Webinar

Hello, and a very warm welcome to everyone joining us for today's select science webinar titled Spatial Analysis and a high-plex immunofluorescence to study human pancreata in type 1 diabetes. My name is Georgina Wynne Hughes and I'll be moderating today's presentation. I'm delighted to be joined by a guest speaker, Dr. Estefania Quesada-Masachs. In this webinar, Dr. Quesada-Masachs will explain the image analysis pipeline that she has developed and applied to study whole tissue pancreatic samples of donors with type 1 diabetes.

Following the presentation, we will have time for a short question and answer session. If you do have any questions throughout the webinar, please feel free to submit these at any time in the box to the right of your screen. Without further delay, I would like to hand over to our speaker, and I would like to thank them again for presenting for us today.

 

Welcome, everybody. My name is Estefania Quesada-Masachs. I am an instructor at Matthias von Herrath's lab at La Jolla Institute for Immunology. And today, I'm going to present our study where we perform a special analysis in a high-plex panel, studying human pancreata, so patients with type 1 diabetes. For today, first, I will provide some perspective about the disease and about why high-plex studies are very interesting in type 1 diabetes. Then, I will show you some of the first initial images that we did in the context of the quality control, because it's interesting to see how the panel look in other tissues that are very rich in immune cells. And then, I'm going to present some preliminary results of the tests that we did in human pancreatic tissue. But more than results, this presentation is focused in our image analysis strategy, in what is the work pipeline that we use in order to get those results.

So, type 1 diabetes is an organ specific autoimmune disease where the immune cells attack selectively, insulin producing pancreatic beta cells located in the pancreas. And I say selectively because other organs or even other parts of the pancreas, most of the times, are totally preserved. For that reason, in type 1 diabetes, it's very important for us to study the organ, the pancreatic tissue. The problem with pancreatic tissue is that its accessibility is very limited, and the availability is limited.

And with accessibility being limited, what I mean is like it's difficult to take a biopsy of a pancreatic tissue from a living individual. And we rely in the organ repositories like import or export that provide actually really high-quality pancreatic tissue from deceased organ donors. But even though these repositories have really helped us a lot with the access to those samples, we still have, in the US, just a few hundred pancreatic organs preserved from non-diabetic and type 1 diabetic individuals.

For that reason, systems like the Orion from RareCyte, where we can study multiple proteins of interest, not just three or four, but up to 18 in this case, are really interesting for us because we can get more information from the same tissues. And those tissues are precious, as you can imagine with the information I provided. So, why is it important for us to study that many proteins? So, with the previous systems that we had at LJI, because, before we purchased the Orion system, we have the Axioscan, it’s a slide scanner, but we can study three or four markers plus nuclei at a time. This means that, in order to study 16 or 18 markers simultaneously, we have two options. We can use consecutive slides, or we can go through multiple cycles of staining, stripping and restraining. Using consecutive slides for 16 to 18 markers, we will need four to seven slides.

Every slide will be six microns thick. And this presents several issues. For instance, once we want to correlate the proteins that we studied in the first slide with the proteins that we studied in the last slide, we already have 24 to 42 microns distance, which means that we may have trouble identifying the same islets. And for sure, we won't be able to identify the same cell. Also, we are going to use a large amount of tissue in order to study this amount of proteins with the previous system.

Alternatively, we can use just one slide of tissue and go through four to seven cycles of staining, stripping and restraining. But the issue with that is that we know that so many cycles will cause some tissue loss, some deterioration. Some of the proteins won't be that well expressed in the last cycles. We will need a lot of optimization. And the more cycles we need, the more artifacts we will be creating.
So, I'm going to show first some of the images in other tissues, not just pancreas, also in spleen and tonsils. In the panel that we tested, in collaboration with our colleagues from RareCyte, using Orion, where in our case we use 16 markers mostly to identify immune cells, even though we also were identifying endothelial cells and other phenotypic markers and autofluorescence. And here, in the first picture, you can see the whole spleen tissue that was in the slide. And in the little squares, you can see details of the expression of different immune markers. And you can see the differentiation clearly between the red pulp and the white pulp. So, we can see even an anatomical, we can differentiate anatomical characteristics easily in a tissue that is very rich in immune cell, similar to what we observe in tonsils.

In the tonsils, we can clearly see the germinal centers. And this is the whole tonsil section in the slide because, with this Orion, the whole point is to get the information of the whole slide. And we can see details of the expression of differential immune cells and some anatomical details in the distribution of these cells.

We also stained pancreas. This pancreas is one of the optimization cases that input provides. And those cases for optimization are non-diabetic. That's the reason why the image looks mostly gray, showing the nuclei. And the reason is because we don't have that many immune cells in a healthy pancreas, at least in theory. Still, we were able to see some expression of CD31 with endothelial cells, some macrophagic markers. And we were able to identify some of the dots. So, we could still see that the panel was working fine in this healthy pancreas. So, when we finished the quality control, we decided to move forward and to stain, and to do a small test staining some of our cases. So actually, we selected, for the pancreatic sections, these were paraffin embedded pancreatic tissue sections of four pancreatic donors, one non-diabetic and three were tissue sections coming from patients with type 1 diabetes.

And we use Orion and we stain for 60 markers. For immunofluorescence, we use, these were cases that we actually knew very well because we previously stained them for insulin, HLA class II, and CD68 using our traditional previous system. And we knew where the islets were located. And we have actually run in those images a machine learning analysis to identify the islets and the tissue outline semi-automatically. And I will show you that, the details of that in a second. And you can ask, "Well, even just for three markers, why were you using a machine learning analysis in order to identify those islets?" And the reason is because the data that was generated, even from this first panel, were already more than a thousand islets and more than 5 million cells.

So, if you really want to get information from every single islet when you get the whole tissue scan, you really need to use a system that makes it semi-automatically. To put some perspective into that, when we do manual analysis, usually, we select something between 30 to 50 islets. If the islets are from patients with type one diabetes, because there is a large number of them that are lost due to the immune attack, we can maybe go and select all the islets. But from a non-diabetic individual, you can find hundreds of islets in a piece of tissue. And actually, it's not feasible to do it manually.

So, what we did actually to identify those islets was to train a pixel classifier using QuPath. And with the pixel classifier, we identified the islet regions. And then, we run a group script when we created like a 10 microns extension around the islet to identify the per islet region, which is like an exocrine region, but in very close proximity to the islet. And also, here in the picture, you can see that the color is more intense. The reason is because this is not the color of the staining, this is the color of the staining. These are the colors of the pixels. After we train the system to identify which cells were positive, which pixels, sorry, were positive for insulin, for HLA class II, or for CD68.
And also, in this study, we run the cell detection and the cell classification in order to identify which cells were positive for HLA class two, which cells were positive for CD68, which cells were negative or which cells were positive for insulin, or which cells were positive for two or three of those markers. And here, in the bigger picture, you can see all the tissue outline in purple, some squares that represent selected areas of exocrine that were very clean that we used to get additional information. And in green, you can see all the islets and peri-islets that were identified semi-automatically with the system with our training.

So, because we use consecutive slides, this is one image of an islet where we stain previously with insulin CD68 and HLA class II, using the Axioscan. And then, we have the same region in the consecutive slide where we stand for these 16 markers, this time using Orion to acquire those images, where you can see some of the immune markers for CD45, CD68, CD4, CD8. This is the original signal and, like that, that this has original signal. And you can see how the quality of the pictures is very good for both of the systems, actually. Before entering into what we did for the image analysis and those details, I would like to give a few, I would like to explain a few considerations first.

So, first of all, the images that we stain with the Axioscan, they were cut relatively recently, but the images that we stain and process using Orion were cut at least with one year and a half of these stains, which means that when one set of tissues is cut, just because of the fact of being cut, even if you preserve it in good conditions, there is certain level of tissue degradation or oxidation. Some of the proteins may lose a bit their expression, so it makes a tissue more difficult to work with. And pancreatic tissue is in general pretty tough to work with, especially in the type 1 diabetic individuals, because there is a certain level of degradation due to the immune attack, and there can be areas that are a bit more fibrotic, etc.

So, another thing that is important for us is to consider that we need images that have a high dynamic range. This means that, when we study markers that are biologically important, like for instance the HLA class II that we studied initially, we don't really want to answer, most of the times, whether a cell is positive or negative for this protein. We actually want to know how much of the protein these cells are expressing. And in order to be able to quantify, we really need a system that has a high dynamic range.

So, one of the advantages of the Orion is that the dynamic range is effectively higher than what we had with the previous system. And it's basically due to how they deal with autofluorescence. They have a system to quench autofluorescence. They use a quench buffer, UV, and white light. And also, the thing that is more relevant is that they scan dedicated channels to autofluorescence that provide tissue architecture, and then they allow this subtraction of the signal so we can identify our proper targets. And here you can have an example of autofluorescence in the channel 445-485. You can see how this auto fluorescence may be annoying the next by channel, the 470-513.
So, with the Orion system, there is a subtraction of the autofluorescent signal, which means that if we then want to identify in this channel, 470-513, targets that are less abundant, we should not have that many issues because autofluorescence won't be kind of bothering how we see this expression.

So, moving forward, this is again the same islet that I just showed you obtained with Orion. Here, I just wanted to show some of the details of the expression of some of the immuno markers like, for instance, CD4 that we can see in yellow, or CD8 that we can see in green. Here we can see CD45 in red. We can also see CD31 for the endothelial cells and Ki67, that is a nuclear marker, for that reason, from cells that are proliferating, it looks like a dot, and then, some of the markers for macrophages, like CD163 and CD68.

So, I'm going to move forward and explain our image analysis strategies step by step. But first of all, I would like to mention that we used, for all of these analyses, two open-source software, QuPath and CytoMap, that are available online. Everybody can download them. And they are actually for free. They have a lot of tools. And every system has their own place. And in order, you may be more interested in one or the other depending on what do you want to which type of information you want to get from those images. And you will see some examples of the type of information that you can get from them really soon. And actually, they are compatible, the systems with each other. Sometimes, you can move data that you get from QuPath and import it into CytoMap in order to do some of the analysis. Anyway, I just wanted to mention that, yes, of course, you need a powerful computer in order to make these systems work, especially if your images are very large. And if you are using images of the whole tissue, like the pancreas that is a large tissue, these usually occupies a lot of space and everything, but those open source software are for free. And if anybody is interested in taking a look and didn't really care before about them, please do it because they have so many tools. They are really, really cool software.

So, I'm going to continue explaining the steps that we use for our image analysis strategy that are four in total. And the first step was the annotation alignment. So, in our panel with Orion, we didn't really have insulin or glucagon or markers to identify the islets. So, what we did was to transfer the tissue annotations manually, using the QuPath software, from the images that we have obtained from the Axioscan.
What is what happened? The Axioscan and the Orion slide scan actually work differently. One of them acquires the images from top to bottom. And the other one acquires the images from bottom to top. And for that reason, the images are mirroring each other. So, in order to transfer the tissue, outline that we got from the Axioscan and the individual islets that you can see here as many dots in purple and know where those islets were located to interpret our Orion data, we needed to flip this annotation and make them be coincidental in the tissue. The tissue these were consecutive slides, so the tissue was six microns apart. But still, six microns is a little difference. And we actually got very nicely coincidental images, as you can see in the picture just as an example. So, once the images were, the annotations, the islet annotations, were transferred, we started with the analysis. This is a machine learning supervised analysis using QuPath, where we calculated the mean intensity and the percentage of positive area based on the fluorescence intensity, pixel by pixel.

And how we did that? So first, we train a pixel classifier. I'm not going to show the details of how this is done, but please feel free to ask me if somebody has a specific questions about that. So first, we train a pixel classifier for every single marker, for every single protein studied. And then, we run a group script that combine all of this information and gave us the information about, in every region of interest, like every islet and every peri-islet, which specific pixels were positive for one, two or all of the markers. So actually, we didn't just get the information about the pixels. We can actually get co-localization information about which of the pixels are coincidental.
So, here in the outside our regions of interest, you see the original signal. Now, you can see the original signal or some of the immune cells. I removed DAPI just to not have the image so crowded. And then, inside the islet and inside the peri-islet, the colors that you see are actually the pixel classification. So, this is how every single pixel was classified as being positive or negative for different markers.
And which information this provides? So, for example, if we study the different regions, we can get the percentage of positive area. In this case, for antigen presenting cell markers like CD11C, CD68, and CD163, we can get the percentage of positive area for every single islet. And as I mentioned at the beginning, this dataset had more than a thousand islets.

So, we know which percentage, for instance, of the islet, which percentage of the pixels were positive for the CD11C marker or the CD68 marker. So, you can see here that, in the first column, and this will be the same for every single graph, are in organs represented the islets of the non-diabetic individual. Sometimes, you barely see them just because everything is like zero or close to zero. And then, it's very close to the line. But all the numbers are there. Then, you can see the type 1, the islets of the patients with type 1 diabetes that were insulin containing islets, ICI, or insulin deficient islets. Insulin deficient islets are islets that lost their insulin after the immune attack happened. So, what we can observe is that most of the insulin containing islets were having more macrophages and dendritic cells, not pixels, inside within the islet.

And when we study what happened in the peri-islet regions, we observe again that, in most of the cases, type 1 D, the islets, insulin containing islets, or even insulin deficient islets of the type 1 diabetic individuals have more of these immuno markers around them.
We studied 13 markers. So actually, we just, we, I'm not going to show everything. I'm just going to show some of these proteins. For instance here, this represents again the percentage of positive area for CD45, identifying most of the leukocytes, and CD8, the T-cells, and CD4, the T-cells. These are not the cells. These are really the pixels. And again, the percentage of positive area for these different immuno markers was higher in the insulin containing islets of the patients with diabetes than within the islets of the non-diabetic. And in the per-islet regions, again, as we observed previously in the insulin containing islets, we observe more of these pixels that were positive for the different immuno markers.

After this analysis, our next step is to go and perform a cell detection and cell classification. This cell detection and cell classification is based not just on the fluorescence intensity, but also in other parameters such as the nucleus size, the shape. And in this case, we perform not just the regional analysis but also the whole tissue analysis. For the regional analysis, regions of interest are the islets, the peri-islets, and the selected exocrine regions. And whole tissue analysis, it would be the whole tissue regardless of which area we are studying.
So, for the cell detection, there are different ways of detecting the cells. We decided to use one tool that is inside QuPath where, based on the presence of the nuclei, there is this system does like an expansion around this nuclei to identify where the cell is located. The issue with that is that, for instance, if you have a very elongated cell, as it happens commonly in endothelial cells and many times for macrophages too, you may have trouble really getting a good shape of this cell because this tool doesn't really tell you where the membrane is located.

It makes this cell amplification, it tells you where the cell is placed and which size it has based on the nuclei staining and based on the proximity of the other cell. So, it expands up to a certain limit, but it stops expanding when it finds the next cell. There are different tools out there that you can use, like StarDist or Cellpose, to identify the cells. We actually tried them. And our results were better with, or look better with the tool in QuPath. And this is how we decided to move forward. But just to let you know, there are other software that are there that are also available, and where you can get results. So, everything depends on what are your needs for the study. So, for us, this was the best option, but actually there are other options depending on your needs.

So, once the cells are detected, what we did was to classify them, but before classifying them, of course, we needed to train a object classifier. So, we told the system, "Would you give the system many examples about which cells are positive for one of the markers and which cells are negative for one of the markers? And you do that for every single marker?"

And then, after that you can create what is called a composite classifier, where the system will put all your training together. And then, it will give you the answer about which cells are positive. For instance, here in blue for S100, or which cells are positive for CD45, or which cells are positive for actually not just one of the marker but for 1, 2, 3, 4 of the markers. So, you get information, you get much information that it can be really interesting.

So, here I'm showing again a regional analysis with the information that we can get from the islet regions. This time, the percent that we see is not the percentage of pixels, but the percentage of cells. So, from the total of cells that was present in the islet regions, how many of them, which is the percentage of them that were dendritic cells expressing CD11C, or were macrophages expressing CD68, or were macrophages probably M2 expressing 163? And again, not that percentages were higher in most of the occasions for the insulin containing islets of the type 1 D individuals, of the individuals with type 1 diabetes, sorry. And for the peri-islet region, again, no, when we analyze which was the percentage of the cell surrounding the islet of the individuals that had no diabetes or that had type 1 diabetes, we see that, around the insulin containing islets, there is a higher infiltration of these, higher percentage of infiltration of these immune cells.

And the same thing was observed when we analyze what we observe in the islet regions and in the peri-islet regions regarding the infiltration of leukocytes 345 or specific T-cell types like CD8 or CD4. And again, in the insulin containing islets of the individuals with type 1 diabetes, we usually see the highest infiltration. And with this cell detection, we cannot just see regions of interest. As I mentioned, we can see the whole tissue. So, we can use the tissue annotation as an annotation itself and analyze and do the cell detection for all the tissue within, and do the cell classification. And here you have another way of visualizing cell classification. But again, those cells were classified for the same markers for expressing 1, 2, 3 or 16 of the markers.

And here, you can get the information of the whole tissue. And here we had just four individuals, right? And every single dot represents the whole tissue, which is the percentage of CD68 positive cells in the full tissue of the non-diabetic individual, or the individuals with type 1 diabetes. And well, we just see graphs with four dots because, of course, we had just four cases in this test that we did. But basically, there is a higher overall infiltration, not just in the islet, but in the whole tissue for most of the immune cells in the type 1 diabetic individuals when compared with a non-diabetic.

And another thing that you can do for instance, in this case, we had the CD45 phenotypic marker that we can use to identify memory cells in black, right? Like, memory CD4 T cells would be expressing CD45RO. So, here we can see that at least two of the type 1 diabetic cases had a higher percentage overall of memory CD4 T cells than the non-diabetic. Or we can use other parameters as, we also had PD-1. PD-1 is a marker, sorry, that we can use to identify exhausted cells. So, here we can see how CD4+, PD-1+, so let's say exhausted cells, were a bit more prevalent in the overall tissue in some of the individuals with type 1 diabetes compared with non-diabetic individual. And in the third diabetic individual, the values were actually very close to the non-diabetic. These results are preliminary and actually not final. Here, I just want to show that, depending on which markers you have in your panel, you can actually get some phenotypic information that gives you an indirect idea about what was going on functionally that can be interesting in some studies.

And it's not easy to incorporate these markers when you really need to select just three, four markers for your study, but it's easier to incorporate them when you have more proteins that come incorporated in your analysis. And finally, as the last part of the analysis, this is still ongoing, so I'm still showing the process, but not really showing any data that is final. It is mostly preliminary stuff. But we also perform this machine learning unsupervised type of analysis with CytoMap. CytoMap is one of the two software’s that I showed at the very beginning. So, from that point, I was just showing the results that I got from QuPath. Now, I'm going to show what you can do with CytoMap.

So, this is one image of a tonsil, of the tonsil that I showed at the beginning. We did the cell detection using the tool inside QuPath. And then, we transfer the data into CytoMap. And we ask CytoMap to analyze the cell clusters and their distribution.

And it's very nice to see in this picture, because this tonsil is very rich in immune cells, how the germinal centers were identified as different clusters than some of the tissue around it, or some of the tissue that maybe was a more supportive tissue in the tonsil. So, it's actually really nice to see that, when you have the proper markers for the tissue, you can even get micro atomic details. And in this case, I did not train the system. I did not tell the system what is this cluster or what is this other cluster. These are the clusters that the system decided automatically. So, you need to introduce the parameters that need to change, depending on the needs of your, depending on the tissue and the data that you are handling. But once you introduce the parameters, the system selects the clusters autonomously, so automatically.

And here, we can see a whole pancreatic tissue section of one of our individuals with type 1 diabetes. And again, we did the same thing, we did the cell detection using QuPath. And we transferred the data into CytoMap. And we asked it to identify the clusters. And in this case, nine clusters were identified. And yes, there is a lot of blood. That means these cells were not really expressing enough immuno markers. But even in the case of infiltrated tissue, like the tissue with type 1 D, we still had no markers to identify the exocrine tissue specifically or to identify the endocrine tissue specifically. And for that reason, we don't really see that many anatomical details here. And we mostly see where the distribution of the immune cells was. And sometimes, well, we had a lot of cells not expressing any of the immuno markers actually.

So, for that reason, we also did a study. This is another of the type 1 diabetic cases, not analyzing the whole tissue, but analyzing our regions of interest. Because you introduce the data into CytoMap, you can introduce which data you are transferring, right? So, in this case, we transfer the data not coming from the whole tissue but coming selectively from the islet and the peri-islet regions. And islet and peri-islet regions were actually more rich in immuno cells. So, you can see how there, there are more of these clusters expressed in those islets. So here, for instance, in this cluster distribution...

And again, those clusters are different because the system decides the cluster, I do not decide the clusters, and depending on the data that you provide, will choose some clusters or others. Now in this case, for instance, we have these nine clusters. Cluster number one is richer in CD8 T cells. Cluster number two is richer in CD4 T cells. And you can see that these clusters are actually still smallish, right? Cluster number four is richer in antigen presenting cells, as you can see here. And there are four clusters that were mostly negative cells, like cluster number 5, 7, 8 and nine are all negative cells, or endothelial cells, or they are cells that were having other characteristics like being maybe out of fluorescent, that this could be maybe some of the beta cells that were CD45- and have a bit more of auto fluorescence. But still, we don't know because we didn't have a marker for that. And we have other clusters, like the number six, that was richer for S100. So here, you see in this first graph the percentage of the different clusters. Here on the very right, you have the explanation on what was its cluster. And then, here in the heat map, you can see the intensity of this mean fluorescence for every of the different markers in its cluster.

And also, with this analysis where we, in this particular case, we transfer the data just for the islets, we wanted to know how the clusters were distributed in the insulin containing islets and in the insulin deficient islets. So, in type 1 diabetes, the islets are mostly attacked while they contain beta cells producing and expressing insulin. So, once this attack has destroyed the beta cells, or the beta cells are dysfunctional and are not producing insulin anymore, the immune attack usually passes away, right? So, it stops.

It's like the arsonist of the first fire are gone and now you don't really see any of the guilty individuals there. You just see the destruction that is left after the fire. So, we wanted to see how these cluster distribution of the immune cells look in the insulin containing islets or the insulin deficient islets. You can see that there is a lot of heterogeneity, right? Like this CD, sorry, this first cluster one and two and four were for immune cells. And whether they are expressing more C8, CD4 or APCs changes. But in the insulin deficient islets, there was more homogeneity. There is little expression of most of the immuno cells there because the immune attack is actually gone. And there is still some heterogeneity in the presence of auto fluorescent cells or endothelial cells or S100 cells.

And then, another thing that you can do, these are just different ways of representing the data. Again, we have the insulin containing islets. The distribution of these clusters, not islet by islet, as we see as we saw in the previous slide, but in all the insulin containing islets and around all the per islets of the insulin containing islets, or in all the insulin deficient islets, or in all the areas around those insulin deficient islets. And we can see that both in the insulin containing islets and around them, we have more CD8 in red and more CD4 in green and more APCs macrophages and dendritic cells in purple.
So, to finalize, I presented the preliminary analysis of these three individuals with type 1 diabetes compared with this non-diabetic individual. Overall, there was a higher percentage of immuno cells, in general. In the whole tissue and in our regions of interest, we could get some information about the phenotypes, not just for the CD4 T cells, is what I show here where there were more memory on exhausted profiles. So, it was more common in the type 1 individuals. We could see this heterogeneity in the insulin containing islets, homogeneity in the insulin deficient islets. But again, those results were preliminary and I did them, went into a lot of depth.

We actually got some interesting insights about other proteins like, for instance, the S100 that now we are studying because the advantage, when you can add more targets, is that you can get to realize or to discover and to correlate some findings that, in other conditions, if you need to be more restrictive, you don't really get the chance to explore and see this type of data.

And then, finally, I would like, well, to kind of repeat or clarify that what I'm showing in all of these analyses pipeline, besides the fact of how we do things, is the fact that the data that we obtained from this Orion platform, it was actually really high quality. It was coming from difficult tissue because it was cut more than one year before. And pancreatic tissue of type 1 diabetic individuals is also relatively difficult to work with. So, even though this was difficult issue, we got really high quality data. We could do a fully quantitative analysis using these two platforms. And we could do cell segmentation without issues. Actually, we tried different platforms with this data and it was working fine for them. And this is important because, if the quality of the data is not good, you may not be able to do cell segmentation at all, which means that everything looked really good on our end.
And then, it's also important that the data that we can get is compatible with open source software or platforms such as QuPath and CytoMap. So, this is really great because it will just allow us, so, these open sources platforms allow us to get so much more information from our tissue. And in diseases like type 1 diabetes where understanding what is going on in situ is so important, it's really critical for us to get as much information as we can from there.

And in our opinion, this technology is suitable for this, for other type of studies, and even for some translational studies. And I just would like to thank all the members in the Matthias von Herrath lab in our microscopy career, especially to Sara McArdle, who is our engineer and who has helped me a lot with all the group scripting and with the machine. And I would like also to thank Erica, who has been help, who is the technician in our lab who has been helping me a lot with all of these CytoMap steps lately, and also help our collaborators, and also thank, sorry, our collaborators at RareCyte, especially Tad, who has been very helpful ally in this process. And thank you all for your attention.

Thank you very much, Estefania, for that really informative presentation. So, it's now time for our question and answer session. So, if you do have any further questions, you can still send these in now. So, you simply need to write your question in the box to the right of your screen.
So, to start off our Q&A today, my first question here is "Are you gleaning new insights from these methods?"

Hi. So yes, definitely. So, studying multiple targets at the same time, like, for instance, with this Orion platform where we can study up to 18 proteins, obtaining really good quality results is very important for us, because it allows us to have more context. And in a disease like type 1 diabetes, where all the immuno attack is happening locally in the pancreas, it is critical for us to understand what is really going on in this pancreas, right?

So, if we can study so many proteins at one time, this means we can't really get a lot of information from different cell types at one time. We can have the local context because we can have phenotypic markers, for instance, to identify what the cells are, what the cells are doing, and what are they in close proximity of, if they are in close proximity of blood vessels, if they are in close proximity of the islets. If they are. we can also understand.

We can also study proteins. And we can see the abundance and the specific location of these proteins. And this gives us a lot of information that mechanistically can be important. And also, finally, when we can study more targets and we don't need to be that restrictive with the number of proteins that we can study at one time, we can actually do correlations to understand locally what is going on.
For instance, we can correlate the results of the type of infiltration and the type of proteins that we see expressed in the islet with the amount of insulin that is left in this islet, or with the amount of SLA class one expression that the beta cells are having in those islets, which are things that are very important for us in type 1 diabetes to understand and correlate. So actually, it allows us also to include targets that otherwise probably we would sacrifice to study proteins that otherwise we would not include. So, it really broadens very much how much we can learn from every piece of tissue.

Great, thank you very much. My next question that I've got here is "What are the features of precision microscopy platforms that are necessary for your work?"

So, in my work in particular, for instance, it's very important for me to get the information of the whole tissue section. We have other systems that provide a really high resolution, but we can just study very small pieces of this tissue. So, with the Orion system, we got a really good resolution that was good for our purpose. But on top of that, we could see the whole section. So, we are not missing information. And in tissues like pancreas that are very precious, you really want to get as much information as you can from every section. And also, additionally, I need to work with systems that have a high dynamic range, this means where I can identify differences in the expression of my proteins of interest.

I mentioned actually that during the presentation. I just don't really need to know, many times, whether one protein is expressed, yes or no, in a cell. I need to know which cells are expressing more, which cells are expressing less. To do this type of quantification, I really need these high dynamic range. And the system provided this dynamic range. And also, I need a system that is sensitive enough to identify proteins that sometimes have lower levels of expression. And it happens many times that these proteins with lower levels of expression are actually biologically very important.

That's great, thank you. Another question I have here is "What is the sample preparation and workflow like with the methods and technology you are using?"

That's a good question. So actually, the system has many advantages, like, for instance, one of the things that I really like is that it's compatible with standard slides, which means that you can use fresh slides of tissue that you just got, but you also can use FFPE sections that you have stored.
For instance, all the results that I showed in this presentation, these were slides that were cut more than one year before I processed them. And it still... I had no issues with, not just pancreas, also with tonsils or with the spleens. And some of them were cut a relatively long time ago.

So, the other thing regarding the processing, so, the slide is treated for antigen retrieval, which is pretty standard. And then, there is the quench of the fluorescence. And then, all the proteins are labeled at once. This is done with a cocktail of fluorescent label antibodies. And this entire process takes just a few hours, so it's a relatively fast protocol. And then, the scanning takes a couple of hours per slide. This means that you can scan several slides per day, and then the data generated in a standard open format that is compatible with many commercially and open access software.
For instance, in my case, I processed the data using both QuPath and CytoMap. And the data was fully compatible with both platforms, but we actually tested other platforms where we were doing the cell detection with Cellpose and StarDist and it was compatible too. So, this data is compatible, as far as I could check, for all the access software or open access software that it was of my interest.

Thank you. I've got another question here which is "Did you get a definite answer to the question of HLA expression by the endocrine islet cells?"

Well, regarding the expression of HLA II, I suppose that you can find this paper published in Diabetologia this year. I think it was in February. Yes, I think so. So, whether pancreatic beta cells were able to express, this is not really related with the Orion thing, for the record. This is a question about the previous study, I guess. Whether pancreatic beta cells are able to express HLA class II or not was a very controversial topic for many years, actually. But in 2019, and because there were some reports supporting that, yes, these beta cells were able to express HLA class II, but there were other reports from the nineties where it was not that clear.

And if you... I'm not going to check all the literature, but you can find both yes and no. It started being like yes with the studies of Folis et al. But then, it was a bit more controversial later on. And then, nobody really was studying this topic for a long time. But in 2019, Russell et al, they published actually that they observed this expression of HLA class II and this also could be induced in the islets. And then, using different techniques, we also observe this expression of HLA class II in some of the beta cells, like something about 30% of them. So, it's not all the beta cells. And the expression of HLA class II was much lower than the one observed in the macrophages, for instance.

That was much intense, which could explain why it may be technically challenging sometimes to detect it, for that reason, not what we were talking about the dynamic range, why it's important to have a high dynamic range. And then, we also used isolated islets, native ones, and spheroids from in situ. And when we stimulated those islets with pro-inflammatory cytokines, we were able to induce the expression of HLA class II by the beta cells of previously healthy islets with these pro-inflammatory cytokines. In both spheroids that are clean, they have no immune cells contaminating because this is how the spheroids are created and with the native islets. So, I think that, at this point, taking the Russell information, this paper from 2019 and our paper from this year, I think that the fact that they can, they don't necessarily always express it, but they can up regulate this expression. It seems pretty clear to me that, yes, this can happen.

Great, thank you very much. We have time for just one more question, I think. So, that is "How do you approach validating or confirming that the machine learning algorithms are accurate?"

Yeah. So, this is a great question. For instance, it depends on which algorithm we are talking about. But for instance, for the algorithm that we run with QuPath, those are supervised machine learning algorithm. This means we need to train the system, right? So, it's not like I give the system the image and it spontaneously will tell me which are the person, which pixels will be positive or negative. No, no. So, I need to train the system channel by channel telling this is a positive pixel, this is a negative pixel. And you need to do that. You need to find the balance between what is enough training and not over-training it. Because when you over-train it, actually, it kind of gets crazy. And then, the results stop making sense.

So, the balance is a bit complicated. We found after, we have been working with this algorithm now for, algorithms, sorry, now, for three years, practically four years. And now, we know how many positive and negative training we need to use to be in the good spot. But still, the validation still and the quality control still require, it's labor consuming. So, you still need to check all your images, make sure that it still makes sense, because the systems sometimes can get creative and you need to identify that. So, the process of training is pretty straightforward. And now, we have an SOP after all the validation, and all the assay and error and developing with, of scripting and stuff with our engineer. But then, all the quality control always needs to be you really checking that everything makes sense.

Great, thank you. So, I think that's all we've got time for today. So, I'd like to thank Estefania once again for that really interesting and informative presentation, and a big thank you to everyone who joined us online. We hope you found this a worthwhile session.

If you do have any further questions, please feel free to email me at editor@selectscience.net and I will follow up with your questions for our speaker. Remember, you can download related resources in the tab to the left of your screen. And this includes the certificate of attendance. If you'd like to listen again to today's webinar or invite a friend to listen, it will be available to watch on demand in just a few day’s time. Goodbye, and thank you once again for joining us.