Nobel Prize lecture: David Baker, Nobel Prize in Chemistry 2024

Honourable laureates, excellencies, distinguished colleagues, ladies and gentlemen, welcome to the Nobel Lectures in Chemistry 2024. So this year's Nobel Prize in Chemistry recognizes groundbreaking achievements in two different areas, computational protein design and protein structure prediction. So it's all about protein structures. As you may know, the polypeptide chain with protein folds itself up in space into a complex three-dimensional structure. Now the remarkable thing is that this 3D structure is somehow encoded in the sequence of amino acids that make up the chain. This finding that a linear sequence of amino acids dictates the 3D structure dates back to the work of Christian Anfinsen, who received the Nobel Prize in Chemistry in 1972, so over 50 years ago. It meant that it should in principle be possible to predict the 3D structure directly from the sequence. Likewise, it should be possible to predict which sequences would produce a certain 3D structure. So it has long been a dream of biochemists to computationally solve these two problems, and it is clear that one needs to use computers to do it. The latter challenge, to predict sequence from structure, that's what protein design is all about. The idea is to draw a new protein, typically one that has never been seen before in nature, and somehow compute a sequence that will give this structure. The sequence could then be made in the lab and completely new proteins with desired functions could thus be created. So this challenge was solved by David Baker and co-workers in 2003. They designed an entirely new protein with 93 amino acids and computationally predicted a sequence that would give this structure. They then made the protein and verified experimentally that the prediction was correct. The reverse problem, to predict 3D structure from sequence, is an astronomically huge search problem. This was pointed out by Cyrus Leventhal already in the late 1960s. Progress in this field was slow indeed for many years, but Demis Hassabis and John Jumper and their co-workers solved it in year 2020 by ingenious neural network engineering, which resulted in the now famous program AlphaFold2, and it was published in 2021. This was undoubtedly the real breakthrough in protein structure prediction, and AlphaFold2 gave unprecedented accuracy. So this network is trained using the data stored in sequence and structure databases to discover correlations and patterns among amino acid sequences, and this allows it to produce accurate distance maps that can then be converted into 3D structures. We'll probably hear more about it from the laureates. So the work of this year's laureates has had an enormous impact. David Baker has opened up completely new possibilities to design proteins that have never been seen before, with many exciting applications in pharmacy, medicine, nanoscience, etc. And likewise the work of Demis Hassabis and John Jumper has caused a revolution in structural biochemistry, and we now have access to predicted structures of all 200 million known proteins. And this is a phenomenal resource for biochemical and biological research, also with many new applications. So our first speaker, David Baker, he was born 1962 in Seattle, Washington, United States. He got his PhD in 1989 from University of California at Berkeley in California, and he is presently a professor at University of Washington, Seattle, and an investigator of Howard Hughes Medical Institute. So please join me now in welcoming Professor David Baker for his Nobel lecture. All right, well first of all, thanks everyone for being here. It's really wonderful. It's a great honor, and it's wonderful to see so many friends and family and colleagues. So today I'm going to tell you about de novo protein design. This is an artist's depiction of the interior of cell in our bodies, and each of these colorful blobs is a protein, and each protein is carrying out a unique job or function. And indeed, in all living things, proteins are carrying out, they're basically the miniature machines that are carrying out all the important functions. And they evolved over hundreds of millions of years or billions of years to be really perfect at solving the problems that arose during biological evolution. But today we face new problems. We live longer, so neurodegenerative disease and other problems now are arising. We're heating up and polluting the planet. So if there was selective pressure over a really long time, new proteins might evolve to solve these problems, but it would take a very, very long time and not be very pleasant. So the goal of de novo protein design is to make new proteins that can solve modern day problems as well as the proteins that came through evolution to solve the problems that were relevant during evolution. First, I need to remind you of a little bit of biology, and you already heard this. I'll just give you the basics again. The genes in our genomes, each gene encodes one protein. It does so by specifying the amino acid sequence of the protein, which then folds up to the protein which has a particular structure and biological function. In de novo protein design, the flow of information is just the opposite. We start with a desired function that doesn't exist. We build a protein which we predict will have this function and then compute the amino acid sequence this protein should have. Then this is a brand new protein, so there's no gene in nature which already encodes it. Instead, we make a synthetic gene, a synthetic piece of DNA that encodes this protein. We put it into bacteria, and the bacteria basically become factories for making the protein. We can get the protein out and then test to see whether it actually does what we designed it to do. Today, I'm going to be giving you a number of different examples of design proteins that we've designed to have different functions. In every case, we've gone through this computation and then made a synthetic gene and then made the protein. I'll be showing you what those proteins do. There is a huge space of possible proteins. There are 20 different amino acids. A typical protein is a sequence of about 100 or more amino acids. That means the number of possibilities is 20 times 20 times 20 a hundred times or 300 times, depending on how long the protein is. This is an astronomical number. It's depicted by this gray space here. What's been sampled during evolution is a tiny, tiny fraction of the possibilities. The proteins in nature are indicated by these red ovals. Evolution proceeds incrementally, so the proteins in us are closely related to the proteins in other mammals and so forth. It kind of bunches up a lot. Traditionally, and this is really up until the present, if an engineer or a biologist wanted to solve a new problem and they thought a new protein could work, the way it was done or is still being done is to look all through nature to see if a protein can be found that has that activity. That's called bioprospecting. The other thing that's often done is to take a naturally occurring protein that does something kind of similar to what you want to do and make small changes to it. You get a protein that's very close to a naturally occurring protein and you've hopefully modified it to do a little bit closer to what you want it to do. What I'm going to talk about today is different. I'm going to talk about building proteins completely from scratch that carry out new functions. The advantage of this is you can build the protein to do exactly what you want to do. You don't have to worry about what the closest nature got or try and modify something that's already very complicated. I'm going to begin with a very brief history. The first stage in de novo protein design was understanding how to build new protein structures from scratch. This really culminated in 2003 with the work of Brian Coleman designing a brand new protein which had a new structure and a new function. Since that time, a growing and now quite large community of people, really wonderful scientists, quite a few of whom are here, have been working to both design new protein structures but also to design new protein functions. In particular, design proteins that solve current day problems. In very recent years, this field has sped with the advent, and you'll hear a lot more about this in Demis' talk, of deep learning methods. I'm not going to tell you much more history in my talk. I'm going to basically fast forward to the present. I'm going to tell you about the deep learning methods that we have been developing for protein design. Then I'll give you examples of what protein design can now do. We've been developing generative methods for creating new proteins. The one I'm going to tell you about, called RF diffusion, is really inspired by methods for generating images. These methods work by taking a large set of data, for example, images off the internet, adding different amounts of noise to the images, and then training a network to remove the noise. Once this has been done well, one can start with completely random pixelated sets of pixels, successively denoise, and generate an image that looks like its image off the internet, but in fact, it's a completely new image. In the same way, we took all the structures in the protein structure data bank. I need to highlight that everything I'm telling you about from now on absolutely relied on the work of the tens of thousands of researchers who solved the protein structures in the PDB, because that's the data that we used to train. Also, the people who organized the protein data bank, it was really an absolutely invaluable resource. We took all the structures in the database, we added increasing amounts of noise, and then we trained a network to predict what the denoise structure at each stage was. At the end, just as in the case of images, we could start with a completely random image and generate something that looked like a very plausible image. We could start with completely random configurations of amino acids, progressively remove the noise, and generate a new protein structure. We found that these proteins folded up pretty well, we could make them in the lab, but they didn't do anything. In the case of images, as you all know, you can condition the generation process with, for example, a text prompt. You can say, generate an image of a cat sitting on the table, and that's what you get. We can do exactly the same thing here, because it's not very interesting just to make new structures. What we want to do is make new functions. We can condition this protein generation process on the function that we want to make. One way to do that is shown here. This is the insulin receptor. We can carry out this generative diffusion process in the presence of a target, like the insulin receptor. During training, the generative method has learned that proteins bind to other proteins through very shape-complementary interfaces. You can see that as this protein gets filled up, how the shape match of this new protein matches the target. We've now used this to design proteins to bind to well over 200 different protein targets. It's pretty much a solved problem now. I'm not going to tell you about all 200 of them, but I'm going to give you some examples of how one can use these binding proteins to solve problems in a wide range of different areas. I'm going to tell you about three types of applications today in medicine, in technology, and in sustainability. I'm going to begin by talking about applications in medicine. Snake poisoning is still a major medical problem, particularly in developing countries. Snake toxins interfere with fundamental biological processes. An antidote for snake venom needs to be a very stable protein or set of proteins. It has to be cheap to manufacture because where they're needed is in countries where there isn't necessarily a cold chain. You need them out in the field. You need to administer them almost immediately. We thought this would be a good match for protein design because the proteins that we design almost always are very stable proteins. You can produce them in large amounts quite cheaply. Here is a design protein which is binding to the toxin, which is here. All this work is highly collaborative. Our collaborators, in this case, tested to see whether this design protein could protect animals against the snake venom. In this experiment here, the animals who got the snake venom alone died almost immediately, whereas ones which got the design protein were completely protected. We're very excited about this as a general way forward for generating anti-venoms that can be more effective and much cheaper than in current approaches. A major problem in current medicine is inflammation, which is associated with many autoimmune and other disorders. Central to inflammation is a protein called the TNF receptor. It's actually the target of many drugs currently on the market. Here you see a diffusion trajectory conditioned on the structure of this target. You can see how it's building a protein which is really exquisitely shaped complementary to the target. This protein is very effective at blocking inflammation in animals. These are animals that have been exposed to an inflammatory treatment or inflammatory signal from bacteria. There's a very large amount of inflammation. Current drugs like Enbrel that are used to treat inflammation have some effect. This design protein, because it binds so tightly, is even more effective. We're very excited now about making drugs to treat a wide range of autoimmune diseases and inflammatory diseases. Cancer, of course, remains a pressing problem. Here we've been designing proteins to activate the immune system to treat cancers. This protein here, this design protein, is bringing together two immune receptors. This results in a very strong activation of the immune system. Again, with collaborators Mike and Stephanie Dugan, they have been looking at pancreatic cancer, which is one of the most difficult forms of cancer to treat currently. What they find is that this design protein is quite effective, more effective than other known treatments in reducing the size of tumors in this model. Again, we're very excited now about getting this protein out in the world for clinical development. Another important application of design proteins is to make antivirals for pandemic preparedness. This is the influenza virus surface protein. Here you see the diffusion process generating a binder here. Here we've added an additional conditioning criterion, which is we want this protein to be an antibody, which is a particular type of protein fold. We wanted to do that because the pharmaceutical industry knows how to develop antibodies and so they like to work with them. Here you see an antibody-like structure being generated that binds to the target. Here I'm showing you an experimentally determined structure of this design protein bound to the flu virus protein. You can see that in one of these colors, let's see, the experimentally determined structure is in purple, the design structure is in gray, and you can see they're almost identical. Those of you who know about antibodies know that they recognize their targets through CDR loops. These loops, which are being built completely from scratch in addition to the docking mode, are coming about almost exactly in the design model. I think that antibodies are currently discovered by immunizing animals or pulling them out of people. I think we're now seeing a transition over to using these kind of design methods to generate antibodies. Another increasingly important medical problem are neurodegenerative diseases like Alzheimer's disease, which are associated with the formation of long amyloid fibrils. There are several proteins, among them amyloid beta or A-beta and tau, which form these fibrils that are observed in the brains of people with neurodegenerative disease. Here what we've been doing is designing proteins that will bind to the disordered portions of these proteins that would normally interact with themselves to form amyloid. These trap these regions so they can't form this extended amyloid. This experiment, again with collaborators, is looking at if you simply take these proteins, put them in the solution, they aggregate as the amyloid is formed. With these trap proteins, this amyloid formation is completely blocked. This is measuring the extent of amyloid or aggregation that's going on. We're very excited about these as a possible way forward with neurodegenerative disease. There are many other applications we're excited about now in medicine, but I'm going to switch over and now talk about technology. One general problem in technology is in sensing. How do we sense arbitrary molecules and how do we determine sequences of things like DNA and proteins? In nature, there are channel-like proteins or pore-forming proteins that look something like this. People have adapted those for problems like sensing and sequencing, but they aren't really optimal for that because they didn't evolve for it. With Anastasia Borobieva's group, we've been developing methods for designing new pores completely from scratch. This has the advantage that we can completely customize all properties of them, of the pore and so forth, to be optimal for the task we want to carry out. This just shows you some examples here. These are three designs with increasing pore sizes. The larger the pore, the more ions that can flow through it, and so the larger the current that is measured going through these. To turn these into sensors, and we're very interested in constructing molecular noses, what we do is to design proteins that bind small molecules that we want to detect. We can put these proteins on top of the pore, like shown here, and then when we put this combined protein into a membrane and we measure the current across it, we see a sizable current. When the small molecule is added, basically it closes up the channel, and that's shown here, and then the current completely shuts off. This can be measured because one can measure electrical signals very, very sensitively. We're excited now about being able to make very general molecular or electronic noses that can tell you what's in the solution by integrating the information from many different pores like this. Now, what you'd really like to do is to couple this kind of sensing scheme with electronics directly. What I showed you before was those were inserting into a hydrophobic lipid layer, so we've been working on methods for inserting these sorts of pores into silicon nitride chips so they can ultimately be, for example, in your cell phone to tell you what's around in the environment. This requires a totally new class of pore, which is an example which shows here. Of course, proteins never evolve to interact with silicon nitride or to insert into it, so we have to design new types of protein structures. Here are some examples of EM structures for this gallery of pores that we've made to try and approach this problem. Again, with collaborators, this is from Jens Gundlach's lab at the UW. Here we have a pore in a silicon nitride chip. There's a lot of current going through. The design protein pore can insert in, and we get a very stable current. Now, we can start modifying these pores, as I showed you a moment ago, by incorporating binding domains and other sorts of domains on the interior to be able to read out sequences and other properties. Another general problem in technology is targeting very specific sequences of DNA. Our genomes have lots and lots of DNA in it, and there are many applications, for example, fixing genetic diseases, where you want to target just a very, very specific region of the DNA. To do this, we've been designing proteins which bind to DNA and have very high sequence specificity, so the protein shown here makes very specific interactions with the DNA bases that make it very sequence dependent. This is now some experimental data showing how, if you take this protein and measure its binding to DNA, it's very sensitive to any changes of the DNA sequence. When the sequence is changed, the red color means that the binding is weaker, so these sorts of designs can read out DNA sequence very precisely. We're excited about using these, for example, to create new ways of smaller, more compact ways of fixing genetic lesions and other diseases. Neil King had this very exciting result last year with the first approved de novo design medicine. This is an interesting case where Neil started off making these self-assembling nanoparticles just to see if we could do it. Then, after he had conceded in building them, he realized that he could make vaccines out of them by putting pieces of viral proteins on them. During the pandemic, he put a part of the receptor binding domain of the COVID surface protein on these and found that this elicited really, really strong immune responses. This is now a clinically approved medicine. It's been approved in several countries. Here is the design protein in a vial. We're very excited about, a year or two from now, being able to have many, many examples like this. These sorts of nanoparticles are very useful also for delivery, but there's a problem in that if you want to deliver a lot of material, let's say a drug, to a specific site in the body, the size of the container matters. With a bigger container, you can make more. This is an interesting geometry problem. The icosahedron or dodecahedron is the largest completely regular polyhedron. You can make it bigger. This is 12 pentagons coming together. You can make it bigger by inserting hexagons between the pentagons as shown here. This is just as in a soccer ball. Here you have a problem because you can see the angle inside the pentagons is different than the angle inside the hexagons. We've been trying to figure out how to solve this problem. There are basically two solutions I'm going to show you here. These are both electron microscopy structures of designs. In this case, we've solved the problem by using just one type of protein that can adopt different conformations in the different sites in this nanoparticle, allowing the interspersal pentagons and hexagons as you can see here. In this case, we've used different protein chains that sit at the different non-geometrically equivalent positions. The advantage of this is that both of these methods, and again this is experimental data, cryo-EM structures, give much larger particles than one would get with pure symmetry. This, for example, is one of the most widely used current delivery vehicles based on a naturally occurring virus, AAV. Everything I've told you about so far is static, but in technology we'd be able to turn things on and turn things off. To do that, we've designed proteins that undergo conformational changes and have two different states. On their own, they fold up to one state, but then when an effector is added, they convert into another state as shown here. If we build larger assemblies out of these, like this triangular one here, in the presence of the effector, these will now bend. You can see when this becomes a right angle, we effectively convert a triangle into a square. This works really nicely. We chose to go for these simple geometric transitions because it's very easy to follow by electron microscopy. Here's a design just like this one that was designed to be a triangle. Then the effector is added and you see you get a square. This is electron microscopy data. You can clearly see that transition. Here's another case where you have a square that's going to a pentagon when the effector is added. For the biochemists in the audience, this is clearly allosteric control because this effector is not acting at the interface between these chains. It's acting quite far away and working through by changing the conformation. Well, that wasn't very particularly useful. That showed that we could build switches, but where might you want actually a switch? Coming back to immunotherapy, we might want to activate the immune system to treat cancer, but then we might want to turn it off because over activated immune system can be a problem. Here I'm going to show you an off switch. Here's a design protein that brings together two receptors activating the immune system as I showed you earlier, but now we've built in a binding site for an effector which turns the whole system off. Here's a cartoon of how it works. Here's the design bringing two chains together. This effector comes in. It causes an allosteric change which kicks off this other receptor subunit. This is again a wonderful collaboration. This is actually going on the surface of a mammalian cell. You're seeing these two receptors are labeled with fluorescent dyes. When we add the drug initially, the two receptors come together. Then when the effector is added, they come apart as you saw here. We can see the receptors coming apart when the effector is added. This has a profound effect on self-signaling. When we add the drug, this protein, it brings these two together. We get an activation of signaling. Then when the effector is added, we can shut it off very, very rapidly. This is useful not only for shutting off the drug when you no longer need it, but for probing exactly what the effect of having these receptors together for different lengths of time is, which has been kind of an outstanding question in immunology. The final area I'll tell you about is sustainability. Some of the most remarkable proteins in nature are enzymes. These carry out chemical reactions. The way they do this is they have these very large complex structures. If you zoom in, they have active sites where there are a few key amino acids that are mediating the chemistry and carrying it out on a bound small molecule. We can now design catalysts completely from scratch. We do that in kind of the opposite way. We start by specifying the active site geometry, what the side chain is on, where they should be, and the small molecule or transition state we want to act upon. Then we use the diffusion process to build up a protein around that active site. This is kind of neat because we can explore in nature, and biochemists and enzymologists have been doing this for many, many years, you just have one thing in your enzyme and you can sort of tell stories about how it works. But with this buildup process, we can change what the active site, what the side chains are in the active site, and what their geometries are, and look at the effect on catalysis. We can build up sites of increasing complexity. What we find is as we adjust the geometry and make the sites more complicated, we get increased activity with a partial site. We get enzymes that carry out a single round of the reaction but then stop. We build up full active sites with sort of more extensive networks of interactions. We get quite active enzymes. We can solve the structures of these by x-ray crystallography. Then we can see to check to see how accurate our designs are. They can be extraordinarily accurate. This is now a superposition of the experimentally determined crystal structure and the actual design. The activities of these are much higher than had been found previously straight out of the computer. That was for a complex multi-step reaction. We can also use metal ions to carry out catalysis. This is actually starting with a quantum chemistry calculation of the reaction pathway. We identify the transition state on this pathway. Then we build a protein around the metal and the binding site. This is sort of a blow-up of the site. As I said, in all cases I've shown you, we are making synthetic genes, encoding the designs, putting them into bacteria, and testing them. Here we tested 96 designs. We got a number of active designs, the most active of which had very much higher activity than anything that had been made straight out of the computer before. This is starting to approach the activities of native enzymes for this type of reaction, which have evolved for hundreds of millions of years. There are many problems. This is a bond-breaking reaction. There are many problems where this is important, including plastic degradation. We're now building completely synthetic enzymes to break down plastic and other polymers using exactly this type of active site. One of the other remarkable things that proteins do is to harvest solar energy. This is done in the photosynthetic process, which is very complex, which involves very big proteins. They have at their core what's called the reaction center, which is made up of two chlorophyll molecules. We've been building from the inside out. Here we've designed a small protein that binds two chlorophyll molecules exactly as in the reaction center. Now we can start building this into larger assemblies. This shows taking this protein here and now designing it into these cubic assemblies. Each one of these has two chlorophyll molecules. We can start building up completely artificial photosynthetic systems. We can also now tune the absorbance profile to cover wavelengths which are not well covered by natural photosynthesis. My final example, again something that nature never tried to do, is in mineralization. In nature we have things like bone and tooth and shells where proteins are directing the mineralization of calcium carbonate and calcium phosphate to make these really remarkable hybrid materials. We've been really fascinated by this and so we've been approaching it by designing proteins which have arrays of amino acids that we predict would drive the nucleation and the mineralization of inorganic compounds. In this example though we chose not to apply this to calcium carbonate or calcium phosphate but to zinc oxide, which is a semiconductor. As you can imagine if you could template semiconductor growth with proteins that would open up a whole new range of interesting materials. Here we've lined this part of the structure with this interface and when we incubate this with zinc oxide we get zinc oxide filling this part of the triangle as you can see here, which is opening up really very exciting new ways of making completely new hybrid materials. I've kind of given you a bit of a whirlwind tour of some of the things that can be done with modern protein design. The applications, there are really a wide range of applications in medicine. We talked about pandemic preparedness and cancer immunotherapy, neurodegeneration. There's applications in technology as I've showed you and many interesting problems as sustainability. My prediction will be that over the next five to ten years we'll see an increasing number of design proteins out in the world addressing these problems. Finally, the most important part, I want to thank all the amazing people who contributed to this work. The prize really celebrates the work of a really large community of people currently in my research group and then a very very large community of former students, postdocs, colleagues and friends. It's really been just absolutely wonderful to work with with everyone and this prize is really all for everyone who contributed to this. I'm really happy because I think nearly 200 people have come to Stockholm to help me celebrate this, which is really really wonderful and I hope I get some time to actually see them.