Scientist Stories: Andrew Kruse, GPCRs & Antibody Discovery

Today our seminar speaker, and actually it's our first seminar speaker for this academic year, and Paul told me to remind everybody that all other speakers will be at 2 PM because we initially were planning the, you know, in-person seminars, but because we're switching to Zoom again, then we're switching back to 2 PM. So other than today, um, so Andrew, I actually wanted to meet him. I actually kind of regret that we're doing it by Zoom because I never met Andrew before, and but I'm really impressed by his work and know him from the papers, and basically actually really wanted to see him in person. But, um, so he is, um, he has actually pretty spectacular career actually. If you look at his CV, it's kind of amazing. So he was undergraduate in University of Minnesota in Minneapolis, and he graduated with, you know, high honors in biochemistry, chemistry, mathematics. And actually, he told me that he already as an undergraduate student, he did quite a bit of research and actually published papers as an undergraduate student in protein crystallography and actually learned protein crystallography as an undergrad. And then from there he went to Stanford University to graduate school, and he did PhD in structural biology with Brian Kobilka. It's actually the same department I did my postdoc, but in a very different time, much later. So Brian, of course, we all know, you know, got Nobel Prize for the studies of GPCR receptors, and that's what Andrew actually was doing in his lab is to do structural biology of GPCR and sort of study GPCR signaling. And actually was one of the people who set up sort of structural biology in Brian's lab. And he was so successful as a graduate student that he actually didn't do postdoc. He actually went directly from after graduation of getting his PhD degree from Stanford in 2014, he went directly to become a faculty member at Harvard. And, you know, I was kind of amazed to see that because I know sometimes it happened, used to happen before, but this is like the, um, very, very rare when somebody was so successful as a graduate student that he was able to immediately go and get a faculty position. And, and he was actually very successful as a faculty, so he very quickly, he went from in a few years he basically already became— 6 years after starting his lab, he became a full professor at Harvard. And all of that is basically because his spectacular research success in several fields, and his sort of research interest revolved around studying of basically signaling by transmembrane proteins. And he's His interest remains to be with the GPCR receptors, and I think today he's going to talk a little bit about— he's going to talk about some of his work on GPCR receptor signaling. And I actually know him from his work on the sigma receptors, because while he was working on GPCR, he basically got to know about sigma receptor field and realized that there's sort of a lot of unanswered questions in sigma receptor. And actually, I know him because in 2016, he published first crystal structure of sigma-1 receptor, which was a very big deal for the field. I mean, the receptor was cloned in 1993, and like people had a lot of sort of speculations about what the structure looked like. And he actually solved the structure in 2016 and solved it in lipid cubic phase format. And that was actually a very, very important advance to the field. Because it was very controversial what the structure is. And he also did another— made another contribution by also determining molecular identity of sigma-2 receptor, which is another sort of mystery in the field about, you know, what the sigma-2 receptor is. And there were several candidate proteins, and there's actually been papers published about claiming that they identify sigma-2 receptor. And actually, Andrew finally sort of resolved that controversy by showing that this protein, TBM97, is a sigma-2 receptor. So, from my perspective, you know, in the sigma receptor field at least, he made two very, very significant major advances in his career already. So, the first crystal structure of sigma-1 receptor and determined molecular identity of sigma-2 receptor. And he just told me that he has a paper actually in revision where he solves also crystal structure sigma-2 receptor. So it's a very, very spectacular contribution, but unfortunately he's not going to talk about sigma receptors today. I hope he will talk about that today, but he'll talk about GPCRs and technology development. But Paul told me that we have more opening for seminar series, so we probably will ask Andrew to come back again and talk about sigma receptors. All right, all yours.

Okay, great. So can you hear me and see me and everything? Okay. So yeah, it's really a pleasure to be here. So thank you so much for the introduction and for the opportunity to tell you about some of our work. You know, as mentioned, I'm not going to talk about sigma receptors today. We've been doing a lot of work on sigma receptors, and I think there's some exciting new work there. If you're interested, we have a preprint on bioRxiv on some of the most recent work that we've been doing there, and a paper hopefully will be published in a journal soon. On that. But today I'm going to focus on some other work that we've been doing, also part of a long-running project where we've been interested in understanding the structural and molecular basis for phenomenon called biased agonism in G-protein-coupled receptor signaling, which is, I think, potentially therapeutically relevant and also really is sort of a mystery from basic science standpoint as to how these receptors can signal in very different ways in response to ligands, even though they are, you know, just a single receptor. And a lot of this work really has led to— has necessitated the development of new tools in the form of antibody fragments and antibody fragment discovery methods that we've applied to GPCRs. And so near the end of the talk, I'll show some more recent work we've done in a collaborative project with the group at UC Irvine using new methods to develop better antibody discovery platforms for in vitro antibody discovery. So just to introduce everyone, I think this is probably review for many people, but G-protein-coupled receptors, GPCRs, are the single largest family of transmembrane proteins in humans, and there are about 800 different members. And with few exceptions, all of these receptors signal through essentially the same mechanism. They reside in the plasma membrane of the cell, they bind to an activating ligand which is called an agonist, and this leads to an allosteric conformational change on the intracellular face of the receptor that essentially turns it into a guanine-nucleotide exchange factor for heterotrimeric G proteins. So it catalyzes exchange of GDP for GTP that leads to structural rearrangements in the G protein, which then causes it to activate downstream effector proteins like adenylate cyclase to elevate cyclic AMP. GPCRs are also regulated by phosphorylation by GPCR kinases or GRKs, which leads to recruitment of beta-arrestin which then results in receptor internalization and either degradation or non-canonical signaling, followed by receptor recycling to the plasma membrane. And so what I'm going to focus on today in the GPCR work is really this phenomenon called biased agonism, in which a receptor can preferentially signal either through the G protein pathway or the arrestin pathway, and this preferential signaling is dictated by the structure, the chemical structure, of the ligand that's bound to the receptor. So I'm going to focus everything today on the angiotensin II type 1 receptor, AT1R. This is a very important receptor both biologically and as a model system for GPCR signaling. The angiotensin receptor is a critical regulator of sodium retention and of blood pressure, and as a result of that, it's become a very important drug target. Something like 5% of the adult population in the United States takes drugs targeting this receptor every day. And a similar number of people, maybe even a large number of people, take drugs targeting this pathway but a bit upstream of the receptor. So it's an extremely important therapeutic drug target for treating blood pressure, but it's also an important model system for understanding this phenomenon of biased agonism. And I'll explain why, and why this receptor is, I think, a uniquely excellent target for looking at this. So this phenomenon of biased signaling or biased agonism occurs when a receptor in response to a certain ligand preferentially signals through one pathway relative to another. So I'm showing data here from Jonathan Bailyn, who's then at Trevina Pharmaceuticals, and basically what they're doing in this experiment is looking at the effect of different ligands through two different assays. They're measuring activation of the beta-arrestin pathway in black and measuring activation of the G protein pathway in red through this IP1 assay. This is a Gq pathway assay. So for conventional agonist in this receptor, the dose-response for signaling basically results in a normal saturable effect, and you can see that this happens for both ligands— for both pathways, sorry— with the native hormone angiotensin II. The curve for beta-arrestin is right-shifted by about 10-fold, and this is common just because this is a less highly amplified readout. For an antagonist or blocking ligand, nothing happens. You can dose it in, there's no effect on G protein signaling, no effect on arrestin. But then for this beta-arrestin biased agonist, TRV-1201, 0027, you can see a very strong response to the arrestin pathway, about 80 to 90% of maximum achievable by angiotensin II, while there's essentially nothing happening on the G protein pathway. So really, this compound, which is— you'll see in a moment— structurally very similar to angiotensin II, has no ability to activate the G protein pathway while it's nearly fully efficacious on the arrestin pathway. And the big question is how this can occur at a molecular level. Part of the reason we care about this is that not only is it interesting from a basic science standpoint, but it's also potentially therapeutically very important. And so this is data again from the same paper from Trevina Pharmaceuticals where they show that if you dose progressively higher levels of either a conventional antagonist telmisartan or of their arrestin-biased drug, you see a decrease in blood pressure. These are experiments done in mice, and if you do the same experiment where you dose these drugs but instead look at a measure of cardiac performance called ESPVR, which is a measure of the pressure that the heart can generate at a given volume, they see that the two curves diverge, and this TRV compound actually results in improved ESPVR as a function of dose, while the conventional antagonist results in a decrease. So this is really exciting because it says that potentially you can imagine targeting a GPCR with a biased agonist and achieving a subset of its effects that might be therapeutically desirable while avoiding undesirable side effects. And in the case of the angiotensin receptor, one of the reasons that this is a really good model is that the ligands for this receptor that have very strong biased profiles are all structural analogs of each other. They're all derivatives of the native hormone angiotensin II, which is an 8-residue peptide. I'm showing the sequence here from N to C. And you can see that all of these different ligands here, including these orange compounds that are very strongly beta-restin biased, show the same basic organization. They typically start with either aspartate or methylated glycine, and then have the same sequence except at position 8, which is the major determinant of signaling profile. There's some differences in the middle, but really this position 8 is the key determinant of the functional properties of these peptides. And you can see that in the case of angiotensin II, there's phenylalanine in this position, and in general, any derivative of this peptide with a relatively large substituent in position 8 results in a peptide that will signal strongly through the G protein pathway and usually also through the arrestin pathway, while smaller substituents result in attenuated G protein signaling while retaining beta-arrestin activation. So isoleucine in this position results in a modest attenuation, complete deletion or replacement with alanine or dealanine results in a peptide that is essentially entirely arrestin biased. And the one at the very bottom here, TRBO27, is the one that I showed you in the signaling assays just a moment ago. So small amino acid deletion in position 8 results in selective signaling through the beta-arrestin pathway. So we wanted to know what's going on at a structural level, and to do this we turned to a technique that was originally developed by my former advisor and his collaborator, Jan Steyrt. And so basically the idea is that when a receptor is bound to an activating ligand, it becomes highly dynamic. The receptor visits multiple conformations with transmembrane helix 6 existing either in an inward conformation or in an outward conformation or various intermediate states. And so to address this and make this receptor tractable for structural study, we use a single domain antibody fragment called a nanobody to to trap the receptor in a defined state that we can characterize pharmacologically. We can determine functionally what the state is doing, and then we can also, by virtue of having a structurally homogeneous sample, we can characterize it through X-ray crystallography, cryo-EM, or other methods. So these nanobodies, like I said, are single-domain antibody fragments, and they're derived from llama, camel, or alpaca antibodies. So conventional antibody, like those that you find in humans and pretty much all other mammals, has two chains: heavy chain and a light chain. And at the tip of this dimeric structure is a VH domain and a VL domain, a variable heavy and a variable light domain, and these collectively determine antigen binding specificity. But in llamas, camels, and alpacas, there's a secondary antibody repertoire in addition to conventional antibody repertoire that consists only of the heavy chain. They have a deletion of this domain here, the CH1 domain. And this eliminates the light chain entirely and means that the entire antigen binding specificity of this molecule is contained within a single VH domain, which is called VHH in this case, and more commonly is called the nanobody. So this can be expressed in isolation as a single 15-kilodalton protein that's structurally rigid and contains the full functional properties of a conventional antibody fragment. And this nanobody, it has its activity determined by 3 different loops called complementarity-determining regions, or CDRs, numbered 1, 2, and 3. And CDR3 is generally the major determinant of antigen binding specificity and is the largest and most diverse in terms of sequence. So these nanobodies have been really powerful tools for cell biology and for investigating GPCR structure and function. But in the case of AT1R, our collaborators in the Leifquist lab had been injecting llamas with purified angiotensin receptor for over 5 years without obtaining useful binders. Useful in this case meaning conformation-specific clones that can stabilize this desired active state that biased ligands recognize. So we needed to find another approach, and so instead we turned to a technique called yeast surface display. It was originally developed by Dane Wittrup and colleagues in the late 1990s and has since become a a really powerful technique for combinatorial biology. And this is a method to allow sampling of very large antibody fragment libraries up to sizes of about 10^10 unique clones per library. One of the really key advantages of this approach compared to techniques like phage display or ribosome display or mRNA display is that yeast display is a cell-based method, which means that we can use fluorescence-activated cell sorting, or FACS, to specifically isolate cells that express antibodies with a desired set of functional properties. So we can design complicated selection experiments to enrich clones that bind to one conformation but not another, or to select for clones that bind tightly relative to their level of surface expression, or that have very high selectivity, or that have cross-reactivity between two targets. Essentially any sort of functional property you can imagine encoding in a fluorescent signal with a cell, you can sort for it and you can identify clones that possess that property. And finally, just as a practical matter, yeast cells grow relatively quickly, and so we can do these selections start to finish in a relatively short period of time. So typically, from beginning a selection to identifying clones for characterization, if you do 4 rounds of selection, you can do this in 2 to 3 weeks. So the basic idea of this approach is shown here. We start with our library of synthetic nanobodies displayed on yeast, and each cell displays about 500,000 molecules of a single nanobody clone on its surface. Each cell displays a different antibody, and we can incubate these then with a purified target protein or receptor of interest labeled with some sort of detectable marker. Most often this is either a covalently labeled fluorophore, or it could be streptavidin— sorry, it could be a biotin, which we can then detect with fluorescent streptavidin, or it could be an epitope tag that we detect with a fluorescent antibody. And then we sort those cells that bind from those that don't using either magnetic cell sorting or FACS. Amplify them, they grow up in culture, and then we can repeat the process as necessary until we have a pool of clones with the desired functional activity. And so in the case of doing this with the GPCR, like angiotensin receptor, we have the receptor embedded in a detergent micelle labeled with a fluorophore either directly or indirectly, and then bound to the nanobody on the cell surface. And so this allows the library to access both the intracellular face of the receptor and the extracellular face of the receptor, so we can make nanobody clones that have very different activities depending on which part of the receptor we're targeting. So we designed a library based on the collection of all antibody structures that have been published in the PDB. The idea here is that any antibody that had been crystallized probably was biochemically well-behaved, well-folded, and stable, and so we use that as the basis for designing our synthetic library and choosing what types of diversity to include in each of the three complementarity-determining regions. The sequence composition ends up actually being very similar to that in the immune repertoire, so there's not really a big difference between modeling off of that versus modeling on the PDB. But in any case, we use the PDB as the basis for our design. And so you can see in some of these positions, like this position here, we allow relatively limited diversity, 4 different amino acids. Here we allow phenylalanine and serine. And then in these asterisk positions that are highly diverse, in the Immune Repertoire and in the PDB, we use a mix of different codons. In this case, we used pre-synthesized trimeric nucleotides, trimerophosphoramidites. So essentially, we're using a mixture of individual codons that have been pre-synthesized so that we can achieve exactly the composition of amino acids that we want. So 14% tyrosine, 12% glycine, 0% cysteine, 0% methionine, 0% stop. So using these pre-made single codons and then mixing them in a defined ratio allows us to achieve this type of design. You can see that CDR3 is the most complex, most of these hyperdiversified positions. And then we also introduced length diversity here with this diverse segment varying from 7 to 15 residues in length. So we have 3 different lengths and we overrepresent the 11-residue length because that's closest to the mean length that's found in the immune repertoire. So in the case of the angiotensin receptor, we designed a selection procedure first to enrich for binders. That's this first part shown in orange in this schematic, but then really the key experiment is doing a selection shown here where we take receptor that's labeled with a red fluorophore and bound to an agonist, and a separate population of receptor labeled with blue fluorophore bound to an antagonist, and we can mix all of this together with the cells in a single pot and take advantage of the fact that in this case we have antagonists with very slow off-rate, and so there's minimal exchange of these two ligands on the time scale of the experiment, and that means that we can specifically enrich for those yeast cells whose antibodies selectively bind the activated conformation of the receptor. And because we have ligands present at high concentration, we can direct the selection to the intracellular face of the receptor. So this is a really nice way of doing a selection where you specifically direct the library to give you clones with a very precisely defined functional activity: binding to the intracellular face, recognizing conformation that's compatible with activating ligand or agonist binding. So we ended up doing this. We selected a bunch of clones, screened them, and I'm just going to focus now on the one that ended up being the most useful. This is is a clone called nanobody AT110. And first of all, you can see that this one preferentially recognizes the receptor in an active state or in an agonist-bound state. And this is an ELISA-type assay looking at either Apo receptor binding, antagonist receptor binding— Losartan's a classic blocker— agonist-bound receptor, angiotensin II, the native agonist, or receptor-bound to S1I8, which is a synthetic agonist. It's a derivative of angiotensin II that's somewhat higher affinity. So you can see it binds preferentially to the receptor when it's incubated in the presence of agonist, and we can also do a more quantitative experiment looking at a radioligand competition binding assay. And in this case, we're using a radioactive probe, tritiated olmesartan, and we're measuring at increasing concentrations of angiotensin II the displacement of this ligand from the receptor. So in black is the control curve, where you can see that if we add in more and more angiotensin II, we compete off olmesartan, and this IC50 value gives a proxy that we can use to calculate the affinity of angiotensin II for this receptor. And since the nanobody stabilizes angiotensin II binding, since it's recognizing the active state, we see a left shift in the curve when the nanobody is present at a fixed saturating concentration. So this means that the receptor is being stabilized in an active state, in the same state that the native hormone angiotensin II prefers. So unfortunately, this initial clone was relatively low affinity. It's about 100 nanomolar affinity, and so that was reasonably good, but it's not good enough to obtain a crystal structure. We tried and didn't work, so instead we turned to affinity maturation, and in this case we just used error-prone PCR to mutate this clone and diversify its sequence. And you can see here a staining experiment with parent clone, the original isolated nanobody AT110, and then in the Error-Prone Library— I should explain, sorry, that angiotensin binding is on the x-axis, expression is on the y-axis, and each of these dots in the flow cytometry plot is a single cell, single yeast cell. So we see this plot where it goes to the upper right, saying basically that the more nanobodies expressed, the more receptor binding you see. And then in the error-prone library, we see a relative attenuation of this. There's only 4.8% of the cells that are in the strong binding gate, just reflecting the fact that on average, if you mutate something, you're more likely to make it worse than to make it better. But some of those clones are better, even though they're relatively few. And so we can do a FACS sort for the cells that are in this far right quadrant, and after enriching several rounds, we obtain a population of binders that are much stronger than the original parent clone. And then we can sequence and identify mutations that are enhancing affinity here. So we focused on a number of different positions, but I'm just going to show 4 that ended up having the biggest effect. These are shown here: position 31, 58, 98, and 113. And in each of these positions, we saw a very strong convergence to either a single substitution or to 2 different possibilities. So in the case of I98V, for instance, which is the strongest affinity-enhancing mutation, We find that position 98 uniformly in all of the high affinity variants is mutated to V. Tyrosine 113 is always mutated to asparagine in all 4 of the high affinity clones that we found with any substitution in this position. And so we took these 4 hotspot residues that showed very clear preference for alternative amino acids and we built a consensus clone combining all of them. We also built several other consensus clones with subsets of these, but the consensus with all 4 of these was ultimately the highest affinity. So we tested it using the same assay I just showed you a moment ago. So we take again a concentration, fixed concentration of a radiolabeled probe compound, treated olmesartan, and then we compete with an agonist. In this case, we're using a different agonist here, V05, for technical reasons. But basically you can see that as you dose in higher and higher concentrations of this agonist, you compete off the radioactive probe that's the curve in black. And if we do the same experiment now in the presence of this non-affinity matured clone, uh, nanobody AT110 in red, we see that the curve shifts a little bit at a low concentration and it shifts a lot at a higher concentration. So at 40 nanomolar, the shift is really quite modest, indicating that we're probably well below the Kd for this parent nanobody clone. We know now that that's the case. If we look in blue at the affinity matured version, which is AT110I1, I1 just meaning improved variant 1. We can see that at the high concentration, we have an even bigger shift, but even at 40 nanomolar in dotted lines, we have a very large shift, essentially no reduction in allosteric effect, which is indicating that at 40 nanomolar, we're still comfortably over Kd for this antibody. So now we have something that's significantly improved in affinity. And once we had this molecule and set trays for the first time, we got crystals almost immediately. It was relatively quick to get a structure of this. And we were able to see, for the first time, how the angiotensin receptor recognizes peptide ligands. So I'm showing here essentially a cross-section through the receptor with peptide shown in the active state and then just modeled into the inactive state. This is a previous structure from Ray Stevens' group showing what this would look like if the peptide were to bind the inactive conformation of the receptor. You can see that the binding site really constricts quite dramatically in the active state. The receptor almost wraps itself around this peptide, really enclosing it entirely. And if we were to look now at how these two states would look with a small molecule antagonist bound, you can see that in the inactive state there's plenty of space, it fits in nicely, but in the active state this contraction of the binding pocket would be incompatible with this drug. So it would have a clash here at this red arrow. So angiotensin receptor blockers are too bulky to accommodate the constriction, which helps to explain how they prevent receptor activation. They also of course prevent ligand binding, they compete with peptides, but they prevent receptor activation even at baseline. We were also able to explain the details of how the peptide interacts with the receptor. I'm not going to go into too much detail here except to say that we can provide a structural explanation now for some of the mutagenesis data that have been developed for this receptor over many decades. In particular, things like lysine 199, which was known to be absolutely essential for peptide recognition. We can see it's making a salt bridge with the peptide C-terminus, and so we can provide a molecular explanation for that effect. So really, in general, I would say very close agreement with prior mutagenesis experiments explaining why these mutations that affect peptide function have their effects. And I'll mention also that the peptide C-terminus sits really deeply buried within the receptor core. So the N-terminus is out here on the extracellular face, C-terminus is buried in the receptor binding site, and this is really where the key determinants for receptor signaling properties are going to be situated. So we were able to describe the activation mechanism in some detail. Again, I'm not going to go through all of this except to say that We can see how basically a series of conformational rearrangements propagate through the receptor, leading to a very classic activated state on the intracellular face with a wide open pocket ready to engage the G protein. So as I said earlier, the C-terminal residue in the peptide is really the key determinant of its functional properties. So large residues here like phenylalanine result in ligands that activate the G protein pathway and the arrestin pathway, while smaller substitutions give you peptides that preferentially signal through beta-arrestin. And so of course, once we had this structure with our peptide analog S1I8, which is sort of an intermediate ligand— pharmacologically it's very high affinity— we were able to dock in phenylalanine in this position to see what would happen. And somewhat surprisingly, we see that you can't actually fit phenylalanine there. It clashes with the receptor. In any of its three rotamers. There's no way you can model this in without forcing a change in the receptor conformation. And so we thought this might explain why angiotensin II can activate Gq very robustly while none of these other peptides can. So we wanted to look at this in more detail, and so of course the first question is, does our antibody bind the receptor when it's occupied with these other compounds, either the balanced agonist angiotensin II, or these arrest and bias ligands. So again, using this ELISA-type assay, this AlphaScreen assay, we're able to see that there's strong binding to angiotensin II-occupied receptor, and also reasonably strong binding to receptor-bound TRVO2-6 and TRVO2-3. This is also true for TRVO2-7, although I don't have data here. So our initial structure was with the receptor bound to this S1I8 peptide, And we were subsequently able to get structures with 3 other compounds, native hormone angiotensin II, and then 2 biased ligands, TRV-026 and TRV-023. And I'll just say that 026 and 023 look almost identical, and so I'm really just going to focus on, on one of these. So first thing you can see superimposing these structures is that at the level of receptor backbone, there's really not much of a difference between these various states. They all look very similar.. But if we start looking more closely in the ligand binding pocket, we can see they really do diverge. So here I'm showing the two different peptides that we're looking at, TRBO2-6 biased agonist in blue and angiotensin II, the balanced agonist, in orange. And you can see that this balanced agonist here has this lysine-199 interaction, phenylalanine-8 sits here in the base of the binding pocket, while this is essentially an empty cavity in the biased ligand state. At the electron density level, we see relatively poor density actually for phenylalanine 8, seems to be somewhat mobile, while the peptide, the entire peptide for TRV026 is very well ordered. And even more surprisingly, if we look at the residues nearby in the receptor, we see that the conformation of this binding pocket is really quite different. So in the angiotensin II bound case, we see leucine 112 is here pointing to the right, it shifts position with the biased agonist. And even weirder is the fact that tyrosine 292, which is normally very well ordered, very clearly resolved in the biased agonist state, is entirely disordered in the balanced agonist conformation. So we also looked at this by molecular dynamic simulation in collaboration with Ron Dror, two members of his group, And actually, these simulations— first round of these simulations was done before we had a crystal structure. And when they had tried just simulating angiotensin II bound to this receptor, they saw that phenylalanine 8 really was visiting multiple conformations. It didn't have a stable state. And actually, this fits really nicely with what we see in the electron density, where it's essentially blurred around position 8, suggesting it's compatible with this model. They also found that in their simulations, this phenylalanine 8 would adopt either a horizontal or a vertical rotamer, or this inserted rotamer, and in particular, the horizontal conformation was strongly associated with G protein biased agonists and with the G protein preferring conformation of the receptor, suggesting that the signaling output for this peptide is dependent not only on the identity of the residue here, but also on the conformation of this side chain. So based on that, we were able to design a compound called 2-aminoindane angiotensin II, and this substituted version has basically a bridging methylene that links the phenyl ring to the rest of the backbone so that it can no longer rotate. So you're basically forcing it to be in a vertical conformation that was more associated with arrest in signaling. And indeed, that's the case. If you test this peptide in a signaling assay, you get relatively little activation of the Gq pathway. It's in orange compared to to the control in black, and in an arrestin assay we see relatively robust activation, about 50 to 60% of angiotensin II, despite not having much activity at all on the G protein pathway. And this is exciting because it's, I think, one of the first, if not the first, example of prospectively designing a ligand with the desired biased signaling profile based purely on structural data, and actually in this case this compound essentially runs contrary to the SAR, the structure-activity relationship you see just looking at these peptide ligands. So it's actually more bulky in position 8 than angiotensin II even, and yet it signals preferentially through the beta-restin pathway. So I think it shows that you can use structural data to prospectively design molecules with the desired functional property in terms of its biased signaling activity. Okay, so I'm going to shift gears real quick and just mention sort of as a preemptive answer to the question that everyone always asks, which is, you know, if you're if you can make intracellular antibodies, can you make extracellular antibodies? And I'll just very briefly say that the answer, of course, is yes, you can make things that bind to the extracellular face of the receptor. And so one of these is called antibody AT118 that we developed targeting the angiotensin receptor. This one is a high-affinity competitive antagonist of the receptor. So in a competition binding assay, dosing in the antibody, we can see that it displaces the radioligand probe. Affinity is about 60 nanomolar in this case. And then we can also look at the effect in a signaling assay. In this case, it's a beta-restin recruitment assay, and we add in the native ligand to angiotensin II. We get a nice signaling response in black. If we do a negative control antibody that binds to the intracellular face— this is the same antibody I showed you the structural data for before— has no effect. But then if we add in this competitive antagonist AD118, we see a right shift in the curve, and we see also a decrease in baseline signaling, suggesting that it's an inverse agonist that's really lowering the basal activity of the receptor. And not only is it active in cells, but we can actually assay this in vivo. And this is a collaboration with Howard Rockman and his postdoc, and so in this experiment they're taking mice and dosing progressively higher amounts of angiotensin II, measuring blood pressure. So in black is the control curve. As they dose more and more angiotensin II, the mice develop very severe hypertension. But if the mice are pre-treated with the drug losartan— this is a conventional clinically approved angiotensin blocker— they see that the effect is blunted. They do the same experiment now with our nanobody in red. You see that the curve is shifted to the right, so the nanobody is really preventing this angiotensin II-induced hypertension, while a negative control nanobody that binds to the intracellular face of the receptor has no effect. So it shows that we can make antibodies now that target receptors not only from the inside to stabilize defined conformations, but also from the outside to control signaling properties. So I'll just briefly summarize this part and then go on to some newer technology development work that we've been doing. So I showed you that we can use our synthetic nanobody library to rapidly isolate conformation-stabilizing antibodies for GPCRs. When we had these molecules, the crystallography became very straightforward. We were able to get structures of the receptor bound to 4 different ligands, in relatively short order. And in the case of angiotensin II, the conformation of this phenylalanine 8 really seems to be a key driver of the conformational changes that selectively activate G protein signaling over beta-restin signaling. And I didn't go into all the details, but we also were able to see further changes in the receptor downstream that helped to explain why these differences in the binding pocket itself are coupled to changes on the intracellular face that ultimately result in a conformation that's more efficient for coupling to G protein, or that can only really couple to beta-restin. Okay, so with that, I want to shift gears to something that's more pure technology development, but before I do that, I just want to set the stage by describing two basic paradigms for antibody purification.

Andrew, do you want to take questions on part one now, or you want to wait till the end?

I can do either, whatever people prefer.

So whatever So if somebody wants to ask questions, either type it in the chat box or you can raise— okay, Alec has a question. So just raise the hand and talk like Alec does now.

Alec?

Alec, you're unmuted.

Yeah, this is Alec. I have a question about the active structure. So Do you think, you know, the antibody stabilized receptor structure is exactly the same as the native agonist bound structure at all? For example, if you have two different antibodies, do you— do they need to the same structure at all?

Yes, I mean, I think that's kind of the standard question, right, is, you know, obviously, we're using a tool that has the potential to perturb receptor conformation. And so, you know, I would say a couple of things. First is that in the case in the case of something like the beta-2 receptor, where we have a structure with an antibody followed later by a structure with the heterotrimeric G protein, they're really very, very similar. They're not exactly identical, there are some subtle ways in which they differ, but the vast majority of the structural features are essentially the same between the two states. So I think that provides some confidence in this basic approach. But the other thing is that we validate the activities of the antibodies pretty extensively by radioligand informatics, so we know that from the perspective of the ligand itself, and in terms of the energetics of ligand binding, the shifts in ligand affinities are essentially identical between G protein stabilization and nanobody stabilization. So that gives us confidence that we're not stabilizing some sort of artifactual state. It really is very similar to the G protein state. And then finally, of course, you can do mutagenesis experiments and so on to validate. To the question of whether two different antibodies would induce the same state, we have multiple antibodies. They tend to show very similar pharmacological properties. And again, in the case of the beta-2 receptor work from Brian Kobilka's group, my former advisor, has shown that multiple structures with the same receptor with different antibodies again are very similar, especially in the ligand binding site. As you get closer to the intracellular face of the receptor, there's probably more nanobody influence on the states there, and so that's part of the reason I didn't talk so much about that region. But one of the things that we can do there is, it's possible to run molecular dynamic simulations and remove the nanobody and let the system relax. And I think that provides a better indication of what might be happening there, but really in the longer term I think we need structures with the native transducer, with G protein and with beta-restin, to be absolutely confident about what's happening in that region.

Thank you.

Don Hilleman. Yeah, just a typical egotistical, egocentric question. So, you know, we have about 10 proteins we would like to be able to inhibit, and we have absolutely no inhibitors for them. A couple we have, you know, the crystal structures are available, and some just speculative structures. So, uh, is this the way— can you sell us— can you tell me this is the way we should go and, uh, we'll If we start working on antibodies, maybe we'll have something in 1 year, 2 years, or 10 years.

What do you think? Yeah, I mean, I can speak to our experience. I mean, it's been, I would say, relatively straightforward for us to make binders for most proteins that we've tried. The, the nanobody library itself is available. We, you know, we provide it under just like a standard MTA, or we have a commercial distributor called Caraphast, which you can just purchase it from them, and they send it out to you. So, you know, certainly you can get the library, you can run selections and so forth yourself, be comfortable doing yeast display. It's not that hard to do, although it does require access to Effex sorting machine to really be able to do it effectively. I would say that depending on the protein of interest, the success rate, you know, might be higher or lower. It's worked really well for us for GPCRs. In general, antibodies have a propensity for binding to cavities, so proteins that that don't present concave epitopes are probably more challenging. The other trend we've seen is that there's a very clear difference in performance of the library with larger proteins, which perform better compared to smaller proteins. And I think that's— I don't know for sure, but I speculate that it's just purely because there are more epitopes in a larger protein with a bigger surface area. So there may not be anything more to it than that. So I think it depends on the details of your protein. The other thing I would add is that there are very good immunization-based approaches for nanobody generation. That you can do on a contract basis or collaborating with someone who has a llama farm. For things like intracellular proteins or proteins that are not, you know, highly homologous to llama, those approaches can work really well. Angiotensin receptor is both biochemically unstable and highly homologous to llama receptor, and so it ends up being a really relatively poor immunization because I think it just denatures as soon as it goes into the animal.

Thanks. Ilya, this is Paul.

Can I ask a question quickly?

Sure, of course. So admittedly, I haven't kept up with the literature, but I remember several years ago, the angiotensin receptor, some people were postulating that it could directly sense tension in the membrane, and its activation was different when it did so. It actually, I think if I remember correctly, activated a different G protein. Where does that literature stand, and is it worthwhile looking to see if, you know, using nanoparticles or something like that in activating, you know, subpanel with lysolipids or something like that, if it would—

if this could be done? Yeah, I'm not so familiar with that for angiotensin receptor. There's work on the— I think on the apelin receptor, APJ, which is a relatively close homolog of the angiotensin receptor. Claiming that there's a difference in biased signaling, essentially in relative signaling through the arrestin pathway versus G protein pathway depending on mechanical force. I have not seen much follow-up on that. There's also work on another receptor on GPR68 suggesting a similar sort of story where mechanical flow seems to activate the receptor in sort of a different way. You know, one of the challenges I think is that in our typical signaling assays we like to to do dose-response of a compound that we can dilute out. And I worry that any sort of mechanical stimulation is complicated by the fact that many different things are changing in the cell under those conditions. So it's something that we just haven't explored in this case. But in principle, I suppose it's possible.

Yeah, OK. Thank you. Yeah, I imagine it would be pretty tough.

Yeah, thanks.

All right, I think we need to go to the second part.

OK. So yeah, the second part is shorter, so I'll be efficient with everyone's time. So just to set the stage, I want to mention two basic paradigms for antibody discovery. The first is animal immunization methods. This is classically done through injecting mice and ultimately making hybridomas, but there are a lot of newer methods relying on single B cell sequencing and other methods. And the other major paradigm is synthetic approaches using combinatorial biology. So that's like the yeast display approach I showed a moment ago, also phage display, mRNA display, and so on. And each of these approaches has advantages and disadvantages. Immunization approaches are really great because they generate high affinity binders, they have access to somatic hypermutation and the full power of the immune system, and they're relatively accessible. You can do a lot of this on a contract basis with a CRO service provider. But they're limited by things like tolerance of self. So epitopes that are found in the animal being immunized often are not very, very good. You can get immunodominance where you get a preferential immune response directed at a particular epitope. This is common for GPCRs to get a strong response to the intracellular face of the receptor, which makes it hard to make extracellular activators or inhibitors. And they're incompatible— this method is incompatible with biochemically unstable antigens because the antigen has to stay folded any animal for some period of time if you want to find confirmation-specific binders, and many GPCRs unfortunately fall into this category. Synthetic approaches, in contrast, have really the opportunity for very precise control over the selection process. You can run the selection in any buffer, any salt concentration, pH you like, within reason. They're relatively inexpensive and scalable compared to animal immunization approaches, but they're limited by the fact that the diversity of the library is fixed at the outset And oftentimes the clones that come out of these selections need further optimization after discovery. So you saw that with the angiotensin receptor where we developed a functionally very interesting clone, ET110, but the affinity wasn't high enough for crystallography to start, and so we had to do affinity maturation to improve it. So really we want to achieve the best of both approaches, so having something that allows a high degree of molecular control, fast turnaround, but also has the power of a somatic hypermutation-like process. And so to do that, we turned to a technique called OrthoRep that was developed by Cheng Gao's group at UC Irvine, and I'll explain a little bit of background of how this technique works and then how we were able to implement it with a yeast display system together with Cheng's group. So basically, the idea here is that certain yeast have selfish genetic elements that replicate as linear plasmids in the cytoplasm. They come as pairs of plasmids called P1 and P2, P1 can be modified to introduce a gene of interest, and then P2 contains a bunch of housekeeping machinery like an mRNA capping enzyme and so forth to process the RNAs that come off of this. And the key point here is that this P1 plasmid is replicated by its own dedicated DNA polymerase, which can be modified to have a very high error rate. And so because these plasmids use their own dedicated DNA polymerases, you can introduce a high error rate polymerase that selectively mutates the gene of interest while sparing the host genome. And as I said, this is work from Cheng Liu's group in developing the system in the first place and moving it into Saccharomyces cerevisiae for enzyme evolution. And so we teamed up with his group to develop a system to do this for antibody display. So the basic idea of the system is shown here. We have a yeast cell displaying your antibody of interest And then we have this orthogonal DNA polymerase encoded in a nuclear plasmid, conventional plasmid, and then this engineered DNA polymerase selectively mutates and replicates and mutates this p1 plasmid, which has the protein of interest, the antibody fragment encoded with an epitope tag, and then a cell wall anchoring sequence, as well as a selectable marker, trip1, in this case. So we can encode our antibody fragments here and use this error-prone polymerase to diversify the antibody fragments with passaging. So essentially, as you grow the cells, the antibodies are continuously diversifying at a low rate, really mimicking the process of somatic hypermutation in conventional immune antibody maturation. And I should mention that this took quite a little— quite a lot of work to actually make this useful in practice. The big barrier was that the display levels of the antibody in the original system were very low, And so we were able to use this system to diversify the system itself. So actually changing things like the promoter here to improve expression was really important. So the system then that comes out of this is shown here. Basically, we have this yeast cell with these properties. We can take a library of antibody fragments or a single antibody fragment, and as we passage the cells, they're continuously diversifying. Incubate them with an antigen of interest, sort them by FACS or magnetic cell sorting, and then passage. So it's the same basic paradigm I showed you before, but the key difference is that at each round of selection, diversity is being reintroduced, and so we can keep iteratively improving the affinity of any given clone. So as a pilot, we started with AT110, the antibody I showed you at the beginning, and if we look at the sequence of this antibody from N to C, in each round and look at the frequency of mutation on the y-axis, we can see that as we evolve it, we observe mutations that first arise, like this I98V mutation, the same mutation we found by conventional error-prone mutagenesis. This mutation first arises at low frequency, rapidly goes to fixation, and then that serves as the background for further mutations like this R66H, which then goes to fixation.. And these mutations iteratively arise and fix, leading to improved variance over time. So we see iterative emergence and fixation of single mutations, which is in part reflective of the still relatively low mutation rate of this error-prone polymerase. So as we go through the selection process, I'm just showing a proxy for affinity here that's log IC50 in the allosteric shift assay. AT110 is okay at about 50 nanomolar. It's improved by this I98V mutation. It's improved further with R66H and finally with Y113H. And one interesting aspect of this is that the later mutations, R66H and Y113H, actually didn't do that much on their own. They have relatively modest improvements in affinity in this assay. So to some degree, it does seem that having this I98V as a starting mutation might help help these other mutations to arise later. So as a model system to look at this in more detail, we looked at SARS-CoV-2 targeting antibodies. So we made 8 different antibodies using our conventional system targeting the SARS-CoV-2 RBD, and then we performed this process to do iterative improvement of affinity. I'm just showing the FACS sorting here for one of these clones where we start with 40 nanomolar of the target. RBD is the receptor binding domain that's required for interacting with the host ACE2 protein, and we can drop the concentration, and yet we still see a progressive increase in the level of binding. So these clones are continuously evolving to become more and more effective binders. If we look at the mutations that occur through this evolution process across all of these different clones, we see that they're generally in the CDR loops or near the CDR loops. We found sort of a hotspot mutation here where the same mutation occurs repeatedly, which probably suggests that something about the library design itself could be improved. And as you'd expect, as these mutations improve the affinity, when we measure them in an SPR assay, we can see that the parent clones show relatively modest binding. We actually deliberately selected relatively low affinity clones to start this process, so this clone called RBD10 has about 400 nanomolar affinity, and after evolution it ends up being sub-nanomolar PD.. And this correlates with improved neutralization potency in a pseudovirus assay. So the parent clone is in dark purple, and it improves through the selection round to become much more effective at neutralizing the pseudovirus. And so we did this across a number of different clones, 8 in total, showing data here from 5 of them that we characterized in more detail. And on the top of each of these little slices here is showing the affinity So just biochemical binding affinity as it changes by round. And so we start here with about a 1 micromolar affinity and improve down to somewhere between 10 and 100 nanomolar. And then on the bottom of each slice is a measurement of neutralization potency, and these axes are matched. So the neutralization potency is generally better than the true biochemical binding affinity.. And we see that in some cases these correlate very strongly with each other. In other cases, we see like in clone 11, the neutralization potency really improves more quickly than biochemical affinity. So since this literally just involves culturing the cells in parallel and then occasionally running a sort, we can evolve many clones simultaneously with relatively little investment of labor. Oftentimes the best clones at the end weren't the best clones at the beginning, and so if we could only do one or a few clones, we may not know which ones to start with. So doing this in parallel, I think, is really powerful. And we found clones that were functionally very diverse. And in the interest of time, I'm not going to go into too much detail here. So last question then before wrapping up is just to say that we'd obviously like to be able to start this with a naive library, so a library that's not just a collection of clones that already have some binding activity, but really a completely synthetic library with no predefined binding activity to any target. The challenge there is that this orthorep system can only accommodate relatively small libraries, and so we need to think about how to very efficiently sample the limited sequence space that we have available, about 10^6 or 10^7 clones. So with the conventional library where we randomly diversify CDR3s, you can sort of imagine a situation like this where the space of possible sequences— if you imagine it as this rectangle is populated by random dots that we actually sample. So the total possible sequence space is enormous. It's many, many billions of fold larger than what we can actually sample, but we can test individual points randomly scattered throughout this. The downside of this is that some clones end up being relatively similar to each other just by chance. They're nearby in sequence space, and so in some sense, this is wasted diversity. We have a big empty space over here that we haven't sampled. So really what we'd like to do is to sample the space in a more uniform way so that each clone is equally dissimilar from the others. And then with the OrthoRep system, we are doing essentially a local sequence search around each of these points in sequence space to try to find any local minima that might be present. So to achieve this, we collaborated with Debbie Marks and her student Jun Shin to develop a machine learning method for training a computational model on a collection of Llama-derived sequences so that we could then generate new sequences that have the same biochemical properties, so the same CDR lengths, the same overall net charge, same distribution in overall hydrophobicity. But we can generate many billions of these sequences computationally and then cluster them, and then from each cluster take a representative example sequence and then use that as the basis for synthesizing a library.. And so you can see if we measure sequence similarity for the natural repertoire in blue compared to our design library in green, on average the pairwise distance between sequences is much larger. So the sequences are less similar from one another than they would be in the immune system, but despite that they retain the same overall properties in terms of charge and CDR3 length and so forth. And so we made a very, very small pilot library. It's a library of about 190,000 clones. Designed in this way, and we were able to find from that one binder for GFP as a model antigen, and then selected it through 6 cycles of this evolution process, and we're able to show that it has a modest improvement in binding affinity. Since this is a model, we haven't really pushed it further, but I think it says that this is feasible at a basic level. And so now we're preparing to go into much larger libraries instead of 190,000 clones, doing something on the order of 10^6 or 10^7 library and probably doing this with several different starting points to find out what the best, most evolvable pool of clones is. So we found binders to GFP and can improve affinity at least modestly. So to wrap up then, I've shown that you can combine yeast display with this OrthoRep evolution system to rapidly optimize many antibodies in parallel. A carefully designed naive library can yield evolvable binders, And I think in the longer term, the hope is that this fully in vitro platform will recapitulate key aspects of the immune system while also having many of the advantages of a synthetic in vitro selection process. And with that, I just want to thank the people involved in the work. The first part of the talk really was led by Meredith Skiva, a postdoc in the lab, in collaboration with Laura Winkler and Dean Staus and Bob Lefkowitz's group. And then all of the work on the orthorep evolution system was done in collaboration with Chang Liu, and Alon Wellner in his lab. And in my group, it was led by Connor McMahon, who's a former postdoc.

And with that, I'd be happy to take any questions. Thank you, Andrew. That was very, very, very interesting. So any questions on the second part, the technology part? I'm sure everybody now thinking how to use it for their system and how realistic. So do you think in terms of sort of application of this, do you think it would be like sort of, it's at the point where a lab without specialized background can sort of run this kind of screen, or it will require some, you know, expertise in like yeast genetics?

It's been varied. I mean, we've distributed the original library now to something like 400 or 500 different groups, and some people have had really a lot of success with it and others have not. I would say in general, the groups that have been most successful are protein biochemists or people who have experience with yeast display. Anyone who has done yeast display successfully before seems to have no trouble with it. It's, you know, you can pretty much plug and play. The biggest technical barrier seems to be getting access to a FACS facility that will allow you to sort yeast. It's kind of a stupid thing, but a lot of facilities are just absolutely adamantly opposed to this. We use a sorter, our sorter is shared with a group that does like mammalian primary cell culture, and that's fine. We've never had any issues with contamination. We flush the lines with bleach after every sort, and that decontaminates it. But for, for many people who've struggled with that, I think that's been one of the major challenges. It's also really important to have good quality antigen. If you don't have biochemically pure, monodispersed, good quality protein, then your selection is sort of doomed from the outset.

Awesome, makes sense. Oh, that's Yeah, I think that's a fantastic tool if people can figure out how to do it easily.

I would add too that we have protocols on my lab website, so if you want, if you're interested in it, you can look and see what's involved. They're overly detailed, I think, by design. We also have a frequently asked questions thing, so hopefully it helps people get a sense of how tractable it is. Anybody?

Everybody's scratching their head and trying to figure out how to get the fox. All right, if there are no any more questions, okay, let's thank Andrew for— it was very interesting talk in terms of technology development and We'll have to talk to him more about sigma receptor once his paper is out.

All right, thank you guys.

Thank you.