Exploiting digital histology approaches to probe the pathophysiology of SARS-CoV-2 infection

Thank you very much. My name's Kelly Hunter, and I work at the Birmingham Molecular Histology Facility here at the University of Birmingham, a core facility where we provide immunochemistry and tissue multiplexing services as well as digital slide scanning. I'm going to introduce our facility and some of the technologies that we offer access to before my colleagues describe some work that they've been doing, including work that's come to our facility. My left arm has just this minute started aching as I have my first Pfizer back in this morning, which I think is going to be quite appropriate.

I wanted to start with this hand scroll diagram. Don't worry about the contents of it, but I like it because it's the first concept of the facility that we drove a few years ago, sat together in the pub, which is such an alien concept now that I thought people would all enjoy sharing that. But effectively, the idea of this facility was to provide a set of complementary platforms that vary in either in throughput and resolution and plexity. That can form a pipeline for discovery and validation of discoveries depending on what the input question is. You can come into this at different points. This was the workflow we did at the time, but I can't really imagine sitting in a pub so casually at the moment. It's very strange.

Obviously our workflows will start with staining. And primarily, our staining platforms comprise of Leica BOND RX auto stainers. We started our BOND journey several years ago, and we have been collecting them ever since. We've got four now, and they are real workhorses. They provide single plex, simplex chromogenic stains, right the way through multiplex chromogenic stains. Not just antibody stains, but ISH multiplexes as well, and IHC-ISH combination assays. And they also provide the stain for our digital spatial profiling assay that I'll detail later. There's also, we're also currently beach testing the Lunaphore Comet, you can see on the bottom right there, which has a staining module built in, which I'll discuss briefly later, and there are other similar products on the market as well. Once we've stained the slides, then generally our slides get digitized. The multiplex slides particularly need to be digitized to be interpreted.

But we have a Leica Brightfield scanner, which we use primarily for digitizing Brightfield slides for collaboration purposes, for sharing with collaborators at different sites across the country and around the world. But it's also used to digitize chromogenic stains that our researchers want to perform image analysis on. Then have the Akoya Polaris seven-color whole slide threshold scanner for our multiplex assays. And the MANTRA 2 parallel assay development station, which is a complementary microscope. We have, as I mentioned, we're testing Lunaphore Comet, which can perform up to 20 staining cycles automated on four different slides, and I'll go into that slightly, I'll go into that later. And the NanoString GeoMx platform, which I'll also talk about more, which has a, it's a hybrid slide scanner and general instrument. All of that, as you can imagine, generates quite a large data challenge. So, we have several different solutions for managing that data. We have a eSlide Manager, which hosts all of this, not just all of the images that come from our Aperio AT2 scanner, but also any other images generated from other facilities or from our collaborators who will send us the images to import into the platform to share alongside the images we're generating locally. So, this gets used an awful lot. It's designed to feel like a standard pathology workflow with a slide tray on the left, and you select your slides to view them on the right, you can make some annotations to the images, as you'd expect. The image, the data behind us is stored in the Birmingham Environment for Academic Research, which is both a research data store and a high-performance computing facility and that powers a lot of what you'll see going on in the background. This is the platform for GeoMx, the NanoString GeoMx instrument, which is also accessed by the web browser, and it stores, it not only stores the freshman scans, but also the data analysis platform, as again, I'll detail shortly.

Then we'll move on to data analysis, which is a key part of most of the technologies that come out of our lab. So here, we're looking at a data analysis workflow from the vector scanner, a seven-color freshman image on the left, which has been spectrally convoluted. And I'll briefly touch on that workflow in a moment. And that deconvoluted image has then had the-- we've performed self-segmentation. On the second image on the left, the green objects are the detected nuclei from the blue DAPB stain on the left image. And the red line around that is an estimated cell border. The next image along is a tissue segmentation image, which has separated the image into different tissue compartments. And then the final image on the right is cell phenotyping, where each of the objects from the second image, each of the cells have been assigned a phenotype represented by different colors there. We also have the VisioPharm software, which can perform the same multiplex phenotyping workflows, as you can see in the top left image there. Except we also have GPU accelerated deep learning capabilities on this, so that really accelerates the tissue segmentation and the cell segmentation required at the beginning of that workflow. But it also provides other outputs such as the heat map you can see in the top right and the t-SNE plot down in the bottom right. And here is the interface for that software. This is the analysis software from the NanoString GeoMx. On the left, you can see the freshman slide scans. Down the center, you can see the various probes that are present in the assay that can be filtered. And then on the right, you can see the outputs from analyses on our combination of those two panes on the left. And I'll go into that workflow shortly.

I'm going to start by talking about the Vector Polaris workflow. I'm not going to go into too many details. I know that Mike Serace has done this exhaustingly in the past for us. But just briefly, the idea behind this technique is that we stain all our markers on a single slide first, label them all with different fluorophores, and then we perform a single scanning step. What you can see here is a representation of all the fluorophores on a single slide and the various overlaps. So, although you can see that the peak expression, the peak expressions are all at discrete wavelengths, that some part of the expression for each of those colors overlaps with other colors. So, then this, the software associated with this platform, the Inform software, is used to at any given pixel to deconvolute how much the intensity captured at a pixel at a given wavelength belongs to which of the markers that are co-expressing in that pixel.

Another representation of that here, and if we choose panCK as an example, the orange bar above the circle I've just drawn shows the range of expression that has been captured at that wavelength and then you can see that represented as signal belonging to three different markers. So, you've got some yellow signal coming from CD68, some orange belonging to pan-cytokeratin, which our peak expression has come from, and then also some signal blind to PD1. And so, the software here will estimate how much of the signal belongs to each of those three markers and bin them together instead of in a per scan channel fashion, but they'll be binned per marker instead. So, then you'll end up with a computed channel that is enriched for your marker. It's also used to remove autofluorescence from the background signal.

We end up with here in the top left sample in image A is a spectrally deconvoluted fluorescence image. In the bottom left, we can see here a simulated brightfield stain, which is sometimes easier to interpret. In image C, you can see the pan cytokeratin has, by itself, you can view it as an individual stain because we've deconvoluted the pan cytokeratin and PD1 stains that were overlapping with it, and you get a clean stain. Then you pass it through the image analysis workflow that I just described, performing tissue and cell segmentation, phenotype your cells, and then perform high dimensional analysis and spatial analyses on the outputs from that.

The Lunaphore is an example of a stain image workflow where a lower number of markers are stained, are applied to the tissue. They're labeled, in this case, with Alexis or fluorescent dyes. You image those markers and then you elute the whole antibody complex and then you start again and you keep repeating the cycles and layer the image up in staining cycles instead of staining them all at once, which means that you don't need to perform a spectral deconvolution because you can use chromogens that are sufficiently far apart on the spectrum that there's no significant overlap. And then you end up with higher dimensional image like this. I think this is a 27 plex image, which you can't view, which is where image analysis becomes essential because we can't, we can't qualitatively interpret this, this image. They'll go through some of the workplaces we've seen before. As well as detecting protein, we do RNA species detection.

These two images here are examples of the RNAscope assay from ACD Bio. On the left, a multiplex chromogen, and on the right, a multiplex fluorescence assay. And we also combine these assays with protein detection on the same slide. And then on the top right here, we have a diagrammatic example of the NanoString distribution profiling assay, where we combine a 3-plex fluorescence stain with targeted ROI-based profiling of a larger pro panel. So the workflow here on the NanoString is we start off by staining our sample with a panel of either antibodies or RNA species and at the moment you can stain panels of around about 100 antibodies or 18,000 up to 18,000 RNA species that are conjugated with barcodes attached by photocleavable links. We stain that at the same time. We stain a four-color fluorescent stain, three probes and a nuclear counter stain. We put the slide in the instrument, and the slide scans the fluorescent morphology stain. And you use this morphology image to select regions of interest that you are going to target to profile the barcoded panel that you've added. The instrument will then iterate around the ROIs you selected, irradiating with UV, cleaving those UV linkers, and releasing the barcodes into the suspension, which are then aspirated off the slide and into a collection plate. We can then count those barcodes and relate the data back to which ROI they came from.

We have a hybrid four-color fluorescence IHC image with a high-dimensional spatial GeoMx output. Here's an example of the ROI selection process. And as you can see, so each of these squares is an ROI, and you're able to cross-section those ROIs, micro-dissect them with the built-in image analysis tools on the platform, which allow you to segment on this reference thing. In this case, the blue marker is a pan cytokeratin, which is highlighting the tumor in the sample. And we've used the onboard image analysis tools to segment each ROI into two, a tumor positive and a tumor negative ROI. And they'll be irradiated and collected separately to have discrete tumor and tumor microenvironment measurements. If the NanoString GeoMx I've just described is the histology lab's approach to spatial genomics, then 10X's Visium assay is the genomics approach to spatial genomics. This involves taking a section and applying it to a slide that has a grid of RNA primers on the slide. And each spot of primers on the plate has a barcode associated with it, which tells us which spot it blank belong to. We also stain it in the same way as we have before with a fluorescence morphology stain and image that. You digest the tissue onto the slide and sequence it. Then what we have is, in this case, sequencing data where each transcript can be related back to where it came from on the slide. And the output data looks like this. You get heat maps for expression in different areas on the slide.

Just a quick summary of the platforms we have. We start off with IHC, which is high throughput. It's good resolution. We've got single cell resolution. It's only single stains and is more of our relation tool than discovery tool. Then we move on to Multiplex, which increases the plexity slightly to maybe three or four colors with the same resolution and comparable throughput. You move on to multiplex IF, you increase the plexity slightly at expensive throughput, but you still maintain that resolution and you really start increasing the discovery power. You move on to multiplex IHC, you lose some throughput in exchange for moving on to 40 or more targets, but that throughput really makes it a good, more relevant discovery tool. And then spatial transcriptomics, which is theoretically infinite flex. We're running a whole transcript home now on several different technologies. But the resolution suffers in different ways between the two technologies I've described, so you don't get quite single cell data from either of them. And the new kid on the block is of course, MERFISH, which answers that problem by giving you back not just cellular, but subcellular resolution, the change of the cost perplexity so far. But I'm sure that will increase over time.

Here's some of the companies that we work with and the sponsors that have helped us get this facility off the ground. The Molecular Study Facility takes both internal academic and external academic projects. But we're also establishing a new entity named Birmingham Tissue Analytics, which is going to be an industry-facing service contract GCP compliant face to the service. And this is going to be part of Birmingham Precision Medicine Centre, which will include other disciplines, such as giant mix and immunology in a multidisciplinary industry contract kind of trial-make facility. And now I'm going to pass on to my colleague, Dr. Graham Taylor, who's going to tell you what they've been doing with these technologies and many others.

OK. Thank you, Kelly. That's a great introduction to the facility. And I'm going to talk now about our project studying SARS-CoV-2 infection in tissue that's used many of the approaches that Kelly has outlined in his talk. So, we studied two different COVID-19-related diseases, the conventional COVID-19 infection in the lungs that can unfortunately cause people to die. And also, COVID-19 placentitis. That's COVID-19 infecting pregnant women, causing them to have problems with their pregnancies, including stillbirths.

We know a lot from analysis of blood samples from COVID-19 patients about what happens in the periphery. Patients often have lymphopaenia, neutrophilia, you see increases in the range of various inflammatory cytokines. You see evidence of T-cell activation. And potentially those T-cells could drive the hyperinflammatory state in the lungs that cause so many problems. And of course, there's evidence of T-cell exhaustion. question. But we know much less about what happens in the tissues of those patients. With that in mind, we studied tissue provided to us from post-mortem investigations performed in South Wales by Dr. Esther Youd and Dr. Gareth Leopold. Patients were included in the study if they've had positive SARS-CoV-2 PCR swab either in life when they were in hospital or a swab taken at post-mortem. And depending on the specific consent that was taken from the relatives of each of those individuals, we took samples from different organs, at least heart and lung in every case.

This has been very challenging, both in sort of technically challenging, but also in terms of the reagents that we have available to us. Of course, this is a new virus, so there's little precedence in the literature for what we're going to find. and the reagents have been untested and poorly validated. We also have additional challenges. The tissue is taken from post-mortem, so it's not only fixed, but there's always a time interval between when the patient dies and the sample is taken, that can result in autolysis and degradation. Therefore, being robust in the validation of the reagents we've used in this study, and hopefully that will cover press in the data I'll show you shortly. Let me start with the study cohort. Focused on eight patients, three at the top in blue, are patients that died in the community from COVID-19 after a very short preceding symptomology. We also have five patients shown in red. These patients were admitted to hospital for quite long periods of time and unfortunately died from their disease. And so, in this talk, we'll be focusing on these early versus late cases.

In terms of their fundamental pathological analysis, both those patients had evidence of diffused alveolar damage in their lungs. As we go from the early patients to the late patients, we see increasing evidence of the proliferative phase, alveolar damage, fibrotic phase in five out of eight patients. All evidence of organizing pneumonia in one out of eight patients. And looking more closely, we see thrombi in four out of eight patients, evidence of secondary infection in five out of eight, here showing neutrophils in the lungs, and vasculitis in three out of the eight patients.

Kelly's already made a great introduction to various platforms we have available to us. Let me tell you a bit about the workflow. So, here's our eight COVID-19 patients in blue on the left-hand side here. We have various controls. We have a rhinovirus pneumonia case, three cases of bacterial pneumonia, 10 cases of non-infectious disease overstained lung tissues, and interestingly, a case of MERS coronavirus from the early 2000s. This is another coronavirus that's somewhat like SARS-CoV-2. As mentioned earlier, we collected lung and heart tissues from all eight patients and a range of other tissues, such as upper airway, brain, liver, kidney, and lymph nodes from the smaller subset of the patients, depending on the consent we were given. On those samples, we've prepared tissue sections, tissue microarrays, and we've taken scrolls of tissue for processing into RNA for genomic analysis.

Here's an overview of the assays we've used. We've used spatial methods here on the left, either single Plex immunohistochemistry or higher Plex, Codex, and Comet multiplex immunohistochemistry. On the right, we've got the molecular methods, which are Quantseq, which is the RNA transcriptome analysis. We've also successfully performed T-cell receptor sequencing on a range of samples. In the middle there are some interesting hybrid methods. It includes RNAscope, which dissects RNA in different parts of the tissue, and the NanoString that Kelly's already outlined. These are interesting to say, combine the power of transcriptome analysis with the power of looking at where these RNA molecules are spatially localized in the tissue. To just add that for all these spatial methods and hybrid methods, we stain tissue successfully, either as that end result for the method or as preparation for that method, into the Leica BOND system, Leica BOND RX.

Here are the outcomes from these different assays. The spatial methods give us high dimensional phenotyping and localization of specific markers we've pre-selected within the tissue. Of course, we can detect viruses in those different regions of tissue. Molecular methods are great discovery tools. They allow us to analyze what genes are expressed differentially in the different types of tissue, we also look at pathways using gene set enrichment analysis. The key point here is that we're cross-validating across all these different methods. The spatial methods, the hybrid methods, and molecular methods all support one another. I'll give an overview of immunohistochemistry. If you use a range of techniques depending on the sample size, on the left here, we've got chromogenic immunohistochemistry using the Leica BOND RX system, we've applied that to large areas of tissue. On the right here, we've used the highly multiplex Akoya CODEX system. But on smaller tissue microarrays, this is sort of balancing cost and the time taken to acquire different sized images. Each of these techniques has their pros and cons, and we've combined them all together in this study.

Here's a range of image analysis techniques to study the data. This is conventional qualitative assessment by pathologists, looking at the tissue onto the microscope. And here you can see a high plex image produced on the Akoya system, the vector system. And on the right here, we're looking at machine learning based approaches, which look at the data and identify patterns using complex learning algorithms such as kidney.

I'll be showing you the genomics workflow. I won't go into too much detail. Kelly's already described it. But essentially, we stain the tissue for PAN-CK, CD45, and for DNA. And then we illuminate different regions of that tissue that we select based on that staining. That cleans off the UV cleavable linkers that are analyzed on the sequencing in Lumina platform. Let me show you some results of this work. First, the virus distribution in the different tissues. So, for a lot of work, we've identified several reagents that work well, both in terms of control tissues, but also have good performance on the FFPE tissues. These have all been stained using the BOND RX system from Leica. We've used here Vero cells, either uninfected or infected with SARS-CoV-2 virus. And here I'm showing you the results from a spike antibody, so either uninfected or infected in brown on the right-hand side, you can see very good staining, going for the nucleocapsid antibody as well. At the bottom using RNAscopes, this is in situ hybridization method that detects particular viral transcripts, either the spike or the OF1AB gene. You can see the red and the green staining at the bottom right of the image here, showing you positive staining of those virus-infected cells.

We've used a range of approaches to identify whether a virus is positive in different tissue sections. These include immunohistochemistry, in-situ hybridization using RNAscope, and DNA expression profiling by Quantseq. We've been very robust in how we score tissue, whether it's positive or negative for the presence of viruses. For example, if we see only one immunohistochemistry marker present at the bottom of this diagram here, we would score that as a negative. We see two immunohistochemistry markers as positive, but no evidence of viruses by RNAscope. We look very closely at the positive control RNAscope probe result. That is negative. That suggests that the RNA may be poor quality. And therefore, we score that as possible. That positive control RNAscope is positive. It shows that the RNA was good quality. And therefore, we think that the virus is unlikely to be present because it's present only or detected only by the protein level and not in the RNA level. We see the tissue positive by two immunohistochemistry markers and by RNAscope in situ hybridization. We score that as likely. And we see it's positive for immunohistochemistry markers, RNAscope, and we see evidence of virus by either quantseq called genomics gene expression profiling and we score that as confirmed. The reason we've been very robust in the way we score virus positivity is with data like this.

These are the antibodies I showed you earlier, working well on the cell line. On the left here, you can see a bit of upper aroid tissue. And you can see good staining for spike, but not nucleocapsid protein. The spike is shown in that purple white color. Now, some groups will score that as positive, we score that as negative. Why? Because if you take the same tissue, which has got good quality RNA, it works on an RNAscope, you can't see any transcripts from the virus in that same part of tissue. So, whether that spike protein works well on the cell lines, it works well in the lung, there are certain types of tissue where it throws up background. One has to be very cautious in interpreting the results of virus detection in the FFPE tissue, particularly from post-mortem cases.

Here's another example. In the middle, on the left-hand side panel, you can see a green cell that's staining positive for viral nucleocapsid. On the right, that same cell is also staining positive for spike, but we've also got other cells that are staining for spike as well. This tells us that spike and nucleocapsid should be coincident in their expression. So again, that spike alone staining doesn't look 100%. We've gone on to score the different tissues, the presence of virus counted virus infected cells, and by that, I mean virus infected cells express both antigens, and where we get good evidence of virus positivity by RNAscope and by sequencing. In the right-hand side, you can see we can only pick up virus positive cells in early-stage SARS COVID infection. We've never picked it up in the late stage in the controls, we never picked it up in the MERS. or the other control tissue.

Another interesting result is shown. Here we're doing RNAscope, and we're looking for pan-cytokeratin by immunohistochemistry. You can see a cell here that stains positive for spike, messenger RNA, but it's negative for pan-CK. There's another cell at the bottom there with the yellow arrow. That again stains positive for spike messenger RNA and positive for Python case. So, we're getting virus detection in different types of cells. And we're still looking into exactly what type of cell can be infected by SARS-CoV-2 in the lung. Our results are backed up by the GeoMx gene expression profiling data. So, each of these columns shows a different type of tissue. You only must pick up SARS-CoV-2 transcripts in the GeoMx data in the SARS-CoV-2 tissue, as you'd expect. And moreover, we only ever pick it up in the early infectious cases and not the late cases. So that light, brown-colored box and whisker plot shows the results for patient SARS-CoV-2A on early pace. Go to places to the right of that are late cases, which are negative for transcripts. Now the GeoMx gives you spatial profiling of RNA, you can see that we pick up transcripts in the alveolus at much higher levels than what we see in the blood vessel. Again, this highlights the fact that we only see viruses early not late. So here we're looking at the Quantseq gene expression data, this is bulk sequencing data. You can see if we pick up multiple virus transcripts, particularly nucleotaxid spike and ORF1AB, in early-stage cases, but not in late stage.

We're now going to move on to talk about the inflammatory response that's present in the lung. Here we show you some data acquired on the Vectra Polaris platform. The slides were stained on the BOND RX system from Leica. You can see very nice aggregates of CD68-positive macrophages in white in the image on the left-hand side. Performs a cyber sort analysis on the bulk RNA sequencing that we formed using Quantseq. You can see that there's a range of different types of immune cells present in the SARS-CoV-2 tissues that are on the right-hand side of this diagram here.

Using computer-aided guided analysis of the immunohistochemistry results, we've identified different regions of the tissue. that we've identified as either lymphocyte-enriched or macrophage-enriched or neutrophage-enriched. And the frequency of these different enriched areas varies depending on the type of disease, whether that's SARS-CoV-2, binder virus or control tissue. You notice these interesting differences in the immune cell neighborhood between early and late. In the bottom right here, you're looking in you're looking at the frequency of CD8 T cell abundance in neighborhood 2. It's much lower in the early-stage cases in red compared to the late-stage cases in blue. You see differences between disease and you see differences over time. Looking at the transcription data, we see evidence of collagen enhancement in SARS-CoV-2. Interesting also in MERS, but not in rhinovirus or bacterial pneumonia. We confirmed that increase in collagen by immunohistochemistry.

Here we're looking at collagen 6 staining. On the left in normal control, you can see a thin ribbon of collagen 6 around the periphery of those cells. On the right, we're looking at a SARS-CoV-2 case. You can see the thickened collagen's positive staining in that lung tissue. In summary, we can say that we see a biphasic pattern of infection and inflammation in SARS-CoV-2 infected patients. End phase is characterized by high virus loads and an increase in infiltrating innate immune cells. Late phase by low virus loads and an increase in infiltrating adaptive immune cells with good evidence of collagen deposition.

We're going to move on now to show you some very emerging recent data from our group on COVID-19 placentitis. So, a small proportion of mothers that are infected with SARS-CoV-2 unfortunately experience complications in pregnancy. That can include intrauterine growth resuscitation, hemorrhage, per into your birth, and, unfortunately, intrauterine death of the infant. There are various abnormalities seen in the placenta when there's occurrence. Here's looking at a piece of tissue from a COVID-19 positive placenta. And on the right here, you can see CD868 positive cells at the fetal internal interface in the placenta. Now looking at another case, it's mainly here by single-plex immunohistochemistry for spite on the left and nucleocapsid on the right. You see good evidence of staining. Here we're using the Lunaphore Comet platform to perform simultaneous spike and nucleocapsid staining on the single piece of tissue, the nucleocapsid in red, spike in yellow. You see that staining is coincident. We're seeing double expression of spike and nucleocapsid protein in the villous trophoblast. Also showing this virus positivity by in situ hybridization using RNAscope. You can see this first spike on the top right and ORF1AB on the bottom right of the slide here. Finally, you can form Quantseq on the placenta samples, and you can see the evidence of virus transcriptomes and a range of different cases of placentitis.

In conclusion, we've shown that you can apply a range of modern immunohistochemistry and in-situ hybridization and sequencing platforms on FFPE tissues. This is particularly impressive given the time between death and taking those samples, and in some cases, several days. It has also shown how important it is to use multiple platforms to cross-validate each other for a key result. This is a fantastic collaboration between myself, my colleague, Dr. Matt Pugh, who's studying there for a PhD in my group, Kelly as well, of course, as part of the IHC facility, with colleagues at the University of Limerick in Ireland, Professor Paul Murray, and Dr. Eanna Fennel. I give special thanks to Dr. Esther Youd and Dr. Gareth Leopold for performing the post-mortem examinations, of course, thank the families of the deceased for making available the tissue for research purposes. This work has been supported by an SFI COVID-19 rapid response grant and thank you for your attention and we'll move to take any questions that you have. Thank you very much.