Distributed Bio’s phage display library has been computationally optimized based on the analysis of thousands of human antibody repertoires. Their library routinely produces over 5000 unique antibodies against any target.
How do they efficiently screen so many unique hits? They use the Carterra platform.
The Carterra LSA high throughout SPR platform is the only antibody screening technology that can deliver high quality kinetic data directly from scFv bacterial supes at a high enough throughput to meet capacity, enabling thousands of clones to be screened in just days. This gives Distributed Bio the confidence to bypass ELISA screening and proceed directly to binding kinetics and epitope binning after panning. Combining Distributed Bio’s phage display library with Carterra’s LSA, enables bioengineering feats that make many other discovery technologies obsolete and can help you achieve a one-week discovery cycle.
Jacob Glanville, PhD
Co-Founder and Chief Scientific Officer
Yasmina Noubia Abdiche, PhD
Chief Scientific Officer
0:00:00.7 Elizabeth: Good day, everyone. On behalf of Cambridge Healthtech Institute's Global Web Symposia Series and our sponsors, Distributed Bio and Carterra, I'd like to welcome you to Accelerate Your Antibody Discovery with High-Throughput Kinetic Screening of Phage Libraries. My name is Elizabeth Lamb, and I'll be the host and the moderator for today's event. Now I'd like to introduce our presenters for today. First is Jacob Glanville, PhD, Co-Founder and Chief Science Officer of Distributed Bio. And our second presenter is Yasmina Noubia Abdiche, PhD, Chief Science Officer for Carterra. Welcome, Jacob and Yasmina. Jacob, the presenter bottle is yours.
0:00:50.4 Jacob Glanville: Hello, everyone. Thank you, Elizabeth, for coordinating the webinar, and thank you, Carterra team, for inviting me here to share some of our work today. I'll be spending a bit of time talking about our efforts to apply computational immunology to optimize antibody discovery. This is a bit of a paradigm shift where our goal instead of identifying antibody hits and then spending a bunch of time trying to correct them and optimize them, is to apply deep computational methods upfront to design libraries that have all the hits you'd need, and you can apply selection and filtering rather than engineering. In order to do this, we ran into a problem of a victim of our own success, where we had a huge number of hits against the antigen tested, and the Carterra LSA platform really came to the rescue for us in being able to screen the vast numbers of hits we were recovering from the library very quickly. And that's the story I'll be sharing with you today.
0:01:51.6 JG: So my interest in this work really started in 2008 when I was comparing the amount of time it took us at the bench to do in-vitro discovery and optimization of antibodies, typically nine to 12 months from initial lead discovery to optimization, to the amount of time it took in vivo, with the particularly impressive system being our own bodies that could go from a small infection to responding antibodies within seven days. And that always made me feel that our technologies were deficient compared to the fully optimized state of our evolved moving systems, 642 million years of evolution. And it made me think that we should be able to get closer to that seven-day goal by applying computational methods and things that we've learned along the way on engineering antibodies effectively to produce libraries where most of the work had been accomplished in advance through the design and it could be avoided downstream. So you could just screen molecules and we're ready to go and you don't have to fuss about trying to correct them.
0:02:52.6 JG: So, at the time, and around 2008, this is the collection of major technologies one could use to generate antibodies. You have hybridoma, traditional phage, and then the emergence of transgenic animals that are producing human antibodies. And I've made a point, an effort to map out roughly how long the development timeline takes from discovering a hit either through immunization, hybridoma generation or phage panning, and then subsequently to make that hit high enough affinity to be in the functional range that one is interested in, maybe make it cross-reactive to the species one would like to be able to run animal studies in, make it more thermostable, remove liabilities, and so forth. And in general, you can see that the additional time taken by immunizations are compensated by having higher affinity molecules downstream. With hybridoma, there's a bunch of extra time taken for humanization and then optimization to the molecule, but you don't need to do it with a transgenic animal. Even then, some work is necessary to make those molecules therapeutic-ready.
0:03:53.6 JG: What I'll be describing today, the consequence of the last nine years of work, is a new paradigm, where within two months can go from taking an antigen, panning it, and then by combining our ability to rapidly pan and produce a very large number of hits, 5,000 to 9,000 against the target, with the Carterra LSA platform's ability to high-throughput kinetic screen those hits, radically changes the speed with which we're performing this kind of development. And at the end, I'll talk a little bit about a seven-day discovery cycle, and that's something we're very excited about, and it's really only made possible because of these advances in the computational immunology of the libraries, combined with the ability to perform rapid kinetic screens.
0:04:38.7 JG: This work really was made possible due to the advent of high-throughput sequencing. Starting in 2009, it became possible for me to high-throughput sequence a library to figure out what diversity was present, and then high-throughput sequence a very large number of hits against a large panel of antigens. And that gave me the ability to understand what parts of our libraries were being successful and what parts were failing against given antigens. What you're seeing here is a progress, the iterative improvement of libraries, four of them built between 2009 and 2013, all panned against the same antigen. This is PCSK9, and you're seeing the nanomolar affinities shown there on the Y-axis. What you can see is that the work helps that each iterative library resulted in more hits and higher affinity hits over time. Some of that work has been published and it's referred to at the bottom.
0:05:24.3 JG: And what I'll be talking about today is the final conclusion of that work, and that's the Distributed Bio SuperHuman Library. I'll talk a little bit about the engineering decisions that went into it, the strategy to optimize the diversity of the CDRs, to be able to produce what we're getting now about 5,000 hits against every target panned, including about 300 unique picomolar hits. That's a number we've been able to discover through the use of the Carterra technology. And then also a bit about how we try to encode the developability upfront, so you don't have to spend time engineering every molecule that comes out, they're ready to go as therapeutics. I'll start with describing the Scaffold Selection. The original libraries I worked with used the whole natural library repertoire, so 70 different VH family scaffolds and germline scaffolds, and about 70 V kappa and V lambda scaffolds.
0:06:12.6 JG: For this library, I wanted to remove most of them because many scaffolds that are found in the human repertoire make terrible choices as drugs. And you can see that at the bottom, a couple of examples of B genes that I just would never wanna build a drug on, things like 4-34 that have inherently terrible half-life because they are inherently utter reactive to red blood cells and lymphocytes, things like igHB2-5 that are inherently unstable. They're difficult to create a stable therapeutic that you can concentrate, and it's non-immunogenic, as a consequence of creating aggregates. And then there's things like igHB4-B, a B gene that's only found in about half of humans, if you build a drug on that, it will appear non-human in the other half of subjects, so you would never want to do that. There are B genes like iglB6-57 that are aggregation-prone, and the list goes on. In the end, I decided I didn't want most of the 3,500 combinations of B genes. I wanted B genes that satisfied six different criteria. They should have been able to have already been put in a human by someone else, so phase one plus. I wanted them to be as non-immunogenic as possible, and I'll talk about what I mean by that. I wanted them to be aggregation-resistant. I wanted them to be able to display well on phage. It doesn't help me if they're good B genes but they don't display well on phage, if I'm using phages by libraries, they just wanna be unproductive in creating binders.
0:07:30.6 JG: I wanted them to be thermostable because that influences expression, the ability to concentrate the drug and the resistance to anti-drug antibodies as a consequence of avoiding misfolding that is more easily captured and presented by the immune system. And finally, I wanted some canonical structural diversity. I didn't just want one scaffold, because every effort to create a library based on one scaffold has a notable side effect that certain epitopes are heavily favored by that scaffold and other epitopes are neglected. I wanted to make sure I could target any epitope. Briefly, here's the kinds of data that we used to generate that Venn diagram and make our ultimate selections. This is a summary of about 400 therapeutics that have made it to phase 1+ trials. You noticed that certain B genes are heavily biased. I've highlighted the four B genes that we've ultimately chosen for the heavy chain and light chain in blue at the bottom. You can see that there's some B genes that really dominate therapeutic development, they've gone into humans over and over again, and no one's observed a problem. I never wanted to build any drug on a scaffold that hadn't been tested in human or had gone into human and there'd been problems with the half-life. Second, I took advantage of my published work and other works and some unpublished data to identify B genes that we know can productively fold up and recognize their cognate antigens when displayed on phage.
0:08:49.5 JG: So I don't wanna use a B gene that if it's in a library but you never see hits coming out of that B gene, that implies that that has been having trouble folding up in E.coli or displaying effectively on phage. Either way, I don't want it. I took advantage of the 1,000 Genomes Project data to make sure that the B genes I chose had a single dominant allele in all human populations. I wanted to avoid variation in the frameworks that might make my drug look more human in one ethnographic group, accidentally making it effectively a racist medicine, or any population might have heterogeneity of alleles, in which case, every population would have a homozygote of the other allele where your drug would look more or less human, depending on the subject. I wanted to avoid that problem and we could easily do that by taking advantage of the 1000 Genomes Project data. I used some aggregation database data to identify certain B genes that I wanted to avoid because they're over-represented in B genes that are known to aggregate.
0:09:45.4 JG: I used thermostability data like most people that are working on synthetic and natural libraries to select B gene families that we know just are inherently more thermostable to begin with, so we're not having to fight an uphill battle on making these molecules therapeutic-ready. And as a result of that selection, I came up with four B gene heavy chain frameworks and four B gene light chain frameworks shown here, with the final criteria being that I wanted different canonical topologies for their CBRs so that there's different shapes of the paratope, and therefore it's more effectively able to target multiple epitopes.
0:10:17.7 JG: This is just a quick summary for the heavy chain. You can see there's a... In each case, many drugs that have reached the point that they've been named. Free scaffolds are available. You notice there's a bias. They're more of 3-23 and hB1-46 and relatively less of 169 and 3-15. And a similar story can be told for the light chain. I chose all kappa rather than lambda for a number of reasons that we can talk about offline. And you notice that again, KV1-39 is the most dominantly utilized in pre-existing therapeutics or drugs in phase trials. Since there are certain B genes that are heavily favored, I built the library to favor those as well, so that the library, rather than them being even combination of all 16 different sub-pools, it is over-represented for igHb1-46, KV139 and 323 KV139. And the idea here is to treat it like a stock portfolio, that if I can get a hit from those two vesting class combinations of scaffolds, I'd like to work with them. But if those two are not good at targeting my epitope of interest, I wanna cover my butt by having other 14 combinations available.
0:11:24.4 JG: And this balance ensures that we can target any epitope and when possible, we use our favorite scaffold combinations as a therapeutic lead. So that's scaffold development. But you can have the best scaffolds in the world, and it's not really gonna help you much if you don't have optimal CDR composition. And that's really the real advantage of our platform, is that I think we've solved the problem of redundancy and non-functionality and like that plagues most antibody discovery libraries. And it's really why we had to start working with Carterra to solve these sorts of problems. That's why we're so glad that their technology has been online. So I'll talk a bit about what we did for diversity optimization of the CDRs. Starting in 2009, it became pretty obvious when we started sequencing natural antibody libraries, that there was a huge amount of redundancy in those libraries. So even though people claim they could be needed a 10 in size, and the fact they were more like needed a seven, and that's because when we SeekDeep sequence them, we found there were many clones that were very high frequencies, they're taking up a lot of space in the library, and that redundancy limited the total library-effective functional size.
0:12:27.6 JG: And what was going on here is that if you take the blood from an individual, and we know that we have about 10 to 11 B cells in an adult, but if I sequence the PBMCs, the peripheral blood, you really only see about 10 million unique naïve clones in circulating periphery and about each of the five memory clones. And those memory clones end up producing more RNA and there's more cells with the same clone. So they end up... The relatively few clones end up dominating a peripheral blood contribution from an individual subject, and that's catastrophic even if you take hundreds of subjects 'cause there's a limited number of clones that end up taking over a lot of that library.
0:13:04.5 JG: We and others attempted to use synthetic methods instead to try to get around this, to use this synthesis by creating combinations of code on. In this case, this was an effort to make a library where each position with CDR, an equal opportunity of having all 20 possible, well, 19 amino acids, we excluded 16. And this turns out to be a pretty bad idea as well, because when we compared the library design of that equal amino acid dispersion in each position to the kinds of hits we were getting out of such a library, we would find that the hits just never had certain amino acids that were built into the library in certain positions. In particular, there were stem positions that were facing inwards towards the rest of the structure, therefore, there was more structural constraint on those positions. And that's pretty bad because imagine we tried to put proline in 5% in position 94 or 95, but proline, if you put it there, the antibody can no longer fold up effectively.
0:13:55.4 JG: What you've just done is you've destroyed 5% of your library, had one position with one amino acid at low frequency, and you're not wrong once, you're wrong many times during these designs, so that means that your library, again, if you run the math, there's not a thousand of slides you think it is and that's mostly because it's no longer a redundancy, it's because of low fitness of many of the molecules. And this has been a, I think a problem observed by many groups that have worked with synthetic libraries and that the library performs worse than it should and the hits that come out sometimes have odd biophysical characteristic issues. This is just summarizing how bad that problem is, which is showing if you're 95% correct in all the amino acids you've encoded in a synthetic library, your library still across all the CDRs will be about a 10th of the size you think it is, but you're not 95% correct. So we look over to the right and log scale and say you're 90% correct, your library's about a hundredth of the size you think it is. And you can't afford to get a B in synthetic library design because then your library ends up being like a thousandth times the size you think it is.
0:14:53.1 JG: So we made some efforts to go around that work back in 2011, we thought, "A-hah, why don't we deep sequence the natural repertoires of humans and other species and learn from them, the fitness landscape, elected by evolutionary forces, which amino acids are tolerated at which positions in the CDRs?" Shown at the bottom is the ZIM one plot distribution of amino acids in CDRH3 from human to mouse. You notice that there's a remarkable convergence of amino acid preferences that allow an antibody to fold up. And our thinking was, "Alright, let's use that distribution to make a synthetic library, but only chooses the amino acids found in nature and roughly the frequency of amino acids per position found in nature." And that helped. It produced better live, better molecules and higher affinity hits, as in that progression I showed at the beginning of the talk. But it still resulted in molecules that had odd biophysical characteristics because the body had never weeded out things with four positive charges in a row, or a bunch of small hydrophobic residues in a row or odd things in the body really would have rejected but in vitro, we ended up selecting four accidentally.
0:15:52.2 JG: That really brought me and Distributed Bio back to the drawing board, and I thought what I really want is to take advantage of those CDRs that have been selected by nature. They're less immunogenic. They've been pre-filtered for all sorts of nonsense that might be generically sticky. I just need to overcome that redundancy problem. So we sequenced a bunch of twins and strangers to try to figure out how much diversity is there and then I even the memory compartment, how much diversity is there across individuals, so how many people should I gather together to get appropriate diversity? How much can I enrich diversity by sort of separating the naïve from the memory, and how much blood should I get per subject? And what we ended up getting is about 140 subjects. We fax and max sorted their naive repertoire, that's about a hundred times more diverse than the memory repertoire, but it doesn't contain any somatic hyper-mutation, from that IGG memory repertoire that is much less diverse, but it has all that great mutations in the H1 and H2, L1 and L2.
0:16:49.8 JG: We took all that blood, we amplified the H3s produced by 140 subjects from the naive repertoire where there's more diverse and flat distribution. And then with specific primers that would amplify out just the CDRs from the frameworks we cared about, we amplified out the H1, H2, L1, L2 and L3 from the memory repertoire, combining VDJ recombination from humanity with somatic hyper-mutation from humanity on the specific frameworks. We then combined that stuff using entirely synthetics framework. So I didn't want any mutations in my frameworks. I wanted variation just in the CDRs and I wanted fixed frameworks just on those 16 scaffold combinations I described previously. Our assembly process ended up multiplying the diversity, which is really great because it means I can take 140 donors and end up with a library after the multiplication process, it's equivalent of about 35,000 donors contributing, but I don't have to deal with that much blood.
0:17:43.8 JG: We checked the assembly process using high-throughput sequencing at every step to make sure that our PCR, overlap extension wasn't causing weird aberration and the composition of the CDR combinations. And that ultimately resulted in a VK diversity. It's about a hundred times for diversity and finding a natural library. And a VH diversity that's about 2,000 times more diverse than what's found in nature, even though we've restricted ourselves to only 16 scaffold combinations out of 3,500 possible.
0:18:12.2 JG: During that process, there was another cute trick we were able to do and that is that biochemical liabilities existed at a certain rate in the natural repertoire. You have the emanation sites, acid hydrolysis sites and ligand constellation sites, persisting , and so forth, all these irritating bugaboos that you have to go in and fix on your therapeutic to ensure stability of formulation. Because we had 140 subjects and we had easy access to deep sequencing, what we did is we checked every subject. We sequenced every subject's diversity contribution from every CDR and we noticed that certain subjects had enrichment of say, 360 on a dominant clone in their H1 of 3-23. So we picked that subject out. We were able to carefully exclude some specific subjects that had an elevated rate of certain liabilities. And that means that our library, even though it's composed of diversity from nature, has actually got way less liabilities when it's found in nature and has the benefit of the most complex reduction of liability and that's the fitness of the CDRs is really good because it's been elected by human bodies as being tolerable.
0:19:12.0 JG: We did a final check, we were building it so that we can do as much work as possible upstream to make sure that diversity is as robust as possible, but whenever possible, I like to do selection pressures in vitro as well. So what we did is we built our library with a light chain, first we dropped in about 100 million different versions of each of our four scaffolds of light chain, and we put it in with a single fixed heavy chain, the 3-23 germ line with a pretty consensus looking CDR-H3. We then attempt to rescue to express all of those antibodies on the surface of phage, we heat stress them to 65 celsius, and then we pull down only the ones that had successfully expressed and hadn't fallen apart under 65 celsius heat pressure, and that resulted in us grabbing about the top 10% of all the light chains, which was still about 10 million different light chains for each one of the frameworks. But what's really cool about this is that now almost every one of our light chains in our library is thermostable, expresses well in E. Coli and functionally active, and then that's the diversity we drop our heavy chains into.
0:20:12.4 JG: And this as the result of producing a library that ultimately with 7.6, even to 10, including the heavy chains, when we dropped them, the number of transformants, for each one of those heavy chains is associated with an excellent expression partner and that increases the overall expression and thermostability of the entire library. This is just a little eye candy showing what those hits look like, so you can see there's almost no mutations anywhere in the frameworks, you don't have to do any framework reversion and the diversity and the CDRs is shown in the white there is enriched from what is produced in humans, so this is a natural diversity in the CDRs, entirely germline in the frameworks, thermostabilized in an expression selected on the light chain prior to assembly. We thought that was pretty cool. We started deep sequencing it, and then we started to run large numbers of screens, and I'll talk about that work and how we collaborated, or how we've been able to take advantage of the technology with Carterra to enable us to start high-throughput characterizing the hits coming out of the library.
0:21:08.1 JG: So the first thing we did is we deep sequenced our library and compared it to previous published libraries, that were not library yet published in the NPNAS in 2009 and then naive and a memory repertoire. And what you're looking at here is basically it's saying if you rank the clones in a library by their frequency, you can ask the top 10 or the top 100, top 1000, top 10000 clones, how much of the total library is that. What you can see is for typical natural libraries, like shown in green, NPNAS library from 2009, the top 10000 clones represent about 70% of everything in that library, so even though you claim you have 3.2 times either of the 10 transformants, almost all of those transformants are occupied by a limited number of clones, and that's catastrophic for the total diversity of that library, that's that redundancy crisis I was telling you about, and that's pretty...
0:21:55.8 JG: You see that naive is better than memory, but both libraries have those sorts of problems, whereas the super human that we built is an entirely different paradigm you see even in the top 100000, most dominant clones represent less than 2% of that entire library, so we're actually finally, taking advantage of the diversity of transformants in library construction. When we do just direct overlaps, we sequence library about four million reads deep twice, we found 99.93% of all clones to be unique on heavy chain and we see about 95% unique on the light chain when we do about a million reads each, so light chain is inherently less diverse than the heavy chain, but both of these numbers are ridiculously nice compared to what is typical of natural and even previous synthetic methods. We started panning it, our first panning was against betagyl, we just took betagyl on beads biotinylated it pan it three rounds in replicate, so independent pannings for three rounds, and then we characterized one plate by ELISA from the panning group A and then one plate by panning group B.
0:22:54.5 JG: And in this cursory test of two plates, we found 20 positives in the first panning group, and 41 in the second panning group, when we sequence them, those 61 hits we found 49 of them to be unique, and there was only one shared clone across the two. And what this told us was that we are getting lots of unique diversity and that screening two plates wouldn't even come close to covering all the unique hits in the library, and we needed to use a different technology to end up really... After answering the question, how many hits do we actually have in this library, how many plates would it be worth screening in order to get exhaustive coverage of all the hits. So to do that, we turn to our Genesis platform, this is a web-based Amazon Cloud-driven high-throughput sequencing and Sanger sequencing integrated system, for if we're looking at millions of antibodies integrating functional data along with sequence data. We can use it in a primer set that we have in-house to look at multiple rounds of selection very easily in our library, and we did that. What you're seeing here are some of the example of antigens that we panned in the first set of experiments, so it's growth hormone receptor.
0:24:00.0 JG: Again, pan and replicate, PD1 pan and replicate. Both of those are Fc-fusions. We also had biotinylated glucagon and biotinylated betagyl and then his tagged, transparent and his tagged human growth hormone. These were all panned on beads on a Kingfisher platform, and I'll talk about that in a minute. So what you're looking at here is a log scale, each little red dot represents a clone, and you see the clone distributions going down to one in 10000 events of enrichment, so we deep sequenced all of these different selections so then we're comparing against each other, and the main takeaways from this figure are that the replicates, most of the hits we come up usually are coming up above replicate C, the two independent panning growth hormone receptor, find a large number of growth hormone receptor hits, about 8500 for growth hormone receptor, 6500 for PD1 and so forth, at least 5000 hits against every antigen we've tested. Most of them only ever show up in the replicates and nowhere else, but we do see some cases of some clones that are emerging Streptavidin beads, so we think they're anti-strep hits or some causes show up at a lower frequency when you're doing nickel, on nickel beads, so they have enrichment of positive charges.
0:25:09.4 JG: That was pretty helpful for us because it lets us computationally pick a clone check to make sure it didn't show up on any other selection or the database hasn't already registered as being, we think this is a nickel binder, we think this is an FC binder and so forth, and also optimize our de-selections. So this is pretty cool. And it put us in a very different position, where we said well, if you have 6500 different PD1 hits, then you should first off, you run into a screening crisis, how the hell am I going to go in and characterize all of those hits. I wanted something faster and more powerful to analyze that, I really wanted kinetics, I wanted to be able to identify different epitope bins very quickly, figure out cross-reactive members. Our strategy was, if there's 6500 different unique PD1 hits, then surely we have PD1 hits like no one else has that crosses from human to cyno, to mouse. I might wanna pick things that are agonists or antagonist, those sorts of criteria where I need a faster screening technology. To do that we turned to the Carterra platform. So the LSA, we'll hear more about the technicals in a second from Yasmina but from my perspective, it was an amazing platform where I could get my panning rounds, it's able to operate on a very low concentration of input analyze, or the antibody and the ligand.
0:26:21.5 JG: And it's able to operate on 384-well format, which was great for me because I could go take a selection from PD1, as I'm about to show you, we'd pick plates of colonies, induce PPE, periplasmic extract overnight, and then they would be able to run it in a day to be able to get kinetics on hundreds of clones, 1000s of clones potentially, that's really a radical improvement over what we need to do to process through say reformatting a bunch of clones, for increasing the scale of culture, in order to be able to operate on some of the other less sensitive instruments, or other BLI or SDR instruments, they just have a way lower throughput. This is an example of some of the data we took from our PD1 selections. This is the primary screen, there was no Eliza done upfront, we basically just don't do analyzes anymore, we go directly to kinetics.
0:27:12.8 JG: This was the first plate gray or negative clones. Otherwise, you're seeing symptograms emerge, which is the second plate. And this gave us the ability to in our library immediately get kinetic information of hundreds of hits. And to understand the distribution of the kinds of affinities we were generating from this library. So this is a summary histogram of our affinities and what we're finding is about 7% of our hits were sub-nanomolar. And some of them were reaching the lower threshold about 500 picomolar that we could assert from the asset. And that was great because it gave us a rank list right away, we knew we were getting hits, we knew what our distribution was and we could pick the ones that have the best affinities or off-rates or range that we're interested in within a 24 hour period directly from the panning. That was very exciting for us. So you see some of the symptograms displayed down below. One of the cool things it also lets us do is we can go in and look for an affinity maturation of a particular clone and understand, okay, what's the relationship between sets of residues that vary across similar versions of the same clone and their kinetics. So that's pretty powerful. Those sorts of questions I've always liked answering, we have really good computational methods to do sequence activity relationship analysis, there's just a pain to do on previous SDR instruments.
0:28:31.4 JG: And the nature of our library is that sometimes you get this for free, so you'll have a dominant hit like the one they're showing at the top that's against betagyl. And then if you look in the deep sequencing data, you'll find multiple versions of that molecule, also enriching. And that's because that given H3 occurred at a higher frequency than the total library diversity, so multiple H1 and H2s were substituted with it, and in fact, we got a free affinity maturation, little mini affinity maturation, and a little bit of information about which residues were important for the high affinity of that molecule when it emerged, beyond just being able to do rapid kinetics, which already is super valuable, we were able to use the platform to ask very quickly for our PD1 hits, for instance, which ones are... We can ask which ones are cross-reactive from human design now and even to mouse. And we can start asking complex questions about specific epitope bins. So in the case of our growth hormone receptor panning, which was very successful, we got almost 8500 unique hits in that case, we screened a number of them, we found blockers, non-blockers, and then Yasmina's team ended up helping us discover some really interesting cases of some antibodies that appear to be able to selectively dislodge the ligand even when it's already bound.
0:29:37.3 JG: So this gives you an ability to search for, where is my hit landing, I want to find the binder that blocks the ligand, or I want to find a binder that doesn't block a competitor's therapeutic epitope. Or in this case, I want to find unique molecules with unique powers. And that's really the paradigm shift. If you have 20 hits coming out of the panning, you're just hoping one of them is functionally active. But if you have 6000 or 8000 hits coming out of a panning, then you're going to start looking for unique functions, you're gonna become more creatively aggressive at trying to find molecules that have unique properties, like an ability to displace a ligand which is already bound at the receptor. So we're very excited about that. And we're still trying to come up with new ways we can leverage the technology to take advantage of that sort of high throughput functional screening of kinetic properties.
0:30:23.9 JG: So that's essentially the summary here, right? We've gone from these more than a year discovery times down to about two months where we do ultra-fast screening, get ultra-fast selections, and then ultra-fast screening on the Carterra platform, and you end up with molecules that don't have all that work ahead of correcting the frameworks, thermo-stabilizing, removing mutations, increasing affinity and the CDRs and creating cross reactivity. You solve all of that in the new paradigm by having so many hits that you just apply additional selection pressure. So you pan against the human version, then the cyno version, then the mouse version. You apply heat between the rounds and you end up with the kind of molecule you're looking for not because you spent a year engineering it but because it was always there from the beginning, and you're just selecting it out, and you need it in tools very quickly to characterize.
0:31:09.1 JG: I'll talk for a minute about where we think this is going. And that's this superhuman zero-day, or seven-day discovery cycle. I kind of say this to get people's attention, but it really is something we are doing now. And the way we're getting a seven-day discovery cycle is that once we realized that this library could produce so many hits against each antigen, we just began industrializing the process, so we picked every immune-oncology target or many of them, they're shown there on the left, we pan them all on magnetic beads, using the Kingfishers that are shown on the right. And then we deep sequence the pool so we know how many hits are available, we know the pannings were successful. At that point, we bank them in the fridge or the freezer. And we have groups coming to us and saying, Hey, I'd actually really like to have a digit or a vista hit or I would like to have something, an agonist or an antagonist or something that doesn't block a certain epitope.
0:32:00.7 JG: And then rather than working nine months with a CRO or internally to see if they get the hits of that type, we say, Well, we already know we have 5000 hits of that class, you can just search through the pool, we've already de-risked the time and the success for you. And now you can just search through our collection of existing hits and we can partner on those molecules as you can have them. And that works like this, they come to us, day one, we take that out of the fridge, the freezer, we plate those clones, we have an agreed number of columns they want to look at. So it could be, we could look at four plates, twenty plates, whatever. Day two, we have colonies. By day three, we have PPE, by day four Carterra kinetics data and then from the kinetics data that quickly winnows down a collection of molecules that are of interest because of certain epitopes or certain cross-reactivities. And those are the ones that end up being characterized further. So it's a super-powerful seven-day discovery cycle because we have a library that's good enough that we can run it all upfront and we have a screening technology that's powerful enough to be able to screen hundreds of hits giving you powerful kinetic data right out of the gate.
0:33:01.8 JG: So I'm almost done. That's the main way that we're using the technology right now. I'll just spend a couple of minutes talking about ways we're excited to keep using Carterra going forward. One of them I alluded to a little bit, is in affinity maturation. So we have a technology called Tumbler that enables us to take a hit and let's say I wanna make it pH sensitive so that it releases its cargo in the end of the zone pathway. Or I wanted to bind conditionally in the presence of a secondary ligand, or do some other enhanced engineering. We have a technology that lets us very quickly take a molecule and produce about 500 near starting molecule variants. So it effectively explores all single, double, triple, quadruple affinity maturation variants. That's great. But at a certain point after the screening process, that library has got a lot of positives in it, and there's gonna be a lot of subtle amino-acid changes, and you'd like to be able to get a lot of kinetic information quickly across all of those.
0:33:49.7 JG: And so we've already done a couple of great programs using the Carterra technology that have given us the ability to have kinetic data for hundreds of versions of a starting molecule. So we can get a range of affinities if someone's trying to assess what the optimal affinity is in vitro or in vivo for a molecule or we can explore sequence activity relationships that quickly teach us which residues and which combination of the residues are giving rise to these new desirable characteristic features. And then, the platform is really great for that because it's not hard to get 384 data points. You can do that multiple times very quickly, so within a week we could have thousands of data points and make smarter decisions 'cause you have more data available to you. And then the final bit... This is a bit of an orthogonal approach.
0:34:31.6 JG: That is our orthogonal technology, but in addition to engineering antibodies, my team does computational engineering of antigens. And the purpose of this work is to try to see, can we engineer better types of vaccine components. We just won a patent recently on an epitope focusing technology, and we've got some very exciting data from immunizing an epitope focusing Influenza vaccine in pigs. That's great. Other people are working on trying to make epitope focusing technologies work, but if you really would like to know whether your vaccine is successfully focusing epitopes, you damn well better have a good technology to tell what epitopes are being targeted by your animals.
0:35:09.2 JG: And again, the Carterra platform for me is game changing here because it lets me look at hundreds of clones from the pigs that receive my vaccine versus a control vaccine to try to figure out what is the bias of epitopes being targeted in a monoclonal level that are being elicited by my vaccine. So I could get a much higher resolution answer than just, do I neutralize against the following viruses. I can ask, is the mechanism true? And I think this is an application area that has been under-utilized and it's critically needed as we make substantive advances in vaccine science to be able to understand it at a molecular level, where our immune systems is focusing on an immunogen and what is that? How has that changed as we engineer the immunogen to try to coerce it to focus on certain sites. Shown here in red is the conserve site on the Influenza viral code protein, hemagglutinin, it doesn't change every year, and that's really... Part of our goal is to make a vaccine that could get the immune system to focus on the sites that don't change every year so that you have a single vaccine that protects you for five years. Like a Tetanus shot. Alright, that's it. So this is our team at Distributed Bio. We'd really like to thank Carterra and at this point, I'm gonna wrap up and hand over to Yasmina to describe a bit about how their technology operates. At the end, I can answer some questions.
0:36:22.9 Yasmina Noubia Abdiche: Hi, everyone. Thanks so much to Jake for that amazing presentation and introducing our collaboration together, which is something that we're extremely excited about. I'm just gonna follow on now. My name is Yasmina Abdiche. I'm the CSO at Carterra. I'm gonna give you an overview beyond kinetics of what our platform can do to really help accelerate your antibody discovery. A little bit about the platform. The LSA, really, that means the Lodestar Array, and the name is really about helping you to navigate through all of your antibodies to really allow you to understand which ones have key properties that are interesting to follow up with. And so the premise of the LSA is array SPR integrated with slow printing. And so we have a system whereby you can either print or array up to 384 different clones. Let's say we have a 96-channel print head that can deposit a whole batch of clones in one print, we can nest up to four prints to create a 384 array, and then we can toggle to a single flow cell system where you can come in with your antigen, and that can flow across the whole surface. And so you can have a one-on-many type essay, which really speeds up the screening process.
0:37:58.7 YA: And so our platform is SPR-based and it will support both capture formats and standard mean coupling. So there's three applications that we really focus on and this one-on-many format is really well suited to kinetics, epitope binning and epitope mapping. So in the kinetic format, as we heard from Jake, that assay, we had actually performed on periplasmic extracts, and these were a single chain of seeds engineered with the V5 tag, and for the capture there, we would have used an anti-V5 lawn on the surface and capture out the SCFEs, using our print head. And the really cool thing about the print head is it really focuses on to a spot, a back and forth mobility of the SCFE allows you to really enrich from a very, very low concentration sample. And I saw one of the questions someone was asking about why the kinetics were drifting up. And this is really just an artifact of the anti-V5 capture surface that we were using, so the reference subtraction for that particular surface wasn't optimal. But actually we've got much better at using the anti-V5 since only a few months of work with Jake. So the next application is epitope binning, and on our array system, we can do a 384 comprehensive matrix of pairwise antibody-antibody competitions. This allows you to really build up a picture very, very quickly of where your epitopes lie within the library, allows you to really survey the landscape within your library.
0:39:56.6 YA: And another application is mapping. And this is where you have an array of peptides that you can put onto the surface. And this could be, for example, a peptide library. And then you can blast through the antibodies in a single flow cell configuration, and then you can pick up linear epitopes and really map where those antibodies are binding. So, with all these applications, together they provide a real comprehensive characterization of your antibody library.
0:40:26.7 YA: So as Jake was mentioning, the epitope focusing that he was mentioning with the flu vaccine, that is something that we're really interested in pursuing with our technology as well. So typically, we have been really targeting therapeutic antibody drug discovery, where the antibodies themselves are the therapeutic entity. But antibody characterization is also extremely useful when you're using antibodies as probes, as reagents to really survey an antigenic response to a pathogen. In this regard, you can use the antibody information that you'll get from, say, vaccination of pigs, and to better understand what epitopes are being targeted, and then you can use that to inform better designs of vaccines. So really it's the same assays that we would use if we were characterizing antibodies where the antibody itself is a therapeutic entity, but this is just another really cool application of our technology to inform vaccine design.
0:41:31.5 YA: So I'm gonna just show you now an example dataset where we can collect really high-quality kinetics on an array of 384 antibodies. In this particular example, I'm gonna show antibodies that were captured via an anti-human Fc lawn with the captures much more stable. So the workflow for this would be to print out the antibodies from a supernatant, they don't have to be purified because the anti-human Fc is itself an inline purification step. And then once you have spotted all of the antibodies onto the surface, in the single flow cell, you'd come in with the antigen a different concentration, then you could get in parallel really nice kinetic data on all 384 spots. Really cool thing about this is that you use very, very little amount of both your antibodies and your antigen. So in this particular experiment, we were using less than 0.1 micrograms of each antibody, and for the entire assay we used about 2 micrograms of antigen.
0:42:43.5 YA: This is what the data looks like from a single unattended run. And all these data were collected in parallel, so every panel that you see here is a unique spot on the array with an antibody on it. In this particular example, we had fewer than 384 different antibodies, which allowed us to array replicates of the same antibody, and then assess spot-to-spot reproducibility also to explore different surface capacity. And I just wanted to highlight here something in our software is that the software very, very quickly can parse out all this data, and identify the good fit, the bad fit, and the ugly fit. And you can see those grayed out or purple sensorgrams or yellow ones, which were the data wasn't good enough quality to justify a kinetic analysis.
0:43:37.8 YA: Now kind of a little bit more up close and personal with the sensorgrams we can see that because we had replicates of given clones across the array, you can see spot-to-spot, you're getting really excellent reproducibility. And this allows you within a single experiment to actually report statistics for your kinetics as well, so this really allows you to have a little bit more confidence in the apparent kinetic rate constants you're reporting. So, just a few more sensorgrams just to show you that we really were able to characterize a really wide range of affinities across a single array. And like I said, this was just basically one experiment, very, very simple to set up. So another application I wanted to touch upon is this mapping, where you have peptides arrayed on the surface and you come in with the antibodies. And again, our software is really elegant in that it can very, very quickly analyze these data, come up with the heat map, allow you to present the data in either dendrogram format or stacked plot format, allowing you very visually to see how the antibodies are clustering into different epitope groups.
0:44:52.4 YA: This is the data from a single experiment, where we actually had about 100 analyte antibodies, and then we tested them over a 384 peptide array. You can see how the blocks kind of parse out, and these are representatives of the different epitopes that are being mapped within this set. So, the last application I wanted to talk about here is epitope binning. And this is an assay whereby you take two antibodies, and you test them for their simultaneous binding to their specific antigen. And you ask the question, can they bind at the same time or it's one blocking the binding of the other. If they bind at the same time, you infer that they're binding non-overlapping epitopes on the specific antigen. If they cannot bind at the same time, you infer that their epitopes are interfering with one another. And so this scale is geometric, see we have a lot of antibodies and we're able to do this in a 384 format. Again, our software is really sophisticated and allows us to very quickly get to the heat map, where you can see the green block in the middle is the establishing interactions and the red.
0:46:09.0 YA: The diagonal shows the blocking interaction, we then also showing that same information in the terms of a network plot, where the different bins are clustered together. And we find the network plots to be really intuitive for people. This is an example of 192 mAb array binning that was just published recently by Ching et al, and what was really cool about this dataset is that if you layer in other information from other assays, you can color the network plots by different parameters, and this allows you to keep the conversation still very epitope-centric, but layer in other information. If you wanted to understand your mouse cross reactivity, for example, you can see which clusters are showing both human and mouse cross and which ones are just human-specific, and say, for example, you had other mapping data with subdomains, you can see if they're lining up with the subdomains.
0:47:16.7 YA: So one more example of how adding in orthogonal data really allows you to really see more information in your panel of antibodies while keeping a very epitope-centric view of what's going on. I wanted to highlight this paper, In Nature by Pfizer Group, authored by Andy Young, and it was very interesting here because we basically did a very agnostic epitope binning experiment several years ago, and we had information of where these antibodies lay in an epitope landscape manner, let's say, and we noticed that when we layered in other data, that we could see that two of the bins were neutralizing bins in the cell-based assay. And then we also could layer in information. These were from B cell donors. But each of these bins were actually contributed by actual different people. And then if we looked further, we could see that there was a real germline bias within each of these bins.
0:48:26.3 YA: If we looked at mutagenesis mapping, we could see that these bins were actually targeting different parts of the protein. So overall what we concluded is that on that project for... On this project, we concluded that the two different bins were actually doing a neutralization by a different mechanism of action. We had crystal data to actually support that conclusion. So really, the epitope binning had really helped to guide where the resources were being put for this project. So in summary, I just wanted to say that, hopefully, I've given you an overview of really the power that the LSA platform can provide in helping you accelerate your antibody discovery through a variety of different applications, and our hardware is also supported by really very powerful software to allow you to see what's going on very quickly. So I wanted to acknowledge our great collaborators. Jake at Distributed Bio, Andy at Pfizer, Yingda at Adimab who provided some samples for our reagent panel, and also the folks at Ligand Pharmaceuticals, formally Crystal Bioscience. And then I just wanted to end with any questions anyone has. We are happy to answer them. Thanks for your attention.
0:49:49.3 Elizabeth: Thank you very much, Yasmina and Jacob. We have had quite a few questions come in. The first question, this one's for Jacob, how are the PPE normalized before putting on the Carterra, and is it done with a single concentration?
0:50:06.1 JG: That's one of the things that's pretty convenient about this process, we don't have to do that. We take the colonies... We pick colonies, we put them into a 1ml induction culture. Our vector is set up such that we have an amber stop that allows our single chain at B to be displayed on phage, but also to be secreted, and it has a mech and a V5 tag on the single chain at B. So that gets secreted into the PVE, we do a PPE extract. And then that PPE extract is gonna be relatively similar concentrations, but there will be variation from well to well due to expression variation of the single chain at B, and then also just settle effects on the 96 well plate we often observe, but the materials present, we send it over to the Carterra instrument, and what happens is the single chain of Bs are captured on the plate or the wafer using the mech or the V5 tag, and then the analytes are flown over of a known concentration. Yasmina, would you like to add anything to that? But the short answer is, it's very easy. We don't have to do a bunch of careful normalization, and we're able to work with relatively low concentrations of soluble single chain of Bs without having to do any complex reformatting or any irritating work right upfront.
0:51:21.2 YA: Yeah, basically as Jake mentioned, the scFvs have been engineered to have a V5 tag, and so we would lawn that onto the surface using anti-V5 antibody to, in essence, inline purify the scFv from the periplasmic extract. And using the continuous flow micro-spotting technology, this allows the periplasmic extract in a compliant volume of, say 100 microliters, to go back and forth over its spot and really enrich on that anti-V5 capture. So we're able to work with very, very low concentrations, probably lower than 0.1 micrograms per mil, get enough scFv onto that spot and do 384 independent spots and then come in with the analyte at known concentrations.
0:52:14.8 Elizabeth: Thank you both very much. This next question is for you, Yasmina. Does array SDR have more sensitivity than SPR?
0:52:24.4 YA: The SDR is SPR basically. [chuckle] So the array, it depends really who you talk to, whether you're talking to a physicist or an optics person, the array is really more about how the image is collected using a CCD camera. So what we find is that with the array SPR, we're able to get a sensitivity plus or minus one response units or so and with the applications that we're working with right now which is not small molecules and we still have to see what our sensitivity is with small molecules, we're actually getting good enough sensitivity to do all the applications pertinent to antibody drug discovery.
0:53:10.2 Elizabeth: Alright. And we've had several questions come in about supernatants. My next one is, have you tried to screen parental hybridoma supernatant for yeast cultures?
0:53:20.8 YA: If this one is for me, hybridomas, we routinely screen the hybridomas using the Anti-Mouse Fc capture. As long as there's a capture system that we can use and that the supernatants are provided in a filtered format, they are definitely possible to capture onto a surface.
0:53:43.9 Elizabeth: Alright, and a follow-up question for Jake. Can you use phage supernatants instead of scFv PPE running array SPR?
0:53:54.1 JG: Sure. So we haven't done that but we should be able to. We have routinely used in the past, prior to creating the Amber stop, we've used ElIZAs that essentially captured by the tag and held the entire phage particle rather than just a single-chain Fv. So Yasmina, correct me on this if I'm wrong, but I think the same principles would hold true. You'd flow the particles, they would get stuck to the chip and then again you'd flow the analyte next, so I don't see why there'd be a mechanistic reason why you couldn't do that.
0:54:20.0 YA: Really just would depend how big the particle is. If it's huge, then we would just be limited by the numbers that could physically actually get onto the spot, but if they are regular SCFE size or antibody size, then that wouldn't be an issue, but if they were cellular size, that would definitely be an issue.
0:54:46.8 JG: Yeah, so these would be full viral part, they're basically asking if they can take a phage display library that has a single-chain Fv attached to P3 and the entire phage particle and it would still be captured by some sort of tag but the whole particle now would be on the chip.
0:55:05.5 YA: I think that would be rather challenging and we haven't tried that, and I think that what you would find is that the phagement may be too large and not compatible with the scale of SPR.
0:55:19.6 Elizabeth: All right.
0:55:20.1 JG: Good. I'm glad we have the Amber stop then.
0:55:23.3 Elizabeth: Okay, we have time for just one more question. We have many more questions that I'll be forwarding on to our presenters for answers via email, but our last question, what tags are compatible with the plates used in LSA? And this is for Yasmina.
0:55:38.4 YA: Pretty much any chemistry can be used as you would on any BLI ElIZA or SPR instrument. So the V5 tag was engineered into the SCFE library because it wasn't present in the antigens. So one could use say his tag, but then that would also be present in the antigens. So the choice of tag, oftentimes, is driven by expression needs but also you have to just be cognizant that it's not in the analyte that you want to analyze later.
0:56:15.5 Elizabeth: Alright. Perfect. Thank you so much. I'd like to thank you again, Dr. Yasmina Noubia Abdiche of Carterra and Dr. Jacob Glanville of Distributed Bio for your presentations today. Most of all, I'd like to thank those of you who came and spent some time with us. We know you've got lots of places to be and lots of things to do so we're very grateful you chose to spend this time with us. So on behalf of Cambridge Health Tech Institute global web symposia series, I'd like to say thank you so very much and have a great