2026-04-21-pm-late-spatial-biomarkers-single-cell-resolution

Date: 2026-04-21 Primary topic: single-cell-spatial-omics Cross-referenced from: bioinfo-tools, single-cell-spatial-omics Word count: ~16,489

Auto-generated captions

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

Okay. So in the interest of keeping this on time, I think we're going to go ahead and get started. I know people are still filtering in a little bit. So I'm Dr. Joshua LaBaer, and this is Dr. Irene Gabriel, and we're here to lead this session. And if I can, yep, switch to this. I'm going to do a couple of quickie introductory notes just because I've spent most of my career looking at biomarkers, and so I thought I would talk in general a little bit about biomarkers just to set the stage for a lot of what we're going to hear today. So biomarkers are basically biometric measurements that convey some kind of information about the subject's either biological or clinical state. We can think about biomarkers in terms of how they're going to be used, which I think is a big part of what we'll hear today, where the material comes from, what type of material people are using, and how thoroughly the biomarkers are validated. In terms of clinical use, a common use is to stratify patients into specific subtypes of disease. You could think of the MSI status in colon cancer, for example. They're definitely useful for looking at prediction in terms of use of therapies. So are they a good companion diagnostic? We'll hear a number of examples of prognostic biomarkers, so essentially assessing how aggressive a cancer will be and where it will end up. We can think about monitoring disease progression with biomarkers. This is commonly done in cancer therapy, looking at things like CEA in colon cancer. The area that I've been interested in is early detection biomarkers. Can we pick up cancer before they're clinically apparent? And then, of course, there are markers that actually predict risk of cancer. So you can sometimes use biomarkers defined for one approach for another approach, but that doesn't always work great. So in terms of types of markers, this is a list of common types of markers. Obviously, today we're going to be thinking about spatiotemporal sampling. So looking at different time points and different locations and also measuring various types of omic responses is what we're going to be focusing on today. Just as a quick note, the success of getting biomarkers through the FDA has not been stupendous. In the past 10 years, there have been five tests that have been cleared by the FDA for cancer detection. So that speaks to the challenges inherent in finding good biomarkers. And just to outline those, first of all, it's very difficult to find a difference in a biological specimen that is predictive of outcome. It's especially hard to find them in accessible fluids, and it's ridiculously hard to find them that are not affected by other related diseases. So, other things that are challenging, as many of you know, just to get these studies off the ground, getting the appropriate clinical samples is a big challenge. And probably one of the biggest challenges is even if you find a difference, getting the funding to do a validation study is especially challenging. You will often get reviews back on your grant proposals saying that this has already been done. Why are you doing it again? So, just to think a little bit about validation. Typically, biomarker studies begin with making some kind of an observed difference between cases and controls. And if you do enough of those together, you will get a subset of those markers that you would think of as candidate biomarkers, once you define their sensitivity and specificity. From there, you really need to verify that biomarker by repeating it in a fully independent set of samples. So nothing in common with the previous discovery set, so you can be sure that they really are not false positives. Then you're set up to do an actual validation study, which should be a prospective blinded study to see if, in fact, the markers are truly predictive. After which you really need to do a utility study to show that by using those biomarkers, you can actually change the clinical benefit of-- you can improve the clinical benefit of people being treated. So it's a long road to get to an approved biomarker, and all of it depends on first defining what's the clinical need that you're going to address with the biomarker. So lots of pitfalls. Discovery without defining the clinical need, doing inappropriate statistics, doing an underpowered study, failure to account for overfitting, which is a huge issue in omics. If we're looking at lots and lots of variables, it's very easy to find something that doesn't hold up when you repeat the study in a different population. Not using the right samples, failing to develop a robust and reproducible assay which can be done over and over again and reliably, and not doing the appropriate validation studies or the utility studies. So today, just coming back to the biomarkers by clinical use, we've got a couple of great talks lined up for looking at disease stratification markers. We've got a talk looking at markers that are good in predicting immunotherapy response, so a companion diagnostic. We've got several talks on prognostic biomarkers and one talk on an early detection biomarker. So I won't spill the beans on what they're going to talk about since they're going to do it in a moment. So I just want to thank you, and we'll start with the first talk. All right. Thank you so much. And our first talk is on spatial transcriptomics profiling of colorectal cancer tumors in Alaska Native people: Discovery of prognostic biomarkers by Geryk Cheng. So thank you very much for coming to my talk, and thank you, AACR, for selecting my talk for this mini symposium. So today I'm going to talk about the spatial transcriptomic profiling of colorectal cancer samples in Alaska Native people and trying to discover prognostic biomarkers. And I have nothing to disclose. So colorectal cancer in Alaska Native people, this is one of the most commonly diagnosed cancer and there exists stark disparities where they are having the highest rates in the world compared to the US average. Their incidence rate is two times higher, and their mortality rates are almost three times higher. And these rates cannot just be explained by the screening because their screening rates are similar to the US public So we want to understand how to have the best available methods to increase their survivability. So the current knowledge gaps and barrier is that there's just a lack of clinically useful biomarkers and tools for predicting clinical outcomes that's tailored specifically for Alaska Native people. And as you all know, most established researchers are focusing on non-Hispanic white people. And by having effective prognostic biomarkers for them, it allows a robust risk stratification and preventing over or under-treatment and overall, increasing their survival likelihood. So, in order to try to find the spatial biomarkers, we decided to look into both the tumor and the tumor microenvironment around the tumor. And, just a quick definition of TME, it's the ecosystem surrounding the cancer cells, and they have different cell types, different soluble factors. And the spatial organization of these cell types of the tumor and TME and how their pathways are enriched can drive the development of the tumor, how the disease going to turn out, and also how the patient's going to respond to therapy. And in here we have a really unique study population, thanks to our close collaboration with the Alaska Native Tribal Health Consortium, where currently we have 217 Alaska Native patients who have colorectal cancer, and we have their FFP tissue blocks. And, the patients who die from colorectal cancer are matched one to two to patients who did not die from colorectal cancer. And, these data have been processed with GeoMix and additional 100 patients are also being generated with GeoMix data that we will be including in our results in the future. And, for these tissue blocks, we will then extract tissue cores for them and arrange them into tissue microarrays. And apart from GeoMix, we also have other spatial omics profiling performed on them, including bulk RNA-seq, PhenoCyte, which is single cell spatial proteomics, and also Xenia, which is single cell spatial transcriptomics. And, this one is our characteristics of our study population, which also includes matched on sex, tumor stage, and tumor sites. So, the way that we have our different cores is that we first, for each patient, we identify cores that are two cores from tumor center and then one core from the invasive margin. And then for each core, we also have identified the epithelium and the stroma, which corresponds to the tumor and TME respectively. And we're using the pencillin marker to identify the different regions, and we also have pathologists to verify that. And in total, we have 1,140 total tissue compartments that is currently used for the GeoMix analysis. And, as I mentioned, we're using GeoMix, and the specific part about that is that it allows for a UV-dependent cleavage of the oligo-labeled RNA probes. So since we know where the epithelial segment and the stromal segment is, we can then just shine the UV light onto there, and then it can release the RNA probes, so we're able to know the spatially specific transcriptomic profiles of those tissue compartments. And the bioinformatic pipeline is relatively straightforward. We first have quality control, where we kick out any segments and genes that are underperforming. And then we also have batch effects correction, as our data is being processed in different batches. We want to make sure all the differences that remains are from biological differences and not experimental differences. So, we did the batch effect correction for that, and then we also did normalization to make sure every gene's in an equal footing and easier for comparison. And then finally, we have the downstream analysis, including differential gene expression, gene set enrichment, and also cell type compositions. And for the results. So, first we have the differentially expressed genes. So we are able to identify spatially compartment-specific expressed genes. So specifically for the epithelial compartment of the tumor center, we see that for the patients who did not die from colorectal cancer, there is a upregulation of SPINK4 and MUQ2 genes. And then for the epithelial compartment of the invasive margin, we see a upregulation of HBA2 for patients who die from colorectal cancer. And all these differentially expressed genes are being processed by using DESeq2, and then it's being adjusted for all the measuring co-variables I mentioned in the study population, including sex, tumor stage, tumor site, age, and year of diagnosis. And we also did a randomization to make sure we don't have any effects of the outliers. And once we have the results from the differentially expressed genes, we can then also try to see which specific pathways are being enriched. And, two things to note here. One thing is that we've identified that in the epithelial compartment, in both tumor center and the invasive margin, there is an enriched pathway of the interferon alpha and gamma response for patients who did not die from colorectal cancer.And that definitely makes sense. And also we see that in invasive margin, we see that both epithelial and stromal compartment, there is an upregulation of epithelial-mesenchymal transition-related pathways that are being enriched for patients who die from colorectal cancer. And that also makes sense because as EMT is being known for epithelial cells losing their cell identity and getting more fluid and easy to get metastasized. So it also makes sense why they are being upregulated for patients who die from colorectal cancer. And, also we are able to determine these pathways based on the reference database using Homologene, KEGG MEDica, and Reactome. And then after that, we also have cell type deconvolution results, where we are able to look at the cell type abundance score for the specific tissue compartments. And this cell type abundance score is being determined by using SpatialDecon, a publicly available library specifically for geomix results. And we're doing this with a custom cell reference based on a publicly available colorectal cancer single-cell RNA sequencing datasets. And what we can see in here is that for the epithelial compartment at the tumor center, we see an increase of plasma cells for patients who did not die from colorectal cancer. And for the stromal compartment of the tumor center, we see that an increase of mast cells for patients who die from colorectal cancer. And just to have a quick validation to make sure our cell deconvolution results are valid, we also check it with the consensus molecular subtypes we determine with our bulk RNA-seq results. And in CMS1, which is known for MSI-H, having a higher immune activity, we see that there is a higher immune cell type in those groups, so meaning that our cell type deconvolution results are valid. So in summary, our analysis have identified spatially specific markers that's tied to colorectal cancer-specific mortality in Alaska Native people. Specifically, in the tumor center, we have a downregulation of MILK2, SPINK4, and the interferon alpha and gamma pathways in the epithelial compartment. And in the cell type, we see that there is a decrease of plasma cells and an increase of mast cells in the tumor center. And also for the invasive margin, we see an increase of HBA2 expression in the epithelial compartment, and the overall invasive margin, we see an enriched epithelial-mesenchymal transition pathways. And for the future direction for this research, as I mentioned, there's still 100 more patients' data being generated for geomix that will be included. And also with the other spatial proteomics and transcriptomics data we have with our sequential TMA slides, we're hoping to drive a more effective and complete prognostic biomarker sets. Thank you very much for your time. Thank you. This is now open for discussion. Thank you very much for your talk. One question is, do you need spatial transcriptomics for this kind of biomarkers? Would you be able to have a pathologist annotate the regions and use some tools to isolate and sequence- I'm sorry, can you repeat that, please? I think I couldn't hear you through the mic. Right. So my question is if you need spatial transcriptomics for this kind of biomarkers- Mm-hmm ... or whether you can use a pathologist, say, to annotate the tumor center and the invasive margin and then sequence these separately in bulk transcriptomics. Yeah. So, ultimately, what we want to do is to present a list of curated biomarkers so that it can be implemented clinically. Because obviously even geomix is a relatively old technology, relatively, compared to the single-cell technologies, but it's still really expensive, so we obviously don't want to bankrupt the patients by performing these geomix datasets for everyone. So yes, ultimately, we'll be having a curated set so that they can use a cheaper way to determine those. Yes. Thank you. Hasan Ishtar, a very nice presentation. Thank you. Questions about, you mentioned two categories, those who died versus those who not died. Mm-hmm. How about the age of the patient? How about the metastasis versus non-metastasis? Because the treatment. Yes. So that is very specific. For the age can be 90- Mm-hmm ... or less than 50, young or late, and then the treatment and the markers, tumor microenvironment, which is hot versus cold. Mm-hmm. You mentioned CMS. Mm-hmm. Especially MSI and MSS, which is now the guideline at the bedside. Mm-hmm. CMS is not really at the bedside yet. Mm-hmm. But that is interesting to know about your- Yes. So thank you very much for your question. That is actually a really good point. So, as I mentioned that our dataset, as we did have a really careful matching for the age, the sex, the tumor stage, and tumor site. Those are the things that we adjust for before getting the results. And yes, you did mention for CMS MSI-H, right now we have some, the subset of the data is being performed with bulk RNA-seq that we are able to know the CMS and the MSI-H. But after that, once we're able to get the additional 100 patients, we can then use, for example, CMS Call, which is a library that can get those results from the geomix data, then we can include them to use for adjusting and also for stratification comparisons. Great. Thanks. Well, thank you for Okay. The next speaker is David Routtenberg from Mesoscale Diagnostics, who will be talking about highly multiplexed immunoassays that enable the discovery of multi-marker EV surface signatures as novel biomarkers. Hi, thanks for coming. My name is David Routtenberg with Mesoscale, and I'm going to talk about a new tool, a new platform that we've put together to enable discovery of multi-marker extracellular vesicle surface signatures, and how we can use these as biomarkers. So I'm an employee of MSD, neither a shareholder nor an officer. Okay, so I'll give a little bit of background on EVs, but I know most of you probably know most of this info. So extracellular vesicles are lipid bilayer vesicles. They're secreted by all cells. They show up in all biofluids. They're typically secreted at higher levels from tumor tissue than healthy tissue, so that helps us if we're looking for tumor biomarkers. And they provide a snapshot of the originating cell. And so that makes them an attractive liquid biopsy target. We're interested in surface signatures because the surface markers reveal information about the cellular origin. They help determine the destination, ultimately, which cells will take these up, as well as they give us clues about the effector function of these EVs. And we're interested in multi-marker signatures because typically, single markers aren't sufficient in order to distinguish a lot of the interesting EV populations. Just as in flow cytometry, where you need two or three or more surface stains in order to define your interesting cell populations, it's no different with EVs. All right. So I'll use an example here to show why I think you should care about EV surface markers. So looking at colorectal cancer, we had a set of stage four patients, and we measured soluble CEA, sort of a classic soluble biomarker for colorectal cancer. And then we also measured that same CEA molecule on the surface of extracellular vesicles. It's actually a multi-marker signature, but it contains the CEA molecule. And what we found is that 16% of the patients were undetectable by the soluble CEA biomarker, and yet were detectable looking at that same molecule on the surface of the EVs. So clearly, there's orthogonal information in these two different measurements, despite being based on measuring the same molecule. So that kind of gives an idea of why we should care. The challenging thing now is to know which EV populations or subpopulations we should be looking at. So one way that you could do that would be to take a whole bunch of different surface markers and look at all the combinations of those surface markers and figure out which ones are of interest. That's challenging to do using typical EV assay measurements. So we developed a new platform for this. We call it combinatorial immunosequencing. So we essentially capture EVs from a sample on a solid phase. We label those EVs using a cocktail of detection antibodies, where each of these antibodies is tagged with a molecular barcode. We then have an assay where if three of these antibodies are co-localized on the same vesicle, those barcodes get stitched together into a sequenceable amplicon. And so we can then put this into an NGS. We can read out all the combinations of three surface markers that were present, along with their relative abundances in the sample. All right. So to show that this is actually reporting sensibly, we ran it on a number of cell lines. Here I'm showing four. But we found that in each case, we're able to pick out combinations of multiple surface markers, these triplets that are both highly enriched in that cell line relative to all the other cell lines in our little universe of cell lines, and also those that are highly abundant on those targets. So those would be the ones in the upper right corner. And what's interesting is we can take two colorectal carcinoma cell lines, so Diffi and HCT15, and those would be really challenging to separate just using a single marker because they do share a lot of the same surface features. But we can easily find triplet combinations that discriminate these quite well. So doing this on cell lines is relatively easy, but the kicker is can we do this in human biofluids? Can we detect these low abundance multi-marker combinations against the background of billions of irrelevant vesicles, or at least relevant to other biology vesicles in something like plasma? So to determine whether this is reasonable, we purified Diffi EVs. We spiked these into healthy human plasma. We did a limiting dilution to figure out what was the minimum number of these spiked-in EVs that we could detect among the background from the healthy plasma. And we found that we can see about one in 10 million. So that's hard to do with single markers, but the more markers you add up in conjunction, the more your nonspecific background can go away, and you can detect these really rare populations. All right. So we showed that we have sensitive-- Actually, one more thing. We have a really nice wide dynamic range, about four logs of dynamic range here, so that's helpful. So now that we have sensitive and specific assays, we wanted to see, can these do something useful in actual patient samples? So we looked at three independent cohorts of colorectal cancer samples. The first was a cross-sectional cohort from Bob Coffey's lab at Vanderbilt, and then two longitudinal cohorts from Andy Nixon's lab at Duke.And despite being quite different populations, and different treatment, we found a lot of similar features, similar three-marker combinations in these three cohorts. And the interesting thing was that these combinations really reflect two distinct tumor compartments. So we have combinations that reflect the epithelial cells, so we can see there are combinations of markers like CEA, E-cadherin, EpCAM, and then we also see stromal combinations. And then there's a few things like CD73 that we know show up on both the stromal population and the epithelial population. So that's where if you just look at the single marker CD73, it's really hard to put this into one of those two compartments. You need the multimarker. All right. So to convince ourselves that what we're seeing here is not just an artifact of the assay, because we're a little worried that if you jam a bunch of vesicles together on a surface, you might get cross-mixing of these signatures, so you might have single positive EVs next to each other, and it look like a double positive. So we used an alternative assay technology, this is based on MST's ECL planar array platform. And we measured the same samples, and we showed really high correlation, so a Pearson correlation of over 0.9. And so this was great. It kind of validated that the markers we're seeing are real, but it also gives us a pipeline so we can do our discovery of multimarker signatures using this high-multiplex combinatorial immunosequencing, and then once we find the interesting populations, we can port those over onto the ECL, the targeted assays for... These are lower plex, but we can run higher throughput and use this for validation. So the next question in our mind was, can we get prognostic information from some of these surface signatures? So we know they were elevated in cancers relative to normals, but that's not that useful. So we wanted to see in these longitudinal cohorts, are they actually predictive of overall survival or progression-free survival? So to sort of limit the scope, we took the top 100 signatures that were overexpressed in the tumors, and we performed Cox regressions on these. And we found there were quite a few after our correction for multiple comparisons, there were quite a few that were still FDR significant. And I guess one thing that's interesting, we're getting a number of consistent three-marker findings between the two different cohorts, but we're also identifying some unique three-marker combinations that only show up in one of these cohorts or that only show up in one context, either in progression-free or overall survival. So this gives us hope that we can make something useful out of this technology. And I guess I want to highlight a few key findings. We found that CD73, CD13, and CEA are kind of recurrent components in these CRC-secreted vesicles. All right. So to kind of sum it up, we built a technology that allows us to profile three-marker EV surface identities. We have currently, this says 63, but at this point, we're well over 100 surface markers, and we can do this by using the three-marker combinations. We can simultaneously assay tens of thousands of subpopulations in a single sample. We've cross-validated the technology by using an orthogonal platform to look at some of these populations that we found. And we've consistently found the same markers across three different CRC cohorts. So obviously, this isn't validation, but it gives us a signal that's at least reproducible in some independent sample sets. So interestingly, we can probe two different tumor compartments. We've seen there's prognostic value of some of these signatures. So what comes next? Well, I had been discouraged about using EVs for early detection simply because there are so many EVs in plasma that are not the ones you're looking for. So it's really hard. You're looking for essentially a few EVs in maybe trillions. But seeing the specificity we have with the multimarker gives me hope that we can go back to early detection. So we've started working with a few different cancer types, not just colorectal cancer. We're also interested in expanding our library of antibodies to look at the broader tumor microenvironment to kind of decode the composition of that microenvironment based on the circulating EVs. And then lastly, we want to try to track the tumor progression by looking at the evolution of these EV surface signatures over time. So overall, we've put together a multimarker profiling technique that we think opens new dimensions for analyzing liquid biopsies. It helps to capture the tumor ecosystem complexity in a way that single analyses can't resolve. So I want to thank my group and our collaborators, the Coffey Lab, the Nixon Lab, and the Cody Lab at Washington University, as well as our funding. A lot of this came from the Common Fund, the Extracellular RNA Communication Consortium, which supported our work for about five years. So thanks for listening. Take questions. Okay. Do we have any questions? I'll jump in right-- And so in terms of validation, what are your plans? Are you going to... Honestly, I'm a technology development guy. It's what gets me excited more so than the actual biology. But Lucy, who's in the audience there, she's a cancer biologist, and so she's going to help with that. Okay. Yeah. I'm not entirely sure. As well as our clinical collaborators will sort of help guide that. Very cool. All right. Thank you. All right. Thank you very much. All right. Our next speaker, Dr. Soto Trujillo. I hope I'm saying that correctly. "Differences in Survival Outcomes Among Metastatic Colorectal Cancer Patients Across Fibroblast Growth Factor Receptor Two and Three Expression Levels." Thank you, Dr. Gulveru, for welcoming me onto stage. My name is Steve Soto, and it is an honor for me to speak before all of you today. I'm a post-bac research trainee at the USC Norris Center, and today I will be discussing survival outcomes in metastatic colorectal cancer patients across FGFR2 and 3 expression levels. This is my disclosure. So colorectal cancer is the second most common cause of cancer deaths in men and women combined in the US. In 2026, it is estimated that over 150,000 people will be diagnosed with colorectal cancer within the US. The rise of incidents, especially in young adults, has only accelerated the demand for more clinical biomarkers. So FGFR2 and 3 belong to the FGF receptor family, which contain a total of four distinct receptors recognized in the literature that bind to a diverse set of 22 FGF ligands. FGFR2 and 3 are receptor tyrosine kinases that dimerize and autophosphorylate upon ligand binding, activating a series of metabolic cascades. Many of these downstream signaling pathways, similar to those activated by other receptors as well, such as the epidermal growth factor receptor and the vasoendothelial growth factor receptor. So FGFR2 is mostly discussed in breast cancer literature, with some germline variants being associated with a predisposition for breast cancer development. FGFR2 has been found to be expressed in the endothelium and has most recently been suggested in allowing dormant breast cancer tumor cells to survive following targeted therapy. The role of FGFR2 in colorectal cancer pathogenesis remains a bit unclear, although some papers do suggest it being linked to a more aggressive type of cancer and consequently worse survival outcomes. FGFR3 is mostly discussed in bladder cancer, with some specific mutations being associated with improved survival. While the underlying mechanisms are a bit unknown, some papers do point towards the possible modulation of the tumor microenvironment through the infiltration of T cells. FGFR3 gene alterations have been found to be increased in a small CRC cohort, although this failed to be statistically significant. So based on the literature for FGFR2 and FGFR3, our team hypothesized that in colorectal cancer patients, FGFR2 gene expression will be associated with worse survival due to a possible contributing factor towards angiogenesis, and FGFR3 gene expression will be associated with improved survival due to this possible modulation of the tumor microenvironment. So the database in which we conducted our analysis in is the CALGB SWOG 80405 Phase III clinical trial, which contained a cohort of newly diagnosed metastatic colorectal cancer patients that were randomly assigned to a chemotherapy plus cetuximab treatment arm or a chemotherapy plus bevacizumab treatment arm. So cetuximab is a monoclonal antibody for the epidermal growth factor receptor, which is responsible for cell survival, and bevacizumab is a monoclonal antibody for the vasoendothelial growth factor, which is strongly linked to angiogenesis. Other covariates that we adjusted for in our additional statistical analysis include the KRAS mutation, BRAF mutation, and MSI status, sidedness, age, sex, chemotherapy as well. So we initially tested for the prognostic potential of FGFR2 by grouping our two treatment arms together. So here on the left side, we have a progression-free survival curve where FGFR2 high expression was somewhat associated with worse survival. However, this failed to be statistically significant. However, FGFR2 did demonstrate to be prognostic of overall survival across our treatment arms, with FGFR2 high expression being associated with worse survival, with the median survival difference between FGFR2 high and low 7.4 months and the hazard ratio of high to low 1.28, meaning that our FGFR2 high expression group had a 28% increased chance of mortality at any time. So to determine if these effects observed in our prognostic curves were more pronounced in a specific treatment arm, we looked at each treatment arm individually. So here on the left side, we found that FGFR2 was not predictive of overall survival in patients being treated with bevacizumab, which was striking because bevacizumab being an anti-angiogenic drug and FGFR2 being strongly linked towards angiogenesis, we expected FGFR2 gene expression to have a greater predictability in clinical response. However, FGFR2 gene expression was predictive of overall survival in patients being treated with cetuximab, with FGFR2 high expression being associated with worse survival and the median survival difference between high and low being 12 months and the hazard ratio of high to low 1.56, meaning that our FGFR2 high expression group had a 56% increased chance of mortality at any time. So we then conducted a similar analysis for the FGFR3 gene expression, but did not find it to be prognostic of overall nor progression-free survival across our treatment arms. However, we did find that FGFR3 gene expression was predictive of overall survival in patients being treated with cetuximab. However, in this case, FGFR3 high expression was associated with improved survival, with the median survival difference of high to low being 12.4 months and the hazard ratio of high to low 0.59, meaning that there was a 41% reduction in mortality in our FGFR3 high expression group. We then tested for differences in FGFR2 and FGFR3 gene expression across ethnicity and race and found no statistically significant differenceNor did we find any difference in early onset versus late onset colorectal cancer patients. So we then moved on to SNPs. SNPs are single nucleotide polymorphisms, which are genetic variants that differ from each other by a single nucleotide and whose functional impact varies depending on the site of the mutation. So SNPs can manifest themselves as either a heterozygous variant if a mutant allele is inherited from one parent, or a homozygous variant if two mutant alleles are inherited from both parents. So here we sourced 12 distinct SNPs recognized in the literature for FGFR2 and FGFR3 and tested their associations with survival. So here we have a visualization of one of those analyses. Here we tested for the prognostic potential of the SNP in FGFR2, RS-3135718, and found that the variant allele groups, the heterozygous variant and the homozygous variant on the left curve, do seem to be associated with improved survival. However, this failed to be significant in the overall survival curve, although there does seem to be a similar pattern. We then looked at each treatment arm individually and found that this SNP was not predictive of survival in patients being treated with bevacizumab. However, we did find that the SNP was predictive of progression-free and overall survival in patients being treated with cetuximab. So here in the overall survival curve, we see that the heterozygous variant group, the blue curve, is associated with improved survival, and the homozygous variant, the red curve, is associated with the highest improvement in survival relative to the heterozygous and the wild type group, with the median survival difference between the homozygous variant and the wild type being three months. So while the difference in survival is not grand, it does suggest that this SNP may have some sort of functional role in predicting clinical response in patients being treated with cetuximab. I think it is important to note that here in the overall survival curve, from zero to 29 months, we don't see a clear distinction in survival between the three groups. It is only after 30 months and forward where we see a clear distinction between the three groups, suggesting that the SNP may be predictive only after a specific clinically relevant time point. So we then tested the SNP's association with its respective FGFR2 gene expression, but did not find any differences in gene expression between the three groups, the wild type, homozygous variant, or heterozygous variant. Nor did we find any differences in gene expression in other SNPs that we tested with their respective gene expression. So overall, in the 80405 Phase III clinical trial, we found that FGFR2 low expression, the SNP in FGFR2 RS-3135718, and the FGFR3 high expression were all strongly predictive of survival in patients being treated with cetuximab, with them all being strongly associated with improved survival. So given the biomarker potential promised by our predictive and prognostic curves, we have transitioned to a pre-clinical platform in an effort to elucidate the underlying mechanisms surrounding FGFR2 and FGFR3 signaling. So here on the right side, we inhibited FGFR2 in various CRC cell lines with unique molecular characteristics and successfully decreased cell viability, reiterating the criticality of FGFR2 in maintaining cell survival in these human-derived CRC cell lines. So thank you so much to my PI, Dr. Lenz, and everyone else on the part of the team that made this project and presentation possible. Thank you to patients and caregivers for their invaluable contribution to this Phase III clinical trial. Thank you. Excellent. It's now open for discussion. Thank you. Very nice talk. I was wondering, you tested also SNPs in FGFR3. Were any of those SNPs associated with outcome? We did not find any of the SNPs that I-- Those 12 SNPs that I mentioned, those of FGFR3, we did not find any to be statistically significant with their- All right. Thank you ... respective gene express-- or survival. Well, a quick question. So, given that FGFR2 and R3 are from the same family, why do you think that they behave differently? So I think that's what makes the findings from this analysis very exciting. The fact that they're part of the FGF, they're both RTKs or receptor tyrosine kinase, but they could be associated with different survival outcomes, makes it even more intriguing, and is what we are basically trying to work on next in these other clinical experiments that we are planning for soon. Great. Thank you very much. See, I think it's me. Is it me or you? It's you. Okay. Okay, so the next speaker is Lubna Ahmad, who is here from the University of Cambridge, and she will be talking about the mapping of spatial organization of B cells and immune signaling in triple negative breast cancer to characterize response to immunotherapy. Good afternoon, everybody. My name is Lubna, and today I'm... So first of all, I'm a postdoc in the Ali Lab at the Cancer Research UK Cambridge Institute, and I'm talking a little bit about my PhD project, which looked at B cells and immune signaling in an immunotherapy trial to characterize response to treatment. So, to set the stage, so the main kind of challenge we have in immunotherapy is how can we better predict which patients are going to respond to this treatment? Essentially, in solid tumors, what we do, we assess PD-L1 expression, and that is usually done by scoring its staining across tumor-infiltrating immune cells. However, emerging clinical evidence has suggested that this does not really capture the complexity of responses and that additional features of the TME might be important to take into account. So in the recent years, B cells have become increasingly associated with immunotherapy response, such as in studies of melanoma, renal cell carcinoma, and sarcoma.Despite the growing interest in B cells, B cells in cancer are generally relatively poorly characterized. For example, in terms of their heterogeneity, very little is known about the composition, localization, maturation, and molecular characteristics. So although we know that B cells seem to matter, we know very little of how, when, or why they contribute to responses in cancer. So to address some of these questions, I use an approach called imaging mass cytometry. So here, I made a targeted panel that looked at B-cell protein markers, cytokine RNA markers, and other contextual protein markers. I used FFPE tissue where I used double Z probes to target the RNA of these cytokines. And I amplified the signal using a branch DNA structure, which was modified to be labeled with metal-conjugated labeling probes in order to detect the signal using IMC. After this, the tissue is then stained with a cocktail of metal-conjugated antibodies. I then identify regions of interest which are ablated at one micron resolution using a laser, and the resulting ion cloud is then analyzed using time-of-flight mass spectrometry. Essentially, a technique such as this allows you to simultaneously detect RNA and protein in tissue. And, in this case, I was able to detect up to 43 RNA and protein. And what you are able to do is you preserve spatial information and that being at single-cell resolution. So in the lab, we have already used this type of technique called IMC, in the setting of an immunotherapy clinical trial called NEOCHIP, which is for early triple-negative breast cancer patients using a broad lymphoid and tumor panel. Here, we identified several key spatial predictors of immunotherapy response, and this is a study that has already been published. So building on this work, I now focus on this NEOCHIP study, but with a focus on B cells and cytokines. So to tell you a little bit more about NEOCHIP, here we have about 280 early-stage triple-negative breast cancer patients, and they are randomized to the chemoimmunotherapy arm, as indicated in red. And then we have the chemotherapy arm, as indicated in gray. So we have samples collected at baseline which are treatment naive, so these are biopsies. On-treatment samples, which is after one cycle of treatment, and post-treatment tissue microarrays. I've applied my targeted panel to all of these samples in order to more comprehensively target B cells and cytokine signaling in the broader TME. So from this, I have resulting pixel-level data which I then segmented into single cells. I compartmentalized these cells into epithelium TME and identified, in total, 22 different TME phenotypes and 14 epithelial phenotypes. So the first question I asked myself was whether any of the TME densities associated with treatment responses. So to begin off with, this is an odds ratio plot at the baseline time point. So any values above one outlined in black are significantly associated with treatment response, where the color of the dot indicates the treatment arm, so gray, chemotherapy, and red, immunotherapy. As we can see at this time point, none of these TME densities are actually associated with immunotherapy response. But then when we go to the on-treatment time point, we see a clear B-cell signature emerging where several B cells, such as germinal center B, plasma memory, and naive B cells, are associated with immunotherapy response, some of which are also differentially associated. Apart from looking at TME densities, I was curious to see whether B and plasma cells exist within niches in triple-negative breast cancer. To find these niches, I used a niche detection method called GraphST and built on this using a sequential hub-based approach to learn about the neighboring cells. To summarize these niches here, I'm showing you that we found germinal center niches, B activated niches, plasma-rich niches, and plasma-maturing niches. On the right, I'm showing you via this dot plot the associations with response on baseline and on treatment. Here, what we see is for this B activated niche, this niche stays significantly associated with immunotherapy response across the neighborhood scales and trends starting to emerge for the germinal center and the plasma-maturing niches, too. Apart from detecting niches, I was wondering whether the niches and their associations or their interactions with the surrounding cells were informative of immunotherapy response. So first, I defined cell-to-cell interactions, which was the direct contact of the segmented cell masks. So at baseline here, I'm showing you a dot plot where the color shows you the odds ratio, and the asterisks shows you the significant associations with response. Here, clearly you can see that all of these associations lie within the immunotherapy arm. For example, the plasma-rich niche interacting with CD15, MHC-1 and 2 epithelial cells being associated with immunotherapy response. If we compare baseline to on treatment, we see a much richer set of interactions being associated with immunotherapy response now. The difference here is now that B cells from the B activated niche interacting with numerous CD4, CD8, plasma, and B cells, as I summarize here, being associated with immunotherapy response. So taken together, these results from niches, their interactions suggest what drives response to immunotherapy is not just the abundance of B cells but also measures of their spatial organization. Apart from this, what is very unique about NEOCHIP is that we have collected samples at different time points, which is what makes it such a nice resource. So I was interested in learning about the cellular composition and how that changes over the course of treatment. In this odds ratio plot, I will show you the enrichment on treatment if I compare them from baseline. So any odds ratios above zero outlined in black and with this darker yellow circle show you significant associations on treatment of cell phenotype abundances. So within the chemoimmunotherapy arm, when we rank them, we see that the most abundantPhenotype is actually the naive and B cells, which is also enriched in the chemotherapy arm, but it's not the most abundant one. Similarly, if we do this kind of comparison from on-treatment to post-treatment, you can clearly see all of the immune cells are now depleted, and the ones that are enriched all belong to the stromal compartment. To get to the final part of my presentation, we now focus on cytokine enrichment, and I wanted to study this in terms of defining cells being cytokine positive or negative for each of the cytokine on my panel. What I was curious about is learning whether these cytokine positive cells were enriched within a vicinity of a niche, so within a given radius compared to its surrounding. So this is what this very complex heat map shows, where we have the cytokine positive cells on the Y-axis and all the niches on the X-axis. As you can quite clearly see, numerous cytokines are enriched in niches across many of the different groups I have here. But the main feature I want to highlight for now is within the chemo immunotherapy arm, in the responding patients, this seems to be a more persistent feature, where this kind of cytokine enrichment persists post-treatment, which is not the case in the other groups. So to summarize my presentation today, what I showed you is that B and plasma cells are differentially associated with immunotherapy response early on treatment, that they exist within niches and that this is a feature of the TNBC TME. That specifically the B cell activation signature remains significant across spatial scales, and its interactions are also having predictive capacity for immunotherapy response. That the cellular dynamics suggest that out of all TME cells, B cells are more significant post immunotherapy, that many cytokines are enriched within niches and also seem to be a persistent feature in immunotherapy responders across time points. So this brings me to the acknowledgments. I would like to take a brief second to thank my lab, the Ali Lab, my supervisor, Razali, our collaborators for the NEOTRIP trial, of course, all patients and families, without whom work like this would never be possible, the funding bodies that make this work possible, and all of you guys for your attention. Thank you very much. Okay. The floor is open for questions. Hi. Really fascinating work. I wonder how you selected your regions, because I think the B cell effects, they seem very believable, but for more subtle effects with cell interactions, you can be impacted by how the regions were chosen. Mm-hmm. Could you say something about that? So essentially what we looked at, so we had adjacent H&E sections to these slides. From these H&E sections, we identified regions of immune aggregation, and when these immune aggregation regions were not present, it would be immune infiltration, but always with the focus on tumor regions. When tumor regions would not be-- When we couldn't identify them, for example, on treatment when the tumor would be gone, this would be in the tumor bed instead. So these were the kind of definitions for sampling. Thank you. I have a question. Were you surprised to think of plasma cells and B cells as an immune reactive and being responsible for this, given that people usually focus on T cells or other cells? B cells or at least plasma cells are not the most critical usually in a tumor microenvironment. I'm a big believer in them. I work in myeloma, but just wanted to know what you thought of it. I think the effect is actually quite striking, and it is quite clear in the data set, also because T cells are the most abundant lymphoid cell in TME, so I was quite surprised by the finding, especially for plasma cells. So yeah, I would agree. Yeah. So I'm going to echo that. So for the long time, our lab studies antibodies and their relationship to responses in cancer, and we've been pushing to try to study that in this context. And recently heard Chuck Perou talk about- Mm-hmm ... one of the best predictors in outcome, in breast cancer was the B cells. Mm-hmm. B cell density. Yeah. This is going to emerge. People have always focused on the T cells. Have you looked at antibodies at all in what these plasma cells are producing? So, there was, of course, interest in looking at antibodies such as IgG or IgA, but at times it is quite difficult to find reliable antibodies for these that are able to work with the assays, especially with the different assay I had. But I do think there's a lot of interesting things spatially that you can learn from targeting IgG or IgA in the TME. Even in the plasma. Yeah. Have you looked at the plasma of those same patients? Uh. With secret antibodies. Yeah. We didn't have information on that data, but yeah, that would be- We talk about using protein microarrays in that setting. Yeah. The other thing to think about would be doing B cell receptor sequencing. Yes, exactly. Yeah. All right. Thank you very much. Thank you. Yeah. All right. Our next speaker is Andreas Mand, and he's talking about AI-powered deep visual proteomics, DVP, links pancreatic precursor biology to PDAC survival phenotypes. Thank you. And yeah, it's a great pleasure to present today an exciting collaboration between Dr. Aniban Mitra from NYU and also from Dr. Michael Hollingsworth from the University of Nebraska, where we used the power of deep visual proteomics to study pancreatic precursor biology. So our mission is to develop and apply mass spectrometry-based spatial proteomics for translational research. We typically utilize formalin-fixated, paraffin-embedded tissue material, these classical surgical pathology specimens, which are millions of them stored in biobanks.And we do so to basically do deep molecular phenotyping of humans to better understand disease initiation, progression, and outcomes, and also to find new and better biomarker patterns to stratify patients and treatment choice. So what is deep visual proteomics? Conceptually, it's mass spectrometry-based spatial proteomics. Deep stands for deep learning and visual for high-resolution, high content microscopy, and proteomics for mass spectrometry-based proteomics. As I just mentioned, we typically take formalin-fixated, paraffin-embedded tissue material. We cut very thin sections, like three to five micrometer thin sections, put them on glass slides, and then stain them with either H&E, immunohistochemistry or immunofluorescence. And then we use the power of computational foundation models to identify interesting cellular phenotypes. In the next step, we then transfer these GPS coordinates from these cells or regions of interest to a laser microdissection microscope, where we, over the years, developed a workflow that automatically isolates single cells or regions of interest. And then we collect the different classes of phenotypes into, for example, a 384-well plate, where we then can collect, for example, class one into well A1, class two in well two, and so on. In the next step, we can then digest the proteins into peptides and use the latest and greatest in mass spectrometry to identify thousands of proteins, which we then can merge together with the images we created, because we preserve throughout the workflow these spatial coordinates. And thereby we create this digital map of the diseased, or these proteome maps of the diseased tissue. So OmicVision platform is based on an emerging and increasingly validated platform. Our initial study was featured in "Nature Biotechnology" in 2022. Then two years later, "Nature Methods" named spatial proteomics the method of the year, and we were very happy to see that our technology was featured there as well. And since then, we have validated and applied this technology in a number of use cases. I want to highlight one example where we identified a drug target for a deadly skin disease called toxic epidermis necrolysis. It's a rare but very deadly skin disease, and here we found a drug target, and through off-label use, we could then went from the discovery all the way into pre-clinical models, down to the patients. And so far, eight out of eight patients were treated with this off-label used drug and cured. So this finding was basically then featured from Eric Topol, a medical AI key opinion leader, and called "The Dawn of Spatial Medicine." So as I said, we typically work with a formalin-fixated, paraffin-embedded tissue material. So we typically stain them with H&E, and then we create these gigapixel large images, which then get divided into thousands or tens of thousands of patches, where each of these patch encode some feature vectors, which then using computation AI, and then we can basically classify AI-based phenotypic clusters and then can analyze them using mass spectrometry-based proteomics. And that's exactly what you can see here. Once we isolate these regions of interest, we can then perform differential expression analysis, for example, between healthy or tumor cells and look for the outliers in order to describe the biology and map them back to the spatial location in the tissue. And we then teamed up with Dr. Anima Mehta from the NYU and Dr. Hollingsworth, and we had the questions, since the fact that 60% of the general population have a prevalence of these incidental precursor lesions we set out to study, and only a subset of those patients develop PDAC, we were interested in understanding why this is the case. And for that reason, we then defined two cohorts. One is the I cohort with the incidental normal ducts and ADMs and also low-grade PANAs. And then we had another cohort, which is the cancer-associated cohort, where we isolated cancer-associated normal ducts, low-grade and very rare high-grade PANAs, and the tumor cells. So in the next step, we then used AI-guided location of precursor lesions and normal ducts. So we used a computational approach to find the optimal number of clusters to identify, which you can see here on the right side, these normal ducts and PANA lesions, and then sent these XY coordinates to the laser microdissection microscope, which then isolated these less than 100 phenotype-matched epithelial cells and analyzed them using mass spectrometry-based proteomics. And here we could, in total, quantify more than 9,000 proteins across the whole cohort, and per mass spec sample, an average of 5,800 proteins and many, many more peptides.So when we take a bird's-eye view on these proteomic changes, we can see that we have massive changes in the proteome, looking at the tumor cells, which is not surprising. But what we can do then is to look what are the drivers between the different disease states. And then, of course, in the case of cancer, we could see that there's a lot of these hallmarks of cancer which drive the separation. But very interestingly, when we looked a little bit closer about the pannus lesion, and particularly the normal ducts, we could already see a separation between the cancer-associated normal ducts and the incidental normal ducts, which is, of course, very interesting. So in the next step, we looked a little bit closer what are actually the molecular changes between these two ducts. And here we did some volcano analysis, and as you can see on the right side, you see significant enriched proteins in the cancer-associated ducts, which are reflecting oxidative stress, metabolic reprogramming, and also immune editing. We then did also pathway enrichment analysis and could show that inflammation, antigen presentation rise early, so as early as in the normal ducts, all the way to low-grade pannus and then sharply drop when they transit to high-grade pannus, suggesting that these cells already undergo immune evasion. So this we then explained like a cancer field effect, where the normal ducts are close proximity to tumor cells, basically undergo or get these immune signatures, which then drive these kind of stress response, whereas the normal ducts not. And then, what comes on top here with the mass spectrometry-based proteomics analysis is that we can not only detect differentially expressed proteins, but we can also detect KRAS mutations without sequencing the DNA. In order to do so, we generate a custom FASTA file and use deep learning-based prediction of peptide properties to generate theoretical spectral libraries. And this allowed us then to identify the different KRAS hotspot mutations. So what this enables us is to combine genotype with phenotype. And just to give you an idea, what we found is that when we look at the ADMs in low-grade, we can see in a number of those hotspot mutations, when we look on the patient level, we can see that some patients have one mutation, others have several different mutations in the different regions of interest. So we can further expand this analysis also to study tumor heterogeneity using, again, the computational foundation models to look at different morphologies. And then we can start to ask what are the different drivers of those different morphologies, and also could see that certain areas are more aggressive than others. So this represents a huge now resource to look at tissue heterogeneity or tumor heterogeneity. And for example, we could see that these two spatially adjacent regions differ in cytoskeletal, metabolic, and RNA processing. Yeah, to make a long story short, I'm very happy to announce that this study is now online at Cancer Discovery, just published today. So if you're interested to know more about what we did in particular, so please use the QR code. And, yeah, with this, I would like to end and acknowledge all the people involved. Most prominently, Dr. Aniban Maitra from the NYU, and also my team at Omic Vision Biosciences, particularly Lisa Schweitzer and Jimin from Dr. Maitra's lab, and also Paul and Michael from the University of Nebraska. Also, our industry partner from Leica Microsystems, and most importantly, the donor patients and their families for their invaluable contribution to scientific research. Last but not least, I would like to announce that we are in the process of rebranding from Omic Vision towards Resolute Bio and have some exciting news soon to share. And if you are interested in partnering with us, please find me. All right. It's open for discussion. Nice work. Thanks. I'm wondering about the diameter of the laser capture size to quantitate 5,000 protein groups in your study, and- Yeah ... have you compared frozen versus FFP tissues with regards to the sensitivity and performance? Yes. So to the first point, in this case, we took less than 100 phenotype-matched cells, which corresponds approximately two to three nanogram of input material. So very little amount. So what we do is that we try to identify computational similar cells from spatially localized regions and then collect them to get statistical power behind it and also the proteomics depth. So it's less of 100 cells.And then the second part is, yes, we have also in a separate study, compared fresh frozen with FFPE. And the good thing is that we only lose marginal proteomic depth. And the reason is that formalin-fixed and paraffin-embedded tissue, we might lose some of the peptides, but we still have enough peptides to then quantify and identify the proteins behind it. Because we have many, many more peptides per protein than we actually need. So loss is below 5% between FFPE and fresh frozen samples. And in fact, we have actually looked in samples that are stored 30 years or longer, and we still get a very robust proteomic signature. Do you see this becoming a diagnostic test that we can use in the clinic? And is there a way for us to see that in a blood-based rather than tissue-based proteomics? Yeah. So a good point, particularly for pancreatic cancer, you don't want to do an ongoing tissue biopsy. But of course, the beauty is here that you can develop, for example, a targeted mass spectrometry-based assay where once you have identified interesting peptides, you can then create marker peptides, which you then can try to refind in the blood, and then even do absolute quantification. If you do some spike-in peptides where you know the exact amount of the peptide of interest, you can then calculate the ratio, and thereby knowing what is the exact amount of any given peptide in your sample. All right. Thank you so much. Okay. Our next speaker is Zach Kinsella from the Royal College of Surgeons in Ireland. He's speaking about spatial immune multiomics refines prognostication in early-stage estrogen receptor-positive breast cancer. Hi, everyone. Can you hear me? Yeah. Thank you for inviting me to talk. I'm here today to just give a background on some of my research, which is used in spatial multiomics to refine prognostication in early breast cancer. I have no financial relationships to disclose. A bit of background. ER positive breast cancer is the most common breast cancer subtype, and in comparison to other breast cancer subtypes, late recurrence is particularly common, where up to a third of patients will experience a recurrence. This really raises the question of how do we know who's going to recur? And it's in part answered by clinical pathological variables, grade, stage, size, et cetera, and genomic variables such as multigene assays. The most widely used and extensively validated is the Oncotype DX Recurrence Score. It may need a little introduction, but essentially it's a test that looks at the expression of 21 genes and algorithmically computes a risk score between zero and 100. It was prospectively validated in one of the largest adjuvant trials in breast cancer to date, and essentially is really quite useful at determining who does and doesn't need chemotherapy. For example, patients with low risk between a recurrence score zero and 10, they're known not to benefit from chemotherapy, whereas those with high risk between 26 and 100 are known to benefit from chemotherapy. The problem is that this intermediate group between 11 and 25, the current benefit of chemotherapy for these patients is relatively ambiguous. So what this reveals is that a test is only really accurate in the extremes, is inaccurate in the intermediate. And that hides a bit of a deeper issue in the sense that the majority of patients, up to 60% of them, actually exist within this intermediate group. To begin to ameliorate this, we have Irish patients that were previously enrolled in TAILORx, and we basically applied our spatial multiomic workflow, which included IHC on full face and on TMAs, we applied spatial transcriptomics using NanoString's GeoMx, their whole transcriptome atlas, spatially segregated from tumor epithelia to the peritumoral stroma, and also spatial proteomics using Akoya's 6-plex immunofluorescence technology. So we essentially wanted to take this together, analyze it, and really try and separate out that intermediate recurrence score using spatial omics data. And the other TAILORx is particularly good for this because the intermediate group we have one to one randomized to receive either hormone therapy alone or chemo endocrine therapy as the major TAILORx trial was as well. Going into our data, we wanted to basically see initially what's going on in the immune microenvironment. So we performed unsupervised clustering with our proteomic data, finding there to be four TME types and three epitypes. These TME types range from T cell low, T cell infiltrating, actually immune hot, and actually immune cold, whereas the epithelial types really only range from cold, intermediate, to relatively warm. What was interesting is that these immune types were not significantly captured by the Oncotype DX Recurrence Score, nor were significantly coupled to one another, potentially showing that there's some orthogonal immune biology going on there to genomic risk, and that the immune information may provide avenues to begin to improve that genomic risk categorization. This is where we started to bring in our transcriptomic data to look at what was actually going on and the differences of what's different between these different TME and epithelial states. What we found among other immune related transcripts was significant upregulation of things like CTLA4 and other checkpoint sites like TIGIT and CD96. Whereas inflamed epithelia were really uniquely characterized by downregulation of ESR1, which encodes ER alpha. This potentially shows maybe a dedifferentiation from luminal-like biology, and we also saw, for example, in gene set enrichment analysis, again, amongst other pathways, downregulation and suppression of things like DNA repair. So from here we wanted to know, are any of these individual immune states in the TME or epithelia prognostic?Unfortunately, not many of them were. Only one was really prognostic, and that was across the randomized intermediate recurrence score. After adjusting for clinical variables, what this shows is that patients receiving chemotherapy on the randomized chemo arm had a poorer outcome, whereas those on endocrine therapy alone actually trended towards improved outcomes. The takeaway here being inflamed TME and epithelia in our cohort, they appear relatively dysfunctional. They significantly upregulate checkpoints. They downregulate the estrogen receptor and suppress pathways such as DNA repair. After that, we wanted to know, okay, there's a lot going on in these environments. Can we decide what's going on and what genes are being expressed and are functionally coupled to infiltrating phenotype densities? So we performed an analysis here, firstly in the epithelia, showing significant upregulation of antigen presentation genes, potentially showing that the epithelia could be engaged by and recognized with the immune response. But also, along with cytotoxic genes like your granzymes and perforin, there's upregulation of CTLA4 and CXCL13, again indicative of a dysfunctional response. We had the same in the stroma for macrophage density, finding increased extracellular matrix remodeling and M2-like macrophage signatures through SPP1, which encodes osteopontin, and trends to downregulation in cytotoxicity. Similarly, for stromal cytotoxic T cell density, we saw significant increases in granzymes and perforin, but also a TCF7, which is a progenitor exhaustion-like signature and non-canonical checkpoints like VSIG which encodes an emergent checkpoint vista. So we are seeing across the transcriptome that there are dysfunctional immune responses, expression of checkpoints, exhaustion, M2 macrophage are engaged, and there's a buildup of extracellular matrix. Can we bring in proteomic data to see something quite similar? So we asked the question, what's the relationship between ITLs or epithelial TILs to STILs, TME TILs as STIL density increases? So what these plots show, one is proportional coupling, where a one unit increase in STIL basically translates to a one unit increase in ITIL. We saw that in those inflamed TME, TME3, that Tregs are much more fluid. As they basically infiltrate into the stroma, they're much more likely to also infiltrate into the epithelia, whereas cytotoxic T cells are relatively decoupled. And we're actually fairly similar to the T cell low environment, despite it being really inflamed. Basically showing that even in the proteomic data, there is, or it suggests there's a level of immune exclusion. So from here, we've seen that the inflamed epithelia are prognostic in the intermediate group and there's dysfunctional responses. It's quite difficult to translate that, though. So are there any individual immune variables that can basically recapture and recapitulate that prognostic signal? The answer is yes. Amongst others tested, this is just as an example. It seems that the high density of T cells, regardless of whether they were scored in the epithelia or TME, conferred significantly poorer outcome of patients randomized to receive chemotherapy. Just to take CD8 density here as an example, and using an arbitrary median cutoff, you can see that there is a significantly poorer outcome of high-density patients receiving chemo, and that the addition of CD8 adds prognostic information significantly to clinical variables. Another way of representing this is an inverse analysis, where you can see patients with low density don't have a significant difference in outcome over 15 years, whereas those with high density, you can see clearly that there is a difference of those receiving chemo. We also performed a formal interaction analysis, again, showing that as CD8 density increases continually, there is a significantly poorer outcome on chemo and basically trying to emulate what the changes in treatment will be like. By using our biomarker, you can see that there is a marked change in treatment, both across pre- and post-menopausal patients from what is known as the current guideline. Next, pathologists are really great. They're regularly known for being the ground truth, but there could be a selection bias in the areas that they used to determine or pick TMAs for sampling. So we performed an orthogonal validation on whole slide images. Again, we're using a similar pipeline of masking the epithelia, doing nearest neighborhood density analysis. And again, we found in the whole slide images that there was a significant formal interaction where, again, increasing CD8 density conferred poorer outcome on chemotherapy. Whenever we looked at absolute risk differences, this was only really significant above the 70th to 80th decile. But we really saw here that, for example, there was about a 16% 15-years survival difference at the 90th percentile of density favoring hormone therapy, which in restricted median survival time corresponded to around a 13-month difference favoring hormone therapy over chemotherapy. So to conclude, ER-positive breast cancer is not truly immune cold. People think it is. It's not. There are unique immune states that exist that differ in their density and composition and range from cold to hot. The inflamed states do appear transcriptomically wired to more dysfunctions, such as the expression of checkpoints, exhaustion transcripts, M2-like macrophage, and remodeling genes, as well as dedifferentiation, potentially from luminal biology. But more crucially is that the inflamed milieu, particularly those with high stromal CD8, at least in alcohol, is a predictive marker of chemo inferiority. It does add significant prognostic information to clinical variables, including menopausal status, which is one of the current guidelines, but one thing is to take this with a pinch of salt. It does need external validation in larger TRX before anything can be really generalized to practice. I want to say thank you for listening, thank you for inviting me to talk, and here's everyone who worked on this with me. Okay, the floor is open for questions. I can start with a question. Very intriguing data, and especially that it's something that you didn't expect to see. I wonder again, how are we going to see that in the clinic? How do you translate that? Would you consider now checkpoint inhibitors in these patients? I like that you looked at immunohistochemistry as a simple, easy way for us to translate that into the clinic. So can you tell us where to go from here? Yeah. I think one thing I'm currently looking at is whether you can use foundational models to score TILs. I don't know yet. I'm still doing the work. As for CD8, let's say it doesn't, and you need CD8. I feel like it could be translated. We use diagnostics all the time, like ER, PR, HER2, Ki67. They come with their own challenges. We're still arguing about the right threshold for Ki67, but I feel like as things progress, foundational models progress, you get more data, more samples, you can kind of find ways to generalize. But yeah, it's still a way off. Well, thank you very much for the talk. Very interesting. I was wondering how your biomarker would correlate in the triple negative breast cancer as well, because there you have the hot/cold classification as well. Have you ever looked at that or? No, this cohort is ER positive, HER2 negative only. The reason why it's relatively unexpected is because CD8 tends to confer improved outcome in triple negatives. I haven't actually looked at that myself, but this is kind of opposite to what we would have expected. So it's most likely due to chemotherapy. I know in triple negatives it does, or CD8 is a marker of improved outcome on chemo. So it was one of the problems with people considering the disease as being immune cold. There's not really been that much investigation. Could it be some kind of, I don't know, interaction with estrogen receptor signaling as well? Because it's not completely, like the intermediate risk score, like this subgroup still has potentially a lot of estrogen receptor signaling. Yeah. It probably is. Again, oncotype, like you said, includes ESR1. It's something that I probably need to look into more. I'd say more so, for example, you can see or we've seen in the data that there is a stronger relationship with post-menopausal women than pre-menopausal. And again, one of the flaws of TAILORx is that they didn't fully account for the fact that chemo was inducing ovarian function suppression. So it probably is there, but only really recent trials are beginning to look at the effect of really taking out estrogen properly has on the immune responses in these different tissues. So yeah, you'd be probably right. Got it. Thank you very much. Very nice work. Thank you. Thank you. Thank you. All right. Our last speaker, Maxime Deynaud, on unbiased phenotyping of CRISPR edited tumors with preserved map and subcellular imaging of the mouse whole transcriptome, 22,000 plex. Yeah. Thank you very much, and good afternoon, everyone. It's a pleasure to be here and represent the whole team at Nordic. I'm actually going to talk today about some recent development that we've been doing at the company, where we are mapping human tumor ecosystems on non-clinical tissue. So an introduction on what Nordic is doing. So we are leveraging foundation models, machine learning models, to develop cancer therapeutics. And it all starts with proprietary and purpose-built data sets that we are generating from large cohorts of patient samples where we generate spatial transcriptomics. We have H&E slides for pathology, and also proteomics on all these samples. And we then feed all this data into foundational models that are going to learn about human biology. And then we have a series of applications for these models, some impacting clinical development here, like clinical response prediction or patient stratification or biomarker identification. But what I want to talk today is how we are leveraging these models as well to investigate and interrogate the causal biology behind cancer and how we can use this to develop novel therapeutics for cases where there is no clinical data yet available. And to be able to do this, you basically need a platform that has a couple of features. The first feature is that you need to be able to ask a lot of questions, and what that means is that you need to be able to have a platform that can do genetic perturbation at scale in the right settings. And this is why we are using the PertiMap platform, which is a scaled in vivo perturbation platform. And what you can see here is an example of a mouse lung that has been seeded with tumors. And you can see that all these tumors are different colors because they've been barcoded previously, and we can now identify directly in the tissue which tumor carries which genetic perturbation based on the barcode that they express. And using this method, we've been able to generate a large data set of more than 600 genetic perturbations across four different pharmacological interventions or therapies. So when you think about it, this kind of represents around 2,500 traditional experiments we would try, like a given drug in a given genetic knockout. So how are we able to do this? We are leveraging protein barcodes technology that I developed back during my postdoc at Mount Sinai. And the protein barcodes are essentially a series of protein reporters that are characterized by a combination of three short epitope tags, and we can generate by combining in triplets these epitope tags. We can generate more than 100 different protein reporters from as little as 10 different tags, and each and any of these reporters can be coupled to a specific CRISPR perturbation, for example.So we are generating cells that express these libraries of ProCode CRISPRs. We inoculate them into mice, and they go and seed the lung and develop these clonal tumors in the lung that we can now resolve spatially directly in the tissue. And so this is essentially how we do this kind of workflow, and the really great thing about this methodology is that it's compatible with multimodal tissue profiling. So you get the high throughput, the high scale of in vivo, but without compromising on any of the phenotyping that you want to do. So, an example here is that you can use antibodies to detect the epitope tags, in which case you can combine this with a whole panel of multiplex IF to study your favorite biology in the TME. And then very recently, we've been partnering with Bluco Spatial Biology, and they have designed probes that can directly recognize the RNA molecules of these barcodes. And as you can see here, you can very clearly also delineate all the clonal tumors. And for each one of these tumors, we have access to the genetic perturbation that is carried by the cells, and we also have access to the spatial transcriptomics data. And in this case, it's actually the whole transcriptome that was generated. And so we have the first piece of the puzzle. We can ask a lot of questions. We can screen all these different genetic perturbations, and we can do that with multimodal tissue profiling. But this is still studying mouse biology. And the second really important element that we want to have is that we want to impact patients' lives at the end of the day. So we need to have a way to analyze this data in a way that will speak to the patients. And this is where we leverage some of our foundation machine learning models. In this case, it's a model that has been trained exclusively on patients' data. And so what the model is seeing to learn biology is aligned multimodal data, which is H&E and then a layer of whole transcriptomics, spatial transcriptomics for the same tissue. And so at training time, the model is going to learn human biology based on these two layers. At inference time, what we can do then is that we can feed the model any H&E image, and the model can predict and infer the expression of every single gene in the spatial proteomics panel. So in this case, the whole transcriptome. And essentially what this does is that it turns your H&E slide into a spatial transcriptomics data set directly without needing to actually run the assay. And when I say any H&E, we really took it to the next level, and we decided that we would just feed this model the H&E samples from the Bertamap high-throughput in vivo platform that we've been using to generate the causal data set. And so we have this way, we feed the model with this H&E. The model has been only trained on human data, so the model is going to make predictions on expression of human gene in the mouse model. So essentially what this does is it humanizes the mouse system, but in silico after the experiment has been run, and you can project human biology directly on mouse tissue. So this is what we have called a MALS for multi-species alignment and reasoning on spatial data, because our model, even though it's trained on human data, is agnostic to the species you run it on. How does that look like? So this is an example where, again, we project the human signature on the tissue, and you can see here on the H&E that you have some tumor nodules in the tissue. And the model is actually very able to identify where the tumors are located because this is where the model is going to predict that there is a very high human keratin-17 expression. And this is because keratin-17 is a marker of the human lung cancer that the model has been trained on. And as you can see, this matches perfectly the ground truth staining of the tumor marker that we use in this mouse model. What's really interesting here is that, again, the model only understands human biology and only projects human biology. So in this case, in this example, what we've looked at is the expression of CCAM6. And you may know that CCAM6 is a gene that does not have an ortholog in mice. So this gene doesn't exist in mice. If you're very interested in that, if you have therapeutics program against CCAM6, you can't really measure its expression in the mouse tissue, in the preclinical tissue. In this case, though, the model is perfectly able to predict what are the tumors that are going to express higher level of this gene. And you can see that most tumors here are comparing to the H&E on the left. Most tumors have very low to no CCAM6 expression at baseline. But we have found that a specific tumor here on the bottom left and magnified here in the center is actually characterized with a high level of predicted human CCAM6 expression. And what we find that we look at the knockouts, it's a knockout of a transcription factor that essentially is going to induce a switch in histology of the tumor. And the resulting histology resembles the tumor histology in patients that have high CCAM6 expression. And so, as I mentioned before, we've generated a very large data set of more than 600 different genetic perturbations across different therapeutics. And so the idea then was, how do we analyze this in a way that we can access the diversity of the mouse phenotype that we have generated? And so the model is only seeing the H&E and is completely blind to the perturbation, but we have the perturbation label on top of this H&E as well. So what we are doing is that we are aggregating the signatures for every tumor that has the same clonal perturbation. And then all these tumors end up being one spot here in this 2D space, which we call an embedding space. And so the way to look at this figure on the right is that each single dot corresponds to the phenotype of one genetic knockout in a particular therapeutic setting. So it could be a beta-2M knockout in an isotype control or a beta-2M knockout in animals treated with anti-PD-1, for example.And then on top of this space, we can add a number of labels. We can just represent, in this case, the expression of predicted human HER2 on top of the space. You see that the knockouts that are here on the right are characterized by a higher expression of HER2, whereas the ones on the left in blue have a lower expression of HER2. And so using this embedding space here, what we can do now is also try to understand, we have a diverse mouse biology, we have the human biology projecting on top of that, and we've done some work to understand what are the main drivers of the diversity of phenotype that we find here. So we know, for example, that there's a given cluster here that is very specific of a certain metabolic signature, for example, that we find in some patients and that we can reproduce here in the mouse model. Another big axis of viability is the immune infiltration gene signature, where you find that the left of the space here is characterized by a knockout that have a very low immune efficient signature, so low T cell activity. And they actually have also an increased fibrotic signature there. And then the ones that are here on the right are characterized by the higher T cell activity. And we can then also map all these knockouts to where they are in the space. So we can identify which are the knockouts, which are the perturbations that are going to drive these phenotypes. And when we do this, what we find is that we find a lot of established biology that we know in patients to be true. We find, for example, STK11 knockout tumors tend to be very cold and have a bit more fibrotic. beta2m knockout tumors have also low T cells. These are both mutations that are associated with lack of response to immune checkpoint blockade, so this is something that makes sense. And on the right part of the space, if you knockout beta-catenin, SOX1, which is a negative regulator in Wnt gamma signaling, or PTPN2, these are also gene knockouts that have been shown to induce recruitment of T cells. And some of these other genes that are anonymized here are things that are in the clinic right now or that we know that those phenotypes are also correlated to what we have in patients. Then one way to use these models, I've mentioned before that we have these knockouts both in the control isotype treatment conditions, but also after immune checkpoint blockade. And so what becomes interesting is that we can look at the trajectories of these different knockouts upon treatment with different drugs. And for example, if you here look at this knockout of VGFR1, the VGFR receptor, when you treat this knockout with PD1, you see a drastic shift towards the right of the space of the phenotype of this tumor. So these tumors under PD1 become tumors that have a very hot TME. And this gives you a very good rationale that it would probably be a good idea to generate a bispecific against VEGF pathway and PD1, which we know is something that has activity in the clinic. And another example here, this is a gene that has an oncogene targeted therapy target that is not typically used in combination with immunotherapy. And we can find the same thing here that we now have a rationale that this is something that probably would be a good idea to target in combination with PD1 and for example, in a bispecific format. So to conclude on this, the BERTA mouse platform that we have developed relies on two main characteristics. The first one is that we can model tumor biology in vivo and at scale, and this is allowing us to generate a large, diverse landscape of tumor biology in vivo in mouse, which reflects the diversity of phenotype that you also find in patients. The second element is that we are able to translate that biology into human biology by humanizing the non-clinical model with H&E inference. So we basically project the human biology on the mouse canvas, and then we can use this methodology to discover next generation bispecific combination targets, biomarkers, and so on. And the example I showed today was on the lung cancer data set and with PD1 blockade, but we are currently actively expanding this on different activation indications and also looking for partners in that space. If you have your favorite drug and you would like to run this kind of assay, for example, we can also run the exact same data set with your drug of choice and look at what are the drivers of response in that case as well. Thank you. All right. This paper is open for discussion. Hey. A very interesting talk. Well, my question is, is the gene prediction, gene expression prediction based on the cellular level or regional level? Yeah. So the model is seeing-- It's trained on COSMIC data, so it's subcellular resolution. Yeah. So how can you validate your prediction? Yeah. So that's a very good question. And I kind of have two elements to this, right? The first one is that we have in a set of the samples, we have the ground truth where we have the COSMIC data, but we can just infer on the H&E and then compare the ground truth towards the prediction of the model, right? The best proof for me at this point as well is that all this that I showed right now, it was really impressive to see because you have an H&E that you feed to the model. All these tumors, to my eyes, I'm not a pathologist, but look fairly much the same. And what we saw very clearly is that the model can give you, okay, I will look at a lobe of lung, and it would say, okay, these three tumors have a very specific expression pattern, and the three tumors will end up being the same exact knockout. So blindly to the knockout, not even knowing there are knockouts in there, the model picked away the different knockouts and the phenotype even for the knockouts are phenotypes we know exist in patients.And one other thing I would say is that we actually just released a demo on this model as well. It's called Target-2 that you can find on the website. You can also play with the data and see. Sorry, last question. Yep. How many samples did you use in the model training? We have thousands of samples. Yeah. Different sample, right? Different results, right? Yeah. The different-- The human samples that you use to train. Yeah. Okay. Yeah. Thank you. Yeah. Hi. Really great talk. I guess I had one question. Maybe the first one is just clarification. So when you said that CCAM6 is not expressed in mouse, when we're looking at that, the expression of predicted CCAM6, was that in mouse cells? Yeah. So there is no ortholog to the human gene. So if you would go and look at the mouse tissue, you can't detect CCAM6. But what the model does, it's projecting the human biology, so predicting, it's like, "Oh, this tumor looks like a tumor that would express a lot of CCAM6." Yeah. So I guess my follow-up question then is have you looked at the attention maps or looked at correlated genes that maybe CCAM6 wasn't the driver, but there is a correlated murine gene with CCAM6 that could explain what you're seeing? Yeah, we haven't looked into that very deeply, but what we can do with this is that the way we analyze this data and the human data is not the same. It's both inference of the human transcriptome, right? So we can identify features on the mouse tumor that are predictive of CCAM6 expression on the human side, basically. Okay. Yeah. Thank you so much. Yeah. I have a question that follows up on getting the gene expression from mouse H&E images. What if you have in vivo data? So let's say you inject some human cells into mouse, so you have both mouse and human cells. Does MAIRS work as well? Yeah. I think we haven't done this because I should have said that all the experiments we're doing is in syngenic models, like immunocompetent and everything. But the model essentially doesn't care about species, right? It would predict anyway what's happening in the sample and just overlay the human signatures. So if you have a humanized mouse, it will give you only human genes. It doesn't know about mouse biology. That's crazy. Is it available for researchers? I agree, it's crazy. Yeah. I was the first one to be mind-blown by this. Yeah, we can talk. Yeah. Okay. Thank you. Thank you. How many genes can we identify in the inferred spatial transcriptomes? It's whole transcriptome, so it's the whole, I believe, 22,000 plaques from CosMx, from NanoString. Yeah. Yeah. Thank you. All right. Thank you so much. Thank you. So many of our speakers finished their sessions very fast, so we still have about 15 minutes if any other questions for the prior speakers are available, it's open for discussion, open for any more questions. Go ahead. Oh, yeah. You're getting a lot of questions. Yeah. I was fascinated by this last talk. This idea about projecting to mouse, it seems to me that you could generalize that quite a bit because imagine that you have two people. Two people will actually have some constraints in the possibility of gene expression. For example, let's say you have a person who's born with a defective gene, then what makes up for that? I think what you're taking is the case where the mouse makes up for it in a certain way. Mm-hmm. But you might have actually the same problem, but to a lesser extent in human, mapping one human to another. I wonder if you've thought about that and how that problem might apply across people. Yeah. It is a big application of the model to also do this from human samples for which we don't have access to the transcriptomics data. So you can also do this in this case where you can, let's say you have a biopsy before treatment, and then you have a bunch of H&E data that you can access, but these are images, and the slide is hard to access. We can now regenerate the whole TX panel from those slides that are difficult to access. I'm not sure if that answers your question fully, but... I would love to talk a little more. Okay. Any other questions for any of the prior speakers? I'll ask, any speaker that's interested. So a lot of these technologies involve measuring multiple things, different parts of tissues, and biopsies. And I'm wondering how you're all thinking about converting this into a clinical assay, because that obviously is going to be challenging from a clinical standpoint. So are you thinking about how to condense it into a simple test? You're welcome to come up, any of you who are interested. Yeah, there are microphones. Or go to the microphones there, yeah. I think I touched a little bit upon that. So, for example, in the tissue, we look for peptides in the case of mass spectrometry-based proteomics, so if you find some disease-specific peptides, you can synthesize them and create a blood-based assay, where you then try to refind these disease-specific peptides in the blood. And in this way, you can not only look for a relative abundance, but you can also design peptides with a known, let's say, weight, and then you spike them in, and thereby you can then really calculate the absolute amount, because in the clinics, often you look for absolute concentrations rather than relative. So in this way, you can design non-invasive tests using these kinds of peptide assays. So you're thinking about doing it in the blood? Yeah, in the blood, for example. Yes. I will caveat this by saying I'm probably the person in the room that know the least about biomarkers. But what I keep learning from the work we're doing at Noetic is that there is a lot of information in the H&E. And I think we are developing tools now that can allow us to detect more and more of this information. So there is certainly a path for me that I think would be interesting to explore, is how we can just leverage H&E more, right? And identify features from this tissue that would be predictive of response. I think this original question was, if you're measuring everything, how do we get this thing to the clinic, I guess, right? How do we actually make it practical? And I think, one of the key thing-- It's no different than what we did in NGS, right? It's like we started out, we sequence everything, and then you hone it down to the key biomarkers that are important. It seems that the same thing will happen here. With Noetic, we're imaging the entire transcriptome. But we're doing that because we wanted to capture all elements of biology before I make the little miniature clinical grade assay, right? I think it seems like we made so many mistakes early. In NGS, we made those mistakes. We went to the small panels too early, and we missed some of the biology that should have been in our clinical biomarker assays that weren't there. So hopefully in spatial, we won't make that same mistake. But just measure everything and make sure we capture the biology in as many different areas where it's happening, and then make the clinical grade assays. Otherwise, our assays are never going to be-- They're just not accurate enough. All right. Well, on that note, then it's a perfect way for us to end this session. We're ending it early, but it's really a nice way for us to see all of the amazing technologies being developed, whether spatial transcriptomics, spatial proteomics, mass spectrometry, whether in the blood or in the tissue, CRISPR screening perturb seeks on the tissue. It's really wonderful for us to see novel biomarkers being developed. But I think the message you're hearing from us is, think of this as this discovery to understand something new. Think of it of not just biomarker or cool technologies, but does it matter for us to identify or risk stratify patients so that we can identify who will respond and who will not respond to certain therapies, and also define new therapies for them in the future. So how does that translate to a patient in the clinic, and how can it identify for us a unique patient population that will benefit from certain therapies is really the end goal here. And thinking of how to bring that to the clinic, diagnostics, predictives, biomarker, and potentially therapeutic selection will be important. Maybe not today with how expensive spatial technologies are, but potentially bringing it as a clinical assay in the future. So thank you so much, and hopefully you're enjoying AACR. Thank you.