Scientist Stories: Michael Birnbaum, Tools to Decode Adaptive Immunity

Okay, I know many others are probably still joining, but we'll get started as it is 10 o'clock. So hello and thank you to everyone around the world for joining us for this week's Global COVID Lab meeting. My name is Kristin Aboud. I am the science editor at the Human Immunome Project, and I will be your moderator today. As you may have noticed, the Human Vaccines Project has a new name. We are now officially called the Human Immunome Project. But this is not just a name change, rather an important next step in the evolution of the project's mission. This week in La Jolla, California, the Human Immunome Project brought together leading minds in the fields of human immunology and machine learning at a summit to develop a roadmap for generating the first AI model of the human immunome. The overarching goal of this work is to transform how we fight our most devastating diseases by unlocking the mechanisms of human immunity, thereby accelerating the development of new vaccines, diagnostics, and treatments for all people. This new name represents that mission. More about that and the progress toward that goal is to come in future lab meetings. For now, it is my great pleasure to introduce today's speaker, Dr. Michael Birnbaum, an associate professor of biology— I'm sorry, of biological engineering at MIT. Dr. Birnbaum received his bachelor's degree in chemical and physical biology from Harvard University and his PhD from Stanford University before joining the faculties of MIT and the Koch Institute in 2016. During his tenure there, Birnbaum has received numerous awards and fellowships, including a Packard Fellowship in Science and Engineering, a Pew-Stewart Scholarship for Cancer Research, a V Scholar Grant from the Jimmy V Foundation, the Michelson Prize for Human Immunology and Vaccine Research, and the Damon Runyon-Rackleff Innovator Award. In October of 2020, he received a New Innovators Award from the National Institutes of Health. He is also co-founder and advisor of both Colonia Therapeutics and Abata Therapeutics. During today's presentation, please send me your questions using the Q&A function in Zoom. I will ask our speaker a broad selection of your questions after the presentation, and then we will have time for discussion. With that, it is my pleasure to turn it over to Dr. Michael Bernbath.

Great, uh, thank you so much, Kristen, and it's good to see all of you here, and, uh, certainly fun for me to present, especially because of, uh, the role that the, the, uh, Michelson Prize played in developing some of the works that I'm going to be able to talk to you about today. So, so yeah, so my name is Michael, and I run a lab at MIT really focused on the intersection of immunology and, and method and tool development. So some of the work that I'll be showing you is some of the new tools that we've been developing to try to better understand how the adaptive immune system works and recognizes antigens. Where we have some applications in COVID among other diseases. And so here I'll be talking mostly about, I think, tools that could be very interesting for ourselves and others to apply to disease, and a little less about uncovering new learnings about the coronavirus itself, although that certainly is on our mind to do internally in the lab and collaborating. So as I said, my lab works pretty much entirely on trying to better understand antigen receptors, and we'll go through just a little bit to bring everybody onto a common footing what I mean here. But we're really interested primarily in what do T cells recognize during the course of disease, and how could we better predict, understand, or affect outcomes based on, on what T cells are recognizing. We also are working to rewire immune signaling, but today we'll focus on the outside of the cell. And so in order for adaptive immunity to work, really what we're talking about are overlapping repertoires or large collections of molecules that need to be bound and recognized by each other in order to make sure that we're protected from disease. So on one hand, essentially every nucleated cell in your body expresses an MHC molecule, major histocompatibility complex, called HLA in humans. So these serve as windows into a cell in order to understand what's been, what's being made inside of the cell. So in the case of a normal cell, this is just like the proteins that are being made all the time. Upon infection or transformation in cancer, the peptides change and are there for patrol by T cells. So this ends up being actually a very large and stochastic collection of molecules. So you could have on the order of a billion peptides presented per MHC molecule, and then each of us are extremely polymorphic in the MHC, something that our lab is thinking more deeply about as time goes on as well. So these are the molecules that make tissue transplants complicated. So for tissue matching is where these were actually first discovered. And it's also across multiple genes. We all have different proteins that are working here. These MHCs are in turn controlled by T cells, and each T cell is denoted by its canonical T cell receptor. And so we have in us roughly 100 million or so unique TCRs, or T cell receptors, out of a possible 10^15, so incredibly large number for, for what's theoretically possible. So our adaptive immune system controlling infectious disease in terms of CD8 killer T cells coordinating the rest of the immune system in terms of CD4 helper T cells relies upon high fidelity recognition of these antigens by these T cell receptors. This occurs at incredibly low affinity, so 1 to 50 micromolar. So this is about 1,000 to 10,000-fold weaker than an antibody would recognize a target. And in spite of this, is incredibly sensitive. And just in order for the books to balance, so for all of these antigens to be able to be recognized by all of these T cells, you need on the order of 100,000 distinct antigens recognized by each T cell. And you need to do this in a way where this cross-reactivity, so one T cell recognizing multiple potential targets, is helpful in terms of recognizing, for example, multiple viral variants, or at the very least neutral but not harmful, where, so for example, a T cell response against a virally infected cell inadvertently recognizes something expressed in a hepatocyte, for example. So we are deeply focused in trying to understand how these interactions work And we are doing it in order to both understand the biology itself, create novel therapies or interventions, and then also, you know, to Kristen's introduction for kind of where the Foundation is moving next, develop better computational tools. Because in order to develop these tools, we're really limited in our ability still to collect and curate high-throughput interaction data. So just like any other computational problem, in general, the places where we have really incredible advances— so whether it's speech recognition or computational picture generation or wherever it is— is because we have truly huge amounts of data to know what these things look like. Where we still remain relatively limited in terms of the amount of data that we have for some of these interactions relative to the theoretical maxima of what could be out there. So another way of looking at this is that you can imagine a kind of an interaction, kind of each side of this double-sided arrow going from T cells on one end and to expressed genomes or proteomes on the other end. So you might be interested in, for example, if you have a viral sequence, understanding what proteins expressed in that viral sequence are able to be displayed on an MHC molecule. So what are possible antigens that, that could be recognized to begin with? You might be interested in a particularly interesting or expanded T cell population. So this is particularly something that comes up in tumor immunology, for example, where you might have a T cell where we're able to sequence it to figure out the identity of its receptor. And we might want to know, okay, well, is there something that is overrepresented compared to everything else in that particular disease context? And then you might be interested in the interaction between these as well. So this is actually where we spend most of our time. So in our lab, we develop tools to address pretty much every step of this process. And so today I'll tell you about two of them. One that we've been using for quite a while now, basically since I started the lab, and one that we've developed over the past few years and partially with support from the Michaelson Foundation that we think is going to allow us to answer some of these questions in ways that we haven't been able to, to this point. So one of the ways that we've approached this over time is to try to treat some of this as a screening problem or an interaction problem. So the idea being that if you have a molecule and you're not sure about its properties in terms of what other molecules it could bind, its stability, or so on, one way to be able to look at this is to make many, many variants of that molecule and then screen for the properties that you might be interested in, usually combined with some type of next-generation high-throughput sequence. So one of the tools we use for this is peptide-MHC-based yeast display. So what we can do is take the MHC molecules that are on each of our cells, format them to be able to display by a yeast molecule, and then randomize the peptides to create on the order of upwards of 100 million sequences. And then we could screen these. So one thing that we could do is screen them for recognition for a TCR that we might care about, and then use this information for bioinformatic inference of other peptides that might match that pattern. So from this way, for example, you could take a T cell not knowing what it recognizes, figure out what it can recognize in terms of just the biochemistry and biophysics, and then figure out what in a proteome, human proteome, bacterial proteome, whatever it might be, is able to match that. So we're still using and developing these tools in the lab. They have some real benefits of being able to do things at extremely high throughput with as few preconceived notions as possible. So one area that we've been working on over the last couple of years is trying to not just look at how a peptide-MHC complex binds to a T cell receptor, but also to look at how a peptide binds to an MHC molecule itself. And so this is an important problem just because a lot of times when we're relying on figuring out what peptide from a pathogen, for example, might be immunogenic, we rely on computational tools, and these computational tools are getting better all the time. But as I can show you in a few slides, we're still far from perfect in terms of how we would ideally have these tools work. This is particularly the case for class II MHCs, so the molecules that are recognized by helper T cells, which are important for a whole host of things, including making sure that we're able to make high affinity antibodies. And then it's also True that our prediction kind of varies depending on how frequent an allele occurs in a population, particularly in Caucasian populations. So for highly represented alleles, our predictions tend to be better simply because we have more training data than alleles that are either relatively rare worldwide or occur in, you know, basically ethnicities found in lower middle income countries. That are underrepresented in the US and Europe. So we created a modification of our yeast display tool in order to try to address this, in order to empirically figure out what types of peptides are able to bind. So in the first iteration of this, what we did is take a Class II MHC, create a library with upwards of 100 million peptides, and then cleave the linker between the peptide and the MHC molecule and try to compete out peptides from binding. We're able to do this reproducing much of the natural antigen processing machinery, so including chaperones like HLA-DM, changing the pH to be more like an endosome, which is where this process generally occurs, and then we can simply ask which peptides are able to stay bound through this entire process in order to understand the motif in terms of what's allowed to bind any given MHC molecule. So we did this for proof of concept with both a relatively well-studied Class II MHC, HLA-DR401, and a less well-studied MHC, HLA-DR402. And you can see, so for our data here is on the right, some data collected by mass spec, which is another state-of-the-art way of doing this, is on the left. And so you can see here that we have a couple of notable things. So one is that by and large the patterns match pretty well here. So it's not as if our current approaches that people use are totally off, which makes sense given computational prediction and all that are imperfect but not random. But you'll note that if you look closely enough, there are residues in both of these patterns that were completely missed in other approaches. And then also, if you just look at the scale of data produced, so mass spec tends to produce on the order of, you know, a few hundred unique peptides, where we're able to produce on the order of thousands to tens of thousands of unique peptides. So what does this mean in terms of making better predictions and understanding? So we did some experimental validation showing that the patterns that we're finding here are real. So the places where we're seeing things that are being missed by current approaches. It seems to be a false negative for the other approaches rather than a false positive for us. And then if we take this data and try to just generate new training algorithms, or rather retrain old algorithms, what we find is that if you use data generated from mass spec as our benchmark, kind of everything does reasonably well. So I think we're the blue curve here. So we do, we do, you know, among the top in terms of approaches, although there's not a huge amount of differentiation. But when you look at this type of data, which is just predicting how good you are at calling something a positive versus negative, there's a couple of important things that are missing. So one is that if there are positives that are missing from your validation set, you're not going to be able to get credit for calling them. And so if we use a validation set that has some of the peptides in there that were being missed previously, you can see now that we actually do significantly better than current algorithms. And then the other thing is that when you have a plot like this, you have a good measure here of— sorry, I accidentally opened a window. There we go. You have a good measure here of yes versus no, but you don't necessarily have a good measure here for, for good versus bad. And so if we look at how the strength of a prediction for binding correlates with the binding itself, the data from our model still is not perfect, but does much better than, for example, NetMHC, which is kind of the state of the art, which, which we use in the lab and pretty much every lab uses. So you can see here just in terms of the relationship between the predicted binding score versus the actual measured score, there's really big variances. So there's actually not much predictive power there. There could be all sorts of reasons why this is the case, which I'd be happy to answer, you know, talk about. But we think that the types of data that we're producing could be helpful just in terms of making these predictions better. So another thing that we're doing with this type of approach is with advances of how we can actually create oligonucleotide libraries, you can create things that are not any longer, say, 100 million completely random sequences, but on the order of tens of thousands to hundreds of thousands, perhaps even into low millions of explicitly defined sequences. So what we were able to do recently is adapt our approach, and instead of building motifs and then use those to predict peptides, we can very explicitly go through and look at every possible peptide from something that's on the scale of a viral proteome. So in a paper we recently published, we did this across the coronavirus along with every serotype of dengue virus, and I believe Zika, all in the span of one experiment. So an experiment that cost us maybe $3,000 to do. We were able to look at the— on the order of tens of thousands of peptides to very explicitly ask whether each and every peptide moving over 15 amino acids by 1 is able to bind to an MHC molecule. And you're able to produce data that looks something like this, where you end up with islands of binding for molecules. And then through this, you're able to make, now empirically— this is, there's no computation here other than the data analysis— predictions of what's able to bind. So for example, here from coronavirus, we're able to see examples of places where there's a sequence that is highly conserved in SARS-CoV-2 that is divergent in SARS-CoV that's able to bind to multiple MHC molecules. Areas where the viruses are highly conserved between the CoV-1 and CoV-2 but can only bind to one MHC, breaks binding to the others. And then places where there's relatively minor polymorphisms that can induce a specific dropout. So here you can see that there's a peptide sequence that between the viruses is able to bind to everything except HLA-DR401. And then, you know, similar to what I showed you before, there's a whole host of peptides. So here, we're not perfect, algorithms aren't perfect, but there's a whole host of peptides that were predicted not to bind via net MHC, so all these ones in teal, that actually do bind well. So there's the ability right now, if there's an over-reliance on computation rather than empirical study, for missing potential peptide epitopes in terms that could be recognized during the course of a viral infection. So just very quickly, I showed you data for class II MHC, where we're able to take advantage of the fact that the peptide tends to sit linearly in the groove, so both ends of the groove are open, which means that you can easily hang linkers off and do a type of peptide exchange in order to study for whether peptides can bind or not. We weren't sure at first whether this would be able to be adapted to, to class I MHC simply because of differences in how the molecules are structured, But it turns out that what a student in my lab, Patrick, figured out is that you can use the same principle but rely on overall stability of the molecule to find peptides that bind well. And so we've now done this with now many, I think on the order of 15 Class I MHC molecules, where it's the same idea, where you're able to pick out motifs And then here, working in a collaborator in the context of a bacterial infection, are able to make a very explicit identification of individual peptides that can bind across a whole host of MHC molecules. So these are the types of experiments that, you know, that many people, Alex Sette, for example, do on a regular basis, and this isn't necessarily a replacement for these, but rather a supplementation. Just in terms of the amount of data that's able to be generated quickly in terms of this type of peptide binding. Over hundreds of thousands of peptides, 15 alleles is something that we were able to do on the order of, of a couple of months for a graduate student and a research technician working together on this. And we're still digging into this data, but there's areas here as well, where for some alleles, especially ones that are very well studied, the enrichment versus what NetMHC says— here antigen prediction tends to be a bit better for Class I MHCs than Class II MHCs— matches quite well. And then there are some places where, for example, it looks like the algorithms have either, compared to our data, a fair number of false positives for the algorithms, which would manifest itself with things down in this lower left quadrant, and then false negatives, or at least some of these, which would manifest themselves in the upper right quadrant. And so we're still in the process of fully understanding the data that we're generating here and doing validations, but we look forward to having this out for the world to see hopefully by the end of the year. So that's the first half of what I wanted to show you, where we're really trying to dig into being able to use these tools in a way to be as explicitly useful as possible in terms of generating datasets that are either interesting on their own or are the types of things that we can deposit into the immune epitope database, a public resource, both for other people to be able to study or to be able to train the next generation of computational approaches. So in the last 10 minutes or so, 15 minutes, I wanted to tell you about a new approach that we've tried to develop that interconnect repertoires right here. So before, I was basically telling you about something that goes here. We could use yeast display here as well, but we end up with a limitation that we're working very, very hard to get around. So the way that you could think about this in terms of this intermediate, in terms of interrelating peptide MHC and T cell receptors is that different tools are helpful depending on exactly the problem that you're trying to solve. So for example, if you had one specificity that you really cared about, so for example, a known immunodominant epitope from a virus, and you're interested in classifying how many T cells a person has that recognizes them, the phenotypes or so on, then we have tools like peptide MHC tetramers, which have been around for nearly 30 years now that are incredibly helpful for this. If you're interested in the reverse problem, so for example, a T cell that's expanded in a tumor and you want to know what it's able to recognize, there's a host of tools including the yeast display approach, which I was just talking about, as well as other approaches by Steve Elledge, David Baltimore, and others that can basically try to probe one T cell or a small number of T cell specificities across a wide range of antigens. And all of these have different strengths and weaknesses and so on, all quite useful in their own way. But the area that we've been trying to put more effort in over the last few years is this, which is trying to understand immunology closer to how the immunology is actually happening. So we typically are interested not necessarily in one antigen across many T cells, or one T cell across many antigens, but some type of connectivity between all of them. So this is a library versus library type problem. So this is not something that is specific to our lab or even our corner of immunology or immunology more generally, but is the type of interaction networks that we would love to build just across biology. And so we sat down, just basically a blank whiteboard, and this is what we came up with in this way to try to answer this. And so a lot of times answering a new question requires the development of a new tool, and so that's where we have found ourselves. So what we've presupposed is that if we take a virus, for example, it actually has a lot of properties that would make this a really good tool in order to address the problem that I had on my previous slide. So why is that? And this is in particular lentiviruses, which are a very common research tool, basically a re-engineered, very safe version of HIV, of retroviruses, and what we use for gene delivery into cells quite commonly in all of our labs. So, so why a lentivirus, and what do we have in mind? So for one reason, viral infection is governed by specific interactions between the virus and the cell. So for coronavirus, for CoV-2, it's ACE2 binding to the spike protein. For HIV, it's GP120 binding to CD4, which is why it ends up being a disease of T cells. Flu binds sialic acid, and so on. So we have many examples of these, and that tends to be something that we think can be engineerable. We know can be engineerable. Retroviruses, including lentiviruses, can incorporate genetic material into the genome of the cell, which would allow us to keep a permanent record of any interaction that occurs. And we know how to make large lentiviral libraries and do this in a way that's economical and safe. So we and many other labs use these types of libraries for CRISPR screening, for variant testing, and all sorts of other approaches. So what we presupposed is, could we engineer a viral particle in such a way that we can have viral infection rely upon a protein on the surface of the cell— of the virus by recognizing something on the cell, and then engineer the virus to basically use its genome as a record-keeping device in order to say, what interaction on the surface of the virus allowed for the infection to take place. So I'll walk you through this in a few different steps. And this is just here to say that this type of viral surface engineering is not a brand new field. So it's a field that there's been a lot of interest in over the last 15 years or so, with all sorts of different approaches by Christian Buchholz and David Baltimore, among others. In order to engineer viruses that would allow different infectivity. So they all have different pluses and minuses, but for our purposes, we found that we needed something a little more bespoke, so something that would be extremely efficient for the affinities and sensitivities that we would be working on. And so what we came up with is actually to turn to the BSV glycoprotein. So this is how— this is a common workhorse in terms of engineering lentiviruses. So it's the protein that we all typically put on our lentiviruses to make them be able to infect cells, because it recognizes naturally the LDL receptor, which is on most cells. So this is a very robust and efficient means of using lentiviruses. So we took advantage of the fact that out of the Albertini group from France had recently published a paper showing mutations which are able to abolish binding of VSVG to its native protein on, on host cells while still retaining its ability to fuse membranes. So what we presupposed is that if we incorporate those mutations into VSVG and then complement this with the binding of another protein, so in our case, for example, a peptide MHC molecule, we could decouple binding and recognition, or sorry, binding and infection. Where binding is through the protein we're dictating, and then the membrane fusion, and therefore the infectious cycle, being driven by, by the viral fusion. So it turns out this works. So here I'm showing you a virus that has a matched peptide MHC on it. We're able to pretty efficiently infect T cells, where a mismatched MHC, you get little to no infection. So this, to our knowledge, is some of the first data showing that you're able to use the T cell receptor as a viral entry receptor, which is, I think, exciting in its own right. And this works across an entire range of affinities, including the very low affinities in the micromolar range of where natural immunity tends to occur. So we should be sensitive enough to be able to recognize naturally occurring peptide-MHC T cell receptor interactions. And we have specificity even in pools. So So what I'm showing you here is we took on-target T cells, labeled them with a dye just so we could tell them apart from the off-target. And you can see that the infections are kind of things that are moving to the right along the x-axis, is quite specific for the cells that it should be infecting in the upper quadrant compared to the lower quadrant. So we have a specificity by our measure on the order of several hundred to one. So with this, we are able to develop a tool that we call RAPTOR, or receptor-antigen pairing by targeted retroviruses. The idea being that we can mix viral libraries, so with all sorts of different peptide MHCs on their surface, with polyclonal T cells, allow for the molecular interactions between the peptide MHC and the T cell receptor to sort out who binds who, And then if you find a good binder, you can then be internalized and complete an infection cycle, where each of these viruses not only displays a peptide MHC on their surface but encodes a different peptide sequence in their genome. And then from there, via pretty normal sequencing approaches, you can then read out either just the peptide if you're interested in just what infection— what antigens allow for infection, or or the T cell receptor as well. So just through adaptations of pretty normal single-cell sequencing approaches. So just a few proof of concepts here. So the first thing that we did is took a library of about 100 commonly identified antigens, mixed it with monoclonal T cell lines, and then read out via sequencing what is able to bind. And we found here that when we look at this, So blue is the background distribution of the library, and then red is what it looks like when we actually do our infection. So you can see pretty clearly that there's one major hit here. So this is a T cell line called C7, which should recognize a peptide called NLB. So mission accomplished with quite good reproducibility and signal-to-noise. And then we also showed that this works in a proof of concept study in terms of pools. So you can do this where if you look at individual T cell MHC pairings and we read out by each individual T cell clone what peptide is able to recognize it, we're able to find peptide specificities— or sorry, T cell specificities. And then when we look at the enrichment of barcodes for each specificity, we're able to find the right pairs. So this shows at least in proof of concept, and we're still working to kind of tune up and then further apply this tool, that we're able to find these types of pairings. And then, you know, something that we realized is that there's nothing that we're doing here that explicitly is only T cell specific. So the overall theory underpinning our system should be able to work for any protein-protein interaction.. And so in order to demonstrate this, we moved one step over in immunology. So we went from, from T cells to B cells. And in order to demonstrate this, we built what we refer to as hybrid pseudotyped viruses. So we have a virus here that, for example, co-expresses the SARS-CoV spike protein with the— our engineered fusogen. We find that for infecting ACE2-expressing cells, so the normal entry receptor for the coronavirus, this actually ends up being a highly efficient way of being able to infect cells still in ACE2-dependent manner. But then we can find that similarly to how a peptide MHC-displaying virus can use the T-cell receptor as an entry receptor, these hybrid pseudotyped viruses can use the B-cell receptor or surface-bound antibody as an entry receptor as well. So here, just in proof of concept, we did this in a few different ways. So we used, for example, an antibody called CR3022, which is one of the earlier described antibody clones which is able to bind to SARS-CoV-2. It's actually an older antibody, so it had— from SARS-CoV-1 that had cross-reactivity to CoV-2. And we can see that we end up with quite efficient and selective infection of cells in a CoV-2 spike-dependent manner. And then for BRC01, which is a canonical, highly studied cross-reactive or broadly neutralizing antibody rather for HIV recognizing gp120, we can see that we can recognize a host of gp120 variants, including with one that is known to have an affinity-ablating mutation being able to be infected much less potent. So there we now applied the same type of library-based approach to show that we're able to do this as well. So for some of this, it ends up quite straightforward. So when we do VRC01, we created a panel of on the order of 50 different viruses. So here red is various different SARS-CoV-2, or at least SARS-CoV-like SARS-like spike proteins. So this data is generated a few months ago, so we don't have the Omicrons in here, but, but many of the other earlier lineage of the virus, along with a host of other spike proteins from commonly occurring viruses—measles virus, influenza, and so on. And so for BRCA1, you can see that HIV is essentially the only hit. For CR3022, using it the way that, you know, it's sold to you from a vendor, we're able to see that essentially all of the different spike proteins are enriched compared to everything else. You can see that all of the red in terms of the rank order moves before the blues or blacks or greens, including with an interesting hierarchy for the ability for these different spike proteins to be recognized by CR3022. And then I think pretty excitingly for us, if we create the germline-reverted variant of CR3022, so basically recreate what would've been on the original B cell that then went through affinity maturation to create the mature antibody, we still see, you know, less but still significant infection of SARS-CoV-1. For example. And so this shows that we're able to potentially use this approach to identify not just memory B cells that might be expressing the mature version of B cell receptors creating antibodies, but to be able to look at precursors as well, or look at the potential reactivity in people who are unexposed to viruses. So looking at antibody sequences that could occur before any affinity maturation happens, which I think is something that's pretty unique for the tool here. And then just more broadly, so it's a step removed from the actual understanding of the immune system, but there's nothing here that we're talking about that is even particular to antigen receptors. So we have examples of all sorts of different receptor-ligand pairs which allow us to specifically infect immune cells, or in theory any cell, in a way that is specific for a particular cell type. So here showing something where we can specifically infect B cells, specific lineages. So for example, cells that are expressing particular growth factor receptors, we can get good efficient infection in a dependent way, or even to provide multiple signals to the same cell through infection. So these are tools that we're very interested in to allow for selective manipulation of the immune system. So you could imagine, for example, trying to use something like this in order to specifically introduce a broadly neutralizing antibody into B cells via a gene therapy, or otherwise manipulate cells, whether to better study them in terms of doing screening or addition of proteins that could be interesting perturbations, or for therapies overall. So with that, we're still very much in process of figuring out exactly what we're going to be able to do with this tool. And just like any other tool, you can imagine that, you know, you end up with approaches that supplement and expand what we can do rather than completely replace everything else that came before it. So we think that all of the tools that are being developed by us and many others in the field are all going to have their particular niches and applications. But for us, we did here a proof of concept to screen on the order of 100 antigens. I think that we can probably get onto the scale of at least into the thousands with a little bit more development and show that we're able to do this for class 1 MHCs, which I showed you in data that we have but haven't published yet, class 2 MHCs, and B cell receptors. So as a way to try to probe immunity very much in large scale in order to understand how all of these cells are potentially recognizing their antigens, both to understand the immunity and to go further to build some of those computational datasets that I had talked about before. So that— I'd just like to thank the lab, it's been a great team, and all of our collaborators and funders. As I mentioned here, this work was really driven in terms of funding by the Michelson Foundation along with the DOD and the NIH. But with that, I'm happy to take any questions. Fabulous.

Thank you so much, Dr. Birnbaum. That's really interesting and exciting work. So we have one of the first questions was, at what point do viral peptides attach to MHC and then trigger antigen-presenting processes. So the question was, could you briefly explain the sequence of events, if known, of viral replication, antigen presentation by innate and adaptive immune cells, and then the subsequent immune response?

Yeah, so it's a great question and is worthy of an entire lecture or two. So we actually— so the short answer is that we know a fair bit about this in terms of basically how peptides get here, and this changes depending on exactly what MHC you're talking about. So the process for Class I MHCs for recognition by— which are on every cell for recognition by cytotoxic T cells— relies on the proteasomal degradation and processing of antigens which are mostly in the cytosol. So it's kind of like a bypass pathway from the normal turnover of protein degradation. And then for class II MHCs for recognition by helper T cells, this is more a process of endocytosis, so eating from the outside. Where Class II MHCs are mostly on professional antigen-presenting cells, so cells that are coordinating the immune system. So just like everything else in biology, it gets much more complicated than we would like. There's exceptions and exceptions to the exceptions, and how immune cells do it is slightly different than how normal cells do it, and how it happens in infection is slightly different than how it happens during just normal life. But really understanding, you know, I think one important point here in terms of the data that I showed you is that when, when I'm showing you some of this, these sequences here, what we've done is taken a very reductionist approach in terms of looking at this, right? Because we're saying, okay, forget about whether these sequences are actually processed in a way that you actually end up with a peptide of this size, which is very much an open question. If that peptide is processed, will it bind? So So we have this one part of the process that we have very carefully isolated that then has to be fed into the larger composite. It's something that we think a lot about, and I think it's important context for what we're doing. Right.

All right. Complexities of the immune system. We had a question about whether any of this work could be viewed as gain-of-function studies.

Yeah, it's a good question. I don't think so on a couple of different levels. So one, in terms of the viral studies, everything that we're doing is working with self-inactivating lentiviruses. And so, you know, there's no ability for any type of loss of containment or anything like that. As soon as these viruses infect once, they're done, much in the way that the tools that we're all using are. In terms of actually studying the proteins, right, like there is information that we will be potentially generating in terms of, say, you know, what sequences are able to evade an antibody versus not evade an antibody or so on. That is, you know, we haven't generated any of that data, but as we or others potentially generate that, there's questions about, you know, how do you publish that type of data responsibly, which is very much of mind. But I don't think that that's unique to us. I think that's true for anybody who's looking at either naturally occurring variants or doing deep mutagenesis scans of the RBD from the coronavirus, or all of that work is something that we should be very mindful of. Right, right.

Absolutely. So we had a question on whether it's becoming possible to computationally fold a newly seen T cell receptor of interest. And predict the peptide HLA ligand, particularly if you have some prior information such as the HLA restricting allele or the name of the virus being recognized, et cetera? Yeah.

So I will get— the short answer is no. The longer answer is that people tend to look silly for betting against computers given a long enough time frame. So the computational tools are getting better all the time. Where we are now is that if you're studying a particular antigen that has been like very well established, so particular things in HIV or CMV or flu, and you have enough data to compare to, you can potentially use that to match with reasonable accuracy whether the T cell that you're interested in also recognizes blue, for example. So that can be done with some reasonable degree of accuracy. Whether that can apply to, for example, an emerging viral disease where it's something that we've never really seen before, that's where, where we really aren't yet. And the tools are getting better all the time, whether it's machine learning models, or whether it's like direct docking models. Like, I'm sure people have heard of some of the advances with AlphaFold, right? So, you know, even since I've been working, there's been a fair bit that's moved from the didn't seem doable to seems pretty commonplace. And so that, that will always keep going. Yeah. One of the things we need is more data, which is one of the things that we're trying to address.

Absolutely, absolutely. So we had a question on how these approaches and data that you're generating can help us predict better vaccine development by optimizing antigen presentation. Yeah, yeah.

So the, so the answer for this differs. So if, for example, if we all were HLA-DR401, then you can imagine that there could be a lot of ability to say, understand exactly what should be in a target. So you could imagine in order to expand immunity to T cells and beyond antibodies, which the current vaccines largely do as well, but you could say if there's a particular peptide that would be recognized by a lot of people, and it doesn't have to be part of spike, it's part of nucleocapsid, or another viral protein, we could add it, and we could potentially add it in a way that more efficiently can be recognized than a natural sequence. Figuring out how to do that or implement that on a population scale, where the peptides that can bind to my MHCs will look quite different than the ones that can bind to yours, right, gets more complex. And so there's, um, so there's— I, I have, uh, a colleague here with me at MIT who's trying to develop computational approaches in order to do this in a way where you kind of maximize individual benefit but also maximize population benefit. But those are some of the balances that would have to be kept to do it.

Right. Okay. So can the methodology you've described be adapted to NHP peptides? And is this something that's being done already?

So like non-human primate? Yeah. Yeah, so in principle it should be. So we, we haven't adapted it to monkey, but we have adapted it to mouse and works fine there.

So, so no reason it shouldn't be. Okay. And then we had a question on the hybrid lentiviral vectors that express either HIV envelope or the CoV-2 spike. Are the viral— is the VSV still needed in that situation, or is the virus entry mediated by the envelope or the spike protein?

Yeah, so we have— so I— the— we probably could do more in terms of actual, like, very deep kind of like mechanistic virology. But what I could tell you is that the coinclusion of the VSVG in our hands at least makes infection much more efficient than if you just have the CoV spike alone or GP120 or GP120/GP41 alone. So we seem to get a benefit by having it there. And then if we use, for example, the pre-fusion stabilized version of the CoV-2 spike, which should be much less efficient if infectious at all on its own, or just the RBD by itself, we still infect quite efficiently. So I think that that shows that the hybrid pseudotype at the very least makes things much, much more efficient if it's not strictly absolutely 100% indispensable. Okay, right.

So then we also had a question on whether you're looking at the variability from person to person in the T-cell receptor and the MHC interactions.

Yep, that's the thing that we most want to do. And so now that we have the tools laid out, trying to do this much more on population scale, either within a particular HLA restriction or across them, is where we're heading.

Okay. And then finally, I think the last question we have is whether you are willing to— if you're looking for any potential collaborators and if you want to share any of your information for people to reach out to you. Yeah, always. Okay, great. Great.

Well, I'm sure— I was going to try to write it on the slide, but I don't think my penmanship is—

Well, I'm sure people can absolutely find you, but there you go.

There we go. Perfect. Yeah, by all means, feel free to email me.

Yeah, no, this is— thank you so much for all of the answering all of the questions, for all of this amazing work and showing us what is possible and what we can can elaborate for the future. So thank you very much for your time. And thank you to all our participants for participating today and for asking such great questions for our speaker. I'd ask that everyone please bear with us as we are transitioning from the Human Vaccines Project to the Human Immunome Project. And you will find some of our social media and things have changed and some have not. And that will be, you know,, you know, updated over the coming weeks. But a recording of today's presentation as well as all the other lab meetings will be available on our website and all social media channels. So with that, I'll say thank you again, Dr. Birnbaum. Thank you everyone for participating today, and we look forward to seeing you soon.

Thank you so much.