Scientist Stories: Aly Khan, Computational Approaches to Study Immune Cells

Welcome to the influenza webinar series hosted by the Human Vaccines Project. The Human Vaccines Project is a nonprofit public-private partnership with the goal of decoding the human immune system to accelerate the development of vaccines and immunotherapies. In order to make real progress toward the development of a universal influenza vaccine, collaboration among researchers around the world is essential. We hope that this monthly webinar will foster the scientific foundation needed to advance influenza research. This series of webinars will be broadcast on the third Tuesday of each month and posted online. For a copy of today's webinar, please visit the Human Vaccines Project website at www.humanvaccinesproject.org/talks. Following the presentation, we will have time for questions from the audience. Please type your questions into the questions box in the GoTo interface. I will read these questions following the presentation. And now it is my pleasure to introduce Dr. Ali Khan. Dr. Khan is a research assistant professor at the Toyota Technological Institute at Chicago, which is a computer science institute located on the campus of the University of Chicago. His research interests are in computer science and immunology, with a focus on developing new bioinformatics and machine learning methods to study the molecular and cellular systems that control immune function. In addition to his academic research, Dr. Khan has also worked on applied math and computational biology research in industry, including Merck and Genentech, and currently Tempus Labs. Today, Dr. Khan will tell us about connecting single-cell gene expression with B-cell receptor specificity.

Thanks, Gillian. So as Gillian mentioned, I'm a computer scientist who's interested in immunology, and I'm really thrilled that this cross-disciplinary series has a specific and I think great interest in utilizing bioinformatics and machine learning methods to better understand vaccine response and influenza. So thanks again for the invitation, and I'm really excited to kind of share today what I think are a series of approaches and techniques that encompass the landscape of computational methods that are useful for influenza research and understanding immune response. So briefly, I also do a lot of industry work, but most of it is unrelated to influenza or vaccine. So I thought I would share 3 short vignettes, and there's a couple themes that connect these vignettes together. The first is leveraging single-cell sequencing technologies, and then the second theme is principled kind of computational techniques for eliciting out subtle signals from single-cell data and improving our understanding of how specific immune processes work. What I'll do is I'll first go over some current and present computational paradigms for how we look at single-cell data, and then I'll move on to the application of some of these techniques for understanding how gene expression can be coupled to receptor specificity. And then I'll end on some open problems or areas that we're working on, which are some new algorithms for modeling B cell function. So single-cell biology, it's a complicated and challenging task from a computational standpoint. So the way we can kind of abstract single-cell RNA sequencing In a way to think about it computationally is that you begin with, on the left-hand side of this diagram, with single cells that have been captured. Those individual single cells are then prepared and sequenced. And mathematically, if we take the top branch of this kind of procedure, we can construct a matrix. And so here the matrix is columns that consist of the different cells that have been captured in a population. So these might be the cells that one captures in a germinal center, or these might be cells that have been sorted and selected for specific cell surface markers such as plasmoblasts. And the rows then are genes that can be— whose expression can be quantified through RNA-seq. And so for, let's say, the human genome, you have about 20,000 genes. And some subset of those are expressed in various cell types. And so on average, you'll get a few thousand genes that are expressed across a number of cells. With that matrix, there's a few computations one can perform. One of them is to kind of just assess the population structure of the cells you've captured. And there in computer science, there's a long litany in history of visualization techniques. So what we can do is we can take this matrix and project it into, let's say, a two-dimensional space where we can kind of look at how different cell types are clustering. And in areas where we've— in problems where we're interested in trying to understand the heterogeneity within a population, this is really great to kind of visualize And then interrogate what cell types are likely present, what are the transcriptional differences that are driving those clusters apart. Oops. So— oh, sorry. So, and then the last part is we can actually take the low-dimensional projection of those single cells, and then we can also infer trajectories, right? In a germinal center, there'll be affinity maturation or there'll just be other types of maturation processes going on in the immune system. And if we sample from specific niches or if we sample across a series of time points, we can actually then reconstruct how those cells are maturing over time and gain an understanding of how certain transcriptional circuits are driving that maturation process. So that was kind of like the top part of this. Schematic. On the bottom part of the schematic where we're referring to something called TCR-BCR assembly, one of the, I think, computational questions that has been neglected a bit was with single-cell RNA sequencing, while this is a really elegant and impressive paradigm for quantifying the expression of genes in various other conditions, at the end of the day, we're sequencing nucleic acid from these cells. And depending on the protocol, such as SMRT-seq2, or even in the more kind of current high-throughput 10x platforms, one can actually exploit that sequencing information to gain an understanding of what the immune cell receptor sequence is. And so here we can unlock an incredibly powerful insight into not only the cell's transcriptional status of what genes are expressed and turned on, but we can also gain an understanding of what does the cell recognize. So in B cells, the BCR sequence, and in T cells, of course, the TCR sequence. So we identified these kind of core computational challenges, and we published this in a review a couple years ago. And today I'm going to focus on the bottom half, which is an algorithmic technique for assembling the B cell receptor sequence. But we've gone into details about some of the opportunities and challenges for each one of these tasks in the review. So when we sought out to determine whether one can assemble the B cell receptor sequence from single-cell RNA sequencing data, there wasn't actually any published work in that area. And it's a challenging problem for a couple of reasons. First, single-cell RNA sequencing can be a bit noisy. Second, it's a challenging task. So the general problem of transcript assembly is challenging, but with single-cell, owing to complexities with coverage of what gets transcribed, this problem hadn't really been approached. So what we thought to do is come up with an efficient and sensitive technique for assembling B-cell receptor sequences when we can from single-cell RNA-seq data. And the basic approach is, if we look at a kind of a generic diagram of a B-cell receptor on the left-hand side of this figure, we can potentially identify regions of the B-cell receptor sequence, such as let's say in the heavy chain, that typically are not subject to significant genetic changes such as somatic hypermutation or other alterations. And that's an interesting kind of insight because what we can do is we can take a published database of known B alleles and constant regions and take the initial data set and try to identify what we call anchor So when we perform single-cell RNA sequencing, you're going to have a library of sequenced reads of a specific length. And we will try to then align those sequencing reads to our reference dataset of variable and constant regions and try to identify specific points along those two regions that can serve as anchors. And the intuition here that I kind of want to convey is that while we have a good understanding of, let's say, regions that are not subject to genetic alterations, there is certainly a gap in between where it's unclear to us what the sequence might be. There's no specific database of all possible BCR sequences with all possible somatic hypermutations, obviously. But what we can do is we can begin by identifying what we think is a good anchor position in, let's say, the constant region. And the gradient that I'm trying to show along that transcript is, let's say, some level of consistency between a given BCR transcript or a heavy chain transcript and what's known in databases. And we can try to identify a second anchor position, which also shares this kind of concept of sequence similarity to some known dataset. Okay, so this is kind of a reference-based initialization where we're beginning to identify two positions on this transcript that maps well to a pre-existing database. The intuition is then what we want to do is stitch across this. So what we would do is we would grab from our bucket of RNA sequencing reads and try to see if we can extend one of these anchor regions towards the other side. And hopefully, like this has become intuitive, what you want to do then is go back into the RNA sequencing read, try to extend further and further and further until you're able to actually bridge across. And so the important point to note here and why we call this a semi-de novo assembly method is that it's not completely de novo. We are initializing based on kind of our known understanding of BCR sequences and the body of receptor sequences that have been collected by many to articulate, you know, what are the constant regions and what are some of the variable regions that are present in human populations. So once we're able to kind of bridge and connect these two anchors together, we're able to then construct a putative VCR sequence. Computationally, a big challenge here that I won't go into a lot of detail, but I hope to kind of, again, convey how these techniques work because I think at the end of the day, methods matter. And it's important for both the people who develop methods and the practitioners who use these methods to kind of understand what were some of the assumptions involved in developing these techniques. So one of the things that computationally is kind of challenging is that as you're extending the sequence— so in stage 2B, what we show is the anchor sequence— at a given position or point in time in the algorithm, there's a variety of ways we could be extending that anchor sequence, just because of the non-uniformity of the reads that are generated from a single-cell RNA sequencing data. Experiment. And so what we used was an idea from information theory. Basically, what we want to do is minimize our entropy, or as every time we extend this sequence from one anchor to the other, we want to be very sure. And that's kind of a generic way of saying this, but we want to be very confident about our selection of reads that we use to extend from one anchor to the other. And this term of being very sure is predicated on the idea that we want to find scenarios where there are a large number of reads that support the read that we're intending to choose for extending an anchor. And at the same time, there isn't a lot of noise associated with it. So it's a very clear signal. So what we're showing in kind of that blue geometric shape is our selection of that read at the bottom over there that has evidence of at least 20 other reads that are identical, and any kind of spurious noise is actually quite small relative to the more abundant read. In other scenarios, you might have an equal number of two possibilities, and so what you want to do is you want to always choose the extension that is most confident from your perspective based on the number of reads that support it and also as an alternative, the amount of noise that might be in the next best alternative option. The thing to also, I think, appreciate is that this is a fairly simple and straightforward idea in computer science where this is a greedy algorithm. So every time we make a choice, we're choosing the best possible local option. So there are two things that we have to kind of consider. One is that it's not guaranteed to always find the optimal answer, which means that sometimes we're not able to stitch across the two anchor sequences, and that's due to some of the issues with it being a greedy algorithm. However, we found empirically and in practice that for the vast majority of time, we are indeed able to assemble a BCR transcript. And the limiting factor is actually just read coverage. So as long as you're able to, let's say with SmartSeq2, get enough read coverage of the BCR transcript, you are able to then successfully assemble the BCR. In order to kind of demonstrate the quality of this assembly, we compared it to a current state-of-the-art de novo assembly program called Trinity that was developed at the Broad. And what we found is that there was a 12% improvement over the Trinity method. So an example is in the single-cell experiment that we did in collaboration with Patrick Wilson's lab at the University of Chicago, where we have a single cell, and when we looked at the assembled sequence, which is that second to bottom gray line, we found that we were able to successfully reconstruct the variable region, including CDR1, CDR2, and CDR3. However, when we use a fully de novo method that does not use kind of some of the known biology, it falls short of assembling the entire sequence and specifically neglects to assemble the CDR3 region. The way we were able to kind of validate this is that Patrick's lab was able to then perform Sanger sequencing on the DNA, and we were able to then use that as a gold standard for confirming how accurate our assembly method is. And we found that we had strikingly less error than typical off-the-shelf tools such as Trinity. And so over time, because this was the first method for BCR assembly, other authors have also put forward their own BCR assembly techniques. And the good news is that we're actually quite good at assembling this. And so overall, this was quite In our mind, an important advancement and proof of concept that sometimes simple methods work very well and it can unlock specific and important insights from immune single-cell sequencing. At the same time, this method gained a little bit more, as it started to gain a little bit more traction, in collaboration with Steven Quake's lab, Derek Crute was able to use our same approach and actually adopt it so that it also assembles TCR sequences. So if you have single-cell TCR sequencing, this technique will also assemble— single-cell T cell sequencing, this method will also assemble TCR sequences. And so the tool is available. It's open source, and it can be easily downloaded and used. So the key points for this were we can now access BCR and TCR information from single-cell RNA sequencing data. And I think in this next part, which I hope to persuade you and demonstrate, we're poised to now connect this receptor information to specific transcriptional signatures. One thing to note is for many people who are adopting, I think, 10x technologies. 10x does have native TCR and BCR assembly for their 5' kit. So BASIC remains quite useful for SmartSeq2 protocols where the entire transcript is being sequenced. And there is some reason to think that the depth of sequencing for SmartSeq2 in some situations might be more useful for answering a specific question than the breadth of cells one can obtain in Metabux. And then just in terms of kind of where we're hoping to go in the future with this algorithmic work, we're actually thinking of adopting this technique now for efficient and fast assembly to bulk RNA-seq data. So if one was to perform bulk RNA sequencing on B cells or some other compartment, we'd like to be able to just use that RNA-seq data and assemble estimates of kind heavy chain and light chain. Of course, it'll be uncoupled, but it may still be quite useful to get a sense of the repertoire of the heavy chain and light chain from a bulk perspective and circumvent the need for doing amplicon-based receptor sequencing. So next, what I'll do is I'll try to share some work we've done in collaboration with Patrick Wilson's lab at University of Chicago on connecting single-cell expression to BCR specificity. So this was work with Patrick that we just published earlier this year, and it was done by a few very talented PhD students in his lab, in particular Carlin Niu, who had worked on this for quite some time. So we were interested in characterizing and understanding receptor-antigen reactivity and clonal— how receptor-antigen reactivity and clonal origin may alter gene expression profiles in plasmablasts. And so the general paradigm was after flu vaccination, 7 days after flu vaccination, IgA and IgG plasmablasts were isolated and subjected to single-cell RNA sequencing. And owing to Patrick's expertise in this area, his lab was able to actually determine which cells could recognize flu. And so that allowed Carlin and everyone in the group to be able to bin all the single cells in this experiment into 3 categories. So IgG vaccine positive, so Ig secreting cells that could recognize influenza, IgA vaccine positive, so for IgA, and then IgA vaccine negative. So as we started on this journey, and I know that we're all kind of, it's a sophisticated audience here, but I think it is important to kind of remember one thing, which is when working with large single-cell datasets, often as it is, in human studies, there are some batch effects. And so one of the things we noticed quite immediately was if we took the single-cell RNA sequencing data and we were to kind of do this low-dimensional projection that I mentioned earlier and we saw a clustering, we could find that the cells were actually clustering quite nicely by experimental batch. And so in the upper left-hand corner, is the initial dataset and PW refers to, let's say, experimental batch. We found that all the cells were actually clustering primarily by batch. And the upper right-hand corner is a different experimental designation. But the primary thing that I'd like to draw everyone's attention to is the graph on the upper left-hand side where the clustering is very clearly delineated by experimental batch. So one of the first things we actually needed to do was do batch correction. And so after batch normalization, where we're using a linear method to adjust for the expression so that it no longer is associated with the experimental batch, we felt a lot more confident and comfortable proceeding with the analysis. And so after batch normalization, you see the transformation that occurs in the bottom left-hand corner, which will show that all the different experimental batches are intermixed. And specifically, as you may remember, these were all specific plasma blast cells that had been sorted and selected. So this population— so the same population was selected from each of those— each of the individuals that comprise across these batches. And so the other thing to also note is that making sure to also do careful experimental design. So if you are going to utilize different batches, it's often useful to include samples across two batches to help facilitate normalization, which is what occurred in this case. One of the things we did straight out of the gate was we applied BASIC to these cells. And here, what we were interested in doing is seeing whether the repertoire information we were able to glean from a single cell, whether or not it matched information that has been previously known by Patrick and others at the population level using receptor sequencing. And so one thing that we found is that 10 to 20% of the clones that we identified in the single cell have both IgA and IgG isotypes. And so this was— and that's been previously seen, and that's also seen in population-level receptor sequencing. And so this was a nice kind of orthogonal kind of validation of the fact that BCR was able to resolve— basically was able to resolve BCR sequences, and also the results were consistent with other or floccinal techniques. So one of the first questions we wanted to answer was, because now we can assemble the B-cell receptor sequence of these individual cells, we're able to determine whether it's the clonal structure of all the cells. Do B-cells from the same clonal family have more similar transcription profiles than unrelated B cells? And so initially when we sought to answer this, what we thought is we would do a clustering or do a projection of all the single-cell data, mark the cells that had— that shared a clone. And so in the upper left-hand corner, we have numerically marked cells that are from the same clonal family. One of the things though that we noticed immediately was that the single-cell data was actually strongly stratified by the receptor isotype. And so if you look in the upper right-hand corner, that same single-cell data shows a strong association with simply whether or not the cell was expressing IgA or IgG. And so in the upper right-hand corner, the black and the blue are all the cells that are expressing IgA. And what you'll notice is that the black and the blue largely cluster together. And when we look at IgG, it clusters slightly apart. However, what we're also able to do then is bioinformatically remove any immunoglobulin genes. So isotype, variable region, B alleles and ask this question now in a completely structured way of saying, without Ig genes, do cells that are within the same clonal family, are there transcriptional profiles more similarly correlated than non-clones? And what we found is indeed that to be the case. It's a subtle but significant significantly important point where cells within the same clonal family are more correlated than non-clonal cells. And if you look then in the lower right-hand corner, this is kind of a revised projection, and it slightly differs from the plots above because we've removed the immunoglobulin genes. So this should kind of result in a slightly different projection of the data. And what you see still is many scenarios where cells from the same clonal family roughly correlate in the same kind of neighborhood and geographies. So overall, although BCR transcripts contribute to transcriptional similarity between clonal plasmablasts, the expression of non-Ig transcripts within a clonal family are also similar. The second question was, we wondered whether, are there significant transcriptional differences between vaccine-positive IgA and IgG plasmablasts? So here we were looking to see whether there's any transcriptional difference between plasmablasts that differ only by isotype of their antibodies. And so similar to kind of that earlier observation, when you look at just the cells that are vaccine positive, you see an immediate stratification between IgA-expressing cells and IgG-expressing cells. And so in the left-hand corner, this is again like a t-SNE plot of all the vaccine-positive cells. We can do a volcano plot, which is we can say, okay, let's actually compare what are some of the significant differences between the IgG and IgA. And perhaps not surprisingly, we found it was largely just the immunoglobulin genes that were different. So then we redid this exercise by removing all the Ig genes, and what you see then is actually a loss of any kind of distinct clustering based on isotype. And so this suggested to us that besides the BCR genes themselves, peripheral blood IgG or IgA plasmablasts induced by influenza vaccine have highly similar transcriptional profiles. The last question that I think we were able to approach because we could combine the receptor information and the expression data was we asked whether there are significant transcriptional differences between vaccine-positive and vaccine-negative IgA plasma blasts. And so here in the upper left-hand corner, we're showing again the dimensionality reduction plot of just the IgA plasmablasts. And here in blue are the IgA vaccine positive cells and in black are the IgA vaccine negative cells. What we find when we do a volcano plot or when we look for significant differences between those two populations are many number of genes besides actually immunoglobulin genes, and in particular two very interesting genes, FOOT8 and V4GALT1. When we remove Ig genes, we still see that the key transcriptional diff— or there are transcriptional differences between IgA vaccine negative and IgA vaccine positive, so it's not just driven by the expression of Ig genes, as it was seen quite clearly with IgA versus IgG. FLT8 and B4GALT1 have been implicated in glycosylation, and so there have been previous work and comments on how glycosylation plays a role in influenza-specific antibodies. And so previous work has suggested influenza-specific antibodies can shift from a pre-vaccine glycan profile towards a distinct glycan profile that may regulate overall immune response. So we found that there are significant transcriptional differences between vaccine-positive and vaccine-negative IgA plasma blasts, including genes associated with glycosylation of antibodies. However, whether this differential expression of glycosylation-associated genes are related to the generation event of plasma blasts or perhaps conversion of memory cells to plasma blasts, it's uncertain. But their differential expression seems to be regulated in an antigen-specific manner. So I think the key points to kind of summarize here are the fact that with single-cell RNA sequencing and key computational approaches, you're able to actually take a multimodal view of immune cells in an important context such as post-vaccination. You can use single-cell data to look at the heterogeneity of the cellular population. You can use the sequencing information from those single cells to obtain a view about the clonal structure of those B cells. And you can actually integrate that then together to identify important patterns and couplings and associations between receptor specificity and gene expression. And so to the best of my knowledge, this is perhaps the first work to kind of integratively develop these techniques and apply it to the specific paradigm. And we're excited to see further and future applications of this integrative approach. So what I thought I would then kind of bring to front are new algorithmic techniques that we're interested in developing and utilizing for studying B-cell function. So I thought I would break this into kind of two parts. First is in understanding— I think there's a distinct and key computational opportunity and challenge in understanding how gene expression changes over time in these immunological processes. And I thought I would also then talk a little bit about some of the questions that we're interested in as it relates to how to think about the space of these receptor sequences. So in regards to the first point about understanding gene expression over time, there's some collaborative work we've been doing with Harinder Singh, who's now at University of Pittsburgh. And so what we've been interested in is using our computational techniques for analyzing single-cell data, assembling the receptor sequence for understanding how B cells behave, let's say, during class switch recombination or epineuronyl maturation. So here's a quick schematic of some data that is unpublished, but it's stuff that we're working on in collaboration with Wender's group. Where we can see across a population of cells going from IgM to IgG the expression decrease of IgM and the increase of the switched isotype. And then we can try to find patterns that are associated with this transition. And so some key computational challenges that still remain to be solved that we're working on includes identifying other patterns in a principled way. So certainly, as one would expect, AID starts increasing as you get closer to completing isotype switching or class switch. And interestingly, it decreases once that's accomplished. Published, but we like to find other transcriptional programs associated with these processes in different ways. In experiments that, of course, involve multiple time points, multiple batches, this can be quite complex, and so developing new techniques to kind of study this more precisely are going to be very important. So I would propose that one open problem for the field is still coming up with effective techniques for integrating multiple datasets to resolve temporal changes with gene expression. And I think what will be very exciting is kind of approaching for the first time our ability to track expression changes during very important B-cell processes such as the 3D maturation. The second kind of area that I wanted to touch base on that I think might be very interesting for the community to think about in terms of advances in computation include works in understanding the structural basis and interaction of the antibodies. So some of you may be familiar with, of course, some landmark work that was presented at an annual meeting called the Critical Assessment of Techniques for Protein Structure Prediction. So this was actually machine learning work work that was performed by a group at Google called DeepMind. And what they showed here really remarkably is in, in blue, their predicted structure of a protein that they predicted from the primary amino acid sequence, and in green was the actual resolved structure of the protein, as let's say through crystallography. And what's incredibly exciting is that they were able to learn on large sets of data the orientation and general structure of these large molecules. And so just given the primary amino acid sequence, they were able to predict that structure in blue, which remarkably corresponds to what that protein actually looks like in a cell. So with colleagues at the Toyoda Technological Institute, we've been also interested in studying protein structure and molecular interactions. And so in the last year, we've been working on techniques to model protein-protein interactions. And I'm starting here because this will lead to, so I guess spoiler alert, this will lead to the question that we're particularly interested in, which is how do we start modeling antibody interaction with antigens. So if you have sequences that recognize just the stalk region of influenza or the head, we'd like to be able to know what sequence properties and what kind of techniques are there for now representing and modeling sequences that discriminatively are associated with different parts of the virus. And or even more simply, whether or not they are likely to interact with flu. So in the earlier work, Patrick's lab had to rely on ELISA to kind of predict whether a cell was vaccine positive or vaccine negative. I think it would be great if we're able to develop computational techniques where after a single-cell RNA experiment, point out distinctly which cells are likely to react to flu versus which ones are not. So here, just as a kind of a general overview of how we've gone about modeling the molecular interaction between two different proteins, and this can be generalized to antibody and antigen, we take, so starting from the left-hand side of this diagram, We begin with two protein molecules as elucidated or defined by their primary amino acid sequence. We find a way to represent those sequences in a specific kind of high-dimensional way, what we call as a sequence profile. And so here we're basically encoding that alphabet of amino acid sequence to a matrix that can represent it in a in a more complex fashion. And we take those matrices, we use neural networks, a specific kind of neural operations called convolutional neural networks, to combine and integrate information about that sequence in a variety of different ways. And we take that integrative representation of those sequences and we learn a model to help identify whether or not those two sequences will interact. And what we found is that this actually works well. So we trained and tested this on a large number of protein-protein interactions. And so we're actually quite excited and hopeful to begin approaching this question now for flu and other immunologically relevant problems. So computationally, I think, we can kind of present this as a series of exciting computational tasks. So beginning with antibody sequences, there's several prediction tasks we could try to approach, including predicting flu strain, whether the antibody is stock reactive or head reactive. More recent computational work has resulted in the ability to actually go backwards and try to learn what parts of the antibodies or even what parts of the target molecule or partner in the protein-protein interaction context are important for predicting, let's say, the strain of flu or other interactions. And so what that allows you to do is actually then develop a roadmap for saying, okay, these are the positions that are quite salient for me to be able to predict what's talk-reactive or head-reactive or associated with one strain or another. What happens if we were to then alter the amino acid in either the structure of the virus, or if we altered the amino acid in the antibody, would it lead to a better binder, or would it actually disrupt any preexisting interactions? So overall, when it comes to computational biology and machine learning, This is an incredibly exciting time. I think novel computational and machine learning methods are really needed to advance our understanding of vaccine response in influenza. And I think this is a really exciting time for computer scientists, but also for the broader interdisciplinary community. Single-cell technologies, large datasets, and machine learning, I think it represents an important confluence of different capabilities that really promises to unlock new immunological insights. So I'd like to kind of end there and thank some students who have worked with me at Toyo Technological Institute. Qingming Tang, he did a lot of the wonderful early work in modeling single-cell data and low-dimensional representation, some of the algorithmic work that went behind BASIC. He's going to be moving to Amazon. So we're losing a really good computational immunologist, but if there are students or trainees or postdocs who are interested in working at this intersection, I'd love to chat. And of course, collaboration with Patrick has been really special. As a computer scientist, it's been fantastic working with him given shared interests. Part of the work that we— that I presented involving connecting BCR specificity with single-cell expression and clonality. That was done also in collaboration with Sarah Kobe's lab at University of Chicago. So yeah, I'm happy to take any questions and go over any of the earlier points. I should also apologize, I'm kind of remote right now, so if there was stuff that I need to repeat or show again, I'm happy to do that as well. Thanks.

Thank you, Dr. Khan. That was really interesting. Um, at this point, if you'd like to ask a question, please just type it into your question box and I'll read it aloud. And, um, Ali, while we're waiting for people to type in questions, I was wondering if you could elaborate a little bit more on the protein-protein interaction modeling for predicting flu strain. I just thought that sounded really interesting and would like to hear a little bit more about it. I think you might be on mute, Dr. Khan.

Can you hear me?

Yeah, there you go.

Thanks. Yeah, thanks for the question. So This work began actually with us thinking about given two sets of sequence information. So in this case, antibodies and let's say sequences known to be associated with certain strains versus others, how do we actually represent these sequences in a machine learning paradigm? There was some work actually that we began playing around with from David Baker's group where they were looking at small, not antibodies, but small peptide inhibitors of different flu strains. And some of those peptides differentially bound to different strains of influenza. And so we were interested in seeing if we could model that high-throughput data by taking the sequence information and finding ways of predicting whether or not those peptides would interact with a given flu strain. So the computational part, I think we have a good idea or I won't say handled until we're able to kind of empirically prove it, but the next part is now actually mining through public datasets with collaborators, getting these types of antibody and strain-associated information and seeing whether or not we can find highly accurate classification results.

Got it. Thank you. Um, and it looks like at this point there are no other questions. Sometimes those come through late, so I'll be happy to email you them if there are any that come in later. Um, and I just want to thank you again.