Modernization of Tissue-based, Biomarker-led Clinical Research

Thank you for the introduction and thank you to Leica Biosystems for giving me the opportunity to speak today. I'm going to present on modernization of tissue-based biomarker-led clinical research using some examples from our laboratory to show, to demonstrate how modernization of these techniques has really helped the research space and tissue-based research. 

One of the things we want to talk about is how modernization has really advanced in the last number of years in tissue-based research. Pathology has been around for a long time, but it's only really been the last 20 years we've seen an advancement in the field, particularly for researchers. And this is mainly because of the introduction of autostainers, so automated systems of being able to stain samples that could be assessed by pathologists. By introducing these other standards, it means now we can look at standardization of protocols and robust methods of analyzing tissues. This has led to the development of new systems, advanced biomarker systems to detect proteins or targets in the tissue. This may be through novel probe designs, through RNAscope and DNA-PAINT, or through advances in polymer technology or fluorescent dyes to detect the tissues through the Opal or UltraPlex systems, or integrating these scaling platforms with sequencing platforms through the AccuCyte or now more increasingly the GeoMx DSP platforms. 

What are the considerations we need to understand to really utilize these autostainers and these systems to their best advantage? As a researcher, we need to appreciate the tissue that we're analyzing. We need to appreciate what type of tissue type it is. Is it from the breast? Is it a brain tissue? Different types of tissue may be affected by the fixation process, has been formed and fixed. Is it fresh frozen? Is it mescarons, some more unusual types of fixation? How old is the sample? Is it prospective cohort we're looking at so we can control all of these factors, or is it an older sample where it may have been taken as a block 20 years ago, or it may have been a cut section dipped in wax. All these factors will affect how the staining process is undertaken, and even then, how the sample is processed in the first place and stored will affect our sealing process.

Due to how long pathology has been around in the clinical practice, a lot of the tissue components we can't really control and are quite standardized within the literature and within this research space. More interestingly for us is the staining procedures that's available. So, this is one of the biggest considerations. What I mean by this is, well, what type of stain are we going to use? Are we using routine H&E stain, structural components of the tissue, or are you going to use a special stain, or are we going to look at using a more targeted method such as immunohistochemistry to look at specific proteins, or in situ hybridization methods to look at specific DNA or protein or DNA or RNA targets in the tissue. Depending on what type of stain we decide to use, this will affect what type of processing we undertake. Do we have to use very thin sections, quite thick sections, they have to consider what sort of stain and what sort of why do we need to cut the sections to prepare for this. From that, well, if we're going to take these sections, how are we going to process them? Are we going to hand dip them using several different baths of dyes? Or are we going to automate this process using the system? If we are using the system, are we going to use commercial reagents? They're more standardized, we can buy them in bulk? Or we're going to develop them in-house to use in these processes, which can have a lot more variation in the stain that's produced in a large sample size. 

If we're looking at more specific techniques, such as immunohistochemistry in-situ hybridization, what's the choice of antibody? What's the other design or probe design choice? This can have a significant impact on how sensitive and specific our assay will ultimately be, and this can be affected by the antigen retrieval method we choose. If we decide to do a manual method, we can use a manual staining, we can use a pressure cooker or more novel techniques of antigen retrieval. If we're using an automated system, we might be limited to heat-based with a pH 6 or pH 9 or enzyme-based retrieval. 

What type of detection system are we using? Are we using a bright field detection system or are we using a fluorescent detection system? All of these things are important to consider whenever we're looking at conducting tissue-based in-situ analysis. If possible, simplifying, standardizing, automating these processes makes it much easier to compare assays and to really interrogate the tissue in an appropriate manner.

In our lab, a great example of this is looking at HPV status in oropharyngeal cancers. Now, if you're not familiar with oropharyngeal cancers, the oropharynx is part of the head and neck, and it specifically is the very back of the throat, the two tonsils, back third of the tongue, and that very back wall, no other part of the mouth. And the reason why this specific tissue type is important in head and neck cancers is because cancer can develop due to smoking and alcohol-related risk factors, but also due to HPV infection. An HPV infection can be detected using a surrogate biomarker, known as P16, allowing us to stratify patients into two distinct phenotypes, HPV negative and HPV positive, and this does have implications for a patient outcome. As I said there, we use a surrogate biomarker and there's been advances in the technology for biomarker detection. In our lab, we looked at different methods of assessing HPV testing and what will be the best tests we could use to identify HPV staff in patients.

Using an epidemiological or pharyngeal cohort from Northern Ireland, we took patient samples and sent them for HPV genotyping to a reference lab within the UK, and also created tissue microarrays in order to do tissue-based analysis using protein-based DNA and RNA-based tests. We ran the P16 through a reference lab and assessed it based on the Western criteria, where we had either moderate or strong staining of greater than 70% of the tissue was considered a P16 positive case and therefore likely to be HPV related. We also ran RNA-ISH in-situ hybridization using the RNA scope ZZ probe design and assessed them for the presence or absence of this brown precipitate on the samples. We also ran HPV DNA-ISH using additional methods and an alkaline phosphatase detection system, which you can see is a slightly different visual appearance. 

After assessing all of these samples using the four tests, we then collated the results for the positive cases and the genotyping results and seeing how closely they were related. And what we found is that for the surrogate biomarker p16, both tissue-based DNA-ISH and RNA-ISH tests were always positive when p16 was in-situ hybridization was positive, or immunohistochemistry was positive. However, the HPV genotype, which was a DNA PCR-based test, was the only test that identified cases that were P16 negative HPV positive. We looked more closely at the tissue-based phenotypes we found. And as you can see here, we had cases that were positive by all three tests, P16 immunohistochemistry, HPV RNA-ISH, and HPV DNA-ISH. We had cases that were P16 positive and HPV RNA-ISH positive, but DNA-ISH negative, which supports the use of the novel probe design from RNAscope. We had cases that were unusually P16 positive and HPV positive or HPV negative by both RNA and DNA tests, in addition to negatives by all three tests. We looked at sensitivity and specificity of these tests against P16 staining. We found that all three tests, the DNA-ISH, RNA-ISH, and genotyping, were all equally sensitive, but the tissue-based tests were much more specific than the DNA-based PCR. 

We wanted to interrogate these a bit further to see, well, did these phenotypes, the P16 positive HPV negative phenotype, really exist? Does the P16 positive, P16 negative HPV positive genotype really exists? Using a time trend analysis, we assessed the occurrence of P16 negative HPV positive patients for all three tests across the time period gathered for the patients. Now, as you can see, there's only one line in this graph, and that's because only the genotyping identified cases like this. And as you can see, there's actually a bias where the older the sample was, the more likely we're going to find it being called HPV positive by genotyping, indicating that actually the HPV genotype may may not be no specific test to use for this assay. In contrast, when we looked at the P16 positive HPV negative patients, we identified this phenotype across all three tissue-based tests in relation to P16, and there's no time bias here in this assessment. And the reason why it is important that we're able to identify this phenotype using these different tissue-based tests is because the surrogate biomarker P16 identifies patients who have really good survival versus poor survival with the relating to HPV status. When we look at the subtype of P16 positive HPV negative in relation to survival, we find that these patients actually behave more like an HPV negative disease. Considering them HPV positive through P16 assessment alone would not be appropriate. In particular, this phenotype is a very rare phenotype across the time period, requiring it in less than 10% cases a year. It is present and we wouldn't have been able to detect it if we hadn't been able to use a modernized system to robustly assess these different assays in conjunction with each other. 

Why is it important to understand what the assay is detecting and the limitations in sensitivity and specificity? Well, in head and neck cancers, and in oropharyngeal cancers specifically, HPV-positive disease tends to present in patients with larger neck and nose compared to HPV-negative disease. Traditionally, the TNN staging for oropharyngeal cancers wasn't able to accurately stratify patients for outcome. Some very clever people in the United States developed a new method of clinically staging patients adjusting for P16 status using the size of the neck nodes. We applied this new staging system to our cohort retrospectively and found that we were now able to accurately stratify patients more for survival outcomes and they were P16 related. However, this subgroup of P16 positive HPV negative patients still existed in this cohort. If only P16 staging, P16 classification was used for these patients for the new staging system, 95% of those patients will be under staged under the new system. This gives you an idea of why it's important to understand your assay limitations, but more than that, why it's important to consider using these novel methods, how we can identify rare phenotypes because it may have clinical implications. Some of this work has went forward to, with others, has went forward to making changes in the UK guidelines to look at HPV status in head and neck cancers. 

This brings a new question. What do we do in terms of controls? We understand from traditional immunohistochemistry when we need to look at positive and negative controls in the tissue. We look at the positive tissue and we run our antibody and we use a negative tissue and we run our antibody. Then we can genuinely understand that our tissue is negative or positive for our antibody of interest. This new technology with RNAscope, or even using in-situ hybridization, how do we know that we can trust the staining being produced when in sequencing analysis we have housekeeping genes? 

Alongside RNAscope's new probe design, they also introduced the use of housekeeping genes through the probes. In our lab, we assessed what would be the best way to assess the housekeeping genes in these tissues so that we can robustly assess them for RNAscope analysis. So, they have three different housekeeping humans available. If we look here at higher modification, you can see that they're varying expression levels, low, moderate, and high expression level. Using digital image analysis, we assess these in tumor regions and in stromal regions. We found most consistently in prospect of samples, which have been controlled for their fixation, that PPIB assessing the tumor was the most robust method of assessing how good a quality the tissue would be for a quality assessment for RNAscope analysis by a novel biomarker. 

We went back to our retrospective cohort of colon samples and used a qualitative metric of PPIB scoring from zero to three. We assessed all of our samples visually for their ability to be included in assessment of a novel biomarker. As you can see here, there is varying quality of the RNA available in each of the tissues. So whenever we assess then our biomarkers of interest, PD-L1 RNAscope and CMET by RNAscope, and compared this to our PPIB scoring, you can see here consistently the low PPIB score was associated with a low PD-L1 and CMET score, indicating that when the tissue quality may not be as good, this affects the ability of the RNAscope probe to be assessed using our biomarker of interest. And as you can see here, we did a time trend analysis of PPIV scoring across that colon cancer cohort. There was no real effect from time, and it may have been these samples were probably fixed in the first instance when they were being created. And in particular, you can see there's a specific, there's a strong association of low PPIV scoring with CMET assessment indicating that if the PPIV score is one or less, it shouldn't be included for the assessment of the novel biomarker. And by doing this type of assessment, it means then we can look beyond what we know clinically that is already accepted as a biomarker. HPV status is well known in the literature as being predictive and prognostic of survival in oropharyngeal cancers. What about if we have a different cancer type we don't know with prognosis of the biomarker? Or there may be, it's going to be very difficult to stratify the patients. 

You can see that as an example. CMET inhibitors have been used with great effect in lots of cancer types, including lung, to treat patients. However, there's been less success of their use in other cancer types, such as colon, for example. Part of this is due to lack of standardization of how to assess CMET status in patients in order to provide their ability to give them the inhibitor. So, we decided to assess using a similar method, the HPV study. The value of CMET staining and expression within a colon cancer cohort. So, using the Northern Irish colon cancer cohort, we assess the patients for targets which are normally utilized in lung cancer, looking at genetic aberrations, so using target sequencing and looking at MET amplification. We also looked then at CMET staining using immunohistochemistry, which has been also used quite widely for patient stratification and when providing the inhibitor, in addition to RNAscope, and we only looked at cases for the RNAscope which have passed our quality assessment by kind of PPIB scoring. From this analysis, we found that in the colon setting, very few patients actually had a met genetic aberration. However, whenever we compared and combined the c-met RNA and c-met protein expression, we were able to identify patients who may, who not only had a very poor prognosis, but also more likely to benefit from met inhibition because they demonstrate true met addiction in the samples. 

Modernization of these techniques not only allows us to robustly identify these rare phenotypes, it's now allowing us to combine the tests to identify these phenotypes. But this is not the only thing that's come along with modernization of the technology. We're also seeing computational advances in addition to the wet lab advancements. These computational advances came alongside the autostainers, the introduction of whole slide scanners to digitize these images. Commercial, then this resulted in the release of commercial image analysis software for whole slide imaging. And then more recently, open-source methods of assessing these whole slide images, such as QuPath and now we're seeing more and more in literature the application of deep convolutional neural networks and more modern methods of AI being applied these whole face histological images to stratify patients. And we've been able to use these convolutional advances in our labs to our benefit to conduct research at scale. 

For example, in this colon cancer cohort, we wanted to look at, well, how is immune status prognostic? What sort of measures of immune status are important to consider? Using 8 different biomarkers across a cohort of around 1,800 patients from three different sites in the UK, we were able to generate tissue microarrays and take slides for staining. Using a Leica BOND RX, stain each one of these sides for eight different biomarkers very robustly. We actually only had to re-stain one sample, so less than 0.5% of samples had to be re-stained to pass quality control. And from this wet lab experiment, we were then able to digitize this information using the Aperio AT2 scanner so that they'd be processed by digital image analysis to create digital results. 

What's great about being able to do this type of research at scale is all these immune biomarkers are quite nuclear. So, we were able to then take these samples, these digital tissue microarrays, de-array them and conduct cell detection and positive cell detection for each of the different biomarkers of interest. We're able to complete this at scale, generating around 50,000 novel data points for IHC data. This digitization of the data meant then we could use sophisticated statistical techniques to look at grouping, subgroupings present in the cohorts and how these potentially could affect survival. And once we've identified a good survival stratifier. Because this data is now digital, we can integrate it with other data sets, such as multi-omic data sets. 

We then combine this data with RNA microarrays. to see, well, is there this Is there a transcriptional subtype with the phenotype we've identified through the tissue-based analysis? It appeared that there was when we did PCA plots. When we did further investigations using heat maps, we found there was a specific number of genes which were associated with hypoxia. When we did a gene set enrichment analysis, this signature was very highly enriched. But because we've started off with a tissue-based analysis and combined this with the other multi-omic data sets, we're then able to go back into the tissue to determine if this is a genuine phenotype we're seeing in the other data sets.

Using this, we could then spatially assess, well, what is the relationship with hypoxia to the immune biomarkers? As you can see here, when the patients are immune cold, the hypoxia was very, tissue hypoxia was very high within the tissue, the tumor regions. What this means now for us as researchers is we can then now begin to combine different types of data, including tissue-based, to start to come up with new hypotheses for how these patients are responding to treatment or how they're prognostically being defined prognostically. For example, here was a stratifier in the sub-cohort of escort patients. Whenever we look at this in relation to hypoxia, we see that there's patients who have very high immune responses and a very low hypoxia who do very well. This may give us a rationale to start looking at different measures of assessing and understanding why patients who have a good immune response do well for treatment.

This brings us to one of our final considerations for tissue-based assessment, and this is the biomarker assessment. We've got a very strong rationale for how we prepare our tissue, whatever our choice of stain is, how do we assess these biomarkers? Do we need a pathologist or could it be a non-expert assessing these for how significant they are? Do we do it on the glass, down a microscope, or can we do it digitally at a screen? If we're doing it digitally at a screen, should it be manual? Should it be digital? And if we're doing it in these methods, what's our actual scoring method? Are we using percentage? Are we using area? Are we using number of positive pixels? What's the difference? What difference does that make for the pathologist or the image analyst conducting this research?

What type of scanner is also important. If we decide to digitize these samples, the different types of scanners may introduce new image artifacts by changing the colors present in the image to make them look more like you're looking down a microscope. What magnification is also important. It depends on how you're scoring it and how you're assessing it. The magnification has more or greater of an effect. And also the long-term storage. If you're digitizing these samples, how much data will this take up? Will it still be accessible in 20 years' time with file types change? We have to start thinking of these types of methods before we're just starting to do tissue-based analysis using digital means.

This is all only for one biomarker. What if we think of doing multiplexing? What if you have multiple biomarkers in a single image and have to consider all these factors? Within our lab there are different methods of multiplexing, but we look specifically at an indirect immuno-labeling using the Opal detection system. And if you aren't familiar with multiplexing, multiplexing involves using indirect immuno-labeling where you take your tissue, you apply your primary antibody followed by your secondary antibody, and then you develop this product. If we're doing DAB detection, it'd be brown, but if we're doing multiplex immunofluorescence, it's a fluorescent label. This method-based in fluorescence allows us to then do repeated rounds of staining by applying heat to remove the previous rounds of antibody staining, but leave the precipitated immunofluorescence present on the slide. This allows us to go back in the second round of antibody staining with primary and secondary, develop a different color, and repeat the process again, so on and so forth.

This isn't novel. This has been around for a long time. We've been able to do multiple rounds of staining as long as we will do in immunohistochemistry. However, as I mentioned earlier, there's been advances in the polymer detection technology and also the dyes that are available to conduct fluorescence. One of the problems with using this indirect immunolabelling method is known as steric hindrance, and this occurs whenever these dye reaction products are developed too close to each other. We wanted to look at this new method that's been produced through Opal to see how difficult would it be to generate a panel design in samples that had co-expression in the same cellular compartment and co-expression in different cellular compartments for immune biomarkers of interest. Using these methods, and then how difficult it would be to assess these digitally.

For this assessment, we were talking about using repeated rounds of staining. I mentioned earlier that antigen retrieval is a very important step to consider with the staining process. The reason why I said this is because whenever we form a fixed tissue, which is what we tend to use for histological research, we create these protein cross-links in the tissue. In order to allow our antibodies to bind to the tissue, we have to unmask the mutin, heat induced epitope retrieval. Whenever we're doing repeated rounds of stain, you can start to denature these proteins present in the slide. Depending on the protein or a biomarker of interest, they might be more or less affected by multiple rounds of antigen retrieval.

In this study, we looked at how we could determine the best method or best order of applying antibodies in order to obtain the most accurate representation of the staining pattern. Using low and high expression tissues of interest, we looked at applying our biomarkers to several different rounds of staining. You'll see here I'm presenting DAB staining. Because this is using indirect immune labelling method, we can actually replace the dye product with DAB for us to better understand how the stain is behaving when we're doing our initial optimization process. As you can see here, in our low and high immune expressing, we have variable expression depending on how stable the epitope is after multiple rounds of heat induced retrieval. This allows us to determine what would be the best order to successfully stain for antibodies in tissue when we're doing our panel designs.

Once we have an idea of what type of order the antibodies will be suitable, we then need to think about, well, how are we going to pair them to the fluorescent probes? We need to consider where they're situated in the cell. Is it in the nucleus, is it in the cytoplasm, is it in the cell membrane? How abundant it was in the tissue, if it's very, very low levels of abundance, we should pair it with a higher expressing fluorophore, or if it's very high expression in tissue, we should express it with a lower expressing fluorophore. If these are in the same, if the sample is expressing in the same compartment, we should try and spectrally separate them then so we're not having overlap of dye products when we're trying to excite the dye for image capture. 

In order, once we've decided, based on what we've done there with the IHC, what was the best order and how we're going to pair them with these fluorophores, we then conducted this assessment with the same tissues and compared it directly to our IHC assessment previously. As you can see here, most of the fluorophores matched up pretty well to their DAB compartment. There are previous DAB staining iterations. However, for some, there was some significant differences. But when we look at these more closely, it's either due to the result of it being from a non-serial section or due to poor cell detection in the DAB immunohistochemistry. 

This gives us an idea of the antibody ordering we apply and also how we pair our opal antibody dyes up for the multiplex panels. It allows us to do our first round of multiplex staining on our system. During our staining process, we're then able to review how well our multiplex designs worked. As you can see here, if we look more closely, our CD20 has some overlap into the CD8 channel. Our cytokeratin has some overlap into our CD4 and CD8 cells. The CD20-CD8 relationship is due to a thing called spectral bleed-through. Both of these antibodies we used two spectral dyes, which were very close in range. So, by changing the two different dyes used, we're able to better respectively separate and not have a seed through in later iterations of staining process. And with the cytokeratin, we were able to change around our antibody order that has been applied.

These fluorescence thermophores are much more sensitive than dab reaction products. So even though something looks as though it might be a good fit, by the DAB assessment, initial DAB assessment, whenever you start to pair it with the fluorophores, you may see it's more sensitive than previously anticipated.

Another thing to look at as well for when you're doing your optimization of the multiplex, you're starting to use panel design and plan on assistance for optimization is does your heat induced epitope retrieval successfully remove your previous round of staining. The reason why this is important, particularly if you're doing manual staining, is whenever you're doing your heat induced epitope retrieval, if you have only partial removal of your previous round of staining, when you do your second round of applying your primary antibody in your secondary, you may then start to develop a secondary product, veriforescent product, on inappropriate targets and non-specific staining. The easiest way to check for this is by running drop-by controls, where you apply your round of antibody staining, your biomarker interest that you're interested in with its fluorophore of interest. In the next round of staining, omit the antibodies for that round, but apply the fluorophore to confirm that you have complete removal of the previous round of antibody staining. After several successive iterations, you'll eventually develop your multiplex panel of interest.

The reason why it's so important to consider all these is this is only the wet lab validation. By using an automated system, we can control a lot of additional factors through using commercial reagents or by having the system using specific times and heat settings. This means it's simpler for us to optimize by just changing small factors like the antibody concentration or the antibody opal pairings. or even the order of staining. 

We also have to start considering the digital assessment, how are we going to create these digital images? I mentioned previously that the importance of magnification when looking at specific biomarkers of interest. In fluorescent staining, the magnification is directly related to the amount of fluorophore emitted. So because the fluorescent light comes through the microscope lens, if you scan at times 20 versus times 40, you'll actually have brighter signal from times 20 scan than the times 40 scan. And depending on the biomarkers of interest, this may be more or less important depending on what you're looking at. We didn't find a difference in our study, but it may be something to consider. 

Another thing to consider relating to the excitation of the fluorophores is your exposure times. If you overexpose in a channel, you can create artifacts of the auto press of staining that may be misinterpreted as genuine staining when you're doing your digital image analysis because it's created on the image as an artifact. You can look at this by looking at individual exposures versus batch exposures in the tissues. If you're like us and use a more advanced system and look at multispectral fluorescent imaging instead of using a fluorescent camera with narrow band pass filters, you can now begin to look at how well your samples have been prepared by looking at the fluorescence channel to see if they're particularly thick or thin sections. This may affect the quality of stain that's produced. And further to start, how effectively spectral mixing allows us to assess separate channels. 

The reason why this is important is because in multiplex and in any fluorescence assessment, unlike bright field images, every channel in a multiplex image has unique information in it. This is quickly important to understand when you think of the digital analysis, because the order in which you apply your levels of self-detection may have a very minor influence on the number of cells detected, or it could have a very major influence, depending on how well you've been able to optimize your panel in the first instance. 

How do we assess this in our study? We decided to look at the best way of doing quality assessment on our multiplex analysis was through using a multi-tissue TMI block, where we stained it for the DAB markers of interest using our single-plex assessment and also our two multiplex panels of interest using different tissue types to get a different range of staining patterns. And we assessed the chromogenic single plex and multiplex as single biomarkers and combined the assessments. And as you can see here, when I talked about our two different panel designs, our multiplex panel one, which had co-expression in the same cellular compartment and required more rounds of multiplex optimization for the wet lab work, was also the most problematic in terms of the order in which the cells were detected and quantified in the tissue, where you may have certain biomarkers going from being 75% of the sample to 44% of the sample. In contrast, our co-expression different study compartments, so we looked at the multiplex panel two, only took two rounds of a multiplex wet lab optimization and consistently demonstrated similar quantities across all different measure orders of biomarkers applied in digital image analysis. 

It's important, as you can see here through this study, it's important to consider how the wet lab work influences your digital analysis. What happens if we can't control our wet lab or our digital components? What if we only received as a researcher an image to work on? Increasingly, we are seeing in literature the use of artificial intelligence being used on H&E images from anywhere in the world, being used to develop models for patient stratification. But one is, but this type of analysis is still hindered by the similar problems we've got from other tissue-based research, where we're having a significant variation in the staining quality produced, just because it's from different labs and different types of hemotoxin and eosin batches could be used. We looked at this in our lab and developed a tool to help clean up these images that we can get from different sources where we can't control the pre-analytical factors or the wet lab factors. Using this tool called HistoClean, we can take the images, break them up into tiles more suitable for you, for applying artificial intelligence. 

We can then take them and start tiling them up by removing the tiles that have got significant areas of white space. We can take the tiles, compare them to a color palette where we know is more relevant to us to adjust the color ranges. We can also start to apply some basic augmentations, which may be beneficial to the artificial intelligence models, depending on what you're looking at. In order to see how beneficial applying these types of augmentations would be in image pre-processing, we took data that was going to be used for artificial intelligence model and applied the different methods available through histone on the images before developing the model simultaneously. 

Taking our oropharyngeal cohort again, we assessed the patients for stromal maturity and stromal immaturity. And using different types of augmentations, found that by just adding some simple augmentations, you get a significant improvement in model accuracy. Just change the augmentations, not change in the image source, not change in the tissue type, just very small amount of adjustments. This instance, we found just by adding, enhancing the features of the stroma, the fibers in the stroma, and by balancing the data set, we saw significant improvement. 

Hopefully today I've been able to demonstrate to you the benefit of using an automated stain platform to perform your tissue-based research. But using these platforms is much easier to optimize complex immunohistochemistry in situ hybridization protocols. Once you have an optimized assay to reproduce this in your large cohort studies at scale. 

In addition to using these automated platforms, it makes it much easier to standardize these protocols, particularly within new technologies, and allowing for direct comparison of the new immunohistochemistry in-situ hybridization and staining procedures. This means now we can look at identifying rare phenotypes with confidence in these rare cohort studies. Further to that, hopefully I've shown to you that whole-slide imaging is of a benefit in this field. It enables quantitative and sensitive analysis of biomarker studies and enables data sharing to further this research. It also allows us, because the data is digitized, to integrate the image data with multi-omic data sets and take it forward for use in artificial intelligence studies. 

With this in mind, the considerations we had previously of tissue, staining and biomarker assessment, prior to optimization and modernization of these techniques within the laboratory, most of our time as a researcher we spent optimizing the tissue and staining process with limited time for biomarker assessment. Whereas now with automation and host site imaging, we're able to spend less time optimizing the assays and more time assessing the biomarkers in novel ways. So with that in mind, thank you for listening. I acknowledge my principal investigators, Professor Jackie James and Professor Manasil Fellas, as well as the researchers involved with data creation and cohort creation in the head and neck cohort and in the colorectal cohorts, as well as the funders. Thank you very much for listening and I'll take any questions.