Hello, everybody. Welcome to the session on Genomic Approaches to Understanding Hormone-Dependent Cancer and Resistance. Today I'm super excited to be able to chair this session, and first off, I'd like to just thank our organizer, Kendall Nettles, for choosing three outstanding speakers. All three of them have applied different kinds of genomic approaches to understanding both therapy resistance as well as trying to identify predictive markers for cancer. So we're going to start off with our first speaker, who is Wilbert Zwart. Wilbert is a professor and senior group leader in the Division of Oncogenomics at the Netherlands Cancer Institute. He's also the professor of Functional Genomics and Oncology at Eindhoven University of Technology and a group leader for Oncocode Institute in the Netherlands. As many of you know, Wilbert has made major contributions to the genomics of clinical samples, and again, today he's going to be talking to us about the androgen receptor enhancers in control of prostate cancer development and progression. So Wilbert, take it away. Thank you for the kind introduction, and I'd like to thank the Endocrine Society for the opportunity to speak. So these are my conflicts of interest, and the work we're doing in my lab is very much focused on hormones in cancer, more particularly breast cancer, endometrial cancer, and prostate cancer. So today's talk is focused on prostate cancer and on androgen receptor biology. Androgen receptor is considered the main driver of disease, which upon hormone binding, dimerizes, translocates to the nucleus, and activates genes which are essential for survival and tumor growth. So for today, I focus on both, not only androgen receptor, but also on FOXA1, which is a transcription factor which opens the chromatin to the androgen receptor combined, but also the active histone modification demarcating promoters and enhancers, H3H7 acetylation. The androgen receptor binds on very specific DNA motif sequences, as can be seen here on the top, and binds also on enhancer regions, as indicated over here, to regulate promoters in 3D genomic space. Now within androgen receptor biology, the key question is always, how variable are those AR sites between tumors, and what does it mean? Now this is a question that we've been studying over the last years, and we know now that androgen receptor DNA binding is highly dynamic, and highly plastic to be regulated between different tumors, being changing in tumor genesis, the transition from healthy tissue towards primary cancers, again in the transition from primary tumors to metastatic disease, and also is differentially occupied in therapy resistance. Now for today's talk, I'd like to focus a bit more on therapy resistance, and explain two unpublished stories, where we look for epigenetic biomarkers in response to therapy in MCRPC setting, and in neoadjuvant setting. And the first story is spearheaded by two talented people from the lab, Yang-Yuan Tsu and Tessa Severson, while the neoadjuvant story is spearheaded by Simon and Marloes, as shown here on the bottom. So first, the MCRPC story. This is a clinical trial performed by Andre, a very close collaborator and friend in the lab, and in this study, 60 MCRPC patients are being recruited. There's an imaging-guided biopsy being taken for treatment. Thereafter, the patient receives either enzalutamide or ibuprofen, and the treatment is continued up until progression, and then for a subset of patients, there's a second biopsy being taken. From these samples, there's a lot of genomic data streams being generated, whole genome sequencing data performed both in the tumor and the germline, there's gene expression data being generated, and promising immunoprecipitation followed by sequencing. I'm going to talk about it later. So for all these 60 patients, we have metastatic lesions analysed for either lymph node, bone or soft tissue, and after confirmation by the pathologist for high tumor cell percentage, we perform an immunoprecipitation for any of our three proteins of interest, where we fish out the protein itself, the piece of DNA where it's been cross-linked to, and then those pieces of DNA are then being sequenced by next-generation sequencing, enabling us to know on a genome-wide scale where the protein was bound. Now this is what the correlation between the different data streams look like, and we can really appreciate that for our histone modification, all the data very nicely clusters, as you can see over here. Well, for AR and FOXA1, we see a very clear distinction, so that all those actually cluster together and AR and FOXA1 are part of the same group. You can also see that there's no clear difference on the treatment groups, either between the non-treated response or intermediate, but also before and after treatment, as you can see over here, there's no clear distinction. Now the biggest group we have is for the active histone modification, so that's where we focus on next, and for those patients, we use the clinical annotations of either responding to treatment, a non-respondent to treatment, or an intermediate group. And for these, we now perform the supervised hierarchical analysis and try to identify those sites which are clearly distinct between respondents and non-respondents, and fair enough, we find over 600 sites which are only active for the active histone modification for non-respondent patients, which is also appreciated here on the right by the volcano plot, and only 25 sites which are active in the non-respondent group, which you can see over there. Now, we believe those sites are real and very strongly differential, as exemplified here on this slide, both for the non-respondent patients as well as for the respondent patients, and it's also quantified here on the right-hand side. Now, these are supervised analyses, so we want to validate this in an independent cohort, and then we sort of run into trouble because this is a unique clinical trial, so there is no validation cohort. But luckily, we got help from our collaborator Eva, who generated patient-derived xenografts from MCRPC patients, where the tumours were implanted into animals, and the animal was either left as control or castrated, and the tumour volume was followed over time. Now, for these very same PDXs, there's also atrial cataracts of anesthetics and ChIP-seq data generated, and those data we generated and published last year together with Matt and Mark, and in those samples, we looked for presence of our 600-something binding sites we identified in our patients. Here we see a very nice distinction of a subset of samples which had a very strong signal at these specific sites versus those which actually had quite a weak signal at these sites. And, as exemplified here, we can actually identify that for samples with a strong signal, we can actually see that there was actually there not really a response to castration, indicative that those sites which were predictive for therapy resistance in patients that were also there, so there was no response to hormonal intervention in the animals. While for those tumours which are associated with the weak sites, you see a very clear drop of tumour growth after castration. Of course, this is only one example, so let me show you all the data. So this is what we had there, that for all the samples which were associated with strong signal at those sites which were associated with a resistance to either astutamide or abiraterone, we can see that for most of those animals there's no response to castration, while we see that for those where there's a weak signal at those sites, we see a very clear drop of tumour growth. Now for this part, of course, we're still working on this study, but for now I've already shown you that there's a distinct subset of over 600 sites which are predictive for resistance to astutamide or abiraterone. OK, so now I'm going to switch gears and I'm going to switch to the next story, how we look for biomarkers in the neoadjuvant setting. Now this is another clinical trial, phase 2 study, neoadjuvant study, which was spearheaded by Henk van der Poel, where high-risk prostate cancer patients after diagnosis had an imaging guided biopsy, thereafter there was three months of astutamide treatment and surgery. After surgery, there's a tumour guided biopsy which is being taken, and now we have both before and after treatment samples. On those samples we performed multi-omics profiling, including gene expression analyses, DNA copy number data, and again promising immunoprecipitation for our three favourite proteins. Now let's first focus on the gene expression data. So here I show you a principal component analysis, which clearly shows distinct groups before and after treatment, and performing a gene set enrichment analysis showed there's a very clear drop of endogen response and cell cycle pathways after treatment, while there's an upregulation of apoptosis and other pathways upon treatment. Now let's first focus a bit more on the imaging receptor pathway. So here I show you a gene set enrichment analysis for endogen response, and you can see there's a very nice downregulation of that pathway, and this appears to be a rather universal feature that every single gene involved in that pathway is all being downregulated in all the patients upon treatment, as you can see here. Now a couple of years ago we reported, using very same technologies, we reported in a primary prostate cancer cohort the existence of three different subtypes. The first of the subtypes was very much following the TEMPRS-to-ERG type of separation, which are also then vermic work previously reported by Mathieu Lippien, but now we also find this third subtype, which is actually quite interesting, because that subtype was hallmarked by a low AR activity while AR was expressed, but a high level of neuroendocrine type features. And interestingly enough, when we look into our cohort we study now, we can see that before treatment we see distribution which is very much following the same distribution as we originally find in our 100 patient cohort, but after treatment we can see that all those tumours are being pushed towards subcluster 3. So three months of insulin-like treatment suffices then to force all those samples towards more neuroendocrine type of behaviour. AR is inactivated, as I've already shown you, but we also see that there's an upregulation of neuroendocrine type genes upon the three months of treatment, as indicated over here. Okay, now let's switch gears and continue now again on the ChIP-seq data. So here I show you examples of peaks for AR, FOXO1 and H3N37 installation, and you can see there's quite a strong robust signal for all of these. When performing a clustering of all the data and scripts are ready to depict what we expect to see. So we expect that AR and FOXO1 are binding together on many of the same sites, which are positive for H3N37 installation, so we expect all three of them to be together, but we also expect H3N37 installation to have a life on its own and activate genes which got nothing to do with AR. And this is basically exactly what we see. So we have, on the unsupervised biological clustering, we can see a very nice correlation of H3N37 installation, as depicted over here, and AR and FOXO1 are completely intermingled between the different samples and there's no stratification on the four phases after treatment, which is then also something nicely visualized here on the right-hand side in the principal components analysis. Now, a bit more of a QC analysis, we can see that AR and FOXO1 are rarely binding promoters, as one would expect, and the majority of the signal is found at intergenic regions and in infants. H3N37 installation is more promoter-bound, as we also would expect, and again, found at enhancers. For motif analysis, we find that AR is occupying sites which are positive motifs for not only AR but also forecast factors, again, what we would expect, and the reciprocal is what we observe for FOXO1. So, so far, so good. So when we now focus specifically on the FOXO1 sites, we find something really interesting. So first of all, the principal component analysis shows unsupervised that we see a difference between pre- and post-treatment samples, but also, for unsupervised analyses, we find there's over 1,100 sites which are very, very different between before treatment and after treatment. Those sites are, again, quite robust and quite reproducible, and we can see a nice drop in the subset of sites while there's a clear gain after treatment with three months of enzalutamide. And also, something we also sort of expect, that if FOXO1 is going down, this is completely followed by a decrease of AR at those very same regions. FOXO1 goes up, also AR is nicely going up. One thing which we did not expect, however, is that for the pre-treatment FOXO1 sites, there's a very low signal of K27 acetylation. So those sites seem to be completely inactive. While at those sites which are enriched after treatment, there's actually quite a strong signal for K27 acetylation. This is what the raw data looks like. Again, there's no K27 acetylation signal at the pre-treatment FOXO1 sites, but clearly present and increased after three months of enzalutamide. Now, this is something we'd like to validate. So for that, we made use, again, of our 100 patient cohort. And those are all patients which did not receive any systemic treatment. And so they should all be considered as pre-treatment patients. And here again, we do see the data at the pre-treatment FOXO1 sites. In those regions, there's no K27 acetylation. Those sites are inactive in contrast to the post-treatment FOXO1 sites. Now, either these sites are epigenetically repressed, or these sites don't have the capacity to be active. So in order to make the distinction, we reached out to a close collaborator, Nathan Lack, who performed an assay called STAR-seq, which is an acronym for self-describing active regulatory region sequencing, where AR binding sites are cloned into a library. And actually, it's indicated over here. It's indicated over here. And once those AR sites act as enhancers, then they boost the expression of the corresponding transgene. And you nicely actually see that there's an increase of transcription, which can then be sequenced and analyzed through conventional RNA-seq. Now, the library is transfected into mRNCAP cells. And those cells are treated or not with DHT. And we can see there's very nicely inactive AR binding sites. But we also find AR sites which are nicely active and constitutively active, irrespective of the former. Now, when we project those data back on our differential FOXO1 sites, what we observe, and this is something we clearly did not expect, is that the inactive AR binding sites are very much enriched for the pre-treatment FOXO1 regions. While for the constitutively active AR sites, those are exclusively observed for the post-treatment sites. So this means that the pre-treatment FOXO1 sites, which are inactive and also don't have any active histone modifications, that those sites are intrinsically incapable to become active based on their primary DNA sequence. Now, what about the genes they regulate? So for this, we made use of K27 installation high-chip data focused on the FOXO1 binding sites, copying those into genes. And for those genes, determined whether those genes were essential to drive prostate cancer cell proliferation. And for that, we made use of data from the broad using CRISPR knockout screens. So at the bottom, you can see the genome-wide data. We can see, make, for example, FOXO1 AR, which are essential for prostate cancer cell proliferation. Those genes, which are under potential control for the pre-treatment FOXO1 sites, we can see no clear difference here. But for the post-treatment sites, we can see a very nice, strong statistical enrichment for being essential to drive prostate cancer cell proliferation. Now, which transcription factors are regulating these genes? In order to identify that, we, again, make use of our pre- and post-treatment FOXO1 sites and then analyze not only the potential motifs which are underlying those sites, but also which transcription factors have previously been reported to occupy these regions. And we wanted to model this also in cell lines. For that, we made use of NNCAPs, either with or without insulinimide treatment, but also an insulinimide-resistant NNCAP derivative, 42Ds, which we received from Aminazabete. Now, first, we observe FOXO1 being differentially recruited, as we would expect. And it's also something we could also validate in our cell lines. There's a decrease of FOXO1 binding post-treatment at the pre-treatment sites. And these sites nicely are being upregulated upon treatment with insulinimide. And this is done even stronger in the resistant cell lines. So far, so good. So, our top hit was ARNTL. This one is very, very unexpected and very surprising, because this is a circadian regulator, a classical dimerization partner with CLOCK, involved in circadian rhythm control. This has never been described before in prostate cancer. But we do find that now ARNTL, to occupy these post-treatment FOXO1 sites, they are increased upon treatment, even stronger in the resistant cell lines. They are also upregulated on protein level after treatment with insulinimide, as observed here in the cell lines, again, stronger even within the insulinimide-resistant cell lines, but also in our patients. We see after three months of treatment that ARNTL is upregulated, and more importantly, that it's associated with outcome. So, while before treatment, there's no difference of ARNTL expression before versus for the patients who do not respond. But after treatment, we can actually observe that ARNTL levels are a lot higher for those patients who respond quite poorly to three months of insulinimide treatment. Also, quite importantly, we see that in our cell lines, that when we get rid of ARNTL in our parenteral end caps, then nothing really happens in the level of cell proliferation in our insulinimide still works. But in the insulinimide-resistant cell lines, we can see that if we get rid of ARNTL, the proliferation is completely abolished. So, with that, I'd like to conclude and state that a new agent of insulinimide treatment enriches for neuroendocrine-like transcriptional programs, which are under control of active FOXE1-bound enhancers, and which are critical to drive tumor cell proliferation. And also that these sites are dictated by circadian rhythm gene ARNTL. So, these are some key references for some further reading. And there, I'd like to conclude, but not before acknowledging all the work from Tessa, Yan-Yan, Marloes, and Simon, Andre Bergman, very close collaborator and a strong friend in the field, Henk van der Poel, all the funding organizations, and of course, all the patients. Thank you very much. Okay, thank you Wilbert. I want to remind everyone that you can ask questions of our speakers down in the Q&A section down at the bottom, and I particularly want to encourage trainees and both graduate students and postdocs to be asking some of these questions. We've seen a lot of questions from senior investigators recently, and we'd like to hear from some of the young folks. So, I guess I'm going to start while we give people a chance to ask questions, Wilbert. One thing I wanted to ask you about was a little bit off topic in the sense that you see this very good concordance between AR and FOXA1. There was an interesting talk presented yesterday about the possibility that the ratio of FOXA1 expression to AR expression may change the binding sites such that FOXA1 may bring AR to sites that are half sites for the AR instead of full sites. So, I'm wondering when you are looking at your sites, do you see full AR sites or do you make that distinction between full and half sites in your analysis? Thanks, Elaine. I think that's a really interesting question. Actually, we have not made that strong distinction per se. So, we were mostly focusing on the classical full AR motif. But there could be something indeed pointing in that direction in the data we have. So, we do see that post treatment FOXA1 expression levels are going down a bit. So, they could give a preference for different types of AR sites. That said, we do feel that those sites where we see FOXA1 signal post treatment that AR, even though it binds there, it doesn't seem to really do anything. So, actually those sites are then not anymore under AR control. So, it could very well be that it then pushes AR to different sites, but not necessarily active in that specific context. But it's a very good point. I think we should reanalyze data on that. Thank you. Okay, so we have a question from Scott Dem. He says, Outstanding work, Wilbur. What is the expression pattern of ARNTL during prostate tumor genesis and disease progression? Thanks, Scott. Yeah, that's also a really, really interesting question. So, we do see there's some differences of ARNTL expression as the tumor develops and progresses, that it starts to go up, basically. And the same goes for CLOCK, which is also getting upregulated while CRY and PUR, most of them are going down. So, actually we see then on the grand scheme of things, a bit more of a circadian rhythm switch more towards the day phase, if you will. But this is all based on a more insidical type of analysis. So, we really need to dive deeper and actually see if this is something we can also validate in some of our own data sets. This is a very good question. Thanks. Yeah, very good. So, Matt Zecora asks, In your CRPC study, is the differential H3K27 acetylation pattern associated with gene expression patterns, or is it more or less predictive of response than RNA-seq data? Yeah, thanks, Matt. Yeah, it's a very interesting question, but it's a bit tough one to address. So, we do have RNA-seq data for a subset of the samples where we also get H3K27 acetylation chips for, but the chip, or the RNA-seq quality is not the strongest in these particular samples. It's basically due to the sample workup. So, we do have some analysis on some of the PDXs that Eva has. So, she also generated some RNA-seq there. And we do see differential expression mostly for some of the, seen for the top differential H3K27 acetylation associated sites, which you also would expect. So, we're still in the process, actually, of further filtering and actually making sure that this is not an over-trained kind of thing, but actually that the differential expression there is real, and then we'd like to validate it in other follow-up cohorts. So, the jury's still a bit out, it seems to be the case, but we need to have a bit more time. Okay. The next question is from Alice Levine. She says, thank you. How do you tie in the neuroendocrine phenotype with findings on FOXA1 clock genes, and are they known to push the neuroendocrine differentiation pattern? Yeah, that's a very, very good one. So, I would like to stress that we don't really see full neuroendocrine disease after treatment, right? So, these are still very much just adenocarcinomas, which have a subtle switch in gene expression profile that they can get more positive for neuroendocrine type gene signatures. To what degree would actually have then a full push towards neuroendocrine disease? There's work from Matt Friedman, which does show that in neuroendocrine prostate cancer, you get a FOXA1 reprogramming to other sites again, which is very much distinct for neuroendocrine prostate cancer as compared to adenocarcinoma. So, it is true that FOXA1 dynamics actually is associated with that. So, I do think there's an effect there. In relation to clock genes, however, that is something that is so new, but I couldn't help, I couldn't resist to show it. But the connection there with the neuroendocrine features, that's really something we need to explore further. So, I'm going to end this, Wilbert, with asking you to speculate a little bit. So, is there, do you know of any information about how circadian rhythms might be affecting prostate cancer tumorigenesis or, so beyond the gene expression, but is there any epidemiological evidence to suggest that circadian rhythms have a part to play in this tumorigenesis or else therapy, right? For tumorigenesis, actually, there's quite some literature out there which points in that direction. So, that indeed that shift workers or pilots or people who have disruptions in circadian rhythm have an increased risk of prostate cancer development, but not only prostate, but also breast cancer and other cancer types. There's been thus far no literature on associations with treatment response and shifts in circadian rhythm. But there's also a paper from Karen Knudsen, who we're also, we're involved in, which came out last year, that CRY1, for example, is under direct AR control. PURR has been described to be interacting with AR. So, there's quite some crosstalk and interplay between circadian rhythm genes and AR, but also with circadian disruptions and potentially prostate cancer development. But, yeah, so there's something going on there, I think. That's very interesting. Thank you very much. So, we're going to move on to our next speaker, who is Wenhui Wang. She is a professor of bioinformatics and computational biology and biostatistics at MD Anderson Cancer Center. Wenhui's done some terrific work very recently in teaching us how to use more bulk RNA-seq and deconvoluted to learn more about the total environment of the gene expression that we're seeing in these tumors. She's going to be talking to us about tumor cell total mRNA expression shapes the molecular and clinical phenotype of cancer. Wenhui. Thank you for the introduction, and I'd like to thank the Endocrine Society for inviting me here to participate in ENDO 2021. So, today I will talk about some recent work in my lab on tumor heterogeneity and tumor cell total mRNA expression. Here's my disclosure. So, there has been a lot of interest in finding tumor cell features such as RNA and DNA content to differentiate atypical hyperplasia from truly malignant cells over the past 250 years. The two main technologies are flow cytometry and fluorescent in situ hybridization. And these are great inventions allowing for the in-depth analysis of individual cells, but are limited in its throughput in a number of cells and the number of targeted transcripts to be studied. The single cell RNA sequencing technology invented just a decade ago allows for a large scale quantification of RNA molecules across cells. Shortly after its invention, researchers have pointed out significant variations existed in total mRNA content that is independent of technical variations and by using spiking quantification standards. So, it has been a routine practice now in analyzing single cell RNA sequencing data to adjust for a cell size factor, which is a combination of transcription size, the total mRNA content, and library size. So, early in 2020, researchers from Stanford looked into 14 single cell datasets and pointed out that the single cell transcriptional diversity as measured by total number of expressed genes, a surrogate for total mRNA content, is a better hallmark of developmental potential than any other specific gene pathways alone. So far, there is no systematic study about the effects of total mRNA expression levels in tumor cells, and little is known about its direct phenotypic impact on cancers. So, our study aims to fill in this gap in knowledge. The work I present today is available on our archive, and it has been led by two very talented young researchers, my bioinformatics postdoc, Xiaowen Cao, and a head and neck surgeon and cancer biologist, Jennifer Wang, Adam Anderson. So, first, to replicate what's observed in normal cells, we obtained single cell data from nine patients in four cancer types, colorectal, liver, lung, and pancreatic. And two of them were generated in-house, and these plus liver have detailed clinical follow-up information. So, we compared the gene expression distributions from 100 random selected cells from tumor cells, stromal cells, and immune cells. And here, we saw that the gene counts for the total number of expression, oh, by the way, so the black means expressed and the gray means not expressed. So, gene accounts are highly correlated with total UMR counts across conditions, whereas in contrast to stromal and immune cells, there is a much larger dynamic range and a dichotomization in the total expression levels in the tumor cell. In colorectal cancer, patient one relapsed at four months, and patient two remained relapse-free at the last scan. We observed three surrogate clusters in the tumor cells presenting distinctive distributions for the, with the gene counts as well as the total UMR counts, with the far right cluster presenting the highest total UMR counts as compared to any other cell types. And then we used monocle two to look into the tumor cells and found that high UMR cells are located on one end of the trajectory and low UMR cells on the other end. And this ordering is consistent with the differentiation score measured by cytotrace, with high UMR cells being less differentiated. Comparing patient one versus two, we saw the high UMR cluster in patient one presented the highest cytotrace score, 0.87 out of the maximum of the two. And we then replicated that observation in liver, lung, and pancreatic cancer. So we conclude that there is much higher diversity in total MRNA expression in tumor cells versus non-tumor cells, which matches the differentiation states. And the higher diversity in total MRNA expression in tumor cells is observed from patients with poor prognosis. However, we're limited with a small sample size to make conclusions on the clinical outcomes. We therefore would like to measure tumor cell total MRNA levels in bulk samples that are made available from large patient cohorts. Two challenges prevented previous studies from addressing this question. The first is total MRNA expression levels are not directly measurable from high-throughput sequencing technology. And this point was made elegantly during the microarray era by Rick Young's group. And then this is their illustration. So when a single gene is elevated in MRNA, the signal remains after normalization. However, when all genes are elevated, then the signal is normalized out together with technical artifacts before the differential expression analysis. And the second challenge is tumor heterogeneity. Two zoom-in views. Here are two zoom-in views of one HME slide of a resected tumor sample used for pathologist report. So here that the large cells with open chromatin and irregular shapes are the tumor cells, whereas the elongated nice-looking cells are the fibroblasts. The small ones are the lymphocytes that are round and solid colored. So sequencing data from bulk tumor samples are measuring mixture of signals from different cell types, and the compositions of these distinct cell types vary largely across different samples from the same or different patients. So we propose a solution to evaluate tumor cell total MRNA expression that uses one challenge to solve the other, deconvolution. Deconvolution contrasts two groups of cells within the same experimental condition, hence canceling out only the technical effects by default. We then use a single cell RNA-seq data to generate pseudobulk deconvolved RNA-seq data to measure, to compare the mean total UMI counts for the tumor cells versus non-tumor cells, which confirms the key assumption moving forward using bulk deconvolution, which is we did observe a significant difference in the average total MRNA levels for tumor versus non-tumor cells. So we therefore propose a model for tumor-specific total MRNA expression, TMS. TMS measures the total MRNA expression per haplogenome for tumor cells over non-tumor cells. Fundamentally, we're interested in measuring the total amount of MRNA across all genes from tumor samples. But to be able to measure that quantitatively across hundreds of samples using high-throughput sequencing technology, we need to take the following considerations. First, we need to use the ratio of tumor versus non-tumor cells to cancel out technical effects and standardization. Second, we need to divide by the total number of cells to calculate per cell total MRNA content. Third, we divide by ploidy to calculate per cell per haplogenome total MRNA content to adjust for dosage effect. So this per cell per haplogenome total MRNA content, is our quantitative metric. So for each tumor cell, we can calculate TMS using matched DNA-seq and RNA-seq data. We first calculate the tumor cell ploidy in proportion using DNA data, and then the tumor-specific total MRNA proportion pi using RNA data. And TMS is then calculated as an odds ratio of the two proportions and further adjusted by ploidy. We have provided more than 50 pages of technical details written up and available on BioRxiv. So here I just mention two features. First, we use the consensus of ASCII-CAT and absolute output to achieve a reliable estimate of tumor cell ploidy proportions. Second, we use our own method DNxT to perform the MRNA proportion estimation. So we have developed and validated the DNxT model to confirm its accuracy estimation for both the two-component and the three-component deconvolution. However, we realized that we're still limited in model identifiability and dependent on the genes that are selected to generate the input matrix. We therefore further developed a profile likelihood approach for optimizing the gene selection, which substantially improved model identifiability and furthered the model accuracy. So we have calculated TMS for 5,000 samples across 15 cancer types in TCGA, where the input data for DNxT were available. Here are the distributions of TMS for cancer types in TCGA. Here are the distributions of TMS for cancer types as ordered from the low to the high median level of TMS. Other than head and neck cancer, the most indolent cancer types are located on the left and the most aggressive cancer types are located on the right. We then ask whether TMS is associated with known prognostic characteristics, which are clinically utilized. So we compared HPV in head and neck to triple negative breast cancer status, hormone status, subtype and amplification in breast cancer, subtype 1 and 2, type 1 and 2 in renal papillary, and Gleason score in prostate cancer. And all of them are positively correlated with TMS. We observed a moderate correlation between cell proliferation markers and TMS, which is expected because more than one biological pathway should affect the tumor cell total mRNA expression. Here's a striking finding that we made that TMS refines prognostication on pathological stages in all 13 cancer types where we have the pathological stage information. So here is the forest plot showing you the estimated hazard ratio and the corresponding 95% confidence interval for the black, the main effect of TMS, purple, the main effect of stage, early versus advanced, and the interaction term. So all but one cancer type showed a significant effect for either the main or the interaction term of TMS. We further grouped the cancer types into three patterns. The first and the most frequent pattern entails that the high TMS presents worse prognosis in both early and advanced stage. So here are two examples, a couple of myocarditis for all cancer types combined and for long adenocarcinoma. So here the blue is early stage, purple advanced stage, and the darker shade is high TMS. So the second pattern includes head and neck squamous, lung squamous, and bladder cancer, which are biologically correlated. Here, high TMS presents worse prognosis in early stage, whereas low TMS presents worse prognosis in advanced stage. The third pattern includes the subtypes of breast cancer as well as renal papillary. So here the low TMS in early stage presents worse prognosis. Probably most clinically relevant is this 33 triple negative cancer patients with high TMS in early stage presented no events over the five years of follow-up after tumor collection. And a plausible explanation for this is transcriptionally active breast cancer cells have been reported to respond better to chemotherapy. So we went beyond TCGA and obtained two additional large cohorts to validate our finding that TMS refines prognostication using Gleason score in prostate cancer. We further built a risk prediction model using TCGA prostate data, which showed good predictive performance in early onset prostate cancer at CGC data set with an integrative AUC of 0.81 for predicting survival outcomes. Now we move from the intra-tumor heterogeneity to intra-tumor heterogeneity and ask how sensitive is tumor cell expression in detecting dynamics in tumor cell phenotype over evolution. So we obtained sequencing data from TracerX multi-region study of early stage lung cancer, 94 regions from 30 patients. And each region has been studied and the phylogenetic structure has been reconstructed. So one highlight of the previous study was that the percent subclonal copy number aberration, which measures ongoing chromosomal instability, is more important for tumor progression than overall copy number aberration burden. So for each region, we calculated TMS as is shown on the right, the overall CNA burden and the percent subclonal CNA. Up to the patient's level across regions, we then calculated the maximum TMS, the median TMS, the total percent subclonal CNA, range of TMS and number of regions, and we associate these with the disease-free survival. So first at the regional level, we observed a much higher correlation of TMS with percent subclonal CNA as compared to the total percent CNA. And then at the patient level, we observed TMS max can be explained by a linear function of percent subclonal CNA, range of TMS and number of regions, with the percent subclonal CNA being the major contributor as is shown here in the regression curve. We further used the TMS max to separate patients successfully into a good and bad prognosis group, which confirms what we observed in the TCGA data. In contrast, interestingly, the median TMS cannot group patients in different prognosis outcomes. On top of subclonal CNA, TMS max contributes uniquely and then further helps patient stratification in terms of risk. So to conclude, we found that regional TMS identified spatial heterogeneity in transcriptome dysregulation across regions and may provide prognostication from subpopulations of tumor cells that are presented in one region. So we believe that tumor cell total mRNA expression is a key feature to track tumor cell phenotype. Then we ask, what are the biological processes that would change tumor cell total mRNA expression? This is clearly a large enterprise, so I'll just show you two examples, which are still ongoing. I'll show you two examples to what you appetite. Our central hypothesis is that total mRNA expression measures the response of tumor cells to modifications in epigenetics, microenvironment, mutation, and chromosomal instability at a critical moment that it sees where the tumor cells are headed to next. So we first look into the mutations. By limiting to driver mutations, we identified 24 cancer gene pairs where there is enough sample size to perform statistical testing. So we found that the TP53 driver mutations are mostly associated significantly with high TMS, whereas MAP3 kinase 1, PIX3CA, and KRAS mutations that are known to associate with group prognosis presented significantly lower TMS. We then did an agnostic search of all non-significant mutations and ended up with 34 cancer gene pairs, which then recapitulated what was found in the driver mutation analysis. The same genes pop up, with just one exception, a new gene, FGFR3, in bladder cancer presented significantly lower TMS, which is also consistent with the ongoing literature on this gene. We then evaluated the genomic features, such as the TMB, the tumor mutation burden, chromosomal instability score, which is the percent of copy number operations across the genome, hypoxia score, and the TP53 mutation fractions, including all non-significant mutations. So here, we compared the distributions of these features across four patient groups, which are defined previously by the TMS stage, and I have shown you that they present different prognosis outcomes. We ordered these patient groups from the highest to the lowest five-year survival probability. Horizontally, we saw that almost all cancer types, the patient groups, are differentially expressed or distributed in at least one genomic feature, except for triple-negative breast cancer. Vertically, we saw that genomic features are associated with at least a few cancer types, except for hypoxia, which are consistently significantly associated with all cancer types, except for triple-negative breast cancer. So we conclude that TMS captures cancer-specific dysregulations of genomic features, whereas the hypoxia signal might be a pan-cancer signal associated with TMS. And triple-negative breast cancer patient groups are uniquely defined by TMS with biological signals that are yet to be identified. So I'll pause here to give you a quick summary. So we have proposed a quantitative metric to measure tumor cell total mRNA expression, or transcriptome ploidy, in analogy to genome ploidy. It provides a key feature to track tumor cell phenotypes and to identify essential biological processes that are cancer type-specific. As compared to total mRNA expression being a hallmark for developmental potential in normal cell types, tumor cell transcriptome ploidy plays a more intricate role in cancer, likely due to its interaction with tumor microenvironment and drug treatments. And the method we used to generate TMS is mainly the DMX-T R package, which is freely available from Bioconductor. I'd like to thank all of my lab members and our collaborators. This is truly a huge team effort, and we have been closely collaborating with other computational biologists group, cancer biologists, and clinicians. And all of them have contributed significantly to this project. And many thanks to my funding sources. Thank you. Thank you, Wenyi. Thank you, Wenyi. That was very interesting. We have one question, and then I will let's see. This is from Ann Danoff. She just says, fabulous talk. Thank you. So, I'm going to ask you, I just have a general sort of ignorance question on my part, which is the fact that you see this high, you know, we have a lot of evidence that pathway specific programs like AR that Wilbur just talked about, ER and breast cancer are very predictive, and what you're finding is that total RNA is equally as predictive for some of these cases. Is there some other thing, like if you were to look for just something simple, like RNA Pol II, do you see an association with TMS with RNA Pol II in these tumors? We looked, no. No. No, my understanding was the Pol II, there's so many subunits. So, we saw, I think 2A is most well understood, and then we didn't really see a clear correlation there, but we did see correlation with other subunits that are less studied. So, there's something, there's definitely something going on with the entire polymerase enterprise. And we did perform, say, a pan-cancer scan of GSEA, and then the RNA regulation pathways definitely came out on the top across all cancers in their, as a major gene sets, the pathways that are explaining this regulation in TMS, but not specifically Pol IIa. Actually, degradation, RNA degradation seems to be more related than the other way, yeah, that kind of increased the transcription. Interesting. We have another question from Joanna Kubo. I'm not going to pronounce her last name because I don't want to mutilate it, but it says, thank you for an excellent and very informative talk. How does your TMS-based risk stratification compare with standard risk stratification systems such as TNM driving oncogene in case of predicting prognosis? Yes, thank you, Joanna. I think I did, these are the two things that I mentioned in the talk. First of all, the TNM, the effect of TMS is actually within the TNM staging. So, for example, within, we first separate the patients in early stage versus late stage, and then within the early stage across the patients, and then we identify further classified patients into high versus low TMS. So that seemed to work the best. So, for example, when I mentioned there are different patterns, like in bladder cancer, there's luminal versus basal type, and basal type are clearly more aggressive. So we separated early stage versus late stage, which is more or less luminal versus basal, and we saw that TMS, high TMS correspond to worse prognosis in luminal, but then in basal and in late stage, the high TMS correspond to better prognosis, which is actually consistent with the fact that basal cells are more responsive to chemotherapy. So that switch is expected in the clinic. And then in terms of driving oncogenes, we did look at the driver mutations that has been provided in the TCGA, and as I showed in one slide, so actually TMS can be used to find, to identify driving genes. So we are able to perform agnostic search and found exactly the same signals as presented by the driver mutation annotated genes. Except for with one gene, FGFR3, again in bladder cancer, that we found mutations in FGFR3 are predictive of better prognosis, which has been shown in literature, but it was not picked up in a driver mutation analysis because of limited amount of driver mutations that were reported in the literature in the TCGA data. Yeah, I have one last question for you. Have you, have you tried to apply TMS in any kind of predictive way in terms of looking at therapy responses? Yes, so actually what Wilbert was talking about, the Allyncap, we have already looked at the measured TMS in Allyncap data before and after enzalutamide treatment, and we saw a huge drop. So that convergence in the clusters on endocrine-like behavior, definitely these are the cells that are associated with low TMS. Yeah. Yeah, in both bulk and single cell data. Lots of opportunities there, I think, for collaboration. We have one last question from Mark Lawson, which, which asked, are there other metabolic conditions of the cells such as hypoxia or energy availability that contribute to driving TMS? Yes. Thank you, Mark. Excellent question. So we have already, I showed about the hypoxia level and which hypoxia score was based on 61 genes. And when, then we further looked at metabolic pathways. So we looked at, there's one annotation, a list of seven different metabolic pathways and found specifically pentose phosphate pathway came out as being a pan-cancer signal that is strongly associated with TMS. And that's 13 genes. And we also saw this really strong networking between these 13 genes and the 61 genes presenting the hypoxia signal. So we have like a total of 79 genes to further mine. Very good. So thank you very much, Wen-Yi. We're going to move on now to our third speaker, who is Felix Fang. Felix is the George and Judy Marcus Distinguished Professor of Radiation Oncology, Urology, and Medicine at UCSF. He is the Vice Chair of Translational Research and the Director of the Ben Haag Initiative for Prostate Cancer Research. And we all know Felix from his outstanding work, again, using genomics and applications to clinical samples and trying to improve our understanding of resistance. Today, he's going to be talking about integrating next generation sequencing approaches to investigate the role of the androgen signaling axis in metastatic prostate cancer. So take it away, Felix. Thank you for the introduction. And also, I'd like to thank the Endocrine Society for inviting me to present. Today, I'll be talking about integrating next generation sequencing approaches to investigate the role of the androgen signaling axis in metastatic prostate cancer. Here are my disclosures. And I should note that I consult for several companies that have androgen-directed therapies for prostate cancer. So, I want to begin by discussing a large, multi-institutional, integrated genomics, epigenomics effort in the context of patients with metastatic, castor-resistant prostate cancer that I've been fortunate to co-lead. And so, this is an effort that was funded by the Prostate Cancer Foundation and the Stamped Cancer Organization. And my co-leader here is Eric Small. And basically, we received funding to biopsy hundreds of metastases from patients with castration-resistant disease. And we have subsequently performed a number of sequencing approaches on these samples, including, as you can see here, whole genome sequencing, transcriptome sequencing, whole genome bisulfide sequencing, and a series of other epigenetic approaches on the tumor tissue and then circulating tumor DNA tissue and then circulating tumor DNA analysis as well. And today, I'm going to go over some of those results. And what we found with whole genome sequencing is that we were able to define the landscape of structural variants in metastatic, castor-resistant prostate cancer. And here, you can basically see a map of all the different chromosomes, chromosomes 1 through 22 and the X and Y chromosome on the X-axis and the location along those chromosomes on the Y-axis. Areas of green are consistent with copy number gain. Areas of blue indicate copy number loss. But you can see areas of gray kind of throughout this map. And these areas of gray indicate areas where structural variants are piling up in the genome. And the darker the gray bars, the greater the prevalence of structural variants at a particular location. And I want to draw your attention to this large dark bar on chromosome X, which indicates the location of the angiotensin receptor, abbreviated AR here. And you can see there's a lot of structural variants piling up there. And so it's not unexpected that we see structural variants or copy number gain or amplification at the angiotensin receptor. It's well known to be amplified in patients with aggressive prostate cancer. But it was always thought that this peak of amplification corresponded with the AR gene body. When we did our deep pole genome sequencing and we went to 100x coverage, what you could see is that there's actually the peak of the amplification is actually 624 kilobases upstream of the AR gene body shown in red. And so it's actually there's something upstream of AR that's important here. That's what we thought. And what we found was that 80% of all patients have an amplification at this particular peak upstream of AR. In 58% of the cases, this amplification was due to tandem duplication. And so tandem duplication is a phenomenon where you take part of the genome and you copy and then paste it in backwards next to this. And then you copy again, kind of forward facing, copy backward facing. And so it's basically like having an aberrant copy paste function on your keyboard and just putting in small pieces of the genome back in over and over at a place. Now, it turns out that 12% of metastasis had a tandem duplication at this peak and lacked amplification of the actual AR gene body itself. And because we went to such depth on our whole genome sequencing, we were able to get relatively nice resolution as to the smallest area of amplification in this area. And what we found was that that area of amplification coincides with ChIP-seq data for H3K27 acetylation. So there's basically a H3K27 acetylation peak right here that basically suggests that it's consistent with an enhancer being at this location. And actually, when we published our paper in Cell two years ago, Matt Freeman's group had actually published a paper roughly at the same time showing that this area functionally contains an enhancer. Really elegant work, so I suggest you look at that paper if you get a chance. And what we were able to do in our clinical data was to show that if you had either amplification of the AR gene body or this enhancer element upstream of AR, this was by itself sufficient to drive overexpression of the antigen receptor. And so here on the y-axis, you can see expression of AR by RNA levels through RNA-seq. And again, just having the AR enhancer by itself increased AR expression. Now, my collaborator Chris Maher and his team subsequently showed that you can detect these AR enhancer amplifications in circulating tumor DNA from patients and that the presence of this enhancer is actually prognostic for treatment with antigen signaling inhibitors. And so here you can see in red are patients who have either amplification of the AR enhancer or the AR gene body or both. And you can see that these patients do worse when treated with AR-directed therapies compared to patients who don't have these alterations as well. Now, I've talked a bit about whole genome sequencing. I now want to transition to our studies looking at methylation. And many of you probably know this, but methylation results in condensation of the chromatin and reduction of gene expression. And what we did was we knocked out the methylation landscape of metastatic prostate cancer using deep whole genome bisulfide sequencing on 101 patient samples, again, all from metastatic prostate cancer. And what we were able to show is that whole genome bisulfide sequencing demonstrates focal hypomethylation marks that are indicative of additional antigen receptor enhancers. And so here what you can see is the AR gene body shown in blue, as well as an enhancer that we have previously described upstream of AR. And what I'm showing you here are tracks of hypomethylation. And so each row represents data from a different patient with metastatic catcher-resistant prostate cancer. And the areas of black that you see here are areas of focal hypomethylation, so lower methylation. And so this area of hypomethylation here represents the AR promoter. You can see that there's an area of hypomethylation that coincides with the AR enhancer that I just showed you by a whole genome sequencing. But you can also see that there are other hypomethylated tracks upstream, and actually downstream, of the antigen receptor in the same area. And this is the data from metastatic prostate cancer. When I now switch to show you data from either primary prostate cancer or benign adjacent tissue, you can see that there are no areas of recurrent hypomethylation through this area. And what's pretty nice, actually, is that we were able to look at associations between these hypomethylated regions as well as expression of antigen receptor. And we can see that for the majority of these hypomethylated tracks, there's actually a correlation with AR gene expression, meaning that the more hypomethylated the area was, the higher the expression of AR. And again, this is consistent with these other areas being enhancers as well. And we then also overlaid ChIA-PET data from prostate cancer cell lines. And what you can see now is that these putative enhancers actually loop back to the AR gene body itself and to other enhancers, consistent with this entire region containing multiple AR enhancers as well. Now, I only have 20 minutes, so I'm going to kind of jump to my next topic. And really, our ultimate goal, the ultimate goal of my lab is to integrate clinical and functional genomics to study disease processes. And so I've shown you some of our studies in the context of sequencing and clinical samples, including whole genome, whole genome bisulfide, deep RNA-seq. But I'm also now going to transition to the studies we've done in functional genomics. And we've been interested in studying specific questions, one of which is, what drives androgen receptor expression and signaling? And really, again, the long-term goal is to integrate clinical and functional genomics to better understand the AR signaling access. So to study androgen receptor biology, we developed an endogenous AR reporter in a prostate cancer cell line. And I should say that this work was led by Haolong Li, who's a postdoc fellow in my lab, and he's co-mentored by Luke Gilbert, one of my close collaborators. And so what Haolong did was he employed a split-fluorescent approach that's previously been developed by Bo Han's lab at UCSF. And this approach basically utilizes the Neon Green 2 system. NG2 is the abbreviation. And so what happens is that you can take the Neon Green 2 protein, and you can split it into two parts. And the first part is encoded by exons 1 through 10, and the second part is encoded by exon 11. Neither of these two components by themselves glow green, but what happens is that when you put them together, they form a non-covalent but quite stable bond that results in a fluorescent protein. And so what we did was we used CRISPR approaches to knock in the 11th exon of Neon Green 2 into the N-terminus of the androgen receptor. And then when you co-express that with the first to 10th exons of Neon Green 2, you now get fluorescent androgen receptor. And the reason we use this system is because the 11th exon is only 16 amino acids long in terms of the tag, and therefore it's minimally disruptive to protein folding. And so here you can see images using our reporter system. On the top left is basically our control cells treated with DMSO. What happens next is we treat with the synthetic androgen R1881. You get increased nuclear signal. We then treat with enzalutamide and next generation AR antagonist, and you can see loss of nuclear signal, and you even see greater loss of nuclear signal when we treat with AR degrader ARD61. And here are some live images from our cells. So this is a cell we just treated with testosterone, and you get nice green nuclear signal consistent with a nice AR expression in the nucleus. And then what happens over time is we're going to subsequently treat with an AR degrader ARD61. Here you see that being applied, and you lose it. You lose the nuclear signal of AR, and in fact, the cell line starts to die. And this is consistent with what we want to see in an AR reporter system. And so what we did with this reporter system was we first characterized it very extensively to make sure that there was no significant differences between this reporter line and its parental line. Again, we just didn't want to change the underlying AR signaling too much. So we did chip assays looking at AR binding to PSA enhancer and actually multiple other AR targets, AREs as well, and we see that there was no big difference between the parental line and the reporter line. We also did RNA-seq of the reporter line and its parental line, and here you can see that the transcriptomes of both cell lines are highly concordant. We then also measured the mRNA half-life of AR using actinomyosin D studies, and what you can see here is that no significant differences in mRNA levels. Again, green is our fluorescent cell line, black is our wild-type cell line. We looked at AR protein half-life with cyclohexamide studies, and again what you can see here is that over time there doesn't seem to be any major differences in the protein half-life of AR in our reporter cell line compared to the parental cell line. And I can show you many other studies that we did showing that this reporter line kind of recapitulates kind of AR signaling and biology compared to the parental line. We then used a genome-wide CRISPR eye screen using the cell line to identify regulators of antigen receptor expression, and so here you can see our reporter cell line. We infect with a lentiviral genome-wide guide RNA library, and then we basically sort for the cells with high AR expression versus low AR expression, and here you see a volcano plot basically summarizing our results. So on these volcano plots, genes that are to the top and left are genes in which when you knock down decreases our AR reporter signal. Genes on the top right are genes which when you knock down increase our AR reporter signal. So I'm going to focus on the ones on the left, and what's very comforting is that when we did a genome-wide screen, the top hit on the on the top left is AR itself, and that's very reassuring because knockdown of AR should really be the top hit in terms of looking at loss of signal from an AR reporter. What's also comforting is that, you know, our other top genes, HoxB13, GATA2, Granny Headlight2, are known AR coactivators. What I'm going to focus on now is this fifth gene, PTGS3, which stands for prostaglandin E synthase 3, which is less characterized compared to some of these other coactivators, and so what we show is that repression of PTGS3 with guide RNAs decreases both RNA and protein levels of AR similar to canonical AR regulators, and so here on the left, and actually on the right, I'm going to focus on the right, you can see western blocks for AR when we knock down AR, Granny Headlight2, HoxB13, and PTGS3 using our decastine CRAB system. We then generalized this finding across multiple cell lines and showed that knockdown of PTGS3 decreases AR protein levels in multiple settings, so on the left, you can see data from the C42B cell line, which harbors an AR mutation. Then we have the MR16D cell line that's enzalutamide resistant, generated by the Vancouver group. We have 22 RV1 cells that harbor AR splice variant and VCAP cells that have AR amplification, and when we use, in this case, siRNAs to knockdown PTGS3, you can see that there's nice corresponding knockdown of AR protein expression as well. We then knockdown PTGS3 and did a gene set enrichment analysis to look at things that came out, and the top theme that emerged from knockdown of PTGS3 was an angiotensin response signature, and here you can see the expression of a variety of AR downstream genes basically goes down when we knockdown PTGS3. Now, we also wanted to see what the effect was of knocking down PTGS3 on proliferation and other oncogenic phenotypes, and here I'm showing you proliferation data, where, again, knockdown of PTGS3 with two different siRNAs shown in red decreases proliferation of a variety of different AR positive cell lines, but if you look at panel E, it doesn't seem to have much of an effect in PC3 and DE145 cells, with both of which are AR negative. And then on panel F, you can see that this results with a xenograft tumor that has an inducible teton system, and when we knockdown PTGS3 with this inducible system, you can see decrease in tumor volume over time. Now, PTGS3 has been reported to have both a chaperone function as well as an enzymatic function. The chaperone function, it's been reported potentially to bind to angioreceptor. The enzymatic function, it's been reported to produce PGE2 from PGH2, and what's also been shown is that there are specific point mutations that can be made that disrupt each of these functions, and so here you can see the Y9N mutation has been shown to decrease enzymatic function. The W106A mutation decreased the chaperone function, and so what we did basically was we created inducible models in which we could induce the expression of either these two point mutations or wild-type AR, and when we do this in the context of knocking down wild-type PTGS3 and then basically inducing the expression of either the wild-type again or the point mutations, you can see that induction of the point mutations here decreases AR expression compared to induction of expression of wild-type PTGS3, and what this suggests is that potentially both the chaperone and the enzymatic function are playing a role in maintaining AR expression, and again, we've done a lot of other mechanistic studies which I can't describe in the interest of time, but one thing that's interesting is that PTGS3 expression appears to be a biomarker of resistance to AR-directed therapies, and so this is data from prostate cancer patients who were treated with surgery followed by androgen deprivation therapy on the left versus surgery without androgen deprivation therapy on the right. You can see on the right, PTGS3 expression is not associated with outcome, but when the patients are treated with androgen deprivation therapy, now you can see high PTGS3 expression is associated with worse outcomes, and I should point out that all these patients were matched for clinical and pathologic features as well. Now, by interrogating the DEPMAP database, what we found was that PTGS3 is not essential for survival in most cell lines, most cancer and normal cell lines, and this is quite important because this suggests that it could be a good therapeutic target, and so in collaboration with Kayvon Shokat and Jim Melnick from UCSF, we're starting to develop lead compounds to target PTGS3. This is a bisulfite tethering screen, and this approach allows us to identify scaffolds that bind to the PTGS3 catalytic site. We've already isolated five promising compound fragments that have actually reasonable binding affinity, and now we're basically optimizing on top of that, but we hope to develop lead compounds here. So, in conclusion, we are using genome-wide sequencing approaches to map out the genomic and epigenomic landscape of castrate-resistant prostate cancer with an emphasis on looking at AR. We've optimized genome-wide CRISPRi screens in MCRPC cell line models to study AR and other drivers of disease progression, but our long-term goal is to integrate our sequencing data from clinical samples with our functional genome-wide screens, but we still have a lot of work to do. I'd like to thank the folks from my lab who really led a majority of these studies. David Quigley was the first author of our cell paper. George Zao was first author of our Nature Genetics paper, and Haolan Li really contributed to developing that AR reporter model in the CRISPR screens. Luke Gilbert is my close collaborator. I'd like to thank the entire SCAM2Cancer West Coast Dream Team and my funding support. Thank you for your time. Thanks, Felix. So, we have a few questions here, so I'll get started reading them. Scott Dam has the first one. Great talk, Felix. A modest two-fold overexpression of AR in LN-CAP cells can drive a CRPC phenotype. Does overexpression of your tagged AR and LN-CAP cells drive this CRPC phenotype? Yeah, Scott, great question. So, you know, we actually selected down to a single clone to match based on total AR expression level between the parental line and the reporter line, and so there's no significant differences in kind of AR expression, so there's not really overexpression. That being said, I would say that the reporter line is slightly, just slightly more resistant to charcoal strip media than the parental line. Right. The next question is from Jay Gertz. He says, great talk. Were there any interesting hits for genes that, when knocked down, increased AR expression? In other words, genes that work to reduce AR. Yes, there were. There were actually plenty of hits on that that far right side as well. We're still working them up. If you ask me, Jay, which ones they are, to be honest with you, I don't quite remember the top hits because we were so focused on genes on the left side of the volcano plot, which, when you knock down, decrease AR expression. But, you know, Jay just reached out to me offline. I'm glad to share those hits with you. The next question is from Jin Dan Yu. Says, great talk. Interesting to see that HOXB13 was as a major inducer of AR expression. Do you think this may explain dual HOXB13 and AR loss in any PC? You know, Jin Dan, it's a great question. It could. For sure, it could, in the sense that our functional data suggests that knocking down HOXB13 decreases AR expression as well. Whether that leads to any PC phenotype per se, I don't think we're positioned to answer that question based on the studies that we've done. But certainly, I think that we do show this relationship between HOXB13 and loss of AR expression, but others have shown that in the past as well. And we have another question from Scott. HSP90 inhibitors similarly reduce protein expression of AR and AR variance. Can PDGS3 knockdown still downregulate AR expression in the presence of HSP90 inhibitors or HSP90 knockdown? Yeah, you know, Scott, it's a great question. To be honest with you, I've stayed away from HSP90 inhibitors because they're relatively toxic, and so we haven't done the experiment you proposed. But certainly, in order to study mechanism, I think it's a worthwhile study to do, but we just haven't done it yet. I have just a simple question for you, Felix. You know, when you're looking at all these metastatic patient samples, you know, we know that when you manifest with these metastases, you get multifocal metastases all over the place, right? Do you ever see samples where you have different foci from the same patient that you can look at to compare? Yeah, we're starting to do those types of studies. Honestly, it's really hard to get the metastatic biopsies because the biopsies aren't the clinical standard of care for these patients in the sense that, you know, if the PSA level goes up clinically and, you know, you do any type of imaging scan and you see a new spot, you just assume that that's a prostate cancer metastasis. And so, these metastatic biopsies have been obtained mostly through research. And so, in the vast, vast, vast majority of time, we get, you know, just one site from a patient, if we're lucky. And I think we've been lucky in about 30 patients, we can get temporally distinguished biopsies, meaning one before treatment and one after treatment at the time of resistance. But the real way to get multiple metastases from the same patient is from the autopsy studies. The issue with the autopsy studies is that, you know, oftentimes, these samples aren't immediately harvested for, you know, reasons that are self-explanatory. And so, you can get like hypoxia signatures on the RNA expression. I think the DNA itself is, you know, is fine. I just don't trust the RNA as much, but we're trying to, you know, see whether we can obtain relatively fresh samples as well to try to look at differences between, let's say, a liver metastasis and a bone metastasis. Jen Ricker asks if GR, if you know if GR is also a target of PDG ES3? Yeah, so Keith Yamamoto back in the day. So, the other name for PDG ES3 is P23. And so, Keith Yamamoto back in the day actually showed that PDG ES3 influences GR expression as well. So, that's been shown. So, it may have some effects on stromal components as well. It might, yep. All right. Well, thank you very much. I want to thank all three speakers for their excellent talks and their contributions to our field in general. They are all continuously contributing. And with that, we'll go ahead and close this session and thank all of the participants. And we'll see you tomorrow at ENDO 2021. Thank you.

Wilbert Zwart

androgen receptor enhancers

prostate cancer development

epigenetic biomarkers

MCRPC

neoadjuvant therapy

therapy resistance

genomic sites

hormone-dependent cancer

predicting therapy response

intra-tumor heterogeneity

cancer prognosis