Scientist Stories: Matthew Krummel, Archetypes of Antitumor Immunity

I am missing coming down to Irvine. I think some know that last time I came down, I flew down, I came a night before and I took a surfing lesson. So I definitely am missing both seeing everybody and that sort of adventure. But I think also I want to just emphasize that I hope everybody can take advantage of this virtual format. I think we should be doing it more, probably flying a little bit less, and trying to figure out how to communicate our science to one another in all of the different ways that are possible now, and certainly this virtual format is one. So having said that, what I was planning on doing— and I don't know how you guys do your seminars there— is I was planning on monitoring the chat at a couple key moments at transitions in this talk and, and maybe answering questions in line. So please feel free to be participating. I think one of the things that's very hard that everyone's finding about Zoom is this sense of being sort of just isolated, listening to a an empty voice. So if you have questions or you have things you want to bring forward, put them in the chat, and, uh, and I will— I'll try to respond to as many as I can, um, at those junction points. Um, and obviously we can do questions at the end just like normal. So anyway, thanks very much for inviting me, and I was going to talk about a topic that's become near and dear to us, which is— which I call— we call the Archetypes Project, and it's based on our, our desire to really understand how, how emergent behaviors happen in the immune system. So from my standpoint, and this is built on giants like Mike Kahalan at Irvine, and the work that has been done over the period of time looking at how the understanding of things in real time can help you see how emergent behaviors happen at the next level of scale is a really, really powerful one. And so if you image molecular events, for example, within cells, you can start to understand how those cells behave at the subcellular level. And so each level of size is really an understanding about how emergent behaviors on molecular levels and molecular time scales play out into the next level of understanding. I'm going to pause here in a moment. I'm going to show you a movie that's done with lattice light sheet that's going to encompass what I want to talk about today, and that is the sort of the partnerships between T cells and antigen-presenting cells, or really between immune cells and one another. So this is a lattice light sheet image. This is done with a microscope that essentially does sort of turf-level resolution, but in 3 dimensions. And what you're seeing in green is a T cell, and many of you may not see a T cell like this normally, but in very high resolution, you can see the microvilli, the little finger-like protrusions by which a T cell palpates opposing surfaces to find antigen-presenting cells. We think it's how they get past the glycocalyx. It's, you know, a big part of every mammalian cell. And in red is a dendritic cell, and dendritic cells actually are better described as having veils. So this kind of ear-like projection you can see over here is the veil that looks like a dendrite in some dimensions. And so as this movie plays, you'll see the palpation of the T cell. This is what we think about, you know, T cells trying to find antigen-presenting cells, and you can see how it's pushing on the surface with little different fingers. But if you break this apart, I think the thing that also becomes apparent is how much the antigen-presenting cell is melding itself to the surface of the T cell. And what's important about that, I think, is to remember that when you want to be a system that's about communication, there's this partnership where the cell, on one hand, this cell can push into this to find information, but this one may want to have that information be seen. And so really, when we think about why the immune system works, it's because these partnerships work. Okay, so going up a level of that, once you start to think about how cells work, then you're really into the domain of thinking about how they work in tissues, how they work as aggregates and collections. This black area here is a tumor surrounded by a series of immune cells, and I'll come back to that a little bit later. And then the final emergent behavior of how that happens is really that of the— you can call it the feet up, feet down, the clinical response. And this would be survival of cancer patients over time, here with checkpoint blockade, giving you the long tail, about 10 to 20% of people with CTLA-4 alone, for example, anti-CTLA-4, or anti-CTLA-4 plus PD-1, giving now in melanoma about 50% of people people responding. And that's obviously a big driver for our interest in this entire spectrum of biology. And today I'm going to focus on the connection really between the multicellular behaviors and the clinical. And when I usually talk about this, I sometimes do or don't show this particular slide that reminds me why it's important to image things and to look sort of at things that you really may think you understand but probably don't. And I'm reminded then of a slide that Art Weiss put together in the '90s that showed all the T cell receptor molecules that were known, the T cell receptor itself, CD4, LFA-1, CD28, et cetera, and their cognate receptors on antigen-presenting cells. And to be honest, this was it. This is what we knew. If you do this same slide right now, we have hundreds of receptors, and to be honest, they're all being targeted by drug companies to see if we can make T cell responses better or worse. But a little one came along that was CTLA-4, and And I'm going to rush through this, but the point is that once you make an antibody to something, you can show, for example, that CTLA-4, if you're stimulating T cells with a T cell receptor signal and you add in CD28 signal, you can get them to co-stimulate, which is what was just being understood at the time. But if you start to titrate in on this axis, CTLA-4 signal, then you could really drive things back down. And that really was the basis for which it made it simple to say, well, what would happen if we put that antibody into mice? And this was an experiment I did that's essentially looking at the response of V-beta T cells that are responding to superantigen. They normally go up over days. This is around 3 days here. They peak, the T cell response peaks, and then it wanes back down. It's pretty typical of immune responses. This is with CTLA-4 blockade. You see things get much better. CD28 blockade, they do much worse. And it's that basis by which, you know, we basically were able to do the experiment to to melt tumors. So here's tumors growing, and these are tumors with Cezolite4D repressed. And of course, that's the source of the kind of emergent behavior I just told you. When I tell this, I like to tell people that the period of time between this conception and these first antibodies was 3 or 4 years of painful postdoc or graduate work. But the period between this was months. And so, you know, the idea that you want to know how something works and then you start to see how it works means that you very quickly can get really good grip on things. And I kind of also say that to motivate graduate students when we give seminars to remind you that, you know, sometimes this long pushing slog to get to these results seem like they last forever, but if you work at it, you can get further along. So today for the tumor microenvironment, when we think about how cells work in collections, it's a little bit like 1993 was for T cells. So I could make you a cartoon that's way more modern than it would have been 5 or 10 years ago, where the tumor now is surrounded by a collection of immune populations. And I can call some of them out. I can say a green cell's a T cell, and a blue cell's a dendritic cell, and purple is a macrophage. But this is going to look as silly as the one that I just showed you from 1993 in a few years, as we start to understand really how these things are arrayed and organized and what the composition really is doing. And it's not until you start to look at these that you start to understand. Now, a lot of us are motivated by now, at this point, by sort of the value of the immune system in oncology. But what I'm going to talk about today, even though I'm going to really focus on, you know, immuno-oncology for most of this, I think one of the things that have motivated us in the last 5 or 10 years is how many cool things has been going on in the immune system outside immuno-oncology. And the value those have to us when we think about what the immune system is about. And so just to give you a sort of sense of that, in the last 5 or 10 years, it's become really apparent that the immune system is essential for normal brain function. This is not about defense, though there are components, obviously, of the immune system that matter for that. But you, everyone on this talk, will not remember the talk I'm giving today if it weren't for macrophages in your brain pruning your synapses. This is not traditional sort of vaccination immunology. This is not tolerance. This is a function of the immune system that we didn't understand before, and we're only now beginning to understand thanks to some of the work that started at Ben Baris at Stanford, for example, with, with C1q molecules and microglia. And of course, we're also understanding similarly that diseases like Alzheimer's are failures of the immune system to function properly in those settings. This doesn't stop with the brain. And in respiratory disease, we've always known that asthma was an immune-mediated, but now we understand that fibrosis fibrosis. Also underneath fibrosis are often Th17 cells and neutrophils. It's not just fibroblasts. A really big one obviously in the last 10 years, 15 years, is the emergence of the microbiota. It's not just that we have a small number of microbiota. We— our immune system essentially has T-cell receptors against most of them and curates them, really keeps collections of them that allow us to have a much broader genome than we possibly could have if we just had the 20,000 or so odd genes that we're born with. Arteriosclerosis, another example. Anyway, the list goes on and on to the point where we're seeing that the immune system has these really divergent functions that are beyond the ones we understand. And a year ago, Adriana Mujal, a graduate student in my lab, and I wrote a perspective on this for Science where we remind ourselves that in the past, if you read textbooks about the immune system, you read this kind of idea that the immune system goes, oscillates in a single line between tolerance, the point where your immune system just doesn't know about things or doesn't respond to them, or destruction or immunity, where your immune system is eliminating cells. And what we're realizing, and what I just showed you, is this huge spectrum that lies between those two poles that consists of, for example, an immune system that regulates commensals by having Tregs and Th17s and cells that are against those. Some of those same cells like Tregs can be involved in another kind of response. For example, we understand now that Tregs and myeloid cells are essential for fat metabolism. So you find them in adipose and you find them in the liver. Macrophages together with T cells can be involved, as, as Zena Werblab has shown, in, for example, mammary development. And I mentioned the brain. So what we've started to think of is moving up the emergent behavior spectrum that I introduced towards something we call archetypes, which here we're defining as not just a single cell type or a single T cell specificity, collections of cells that exist in linked states across tissue types, probably following some sort of evolutionary design that's related to the things that are shown at the top of these. And it's really a higher level of abstraction than a single cell. It's kind of a recurrent motif. And as we go into cancer immunotherapy and we think about the immune system that's been built around tumors, it really becomes important to understand which of these, if one or more, the immune system has tapped. So while that's going on in this sort of inner zone, the things that we really want in tumors consist of various forms of antiviral immune responses or ones that we traditionally classified in that way. So for example, I'm going to show you some of the data on CD8 T cells and CDC1s and NK cells as a collection of cells that together give you the potential for antitumor immunity. And we'll look at that as it sort of relates to the idea that there's other ways to get rid of tumors that might involve, for example, CD4 T cells, stimulatory CDC2s and others. Other populations. So this idea of thinking of the immune system as potential modulatory sort of allied systems is what we're referring to here as these archetypes. So this started for us in— about 7 or 8 years ago, we first published on this when we were trying to figure out whether there were any actual allies in the tumor microenvironment. At that time, we knew of macrophages as being sort of suppressive which we still don't fully understand what we mean by that. But we knew the T cells were in the tumor microenvironment, and we've known for 30-odd years that the T cells that you pull out of tumors are tumor-specific. So the question was, are there any allies, or were there any allies in the tumor microenvironment that could actually stimulate T cells? And this is an experiment that Miranda did that really shone light on a cell population called conventional dendritic cell type 1, CC1s, and in mouse they express this gene, CD103. And the result was that if she took out all the different myeloid and B-cell populations that she could get out of tumors and used them to stimulate T cells that were against an antigen that was expressed by the tumor— in other words, asked which one was actually a bona fide APC— it was really only the CDC-1s that would drive upregulation of Nrf77 or CD69 in an in vitro assay, whereas all of the other ones failed to do so. And she showed that those cells are primary producers of things like IL-12. They can upregulate CCR7 as they go to the lymph node. And they really represent, you know, what we think of as a primary ally. Importantly, if you get a signature for the genes expressed by this and you throw this at human datasets and look at overall survival versus time, you find that the people that have large collections of gene expression suggestive of CDC-1s tend to do a lot better even without therapy than those without. And there's been a really a cottage industry. Lisa Cousins and Brian Raffael published about the same time as us, same time as us, also showing these populations. And Brian has built a really nice set of stories about those. Nouria Morad, David Sancho, Tom Gajewski, Kitana Reis-Issouza have all followed with these. And this is really a big paradigm now that the CDC-1s are a big part of a good immune response. And what I'm going to show you very briefly is the way that we've been building on this as an archetype. So we knew that these cells were important, you know, from some of the data that I just showed you. And we knew it from experiments that essentially said that those— the cells, the CC1s in the tumor microenvironment here in the top ultimately travel to the draining lymph node where they can act as APCs for new CD8s. And they also have a secondary role that when the CD8s come back, they can re-interact, interact with those CDC-1s. So this is the tumor microenvironment again as the CD8s come back to it. Those CDC-1s act as essential APCs to keep the CD8 T cell response going. So we wanted to understand more about that. And one of the ways that we've done that has been something that Francesco alluded to is that we've really been trying to work together with a clinic at UCSF to build ways to profile immune responses. And as he alluded to, we've really been trying to get, and we've really been successful at getting, a lot of the groups that study everything from fibrosis to Alzheimer's to do some standard profiling of the tissue so that we can start to see where these immune states are similar. But the first step in that was just to take tumor biopsies. And to dissociate them and to do flow cytometry to quantify all the different populations you could and to do flow cytometry also to do sorting of the populations into individual populations like T cells and myeloid cells and stromal cells and tumor cells themselves and do RNA-seq. And that became valuable in this question that I was just telling you about because we were able to take from a series of melanoma biopsies shown here Every patient is a dot. And we were able to ask if we just look at the population frequency of the CDC1s, what genes when expressed in the tumor microenvironment mirror that, correlate with that? So this is an example of a gene called FLT3 ligand, which is interestingly the ligand that is involved in making these cells. And this is expression in the tumor microenvironment. You see this nice linear relationship. So this was really intriguing to us because it implied that within the tumor microenvironment, you could not only have these cell populations in these people that were at higher levels, and those people would live longer, and they'd respond better to therapy, but you also had some cell that was making this FLT3 ligand. You had an ally for the ally. And so it just so happened that we were— we just, for kind of almost different reasons, had made a Ritz-Loxley-style FLT3 ligand reporter mouse. And that's basically using the FLT3 ligand locus and importantly inserting TFP, the teal fluorescent protein. And so when we put tumors into mice that bore this, we found really no expression in the cell types that people had already postulated that FLT3 ligand would be important in. But we did find them really highly expressed in lymphocytes. And so the green here is the TFP and the TFP reporter, and the gray is a control mouse, non-transgenic or non-knock-in. And this is just quantifying that. So you can see that the NK cells, you found profound RNA expression for the Flitzer ligand, and CD4s and CD8s some, and B cells a tiny bit. So this brought up the idea that the very cell types that CDCC1s should try to stimulate, notably CD8s and NK cells, might actually then reciprocally carry the ligand that keeps those guys in large numbers. So we wanted to test that. So we're in mice now. And there's a convenient strain, IL-2 receptor gamma C knockout, the common chain that loses all lymphocytes. And if you give those tumors, you find the IL-2 receptor gamma C knockouts really don't make CD103 dendritic cells. So you do need lymphocytes to get that population. You don't, as a control, for example, need those to get other kinds of dendritic cells like CD11b's. Can reduce that down if you use the RAG knockout mouse. You will have a mouse that doesn't make T cells or B cells. That still can make NK cells. And in those settings, there's really no effect. So we get similar frequencies of the CD103 dendritic cells, and again, the control populations in the RAG. So that says that, that it's looking like it might be NK cells. And indeed, if you do an NK1.1 depletion, which is not complete but is a pretty decent way to go, you get a pretty good diminution in the CD103 dendritic cells again. So it looks like NK cells are required. And there's a few other experiments that were done in that paper that that also support this. But perhaps the most interesting part of this in respect to the idea of the archetype then is that if you go back to the frequency of those cells that I mentioned that we get from the biopsies and you look in melanoma, you find this very nice relationship that if you have more CDC1s, you tend to have more NK cells. This is a ridiculously good relationship in head and neck with a Pearson value of 0.9, which in clinical data is very, very hard to see. So with the big dataset that we've been collecting about immune sort of systems and compositions in tumors, we could also do the following analyses. And this is essentially where we took melanoma patients that we would profile their biopsy, and then they would go on anti-PD-1. And we could basically binarize them into a 1 or a 0. They respond to the therapy or not respond. And then we can ask which cell population along this axis— so patients are columns here, and cell populations are rows. And we could ask which population has a very high p-value in predicting or correlating with responsive status. And sort of consistent with what I've been telling you all along, the CDC1s—and I should clarify, they express BDCA3 as their primary antigen that we find them by in human—the non-responders really don't have those, and responders have of, you know, solid numbers of them. And the correlate of NK cells also has a very high— if you look at the p-values over here, this is the relationship, the statistical significance of this relationship tested. And so you have this situation where a lot of the patients are expressing both of those things. So I want to pause here and go back to movies for just one moment because This brings up an interesting problem that most of the tumor microenvironment is somehow supportive of the tumor, but I've just shown you about a population that can exist in them that is sort of immunostimulatory. So there's the question about how these things are parsed out. And the first point to make is that the CTC-1s and the NK cells are really rare as a fraction in general. No matter what you look at them, they're sort of 1 or 2%. But the other is the distribution and the fact that these do look like modules. So this is a movie in which the tumor is in black. This is a mouse model of breast cancer called PyMT Cherry OVA. It's a variant of the famous Muller line, MMTV PyMT. So these mice develop tumors around 6 to 8 months of age, specifically in the mammary. And so here's an emerging tumor. It's black here. It's not given any fluorophore. And what we did is we used two different alleles to effectively parse out the different myeloid populations, and they're named here. The main point you need to understand is that yellow, orange, green— these hues, all of these hues essentially represent forms of tumor-associated macrophages. And the only two red cells in the entire field— and I mentioned dendritic cells are rare— and the only two that you can find are right here. And into this experiment, we had put in OT1, cells that are against the— I should mention, as I probably didn't emphasize, that the tumor expresses ovalbumin, a popular trick. And so the CD8 T cells are visualized there. And so as this movie begins to play, you can see the swarms that we defined some, some years before of CD8 T cells that are swarming around tumor-associated macrophages. And we'll see this for the first 2 or 3 days of residency. The T cells will stay pretty much confined. They do this very slow-range movement. But we also found CD8 T cells around the CD8s, these dendritic cells. And it's actually interesting because it has a slightly different hue of blue in our— we have a 6-channel microscope, and we think this is because CFP is actually pH sensitive. And we know that these cells, these dendritic cells, drive pretty strong signaling fluxes within the CD8 T cells. So the point of this, of showing you this, is that two things can coexist in the immune system at the same time. And I think probably everybody on this call knows that. The most classic example I know is Th1, Th2 biology, where we think of Leishmania responses in some strains going to Th1 and other strains going to Th2, with the actual reality being that when you look in any Th1-biased lesion of Leishmania, you always find Th2 cells. It can vary in number, but you always find them there. Another example would be that even in a really activating immune response, you always find regulatory T cells. And personally, I think that the reason for that is that biology almost always has the seed of the opposition of the destruction of the immune response ready, because at some point the antigen will run out and you're going to want to dampen down that immune response or change it. Okay, so what I've shown you so far with the CD8 story is this idea that there is a definable collection of cell types. And this is— I'm now going back to the cartoon I've shown you before and emphasizing now this top left. That NK cells through the delivery of FLT3 ligand can give rise to this kind of like archetypal selection of cells that work together with CD8 T cells both to prime them and to restimulate them. And for the moment, I'm going to consider that as the beginnings of a bigger archetype that we're now getting the next pieces for that I'm going to call antiviral class I responses. And it's about 30 to 40% of melanoma patients have that. That response. It looks like it's a better immune system overall. It looks like it's also the one that responds to checkpoint blockade from data that I didn't show you. And it really involves this partnership. So the eagle-eyed people that are out there may have noticed that I glossed over these 4 or 5 patients in this relatively small dataset that didn't meet the criteria that I've just I told you about. That is, these people generally have higher numbers of the CDC-1s and the NK cells and their responders. But there's a few responders out here that kind of look like non-responders. They don't necessarily have a lot of those cell populations. But what's interesting about them then is what their next highest association is. What do they have that's sort of in common to one another? And the thing that we noticed right away was that they have tons of CD4 T cells. And correspondingly, they also have a lot of these dendritic cells that had previously been at least hinted as being the more dominant dendritic cell. These are called CDC2s. Unfortunately, to make things confusing, they're marked by BDCA1. They're called CDC2s. And we found that those were pretty highly expressed. And I'm going to show you a little bit about that archetype now. So that's another group of people that respond that really don't match what I just told you. So in order to move forward into that, I have to tell you one other thing is that this is largely melanoma. And melanoma doesn't have a lot of regulatory T cells. That's just a fact that's established. And on the other side of things, head and neck tumors, as an example, can often have up to 50% of their T cells are regulatory T cells. So you'll see why that's important, that distinction between these two classes or these two types of cancer. Because I'm going to tell you about this— we've just gone over this archetype, the CD8 response that involves NK cells and CC1s, but I'm going to tell you about another one that involves now Cdc2s that are able similarly to go to the lymph nodes, stimulate CD4 T cells, but that their ability to stimulate them is really gated by regulatory T cells. So I'm not going to show you any of the mouse data, I'm just going to jump in a moment to the human correlate of this to explain this, but the result in mouse was pretty clear that if regulatory T cells are present, they suppress these Cdc2s not so much their complete ability to traffic, but when they traffic, they produce dud CD4 T cells. And when you get rid of Tregs, these CD4 T cells are made in higher numbers, and they come out with a phenotype that's much better in terms of being ICOS high and PD-1 low, and they lead to rejection. And so a mouse model basically produced sort of a picture of that archetype. And we started to then go back into now head and neck cancer where we have Tregs can be present. If you looked on that graph that I just showed you with the dots for the head and neck, there's a pretty wide distribution. Some people have low numbers of Tregs and some people have high. And so now what I'm going to show you is sort of an elaboration on, again, the sort of diversity of tumor microenvironments that exist out there. So on this axis in head and neck patients, are the presence of these cDC2s. These are the, again, the dendritic cells that can stimulate CD4s and are subject themselves to regulation by Tregs. And on this axis are the percentage of Tregs. I'm going to use this, I'm going to walk through the various different panels of this slide, but I'm going to use it as a demonstration of something that also has been a conundrum. And that has been that in the past, if you look in tumors, and you ask about those tumors that are Treg-rich, here is blue. Every dot is a patient again. You can see patients again that are Treg-positive, those that are Treg-negative. And in the Treg-negative, you can see that there's— it very easily parses into those that are high for the CDC2s and those that are low. And we've just color-coded that to make that easy, but I think even without the color coding, you'd be able to see this sort of square matrix. So if we, if we take the patients and we just consider the ones that are, that have CDC2s or don't have CDC2s as a single class, and we call them green, and we plot that down here, we get this result that a lot of people have seen in the last 10 years about Tregs and has been a huge conundrum. And that is that if you look at Treg negatives versus Treg positives, their survival is often very similar. This makes zero sense if you just think of Tregs as being regulatory cells, because you would expect that Treg-positive people would do a lot worse than Treg-negative. In other words, that their immune system would be more dampened. But if you take this apart instead the way that I've just been describing and say, look at this populations that have the CDC2s that they need in the first place— these are the red people— versus those that actually for whatever reason don't even have that, they don't have the ability to build a good CD4 response, then the, the survival of the same patients— it's the same patients that's just now parsed— those people that have cDC2s, have the cells that can stimulate CD4 immunity, do way better than those that are also Treg-negative and don't. In fact, they do way better than the Treg-negatives now. So the world starts to make sense if you understand what it is that's being regulated by Tregs and whether that cell population is there or not in the first place. And as promised, I'll take you through this. So if you look at these same people right? So the blue people are ones that have their Cdc2s, but they're regulating the CD4 numbers. They're here. If you get rid of Tregs effectively, you find the same result that we showed in mice, that you get more CD4s made. If you don't have the Cdc2s to make CD4s, you get less CD4s. So this is fraction of CD4s versus those people. The phenotype is really demonstrably different. So you can really see how the red people, their CD4 populations cluster as being PD-1 low, and typically ICOS high. And you can really see how the blue and the purple— pink people cluster really similarly. And in interest of time, I'll skip this panel, actually. Okay, so this brings up this concept of archetypes of immune systems and tumors in general. So I told you we started with melanoma, the idea of CD8 T cells having partners, CDC1s and NK cells, and that gives good responsiveness and overall responders to checkpoint. And I've just told you about this archetype that's the antiviral class 2, that's CD4 T cells, stimulatory CDC2s, typically few Tregs, and these guys also look like they're responders to checkpoint blockade and have decent overall survival. And then there's this population of regulated versions of this. And then there's the rest of the world. And what's interesting about this is how this parses out then in different diseases. So in melanoma, for example, about 50% of people respond to checkpoint blockade. And those 50% look like they're these two classes. They have people that have decent antiviral class 1 sort of archetype presence, this combination of cells, or similarly for class 2. Head and neck, we never find this antiviral class 1 population. Head and neck seems to be a very CD4-based tumor type. And what we do find instead is a population that's around 30%, 25 to 30% of patients look like they should be responders. And so far, that's about the frequency that we're seeing with checkpoint blockade in head and neck. And then there's a population up here that we could at least imagine that if we could get rid of their Tregs, they would actually convert to this and then they would be responders. This is a way to start to understand cancer then not as a collection of individual diseases as much as individual sort of immune compositions and thinking about how those work. So I'm going to pause here. And just see if— give people a moment. Where's the chat gone? And see if people want to stretch and pose questions. So Mike asks, what are the therapeutic implications for manipulating Tregs or DCs given the results you presented in the Cell paper? Anything cooking in the clinics with adoptive cell therapy or checkpoint inhibitors? Yeah, really good question. So So if you go to the slide that I have up at the moment, there are drugs that are being made. For example, BMS has devised a form of IPI, which is the CTLA-4 drug, that mimics a feature of the antibody that I made that both blocks inhibitory signals and actually turns out to have a secondary effect of depleting Tregs. And there's a separate company that's doing this using the IL-2 receptor antibodies. So to me, those are interesting combination therapies because that could convert this class 2 regulatory component where you have CD4 immunity is the predominant grip you seem to have on the tumor. You have CDC2s being made in those patients, but they're regulated. So those drugs that hit Tregs could be very useful even if you use them at really low dose or you use them in sequence so that you proximate similar to giving traditional checkpoint blockade, you drive down the Tregs even if only temporarily, and in that window of time, you come in with traditional checkpoint blockade. So I think those are examples of therapeutic applications of this that I think we can start to think of rational drug design. And one way I think I conceptualize that is I think of some of these states as being ones that, like puddles that you can jump between. And, you know, this one is a reasonably easy one to imagine how you would transition between that. So as we think about how we're going to get somebody past all the puddles to cure, we know how to cure a certain thing, you know, we know that checkpoint blockade can work in this setting. So then how can we get these patients to jump into this state? Is head— Nilamani, good question. Is head and neck tumors associated with any viral infection? So head and neck, for sure, a good portion of them, sometimes 50% or so, are papillomavirus. And of course, we may see fewer of those in the next generation since we have vaccines that lower that. Now, having said that, the relationship between HPV between positivity and negativity for the things that I'm showing you is relatively limited. There's not a big connection. What there is, and I'm not going to show this data today, but I can tell you, there is an interesting relationship between HPV positivity and negativity and the sort of the purple state. That is the— we kind of think of these ones that I've just shown you as the accessible, you know, checkpoint responsive or sort of the states the archetype that's the little gem that we want to get out. But at the background behind that is the immune system that is there, that's in place, and there are connections between that and whether the patient had a history or shows gene expression still from HPV positivity. Francisco asks, do you know which DCs activate Tregs within tumors? Does it vary by tumor type? So we think it's cDC2s as well. So this is a source of what is almost certainly a bifurcation, that if Cdc2s can activate Cd4s or could activate Tregs, then you have the situation, of course, where the more Tregs you have, they could dominate that Cdc2 such that the Cd4s come out in this sort of crappy state, and that may be evolutionarily really important to make an immune response that is maybe like the one you want against commensals, that curates the— curates the peptide-producing, whatever it is, organism, but doesn't necessarily get into a cytolytic state against it. And the reason we think it's Cdc2 is because if you give GM-CSF, which upregulates Cdc2 numbers, the very first thing that rises is Tregs numbers. And that was in our most recent paper on this. So it's not a direct demonstration, And, you know, I think somebody will have to do the experiments that we did of pulling out CDC2s in the various different populations from tumors and seeing which ones directly stimulate Tregs, which may be an imaging experiment because, of course, they don't proliferate very well. But we think it's CDC2s. Now, whether there's subsets of CDC2s that do, that are maybe a little bit better at stimulating conventional CD4s or not, we don't know. Amit asks, would simply getting more NK cells in the tumor lead to an increase in the right kind of CDC-1 cells, which can then activate T cells to clear tumors? We think so. That's the implication anyway. But there is the question, of course, of what kinds of NK cells and what the NK cells feature list needs to include. And I'll point out that when we published our story about FLT3 ligand, being an important feature of the NK cells' ability to stimulate the CDC-1s. Kitano and Reisisuza published a very parallel paper that says that NK cells and CDC-1s communicate by virtue of sharing chemokine-chemokine receptor pairs. So whereas the CDC-1s are XCR1 positive, the NK cells are XCL1 positive. And so, you know, the idea that I'm pointing out, of course, is that the NK cells may need to be XCL1 positive, and they're not all. So I think getting more NK cells in there may be a solution, but they may have to be a particular sort of differentiation state. Robert asks, where do you think the NK cells are providing the FLT3 ligand, lymph node, tumor, bone marrow? Well, we're pretty sure that it's in the, in the tumor microenvironment by virtue of the fact that that's where the— that's where these cells are, the NK cells are. But you're right, in, you know, in form, we don't know that they don't also play a role, NK cells. And in fact, we don't— we think that CD4s and CD8 cells that express Flutirigin can also be involved in generating CDC-1s, but elsewhere. They don't seem to be essential for doing it in the tumor microenvironment. So we think it's happening there. And this brings up the question of how much we don't or do understand about FLT3 ligand. In most studies in the past, it's been produced as a soluble cytokine, but 99% of the transcript in cells is actually transmembrane bound. Can be cleaved, which could make it again soluble. But there's, to my mind, I think there's a pretty big question about all the different ways that FLT3 ligand can act locally versus process globally. And Francisco asks, easy experiment to test Roberto's point, treat mice with FTY720, do tumor DC1s decrease? Have we done that? Let's see. I don't think so. Yeah. I can't think of any. Okay, all right, I'm going to move ahead, but particularly students, postdocs, everyone that has questions, feel free to enter them. You can say don't say my name if you don't want to be named. So what I want to move to is a technology story that will be the last part of this talk. And this is has been a longstanding problem. Mike and Francesco and others at Image will recognize this problem, that you can see what you've set out to see in imaging. You can label the things that you're interested in seeing, and you can sometimes see more by inferring the brains of your T cells, the brains of your myeloid cells, the brains of the tissue. But it's always been hard for us to really just to get a holistic sense of of everything that's going on when we watch something. And so I'm going to start you with the sort of the very— one of the more recent ways that we understand cell populations, and that's by high dimensionality reduction, UMAP or t-SNE representations. So here, for example, if we take all of the myeloid cells from a B16 tumor, we can UMAP represent them and really easily call out by knowing our favorite genes the monocytes that lie over here, the tumor-associated macrophages that lie over here, some very sort of stress-activated macrophages that lie here, various CDC populations including the activated CDC-1s that express CCR7. And we can do very cool tools to understand how the cells are sort of exist in continuum with one another. For example, if you do what's called a monocle analysis or a pseudotime analysis, you can You can lay out all the cells and say these are the ones that are most similar and they get different and are different and different and then we start to see these populations play in and then we start to see the tumor-associated macrophages. And you can obviously represent that as a single pseudotime line if you want. So this is the level to which things like single-cell sequencing are allowing us to see the diversity of populations that exist and maybe how they're most similar either by some aggregate measure of gene expression or if you want you can have these obviously anchored on your favorite genes. The problem is that if you go into image, into imaging space, you can put— only put so many colors typically into the imaging space. And you can see that in a, for example, a colorectal— this is a human colorectal biopsy, and here's some antibodies that we use to live image this. You can find that the tumor microenvironment is really different depending on where you look. And so here's an example of a space where the HLA-DR-positive myeloid cells are really distinctly segregated from the CD8 cells. Here they're more mixed. Here they're more mixed, but they have collagen surrounding them. There's at least 3 microenvironments that you can pull out just really quickly in there. And it— and then the question is like, which of the populations that I just told you about that are distinct are represented in here? So we wanted to understand the relationship between identity of the populations in the tissue and and their neighbors, and we wanted to understand the spatial localization of gene expression. And single-cell sequencing doesn't allow you to get spatial. So what we devised is a way that we can put barcodes onto every cell, print a barcode that later tells us where that cell came from, do whatever imaging we want either before or after that, and then dissociate for single-cell sequencing. We can go back and say, what was that cell thinking? Or who were its neighbors and what were they thinking? Why did this biology happen here and not over here? So the way that we've done that is to essentially take advantage of, um, of an antibody or a linoleic acid, uh, moiety to insert a heteroduplex piece of DNA to which we can print and base pair, uh, other oligos on demand. And the way that we make this on demand is that the 5' overhang of the oligonucleotide that we first label in the cells is multiply blocked with a photocage. And so what that effectively means is that were we to bring in this O1' region that would normally bind to the red O1 region here, it won't bind because it's blocked. So if we lay this— a lawn of cells down with all this on it, you know, with any— we wouldn't be able to print them. But what we can take advantage of is spatial illumination in the microscope. So normally when you illuminate, you will shine a beam of light onto your sample that, you know, evenly illuminates the sample, say with blue light here. And then if it fluoresces, it comes back and goes through a dichroic to a camera as green light, the fluorescence wavelength. But the same microscope that's meant to image in this this way can also send an image. So for example, if we send an ultraviolet 365 light into a digital micromirror, the same kind that would be used in your LCD projector if we were live right now, and you can send a pattern of on-dark, light-dark, light-dark that is spatially constricted to the sample, and it will look like this. So the various regions will be illuminated and other regions will be not so well illuminated. And so that allows us basically to do printing on demand into various different spaces. So for example, if you take a rectangular region and you illuminate only the left third and add your first zip code, the illumination will strip off the photo cage, and now this zip code 1, SI5, can label here. We follow that by an illumination in the middle third and add a red zip code labels here and green labels here. And by the way, these are— this is a lawn of T cells, so these individual guys are floaters, T cells that have left the lawn and gone to migrate away. So this allows us basically to isolate cells that are in different positions. And what's really cool is you can do this as a chain reaction. So I showed you where the photo cages are initially laid down. If the incoming print the zip code that carries obviously a sequence that we'll be able to use later, can contain the fluorophore that I just showed you, what have you, is itself carries an overhang region that is blocked, then you can come in and do this repeatedly and add on. So the way that works is if you have a lawn of cells and you eliminate exactly half of the lawn, let's say, so the left-hand side, and you bring in a blue zip code that's got its own overhang, you get this. So the left-hand side has been, uh, has been coded now, is being illuminated in the first round. But now if I illuminate again half, but I do it as a bumblebee stripe and bring in a red print, I end up with a nothing, red only, a blue only, and a red blue. And if I do that again and I bring in a log only to the top half, I end up with 8 regions. And the formula for this, by the way, is 2 to the n. Where n is the number of print cycles. So if I wanted to get 256 regions, I would have to do 8 prints. So that's way better than 256 prints. And this works to really arbitrary dimensions. This is printing where, for example, we printed it something more like 20-micron squares. See the blue cells here? They're really only expressing the blue zip code. These ones never got printed to, as you can— if you look at the pattern up here, they don't express any zip codes. These ones are blue-green, green, green only, et cetera. So we can basically— this measurement here is something that we can do in single-cell sequencing space as we pull the cells back out and it can tell us where they came from. So I'll just show you a brief couple proofs of concept and I'll come back to the tumor microenvironment in a few. So the basic zip code workflow is you make a movie, you print barcodes, You dissociate the cells and then you use sequencing to basically both read the RNA of each individual cell, but also the zip codes to tell you where that cell came from. And since that all happens in a single droplet, you end up with them being linked. So you can go back and forth between physical space and gene expression space. I'm going to skip through these. These are sort of the GUIs that we wrote for that. And I'm going to tell you about a model of wound healing as a first run at this. And so in wound healing models, you might take an NIH 3T3 monolayer and scratch it. Over the next 12 hours, gene expression will upregulate and the cells will start to migrate to close the wound. So what we're going to do is we're going to label those regions and watch them. So if you watch the actual migration, you'll see something that the gene expression will, I think, corroborate. So the cells at the leading edge of this migrate and they migrate into the space. But if you look behind the leading edge, you find a lot of cells that are boiling. And they're boiling because they're actually pulling up off the plate and they're dividing, and then they set back down. So the strategy for healing a wound at this level seems to be migrate into the open space and then make more cells behind that border in order to fill in the volume. So we took this at 12 hours, and we printed to the very front one blue barcode, and we printed behind it a red barcode. Those having different actual barcode sequences. So that when we then took the cells and isolated them and put them through the DropSeq, you could really see this difference that the blue cells and the red cells segregated differently. This is just UMAPs made based on differential gene expression, the top 10 principal components. But you can really see the blues and the reds are different. And you can do a volcano plot that'll build really quickly here for you, which is essentially— these are genes— this is the fold change of gene expression versus the p-value for blue versus red. So cells that were at the front of the leading edge of the wound versus those that were behind. And the key feature you find here is a lot of motility-associated genes. An example would be actin, ACTA2. And then if you look just behind that, 100 microns behind that, you found all kinds of genes like centrin and Staph men then that are involved in cell division. So that's sort of borne out by doing GO analysis, gene ontology analysis of this volcano plot, and shows you that the red genes tend to be the ones behind, tend to be involved mitotic cell cycle, cell division, whereas the ones at the front, everything from cell migration motility to extracellular matrix reorganization and indeed wound healing. So we can validate these genes This is just an antibody now against the protein. And if you do a line plot distance from wound edge, you find that Staphman associated with cell division is found behind the wound edge, whereas ACTA2, for example, involved in migration, is located much more higher density near the wound edge. Okay, so how does this play out into tissues? So I'll show you first a lymph node example here where we Similarly to what I just showed you, we put a blue barcode on the outermost ring of cells and then a red one on the innermost cells. This is— we were first imaging, live imaging CD3 and B220. So obviously you see the B-cell follicles out here and you see the T-cell zones on the inside. If you first just lay those out, you do see a preponderance obviously of B-cells that are blue. These are individual cells that are now recovered and they're assigned based on the barcode, they came from the blue region. Indeed, B cells are much higher out there, so that's a nice reality check. But there's other things that you can find then. So for example, I'll just show you one of these, which is that if you look in the T cell zone at in B cells, so we zoom in on the B cell cluster and we can see this beginning of a distinction, more red here and more blue here. If you do the volcano plot again, you find a gene, for example, KLF2, that is much more highly— it's highly expressed in B cells that are in the T cell, in the center of the lymph node. And indeed, we got a mouse from Steve Jameson that expressed GFP-KLF2, and consistent with that, you can find that KLF2 is much more expressed in the center of lymph nodes. And in fact, KLF2 is interesting because it's generally much higher on immune cells always in the center of every organ now that we've looked at, which is a half dozen or so. So again, this is KLF2 over time. I'm going to skip— the same thing happens essentially for T cells where we find CCR7 and KLF2 are much more highly expressed in the center. And I'm going to go to tumor volumes to wrap us up here. So this is now using a model that I told you about before where At day 0, we put in a PyMT cherry ova cell line into the fat pad. It's a mammary fat pad of a mouse. And 14 days later, we put in— so the tumor is now grown— we put in some GFP-labeled OT1 cells. And then 4 days later, we harvest, and it looks like this. So the tumor itself was red, and at the border, you see all kinds of T cells lining up. And you also find some T cells that have already made it in. You barcode those cells that are sort of in the margin, and you separately barcode the ones that are inside, and you do your UMAP analysis, and you can see decent representation of reds and blues in all your major populations. The UMAP obviously divides everybody up into— if you do it this way, we'll find populations, neutrophils, T cells, DCs, et cetera. But what gets interesting is when you divide that into individual cell populations. So if we focus on the T cells, we take the T cell object out and we recluster, what we find is the reds and the blues start to parse a bit. And the— one of the biggest ways they parse from one another is that the reds is on the inside. You find higher exhaustion scores and terminal differentiation scores. And you can see that in genes that many of us know that if you study T cell differentiation. So for example, TCS7, SLAMF6 are the most highly differentially regulated genes in the outside, and things like ID2 and PD-1 itself are much more high on the inside. So this looks like a gradient, or at least a differential distribution of exhausted cells that are in the inside. So terminal exhausted, exhausted, versus ones that are still sort of more stem-like that are on the outside. And meanwhile, if you do the same thing on monocytes and macrophages, you find a gradient of tumor-associated macrophages that moves inward as well. And that's shown here. This is the UMAP you'll get for just the monocytes and the macrophages. Cells at the center, the dots tend to be on the right of this UMAP and blue on the left. If you do an analysis of how these populations lay out, you find that C1q indicative of tumor-associated macrophages tend to be on the right. And the Ly6C indicative of monocytes tends to be on the left. And so if you make a monocle of that, you would go from monocytes that are green here to TAMs here. That's the differentiation trajectory. And if you look at their distribution, there are much more of the monocytes— the monocytes tend to be in the blue zone, which is the margin, and the tumor-associated macrophages tend to be in the center. So it looks like this is core coordinate differentiation of the immune system over space that one can sort of tease out in this way. So I'm going to end there because I think I'm certainly at or possibly even over time. I apologize. We took a few questions earlier. And I just need to really thank a lot of the people in the lab that have done this work. John started the tumor imaging. Miranda did the CDC1 work. Mikhail did the CDC2. Adriana I told you about did the work on the on archetypes in general. Kevin Barry did the most recent CDC1 NK cell work. Ken has done ZipSeq. Arjun has done a lot of the imaging that I showed you. Steve Jameson gave us the KLF2 mouse. Kyle runs our imaging facility. And Vincent really coordinates a large number of people, including clinicians, for the ability to profile material from patients. And again, I'm happy to stay a little bit over if people want to ask more questions. I realize that I've gone over in time and we had some questions intermediate, but why don't we see what the chat gives. And again, one of the nice things about this being virtual is that if people want to connect afterwards, you know, there's really no barrier for us to do that. It's just a matter of time.

Thank you so much, Max. That was awesome, an awesome talk. We do have time for a few questions, if especially students want to— they have curiosities and stuff. I see one appearing. So that is a good question. You go ahead.

Yeah, thanks so much. And I want to emphasize what Francisco brought up is that this is a very low barrier to— if you have a question, there's really not any dumb ones in this setting, and you should really use this as an opportunity to ask the sort of ideas you might have. Marin asks, great— or says, great— thank you, Marin. Great talk. Did you investigate where CDC1, CDC2 cells locate within melanoma or other tumors? So we haven't done so much on CDC2s. We did look in The polyoma middle T is a beautiful model to study from spatial perspective because it ends up being very radially symmetric. Some of the models where you inject them, it's really, you know, unclear what zones, how to name zones, how to call things out. But in the polyoma middle T, it looks a little bit like the model I just showed you. It is, in fact, that model where the center of the tumor radiates outward and you have margins that are really highly enriched for T cells. In that setting, CDC-1s, they're— as I mentioned, they're really rare overall, but they're almost all outside. They're very rare on the inside, and that's why it was a— it's a kind of a unique movie that I showed you where the CDC-1s were having their little niche in one corner and the TAMs, the tumor-associated macrophages, and the T cells, which are the dominant sort archetype immune response that's ongoing, were right up against the tumor itself. And so we've thought that possibly this is also an important feature, that the CDC-1s are not only low in frequency in cancer, and so the dominant immune response is driven by things like tumor-associated macrophages, but that they're also marginalized, literally, that they're pushed to the outside. So if a CD8 T cell wants get restimulated to go and become more and more cytolytic, it essentially has to go, almost leave the tumor, you know, go to the outside and then get restimulated. It's not going to find a lot of good APCs for itself very locally. CDC2s we don't know as much about. We just haven't done as much of the spatial studies of those yet. But that's a really— I think the spatial question is one that is super important. Francesco, I'm sure, Joshua Elkington: concurs, and Mike too, that this idea that immune— some immune populations and their cytokines are like carpet bombed, you know, they're going across the whole entire organism. But some things, it really matters whether one cell and another cell have an axis of communication and, and how profound and frequent that is, you know, I think is a big part of, of why immune systems work. Okay. Roberto, do you think the exhausted T cells positioning stem versus terminal also occur in other tumors like melanoma? We do not know. It's a good point. We've really only done it in this particular model that, as I mentioned, is also pretty highly stylized in terms of being somewhat symmetrical. And there are definitely tumors, especially early ones, where that is harder to work out. I don't know. Thanks, Michael. See you tomorrow. And yeah, so this issue of spatial distribution is seemingly really important in terms of why biology works on the whole. Some people say that biology is just spatially organized chemistry. Yep. All right, maybe I—

may I ask a question by voice? It is actually— of course—

easier to me.

A couple of curiosities about the platform that you presented in the last part of the talk. So it looks to me that— I mean, it would be, it would be very, very cool if one could actually apply, you know, all this zip code DNA in vivo, but then I don't know if you can do that sequentially? In other words, can you wash out the first zip code, then inject the second one, and so on and so forth, to recapitulate the experiments you did in vitro?

Yeah, so we can. So, but, you know, I should just be— well, two things, by the way. If anybody's interested in that, this is you too, Francesco, we have a little bit of a— we have a kit that we're making to give people a few, you know, a few oligos in the starting point. You got— obviously, once you know the sequences, you can just build your own. So there is— there are two things that are important about what you said. About the spatial— this— the temporal absolutely works. So you can definitely take tissue and you add one oligo, you uncage where you want, you add oligo 1, and then you wash it out, and then you add oligo 2. So that— there's no, as far as we can send, real difficulty there. There is, however, two things. One is an opportunity, is that if you add a label to the region over here, you can do a slightly different experiment than one we had— we said before. You could wait a while and see if some of those cells migrate and then relabel. And so you can say which cells that were originally in a red zone are now in a green zone, right? Because they'll be red-green. However, the second point is that this is DNA. And DNases are around, and, and DNase is also a component of some people's digestion cocktail because it keeps cells from clumping. So this does not come with no cost. You have to be, you know, judicious. And some tissues have more inherent DNases floating around that will eat up your label, just like any other fluorophore, any other way you label things. So there is a, there's a temporal limit to how many kind of cool things you can do that are sequential where you're using time as part of the labeling method. And there is also just some issues that you need to be aware of when you think about, will my tissue do this? Like if you need to have DNase to do your digestion, then you— well, if this continues to build and we want to do more, we'll just have to start using some of the very expensive. Less digestible oligonucleotides. But for the moment, we're using conventional ones. And so it has, it does have that caveat. Yeah.

I think we have one last question from Kirsten. Hi. And to everybody else, I will forward the further questions to Max directly. So yeah, that's great.

Yeah. And feel free to email me, others that want as well, you can go through Francesco or me if you have questions. So there is something nice about these virtual things is that there's no reason why you can't do that. Okay, does the entire tissue get printed as opposed to a single plane? Ah, good question. And can you use that to study tumor heterogeneity and the immune response to that? Very good question, Kristen. Yes, it— you're limited in single photon excitation as to how far you can even get in. But you do label, as a result, a depth. If you do single photon, you will label, you know, the first layer very well and the second layer a little bit and all that third layer really badly. So that both is an opportunity and a cost. It means that sometimes in tissues when you're looking at things, you're looking at a kind of a cone of cells that goes down from where you originally shone your light. However, It does work in two-photon, so the N-POMs are decent, and we've made a coumarin version, caged version, that is much more amenable to two-photon excitation. We haven't really had enough reason to use it, but you could imagine an XYZ printing then, right, where you print this plane certain patterns, you print this plane different. That, you know, may be something we do in 5 or 10 years. But right now it's a lot more work and it's a lot more total labels. But again, I think the binary scheme helps you there to get more labels out of less. But it's a really good question because light, as you know, does penetrate to varying degrees. All right, well, thank you everyone for coming. I can't see anybody, but thank you so much. And again, via Francesco or directly if you'd like. I'd be happy to answer more questions and chat at more length.

All right, thank you so much, Max. That was awesome and enlightening as well. And we will see you tomorrow for some meetings and meetings with the students as well.

Perfect. Thanks, Francesco.

Thanks a lot, guys. Thanks everyone.

Bye, bye.