In the chat to let us know where you're tuning in from, and it's always great to see the reach of this community. Before we dive into today's presentation, just a quick bit of housekeeping. We'll be holding a Q&A session at the end of the presentation, so if you have a question that comes to mind at any point, please type it into the chat and then I'll be collecting those throughout the session to ensure that we get to as many as possible during the final 10-15 minutes. It is truly an honor and a privilege to introduce our guest speaker, the Director of Machine Learning at iBio, Alex Taguchi. Alex has built his career around solving complex problems in biology with machine learning.
He has over a decade of experience in computational biology and more than 30 publications spanning work at research organizations, including the Massachusetts Institute of Technology. At RubrYc Therapeutics, Alex managed antibody discovery campaigns using Machine Learning to produce drug molecules with exquisite epitope selectivity, helping drive the company's acquisition by iBio in 2022, in which he now leads the development of iBio's machine learning platform for epitope-steered antibody discovery. Alex, we are so grateful to have you here to discuss Generative Antigen Design for GPCR Antibody Discovery . The floor is now yours.
Thank you. It's always a pleasure to work with you guys. You're a tight-knit team, and it's always been fun to interact with people like Rob, too. Generative AI for drug discovery. It's something that pops up on our newsfeed probably at an increasingly exponential rate, making it harder and harder, I feel these days, to parse out what is real and what is just hype. What I wanted to center this talk around is this problem I'm observing, which is the increasing or growing trust gap between the promise of generative AI to solve effectively all of our problems in drug discovery and R&D cost reality. Right? I'm sure many of you have had this thought before, but how can I trust generative AI or AI in general to guide such a time-consuming and expensive process like drug discovery?
I mean, the path to FDA approval encompasses something on the order of more than $0.1 billion and more than 10 years of effort to get there. We're putting an awful lot of faith in these technologies to guide this process. Over the several years that I've been a Director, I've had the honor to work as a Director of Machine Learning at my company, I've been asking these questions, and the answers have been evolving over the years, and this is what I want to provide to you, what my answers to these questions are. Is AI the right tool for my problem? It may be different for mine than yours. Even if I decide I want to use artificial intelligence for my drug discovery platform, which model or models should I be using, and how do I trust the model outputs?
For us to really answer these questions, we've been spending the past 1.5 year , two years, deeply integrating our artificial intelligence platform into an integrated experimental structural and functional validation pipeline. What this does, it allows us to pre-validate the molecules and predictions that are coming off of the AI platform before implementing them into actual drug discovery campaigns, thus de-risking their use. That platform, that integrated platform, is what I'll really be going over today. First, let's back up. Who are we? I work at iBio. We're an antibody drug discovery company focused on delivering high-quality therapeutics, antibody therapeutics, to treat obesity. Our claim to fame is our ability to go fast while still delivering high-quality molecules.
We've been able to, in certain cases, go all the way from antibody discovery to development candidate in as little as seven months, like in this IBIO-600 campaign. It's our lead campaign here, which actually, just I think yesterday or this morning, yeah, this morning we press-released, is now approved to go into clinic. That was from inception all the way to clinic approval in two years. Hot on its heel is our other lead program, IBIO-610, which is a potential first-in-class Activin E antibody, both with the idea of treating different forms of obesity. What is IBIO-600? This one that we really move fast. I like to think about as our Arnold Schwarzenegger drug. Now, you can, from a single injection of 5 mg per kg IV dose in monkeys, we're able to observe a significant fat loss coupled to muscle gain, right?
The power of this drug is, not only is it very efficacious and more or less converting fat to muscle, we're observing a really long half-life in monkeys, which we're predicting will result in something on the order of 74-147 days. You're looking at really low, infrequent dosing with these really nice benefits. Now, maybe you're not interested in looking like Arnold Schwarzenegger. I get it. You just want to cut down the fat and leave it that way. Well, then you might be more interested in IBIO-610.
It's our potential first-in-class Activin E antibody, where what you would do is you'd get onto a GLP-1 like therapeutic to cut down your weight to where you want it to be, but then instead of being on a GLP-1 for the rest of your life, you switch over to a maintenance dose of IBIO-610, where, because it's an antibody now, you can dose much less frequently and hopefully have a lot less of those nausea-like side effects from the GLP-1 therapies. This one has, again, a really nice long predicted human half-life, suggesting infrequent dosing up to maybe even 100 days. What is the platform that's powering these discoveries? I like to think in terms of what is it that differentiates us from others first before actually going into the AI technology itself.
I like to think about antibody drug discovery as going on a fishing expedition effectively. Right? Now you can do what the big pharma does. Big pharma effectively hires an army of fishermen to effectively go out there, scour the seas to find that rare catch. That's sort of the brute force method. Sometimes that works, but a lot of times it doesn't, depending upon how rare that fish is and what unique characteristics of that fish you want to catch are. Now, another field that's really booming and popular right now is de novo antibody design. Right? The idea here is it's akin to not going fishing at all. Instead, you're staying in the lab and you're more or less building a lab-grown version of a fish. The advantage there is you're not spending the manpower to go out, but it's not the real thing, right?
We don't know how safe it is for humans, and we also don't know how reproducible this process will be for all types of antibody fish out in the sea. Now, at iBio, we don't have the capital to hire an army of fishermen, nor are we interested in reinventing the fish. What we do is we actually use machine learning in a very different way. We actually build personalized baits that we arm our fishermen with. What that does is it allows them to go out there and find that fish with a higher degree of efficiency, oftentimes orders of magnitude more efficient than without those baits, thus effectively beating big pharma at its own game.
Looking under the hood, this is really what it looks like. That special bait that I'm talking about is effectively an engineered version of the antigen, usually focused on just the epitope that we want to hit. That allows for epitope selective antibody discovery. The fish that we're going to catch, these are human, right? These are not in silico designed. We have our own proprietary human antibody library with a human CDR diversity and a clinically validated framework. Then in order to ensure that the developability characteristics of these molecules coming out are very high, we actually do our selections in a combination of a phage display and mammalian display, where that mammalian display is done in the production cell line. That's really nice for developability optimization of your molecule, resulting in on epitope developable clinical candidates in a fast timeline.
We're going to focus predominantly on the AI side of things, where I'll talk about how we started spinning up the engineered epitope engine and how that has evolved into a platform where we can now start to look after going after GPCR targets for antibody discovery. What is this engineered epitopes engine? Well, starting with a naive in vitro or in vivo source of antibodies like that on the left-hand side, we use generative AI to build peptide scaffolds that support an epitope, either linear or discontinuous, in a native-like conformation. Then using this small structural mimic as a bait, we fish out those antibodies that bind to the epitope of interest, and that's what greatly increases the efficiency with which we can discover epitope selective antibodies against a target of interest, like maybe a GPCR on the right-hand side.
We can do a whole lot more, actually, with this generative AI than just go after epitopes of interest. If you imagine this really complicated heterodimeric GPCR on the left-hand side, we can use the generative AI to scaffold individual epitopes on that to do epitope selective antibody drug discovery like that on the upper right-hand side. Let's say you're interested in something much more complex. Let's say you want to develop an antibody that binds to or stabilizes a protein-protein complex or protein-protein junction. We can actually use this generative AI technology to glue together different subunits in the right orientation with respect to each other to enable a junctional discontinuous epitope antibody discovery. Maybe you're just interested in developing an antibody against the GPCR itself, right?
We can actually use this technology to re-engineer and reimagine the transmembrane domain of these proteins into a soluble format, and then that makes in vitro selections much, much easier without having to worry about the cell. These antigen designs are really optimized for antibody drug discovery. They are water soluble, they are structurally stable, and they have minimal scaffolds to avoid off-target binding, as well as the chemical composition of the scaffolds themselves being predicted to have low immunogenicity. All right, let's first talk about using this technology to target individual epitopes. We'll focus this specifically on tackling an age-old problem of immune tolerance. We had a collaborator that wanted to go after this target called latent TGF- beta 1. It's a potential oncology target for immune modulation, where TGF- beta release from the latent complex is immunosuppressive.
The idea is here, can we build an antibody that locks TGF-beta into its latent state so it doesn't engage in the downstream signaling pathway? The problem that our collaborator had was, well, latent TGF-beta 1 shares 89% sequence identity between human and mouse. Because of that, immunizations in mice failed just because they couldn't overcome the immune tolerance, right? Mice don't want to build antibodies against things that look like themselves. We thought, well, maybe we can help you by engineering protein mimics that can help you overcome and break the immune tolerance.
Now, we didn't know ahead of time which epitope would be most efficacious for locking the complex into its latent state, so we just built a bunch of engineered epitope designs like what I'm showing here, where the colored regions are the epitopes that we're representing, and in gray are the protein scaffolds that are supporting the epitope in a native-like conformation. Like I mentioned earlier, we deeply integrate our AI into an experimental validation pipeline. Here, for example, we had evidence from previous HD Exchange data that epitope 3 in particular bound to a benchmark antibody. It's SRK-181. It's the competitor molecule, actually, to the collaborator. Indeed, we find that our miniaturized representation of that epitope does indeed bind to the benchmark antibody, SRK-181, via Octet and ELISA binding. Similarly, we built an engineered epitope that represents the integrin binding site. That's epitope number 5.
As we like to see, we were able to functionally validate these designs by observing Octet and ELISA binding to the corresponding integrin for that site. How we actually did the immunizations, and this is actually a nice tip for all you out there. If you're doing peptide immunizations, one of the big problems with peptide immunizations, if you just rely upon the peptide itself, since that peptide doesn't have the full molecular context of the full-length protein around it, you'll oftentimes get a lot of hits to the peptide and not very many hits that actually translate to full-length binders. To overcome that problem, what you can do is alternate between peptide immunizations and full-length protein immunizations.
This steers antibody discovery to the epitope of interest while still filtering out those that only bind to the peptide and ensuring you get epitope-selected binders that also translate to full-length binders as well. This was a really successful campaign for us. It was a big turning point. We saw good serum titers across the board for latent TGF- beta 1, where they weren't able to observe any with just latent TGF-beta immunizations. Really remarkably, and this opened up a lot of doors for us, is we observed for the very first time, we thought it was actually just luck, but it wasn't that just immunizing with the engineered epitopes alone actually elicited a similar immune response to if we co-incubated with the latent TGF-beta 1. Strongly suggesting that these designs are actually now, the AI is so good at making these structural mimics.
They're good enough structural mimics of the real thing that maybe you don't even necessarily need to immunize with the full-length protein. That was sort of eye-opening. That's that blue trace right there. Usually, you saw what's on the left-hand side, where it's not as good as if you were to immunize with the full-length protein. On the right-hand side, sometimes it was actually as good, if not better, just to rely upon the engineered epitopes. This is a complicated target where hybridoma screening actually had to be done via a flow-like experiment, where they got hundreds of binders. Even from that first pass, they were able to discover hits that had comparable TGF-beta release inhibition compared to their competitor, which is SRK-181, shown in purple. Sorry, in orange, excuse me. The hits that were comparable to that benchmark are shown in orange.
Then moving those orange hits into purifications, we're actually able to observe that a lot of them really are comparable in their ability to inhibit release of TGF-beta from the latent complex compared to SRK-181. This is really amazing. We were really surprised by this result because SRK-181, you have to understand, is a highly matured molecule. It's gone through multiple rounds of maturation to get to where it is. These are just clones that we just popped out of a mouse, right? There's clearly an epitope privilege going on here, where by steering to individual epitopes, we can unlock novel biology that you can't otherwise. That was really encouraging.
We want to step up our game a little bit, and especially since we're interested in the obesity market, we wanted to go after a more challenging target to see how far we could break immune tolerance, and that next case study is Activin E. It's been a long sought-after target, we've heard, under the radar. It inhibits fat metabolism, and targeting Activin E may selectively reduce fat while retaining muscle mass, which we're now starting to see a lot of evidence for, which is a huge plus over the GLP-1 like therapeutics. The problem is Activin E in human and mouse are nearly sequence identical. They're only off by three residues, so it's roughly 97% sequence identical. In doing immunizations, you have to overcome the immune tolerance.
The other challenge of this project is it's just very difficult to produce Activin E itself in an active state just because it has so many disulfide bonds. Our answer to this is, well, given our success in immunizations in certain cases with just the epitope alone, what if we could reproduce that here? The idea is here, let's chunk up the Activin E into six individual epitope design regions, and let's do immunizations similarly to how we did before and see whether or not those immunizations can elicit an Activin E response. For this task, because we knew it was going to be really hard, we actually had built a novel nanoparticle display system.
Moving away from just what you can buy off the shelf to a ferritin nanoparticle display system, which can now present up to about 50 epitopes per nanoparticle, really improving the immune response and B cell activation of our epitopes. Because we certainly don't want to just immunize with the small peptides by themselves. That would just flag the regular immune system and do pretty much nothing. We set it up in a way to really interrogate whether or not these engineered epitopes have the ability to break immune tolerance. Because that's the question we really want to answer. The control is on the top. Control is just immunization with the full-length Activin E in orange at the end.
Everything below, we're doing a pre-immunization with the engineered epitopes that's in light blue, and we're measuring serum titers against Activin E at three different time points shown in purple before post-engineered epitope immunization, post-Activin E injection. Okay? What we saw was remarkably in line with what we expected from computational predictions. All right? The control was not surprising to us at all because of the nearly identical sequence identity between the mouse and human forms of Activin E. If you just immunize with Activin E, you don't elicit an immune response. Remarkably, there are a few epitopes for which when we immunize with them, that you get a significant immune response against Activin E, even though the mouse has never seen that form of Activin E.
What's really neat is if you actually take the predicted structure of Activin E on the right-hand side, and you predict what is the immunogenicity as a function of residue position, you see there are two hotspots that would have been highly immunogenic had it not been for the immune tolerance issue that you have to break, and that's epitopes 2 and 6. Indeed, what we see is our technology allows us to break immune tolerance so that we can observe the expected immunogenicity pattern that you see from predictions here. Again, I can go into it, but we were able to observe many hybridomas that were positive on Activin E. We were also interested to ask the question, well, does this immune tolerance breaking also translate to VHH discovery? The answer, thankfully, is yes.
We tried 2 cohorts of llamas, and we were able to observe again that after immunization with our cocktail of engineered epitopes, you see serum titers generate against Activin E, even though it's never been immunized with Activin E in this case. Although I think what was really nice to see is for the first llama, you actually see a nice synergistic effect so that you actually get a boost in titers after doing an immunization with the engineered epitopes, followed by immunization with that Activin E, now that it has antibodies that can lock onto that complex. This is a nice stopping point for me. I can talk about the key learnings from the engineered epitope work for immunizations. We were really surprised by how remarkably structurally accurate the diffusion models appeared to be, especially for the Activin E project.
Only a single design was made per epitope, and that was able to deliver the efficacy we observed in the previous slides in our immunization studies for breaking immune tolerance. That was really neat because I've lived through the Rosetta years, right? Where you'd be lucky if anything out of 100 designs produced anything, right? Now you can almost expect to see an efficacious result from these diffusion models just from a zero-shot strategy. Another thing that's sort of a tip in general is we did observe that the small size of the engineered epitopes is actually a pretty big advantage for immunization. Because they're small, we can pack a lot more onto our nanoparticles. That's why we can get 50+ epitopes presented per nanoparticle immunogen, which we found was important for overcoming the immune tolerance.
Making these small, compact representations of individual epitopes is really synergistic with immunizations, we've found. It also just opens the door to, well, wait a second. Does this make any target druggable in principle? Because there's a lot of targets you simply can't just immunize with because they're insoluble or maybe they're complex, and they'd fall apart as soon as you immunize a mouse with them. But if engineered epitopes can be sufficiently good sequence and structure representations of that portion of the full-length protein, maybe the full-length protein isn't necessarily needed for immunizations, and these are the questions we're starting to ask now, and can we go after these targets now that we couldn't otherwise in immunizations? Now I'll migrate back to the title of my talk, which is, how do we use this technology to unlock GPCRs? Because we were really excited. Okay.
We seem to have a technology now, and this is a very custom-built technology in-house. We have a way of building structural mimics that appear to be very efficacious and are experimentally, functionally validatable. How do we apply this to the long sought-after space of GPCRs now? This is actually a fairly early study that we did in this topic. We went after CCR8 as sort of first pass on this. It's an emerging immuno-oncology target where CCR8 is upregulated on Tregs within the tumor microenvironment, where it promotes Treg recruitment expansion and overall immunosuppression. The idea is if you could build an antibody that binds to the extracellular loops of CCR8, well, that could potentially serve as a Treg depleter for oncology. In order to do that, right, we adopted a similar strategy like I showed on the previous slides.
Can we ask the AI to imagine a peptide scaffold that supports the extracellular loops in a native-like conformation? I won't go through the campaign itself, I just want to talk about the experimental validation we did. It was nice to see this is the very first time we actually did a full structural biology study here that the NMR structure, the experimental NMR structure, which is shown as a transparent overlay, aligned really well with the generative AI design. That was cool, but also limiting, right? We've been focusing thus far on building small structural mimics for portions of an antigen target. What if I'm actually not interested in going after just little portions? I want an antibody that binds to and potentially even regulates GPCR function.
Now what I need is a way to build an engineered antigen that encompasses the full extracellular domain of a GPCR. To go after this challenge, because we're in the obesity space, we thought, well, maybe we should try targeting the GIP receptor, right? The GIP receptor is a target of a lot of weight loss drugs now at this point, but one of them being tirzepatide, otherwise known as Zepbound. The challenge we posed for ourselves here was, could you design a soluble version of the GIP receptor and then show that it specifically binds to the native ligand it should, which is GIP, and not to its related ligand, GLP-1. This is how we formulate the problem for the AI.
We have it as an input to the AI, a cryo-EM structure on the left-hand side, and we ask the AI, "Hey, can you redesign and reimagine the transmembrane domain into something soluble?" Right? We asked it to make six designs, if I remember correctly. The designs look like that on the right-hand side, where now the transmembrane domain is much more packed. It doesn't look exactly like the original input and in sequence space is certainly completely different. That scaffold now supports a native-like extracellular region. It's water-soluble now, which is great for in vitro selection and in vivo as well. It's stable, it expresses well, and then we can even reproduce PTMs by expressing these guys in human cell lines too. A nice soluble representation of a GPCR. Does it work?
Yeah, it binds to GIP, which is great, and we show it on the right-hand side. In fact, most importantly, it doesn't bind to the individual components, at least not very well. The GIP can't bind to just the extracellular domain by itself. We had to make sure. It also doesn't bind non-specifically to the transmembrane domain either. Only when those pieces are put together does GIP bind to the full complex, thus functionally validating this as a soluble GPCR that binds to its native ligand. Binding alone, it's insufficient. Does it bind GIP specifically? The answer is yes. Right? We tested against a whole panel of compounds that should and shouldn't bind. It should bind GIP. Good, it binds GIP. It shouldn't bind GLP-1, its related cousin. Good, it doesn't bind to that.
Actually, also really neat is we demonstrated it binds to Zepbound and retatrutide, which are current peptide therapeutics that are already out there, suggesting this could be really useful as a surrogate for drug discovery, maybe even beyond just antibody drug discovery. Maybe it'd be useful for peptide discovery too. We were really excited about this. We want to do some degree of structural characterization, even though we didn't have access to a cryo-EM at the time. We did have access to negative stain electron microscopy just to take some images of it as an FC fusion. We were able to observe at low resolution, yes, the transmembrane domain adopts a similar morphology as we would expect from the 2D class averages.
Despite the fact that there is clear flexibility at the hinge between the FC and the transmembrane domain itself. Now, I have to let you in on a dirty little secret. In reality, right, that project was fairly biased to success because there already existed a cryo-EM structure of the GIP receptor in complex with GIP. So we knew exactly which residues we should keep in our designs to make sure that it still retains binding GIP. So as a final challenge to my team, I was like, "Okay, that was great, but what if we didn't have the structure ahead of time? Would it still be possible to use this technology to make a soluble surrogate?" GPR75 is another hot target in the obesity space. It's a genetically validated obesity target.
At the time of this study, it's different now, no experimental structures were available, and there were only two potential ligands that we could use to validate these designs. One is 20, but it binds the transmembrane domain, so it's not useful. That's what we actually want to redesign and re-engineer. There's also CCL5, which is known to bind extracellularly, but again, no experimental structural insight into exactly how it binds, what residues are important, and even just overall what GPR75 looks like. This is much different, right? Now we're asking the generative AI to design a soluble version of the GPCR based on an experimental structure, but based on an AlphaFold structure of this GPCR. The input target's on the left here. It's an AlphaFold structure, and this GPCR actually has a fairly large intracellular domain.
It's not quite half of the overall mass, but it's getting pretty close. When we're asking it to make these soluble designs, we're asking it to not only chop off the entire intracellular domain, but to reimagine that transmembrane domain, resulting in something that looks very, very different now from the original input. Because of the challenge of this task, we built, I think, 12 designs, purified and tested them for binding to CCL5. Hit rate was definitely lower here, but we actually observed CCL5 binding, which was really remarkable to us, considering that we're starting from just sequence alone here. It's models built on top of models. What was additionally promising is that design number eight, which is the one that showed significant CCL5 binding, also passed with flying colors the binding test to a commercial anti-GPR75 that's out on the market.
We got excited again. We wanted to just at least take some images to see whether or not the overall morphology agrees with our expectations from the diffusion model's predicted output. Indeed, you see this sort of bunny-like pattern, right? Where the ears of the bunny in this image here are the transmembrane domains in that packed helical fashion. This is great, right? We now have a platform where we have an AI system that's deeply integrated into an experimental validation pipeline so that the things that come out of our AI now get experimentally validated before going into an actual drug discovery campaign. Alex, where is the drug discovery campaign? That's the next question. Do these things actually now work in antibody drug discovery? For that, we're going to look at a case study with amylin receptor agonism.
Why target amylin? Amylin is an orthogonal pathway to the GLP-1 based obesity treatments, but it's been a very, very hard egg to crack. Why? I mean, there are a lot of peptides actually now that bind to and agonize this receptor pathway. The problem is that none of them are exquisitely selective for amylin receptor. Amylin receptor itself is comprised of two different proteins. One, the calcitonin receptor, showing gray here, and two, one of three ramp accessory proteins shown in these colored images on the right-hand side. You have the peptide therapeutics that bind to the receptor, but because it's activating multiple pathways, it's not exquisitely selective to what we believe is the fat selective loss and safe pathway.
Our goal here is instead of building what's called a DACRA, dual amylin and calcitonin receptor agonist, could we build a SARA, S-A-R-A, profile, something that's actually selective amylin receptor agonists using our protein engineering. Again, peptide drugs exist, but they're limiting, right? pramlintide has been FDA approved but requires multiple dosing a day, so no one's going to want to do that. You still have weekly dosing even for the next gen peptide therapeutics like cagrilintide. Not to mention they're not specific or exquisitely selective for the actual amylin receptor pathway. The goal here was, could we build an antibody that was selective for amylin receptor, meaning it doesn't bind to red alone, doesn't bind to blue alone.
It only binds in the presence of red and blue together, and then couple that to a synthetic amylin peptide agonist, thus giving you an exquisitely selective amylin agonist. Again, we didn't know where on the protein would be the best place to target, so we designed the full spectrum of epitopes, small epitopes to large. We did the same sort of experimental validation, which means that we observed and validated that our transmembrane encompassing designs on the right-hand side bind to salmon calcitonin as you expect, and don't or weakly bind to the individual components. Then when we're looking at these junctional designs here, there is no antibody that's selective for the junctional designs. At least we're able to observe that anti-CTR antibodies bind to our designs in the way we would expect.
Now, we're fairly agnostic as to whether or not the right answer would be immunizations or phage panning. We subjected these designs to both. Just like I showed you before, we were pleasantly surprised to see that the engineered epitopes were able to generate an immune response selective for the full-length protein, even without immunizing with that full-length protein. Again, so you're seeing we have day one, no immune response against the full-length protein. Day 21, you see a big jump even though you haven't immunized with it yet. You just immunized with the engineered epitopes, and that's further improved as you get to day 49. Here's a more challenging case of a helical protein, which is just more naturally lower and not as immunogenic. Still, you see a meaningful jump between day one and day 21.
Now, that said, to actually get to that SARA-like profile, it turned out that immunizations didn't give us the amount of levers and knobs that we needed to control the antibody drug discovery process as in vitro discovery could to get us there. I'll now focus on the in vitro side. Right off the bat, though, we noticed from phage panning against the engineered epitopes that the engineered antigens like that in the upper left-hand corner that we panned against gave a lot of antibody hits, tens of antibody hits that translate immediately to cell binders. Whereas if you do phage panning on just the cells alone, you get no specific binders, right? You get things that are binding off target, but to the cell.
Given that success, Cody's the guy that actually runs our in vitro discovery engine, our platform engine in general, had a really brilliant idea. I thought it was crazy at the time. He said, "Well, now that we're so good at building these structural analogs or surrogates for these things, would it be possible to bake in cross-species reactivity into the selection process itself?" This was his idea, and I made the designs for him. He said, "Well, Alex, can you build me human amylin receptor 1, human amylin receptor 3, and rat amylin receptor 3 engineered epitopes? And I will see whether or not with four-dimensional mammalian display sorting experiments, I can select for a triple positive on these, giving a multi-specificity profile we want, while also being negative on the human calcitonin receptor." It looks like this on the right-hand side.
In the lower left-hand graph of the FACS plot, we're looking at low human CTR binding, high human amylin-1 binding, and then taking that subpopulation, we're looking to see are there any clones in our naive library that are also positive on rat and human amylin receptor three binding. Out of the millions and millions of dots on the screen, there were 10. He was able to actually recover them, and now we have for the first time, I think, it's very interesting. It's actually a VHH, if I remember correctly, but an antibody that binds selectively to amylin receptor one and three and not calcitonin receptor and is species cross-reactive with rat, all from a single experiment, a single selection experiment. Okay?
When you move that into an agonism assay by tethering a peptide onto it, you see now for the first time, unlike a cagrilintide, which is nonspecific to amylin versus calcitonin receptor cells, we see that exquisite selectivity profile. We're not seeing activation of the calcitonin receptor cells, but only activation of the amylin receptor cells. Just wrapping up here, the key learnings from the soluble GPCR work, again, I have the same thing to say about diffusion models, that they're just maturing at a remarkable rate. We're observing that we can build soluble GPCRs now with a design success rate of something on the order of about 50%, which is just really amazing.
Using these as soluble surrogates for the GPCRs themselves, we're learning that we can actually bake in species cross-reactivity, subtype cross-reactivity, into the early selection discovery phase of the antibody discovery campaign via a combination of these soluble antigens and multi-dimensional mammalian display sorting. Finally, I just want to leave off with, I know we're an antibody drug discovery company, but the utility of these soluble surrogates may not be limited to just antibody discovery, right? I mean, we're functionally validating these soluble surrogates as binders to the peptides and to peptide drugs as well that have been delivered against them. With that, I want to thank the team.
Our CEO is in the upper left-hand corner. I don't know why he's black and white. He's not dead. But he's alive and well. I'll make him colored next time. One of the primary drivers of the work that we saw here is in the lower left-hand corner. He's the guy with the glasses kind of pointing, holding a piece of paper. His name is Arjan. A wonderful guy to work with. The genius who thought of that sort of multi-dimensional FACS sorting experiment to get cross-reactivity in a single experiment is Cody and Hongyu in the middle. Of course, thanking the rest of the team.
You know, finally, I just want to leave off with we really love hard problems. This is what drives the science, this is what drives our innovation. If you have really interesting ideas or problems you think could be addressed with GPCR solubilization or protein engineering in general, similar to what we're doing here, the door is always open to a conversation at the least. Anyway, I just want to thank everyone for their time. That's it for me.
Thank you so much, Alex. That was such an amazing presentation. Now we'll get into the Q&A, so if anyone has any questions, please drop it in the chat and then we'll cover those. As we wait for those to roll in, I actually have one for you, Alex.
Sure.
Now that iBio is using this platform, what do you see as kind of the next big hurdle for machine learning in this antibody discovery over the next three to five years, you would say?
It's probably the intersection of de novo antibody design and antibody antigen engineering. I think right now the world is way too focused on the antibody engineering side, meaning there's a bunch of companies out there who are just de novo designing antibodies to bind to targets. Why not combine the pieces, right? If you can build efficient soluble surrogates for the antigen side and the paratope side, I think that unlocks the next wave of innovation for antibody drug discovery in machine learning. I think that people haven't quite gotten there yet, and I'm hoping that's going to be the next wave of innovation, engineering both the epitope and the paratope simultaneously to unlock that next gen.
Okay. To kind of follow up with that.
Right now most people are just engineering the paratope side.
Yeah.
Yeah.
Okay. To kind of follow up with that, it's kind of impressive how far these generative models have come, and so I'm kind of interested in the trial and error side of the process. From a practical standpoint, was there a most common hallucination or failure mode that you saw when the generative model designs these?
Yeah. That's actually a really good point. I would say the largest failure mode when using any diffusion model is not constraining well enough the, I want to call it the blueprint, but maybe if I can go back a little bit. Yes. Let's use this as an example. It could have come up with many things, right? Constraining these models to create a compact structure is actually still a challenge.
A lot of times, these models will hallucinate things where you'll get things that, I guess to a structural biology person's eyes, obviously don't look like they'll fold well because they're not compact. I think it's really important when using these diffusion models to enforce global high-level structural constraints that are weak but still ensure that it creates a globular structure as opposed to one that has things shooting out left and right. They're still not perfect. They still need to be babysat.
All right. Well, we're kind of reaching the end towards our webinar, and so if you could just go to the last slide for me.
Of course.
I want to thank you, Alex, so much for sharing your expertise with us today and just seeing how generative design is practically applied to the GPCR discovery. It just really highlights this lock-and-key series of drug discovery. Before we sign off, I'd really like to share to our audience about our upcoming symposium, which is Cell Signaling Breakthroughs, and you don't want to miss this next major event. We are officially hosting this on April 25th in Seattle, Washington, and it's going to be a deep dive into the cell signaling pathways and processes.
Currently, we're actually running a two-for-one promotion, so this is the perfect opportunity if you want to bring a colleague or a member of your lab along to join the conversation at no extra cost. To learn more, just scan the QR code and be sure to grab that before we close the room, for this deal will be only a limited time. Thank you so much for your time, Alex, and everyone who has joined us from all over the world, and we really look forward to the next webinar. All right. Thank you, Alex. You guys have a wonderful day.
Thank you.