BRD9-SMAD2/3 orchestrates stemness and tumorigenesis in pancreatic ductal adenocarcinoma

The dismal prognosis of pancreatic ductal adenocarcinoma (PDAC) is linked to the presence of pancreatic cancer stem-like cells (CSCs) that respond poorly to current chemotherapy regimens. By small molecule compound screening targeting 142 epigenetic enzymes, we identified that bromodomain-containing protein BRD9, a component of the BAF histone remodelling complex, is a key chromatin regulator to orchestrate the stemness of pancreatic CSCs via cooperating with the TGFβ/Activin-SMAD2/3 signalling pathway. Inhibition and genetic ablation of BDR9 block the self-renewal, cell cycle entry into G0 phase and invasiveness of CSCs, and improve the sensitivity of CSCs to gemcitabine treatment. In addition, pharmacological inhibition of BRD9 significantly reduced the tumorigenesis in patient-derived xenografts mouse models and eliminated CSCs in tumours from pancreatic cancer patients. Mechanistically, inhibition of BRD9 disrupts enhancer-promoter looping and transcription of stemness genes in CSCs. Collectively, the data suggest BRD9 as a novel therapeutic target for PDAC treatment via modulation of CSC stemness.


Introduction
Pancreatic cancer with its most common type, pancreatic ductal adenocarcinoma (PDAC), is one of the most lethal human malignancies [1][2][3] . It has an overall median survival time of 6-9 months and a similarly 5-year survival rate of 6%, making it currently the fourth leading cause of cancer-related deaths in western countries 4 5 . Due to the increasing incidence of risk factors including obesity and other metabolic traits, pancreatic cancer is projected to overtake colorectal, breast and prostate cancer and become the second leading cause of worldwide cancer-related deaths by 2030 6 . The disease owes its exceptional level of lethality to multiple factors. Pancreatic cancer often has no early symptoms and presents itself in an advanced stage at diagnosis: only 20% of newly diagnosed pancreatic cancers are amenable to surgery 7 . In turn, the disease's poor response to chemotherapy and radiotherapy results in disease re-emergence for 90% of the surgically treated patients 8 .
Treatment options for PDAC are limited and inefficient. Gemcitabine, an antimetabolite drug of the nucleoside analogue class is the standard of care in PDAC therapy 8 . It is currently the only approved single-agent drug for pancreatic cancer, with still modest improvements in survival rate 9 , whether the drug is administered alone or in combination with adjuvant drugs such as the EGFR inhibitor erlotinib 10 or the tubulin-targeting drug Nab-paclitaxel 9 . A combination therapy such as FOLFIRINOX is superior to gemcitabine-based regimens in restraining the progression of metastatic PDAC but has lower tolerability 11 .
Precancerous lesions and dedifferentiation of the cells to a progenitor-like or stem cell-like state with increased cellular plasticity frequently occur during pancreatic tissue transformation 4 12 . A distinct cell population often referred to as cancer stem cells (CSCs), seems to acquire a stem cell-like state partially resembling naturally occurring stem cells [13][14][15] . This phenotype allows them to give rise to the whole tumour with its entire cellular heterogeneity and thereby supports metastases formation and development of resistance to current cancer therapeutics. The existence of developmentally plastic cancer stem cells has been discovered in the brain, breast, colon, oesophagus, liver, lung, ovarian, prostate, stomach and thyroid cancers, among others. In the case of PDAC, the first reports of cancer stem cells date back to 2007 13 14 . Since then, pancreatic CSCs have been conclusively shown to be involved in PDAC resistance to chemotherapy, displaying increased prevalence within the tumour after treatment with gemcitabine 16 17 . Annihilating CSCs is thus emerging as an essential aim of PDAC therapeutics. CSCs are thought to have specific epigenetic mechanisms 18 that regulate their self-renewal, and the formation of CSCs has been postulated to occur as a result of epigenetic events 19 . Accordingly, cancer epigenetics has established itself as a promising area of oncology research 20 .
After a decade in which epigenetic cancer drugs were approved only for haematological malignancies, the first FDA approval for an epigenetic drug targeting solid tumours was granted in 2020 for an EZH2 inhibitor. Despite their paradigm-shifting novel mechanism of action, early results of epigenetic modulators have identified the need for better selection of targets, improved intratumoral drug penetration and elimination of CSCs.
Pancreatic cancers are complex tumours with significant heterogeneity in their molecular and cellular make-up that are controlled by various signalling pathways that crosstalk with epigenetic regulators. Among these pathways is the TGFβ/Activin/Nodal-SMAD2/3 pathway. This developmental signalling pathway plays a central role in early development by regulating the self-renewal of human pluripotent stem cells (hPSCs), Epithelial-to-mesenchymal transition (EMT) and pancreatic tissue homeostasis [21][22][23] . The TGFß/Activin/Nodal-SMAD2/3 pathway regulates epigenetic mechanisms, for instance by cooperating with the core pluripotency protein NANOG and epigenetic modifiers such as DPY30-COMPASS to control pluripotency and differentiation of hPSCs 24 . Aside from the key role of TGFβ/Activin/Nodal-SMAD2/3 in pluripotent stem cells and developmental processes, this signalling pathway is directly involved in the formation of PDAC 22 and is frequently deregulated in PDAC 25 26 .
The function of the pathway in PDAC is particularly interesting, because it confers dedifferentiated stem cell-like features to CSCs in PDAC 15 , although the underlying mechanisms are still largely unknown.
Our hypothesis was that epigenetic or chromatin-templated mechanisms are essential to main CSC properties and that inhibition of these mechanisms may provide a path forward to modulate CSC phenotypes. Using a focused compound library of epigenetic inhibitors we performed a small molecule compound screening and identified the BAF chromatin complex component BRD9 as a critical regulator of CSC behaviour. Our results indicate that BRD9 is an attractive therapeutic target for specifically eliminating CSCs in PDACs.

Development of a screening platform to target PDAC CSCs.
Since PDAC cells comprise a heterogeneous population of CSCs and non-CSCs, we first decided to establish a suitable small molecule screening platform in PDAC cells ( Figure 1A). To uncover novel regulators of CSCs, we used three CSC markers (OCT4, CD133, SSEA4) for identifying stem cell-like cells in the experiments. OCT4 is a transcription factor crucial to the selfrenewal and pluripotency of embryonic stem cells 27 and is associated with worse outcomes in PDAC 28 . Besides OCT4, we used two other CSC markers: surface protein CD133 and glycolipid carbohydrate epitope SSEA4. A high expression of CD133 in PDAC patients (over 5%) has been linked to a 0% 5-year patient survival rate, compared to 23.5% for CD133-negative tumors 29 , and PDAC metastasis was found to depend upon a subset of CD133-expressing cells 13 which trigger early metastasis to lymph nodes 29 . Finally, the third stem cell marker, stage-specific embryonic antigen-4 (SSEA4), is known to be expressed in a wide range of cancers, including PDAC 30 . The pancreatic cancer cells used for this screening were genetically engineered, with the sequence coding for green fluorescent protein (eGFP) integrated in the endogenous locus via TALEN-mediated recombination, resulting in controlled expression of a OCT4-eGFP fusion protein driven by the endogenous OCT4 promoter (Supplemental Figure 1A). To validate the importance of CD133, SSEA4 and OCT4 as markers for the CSC population in our PDAC cells, we treated the cells with chemotherapy reagents gemcitabine (GEM), paxclitaxel (PAX) and 5-fluorouracil (5-FU) that are currently in clinical use as PDAC patient therapeutics. Gemcitabine, paxclitaxel and 5-fluorouracil treatment of PDAC cells for 5 days enriched for cells expressing OCT4-GFP, CD133 and SSEA4 from ~1% triple positive cells to ~75%, ~25% and ~60%, respectively (Supplemental Figure 1B), by eliminating most of the PDAC cells that do not express these three CSC markers (Supplemental Figure 1C). This results in selective survival and enrichment of the rare OCT4-GFP+/CD133+/SSEA4+/EPCAM+ CSCs.
Since the TGFβ/Activin/Nodal-SMAD2/3 developmental signalling pathway regulates pluripotency via epigenetic regulatory complexes and impacts stem cell-like characteristics of CSCs 24 31 , we also tested the impact of TGFβ/Activin signalling on CSC resistance to currently used chemotherapeutics by treating the cells with the TGFβ/Activin signalling inhibitor SB431542 in combination with Gemcitabine, Paxclitaxel and 5-FU (Supplemental Figure 1B-C). Inhibition of TGFβ/Activin signalling strikingly reduced the chemoresistance of PDAC cells as indicated by reduced numbers of OCT4-GFP+/CD133+/SSEA4+ CSCs (Supplemental Figure 1B) and therefore the overall number of surviving PDAC cells (Supplemental Figure 1C). These results emphasize the crucial importance of TGFβ/Activin signalling on CSC maintenance and their elevated chemoresistant characteristics. Furthermore, PDAC patients with high expression of CD133 and OCT4 have lower overall survival (Supplemental Figure 1D) and lower disease-free survival (Supplemental Figure   1E). These results underline the suitability of our screening platform and the translational relevancy of this preclinical model for compound screening.
A focused compound library screen of epigenetic inhibitors identifies BRD9 as an important factor for governing CSC characteristics.
Due to the importance of epigenetic pathways in tumorigenesis, we hypothesized that the formation and maintenance of pancreatic CSCs are controlled by epigenetic mechanisms, and these could be used as therapeutic targets for eliminating CSCs. To address this, we performed a focused library screen consisting of validated small molecule inhibitors (Supplementary Table 1) targeting epigenetic regulators such as 'readers, writers and erasers' of a histone code 32 . These experiments aimed to identify molecular targets of small molecule compounds which specifically affect CSC marker-expressing cells (Figure 1A-C). We measured CSC marker expression, cell growth and cell death by flow cytometry (CD133+/OCT4+/SSEA4+) after incubating the cells with the compounds for 5 days, which allowed us to identify effective compounds that impact CSC marker expressing populations while also detecting cells that do not express these CSC markers. The compound library used in our experiments consisted of 142 compounds that have been verified to be active and targeting specific epigenetic modifying enzymes ( Figure 1C). A substantial fraction contained in the library were acetyl lysine (bromodomain) inhibitors that represented approximately 26% of the compounds, followed by histone deacetylase (HDAC) inhibitors (14%), Histone methyltransferase inhibitors (11.3%), compounds targeting methyl lysine binders/tudor domains (9.9%), arginine methyltransferases (9.2%) and lysine demethylases (8.5%). Overall, this screening identified compounds that significantly and reliably reduced the relative percentage of triple+ marker (CD133+/OCT4+/SSEA4+) expressing CSCs and also reduced cancer cell survival ( Figure 1C). Our study identified the BET bromodomain inhibitors I-BET 672 and JQ1 as candidate compounds, both of which have previously shown antitumour activity 33 and synergy with gemcitabine 34 , confirming the suitability of our screening platform.
Importantly, the screening identified novel compounds that target distinct epigenetic regulatory components and strikingly reduced the percentage of CSCs. Among the top candidate compounds with a distinct effect on CSCs we identified BRD9 inhibitors (I-BRD9, dBRD9, TP-472, BI-7273, LP99) while the corresponding inactive negative control compound (TP-472N) did not reduce CSC marker expression ( Figure 1C-D). These results suggested that BRD9 inhibition could induce differentiation of therapy-resistant CSCs into a more therapy-responsive population and thus could possibly sensitize PDAC cancers to conventional therapy. BRD9 is a component of the non-canonical BAF chromatin remodelling complex (ncBAF) 35 and has recently been shown to constitute a barrier in reprogramming of somatic cells to induced pluripotent stem cells 36 . The catalytic component SMARCA4 of the BAF complex regulates stem cell properties in pediatric gliomas and provides a possible therapeutic angle in glioblastoma treatment 37 . Collectively these data provided evidence to focus our investigation on the molecular role of the BAF complex and BRD9 in particular, in PDAC CSCs.
The bulk cell population of the PDAC is sensitive to chemotherapy while the CSC subpopulation of PDAC cells is the reason why PDAC is therapy recalcitrant. Therefore, we investigated the effects of compounds on CSCs by 3D tumour sphere assays. To study the impact of BRD9 inhibition on CSC self-renewal capacity we performed tumour sphere assays with three different BRD9 inhibitors (I-BRD9, TP-472, dBRD9) in five different PDAC cell lines and primary cells from surgically resected PDAC tumours ( Figure 1H). Inhibition of BRD9 with these three inhibitors showed a significant reduction in CSC sphere numbers while the negative control compound of TP-472 (TP-472N) did not have such effects in any of the cell lines or tumour-derived cells. To investigate the effect of BRD9 inhibition by genetic means, we performed stable BRD9 knockdown (KD) in two PDAC cell lines (Supplemental Figure 1H-K). Tumour sphere assay revealed a reduction of spheres in both cell lines by two different shRNA constructs (Figure 1J), indicating that BRD9 loss-of-function via knockdown phenocopies the reduction of CSC self-renewal as seen with chemical inhibition of BRD9 ( Figure 1K). BRD9 inhibition reduces pancreatic CSCs entry to G0 cell cycle phase and chemoresistance.
Since CSCs have lower sensitivity to chemotherapeutics besides their self-renewing capacity, we investigated the impact of BRD9 inhibition on the chemoresistance of CSCs in combination with Gemcitabine treatment (Figure 2A). BRD9 inhibition together with Gemcitabine treatment led to a stronger reduction in CSC sphere formation compared to either I-BRD9 or Gemcitabine treatment alone, suggesting that BRD9 inhibition sensitizes CSCs for Gemcitabine-mediated cell killing. The TGFβ/Activin pathway inhibitor SB431542 reduced the number of CSCs compared to Activin A treated cells (Supplemental Figure 1L-M), indicating that the TGFβ/Activin pathway promotes CSC characteristics. To understand whether BRD9 inhibition engages in crosstalk with TGFβ/Activin signalling, we co-treated the cells with I-BRD9 and Gemcitabine along with Activin A or the TGFβ/Activin pathway inhibitor SB431542 during tumour sphere formation. BRD9 inhibition as well as TGFβ/Activin pathway inhibition strikingly increased the sensitivity of CSCs to Gemcitabine (  Figure 1N). The effect of BRD9 inhibition on sensitizing cells to Gemcitabine was further validated by three different BRD9 inhibitors (Supplemental Figure 1O). Furthermore, BRD9 genetic knockdown in two different PDAC lines reduced CSC chemoresistance not only for Gemcitabine but also 5-FU (Supplemental Figure 1P-R).
Due to the plasticity of the cell state, CSCs and non-stem cancer cells have been proposed to form a dynamic balance between differentiation and dedifferentiation. This process could be mediated by an epigenetic plasticity of the cells through the erosion of epigenetic barriers that would otherwise keep cells in a more constrained cell identity. Since we observed that BRD9 treatment abolished CSC self-renewal, we decided to study whether BRD9 inhibition could impact the re-emergence of CSC marker-expressing cells. To study this aspect, we sorted cells into CSC marker OCT4-GFP/CD133/SSEA4 negative population and analysed the re-emergence of these CSC triple markerexpressing cells upon temporary or continuous inhibition of BRD9 ( Figure 2E). In the steady-state cell culture, the population contains approximately 1.5% OCT4-GFP+/CD133+/SSEA4+ CSCs. Control cells treated with DMSO showed the beginning of the re-emergence of OCT4-GFP+/CD133+/SSEA4+ CSCs after 3 days and reached the steady-state level similar to the untreated and sorted cells after 2 weeks. Pre-treatment of cells with I-BRD9 for 5 days significantly slowed down the re-emergence of OCT4-GFP+/CD133+/SSEA4+ CSCs, while continuous I-BRD9 treatment of cells had an even more striking effect similar to continuous treatment with TGFβ/Activin pathway inhibitor SB431542. These data suggest that BRD9 inhibition could promote CSC differentiation or decrease cellular plasticity by increasing the epigenetic barriers that are necessary for the cells to dedifferentiate from non-stem cancer cells to CSCs.
The fluorescent ubiquitination-based cell cycle indicator (FUCCI) system is a powerful tool to assess cell cycle-dependent responsiveness to drugs and the effect of drugs, gene silencing or activation on the cell cycle without the need for synchronisation 39 . We have previously used the dualcolour FUCCI system in hESCs 40 . The basic dual-colour FUCCI system uses truncated hCdt (DNA replication licensing factor) and Geminin (inhibitor of hCdt) to detect cells in G1, S and M-phases 40 41 .
More recently, the presence of the cell cycle inhibitor p27 with distinct site-specific mutations has been shown to characterise G0 cells 42 . To study the effects of BRD9 epigenetic machinery inhibitors on CSC proliferation by small molecules we developed an advanced three-colour FUCCI system (RFP-hCdt1(30/120)_mAG-hGEM(1/110)_mKate2-p27(mut)) by combining the truncated hCdt1, truncated hGeminin and p27K(-). We established pancreatic ductal adenocarcinoma cell lines by TALEN-mediated targeting of the three-colour FUCCI construct into the AAV1 locus for stable genomic integration. This FUCCI system enables distinguishing between the cells in early G1, late G1, early S, S/G2/M and G0 phases (Figure 2F-H). We first aimed to determine the cell cycle kinetics in FUCCI cells after BRD9 inhibition. We treated the CSC-FUCCI spheres from two different PDAC cell lines for 5 days with I-BRD9, Gemcitabine or a combination of I-BRD9 and Gemcitabine, and performed flow cytometry analysis in two PDAC cell lines (Supplemental Figure 2A-D). In both cell lines, Gemcitabine treatment led to an increased fraction of cells positive for Cdt1-mRFP and p27k(-)-mKATE2 signals which marks G0-phase cells. Of note, the accumulated expression of p27k(-)-mKATE2 is not high in G0, which could indicate a shallow and transient quiescence with cell entry into the G0 condition while maintaining its readiness to re-enter the cell cycle phase due to KRAS activation 43 . In contrast, I-BRD9 treatment led to a reduction in G0-phase cells and also blocked cells from accumulating in the G0-phase upon Gemcitabine treatment ( Figure 2I). This suggested that BRD9 inhibition reduces the capacity of cells to enter the G0 phase and this could be particularly useful in eliminating CSCs upon combination treatments with DNA damaging reagents such as Gemcitabine that prevent the cells from escaping genotoxic insults through temporarily dormant or quiescent states by entering the G0 phase.
PDACs are highly aggressive cancers in part due to the ability of cancer cells to migrate and initiate metastatic processes through their invasive capacities. Since BRD9 inhibition led to a reduction in CSC self-renewal and chemoresistance, we decided to investigate its impact on other key characteristics of PDAC cells such as migration capacity and invasiveness. Wound-healing assays with BRD9 genetic knockdown, chemical inhibition and protein degradation indicated reduced migration of PDAC cells compared to control cells ( Figure 2J-K). Importantly, BRD9 knockdown also reduced PDAC cell invasiveness by trans-well assays in two separate PDAC lines ( Figure 2L-M), indicating that BRD9 inhibition could be useful for reducing the metastatic capacity of PDACs since BRD9 inhibition decreases PDAC cell motility and invasiveness.
Collectively, chemical inhibition and knockdown of BDR9 blocked the self-renewal of CSCs, reduced CSC invasiveness and re-sensitized PDAC CSCs to conventional chemotherapeutic reagents suggesting that BRD9 could be an attractive therapeutic target in PDAC.
Our previous experiments indicated that BRD9 as well as the TGFβ/Activin-SMAD2/3 signalling pathway both regulate the characteristics of CSCs and PDAC cell invasiveness. This led us to hypothesize that BRD9 and TGFβ/Activin-SMAD2/3 signalling pathway could cooperate in CSCs at the molecular level by protein-protein interactions. Therefore, we proceeded with identifying the binding partners of SMAD2/3 proteins in pancreatic CSCs by performing SMAD2/3 coimmunoprecipitation followed by mass-spectrometry ( Figure 2N). This unbiased proteomic approach identified SS18/SSXT, BAF180/Polybromo-1 and ARID1A peptides indicating that these proteins could be co-factor candidates of SMAD2/3 ( Figure 2O). STRING analysis of protein interactions confirmed that these three proteins are all part of the ATP-dependent chromatin remodelling complex BAF (BRG1/BRM-associated factor) that corresponds to the mammalian SWI/SNF complex (Supplemental Figure 2E). The BAF complex has tissue-specific functions that arise from the combinatorial assembly of distinct subunits such as an ES cell-specific BAF (esBAF), the neural progenitor BAF (npBAF) and neuronal BAF (nBAF), as well as cardiac progenitor-enriched (cBAF) and hematopoietic stem cell BAF 44 . More recently, a non-canonical (ncBAF, also called GBAF) complex has been described 45 46 . However, BRD9 protein is a common subunit of esBAF, npBAF, nBAF, GBAF, cBAF and the BAF complex observed in hematopoietic stem cells. Therefore, we investigated further the distinct complex composition of BAF complexes that interact with SMAD2/3 (Supplemental Figure 2F). SMAD2/3 co-immunoprecipitation experiments indicated an interaction of SMAD2/3 with general subunits SS18 and BRD9 (Figure 2P), and subunits that are specific for esBAF (BCL11a), npBAF (BAF180) and ncBAF (GLTSCR1).
These results suggest that SMAD2/3 forms a complex with several of the distinct BAF complexes (Figure 2R), which have all BRD9 enzyme as a subunit, and suggest that SMAD2/3 could be involved in targeting BRD9 as a subunit of the BAF complexes to distinct chromatin regions to regulate the chromatin landscape and gene expression in pancreatic CSCs.

Inhibition of BRD9 suppresses in vivo PDAC tumour formation.
To study the effect of BRD9 on PDAC cell growth in vivo, xenograft PDAC mouse models were developed by subcutaneous injection of CSCs from the PDAC line A13A with different transduction/treatment regimes (Supplemental Figure 3A). At 2 months post inoculation, only mice implanted with control adenovirus transduced cells (containing scrambled control shRNA, A13A Null ) developed tumours. At this point, no tumours were observed in other mice implanted with Gemcitabine-treated cells (A13A Gemcitabine ), shRNA-BRD9adenovirus-transduced cells (A13A BRD9-KD ), or A13A BRD9-KD plus with Gemcitabine treatment (A13A BRD9-KD + Gemcitabine ), indicating the slower growth rate of these cells in contrast to A13A Null . After 3 months of cell inoculation, although no tumours were observed in mice with A13A BRD9-KD + Gemcitabine , the other mice implanted with A13A Null , or A13A Gemcitabine , or A13A BRD9-KD developed tumours respectively. To further characterize these tumours, calliper measurement and PET- 18 F-FDG were performed to analyze the size and metabolic activity of these tumours. As expected, the A13A Null tumours grew much faster than other tumours from month 3 to month 4, as evidenced by the significantly increased tumour volume (Supplemental Figure 3B).
Although tumours formed by A13A Gemcitabine or A13A BRD9-KD were detected at month 3, they were moderate and grew slowly from month 3 to month 4. Importantly, BRD9-KD further delayed tumour growth compared with A13Al Null and A13A Gemcitabine groups. These results were consistent with the observation obtained from PET- 18 F-FDG. At month 4 post inoculation, the 18 F-FDG uptake in tumours formed by A13A Null was much higher than in other tumours (Supplemental Figure 3C). Gemcitabine significantly decreased the  in A13A Gemcitabine tumours compared with A13A Null tumours, and BRD9-KD further reduced the 18 F-FDG uptake of A13A BRD9-KD tumours compared with Gemcitabine treatment. Consistently, the tumours harvested from A13A BRD9-KD -implanted mice at month 4 were markedly smaller compared with tumours from A13A Null or A13A Gemcitabine implanted mice (Supplemental Figure 3D). The lower tumour-weight to body-weight ratios of A13A BRD9-KDtransplanted mice further demonstrated a progressive decrease in tumour mass (Supplemental Figure 3E). In contrast, no tumours were detected in any mice with A13A BRD9-KD + Gemcitabine and these mice were alive without any signs of sickness at month 4. When we further extended the time window to 6 months, only one of 6 mice in the A13A BRD9-KD + Gemcitabine groups developed a small tumour (data not shown). At necropsy, no metastases were observed among these tumour-bearing mice and nontumour bearing mice. Histological examination of these tumour sections by H&E staining found prominent heterogeneity and extensive necrosis in A13A Null tumours than A13A Gemcitabine tumours, whereas A13A BRD9-KD tumours only showed a small area of necrosis (Supplemental Figure 3F).
These data suggested that BRD9 activity drives the tumour progression through regulating PDAC cell growth.
Although BRD9 inhibition suggested tumor progression through regulating PDAC cell growth in our cell-line xenografts, the cell-line xenografts do not accurately recapitulate the histopathological and molecular characteristics of the human parental tumor. Therefore, human xenograft PDAC models (PDX models) were employed to further evaluate the therapeutic potential of BRD9 inhibitor in vivo.
Human tumors from a PDAC patient were implanted into NOD.Cg-Prkdc scid Il2rgtm1Wjl/SzJ (NOD Scid gamma) host strain. Following transplantation, various treatment schedules were initiated when the tumors reached an average size of 150 mm 3 with a repeated injection schedule (once every 3 days for a total of six treatments within 15 days) ( Figure 3A). As expected, Gemcitabine (Gem) or 1 combination with BRD9 inhibitor (IBRD9 + Gem) significantly delayed the tumor growth as evidenced by the tumor growth curves and tumor volume ( Figure 3B) when compared with DMSO treated control group (Ctrl). In contrast to the Gem group the most striking tumor regressions were observed in the IBRD9 + Gem group. Notably, from day 27 to day 42 after completing treatments, the tumors in the Gem group recurred with rapid growth, whereas most of the xenograft tumors in IBRD9 + Gem group grew slowly over a prolonged period without significant changes in tumor volume ( Figure 3B).
To monitor the treatment response and tumor progression in living mice, PET-18 F-FDG was performed ( Figure 3C). In the contol groups, higher metabolic activity associated with higher 18 F-FDG uptake was found in these highly proliferative tumors. In contrast, treatment with either Gem or IBRD9 + Gem decreased levels of tumor 18 F-FDG uptake. In particular, 18 F-FDG uptake in the IBRD9 + Gem group was significantly lower than in any other group ( Figure 3C) indicating that the combination of IBRD9 + Gem significantly delayed the tumor growth. Consistent with reduced tumor growth, the tumor size and tumor-weight to body-weight ratios in IBRD9 + Gem group were significantly lower than those originating from Ctrl and Gem treatment groups ( Figure 3D-E). Importantly, the combination therapy with IBRD9 did not show any toxic effects as no significant alteration of body weight was observed in IBRD9 + Gem treated mice (data not shown). Histology analysis of these tumor sections revealed a larger necrosis area in Ctrl and Gem groups as compared with IBRD9 + Gem groups (Figure 3F), indicating rapid growth of these tumors in Ctrl and Gem groups. In line with these findings, Ki67 + proliferating tumor cells were substantially increased in Ctrl and Gem groups, whereas they were reduced in IBRD9 + Gem treated tumors ( Figure 3G-H), suggesting a potential additive effect of BRD9 in controlling primary tumor growth.

BRD9 inhibition eliminates CSCs from patient tumour.
CSCs are particularly challenging to eliminate due to their chemoresistant phenotype. Our  Figure 4C). The combined treatment of I-BRD9 and Gemcitabine resulted in a synergistic reduction of these CSC factors ( Figure 4E and Supplemental Figure 4C).
Altogether, the results using resected patient tumour samples indicated that BRD9 inhibition by a small molecule compound can efficiently target and eliminate the CSC subpopulation of pancreatic cancer cells, thus confirming our prior discoveries on PDAC cell lines.
Next, we performed bulk RNA-sequencing upon I-BRD9 inhibition in A13A CSCs which In a cellular context, transcriptional output is largely orchestrated by cis-regulatory elements (CREs), in particular enhancers and promoters. By performing ATAC-seq assay, we identified that I-BRD9 treatment in A13A CSCs results in extensive loss of CREs (n=1,609) compared to control samples (Supplemental Figure 4D). In line with this, Western blots demonstrated that acetylated lysine27 at histone 3 (H3K27ac) 47 , an important active enhancer and promoter histone mark, was significantly decreased in response to BRD9 inhibition ( Figure 4K). In most cases, enhancers control gene expression through long-range interactions with promoters 48 49 , but very little is known about the enhancer/promoter connectome in pancreatic CSCs. In order to study the effects of BRD9 inhibition on the enhancer/promoter connectome of CSCs, we performed H3K27ac In-situ ChIA-PET 50-52 experiments, which capture H3K27ac-centric chromatin interactions (i.e. enhancer/promoter 1 3 connectome) ( Figure 4L). These data suggest that the transcriptional downregulation of stemnessrelated genes is possibly due to the loss of long-range enhancer-promoter connectome. For example, we found that in Control CSCs, the NBAT1 promoter is characterized by H3K27ac enrichment on the promoter but is transcriptionally silenced. In addition, the NBAT1 promoter has strong chromatin interaction with the SOX4 promoter, suggesting that this non-transcribing promoter may function as an enhancer 53 , to regulate the expression of SOX4. However, I-BRD9 treatment diminished the H3K27ac enrichment on the NBAT1 promoter and its physical contact with the SOX4 promoter, which possibly contributes to the transcriptional downregulation of SOX4 ( Figure 4M). Similarly, the transcriptional downregulation of other stemness-related genes (e.g. CD133/PROM1, SMAD1 and SNAI2) are linked to the loss of enhancer-promoter connectome (Supplemental Figure 5B). ATACseq data showed SMAD2/3 footprints at the anchor regions (Supplemental Figure 5C-D). SMAD3 has several TF motifs in the JASPAR motif database. We identified MA0513 in Ctrl-specific ATAC peaks (254 hits) and 211 hits in IBRD9-specific ATAC peaks, while MA1622 has 1370 hits in total on anchor regions with approximately similar number of hits in both conditions. Hence, SMAD2/3 footprints were not significantly lost upon BRD9 inhibition. To further study the binding dependency of BRD9 and SMAD2/3 on TGFβ/Activin pathway we performed BRD9 and SMAD2/3 ChIP-qPCR on representative SMAD2/3 binding regions near SOX4, CD133/PROM1, SNAI2 and SMAD1 loci upon TGFβ/Activin pathway inhibitor SB431542 treatment of pancreatic CSCs (Supplemental Figure 5E).
BRD9 and SMAD2/3 bound to the anchor regions, while the binding of BRD9 on the anchor regions was lost following TGFβ/Activin pathway inhibition.
These results indicate that BRD9 and SMAD2/3 are co-binding to these regions and SMAD2/3 transcription factors recruit BRD9 to the gene promoter and enhancer regions (Supplemental Figure   5F). Collectively, our results revealed the cooperation of BRD9 and SMAD2/3 in regulating the enhancer-promoter connectome and gene expression of stemness-related genes in pancreatic CSCs.

Discussion
Epigenetic regulation of gene expression is essential for guiding developmental processes throughout embryogenesis and tissue homeostasis in an adult organism. The cellular identity of differentiated somatic cells can also be epigenetically reprogrammed through erasure and rewriting of chromatin marks resulting in a pluripotent stem cell state. Similar differentiation and dedifferentiation processes mediated by transcription factors and epigenetic regulatory proteins can contribute to tumour formation, invasiveness, metastatic processes, dormancy and reactivation of cancer cells, clonal evolution of tumour cells and the development of therapeutic resistance in cancers 4 . Of particular importance to all these processes, epigenetic mechanisms regulate phenotypic plasticity and the selfrenewal capacity of cancer stem cells. Accordingly, targeting epigenetic mechanisms offers an attractive strategy to eliminate cancer stem cells through reducing their self-renewal characteristics or re-sensitizing them for chemotherapeutic drugs.
We used a compound screening approach to identify candidate targets that regulate stem celllike characteristics of pancreatic cancer stem cells. The compound library used consisted of extensively ex vivo validated small molecule compounds covering major epigenetic target classes.
For several of the targets, including BRD9, the inclusion of multiple chemical scaffolds for a given target and availability of negative controls (similar chemical scaffolds with significantly reduced activity) provided a robust approach for target identification. Using this strategy we identified BRD9 as binds several of the BAF complexes in CSCs: ncBAF, esBAF and npBAF. Hence, the dynamics of different SMAD2/3-BAF complexes could provide the necessary stem cell-like developmentally plastic or "metastable" capacity of CSCs. We have previously shown that TGFβ/Activin-SMAD2/3 regulates pluripotency in human embryonic stem cells through MLL/COMPASS mediated histone modifications 24 . Based on our current data it seems that TGFβ/Activin-SMAD2/3 also promotes the stem cell-like characteristics of pancreatic CSC. In the latter case, SMAD2/3 seem to direct the BAF complex to stem cell loci to induce their expression.
Many cancers have been found to be almost separate diseases due to different mutations and epigenetic effects that impact divergent mechanisms during tumorigenesis. This heterogeneity in the Our findings can potentially translate into clinical benefit for PDAC patients that have a surgically unresectable cancer and even undergoing relapse.

Materials and Methods
For full materials and methods please see Supplemental Information.

The small molecule screening library
The screening library contained concentrated small molecule compounds with verified biochemical activity against their targets. Most of the compounds target epigenetic regulators with high specificity (Supplemental Table 1).
Cell lines and cell culture 1 7 The A13A, A13D, A13B cells were provided by Christine

Knockin cell lines for Oct4
pCCC construct and OCT4 TALEN constructs (pTALEN_V2-OCT4F, pTALEN_V2-OCT4R) constructs were a gift from Francis C. Lynn and have been published 1 . OCT4-eGFP-PGK-Puro was a gift from Rudolf Jaenisch (Addgene plasmid #31937) and have been published 2 . Cells were transfected with Lipofectamine 3000 (ThermoFischer Scientific) and cultured for 4 days after transfection before selecting with 0.25 µg/mL puromycin (Sigma). Colonies were individually picked, trypsinized and placed into 24-well plates with 500 µl of media. Once clones were close to confluent, cells were replica plated for genotyping, freezing and for expanding the correctly targeted clones. Genomic DNA was extracted using Promega Wizard SV Genomic DNA Purification System (Promega) and genotyping was performed as described 1

Screening of the chemical compounds
The cells were grown in 96-well plates in standard growth medium with puromycin (1 µg/ml stock).
Three technical replicates and three biological replicates were used for the screening. Cells were plated at a concentration of 10,000 cells in 100 µl of media per well in a 96-well plate. One day after plating the cells, the medium was changed to 90 µl standard growth medium supplemented with puromycin (0.5 µg/ml) and Activin A (10 ng/ml). On the same day, the compounds were added: first, 100x compound library dilutions were made, and 10 µL of 100x diluted chemical was added to each well to obtain 1000x final dilution of the compounds. Cells were then cultured with chemical compounds for five days with media change at day 0, day 2 and day 4 supplemented by fresh compounds. Each replicate was analyzed using Celigo Image Cytometer (Nexcelom) and flow cytometry. Cells were lifted and dissociated into single cells with Trypsin. Details on the antibodies that were used for flow cytometry are listed in Table 2

RT-qPCR
2ng of synthesized cDNA was added to 5µl Power SYBR Mix (Life Tecnologies, 4368708 (Master Mix)) and 1.5µl 2µM of forward and reverse primers. RT-qPCR was performed on ViiA 7 machine with the following intervals: denaturation (95 C) for 15s and a total of 40 cycles, annealing/extension (60 C) for 60s, final extension (60 C) for 10 minutes.

Flow cytometry for cell cycle analysis
A13A and FG PDAC cells in which the FUCCI construct was incorporated were taken from adherent conditions and counted, then plated in spheroid conditions at a density of 5,000 cells / 1 mL medium for 10 days, before the cells were collected and analysed using Fortessa (BD Bioscience). Passaging was performed a day 5, after which cells were plated again in spheroid conditions, with the same initial density of 5,000 cells / 1 mL medium. Compounds were added in

Preparation and Sequencing of Illumina RNA libraries
RNA-Seq libraries were created using the NEBNext Ultra RNA library prep kit using TruSeq indexes, following the manufacturer's recommendations. In summary, 500 ng of total RNA was used to isolate mRNA poly(A) by two rounds of purification using oligo dT magnetic beads followed by fragmentation and cDNA synthesis by random primers and reverse transcriptase. Bar-coded adapters were ligated to the cDNA fragments and a PCR reaction was performed to produce the sequencing libraries. To verify the library concentration and the library fragments length was used Agilent 2200 Tape-Station System. Adapter-ligated cDNA fragment libraries were sequenced on a NextSeq 500 platform (Illumina) using a paired-end run (2 × 41 bp).

RNA-sequencing analysis
Sequencing reads from the RNA-seq experiment were aligned to the human genome (hg38) using HISAT2 with the default parameters. FeatureCounts was used to assign mapped reads to genes with annotation gtf file Ensembl94. Differential gene expression analysis was performed using DESeq2 with IHW method for p value adjustment, apeglm method for effect size shrinkage, FDR-adjusted p value < 0.05 and log2 fold change > 0. Functional analysis of differential expressed genes was performed using Metacore. All plots were generated using R package 3.6.

Western blot analysis
Protein was isolated by lysing cells with RIPA Buffer (Sigma-Aldrich) supplemented by cOmplete EDTA-free protease inhibitor (Roche) and PhosSTOP ™ (Sigma-Aldrich) and extracting the supernatant after high-speed centrifugation at 4°C. Protein quantification was performed using the  Figure 4) were diluted in 5% milk in PBS and 0.05% tween 20. Quantification was performed using ImageJ gel analysis tool.

Immunostaining
The immunostaining method has been described previously [3][4][5] . Cells were fixed for 20 minutes at 4°C in PBS 4% PFA (electron microscopy grade), rinsed three times with PBS, and blocked and permeabilized at the same time for 30 minutes at room temperature using PBS with 10% Donkey Serum (Biorad) and 0.1% Triton X-100 (Sigma). Incubation with primary antibodies diluted in PBS 1% Donkey Serum 0.1% Triton X-100 was performed overnight at 4°C. Samples were washed three times with PBS, and then incubated with AlexaFluor secondary antibodies for 1 hour at room temperature protected from light. Cells were finally washed three times with PBS, and Hoechst (Sigma) was added to the first wash to stain nuclei. Images were acquired using a LSM 700 confocal microscope (Leica).

Chromatin Immunoprecipitation (ChIP)
All steps were performed on ice or at 4°C and ice-cold buffers and PBS were supplemented with 1mg/ml Leupeptin, 0.2mM PMSF, and 10mM NaButyrate were used unless otherwise stated.

ATAC-sequencing analysis
Sequencing reads from the ChIP-seq and ATAC seq experiment were aligned to the human genome (hg38) using bowtie with reporting mode," -best -strata -v2". Deeptools was used to generate covergae track(bigwig). Coverage track was visualized by using UCSC genome browser.
Peak calling was performed by using macs2 peak caller with default parameters for ChIP seq, and with parameter "--nomodel --shift -100 --extsize 200" for ATAC seq. Peaks annotated with nearest gene information by using BEDTools. Peak distribution over different genomic features were summarized by using Bioconductor package ChiPpeakAnno. Motif enrichment analysis within peak regions was performed using HOMER. All plots were generated using R package 3.6.
For the TF footprint analysis, motifs in JASPAR database were scanned using ATAC bam file and peak coordinates for Ctrl and IBRD9 separately. We then compared the cleavage profiles from Cells were transferred to low-adherent flat-bottomed plates. Cells were incubated with four different conditions for 72 hours. Cells were collected for single-cell RNA-sequencing.

Preparation of Single Cell Samples
After collection, the cells were centrifuged at a speed not exceeding 400rcf. The supernatant was discarded. The cell pellet was resuspended in 1mL 1X PBS containing 0.04% BSA and the washing procedure was repeated twice. After washing, appropriate volume PBS was added to the cell precipitation to obtain single-cell dispersion suspension with a concentration close to the goal number. A wide-bore pipette tip was used for pipetting cell resuspension for lower cell damage.
40µm Cell Strainers were used for removing cell debris and cell clumps. Automatic cytometry was used to determine the cell concentration. The sample volume was calculated based on the optimal cell sampling concentration supplied by the 10X official website and the target capture number. If the calculated concentration was too high, the liquid volume was adjusted to the appropriate concentration and the counting was repeated. Once the desired cell suspension was obtained, it was immediately placed on ice for subsequent GEMs preparation and reverse transcription.

Library Construction
The library construction was performed as follows: (1) Interrupt, end repair and add A base. Finally, the SPRI select beads were used to purify the product.

Single-cell gene expression analysis
Sequencing data was aligned and quantitated using CellRanger (10x Genomics) and hg38 reference genome sequence. Raw gene expression matrices were then analysed using Seurat R were used to cluster cells. Clusters in 2D UMAP were used to identify cell types based on marker genes. Unless specified, default parameters were used for each function.

ChIP-sequencing data analysis
The adapters were removed and the filtered reads were mapped to hg38 reference genome using bwa 6 mem. The duplicated reads were removed using Picard (https://broadinstitute.github.io/picard/) and the reads with quality lower than 20 were filtered. Bam files were converted to bigwig format file using deepTools 7 bamCoverage, The parameter 'scaleFactor' for each sample were determined by edgeR 8 function calcNormFactors using "TMM" normalization method. The ChIP-seq peaks were called using the default (narrow) setting in MACS2. The promoter regions were defined as +/-2.5Kb around the TSSs regions.

.
Chromatin interaction loops from all samples were merged to form a union set. Read counts for each merged loop were counted across all samples to form a loop -sample count matrix. To identify control-specific and IBRD9-specific loops, the read matrix was imported into R and a negative binomial model was fitted using glmFit function in edgeR 8 R package. The loops with pvalue <0.1 were determined as differential loops 11 . Loop anchors proximal to (less than 3Kb) known transcription start sites (TSSs) were defined as promoters and the remaining loop anchors were defined as distal regulatory elements.

Functional enrichment of downregulated genes with IBRD9 treatment
Functional enrichment of genes network was generated using the ToppFun application of the Toppgene Suite using a corrected (Benjamini and Hochberg) p-value cut-off of 0.05 (https://toppgene.cchmc.org/) 12 . Cytoscape application (https://cytoscape.org/) 13 was used to generate the functional enrichment network.

Visualization
Statistical plots were created using R. Screenshots of chromatin interactions was visualized using WashU Genome Browser.

Animal studies
All Medicilon Inc) were maintained in a specific pathogen-free facility with high efficiency particulate air (HEPA)-filtered air and given autoclaved food and water.

Establishment of xenograft tumor models in nude mice
The A13A cells for subcutaneous injection were divided into the following groups, 1) Null group,

Tumor growth and mouse survival
For tumor growth, mice were monitored twice a week or every 3 days by observation and palpation. Tumor size was measured with a digital caliper and tumor volume was calculated using the formular LxW2x0.5. Mouse body weight was measured twice weekly. Death, weight loss of 15% or more of body wight, tumor volume with 2,000m 3 or more, or severe labored breathing or dyspnea were considered endpoints and all mice were euthanized.

In vivo PET imaging
Mice were fasted for 12 h before 18F-fluorodeoxyglucose (FDG) injection but allowed free access to water. After anaesthesia with 2% isoflurane, mice were injected with 18F-FDG (11 to 16MBq, German Cancer Research Center, Heidelberg, Germany) by tail vein and kept at 37oC until imaging. Imaging was started at 60 minutes after 18F-FDG injection with non-invasive micropositron emission tomography (μPET, Siemens Inveon). A CT scan (80 kVp, 500 μ A, at 120 projections; approximately 4-minutes) was acquired for anatomical reference and enabled PET attenuation correction during reconstruction.
Postprocessing was carried out with Inveon Research Workplace.

Therapeutic studies in PDX models
Therapy was initiated in PDX mice after the tumor reached an average volume of 150mm 3 with a repeated injection schedule (once every 3 days for a total of six treatments within 15 days). The PDX mice with various treatment were divided into the following groups, 1) Control treatment group (Ctrl), the mice were administrated with DMSO in the same manner as the treatment groups; 2) Gemcitabine treatment group (Gem), the mice were administrated with Gemcitabine (50 mg/kg i.p.); 3) Combination of BRD9 inhibitor with Gemcitabine (IBRD9 + Gem), the mice were administrated with Gemcitabine (25 mg/kg i.p.) and BRD9 inhibitor (10 mg/kg i.p.).

Histology analysis with H&E staining
Mouse tumor tissues were dissected, rinsed with PBS and fixed in 10% formalin for 48 h. Following dehydration through a series of ethanol solutions, the tissues were embedded in paraffin wax according to standard laboratory procedures. Subsequently, they were sectioned (5 µm thick) using a microtome. Following deparaffinization and rehydration, H&E staining was performed on the sections. In brief, the sections were stained in Mayers Hematoxylin for 1 min. Following rinsing in tape water, the sections were stained in Alcoholic-Eosin for 1 minute and dehydrated and cleared with xylene. The images were examined with an Olympus BX41 microscope equipped with a CCD (Magna-Fire TM) camera.
DAPI was used for nuclear counterstaining. Four fields of each section were examined for quantification. Fluorescent imaging was performed with an Olympus BX41 microscope equipped with an epifluorescence attachment.

ChIP-qPCR of anchor regions
The cells were treated as for RNA-sequencing experiment and the ChIP was performed as described above. The primers used for ChIP-qPCR of anchor regions are listed in Supplementary