Three-dimensional analysis reveals altered chromatin interaction by enhancer inhibitors harbors TCF7L2-regulated cancer gene signature

Abstract

Distal regulatory elements influence the activity of gene promoters through chromatin looping. Chromosome conformation capture (3C) methods permit identification of chromatin contacts across different regions of the genome. However, due to limitations in the resolution of these methods, the detection of functional chromatin interactions remains a challenge. In the current study, we employ an integrated approach to define and characterize the functional chromatin contacts of human pancreatic cancer cells. We applied tethered chromatin capture to define classes of chromatin domains on a genome-wide scale. We identified three types of structural domains (topologically associated, boundary, and gap) and investigated the functional relationships of these domains with respect to chromatin state and gene expression. We uncovered six distinct sub-domains associated with epigenetic states. Interestingly, specific epigenetically active domains are sensitive to treatment with histone acetyltransferase (HAT) inhibitors and decrease in H3K27 acetylation levels. To examine whether the subdomains that change upon drug treatment are functionally linked to transcription factor regulation, we compared TCF7L2 chromatin binding and gene regulation to HAT inhibition. We identified a subset of coding RNA genes that together can stratify pancreatic cancer patients into distinct survival groups. Overall, this study describes a process to evaluate the functional features of chromosome architecture and reveals the impact of epigenetic inhibitors on chromosome architecture and identifies genes that may provide insight into disease outcome.

1 |. INTRODUCTION

The compartmentalization of the eukaryotic genome into highly organized chromatin domains is central to the regulation of gene expression and to cellular homeostasis.¹ Until recently, the genome and its structural organization have largely been studied as a unidimensional entity where local chromatin structure is regulated by epigenomic mechanisms such as post-translational histone modifications, DNA methylation, and chromatin-binding proteins. However, advances in genome-wide chromatin conformation capture (3C) methods have enabled the study of the threedimensional (3D) organization of the genome. Studies employing various 3C-based methods, including 4C, 5C, ChIA-PET, and Hi-C, have been developed to map longrange chromatin interactions, and have provided experimental evidence to explore the principles of 3D genomic architecture.² Collectively, these approaches support a model that interphase chromosomes occupy distinct chromosome territories and provide insight into how chromosomes fold within these territories.^3–6 However, the mechanisms that underlie the partitioning of the genome into these domains and their functional importance remains poorly defined.

Analysis of Hi-C data has revealed characteristic structural features of the genome, including chromatin compartments, topologically associated domains (TADs), and chromatin loops.^5,7–9 These distinctive higher order chromatin structures frame long-range enhancer-promoter interactions for epigenetic gene regulation.^10–13 However, large-scale structural studies generally provide little mechanistic detail regarding the functional relationships between higher order chromatin structure and cell-specific gene regulation. Recent computational and statistical approaches demonstrate that Hi-C data can be used to identify interacting genomic loci at a resolution of 8–20 Kb,^14–16 providing sufficient resolution to integrate higher order chromatin organization and gene expression data.

In cancer, altered regulation of epigenetic networks plays a central role in tumorigenesis and metastasis. While DNA methylation and histone modification patterns are frequently associated with both solid and hematological malignancies, it remains to be determined if 3D chromatin states are characteristic to specific cancer types and their gene expression programs.^11,17,18 The reversibility of histone modifications makes them an attractive target for cancer therapy and thus defining the epigenetic landscape of specific cancer types may provide important insight into the development of new therapeutic targets. Small molecule inhibitors that target histone modifying enzymes to disrupt the cancer cell epigenome are being developed for the treatment of cancer.¹⁹ In particular, histone acetyltransferases (HATs) are emerging targets in drug discovery with potential applications in cancer and other disease models.²⁰ HATs catalyze the acetylation of lysine residues on histones during the epigenetic regulation of gene transcription.²¹ In addition to histones, HATs mediate the lysine acetylation of transcription factors, which is important for their function.^22–24 However, currently the role that HATs, histone acetylation, and histone deacetylases (HDACs) play in regulating higher order chromatin structure remains unknown.

In this study, we investigate the relationship of higher order chromatin structure, histone modification, and gene expression using the human pancreatic cancer cell line PANC1. We conducted tethered chromatin capture (TCC), a modified Hi-C protocol,²⁵ to identify and characterize chromosome interactions and domains in PANC1 cells. We integrated the interacting regions with chromatin state information (histone modifications, DNase hypersensitivity, and RNA Polymerase II binding) to uncover distinct types of subdomains associated with specific epigenetic states. We then determined the impact of two epigenetic inhibitors that target the histone acetyltransferases CBP (ICG-001) and EP300 (C646) on chromatin architecture.^26,27 Finally, we incorporated chromatin binding and gene expression data for the transcription factor TCF7L2 to examine the association of chromosome architecture and TF-mediated gene regulation. Overall, our analysis highlights (a) a process for evaluating chromosome architecture and epigenetic states; (b) the impact of two histone acetyltransferase inhibitors on chromosome architecture; and (c) chromatin architecture associated with TCF7L2-mediated gene regulation.

2 |. RESULTS

2.1 |. Identification of chromosomal interacting regions in PANC1

We conducted our studies of higher order chromatin structure in the human pancreatic ductal adenocarcinoma (PDAC) cell-line PANC1, which is a model used for a variety of mechanistic and functional studies of pancreatic cancer. We identified the interacting regions of chromatin via tethered conformation capture (TCC) using 2 biological replicates (Supplemental Table S1).²⁵ The TCC protocol decreases random intermolecular ligations between DNA fragments, particularly from interchromosomal interactions. We assessed the TCC data quality by comparison to available ENCODE HiC datasets for PANC1. Figure 1A compares the genome-wide and chromosome 17 contacts for the TCC and HiC datasets binned at 1 Mb resolution, respectively, where the heatmap color indicates the contact frequency. Both interaction maps exhibit comparable patterns of the regional enrichment of long-range interactions. However, the TCC dataset has a notable depletion of interchromosomal interactions compared to the HiC dataset, with a similar percentage of cis interactions greater than 20 Kb (cis and trans interactions, respectively; Figure 1B). The correlation of the TCC replicates for each chromosome binned at different resolutions (200 Kb, 500 Kb and 1 Mb) correlate well, except for chromosome 9 (Supplemental Figure S1). Poor correlation for chromosome 9 has been found in other cell types (Rao et al., 2014). Further, the PANC1 TCC and HiC datasets have a comparable number of corresponding topological associated domains (TADs) and TAD boundaries (Supplemental Figure S2). Overall, these results indicate a high degree of similarity between the TCC and HiC datasets. Thus, TCC replicates were combined for downstream analyses of PANC1 chromatin structure.

FIGURE 1.

Characteristics of interacting chromatin regions in PANC1 cells. A, Genome-wide and chromosome 17 interaction matrices for PANC1 HiC (top) and TCC (bottom) datasets. The color intensity represents the normalized number of contacts between a pair of loci, chromosome numbers are indicated on the outside of the matrix. B, The observed proportions of intra- and interchromosomal interactions in the valid HiC pairs using HiC or TCC (cis and trans, respectively). C, Histogram displaying the size distribution of TADs within each individual chromosome. TCC, tethered chromatin capture; TADs, topological associated domains

Using the merged TCC replicates, we defined a total of 1371 TADs, 709 boundary regions and 71 gap domains. Boundaries are interaction-sparse regions that lack interdomain chromosomal interactions with neighboring TAD regions, whereas gaps are regions that lack interactions that are located between two identified domains. Gaps occurred in gene deserts or centromeres and few boundaries or gaps were found between two adjacent TADs. TADs are distributed across the genome and majority of TADs are smaller than 2 Mb (Figure 1C). This is consistent with other recent studies showing that the genome is partitioned into Mb-sized local chromatin interaction domains.^7,8,10 60% of genes are contained within TADs whereas boundaries and gaps contain 38% and 2% of genes, respectively. The TAD domains along each chromosome are represented in Supplemental Figure S3.

2.2 |. Classification of domains with epigenetic marks

We next characterized the epigenetic states associated with the different types of structural domains (TADs, boundaries, and gaps). We applied a Hidden Markov Model (HMM) to segment the genome based on combinatorial epigenetic states using histone ChIP-seq data.²⁸ A 12-state HMM with a 1 Kb bin size and an optimized emission probability matrix using the best Bayesian Information Criterion (BIC) scores was used Figure 2A). The resulting 12 epigenetic states are referred to as S1-S12, and can be categorized by regulatory potential by the emission probability values. [n particular, values greater than 0.5 are considered valid marks for that state and values larger than 0.5 represent dominant marks. S1 and S7 are one-mark states enriched with the repressive mark H3K27me3, whereas S9 is a two-mark state having both H3K27me3 and H3K9me3. Both S2 and S11 represent regions that are depleted of any epigenetic marks (all emission probabilities less than 0.015), thus are termed depleted states.

FIGURE 2.

Classification of PANC1 domains with epigenetic marks. A, Emission probabilities of 12 epigenetic states trained by an HMM model on seven histone modifications, DNase and POLR2A. Marks containing emission probability values greater than 0.1 for a given state are considered to be valid and values greater than 0.5 are considered valid marks for that state and values larger than 0.5 represent dominant marks. B, Transition probabilities of the 12 epigenetic states mentioned in (A) with a high transition indicating a higher probability that a state is assigned to a given bin due to the state of the previous bin. C, Genome-wide location analysis of the 12 epigenetic states defined in (A). D, Illustration of one genomic region along chromosome 7 displaying a TAD, boundary and gap domain with the corresponding IGV snapshots of 7 histone modifications, DNase and POLR2A. E, Heatmap displaying clustering of the 12 epigenetic states within the corresponding domains. Each row corresponds to one domain and each column represents the percentage of each epigenetic state in each domain. Columns are clustered based on the TAD domains. F, We then categorized the 12 epigenetic states based on their regulatory potential, these new categorizations are referred to as “sub-domains.” We identified six sub-domains (SD1–6). The epigenetic mark-depleted states S2 and S11 are merged into SD1, the interspersed S6-S7 transition regions are merged into a repressed SD2, the interspersed S1-S7-S9 regions are merged into a repressed SD3, and the regions having active states are merged into a genebody SD4 and two active enhancer/active promoter SD5 and SD6. TADs, topological associated domains; HMM, Hidden Markov Model

We determined the proximity of the epigenetic states by evaluating their transition probabilities (Figure 2B). Three states (S1, S7, and S9) have relatively high transition probabilities to each other, indicating a strong neighborhood of interspersed H3K37me3 and H3K9me3 repressive marks. The S6 state is a one-mark state enriched only with he repressive H3K9me3 and the S8 state is enriched with both H3K36me3 and POLR2A. S3 and S4 are enriched with H3K27ac/H3K9ac/H3K4me1 and DNase/POLR2A. S10 is an intermediate active state, with a pattern similar to S3 and S4, but only highly enriched with H3K4me1. S5 and S12 are two mixture states showing enrichment of both active and repressed marks, as well as high POLR2A.

We categorized genomic regions into 8 different categories and determined the distribution of epigenetic states within each region (Figure 2C). Non-promoter regions, including 5′ and 3′ distal and gene body intragenic) categories are enriched in repressive states S1, S7 and S9). Active states are enriched in 5′ TSS and 3′ Proximal regions (states S4 and S5). Figure 2D shows a region that contains a gap, boundary and TAD, with the corresponding epigenetic state. The bulk of gaps are S2 domains and are depleted of any epigenetic mark Figure 2E). Interestingly, there are subgroups of boundaries and TADs that have varied patterns of histone modifications. We therefore further divided these into different categories; S1/S7 dominant (repressive marks), S4/S5/S10 dominant (active marks, near a TSS), or a mixture (mixed percentage of active and repressive states). This characterization indicates that distinct epigenetic states are physically connected and certain domains contain interspersed repressive epigenetic patterns (Supplemental Figure S4).

Based on the association of epigenetic states, we classified adjacent or intra-domain states by defining subdomains. These combinations defined six sub-domains, referred to as SD1 to SD6 (Figure 2F). The average length of the sub-domains is 60 Kb. SD1 is a depleted sub-domain (comprised of S2 and S11) and lacks marks. SD2 and SD3 are repressed sub-domains (S6/S7 and S1/S7/S9, respectively) and SD4 is a gene body sub-domain (S5/S8/S12). SD5 is an active enhancer sub-domain (S4/S5/S10/S11 or only S10/S11), and SD6 is an active promoter sub-domain (S3/S4/S5 states), which are centered by a promoter (S4) and extend up/downstream of the 5′ TSS (S3).

2.3 |. Correlation of sub-domains and interacting peaks

To explore the relationships between chromosomal loops, epigenetic states and domains, the loci of interacting chromatin regions were determined at a 10 Kb resolution in 40 Kb overlapping windows (interaction peaks (IPs)). We identified 30297 significant IPs (false discovery rate (FDR) < 0.1 with a peak pair distance >20 Kb; Supporting Information File 1). 90% of filtered IPs are intra-domain interactions, whereby the two different loci are located within the same domain. Nearly 80% of the IPs are within a TAD, 19% of the IPs are in a boundary, and very few IPs are in gap regions. Since cancer cell-lines typically harbor chromosomal abnormalities, including chromosomal amplification, we investigated whether amplified regions contribute to the set of IPs. Only 0.72% of the IPs for PANC1 are in amplified regions of PANC1 cells, confirming that amplified regions are not enriched in the set of identified long-range interactions.²⁹ Since the small number of IPs in the amplified regions of PANC1 cells may play important roles in gene regulation, we included them in further downstream analyses, which has been done previously.³⁰

The heatmap in Figure 3A demonstrates that the specific subdomains of interacting loci tend to be the same on either end. For instance, an IP having SD4 at one end usually has a matched SD4 at the other end. This result is consistent with the hypothesis that the two ends of an IP are indeed physically close to or interacting with each other and thus have similar epigenetic states. We found that many IPs have at least one locus in a depleted or repressed subdomain (SD1, SD2, or SD3) (summarized in Figure 3A-right panel).

FIGURE 3.

The relationship between interaction peaks and sub-domains. A, Subdomains at each locus of interaction peaks (IPs) (left). ‘Peak Loci 1′ and ‘Peak Loci 2′ represent the two ends of an IP. The table on the right summarizes the sub-domains identified in (Figure 2). B, Classification of promoter-centric IPs based on nearest genes. We defined a promoter region (P) to include 5 Kb upstream to 1 Kb downstream of a TSS, a distal region (D) as 100 Kb upstream or downstream of a TSS, and any region beyond 100 Kb from a TSS as a far (F) region. IPs for which the same gene is the nearest gene to both ends are defined as PP1, PD1, and PF1 whereas loops in which the nearest gene is different for each end are denoted as PP2, PD2, and PF2. C, Boxplots of expression for genes associated with promoter-centric IPs in PANC1 cells. For PD1 genes, only one gene is involved and that data is plotted in the PD1 panel, for PD2 genes, the expression of the gene at the promoter end is plotted in the PD2-P panel and the expression of the gene at the other end is plotted in the PD2-D panel. P-P is PP1 and PP2 combined. The expression of all other genes that are not involved in IPs in PANC1 cells are plotted in the Non-IP panel. The genes are grouped by the type of subdomain where the promoter is located. PP, promoter-promoter; TSS, transcription start sites

We annotated the IPs according to gene regions and defined six promoter-centered and one non-promoter related interaction groups (Figure 3B); Promoter (P; −5 to +1 Kb of a TSS), Distal (D; 100 Kb upstream or downstream of a TSS), and Far (F) regions (greater than 100 Kb from a TSS). We defined promoters that interact with each other. Because the distance between two loci was calculated using the center of each locus (as opposed to the boundary), in certain cases, the two loci of one IP could actually overlap with each other. If this occurs, then it is possible that the same promoter is identified by both loci; these are designated PP1. There are also promoter-promoter (PP) IPs, where different promoters are at each loci (PP2). If one end of an IP is in the promoter region of one gene and the other end in the distal region of the same gene, this IP is categorized as PD1. PD2 is an IP that has one end in the promoter region of gene1 and the other end in the distal region of gene2. Similarly, if one end is in the promoter region of one gene and the other end in the far region of the same gene, this IP is categorized as PF1 and PF2 is an IP that has one end in the promoter region of gene1 and the other end in the far region of gene2. If neither end is in the promoter region of any gene, that IP is classified as non-promoter-related and given the designation O-O (total number is 19035).

In total, we obtained 11262 IPs associated with at least one promoter, and thus referred to them as P-centered IPs or looping (Supporting Information File 2). We found that approximately 26% of looping events represent P-P interactions with 951 PP1 and 2059 PP2 (interactions between the promoters of different genes), respectively. Further, 24% promoter IPs (2724) occur with distal regions of the same gene (PD1; 2,898 genes), whereas 27% of IPs (3,045) are between the promoter and distal regions of different genes (PD2). The expression level of genes in PANC1 cells linked to each IP categories was determined (Figure 3C).³¹ Interestingly, promoter-centered loops either contain genes that are in the repressed states (low expression and in SD1–3) or genes that are in the active states (higher expression and in SD4–6). Genes in IPs corresponding to SD5 and SD6 are more highly expressed as compared to any other types of sub-domains. Overall, these results indicate that connecting epigenetic state to topological structure can identify epigenetic sub-domains that have distinct patterns of gene expression.

2.4 |. Enhancer inhibitors affect chromosomal organization in PANC1 cells

We previously reported the impact of the histone acetyltransferase (HAT) inhibitors ICG-001 and C646 on global gene expression in PANC1 cells.³² To examine whether PANC1 sub-domains are functionally linked to changes in gene expression, we treated cells with ICG-001 and C646 for 24 hours and performed TCC on control and treated cells. The total number of TADs within each chromosome is equivalent between drug-treated and control-treated PANC1 cells. A large proportion of the TADs, boundaries and gaps do not change with treatment (50%, 40%, and 80%, respectively; Supplemental Figure S5). However, this analysis uncovered chromatin domains that are sensitive to HAT inhibitor treatment. We therefore classified the domain changes (Figure 4A). The most frequent type of change occurred within TADs, whereby treatment increases the TAD length by a maximum of 300 Kb (Figure 4B; conserved-expand category, yellow bar). In contrast, a TAD in treated cells that overlaps with a TAD in untreated cells but the position shifts by more than 300 Kb occurs much less frequently (Figure 4B; the shift category, blue bar). The boundaries were most sensitive to treatment and were more prone to change than either gap or TADs (Figure 4B, purple bar). Pearson correlation showed that domain type changes are associated with boundaries and represent the active sub-domains (SD4–6; Supplemental Figure S6). This suggests that changes of a domain type, especially the transition of boundary to TADs are linked to epigenetically active regions.

FIGURE 4.

Effects of histone acetyltransferase inhibitors on chromatin loops and gene expression in PANC1 cells. A, Diagram of observed types of domain alterations after HAT inhibitor treatment. (1) No change – regions which match exactly between control and treated cells, (2) conserve-expand – regions identified as a TAD in both control and treated cells, with the length of the TAD increasing by at most 300 Kb in treated cells, (3) conserve-shrink – regions identified as a TAD in both control and treated cells, with the length of the TAD decreasing by at most 300 Kb in treated cells, (4) shift – a region identified as a TAD in treated cells that overlaps with a TAD in control cells, with the position shifting by more than 300 Kb, (5) split – a region identified as one TAD in control cells but covers multiple TADs in treated cells, and (6) type-change – a region identified as a TAD in control cells but has switched to a gap or boundary in treated cells. B, Percentage of domain changes after treatment for 96 hours with ICG001 (left) and C646 (right). C, Log2 fold change of H3K27ac levels after treatment, separated by type of domain changes described in (A) (left panels) and type of sub-domains (right panels). D, Overlap between IPs identified in untreated and ICG001 treated (left) or C646 treated (right) PANC1. E, Number of differentially expressed genes after ICG001 (left) or C646 (right) treatment, separated by type of domain change or sub-domain. HAT, histone acetyltransferase; TADs, topological associated domains

To examine how the HAT inhibition impacts histone acetylation within chromatin domains, we conducted ChIP-seq for H3K27ac in ICG001- and C646-treated PANC1 cells. We calculated the log2 fold change (log2FC) of normalized and averaged H3K27ac read signals at 100 base pair (bp) bins in drug-treated versus untreated PANC1 cells for each domain and sub-domain. We observed a minor decrease of H3K27ac at altered categorical domains (shift/split/type-change) and a slight increase of H3K27ac in the no change category (Figure 4C, left panel). While we did not observe major alterations in these domain categories, there were significant differences in the sub-domains. Specifically, all of the active subdomains (SD4–6) showed loss of H3K27ac signal in the treated cells (Figure 4C, right panel). H3K27ac is enriched at promoter and at distal regions. We found that the loss of H3K27ac is more profound at the promoter active subdomain SD6 than at the enhancer active subdomain SD5, suggesting that these inhibitors may affect HAT activity at promoters more than at enhancers. To assess how IPs are altered in drug-treated cells, we performed an IP analysis (described above) and identified 10787 IPs in ICG001-treated and 13 773 IPs in C646-treated PANC1 cells (Supporting Information Files 3 and 4, respectively). This represents an approximate 50% reduction in total IPs in treated PANC1 cells compared to the untreated control. Additionally, we identified that only approximately 50% of IPs in drug-treated cells were concordant with the control. Thus, treatment with ICG001 and C646 results in a decline in total IPs and also a generation of new IPs (Figure 4D).

We previously identified 2029 differentially expressed genes (DEGs) in ICG001-treated and 1740 DEGs in C646-treated cells compared to control cells treated with dimethylsulfoxide (DMSO) (using a log2FC cutoff of 0.5 and a detection P value ≤ 0.05), with an overlap of 754 DEGs common to both drugs.³¹ We integrated expression data with domains and found that approximately 70% of the genes that respond to drug treatment are located in conserved domains. Strikingly, the sub-domains SD5 and SD6 contain a large number of DEGs, regardless of the type of domain or domain change they are associated with (Figure 4E). After further associating DEGs with regions of differential H3K27ac enrichment (based on average H3K27ac reads in the sub-domains) and with looping events, we derived a list of 784 genes for ICG001-treated cells and 380 genes for C646-treated cells.

2.5 |. TCF7L2-regulated genes are involved in altered chromatin interactions

ICG001 and C646 inhibit the activity of CBP and P300 HATs and likely alter key signaling pathways. ICG001 was developed to be a specific inhibitor of the Wnt signaling pathway, which is important for developmental and disease processes.^26,32 A key transcription factor involved in this pathway is TCF7L2, which recruits CBP/P300 to its target gene regulatory elements. Our previous study assessed the impact of TCF7L2 and HAT inhibitors in PANC1 cells; however, the relationship between these processes and chromatin interactions as well as epigenetic modifications is still unknown. TCF7L2 has been linked to a variety of human diseases such as type II diabetes and cancer.^33,34 In a previous study exploring cell type-specific binding patterns of TCF7L2, we showed that the majority of TCF7L2 sites colocalize with H3K4me1 and H3K27ac,³⁵ Given the relationship between TCF7L2 and H3K27ac marked distal regulatory elements, we hypothesized that drug treatment would affect TCF7L2-associated chromatin loops in PANC1 cells. We therefore identified promoter-distal (PD) IPs that were bound by TCF7L2 in PANC1 cells that are no longer classified as IPs in the drug-treated cells. We isolated the genes associated with these IPs and compared them to genes differentially expressed upon drug treatment or upon TCF7L2 knockdown in PANC1 cells, which we identified in a previous study (Figure 5A).³¹ We found that the highest fraction of these IPs were those containing interactions between promoter and distal regions of different genes (PD2-D). We derived a list of 39 genes that are differentially expressed in drug-treated PANC1 cells and are also regulated by TCF7L2 (Supplemental File 5). Pathway analysis using GSEA (Figure 5B)³⁶ reveals enrichment in several cancer-related pathways, including Wnt signaling. We used SurvExpress³⁷ to determine if these genes can stratify survival risk of pancreatic cancer patients and found that this geneset predicts a significant survival correlation (Figure 5C, left panel, p-value 2.5e-07), with high-risk patients displaying a probability of an overall worse survival rate.³⁷ Specifically, 25 of the candidate genes showed differential gene expression between the high- versus low-risk patient groups (Figure 5C, right panel). Thus, our results demonstrate that the HAT inhibitors not only alter chromatin interactions but also distinguish TCF7L2-regu-lated genes for potentially useful clinical signatures.

FIGURE 5.

Effects of histone acetyltransferase inhibitors on TCF7L2-mediated looping in PANC1 cells. A, Number of differentially expressed genes within promoter-centric IPs that were bound by TCF7L2 in untreated PANC1 cells that are no longer classified as IPs in the drug treated cells. These differentially expressed genes were altered in siTCF7L2 knockdown cells as well as drug treated cells. PP2-P1 and PP2-P2 are genes of which the promoters are associated with a PP2 IP, PP1 is the gene of which the promoter is associated with a PP1 IP, PP2-D is the distal gene, while PP2-P is the promoter gene that are associated with a PD2 IP and PD1 is the gene that is associated with PD1 IP. B, KEGG pathway analysis of the genes (n = 39) that are associated with promoter-distal interactions and that are differentially expressed in drug treated PANC1 cells that are regulated by TCF7L2. C, Survival analysis of 176 TCGA pancreatic adenocarcinoma patients of the 39 genes identified in (A). The red line is the survival of the High Risk group, and green line is the survival of the low risk group patients. “+” in the legend stands for the censored patients in each risk group (left). Boxplots displaying the expression of the 39 genes in the two risk groups (right). CI, confidence interval; HR, hazard ratio; P, P value

We identified promoter-distal (PD) IPs that were bound by TCF7L2 in PANC1 cells that are no longer classified as IPs in the drug-treated cells. We isolated the genes associated with these IPs and compared them to genes differentially expressed upon drug treatment or upon TCF7L2 knockdown in PANC1 cells, which we identified in a previous study (Figure 5A).³⁴ We found that these genes are more distributed in the PD interactions, with the highest amount located within genes containing interactions between promoter and distal regions of different genes (PD2-D). We derived a list of 39 genes that are differentially expressed in drug-treated PANC1 cells and are also regulated by TCF7L2 (Supplemental File 6). We performed a pathway analysis using GSEA (Figure 5B)³⁹ on genes are enriched in several cancer-related pathways, including Wnt signaling. We used SurvExpress⁴⁰ to determine if these genes can stratify survival risk of pancreatic cancer patients and found that this geneset predicts a significant survival correlation (Figure 5C, left panel, P-value 2.5e–07), with high-risk patients displaying a probability of an overall worse survival rate.⁴⁰ Specifically, 25 of the candidate genes showed differential gene expression between the high- versus low-risk patient groups (Figure 5C, right panel). Thus, our results demonstrate that the HAT inhibitors not only alter chromatin interactions but also identifies TCF7L2-regulated genes as potentially useful cancer signatures.

3 |. DISCUSSION

Despite technical advances in 3C-based chromatin interaction mapping,^38,39 there is a lack of understanding of how nuclear architecture affects gene expression and cellular function. In particular, our knowledge of how the 3D chromatin architecture of cancer cells contributes to cancer cell-specific gene expression programs is limited. Due to the limitation of sequencing depth and the use of 6-mer cut sites of restriction enzymes, most studies of 3D chromatin architecture thus far have focused on characterizing very large 0.7–2 Mb TADs. Although such studies provide important insights into chromosomal architecture,^29,40 studies of large domains do not address the challenge of associating chromosomal interactions with transcriptional control at the individual gene level. Recent advances in both the experimental and computational aspects of chromosomal interaction analyses now enable the exploration of the 3D chromatin architecture of the human genome at a much higher resolution than previously possible, allowing for the construction of a detailed genome-wide interaction map.^14,15,29 A recent study used an in situ Hi-C protocol to achieve 1–5Kb resolution of interacting genomic segments and linked chromatin loops with promoters, enhancers, and CCCTC-binding factor (CTCF) sites⁸; however, it did not address the relationship between gene loops and gene regulation.

In this study, we demonstrated that the method of TCC can partition the PANC1 genome into three types of structural domains termed gap, boundary, and TAD. Our results are similar to previous studies of other cell types that used different experimental chromatin interaction methods.^7,41 Interestingly, we observed that both TAD and boundary domains (which are 1–5 Mb in length) were embedded with approximately 170 000 intra-domain chromatin interactions or IPs as defined by HOMER. We found that these domains could be further categorized into six types of sub-domains, each with distinct epigenetic characteristics. We note that similar types of subcompartments were defined in a previous study.⁸ However, there are notable differences between the method we present here and that described in the previous study. The previous method divided each of two compartments with histone marks based on underlying interaction intensity and patterns. In contrast, we first used an unbiased training process in which we trained epigenetic states on the whole genome. We then associated the states with gene structure, expression, and other features resulting in the derivation of the six sub-domains. Using this approach, we found that promoter-centered looping genes within the three active sub-domains (SD4–6) showed much higher expression than those in the two repressed sub-domains (SD2–3; Figure 3C), suggesting these newly defined subdomains have functional distinctions.

We further examined the relationship between histone acetylation and chromatin architecture. Although two histone acetyltransferase inhibitors, ICG001 and C646, have been previously shown to alter gene expression in cancer cells,^31,32,42 their impact on the 3D genome and epigenome structure has not been studied. Therefore, we conducted TCC and ChIP-seq of H3K27ac in ICG001- and C646-treated PANC1 cells. Interestingly, we uncovered five major types of domain changes that occur upon treatment of PANC1 cells with ICG001 or C646 (Figure 4A). In regard to Type-Change domains, we found that TADs are largely conserved and stable with drug treatment whereas boundary domains tend to switch to TADs. We also found that the drugs altered chromatin structures associated with positive regulatory elements. The H3K27ac enrichment is reduced predominantly within the active enhancer sub-domains (SD6; Figure 4C) and the most significant gene expression changes occurred in the active-promoter sub-domains (SD6; Figure 4E). We were able to link loops that are lost upon drug treatment with a list of 39 coding genes regulated by TCF7L2, a transcription factor important for developmental processes and implicated in human disease. This subset of genes is associated with cancer-related pathways and could separate pancreatic cancer patients into distinct survival groups.

In summary, we have developed a computational analytical approach for analysis of HiC/TCC data that can identify domains and subdomains and can classify chromatin looping events. Through the use of epigenetic inhibitors, our work also provides insights into the interdependence of 3D chromatin looping and transcriptional control. We recognize that our current studies cannot determine if the enhancers that are affected by the epigenetic drugs are the same enhancers as identified by the chromosomal looping method. Future work using CRISPR/Cas9 to delete the TCF7L2-associated enhancers within the identified promoter-enhancer loops is needed to fully elucidate the mechanistic involvement of enhancer-mediated looping events in the regulation of drug-responsive genes. Nevertheless, our work provides genome-wide evidence that a strong association exists between a subset of enhancer-associated loops and enhancer-regulated genes.

4 |. METHODS

4.1 |. Tethered chromatin capture

TCC was performed as described.²⁵ Briefly, approximately 5e+07 PANC1 cells (ATCC, Manassas, VA, USA) were crosslinked with 1% formaldehyde for 10 minutes at room temperature, crosslinking was quenched with 0.125 M glycine for 5 minutes at room temperature, and cell pellets were collected and stored at –80°C. Nuclei were digested with 2000U HindIII (New England BioLabs, Ipswich, MA) before dilute solid-surface ligation reactions and TCC library preparation as described.²⁵ For drug treatments, PANC1 cells were grown to 60% confluency before a 24 hour treatment with 10 μM ICG-001 (Tocris, Bristol, UK), 10 μM C646 (Sigma-Aldrich, St. Louis, MO) or DMSO and fixed and harvested as described above. TCC libraries were sequenced using an Illumina HiSeq 2000.

4.2 |. Frequency contact matrix of TCC data

To assess the quality of our PANC1 TCC dataset, we compared our data to the publicly availably PANC1 HiC dataset on ENCODE (ENCSR440CTR). We processed the data using HiCPro⁴³ binned at 1 Mb resolution and visualized the contacts using Juicer tools.⁴⁴ For the remainder of the analysis, paired raw reads of TCC data for the PANC1 cell line were aligned to the human reference genome (hg19) by BWA⁴⁵ with default parameters. Reads were trimmed by 5 bp until 25 bp and aligned iteratively. Multiple aligned reads and reads with MAPQ less than 30 were removed. After performing fragment filtering (such as removing self-circles, error-pairs, and polymerase chain reaction duplicate reads), the reads were binned into either 500 Kb or 1 Mb size bins, where the sum of interaction pairs within the bins is used for all bin-bin interactions. The construction of a frequency contact matrix was as described in the previous publications.^46,47 Briefly, binned data was first subjected to normalization and transformation to Z-scores, the distribution of chromosomal interaction frequencies of both cell lines was then examined using a 500 Kb or 1Mb resolution for intra-chromosomal and a 1 Mb resolution for whole-genome contact matrices. More specifically, for every 500 Kb or 1 Mb bin of chromosome regions, the number of interactions (ie, Z-scores not equal to zero) between each chromosome region and the rest of the chromosome regions was counted. The chromosomal interaction frequency of the region was then calculated as the counted number of interactions in the region divided by the total number of chromosome regions (eg with a 1 Mb resolution there are 3029 chromosome regions in the human genome, and with a 500 Kb resolution there are 498 chromosome regions in Chr1). Z-scores of intra- or inter-chromosomal interaction matrices were then constructed as either a genome-wide contact heat map or a chromosome-specific intra-chromosomal contact heat map.

4.3 |. Topological domains of TCC data

A raw interaction contact matrix of each chromosome at 100 Kb resolution was normalized using Hi-Corrector,⁴⁸ which implements a set of scalable algorithms adapted from the original IC algorithm⁴⁹ for parallel computing. Domains were detected using TopDom⁵⁰ based on the local minima of normalized contact matrix. For two consecutive local minima, if any bin does not show a significant difference between the contact frequencies of within interactions and between interactions, they are defined as being within a topological domain (TAD); otherwise, they are either a boundary or a gap. The boundary and gap regions represent TAD-free chromatin at the given sequencing resolution and current parameter settings. We note that a boundary does not refer to the left or right side of a TAD, but is a specific region that has low interactions within itself and also between neighboring regions. Thus, based on this definition, there is not always a boundary between two TADs. A Gap is a region depleted of interactions.

4.4 |. HMM epigenetic states and subdomains

Histone modification marks (H3K27ac, H3K4me1, H3K4me3, H3K36me3, H3K27me3, H3K9me3), RNA Polymerase II, DNase-seq and TCF7L2 datasets in PANC1 cell lines were obtained from the ENCODE Project^35,51 (Supporting Information Table S2). The aligned (hg19) BAM files were downloaded from the University of California at Santa Cruz (UCSC) genome browser database.

The data was trained by a univariate first-order HMM⁵² to identify combinatorial epigenetic states. For each bin on the genome, the reads of each epigenetic mark were evaluated to determine if that mark is enriched in that bin (1) or not (0). We then used this binarized information of all epigenetic marks to train the HMM model for the default 300 iterations. For each combination of bin size (1 Kb) and number of states (8, 10, 12, and 20), five trainings of the HMM model were performed, and the best model was selected based on the BIC. The outputs of emission and transition matrices or states were visualized using the commercial MATLAB program. Consecutive bins of the same HMM states were merged into a single region, given that the bins were within the same domain defined by TopDom.

The emission probability of the HMM represents the distribution of the epigenetic marks in that particular bin, whereas the transition probability represents the possibility that a certain state should be assigned to a specific bin given the known state of the previous bin. If in a given state there were marks with an emission probability greater than 0.5, only these marks are considered as dominant marks for that state. For states without dominant marks, we used an emission probability cutoff of 0.1 for a mark to be considered as valid to identify a corresponding state.

The transition matrix indicates which states are frequently neighbors. In addition, states that have very low values in the emission matrix, such as S2 and S11 in Figure 2A, may represent epigenetic mark-depleted states. Therefore, based on the transition and emission matrix, as well as the other genomic features, certain epigenetic states were merged together to a single region and biologically defined as a sub-domain. These sub-domains reflect the epigenetic modification context over a chromatin structural domain.

4.5 |. TCC data modeling using HOMER

We used HOMER⁵² to find significant interactions or IPs in our TCC data. HOMER can search for pairs of loci that have a greater number of reads in interaction data than would be expected by chance using a background model. HOMER defines the expected number of reads as

$${\mathit{e}}_{\text{ij}}=\frac{\mathit{f}(i\mathbf{-}\mathit{j})\left({\mathit{n}}_{\mathit{i}}^{\mathbf{*}}\right)\left({\mathit{n}}_{\mathit{j}}^{\mathbf{*}}\right)}{\mathit{N}*}$$

where f is the expected frequency of reads, N* is estimated total number of reads, and n* is the estimated total number of interaction reads at each region. HOMER uses the actual number of interaction reads at each region as the initial value and then iteratively calculates the expected number of reads using the above model until the error between expected and observed reads totals per region is near zero. We examined genomic regions at 40 Kb resolution to find significant interactions, using a minimum distance of 10 Kb to consider an interaction between regions. The peaks were then further filtered using an FDR cutoff of 0.1 and distance cutoff of 20 Kb between loci centers.

In addition to the genomic location of the two interacting regions, HOMER also outputs a binomial p-value and FDR based on Benjamini correction. We further filtered the peaks using FDR ≤ 0.1 and loci distance greater than 20 Kb to isolate a more stringent set of IPs.

4.6 |. Associating chromatin interactions with epigenetic sub-domains

For each IP, the sub-domains overlapping (1bp) with the two peak loci were extracted. The average sizes of IPs are longer than that of sub-domains, so one peak locus may cover multiple sub-domains. Changes in histone modifications in treated PANC1 cells were calculated by first extracting and averaging reads in the sub-domains within a 100 bp bin, then dividing the averaged reads in drug-treated PANC1 cells by control PANC1 cells.

4.7 |. Associating chromatin interactions with annotated genes

For all annotated human RefSeq genes, we defined five genomic regions (Figure 3B) relevant to TSS, which are promoter (–5 Kb to +1 Kb), distal (±100 Kb), and far (beyond 100 Kb). Then, we defined the following seven categories of IPs: (a) PP1—any IPs with both ends within the same promoter; (b) PP2—any IPs having two ends located in two different promoters; (c) PD1—any IPs between a promoter and a distal region, with the closest TSS to the distal region being the same gene as for the promoter end; (d) PD2—any IPs between a promoter and a distal region, with the closest TSS to the distal region NOT being the same gene as for the promoter end; (e). PF1—any IPs between a promoter and a far region, with the closest TSS to the far region being the same gene as for the promoter end; (f) PF2—any IPs between a promoter and a far region, with the closest TSS to the far region NOT being the same gene as for the promoter end; and (g) Other—any IPs that do not involve a promoter.

4.8 |. Integrating gene expression datasets

All of our expression datasets were obtained from our previous study.³¹ Further details for each component of our analysis are described. For assessing the differential expression analysis for IP alterations within untreated and HAT-inhibitor treated PANC1 cells, total RNA was collected for untreated PANC1 cells and cells treated with either epigenetic inhibitor for 96 hours. Total RNA was collected for untreated PANC1 cells and cells treated with either epigenetic inhibitor for 96 hours. Total RNA was collected using Trizol (Life Technologies, Carlsbad, CA). Ultimately, these RNAs were labeled, hybridized and analyzed with Illumina HT-12 v4 Expression BeadChips (catalog#: BD-103–0204; Illumina, San Diego, CA) with the Direct Hydridization Assay and then scanned on an Illumina HiScan (catalog#: BD-103–0604). We analyzed the data as described.³¹ For this analysis, we used a log2FC cutoff of 0.5 and P-value < 0.05 for further analyses.

To incorporate TCF7L2 regulation we utilized our knockdown RNA-seq data from our previous study.³¹ Total RNA after knockdown with 40 nM siRNA targeting TCF7L2 (Life Technologies, Carlsbad, CA). We then performed RNA-sequencing on the polyA+RNA selected True-Seq libraries (Illumina, San Diego, CA) using the Illumina HiSeq. 2000 platform and differential expression was determined as described.³¹ For this analysis we used a log2FC cutoff of 0.5 and P-value < 0.05 for further analyses.

Ontology analyses were performed using GSEA³⁶ with default settings and survival analysis was done using SurvExpress.³⁷