Transcriptome analysis of pancreatic cancer reveals a tumor suppressor function for HNF1A

Summary

Transcriptome analyses indicate HNF1A is a key regulator of dysregulated gene expression sub-networks in pancreatic adenocarcinomas as compared to normal pancreatic tissues, and functional experiments suggest HNF1A down-regulation in pancreatic tumors benefits their survival and proliferation.

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is driven by the accumulation of somatic mutations, epigenetic modifications and changes in the micro-environment. New approaches to investigating disruptions of gene expression networks promise to uncover key regulators and pathways in carcinogenesis. We performed messenger RNA-sequencing in pancreatic normal (n = 10) and tumor (n = 8) derived tissue samples, as well as in pancreatic cancer cell lines (n = 9), to determine differential gene expression (DE) patterns. Sub-network enrichment analyses identified HNF1A as the regulator of the most significantly and consistently dysregulated expression sub-network in pancreatic tumor tissues and cells (median P = 7.56×10⁻⁷, median rank = 1, range = 1–25). To explore the effects of HNF1A expression in pancreatic tumor-derived cells, we generated stable HNF1A-inducible clones in two pancreatic cancer cell lines (PANC-1 and MIA PaCa-2) and observed growth inhibition (5.3-fold, P = 4.5×10⁻⁵ for MIA PaCa-2 clones; 7.2-fold, P = 2.2×10⁻⁵ for PANC-1 clones), and a G₀/G₁ cell cycle arrest and apoptosis upon induction. These effects correlated with HNF1A-induced down-regulation of 51 of 84 cell cycle genes (e.g. E2F1, CDK2, CDK4, MCM2/3/4/5, SKP2 and CCND1), decreased expression of anti-apoptotic genes (e.g. BIRC2/5/6 and AKT) and increased expression of pro-apoptotic genes (e.g. CASP4/9/10 and APAF1). In light of the established role of HNF1A in the regulation of pancreatic development and homeostasis, our data suggest that it also functions as an important tumor suppressor in the pancreas.

Introduction

Pancreatic cancer is a highly lethal cancer with a median survival of less than 6 months, and 5 year survival rate of <5% (1). Early stages of pancreatic neoplasms are usually asymptomatic, and thus difficult to detect. By the time pancreatic cancer is diagnosed the tumor has progressed to an unresectable and aggressively metastatic state. It has been estimated that it may take at least 15 years from the time of the initiating mutation to metastasis (2). Furthermore, pancreatic ductal adenocarcinoma (PDAC), the most common sporadic pancreatic cancer, rarely responds to modern chemotherapeutic regimens.

Risk factors for PDAC include family history, tobacco smoking, body mass index, chronic pancreatitis and diabetes (1,3–7). The epidemiologic association between diabetes and PDAC is complex, but long-term type-2 diabetes appears to confer up to a two-fold increased risk of pancreatic cancer (4,8). Hyperinsulinemia in response to insulin resistance may be partly responsible for this increased risk, as insulin promotes cell proliferation and increases glucose utilization (8). The association between pancreatitis and pancreatic cancer suggests that inflammation also contributes to carcinogenesis of PDAC (9,10).

PDAC arises from precursor lesions called Pancreatic Intraepithelial Neoplasia-1–3, as proposed in the Pancreatic Intraepithelial Neoplasia progression model (11,12). Somatic mutations and genomic rearrangements have been identified that drive tumorigenesis in pancreatic cancer (13–15). The best example is the mutation of codon 12 of KRAS observed in ~30% of Pancreatic Intraepithelial Neoplasia-1 neoplasms and nearly 100% of advanced PDAC (3,13,16). Additional somatic mutations have been observed in CDKN2A, TP53 and SMAD4, though none are as frequent as KRAS (3,13,14).

Approximately 5–10% of pancreatic cancer patients have a family history of the disease (3). Highly to moderately penetrant germline mutations have been found in the tumor suppressors CDKN2A, BRCA2, PALB2, ATM and STK11, the DNA mismatch repair gene MLH1, as well as in the hereditary pancreatitis associated genes PRSS1 and SPINK1, and the cystic fibrosis gene CFTR (1,3,17). Furthermore, recent genome wide association studies (GWAS) have identified multiple loci that harbor common germline susceptibility variants with small effect sizes located in intergenic or intronic regions on chromosomes 1q32.1 (NR5A2) 5p15.33 (TERT/CLPTM1L), 9q34.2 (ABO), 13q22.1 (nongenic) in populations of European descent, and on 3q29 (TFRC), 5p13.1 (DAB2), 6p25.3 (FOXQ1), 7q36.2 (DPP6), 10q26.11 (PRLHR), 12p11.21 (BICD1), 21q21.3 (BACH1), 21q22.3 (TFF1) and 22q13.32 (FAM19A5), in Asian populations (18–21).

With the rapid growth of publically available datasets and literature mining databases, it has become increasingly feasible to consider the effects of such tumorigenic influences in the context of regulatory networks (22). Identification of commonly dysregulated gene expression networks in PDAC could provide insight into the mechanisms of tumor progression. Towards this end, we have used a genome-wide approach to profile gene expression changes between tumor and normal derived pancreatic samples. An examination of these data in the context of gene expression sub-networks implicated a novel, central role for the transcription factor HNF1A in key gene regulatory programs of the exocrine pancreas, and functional experiments suggest HNF1A down-regulation benefits proliferation and survival of pancreatic tumor cells.

Materials and methods

Cell lines and tumors

Nine pancreatic cancer cell lines (AsPC-1, BxPC-3, Hs766T, SU.86.86, SW1990, CFPAC-1, Capan-1, PANC-1, MIA PaCa-2) were purchased from ATCC and cultured as recommended (http://www.ATCC.com). Eighteen fresh frozen pancreatic tissue samples (n = 8 tumor and n = 10 normal derived) were obtained from the Mayo Clinic in Rochester, MN. The project was approved by the Institutional Review Boards of both participating institutions. In light of accumulating evidence for acinar cells as potential progenitors of PDAC, we analyzed normal pancreatic tissue that consists mostly of acinar cells rather than microdissected ductal tissue (10,23). All tumors and cell lines were derived from the exocrine pancreas, and classified as PDAC.

RNA-sequencing

Transcriptome analysis was performed for the 9 pancreatic tumor-derived cell lines, 10 pancreatic tumor-adjacent normal and 8 pancreatic tumor tissue (60–90% tumor) samples listed above using next generation sequencing. One µg RNA (RNA integrity number scores >8.0) isolated with the Ambion mirVana kit was poly-A enriched with oligo(dT) beads, fragmented and subjected to complementary DNA (cDNA) synthesis, end repair, library construction and massively parallel sequencing at National Cancer Institute’s Center for Cancer Research Sequencing Facility (Supplementary Table 1, available at Carcinogenesis Online). Reads were aligned sequentially to RefSeq and Ensembl databases (NCBI Hg19) using BWA as previously described (24,25).

Differential expression analysis

Analysis of differential expression (DE) was performed with the EdgeR package for R from Bioconductor (http://www.bioconductor.org/packages/release/bioc/html/edgeR.html (26)). The following workflow was carried out separately for the comparisons of tumor tissues (n = 8) versus non-tumor tissues (n = 10) (TvN), and cell lines (n = 9) versus non-tumor tissues (n = 10) (CvN). First, genes were filtered out if fewer than three samples had ≥1 counts per million. Raw read counts were scaled based on total sample counts, adjusted for RNA composition bias, quantile normalized and fitted to a negative binomial distribution yielding pseudo-counts for each gene. Samples were grouped into ‘normal’ and ‘tumor’, and the magnitude (log₂[tumor/normal]) and significance (P-value) of differential expression between groups were calculated by an exact test using the pseudo-counts and tagwise dispersion estimates per gene. Genes with false discovery rate-corrected P-values < 0.05 were considered significant. We furthermore calculated log₂(ratios) for the combined tumor tissues and cell lines, as well as for each individual tumor-derived sample. For the combined DE data, pseudo-counts for all tumor tissues and cell lines were averaged for the numerator in the log₂(tumor/normal) DE calculation per gene. The individual tumor samples’ per gene DE values were calculated as the log₂(ratio) of the tumor sample’s pseudo-count to the average of the pseudo-counts across all normal samples.

Gene set enrichment analyses

Pathway and gene set enrichment analyses of DE transcripts were performed with the GeneRanker tool within the Genomatix Software Suite using the FuncAssociate algorithm (27). The pre-annotated pathway collections included Gene Ontology’s Biological Processes, canonical signal transduction pathways collected from National Cancer Institute-Nature’s Pathway Interaction Database and The Cancer Cell Map (http://cancer.cellmap.org), and Genomatix’s self-curated Disease associations (28,29). Supplementary Tables 2–4, available at Carcinogenesis Online include the top 10 most significant pathways, while Supplementary Tables 5–16 and 18, available at Carcinogenesis Online include all significant pathways based on an false discovery rate controlled P-value threshold of 0.05.

Examination of the genes annotated in the epithelial-mesenchymal transition (EMT) Gene Ontology Biological Processes pathway led us to conclude that this pathway is under-annotated with only 30 human proteins, and does not include universally recognized EMT factors, such as E-cadherin, ZEB1 and ZEB2. Therefore, we generated our own EMT pathway through manual examination of the literature (Supplementary Table 19, available at Carcinogenesis Online). For custom pathway analyses we used the FuncAssociate 2.0 tool (http://llama.mshri.on.ca/funcassociate/) with default settings (27).

Expression regulatory sub-network enrichment analyses

Expression regulatory sub-network analyses were carried out with Elsevier’s Pathway Studio 8.0 on the TvN, CvN and individual tumor DE data (30). The analysis method is described here, http://help.elsevier.com/app/answers/detail/a_id/2969/p/9047/c/9035. In brief, sub-networks were first dynamically generated based on expression regulation and promoter binding relationships in the ResNet 9.0 database. A sub-network was defined as a group of genes with a common upstream expression regulator (sub-network seed). Then the unfiltered gene lists with corresponding log₂ expression ratios for each comparison was tested for enrichment among the generated sub-networks. An expression regulatory sub-network is more significantly enriched when a greater number of downstream target genes are differentially expressed, and when the magnitude of their differential expression is larger on average. The threshold for significance was set at P < 0.05. We locally modified the ResNet database to remove false HNF1A relationships that were due to the ambiguous alias TCF1 in the literature, which sometimes refers to T-cell specific factor 1 that should now be referred to as TCF7. The HNF1A sub-network was the most significantly dysregulated in both TvN and CvN comparisons prior to removal of the false relationships. Images of sub-networks were made with Pathway Studio 10.

Gene expression correlation with HNF1A

Pearson and Spearman correlation coefficients and P-values were generated in R from log₂(ratios) of individual tumor tissue samples for each gene with HNF1A. Genes with both Pearson and Spearman correlations greater than 0.7 or less than −0.7 were subjected to gene set enrichment analysis as described above.

HNF1A-inducible MIA PaCa-2 and PANC-1 clones

MIA PaCa-2 and PANC-1 cell lines were stably transfected with Clontech TetOn3G transactivator (pLVX-Tet3G) via lentiviral transduction (MOI = 3), and selected with 5 µg/ml blasticidin. Human HNF1A cDNA (Thermo Scientific MGC cDNA clone 8143911) was sub-cloned into the lentiviral Tet response vector (pLVX-TRE3G). The resulting construct or empty vector were stably transduced into the TetOn3G expressing cells (MOI = 1), and selected with 2 µg/ml puromycin. Clones were isolated, expanded and tested for HNF1A inducibility by Western analysis.

Growth rate analyses

Cells were plated (3000 per well for PANC-1 derived cells; 1500 for MIA PaCa-2 derived cells) in 96-well plates, and treated (in triplicate or quadruplicate) with 0–1000ng/ml doxycycline (Clontech) for up to 8 days. WST-1 (Roche) was added daily to 1% to the assayed wells and incubated for 30min at 37°C. Absorbance was measured at 450nm and 600nm on a Promega GloMax Multidetection System plate reader. After reference correction, 600nm absorbance values were subtracted from 450nm values, and the averages and standard deviations of replicate wells were calculated.

Cell cycle analysis by flow cytometry

Cultures of HNF1A inducible clones were grown to ≤90% confluency for 0–3 days with 10 or 100ng/ml doxycycline (Clontech). Cells were trypsinized, washed in 1× phosphate-buffered saline (PBS); (Corning) and fixed in 70% ethanol in 1× PBS at −20°C overnight. After washing twice in 1× PBS, cells were resuspended in 0.5 or 1.0ml 1× PBS. RNase A (Ambion) was added to 1 µg/ml and cells incubated for 1 h at 37°C, after which propidium iodide (Sigma-Aldrich) was added to 40 μg/ml. At least 30 000 cells for each sample were counted and measured for propidium iodide fluorescence on a Becton Dickson FACSCalibur. Cell cycle phase populations were estimated after doublet discrimination using ModFit LT (http://www.vsh.com/products/mflt/).

RT-qPCR array analyses

Logarithmically growing cultures of control (empty vector) or HNF1A-inducible PANC-1 or MIA PaCa-2 clones were grown for 2 days ± 10 or 100ng/ml doxycycline, respectively, and harvested for RNA extraction with the Ambion mirVana kit (25). RNA was reverse transcribed to cDNA with the QIAGEN RT² First Strand kit. Expression was assayed on QIAGEN’s Human Cell Cycle RT² Profiler PCR Array and Human Apoptosis RT² Profiler PCR Array with QIAGEN RT² SYBR Green ROX qPCR Mastermix, and run in an ABI 7900HT thermal cycler. Relative Quantities (RQ) of transcripts were calculated by ΔΔC_t using stable housekeeping genes as endogenous controls and comparing HNF1A-induced to uninduced samples. HNF1A-dependent differential expression was determined by unpaired student t-test between the log₂ expression ratios of the four HNF1A-inducible clones (two each of MIA PaCa-2 and PANC-1) versus the four uninducible clones (two each of MIA PaCa-2 and PANC-1).

Immunohistochemical analysis

Normal pancreas tissue (n = 3) was obtained from organ donors through a transplantation program (Hospital Germans Trias i Pujol, through Dr. R.P-Borrell (31)). Tissue microarrays (TMAs) with formalin-fixed tumor-derived pancreatic tissue samples were obtained from the Mayo Clinic (32). Two TMAs were used as previously described (33): PDAC TMA1 (n = 128), PDAC gemcitabine (n = 151). Immunohistochemistry was performed with a rabbit polyclonal anti-HNF1A antibody (Santa Cruz Biotechnology: sc-10791) at 1:75 dilution (in PBS with 1% bovine serum albumin). The specificity of the antibody was demonstrated using in vitro knockdown of HNF1A in cultured AsPC-1 pancreatic cancer cells (not shown). Sections were deparaffinized, and hydrated. Antigen retrieval was performed using sodium citrate (pH 6.0). Slides were incubated with H₂O₂ to block endogenous peroxidase. Primary antibody was incubated overnight at 4°C; after washing with PBS, the avidin-biotin immunoperoxidase complex method was applied using diaminobenzidine as substrate (Dako). Sections were lightly counterstained with hematoxilin. Histological images were acquired with a Nikon TE2000E microscope. HNF1A staining was scored based on intensity (on a scale from 0–3; 0, negative; 1, weak; 2, positive; 3, strong) and the proportion of reactive cells (0–100%); histoscore was determined by multiplying both parameters (range 0–300). When more than one core was available from a given tumor, the mean score was used.

Differences in clinical parameters (survival, age at surgery, age at onset, sex, body mass index, previous diagnosis of diabetes, personal history of pancreatitis and tumor grade) were evaluated across quintiles of average histoscore. A log-rank test was used for censored data (survival), a Kruskal Wallis test for continuous variables (age at surgery, age at onset) and a Chi-Square test for categorical variables (sex, body mass index, previous diabetes diagnosis, pancreatitis, tumor grade).

Sample collection and studies were approved by the Ethics Committees of the participating institutions.

Results

Identification of differentially expressed (DE) genes

Paired-end messenger RNA-sequencing was performed in pancreatic tumor tissue samples (T; n = 8), non-tumor pancreatic tissue samples (N; n = 10) and pancreatic tumor-derived cell lines (C; n = 9) (Supplementary Table 1, available at Carcinogenesis Online). Gene expression was compared in tumor versus non-tumor tissue samples (TvN) and in cell lines versus non-tumor tissue samples (CvN). Genes were designated differentially expressed (DE) if the false discovery rate-adjusted P-value was ≤0.05 (see Materials and Methods and Supplementary Figure 1, available at Carcinogenesis Online). As each comparison (TvN and CvN) emphasizes important features of the tumor microenvironment and cellular homogeneity, respectively, there were clear differences between the two (Supplementary Figure 2, available at Carcinogenesis Online). However, there was also considerable overlap with 4147 genes with statistically significant DE in the same direction in both comparisons (1792 up-regulated and 2355 down-regulated in tumor).

Gene set enrichment analyses of DE genes

We next performed pathway analyses using the GeneRanker tool within the Genomatix Software Suite. Separate analyses of up- and down-regulated DE genes indicated clear patterns of enriched pathways for TvN versus CvN (Supplementary Tables 2–16, available at Carcinogenesis Online). In TvN there was an enrichment of up-regulated genes in pathways related to immune response, inflammation, epithelial-mesenchymal transition (EMT, inferred from enrichments in adhesion, motility and integrin signaling pathways; see Supplementary Table 19, available at Carcinogenesis Online), while signaling pathways involved in pancreatic development and homeostasis were down-regulated. In CvN, the up-regulated pathways more prominently featured cell cycle and mitosis, while a significant down-regulation of genes involved in pancreas development and cell adhesion was noted. Both TvN and CvN comparisons showed significant up-regulation of genes associated with pancreatic neoplasms and carcinomas in general, while notably having significant down-regulation of genes associated with diabetes and pancreatitis. These pathway enrichments underscore the similarities and differences between in situ tumors and cultured cells.

Expression regulatory sub-network enrichment analysis

Next we performed expression regulatory sub-network enrichment analyses using the log₂ expression ratios of all expressed genes. A sub-network was defined as a group of genes with one common upstream expression regulator (direct or indirect) using Elsevier Pathway Studio 8.0. In TvN, the most significantly dysregulated sub-networks are those controlled by cytokines (e.g. IL1A, IL13 and IL4) and transcription factors involved in pancreatic development and homeostasis (e.g. HNF1A, HNF1B and FOXA3) (Table 1, Supplementary Table 17, available at Carcinogenesis Online). Consistent with the pathway analyses described above, the CvN sub-network analysis did not reveal significant alterations of immune or inflammatory related sub-networks, but instead, featured pancreatic development and homeostasis factors, including HNF1A, GATA6, NEUROG3, PDX1 and NEUROD1 (Table 1, Supplementary Table 17, available at Carcinogenesis Online).

Table I.

The most significantly dysregulated sub-networks in pancreatic tumors and cells

Tumor Tissue versus Normal (TvN)						Cells versus Normal (CvN)
Rank	Total neighbors	Measured neighbors	Gene set seed	Median change	P	Rank	Total neighbors	Measured neighbors	Gene set seed	Median change	P
1	301	259	IL1A	2.07	8.16×10−6	1	197	139	HNF1A	−2.29	8.55×10−7
2	197	146	HNF1A	−1.05	1.92×10−5	2	93	70	GATA6	−2.96	5.13×10−5
3	510	431	inflammatory cytokine	1.84	3.58×10−5	3	10	9	MLN	−132.96	1.53×10−4
4	19	16	FOXA3	−1.79	4.93×10−5	4	37	26	NEUROG3	−8.99	1.74×10−4
5	26	25	MMP1	4.43	5.11×10−5	5	75	58	PDX1	−2.26	2.20×10−4
6	77	59	HNF1B	−1.42	5.42×10−5	6	431	321	CEBPA	−1.64	2.72×10−4
7	308	244	IL13	2.27	1.41×10−4	7	8	7	SSTR5	−37.24	3.02×10−4
8	163	131	allergen	2.20	1.43×10−4	8	58	43	NEUROD1	−4.23	3.43×10−4
9	572	490	IL4	2.07	2.79×10−4	9	63	41	ASCL1	−2.87	3.73×10−4
10	84	70	JUND	2.56	2.97×10−4	10	73	71	E2F4	5.74	4.28×10−4
11	29	22	WNT7A	2.54	4.09×10−4	11	70	53	CCK	−2.81	4.78×10−4
12	30	23	ONECUT1	−3.62	5.62×10−4	12	43	26	TRH	−14.23	4.92×10−4
13	210	172	LIF	1.56	6.62×10−4	13	385	295	CEBPB	−1.33	6.60×10−4
14	52	46	MAP2K3	2.60	7.22×10−4	14	19	15	FOXA3	−8.02	7.56×10−4
15	660	544	IL6	1.61	8.09×10−4	15	19	12	PTF1A	−32.38	7.78×10−4
16	50	35	MSX2	3.33	8.17×10−4	16	121	96	AVP	−2.72	9.15×10−4
17	163	145	OSM	2.03	9.18×10−4	17	77	58	HNF1B	−2.53	9.70×10−4
18	89	67	JUNB	2.30	9.37×10−4	18	24	13	SCT	−1136.59	1.52×10−3
19	5	5	TAB2	11.08	1.03×10−3	19	26	22	IHH	1.97	1.54×10−3
20	1284	1100	TGFB1	1.72	1.08×10−3	20	11	6	BGLAP	−17.27	1.71×10−3
21	29	20	KLF15	1.57	1.10×10−3	21	146	118	FOXM1	1.35	1.74×10−3
22	75	60	PDX1	−1.78	1.11×10−3	22	54	47	FGFR1	1.03	1.99×10−3
23	339	279	NR3C1	1.51	1.27×10−3	23	339	257	NR3C1	−1.32	2.07×10−3
24	63	38	ASCL1	−1.46	1.27×10−3	24	78	61	IL11	−1.16	2.11×10−3
25	86	77	collagen type I	2.27	1.56×10−3	25	112	97	PTHLH	−1.01	2.28×10−3

Full lists of significant sub-networks presented in Supplementary Table 17, available at Carcinogenesis Online

To determine the contribution of each individual tumor sample to the sub-network results, we calculated DE for each tumor sample and cell line individually, as compared to the average of normal-derived samples. The most consistently dysregulated sub-network was regulated by HNF1A (Table 2 and Figure 1; median P = 7.56 × 10^{− 7}, median rank 1; rank range 1–25). The second sub-network ranked by consistency was regulated by HNF1B (median P = 2.14 × 10^{− 4}, median rank 10, rank range 2-not significant), which is paralogous to HNF1A (34). Both HNF1A and HNF1B were down-regulated in all tumor tissue samples and all but one cell line. The only sample showing up-regulation of HNF1A was the pancreatic cancer cell line with the least significant dysregulation of the HNF1A sub-network (AsPC-1, see Table 2); this cell line showed the highest levels of HNF1A protein among all lines tested by Western blot (Figure 2A). These analyses suggest that reduced HNF1A activity may be an important step in tumorigenesis, tumor progression and/or maintenance.

Table II.

The most consistently dysregulated sub-networks in pancreatic tumors and cells

Samples	HNF1A		HNF1B		PDX1		PTF1A		FOXA3
	Rank	P	Rank	P	Rank	P	Rank	P	Rank	P
All tumor tissues	2	1.92×10−5	6	5.42×10−5	22	1.11×10−3	78	8.07×10−3	4	4.93×10−5
Tumor tissue T1	1	4.18×10−14	6	2.08×10−6	3	5.40×10−8	10	1.34×10−5	9	9.79×10−6
Tumor tissue T2	1	1.36×10−6	10	3.06×10−4	3	1.17×10−5	4	2.78×10−5	2	6.04×10−6
Tumor tissue T3	1	2.03×10−9	6	1.32×10−5	2	4.47×10−9	7	2.60×10−5	9	4.83×10−5
Tumor tissue T4	2	6.87×10−6	7	4.74×10−5	NS	NS	106	1.06×10−2	184	2.49×10−2
Tumor tissue T5	6	3.57×10−5	4	1.96×10−5	195	2.48×10−2	69	6.73×10−3	31	2.43×10−3
Tumor tissue T6	1	1.70×10−7	3	7.37×10−5	NS	NS	NS	NS	16	5.43×10−4
Tumor tissue T7	1	1.39×10−7	2	6.83×10−7	10	3.66×10−4	76	6.37×10−3	3	1.94×10−5
Tumor tissue T8	2	2.46×10−7	10	3.31×10−4	1	1.24×10−8	4	2.54×10−4	18	9.87×10−4
All cell lines	1	8.55×10−7	17	9.70×10−4	5	2.20×10−4	15	7.78×10−4	14	7.56×10−4
PANC-1	1	3.35×10−8	18	2.46×10−4	86	7.53×10−3	11	1.19×10−4	NS	NS
AsPC-1	25	5.79×10−4	NS	NS	NS	NS	59	2.24×10−3	NS	NS
BxPC-3	1	2.08×10−9	8	1.20×10−4	10	2.32×10−4	31	1.76×10−3	80	6.98×10−3
CFPAC-1	12	3.36×10−4	204	3.44×10−2	3	9.86×10−5	8	1.88×10−4	39	1.77×10−3
SU.86.86	3	1.56×10−5	43	2.34×10−3	17	3.97×10−4	6	5.30×10−5	58	3.61×10−3
MIA PaCa-2	2	1.24×10−6	90	4.73×10−3	NS	NS	45	1.15×10−3	279	2.95×10−2
SW1990	1	1.87×10−8	24	1.84×10−4	25	1.93×10−4	23	1.78×10−4	12	5.25×10−5
Capan-1	1	6.58×10−7	94	8.18×10−3	18	6.26×10−4	10	4.02×10−4	92	8.03×10−3
Hs766T	1	6.92×10−12	7	3.31×10−5	22	7.68×10−4	4	1.88×10−5	2	7.41×10−6
Median	1	7.56×10−7	10	2.15×10−4	17.5	3.81×10−4	19	5.90×10−4	17	9.03×10−4
Rank Range	1–25		2–NS		1–NS		4–NS		2–NS

NS = not significant

Fig. 1.

Dysregulation of the HNF1A expression regulatory sub-network in pancreatic tumors. The HNF1A sub-network shown contains all downstream targets of HNF1A expression regulation and/or promoter binding in the TvN differential expression (DE) analysis. The intensity of bubble shading indicates the degree of differential expression (blue = down-regulated, red = up-regulated, white = unchanged). Blue lines represent expression regulation relationships and green lines represent promoter binding relationships based on Elsevier’s ResNet9.0 literature mining database. Genes were grouped according to the indicated Gene Ontology Biological Processes. Oval = protein; Diamond = ligand; Light bulb = receptor (e.g. FGFR4); Crescent = kinase (e.g. SRC); Combed circle = transcription factor (e.g. HNF1A); Irregular box = phosphatase (e.g. PTPN6).

Fig. 2.

HNF1A induction arrests proliferation of pancreatic cancer cells. (A) Western blot analysis of pancreatic cancer cell lines indicating HNF1A protein levels consistent with RNA-seq results. (B) Western blot analysis of two representative HNF1A-inducible PANC-1 clones indicating HNF1A induction is robust. See Supplementary Figure 3, available at Carcinogenesis Online, for full blot images. (C and E) Growth rates were measured in vitro over 4 days. Growth curves are based on normalized WST-1 absorbance as indicator of viable cell numbers. (C) Growth curve for representative MIA PaCa-2 HNF1A-inducible clone with increasing doxycycline concentrations. By day 4, there are 5.3-fold fewer viable cells with 100ng/ml doxycycline compared to no induction (P = 4.6×10⁻⁵). (E) Growth curve for a representative control (empty vector) MIA PaCa-2 clone treated with increasing doxycycline concentrations. (D and F) Flow cytometry analysis was performed on propidium iodide stained clones grown in 100ng/ml doxycycline for the indicated number of days. Population percentages were estimated by cell cycle phase modeling with ModFit LT. (D) Cell cycle analysis of a representative HNF1A-inducible MIA PaCa-2 clone. (F) Cell cycle analysis of a representative control (empty vector) MIA PaCa-2 clone. Other clones showed similar results. Error bars indicate 1 s.d.

Correlation of differentially expressed genes with HNF1A

In light of the apparent role for HNF1A in the altered genetic programming of pancreatic tumors, we wanted to identify genes whose expression changes were highly correlated with that of HNF1A, which was down-regulated in both TvN and CvN (log₂ expression ratios of −1.99 and −1.60, respectively). The log₂ expression ratios for 811 genes were highly correlated with HNF1A (both r and ρ > 0.7 or < −0.7) in the tumor tissue samples. Pathway analysis of these genes revealed enrichment in β1-integrin, discoidin domain receptor and TGF-β signaling, and biological processes related to extracellular matrix organization and cell adhesion (Supplementary Table 18, available at Carcinogenesis Online), which are all related to EMT (see Supplementary Table 19, available at Carcinogenesis Online). Since HNF1A regulates drivers and markers of EMT in liver cells, we considered a general EMT pathway (35). Due to the clear lack of well-known EMT factors in the Gene Ontology Biological Processes EMT pathway, we generated a custom EMT pathway (see Materials and Methods and Supplementary Table 19, available at Carcinogenesis Online). Enrichment analysis indicated a significant (P _unadjusted = 1.28 × 10^{− 6}, P _adjusted < 0.001) over-representation of genes highly correlated with HNF1A in this pathway.

Phenotypic effects of induced HNF1A overexpression

To experimentally validate the importance of HNF1A activity, we generated MIA PaCa-2 and PANC-1 clones with inducible (Tet-On) HNF1A expression, and assessed phenotypic effects of HNF1A overexpression. Endogenous HNF1A was low in these cells, and induction of HNF1A was robust, even at low doxycycline concentrations (Figure 2A and B). After 4 days at 100ng/ml doxycycline, HNF1A-inducible MIA PaCa-2 clones (n = 2) displayed slower growth rates in induced as compared to uninduced cells (5.3-fold reduction on average, P = 4.6 × 10^{− 5}) (Figure 2C). The slower growing HNF1A-inducible PANC-1 clones (n = 4) displayed an average 7.2-fold decrease in cell growth between induced and uninduced cells on day 8 (P = 2.24 × 10^{− 5}) (Supplementary Figure 4, available at Carcinogenesis Online). Both cell lines showed evidence of viable cell loss.

To explore the cause of this striking growth inhibition, cell cycle analysis was performed by flow cytometry. By 2 days after HNF1A induction in MIA PaCa-2 and PANC-1 clones, the average ratio of cells in G₀/G₁ phase to S phase increased 5.0-fold, and this further increased to 7.3-fold by day 3 (Figure 2D, Supplementary Figure 5, available at Carcinogenesis Online). This is consistent with the observed growth arrest, and suggests a G₀/G₁ cell cycle block. Interestingly, the G₂/M phase population did not decline, suggesting a possible G₂/M cell cycle block as well. The MIA PaCa-2 clones also accumulated a sub-G₀/G₁ cell population, indicating apoptosis may play a role in the reduced cell viability observed (Figure 2D). The uninducible control clones did not show the same effect on cell growth or cell cycle arrest upon doxycyline treatment (Figures 2E and F).

To investigate the genes and pathways underlying the growth arrest triggered by HNF1A, we performed RT-qPCR with PCR arrays that include assays for 84 known cell cycle- or apoptosis-related genes (Supplementary Table 20, available at Carcinogenesis Online). After 2 days of treatment with doxycycline, RNA from MIA PaCa-2 and PANC-1 derived HNF1A-inducible clones (2 each) were compared to uninducible control clones (2 each). Fifty-one of the assayed cell cycle genes were significantly (P < 0.05) down-regulated upon HNF1A overexpression, including several that may be relevant to the observed G₀/G₁ arrest (e.g. E2F1, CDK2, CDK4, SKP2, CCND1, MCM2, MCM3, MCM4, MCM5, TFDP1, and CCNA2) (Figure 3A). CCND2 and CDK7 were the only cell cycle genes significantly up-regulated upon HNF1A induction. On the apoptosis array, 11 of the assayed genes were significantly (P < 0.05) down-regulated, while 10 were up-regulated (Figure 3B). Notably, the down-regulated genes included IAP (inhibitors of apoptosis) family members BIRC2, BIRC5 and BIRC6, as well as AKT, which coordinates survival programs through inhibition of apoptotic effectors and reduced expression of pro-apoptotic genes (36). Furthermore, expression of genes encoding Caspases (CASP4, CASP9 and CASP10) was higher upon HNF1A induction, along with APAF1 that forms the apoptosome with Caspase-9 and Cytochrome c. These gene expression changes suggest HNF1A overexpression erodes anti-apoptotic buffering capacity while increasing the pro-apoptotic load.

Fig. 3.

RT-qPCR reveals HNF1A-induced changes in cell cycle and apoptosis gene expression. Graph displays the average log₂ expression ratio between HNF1A induced and uninduced PANC-1 and MIA PaCa-2 clones (two clones each) for all genes with significantly altered expression (P < 0.05) in inducible versus uninducible control clones. (A) 51 of the 84 assayed cell cycle genes were down-regulated, and two were up-regulated. (B) 10 of the 84 assayed apoptosis genes were down-regulated, while 11 were up-regulated. Error bars indicate 1 s.d.

Immunohistochemical analysis of HNF1A in pancreatic tumors

Based on the observed dysregulation of the HNF1A expression sub-network and its probable tumor suppressive effects, we looked for associations between HNF1A protein staining in pancreatic tumors (n = 279) and clinical measures by immunohistochemical staining on TMAs (Supplementary Figure 6, available at Carcinogenesis Online). Low HNF1A expression (histoscore < 106) was found in 43% and 36% of tumors present in each of the two microarrays, respectively, supporting the notion that HNF1A expression is lost in a subset of PDAC. Furthermore, cytoplasmic immunostaining was observed in 34% of tumors but not in normal tissue (not shown). We examined potential associations between HNF1A histoscore (range 5–260) and survival, age at diagnosis, sex, body mass index, history of diabetes, self-reported pancreatitis, and tumor grade (Supplementary Table 21, available at Carcinogenesis Online). The only significant association noted was between HNF1A staining and pancreatitis (P = 0.047). Specifically, tumors from pancreatic cancer cases with self-reported pancreatitis had lower levels of HNF1A.

Discussion

Expression regulatory sub-network analysis allows the identification of important regulators with broad impacts on genetic programs regardless of expression changes in the regulators themselves (30). While the down-regulation of HNF1A expression is unremarkable, its sub-network stands out as the most significantly and consistently dysregulated sub-network in our analysis, which is notable because HNF1A is implicated in pancreatic development, pancreatitis and diabetes (37,38). The second most consistently dysregulated sub-network is regulated by HNF1B, a paralog of HNF1A that can act both as a co-factor with HNF1A or in a compensatory manner after loss of HNF1A activity (34).

HNF1A encodes an atypical homeodomain-containing protein (hepatocyte nuclear factor 1α) initially characterized as a liver-specific transcription factor, but now known to play important developmental and/or homeostatic roles in liver, pancreas, intestine and kidney (34,39–41). Well over 100 germ-line mutations in HNF1A have been identified as causes of maturity onset diabetes of the young type 3, which account for 21–64% of all maturity onset diabetes of the young cases (42,43). Also, a GWAS identified a common, non-synonymous variant (rs1169288, I27L) in the HNF1A gene associated with type-2 diabetes risk (44). Interestingly, pathway-based and pleiotropic analyses of pancreatic cancer GWAS data have also identified potential pancreatic cancer risk SNPs in the HNF1A gene (45,46).

The majority of HNF1A studies have been conducted in liver and/or pancreatic islet cells (40,47). They revealed direct expression regulatory effects of HNF1A that are tissue specific and, sometimes, opposite in direction for liver and islet cells. A role for HNF1A in acinar cell homeostasis and regeneration after caerulein-induced pancreatitis was recently discovered using mouse models (37). Pancreatitis led to a transient decrease in HNF1A levels mirrored by a decrease in acinar transcription factors and digestive enzymes, and increased acinar cell proliferation. This is consistent with both the inverse relationship between HNF1A staining and self-reported pancreatitis we observed in pancreatic tumors, and the growth arrest caused by HNF1A induction in pancreatic cell lines. Furthermore, Molero et al. found Hnf1a null mice exhibited reduced expression of the master acinar gene regulator PTF1A (pancreas transcription factor 1 alpha) and digestive enzyme messenger RNAs, and higher basal acinar cell proliferation that was further increased by damage (caerulein-induced pancreatitis). In silico examination of PTF1A and digestive enzyme promoters did not reveal HNF1A binding elements, but instead an HNF1A binding site was found in the Nr5a2 promoter and confirmed by ChIP analysis. Additionally, NR5A2 levels were strongly reduced in Hnf1a null mice and in wild type mice with induced pancreatitis. Thus, HNF1A may in part exert its influence on the exocrine pancreas through regulation of NR5A2. This is intriguing considering that a pancreatic cancer risk locus resides within the first intron of NR5A2, and a full NR5A2 dose is required for effective recovery from induced pancreatitis and suppression of tumor progression in an oncogenic Kras mouse background (19,48). HNF1A also directly regulates PDX1 expression, which encodes another important regulator of both exocrine and endocrine pancreatic function (37,38). It is interesting that, along with HNF1A and HNF1B, PDX1 and PTF1A are among the top five most consistently dysregulated sub-networks in our analyses. Furthermore, mutations in HNF1A, HNF1B, HNF4A and PDX1 can each cause maturity onset diabetes of the young, and SNPs localizing within the NR5A2, HNF1A, PDX1 and HNF1B gene regions have been identified as markers of pancreatic cancer risk loci through GWAS or pathway-based GWAS analyses (19,46). Therefore, HNF1A holds a central position in a regulatory network of transcription factors required for both exocrine and endocrine pancreas development, homeostasis and regeneration (Figure 4). It is worth considering whether dysregulation of this highly interconnected network might underlie a molecular mechanism behind the associations between pancreatic cancer and both diabetes and pancreatitis.

Fig. 4.

HNF1A is central to a highly interconnected network of transcription factors regulating pancreatic development and homeostasis that are also implicated in pancreatic cancer, pancreatitis and/or diabetes. Blue internal shading indicates down-regulation in TvN comparison. Lines between nodes indicate verified relationships from Elsevier’s ResNet 9.0 literature mining database (blue = expression regulation; green = promoter binding; purple = protein binding). External highlights around nodes indicate notable attributes (blue = harbors potential pancreatic cancer risk locus; red = mutation causes maturity onset diabetes of the young; yellow = implicated in recovery from pancreatitis; green = consistently dysregulated sub-network in our study).

Our results support the hypothesis that HNF1A is important for maintenance of the normal epithelial acinar phenotype. A recent study reported forced HNF1A expression in hepatocellular carcinomas restored liver-specific expression patterns and arrested proliferation, thus reprogramming the tumor cells to a normal-like state (49). Furthermore, HNF1A expression in hepatoma cells in vivo reduced tumorigenicity and inhibited tumor growth. These effects corroborate our observations that HNF1A inhibits growth of pancreatic cancer cells. Targeted analysis confirmed consistent changes in the expression of cell cycle genes upon HNF1A induction, potentially contributing to the G₀/G₁ arrest. In addition to halting proliferation, HNF1A induction was associated with down-regulation of inhibitors of apoptosis and concurrent up-regulation of apoptosome components. These effects represent an erosion of the anti-apoptotic buffer with a concomitant increase in the pro-apoptotic load in the MIA PaCa-2 and PANC-1 inducible clones (36). Interestingly, a previous study indicated that knockdown of BIRC5 alone caused a significant increase in apoptotic MIA PaCa-2 cells (50). These data suggest that down-regulating HNF1A could be a critical step in pancreatic tumor progression, tipping the scale toward proliferation and survival.

In conclusion, our transcriptomic analysis of differential gene expression in pancreatic tumors identified altered pathways and sub-networks, some of which are consistent with known pancreatic tumor biology. The most notable novel observation is the dysregulation of the HNF1A expression sub-network. Little is known of the role of HNF1A in the exocrine pancreas or PDAC, but bioinformatics and functional analyses, combined with knowledge of HNF1A, strongly suggest it acts as a pancreatic tumor suppressor that is important for normal exocrine pancreas development, functioning, regeneration and maintenance of the epithelial state. The down-regulation of HNF1A protein in PDAC further supports the notion that HNF1A acts as a tumor suppressor in the exocrine pancreas. Since neither HNF1A mutation nor gene loss has been reported in PDAC, functional restoration may provide a novel approach to inhibiting tumor growth.

Supplementary material

Supplementary Tables 1–21 and Figures 1–6 can be found at http://carcin.oxfordjournals.org/

Funding

Intramural Research Program of the Division of Cancer Epidemiology and Genetics; National Cancer Institute; National Institutes of Health (HHSN261200800001E).

Glossary

Abbreviations:

cDNA

complementary DNA

EMT

epithelial-mesenchymal transition

GWAS

genome wide association studies

PBS

phosphate-buffered saline

PDAC

Pancreatic ductal adenocarcinoma

TMA

Tissue microarray.