Scientist Stories: John Wherry, Systems Immunology and T-cell Exhaustion

All right, given that, um, John has quite a few slides as usual and, uh, that we were chatting it up way too before we started, let's get started. Uh, so today it's my pleasure to introduce Dr. John Wehry for our CTSI I3T seminar series. He holds the Barbara and Richard Shiffrin President's Distinguished Professorship at the Perelman School of Medicine at the University of Pennsylvania. He's also chair of the Department of Systems Pharmacology, Translational Therapeutics. He's director of this Institute of Immunology and co-director of the Penn's Parker Institute for Cancer Immunotherapy. He did his undergrad education at Penn State, his PhD at Thomas Jefferson in Philadelphia, which was then followed up by a very productive postdoctoral fellowship training with Rafi Ahmed at Emory in Atlanta, And then he came back to Philly, 2004, is that about right? Something like that, to start his own lab at the Wistar Institute, and then moved up the academic ladder very quickly. So I think for the trainees in the audience, and I think this is, this is neat to show, because Dr. Weirich is a great and fantastic example of someone who trained in fundamental immunology, but now really is a leader in the translational study of antitumor T-cell responses in cancer patients, especially undergoing immunotherapy. So he's arguably one of the world's experts in the transcriptional control of T-cell exhaustion. But he began all of this in mouse models of chronic viral infection where T-cell exhaustion was really first identified and where co-inhibitory PD-1 antibody blockade was first identified. In the last 10 years, he's really defined the molecular epigenetic programs that control T-cell exhaustion. This is obviously important in checkpoint blockade in cancer patients. He was one of the first initial groups to identify the transcription factor Tox and how it plays in T-cell exhaustion. And he's also identified these precursor populations of exhausted T-cells that preferentially respond to checkpoint blockade. So I think the common underlying biology of T-cell exhaustion and chronic viral infection and tumor growth, you know, has really spurred a lot of this in-depth understanding of how checkpoint blockade can reinvigorate T-cell functionality. It's, you know, it's not only confirmed or confined to this transcriptome. He's also defining this at the epigenetic level of exhaustion as well. So I won't take too much of the time, but to say, His talk today is entitled Mechanisms and Clinical Relevance of T Cell Exhaustion. Please welcome and join me in welcoming him for the CTSI Seminar Series today. Thank you. OK, John.

Thanks, Rob. I appreciate the very kind introduction. And let me just share this. And hopefully—

You're in the back of the room. Yeah, there you go.

All right, so thanks for the invitation and the introduction that sort of sets the stage, I think, nicely for what I want to talk about. What I really want to do today is talk about a new set of data we have that we generated with the goal of providing an atlas or a broad foundation for understanding T cell biology in humans. And the orientation of this is, as Rob said, really coming from the standpoint of trying to understand T cell exhaustion in humans and the relevance for immunotherapy. So I'm gonna give just a little bit of background on that, and then I'll talk about what we've learned from this profiling of human T cells. A few disclosures, I don't think any of this is directly relevant for what I'm gonna talk about for today, but there it is. All right, so I like to start with just a general slide about T cell exhaustion. And I know many of you have seen this and heard me say these things before, but just to get everybody on the same page, what we discovered many years ago in chronic viral infections was this population of T cells illustrated by the functional profile here on the upper left of cells that, unlike memory cells that can make cytokines very rapidly when you stimulate them, there are these cells in chronic infection that we could detect by tetramers, but when you stimulated them, they didn't behave like memory cells. They made a little bit of interferon gamma. They actually make quite a lot of chemokines, but they really weren't polyfunctional like you see for memory cells. These cells express high levels of inhibitory receptors, especially PD-1, but many others. And we now know that, that those inhibitory receptors are very important because when you block them, you can at least partially reinvigorate these exhausted T cells. And that was first shown in mice by Dan Barber and others in tumor models. And we now know from Tony's work, from Suzanne Topalian's work, from many, many others, that when you block PD-1 in humans, you can improve tumor immunology, tumor immunotherapy. And that operates at least partially by the reinvigoration of exhausted T cells illustrated here just by an increase in proliferation of cells expressing multiple inhibitory receptors. But these cells probably didn't evolutionarily evolve to deal with tumors. They evolved to actually establish a long-term stalemate between probably a chronic viral infection or chronic infection in the host. And they do so by basically establishing temporary or partial control of disease progression. We know that these cells are important in those settings because one, they drive escape mutation in things like HIV and hepatitis C infection. And two, if you eliminate these cells by depleting them with depleting antibodies in monkeys infected with SIV, viral load goes up and the monkeys progress rapidly to AIDS and die. So these cells are very— they're functionally relevant in these settings, but they're not functional enough to achieve a cure. And there are a variety of reasons why that might have evolved that we could talk about later if people are interested. But the fact that you can reinvigorate them has led to the understanding that these cells likely play an important role in cancer immunotherapy in many settings, at least settings where the tumors are immunogenic. We also learned over the years that these cells have a very distinct transcriptional program, which has had us thinking for a long time that they're really a distinct type of cell compared to an effector cell or memory cell. And then really in the last handful of years, we've been able to provide hard evidence behind that, looking at the open chromatin landscape of exhausted T cells that I'm illustrating here just at one locus in the mouse, the locus encoding PD-1, where these peaks represent areas of open chromatin, where transcription factors can come in and work. And in naive T cells, that gene is closed just with the boundary elements open here. But as cells get activated and become effector and then memory cells, the gene becomes open. These peaks become accessible— or these sites become accessible. And in fact, in exhausted T cells, you can see some of these sites are more accessible. The transcription factors we know bind there. But importantly, you see that there are also these unique places that are only open in exhausted T cells. This one I'll talk about later a little bit more, about 20 kb upstream of the start site for PD-1. But overall, there are about 18,000 open chromatin regions that are different from naive T cells in exhausted T cells. And about a third of those, or 6,000, are changes that are unique to exhausted T cells. So that's about the same magnitude of difference that you get if you compare a B cell and a macrophage or a macrophage and a T cell. So it really kind of illustrates that these exhausted T cells are not simply effector T cells that are just tired or memory cells that just, you know, have some disruption in cell signaling. They're actually a distinct lineage or distinct fate of post-naive T cells. So how can we use this information to understand immunotherapy to cancer? That's really the goal of all this. We want to understand, you know, which cell types are responding, which ones are important, you know, when does that response occur, and really what are the mechanisms. And I think underlying this has been something that has troubled me and the lab for a long time, and that is that we have all of these different flavors of T cells, especially in humans, where the more markers you use, the more subsets of T cells you might find. And what we often struggle with is the fact that many of us as immunologists, when we say one of these terms, we may not mean the same thing as the person sitting next to us. And just a good illustration of this is My good friend Mike Betz at Penn and I argue constantly about the definition of an effector memory cell. He likes to use CCR7, I like to use CD27. They don't overlap perfectly, but to both of us, those are effector memory cells, and yet they're not describing exactly the same population. And so part of the goal of the work I'm going to tell you about is to try to provide a better definition of these things that we talk about, but to provide those definitions in molecular terms. All right, so first, just a little bit of background. So one of the things we wanted to try to understand very early on about cancer immunotherapy was how quickly does this response occur? And, you know, what happens when we block the PD-1 pathway? We went in sort of trying to ask first, you know, what are the cell types that respond? And if they respond, what do we do? When we first started this now, you know, 7, 8 years ago, clinically, patients were put on anti-PD-1, and come back a few months later and look at what happens to the tumor. And there were, of course, people like Tony and others doing much more sophisticated work at the time, but it took us some time to actually get our folks oriented to the idea that the immune response might actually occur more quickly. And in fact, working with Dara Mitchell, Alex Huang in the lab did a lot of work trying to understand the kinetics of the immune response after initiating PD-1 blockade. In melanoma patients, and just illustrated from a couple of papers here, we now know that this response occurs very rapidly. There's a single peak of this activity, and then the cells basically return to baseline. And we can see typically this peak occurs within 3 or 4 weeks, but even as early as 1 week in many patients, we see very robust activation of CD8 T cells that turn out to be, for all intents and purposes, exhausted T cells circulating in the blood. So we of course wanted to know when this was happening in the tumor. The blood is very useful. We knew from Tony's work that there were T cells in the tumor pre-therapy, and that beneficial clinical responses correlated with activation of those cells after you initiated PD-1 therapy. We wanted to try to understand, you know, how quickly that was happening and what were the other associated events. And so we designed a similar study where we got a biopsy before therapy and a resection 3 weeks later to try to ask what was happening to the T cells in the tumor tumor and also ask how that might relate to what we see in the blood. In this kind of stage 3 and 4 neoadjuvant kind of trial, the response rate is about 55% overall. But what you noticed, or what Alex noticed, was that there were really different phenotypes of the resected tumor material at this 3-week time point. About a third of the patients at tumor resection had either no tumor on pathological evaluation or less than 10% viable tumor. When you analyzed the resection. That 30%, we know they had tumor at the beginning because we have the biopsy, but within 3 weeks, many of those patients, the tumor was completely eliminated. And as predicted from earlier work, that decrease in viable tumor was associated with an increase in the infiltration or the number of T cells in the tumor bed following therapy. In fact, if you had that kind of rapid response early, you were essentially 100% disease-free for, it's now out well past 3 years. So the very rapid activation of the immune system, in this case exhausted T cells, led to a very rapid cure, whereas if you did not have that, you had about a 50% chance of disease recurrence. And actually, disease recurrence occurred typically within the first year. All right, so what we know from this and from lots of other studies in the field is if you sort of examine the cells that are responding to PD-1 therapy, do a whole bunch of sort of detailed immune profiling that I won't go into in a lot of detail, sort of crunch the data, to understand what kind of T cell and what kind of interactions are occurring. What we found is it's mainly exhausted T cells that are responding to PD-1 blockade, at least in melanoma patients. And you could see that both phenotypically and transcriptionally, just showing you sort of a transcriptional enrichment for the exhausted signature in these responding cells. They had markers of those exhausted T cells actually being reinvigorated based on proliferation, 67 and other transcriptional features. The cells in the blood also overlapped in TCR sequence with those found in the tumor. So now extending what we knew was happening in the tumor out into the blood, many of the cells that increase in this proliferation signature in the blood at least share TCRs with the tumor. At least some of them we now know are tumor reactive. Whether they're specific for mutational neoantigens or tissue antigens, we don't know at this point. But what we do know is that exhausted T cells are a major responding cell type to immunotherapy. We can see that both with phenotypic markers and transcriptional signatures. There's overlap in TCR usage and specificity in blood and the tumor. And we could also use this to identify some resistance mechanisms, including seeing things that others had seen, like loss of beta-2m, induction of other kinds of signatures in myeloid cells, and also other types of interferon pathway mutations. But it clearly raised new questions that we thought needed to be addressed. This is probably the easiest setting of trying to identify cell identity and what cell types matter in melanoma. The major questions we were left with was really, how do I identify exhausted T cells in other settings in humans? These cells may be heterogeneous. We now know that there are different subpopulations of exhausted T cells. And we don't really know that these are the relevant cell type in all kinds of immunotherapy. So what if other cell types matter? How can we identify what they are? And the underlying mechanisms of how they work. And of course, the semantics here are terrible. We sort of interchangeably use the word exhaustion and dysfunction, sometimes senescence, sometimes other terms. And I think part of that is, even for us, a lack of precision on exactly what we mean when we say exhaustion or senescence. So we wanted to understand how human T cell differentiation states were related, and could we actually put a molecular what molecular signatures and also molecular mechanisms behind those different states of human T cell differentiation. So how do I identify basically T cell intrinsic reasons why immunotherapy fails? Well, we reason that we can only do this if we have an accurate map of those human T cell differentiation states. So how can we start to build the tools to really understand this in detail, taking into account the fact there's going to be heterogeneity and underlying gene regulatory landscapes that control that heterogeneity. One of the things that sort of guided our set of choices was a set of data from mice. And this is looking at the global view of that ATAC-seq data, the open chromatin landscape of naive, effector, memory, and exhausted T cells from mice. And one of the key points you can see is that exhausted T cells are as different from naive T cells as they are from effector memory T cells, so indicating that sort of divergence and differentiation state. But the second point I think is really important, and that is when we block the PD-1 pathway in this model, which is chronic LCMV infection, we actually achieve really significant therapeutic benefit. We reinvigorate the cells functionally and phenotypically, but we really don't change them epigenetically. In other words, they stay very, very similar to their parent exhausted T cells despite the fact that we've changed their activity, we've changed the way they look. So we sort of reasoned that one of the things that was going to be important in understanding the biology of human T cells was really to create maps of this epigenetic landscape. One of the things guiding that was the idea that at least some of these features in an open chromatin map actually are causal to the way the cells behave. So I pointed out this open chromatin region upstream of PD-1 that's really distinct and found only in exhausted T cells in this dataset. What we wanted to do, and this is work that Deb Seddon did in collaboration we had with Nick Caning's lab, is she actually CRISPR-engineered cells to remove this enhancer element. When she did that, she could actually select a cell line that had that deleted. And when she looked at the expression of PD-1 in that cell line that was now only lacking this small enhancer element, she actually decreased PD-1 expression. But she didn't ablate it. What she did is she lowered PD-1 expression level from the very high level that you see in exhausted T cells down to the level that you see typically in an effector cell. And that's because these effector-type open chromatin regions were still intact. She did something else very cool, which she did sort of a tiled CRISPR screen through the whole 1.2-kb enhancer, and you could actually identify transcription factor binding sites that may be important. I'm not going to talk much more about this, but this just sort of illustrates how you can go one layer deeper into this kind of analysis, and that might be useful. So the key question that we want to address is whether similar principles like this existed in human T cells. And so what I'm going to tell you about for the last part of the talk is the development of a human T cell molecular reference atlas that Josephine Giles has generated over the past several years in the lab. So what Josephine wanted to do was develop a very, very deep and very robust atlas of both transcriptional profiles and open chromatin landscape in a very, very well-annotated set of T cell subsets. She used about 24 healthy adults. They were mixed young and old. There's some age-related things that are interesting here that I'm not going to talk about today. But she sorted 14 different human T cell subsets they are listed here, and performed high-resolution ATAC-seq and RNA-seq. Now I'll point out that we have a bit of CD8 T cell bias in the lab. So we have about 10 subsets of CD8 T cells, only 4 of CD4s. I apologize to the CD4 biologists. We can go back and do more there. But we sorted a variety of well-known and well-defined subsets that are reported in the literature. And we did that so we could start with known biology and then work from there. The other reason we did this by bulk RNA-seq and ATAC-seq rather than single cell is that some of these populations are very, very rare, and by single cell analysis, we simply wouldn't get the data. The one at the bottom is one very relevant for this discussion. This is from healthy donors. We sorted out a population of putatively exhausted T cells based on high expression of PD-1 and CD39. This is just the sorting strategy. I won't belabor all of this because I think the details aren't terribly important, but the sorts were better than 98% pure for all these populations. And showing you which populations, we have a sort of crudely purified naive and non-naive, just gated CD27, CD45RA. But then we also sorted a rigorously pure population of naive T cells excluding the stem cell memory populations. We sorted both of them based on CXCR3 positive or negative, central memory, two populations of effector memory based on differences in CD27 and CCR7. Temeras, that's sort of terminally differentiated population, and then exhausted T cells. And you can see that these putative exhausted T cells and the stem cell memory cells would simply have been lost by a single-cell approach. Josephine then performed ATAC-seq and RNA-seq. And when you just lay this out in principal component space, several really clear things are obvious and quite interesting. One, you very easily separate CD4 and CD8 T cell lineages in principal component space with Principal component 2 contributing really a lot of that information. You also very clearly see progression from less differentiated on the left-hand side with the naive populations and more differentiated on the right-hand side with the so-called TEMRA populations, the sort of more senescent populations. And then almost right in the middle, you find your exhausted T cells. And so this, I think, is what we expected but may not always be obvious, that these cells are very distinct from your sort of terminally differentiated or senescent-like effector cells. You can also see that those exhausted T cells sit about in the middle, not quite in the middle, but they're somewhere in between the stem cell memory and the effector memory. And it's not your eyes fooling you. Actually, the resolution of this is actually much tighter by ATAC-seq than RNA-seq. And that's a point I'm going to make on a couple different slides. You can sort of cluster this in a lot of different ways. There's only one main point I want to make from this is that you can cluster it. You can see some cell types actually cluster very neatly together. And you can see those RNA signatures. You can see it by ATAC-seq, and you can separate ATAC-seq into the open chromatin regions that are near the promoter, so the ones that are what we call promoter proximal. And you can remove the ones that are more distal. Now, what's interesting about doing this— and this is set with about a 2-kb cutoff— the more distal open chromatin regions have been reported in other settings to actually contain more cell-type-specific information than the promoter-proximal open chromatin region. You can think about a gene that's expressed in 5 cell types. The enhancer elements near the promoter are more likely to be used in all 5 cell types, whereas the cell-type-specific regulation is more likely to be contained in things that are more distal to the promoter. And so you can see it quite easily visually, but you can also measure it with something called an information quotient. In fact, the distal open chromatin regions convey a lot more information about the cell type identities here than do RNA-seq or the proximal open chromatin regions. So how do these patterns compare for different cell types across this atlas so we can do some benchmarking and some quality control? So first of all, if you compare— the way this graph works is this is taking the transcriptional signature or the ATAC-seq signature at the top, genes at the bottom, of naive T cells and asking how different that is from each of the cell types listed across the bottom. And you can see there's a different fold change in the different colors going up and down. So what this tells you is that naive T cells are most similar to your stem cell memory populations and most different from these mRA populations on this side. And in fact, you can see the gradient of that for ATAC-seq and RNA-seq. If you take stem cell memory, the ones that express CXCR3 are almost identical to the ones that do not by ATAC-seq, even though they differ in some phenotypic marker that we can sort. And they're a little bit different from central memory cells by RNA, but not by open chromatin. And they're actually most different in chromatin landscape from the TEMRES and the naive, and similar by RNA-seq. And you can do this across the whole spectrum of things. You can take, for example, these exhausted phenotype T cells, and you see they're actually most similar to your effector memory cells. They're actually most different from your naive and actually most different from these MRA or the senescent cells over here. And you can see it by RNA-seq, it's roughly the same order, although not exactly. So this starts to give you a sense of how related these cells are, both in RNA-seq space and in chromatin landscape. You can look at this in a little more detail and ask what's underlying those transcriptional circuits. This is just two different ways of actually viewing that data scaled by column or row. And the reason it's scaled differently is so you can see actually different patterns here depending on which feature you're highlighting to be most relevant. You can then ask in those clusters of open chromatin landscape, what are some of the transcription factor binding sites that are enriched in a feature like this cluster 5 that's actually most preferentially found in exhausted T cells. So you go over here, and you can see that the transcription factor binding sites that are really preferentially found in just that set of open chromatin is highly enriched in the retinoic acid receptor family, ROR family of transcription factors. You can take, for example, cluster 6 here, which is now enriched in your effector memory. You find a lot of effector-like transcription factors, including members of the AP1 family, but also some overlap with this ROR signature. You can go down here to cluster 7 or 8, where you have a lot of these MRA cells. And what you find is actually a lot of transcription factor binding sites for things like T-box factors, as well as NF-kappaB and some other settings and AP1 family here. So you can look at this in more detail and try to identify which transcription factors might not just have binding sites but actually have the most activity by linking binding site accessibility to RNA, you come up with another heat map. But if I just call your attention to a couple of specific things that fall out of this, what you find is in those mRAS cells, more effector-like cells, really high activity predicted for Tbet and also p53, which might relate to the senescence of these cells. If you look in the middle here, these are the effector memory populations, you see strong signatures of effector-like transcription factors, including AP-1 family members and also NF-kappa-B. And then if you compare down here, the exhausted T cells and naive T cells, there's this pair, this E2F pair, E2F4 and E2F2. E2F4 actually plays a very important role in restraining activation and restraining proliferation, whereas E2F2 is an antagonistic partner that actually promotes cell division. And we'll talk about how that relates to the biology of exhausted T cells. The other thing you can do with this data is actually start to predict developmental relationships. And so you can do sort of correlation between the transcriptional or epigenetic signatures. And you can get a predicted sequence or predicted relatedness, let's call it, between these populations where the stem cell memory come before the central memory, comes for the effector memory with the, the mRA on the far most differentiated side. We've removed the exhausted T cells and one of the populations of stem cell memory here because we expect that there's some branching in this differentiation that really won't be picked up by this sort of simple correlation analysis. You can also do this another way, and you can use a pseudotime algorithm. There are many of them. Most of them that we tried give you a very similar picture of basically a progressive relationship, perhaps a lineage relationship, but certainly a progressive relationship in terms of transcriptional program going from naive to stem cell memory to central memory and so on. None of this is very surprising, but I think that what this does is it creates a foundation that we can do a lot of cool things with the dataset that we have. So can we use this to understand the regulatory mechanisms that might control the transitions between those different human in T cell subsets. So this is now taking that kind of predicted lineage relationship and projecting it across a set of data. Now, this set of data is the number of genes that change across a pattern with these cell types at the bottom. So for example, 1,087 genes are low in naive T cells and are turned on in anything non-naive. On the flip side, 657 are expressed by naive cells and then get turned off in anything that comes after. You can see that there's— that's the major inflection point of about 1,700 genes have that pattern. They're either expressed or not in naive T cells and then transition after that. There's another set of about 1,400 genes that stay either off or low and then transition as cells go from effector memory— from central memory to effector memory. And you can see the patterns of other types of gene regulation, genes that peak really in the stem cell memory or genes that really progressively increase as you go across to the mRA population. Overall, though, 79% of the changes in gene expression are explained by the top row of 6 different patterns here. We thought this was very interesting on the RNA-seq side, but actually what we were really interested in was how this works for the open chromatin landscape. And there are two reasons for that. One, we think we have more information about cell identity here, and that's based on that information quotient feature that I told you about earlier. But also now with each of these transitions between developmental states, we can identify the underlying transcription factors that bind to the sites that either become more or less accessible. So for example, as cells transition from central memory to anything more effector-like, the transcription factor binding sites that become more accessible are for ROR gamma T, T-bet, and ETS family members. Conversely, you can see things that become less accessible as cells become less and less or more different from naive T cells. You lose binding sites for LEF1, TCF7, EGR, ERG, and ZEB1. And you can do this across this whole pattern, again, identifying transition associated transcription factor binding sites. But like the RNA-seq, about 80% of all of the variance that occurs here occurs with the patterns listed across the top, with more fine-tuning of the patterns down below. And just to show you a couple of examples of what this kind of looks like for a gene or two, Granzyme B is a gene controlled by a handful of open chromatin regions. Down on the right-hand side, you can see the RNA pattern. And what you see is that there are 6 open chromatin regions that become more accessible right at the transition between central memory and effector memory. There's 1 that changes anytime the cell becomes non-naive and 2 that sort of progressively become more accessible across the whole pattern of cell differentiation. So you can envision that this 1 site here might control a little bit of this increase in variance that we see as the cell goes from naive to non-naive. But really, it's likely these 6 sites that control this transition from the central memory to the effector memory, and then these 2 sites fine-tune the increase in expression that you see after that. So having sort of the logic of chromatin accessibility sites for an individual gene starts to provide context for the fine-resolution control of gene expression. And you can see this for other genes. Granzyme B is actually relatively simple in terms of the number of open chromatin regions, But there are other genes like LEF1 that really have— in this case, I think it's around 27, 29 different open chromatin regions that seem to have some contribution to the fine-tuning of regulation of this gene. All right, so that's all great, and it sort of gives us some backdrop for thinking about T cell biology. But what can this atlas tell us about T cell exhaustion? I started off with that being really a major goal of understanding more about T cell exhaustion in humans. Well, because we think that might branch off of this sort of linear differentiation, we decided to just take a deep dive into these cells that we sorted from the blood that express PD-1 and CD39 that we thought might represent putatively exhausted T cells and sort of dig into the biology there. If we just do sort of a sort of user-defined look at what genes are expressed, these putative exhausted T cells from the blood in fact express PD-1, CD39. That's good because that's what we sorted on. But they also express a lot of other genes that are more highly expressed in exhausted T cells in other settings, where you can see that you can benchmark the other cell types based on genes associated with naive and quiescence, although you see a hint of those over here in exhaustion. More effector-like genes that you see really rising in the effector memory and the TEMRs. The cytotoxic circuits that are really very high in the TEMRs and maybe a tiny bit here in the exhausted T cells and so on. You can do this in a more agnostic way and just let the genes tell you what's associated with each cell type. You can see very, very clear patterns over here for this exhausted population in cluster 12, cluster 13, maybe cluster 15. You can then benchmark this also against other known things. So first of all, you can look at the GO terms and things that these clusters might tell you. The exhaustion-associated clusters are really enriched in features of cell cycle and perhaps features of response to inflammation.. And indeed, the exhausted T cells from the blood have the highest Ki-67 of any of the populations you sort from human blood. Now, there are a variety of questions that arise from this. One, what— how do you know these aren't just recently activated effector T cells? I'm going to tell you a little bit about that in a second. But also, we know from mouse models that, in fact, the only way to maintain a population of exhausted T cells is if the population is able to undergo proliferation.. And that proliferation actually maintains sort of a proliferative hierarchy where a population of progenitor-exhausted T cells is constantly undergoing proliferative differentiation into a more terminal population of cells. Once these cells have divided, though, they become basically postmitotic and can't divide anymore. And we sort of benchmarked this against data from Zemin Zheng's group and, you know, Amit's group. And in fact, the population of cells we sort from the peripheral blood of healthy humans has the most overlap both transcriptionally and epigenetically with basically exhausted T cells from tumors in Zeman's data and also in Ido's data. So what else can we learn about this population of exhausted-like cells from the peripheral blood? Well, if we look at differentially accessible peaks in these cells, you can sort of see how you get clusters that overlap for each of these different populations listed on the side. And you can then ask again about transcription factor binding sites underneath these open chromatin peaks. And you find cluster 4 here that actually is very enriched in exhausted T cells has binding sites for many effector-like transcription factors, but it's missing a few of the canonical ones like canonical AP1 and JunB that you find in these sort of more effector-like TEMRA cells. You also have openings for other types of bZIP or AP1 family combinations and a few other features that are quite interesting there. There are a few other smaller clusters here that have other unique binding sites that may be interesting. And you again see the ROR family popping up as being perhaps relevant in exhausted T cells. And you can just look at the patterns of open chromatin. This is under the— for the gene CD27, where you see CD27 is increased in exhausted T cells, or at least the same as you have in central memory and naive cells.. And the open chromatin landscape sort of fits with that, with these 3 sites highlighted in purple that are lost from the more effector-like sites but are shared with the sort of central memory stem cell memory. All right, so PD-1/CD39-positive cells in the peripheral blood of healthy donors have similarities to exhausted T cells. So how do they compare phenotypically and functionally? And so if you look at sort of benchmark molecules expressed by exhausted T cells, including a little bit of TCF, TCF1, a little bit of granzyme B, TOX, and TIGIT. The pink here are these exhausted T cells from the peripheral blood. They have some TCF7. I'm comparing here to basically a TIL population, some of which have TCF7. They have some TOX, although they're not the highest TOX in the peripheral blood, and there's much higher TOX in the tumor. And they do have high TIGIT. So you can do this for a variety of other markers. I won't go through all of these except to say that they're more activated in the peripheral blood, have higher K67, and lower expression of many of the effector molecules like T-bet, CX3, CR1, and so on. And then functionally, you can take out each of these populations and stimulate them in vitro and compare them to unstimulated cells. Naive T cells stimulated ex vivo don't make interferon, TNF, or IL-2, whereas central memory cells, some of them at least, make interferon, TNF, your effector memory, your TEMRA cells make lots of cytokines. And you can do this for lots of the other populations, and they behave just like you would predict from even old literature about the functional properties. These exhausted T cells in the peripheral blood are very functionally poor. They make a little bit of interferon gamma. I'm sure you can't see it that well, but there is a little bit of interferon gamma being made here. And you can distinguish these exhausted T cells because they express PD-1 and CD39 from other activated T cells that simply express PD-1. And you can sort of quantify this here. These cells are poorly polyfunctional based on these markers. So all right, that's great. So how can we use this to understand the identity or the biology associated with unknown T cell types that you might have from some, you know, clinical trial or some other setting in humans? So how can we discover new biology about disease? So I'm going to give you overall, if I have time, I'm going to give you 3 examples, but let me give you 2 here. So the first thing we did is take a dataset from Anshu Sipati at Stanford where he did single-cell ATAC-seq in basal cell carcinoma patients after they received immunotherapy. He identified 3 populations of TILs and identified them based on function, phenotype, and a little bit of epigenetics as being putatively early exhausted cells, intermediate exhausted cells, or terminal exhausted exhausted cells. So the first thing we wanted to know is in that ATAC-seq landscape, how much of their epigenetic landscape overlapped with what we captured in the peripheral blood in any of the subsets in the peripheral blood. Well, almost 99% of the epigenetic landscape of those early exhausted T cells overlapped with something that we saw in the peripheral blood. 85% of the intermediate exhausted T cell epigenetic landscape overlapped. But only about two-thirds of this terminal exhausted population had similarity to what we saw in the blood. So this is the one we wanted to ask whether we could understand more about that. You can see this nice overlap with this intermediate exhausted population here with the exhausted population from the peripheral blood, those PD-1 CD39. That fits with a lot of the biology that we know about these cells. Those healthy donor exhausted T cells in the blood are most like Aun Su's intermediate exhausted T cells. That fits with studies from mice that show that intermediate exhausted T cells are, one, the ones that are currently undergoing the proliferative expansion, and two, are the population in mice that actually has the best recirculation capacity and are the most abundant in the blood. So that fits, fits really nicely with the very strong enrichment there. But we wanted to know is whether this atlas might provide insight for a population that actually didn't resolve very well. Can the fact that we have high-resolution signatures of multiple cell types in humans help us deconvolve a more complex picture? So one of the things we did is take the open chromatin landscape from Aung Suu's terminal exhausted T cells and look at which features overlap with the healthy donor atlas and which don't. And what we're doing here specifically is looking for transcription factor binding sites that fall off the diagonal. That is, they are more enriched in Aung Suu's dataset than they are in the healthy donor dataset.. And one of them popped out very, very clearly. And this— I can't believe how robust the signature was. That is the signature of NR4A1, so basically NERF77. So this is a transcription factor that operates downstream of very proximal or very recent TCR signaling. And what this suggests is that one of the major features contained in ANSU's terminal exhausted population in the tumor that we do not see in the blood is a signature of recent transcriptional activation downstream of the TCR activating those genes. So it sort of suggests that what we're missing in the blood was actually the signature of a cell that was getting actively signaled. You could see some evidence of that in other genes like TOX, where you can see there are binding sites that are— or chromatin accessibility sites that are really preferentially enriched in these terminal exhausted T cells that are not found in the blood.. And you can then look at the— basically the clusters of open chromatin landscape in those terminal exhausted T cells. And what you get is sort of a composite picture that might explain what's happening in these cells. There's clearly an overlap with one part of the signature of these terminal exhausted cells with the PD-1 CD39 cells from the peripheral blood. But actually a lot of other signature comes from these sort of central memory stem cell memory and maybe even that naive signature epigenetically. These are just two genes illustrating sort of an overlap with a naive-like pattern and a more effector-like pattern here. But this fits very well with the biology that Ansu reported in that paper, whereas in these terminal exhausted cells in basal cell carcinoma, because of the TCR clones, there was strong evidence of recruitment of new TCR clones, perhaps new priming playing a role. And I think what we would interpret here from this data is that this is evidence of actually residual signature of early differentiated cells as they come into the tumor and begin to acquire features of exhaustion. All right. So what about in a dataset where we have evidence for clinical progression versus clinical response? So we turned to an existing dataset from our lab on melanoma where we knew some of the reasons for treatment failure, but not all of them. So I'm going back to that neoadjuvant trial that I told you about in the beginning where we actually had tumor available both pre- and post-anti-PD-1. We wanted to ask whether there are signatures of clinical responders versus progressors that overlap with what we saw in this healthy donor atlas. And indeed, in the clinical non-responders, there was a very strong transcriptional signature that included very prominent CCR7 and Lef1 expression and a few other genes that sort of had a flavor of poorly activated activated T cells. If you then take the signature, in this case the ATAC-seq signature from these tumor-infiltrating T cells from clinical responders and non-responders and ask where they overlap in this healthy donor atlas, what you find is the vast majority of those TIL actually overlap almost dead on top of where the exhausted T cells are. But there are a few important outliers here. And those outliers, if you look in higher resolution, are actually mapping to the place in this UMAP where our central memory or stem cell memory populations are. If you look in more detail, what you see in the open chromatin landscape of these cells and also some of the other clinical progressors is a lack in 4 of the 6 clinical progressors of a signature of accessibility at cytotoxic genes, but a greater accessibility of genes involved downstream of Wnt signaling. In at least 5 of the 6. In fact, while the cytotoxicity signature is not significant based on the number of patients we have, the Wnt signaling signature is very significant, where it correlates with lack of infiltration into the tumor. Now, Wnt signaling is known to be a negative prognostic marker in a lot of tumors, but what's thought to drive that in other settings is actually either tumor-intrinsic Wnt signaling or actually Wnt signaling in the dendritic cells that causes them to be poor primers of T cells. Here, however, we're seeing this Wnt signaling signature very prominently in the CD8 T cells themselves. That signature correlates actually with activation of Tregs, suggesting that while the CD8 T cells might be negatively regulated by this Wnt signaling signature, the Tregs may actually be benefiting from it. Okay, so I'm going to run out of time to tell you about the third example, but I'll mention it in passing. What we wanted to know is whether any of this was functionally relevant. We have these signatures, we might be be very, very useful biomarkers in the open chromatin landscape, but is any of this actually physically controlling the patterns of gene expression in these cells? So what Josephine realized is that with the number of replicates that we have, the number of cell types we have, and the variance in accessibility at an individual site, we could actually link that to the variance in gene expression and build computational models to predict which open chromatin regions might control gene expression in a cell-type-specific manner. I won't go through all the computational biology, but basically you can derive a lot of really interesting information from this, like each gene is controlled or— yeah, is controlled or associated with 6 to 10 open chromatin regions. Each open chromatin region is associated on average with only one gene. And in fact, each open chromatin region explains on average 25% of the variance in gene expression. There's some that are actually very high, but on average, each accessible chromatin region or enhancer contributes to how a gene is controlled. So you can build these really interesting maps for different genes of which open chromatin regions are predicted to have the most influence on control of expression of that gene. And you can do this because you can look at the variance in accessibility here and relate that to the variance in RNA expression across these different cell types. So you can create these maps, in this case, of these 5 open chromatin regions that might contribute. You can then see that the model predicts that actually site A and B control 50% of the variance in expression in CXCR3, which site— with sites C, D, and E controlling a smaller amount of the variance. For granzyme B, you can do the same thing. You actually predict that the top 4 sites control about 60% of the variance in gene expression, with really the first site controlling a little bit more than the next 3. This model actually works really well. You can do sort of leave-one-out controls. We had cells that we had sorted that we didn't use for the model that we could use, and the correlation is very good. But you can also do a leave one sample out, redo the model, and then ask how the model predicts that missing sample. And actually, the p-values are quite strong, quite robust across models. You can do this for lots of other genes. I'm going to skip this here. I will just mention in passing and then skip to the last experiment. One of the other ways that you can use this is to actually overlap with SNPs, with genetic variants. Genetic variants occur mostly in non-coding regions of the genome, and the thinking is that they probably control or probably influence open chromatin sites or enhancer elements in the way those enhancer elements function. So you take this now whole database and take the, the open chromatin regions and ask which SNPs map to those open chromatin regions and then also derive where those open chromatin regions are most accessible in which T-cell subset and then perhaps derive information about the relevance of that subset for disease. So I'm going to skip over that for now. Okay, so can we functionally validate any of this? So the modeling is great, but can we actually directly test whether these open chromatin regions that predicted to control gene expression actually do? So what we did is we designed a way where we could introduce a guide RNA to one of those open chromatin regions and and drag using dead Cas9 a chromatin repressor, KRAB domain. So targeting an open chromatin region predicted to control gene expression and then closing that region with KRAB, does that influence gene expression? This is an in vitro co-transduction. We're looking for our cells in the upper right quadrant here that are going to be co-transduced with a guide to either the promoter or the enhancer and ask what happens to gene expression. We first picked CXCR3 because it's actually a very simple gene architecture. We have the promoter and we have these two enhancers, A and B, that are predicted to control 50% of the variance in gene expression. We targeted one guide to the promoter. We chose not to try enhancer B because it's too close to the promoter and the KRAB repression is going to spread into the promoter. But we tested this enhancer element up here about 2.5 kb upstream our 3 kb upstream to see how that worked. So here, if you look at targeting the promoter, we can actually use protein expression here because we can stain for CXCR3. If you close the promoter, actually protein expression goes down dramatically. If you target that candidate enhancer upstream, you actually see that you also dramatically reduce expression of protein. In fact, if you quantify this at the RNA level, closing that one enhancer actually reduces RNA expression almost as well as targeting the promoter. Targeting those guides to some other part of the genome doesn't have any effect. You can do the same thing for Granzyme B, where the architecture is just slightly more complex. You actually have 4 sites that we targeted on the promoter. You get good repression of gene expression. And then actually these first 2 sites, B and C, that are actually further from the promoter, give us better repression of gene expression than site D, here, which seems to have a relatively modest effect on control of gene expression. So I know that was probably talking too fast, but let me just sort of slow down and give sort of a summary here. So this human T-cell subset atlas can resolve differentiation states and also point us to the underlying transcriptional circuitry. We think that it also points out how epigenetic identity is more robust and contains more information than gene expression. We think that's because cells with a similar epigenetic landscape can actually exist in multiple RNA states just based on extracellular cues and turning genes on and off. Human exhausted T cells, even in the peripheral blood of healthy humans, are distinct from other T cell subsets. And human TIL in at least a couple of different settings have an epigenetic signature of exhausted T cells. I didn't show you, but we can do cross-species mapping to identify some of these conserved enhancer elements that may be more evolutionarily important or fundamental to control of gene expression. We think that this atlas now provides some ground truths for understanding the biology of key T cell subsets and perhaps even for deconvolving complex signatures of cell types where— that don't fit neatly into a box, perhaps allowing us to identify the underlying circuitry that those cells use in combination to achieve some new state. We also think that this epigenetic mapping can identify state-specific enhancers, so controlling gene expression in individual T cell subtypes, and that this open chromatin landscape may, on one hand, be a more stable molecular biomarker or barcode of a cell identity. But we also think this is an opportunity to think about cellular engineering in the non-coding space, if we can use manipulation of these open chromatin regions to guide cells down different paths of different differentiation. All right, so I'll end here. I want to acknowledge Josephine, who did all of that work and who is truly amazing. This is her here in the middle. I started off talking about some of our melanoma work, which was really spearheaded by Alex Huang when he was in the lab, with a lot of help from the Penn melanoma team, a lot of funding from a lot of folks here, Stand Up to Cancer and PISCY most, but also NIH. So I'll stop there. Hopefully I left at least a couple minutes for questions.

Thank you, John. Boy, that's a lot of data. That's just amazing. Tell me, just as long as we're thinking about it here, when you were designing these sorting panels for the different subsets and you chose the exhausted one, what made you go with 39 and PD-1?

I mean, it obviously worked, but it's kind of not what I would have I thought originally— yeah, there are 3 considerations. One, you could have picked any inhibitory receptor, right? Because who knows why one is more important than the other? We still don't know after, you know, almost 20 years of studying this stuff. So that wasn't a reason. The one practical reason is that it's a really clean stain. It separates really well. But the more important biological reason is there are good, very good data in human HIV and hepatitis C, that that population absolutely contains exhausted virus-specific T cells. So this is more that we wanted to know that what we captured really fit that definition. I am positive that we are missing other exhausted T cells in the blood. So I don't think we're capturing all of them. I think if you chose TIGIT and PD-1, you'd get an only partially overlapping Venn diagram and so on down the line. Yeah, that's good.

And I think, and related to that, it's interesting from the gene expression of how the exhausted cells overlap with some of these others, especially the TIL, but in the central memory. And I'm curious, because I know you didn't really go through it, but what is— how do you see it as the developmental trajectory and how they end up exhausted in the tumor that, and how that might relate to what you're seeing in the blood?

Yeah, so there, there, I think, two important points here. One, from a developmental standpoint, in mice, the exhausted T cells sort of branch off that differentiation scheme around the stem cell memory population, and then they go down a separate path from there. And that's actually where Tox plays a role. It fosters that positively, and it represses the development into the effector memory or more effector like populations. So that's the developmental biology question. You know, I think that what that overlap also tells us is that exhausted T cells have to have the survival circuits to withstand stressful environments. And the other cell types that have solved that are the sort of stem cell memory, central memory cells that use sort of at least a partially overlapping set of survival circuits to last a long time. Yeah.

All right, there are a couple of questions here. Some of them— here's one from Anusha, who I think you may know from Piper. If PD-1 therapy reprograms cells transcriptionally, but not at the chromatin level, are they more poised to return to an exhausted state after a certain time period or once PD-1 antibodies are no longer given?

Great question. So we published a paper in 2016 where we looked at this very directly. And that's exactly what happens. After some period of time, they basically collapse back down transcriptionally right back where they were. And we actually can't distinguish them in principal component space transcriptionally. Now, we think that's because you're not getting substantial remodeling of the chromatin landscape. And maybe that's, that's what this is pointing to, is that that's our opportunity, is to coax the cells epigenetically. The second part of the question is really important. Does it go away when you remove PD-1 blockade? It goes away regardless of whether you have PD-1 blockade on board or not. And I think this is a critical question about the pharmacology of PD-1 blockade that we've missed as a field. You can keep patients on anti-PD-1 and you collapse back down to the original transcriptional cell type regardless of continuing to block PD-1. We think that's because the cells basically retune their sort of signaling apparatus to a new set point threshold. And so we're starting to think that continuous blockade of the PD-1 pathway may not be the best strategy because the cells just adapt to that new reality.

So food for thought. Yeah. And then it sort of brings up this idea of interferon temporal nature, right? I mean, you can imagine how chronic interferon signaling and PD-1 versus, you know, the kind of the temporal nature of interferon signaling.

Yeah, how the ISG circuits fit in here we don't understand very well yet.

Yeah, right. I mean, that's just going to be It's obviously important, but beyond. So I think that a couple other questions here. How do exhausted T cells come from the blood?

I think they're sort of asking about trafficking and how do they end up in the tumor? Yeah, we don't really know. What we know from mice, and, you know, take it with a grain of salt because it's it's this LCMV system, but it fits with other biology out there, is that the population of progenitor-exhausted T cells sits in a tissue somewhere. In LCMV, it's sitting in the spleen or the lymph nodes. Every so often, sort of at steady state, some of those cells get activated, get kicked out of the lymphoid environment. They lose TCF1. They actually gain a migratory program, come spilling out into the blood., and then have access to all the tissues because they do express these inflammatory chemokine receptors. So in a steady state, in LCMV at least, and I think Matt Spitzer's data in tumor models would say the same thing, you have this sort of migratory phase of exhausted T cells recirculating and then ending up in tumors or in tissues. Why they pick a tumor rather than a non-tumor tissue may have to do with the inflammatory microenvironment of that tumor. But we can capture these cells in the peripheral blood at steady state in mice and humans as probably part of their normal biology.

Another question from Bill Mu. He's pointing out the NR4A1 might be a marker for the late stage of exhaustion. He was asking if there's any other biomarkers for this early stage or intermediate stage of exhausted T cells from your data.

Yeah, so humans is much harder. In mice, we map this out, like, with a lot of detail. It's in a paper from last year from Jean-Christophe Beltra in Immunity. There are a lot of great markers there for each of these discrete stages, both transcriptional markers and epigenetic markers, and phenoprotein as well. NR4A1 is a pretty good marker of the most terminal because they are getting constant TCR signaling. But interestingly, you also see NERF77 at the first stage. So the most progenitor and most quiescent exhausted T cells are also actively getting TCR signaling, probably sitting next to these XCR1-positive dendritic cells. But they're being held in check, very likely by PD-1 and other inhibitory receptors.

So this whole thing is balanced on a knife's edge in terms of signaling threshold.

And in between there is— in between there you see TCF7 and TCF7 going down. CX3CR1 is a great marker of this intermediate migratory, more effector-like subset.

And then T-BET is there, and then so on. So is it— here's a good one— is it possible to use some of these epigenetic drugs modulators like HDAC inhibitors or maybe hypomethylating agents to intersect or interfere?

Yeah, we were just playing around with this. It's a great question, and it's one that many of us in the field are trying. They're not as good as we would have hoped or predicted. I think most of them that you can buy just commercially are a little bit messy. There are a couple— the EZH2 inhibitors have some effect, and I think some of that's predicted from work that you know, Jeff Bluestone did and Pam Sharman did and others. But we've not found sort of a home run here yet. It may be more amenable to go in to understand the biology, to go in genetically and manipulate some of the epigenetic enzymes genetically rather than use these inhibitors that have PK/PD problems and sort of off-target effects. But the answer to your question is it's a good idea.

The data is still kind of messy. Have you found— I mean, we've certainly played around with this a little bit too, but many of these HDAC inhibitors or hypomethylating agents like decitabine, they kill T cells, like, you know, just a sniff of it.

And in vivo seems to be different, right? So in vivo, maybe it may be a little bit easier to tolerate, but in vitro, when we've done these kind of in vitro screens like you're referring to, like 80% of the compounds we use kill the cells and we can't find a dose that that doesn't kill the cells but still has an epigenetic effect.

Epigenetic effect that closes all chromatin. Ah, that's fantastic. All right, well, let me see here. Basic question on CD39. Are the— oh, this is interesting from our area.

The protein and the mRNA correlated pretty well? Reasonably well. I won't say that we've actually done a really, really careful correlation there over lots of samples, but yeah, I think most of the time the RNA and the protein are pretty well correlated in our hands.

And then there's a couple here. This is actually from Allison, who's a grad student in the transplant group, and Obviously in transplant, having exhausted T cells is beneficial for decreasing alloimmune responses. So she's interested whether there's any potential for some of these chromatin-targeting treatments to enhance the exhausted T cell populations in transplant patients.

Yeah, we're super interested in this. And I think that the answer is, I think, yes. I think autoimmunity is the other place where this would have very similar application. We've not thought too much about the epigenetic drugs there, although that's a good idea. We've thought more about the NFAT targeting, where we actually make things more exhausted by actually keeping NFAT on longer, which is not typically what you do in transplant treatment.

So the NFAT circuit might be a good one to think instead of calcineurin inhibitors.

You're looking for calcineurin activators? Yeah, you know, maybe a calcineurin activator with like an NF-kappaB inhibitor. Or an AP-1 inhibitor. I think it's all about the balance there. And keeping NFAT on, just to fill in the dots here, keeping NFAT on will drive more TOX, which will drive the cells into a more severe state of exhaustion. Interesting.

All right, maybe one last one here. This is from Luke Regan. Do you think that the dead Cas9 CRAB approach to look at enhancer roles may be limited to surface proteins, which stain well in flow, or could intracellular proteins be analyzed as well?

Good question. So the CXCR3 was just a convenient tool. The, the granzyme B we actually just did by RNA. We didn't even look at protein there. So you can certainly do plenty of things by RNA. We've probably done about another 5 to 10 genes, many of which we don't even have protein, you know, antibody markers to. So absolutely, absolutely applies to just about anything you can assay.

That's amazing. Okay, I think we have kept you long enough. Um, so without any ado, we will thank you very much for coming to speak to us today. I know you got one more meeting with Manish, uh, after this in a few minutes, so that means you'll get, you'll get off this, get onto a separate one, practically speaking. But, uh, I do appreciate you coming to talk, and hopefully We will be able to have you back in person one of these times, and certainly at one of the Parker meetings where there will be a drink involved.

Absolutely. Thanks for having me.

Good to see you, Rob. Thank you. Thank you.

All right. Thank you, everybody.