A chromatin basis for cell lineage and disease risk in the human pancreas

SUMMARY

Understanding the genomic logic that underlies cellular diversity and developmental potential in the human pancreas will accelerate growth of cell replacement therapies and reveal genetic risk mechanisms in diabetes. Here, we identified and characterized thousands of chromatin regions governing cell-specific gene regulation in human pancreatic endocrine and exocrine lineages, including islet β-cells, α-cells, duct and acinar cells. Our findings have captured cellular ontogenies at the chromatin level, identified lineage specific regulators potentially acting on these sites, and uncovered hallmarks of regulatory plasticity between cell types that suggest mechanisms to regenerate β-cells from pancreatic endocrine or exocrine cells. Our work shows that disease risk variants related to pancreas are significantly enriched in these regulatory regions, and reveals previously unrecognized links between endocrine and exocrine pancreas in diabetes risk.

Graphical Abstract

INTRODUCTION

The pancreas is a vital organ whose disorders include diabetes mellitus, pancreatitis, cystic fibrosis and adenocarcinoma, estimated to affect over 10% of the world’s population (Lévy et al., 2014; NCD Risk Factor Collaboration, 2016; Wilschanski and Novak, 2013). During development, pancreatic epithelia are thought to derive from a multipotent progenitor formed in primitive ductal epithelium. Advances in the past two decades have identified cellular, genetic, signaling and molecular pathways that lead to formation of pancreatic exocrine and endocrine cells, the two major lineages in the pancreas. The exocrine pancreas is comprised of acinar cells that produce zymogens that hydrolyze macromolecules to aid digestion, and ductal cells that form branched tubules whose principal products include bicarbonate. The endocrine pancreas is comprised of hormone-secreting epithelial cells organized in structures called Islets of Langerhans. Islet cells synthesize and secrete hallmark hormone products, and include α-cells (glucagon), β-cells (insulin) and δ-cells (somatostatin). Our understanding of the mechanisms underlying development is largely based on studies of rodents and other vertebrate models. While the adult mouse and human pancreas share similar functions and morphological features, much remains to be discovered about the mechanisms regulating human endocrine and exocrine cell development and diversification (Arda et al., 2013; Benitez et al., 2014; McKnight et al., 2010).

Enhancers are noncoding genomic elements that harbor regulatory information for specifying gene expression patterns, and prior studies show that enhancer activity is a primary determinant of cell type specificity and diversity (Heintzman et al., 2009; Heinz et al., 2015; Stergachis et al., 2013). A growing body of evidence strongly argues that thousands of these regions in mammalian genomes function to govern spatial and temporal control of gene expression, and demonstrates a direct link between abnormalities in enhancer function and disease (Bauer et al., 2013; Chatterjee et al., 2016; Maurano et al., 2012; Weedon et al., 2014). Understanding how regulatory programs are organized and deployed through enhancers and other cis-regulatory elements during cellular differentiation could accelerate efforts to create regenerative therapies for pancreatic diseases like diabetes. For example, generation of replacement β-cells from renewable sources like multipotent human stem cells remains a focus of intensive effort (Pagliuca et al., 2014; Rezania et al., 2014; Russ et al., 2015). A growing consensus is that this effort could benefit from improved understanding of the gene regulatory mechanisms governing native human islet cell fate, diversification and function. Recent studies in mice suggest that endocrine cells like pancreatic a-cells or δ-cells can interconvert toward a β-cell fate, after profound β-cell destruction or genetic loss-of-function (Chakravarthy et al., 2017; Chera et al., 2014; Thorel et al., 2010). Similarly, exocrine ductal or acinar cells have been shown to interconvert toward a ß-cell fate after genetic gain- or loss-of-function (Lee et al., 2013; Zhou et al., 2008) or cytokine exposure (Baeyens et al., 2013). Thus, pancreatic cell transdifferentiation has emerged as another possible means for generating replacement islet cells. Findings from prior studies using rodent models suggest that cell interconversion is governed and can be manipulated through epigenetic mechanisms (Chakravarthy et al., 2017; Dhawan et al., 2011). Harnessing the therapeutic potential of native pancreatic cells could benefit from improved characterization of the transcriptomes, epigenomes and the chromatin landscape of native pancreatic cells in development or disease.

Recent progress has been made in identifying human pancreatic cell transcriptomes and epigenetic features like histone modifications (Arda et al., 2016; Bramswig et al., 2013; Enge et al 2017; Gaulton et al., 2010; Parker et al., 2013; Pasquali et al., 2014; Thurner et al., 2018; Varshney et al., 2017; reviewed by Grapin-Botton and Serup, 2017). The advances in those studies derive from powerful high-throughput sequencing approaches like RNA-Seq (Mortazavi et al., 2008), ChIP-Seq (Johnson et al., 2007) and ATAC-Seq (Buenrostro et al., 2013). However, practical challenges like rapid autolysis of cadaveric human pancreas, optimized primary cell isolation schemes based on flow-cytometry, and a need for large numbers of cells have limited genomic analysis of primary human pancreas cells. For example, a prior study using ATAC-Seq in human islet α- and β-cells from a limited number of donors unexpectedly did not observe cell type-selective chromatin regions that support α- or β-cell specific gene expression (Ackermann et al., 2016). Thus, chromatin-based models for cell-specific gene regulation in pancreatic endocrine or exocrine lineages have not yet been achieved.

To overcome these challenges, we combined a robust cell purification strategy with chromatin analysis using ATAC-Seq and ChIP-Seq assays to generate cell type specific chromatin maps of the human pancreas. Our analysis revealed thousands of novel genomic regions that serve as putative enhancers, and are specific to β-, α-, duct and acinar cells. These regions are linked to genes expressed in specific cell types, contain binding sites for lineage-specific transcription factors, and overlap with noncoding variants that are associated with pancreatic disorders. Integrating gene expression with chromatin accessibility data reveals mechanisms for cell type specific gene regulation in human pancreatic cells.

RESULTS

Chromatin landscape of human pancreas cell subsets

To generate comprehensive regulatory maps of genomic regions governing gene expression in distinct pancreas cell subsets, we performed ATAC-Seq (Buenrostro et al., 2013) on primary human pancreas cells isolated using cell surface antibodies and fluorescence activated cell sorting (FACS, Figures 1A and 1B; Arda et al., 2016). This included highly purified populations of β-cells, α-cells, duct and acinar cells from human donors (Table S1). In addition to ATAC-Seq, we generated genome-wide primary pancreas islet β- and α-cells, duct and acinar cells from previously healthy donors. Altogether we performed ATAC-Seq and ChIP-Seq on cells obtained from 13 donors, generating 46 new maps of accessible chromatin and histone modifications (Tables S1-S2, Figure S1).

Figure 1.

(A) Experimental approach used in this study. Pancreatic islets and exocrine tissue were obtained from human donors. α-, β-, duct and acinar cells were purified using FACS and followed by high-throughput sequencing assays. * RNA-Seq data reported in (Arda et al., 2016). See also Figure S1, Tables S1-S2.

(B) FACS plots display distinct populations of pancreatic cells using the indicated cell surface markers. UCSC genome browser display of the PAX6 (C) and NR5A2 (D) locus. Normalized signal profiles of ATAC-Seq and H3K4me3, H3K4me1, H3K27ac, H3K27me3 ChIP-Seq obtained from sorted pancreatic cells are shown. Regions corresponding to cell type-specific signals are highlighted in blue.

We focused our initial analysis of ATAC-Seq and ChIP-Seq signals at genomic loci containing genes that are essential for pancreas cell identity (Figures 1C-D and results below). For instance, PAX6 is necessary for endocrine cell differentiation, and is expressed in pancreatic endocrine cells but not in exocrine cells (Sander et al., 1997; Swisa et al., 2016; Yasuda et al., 2002). We identified distinct ATAC-Seq peaks in a-cell and β-cell samples at the 5’ proximal region of PAX6, as well as H3K4me3, H3K4me1, and H3K27ac ChIP-Seq signal around these peaks (Figure 1C). However, these marks were absent in duct or acinar cells; instead, the same region had the repressive H3K27me3 signal in exocrine cells. Similarly, NR5A2 has been reported to control early pancreas development and exocrine cell differentiation (Hale et al., 2014). We detected several ATAC-Seq peaks at the NR5A2 locus in duct and acinar cells which were absent in α- and β-cells. Furthermore, these ATAC-Seq peak regions were marked by H3K4me3, H3K4me1, and H3K27ac ChIP-Seq signal in exocrine cells but not in endocrine cells (Figure 1D). Thus, combining a robust cell purification approach with powerful chromatin assays successfully generated comprehensive, high-quality maps of open chromatin and histone modifications corresponding to major cell lineages and lineage regulators in the human endocrine and exocrine pancreas.

ATAC-Seq identifies gene regulatory elements in human pancreatic cells

Mammalian genomic regions that control cell type specification are marked by chromatin features detectable by their accessibility to enzymatic digestion or transposition (Buenrostro et al., 2013; Roadmap Epigenomics Consortium et al., 2015; Thurman et al., 2012). To identify genome-wide accessible regions systematically in human pancreatic cell subsets, we used the ZINBA algorithm to interrogate the peak regions in our ATAC-Seq datasets (Buenrostro et al., 2013; Rashid et al., 2011). We found a total of 170,361 open peaks after stringent filtering and scoring (FDR 0.05, Table S3). We then quantified the ATAC-Seq reads in each peak across donors and cell types and used DESeq to identify differentially open regions (DORs) (Anders and Huber, 2010). k-means clustering of these DORs revealed five distinct clusters corresponding to different cell types and groups (Figure 2A). For example, we found DORs unique to β-cells (3999 regions), α-cells (5316 regions), both β- and α-cells (‘endocrine’, 5836 regions), duct cells (3871 regions), or both duct and acinar cells (‘exocrine’, 3444 regions: Table S4, Figures 2A, S2A). Despite showing hallmark gene expression profiles (Figure S2B, FDR <0.05), we did not observe a distinct acinar cell open chromatin cluster, and found a more variable ATAC-Seq signal within exocrine cell samples, independent of the age or source of the tissue (Table S1, STAR Methods). This variability may reflect interference by acinar cell enzymes, an inherent technical challenge of exocrine tissue culture. However, recent evidence suggests that open chromatin profiles may be better predictors of cell identity than gene expression profiles (Corces et al., 2016, Shih et al., 2016). Thus, chromatin accessibility profiles may reveal exocrine cell subtypes that are not distinguishable at the gene expression level.

Figure 2.

(A) Heat map shows distinct clustering of differentially open regions (DORs) identified in ATAC-Seq assays. Each column is a unique genomic region corresponding to an ATAC-Seq peak. Individual samples are organized in rows; magenta-α-cells, cyan-β-cells, black-duct cells, and orange-acinar cells. DOR clusters and the number of genomic regions in each cluster are indicated at the top of the heatmap. See also Figure S2 and Tables S3-S4.

(B) Distribution of pancreas DORs with reference to genomic features. TSS; transcription start site.

(C) Bar graphs represent the enrichment of GO Biological Process terms for genes associated with each DOR cluster. Select genes linked within these clusters are listed on the right.

Findings by the ENCODE consortium showed that 5% of open chromatin regions localize to the transcriptional start sites (TSS), while the remaining 95% of open regions map nearly equally to either intronic or intergenic regions (The ENCODE Project, 2012; Thurman et al., 2012). To determine the distribution of open chromatin regions identified in our assays, we calculated the overlap of these regions with specific genomic features, such as TSS, genic or intergenic regions (Figure 2B). We found that the majority of lineage-specific DORs reside in intergenic or intronic regions (87–92%), while 3–10% overlapped with promoters (Figure 2B). Consistent with the view that distal enhancers are highly cell type specific (reviewed in Bulger and Groudine, 2011; Visel et al., 2009), pancreatic DORs were enriched in distal regions an average of 20 kilobases (kb) away from the TSS of neighboring genes (see STAR Methods). Thus, our experimental approach yielded quality datasets identifying accessible chromatin regions of human pancreatic cell subtypes, whose essential features were in accord with prior work on human genome regulation.

To analyze functional enrichment of genes associated with pancreatic DORs, we used the Genomic Regions Enrichment of Annotations Tool (GREAT: McLean et al., 2010). Genes associated with endocrine cell DORs were linked to relevant GO terms such as regulation of peptide hormone secretion or transport (Figure 2C). Genes nearest these endocrine-specific open regions encoded known physiological regulators like GCK and SLC2A2 as well as transcription factors regulating endocrine cell fate and function like TCF7L2, NEUROD1, and PAX6. Moreover, we found that α- or β-cell DORs are linked to genes with cell type-specific functions, like GCG, GLP1R and IRS1/2. In contrast, the duct and exocrine cell DORs neighbored genes enriched for GO terms like TGFβ signaling pathway, cell cycle control and exocrine development, and included known exocrine cell regulators such as PTF1A, HES1 and FOXA1 (Gao et al., 2008; Jensen et al., 2000; Weedon et al., 2014). Thus, our differential open chromatin analysis revealed putative regulatory regions corresponding, with remarkable specificity, to distinct pancreatic cell lineages.

DORs reveal the regulatory logic and lineage history of human pancreatic cell subsets

To begin elucidating the c/s-regulatory logic governing human pancreatic cell fate and function, we characterized the relationship between DORs (Figure 2A) and genes that are potentially regulated by these genomic regions. We first analyzed the expression of genes assigned to DORs by the GREAT algorithm. We derived gene expression specificity scores (‘ESS’: Table S5) based on RNA-Seq data (see STAR Methods, Arda et al., 2016; Julien et al., 2012; Scott et al., 2016). Thus, a hallmark cell type specific gene like insulin (INS) had a high ESS in β-cells (0.95), but not in other cell types (e.g. 0.002 in duct cells). Confirming our prediction, we found that DORs are associated with genes that have higher expression specificity in the appropriate cell type(s) (Figure 3A). For instance, β-cell DORs are associated with genes whose expression is higher or specific to β-cells (GLP1R, DACH2, INS, IAPP). As a control, when we considered the genes linked to all ATAC-Seq peaks, i.e. not cell type specific, the median score was at or below the threshold for ubiquitous expression (0.25: see STAR Methods) in all cell types (Figure S3). Thus, DORs identified here provide an unprecedented resource, and can explain cell type-specific gene expression in the human pancreas, with more than 50% of DORs in each cluster linked to high cell type-specific gene expression.

Figure 3.

(A) Box plots show the distribution of expression specificity scores (ESS) of genes associated with DORs for each cell type (acacinar cells). Dashed red lines indicate the 0.25 threshold of specificity. See also Figure S3, Table S5.

(B) 95% confidence intervals of expression specificity scores of genes associated with DORs are plotted. See also Table S6.

(C) Model suggesting regulation of cell type-specific expression in human pancreatic cells. The dendrogram on left represents pancreatic lineages. Black tick marks indicate putative regulatory regions in the genome. Combination of lineage and cell type-specific regulatory regions confers cell type-specific gene expression.

To further investigate the connection between DORs and cell type-specific gene expression, we examined how the ESS of a gene changes with the number of DORs associated with it (Table S6). We found that genes with the highest cell type-specific expression were those linked to multiple DORs in the same cluster (Figure 3B). For example, we observed two distinct β-cell DORs near GLP1R, which encodes a G protein-coupled receptor expressed in human β-cells but not in α-cells, ducts or acinar cells (Dai et al., 2017). For DACH2, which encodes a transcription factor expressed in human β-cells, we detected five distinct DORs. This suggests that in pancreatic cells, one mechanism for achieving cell type-specific expression is to increase the number of cell type-specific accessible chromatin regions (Figure 3C). We speculate that this mechanism could provide additional binding sites and a combinatorial basis for lineage-specific transcription factors to ensure cell type-specific expression. If so, DORs associated with genes expressed in specific cell types should be enriched for cell type-specific transcriptional regulators.

DORs are enriched for lineage-specific transcription factor binding sites

Accessible chromatin regions are sites for transcription factor (TF) binding (Corees et al., 2016; Neph et al., 2012; Stergachis et al., 2013) and our chromatin-based assessment of gene regulation in pancreatic cells predicted cis-regulatory sites in cell-specific DORs. To discover TF motifs enriched in DOR clusters, we used the HOMER algorithm (Heinz et al., 2010). Consistent with our GREAT analysis, we found over-represented motifs in endocrine specific DORs for factors like FOXA, RFX, and PAX6, known regulators of pancreatic islet fate and function (Figure 4A, Table S7; Gao et al., 2008; Sander et al., 1997). Likewise, we identified enriched TF motifs for lineage markers of β- and α-cells, such as PDX1 or ARX in respective β-cell or α-cell DORs. In contrast, our analysis of exocrine-specific regions yielded a very different set of motifs, demonstrating the specificity and sensitivity of our experimental approach. For instance, we found representative motifs for SMAD, STAT, GATA, RBPJL and PTF1A. Both RBPJL and PTF1A are established regulators of exocrine pancreas development and function (Sellick et al., 2004; Masui et al., 2007; Figure 4A; Table S7). In addition, we identified a motif for TEAD, a recently-discovered regulator of human multipotent pancreatic progenitors (Figure 4A; Cebola et al., 2015). Thus, our approaches led to de novo discovery of sequence motifs nominating candidate TF regulators of human pancreas cell diversification and function.

Figure 4.

(A) Position weight matrices of TF motifs enriched in each DOR cluster. See also Table S7.

(B) Bubble plot showing the motif enrichment scores and expression levels (rpkm) of TFs in pancreas cells.

(C) Association index analysis reveals TF modules, marked with black rectangles (See STAR Methods for details). Matrices show the similarity between TF pairs based on their motif cooccurrence in DORs. See also Figure S4, Table S8.

In many cases, TFs like the bZIP factors MAFA and MAFB or the GATA-box binding factors GATA4 and GATA6, share structural motifs and DNA binding specificities, making it difficult to assign specific TFs to regulatory genomic elements. To address this issue, we integrated our ATAC-Seq and RNA-Seq data (Figure 4B). Integration of these two datasets helped resolve instances where the predicted motif enrichments are too similar to reliably predict the corresponding trans-acting regulator. For instance, both MAFA and MAFB recognize the same DNA sequence; however, MAFA expression is specific to β-cells, whereas MAFB in humans is expressed in both α- and β-cells. Similarly, our analysis revealed that GATA6 is enriched in α-cells, whereas in pancreatic exocrine cells both GATA4 and GATA6 are likely functional. To delineate cell type-specific TF networks further, we also performed an association index analysis followed by hierarchical clustering (Bass et al., 2013). We calculated the similarity between TFs based on motif co-occurrence in a given peak (Figures 4C and S4A, Table S8) or based on target gene overlap as assigned by GREAT (Figure S4). As expected, TFs with similar motif preferences clustered together; however, we also found clustering of distinct TF motifs within the same DOR. For instance, in α- or β-cells, NEUROD1 motifs appeared in the same modules as MEIS and RFX motifs (Figure 4C). In addition, our analysis revealed possible coregulation of genes through distinct genomic regions (Figure S4). For example, GATA and STAT motifs appeared to be frequently linked to a shared group of genes (Figure S4B). Thus, our integrative analysis permitted further delineation of pancreatic TF networks and refined the list of possible regulators governing cell type-specific gene regulation.

Identifying chromatin regions underlying the competence for pancreatic cell conversion

Recent mouse studies have revealed conditions, like extreme β-cell loss or genetic loss-of-function, that promote conversion of pancreatic endocrine cells into β-cells (Chakravarthy et al., 2017; Chera et al., 2014; Thorel et al., 2010) or into a-cells (Dhawan et al., 2011). In mice and humans, adult exocrine cells can also be reprogrammed into insulin-producing cells resembling β-cells (Lee et al., 2013; Zhou et al., 2008). While the molecular mechanisms of pancreatic cell fate conversion are not entirely understood, evidence in other systems point to chromatin structure as a major determinant of successful reprogramming outcome (Soufi et al., 2012; Xu et al., 2011). Supporting this view, loss of chromatin regulators like DNA methyltransferases has been shown to promote islet cell type interconversion (Chakravarthy et al 2017; Dhawan et al., 2011). To examine the potential for cell fate conversion at the chromatin level in α-, β-, duct or acinar cells, we calculated the correlation between the DOR clusters based on the genes that are assigned by the GREAT algorithm (Figure 5A, Table S6). Unexpectedly, we found evidence of shared DORs in α- and β-cells (Figure 5A) near genes with restricted expression, like PCSK1 (restricted to β-cells) and PCSK2 (restricted to a-cells: Figure 5B). We observed a similar relationship with genes linked to duct and exocrine DORs; by contrast we found an anticorrelation between endocrine and exocrine lineage DORs (Figure 5A). Altogether these findings suggest a model of enhancer usage in which endocrine and exocrine lineages adopt distinct chromatin profiles after diversification. After further exocrine or endocrine diversification (for example development of endocrine α- and β-cells), lineage-related cells continue to maintain shared control regions (Figure 3C). Thus, there may be distinct thresholds for interconversion between pancreatic lineages established by chromatin structures during development and cell differentiation.

Figure 5.

(A) Correlogram of Spearman rank correlation coefficients for cell type-specific DORs is shown. Circle size reflects the absolute coefficient value. See also Table S6.

(B) Scatter plot shows the ESS values of genes linked to endocrine DORs in α- and β-cells. The orange contour lines indicate data density. Select genes (dots) are highlighted with red on the graph.

(C) Heat map shows the chromatin states identified in purified human pancreas cells based on ChromHMM analysis. 10-state model was built using the histone mark ChIP-Seq data. The white-blue color bar represents the emission probability. States are numbered, colored and annotated as suggested in (Ernst and Kellis, 2012).

(D) Bar graphs show the cumulative fraction of DOR overlap with each of the annotated chromatin state.

Open chromatin-associated histone modifications identify enhancers in human pancreatic cells

Prior studies have revealed that pancreatic lineages arise from a common multipotent progenitor during development (reviewed in Arda et al., 2013; Benitez et al., 2012). However, chromatin regulation underlying lineage specification in human pancreas development remains unknown. After the establishment of distinct cell lineages, cell identity is thought to be maintained by histone modifications at cis-regulatory regions or enhancers (reviewed in Calo and Wysocka, 2013). Recent work has shown that specific patterns of histone modifications are reliable predictors of enhancers that regulate cell identity (Heintzman et al., 2009; Rada-Iglesias et al., 2011; reviewed in Ong and Corces, 2011; Shlyueva et al., 2014). To further characterize the lineage-specific regions identified in our study, we integrated the DORs with our ChIP-Seq results, which include data obtained from multiple donors, cell types and histone marks (Tables S1-S2, S9; Arda et al., 2016). First, we used the ChromHMM algorithm to identify combinations of histone modifications and their spatial distributions from pancreas cell ChIP-Seq data (STAR Methods, Ernst and Kellis, 2012). This approach permitted de novo discovery of cell type-specific chromatin states in human pancreas cells, and led to “functional annotation” of genomic elements, including active TSS, enhancers or repressed regions (Figure 5C). We then quantified the extent of DOR overlap with these annotated chromatin states, and found that in every cell type, 50–90 % of DORs were associated with active chromatin marks, including genic and active enhancers (Figure 5D). Thus, integration of histone mark datasets with open chromatin regions produced an augmented ‘chromatin state’ map and further strengthened the view that DORs could function as enhancers in human pancreas cells.

DORs stratify noncoding variants linked to pancreatic disorders

Genome-wide association studies (GWAS) have identified genetic variants called single nucleotide polymorphisms (SNPs), linked to disease risk. In studies of diabetes risk, most SNPs localize to non-coding genomic regions (McCarthy, 2017; Thomsen and Gloyn, 2014). However, cell type-specific stratification of these variants remains problematic. To investigate associations of SNPs linked to diabetes and associated traits with pancreatic DORs (see STAR Methods, Welter et al., 2014), we computed the localization of SNPs and pancreas cell type-specific open chromatin regions. We found 90 (out of 623) risk SNPs that overlapped with one or more open regions. Thus, diabetes and related trait risk variants are significantly enriched in open chromatin regions in human pancreatic cells (P-value 1.4E-06, Figure S5). Of 90 risk SNPs, 15 overlapped with cell type-specific DORs (Figure 6A). 80% of these variants were associated with diabetes or related traits and overlapped with an α-cell or 0-cell DOR. This may suggest that anomalies in non-β islet cells might contribute to diabetes disease risk as well. Unexpectedly, we also found SNPs previously linked to diabetes or fasting glucose traits that mapped to pancreas exocrine cell-specific DORs (Figure 6A). One of these, rs11920090 resides in the 8^th intron of the SLC2A2 gene, which encodes for a glucose transporter also known as GLUT2. Our data reveal an open chromatin region in this intron specifically in exocrine cells (Figure 6B), despite the normal expression of this gene in β-cells, but not acinar or duct cells (Arda et al., 2016). Thus, obtaining chromatin maps at cell-specific resolution has identified unrecognized possible epigenetic connections between pancreatic exocrine cells and diabetes.

Figure 6.

(A) Hierarchical clustering of the ATAC-Seq signal corresponding to DORs that overlap with SNPs reported in GWAS studies related to diabetes and associated traits or pancreatic cancer.

(B) Genome view of the SLC2A2 locus highlighting ATAC-Seq peaks found in α-, β-, duct and acinar cells and the location of the risk SNP that is linked to fasting glucose traits. On the right, the bar graph shows normalized RNA-Seq read counts representing gene expression of SLC2A2 in pancreatic cells.

DISCUSSION

Here, we sought to delineate the regulatory genomic regions of pancreatic cell subtypes isolated from human donors, and to understand the relationship between these elements and cell-specific gene regulation. To achieve this, we combined a reliable organ procurement framework and cell purification strategy with ATAC-Seq, ChIP-Seq and RNA-Seq assays to generate high-quality open chromatin and histone maps linked to expression profiles of primary human pancreatic cells. Our datasets and analysis provide distinct cell type-specific regulomes compared to prior maps of human pancreatic endocrine or exocrine cells, thereby providing unique resources and knowledge about regulatory elements governing pancreatic cell identity and function.

Recent large-scale epigenomic studies showed that most gene promoter regions appear accessible in diverse cell types, but distal open chromatin regions are highly cell type selective, and increasing evidence indicates these distal sites may function as enhancers. Our DOR analysis of human pancreatic cell subsets is consistent with this canon, with the majority of DORs identified in our study overlapping with distal intergenic or intronic regions. Our prior work investigating age-dependent gene expression differences between juvenile and adult pancreatic cells (Arda et al 2016) suggests that histone modifications could modulate age-dependent gene regulation. While our datasets here include cells isolated from children and adults, the sampling was comparatively smaller than those in prior studies (Arda et al 2016), and did not identify age-dependent changes in open chromatin of pancreatic cells. Thus, larger cohorts may be necessary to unveil possible age-dependent chromatin accessibility in postnatal human pancreatic cells. In addition, delineation of rare islet δ- and PP-cell chromatin profiles remains a challenge, and emerging single-cell technologies may overcome these technical limitations in future work.

A prior study using ATAC-Seq to investigate accessible chromatin in human α- and β-cells from three donors, identified cell-selective open regions that were enriched 14-fold in α-cells compared to β-cells (Ackermann et al 2016). In addition, Ackermann et al. reported that only 5% of α-cell specific peaks and 12% of β-cell specific peaks were linked to differentially expressed genes, and the preponderance of these peaks was associated with promoter regions. In our study, the number of α- and β-cell specific open regions was nearly equal, and these regions largely overlapped with genic regions, not promoter regions, in line with precedents like the entire ENCODE dataset (The ENCODE Project, 2012). Moreover, we found that genes with cell type-specific expression were linked to multiple cell type-specific DORs, with more than 50% of peaks in all cell type specific clusters associated with cell specific/high expression. These observations are similar to recent findings showing the association of multiple enhancers to specific expression in hematopoietic lineages or providing phenotypic robustness during limb development (Javierre et al., 2016; Osterwalder et al., 2018). Our results are also in accord with the known linkage of active enhancer elements to nearby gene expression (Creyghton et al., 2010; Zentner et al., 2011). Thus, the datasets we provide in this study achieve cell type resolution not previously observed. We conclude that multiple cell type-specific regulatory regions regulate gene expression in human α-cell or β-cells. Moreover, HOMER analysis identified several lineage specific TF motifs in DORs (Figure 4). For instance, we observed (1) enrichment in β-cell DORs for sequence motifs bound by PDX1, NEUROD, MAF, RFX and NFAT, (2) enrichment in α-cell DORs of motifs bound by ARX, RFX and MAF, and (3) enriched motifs for ISL1 and PAX6 in ‘common’ endocrine cell DORs. These findings corroborate multiple prior studies in pancreas development, transcriptional regulation and genetics (reviewed in Benitez et al., 2012 and in Arda et al., 2013).

An important implication of our findings for pancreas development derives from the high degree of overlap in genes linked to endocrine specific DORs and α- or β-cell specific DORs. Based on these data, we speculate that specific regulatory regions in the endocrine lineage become accessible in a stepwise manner during development, with genomic regions governing common endocrine features made accessible first, followed by increased chromatin accessibility in regions regulating cell type-specific features. If so, this highlights mechanisms that could be exploited to direct conversion between endocrine cell types, with a focus on manipulations impacting cell type-specific DORs. Likewise, non-endocrine to endocrine cell conversion might be facilitated by manipulations that sequentially target common endocrine DORs, followed by cell type-specific DORs. Thus, future work to assess the contribution of enhancers to gene regulation could advance our understanding of pancreatic cell identity and function, and transdifferentiation.

In diabetes and other diseases investigated by GWAS, the vast majority of disease-linked variants are located in the non-coding genome. An outstanding challenge in diabetes genetics is to identify the cellular, physiological and molecular mechanisms linking these genetic variants and disease risk. Annotating and parsing the genome into differentially functional states in distinct cell types should facilitate the discovery of these genome-disease links. In addition to enrichment of GWAS SNPs to islet endocrine gene DORs, we found diabetes-related SNPs overlapping with exocrine specific DORs, suggesting future studies to assess the possibility that pancreatic exocrine cell dysfunction could impair islet glucose sensing or insulin output. There is strong evidence that exocrine pancreas growth (Campbell-Thompson et al., 2012), or diseases like chronic pancreatitis, cystic fibrosis and pancreatic adenocarcinoma are linked to multiple forms of diabetes (Hart et al., 2016). Our findings provide a chromatin-based framework to investigate the genetic basis of these pancreatic exocrine-endocrine disease associations.

STAR METHODS

Contact for Reagent and Resource Sharing

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Seung K. Kim (seungkim@stanford.edu).

Experimental Models and Subject Details

Human Pancreas Tissue Procurement

All studies involving human pancreas tissue were conducted in accordance with Stanford University Institutional Review Board guidelines. We established a procurement protocol with strict criteria to obtain pancreatic tissue from healthy, non-diabetic organ donors in a timely manner. De-identified human pancreata, islets or acinar tissues were obtained from healthy, non-diabetic organ donors with BMI<30, deceased due to acute trauma or anoxia. Organs and islets were procured through Integrated Islet Distribution Network (IIDP), National Diabetes Research Institute (NDRI), and International Institute for the Advancement of Medicine (IIAM). Pancreas tissue was processed and shipped using standardized procedures to reduce variability between shipments, which include 48 hours of pre-shipment culture time, shipment packaging to maintain internal temperature at 8°C and overnight delivery. Upon receipt, the tissue was processed immediately, without additional culture time. For ChIP-Seq and ATAC-Seq studies, purified cells were obtained from islets or exocrine tissue of 7 adult donors (ages 19, 20, 33, 40, 48, 54, 66) and 6 juvenile donors (ages 0.8, 1.4, 1.5, 5, 5, 9). Further donor details are provided in Tables S1 and S2.

Method Details

Isolation of Human Pancreas Cells Using Flow Cytometry

Preparation of islet or exocrine tissue for flow cytometry, including antibody staining and FACS settings was performed as described in (Arda et al., 2016). Briefly, the tissue was rinsed with phosphate buffered saline (PBS, Gibco # 10010023) and dispersed into single cells by enzymatic digestion using Accutase (Life Technologies) following manufacturer’s protocol. Prior to antibody staining, cells were incubated with a blocking solution containing FACS buffer (2% v/v fetal bovine serum in PBS) and goat IgG (Jackson Labs, 11.2 μg per million cells). LIVE/DEAD Fixable Aqua Dead Cell Dye (Life Technologies) was used as a viability marker. Cells were then stained with appropriate antibodies at 1:100 (v/v) final concentration. The following antibodies were used for FACS experiments: HPx1-Dylight 488 (Novus, NBP1– 18951G), HPi2-Dylight 650 (Novus, NBP1–18946C), HPa2-Biotin (Novus, NBP1–18950B), CD26-PE (BioLegend, 302705), CD133/1 - Biotin (Miltenyi Biotec 130–090-664), CD133/2 - Biotin (Miltenyi Biotec 130–090-852), streptavidin-eFluor780 (eBioscience, 47–4317-82), streptavidin-APC (eBioscience, 17–4317-82). Inclusion of the CD26 antibody greatly improves separation of α-cells from β-cells, also reported in (Arda et al., 2016). All antibody incubation steps were performed on ice for 30 minutes. Labeled cells were sorted on a special order 5-laser FACS Aria II (BD Biosciences) using a 100 μm nozzle, with appropriate compensation controls and doublet removal. Sorted cells were collected into low retention tubes containing 50 μL of FACS buffer. Cytometry data were analyzed and graphed using FlowJo software (TreeStar v.10.8). After each sort, a diagnostic reverse transcription quantitative PCR was performed to confirm enrichment of appropriate cell populations before proceeding with ATAC-Seq or ChIP-Seq assays. The following TaqMan probes were used to test for marker transcript enrichment: INS (Hs00355773_m1) for β-cells, GCG (Hs00174967_m1) for α-cells, KRT19 (Hs00761767_s1) for duct cells, CPA1 (Hs00156992_m1) for acinar cells. ACTB (Hs01060665_g1) was used as endogenous control. Enrichment was calculated using the comparative ΔΔCT method based on relative mRNA abundance in pre-sorted versus sorted cell populations (Arda et al., 2016).

ATAC-Seq assays

On average 50,000 sorted cells were used for each ATAC-Seq assay following the protocol described in (Buenrostro et al., 2013). After flow cytometry, cells were centrifuged for 5 min at 500 g in a micro-centrifuge tube. Sorting buffer was removed, and cells were washed twice with PBS. Nuclei was isolated using cold lysis buffer (10 mM Tris-HCl, pH 7.4, 10 mM NaCl, 3 mM MgCl2 and 0.1% IGEPAL CA-630). The nuclei pellet was resuspended in the transposase reaction mix; 25 μL 2× TD buffer, 2.5 μL transposase (Illumina) and 22.5 μL nuclease-free water. The transposition reaction was performed at 37C for 30 min. Transposed DNA fragments were purified using the Qiagen MinElute kit and amplified 6–8 cycles using the Nextera (Illumina) PCR primers. Libraries were sequenced as 2×50 on HiSeq2000 platform.

ChIP-Seq assays

ChIP assays were performed using protocols described (Arda et al., 2016; Rada-Iglesias et al., 2011; Wapinski et al., 2013). On average 2.5×10⁵ sorted cells were used for each ChIP-Seq experiment. 0.5 mg of the following antibodies were used to immunoprecipitate modified histone proteins: anti-H3K4me3 (Abcam, ab8580), anti-H3K4me1 (Abcam, ab8895), anti-H3K27me3 (Diagenode pAb-069–010), anti-H3K27ac (Abcam, ab4729). NEBNext library prep kit (NEB, E6240) was used to prepare sequencing libraries of ChIP fragments. Barcoded ChIP-Seq and input libraries were multiplexed and sequenced as single-end 50 bp reads on Illumina HiSeq2000 platform.

Quantification and Statistical Analysis

ATAC-Seq data processing and analysis

Initial processing of ATAC-Seq reads were performed as described (Buenrostro et al., 2013). Briefly, sequencing reads were trimmed and aligned to the human genome (hg19) using Bowtie (version 0.12.6) (Langmead et al., 2009) with -ml and -X2000 options. Peaks were called on each sample separately using the ZINBA algorithm (Rashid et al., 2011). These peak regions were then merged to generate a non-overlapping union peak list for downstream analysis. ATAC-Seq fragments corresponding to the identified peaks were quantified by running annotatePeaks.pl option in the HOMER suite (Heinz et al., 2010). Peaks were further filtered by excluding those that have less than the median ATAC-Seq signal in each sample. At the end of this primary data processing, we obtained a final list of ~170,000 peaks (Table S3) encompassing 14 ATAC-Seq samples (Table S1). The raw ATAC-Seq read counts of these ~170,000 peaks were passed onto DESeq R package to identify differentially open peaks (Anders and Huber, 2010). DESeq detects and corrects dispersion estimates that are too low by modeling dispersion dependence on the average signal strength over all samples. Negative binomial generalized linear models were fitted to evaluate the significance of cell type effect and calculate P-values. ATAC-Seq peaks that passed this significance threshold with an FDR <0.05 were considered “differentially open regions” (DORs) between cell types. To reveal the clusters within the DORs, we used k-means clustering method in Cluster 3.0 (de Hoon et al., 2004) and visualized the clusters using TreeView (Saldanha, 2004). Sample correlation was calculated and visualized using the R ggcorrplot () package with hierarchical clustering. To assign regulatory domains to pancreatic DORs (Table S4), we used the GREAT algorithm (v3.0.0) with the single-nearest gene association rule (1000 kb max extension, curated domains included). The region-gene association results were used for subsequent ESS analysis. Each DOR is associated with a single gene, however different DORs could be associated with the same gene. For the GO Term enrichment analysis, DORs were used as test regions against whole genome (hg19) as background. To obtain the correlogram in Figure 5A, the data listed in Table S6 were passed into the R corrplot package to calculate the Spearman rank correlation coefficients between DORs.

RNA-Seq data analysis

Differential gene expression analysis was performed on the RNA-Seq dataset (GEO SuperSeries GSE79469) which was obtained from sorted pancreas cells as described in (Arda et al., 2016). Briefly, gender specific genes, based on the likelihood ratio test with an FDR of < 0.1 or are on chromosome Y, were removed from the dataset. FDR was computed by p.adjust function of R using the Benjamini and Hochberg method. Differential gene expression calls were done using the DE-Seq R package (Anders and Huber, 2010). Raw counts of genes with interquartile range > 2, and factors of interest were passed to DE-Seq to estimate the size factor of each sample for normalization and the dispersion of each gene. The factors of interest were cell types: alpha, beta, duct and acinar. To cluster differentially expressed genes, normalized counts were shifted by +1, log base 2 transformed and mean centered. Hierarchical clustering of differentially expressed genes was done using the average linkage method of Cluster 3.0 software (de Hoon et al., 2004) and visualized with Java TreeView (Saldanha, 2004).

ChIP-Seq data processing and analysis

Sequencing reads were mapped to the human reference genome hg19 using Bowtie (Langmead et al., 2009) with default parameters. HOMER suite was used to quantify ChIP-Seq tag densities in the 4 kb region centered at ATAC-Seq peaks. Tag directories were generated by running makeTagDirectory on the ChIP-Seq SAM files with default options. annotatePeaks.pl script was used to count tags from individual ChIP-Seq samples with following options: hgi9 -size 4000 -noann (Table S9). These tag counts were used as input for the principle component analysis (PCA). To calculate and visualize the PCA of ChIP-Seq samples, R prcomp () function and ggplot2 package were used. To assign chromatin states in each cell type, we used ChromHMM (Ernst and Kellis, 2012). These states are ‘learned’ from user provided data based on an unbiased machine-learning procedure. 6-,8-,10-state models were tested and compared. 10-state model yielded the highest resolution of different chromatin states and was selected for subsequent analysis. The resulting states were annotated based on the analysis reported in prior work (Roadmap Epigenomics Consortium et al., 2015; Ernst and Kellis, 2012). To calculate the overlap between DORs and ChromHMM states, we used the intersectBed command from the BedTools suite with -f 0.5 optional parameter, which requires at least 50% overlap for an intersection to be reported between two genomic regions (Quinlan and Hall, 2010). ChromHMM also generates custom genome browser tracks with the resulting chromatin state segmentation. BED files with segmentation information generated in this study are available at the GEO Series GSE79468 for uploading and visualizing on a genome browser.

TF motif enrichment and Association Index analysis

Motif enrichment analysis was performed individually on each DOR cluster using the HOMER suite (Heinz et al., 2010). For de novo and known motif discovery, we used findMotifsGenome.pl command with the hgl9 -size 500 -len 6,8 options. As a background control, HOMER chooses a set of sequences from the same genome build that are matched in size and GC content to the peak set of interest. To obtain enrichment P-values of the de novo predicted motifs in all other DOR clusters (excluding the cluster that the motif was predicted), we used the findMotifsGenome.pl command with the -size 500 -len 6,8 -nomotif -mknown options and directing the program to use the position weight matrices corresponding to the predicted TF motifs from the step above. The association index analysis, which is a similarity measure between TFs based on connectivity profiles (Bass et al., 2013), was performed as follows: HOMER’s annotatePeaks.pl command was used to find the TF motif instances in DORs (Table S8). The TF motif occurrence in a given peak was defined as motif being present (1) or absent (0). Pearson correlation coefficient method was used to determine the similarity between TF pairs based on the present/absent calls. The results were visualized using the R package ggcorrplot with hierarchical clustering. The analysis was repeated for clustering TFs on the basis of linked genes as determined by GREAT algorithm, with the exception that the Pearson correlation coefficients were scaled before visualizing as heat maps.

Expression specificity score (ESS)

We derived the cell type expression specificity scores using a method similar to that described for tissue-specific expression in (Julien et al., 2012). ESS was calculated as follows: x_i is the expression of the gene in cell type i n is the number of cell types

$$ESS=\frac{median(x_i)}{{\displaystyle {\sum }_{i=1}^{n}median(x_i})}$$

To obtain x_i, we used the normalized and filtered RNA-Seq counts as gene expression values reported in (Arda et al., 2016), in which pancreatic β-, α-, duct and acinar cells were purified using the same FACS strategy. Thus, a gene with an ESS of zero would indicate no expression in that cell type, 0.25 would indicate ubiquitous expression across cell types and an ESS of 1 would indicate cell type-specific expression, i. e. the gene is only expressed in that cell type.

Cell type-specific overlap of disease-associated regions

To calculate the overlap between SNPs reported in individual GWAS and DORs, we downloaded the lead SNPs from the NHGRI GWAS catalog (Welter et al., 2014), dbSNP Build 147 for traits related to diabetes, glucose or pancreatic cancer. UCSC Table Browser was used to convert GRCh38.p7 coordinates into hg19. A non-redundant list of GWAS SNPs were created based on a unique combination of SNP, annotated trait and PubMedID. The overlap between the SNP and ATAC-Seq DOR coordinates were computed using intersectBed from the BedTools suite (Quinlan and Hall, 2010). We used the GREGOR tool to calculate the enrichment of GWAS SNPs associated with diabetes and related traits (as listed above) in ATAC-Seq peak regions (Schmidt et al., 2015). This tool generates a null-SNP set matching the minor allele frequency, distance to nearest gene TSS and the number other variants in the linkage disequilibrium (LD) of the test set to calculate the enrichment in the regulatory regions. Specifically, we used following parameters in GREGOR: r² threshold (for inclusion of SNPs in LD with the diabetes associated GWAS SNPs) = 0.99, LD window size = 1 Mb, and minimum neighbor number = 500. ATAC-Seq signal corresponding to the overlapping DOR coordinates were log₂ transformed, clustered using hierarchical clustering method in Cluster 3.0 (de Hoon et al., 2004), and visualized as a heat map using the Treeview software (Saldanha, 2004).

Data and Software Availability

Data Accession Numbers

The data discussed in this publication have been deposited in NCBI’s Gene Expression Omnibus (Edgar et al., 2002) and are accessible through GEO Series accession number GSE79468 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE79468).

Supplementary Material

2

Table S3. (Related to Figures 1–2) List of all ATAC-Seq peak coordinates. This list includes all ATAC-Seq peaks that were detected in one or more cell type, and passed the filtering process.

3

Table S4. (Related to Figure 2) Chromosome coordinates of differentially open regions (DORs) identified in this study.

4

Table S5. (Related to Figure 3) Expression specificity scores (ESS) of genes detected by RNASeq in human pancreatic cells.

5

Table S6. (Related to Figures 3 and 5) Number of cell type-specific DORs associated with genes in GREAT annotated regulatory domains.

6

Table S7. (Related to Figure 4) Lists of enriched TF motifs using HOMER’s known motif analysis.

7

Table S8. (Related to Figure 4) TF motif occurrences in pancreatic DORs.

8

Table S9. (Related to Figure 5) Read counts obtained from histone ChIP-Seq assays that map to DORs.

• We report chromatin maps of purified human pancreatic α-, β-, duct, acinar cells

• Open chromatin regions are linked to genes with cell type-specific expression

• Lineage-specific transcription factor motifs are enriched at distal open regions

• Some diabetes risk SNPs localized to exocrine cell regulatory chromatin

Arda et al. identify and characterize chromatin features underlying cell type-specific gene expression in human pancreatic α-, β-, duct and acinar cells. This work provides reference chromatin maps to guide regenerative medicine approaches for generating functional replacement cells, and prioritizes candidate risk genes for pancreatic diseases like diabetes mellitus.