and a little bit of actually what we've talked about today. But really what I wanted to focus on were vignettes on some of the different people who have come through because I really wanted the message that anyone anywhere could become a great scientist and to have some idea of what the paths are that could be taken. So I actually started writing this a little bit, but then I talked to agents who said, "You know, David, what people are going to want to hear about is your story," but I wasn't really interested in telling my story. So that sort of just got put on the back burner.

But then we started talking about this and realized that a podcast could be a really good format for this because then everyone gets to tell their story, and there are many different stories, and a podcast, we can keep having new installments. So I'm really excited about this as a way of, you know, no matter where you are in the world, you can tune into this podcast and you're going to find someone whose path you could imagine being yours. And so that I'm very excited about.

Beatrice Erkers: Yeah, I listened to the... There are only a few episodes out for now, but they're really nice. And they're very... yeah, the human side of science is, I think, what you're pushing forward. Well, I think that's it. It's been just such a pleasure to talk to you and thanks for all the work that you're doing. Maybe one... I'll sneak one last question in, which is: what is the best advice you ever received?

David: Yeah, exactly. I don't know. I'm not very good at listening to advice, but I'm good at giving advice. I mean, I can say what I advise people now is to do what you're most excited about and most passionate about at any given time and not to worry about the future and don't plan too far ahead because the world changes. And that's what I've always tried to do. Do at any time what you're most excited about, and it can take you places that you never expect, and the world will change in ways you never expect. So that's my primary piece of advice.

Beatrice Erkers: I think that's great advice, especially in fast-changing times like this. Well, thank you so much, David, for joining us.

David: Yes. Okay, and thank you. It was really fun to talk and those were great questions. Thanks!

Beatrice Erkers: Thanks.

‍