HNF1A recruits KDM6A to activate differentiated acinar cell programs that suppress pancreatic cancer

Abstract

Defects in transcriptional regulators of pancreatic exocrine differentiation have been implicated in pancreatic tumorigenesis, but the molecular mechanisms are poorly understood. The locus encoding the transcription factor HNF1A harbors susceptibility variants for pancreatic ductal adenocarcinoma (PDAC), while KDM6A, encoding Lysine‐specific demethylase 6A, carries somatic mutations in PDAC. Here, we show that pancreas‐specific Hnf1a null mutant transcriptomes phenocopy those of Kdm6a mutations, and both defects synergize with Kras ^G12D to cause PDAC with sarcomatoid features. We combine genetic, epigenomic, and biochemical studies to show that HNF1A recruits KDM6A to genomic binding sites in pancreatic acinar cells. This remodels the acinar enhancer landscape, activates differentiated acinar cell programs, and indirectly suppresses oncogenic and epithelial–mesenchymal transition genes. We also identify a subset of non‐classical PDAC samples that exhibit the HNF1A/KDM6A‐deficient molecular phenotype. These findings provide direct genetic evidence that HNF1A deficiency promotes PDAC. They also connect the tumor‐suppressive role of KDM6A deficiency with a cell‐specific molecular mechanism that underlies PDAC subtype definition.

Cooperation between a transcription factor and an epigenetic modifier reveals how defective transcriptional control can contribute to development of pancreatic ductal adenocarcinoma.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is a leading cause of cancer mortality (Garrido‐Laguna & Hidalgo, 2015). The incidence of PDAC is rising, yet current chemotherapies are generally ineffective (Ryan et al, 2014). Genomic analysis of PDAC has identified almost universal driver mutations in KRAS, TP53, SMAD4, and CDKN2A, among a long list of loci that show recurrent somatic mutations and structural variations (Jones et al, 2008; Biankin et al, 2012; Waddell et al, 2015; Bailey et al, 2016; Notta et al, 2016). A small subset of tumors is caused by germ‐line mutations in DNA‐repair genes (Waddell et al, 2015; Roberts et al, 2016), whereas GWAS have identified dozens of common variants that impact PDAC susceptibility (Childs et al, 2015; Klein et al, 2018). Genetic studies have therefore uncovered leads that promise to define molecular targets for future precision therapies.

Up to 18% of PDAC tumors carry mutations in KDM6A (Waddell et al, 2015), which encodes Lysine‐specific demethylase 6A (KDM6A), a component of the MLL/COMPASS transcriptional co‐regulatory complex (Cho et al, 2007). KDM6A catalyzes demethylation of histone H3K27me3, a modification associated with polycomb‐mediated repression (Agger et al, 2007; Lan et al, 2007; Lee et al, 2007). Most somatic pathogenic KDM6A mutations are likely to result in a loss of function, and mouse genetic studies have shown that Kdm6a and Kras mutations cooperate to promote PDAC (Mann et al, 2012; Andricovich et al, 2018). How KDM6A is recruited to its genomic targets in pancreatic cells, and the direct mechanisms through which it controls PDAC‐relevant genetic programs are still poorly understood (Wang & Shilatifard, 2019).

There is increasing evidence that the transcriptional regulation of differentiated pancreatic exocrine cells is tightly linked to PDAC development and subtype definition (Stanger & Hebrok, 2013; Krah et al, 2015; Diaferia et al, 2016; Martinelli et al, 2016; Cobo et al, 2018). Little is known, however, about the underlying molecular underpinnings. We have examined HNF1A, a homeodomain transcriptional regulator of liver, gut, kidney, and pancreas, which has been proposed as a candidate pancreatic tumor suppressor (Molero et al, 2012; Stanger & Hebrok, 2013; Hoskins et al, 2014; Luo et al, 2015). Human heterozygous HNF1A loss‐of‐function mutations cause diabetes, in part because HNF1A promotes pancreatic β‐cell proliferation, and mouse Hnf1a mutations prevent the formation of large T antigen‐driven β‐cell tumors (Servitja et al, 2009). The function of HNF1A, however, is cell‐type specific (Servitja et al, 2009), and both co‐expression network analysis of PDAC samples and in vitro studies suggest that HNF1A has a tumor‐suppressive function in pancreatic exocrine cells (Hoskins et al, 2014; Luo et al, 2015). Furthermore, GWAS suggest that genetic variants in the HNF1A locus predispose to PDAC (Pierce & Ahsan, 2011; Klein et al, 2018). Despite these observations, there is currently no direct mouse or human genetic evidence to incriminate HNF1A deficiency in PDAC.

Here, we combine mouse genetics, transcriptomics, and genome binding studies to show that HNF1A is a major determinant for the recruitment of KDM6A to its genomic targets in acinar cells. This remodels the enhancer landscape of acinar cells and activates a broad epithelial cell transcriptional program that inhibits tumor suppressor pathways. We demonstrate that Hnf1a inactivation promotes Kras‐induced PDAC, and partially phenocopies morphological features of Kdm6a‐deficient PDAC. Finally, we define a subset of human tumors that exhibit HNF1A/KDM6A‐deficient transcriptional programs. These findings, therefore, provide a molecular mechanism that connects the tumor‐suppressive functions of KDM6A and pancreatic differentiation transcription factors.

Results

Hnf1a deficiency promotes Kras‐induced oncogenesis

To directly test the role of Hnf1a in pancreatic carcinogenesis, we created a conditional Hnf1a loss‐of‐function allele (Hnf1a ^LoxP) (Appendix Fig S1A) and used a Pdx1 ^Cre transgene to delete Hnf1a in all pancreatic epithelial lineages (hereafter referred to as Hnf1a ^pKO mice, Appendix Fig S1B). HNF1A is normally expressed in pancreatic acinar and endocrine cells, but not in duct cells (Nammo et al, 2002), and Hnf1a ^pKO mice showed disrupted HNF1A expression in both acinar and endocrine cells (Appendix Fig S1C). As expected from previous studies of Hnf1a germ‐line null mutants, this did not produce gross defects in pancreas organogenesis or tissue architecture (Appendix Fig S1D) although acinar cells displayed signs of markedly increased proliferation (Pontoglio et al, 1996; Lee et al, 1998; Boj et al, 2001; Molero et al, 2012) (Fig 1A).

Figure 1. HNF1A deficiency leads to increased proliferation and promotes Kras‐induced oncogenesis.

• A
  Representative immunofluorescence images and quantifications showing that 3‐month‐old Hnf1a ^pKO mice have increased number of KI67⁺ (red) acinar cell nuclei co‐staining with DAPI (blue) and Amylase (green). Arrows point to KI67⁺ acinar cells in Hnf1a ^pKO mouse. Acinar proliferation is represented as the average of the KI67⁺/Amylase⁺ cell ratio. Quantifications were performed on 3 random fields from 3 Pdx1 ^Cre and 3 Hnf1a ^pKO mice. P‐values are from two‐tailed Student's t‐test. Representative H&E stainings of pancreata from Kras ^G12D and Hnf1a ^pKO;Kras ^G12D mice.
• B–D
  Kras ^G12D and Hnf1a ^pKO;Kras ^G12D mice have normal morphology at 7 days.
• E–J
  At 21 days, Hnf1a ^pKO;Kras ^G12D mice show acinar‐to‐ductal metaplasia (dashed encircled areas) and regions with desmoplastic reaction (asterisk), which are not observed in Kras ^G12D mice (E, H).
• K–P
  At 8 weeks, Kras ^G12D pancreas show occasional abnormal ductal structures (dashed encircled areas in N, which is a magnification of squared dotted box in K) and Hnf1a ^pKO;Kras ^G12D mice (L, M, O, P) present mucinous tubular complexes (black arrows), and more advanced PanINs with luminal budding (open arrows) including foci of spindle cell proliferation (asterisks) and incipient infiltrative growth (black dashed box area in O).

Data information: Black dashed boxes in (E, F, K, L and O) indicate magnified areas in (H, G, N, M and P) respectively. Scale bars indicate 200 μm (A), 100 μm (C, E, F, K, L), 50 μm (O), and 20 μm (B, D, G, H–J, M, N, P).

To determine whether Hnf1a interacts with Kras‐induced carcinogenesis, we created mice with combined conditional Hnf1a and Kras ^G12D mutations, hereafter referred to as Hnf1a ^pKO ;Kras ^G12D mice (Appendix Fig S1E). In the absence of Hnf1a mutant alleles, Pdx1 ^Cre‐induced Kras ^G12D activation expectedly gave rise to occasional low‐grade PanINs or acinar‐to‐ductal metaplasia (ADM) lesions by 2 months of age (Hingorani et al, 2003) (Fig 1B, E, H, K and N). Hnf1a ^pKO ;Kras ^G12D mice showed no lesions at 7 days of age (Fig 1C and D), yet by weaning they had already developed focal ADM and desmoplastic reactions, which became more prominent as the mice aged (Fig 1F, G, I and J and data not shown). Eight‐week‐old Hnf1a ^pKO ;Kras ^G12D mice additionally showed non‐invasive atypical tubular complexes, higher‐grade PanINs with luminal budding, desmoplastic reaction, and foci of spindle cell (mesenchymal) proliferation, some of which showed incipient infiltrative growth (Fig 1L, M, O and P). These findings indicate that pancreatic Hnf1a deficiency cooperates with Kras to promote sarcomatoid forms of PDAC.

HNF1A activates an acinar differentiation program that inhibits oncogenic programs

To understand how Hnf1a deficiency promotes pancreatic cancer, we examined the transcriptional programs controlled by Hnf1a in pancreatic exocrine cells. Genetic lineage tracing studies in mice have shown that, despite the ductal morphology of PDAC, Kras ^G12D‐induced PDAC can originate from acinar cells that undergo ADM and PanIN, contrasting with intraductal papillary mucinous neoplasms that arise from duct cells (Kopp et al, 2012; von Figura et al, 2014). We therefore used a Ptf1a ^Cre allele, which ensured high‐efficiency recombination of Hnf1a in mouse acinar cells, and more limited recombination in endocrine cells (Ptf1a ^Cre; Hnf1a ^LoxP/LoxP; hereafter referred to as Hnf1a ^aKO) (Fig EV1A and B). Eight‐week‐old Hnf1a ^aKO mice were normoglycemic, and like Hnf1a ^pKO mice showed normal pancreatic histology (Fig EV1C). We profiled transcripts in pancreas from 8‐week‐old Hnf1a ^aKO mice and, despite the normal histology, found profound transcriptional changes (Fig 2A, Dataset EV1). We observed decreased expression of genes specific to differentiated acinar cells, including Ptf1a, Pla2g1b, Serpini2, and Ctrb1 (Figs 2B and EV1D), and increased expression of genes specific to pancreatic mesenchymal cells (Fig 2B, Dataset EV2). Downregulated genes were enriched in metabolic processes such as inositol phosphate turnover, amino acid metabolism, and protection against oxidative stress, whereas upregulated genes were enriched in annotations associated with the extracellular matrix (collagen formation, ECM‐receptor interactions, integrin cell surface interactions) and complement activation (Fig 2C, Dataset EV2). We also observed activation of cell cycle‐related pathways, cholesterol biosynthesis, and known oncogenic programs such as EMT, RAS, PI3K‐AKT, STAT3, WNT, and MAPK signaling (Figs 2D and EV1E and F, Dataset EV2).

Figure EV1. Hnf1a deletion leads to abnormal acinar cell identity and increased expression of genes related to oncogenic pathways.

• A
  Breeding strategy to generate Hnf1a ^aKO and Ptf1a ^Cre ;Hnf1a ^+/+ control mice using Ptf1a ^Cre and Hnf1a ^LoxP alleles.
• B
  Ptf1a ^Cre deletes HNF1A efficiently in acinar cells but to a lesser extent in islets of Langerhans. HNF1A IHC and hematoxylin staining in pancreas of control and Hnf1a ^aKO mice. HNF1A is expressed in acinar and islet cells, but not in ductal cells in normal pancreas (left). HNF1A expression is depleted in acinar cells but largely not in islets in Hnf1a ^aKO pancreas (right). The squared dotted boxes (top) indicate magnified areas (bottom). Arrows point at ducts, arrow head at HNF1A‐positive acinar cell, and open arrow head at HNF1A positive islet cell. The dotted encircled areas indicate islets of Langerhans. Scale bar (top) 300 μm (bottom) 50 μm.
• C
  H&E stainings in pancreas of control (left) and Hnf1 ^aKO mice (right) showing unaltered pancreatic morphology. Scale bar 300 μm.
• D
  Expression of acinar differentiation genes in pancreas from Hnf1a ^aKO and controls, depicted as box plots with median values and IQR of TPM values. Whiskers extend to highest and lowest data points within 1.5× IQR outside box limits. P‐values were determined by two‐tailed Student's t‐test and n = 3 replicates per condition.
• E
  GSEA showing increased expression of oncogenic pathways in Hnf1a ^aKO pancreas.
• F
  Western blots (top) and quantifications (bottom) showing increased phospho‐p42 levels in Hnf1a ^aKO pancreas. Quantification of signal intensities of phospho‐p44/p42 normalized to total‐p44/p42 levels. Data are shown as dots with mean and error bars ± SD. P‐values were determined by two‐tailed Student's t‐test.
• G
  Distribution of pancreatic HNF1A binding sites in annotated genomic regions.
• H
  Venn diagrams illustrating that HNF1A‐bound regions are enriched in regions of active promoters and enhancers. P‐values and odds ratios were calculated by Fisher's exact test.
• I
  Enrichment of known HNF1 motifs in the top 500 most significant HNF1A‐bound ChIP‐seq regions and percentage of regions containing each motif. The “union” is the percentage of regions with at least one motif sequence occurrence. Enrichment P‐values are calculated using the one‐tailed binomial test.
• J
  Genome browser track for Fn1 and Timp1 genes showing upregulated expression in Hnf1a ^aKO and Kdm6a ^pKO pancreas, and absence of HNF1A or KDM6A binding in adjacent regions. Plots show TPM values normalized to Hprt with mean and error bars ± SD. P‐values were determined by two‐tailed Student's t‐test.

Figure 2. HNF1A regulates gene programs essential to acinar cell identity, metabolism, and proliferation.

• A
  Fold change (FC) in transcripts in Hnf1a ^aKO versus control pancreas, plotted against significance (−Log₁₀ q; genes significant at q < 0.05 are shown as colored dots above the horizontal line).
• B
  GSEA showing that genes specific to differentiated acinar cells were downregulated in Hnf1a ^aKO pancreas, but not genes specific to islets or duct cells. Upregulated genes were enriched in genes specific to mesenchymal cells. Lineage‐enriched genes were obtained from Muraro et al (2016).
• C
  Top functional annotations for differentially expressed genes in Hnf1a ^aKO pancreas.
• D
  GSEA revealed that Hnf1a ^aKO pancreas showed increased transcripts involved in oncogenic pathways such as EMT, MAPK, KRAS, PI3K‐AKT.
• E
  HNF1A promotes transcriptional activation of direct target genes. Left: HNF1A‐bound genes were enriched among genes that showed downregulation in Hnf1a mutants, but not among upregulated genes. P‐values and odds ratios (O.R.) calculated by Fisher's exact test. Right: Venn diagrams showing overlap of HNF1A‐bound genes with genes that were downregulated and upregulated in Hnf1a mutant pancreas.

To assess which of these transcriptional changes reflects a direct function of HNF1A in acinar cells, we profiled genome‐wide binding sites of HNF1A in adult pancreas (Fig EV1G), as well as H3K27 acetylation to mark active enhancers and promoters (Fig EV1H). HNF1A‐bound genomic regions had canonical HNF1 recognition sequences in 405 out of the top 500 most significant binding sites (81%), confirming the specificity of the assay (Fig EV1I), and they were expectedly enriched in enhancers and promoters (Fig EV1H). HNF1A binding was specifically enriched among genes that showed downregulation in Hnf1a ^aKO pancreas (odds ratio = 3.6 relative to all other genes, P = 10⁻²², Fig 2E). This was consistent with HNF1A acting as an essential positive regulator of some of the genes to which it binds. HNF1A did not, however, show a higher number of binding near genes that were upregulated in Hnf1a ^aKO pancreas, compared with active genes that did not show altered regulation (odds ratio = 0.4, P = N.S.; Figs 2E and EV1J). These studies, therefore, uncovered direct HNF1A‐dependent genetic programs. They show that the function of HNF1A in pancreatic acinar cells entails direct transcriptional activation of a broad differentiated cell program that controls metabolic functions. They also revealed that HNF1A suppresses growth‐promoting pathways and that, given the absence of enriched HNF1A binding to upregulated genes, this is largely mediated through indirect regulatory mechanisms.

HNF1A‐dependent programs in human acinar cells

We explored the relevance of mouse pancreatic exocrine HNF1A‐dependent programs to human acinar cells. Given the lack of human acinar cell models to perform genetic manipulations, we examined transcriptomes from 328 human pancreas samples from the Genotype‐Tissue Expression (GTEx) project (Consortium et al, 2017). We defined GTEx samples with highest and lowest deciles of HNF1A mRNA expression (n = 33 per group) and asked if human orthologs of genes that were deregulated genes in Hnf1a ^aKO pancreas showed concordant changes. We observed that Hnf1a ^aKO up‐ and downregulated genes showed consistent up‐ and downregulation in HNF1A‐low versus HNF1A‐high human pancreas (median Log2 fold‐difference [IQR]: upregulated genes 0.44 [0.05–0.81]; downregulated genes −0.16 [−0.38 to 0.05]; control genes 0.03 [−0.2 to 0.36]; Kruskal–Wallis P < 10⁻⁴) (Fig 3A). Accordingly, 70 and 75% of Hnf1a ^aKO up‐ and downregulated genes showed differential expression in the same direction at nominal P < 0.05 in HNF1A‐low versus HNF1A‐high human pancreas, respectively. These observations suggested that HNF1A‐dependent programs are largely conserved in mouse and human pancreatic cells.

Figure 3. Human non‐classical and KDM6A‐deficient PDAC exhibit HNF1A‐deficient phenotypes.

• A
  Human orthologs of genes that were up‐ and downregulated in Hnf1a ^aKO pancreas were also up‐ and downregulated in human pancreas with low versus high HNF1A expression (lowest versus highest expression deciles, respectively). A random list of 717 genes controlled for similar expression levels was used for comparison. Violin plots include median and interquartile ranges. Dots are average values for each gene. Kruskal–Wallis P < 0.0001.
• B
  GSEA demonstrates that down‐ or upregulated genes in Hnf1a ^aKO mice (downward or upward arrows) showed down‐ or upregulation, respectively, in gene lists ranked by differential expression in non‐classical versus classical PDAC molecular subtypes from the TCGA‐PAAD study (Cancer Genome Atlas Research Network, Electronic Address Aadhe, Cancer Genome Atlas Research N, 2017). All enrichments had GSEA FDR q‐values < 0.01.
• C
  Analysis of HNF1A function in 121 high‐purity cases of the ICGC‐PACA‐AU cohort identified tumors with most pronounced downregulation of direct HNF1A target genes. We performed GSEA with a gene set of 106 human orthologs of HNF1A direct targets showing downregulation in Hnf1a ^aKO pancreas. For each tumor sample, we performed differential expression against all other samples and used GSEA to ascertain abnormal expression of the mouse HNF1A‐dependent gene set in the tumor. Samples were ranked by the resulting normalized enrichment score (NES) and classified as either HNF1A LoF samples (purple, NES < 0; P < 0.05), or Control 1 (beige, NES < 0; P > 0.05) and Control 2 (gray, NES > 0). HNF1A LoF samples were predominantly non‐classical tumors (Collisson et al, 2011; Moffitt et al, 2015; Bailey et al, 2016). Putative loss‐of‐function KDM6A mutations (KDM6A LoF mutants) were found in 19% of HNF1A LoF tumors versus 2% of all others (Fisher's P = 0.005). KDM6A mutations were considered functional if classified as “high” functional impact in ICGC (small ≤ 200‐bp deletions/insertions, single base substitutions), or as likely loss‐of‐function structural variants in Bailey et al (2016), all of which were frame‐shift mutations. Other KDM6A mutations were classified as unknown. Heatmaps show Z‐score‐normalized expression of deregulated genes in Hnf1a ^aKO pancreas. We confirmed that 85% of 106 downregulated and 60% of genes of 146 upregulated human orthologs showed differential expression across the 3 HNF1A profiles (q < 0.05, SAM multiclass analysis).
• D
  HNF1A mRNA levels differed in HNF1A LoF and control groups (Kruskal–Wallis, P < 0.01), despite considerable variability and overlap between groups.
• E
  KDM6A mRNA levels were downregulated in HNF1A LoF tumors (Kruskal–Wallis, P < 0.001).

Data information: Box plots in (D and E) show HNF1A and KDM6A expression in HNF1A LoF tumors (n = 26) and Control 1 (n = 39) and Control 2 (n = 57) tumors. The horizontal central line marks the median. Box limits indicate the first and third quartiles, and whiskers extend to highest and lowest data points within 1.5× interquartile range outside box limits.

Deregulation of the HNF1A‐dependent program in non‐classical PDAC

We next studied HNF1A‐dependent genes in human primary PDAC. We examined transcriptome data from the TCGA‐PAAD study (Cancer Genome Atlas Research Network. Electronic address & Cancer Genome Atlas Research, 2017) and found that human orthologs of deregulated genes in Hnf1a ^aKO pancreas showed concordant down‐ or upregulation in non‐classical tumor molecular subtypes—variously defined as quasimesenchymal, basal, and squamous‐like PDAC (Collisson et al, 2011; Moffitt et al, 2015; Bailey et al, 2016) (Fig 3B). We employed these same Hnf1a ^aKO deregulated gene sets to cluster human tumor samples from TCGA‐PAAD and ICGC‐PACA cohorts (Bailey et al, 2016) using non‐negative matrix factorization. This exposed a cluster of predominantly non‐classical tumors that showed transcriptional changes in the same direction as Hnf1a‐deficient pancreas (HNF1A cluster 3, Fig EV2A).

Figure EV2. HNF1A signature correlates with non‐classical PDAC subtypes.

• A
  Consensus clustered Z‐score‐normalized gene expression heatmaps of high‐purity TCGA‐PAAD and ICGC‐PACA‐AU human PDAC samples. Clustering was performed with non‐negative matrix factorization based on expression of significantly down‐ and upregulated genes in Hnf1a ^aKO pancreas. This revealed a cluster (HNF1A cluster 3) with concordant up‐ and downregulation of genes in Hnf1a ^aKO pancreas, which predominantly matched non‐classical PDAC molecular subtypes (quasimesenchymal, basal, squamous‐like, pink in top tracks), as opposed to classical PDAC subtypes (green in top tracks). Multiclass SAM differentially expressed genes (q < 0.05) between HNF1A clusters are shown. Genes were hierarchically clustered using complete linkage with one minus Pearson correlation metrics. Along the right side of the heatmaps are green and red indicators of down‐ and upregulated genes in Hnf1a ^aKO pancreas, respectively.
• B
  TP63 expression was increased in HNF1A LoF tumors compared to control tumors. RSEM normalized count data are shown as box plots with interquartile range, median, and whiskers. Box limits indicate the first and third quartiles and whiskers extend to highest and lowest data points within 1.5× IQR outside box limits. HNF1A LoF (n = 26), Control 1 (n = 39), and Control 2 (n = 57) tumors (P, Kruskal–Wallis).
• C, D
  Expression of HNF1A and KDM6A, showing downregulation in non‐classical PDAC subtypes (P, Kruskal–Wallis). Dots are RSEM normalized values presented with mean ± SD. Collisson subtypes: Quasimesenchymal (QM, n = 34) and Classical (CL, n = 54). Moffitt subtypes: Basal (BA, n = 65) and Classical (CL, n = 85). Bailey subtypes: Squamous‐like (SQ‐like, n = 31) and Pancreatic Progenitor (PP, n = 53).
• E, F
  HNF1A levels are not lower in high histological grade PDAC (E), while KDM6A levels are (F). To determine whether histological grade of human PDAC was associated with expression levels of HNF1A (E) or KDM6A (F) proteins, we evaluated contingency tables of tumor grades versus staining intensities of each case in tissue microarray (TMA) IHC. Tumor grades were scored as either moderately differentiated (G2) or poorly differentiated/high grade (G3), and staining intensities were expressed as an Immuno Reactivity Score (IRS) reflecting either No, Weak, Moderate, or Strong staining intensities (see material and methods for details). Numbers of cases and percentages (in brackets) out of total cases are indicated for each tumor grade and staining intensity. The Chi‐squared test was used to determine the probability of a significant relationship. Chi‐square and P‐values are shown. N = 217 patients for HNF1A and N = 208 patients for KDM6A.

We next sought to identify tumors with most pronounced HNF1A‐deficient function. Because our findings showed that HNF1A primarily functions as a direct activator, we defined tumors with most pronounced downregulation of genes that were both HNF1A‐bound and downregulated in Hnf1a mutant mice (hereafter referred to as HNF1A loss of function, or HNF1A LoF tumors). We found that HNF1A LoF tumors showed a remarkable concordance with tumors that were consistently classified as non‐classical molecular subtypes in independent studies, and were therefore significantly enriched in quasimesenchymal (Fisher's P = 1.7 × 10⁻¹⁴), basal (P = 2.9 × 10⁻⁶), and squamous‐like (P = 1.1 × 10⁻⁹) subtypes (Fig 3C). They also had increased TP63 mRNA, a marker of squamous‐like PDAC molecular subtype (Bailey et al, 2016; Andricovich et al, 2018; Somerville et al, 2018) (Fig EV2B).

HNF1A LoF tumors—as well as non‐classical PDAC—showed significantly lower HNF1A mRNA levels (Kruskal–Wallis P = 0.0064), although there was considerable overlap with control tumors, suggesting that abnormal HNF1A function cannot be exclusively explained by differences in HNF1A expression (Figs 3D and EV2C). Furthermore, although tumors with an HNF1A LoF signature also had non‐classical molecular signatures—which correlate with high histological grade—HNF1A immunoreactivity was not significantly lower in high histological grade (poorly differentiated) tumors in tissue microarrays of human PDAC (n = 102) (Fig EV2E). These results, therefore, revealed that a subset of non‐classical PDAC tumors had a transcriptional signature that was consistent with abnormal HNF1A function.

HNF1A LoF signature in KDM6A‐deficient non‐classical PDAC

Although a subset of non‐classical PDAC samples showed a gene expression profile that resembled that of pancreatic Hnf1a‐mutant mice, HNF1A mRNA levels were not invariably altered, and so far, recurrent somatic HNF1A mutations have not been reported in PDAC. This raised the question of why non‐classical PDAC shows abnormal HNF1A function. Non‐classical (e.g., squamous‐like) PDAC have been shown to express low KDM6A mRNA and are enriched in KDM6A somatic genomic defects (Bailey et al, 2016; Andricovich et al, 2018). Our studies also showed decreased KDM6A mRNA in non‐classical PDAC (P < 0.001) (Fig EV2D) and decreased KDM6A immunoreactivity in poorly differentiated tumors (P = 0.03, n = 94) (Fig EV2F). Importantly, tumors with HNF1A LoF phenotypes showed decreased KDM6A mRNA (median [IQR]: 41.9 [32.2–50.5] in LoF tumors, versus 57.3 [44.3–64.3] and 79.4 [53.1–104.2] in control tumors, Kruskal–Wallis P < 0.0001) (Fig 3E). Furthermore, analysis of the Australian ICGC‐PACA data revealed putative loss‐of‐function KDM6A mutations in 19% of tumors showing HNF1A LoF phenotypes, compared with 2% of all other tumors (Fisher's P = 0.005) (Fig 3C). Likewise, tumors with KDM6A putative loss‐of‐function mutations showed abnormal HNF1A‐dependent programs (Fig 3C). Collectively, these correlations hinted at a mechanistic link between KDM6A‐ and HNF1A‐deficient phenotypes in non‐classical PDAC.

Kdm6a‐dependent Kras‐induced oncogenesis

To examine the relationship between Hnf1a and Kdm6a deficiency in pancreatic cancer, we generated mice with pancreas‐specific inactivating Kdm6a mutations and oncogenic Kras mutations [Pdx1 ^Cre, Kdm6a ^LoxP/LoxP, Kras ^G12D, hereafter referred to as Kdm6a ^pKO ;Kras ^G12D mice (Appendix Fig S2A and B)]. KDM6A is normally expressed in all pancreatic cell types and was efficiently excised in Kdm6a ^pKO pancreas (Appendix Fig S2C and D). We focused on female mice because KDM6A is encoded in the X chromosome, and males harbor a Y chromosome paralog named Uty. Kdm6a ^pKO ;Kras ^G12D mice showed normal pancreas morphology at 7 days of age (Fig 4A–C), but subsequently rapidly developed invasive PDAC. Early tumors were apparent by weaning, showing prominent signs of spindle cell proliferation, sarcomatoid morphology with occasional glandular tumor components, regions with widespread signs of ADM and abundant desmoplastic reaction (Fig 4E, F, H and I). Subsequently, tumors showed massive remodeling and very rapid infiltrative growth (Fig 4K, L, N and O), leading to the death of all Kdm6a ^pKO ;Kras ^G12D female mice by 15–16 weeks of age (Appendix Fig S2E). This contrasted with control Kras ^G12D mice, which showed normal morphology at weaning (Fig 4D and G) and expectedly only presented occasional acinar‐to‐ductal metaplasia at 8 weeks (Fig 4J and M). These findings were generally consistent with recently reported observations in 6‐week‐old Kdm6a mutant mice (Andricovich et al, 2018), with the exception that we have not observed histological signs of squamous differentiation. Male mice presented delayed mortality compared to females, suggesting partial Uty compensatory tumor suppressor functions, as reported previously (Andricovich et al, 2018; Gozdecka et al, 2018). These findings, therefore, showed Kdm6a‐dependent sarcomatoid PDAC lesions that are reminiscent of those observed in Hnf1a ^pKO;Kras ^G12D mice, although Kdm6a‐deficient tumors showed earlier age of onset and much more rapid growth. These results were consistent with the overlapping KDM6A‐ and HNF1A‐deficient molecular signatures of human tumors.

Figure 4. Kdm6a‐dependent Kras‐induced oncogenesis.

H&E staining of pancreata from Kdm6a ^pKO;Kras ^G12D and Kras ^G12D mice.

• A–C
  Seven‐day‐old Kras ^G12D and Kdm6a ^pKO;Kras ^G12D mice showed normal morphology.
• D–I
  At 21 days, Kras ^G12D mice showed normal pancreas morphology (dotted area in D is shown at large magnification in G), whereas Kdm6a ^pKO;Kras ^G12D mice already show acinar‐to‐ductal metaplasia (dashed encircled area in F), spindle cell proliferation (asterisks in E and H), sarcomatoid architecture (I), and desmoplastic reaction (black arrowhead in H).
• J–O
  Kras ^G12D mice present occasional acinar‐to‐ductal metaplasia and low‐grade PanINs at 8 weeks (M is a magnification of squared dotted box in J and see Fig 1N), whereas at the same age pancreas from Kdm6a ^pKO ;Kras ^G12D mice show massive remodeling (K), extensive acinar‐to‐ductal metaplasia (L and dashed encircled area in N), cancer with prominent spindle/mesenchymal proliferation, infiltrative growth (black arrows and asterisk, respectively, in O) and abundant stroma (black arrowheads in N).

Data information: Scale bars: 100 μm (B, D, E, J, K) or 20 μm (A, C, F–I, L–N).

KDM6A activates an acinar cell growth‐suppressing program

The analysis of molecular signatures in human tumors, and the semblance of mouse genetic phenotypes, suggested that KDM6A and HNF1A might control common tumor‐suppressive programs in pancreatic exocrine cells. We thus studied the transcriptional function of KDM6A in non‐tumoral, differentiated pancreatic exocrine cells, following a strategy similar to that described above for HNF1A. We studied female Kdm6a ^pKO mice, in which the Pdx1 ^Cre transgene led to pancreatic inactivation of Kdm6a in most cells from all pancreatic epithelial lineages as early as e15.5, and a marked reduction of KDM6A protein at weaning (Fig EV3A, Appendix Fig S2D). Kdm6a ^pKO mice were born at the expected Mendelian ratio and appeared healthy. They had normal glycemia at 12 weeks of age (Fig EV3B) and unaltered pancreas morphology at weaning, indicating that, like HNF1A, KDM6A is dispensable for pancreas organogenesis (Fig EV3C–F). In analogy to observations in Hnf1a ^aKO mice, acinar cells showed increased proliferation at weaning (Fig EV3I), although by 8 weeks of age some mice showed acinar cell attrition and atrophic pancreatic lobules (Fig EV3G and H).

Figure EV3. KDM6A suppresses growth and oncogenic pathways and maintains acinar cell integrity.

• A
  Efficient deletion of Kdm6a in the pancreatic epithelium at E15.5. KDM6A (red) is ubiquitously expressed in all pancreatic cells. CDH1 (green) marks epithelial cells. Upon deletion, KDM6A staining is lost specifically in CDH1‐expressing epithelial cells but not in mesenchymal cells or in the stomach epithelium (white arrow heads). Scale bar indicates 100 μm.
• B
  Kdm6a mutant mice show normal fasting and fed glycemia. The horizontal stroked line indicates blood glucose levels at 250 mg/dl as a reference.
• C–H
  The pancreas of Kdm6a ^pKO mice were histologically normal until 8 weeks of age. At 8 weeks of age, some signs of acinar cell attrition and fat replacement could be observed, as shown in this representative image. Scale bars: 250 μm (10× magnification), 50 μm (40× magnification).
• I
  Representative picture (left) showing increased number of KI67 (red) amylase‐expressing acinar cells (green) in Kdm6a ^pKO pancreas. Scale bar, 250 μm. Quantifications (right) were performed on three pancreatic sections separated by at least 100 μm from 4 control and 4 Kdm6a ^pKO mice. Acinar cell proliferation was represented as the average of the KI67⁺/Amylase⁺ cell ratio ± SD. P‐values were determined by two‐tailed Student's t‐test.
• J
  GSEA plots showing enrichment of Oncostatin M and “TNFA signaling via NFKB” gene sets among genes upregulated in Kdm6a ^pKO pancreas.
• K
  Western blots (top) and quantifications (bottom) showing increased phospho‐p44/p42 levels in Kdm6a ^pKO pancreas. Quantification of signal intensities of phospho‐p44/p42 normalized to total‐p44/p42 levels. Data are shown as dots with mean and error bars ± SD. P‐values were determined by two‐tailed Student's t‐test.
• L
  Most significantly deranged REACTOME pathways in both KDM6A‐ and HNF1A‐deficient pancreas (see also Dataset EV5).
• M
  Kdm6a ^pKO down‐ and upregulated gene sets showed concordant deregulation in KDM6A LoF mutant tumors versus classical PDAC (based on Bailey et al's signature) (Bailey et al, 2016).
• N
  Tumors with KDM6A‐deficient phenotypes showed decreased KDM6A mRNA. We created a gene set of human orthologs of Kdm6a ^pKO downregulated genes, and for each high‐purity tumor sample in the ICGC‐PACA‐AU, we used GSEA to test for enrichment of this gene set in gene lists that were rank‐ordered by differential expression in the individual sample versus all other samples. Samples with NES < 0 and P‐value < 0.05 were considered as having KDM6A LoF phenotypes and were compared against all other samples. Z‐score‐normalized count data are shown as box plots with IQR, median, and whiskers. Whiskers extend to highest and lowest data points within 1.5× IQR outside box limits. P‐values were determined by two‐tailed Student's t‐test.
• O
  Gene sets that showed up‐ or downregulation in non‐classical human PDAC showed concordant enrichment in up‐ or downregulated genes in Kdm6a ^pKO versus control pancreas. GSEA NES and FDR q‐values are shown.
• P
  Genomic distribution of KDM6A binding sites in mouse pancreas.
• Q, R
  Top: ChIP‐seq and RNA‐seq tracks in control and Kdm6a ^pKO pancreas, in two loci harboring downregulated genes (Kif12, Gprc5c) in Kdm6a ^pKO pancreas. Bottom: ChIP‐qPCR validations for regions highlighted in green (R1, R2, R3), showing that Kdm6a mutants have increased H3K27me3 and decreased H3K27ac in most regions. H3K4me1 was also decreased in mutants in distal sites. Error bars show SD, and P‐values were determined by two‐tailed Student's t‐test, n = 3.
• S
  KDM6A‐bound regions are enriched in active pancreas promoters and enhancers. P‐values are calculated by Fisher's exact test.
• T, U
  Genome Browser examples (top) of HNF1A and KDM6A binding to genes known as negative regulators of EMT: Gstp1 (T) and Deptor in (U) that are downregulated in Hnf1a ^aKO and Kdm6a ^pKO pancreas (bottom). Plots show TPM values normalized to Hprt with mean and error bars ± SD. N = 4 per condition and P‐values were determined by two‐tailed Student's t‐test.

To assess the transcriptional function of KDM6A in the pancreas, we analyzed pancreatic RNA (> 98% of which originates from exocrine cells) from 4‐day‐old female Kdm6a ^pKO mice. This revealed profound transcriptional changes, with downregulation of acinar‐specific genes (but not duct or islet genes) and upregulation of mesenchymal genes (Fig 5A and B, Dataset EV3). Downregulated genes were associated with metabolic pathways including glutathione, ether lipid, and NRF2 (antioxidant) metabolism (Fig 5C). Upregulated genes in Kdm6a ^pKO mice were enriched in programs regulating cholesterol biosynthesis, extracellular matrix organization, and the innate immune response (Fc gamma R‐mediated phagocytosis) (Fig 5C). GSEA showed prominent enrichment in additional annotations related to oncogenic pathways, including oncostatin M, NFKB, Wnt, EMT, ErbB, and MAPK signaling (Figs 5D and EV3J and K, Dataset EV4). The biological pathways that were down‐ and upregulated in Kdm6a ^pKO mice were generally deregulated in the same direction as in Hnf1a ^aKO (Fig EV3L, Dataset EV5). This suggested that Kdm6a could exert its tumor suppressor function through broadly similar pathways as Hnf1a, namely through the maintenance of acinar differentiated cell programs and inhibition of growth‐promoting pathways.

Figure 5. KDM6A promotes acinar cell differentiation gene programs and suppresses oncogenic pathways.

• A
  Fold change (FC) and significance for transcripts in Kdm6a ^pKO versus control pancreas. Genes significant at q < 0.05 are shown as colored dots above the horizontal line.
• B
  GSEA showing that genes specific to differentiated acinar cells were downregulated in KDM6A‐deficient pancreas, but not genes specific to islets or duct cells. Upregulated genes were enriched in genes specific to mesenchymal cells. Lineage‐enriched genes were obtained from Muraro et al (2016).
• C
  Most enriched functional annotations in genes differentially expressed in Kdm6a ^pKO.
• D
  GSEA showed increased expression of indicated oncogenic pathway genes in Kdm6a ^pKO pancreas.
• E
  KDM6A functions as a transcriptional activator of direct target genes in pancreatic cells. Left: Downregulated, but not upregulated, genes in Kdm6a ^pKO pancreas were enriched for KDM6A binding. P‐values and odds ratios (O.R.) were calculated by Fisher's exact test. Right: Venn diagrams showing overlap of down‐ and upregulated genes with KDM6A binding.
• F
  KDM6A‐bound enhancers and promoters showed increased H3K27me3 and decreased H3K27ac in Kdm6a ^pKO pancreas.
• G
  Genes that were downregulated in Kdm6a ^pKO pancreas (KDM6A‐dependent genes) showed greatest changes in histone marks. Box plots show fold‐changes of H3K27me3 (left) and H3K27ac (right) signals in Kdm6a ^pKO compared to control. Signals were analyzed in KDM6A‐bound regions in promoters and enhancers of genes that were KDM6A‐dependent (gray; n = 420 regions) and KDM6A‐independent (white; n = 8,035 regions). The signals are average values from ChIP‐seq experiments in two biological replicates. The horizontal central line marks the median. Box limits indicate the first and third quartiles, and whiskers extend to highest and lowest data points within 1.5× IQR outside box limits. P‐values were determined by two‐tailed Mann–Whitney U‐test.
• H
  Changes in KDM6A, H3K27me3, and H3K27ac ChIP‐seq profiles in the Ppp4r4‐Serpina10 locus in Kdm6a ^pKO pancreas.
• I
  qPCR of H3K27me3, H3K27ac, and H3K4me1 changes in promoter and enhancer regions highlighted in (H) (R1, R2, R3), and Serpin10 mRNA. Error bars show ± SD, and P‐values are from two‐tailed Student's t‐test, n = 3.
• J
  KDM6A‐bound genes that showed decreased expression in Kdm6a ^pKO pancreas (functional KDM6A binding sites) showed simultaneous gain of H3K27me3 and loss and of H3K27ac.

Shared transcriptomes in Kdm6a ^pKO mice and KDM6A‐mutant human PDAC

The existence of KDM6A mutations in some human tumors allowed us to directly assess the relevance of the transcriptional phenotype of Kdm6a ^pKO mice to human PDAC. GSEA showed that up‐ and downregulated genes from tumors with KDM6A putative loss‐of‐function mutations were concordantly up‐ and downregulated in Kdm6a ^pKO pancreas (Fig EV3M). Furthermore, KDM6A mRNA was decreased in human tumors that showed greatest deregulation of genes downregulated in Kdm6a ^pKO pancreas (Fig EV3N). Likewise, non‐classical PDAC tumors, which exhibit decreased KDM6A expression, showed down‐ and upregulation of deregulated genes in KDM6A ^pKO pancreas (FDR q < 0.0001 and 0.04, respectively; Fig EV3O). The transcriptional changes observed in Kdm6a ^pKO mice are, therefore, directly relevant to a subset of human PDAC tumors.

KDM6A co‐activates acinar differentiated cell programs

To further understand the direct molecular mechanisms whereby KDM6A regulates genetic programs relevant to PDAC, and the potential link to HNF1A, we profiled KDM6A‐bound genomic sites in pancreas from 4‐day‐old mice, and integrated these data with transcriptome changes in Kdm6a ^pKO pancreas. We identified 8455 KDM6A binding sites (Fig EV3P) and confirmed their specificity by showing the absence of binding in Kdm6a ^pKO pancreas (Fig EV3Q and R). The vast majority of KDM6A binding sites were located in active enhancers and promoters (Fig EV3S), and they were enriched in genes that were downregulated in Kdm6a ^pKO mice, but not in those that showed upregulation (Figs 5E and EV3Q and R, and EV1J). KDM6A was, consequently, preferentially bound near acinar‐enriched genes, but not to genes involved in EMT or other pathways that showed induction in Kdm6a ^pKO pancreas. Thus, although KDM6A has been reported to have transcriptional activating and repressive functions (Gozdecka et al, 2018), its predominant direct function in pancreatic cells appears to be the transcriptional activation of gene targets, and indirect inhibition of many other genes.

KDM6A is a H3K27 demethylase, although previous studies have also demonstrated KDM6A functions that are independent from this catalytic activity (Miller et al, 2010; Shpargel et al, 2012, 2017; Vandamme et al, 2012; Andricovich et al, 2018; Gozdecka et al, 2018). We observed that in pancreas from Kdm6a ^pKO mice KDM6A‐bound enhancers and promoters showed increased H3K27me3 (Fig 5F), particularly at genes that were downregulated in Kdm6a ^pKO mice (Figs 5G–I and EV3Q and R). Consistent with KDM6A forming part of complexes that contain histone acetyl transferases P300/CBP and histone methyltransferases MLL1/2 (Tie et al, 2012; Zha et al, 2015; Wang et al, 2017), Kdm6a mutants also showed decreased H3K27ac and H3K4me1 in KDM6A‐bound regions (Figs 5F–I and EV3Q and R). KDM6A‐bound sites, therefore, often showed concomitant H3K27me3 gain and H3K27ac loss in mutant cells (Fig 5J). Collectively, these results indicate that KDM6A has a profound direct influence on chromatin states and transcriptional activity in pancreatic cells.

HNF1A and KDM6A share functional targets in acinar cells

KDM6A is not a sequence‐specific DNA binding protein, and the mechanisms that recruit KDM6A to its genomic sites are poorly understood. We therefore examined KDM6A‐bound sequences to identify candidate transcription factors (TFs) that recruit KDM6A in pancreatic cells, and specifically focused on KDM6A‐bound sites associated with transcriptional changes in Kdm6a ^pKO mice. These functional KDM6A‐bound regions were enriched in in silico‐predicted DNA recognition sequences for canonical acinar TFs such as FOXA, GATA6, NR5A2, and RBPJL, although the most enriched sequence was the HNF1 recognition motif (Fig 6A). Consistently, functional KDM6A‐bound regions were enriched for in vivo binding sites for multiple acinar cell TFs, yet showed the highest enrichment for HNF1A binding (Fig 6B). These findings, therefore, pointed to a strong overlap between KDM6A and HNF1A occupancy in pancreatic cells.

Figure 6. KDM6A and HNF1A regulate shared target genes.

• A
  Motif analysis in functional KDM6A‐bound regions, showing top ten de novo motifs ranked by P‐value determined by HOMER software.
• B
  Co‐binding analysis in functional KDM6A‐bound enhancer and promoter regions revealed that HNF1A was the most enriched co‐bound TF among three other acinar cell TFs. Binding regions of TAL1 in a non‐pancreatic cell type and random binding sites were used as negative controls. P‐values were determined by Fisher's exact test for peak comparisons using all enhancer and promoter regions as background.
• C
  The most downregulated genes in Kdm6a ^pKO pancreas are shown ranked by q‐value and are almost invariably bound by HNF1A and downregulated in Hnf1a ^aKO pancreas, or known to be direct HNF1A‐dependent target genes from other studies (red and purple, respectively).
• D, E
  GSEA analysis on the Hnf1a ^aKO and Kdm6a ^pKO ranked‐ordered gene lists versus their reciprocal up‐ or downregulated gene sets, demonstrated that KDM6A and HNF1A regulate similar genes.
• F
  Expression changes in Hnf1a ^aKO and Kdm6a ^pKO pancreas, showing that genes bound by KDM6A and downregulated in Kdm6a ^pKO pancreas (red dots) were generally downregulated in Hnf1a ^aKO pancreas.
• G
  HNF1A and KDM6A co‐occupy the same regions in Pah, which is downregulated in Hnf1a and Kdm6a knock‐out pancreas.
• H
  Genes that were co‐bound by KDM6A and HNF1A showed greatest downregulation in Kdm6a ^pKO pancreas, compared with KDM6A‐bound genes that were not bound by HNF1A. Box plots show median and IQR of Log₂ TPM fold‐changes and whiskers extend to highest and lowest data points within 1.5× IQR outside box limits. P‐values were determined by two‐tailed Student's t‐tests and n = 4 replicates per condition.

Consistent with KDM6A and HNF1A co‐occupancy, 31 of the 40 most downregulated genes in Kdm6a ^pKO pancreas were also downregulated in Hnf1a ^aKO pancreas or are known to be HNF1A‐dependent target genes (Fig 6C). Accordingly, gene set enrichment analysis (GSEA) showed that genes that were significantly down‐ or upregulated in Kdm6a ^pKO mice were preferentially down‐ or upregulated in Hnf1a ^aKO mice (Fig 6D and E for the converse comparison). We verified downregulation of selected HNF1A/KDM6A co‐occupied genes by quantitative PCR in Hnf1a knock‐out pancreas (Fig EV4A–D) and in Kdm6a knock‐out clonal cell lines (Fig EV5B–F).

Figure EV4. KDM6A binding to target genes is dependent on HNF1A, while HNF1A can bind to chromatin in the absence of KDM6A.

• A–D
  Left: Genome Browser examples of loci co‐bound by HNF1A and KDM6A, showing loss of KDM6A binding in HNF1A‐deficient pancreas (region highlighted in green) and decreased RNA levels in HNF1A‐deficient pancreas. Right: ChIP‐qPCRs showing loss of KDM6A and HNF1A binding in highlighted regions in left and qPCRs show downregulation of target genes in Hnf1a‐KO pancreas. Error bars show SD, and P‐values were determined by two‐tailed Student's t‐test, n = 4 for ChIP‐qPCRs and n = 3 for qPCRs.

Figure EV5. KDM6A dependency of HNF1A‐mediated transcription.

• A
  HNF1A binding to chromatin is unaffected in Kdm6a ^pKO pancreas. Scatterplot showing unchanged HNF1A binding in Kdm6a ^pKO versus control pancreas (e.g., see also Appendix Fig S6A–D).
• B
  Western blot showing KDM6A depletion in two clones from Kdm6a‐KO acinar cell lines.
• C–F
  qPCR in Kdm6a‐KO acinar cell lines (KO1 and KO2) shows reduced expression of HNF1A bound genes, while ChIP‐qPCRs for HNF1A show that its binding to those genes is unchanged when KDM6A is depleted. Selected genes and HNF1A binding regions were from Appendix Fig S6A–D. qPCR data are relative mRNA expression of indicated genes versus Hprt. ChIP‐qPCR values indicate fold enrichment relative to control region. Error bars show ± SD, and P‐values were determined by two‐tailed Student's t‐test.
• G
  GSEA analysis on ranked‐ordered gene list from Kras ^G12D ;Kdm6a ^pKO data from Andricovich et al (2018) versus gene sets from up‐ or downregulated genes in Hnf1a ^aKO and Kdm6a ^pKO pancreas demonstrates that KDM6A and HNF1A regulate similar genes in normal and Kras ^G12D‐transformed pancreas.
• H
  GSEA, comparing rank‐ordered expression data from Kras ^G12D ;Kdm6a ^pKO, Hnf1a ^aKO, and Kdm6a ^pKO mice with gene sets from Andricovich et al (2018), shows that most of the pathways that are enriched in Kdm6a‐deficient pancreatic cancer are dependent on HNF1A and KDM6A function in the non‐tumoral pancreas.

We found that KDM6A‐bound genes that showed downregulation in Kdm6a ^pKO mice were largely downregulated in the pancreas of Hnf1a ^aKO (Fig 6F and G). Importantly, KDM6A‐bound genes showed downregulation in Kdm6a ^pKO if they were co‐bound by HNF1A, but showed more marginal changes when they were not co‐bound by HNF1A, suggesting that HNF1A is critical for KDM6A function in the pancreas (Fig 6H).

Among genes that were co‐bound by HNF1A/KDM6A and downregulated in Kdm6a ^pKO and Hnf1a ^aKO pancreas, we identified several negative regulators of EMT such as Foxa3 (Jagle et al, 2017) and Deptor (Chen et al, 2019) as well as negative MAPK regulators such as Gstp1 (Ruscoe et al, 2001; Xue et al, 2005) and Ptprj (Sacco et al, 2009) (Fig EV3T and U). These findings, therefore, indicated that the derangement of shared biological pathways in Kdm6a ^pKO and Hnf1a ^aKO mice (shown in Fig EV3L) was largely due to the deregulation of common direct target genes.

We further found that transcriptional changes in HNF1A‐ and KDM6A‐deficient pancreas were also recapitulated in Kdm6a‐mutant Kras ^G12D tumors (Andricovich et al, 2018) (Fig EV5G and H). This provided further evidence that HNF1A function is not only relevant to Kdm6a‐regulated programs in the non‐tumoral pancreas, but also to Kdm6a‐deficient cancer.

Collectively, these findings indicate that HNF1A and KDM6A target common genomic sites and regulate shared genetic programs in pancreatic acinar cells.

HNF1A recruits KDM6A to genomic targets in pancreatic cells

To define the molecular mechanisms that link KDM6A and HNF1A function, we performed co‐immunoprecipitation experiments using purified nuclei from a mouse acinar cell line. This showed that HNF1A and KDM6A form part of a common complex in pancreatic cells (Fig 7A). To test whether this interaction could mediate the recruitment of KDM6A by HNF1A to its genomic targets, we performed ChIP‐seq for KDM6A using chromatin from pancreas of wild‐type and Hnf1a ^−/− mice, which had unchanged KDM6A levels (Fig 7B). This showed that KDM6A binding to genomic targets was drastically reduced in Hnf1a ^−/− pancreas (Fig 7C). Importantly, regions with reduced KDM6A binding in Hnf1a ^−/− pancreas were bound by HNF1A and carried HNF1 motifs (Figs 7D–F and EV4A–D). These findings, therefore, showed that HNF1A recruits KDM6A to genomic targets. Consistent with the functional importance of this recruitment, most genes associated with reduced KDM6A binding in Hnf1a‐deficient pancreas were direct HNF1A target genes with decreased expression in Hnf1a‐deficient pancreas (Figs 7G and EV4A–D). By contrast, HNF1A binding was not affected in Kdm6a ^pKO pancreas (Fig EV5A) or in co‐bound genes in Kdm6a knock‐out cells, indicating that, despite the functional interdependence of both proteins, KDM6A is not required for HNF1A binding to chromatin (Fig EV5B–F).

Figure 7. HNF1A recruits KDM6A to activate transcription of its target genes.

• A
  Co‐immunoprecipitation of endogenous HNF1A and KDM6A followed by Western blot demonstrated that HNF1A is in the same complex as KDM6A.
• B
  Western blot showing loss of HNF1A and unchanged KDM6A in Hnf1a ^−/− pancreas.
• C
  Differential binding analysis of KDM6A in Hnf1a ^−/− versus wild‐type pancreas. Pink dots below zero (1,873 sites) show regions with reduced KDM6A binding, and pink dots above zero (118 sites) are regions with increased binding at FDR < 0.05.
• D, E
  Regions that show reduced KDM6A binding in Hnf1a ^−/− chromatin are strongly bound by HNF1A and are highly enriched in HNF1 motifs. P‐values in (D) were calculated with two‐tailed Mann–Whitney U‐test and in (E) with Fisher's exact test.
• F
  KDM6A binding is markedly reduced in HNF1A‐ and KDM6A‐co‐bound regions in Hnf1a ^−/− pancreas, but not in other KDM6A‐bound regions.
• G
  Genes that loose KDM6A binding in Hnf1a‐mutant pancreas are predominantly downregulated in Hnf1a ^aKO pancreas and are direct HNF1A target genes (red dots).
• H
  Summary model depicting that HNF1A recruits KDM6A to genomic binding sites, activating an acinar differentiation program that indirectly suppresses core oncogenic pathways. Defective HNF1A or KDM6A function results in failure of this shared program, with increased activity of pathways that, in the presence of KRAS mutations, promote high‐grade non‐classical PDAC with sarcomatoid features.

These results, therefore, show that HNF1A and KDM6A interact in a common complex, and this enables the recruitment of KDM6A to its functional targets in pancreatic cells. Collectively, these findings suggest a mechanistic model (Fig 7H) that explains the overlapping genomic phenotypes of HNF1A and KDM6A deficiency in genetic mouse models and human tumors, and their role in the determination of PDAC subtypes.

Discussion

In this study, we provide a direct genetic demonstration that HNF1A, which encodes for a homeodomain transcription factor best known for its causal role in autosomal dominant diabetes (Yamagata et al, 1996), has a tumor‐suppressive function in the exocrine pancreas. We further provide epigenomic, biochemical, and genetic evidence that mechanistically link HNF1A function in pancreatic exocrine cells to KDM6A, an established tumor suppressor (van Haaften et al, 2009). These results in turn indicate that the tumor‐suppressive function of KDM6A is tightly linked to its role as a transcriptional co‐regulator of epithelial differentiation. We demonstrate that HNF1A deficiency partially phenocopies KDM6A‐driven tumorigenesis in mice, and show that HNF1A/KDM6A deficiency is a prominent feature of the genetic programs of non‐classical human PDAC tumors. These findings, therefore, shed light on mechanistic underpinnings of emerging sub‐classifications of PDAC subtypes.

Our analysis of Kdm6a‐deficient Kras ^G12D mice is consistent with a recent report showing aggressive PDAC in Kdm6a‐deficient Kras mutant mice (Andricovich et al, 2018), and extend those studies by defining the genetic programs that are directly regulated by KDM6A in the differentiated pancreatic exocrine cells that give rise to PDAC. The KDM6A‐deficient transcriptome phenotype observed in non‐tumoral pancreas was broadly consistent with that reported in tumors from Kdm6a‐deficient Kras mutant mice (Andricovich et al, 2018). Our integrated analysis of genomic binding sites indicated that the positive regulatory effects of KDM6A (and HNF1A) on acinar differentiation genes are largely mediated through direct mechanisms. By contrast, genes associated with EMT, remodeling of the extracellular matrix, and oncogenic pathways that were upregulated in mutant cells were not frequently bound by HNF1A and KDM6A, whereas both proteins bound to known negative regulators of EMT. These findings strongly suggest that the suppression of oncogenic and EMT transcriptional programs is predominantly indirect.

Our studies, therefore, exposed a molecular mechanism whereby HNF1A recruits KDM6A to activate differentiation programs of pancreatic exocrine epithelial cells, thereby suppressing oncogenic and EMT pathways (Fig 7H). This mechanism aligns well with the observation that Kdm6a‐ and Hnf1a‐deficient tumors in mice featured sarcomatoid architecture. We further showed that HNF1A‐ and KDM6A‐dependent pancreatic programs are defective in non‐classical PDAC tumors, variously defined as human basal, quasimesenchymal, or squamous molecular types (Collisson et al, 2011; Moffitt et al, 2015; Bailey et al, 2016; Cancer Genome Atlas Research Network, Electronic Address Aadhe, Cancer Genome Atlas Research N, 2017, Puleo et al, 2018). Interestingly, abnormalities in the HNF1A/KDM6A‐dependent pancreatic program in human tumors were associated with either somatic KDM6A mutations, or to decreased expression of KDM6A or HNF1A, in consonance with the proposed molecular mechanism.

These findings indicate that the disruption of HNF1A‐ and KDM6A‐dependent programs not only promotes the formation of PDAC, but is also instructive for the definition of PDAC subtypes. HNF1A is expressed in normal acinar cells, the presumed cell of origin of most PanIN precursors and PDAC (Kopp et al, 2012; von Figura et al, 2014), but not in duct cells. The results suggest that in the presence of other oncogenic events, defective HNF1A‐ or KDM6A‐dependent gene regulation can promote the development of an undifferentiated high‐grade PDAC cellular phenotype (Fig 7H). More generally, our results open avenues to study the precise manner in which HNF1A‐ and KDM6A‐dependent regulation of epithelial differentiation programs fails and thereby contributes to the development of PDAC.

Materials and Methods

Animal studies

Animal experimentation was carried out in compliance with the EU Directive 86/609/EEC and Recommendation 2007/526/EC regarding the protection of animals used for experimental and other scientific purposes, enacted under Spanish law 1201/2005.

The Kdm6a ^LoxP (Shpargel et al, 2012), Pdx1 ^Cre (Hingorani et al, 2003), LSL‐Kras ^G12D (Jackson et al, 2001), Hnf1a ^−/− (Lee et al, 1998), and Ptf1a ^Cre (Kawaguchi et al, 2002) mouse lines have been described.

To generate a conditional Hnf1a allele, a floxed Hnf1a exon 2 was generated in C57Bl/6N JM8.F6 embryonic stem cells (ES cells) (Pettitt et al, 2009) by homologous recombination. Briefly, ES cells were electroporated with a targeting plasmid construct linearized by PmeI and MluI, containing the floxed exon and a PGK/Neomycin selection cassette flanked by FRT recombination sites, and then selected with Geneticin. Several correctly targeted ES cell clones were identified by Southern blot using external probes detecting a 9.2 kb 5′ ScaI fragment, an 8.7 kb 3′ BamHI fragment from the recombined allele, and a neo probe detecting a unique 11.7 kb SpeI Neomycin fragment indicating no additional construct integrations. The presence of the distal LoxP site, upstream of exon 2, was confirmed by PCR using the primers indicated in Table EV1. Correctly targeted ES cell clones were injected into C57BL/6BrdCrHsd‐Tyrc morulae (E2.5) to create chimeric mice that transmitted the recombined allele through the germ line. The PGK/Neo cassette was excised by crossing heterozygous mice with Tg.CAG‐Flp mice (Rodriguez et al, 2000), which express the FLPe recombinase in the germ line to create the Hnf1a conditional knock‐out allele. Lines with floxed alleles without Cre, Cre lines without floxed alleles, Pdx1 ^Cre ;LSL‐KRAS ^G12D or wild‐type littermates served as controls. Gender was considered as a variable and included in the study design. Genotype was confirmed by PCR using the primers provided in Table EV1.

Prior to decapitation, mice were anesthetized using isoflurane (Zoetis), and pancreas was collected quickly and placed in paraformaldehyde 4% (Sigma‐Aldrich) for paraffin embedding or processed for RNA isolation.

Statistical methods

Statistical tests are described throughout the materials and method section and mentioned in figure legends and main text where appropriate. Two‐tailed Student's t‐test was used to compare two sets of data from qPCR experiments. For comparisons between two independent groups of genome‐wide data with continuous variables, we used two‐sided Mann–Whitney U‐test. Non‐parametric Kruskal–Wallis test was used to assess more than two groups with dependent values, including RNA‐seq and microarray data. Fisher's exact test was used to determine whether there were nonrandom associations between two categorical variables. Data are presented as means with their standard deviations, or as box plots with median, IQR, and whiskers that extend to data points that are < 1.5× IQR away from the first and third quartile. Outliers are defined as values larger than the third quartile by at least 1.5× IQR, or smaller than the first quartile by at least 1.5× IQR.

Histology, TMA preparation, and Immunohistochemistry

Fixation of mouse pancreas was performed using paraformaldehyde 4% (Sigma‐Aldrich) or 4% buffered formalin overnight at 4°C followed by dehydration using increasing concentrations of ethanol. Dehydrated samples were embedded in paraffin blocks and cut in 4‐μm sections. Sections were then deparaffinized and rehydrated. For hematoxylin and eosin analysis, the pancreata were stained, dehydrated, mounted with DPX (Panreac), and photodocumented using optical microscopy (Olympus BX41TF).

Tissue samples for tissue micro arrays (TMAs) were obtained from patients who underwent pancreatic resection for PDAC (n = 223) at the Department of Surgery of the University Hospital of Duesseldorf, Germany. The samples were fixed in 4% formalin and subsequently routinely embedded in paraffin. Three tumor samples per case (two samples from the tumor center, one from the periphery) with a 1‐mm core size were selected and assembled into the TMA (Manual Tissue Arrayer MTA‐1, Beecher Instruments, Inc., Sun Prairie, WI, USA). The use of human tissue samples was approved by the local ethics committee at the University Hospital of Duesseldorf, Germany (study number 5387). The TMAs were then cut in 2‐μm sections for subsequent procedures.

Immunohistochemistry on mouse pancreas or TMAs was performed according to manufacturer's instructions using VectaStain^® ABC HRP Rabbit IgG kit (Vector Laboratories, Peterborough, UK). Epitope retrieval was achieved by pretreatment with boiling citrate buffer (10 mM citric acid, pH6) for 15 min using a pressure cooker. Endogenous peroxidase and protein blocking was performed with 3% H₂O₂ diluted in PBS for 10 min and with 1% BSA, 10% normal goat serum (Abcam, Cambridge, UK), and 0.1% Triton X‐100 (Merck KGaA, Darmstadt, Germany) for 60 min. Anti‐HNF1A and anti‐KDM6A stainings were performed at a dilution of 1:250 (Anti‐HNF1A, Abcam ab204306, Cambridge, UK), 1:200 (Anti‐HNF1A, Cell Signaling Technology, 89670, Leiden, The Netherlands), and 1:100 (Anti‐UTX, Cell Signaling Technology 33510S, Denver, USA), respectively. Diaminobenzidine (Abcam ab64238, Cambridge, UK) was used as chromogen to visualize protein expression, counterstaining was achieved with Mayer's hemalum solution (Merck KGaA, Darmstadt, Germany).

Analysis of TMAs

The semi‐quantitative analysis of the stained sections was done by light microscopy according to the immunoreactive score (IRS) of Remmele and Stegner (Remmele & Stegner, 1987). Both the staining intensity and the number (%) of positive stained cells were evaluated for each case. If more than one tumor core was available in one case, a mean value was calculated. Staining intensity was scored as no staining (0), weak staining (1), moderate staining (2), or strong staining (3). The proportion of stained cells was scored as follows: no staining = 0, 1–9% = 1, 10–50% = 2, 50–80% = 3, > 80% = 4. The IRS was calculated by multiplication of the intensity value and the percentage value, resulting in a score between 0 and 12. Finally, IRS results were classified into four groups as follows: 0–1 = negative (group 0), 2–3 = weak (group 1), 4–8 = moderate (group 2), 9–12 = strong (group 3). Each tumor case was graded according to UICC and WHO classifications. Two cases with undetermined grade (Gx) were omitted from the analysis. The rest of the cases were either scored as G2 or G3 grades. Contingency table analysis and Chi‐square test were used to evaluate the association between IRS classifications and tumor grades.

Immunofluorescence

Immunofluorescence analysis of murine pancreata was performed on paraformaldehyde‐fixed paraffin‐embedded tissue cut into 4‐μm sections, as described with modifications (Solar et al, 2009). Antigen unmasking was performed boiling sections on 10 mM citrate buffer pH 6 for 10 min, and tissue was permeabilized with PBS 0.5% Triton X‐100 (Sigma). Blocking was performed by 1‐h incubation in humid chamber with 3% of normal donkey serum (Vendor) in blocking solution (DAKO). Slides were incubated in humid chamber with primary antibodies overnight at 4°C and washed in PBS 0.2% Triton X‐100. Slides were incubated with secondary antibodies for 1 h in humid chamber at room temperature, washed for 5 min in PBS 0.2% Triton X‐100, and mounted using fluorescence mounting medium (DAKO). Stained slides were analyzed using a confocal laser scanning microscope (Leica CTR6500). Primary antibodies used are indicated in Table EV2.

Western blotting and co‐immunoprecipitation

All protein isolation steps were performed at 4°C or on ice, and protein concentrations were assessed with the Bradford assay (Bio‐Rad). Protein from total pancreas lysates was isolated from a small fragment of mouse pancreas and minced in lysis buffer (20 mM Tris–HCl, pH 7.5, 150 mM NaCl, 1 mM Na₂EDTA, 1 mM EGTA, 1% Triton X‐100, 2.5 mM sodium pyrophosphate, 1 mM b‐glycerophosphate, 1 mM Na₃VO₄, 1 μg/ml leupeptin containing 1× phosphatase inhibitor cocktail, Sigma‐Aldrich, and 4× EDTA containing complete protease inhibitor cocktail, Roche). Lysates were freeze‐thawed twice, cleared at 15,000 g for 15 min, and then, the supernatant was recovered. For Western blots, lysate supernatants were denatured in 2× NuPAGE LDS sample buffer (Thermo Fisher) containing 0.1 M DTT at 95°C for 5 min. Equal amount of lysates were loaded in NuPAGE 4–12% Bis‐Tris Protein Gels (Life Technologies) and run at 150V for 90 min. Gels were transferred to 0.2‐μm pore PVDF membrane (Life Technologies) and blocked with 5% fat‐free milk powder for 60 min and probed overnight (ON) with primary antibody. Horseradish peroxidase conjugated secondary antibodies were used at 1:10,000 in BSA 5% and incubated 1 h at RT. The blot was developed using ECL detection reagent (Amersham ECL, GE Healthcare). The pancreatic acinar cell line 266‐6 (ATCC CRL‐2151) was grown in DMEM (Thermo Fisher) with 10% fetal calf serum, 1% pen/strep, and passaged twice per week. For co‐immunoprecipitation, the cells were washed in ice‐cold PBS and resuspended in hypotonic lysis buffer (10 mM HEPES, pH 7.9, with 1.5 mM MgCl₂ and 10 mM KCl). After plasma membrane disruption, the lysate was centrifuged at 4,000 g for 5 min and the nuclei lysed for 20 min with gentle shaking in nuclear extraction buffer (20 mM HEPES, pH 7.9, with 1.5 mM MgCl₂, 0.42 M NaCl, 0.2 mM EDTA, and 25% (v/v) Glycerol) before removing cell debris by centrifugation at 15,000 g for 15 min. The cleared nuclear lysate was equilibrated to 150 mM NaCl and pre‐cleared with Dynabeads Protein G (Thermo Fisher) for 1 h at 4°C with rotation, and 4 mg of extract was incubated overnight with 2 μg of anti‐KDM6A, anti‐HNF1A, or normal rabbit IgG. Immune complexes were precipitated with 20 μl Dynabeads for 2 h and washed four times with IP wash buffer (10 mM Tris–HCl, pH 8, 150 mM NaCl, 0.1% NP‐40, and 1 mM EDTA) supplemented with protease inhibitor. Proteins were eluted with 2× NuPAGE LDS sample buffer (Thermo Fisher) containing 0.1 M DTT at 95°C for 5 min and analyzed by Western blotting.

RNA isolation

For gene expression profiling, we implemented a method for purification of intact pancreatic RNA according to a modified guanidinium salts method (MacDonald et al, 1987). Briefly, a small piece of pancreatic tissue was immediately homogenized with a Polytron (VWR) in pre‐cooled 4 M guanidinium thiocyanate with 112 mM beta‐mercaptoethanol, centrifuged at 5,000 g for 5 min at 4°C to remove insoluble debris. RNA was precipitated from the supernatant with pre‐cooled 75% ethanol, 0.1M potassium acetate, pH 5.5, and 75 mM acetic acid at −20°C for 2 h. The precipitate was pelleted by centrifugation at 10,000 g for 10 min at 4°C and resuspended at room temperature in 7.5 M guanidinium HCl and 10.5 mM beta‐mercaptoethanol. The RNA was re‐precipitated twice with 0.1 M potassium acetate, pH 5.5, and 50% ethanol to remove residual RNases, followed by purification with TRI Reagent RNA Isolation Reagent (Sigma‐Aldrich). The concentration and quality of RNA was measured with a NanoDrop spectrophotometer (ND‐1000, Thermo Scientific) and an Agilent 2100 Bioanalyzer. RNA integrity numbers ranged from 7.8 to 9.3.

Alternatively, we used a previously described protocol (Cobo et al, 2018).

RNA‐seq

One μg of total RNA was used to make RNA‐seq libraries using the Truseq Stranded mRNA Sample Prep Kit (Illumina) following manufacturer's instructions and sequenced by an Illumina HiSeq2500 platform with single‐end reads of 50 bases. Conversion to FASTQ read format was done using Illumina's bcl2fastq algorithm. Four or three pancreases from, respectively, 4‐day‐old or adult female mice of each genotype were analyzed: Kdm6a ^LoxP/LoxP and Pdx1 ^Cre;Kdm6a ^LoxP/LoxP or Ptf1a ^Cre and Ptf1a ^Cre;Hnf1a ^LoxP/LoxP. Raw RNA‐seq reads were aligned to the mouse transcriptome (a combined build of cDNA and ncRNA from Mus musculus v.GRCm38.p5, release 87) and quantified using Salmon v0.7.2 (Patro et al, 2017), with the following parameters: salmon quant –gcBias, – libType A, and –fldMean and –fldSD. The latter two were set to the average size and standard deviation of the fragment length distribution of the given library, respectively. The counts were used in DEseq2 (v1.14.1) (Love et al, 2014) with R (v3.3.3) (RStudio, 2015) for normalization and identification of significant differential expression (q < 0.05) between controls (n = 4) and mutants (n = 4). Only genes with log2 (basemean) > 2.5 were retained for downstream analysis. For visualization, STAR (v2.3.0) (Dobin et al, 2013) was used to align reads to the mouse genome assembly GRCm38.p5 (mm10) and resulting BAM files were converted to bigwig with reads per genomic content (RPGC) normalization by Deeptools (v2.4.2) (Ramirez et al, 2016).

ChIP

ChIP experiments were performed on pancreas from 4‐day‐old or 3‐week‐old female mice. A Polytron was used to quickly mince 1–2 dissected pancreas in PBS with protease inhibitors followed by crosslinking with 1% formaldehyde for 10 min at room temperature on a rotator and then quenched with 0.125 M glycine for 5 min. Tissue pieces were washed in PBS and lysed in 130 μl lysis buffer (2% Triton X‐100, 1% SDS, 100 mM NaCl, 10 mM Tris–HCl, pH 8.0, 1 mM EDTA) with protease inhibitor on ice for 10 min and finally resuspended with a 30G needle syringe. The chromatin preparation was then sonicated to fragments enriched in the size range of 150–500 bp using S220 Focused Ultrasonicator (Covaris) and centrifuged at 18,000 g for 10 min at 4°C to pellet insoluble material. The supernatant was incubated for 1 h with 350 μl RIPA‐LS buffer (10 mM Tris–HCl, pH 8.0, 140 mM NaCl, 1 mM EDTA, pH 8.0, 0.1% SDS, 0.1% Na‐Deoxycholate, 1% Triton X‐100) with protease inhibitor and 20 μl Dynabeads Protein G for pre‐clearing. After preserving 1% as an input sample, the pre‐cleared chromatin was incubated with 2 μg antibody, 50 μl 10% BSA, and 5 μl tRNA (10 mg/ml) overnight at 4°C with rotation. Immune complexes were retrieved with 20 μl BSA blocked Dynabeads Protein G for 2 h and washed 2× with RIPA‐LS, 2× with RIPA‐HS (10 mM Tris–HCl, pH 8.0, 500 mM NaCl, 1 mM EDTA, pH 8.0, 0.1% SDS, 0.1% Na‐Deoxycholate, 1% Triton X‐100), 2× with RIPA‐LiCl (10 mM Tris–HCl, pH 8.0, 250 mM LiCl, 1 mM EDTA, pH 8.0, 0.5% IGEPAL, 0.5% Na‐Deoxycholate), and 1× with 10 mM Tris, pH 8.0.

ChIP‐seq libraries were prepared directly on the bead‐bound chromatin according to a ChIPmentation procedure that is optimized for low cell number samples (Schmidl et al, 2015). Briefly, beads were resuspended in 25 μl of tagmentation reaction buffer (10 mM Tris, pH 8.0, 5 mM MgCl₂, 10% v/v dimethylformamide) containing 1 μl Tagment DNA Enzyme from the Nextera DNA Sample Prep Kit (Illumina) and incubated at 37°C for 10 min followed by 2× washing in RIPA‐LS and TE (10 mM Tris–HCL, pH 8.0, 1 mM EDTA, pH 8.0). The DNA was then de‐crosslinked and eluted from the beads in ChIP elution buffer (10 mM Tris–HCL, pH 8.0, 5 mM EDTA, pH 8.0, 300 mM NaCl, 0.4% SDS) with proteinase K at 55°C for 1 h and 65°C overnight. ChIP DNA was purified with Qiagen MinElute kit (QIAGEN), and enrichment cycles for library amplification were assessed by qPCR. The libraries were PCR amplified with KAPA HiFi Hotstart Ready mix (Sigma‐Aldrich) and barcoded Nextera custom primers (Schmidl et al, 2015) and finally size‐selected (250–350 bp) using Agencourt AMPure XP beads (Beckman Coulter) and validated using the Agilent High Sensitivity DNA Kit with Agilent 2100 Bioanalyzer. Equimolar quantities of libraries were combined for multiplexing to obtain 40 million reads per library. ChIP‐seq libraries were sequenced using a NextSeq platform with single‐end reads of 75 bases.

For ChIP‐qPCR, DNA was de‐crosslinked and eluted from the beads in ChIP elution buffer (10 mM Tris–HCL, pH 8.0, 5 mM EDTA, pH 8.0, 300 mM NaCl, 0.4% SDS) with proteinase K at 55°C for 1 h and 65°C overnight, and purified by phenol chloroform extraction according to standard procedures. Quantitative PCR was then performed on a QuantStudio 6 Flex Realtime PCR machine (Applied Biosystems) using GoTaq qPCR reagent (Promega). Fold enrichments were visualized relative to input and negative control regions using primers available in Table EV1.

ChIP‐seq alignment and peak calling

ChIP‐seq reads were aligned to the mouse genome (M. musculus, UCSC mm10) using Bowtie2 followed by exclusion of mapping quality scores < 30 by Samtools (v1.2) (Li et al, 2009) and removal of duplicate reads with Picard (v2.6.0) and blacklisted regions with Bedtools (v2.13.3) (Quinlan & Hall, 2010). The processed BAM files were converted to bigwig files using Deeptools. Peaks for histone marks were called with MACS2 (v.2.1.0) with settings: –extsize = 300 –q 0.05 –keep‐dup all –nomodel –broad and for KDM6A and transcription factors with MACS1.4 (v1.4.2) (Zhang et al, 2008) using default parameters and P < 10⁻¹⁰. Enriched regions were scored against matching input libraries. All experiments were done as biological replicates. For histone marks and KDM6A ChIPs, only enriched regions that overlap in replicates were retained as consistent peaksets. Publically available datasets were processed identically. For HNF1A ChIPs, peaks from two replicates that overlap by at least one base were merged and replaced with a single peak. Library information of public and internal datasets is provided in Table EV3.

Integrative analysis

Active promoters were defined as consistent H3K27ac peaks occurring within 1 kb of an annotated transcription start site (GENCODE GRCm38.p5), and all remaining peaks were considered as active enhancers. HNF1A and KDM6A peaks were annotated with HOMER (v4.10.3) (Heinz et al, 2010) to define TSS‐proximal, TSS‐distal, intronic, exonic, and intergenic regions, and were assigned to genes using the “closest” function in Bedtools with default parameters. Fisher's exact test in R was used to define enrichment of HNF1A‐ and KDM6A‐bound regions and active promoter and enhancer regions, or differentially expressed genes in the Hnf1a ^aKO and Kdm6a ^pKO datasets. Expressed genes (Log2 TPM > 2.5) were used as background. Aggregation plots were calculated with Deeptools. Briefly, all KDM6A‐bound regions were extended ± 10 kb or ± 5 kb from the center of the peaks for analysis of H3K27me3 or H3K27ac signals, respectively. The resulting regions were then divided into 10‐bp bins, and the mean value in each bin was calculated based on the normalized values given in the bigwig files for the histone marks. To assess the relationship between functional KDM6A binding sites and KDM6A‐dependent histone modification changes, the Log₂ fold change of the averaged normalized H3K27me3 signals was plotted against the Log₂ fold change of H3K27ac signals in each KDM6A‐bound region. Differential binding of KDM6A and HNF1A was analyzed with DiffBind (v2.6.6) (Ross‐Innes et al, 2012) using the DEseq2 method without control input read counts. Differential peaks with FDR < 0.05 were defined as significant.

Functional annotation and gene set enrichment analysis (GSEA)

The ENRICHR tool (Chen et al, 2013; Kuleshov et al, 2016) was used to functionally annotate differentially expressed genes. Significantly enriched signatures were identified using Bonferroni corrected P < 0.01. The GSEA pre‐ranked analysis (Subramanian et al, 2005) was used to determine whether predefined sets of genes show significant concordant differences with gene expression changes in RNA‐seq datasets. For GSEA, we used gene sets from the Molecular Signature Database (Liberzon et al, 2015), custom‐made sets of differentially expressed genes from our dataset, gene sets of pancreatic cell types (Muraro et al, 2016), and human PDAC (Collisson et al, 2011; Moffitt et al, 2015; Bailey et al, 2016). All were analyzed with the parameters: numbers of permutations = 10,000 and scoring scheme = weighted.

Motif analysis

To find enriched de novo transcription factor DNA‐binding profiles, we used HOMER (Heinz et al, 2010) and searched for 6‐, 8‐, 10‐, and 12‐bp sized motifs in ± 250‐bp regions surrounding the center of KDM6A‐bound regions within promoters and enhancers that were functionally associated with transcriptional changes. As background sequences, we used all H3K27ac regions. DNA motifs were annotated with HOMER and retrieved if the HOMER score was more than 0.7. Centrimo from the MEME suite (Bailey & Machanick, 2012) was used to compute enrichment of known HNF1A motifs in HNF1A‐bound regions (Fig EV1I) and in KDM6A‐bound regions (Fig 7E). Position weight matrices for HNF1 were retrieved from JASPAR (Sandelin et al, 2004) (http://jaspar.genereg.net/) and HOCOMOCO (Kulakovskiy et al, 2018) (http://hocomoco11.autosome.ru/) databases.

Data visualization

Bigwig files from RNA‐seq and DNA‐seq experiments were visualized in the Genome Browser (http://genome.ucsc.edu/). R packages were used to create boxplots (BoxPlotR) (Spitzer et al, 2014), violinplots (PlotsOfData) (Postma & Goedhart, 2019), and bar and scatter plots (ggplot2) (Wickham, 2009). Enrichment plots from GSEA were drawn by Genepattern (Reich et al, 2006). Aggregation plots and heatmaps of ChIP experiments were drawn with Deeptools. GraphPad Prism 6 was used for bar graphs, calculation of contingency tables and piecharts. To draw heatmaps, we used Morpheus (https://software.broadinstitute.org/morpheus).

GTEx analysis

Quantile‐normalized read counts of 328 pancreas samples from GTEx version 8 dataset (GTEx_Analysis_2017‐06‐05_v8_RNASeQCv1.1.9_gene_tpm.gct.gz) were retrieved from https://gtexportal.org/home/datasets. Samples were ranked according to HNF1A levels, and fold differences between bottom and top expression deciles (n = 33 per group) were calculated. We then examined differential expression of orthologous genes in Hnf1a ^aKO versus control pancreas. A random list of 717 genes controlled for similar expression levels was used for comparison.

Analysis of human PDAC genomic data

Data acquisition and preprocessing

RSEM normalized RNA‐seq data from the TCGA‐PAAD cohort was downloaded from the Firehose browser (https://gdac.broadinstitute.org/). Molecular subtypes and purity class were assigned to each sample as defined by the Cancer Genome Atlas Research Network (Cancer Genome Atlas Research Network, Electronic Address Aadhe, Cancer Genome Atlas Research N, 2017). Samples with high purity (76 samples) were retained for further analysis. Quantile‐normalized background‐adjusted array expression data from the ICGC‐PACA‐AU cohort were downloaded from the ICGC data portal (https://dcc.icgc.org/). Probe sets (ILLUMINA HumanHT 12 V4) were assigned to gene names. Multiple probe sets mapping to the same gene were collapsed using the collapseRows algorithm from the WGCNA package (Langfelder & Horvath, 2008) keeping the gene‐probe set combination with highest variance across all samples. Purity class was assigned to each sample according to Bailey et al (2016) retaining 121 high‐purity samples for further analysis.

Consensus clustering

Samples from the ICGC‐PACA‐AU study were classified into molecular subtypes based on signature genes from Collisson et al (2011), Moffitt et al (2015), or Bailey et al (2016). Out of the 62 PDA assigner genes identified by Collisson et al, 57 matched with our preprocessed data. We also selected 50 genes with the highest gene weights for the basal and classical PDAC subtypes (100 genes in total) from Moffitt et al's gene factorization analysis, out of which 94 genes matched with our data. From Bailey et al's 613 differentially expressed genes across their defined classes (Squamous, ADEX, Immunogenic, and Pancreatic Progenitor), we retained 457 matching genes. We then applied consensus clustering to the ICGC‐PACA‐AU data with the signature genes from the three studies using ConsensusClusterPlus (CCP) v.1.24.0 (Wilkerson & Hayes, 2010). Each gene expression profile was first Log₁₀‐transformed and median centered. We next performed 1,000 iterations of CCP using Pearson correlation as the distance metric, partitioning around medoids, and a random gene and sample fraction of 90% in each iteration. We verified that the obtained groups and their gene expressions reflected the up and down relationships for each gene in each group described in the three studies. We next used human orthologs of differentially expressed genes from our study in Hnf1a‐deficient mouse pancreas to cluster the TCGA‐PAAD and ICGC‐PACA‐AU cohorts. We applied consensus clustering with non‐negative matrix factorization, using Pearson distance metrics and 2,000 descent iterations (Kuehn et al, 2008), and thereby identified a group of samples with similar expression signature to the Hnf1a ^aKO pancreas. To generate expression heatmaps, we identified genes that were differentially expressed across the clusters (FDR < 0.05) with a SAM multiclass analysis (samr v.2.0) (Tusher et al, 2001), Z‐score transformed each row of the matrix, then used Morpheus (https://software.broadinstitute.org/morpheus) to cluster only the rows, using a one minus Pearson correlation distance metric, average linkage method, and Hierarchical clustering.

To identify tumors with most pronounced HNF1A‐deficient function, we first defined a gene set containing 106 human orthologs of HNF1A‐bound genes that were downregulated in Hnf1a ^aKO mouse pancreas. We then interrogated that behavior of this gene set in every tumor sample. This was carried out by GSEA (see also GSEA subheading), using the HNF1A‐dependent gene set, and testing the enrichment in each tumor's gene lists rank‐ordered by differential expression of all genes in the tumor versus the median expression of all genes across all samples. This analysis revealed 3 groups of tumor samples: (i) HNF1A loss of function (LoF) with NES < 0, P < 0.05, (ii) Control 1 with NES < 0, P > 0.05, and (iii) Control 2 with NES > 0.

KDM6A mutations

Information on KDM6A mutations was retrieved from the ICGC data portal (https://dcc.icgc.org) and Bailey et al (2016). 15 samples in the high‐purity ICGC‐PACA‐AU cohort had KDM6A mutations, which we separated into three groups according to mutation types defined by ICGC: (i) Deletions (small ≤ 200 bp), (ii) Substitutions (single base), (iii) Insertions (small ≤ 200 bp. The fourth group, (iv) Structural Variants, was from Bailey et al Mutations defined by ICGC as having “high” functional impact in (1–3) and having a “loss‐of‐function” consequence in (iv) were all frame‐shift mutations. We considered those mutations as being functional deleterious and defined samples with such genetic alterations in KDM6A as KDM6A loss‐of‐function (LoF) tumors. The rest were defined as samples with KDM6A mutations of unknown consequence. Significantly mutated genes (SMGs) were from Bailey et al (2016) SMG analysis using Intogen, Mutsig, and HOTNET. Genes considered significantly mutated were significant in > 1 analysis.

Author contributions

MK and JF conceived the study. MK, MAM, and EB analyzed mouse phenotypes. AB designed the Hnf1a mutant allele that was generated by SO. MK and IM performed bioinformatic analysis. NP and VG maintained mouse colonies. KBS and TM created the Kdm6a mutant model. EV provided the LSL‐KrasG12D mice and MV provided samples. IE carried out mouse histopathological analysis. IE, LH, MS, SAS, and WTK analyzed TMAs. FXR provided expertise and resources for conducting experiments during the revision phase. MK, EB, FXR, IE, and JF wrote the manuscript with input from all authors.

Conflict of interest

The authors declare that they have no conflict of interest.

Supporting information

Appendix

Expanded View Figures PDF

Table EV1

Table EV2

Table EV3

Dataset EV1

Dataset EV2

Dataset EV3

Dataset EV4

Dataset EV5

Review Process File

The EMBO Journal (2020) 39: e102808

Data availability

RNA‐seq and ChIP‐seq data sets generated here are available in the ArrayExpress database at EMBL‐EBI (http://www.ebi.ac.uk/arrayexpress) under accession numbers E‐MTAB‐7944 and E‐MTAB‐7945.