Scientist Stories: Bruno Correia, De novo Design of Proteins

Thanks a lot for having me. It's a big pleasure to be here today to tell you about some of the things that we've been doing in Lausanne. And thanks a lot for the invitation again. So let's go straight to business. Oh, you know, before we go straight to business, just to sort of like mention that, yeah, our regular home is at EPFL, but as it turns out, I've been spending some time in the Boston area these days. And so during this last few months, or during this last month and the next few months, I'm still going to be around. So if you want to, so like, get in touch and talk about stuff, that it's a good time to do it. So let's see. So I will tell you about some of the work that we've been doing in our lab, which the main theme is really this idea on how to engineer new molecules. The bulk of it is really going to be in terms of engineering new proteins. So as you all know, sort of like I'm sure you are aware and you have heard even in the different seminars that you have had in this series, proteins are very important for a number of different reasons. But perhaps the common denominator that I want to have in the different projects that I will tell you about is related to different ways and different representations that we can actually use to try to reduce the space of proteins that we actually have to engineer in order to find proteins that have the correct set of properties. And so I'll guide you through this as we go along. And so today I'll tell you 3 different topics that we've been doing in the lab, one of them related to the novel design and how we can use different types of machine learning approaches to help us with the structural fold exploration problem. Then you're going to have a follow-up from Caspar, which is going to be focusing a lot on this problem and is going to tell you more or less the latest things we've been doing. And then in the two other topics, I'll tell you about our work that we've been doing in terms of representing protein molecules or protein structures, if you will, using surface-type representations and how we've been exploring that concept both for the prediction of function of these proteins in a broad sense, but also for their design. And then in the latest points, I'll show you just some brief results on how we've been using this to start exploring the space of new chemical molecules or new ligands. Okay, so just to sort of like get us in the mood, I'm sure that many of you are aware of what proteins are, but ultimately they are the fundamental molecular units of life, and they perform a lot of functions. And all that this slide will try to convey you is that information about the proteins that are used in nature is encoded in the genetic information in our genomes, and then this genetic information gets translated to amino acids, And, you know, of course, the most— one of the most interesting steps of all this process is that these different strings of amino acids then adopt three-dimensional structures, which will give some biological function into these molecules. And now the biological function of these molecules can be many different things, being that binding and recognition, catalysis, mechanical functions. And you'll see a couple of examples in the next slide as well, right? So speaking about functions. You know, proteins really do pretty much anything and everything that you can, that you can think in biology, both in terms of defense as well as storage of molecules, different types of transport, communication, also the actual real structural aspects of tissues and cells. And ultimately, you know, perhaps one of the ones that I find the most incredible is this capability of performing chemistry, right? So like these polypeptides that can perform chemical reactions. Now, if you are a protein designer, and I'm sure that many of you are, the first problem that you encounter when you're trying to design new proteins is a space problem. The number of possibilities that you have when you're trying to engineer these polypeptides is essentially, you know, a mind blow, by the way, where you very quickly, even if you're just designing very small proteins of the size of 100 amino acid polypeptides, thinking that you have 20 different building blocks for each one of these positions, the space of combinations is more than what our brain can comprehend. And to give you sort of like an idea on how large it is, you can try to compare it to the estimated number of atoms in the universe. And you see that we are many, many orders of magnitude. Even with polypeptides, which for all purposes and means in the biological scenario, they're actually small, right? So most of the proteins in nature are actually much bigger than 100 amino acids. That's something to think about. Now, if you would like to frame this problem in an idea on a sequence and a structure space— well, let's for now perhaps think about the sequence space. You can think that, for instance, nature has sampled this space only at small islands, and that's basically what the exercises of sequencing and genomics and metagenomics have actually taught us through them, through as we gather more and more sequence data, that we see a lot of new proteins. But as it turns out, many of those proteins are in a way sort of like related to each other. And of course, this can be due to a number of different things. Perhaps nature has never really needed to sample outside of those islands, or it just basically, you know, It got stuck in those islands and those islands were sufficiently good for all that nature needs. But now, as it turned out, humans these days have a lot of needs and they have a lot of needs that nature didn't account for, you know, and you have to fight cancer and you have to make vaccines and you have to make drugs and whatever. So there's a lot of needs that nature didn't account for. And as such, the conversation is in this uncharted space, are there sequences which number one, are they folded? And two, are they functional? So those are two different questions, but two interesting, very philosophical questions to understand what is really hidden in this uncharted sequence space. Okay, so now if you are an old person like me, computational design of proteins looks a lot like this. This is basically in the good old times. You would give it a structure. And essentially what the structure looks like is sort of like the yellow part of this protein. So you see sort of like these ribbons. This basically gives you an idea of what the 3D shape should look like. And when this movie was generated, this was basically using a Rosetta Monte Carlo algorithm, which would model explicitly every side chain. And they would ultimately spit you out a score to tell you how likely this sequence was to adopt this structure. And essentially what you get out of this process is it's just a bunch of sequences which your algorithm predicted that they would be going to fold into this particular structure. Now, so things have changed a lot, but essentially the exercise is still the same, right? Maybe your starting point is not it's truly a three-dimensional structure. Maybe there's a lot more now that goes into the starting point, but the endpoint is still the same, right? You have an idea of a protein that you want to design, and you want to have a sequence that it's going to realize that protein. Now, what we hope to learn during this process, there's two main things I feel besides all these more philosophical questions about are there new sequences out there, what do they do, are they functional? But really through this design exercise, one of the ideas is to try to understand the fundamental principles that guide protein structure and function, which remains an extremely hard thing to do even as we get more and as we improve on the design exercises that we've been doing. But this aspect of understanding the fundamental principles, I think it still remains very hard. And at the same time, and as I mentioned, you also hope that you are creating things that are useful in general, right? You know, that do things for you. And here, you know, I'll just list you a couple of biotechnological applications now. So what have been really some of the challenges, or at least one of the main challenges in computational design? So we tend to use this umbrella term which is called de novo design. To basically every kind of exercise which is trying to generate a protein which has not been seen by nature, and where we start in the computer from scratch. And basically, this is what we typically tend to mean to call de novo design. Now, a lot of efforts have been done in de novo design exercises using computational approaches. And here you see now perhaps a slightly outdated slide. But one of the things that we've learned through the years is that we seem to be pretty good at finding sequences which adopt the folds that we are trying to design. So that seems to be something that we have had a handle on and that we keep improving with more data and more sophisticated algorithms. What still remains, though, less easy to do is to actually endow these molecules with functional components. And again, I will acknowledge from the beginning that the word function is perhaps one of the most— one of the less well-defined words that we use in biology. And you say function, well, what's function? Function means something very different for many of the people that use it. But in our context, So like, we're biochemists. We think about function as, for instance, something that has a molecular recognition event and that it's able to bind to another protein. Well, that's a level up than just folding. So of course, here, function is something very broadly defined. OK, and so to give you also— since I'm the one that is doing the primer, I suppose, I'll give you some— of the now historical perspectives on how we've been doing this in the past and what were the kind of techniques and the kind of approaches that we were using. And so back in the day, which was like 3 years ago or 4 years ago, what— how we started handling this problem is we put together this algorithm which we called the TopoBuilder. And the TopoBuilder was essentially conceptualization of how we could actually sample different structures with different levels of resolution. And so, and here I'm giving you an example where our idea is, okay, well, so let's say we have a functional site here. So for those of you that are a little bit far in the room, what basically this is, is a segment of a protein which is able to be recognized by another protein. And so we take this as our functional motif. And then we start defining how can we actually build proteins around a particular functional motif. And at this point, we were just sort of like sketching out 2D topologies. So basically saying, OK, well, should we put the helix? Should we put a strand? How many helices? This kind of idea. And then in the next level, we could actually sample very quickly how can we actually connect the different elements. And then in the next level, we would project all this information into the 3D space and have something that looks more like a protein. Then what would happen is we would pipe all this information— oh, let's see— we would pipe all this information to the computer and actually start designing things. But let me give you an example in this movie so that you can have a perspective. So what you see here is basically the epitope or the functional motif. And as we go along, you start seeing the protein being sketched out and appearing, and all the architectural elements of the protein arranging so that this motif can be stabilized. What we will do after this as a follow-up is to basically sample different configurations to stabilize this motif and then generate sequences which would then be tested. Okay, so we use this for a number of different purposes, and more notoriously, we have used this to essentially present epitopes in designing proteins. And what you see here is an exercise that took us many years to do, but nevertheless, I will just summarize it to you in like 3 sentences. We took a bunch of structural motifs from viral proteins and we made synthetic antigens or immunogens around this viral protein, around this epitope, and ultimately obtaining computationally designed immunogens. Injected these immunogens into animal models and studied their antibody specificities. So let's just— I won't tell you too much about this today. Let's just say that we would we would have liked to see stronger immune responses when we did this exercise. And we'll leave it at that for now. And if you have more questions, we can go at it. Do you have a quick question? Please go ahead. So it seems like, you know, it's a nice— you presented a nice framework for like once you know what the constraints are to get the the proteins that fit those constraints. Is it the case that in the protein design field, as in the knowledge of what we actually want to design is there, and it's just a question of finding the sequences that fit that design? Or is— yeah. Yeah, so clearly here, our approach favors this idea of defining the constraints of the problem by limiting the structure that we want to design. So that's been our strategy. To go at this problem. Now, there are different sets of techniques which can give you more degrees of freedom, which are basically sort of like, you know, you might have heard about this hallucination kind of approaches where essentially you just go and you sort of like sample randomly from these types of structural databases or structural networks and you start generating new proteins from there. So I do think that oftentimes If you are trying to design some type of specific function, it does help you to define at least some constraint of the problem. I'm not saying that it's impossible to go for like from, you know, of course the dream would be, you know, you type design me an enzyme that performs whatever reaction and sort of like, you know, a sequence would come out. I think we still need a few steps to get there. Right, right, right. Yeah. So, so, okay. So that answers that part of the question, but I guess I'm saying, do we know what functions we want to design at all? Or is it? Yeah, so I think that in general you do. In general, that's predefined before you start the project. Just say, OK, it would be nice to have an amazing vaccine. Or it would be nice to have an enzyme that shoots plastic. Or that kind of thing. And then you go to the next level. OK, so if I need this, what kind of elements do I need to build in my model? So in general, you do. Does that make sense? Very good. All right, so yeah, so we've heard a little bit about the TopoBuilder. And so to give you a better idea of what this TopoBuilder was, and this ultimately is going to be important for the kind of stuff that we're doing right now, this is basically a way to just compress protein structure into string formats. And so basically, you know, we very quickly— and again, we didn't invent this, this was actually invented by, or proposed by, by Willie Taylor. And essentially with this type of format, you can actually then translate this information into the connectivity that you have between the secondary structure elements and in which layer the secondary structural elements are found. And so, and from here, what we started doing basically to implement computational workflows that would look like this. And so you basically start with a string format definition of your protein. So string formats are nice because then you can just sort of like randomly generate a lot of strings, and that's good. So it gives you a very compressed way of sampling new protein structures. And then we would basically translate or project this protein these strings into what we called by that time— oh, please go ahead, sorry. Could you explain like what A2E.A1E? So if you're an A, that, you know, I always get the question and I always answer it wrongly, but I'll give you sort of like the guidelines so that you get an idea. So So if you're in layer— so basically, this only works for proteins which you have multiple layers. So for instance, if proteins have very complex topologies, like sort of like these round topologies and whatever, it doesn't really work. So this has to be very simple proteins that you can break down into layered topology, a bit like what I'm showing you here. Now, if you have multiple layers, you know which layers or you know how many layers you have in your protein. So in this case, you have layer A, layer B, so that's all fine. You could have more layers if you wanted. And then depending on what layer you are, you put that letter into your string. And then here, it basically tells you what is the order of this this element of secondary structure in the construction of your protein. And then it tells you the third element. It tells you if it's a strand or a helix. So with this small set of descriptors, you can actually build many different proteins or describe at least their topology. Does that make sense? So it's not the— let me tell you, it's not like you look at these things and suddenly the structure of a protein comes into your mind. But you can decode it very easily with just some very simplistic principles, ultimately. And of course, the advantage of this is that then you can actually play around with many different layers, with many different combinations of secondary strands and whatnot. Now, what we learned from this exercise was that, well, we could translate this information into through these structures, and we call these structures sketches. But these sketches, they were not real structures. That's something that I'm going to talk a lot about today. So you can create this structure. They tell you more or less how a protein looks like. But in fact, real proteins don't have this type of geometry. You see sort of like these strands are all sort of like flat. The way that the helix is actually packing into the strands doesn't look super great. And basically what we've learned from here is that, well, we could pipe in these sketches to guide the sequence search, but the sequences that would be produced, they were not good at all. Because this is the typical principle that if you have garbage in, garbage out. You show a bad structure to a sequence search algorithm, it will give you bad sequences. And that's life. So what we thought by the time— this is already a few years back— Well, can we actually— can we somehow make these sketches more native-like so that then when we go to the sequence search stage, we actually can generate good sequences? That's basically a simple principle. And the way we did this, we essentially used a lot of the tools developed by Gevork Grigoryan and the master searches. And basically by dividing— this structure into subsegments, we could then derive parameters to basically do refinements of the structure, if you will, so that it would become a more native-like structure. And then ultimately, we started using, of course, different types of tools to try to validate our design. And in this case, you start seeing that TR Rosetta and AlphaFold models start coming in. Okay, so, but of course, then, you know, there was the advent of structure prediction networks and how they were being used for protein design. And here we're looking at some of the work that Sergei, that is also a local, has been highly involved in developing in its first steps, which was PR Rosetta. And basically what people had shown was that, in fact, you could actually bias these structure prediction networks to create the type of design backbones that you aim to have by piping in a contact map into these networks. And so going back to this concept of sketches, we started looking at the sketches more as descriptions of their features in terms of geometry. And there was something that was very interesting, and I've already mentioned to you that this these naive sketches, they're not real proteins. And I think when you start looking at these contact maps, you see why they're not real proteins, right? So you have now a real protein, and you derive the contact map in this first row. You have a few other parameters which are based— which are essentially mappings of the dihedral angles between side chains. And you see that there are patterns in this native structure all across the board, independently of which parameter you're looking. That when you have a sketch, you just basically lose them, right? They're no longer there. So clearly meaning that you are in a space where this is not a native structure at all. So what we thought that we could— and maybe I'll stop here before I tell you what— how we approach this problem. Also to tell you that many of these ideas that come, they have a name in the protein design field. Field, and they're called designability. We also didn't really— didn't really came up with this term. This has been work particularly that Bill deGravel has been pushed a lot, right? And the idea is the following, right? So there are particular geometries of bad bones which the 20 amino acids or the 20 building blocks that are found in nature allow you to have. And there's a finite number. It's not like suddenly— the extreme example And again, this also remains an unanswered question. But the extreme example is not like with these 20 building blocks that we have, you might be able to do a whole new set of secondary structures. Perhaps what's possible with what we have— I don't know the answer to this question, but perhaps from what we see, that the secondary structure elements that are possible to do with these 20 elements, with these 20 building blocks, are beta strands and alpha helices. How do you know that a structure is corrupt? Which ones?

Well, do you have people labeling the— in the previous slide, you had the corrupt structure.

Yes, corrupted structures. Yeah. So if I know— how I know that those structures are correct. So I know that the native is correct. And I know that the representation that we were using in order to guide the sequence search, it's definitely not correct. If you see what I mean.

I mean, how do you know it's not another confirmation of a real structure?

Well, because we made it in a computer in a completely artificial way. So it can't— now, I don't know if it can or it cannot exist. That's a different conversation. But at least when we compare the kind of template that we were using to guide our sequence search with a native structure, it simply was just too far, if that makes sense.

Thanks.

Um, just on the topic of the patterns, is it always the case that the, um, like the distance maps with respect to the distance and these different angles look the same? Yeah, so from the, the, the sketches that we generated from native structures, you know, they were, they were always very poor in terms of structural information. That's basically the only message that I'm trying to pass here. Yeah. So any native protein, if you look at these contact maps— OK. So for many native proteins— I cannot tell you for all of them, OK— but for many native proteins, when you look at this contact map, they will have very well-defined patterns. And they'll be the same across the different angles? So like the first, the first row is the same as the second row in the native, or like approximately? Uh, no, it doesn't necessarily have to be the same, right? You know, because here you're actually checking distances and here you're checking angles. So, so it doesn't have to be the same. No, no, it doesn't have to be the same. Okay, but they, they do have to be, you know, in a way, if you want to make an analogy that it's kind of easy, it seems like this, this native apologies always have this very sharp pattern that you see in these maps. And essentially when you start distorsing these structures, they become very blurred. So that's ultimately, I think it's an analogy that helps us. Okay, very good. So here's the point about designability and the point of making that, again, this idea of trying to guide a sequence search with a structure, I think part of the difficulties that we have had doing that exercise are because our initial templates on structures were bad, to put it in a very simple way. And I think now we have different sets of tools where we can actually have templates. We can actually generate structural templates, which allow us to do the sequence search in a much more efficient way. Okay. So how did we actually think about this now more in a 21st century type of solution, right? So what we created was was basically this genesis network where it's basically just a variational autoencoder architecture. And basically the only thing that it was used to do was— or that it had to learn— was to basically sharpen these topological features. So the idea is you start from a sketch of a structure that looks blurred, and you need to walk it until a sketch of a structure that actually starts define. And the question is, you know, could it even do it? And we didn't know when we started. Now I think we have an idea, but we didn't know. And so, in fact, we had to play a few tricks. It wasn't like we could just go strictly from corrupted structures to final refined maps, but we actually had to pre-train it with only an intermediate level of blurriness, and then actually pre-train it in, and then train it finally in two steps so that it could go from completely blurry maps to much more well-defined maps with different features. And so the way that the pipeline currently looks, or at least it looked at some point, I think it's already changing again, but the idea is that you have some like this form, this string. You can translate this string into the sketches, blurry features, and you sharpen those features, and then you can use TR-OZ in order to create these 3D structures. And we've done a few types of tests to try to figure out if things actually— if this pipeline, the genesis step, in fact helps all that much or not. And And so I will just pull up your attention for the different types of pipelines that we have here, where we have generating structures directly from the sketch to Genesis to the PyRosetta, which is just basically the sequence generation step, or going from the sketch to the TiRosetta, or now the full pipeline that you saw in the previous slide. And so here, this score is— If you have higher PM score, it means that you're closer to your target structure. So when you see sort of like the points going out, going up, it means that in fact you are generating sequences that are predicted to fold to your target structure. So you see that in helical proteins, that's what we have here, we're doing very well. In strand proteins also, we can see major gains. In mixed proteins, things don't look so good, which basically also tells you that there are some limitations in the geometric patterns that these networks can actually learn. And I think here, it seems very clear that in some particular types of folds, it wasn't doing so well. But ultimately, where we wanted to get at was at this problem here. And we really wanted to use this pipeline to basically create this fold which had not been seen in nature, or that very rarely have been sampled in nature. And here again, we went to the work of Willie Taylor, which back in the day had mapped a lot of the folds which weren't very common in nature. And basically, he coined this term of the dark matter of the protein fold space. And so that's what we wanted to do. And so what we did here was to basically take some of Willy's proposals for the dark matter of the fold space. And so here you see 3 different examples of folds. You only see their C-alpha traces. We check if, in fact, there are some similar proteins to date in the databases. So you see here the PDB, the AlphaFold database. The matches that we find in terms of root mean square deviation are only at the level of the 6 angstroms, which basically means that these structures don't look anything like that. We also have data for this in terms of TM scores. They basically show you the same type of idea. And ultimately, we sort of like plug in these structures through the Genesis pipeline, create a bunch of sequences, predict those sequences with AlphaFold. AlphaFold seems to think that many of them, in fact, have very high TM score. With high confidence as well. So for this threefold, and now it's time to basically check, is this real or is this not real? And at least one level of validation is what you see here. And again, this slide is probably very horrible for those of you that are not very used to look at experimental data, but I'll guide you through this. So essentially here, what we did, we took the sequences, we expressed them in bacteria, purified them, and saw and checked are they folded? First of all, are they monomers in solution? So that's the first type of data that you see here. So many of them are well-behaved proteins in solution. They're actually not monomers. They turn out to be dimers, many of them. Many of them are also not so great proteins in solution, as you see here. So if you don't see a very sharp peak, this means that this is not a very well-behaved protein in solution. So you see here, —some of them very good. This one, double peak, not so good. And then the next question that we ask is in this middle row. We use this technique called CD spectroscopy, and basically just tells you, is a protein folded or is unfolded? Which kind of secondary structure does it have? And then you can actually test these proteins in terms of their ability to resist to heat. So you boil them up, and you see, do they unfold or they don't unfold? And what you see— in the middle row is that many of these proteins are in fact folded. And many of them are also extremely stable to heat, which is actually a phenotype of computationally designed proteins, which very often this happens. Well, very good. So that was good. So now, of course, the question is, if you actually solve a real structure, is it is it the one that you predicted? And this question, I don't have an answer for you because crystals have been very hard to obtain for these proteins. But at least according to the AlphaFold predictions, that's basically what we see right now. OK, so I would like to conclude this part and just give you a few thoughts in terms of trying— how we've been handling this problem of sequence search by guiding the sequence search using structural templates and how we've been trying to develop this idea that if we sharpen these topological features and if we increase the designability of such structural templates, can we actually improve or enhance the way that we search for sequences. And so, and I think that some of these results go in that direction, showing that in fact we can learn this geometric geometrical features and improve that sequence search. Okay, so now I will change topics for a different set of our work. I don't really know how much time do I still have because I feel— oh, like 20 minutes. Okay, that should be fine. So yeah, the idea here, and I mentioned at the beginning of my talk that I was going to tell you about protein representations. And what we are doing here, or what we've started doing a few years back was to pose this question about what can we learn from protein structure and function just by studying the surface features. So as you know, we can represent protein structures in many different ways. I showed you how— the richness of different ways that you can represent protein structures, and surface is also one of those ways. A surface representation is also one of those ways that we can represent this molecule. And you can think about this surface as basically the solvent-excluded surface where the solvent of the protein cannot penetrate. And also in this one interesting aspect, and this will come also become very clear during the talk, is that these surfaces have inherent features to them which are both geometric and chemical. And that geometrical— those geometrical features, you can in fact think about them as, well, am I a flat surface? Am I a concave surface? Am I a convex surface? And in the chemical features is more so like the typical thing that you're used to, you know, do I have a positive, negative charge? What kind of potential I have here? Now, one thing that has been— that drove the way that we thought about this problem is that it was known that for proteins, for instance, to perform molecular recognition events, it is necessary and required that there has to be chemical complementarity and shape complementarity. And only if molecules in fact fit a bit like a 3D puzzle, they are actually able to interact with one another. And of course, as we know, interactions between proteins is one of the initial aspects that drives their function. Okay, so ultimately the way that we posed this question was, well, can we actually identify the surface patterns which help us to determine functional features of proteins? And you'll see an extension of this in a few slides on how we can actually do this to change the surface pattern. And this also, I will only give you a glimpse of the work that we've done that has been published a couple years ago now. And so what we implemented here was this MASIF framework, which stands for Molecular Surface Interaction Fingerprint. And what this framework does is to basically generate fingerprint descriptors, which are nothing less or nothing more than numerical vectors that encode the surface features. And so the way that this goes is you have a protein structure. You derive its solvent-excluded surface, and then you start segmenting that surface into geodesic patches just like the ones you see here. And once you have those patches, then, you know, each of those patches— there are also points that compose the patch. So you see these little points that are inside, and what you start doing is enriching each one of these points with features. And these could be— well, these will be both geometric features as well as chemical features. In addition to this, to this set of features, you also provide this basic data unit, which is this geodetic patch, a set of polar coordinates, so that basically each point knows where it is relative to one another, and so that we don't lose the information in the 3D space. And then once you have this, you can actually pipe this information into geometric deep learning layer where you can train it for different tasks and create these fingerprint descriptors, which then you can use for different applications. So which type of applications do we actually use this for? Typical prediction tasks, I would say. So let's say that you have a structure that has a clear protein pocket. One of the tasks that people in general are very interested in is, well, which kind of substrate can bind to this pocket. And so we've done a very limited type of exercise here to try to do this kind of prediction. Another thing that we thought that was kind of interesting is, again, you know, with this idea that you start with a protein structure, can you actually pinpoint which sites are responsible by the interactions that this protein has with other proteins? And again, here the assumption was that there are some some fraction of the sites in a protein surface that are there for solubility, but some others that are there for interaction with other molecules. And this turned out to be a very interesting exercise and perhaps one of our best predictors, in fact. And then a more canonical exercise of protein docking. Can we actually use these surface fingerprints to understand how proteins form complexes? So I won't guide you through all the results, but I'll just highlight a few things that we've learned in that particular paper. And so we learned that it was actually that we could very quickly learn which patches interact with which and which patches don't interact with one another. And so, and basically by just training By just training the algorithm with the ground truths for patches that are interacting with one another versus random patches, you could very quickly get to very high performances. Now, so that was interesting because it means that the algorithm could learn this type of information, but then when we actually used it to compare to docking protocols, —and in this case, we used ZDOT by the time, which was one of the best—what we saw was that we weren't necessarily better. So here, a higher number means that you're solving, that you're getting the correct answer for most of the complexes. So we were not necessarily better than the typical algorithm. What we were, though, was really, really fast. So we were just as bad, but at least we were faster. So that was basically the conversation. And this will become also important for a different type of exercise that you'll see in the new slides. Now, as a second incarnation, let's say, of the massive framework, we realized that our piece of software was so bad that it was so clunky. It had so many dependencies that nobody really wanted to touch on it. And neither did we, by the way. And so what Pryor started doing was, well, OK, let's make something that is a bit more end-to-end and that, in fact, we can compute the electrostatic features on the fly just based on the atoms that are underneath the surfaces. And this turned out to work pretty well, and it has been a very interesting direction for development to make these algorithms a bit more or a bit lighter in terms of usage. Now, what I'll spend the last part of my talk is to tell you how we actually use this type of framework to design new protein-protein interactions. So for those of you that follow the literature, you will know that designing protein-protein interactions is actually very hard. And it's hard for a number of reasons, which go well beyond this talk. But it's hard because in protein-protein interactions, there's only a very restrictive set of molecular interactions which make this interaction be possible. Only a very restricted set of amino acids that does this. There are many factors that are important for protein interaction, let's say like solvent, which oftentimes we don't use in our representations. And there are also aspects, for instance, related to the dynamics of the proteins, which are absolutely key for any kind of interaction. And in fact, when we formulate this problem in the computer, We don't really account for any of those things. It's just sort of like, well, you know, we just have a very simplified perception of what the— or description of what the problem is. And so it is not very surprising that it is a very hard problem, right? Okay. And so how did we do this, right? Or how did we try to implement things? So the exercise here was, or is, to start with a target protein. Where you define a site where you want to design a protein to bind to that specific site. And so what we did here was to basically define the site, drive a target surface fingerprint, and then because we have a very fast docking program, what we did was to assemble what you can call a fragment database with just protein motifs where we could drive fingerprints from all those motifs and just sort of like very quickly go over databases, sort of like, let's say, 100 million fragments, which basically made this search very, very broad. And ultimately what you would get out of this search is basically a putative binding interaction between a motif and your target of interest, which we then would transfer into a protein scaffold so that we could present that surface at, um, in the context of a folded, of a folded structure. So, and then of course it was time to test it. I won't tell you all the sordid details, um, of what we have done experimentally, but I will tell you how things have been evolving in our hands. And so one of the first targets that we decided to go after was SARS-CoV-4. Reasons that it was very easy to access reagents, and we had many things that we could do. And so we defined one of these target sites that is here according to the predictions from the massive site prediction, and we started figuring out, okay, are there any motifs that could bind to this site? And so we generated 63 designs, and ultimately what we got straight out of the computer was basically one of these very residual signals where you could see, oh, there's a hint of binding, but it could also be just a human. And so this is where experiments start sort of like killing you because you don't really know if the right thing is happening or not. So what you do in this situation is, well, you start like asking questions to your experiment. And what we did in this case was, okay, so if this little signal, little binding signal is competed by one other binder that we would predict that would compete, would the binding signal go away? And so, and it did go away. So we're mildly confident that whatever very low affinity interaction we were seeing was in the predicted site, or at least close to the predicted site. So, and that, as it turns out, what happened here is that ultimately we had to have an original computational design and then improve it with a number of random mutagenesis that— well, I shouldn't say random. In fact, it wasn't random because we knew which mutations we wanted to use, but combinatorial libraries and site saturation mutagenesis in order to bring a scaffold that had very low affinity. So let's say 4.6 micromolar to something that would be 80 micromolar. So here clearly this was still an exercise where we use both computation and experiments to basically reach high affinity. A few more functional characterizations here in terms of inhibiting virus replication and all those things, but I will not spend too much time on that. Ultimately, and I don't have a slide here, ultimately we did solve the structure of this design and was in fact in the right side with the predicted orientation. So giving you another example, we go to this PD-L1 target. And basically this PD-L1 target had something very interesting to it, which was an extremely flat surface. And so for us, this was interesting because we wanted to know if we could design molecules that could bind to a very flat surface. And so again, we do the same kind of search, right? We start seeing which kind of clusters in terms of protein interactions emerge from our motif search. We select one of them. We also start to see that this motif search then comes with some sort of sequence preference as well. So there is convergence both at the structural level as well as at the sequence level, which is good. Again, we do a few designs. In this case, I think we did around 60. And again, what we observe is basically we get this very residual binding binding signals. I didn't explain you much what this experiment is, but on the y-axis you have binding, on the x-axis you have display of the scaffold or of the design in this case. So you can see that we're having sort of like very residual binding sites. However, when we actually went and mutated the binding site that we have designed into this protein, we could see this signal going away. So we saw this once. We saw this twice. It was again the same story. And then you start doing a different experiment, which was, well, so if this design, when we engineer them in the computer, they don't really come out with high affinity, how could we make them high affinity? And so the experiment that you're seeing here is basically this high-throughput experiment where we mutated each residue of the design and asked the question, do you improve binding? Or you don't improve binding. And we've learned a couple of interesting things that then guided the next improvement. So what we've learned is that— so the positions that you see here that are boxed are all the positions that came from this binding seed, so sort of like where the core interaction was happening. And these positions were mostly optimal. So you could not mutate them for any other residue, because otherwise this would just basically be deleterious for binding. Where we saw that there was a lot of improvements to be made was on what we call the ring positions, which are sort of like these positions which are half solvent exposed, half buried in the interface. And we saw that here many other residues in fact would help us to improve the binding affinity. So to give you an idea— oops, sorry— to give you an idea on how we went through this, we started with something that we could barely measure the affinity experimentally, which is like around 2 micromolar, and then we could get to something that was already 265 nanomolar. Now maybe something that is coming into your mind as well, but is this really binding the right way? And so here we solved two crystal structures together with our collaborators showing you the difference between the crystal structure and the model and that impact. The complex is essentially dead on. So the prediction that we have designed and the mode of interaction that we had engineered and predicted was in fact what came out of the real crystal structure. Maybe one more thought before I tell you what the next steps for this design of new protein-protein interactions were. Another question that came to our mind was, well, look, We're sort of like docking the seed, but are these seeds really novel? Are they in fact just picking up motifs that nature is already using to do protein-protein interaction in this particular target? Or are we really innovating here? Are we really sort of like sampling new stuff that we didn't know that it could do it? And this actually here is a representation of the native binder. So in this case, you have PD-L1, our target. You have PD-1, the native binder. And what you see is that the native binder, in fact, uses a strand hairpin or a strand turn, a beta turn— that's the right answer— to establish this interaction. And what we did is a completely different secondary structure and uses completely different types side chains to basically create a functional protein-protein interface, really telling us that in principle, we are exploring new motifs for protein-protein interactions, which we are accessing through this finger, through the surface fingerprints. Okay, so again, I'm kind of running out of time here very quickly, but our idea and our dream is always, well, we want to press that button in a computer and just generate binders, right? So from doing, from looking at everything we learned and all this mutagenesis, how could we actually improve this pipeline? And essentially we started doing work in trying to improve the pipeline by enriching our interfaces with hydrogen bonds so that in fact we could now try to get more binders straight out of the computational design. We don't have to to keep improving them experimentally. And so here, just to give you an example where what we did is we went to a different target. So now we're talking about PD-1. And we designed 1,000 different sequences. And then we just asked, from this 1,000 different sequences, can we actually find binders? And the answer in this case was yes, we could actually find binders. In fact, we found 2 binders. This one clearly should not classify as a binder. And what we also saw was, again, the same kind of message when we— and here what you have is an AlphaFold model telling you what the interaction looks like. And this is in agreement with our model. And we kind of saw, again, yet the same kind of message where we were doing very well in terms of the seed residues, but not on the outside. So again, there's still work to be done there. Although it seems like at least we got something to work a bit faster this time. Biochemically, we measure this affinity, and it was a low affinity interaction as well. It was sort of like a 6.5 micromolar. But if in fact we compare this 6.5 micromolar with a native affinity from the native complex, this is 8 micromolar, and it came straight out of the computer. So in that sense, it was a success. Now, still very far from what monoclonal antibodies do. Okay, so I think I will just stop here for today because I think I gave you enough. And if we want to discuss, I hope that I gave you enough things that you have doubts about and questions about. But ultimately, what I've told you about in this second part was related to this idea of representing surfaces and surface features using vector fingerprints and how we could actually learn from this data representation aspects about the function of these proteins? How could we actually use this for the efficient design of site-specific binders with high structural accuracy? And then again, as in the future, there's a lot of work to do. So in terms of getting architectures that are much more properly designed in terms of being end-to-end rather than what we have right now, improving the surface descriptors for different types of tasks, and also using this kind of idea for drug design. So I won't tell you much about drug design today, and I will essentially jump to the most important slide. So I've been extremely fortunate to have worked with the people that I've worked in Lausanne. They really have been an incredible group of people. Highly motivated, never turning their backs to hard problems, and it's really been a lot of fun to have had this team in Lausanne through the years. And so if it wouldn't be for them, clearly we would have not been able to do many of the things that I've told you about here today. So with this, I thank you for your attention and your resilience in this morning, and I'm happy to take any questions you might have.

Um, so yeah, as you know, Bruno, my PI, he's doing a sabbatical here and he said, why don't you come over for a few months? And of course I cannot say no to that. Um, I'll be talking a little bit about the research that I'm doing, and I'm now in my second year of my PhD, almost third. In 2 weeks. And yeah, I'll just explain a little bit about what I've been working on, the problems I'm running into, how I've been solving them. And please feel free to interrupt me and start a discussion with me and ask as many questions as you can. So let's zoom out a little bit. Let's do a recap. So proteins, they perform various functions, and they're built up out of amino acids, which are their building blocks, which come in 20 different flavors. And the combination of these amino acids, they determine the 3D structure these proteins fold into. And this structure in return determines its function, which can be cell signaling, immunity, it gives structure to cells, it performs biocatalysis, and it can perform transport. So it's essential in all living organisms that we know of that they have proteins that perform these functions. Now, as Bruno explained earlier, what we can do is we can think of these protein sequences and these proteins as a protein universe, and nature has already explored a part of this protein universe, which you see here in green. I don't know how well you can see it, but here in the center, this is the part that is explored by nature. And what you can do is you can introduce— you can take these natural proteins and introduce some mutations. And these methods are, for example, guided evolution. And then you're pushing the boundaries of this green point in space here. But it's not really what we want to do. We want to engineer really, really novel functions, new protein folds, and yeah, new functions. So we want to explore these red parts of the protein space, which we'll call the protein dark matter. So exploring the— why is this important that we expand this protein space? It can potentially lead to new therapeutics. For example, think of CAR-T cell therapies. It can lead to the detection of new biosensors. Think, for example, of the COVID-19 self-test. It's an example of a biosensor. It can lead to the development of new enzymes or new biomaterials. So generally, we think of proteins in 3 different domains. We think of proteins in the sequence space. In the structure space and in the function space. So the sequence determines the structure and the structure determines the function. However, as a protein engineering lab, which we are, we're interested in going from function to structure back to sequence. So we want to go the other way around. So Bruna already gave a good introduction this morning on how this would go traditionally. So just a quick recap, what you can do is you can take a look in the PDB, which is a protein database full of natural proteins. We can select, we can take a functional motif, which is a part of a protein that performs a certain function, and then we can check this protein database and see which part fits our functional motif, can best support our functional motif, and we just grafted in there, we transplanted in there. And then you also have more ground-up methods, because with these grafting methods you're really, really dependent on the proteins that nature has provided to you. But you get a lot more degrees of freedom when you start building up your proteins from protein fragments. So you can fragment this protein database and take your functional motif and build a protein around this functional motif to stabilize it, given your database. So you already increase the space that you can explore this way. And recently there's been a breakthrough in machine learning techniques. So these days, these classical methods, even though they're only 4 years old, they're rarely being used anymore. And most of it is now happening using machine learning techniques. So I'll be telling you a little bit more about that. But first I want to make a little sidestep. I don't know how many of you are protein engineers and familiar with proteins, but the way how we determine what a protein looks like is the gold standard method is X-ray crystallography. So if you express proteins in a high concentration, you can put them under certain conditions and then they might form crystals. And then when you shoot these crystals with very powerful X-rays, you get a diffraction pattern. Then this gives you an electron density map when you release some clever algorithms on them. And then you can, since generally you know the sequence that this protein has, you can fit an atomic model to this electron density map. And over the last few years, there's been a significant increase, an exponential increase in the amount of protein structures that are available to us. So as you can see, in 2021, we already hit 180,000 entries in the PDB. So yearly more and more structures are being added to this database, and I think at the moment we even surpassed the 200,000 entries in the protein database. So this is still, I should update these numbers a little bit, these are still a bit outdated. And the same we observe in the sequence space. So when we take a look at some popular sequence databases, like UniParc, UniRef, you see that there's also an exponential increase in the number of sequences, the data that is available to us. And why are these sequences so interesting to us as protein designers? And since mostly we're interested in the structure of these proteins, so I need to make another little sidestep. I need to tell you something about multiple sequence alignments. So what you can do is you can take a protein sequence. So for example, we take a protein that occurs in a human, and then we can do a database search in all these sequence databases, and we find homologs of these proteins. So we find the same protein in, for example, a mouse, a pig, a cat, a donkey, and we can align these sequences. And what you would find is that sometimes these sequences, they have conserved residues. For example, if this residue, so this amino acid would mutate, the protein would lose its function and the organism would die. So these are conserved. But you also see something, something else. For example, here you have arginine and asparagine. Aspartic acid. So arginine is positively charged, aspartic acid is negatively charged, and what you see is that there's a pattern and evolution between the two, which generally means if two proteins are close together in 3D space, they tend to mutate together. So if you have this positive and negative charge close together, if one of them would mutate, they both would become positive or both negative, they would repel, and the protein might unfold and become unstable. So these tend to mutate together in a correlated manner, and this is something we call co-evolution, and this is very, very powerful. This contains a lot of information about protein structures. So determining the protein structure using using crystallography is very, very difficult. It's, first of all, it's very expensive. Second of all, it takes a lot of time to grow these crystals. And third of all, if it's even possible to grow the crystals in the first place, because sometimes they just won't grow. So they came up with something called CASP, the Critical Assessment of Protein Structure Prediction, where basically a bunch of biomolecule modeling nerds, they come together. I'm also included. They come together and they start having a competition to see which protein, which group is best at, given an amino acid sequence, to predict the protein structure. So what they do is they release a they have a group that wants to apply a protein structure to the PDB, but they hold it back. They hold it back for CASP, and they only release the sequence, and then a research group can enter this competition. They can give a prediction of the structure, they send it to the jury members, and then they will measure how accurate they are in their prediction. And this is considered the world championship of protein folding. So as you can see, over the last few years, the CAS progress has been more or less limited. It's been stuck on a plateau. So GDT tells you something about how accurate the predictions are of the best group. However, now something happened in 2018. And that was AlphaFold. So this is an algorithm developed by Google DeepMind which made use of this enormous amount of sequence and structural data by training a neural network on this data to predict protein structures. And then we thought, wow, the GDT of 60 is reached, this is incredibly high already. What are we going to see next year? Are we going to see an increase of maybe again, maybe 5% or something? But then AlphaFold 2 was released and they surpassed the GDT of 80 threshold, which means that for some structures they were so accurately predicted it becomes hard to differentiate them from crystal structures, that they become so accurate. So what does the AlphaFold 2 pipeline look like? So you take an input sequence, when you, yeah, you take an amino acid sequence, your input sequence, and you want to, of which you want to know the structure of. Then you do a genetic database search and you generate an MSA, like I explained earlier. And what you also do is you check the sequences of the MSA, and you do a structural database search. So basically you're finding proteins that are very, very similar to your query sequence, and then you're going to check, hey, are there also structures available? So we can maybe copy some parts or already have an example of what this protein should probably look like, or what a certain protein domain should look like. And These are encoded as the MSA representation and the pair representation, and this goes into the AlphaFold model, which exists out of the Evoformer. So the Evoformer allows reasoning between sequential and structural information. So this uses a lot of transformer architectures to do its reasoning. And then the output of this Evoformer, it goes into the structural module, which eventually takes this latent representation of the protein and maps it into 3D space. And the great thing about this pipeline is that it does not only predict the protein structure, but it also predicts how confident it is in this protein structure. So as you can see here, the blue parts of this protein means that it's very highly confidently predicted, and the yellow parts mean it's lower confidently predicted. Then you see that the helices, which are generally stable secondary structural elements, they're very confidently predicted, and the loop regions, which generally just flop around, they are predicted in lower confidence. And on top of that, what they do is they introduce recycling, So basically what that means is they take their outputs and they put it back as the input. So it's just a fancy word for iterating, and this seems to also increase the accuracy of the prediction. So now, as a protein designer, and based on previous work with other structure predictions networks, we thought, okay, can we apply this to AlphaFold 2? Can we use AlphaFold 2 to design new proteins? So AlphaFold 2 takes as an input a sequence, which you can see here as a one-hop encoded matrix. So this matrix is 20 because you have 20 different amino acids by the length of the protein, and we pass it through AlphaFold 2 and we get a predicted structure. So we initialize with a random sequence, and then we can take a target structure. For example, we're trying to design for this red protein structure, and we can calculate the structural loss, which in the case of AlphaFold is calculated as the FAP loss, the frame-aligned point error. And what we can do next is we can take the derivative of the input over the FAP loss. So the way how you train neural networks is you take the derivative of your weight over your over your loss, and then you use it to update your weights. But you can also retrace this all the way back to your input and keep the weights fixed. So you start updating, updating your input. So by taking this derivative, you get an error gradient, and this error gradient can be used to create a PSSM, which stands for position-specific scoring matrix. So again, this is a matrix of 20 by the length of the protein, And this basically tells you, for each of the residues, it gives you a probability distribution of which residues are most— you're most likely to find in that position. And what you can do is you can take them for each position, the most likely residue, and plug this back into AlphaFold and then iterate this a couple of hundreds of times until you converge to to a target structure. So one of the things that we noticed when we went after 5 design targets, so here we have TOP7, protein A, protein G, ubiquitin, and 4H, and for the people who followed the talk this morning, 4H is actually based on a sketch, so it's not based on a natural protein from the PDB. And we noticed something peculiar, and that is that there are a lot of hydrophobics on the surface, which are colored red here. And generally the number one rule in protein folding is that the hydrophobics, they, they go on the inside, because they don't like to be in contact with water. So what would most likely happen if we would try to express this protein in cells, it would either not fold or it would aggregate. It would bind to all the other hydrophobic proteins. So we had to implement a Rosetta surface redesign step to fix this surface. However, at the moment, a new improved method came out, which is called ProteinMPNN, which is a message-passing neural network. Which had an incredibly high success rate. So by now we've replaced the Rosetta surface redesign step with protein MPNN. So let me, let me go into a little bit of detail on how this works since we're at a talk about algorithms. So you can represent biomolecules as a graph. So for example, if you take a molecule here, and you take this carbon atom here in red, you can represent the atoms as nodes, as the circles here, and you can represent the distances between them as edges, which are the arrows. And the nodes, they can contain information about the atom type or other chemical features. And the same we can do, this is just an example of a molecule, but we can do the same for proteins. And this is exactly what ProteinMPNN does. So ProteinMPNN stands for Protein Message Passing Neural Network. So we can encode the protein structure as a graph, and then we first do an encoding step of the structure. So we represent each amino acid as a node, The distances between the nodes are represented as edges, and then we allow each residue to see the 48 closest neighbors, and we iterate and we encode this structure as a latent vector. Then next, we start to decode this structure. So what we do is we start at a certain position, and then we let the message-passing neural network predict which residue is most likely to be found there. And this is an iterative decoding step. So once it determines that a certain residue is placed at a certain position, we will fix it in the network and we will iterate again and then ask, okay, so what is the next one and what is the next one? And we iterate this until the entire sequence is is recovered. So yeah, let's move this into practice. Let's see if this pipeline can indeed design some novel proteins. So I want to give as an example, zoom in on a protein topology called the TIM barrel. So TIM stands for triophosphate isomerase, and this is basically the first enzyme that was found that had this topology. But there's something very interesting about this topology, it is because 10% of enzymes, they have this topology. So there's something about this topology that makes it a very good enzyme, which most likely is that it has a nice pocket on the inside where it can perform all kinds of biochemical reactions, or it can evolve to perform these chemical, biochemical reactions. But the design of these tin barrels traditionally before deep learning methods, it has been extremely challenging. So for 25 years they've been trying to design a tin barrel, but only in 2016 there was some success by Poh Suhuang. And the way he did it, he applied fourfold symmetry to the tin barrel. So what you do is you divide your thin barrel into 4 pieces and you only define, you only design for a quarter and you apply symmetry rules. So this really crunches down your sequence space, the design space, so it makes the design process easier. Then they use a software program called Rosetta and they have to add all kinds of constraints. For example, all helices are N and C capped, which means that the hydrogen bond on the the backbone needs to have a hydrogen bond on the end of PDCs, the sheet-facing alpha helices cannot be all alanines, all backbone hydrogen bonding groups in the loops must be satisfied, and on top of that there was manual design needed afterwards. To be fair, the manual design steps, they have recently in another paper by Po-Se Huang's group been removed, where they also apply a convolutional neural network to design these, redesign these structures. But overall, what I'm trying to tell you here as a takeaway is it's been challenging to design these structures and you need to apply some constraints on the algorithms to get success. So we plug this into into the pipeline. And first we use our AlphaFold gradient descent, and then what you can see is we made about 50 designs here. And what you can see is that they have a TM score, the general bulk has a TM score above 0.8. So TM score, it tells you something about how close two protein structures match, and generally we say if it protein structure has a TM score above 0.5, then it's the same structure, which is still very, very, very much debated definition. But if you're above 0.8, you can clearly say, well, these are very similar. Then after MPNN redesign, you see that we get more sequences, but they more or less, they have the same TM score. And what we also did is we re-predicted these structures, but with RosettaFold, which is also a structure prediction network independently trained of AlphaFold. And they gave back the structures as well for most of them. But as you can see, we get a lower success rate with RosettaFold, which makes sense because where you do our design and evaluation with AlphaFold, right? So you would expect that RosettaFold, a different network, would pick out these adversarial sequences and tell you that some of them are wrong. And then when we take a look at the E-values, which tells you something about how close these sequences are to natural sequences. So generally we say if an E-value is above 0.05, it's a de novo sequence. It's so different from the one in nature we consider it something new. So when you look at the designs we got from the AlphaFold pipeline, you see that they're indeed de novo sequences. And the same when we look at the redesigns from MPNN. There are a few that are considered natural, but we filtered them out. So we ordered 40 of them, and 5 of them, they they showed to be expressing. So I'll guide you through it a little bit. They showed to have a single peak in the SEC-MALS curve, which basically means that they're monomeric and expressing, that they are folded, which you can see in the circular dichroism curve. So this means that if you shoot light in a certain polarization at proteins, they will the polarization will change depending on the secondary structural element composition that this protein has. So what we saw is that indeed alpha helices and beta sheets are present in this protein, and when we increase the temperature of this protein, we see that this CD spectrum, it barely changes, which tells us that this protein is very, very stable. So what we thought is, all right, it looks like this protein is expressed and biochemically well behaved, let's, let's try to make some crystals. And we seeded these, these proteins for crystallization, and within 2 weeks we got thin barrel crystals. So we got incredibly lucky there, and we shot these crystals, and eventually we got a structure. And this the design is shown in grey, and the crystal structure itself has these rainbow colours. And what I found very interesting about this prediction and this crystal is how closely not only the backbone atoms match, but also how closely the side chains match to to the design. So we got a full atom RMSD of about 2.1 angstrom, which is about 2 hydrogen atoms in accuracy away. So yeah, we were about 2 hydrogen atoms off from our design. But now you can ask me the question, but Kasper, why do you need this AlphaFold redesign step? Why don't you just take this a tin barrel architecture and you just pass it through this protein message passing neural network, why do we need this AlphaFold step in between? So let me take you back to the E-values on how close they are to natural sequences. And I would also want to show you on the right here a figure that was in the supplementary information of the protein MPNN paper. And what you can see is that for— monomers, homomers, and heteromers, the sequence recovery where there are a lot of neighbors, so the average C-beta distance can be seen on the x-axis, so the average C-beta distance for the 8 closest neighbors, you see that the more neighbors it has, the higher the sequence recovery. So in the core of the protein, you can have up to 95% sequence recovery, which drops steeply the less dense packed the core is, which means that the core is overall very well preserved when you redesign with MPNN and the surface is being completely redesigned. So what this means is when we take a natural protein and we pass it through MPNN, which you can see here with the axis, so this is in the case where we did these two examples, the core will remain more or less the same. So we get very, very low e-values, and note that this is on a logarithmic scale. So these sequences still have a natural core, and this matches against the natural database. So we also tried something else. So we thought, okay, we check the sequence recovery when we take the original crystal, but what we can also do is we can put it through an MD simulation. So we do some molecular dynamic simulations and then we sample some of these states from the MD simulation and we pass it through ProteinMPNN and then we check what the sequence recovery is. So from the original crystal it's around 45%, and then when we move to the MD simulation it's still around 40%. But then when we apply the AlphaFold design pipeline and also MPNN afterwards, you see that the sequence recovery is way lower, which basically what this pipeline is doing is it's redesigning the core as well. And I think that's where the power comes from with this pipeline, where we're also really exploring the unnatural protein sequence space. So what else can we do with this pipeline? So one idea that we had is we can take transmembrane proteins and we can try to make them soluble. So for example, these transmembrane proteins like GPCRs, claudins, and rhomboid proteases, these protein folds they only occur in the membrane, which raises interesting questions. Are these topologies only possible in the membrane, or are they also possible to have them in a soluble form? So I want to give you an example of the GPCR, what we've been working on, and we actually got some very exciting results last week. So for the GPCR, it's a 7— the topology exists out of 7 transmembrane helices, and this topology we don't see in soluble proteins. And what would be very interesting is if we can make this protein soluble again. Since GPCRs, they are excellent small molecule binders, which also gave us the idea is we can then repurpose this topology as a small molecule binder, which we can do by either designing a new small molecule binding site de novo, or maybe transplant back a small molecule functional, small molecule binding domain from an existing GPCR. So here we have the original GPCR, and I colored the hydrophobic residues orange. So what you can see is that on on the surface, it's completely covered in hydrophobic residues. And when we redesign it with AlphaFold and we count the fraction of surface hydrophobic, so the fraction of hydrophobic residues on the surface, what you see is that this is around, well, 15 to 20% of the surface is covered in hydrophobics, which is what you would expect in a soluble protein. But then, when you pass it through protein MPNN, it recognizes the folds as a transmembrane fold, and it places the hydrophobics back on the surface. So this was one of the problems we first ran into. So we thought, okay, what we can do is we can bias MPNN, so we can bias it towards sampling hydrophilic amino acids, and this, as you can see, this shifts the distribution a little bit downwards, so we get more hydrophilics on our surface, but yet it's not great. So then I met with Justus, Justus de Parros, who is the first author of the protein MPNN paper, and he said, well, how about we try to retrain on a dataset with only globular folds, so no transmembrane pulses, which results in these two networks. So we have an MPNN version trained on non-membrane proteins with two different types of noise level that they were trained on. And as you can see now, our distribution shifts even more towards globular proteins. So we picked some of the most promising designs and we started seeding them as crystals. And again, I think I'm one of the luckiest PhD students because within 2 days I had crystals. And yeah, I have no idea whether this is because of ProteinMPNN that is trained on protein crystals, so it's really good in producing structures that are crystallizable, or I'm just really lucky, which Bruno seems to think I am.

I'm, well, yeah, I'll take it, I'll take it.

And we shot these crystals actually last Thursday, and we can already show you the backbone. The side chains haven't been optimized and fitted yet. So on the left we see the design, so you can see all the hydrophilic residues on the GPCR, and on the right you see in gray, you see the design, and in colors you see the the crystal structure. So we have a C-alpha RMSD of about 0.75 angstroms. So this is really cool to see that this pipeline is working. Yes. Just a quick question here. Does the GPCR retain any of its other functionality? No, no, no. So I might be saying we're redesigning a GPCR here. We're not. We're redesigning a GPCR topology because the sequence is nothing like the original GPCR, so it will have nothing of its function, which will probably be the best— the next step to bring some of the function back. We'll probably not bring the full function back because the dynamics will be different, of course, than in the membrane. And second of all, we can only bring part of the function back. So some of the ideas that we were having is we can, for example, functionalize these GPCRs by adding back the small molecule binding part. So we can check, we can transplant back residues of the small molecule binding domain and check if it still binds small molecules. So this is what we're working on right now. What we also can do is we can transplant back part of the extracellular domain. And this can be, for example, be very useful for the pharmaceutical industry, because they're very interested in antibody therapies, and especially antibodies that target-specific GPCRs. And at the moment, it's really difficult to get these GPCRs expressed in high enough concentrations and get them out of the membrane to do your experiments on. Yes. Follow-up: it's a really cool proof of principle, but why do you actually want the soluble GPCR? Yes, yes. So right now, if you want to Well, the end goal is to transplant back function or parts of the function. So first of all, we can repurpose these GPCRs as small molecule binders, which is a lot easier when they're out of the membrane since you can acquire them in, in high concentration. And second of all, we, we want to transplant back the extracellular part on the protein, and then it can be used for, for example, antibody testing. If you want an antibody that targets a specific GPCR, you can transplant back the cellular part, see if it binds. And it's a lot easier because if you want to do this on a GPCR that's in the membrane, you only have 2D expression versus 3D expression when it's dissolved in the cytosol. So you can get it easier in way higher concentrations. I hope that answered your question. Okay, let's have the discussion. So yeah, let me conclude. So what we can do at the moment is we can design complex topologies quickly without design constraints, and we have shown that you can take this GPCR membrane topology and you can make it soluble. And what What I also want to be— what I want to do next is to go after one-shot functional design. So we had a 60% success rate with these GPCR designs, so it would already have been interesting actually, and then probably possible from the beginning to already design for function and then get a functional protein from your pipeline from the beginning instead of having to go back to the drawing board and redesigning these these GPCRs. So I would like to thank everyone who helped me with this project, and especially Bruno for giving me the opportunity to be here and to speak here, but also to Martin Pratsessa and Lars Dornfeldt who've been helping me with this project in the wet lab. And with helping, I mean they did the wet lab part I'm just a computer guy. So thank you very much.