Genome-Wide Analyses of ChIP-Seq Derived FOXA2 DNA Occupancy in Liver Points to Genetic Networks Underpinning Multiple Complex Traits

Abstract

Background:

Forkhead Box A2 (FOXA2) exerts an influence on glucose homeostasis via activity in the liver. In addition, a key genome-wide association study (GWAS) recently demonstrated that genetic variation, namely rs6048205, at the FOXA2 locus is robustly associated with fasting glucose levels. Our hypothesis was that this DNA-binding protein regulates the expression of a set of molecular pathways critical to endocrine traits.

Methods:

Drawing on our laboratory and bioinformatic experience with chromatin immunoprecipitation followed by massively parallel sequencing, we analyzed our existing FOXA2 chromatin immunoprecipitation followed by massively parallel sequencing data generated in human liver, using the algorithm hypergeometric optimization of motif enrichment, to gain insight into its global genomic binding pattern from a disease perspective.

Results:

We performed a pathway analysis of the gene list using the gene set enrichment analysis algorithm, which yielded a number of significant annotations. Motivated by the fact that the FOXA2 locus has been implicated by GWAS, we cross-referenced the occupancy sites with the National Institutes of Health GWAS catalog and found strong evidence for the enrichment of loci implicated in endocrine, neuropsychiatric, cardiovascular, and cancer trait categories, but interestingly there was no evidence for enrichment for inflammation related traits. Intriguingly, a FOXA2 occupancy site coincided with rs6048205, suggesting that this variant confers its effect, at least partially, via a perturbation of a FOXA2 feedback mechanism.

Conclusion:

Our data strongly suggest that FOXA2 is acting as a master regulator of key pathways that are enriched for loci implicated by GWAS for most trait categories, with the clear exception of inflammation, suggesting that this factor exerts its effect in this context via noninflammatory processes.

The Forkhead box A (FOXA) family of proteins, FOXA1, FOXA2, and FOXA3, have been implicated as pioneer transcription factors in the initiation of chromatin remodeling during organ-specific transcriptional programing (1), including the liver, pancreas, lung, and dopaminergic neurons in the midbrain (2).

A recent genome-wide association study (GWAS) of fasting glycemic traits revealed multiple loci associated with glucose levels, including FOXA2 (3). This finding reinforces an existing body of evidence that FOXA2 plays a key role in glucose homeostasis. By defining the pattern of genomic binding sites occupied by this transcription factor, one has the opportunity to gain insights in to molecular pathways involved in disease processes.

We previously carried out chromatin immunoprecipitation followed by massively parallel sequencing (ChIP-Seq) with another GWAS-implicated transcription member, namely TCF7L2, which is widely considered the most strongly associated locus for type 2 diabetes (T2D) reported to date (4), to similarly elucidate its genome-wide occupancy (5, 6). Unexpectedly, and despite using the carcinoma cell line HCT116, our initial data suggested that the genes coinciding with TCF7L2 occupancy sites are strongly enriched in pathway categories related to metabolic-related functions and traits. It was also noted that TCF7L2 binds to many more genes implicated in the GWAS of metabolic and cardiovascular traits than would be expected by chance (5).

Following this notion, we aimed to similarly elucidate such patterns for FOXA2, using the algorithm hypergeometric optimization of motif enrichment (HOMER), to address our hypothesis that this transcription factor regulates the expression of a set of molecular pathways critical to disease pathogenesis. However, in this case we leveraged primary tissue, ie, liver, to elucidate a potentially cleaner signature.

Materials and Methods

Sequence alignment

We leveraged our previously generated FOXA2 ChIP-Seq and corresponding raw sequence files derived from the liver (7). After aligning the reads with Bowtie to hg19, binding sites were defined at a false discovery rate of 1%, cumulative Poisson value of P = .0001, and fold coverage threshold of 4 times the normalized sequence tags in the target experiment comparable with random background sequence tags using HOMER (8) analysis package. The candidate target gene was the closest gene to the transcriptional start site (TSS), regardless of direction from binding site. In all cases, the TSS of the aligned transcript was used as the anchor point for distance measurements.

Pathway analysis

Data were analyzed through the use of the gene set enrichment analysis (GSEA) algorithm (Broad Institute, Cambridge, Massachusetts, http://www.broadinstitute.org/gsea/index.jsp) specified for human tissue compendium (9, 10). The genes assigned with at least one function or pathway annotation in the GSEA knowledge base were eligible for the analysis. The P values associated with functions and pathways were calculated using hypergeometric distribution. Subsequent Ingenuity-based pathway analyses (Ingenuity Systems, www.ingenuity.com, Redwood City, California) were specified for human. Genes assigned at least one function or pathway annotation in the Ingenuity knowledge base were eligible for analysis.

GWAS data analysis

We based our analysis on all GWAS-implicated genes summarized in the National Human Genome Research Institute GWAS catalog (http://www.genome.gov/gwastudies) from February 19, 2013. The component diseases making up each of the disease categories are outlined in detail in Supplemental Table 1. Further analytical details are provided in the Supplemental Methods.

Results and Discussion

After overlaying a total of 20 603 127 sequence reads from two independent replicate experiments, a total of 1106 binding sites was observed at a false discovery rate of 1% using HOMER (8) (see Supplemental Figures 1 and 2 for genomic distribution and consensus site information).

Next, we performed a pathway analysis of the 967 genes colocalized with a FOXA2 occupancy site that could be adequately functionally annotated. Data were analyzed with GSEA specified for human tissue compendium. Eight hundred thirty-two of these genes had at least one functional or pathway annotation and were thus eligible for the analysis. Forty-six of these genes are known in the context of metabolism of lipids and lipoproteins, leading to this category being ranked the most significant annotation with a P = 9.74 × 10⁻³⁰. This category could conceivably have a knock-on effect with respect to the disease categories we observed overrepresented in our GWAS-related analyses (see below). Further details are provided in Table 1.

Table 1.

Top 50 FOXA2 GSEA Conical Pathways, Including Adjusted P Values for Multiple Category Testing

Conical Pathways	Genes in Gene Set (K), n	Genes in Overlap (k), n	k/K	P Value	FDR Q Value
Genes involved in metabolism of lipids and lipoproteins	478	46	0.0962	9.74E-30	1.29E-26
FOXA2 and FOXA3 transcription factor networks	45	14	0.3111	3.62E-14	2.39E-11
Genes involved in metabolism of amino acids and derivatives	200	24	0.12	3.55E-13	1.56E-10
Genes involved in SLC-mediated transmembrane transport	241	25	0.1037	3.10E-12	1.02E-09
Genes involved in fatty acid, triacylglycerol, and ketone body metabolism	168	21	0.125	4.76E-12	1.26E-09
Genes involved in transmembrane transport of small molecules	413	31	0.0751	3.98E-11	8.75E-09
Genes involved in biological oxidations	139	18	0.1295	8.74E-11	1.54E-08
Validated targets of C-MYC transcriptional repression	63	13	0.2063	9.31E-11	1.54E-08
FOXA1 transcription factor network	44	11	0.25	2.86E-10	4.11E-08
Complement and coagulation cascades	69	13	0.1884	3.11E-10	4.11E-08
Pathways in cancer	328	25	0.0762	2.23E-09	2.68E-07
Validated transcriptional targets of δNp63 isoforms	47	10	0.2128	1.01E-08	1.12E-06
HIF-1α transcription factor network	66	11	0.1667	2.79E-08	2.83E-06
Genes involved in hemostasis	466	28	0.0601	4.50E-08	4.24E-06
Genes involved in signaling by NOTCH	103	13	0.1262	4.90E-08	4.31E-06
Genes involved in neuronal system	279	21	0.0753	5.24E-08	4.32E-06
Genes involved in immune system	933	42	0.045	9.40E-08	7.30E-06
Validated nuclear estrogen receptor-α network	65	10	0.1538	2.61E-07	1.92E-05
Genes involved in cytochrome P450, arranged by substrate type	51	9	0.1765	3.08E-07	2.14E-05
Genes involved in triglyceride biosynthesis	38	8	0.2105	3.38E-07	2.23E-05
Genes involved in bile acid and bile salt metabolism	27	7	0.2593	4.03E-07	2.53E-05
Genes involved in PPARα activates gene expression	104	12	0.1154	4.32E-07	2.59E-05
PPAR signaling pathway	69	10	0.1449	4.65E-07	2.67E-05
Validated transcriptional targets of TAp63 isoforms	54	9	0.1667	5.14E-07	2.82E-05
Genes involved in phase 1, functionalization of compounds	70	10	0.1429	5.34E-07	2.82E-05
Steroid hormone biosynthesis	55	9	0.1636	6.04E-07	3.07E-05
ATM signaling pathway	20	6	0.3	1.08E-06	5.28E-05
Genes involved in transport of vitamins, nucleosides, and related molecules	31	7	0.2258	1.12E-06	5.28E-05
Genes involved in NOTCH1 intracellular domain regulates transcription	46	8	0.1739	1.59E-06	7.22E-05
IL-6-mediated signaling events	47	8	0.1702	1.88E-06	8.02E-05
Regulation of nuclear β-catenin signaling and target gene transcription	80	10	0.125	1.88E-06	8.02E-05
Genes involved in posttranslational protein modification	188	15	0.0798	2.00E-06	8.27E-05
Genes involved in asparagine N-linked glycosylation	81	10	0.1235	2.11E-06	8.46E-05
Regulation of nuclear SMAD2/3 signaling	82	10	0.122	2.37E-06	9.20E-05
Genes involved in metabolism of carbohydrates	247	17	0.0688	3.25E-06	1.23E-04
Genes involved in synthesis of very long-chain fatty acyl-CoAs	14	5	0.3571	3.36E-06	1.23E-04
Genes involved in metabolism of proteins	518	26	0.0502	3.89E-06	1.39E-04
Wnt signaling pathway	151	13	0.0861	4.14E-06	1.44E-04
Genes involved in developmental biology	396	22	0.0556	4.35E-06	1.47E-04
Genes involved in PI3K/AKT activation	38	7	0.1842	4.81E-06	1.47E-04
Genes involved in circadian clock	53	8	0.1509	4.82E-06	1.47E-04
Long-term depression	70	9	0.1286	4.85E-06	1.47E-04
Genes involved in signaling by NOTCH1	70	9	0.1286	4.85E-06	1.47E-04
Genes involved in response to elevated platelet cytosolic Ca2+	89	10	0.1124	5.03E-06	1.47E-04
Genes involved in transport of glucose and other sugars, bile salts, and organic acids	89	10	0.1124	5.03E-06	1.47E-04
Genes involved in diabetes pathways	133	12	0.0902	5.96E-06	1.71E-04
Genes involved in transcriptional regulation of white adipocyte differentiation	72	9	0.125	6.15E-06	1.73E-04
Genes involved in platelet activation, signaling, and aggregation	208	15	0.0721	6.93E-06	1.91E-04
Direct p53 effectors	137	12	0.0876	8.08E-06	2.18E-04
Genes involved in lipoprotein metabolism	28	6	0.2143	9.27E-06	2.45E-04

Abbreviations: HIF, hypoxia-inducible factor; PI3K, phosphatidylinositol 3-kinase; PPAR, peroxisomal proliferator-activated receptor; SLC, somatostatin-like receptor; SMAD, phosphorylated mothers against decapentaplegic. Metabolism of lipids and lipoproteins was ranked the most significant annotation. The next highest categories were FOXA2 and FOXA3 transcription factor networks, metabolism of amino acids and derivatives, SLC-mediated transmembrane transport, and fatty acid, triacylglycerol, and ketone body metabolism, all of which easily survived the correction for multiple testing.

We also observed a significant overrepresentation of genes with a FOXA2 occupancy site encoding transcription factors [12.7% (106 of 832); P = 1.2 × 10⁻²⁴]; indeed, of these transcription factors, 53.8% (57 of 106; P = 1.1 × 10⁻¹⁸) are known GWAS disease-implicated loci (Supplemental Table 2). As such, FOXA2 appears to act as a master regulator for more than a hundred other transcription factors, resonating with existing knowledge on FOXA2 behavior (11), and may themselves be good candidates for follow-up to gain insights in to disease-related pathways.

Motivated by the GWAS-implicated locus enrichment from our previous TCF7L2 work (5), we elected to query the results against all GWAS signals reported to date. Initially, we examined 154 genes with FOXA2 binding sites in close proximity (5 kb) to their TSS, which represents 0.8% of all reference sequence (RefSeq) genes (n = 19 015), and indeed, after the correction for multiple testing of categories, there was a significant overrepresentation of GWAS-implicated cardiovascular loci (P = .001), whereas there was nominal evidence enrichment of GWAS-implicated endocrine loci (P = .012). Conversely at this distance, GWAS loci for cancer, inflammation, and neuropsychiatric categories showed no significant evidence for enrichment; in addition, the overall combination of GWAS-implicated loci showed no evidence of enrichment (Table 2).

Table 2.

Enrichment of GWAS Signals for Nearest RefSeq Genes to a FOXA2 Binding Site

	HSA Liver FOXA2 5-kb Genes: 154 Genes				HSA Liver FOXA2 50-kb Genes: 592 Genes
	Total hg19 Gene List, %	ChipSeq Gene List, %	P Values: χ2		Total hg19 Gene List, %	Chip-Seq Gene List, %	P Values: χ2
Endocrine	0.8% (154/19 015)	1.5% (13.7/888)	.021	Endocrine	3.1% (592/19 015)	4.9% (43.86/888)	.004
(T2D)	0.8% (154/19 015)	1.2% (1/82)	.683	(T2D)	3.1% (592/19 015)	4.7% (3.8/82)	.436
(Glycemic Traits)	0.8% (154/19 015)	3.6% (2/56)	.025	(Glycemic Traits)	3.1% (592/19 015)	10.7% (6/56)	.002
Cancer	0.8% (154/19 015)	1.3% (4.25/335)	.36	Cancer	3.1% (592/19 015)	5.3% (17.83/335)	.028
Cardiovascular	0.8% (154/19 015)	2.2% (10.28/463)	.001	Cardiovascular	3.1% (592/19 015)	6.9% (31.74/463)	1.69 × 10−5
(Lipids)	0.8% (154/19 015)	6.4% (8.25/127)	1.77 × 10−11	(Lipids)	3.1% (592/19 015)	11.3% (14.4/127)	8.63 × 10−7
Inflammation	0.8% (154/19 015)	1.3% (7/521)	.173	Inflammation	3.1% (592/19 015)	5.1% (26.41/521)	.016
Neuropsychiatric	0.8% (154/19 015)	1.1% (6.33/582)	.468	Neuropsychiatric	3.1% (592/19 015)	3.1% (18.16/582)	.993
All	0.8% (154/19 015)	1.0% (35.7/3607)	.284	All	3.1% (592/19 015)	3.9% (140.33/3607)	.02
HSA liver FOXA2 all genes: 967 genes
Endocrine	5.1% (967/19 015)	9.8% (87.27/888)	9.90 × 10−9
(T2D)	5.1% (967/19 015)	7.7% (6.3/82)	.309
(Glycemic Traits)	5.1% (967/19 015)	16.1% (9/56)	7.55 × 10−4
Cancer	5.1% (967/19 015)	10.4% (34.8/335)	5.67 × 10−5
Cardiovascular	5.1% (967/19 015)	10.0% (46.26/463)	1.27 × 10−5
(Lipids)	5.1% (967/19 015)	15.2% (19.4/127)	2.63 × 10−6
Inflammation	5.1% (967/19 015)	7.7% (40.36/521)	.011
Neuropsychiatric	5.1% (967/19 015)	8.3% (48.33/582)	.001
All	5.1% (967/19 015)	7.7% (277.04/3607)	4.01 × 10−9

We based our analysis on all GWAS genes summarized in the National Human Genome Research Institute GWAS catalog (http://www.genome.gov/gwastudies) from February 19, 2013. Enrichment was investigated using χ² analyses. Our method of scoring GWAS ChIP-Seq gene overlap was to assign 1 point to a GWAS region in which all the genes in the region were found in our list and a fraction of a point determined by how many genes were found in our gene list divided by the total genes in the GWAS region. This analysis model would equally weight a GWAS region with one gene the same as a region with eight genes as a single region. Specific diseases are shown in parentheses to indicate they are a subcategory of the relevant main general disease category, ie, T2D, glycemic traits, and lipids.

Of the 592 genes with FOXA2 binding sites within the greater distance of 50 kb, which represents 3.1% of all RefSeq genes, there was again significant overrepresentation of the GWAS-implicated cardiovascular (P = 1.69 × 10⁻⁵) and endocrine (P = .004) categories, with both surviving the correction for multiple testing of the categories. Although cancer (P = .028), inflammation (P = .016), and overall (P = .020) categories yielded evidence of an enrichment, these observations did not survive the correction for the number of categories tested (Table 2).

Finally, moving out to all binding sites, there were 967 genes (with only 37 being beyond 500 kb), representing 5.1% of all RefSeq genes. There was a highly significant overrepresentation of overall GWAS loci (P = 4.01 × 10⁻⁹), with most of the individual disease categories revealing highly significant enrichment, namely endocrine (P = 9.90 × 10⁻⁹), cardiovascular (P = 1.27 × 10⁻⁵), neuropsychiatric (P = .001), and cancer (P = 5.67 × 10⁻⁵). The exception was inflammation, which showed only evidence of nominally significant enrichment but did not survive correction (Table 2).

Although there are challenges to breaking out into specific diseases due to the statistical power challenges, we did investigate some key traits with sufficient loci reported, ie, T2D, lipids, and glycemic traits (Table 2). We did observe substantial enrichment for lipids and glycemic traits, but this was not the case for T2D, which may resonate with the fact that rs6048205 has not been associated with this latter trait to date.

The inflammation observation is particularly surprising because FOXA2 has been shown to play a critical role in regulating genetic pathways influencing Th2 cell-mediated pulmonary inflammation through the expression of genes regulating Th2 cell-mediated inflammation and goblet cell differentiation (12). Also, FOXA2 has been implicated in the mechanism of TNFα-mediated tumor development through the phosphorylation of FOXA2, thereby suppressing FOXA2 transactivation activity and leading to decreased NUMB expression, which activates the downstream NOTCH pathway, promoting cell proliferation and tumorigenesis (13). As such, we can hypothesize that these genes near FOXA2 occupancy sites are acting through several not-yet-understood noninflammatory pathways associated with these related traits.

The use of these data resonates with previous studies that suggest an enrichment of GWAS-identified single-nucleotide polymorphisms (SNPs) in transcriptionally relevant genomic locations (14, 15). These findings should also be contrasted with the occupancy recently reported for vitamin D receptor, in which the investigators found relatively restricted enrichment, ie, near inflammatory and cancer GWAS categories only (16), whereas FOXA2 binding suggests more widespread effects.

Although much is known about the biological functions of FOXA2, this is the first time it has been reported to have a strong affinity for specific categories of loci uncovered by GWAS and brings together variants that are largely additive in effect.

With the observations made above and to better understand the GSEA output, we used Ingenuity to visualize the connections between the genes making up the top three GSEA pathways. This led to general agreement between the two algorithms (see Supplemental Figure 3 for further details).

Next, we wanted to investigate FOXA2 occupancy relative to both the FOXA2 gene itself and the nearby SNP (rs6048205) that implicated this locus in a GWAS of fasting glucose (3). Our analysis identified two FOXA2 occupancy sites in the intergenic sequence 73.5 and 2.2 kb proximal to the FOXA2 TSS. Interestingly, the occupancy site 2.2 kb from the FOXA2 TSS directly overlaps rs6048205; indeed, independent FOXA2 ChIP-Seq generated by ENCODE in the carcinoma HEPG2 cell line also yielded the same finding (17). We also checked whether any SNPs in linkage disequilibrium to rs6048205 overlapped a FOXA2 occupancy site; however, despite checking a total of 141 SNPs with an r² > 0.5 to rs6048205 (note: no known SNP was in perfect linkage disequilibrium), no other variants coincided with either of these FOXA2 occupancy sites. Our data therefore suggest that rs6048205 perturbs an autoregulation mechanism; indeed, despite such a process having been largely ruled out for FOXA1 (18), there is substantial evidence that FOXA2 expression is at least partially controlled in this manner (19).

We also noted that a second GWAS-implicated SNP from a genetic study of plasma levels of liver enzymes (20), rs11597390, coincided with a FOXA2 occupancy site located 2.2 kb from the TSS of CPN1; again, this agreed with the carcinoma observations from ENCODE and thus also warrants further investigation.

We then elected to investigate whether any of the FOXA2 occupancy sites coincided with a subgenome-wide significant SNP (P > 5 × 10⁻⁶ but < 5 × 10⁻⁸) associated with relevant traits to implicate novel loci. We took all SNPs and their proxies (r² ≥ 0.7) that yielded such P values in either the Diabetes Genetics Replication and Meta-Analysis T2D study or the Meta-Analyses of Glucose and Insulin-Related Traits Consortium study fasting glucose GWAS data sets and cross-referenced them with all the binding sites we characterized; however, no positive findings were made.

Overall, one obvious strength of this study was that hepatic tissue was used in this approach. If one were to analyze FOXA2 ChIP-Seq data derived from other tissues, one would see either discrete patterns or robust consistency; as such, more efforts will need to be made in the future to ascertain this aspect of FOXA2 behavior.

In conclusion, we have gained insight into FOXA2 occupancy by leveraging the primary tissue of human liver. Our analyses suggest that FOXA2 acts as a master regulator by transcriptionally controlling multiple other transcription factors. There is also evidence that FOXA2 binds to the key SNP within the FOXA2 gene itself and may point to an influence on autoregulation that in turn may influence basal glucose levels. It also has an affinity to bind to genes found by GWAS but is constrained to the loci uncovered in the fields of endocrine, neuropsychiatric, cardiovascular, and cancer related disease whereas at the same time pointing to the possibility that it does not confer its effect via inflammatory processes.