High resolution epitope binning informs antibody drug discovery and enhances next generation antibody workflows

0:00:15.8 Daniel Bedinger: All right, well, thank you all for sticking it out to the last talk of the day, and thanks FairJourney Biologics for inviting me and Carterra to be part of this event. We really appreciate it. So my goal for today and this talk is to introduce the idea that high-throughput epitope binning early in antibody discovery process is both possible, a value-added component, and something that we believe is well on its way to becoming a standard aspect of probably the majority of antibody discovery efforts going forward.

0:00:52.7 DB: So if you go back, say, 10 years ago and look at who was doing competition-based fully symmetric pairwise SPR epitope binnings at a scale of about maybe 30 by 30 or larger, probably 50% of those people are in this room right now. So it wasn't a broadly adopted thing. And it's still, so over the last decade, there's been some technologies that have made this possible, and it started to be introduced. But really in 2018, when Carterra started installing LSAs in the field, which is our high-throughput array-based SPR system, I think this technology or approach has really become more widespread.

0:01:37.3 DB: So one of the things I'm going to do early in this talk is show a number of published examples in the last few years of people that have integrated this into their workflows and how it's impacted their antibody discovery efforts.

0:01:50.0 DB: So obviously, if you start something in 2018, you know what happens in the end of 2019 with the SARS-CoV-2 or COVID-19 pandemic arises, and as I'm sure you're well aware, there was a massive pivot in the antibody discovery area towards making antibodies against SARS-CoV-2. So most of the publications that are out on the LSA are antibodies to SARS-CoV-2. But we'll go through a few examples, highlight some of those aspects. Also we're going to talk about a paper that I was involved in with the CoVIC Consortium, which was looking at a nonprofit collaboration of a global consortium study to make antibodies against COVID-19 and evaluate them in a shared mode where we are all blinded and analyzing all the antibodies through a number of assays to pick which were the best ones, which were most likely to be broadly neutralizing cocktails, et cetera.

0:02:50.1 DB: And also we'll point out now that high throughput kinetics and epitope binning with the LSA is something that's offered through Fair Journey Biologics. They were an early adopter of the platform. And so as you get antibody discovery projects through them, that's something they can include. So the first paper I'll highlight here is a science translational medicine paper.

0:03:15.6 DB: And this was a collaboration between Eli Lilly and Abcellera. So basically one of the first patients in Canada that recovered from SARS-CoV-2, Abcellera took their blood, identified spike-binding B cells, and cloned many of those and started expressing them as constructs. And then so both Lilly and Abcellera have LSAs, so they were able to sort of unite on this effort. And in about seven days, they screened 187 antibodies for affinity to RBD and spike. They did ACE2 competition and neutralization in vitro studies, and then did full pairwise competition-based epitope binning for these.

0:04:02.4 DB: So they were able to take 24 clones that had neutralization, high affinity, and epitope diversity into their sort of next downstream processes, which was a simultaneous attack on manufacturability, cell-based activity, some scale-up into early animal models, et cetera, so that 90 days later, they filed an IND, and four days later were in patients. So this is kind of a big record for, I think, biologics discovery in terms of speed and efficiency. And we feel like the LSA and the high-throughput SPR platform really helped enable that great achievement.

0:04:41.7 DB: So the next paper I want to mention is a group that was at NIAID, which is part of the US NIH. This is Josh Tan's group. And they developed bispecific DVD-IGG or DVD-Ig molecules to target the SARS-CoV-2 RBD molecule. So these are two independent binning sites on the same molecule to make an antibody that was extremely potent neutralizer and high affinity through avidity to that. And also, there was an element of mutational resistance that comes along with that, because each of those binding domains that they targeted with the bispecific had some element of neutralization on its own.

0:05:26.0 DB: So they gave them a little bit more leeway. And the antibody is potent at neutralizing all of the SARS-CoV-2 strains through at least Delta, which was the state of the art, I guess, when this paper was published. Also, I don't have a slide on it, but this group just last month published another paper in Science that is actually really fascinating. They looked at coronavirus broadly neutralizing antibodies, so antibodies that bound and neutralized all seven of the human infectious coronaviruses and were able to find antibodies, a couple of them that bind to the splice domain regions that's broadly conserved.

0:06:07.2 DB: So they have an antibody, or I think two antibodies, that will neutralize all known infectious human coronaviruses and all COVID-19 variants. So that study, I didn't include it, it doesn't have a lot of epitope binning. It actually does have another application we use, which is called peptide mapping, where you can look at the subpeptides recognized on the probe. So they're great LSA users and publishing broadly on it.

0:06:36.3 DB: So this Cell Reports paper is another paper from the Lilly Abcellera collaboration. So I think it was obvious to everyone relatively early on that the COVID-19 was going to continue evolving and bringing out new mutations. So they continued screening patient blood and developing new antibodies until they discovered the, it's hard to say, bebtelovimab. And this antibody binds to a highly conserved region sort of on the lower end of the ACE2 binding domain, and it neutralizes all known, currently known COVID-19 variants. So this is actually currently available in the US on emergency use authorization since February.

0:07:25.8 DB: I'm afraid it's not available in Europe. This is a great tool in the toolbox to fight COVID-19. And again, it was part of their similar workflow to the previous one where they did kinetics, neutralization, epitope binning in the screening of these clones. So two more examples here. The first one was a study at Twist where they were looking at four libraries. Two of them were human IgG, two of them were VHH. They tried to find antibodies that were broadly neutralizing to SARS-CoV-2.

0:08:01.1 DB: Things they found from the epitope binning studies is that they had a couple of rare NTD binders, which were quite good, at least at neutralizing the variants they had at the time. Also, they were able to use the epitope binning data to find noncompetitive clones to use in a cocktail that was actually very potent in rodent models of sort of reversing COVID-19 infection, so rodents that already had high viral loads. They were able to increase their fitness and survival, prevent weight loss, et cetera.

0:08:37.8 DB: Another example is in the antibody therapeutics. This was at Varis, which I guess is now also part of Twist, but I think this was kind of an independent work at the time. And they were comparing antibodies made from two different methods. So it was a B-cell-based identification method using, I believe, the Alloy mouse, and then a hybridoma method using, I believe, the DiversimAb Mouse. So two different transgenic animals. They compared the neutralizers that they generated from both campaigns in epitope competition experience and found that they actually had quite...

0:09:14.0 DB: Oops. I don't know if this works. Well, anyway, quite... I'll figure it out later. Quite distinct binding responses between those two animals. So I think that probably goes to show if you're going to use transgenic mice for your antibody discovery efforts, you might want to try more than one flavor.

0:09:34.5 DB: So the last COVID-19 paper I'll mention is this science paper. So this is one that I was a part of. And I'll talk more... I actually have multiple slides on this later on. But I think this is a really interesting paper and that it's probably the richest example I've ever seen published that compares high throughput epitope binning data with structural data. The team at La Jolla Institute for Immunology has a cryo-EM facility. And I believe by the end, they've done over 40 solutions of structures of monoclonal binding to the COVID spike. So it's a really interesting mapping of what competition-based epitope binding data looks like on a structural aspect.

0:10:20.0 DB: Okay, and it's not entirely COVID-19. So this is an earlier study that was going on sort of shortly before the pandemic at Twist with collaboration from Yasmina, where they took a large number of published antibody sequences from recovered Ebola survivors. And they wanted to understand sort of the epitope diversity within that panel. And so they did a large scale epitope binning. They also included some already published known structure antibodies. So they were able to map a large portion of the patient or the person's immune epitope space physically on the molecule based on the competition studies. Also they were able to identify some antibodies that had unique epitopes that had not been sort of mapped or crystallized previously.

0:11:16.6 DB: So those were very interesting. And obviously if you had this kind of data and you were going to go forward, especially in a rapid fashion, to analyze cocktails, things like that, you could pick very easily which clones are likely to play well together for your in-vivo models. Two examples from bacterial proteins. The first one was a study at Pfizer. They were including epitope binning and affinity characterization in antibody discovery efforts. I think the interesting punchline from this is, so this was antibodies generated against staphylococcus aureus, which is both a commensal and a pathogen.

0:11:58.8 DB: It's something that's been around with people for a long time. And in the four donors of human antibody sequences that they had, they found that all four donors had the same two antibody sequences. And when they analyzed those, they could take the germline sequences, it was a human heavy chain CDR2 domain, but the germline sequence conferred relatively high affinity and neutralizing binding to two different epitopes of the, it was the ISDP protein, which is a iron transport protein on the surface of gram, many gram negative bacteria. So this I think represents sort of a convergent or a parallel evolution of humans and our commensal bacteria where in your genome you carry two antibody sequences that neutralize two different domains on this common gram negative bacterial code. So I thought that was just very interesting.

0:13:01.1 DB: The last paper I'll mention here is one from Genentech published in eLife. And this was, they were looking at this LPTD protein, which is an outer membrane protein on E. Coli, and it loads lipopolysaccharides into the membrane, so it's sort of a structural building block protein. They made about 3,000 clones and used some deletion mutants and these shared competition profiles to map which ECL loops all the antibodies bound to or interacted with. And what was really interesting in this is they covered most of the ECL loops, except really two, two and ten, which were the two that are fundamental for the activity of the protein.

0:13:51.9 DB: So kind of the hypothesis is there that the evolutionarily the bacteria had developed decoy loops that give the protection to antibodies that allow them to grow in mammals, and that there was no, in the 3,000 antibodies they found using multiple different immunogen strategies, no antibodies bound to the functional ECL loops, which is two of 13. So that was really interesting too, and Genentech's been heavily into high-throughput epitope binning for a number of years.

0:14:30.6 DB: So that was the brief literature introduction here, so I have a few other topics I want to talk about. So I want to briefly introduce what is the Carterra LSA, why are reasons why you would want to perform high-resolution epitope binning, what is our assay strategy for addressing that, and then I have a few slides about the CoVIC consortium study in the science paper that I was involved in at the end. The LSA is an SPR-based array instrument, and we have sort of an ecosystem, we have our software, the control software, analysis software called Kinetics, and an analysis software called Epitope, relatively straightforward what they do, and then we have a whole line of chips and consumables.

0:15:13.7 DB: So the LSA works by having a sensor chip sort of in the middle, and a robotic fluidic system that can come on and off that has a 96-channel flow cell that can be used to create arrays of up to 384 MAPs, so four independent captures or mobilizations from that 96-channel flow cell, and then it has a single flow cell that can dock over that same array area that was created and flow one sample to do what we call one-on-many.

0:15:43.5 DB: So this is parallel analysis, one sample volume gives you kinetic or competition data to everything in the array. So this is what makes it sort of very high throughput and very sample efficient. So this platform and architecture is really ideally suited for some of the core applications in antibody characterization, especially kinetics and affinity analysis, this competition-based pairwise epitope binning, mutant or peptide mapping, and quantitation. And so in addition to having sort of a hardware that is very efficient and scalable for these applications, we put a lot of effort into having analysis software that makes analyzing these large data sets very rapid, but also gives you really visually rich and useful tools for interrogating the data.

0:16:33.3 DB: So real quick on just the kinetic side, you know, we think high-throughput kinetics should be accessible. This is a run that was set up in the afternoon, run overnight, used seven micrograms of antigen, this was PD-1, so this is a full starting at one micromolar concentration series, seven dilutions, antibody two, 384 interactions. You can see there's actually probably, I think, 38 clones in this set. There are eight to 12 replicates each in this experiment.

0:17:02.1 DB: A more specific example, that was PD-1, antibodies binding PD-1 on the previous slide. This is some data generated by Denisa Foster at Lilly, and this was, they had taken the SARS-CoV-2 RBD protein and mutated in 96 different single amino acid mutations and expressed them as supernatants from 293 cells in small-scale culture. So in a day they were able to capture them all to the array, I think this was his tag capture, and then inject a concentration series of fabs over them and get full kinetic profiles to all of the mutants. So you get a good map of sort of mutational sensitivity of your antibodies of interest to that. They did it to multiple clones. This is the bamlanivimab, I can't say that either, the first therapeutic they made is mAb.

0:18:00.7 DB: So why competitive epitope binning? So obviously the function of an antibody is linked to its epitope, and we all know we can pretty readily address, assign, alter the affinity, but really that epitope is the innate property. And like I say, it can't be engineered, at least not from an existing clone very easily. It must be screened or selected. Obviously there's a whole AI, ML aspect to that where this may be changing in the ability to design these proteins that bind to specific epitopes may be increasing, but that's okay.

0:18:34.0 DB: We love the AI, ML customers because they're some of our best because they need a lot of data to train models for things like affinity predictions and things. So the LSA is probably the best way to get to that endpoint too, but nonetheless. So early epitope characterization can serve as a surrogate for functional diversity. Sequence data by itself doesn't necessarily ensure broad epitope coverage, but if you have a demonstration that you have many independent epitopes recognized on a protein and a panel of antibodies, you can assume that you have quite a bit of good epitope diversity and you can ensure that you move clones with epitopic and functional diversity forward in your funnel.

0:19:20.2 DB: Doing this type of mapping, but especially if you couple kinetics and high throughput epitope binning and you have a large sequence set, you can use that to infer some activity, at least to some extent, on your database. Especially things like if you scrape a panning and phage display and do next gen sequencing, you end up with a lot of closely related clones or shared CDR sequences. You can use an understanding of epitope recognition and infinity potentially to inform the understanding of the whole set.

0:19:52.9 DB: And then lastly, there can be some IP implications. There's a lot of especially older antibody claims where they claim an epitope or competition profile. The more clones you include in a competition experiment, the more likely you are to differentiate things. So you may be able to demonstrate that some two clones that block each other actually bind to slightly different epitopes at least and are differentiated that way. So the larger your binning assay is, the more resolution you have in it. If you have a 96 by 96 competition profile, you have 9,000 interactions to interrogate. If you did a 384 by 384, it's almost 150,000 interactions to look at.

0:20:35.8 DB: And each unique interaction can be thought of as a probe. It's a point you're using to interrogate the behavior of those clones. And so the more clones you include and the more diverse they are, the more complete of a map you get. And this is one of the things that's really interesting about the COVIC study is that it was a very diverse panel of antibodies. And so the epitope resolution from the resultant map was pretty exquisite.

0:21:04.0 DB: We think that LSA is the only platform, definitely the only SPR-based platform that will allow you to scale these assays in a linear fashion. So on the LSA, if you want to go from 95 to 96 clones, it's just one more well on a ligand plate, one more well on an analyte plate. And it doesn't scale geometrically in complexity like it would on most other platforms.

0:21:29.1 DB: So the way we typically run these assays is we mobilize an array of clones. We inject the antigen, then we inject a sandwiching antibody. Or if it's a multivalent target, you would inject the antigen premixed with the competitor clone in series. Then the analysis software has tools for setting normalization points, report points, cutoffs. It will then sort them and develop these competition matrix, we call them heat maps, and build visualization and network plots. So the network plot on the left is what we call the Bin-Level network plot. So every clone is a node, a line connecting them means they're competitive with each other. If they're covered within a shaded circle, it means that they're in the same bin and they have the exact same competition profile in that heat map.

0:22:17.3 DB: The software will also build these dendrograms, which show the degree of shared competition relationships between clones. So the more differences there are, the higher the branching point. You can set a manual cutoff and build what we call community plots. These essentially forgive degrees of difference between clones and allow you to cluster clones that are binding to say similar faces of a molecule or have shared competition profiles together. So now to talk briefly about this CoVIC study.

0:22:46.1 DB: So the CoVIC consortium or the Coronavirus Immunotherapy Consortium was set up by Erica Ollmann Saphire at LJI and funded by the Bill and Melinda Gates Foundation, Wellcome, Mastercard, and a few other foundations. And the idea was that contributors would give antibodies to the effort. They would be anonymized by LJI and distributed to partner labs to do analysis. So two of the partner labs, Carterra, so my own lab, and then Duke had LSAs and were doing competition-based epitope binning, kinetics, and ACE2 competition. And then there were several other labs doing viral neutralization, like Ragon Institute was doing FC receptor functional assays. They're doing escape resistance. And then LJI itself was doing the cryo EM work on a representative subset of clones.

0:23:45.9 DB: So we found that... So by the end of the consortium, well, I don't know if it's the end, but current status is there's about 370 clones included. 75% of to bind to the... I don't know how to get this to work. The RBD, the receptor binding domain of the spike protein, and there's also a handful of NTD binders and ones that only bind to intact spike trimer. About 65% of the antibodies came from industry labs.

0:24:20.7 DB: There was a subset from academia and a small percentage from government too. So the Tomaras lab at Duke with Moses and Jonathan did the kinetics, so they compared the RBD and the spike. Obviously, we had broad kinetic diversity to the RBD within the panel, and everything looks much higher affinity to the spike because it's trimeric and everything was binding with avidity, but at least they binded to both. This is a representative competition binning assay that I did sort of near the end of the effort. This is 170 ligands. So every row is a ligand, a mobilized antibody, and every column is an injected antibody. And so what you can see is that the binning map is highly symmetrical. We're getting good recapitulation of competition profiles in both orientations. All of the clones are self-blocking, so none of the clones included in this analysis were self-sandwiching.

0:25:30.1 DB: And you can see distinct clusters of behavior. So these were what was used to establish these communities. So highlighted sort of in both orientations is a set of clones that are part of one what we call community. And you can see they're all competitive with each other. They share many other relationships both with sandwiching and blocking. And if we were to move over one bar on the dendrogram, we would get highlighted this group. And you can see it's competitive with the other group. So they block most of the same things, and they sandwich with most of the same things. So the epitope is not totally independent. It's closely related, but it's also well differentiated.

0:26:10.1 DB: We have a small cluster, sort of the yellow and the blue community that sandwich with it, where the other one's competitive. And that's true in both orientations. This is an independent community that's distinct from the other one.

0:26:23.9 DB: And so when the dendrogram, and this is the dendrogram from the paper. So when the science paper published, we were about 100 clones deep on the RBD binding side into this. And it's grown since then. The work kind of continued over the next year. We divided it into seven super communities, and really about 13 kind of more nuanced communities that are subgrouped here. And we were able to sample pretty well into each of these communities and generate the cryoEM structure data. So there's sort of three different rotations of the RBD shown below. The dotted region is the ACE2 binding domain. So you can see we have inner and outer face binders, clones that bind sort of all over the RBD interface shifting in various ways.

0:27:21.1 DB: So you can kind of map these profiles onto the RBD. This would be really useful if you wanted to come up with a cocktail. You can know right away which clones are likely to play well with each other and which ones would be competitive with each other.

0:27:37.0 DB: So we use this to look in the context of mutation resistance. So we found that at least through mu, so alpha, beta, delta, and mu, the branches of the dendrogram were usually very predictive of sensitivity to mutation. And so there were several quite resistant mutational clusters in the set. Then omicron happens. So we have the lower portion is through mu, and then the omicrons are at the bottom. The mutational space just is way different. There's spread out throughout the entire RBD a number of mutations, additional deletions. It was a major change.

0:28:25.0 DB: And so this sort of made the mapping of which regions were less resistant to mutation more of a challenge. We dropped from a 23% pan neutralizers with delta and mu to a 5% pan neutralizer, which is still, I guess, the good thing about this is 5% of the antibodies still neutralized all of the strains that were identified in this pseudovirus assay. But they are kind of scattered around. I think the specific contact residues that determined how each antibody docked onto the RBD became much more important than sort of the face because there was mutations everywhere at this point. So yeah, so I can basically end here.

0:29:16.4 DB: So we think high throughput epitope binning is broadly becoming an integral part of early antibody discovery projects, and we hope this is a trend that continues. Epitope competition data is crucial for selecting antibody cocktails and having a good prediction of things that are likely to play well together with bispecifics, especially in cases where you're targeting multiple things to the same target, like the NIAID example. The LSA allows kinetic and epitope binning analysis at an unrivaled scale with very modest sample requirements. You can use crude samples for most of these applications. And so with that, I just want to thank FairJourney again for the invitation to talk here. Like I said, they're an early adopter. They have the platform. They can use it to support your discovery efforts. So also I'll end with my thank you slide to all the members of the CoVIC consortium and the teams at LJI.