The goal of this study was to understand the integrated transcriptional function of Hnf1α and Hnf4α in pancreatic islets. To study Hnf4α function, we generated pancreas-specific Hnf4a knock-out mice (Hnf4a^pKO), and confirmed that this targeted deletion caused a marked reduction of Hnf4a gene mRNA in islet-cells (Figure 1A).

In keeping with previous studies of mice with β-cell specific ablation of Hnf4a (Hnf4a^betaKO) [28]–[30], Hnf4a^pKO mice developed a mild complex phenotype, with very subtle glucose intolerance and a slightly reduced fasting glycemia (Figure S1).

Despite this mild metabolic phenotype, Hnf4a^pKO mice showed a clear islet transcriptional phenotype (Table S1). Downregulated genes encoded for varied cellular roles, including the metabolism of steroids, glucose, and amino acids (Table S1 and Table S2). Others encoded for regulators of signal transduction and cell growth, consistent with a previous report in Hnf4a^betaKO islets [31], or were in keeping with the proposed role of Hnf4α in epithelial differentiation [32] (Table S1 and Table S2). Upregulated genes included genes known to form part of the epithelial mesenchymal transition process (Table S2). Overall, the functional classes that were perturbed in Hnf4a-deficient islets were remarkably similar to those reported in Hnf1a^-/- islets[11].

To study Hnf1α transcriptional function, we used Hnf1a^+/− mice. As opposed to mice with homozygous Hnf1a mutations, which develop diabetes, this model has no documented in vitro or in vivo metabolic disturbances (Figure S2)[33]–[36]. Also in keeping with previous studies, Hnf1a^+/− islets exhibit only marginal downregulation of Hnf4α (∼70–90% of normal values) (Figure 1B)[36]. Hnf1a^+/− mice thus lack two elements that are thought to exert an indirect impact on islet gene expression in homozygous Hnf1a mutant islets. Array analysis revealed a transcriptional phenotype in Hnf1a^+/− islets, with 196 non-redundant genes downregulated >1.5-fold in Hnf1a^+/− islets at a nominal P value<0.01. We validated this dataset with gene-specific assays in 20 genes from independent Hnf1a^+/− mice (Figure S4A). Furthermore, genes that were bound by Hnf1α and downregulated in homozygous Hnf1a mutant islets were significantly downregulated in Hnf1a^+/− islets (Figure S4B). Thus, expression profiling in Hnf1a^+/− islets provides a tool to assess the transcriptional function of Hnf1α in this tissue.

We next compared expression changes in Hnf1a^+/− and Hnf4a^pKO islets. This revealed a striking correlation between the two models (r = 0.57, P = 10⁻⁶) (Figure 2A–2C). Gene set enrichment analysis (GSEA) showed that genes that were significantly downregulated in Hnf4a^pKO islets were downregulated in Hnf1a^+/− islets (Figure 2D, P<0.001). Conversely, genes downregulated in Hnf1a^+/− islets were downregulated in Hnf4a^pKO islets (Figure 2E, P<0.001). Not surprisingly, gene expression changes were consistently lower in the Hnf1a haploinsufficient islets compared with islets with biallelic inactivation of Hnf4a (Figure 2A–2E). We confirmed the correlation with gene-specific assays (Figure 2B), and with an independent comparison of Hnf1a^+/− versus Hnf4a^pKO mice of 16 rather than 8 weeks of age (not shown, and Table S1).

These common gene expression changes were unexpected, because Hnf1a expression in Hnf4a^pKO islets was unperturbed (Figure 1A) (as previously shown for Hnf4a^betaKO mice[28]–[30]) and Hnf4a expression was only marginally reduced in Hnf1a^+/− islets (Figure 1B, and [36]). In conclusion, the analysis of models that minimize the impact of indirect perturbations showed that Hnf1α and Hnf4α regulate a common set of genes in pancreatic islets.

Previous studies provide two possible mechanisms whereby Hnf1a^+/− versus Hnf4a^pKO islets could exhibit a similar transcriptional phenotype (Figure 3A). One is that Hnf1α and Hnf4α frequently bind the same genes in human liver and islets [6]. We confirmed this finding using mouse liver binding datasets reported elsewhere [8], [11] (Figure 3B), after estimating that ∼75% of Hnf4α-bound genes in islets may also bound in liver (Figure. S8). However, co-occupancy does not per se explain the similar gene expression changes in the two mutant models, because for most genes bound by Hnf1α or Hnf4α, gene expression is not altered in the respective knock-out tissues [8], [11]. An alternate explanation for the similar transcriptional phenotypes is that Hnf1α regulates Hnf4a gene transcription in islets (Figure 3A) [11], [19]–[21]. In the current study we used heterozygous Hnf1a mutant islets because Hnf4a was not significantly altered (unlike homozygous Hnf1a mutants). However, it remained possible that a subtle decrease in Hnf4a expression in Hnf1a^+/− islets caused the common transcriptional phenotype (which would thus result from a perturbation of Hnf4α in both models). We predicted that if this were true we should observe impaired expression of direct targets of Hnf4α in the two models, whereas the direct targets of the upstream factor of this hierarchy, Hnf1α, should be impaired only in Hnf1a-deficient islets (Figure 3A).

We thus tested if transcription of Hnf1α-bound genes was impaired in mice deficient for either factor. Hnf1α binding frequency was increased 5-fold among genes that were downregulated in homozygous Hnf1a mutant islets (Figure 3C). Remarkably, Hnf1α binding frequency was also increased 5-fold in promoters of genes that were downregulated in Hnf4a-deficient islets (Figure 3C). We also found that Hnf1α bound genes that were downregulated in Hnf1a^-/- islets were downregulated in Hnf4a-deficient islets (Figure S5). In silico studies confirmed these findings, as the most overrepresented conserved motif in genes downregulated in Hnf4a^pKO islets was identical to the canonical HNF1 binding sequence (Figure 3D). Thus, Hnf1α targets are impaired in Hnf4a deficient islets to a similar extent as in Hnf1a deficient islets.

Genes that were downregulated in both Hnf4a- and Hnf1a-deficient islets also showed significantly enriched Hnf4α binding and a 4 to 6-fold higher co-occupancy rate than non-regulated genes (Figure 3C). Collectively, these results argue that the shared transcriptome in our models is not due to the known Hnf1a-Hnf4a transcriptional hierarchy, and instead support that it is linked to the regulation of common target genes.

The above findings were consistent with an interdependent function of Hnf1α and Hnf4α, or alternatively with the regulation of common genes in a mechanistically independent manner. To discriminate among these possibilities, we compared expression profiles in single mutant (Hnf1a^+/− and Hnf4a^pKO) versus double mutant (Hnf1a^+/− Hnf4a^pKO) islets (Figure 4A). We predicted that if Hnf1α and Hnf4α act independently in any gene that is downregulated in both single mutant islets, the expression ratio in double mutant (Hnf1a^+/− Hnf4a^pKO) islets should reflect the product of the two single mutant expression ratios. By contrast, if the two factors regulate a gene in an interdependent manner, the expression in double mutant islets should differ from this expectation (Figure 4A).

For each gene that was downregulated in both Hnf1a^+/− and Hnf4a^pKO single mutant islets we calculated an epistasis ε value that measures the deviation from expectation. An ε value > 0 indicates that the expression ratio in Hnf1a^+/− Hnf4a^pKO islets is higher (less perturbed) than expected from the independent effects of the two single mutant values. As shown in Figure 4C, the distribution of ε values was unambiguously greater than 0. Figure 4D further illustrates this concept, showing that the perturbation of individual genes in Hnf1a^+/− Hnf4a^pKO islets was epistatic (P = 10⁻³⁵). Similar findings were confirmed in independent mice using qPCR (quantitative PCR) rather than oligonucleotide chips (Figure S6).

The high co-occupancy rate of Hnf1α and Hnf4α also raises the question whether these two factors might exert redundant functions at some targets. Because selecting genes that are downregulated in the single mutant islets can represent a bias against redundancy, we also performed this analysis in all genes that were downregulated >3 fold in double Hnf1a^+/− Hnf4a^pKO mutant islets. Most such genes were markedly downregulated in the single mutants, and again had ε values exceeding 0 (Figure S7), thus showing that although Hnf1α and Hnf4α often bind to the same genes, their function is not redundant in islets. In keeping with these findings, the mild glucose intolerance phenotype observed in Hnf4a^pKO mice was not further impaired in Hnf1a^+/− Hnf4a^pKO mice (Figure S3). In summary, these results indicate that Hnf1α and Hnf4α bind to similar targets and act through interdependent regulatory mechanisms in islets, thus leading to a common transcriptional phenotype in Hnf1a^+/− and Hnf4a^pKO islets.

To assess the mechanisms underlying the interdependent function of Hnf1α and Hnf4α in common target genes, we examined whether their binding is interdependent. Because Hnf1α expression is unaltered in Hnf4a-deficient mouse islets, we were able to use this model to study whether Hnf4α is required for Hnf1α binding to common targets. We selected eight genes that we have previously shown are bound by Hnf1α and are functionally dependent on Hnf1α in islets [11](Figure 5A). All of them were also directly bound by Hnf4α in wild type islets (Figure 5C), and were markedly downregulated in Hnf4a-deficient islets (Figure 5B). Thus, all 8 selected genes were co-occupied and were functionally dependent on the two factors in islets. We next examined Hnf1α binding to these sites in Hnf4a-deficient islets. We found that in 5 of 8 genes, Hnf1α binding was maintained in islets that lack Hnf4α (Figure 5D). In two other Hnf4a-dependent genes, Hnf1α binding was significantly reduced in Hnf4a-deficient islets, and was completely abrogated in one case (Figure 5D). The difference between genes concerning Hnf1α binding in Hnf4a-deficient islets could not be linked to differences in the affinity of Hnf1α and Hnf4α binding sites. Thus, the synergistic activation of common targets by Hnf1α and Hnf4α can reflect gene-specific interdependent mechanisms at both binding and post-binding levels.

We have addressed how transcription factors establish functional interactions in an in vivo context. To achieve this, we studied epistatic relationships of transcriptional phenotypes. Our approach follows recent studies that used epistasis of transcriptomes to study functional interactions between regulators of protein kinase A in Dictyostelium, Mediator subunits in yeast, and most recently to unravel yeast transcription factor networks [37]–[39]. Our use of epistasis is analogous to classic studies that studied synergism or redundancy between transcription factors by comparing cells transfected with reporter minigenes along with single versus multiple transcription factors[1]. In the reverse approach we employed, we studied the transcriptome of mice with single and double transcription factor mutations. This allowed us to study combinatorial function in vivo, in endogenous genes of primary mammalian cells. It also enabled a global analysis, rather than studying specific gene targets that do not necessarily reflect a predominant regulatory strategy. This approach complements studies that compare the genomic location of different transcription factors without assessing their functional interactions. Our analysis thus confirmed previous observations that Hnf1α and Hnf4α bind to common targets, and suggests that these two factors function as obligate interdependent regulators in pancreatic islets. We thus demonstrate a role for epistasis to unravel the function of transcription factor networks in mammalian cells.

Previous studies showed that homozygous Hnf1a mutant mice exhibit full blown diabetes, in contrast to β-cell and pancreas-specific Hnf4a mutations which only result in glucose intolerance [21], [28], [29], [31], [34], [40]. In light of these differences it was somewhat unexpected that Hnf1α and Hnf4α regulate similar islet genes. We observed this coregulation using heterozygous Hnf1a and pancreas-specific Hnf4a mutations, which do not have common metabolic disturbances that could confound the comparison. They are also selective models: four studies have now shown that Hnf4a-deficient islets express normal Hnf1α levels (this study, and [28]–[30]), and, in contrast to Hnf1a^-/- mice, Hnf4α levels are only minimally altered in Hnf1a^+/− islets (this study, and [36]). Plausibly, differences in the phenotype of Hnf1a and Hnf4a-deficient models reported so far are due to the use of different types of genetic inactivation systems. More generally, we believe that weak genetic perturbations can be of great interest in studying transcription factor function, because although they only cause mild target expression changes, they are also less likely to disrupt downstream regulatory networks, thus limiting the magnitude of indirect effects.

Most genes bound by Hnf4α or Hnf1α are not affected by mutations of these two factors[11], [19]. It was thus necessary to integrate binding studies with genetic perturbation models to understand the functional interactions between these transcriptional regulators.

We observed that functional Hnf1α targets were similarly perturbed in Hnf4a-deficient and Hnf1a-deficient islets. This suggests that Hnf4α regulates Hnf1α function, and discards that epistasis was simply due to the known transcriptional hierarchy in which Hnf1a is upstream of Hnf4a. Together with the high Hnf1α/Hnf4α co-occupancy rate reported here and previously in human liver and islets[6], these findings suggest that epistasis between Hnf1a and Hnf4a is at least in part due to interdependent interactions at common direct targets. We thus propose a model for the integrated function of Hnf1α and Hnf4α in islets whereby Hnf1α regulates Hnf4a transcription [19], [20], and furthermore both factors act as interdependent transcriptional partners in islet-cell targets (Figure 6). Importantly, this network model is not based solely on binding studies, but integrates information on the combinatorial functional interactions between these factors.

Several reports have demonstrated protein-protein interactions between Hnf1α and Hnf4α [22]–[24]. Such interactions can lead to functional inhibition [23], [24]. We show that although inhibitory consequences may be prevalent in other tissues or may occur in a small subset of genes, Hnf1α and Hnf4α largely activate genes synergistically in pancreatic islets. Earlier gene-specific studies have shown cooperative binding of Hnf1α and Hnf4α [2], [26], [27], while Hnf4α has been shown to co-activate a gene that is only directly bound by Hnf1α[22]. Our in vivo data showed that binding was interdependent in a subset of targets, but also showed that in many targets Hnf4a-deficiency does not entail decreased binding of Hnf1α. In the latter genes Hnf4α is likely required for post-binding functions of Hnf1α, such as the recruitment of co-regulatory complexes required for chromatin remodelling and/or assembly of the preinitiation complex. The mechanisms underlying functional synergism between Hnf1α and Hnf4α therefore appear to vary across target genes.

Our study predicts a common islet transcriptome defect in the pathophysiology of HNF1A and HNF4A diabetes. This is consistent with the clinically indistinguishable diabetic phenotype of adult HNF1A and HNF4A patients[18], [41]. An exception to this notion is that HNF4A mutations cause transient in utero and neonatal hyperinsulinism, which later evolves to decreased insulin secretion, whereas HNF1A mutations develop the latter phenotype without early hyperinsulinism [30]. This may result if Hnf4α has Hnf1α-independent functions during prenatal and neonatal developmental stages.

The interdependent function of Hnf1α and Hnf4α is also relevant to our understanding of how haploinsufficiency of HNF1A and HNF4A leads to β-cell dysfunction and diabetes. We previously proposed that HNF1A or HNF4A haploinsufficiency could lead to the disruption of a Hnf1α/Hnf4α positive cross-regulatory network in β-cells[42]. Our current model depicted in Figure 6 provides new elements to assess the consequences of haploinsufficiency for the complex Hnf1α/Hnf4α network. Synergistic Hnf1α/Hnf4α-dependent activation is expected to result in steeper activator-response curves, and thus greater vulnerability to decreased gene dosage. This may be more pronounced in islet-cells, where Hnf1α and Hnf4α concentrations are much lower than in liver and other tissues that are not clinically afflicted in MODY[11], [43]. Haploinsufficiency of either HNF1A or HNF4A may in this manner disrupt the function of both HNF1α and HNF4α and the common transcriptional program, which include essential genes for the proper function of pancreatic islets. In conclusion, these studies provide an approach to understand the in vivo function of a regulatory network, and increase our understanding of the mechanisms underlying monogenic diabetes.

Hnf4a^LoxP mice were obtained from The Jackson Laboratory[44], Hnf1a^+/− mice were provided by Frank Gonzalez (NCI)[34]. Pdx1^Cre transgenic mice were provided by Pedro Herrera (U. Geneva) [45]. All studies were performed according to procedures approved by the institutional animal care and use committee. Animals were maintained on C57B/l6 backgrounds and genotyped as described [40]. For Hnf4a studies, Hnf4a^LoxP littermates lacking Pdx1^Cre and Pdx1^Cre mice lacking Hnf4a^LoxP alleles were used as controls unless stated otherwise. Pancreatic islets from 2- to 4-month old mice were isolated as described [40]. Islets were cultured for 48 hr at 37°C, 5% CO2 in RPMI (Invitrogen) containing 11mM glucose supplemented with 10% FCS.

Animals were fasted overnight and injected glucose intraperitoneally (2 gm/Kg). Glucose was measured from the tail vein at 0,15,30,60 and 120 min. Fasted plasma insulin was measured by ELISA (Mercodia).

RNA from purified islets was isolated with Trizol (Invitrogen) and tested with an Agilent 2100 Bioanalyzer to ascertain RNA integrity. The reduction of Hnf4a mRNA in Hnf4a^pKO mice varied between 5–50% of wild type islets, most likely due the inherent variability of Cre-based recombination. We thus assessed Hnf4a mRNA by semiquantitative PCR [46] and used samples with >80% reduction for further analysis.

For each array replicate, RNAs from 2–4 male mice were pooled, and 50 ng was used in two cycles of cDNA synthesis for hybridization of Affymetrix 430 2.0 arrays. For epistasis experiments we used 8 week-old male mice. We separately compared 16 week-old male Hnf4a^LoxP and control mice. Three arrays (a total of 8–12 mice) were analyzed per genotype, normalized with RMA, and analyzed with the LIMMA package to identify downregulated genes using a multiple test adjusted P value <0.05. To select genes downregulated in Hnf1a^+/− islets we used a nominal (unadjusted) P<0.01 threshold, and validated this set with gene-specific assays. Gene-specific expression was assessed either with Taqman Low Density Arrays (Applied Biosystems) using unpooled islet RNA samples from 2–3 additional mice per genotype, or by qPCR using SybrGreen detection system as described [11]. Gene expression datasets are available in ArrayExpress (http://www.ebi.ac.uk) (Accession number: E-MEXP-1729).

Approximately 2000 purified islets from Hnf4a^LoxP Pdx1^Cre and littermates lacking the Pdx1^Cre transgene were used for ChIP assays essentially as described [11], [19]. For each genotype we processed islets from two independent pools of ∼10 mice separately, we measured in duplicate the enrichment of immunoprecipitated DNA relative to input DNA, and corrected for the same values obtained in Actb as a negative control gene.

We used oPOSSUM [47], which computes a Fisher exact test to measure over-representation of sequence elements in a gene set relative to a background comprising all genes. We analyzed evolutionary conserved sequences 5 Kb upstream and 5 Kb downstream of transcription start sites of all downregulated genes (M<−0.6), and searched for conserved sequence elements as described in [48]. The empirical recommendations to identify binding sites oPOSSUM are a Z-score>10 and a Fisher P value<0.01. Overrepresented motifs were tested against the JASPAR CORE database of binding site profiles.

Our analysis of epistasis of transcriptome phenotypes is based on previous large-scale studies of epistasis among yeast mutants regulating cell growth [49], [50]. We selected 105 genes that were downregulated (M<−0.6) in both Hnf1a^+/− and Hnf4a^pKO islets, and compared gene expression changes with that of double mutant islets. For each gene we calculated an ε epistasis value that measures the deviation of the observed mutant/wild type expression ratio (R) in Hnf1a^+/− Hnf4a^pKO double mutant islets from the expected ratio based on the product of the two single mutant values (ε = R Hnf1a^+/− Hnf4a^pKO - [R Hnf1a^+/− x R Hnf4a^pKO]). We performed a similar analysis for a control set of genes that showed no regulation in the two single mutant islets (M-1.1 to 1.1, p>0.2,). We calculated statistical significance with Student's t test, comparing the experimental and control ε value distributions, or with a paired Student's t test, comparing Hnf1a^+/− Hnf4a^pKO R values to expected [R Hnf1a^+/− x R Hnf4a^pKO] values in each gene.

Enrichment of functional annotations was examined with GSEAv2.0 (http://www.broad.mit.edu/gsea/) using gene sets as the permutation type and 1000 permutations, and with DAVID (http://david.abcc.ncifcrf.gov/). Statistical significance in binding comparisons was calculated with two-sided Fisher's exact test, or by testing the Hypergeometric distribution.