2026-04-18-mapping-neural-immune-circuits-spatial-multi-omics

Date: 2026-04-18 Primary topic: single-cell-spatial-omics Cross-referenced from: bioinfo-tools, single-cell-spatial-omics Word count: ~16,508

Auto-generated captions

This transcript is Vimeo auto-generated English. Expect misheard names and technical terms. Use the ripgrep/search approach for exact quotes.

Okay. It's 10 on the dot. So good morning, everyone. It really warms my heart to see so many people in the room talking about cancer neuroscience. A few years back, we had to convince people that nerves exist in cancer. So that's a great thing to see, at least for me personally. My name is Maron. I'm a head and neck surgeon at MD Anderson. I'm also studying the intersection between nerves and cancer. I'm really honored to have Dr. Will Wang and Dr. Logan Walsh here with me. I have no financial disclosure. Just having you here means that you all believe that there are nerves in cancer. So this is the first documentation of nerves in cancer over 100 years ago. I used to say that not a lot has been done since, but in the last five years, we really see a booming of a lot of papers and data that keeps flowing that talks about the role of nerves in cancer and, not less importantly, how nerves actually regulate the anti-tumor immunity. When we're talking about nerves, it's important to remember that nerves come in different forms and flavors. For those of you who don't remember from medical school or physiology class, we have the central nervous system, brain, spine. We have the peripheral nervous system, where about 95% of the human tumors arise. And it splits into the somatosensory system, sensation, motor movement, and also the autonomic nervous system that controls everything in the background, our heart rate, our perspiration. And they have different functions, and they have different modus operandi. Some of them have more electrical activity, some less, and they're using different neurotransmitters. This is a small study that we did at MD Anderson, looking at patients with oropharyngeal cancer, stratifying and trying to identify the different types of nerves that exist in the tumor microenvironment. And we've seen that it's not just that their nerves are important, but the different types are important in order to understand the clinical outcome, the pathological features like invasion, progression. But today, we're going to focus our discussion on the crosstalk between nerves and cancer in the context of immunity. We know that nerves and cancer actually interact in order to support tumor growth and support blood vessels to the tumor, but it also regulates the anti-tumor immune response. So, we're going to hear now from Dr. Walsh about how we're going to analyze those interactions in the tumor microenvironment on the protein level, and then from Dr. Wang on the RNA level. And then we're going to have a QA and going to look how we can also look at clinical databases and bulk approaches in order to validate our findings. So Dr. Walsh. There you go. Thank you. Well, thanks very much. Good morning, everyone. Thank you very much, Dr. Amit, for the invitation to speak at this session, and for this cool title that I didn't provide, but it's a pretty cool title, so I'm very happy with it. I have nothing financial to disclose except for the fact that it kills me to say that I actually don't study nerves and cancer. I know it warms your heart that it's so important. I've recently sort of embarked on the field, thanks to these guys and collaboration through Sebastian Talbot. But I do study spatial biology, and so since this is one of the educational sections, we thought it'd be kind of important to talk a little bit about spatial omics. Really, in the last five years, we've accelerated into this spatial omics era, and everyone seems to be talking about spatial transcriptomics, spatial proteomics, metabolomics, et cetera. And there's a lot of commercial solutions that are available out there, but I think a lot of people dive right in because they're commercially available. But there's a lot of bottlenecks and things that I think people need to consider. So as this is an educational session, we thought it was kind of important to build up this concept of how you're going to look spatially at the tumor immune microenvironment, and then how we can then layer in the nerve part of that for studying anti-tumor immunity. And so this is a review that I wrote a couple of years ago now, which is slightly outdated, but even a few years ago, it highlights just the breadth of commercial tools that are available now to study spatially at the transcriptomic level, the proteomic level, the epigenetics, and metabolomics. And what's kind of cool about these technologies, or some of these technologies, is they're actually non-destructive. In the past, when you wanted to study a tumor, you'd take a section, and then you had to take a next section for the next tool, and the next, and the next. In this context, what you can do is you can actually use these non-destructive technologies, and on the same piece of tissue, you can run spatial transcriptomics, then you can run metabolomics, and then you can run proteomics on the same tumor tissue, which we think is really cool, and we think it adds a lot of value for doing multi-omic profiling. The big bottleneck, as I mentioned, in the field, has been on the analysis side. So these references, a couple of years old, are already completely outdated, I feel. I think what's great is with theincrease in the use of bioRxiv. We have every single week a new bioRxiv paper, it's really hard to actually keep up with, by various groups who are solving a lot of the problems that I'm going to sort of highlight in the next 15 or 20 minutes. And so people are doing amazing work to solve some of the basic challenges with spatial biology. And so I think if you are looking for tools and the latest and greatest, I think you should also consider a lot of the work that's on bioRxiv, and I'll reference a few of those studies as we go forward. And so for the purposes of my talk, just because of the most sort of advanced, and I think it's the broadest category which most people work on, I'm going to focus on looking at spatial transcriptomics and the proteomic side of things. And both of these methods were named Method of the Year for Nature Methods. It's been about five years since spatial transcriptomics was Method of the Year, and then just a little over a year ago, Nature Methods named spatial proteomics. This cover's actually an image from my group. And there's so many different types and instruments and companies, and everyone's pushing you, telling you that it's the best, and you should do this panel, and you should do that panel, and 200 genes and 500 genes and 5,000, five million and five billion. And everyone's just getting worked all the time from every different company about what's the biggest and the best, and what's the difference. And people always ask me, what's the difference in the different things? So I want to just go sort of into broad categories. You have on the spatial transcriptomics side of things, we have these imaging-based or optical in situ transcriptomics, where you're directly visualizing RNA in place, and some examples of those are your MERFISH, your seqFISH, your Xenium, CosMX, et cetera. So you have on one side, you have this sort of direct measurement, and on the flip side, you have a sequencing-based or NGS type of approach, where you have some barcodes, and from those barcodes, you'll then sort of mash up your tissue with those barcodes. You have them fixed in some sort of spatial orientation that you can co-register to them later on, and I'll give you a couple of examples of that later. Examples of that are Visium, Visium HD, Stereo-seq, Slide-seq, et cetera. On the proteomic side of things, again, we sort of have two broad categories. The first one is, again, an optical-based approach where you have either DNA barcodes or fluorophores that are visualized, usually cyclically. So you stain with a few, you wash them off, you stain with a few more, you visualize, you wash them off, and examples of this are your CODEX, PhenoImager Opal, IBEX. And then you have your sort of mass spectrometry-based imaging. So in this case, what you do is your antibodies are labeled with various metal isotopes of a different weight. You stain your tissue with all your antibodies, and then you either excite them or you ablate them, and you measure the amount of antibody using mass spectrometry. So those are the broad categories. So you usually have optical or direct, and examples are IMC and maybe TOF. Okay, so we get these beautiful pictures. So here's just a few pictures that have come out of my group looking at spatial omics in a variety of different settings from different tumor models. You can see on the bottom left corner, we have actually a DRG with a number of different nerves, which I hope warms your heart a little bit. And so we have been working on this, and I think one of the big travesties, though, is that I've seen a number of publications that people have these beautiful clinical cohorts, and you could tell that they ran these commercial technologies, but they just didn't really know what to do with them. And so they put these really sort of sexy pictures in their papers as just sort of an add-on bit, when the reality is, is there's so much rich, amazing information here, and we just need more tools and technology to be able to sort of unlock what's available within these sections. And so in order to do the neuroimmune characterization, we first need to go and consider all the things that are dealing with immune characterization inside of these datasets. And so the workflow is really quite simple, and when I made this slide, I thought, "This is way too stupid for me to put into a presentation here." But the reality is, every single dataset that we have in my lab, and we've done hundreds of them, we first, and the most important thing, is to identify where the cells are, so cell segmentation. You need to know where the cells are and where the boundaries of the cells are in order to then go on to the next step, which is what are the cells or doing immune phenotyping. And this is so important, and I think this is where a lot of groups actually get it wrong, and it affects all of the downstream interpretations. You can have the best clinical cohort you want, but if you don't accurately identify where the cells are and what the cells are, it's going to mess up everything that you infer afterwards. And so think about a piece of tissue. Okay? So tissues are complex. They're very compact. They're dense. And we're often cutting these tissues in just a small two-dimensional slice. Now, when you have some sort of normal piece of tissue, like skin, for example, here, and you can sort of slice it in an orientation that you know, and you have, here's the epidermis, the dermis, the muscle, the fat. You can have some sort of orientation, but in tumors, typically, you're just sort of getting a random disoriented bunch of cells and you're cutting through it. Okay? So you don't really have an orientation, an up, a down. You're just sort of cutting directly a small section through the tumor. So if you're trying to track nerves in that tumor, it's not so simple. Now, imagine that you have a single cell, so this is just sort of to represent a nucleus that exists within any piece of tissue. Depending on where you cut that nucleus, you're going to get a different size or a different piece that could be represented across different scales.And we all know that nuclei aren't also just round circles, right? They come in all kinds of different shapes and sizes, and sort of a classical example of that is neutrophils. So this is an example of a neutrophil where the nuclei is this sort of multi-lobular shape, and here's using one of the sort of leading spatial technologies. You do an H&E of a neutrophil, and then when it segments using its sort of default segmentation tool, what it does is it picks up each of these little lobes to segment into the three. And then you layer on the transcript, so this is using spatial transcriptomics. Of course, you're going to triple the number of neutrophils in your sample if you're counting just neutrophils, but then also you're going to be ascribing certain transcripts to each of these individual cells. And so it's extremely important to understand that getting it right, getting the exact nucleus right is very important. And it's not easy. So I've done this exercise. When I first started doing this work more than five years ago, I actually sat in a room, I brought a few pathologists in the room, and I gave them a couple of these type of challenges, and I wanted to just see what they thought of various images. So this is just nuclei from a tumor, and often I ask, I say, "Okay, so this little yellow spot here, is that one or two cells? What about here? Is that a smudge? Is it a piece of debris? Is it a nuclei that we just caught the tip of? How about over here? Or in this little blob down here, is it two cells? Is it three cells? Is it four?" Usually, we split the room with a bunch of trained pathologists. It's really, really difficult. We have these datasets out there that are sort of ground truth that you benchmark against. But the reality is we don't really know for sure what is exactly ground truth. Okay? These are really complicated problems because, again, you're just randomly slicing through the tip of different nuclei and different cells in the tissue. And so we think that machine learning can solve these things, and there is some wonderful machine tools out there, but often people are just like, "Yeah, I just use machine learning." And as you can see in this example, this is one of the leading machine learning tools that does well for some cells but not for others. And in fact, here's a bunch of machine learning tools on the left, some classical computer vision type tools on the right, just highlighting that they all have their strengths and weaknesses, and they do well in certain places, but none of these tools are perfect. So we went after this about five years ago. We worked on this for a couple of years for our datasets to see if we can develop something that would work quite well. We developed this thing called Circle Cell Segmentation Across Scales. It's not super important other than to say that it does work quite well qualitatively and quantitatively compared to a number of different tools out there. And so suffice to say that there is now a lot of tools that exist for segmentation. Everything from very, very specific tools that are cell-specific, biology-specific, all the way to these sort of broad industry-led models like Segment Everything, SAM from Meta. So we have lots of tools out there that are available, and I just encourage everyone to take a really close look at your segmentation because it really, really, really matters. And because all of these different tools have various layers, so this is just a new tool that we developed, and this is just showing you examples. It doesn't matter what we're doing here, but there's a lot of different computer vision, there's a lot of different math, there's a lot of different cool things that go on within these tools to both segment the nuclei and to segment the cell membranes. We do all these fancy tools and transformations, all this cool computer vision stuff, and we get a nice picture of the nuclei. But we also need to capture for most of these technologies where the cell membrane boundaries are as well. And then beyond capturing these two pieces of information, one of the things I see in the literature quite a bit is this sort of people are torn about whether or not you need to harmonize these two. Meaning, of course, when you're cutting through a 2D piece of tissue, you're going to have cells that are coming from above and below, and you're going to be getting transcripts and proteins that are coming from cells that are not within that plane. Do you include them? Do you not? Do you call them cells? Do you not call them cells? These are really, really difficult challenges. We've been quite conservative by saying that if you don't have a nuclei, if we can't identify it's a cell, we ignore anything from there, but this is just the reality of the type of technologies that we're using. And how do you then attribute those different signals from cells that are in or out of that plane? I'll give you a couple of examples in a moment. Okay. So once you sort of are comfortable or confident that you've identified to the best of your ability where the cells are, the next question is what are the cells? So immune phenotyping. Now it's been well more than 10 years that we've had hundreds of millions of single-cell sequencing datasets or cells and many, many, many datasets out there, and most people are still using just simple dimensional reduction, t-SNEs, UMAPs to identify cell populations. But if you look at this-- I love the name of this paper. I loved it ever since I saw the art of using t-SNEs for spatial transcriptomics-- or sorry, for single-cell technologies. But if you look at the pink and the yellow, are those two different cell types? Are they the same cell type? Are they different activation states? Is it someone just not changing the K means when they do the actual analysis? It's a good question. We don't necessarily know for sure. Now in these contexts or in this context, when you have single-cell sequencing data, you have maybe 1,000 to 2,000 data points per cell. Really clean, high-resolution information in each cell. When you're doing these spatial technologies, you have maybe in the protein, you have maybe 50, 40, 30. Transcriptomics, even with these big 5K panels that you can do for spatial transcriptomics, most cells on average are going to have 100, 200, 300 transcripts per cell. So if it's really difficult to do this kind of thing with 1,000 to 2,000 data points per cell and to identify different cell populations, it's really, really hard to do it with 30.And when we started doing spatial proteomics many years ago, this is what people were doing. They were just doing simple dimensional reduction with 14 lineage markers in their antibody panel, and they're staining with 30 antibodies, which we don't think was very accurate. From the spatial-- Oh, sorry. One thing I'll mention is there are some really great new tools sort of using AI and using these really robust data sets that are available out there for phenotyping. But again, this is on the sort of single-cell sequencing perspective, whereas spatial is actually quite different from this. The one thing that we thought was very similar to this sort of spatial proteomic phenotyping was that if you have maybe 15 or 20 lineage markers in your panel, it was very reminiscent of just sort of doing spectral flow cytometry or some form of flow cytometry where you could actually threshold each of your markers and then do some sort of assignment. And so we came up with this sort of ranked hierarchy approach, which was in the backbone of many of our publications. So this is just one that we published a couple of years ago. Essentially, we do this on a pixel-by-pixel level. So we do a pixel-level cell classification using a certain threshold for every antibody marker. And that means, let's say that you have a pixel that's stained both for CD4 and CD8. Well, if we think that the CD4 or CD8 stain is much cleaner and it's higher up on the hierarchy, it will be attributed to a CD8 versus CD4. And each cell boundary has to have more than 50% of that specific marker to be called that cell. So this is just one approach, but it's got sort of ranked in that type of logic where it's sort of got a flow cytometry type of background. Here's just an example of a picture of taking a raw staining, and this is of a GBM tumor, and separating into about 20 different cell types. A couple control examples, appendix and spleen as well, with the raw staining on the left and the processed images on the right. Okay, so on the transcriptomic side, again, you have anywhere from a few hundred to 5,000, 6,000 transcripts in your antibody panels. You can do this. So this is data from my lab on the right here. This is from neoadjuvant-treated, chemo-naïve treated lung cancer patients. And so we had these sort of non-responders that fit into these different archetypes that we identified in our group, showing that there's these B cell, T cell, macrophage-enriched and non-responding lesions. So you can phenotype them, but this is difficult, and I'll show you a couple of the challenges that come with doing this. If you look on the far right, so this is just an increasing blow-up of how these data sets come off the machine. What you'll see is that each of those little spots represents a single transcript. And so it should be pretty obvious to you right away that the boundaries that you draw are extremely important for which transcripts belong to which cell. This is just a depiction or a figure showing you that, where you can have sort of boundaries that will attribute something as a false positive or a false negative. This is sort of a better picture of it, where if you over- or under-segment, you're going to get an increase or decrease in false positives and false negatives that are attributed to that cell. And when you start doing phenotyping, if you have two cell types that have the wrong transcripts attributed to them, it becomes very difficult to sort of disentangle what's happening in that immune microenvironment. And again, these cells are really, really densely packed up next to each other, so it's not so simple to just-- it's not like you're doing this in cell culture where you have these massive boundary or barriers or spaces between these different cells. So what do you do? Well, there's a number of different tools that have come out. This is one called MYSTIQ. There's ResoIVI, a few fairly cool tools. And what they essentially do is they do some complicated math to look at what is the likelihood that some transcripts should or shouldn't belong to a certain group of neighboring transcripts. And you can do these various sort of stochastic variation inference models. And what they'll do is they'll essentially reattribute these transcripts, or they can redefine the boundaries of where these membranes are being drawn, which I think is really cool and quite necessary actually, in order to accurately try and ascribe these transcripts, ones that are coming perhaps from a different plane and look totally different than the cell that you're interested in. Those can be removed and/or you can redefine the boundaries for any cells that you're interested in. So there is a number of different tools, and if you are doing a lot of this spatial transcriptomics, I really employ you. These are mostly on BioRxiv, but they really clean up the data. We've done a lot of QC in background to see how these tools work and benchmark them against each other. And they're really, really phenomenal in terms of cleaning up some of these data sets to more accurately define the immune microenvironment within the spatial context. From the sequencing-based approach, so that was all sort of the in situ optical approach. You got to keep in mind that how these things work is that you have this sort of continuous lawn of barcodes on a slide. Right? And you take that lawn of barcodes, and you know where those barcodes are, and you just kind of hybridize the tissue to it. And when you do that, you then co-register the tissue, and you know where those barcodes are with respect to where your tissue was. However, the cells themselves are bigger than the individual barcodes. And so how do you then ascribe these individual barcode squares that are about two microns to a given cell? If you do it on the membrane side of things, you can clearly see that there's a number of membranes that will actually be-- or barcodes that will cover multiple membranes. Or you sort of depict it on the right, which is something that we've been doing, again, we're very conservative in this space, where we eliminate anything that doesn't cover a nuclei. So we find nuclear segmentation, and then we look for barcodes that cover just that nuclei. We concatenate those, and we ascribe those to one cell, and everything else we remove. But then again, you're losing a lot of information when you do that. But again, what's the lesser of two evils? So something to consider. If you're doing some of these type of approaches, you're doing these sequencing-based approaches, how do you know which barcodes to ascribe when you co-register your tissue? Something to consider, and I think this approach is actually a really good one.Okay. So I mentioned before that these technologies are great because in a non-destructive way, you can use some of these optical instruments. You can combine them with spatial metabolomics, like a DESI or a TOF SIMS. These are spatial metabolomics that are non-destructive as well, and then you can go on and do spatial proteomics. But the workflow is really sort of the same across the board. So whether you add in some of these nerve markers, which is kind of what we're doing, you follow the same workflow. So all of our work is sort of the same thing. So you build in antibodies or transcripts or genes that are within the context of understanding the tumor microenvironment in your context. And then you add in some of these nerve markers. So here's just a few that we did in some DRG that we've added to some of our panels. But essentially the workflow is the same. So you get all the same outputs. This is just some spatial proteomics that we did where we have 35 markers in here that we also did transcriptomics on these samples as well. But you can see in the top left in pink, we've added a nerve marker, so you can see how it just sort of layers in there. And you can look at the spatial context of any of these markers with respect to the question that you're looking at. And the workflow's the same. So everything we do starts off with just simply population dynamics as sort of the simple low-hanging fruit that you have there. So you do population dynamics, then you do pairwise or ternary structures, and then we finish usually doing more complex cellular neighborhoods. And to give you a couple examples from some of our work, so here's just an example of profiling primary versus metastatic brain tumors. You can see that in metastatic brain tumors, just from the broad population dynamics, they have exceedingly more both myeloid and lymphoid cells just by population numbers. Simple stuff. In lung cancer, we profiled over 500 lung tumors, and we looked at... So this is just, for example, CD8 T cells. You can correlate the amount of CD8 T cells with according to any of these specific clinical markers you're interested in. You can make that more broad and look at these 15 to 20 different immune populations or cancer cells, et cetera, and look at how they correlate with any clinical variables. You can see that on the top from the right. And it's just a bubble plot showing that the difference between the two is larger. It's not really important. The important point is that you can correlate just the population dynamics to anything clinical that you're interested in, and then you can layer in any nerve information that you want. So this is one of the powers of this technique, is just first things first. It's not why you do the spatial, but it is the easiest and simplest low-hanging fruit that you get out of these data sets. So you really do the spatial to look at spatial relationships. So the first thing that we always do is we always look at pairwise interaction. So you can see this is an example where you have two tumors, and if you're looking at, in this case, we're looking at CD8 T cells and cancer cells, and if they're interacting. So you have a cohort of samples that you want to compare A versus B, and you can quite see on the tumors on the left, the green and the white are a lot more interspersed qualitatively than the ones on the right. You just see it. It's just pretty obvious. But the way you quantitate it is quite simple. For any image, you just ask how many white and green are within four pixels of each other or six pixels? Are they what you call them touching? And then you say, in that image, there's 10 T cells touching cancer cells. And then you scramble the image. You randomly scramble where the cells are in the image, but you maintain the same proportion of cells. So you randomly scramble the image, and you ask now, how many are there? Are there more than 10 or less than 10? And you do that permutation 50,000 times, and you can quite simply say, based on random chance, is there more or less likely a chance of those two cells interacting or avoiding each other? And when you do it like that, you can also compare between two groups. And so we have the code for all this on my GitHub. If this is something that you're interested in doing, you can just run this very simple permutation test, and it gives you an absolute quantitative and between two groups of whether something's more or less likely to be interacting with each other. So just an example of this from some of our work. Quite simply, ignore sort of the big plot on the right, but on the left. In brain tumors when you have cancer cells touching CD31, so endothelium, there are significantly more Ki-67 positives that are more likely to be proliferative, suggesting that the tumors are promoting angiogenesis. Or even more complex inference from this type of analysis where when patients came in with brain tumors who had moderate to high edema, so they had swelling in their brain indicative of what you think is sort of a broken-down blood-brain barrier. When cancer cells were touching the vasculature, they had significantly less claudin-5. So just again, doing this pairwise analysis, you can just sort of see what's manifesting ultimately clinically, or you can infer this type of thing, which I think is really powerful from a spatial perspective. We also have a little tool on our GitHub as well if you're interested in doing IO and ternary complex formation. So you can quite simply run this, any three cell types of interest, any antigen-presenting cell that you're interested in, cancer cells, T cells, et cetera, and you can just look for triads or anything that's of interest to you, and then you can correlate it with these nerve markers that you think are of interest within your set, and this is something that we've been doing with our IO cohorts. And then finally, the sort of more complex cellular neighborhood, cellular niche. I don't know what other names people have for this, but essentially what you're looking at is how the tissue is organized at a larger scale. There's a number of different tools that have been developed to do this type of thing. This happens to be one that we're involved in a little bit called Cell Charter. If you want to know how any of these tools work, because they're all quite similar, there's this very nice little in silico talk from Marco. So if you type in Marco Varone and you put in silico talk, it's like an eight-minute video. It's really short.But it's a really nice, beautifully illustrated, and spoken explanation of how essentially you're considering how many neighbors, and what the identity of those neighbors, and how you can change those two variables to understand the sort of topology of any different tumor. So if you're ever confused about how people do these different neighborhood analyses, I highly recommend it. It's a really short video, and really beautiful. And essentially all you're doing is saying, "If I don't look at the individual cell types," and you redefine the tissue as these communities that exist within the tissue, does it give us a different way of looking at how these communities and how they're organized may correlate with outcomes clinically? And so this is just a very simple example of brain tumors, and you can see we've done cellular neighborhood analysis in primary versus metastatic brain tumors again, and just the easy thing is to look at the sort of plots on the right, those little circle plots, and how different organizationally are these primary and metastatic brain tumors. And we've gone on to correlate the number of different clinical outcomes. But the point is that these communities hold a lot of different, very vital information when you're looking at with outcomes that you may not capture on the single-cell level. Okay, so just bringing it back in the last minute or two, just to the overall point. We do have a ton of these different commercial solutions. There are companies here today that you can go down or tomorrow in the vendor showcase and see a lot of these different technologies. They are wonderful, but I think you need to definitely be cognizant of the starting point, how you're segmenting the cells, how you're assigning these different transcripts. Get into the data because there's a wealth of information of these various different tools, and there's a huge power to use spatial biology for discovery. But you have to really, really work the really fundamental part of these data sets because this is the key. The segmentation, the assignment, and how you're actually ascribing these cells ultimately dictates how your data is going to come out in the end. And so just one last example to show you, again, just to warm Dr. Amit's heart. Here's another example of a melanoma tumor with nerve markers looking at vasculature and macrophages, et cetera. So with that, I'm going to pass it to Dr. Wang to give you some actual practical examples of how he uses spatial transcriptomics to actually functionally look at nerves and tumors and what their clinical associations are. Thank you. Thank you, Dr. Walsh. I just encourage you all to submit your questions through this QR code right here. Just scan it, and if you have any questions. Without further ado, Dr. Will Wang. All right. Thanks, Dr. Amit, for the invitation to speak, and Dr. Walsh for teeing it up so nicely. I apologize my voice doesn't normally sound like this. My son got sick a few days ago, so I pretty much knew my fate was sealed at that point. But happy to be here with you anyways and hopefully my voice will hold up for the whole 30 minutes. So as Dr. Walsh mentioned, I'm going to be giving a few examples about how my lab and some collaborating labs have used various spatial multi-omic, mostly transcriptomic, but also some proteomic work I'll show you, to try to understand the role of nerves in the tumor microenvironment. So these are my disclosures. Okay, so the two examples I'm going to focus on today, the first story is a little bit more mature. And the second example is a mix of a lot of ongoing work, both from my lab and some other labs. The first one, we're going to focus on the use of spatial transcriptomics to identify paracrine mediators of this phenomenon called perineural invasion, which many of you may be familiar with. As Dr. Amit discussed in his opening, this is one of the many ways in which cancer cells and nerves can interact, and this is just a particular way that's very histopathologically obvious. So it has been annotated by our pathologists and part of staging and adjuvant treatment decisions for a long, long time. So it's a very useful anchoring place when we were starting to look at how cancer and nerves interact. And then for the second example, I'm going to look at integrating spatial proteomics and transcriptomics, to both be able to do high-fidelity immunophenotyping as well as characterize nerves and cancer cells in the tumor microenvironment to understand how these three different players may be interacting to lead to emergent properties like therapeutic resistance, metastasis, et cetera. Okay. So perineural invasion, as I mentioned, is a histologically apparent form of cancer nerve interaction. These are some images that my grad students took in the lab with some simple immunostaining. So the pink represents a neurofilament, which is going to mark the nerve bundles. This is pancreatic cancer. So I'm a GI radiation oncologist clinically. So most of what I'm going to show you today is going to be focused on pancreatic cancer, although we're also very interested in other tumor types and which of these kind of principles that we find in pancreatic cancer are generalizable and which ones are disease-specific. So here's an example where you see nerves in the tumor microenvironment, but there are no pan cytokeratin green immunostained cancer cells nearby. And then as we progress along the spectrum of severity of perineural invasion, you see abutment. So here you see some cancer cells that are right up against this nerve bundle, complete encirclement, and then finally intraneural invasion, where you actually see these cancer cell glands within the nerve bundle splaying apart these nerve fibers.So we and others have found that perineural invasion is associated with poor clinical outcomes. If you look at epithelial cancers in general, even the ones that are associated the least with incidents of perineural invasion have up to about a third of cases with evidence on histology of PNI. And then, cancers that are near and dear to Dr. Amin and my heart, head and neck, pancreas, very high in terms of the rate of perineural invasion. And this is from a study that we did a few years ago, looking in a pan-cancer perspective. But if you just look at all the GI cancers and you isolate perineural invasion at the bulk histology level, there's definitely a negative association with survival, both overall and disease-free survival. We also found in that study that, this is from TCGA, we looked at bulk transcriptomics, and we were curious whether or not, even at this relatively coarse level of analyzing these tumors, if we could predict perineural invasion. So this is a volcano plot from a DE analysis. On the right, these are genes that are enriched in tumors that were annotated as perineural invasion positive. On the left, these are depleted in those PNI positive tumors. And we were reassured to see that some of these hits matched what was already known in the literature. So we saw some EMT-associated genes like Vimentin and MET and MEST. We also saw some neurotrophins like GDNF and Sema4B. We also then took this transcriptomic data, and we developed a random forest classifier. We did a grid search to optimize parameters, and we ultimately came up with a classifier that could take every one of these tumors and put it into a high or low classifier score category. We then took a bunch of cases, 66 to be exact, based on a power analysis. All of those had been annotated by the TCGA pathologist initially as being PNI negative. But we blinded our pathologist at MGH to these outcomes, so they didn't know if they were PNI positive or negative on initial annotation. And they also didn't know the classifier score. And it turns out that the blinded pathologist ended up overturning more cases from PNI negative to positive that had a high classifier score than a negative classifier score. And here are just four examples of some of the PNI that was missed on initial H&E. This is just the pseudocoloring. These are masks that we put on after the fact. The initial TCGA pathologist did not have the benefit of having these. But some of the cases that our pathologist reviewed that had a high classifier score marked PNI negative initially. You can see, once you see it, it's hard to not see some of these more clear cases of PNI. So we were pretty excited about this. This told us that bulk transcriptomics on some level, seems like there's definitely enough signal there to predict PNI. But there were several limitations with both this study and also a lot of prior studies on perineural invasion if we really want to understand this at a more mechanistic level. So the three limitations that all these studies had, some combination of, were lack of comprehensive examination. A lot of studies, not this one in particular, but would focus on specific pathways that were hypothesized to be involved in perineural invasion. So afferents, efferents, plexins, et cetera, Robo, Slit. On the other hand, most of these studies also lacked either cell type specificity, like ours did, and/or spatial context. Which is really important because, from a mechanistic standpoint, we don't want to just be able to say which tumors have PNI or not. We want to be able to understand what crosstalk is happening between the nerves and the cancer cells that are leading to this kind of morbidity and mortality associated phenomena of PNI. And for that, we really need to be able to separate out regions of the tumor that are engaged in perineural invasion versus those that are not. So to rectify these shortcomings, we collaborated with our pathologists to develop these tissue microarrays of primary human PDAC for spatial PNI analysis. You're looking at these 12 different TMAs here. And within these TMAs, we had matched intrapatient regions where we had nerve control areas with just nerve, no cancer. We had areas where there was both nerve and cancer engaged in what would be histologically defined as perineural invasion, and then areas with cancer cells and the cyan color is representing fibroblasts as labeled by alpha smooth muscle actin, but no nerve. So we had regions like this in each of our 29 patients that went into these 12 TMAs. And as Dr. Walsh really nicely highlighted, we can divide our spatial omics technologies, especially on the transcriptomic side, largely into these two big categories, sequencing-based versus imaging-based. And we leverage both of these in a complementary fashion to answer this question of what paracrine interactions may be facilitating the perineural invasion that we see in our tumors. My grad students wrote a review, now a couple of years ago, that I pulled a couple images from, highlighting examples of these sequencing and imaging-based approaches. So for the sequencing-based approaches, you see a couple of examples here. Dr. Walsh mentioned Visium, SlideSeq, StereoSeq. There's some debate about whether this really fits perfectly into the sequencing-based category, but I think it fits more in this category than imaging-based, this digital spatial profiling or commercially known as Geomix method, which is ROI based. And it is probe based, but at the end of the day, you do have an NGS readout. And then I showed you in our TMAs the three different regions of interest: nerve control, nerve positive, nerve negative. We were also able to segment these into areas of illumination for the purpose of the digital spatial profiling. We ended up doing this because when I was a fellow, we were a beta testing site for Geomix at the Broad Institute, and this, early on, from my testing, was one of the best technologies in terms of working with FFPE tissue at the whole transcriptome level. And because our question is inherently a pathologically anchored question, an ROI selection-based method was not actually a hindrance. It was actually a benefit because we want to be able to select regions with perineural invasion versus not. We don't need to profile in a grid fashion the entire tumor space.So we got a little creative in terms of how we did the optical segmentation. So here's an example of a field of view with nerve invasion, and we wrote some code on top of the GeoMix software where we were able to capture the whole transcriptome profile from the nerve in the middle there. And then in this case, fibroblasts within 100 and 300 microns of the nerve, and then the cancer cells within 100 and 300 microns of the nerve. All right, flipping to the imaging-based approaches, so one of the main limitations, as you can gather from the sequencing-based approaches, is that you don't have single-cell or subcellular resolution, right? Even a technique like Visium HD, where your spot size is subcellular because you're not directly visualizing where the transcripts are relative to the cell segments. And as Dr. Walsh really highlighted very nicely, cell segmentation is really key. But with the sequencing-based approaches, you don't have that to anchor on, so it's pretty much impossible. You can try to make some inferences and do some deconvolution, et cetera, but you can't really assign transcripts to individual cells as accurately with the sequencing-based approaches. So the imaging-based approach is a really nice complement. The trade-off typically is less molecular plex. So whatever your biomolecule interest is, generally, you can't go as high with the imaging-based approaches. There's obviously many exceptions to this now. But on the other hand, you tend to get better spatial resolution. So here are some of the different techniques. Merscope, Xenium, CosMx on the transcriptomic side. And then there's various techniques on the protein side as well, like CODEX, COMET, et cetera. And here's an example from our data. So those same specimens that we ran GeoMix on, we also did spatial molecular imaging. Here's a field of view after segmentation and annotation of intraneural, perineural invasion. You can see these green annotated cancer cells amongst this nerve fiber. The pink cells are going to be mostly glial cells because, of course, we don't have the nerve cell bodies. Those are in ganglia, generally pretty far away from the tumor. But you can still infer where the nerve fibers are and still get some transcriptomic information from the fibers, but mostly the glial cells associated with them. Okay. So amongst the many methods that Dr. Walsh highlighted, my group collaborated with Martin Hemberg's group a couple of years ago and developed a way to predict cell-cell interactions from single-cell resolution spatial transcriptomic data that we called SCOTIA, or Spatially Constrained Optimal Transport Interaction Analysis. This is published, so I won't go into the details of this. But in broad strokes, we first identify adjacent source and target cluster pairs. So you can pick any source or target cell pairs of interest. So cancer Schwann, cancer fibroblast, macrophage fibroblast, whatever the case may be. And even though we're looking at paracrine interactions with this method, our hypothesis was that clusters of source and target cells that are closer together or adjacent are more likely to interact than those that are not. So this limits our search space, and then we use optimal transport theory developed various cost functions. We found that our results were largely robust to the cost function, so we ended up going with the simplest one, which is where the cost of transporting a ligand from a source cell to a receptor on a target cell is going to be proportional to the spatial distance between the cells, their centroids, and inversely proportional to the ligand and receptor gene expression, the product of them. So from this, we can derive the spatially constrained cell-cell interaction score matrix and then do various statistics to predict which ligand receptor interactions are most likely to be happening. I will mention that another layer that you can do on top of this that can be really helpful is you can look at the supposed target cell that is expressing the receptor and look beyond receptor expression. Also look at the downstream signaling pathways that you would expect if that receptor were to be activated by the ligand. And that can also be a really nice sanity check and help increase the likelihood that your predicted interactions are going to be biologically significant. So putting the sequencing-based and imaging-based approaches together, we found some interesting hits. So first, the digital spatial profiling, sequencing-based approach. Just a reminder, this is whole transcriptome data, but multicellular resolution. So we looked at the cancer cells that were most proximal to nerves, and we found that on top of various adhesive extracellular matrix EMT-associated genes, which were very much expected, we also saw this strong enrichment, as you can see from this GSEA plot and in this volcano plot, for a platelet-derived growth factor family hits several members. And this was intriguing. But then we wanted to look at our orthogonal method because with spatial data like single-cell data, you have to really be careful what you run with because depending on how you analyze things, sometimes a lot of spurious results can be had. So we looked at our SMI SCOTIA data. So this was 1,000 plex. We also had some 6,000 plex data, but this was subcellular single-cell resolution. And of the three predicted ligand receptor interactions that were above the significance threshold, only one overlapped with our GeoMix data. And that was PDGFD, PDGFRB. We're very interested in exploring this, both because we saw this in two different orthogonal spatial methods, but also because this pathway had been described as having EMT-inducing effects in cancer cells, in other contexts, even therapeutic resistance, but hadn't really been robustly associated with perineural invasion. So before getting too carried away, we first wanted to look at the protein level data. Everything I showed you before was RNA, so we did some immunostaining for PDGFD. We found that indeed it was enriched in cancer cells proximal to nerves at a statistically significant level. We then decided to do a lot of functional validation to see if this pathway plays a role in perineural invasion. So we first developed a 3D neural invasion assay where we took a ganglia. So I don't have time to get into this today, but sensory nerves as well as sympathetic, but sensory in large partPlay a very important role in the development and progression of pancreatic cancer. So it felt like the best initial model to start with. So we took dorsal root ganglia, which is a major supplier of sensory nerves to pancreatic cancer. We put it in the middle of this Matrigel dome, and we had these tdTomato positive cancer cells on the outside. And if you look at this movie, over time, these tdTomato positive cancer cell aggregates will invade radially into this ganglia in the middle. This is what the projection image looks like over a 72-hour experiment. And then when you take out the ganglia at the end of the experiment, you can see these tdTomato positive cancer cells in the ganglia. So we then turn to the well-established KP mouse model of pancreatic cancer with the addition in the Rosa26 locus of this dCas9 Vpr mNeonGreen cassette, which would allow us to do CRISPR activation, using various guides against PDGFD, the ligand of interest PDGFRB, the receptor, and then a scramble control. And long story short, when we looked at these three different lines in the 3D neural invasion assay, we saw that the ligand overexpressing cancer organoid line had significantly higher invasion index. And this, if we gave a small molecule inhibitor against the receptor PDGFRB, we saw a dose-dependent abrogation of that increase in nerve invasion. And then finally, if we took the receptor overexpressing line, which at baseline doesn't show increased invasion index compared to the scramble, but if we now gave it recombinant PDGFD, we saw an increase in invasion, and then we could also largely abrogate this with the small molecule inhibitor against the receptor. What about in vivo? So we did orthotopic transplants in a syngeneic fashion, and we saw increased invasion area with PDGFD line compared to scramble. We also noticed that in the neurons themselves and the Schwann cells, there was a subpopulation that expressed high levels of the PDGFRB receptor. So we hypothesized that perhaps this signaling axis wasn't just acting on the cancer cells to increase invasion, but also acting on the neurons and the glial cells as well. So we first explored this in the neurons. So we took a Nav1.8 Cre, so this is going to be fairly specific to peripheral sensory neurons, tdTomato mouse. We took their DRGs, we dissociated them, and then we co-cultured them with the various CRISPR activated cancer organoids. And we saw that with scramble versus PDGFD, the PDGFD co-culture showed increased neurite arborization. So you see the neurite length per cell body area is significantly higher when co-cultured with the PDGFD overexpressing line. When we took away the cancer cells, just gave conditioned media from these various lines, we saw the same effect. And this could be, again, blocked with a small molecule inhibitor to the receptor PDGFRB. And finally, if we just gave recombinant PDGFD, we saw a similar dose-dependent effect in neurite outgrowth. Now, in case this was an artifact of these dissociated neurons, we also looked at intact whole ganglia explants. So this is from the same Nav1.8 Cre tdTomato mouse, these DRGs with carrier. You see a certain amount of neurite arborization. When you give recombinant PDGFD, you see increased arborization. This could be reduced with the inhibitor SU16f. All right. Finally, with the glial cells, here is the 3D neural invasion assay again, except now we're using a PLP1 Cre EGFP mouse. So the PLP1 is going to mark the Schwann cells. So all these cyan cells, you're seeing a DRG in the middle, and the Schwann cells are labeled in cyan. And the yellow cancer aggregates you see are radially invading in. And if you zoom in on some of these arrowheads, you'll see that these glial cells, and Richard Wong's group at Memorial Sloan Kettering, has shown that these glial cells can form these neurotracks, and actually help facilitate neural invasion into these ganglia or nerve bundles like the sciatic nerve. We're seeing something similar where we see these glial cells will migrate out, make contact with the nerve-- sorry, with the cancer aggregate, and seemingly pull these cancer cell aggregates towards the DRG. We're now doing some biophysical experiments with optical tweezers and other techniques to actually try to measure the forces exerted between these two cell types. But for purposes of this mechanism, we found that the distance seemingly pulled by the glial cells acting on the cancer cell aggregates was also dependent on PDGFD signaling. And finally, we went in vivo, the simplest model, because we weren't, as other groups have also shown with the orthotopic model, we don't see a lot of intratumoral innervation. We're now doing various techniques to try to overcome this limitation. But a very simple in vivo model, even though this doesn't totally mirror what PNI is, is the sciatic nerve invasion model. So in this model, when we injected our cancer cells and treated with vehicle control or the PDGFRB inhibitor, SU16f, we saw that by day 14, there was a significant inhibition in longitudinal invasion of the cancer cells in the sciatic nerve when you block PDGFD, PDGFRB signaling. So we're now doing this with intrapancreatic tumors, trying to solve that problem of inadequate innervation, as well as looking at genetic ablation models on top of small molecule inhibitors. So just to summarize this first part of the talk, we found that the PDGFD, PDGFRB axis seems to act on cancer cells, neurons, and glial cells to facilitate PNI in this multimodal fashion. And we think this has potentially some therapeutic implications moving forward. Okay. So Dr. Walsh kind of mentioned some of this, but just to kind of make this a nice segue, I also wanted to highlight my thoughts on kind of what the next things in spatial omics that are most directly applicable to what we're interested in, kind of the neuroimmune cancer axis. So of course, there's going to be improvements in performance, sensitivity, plex, and throughput. Really important, right? Some of these experiments that I showed you take like a week to do, depending on your plexity and how many regions of interest you have. If you have to go offline NGS, that takes a couple of days at the very least as well. Higher dimensionality I think is really importantThree-dimensional space to look at these tumor-nerve interactions are really important because nerves are inherently three-dimensional structures. If you're just looking at one 2D plane, you're going to miss a lot of things. I think the fourth dimension is also critically important, temporal dynamics, things beyond just intravital microscopy, which is cumbersome, has a limited field of view, and number of markers you can look at. I think innovation there is really going to help us understand the dynamics of these interactions, not just these static snapshots. Spatial multi-omics, I think, is important from the perspective of more information, but also allowing us to get more out of very scant patient samples. And then finally, clinical translation as we look ahead. More and more studies are starting to look at matched samples, which allow us to infer causality beyond just kind of simple associations. Correlating a lot of what we're finding in situ with these very resource-intensive spatial multi-omics methods to things like liquid biopsies, clinical imaging, radiomics, I think, will become critically important to translate some of these findings in a way that doesn't bankrupt everybody. And then similarly, high-plex discoveries, moving those to low-plex assays. So how much can we glean from H&E if we're going to compare tissue to tissue, or kind of lower plex IF or IHC? Okay. So for the last few minutes, I want to talk about the second example that I highlighted before. So my lab, a few years ago, we started developing approaches to look at spatial proteo-transcriptomics on consecutive sections. Our motivation for doing this initially didn't actually have anything to do with cancer neuroscience, but we were discovering through single-cell, single-nucleus RNA sequencing in some of our early spatial studies that different cancer cell states in pancreatic cancer seem to have very different immune microenvironments around them. But we struggled a little bit at the transcriptomic level because while those characterized the cancer cell states the best, they did not do immunophenotyping justice. We all know with immunophenotyping, flow cytometry, antibody, protein-based methods are still kind of the gold standard. So by combining these two, we felt like we had our best chance at giving the best immune characterization as well as cancer and stromal characterization at the RNA level. So we developed a way to do high-plex spatial proteomics and transcriptomics on consecutive sections, and then we did gland-level integration. So I'll mention the same section in a second, but two and a half years ago when we started this, there weren't really robust methods to do this on the same section. So here's an example. So we started actually with CosMX 1000plex and CODEX, a koala phenocycler 96-plex protein. So pretty high plex at the time, even now, still pretty high plex, but we can obviously do better than this now. But we were able to integrate. So you can see here, this is the CosMX level immunostaining. You can see the glands after alignment with the protein-level data. Here is a pilot that we did. We did two untreated patients, two treated patients, all with pancreatic cancer. Many of them had other spatial or single-nucleus data matched to them. We did 20 to 25 fields of view, and three-quarters to a million cells were annotated. And we made some interesting observations. I'm going to just highlight a lot of examples and really not get into the deeper biology. This is all unpublished right now, but we're very actively working on this. So we actually found what was really interesting was that across these four patients, huge diversity in number and the quality of immune aggregates. So we often think of pancreatic cancer as being a little bit more immune cold, or certainly the immune cells there are relatively suppressed and not very effective at eliminating cancer cells. This doesn't take away from that overall observation, but we were quite surprised by just the diversity. We would see structures that look like more mature tertiary lymphoid structures, B cell aggregates, myeloid aggregates, even mast cell aggregates. And then some of these, and as others have observed, and I'll get to in a few slides, we saw some very prominent examples of nerve association with tertiary lymphoid structures. Or, sorry. Immune aggregates, I should say, more broadly. And across these patients, we also did GSEA analysis, and we saw that they had very different geo terms associated with their immune microenvironment. So various things can be done at the kind of low-hanging fruit level. We were very interested in differential malignant expression based on immune neighborhood composition. So here are just some of the different immune cell states that we were able to identify with this high-plex spatial proteomics. And then our malignant cell states of interest, we had characterized very nicely in detail at the RNA level before. So you can ask questions like, what is the differential expression of malignant cells that are close to macrophages versus further away? So here you're seeing the kernel density of macrophages shown in grayscale, green cancer cells. And then if you look at this question for malignant cell expression as a function of proximity to all these different immune cell types that have been identified by these markers along the rows you start to get all the data encapsulated in this heat map. So lots of questions can start to be answered with a technique like that. We started to make a lot of different observations, and we only had four patients that we had profiled pretty deeply. So one of the directions we've taken this now in a collaboration with MD Anderson, Ryan Park who used to be a resident in the same program as me at Harvard, is now faculty there. We've been collaborating on this 40-plex plus Lunaphore COMET panel that does a really good job of immune phenotyping based on the things that we were interested in from that pilot, as well as nerve phenotyping we're starting to do now as well. And here, just an example of some of these cores from this TMA. So I'm going to show you some of the statistics from the TMA later, but this is allowing us to now validate and further explore some of these associations that we saw. So, 405 different cores from 158 patients, over 4.2 million cells. You can see from our annotation that on the cancer cell side, we see a mixture of classical, basal, neural-like progenitor, as well as unclassified cancer cells. We were quite conservative with the annotation. Only the highest confidence programs were annotated as such. And then on the immune side, because we have really niceProtein level information, we were able to annotate a lot of different immune cell subtypes. This allowed us to ask questions like, in space, what are the different cell types that are associated with classical cancer cells or basal cancer cells or neural-like progenitor? And you can see from this plot some of those associations. We also noticed, as I mentioned, that there were a lot of nerves in our tissue microarrays as well as in our pilot. This is just showing circularity versus area. And then this is an example of a nerve where there's some immune cells in here. There's no tertiary lymphoid structure. This is actually a nerve that's been intraneurally invaded with cancer cells shown here in green, and there's this dense collagen 1A1 positive stroma on the outside. And when we look at nerve bundle spatial correlations, we see again, this is an ANOVA analysis looking at what immune cells, stromal cells are most associated. So we see a strong enrichment in FoxP3-positive cells as well as endothelial cells of Vimentin-positive cells. And some ongoing work right now is we have added a nerve subtyping panel of sympathetic parasympathetic sensory nerves, and we're looking in more detail at immune colocalization. I wanted to highlight this paper that was from another group that we're doing some secondary analysis on. So this was Visium data on a bunch of pancreatic cancer where they had a pretty robust number of tertiary lymphoid structures as well as nerve bundles. And when we did a secondary analysis and did a consensus NMF analysis here, you see that program seven really falls out as expressing a lot of markers characteristic of TLSs and then program 12 of nerves. And here's an example of four patients where on the top here, you can see where the TLS structures are. So the red color is where the TLSs are, and then on the same exact sections, MP12 represents the nerves. So you can see if you just transpose with your mind's eye, you can see that the nerve are associated in many cases with this tertiary lymphoid structure that are seen in these patients. And you can start to look at TLSs that are associated with nerves versus not associated with nerves and what types of immune cells are enriched in each category, which is what this kind of preliminary analysis is showing here. And then just summarizing what the authors showed, and we're providing an extra layer on top of this. They found that non-invaded nerves, so non-cancer invaded nerves, were the ones that were most associated with tertiary lymphoid structures, whereas invaded nerves tended to be associated with a very myeloid or immunosuppressed immune microenvironment. So I'll just leave it at that. But this is another example of how you can use spatial omics to really look at these tumor nerve interactions. And then do I have two minutes or should I stop? If you're going to talk about this one, I'm going to talk about it. Oh, you're going to have it? Okay. Yeah. I'm going to skip. I'm going to let Marwan talk about this. I'm just going to leave you with one thought, just future-looking of what we're able to do now. So I showed you spatial proteo transcriptomics on consecutive sections. We can now do this on the same section. This is some pilot data that we have where we do on the same section, whole transcriptomics using CosMx, and then on the same section, 64plex immuno-oncology panel with protein. The data actually looks quite good. The protein level data looks fantastic. And we were pleasantly surprised, even at this combined multimodal level, we were seeing over 100 million transcripts per cell, 750 to over 1,000 unique genes per cell. And this has allowed us to really improve our cell annotation with RNA-only, where we really struggle with some of the immune subtypes without additional protein information or lowplex RNA combined. But now using a technique called Hirotype, we're able to do much more fine cell annotation of immune subpopulations with RNA and protein. And then we're also now, as Dr. Walsh alluded to, doing 3D spatial proteo transcriptomics. These are some tumor slices that we have fresh from our patients. We've been able to keep these structurally largely intact with minimal apoptosis for three to five days. We can treat these with chemotherapy, radiotherapy, KRAS inhibitors, whatever the case may be. Here, you're seeing day zero, day one, and day three. You see you don't really lose a lot of the architectural features of the tumor. And then in collaboration with Singular, with their G4X platform, we've been able to do serial sectioning, 3D spatial proteo transcriptomics. And here's just an example of what one of these tumor slices after it's been out of the body and treated for several days looks like just four slices here. But the yellow represents some of the lymphoid aggregates, and some of these are TLSs, others are just less mature immune aggregates. So I think this will also really open the door to studying tumor nerve interactions in more detail. All right. Thank you. Sorry for running over time. Thank you so much, Will. Please, again, don't forget to submit your questions. Okay. So what I'm going to talk about is basically the integration of all the spatial analysis and the neuroimmune characteristics that we've seen in tumor microenvironment into the clinical environment. We need to remember that both the nerves and the immune system from an evolutionary standpoint were evolved together to protect the host. There is a lot of shared mechanism that we see in inflammatory response that were carried by nerves. Pain is one of them, but also everything downstream to the nerve activity in the inflamed tissue, and we're considering cancer as an inflamed tissue.And there is a lot of shared receptors and transmitters between nerves and immune cells, and you just heard in the last hour on how to identify those in the tumor microenvironment. Again, from an evolutionary standpoint, those two system have evolved together to protect the host. They're doing this in a very different way. I actually took this slide from one of my colleagues, Dr. Talbot, who is here with us. The immune system is very mobile, it's very specialized, while the nervous system is more fixed, but it allows a faster information transfer. So when you keep those things in mind, you can actually take your cohorts, whatever they are. And this is a case study of a clinical trial that we held in MD Anderson. Patients with surgically resectable skin cancer in which we treated them neoadjuvantly, meaning before surgery with the anti-PD-1. We have pre-treatment, post-treatment samples, and we've run those samples through the entire pipeline that you just heard about, ranging from spatial transcriptomics, proteomics, and I'm going to show you the data in a minute. But those techniques are extremely expensive, and we have access to RNA bulk sequencing, either publicly available or in our core facilities, which is more accessible for most of us. And I want to show you now how we can actually zoom out and take this beautifully generated spatial data and apply it to our bulk techniques. So in this study that we've seen, that we've analyzed, we've seen that about half of the non-responders to immunotherapy had perineural invasion. You heard about perineural invasion. I'm not going to repeat that, but the patients that responded really well had significantly lower rates of perineural invasion, and that kind of raised a flag about whether or not nerves have a role in actually mediating resistance to immunotherapy. So the first thing that we did was to take our bulk RNA and dump it on IPA, Ingenuity Pathway Analysis. You're taking the full transcriptome for RNA-Seq. You're kind of melting the tumor, taking the RNA. And we found that in non-responders, there was actually enrichment of pathways associated with nerve injury, nerve response to injury, and also Schwann cell or glial cell response to injury. That means that just by taking the entire tumor microenvironment and looking at the RNA, you can actually pinpoint if you're looking at the pathways that are either enriched or inhibited in the transcriptional analysis, and you can pinpoint those pathways that might be relevant to your clinical features. Perineural invasion is characterized by damage to those nerves caused by the invasion of the tumor cells into the nerves. We've just seen beautiful images from Will's group. But not less importantly, how the microenvironment respond to it. You need to remember that the nerves do not really exist in the tumor. It's the nerve endings. The soma, and you've heard about segmentation of the nuclei, the soma where the nucleus is, is far, far away. If we're looking at the tumor that evolves at the tip of my finger, the cell body of the nerve is actually about half a meter away. So we're seeing the nerve endings. This is why spatial is very critical for us. But we can actually tease out some cues from bulk approaches like RNA, like the glial cell response, like the immune response, that might indicate that what we're seeing in the spatial is actually real, but not only happening in the nerve, but everything around it. So when we took those samples and we've seen this enrichment, the next thing we asked ourselves, "Okay, if I'll go now and I'll do reverse engineering, I'll figure out the signature of nerve injury, for example, Alzheimer disease, neurodegenerative disease, or axonal degeneration, or ALS, and I'll take those preexisting GO terms, and I'm going to apply them to existing cancer cohort." In this case, we used the TCGA, the Cancer Genome Atlas. So we took the cohort of head and neck patients there, and we did analysis of enrichment for those neurodegenerative nerve injury using bulk RNA. And look how beautifully actually the cohort is stratified. Those patients that had enrichment combined or composite score enrichment for those nerve injury-related, the neurodegeneration-related pathways actually separated from those who didn't. And when we looked at the outcomes, now we saw that those patients that have nerve injury enrichment actually had worse outcome. So the next logical question in the context of immunotherapy was whether or not this is associated with immune response. So you can take your existing cohort now, and you can further stratify it according to the immunity of those tumors. And again, this is all bulk RNA data. No spatial just yet. We took those tumors, the head and neck tumors, and we separated the cohort into all tumors or high immunity and low immunity.And then we did the exact same analysis using the same signature based on the GO terms. Again, reverse engineering, and you're taking the neurodegeneration pathways that exist out there for other neurological diseases, and you apply it on your cohorts. And we found that the phenotype that we've seen, meaning worse outcome when you have nerve injury, applies only when you're looking at high immune or hot tumors, and that told us that there is a strong crosstalk between the nerves and the immune system within the tumor microenvironment. Not a big surprise, but a novel finding in the context of cancer. So we went back to our post-treatment samples to see if this difference actually is being accentuated under the immunological pressure induced by immunotherapy. And we found you see that in non-responders, those samples or, sorry, those transcriptional program are hyperactivated in patient that did not respond. So again, everything I'm showing you now is bulk. So we took cohorts that existed out there of other patients that were treated with PD-1, gastric cancer, melanoma, metastatic melanoma. But this time we used gene enrichment analysis or GESA, for those of you familiar with that, and we did the analysis backward. Instead of looking at every single sample, we looked at those groups of responders, non-responders reported in those studies, applying those signatures, and looking at the gene enrichment there. And we found that in non-responders across the different publicly available cohort, you see an enrichment of neurodegeneration, neural injury signatures compared to responders. So, the next step was for us to actually validate, again, those findings. So now we have bulk data that shows that nerve injury is a thing, and it's important, and we see correlation with non-responders. And you've heard about the NanoString Geomics. So we did a NanoString Geomics, which actually allows you, by different areas of illumination, to focus your beam of interest on either the nerve or what's around the nerve. So this is a sample of a skin cancer. You can see the nerve marker, the NFH here, and you can see that we are one time looking at the nerve and its phenotype using a preexisting panel of analytes or antibodies, and that are specific for neurodegeneration. And then we're looking at what's around it, and we're analyzing the immune cell around it. And when we did that, we found that in non-responders, there is a strong crosstalk between neurodegeneration marker and immunosuppression immune checkpoint markers, and this did not exist in responders. Now, we couldn't do that without the spatial context because, again, when you're doing bulk, you're taking the entire thing, and you can't really look at those niches or neighborhoods. We confirmed that with spatial transcriptomics. This is a Visium image, and you can see that we had three different signatures that, again, we developed based on the RNA and also based on the data from the literature. Nerve injury phenotype, immunosuppressive phenotype, and effective anti-tumor phenotype. And then you don't have even to do a robust analysis. You can qualitatively look at the overlap between the nerve injury and the other immune phenotypes, and you clearly see that nerve injury phenotype correlates really nicely with the immunosuppressive inflammatory response, but not really with the anti-tumor effective immune response. For the sake of time, I'm going to skip that, but those are lunifer images looking at those niches. And what we found in those analysis is that the nerve niche within the tumor microenvironment is composed of many different cells, and there are immune cells. We just heard about the immunosuppressive myeloid cells and also the glial cells, and we should analyze each one of them separately, and we should actually try and go and model this. We just heard about the fourth dimension, which is the temporal dimension, and of course, in humans, it's a little bit challenging. We cannot sit on the tumor and let it grow in patients. So we modeled this in animals, and we induced perineural invasion by injecting those mice at the whisker pad, highly innervated region, and inducing the invasion and monitoring and harvesting the tissues at different time points. And we've seen that as the time goes by, you see more invasion, you see more nerve injury, and you can look at markers like ATF3 or Cjun, which are known for being as a master regulators of nerve response to injury. And then you can also look at the evolution of the immune response around those invaded injured nerves. I want to just step back and talk about using clinical databases that exist out there. You can do that either by usingDifferent sets of gene, and you can see this is TCGA data. The Y-axis is the different tumors that we're looking at, and you can either look at the signatures and see the enrichment of the different signatures, like adrenergic signaling, axonal growth, cholinergic signaling, and so on and so forth. And you can see by looking at the enrichment of those different pathways, what exactly those tumors are addicted to in a sense of neuronal signaling. And you can also do it on the single-gene level, again, looking at families like adrenergic, cholinergic, GABAergic. And this will give you an indication what kind of signals those nerves might be sensitive and susceptible to. And this is important because we can actually target each one of those interactions with preexisting drugs that we have from many years of neurology and psychiatry research. We basically have compounds to either block or stimulate every branch of the nervous system. So once we're going to figure out what those tumors are addicted to in the sense of neuronal signals, we would be able to target that. But it's also important to say that when you're doing this with the bulk signatures or bulk analysis, you don't really know if it's coming from the nerve or from other cells like Schwann cells or immune cells. And by using single-cell sequencing, you can also zoom in a little bit. You're losing the spatial context, but still, you know if the signals that you're looking at, like adrenergic receptors or cholinergic or serotonergic receptors, are existing on the tumor cells or any of the stromal cells around the tumor. This is a beautiful example that Dr. Netto from my lab generated by looking on one end, what those tumors are addicted to, and on the other end, using the drug databases and crossing them to see if we have agents to target those specific pathways in those tumors. And you can see that many of those cancers actually have a pretty good amount of compounds that are publicly available, affordable, not patented anymore, and relatively safe, that can be used to target those different mechanisms of actions within the tumor microenvironment. I'm not going to go any further into that. We had quite a few questions, so I'm going to leave it out there. I'd like to, again, thank our panelists and speakers, and we're going to take questions now. So we had quite a few questions. First question to you, Logan. So- The second one's a good one I have. Okay. So, the first question is, is there a pipeline that QCs multiplex immunofluorescence images for staining quality right at the beginning of phenotyping rather than manual QC? The answer is no. When we started doing this work, we realized to build these big panels, it's quite expensive. When you have one antigen retrieval often for some of these steps, and you want to make sure that antibodies work together with the same antigen retrieval, and antibodies are antibodies, and they work well in some contexts and some in others. What we ended up doing was after we spent a small fortune testing and ordering a bunch of antibodies, one of the students in my group said, just randomly, "What if we just asked one of the antibody companies to send us 1,000 antibodies?" And we kind of laughed about this, and we thought it was a stupid comment at the time and kind of funny. And then I went home that night, and I thought, "What if we just asked them to send us 1,000 antibodies? What do we have to lose?" And we did, and they did. So they sent us five microliters of nearly 1,000 antibodies, which we tested, and we sent them back the images from the immunofluorescent staining and testing. So suffice to say that there is no automated way of doing this, and I think you have to spend the time and really add control tissues to your TMAs. If you're doing this on a tissue microarray basis, add proper control tissues to the TMAs, which allows you to do this simultaneously, which is nice. And really, one of the frustrating things is that sometimes when you have a cohort of tissue, some tissue is preserved better than others. And so you'll realize that on some tissues within your cohort, that antibody will work really well, and on others, it doesn't. And unfortunately, if it's not universally staining well across the tissue, sometimes you either got to eliminate cores or not. So the reality is you really got to spend that time to get into the weeds and do the QC manually, unfortunately. Thank you. We have a question there. Yeah. This is for the second talk, and all the talks were really cool. Thank you. What would it take to rescue T-cell suppression in the nerve-induced setting? It was on the later slides that you presented. You talked about this nerve invasive setting for tertiary lymphoid structures in a non-nerve invasive setting. So what would it take to get TLS in the nerve invasive setting? And what do you imagine it would look like if theoretically you can drop in a preexisting TLS into a nerve-invasive setting? Would it persist, or would it then suppress and sort of die out? Yeah. Thanks for the question. The short answer is we don't really know yet, but I think it's a really interesting question. I think that if you look at the cell types that are surrounding the invaded nerves, right? And Dr. Amit kind of alluded to this, too. There's a lot of interest right now in nerve injury, which can be both induced by cancer itselfWhether it's invading or not, right? There can be some longer-distance signaling. It seems like that induces these nerve injury patterns even without physical invasion. But also a lot of the therapies that we give, like chemotherapy, can be very damaging to the nerves, for example, and that can trigger reprogramming of some of the myeloid cells around, right? So we see these Spp1-positive macrophages, for example, that become very enriched around these injured nerves, and that can ultimately lead to some of the downstream consequences you're alluding to, T-cell dysfunction, et cetera. In terms of dropping in, so I think that that cascade of what happens with the invaded setting is likely in large part from nerve injury and myelin degradation and all that. So some ideas are, well, can you block the nerve degeneration, right? So we're definitely interested in that. Have some ongoing work there. I don't know what would happen if you dropped a TLS right next to a... My hypothesis would be that it would not remain intact and functional, right? I don't know how quickly it would degenerate, but that would be my hypothesis. Of course. Thank you. So, the question we're dealing with almost every day, how do you improve the statistical sound of spatial technology since one slide represents only a 5 to 10-micron section that might not represent the entire heterogeneity of the tumor microenvironment? So heterogeneity is a question we get all the time. It's the reality of the tumors we work on. I'll give you two sides of the argument, so I'll probably just argue against myself and shoot myself in the foot right now. But the reality is, on one side, we constantly go in, we take a random biopsy from a patient, just of one region of a tumor often. We send that off clinically, and that's how we make clinical decisions right now off of some random piece of tissue. So, there's that side of the coin. On the flip side, we often include multiple cores from various regions within our data sets to try and account for heterogeneity. The fact is that these technologies are so expensive, and Will alluded to the fact that we really need to decrease the cost of these so we can start doing them on whole sections to be able to encapsulate that. My group's actually shifted away from doing a lot of these for predictive and clinical values and moved to whole sections on H&E's just for that region or for the reason of the heterogeneity. But I will say that we published this a couple of years ago, that we developed deep learning models using these spatial multiomics to predict clinical outcomes within patients. And when we take vastly different regions of a tumor that are actually quite unique with the heterogeneity and the composition, these models are capable of predicting clinical outcome with these different samples from different regions of the tumors with high accuracy in the mid-90s. So there is differences, but there's something unique about tumors overall and their architecture that can be picked out with some deep learning models. So those are just my unfiltered thoughts. It's not going away, but this is what we're limited to given the cost of these. So we got this question from multiple people in different kinds of variation. How do you guys tackle the heterogeneity issues in your lab from sample to sample, from slide to slide? Yeah. I echo everything that Dr. Walsh said. I think this is the ever challenge for everybody. I think there's other techniques you can do. So you can try to certainly minimize heterogeneity that's ateriogenically introduced by reducing batch effects. A lot of times, if we're comparing things directly, we'll try to get them onto the same experiment, onto the same slide. So at least you're not introducing other forms of, I guess, you could consider heterogeneity. I think powering these studies, knowing what your question is, and what power really means in that context. I think with all the omics, spatial included, thinking about multiple testing correction is just, I think, honestly not even where the field is at, but there's a lot of potential for spurious results, and heterogeneity can contribute to that, especially if you don't think about the question and how best to power the question and how you're selecting specimens and the number of specimens and things like that. And then I think the other thing is trying to, before you really believe in a result enough to then do the functional validation and then eventually, hopefully, get to clinical translation, trying to, with these more descriptive approaches, look at it from as many different angles as possible. Yeah. And if things converge on the same thing, regardless of which technique you're using, which specimens you're using, then you're very likely-- Maybe you miss some more subtle biology- Yeah ... but you're more likely to hit on real biology. I think functional validation is key here. Yeah. Yeah. Yeah. Thank you. Sorry. No worries. Go ahead. Sorry. I asked that question about improve the statistical sound. I got lots of inspirations from all the talks, and I'm wondering, is it possible to improve the statistical sound by comparing certain workflow, like start from the bulk transcriptomics, like from the last talk, and at least we can focus on the certain target. We already know what target we want to be targeting from the bulk sequencing, and then from the spatial, we can maybe focus on the imaging-based and just that few proteins. And then, if we cannot get what we want, we can slice deeper and get more slides until we really visualize every protein and the cell-cell interactions. So yeah. Is there any certain workflow can help us to improve the understanding of the heterogeneity? I would say that ironically, we're talking about all the nuances of these technologies, but I review a ton of spatial papers, and the main reason why I think some of these papers don't make it is not in the rigor of the analysis or not because they're accounting for heterogeneity. I think it's actually the clinical cohorts that they choose. So sort of echoing what Will just said, it's super important that you're powered, and you're powered with apples-to-apples comparators. If you're looking at response to therapy, or you're looking at survival, it's probably the most common thing people use. They often have a mixed bag of cohort of what they can get from the clinic, and they're taking these samples, and they're doing all these amazing, cool profiling things that are very statistically strong. But you're enriching with some outcome with a late-stage versus an earlier-stage patient or some other super compounding variable. So in my opinion, what I see is to really focus on the clinical question and make sure that your cohorts, if you're looking for to compare some sort of outcome, response to therapy, et cetera, are significantly powered in an apples-to-apples way so that you're looking at these things directly and not comparing something that really shouldn't necessarily be comparable. Thanks. That's cool. Thank you all for three incredible talks. I want to ask a question about some of the biology. The techniques are awesome, but we know that in lung cancer, palliative care improves results. We know that this neuroinflammatory signaling is bi-directional. So the inflammation makes certain nerve signaling worse. Nerve signaling makes this inflammation. It can change the inflammatory state. So I guess my question for you is, is palliative care, where we're using a lot of things that modulate nerve signaling, is that also oncology care? But just speculating, if you don't mind. There's a lot of activity in this field looking at things like sodium channel blockers, like lidocaine, whether or not they're affecting the tumor progression. What's the role of opioids, right? What's the role of beta blockers? All of those things are matters of ongoing investigations in different groups. We know that anxiolytics like benzodiazepines, now the group from Guido Kramer showed that they are associated with worse patient survival. So there are a lot of confounders to those things, and it's very muddy. But this is an excellent question and an active matter of investigation by many groups around the world. Thank you. It's great talks. I was wanting to zoom in on the perineural invasion talk about most of the data that you showed was that in primary versus metastatic tumors, and thinking about site-specific metastases and how perineural invasion could be different and be contributing to the unique metastatic niche versus primary disease. You can comment on that. Yeah, thank you. So the data I showed was in primary disease. I think that's a really important question. We are starting to look at that. And I know there's other groups interested. I'd say it is a little bit tricky in pancreas cancer to certainly have matched samples because we often don't resect metastases, right? And usually, the patients, their functional status doesn't usually help them. But for certain diseases, like other GI cancers where we do that, like colorectal, there's maybe an opportunity to be able to study matched analyses in primary met of neural environments. We're very interested in also, as you said, organ tropism for metastases, but also looking at neuronal subtypes, right? We didn't really talk about it very much, but it's clear that it's not one size fits all for these different nerve bundles. They're usually mixed, right? If you look at these larger fibers, right? There's going to be mixed between the different neuronal subtypes, parasympathetic, sympathetic, sensory, and then even finer categories beyond that. So I think all those things will likely have nuanced roles. We don't generally see as much perineural invasion in, say, liver mets as we do in the pancreatic primary in the cases where we have seen it. But the innervation patterns are very different, right? So how much of, if you hold the cancer cell state as constant as possible, you just change it to a different microenvironment, or vice versa, right? You take two different cancer cells in terms of very different phenotypes and then put them in the same neural microenvironment and see. Those are the types of experiments I think we need to do to really get at your question. Okay. Thank you. Thank you for a very impressive talk. I have a question in terms of you look at most gene expressions abroad could just follow up with just said that's the perineal invasion there, but mostly just total. Do you have sample you can dissection out the efferent sensory nerve, sympathetic nerve, parasympathetic nerve? What is the nerve type that's most likely to be invaded via the tumor cells? This first question, maybe ask first question. So most likely nerve subtypes to be invaded? Yeah. So most of the evidence is pointing towards sensory and sympathetic subtypes. So because you only have small amount of transcription because there's no nuclear, so you're saying nerve fibers, from limited amount of message RNA, you can get this data from the- So that's mostly at the protein level. Oh, it's look at the protein. Protein level. Okay, got it. There is also support by Nishida Silver Group in University of Michigan. She published this in Clinical Cancer Research and follow-up paper that looked exactly on that. And she showed that sensory and sympathetic, like we just heard, she used a bunch of markers for sympathetic. But, yeah, those are the most common nerves. Yeah. My second question that because nerve injury is first initiation of the nerves in regeneration and infiltration, and these are major because of Schwann cells or it's because of direct damage to nerve fiber bundles? Both. Both? Yeah. So if we develop therapeutic agents, do you think a more effective target to Schwann cells or target to both Schwann cells and nerve fibers? Again, probably both, but we currently don't have those means to mitigate directly the perineural invasion. Yeah. Okay. Thank you. Sure. Mm-hmm. Have you looked at nerve growth factor in terms of is it a positive thing for the immune circus or negative? I mean- Yeah ... it's developed neuropathy. There's a lot of work in the perineural invasion around different types of nerve growth factor. It's a family of actually four ligands, NGGF, BDNF, NTF3, and GDNF. And they all have been linked to more rapid tumor progression and perineural invasion, not necessarily neuropathy, that's a whole different question. But this family of ligands have been associated with tumor progression, yes. Thank you. Sure. Thank you everyone for staying with us. Have a great meeting.