Introduction

Transcription factors control the ability of stem cells to maintain their undifferentiated state [1, 2]. They also control the developmental process that leads to the formation of beta cells [3], and provide the genetic program for adult beta cells to function properly [4, 5, 6, 7]. It was originally thought that single dedicated transcriptional activators might be responsible for directing the specific phenotype of each cell type. The accumulated evidence points instead to a combinatorial mode of action, whereby unique sets of transcription factors determine the identity and function of a cell. For a set of transcription factors to be expressed in a specific cell, mechanisms need to be in place to initiate and maintain the transcription of their genes. This is achieved through networks formed by transcription factors that directly regulate the promoters of genes encoding other transcription factors [8, 9]. Thus, each cell type possesses not only a unique combination of transcription factors, but also a specific gene regulatory network structure.

With the advent of the sequences of entire genomes of key model organisms and the technologies to analyse them, we can realistically expect to obtain most of the relevant experimental information pertaining to pancreatic transcription networks. Deciphering the logic of such networks is highly relevant to the pathophysiology and treatment of diabetes. Knowledge of pancreatic gene programs could enable in vitro recapitulation of embryonic development to produce functioning insulin-producing cells. It could also enable in vivo stimulation of beta cell regeneration for the treatment of Type 1 diabetes [10, 11, 12]. In addition it could help us to understand the pathophysiology of monogenic forms of diabetes caused by mutations in genes encoding transcription factors [7, 13, 14]. An in-depth understanding of beta cell transcription programs could also help in the development of drugs to treat insulin secretory dysfunction in Type 2 diabetes.

In this article we discuss the process of pancreatic development and beta cell differentiation from the perspective of its transcriptional control. Particular focus is placed on emerging evidence revealing the in vivo structure of transcription networks within distinct cells of the developing and adult pancreas, and on providing insights into how such knowledge can affect our understanding of pancreatic gene regulation.

Interpreting experimental data on transcription networks

The information currently available on pancreatic gene networks is still very incomplete. As this information increases, so will its apparent complexity. To interpret the functional logic of such large networks, it will ultimately be crucial not only to obtain sufficient experimental data, but also to define the significance and level of certainty for each of the network components.

A major source of information on transcription networks is studies using mice with targeted inactivation of a transcription factor gene. Using either microarrays or more focused analysis of candidates, such knock-out models can identify which genes are activated or repressed by a transcription factor of interest in specific tissues and developmental stages. Yet these studies do not distinguish between direct and indirect effects. The inactivation of a transcription factor in an animal can modify the function of a cell and of the entire organism, and this can exert profound changes in the expression of genes that are not directly bound by the transcription regulator. Thus, to understand the structure of a gene network, we need to complement differential expression data with biochemical evidence indicating which effects reflect direct regulation.

Traditionally the analysis of direct interactions between transcription factors and promoters has been carried out through in vitro binding assays and transient transfection studies in cultured cell lines. Such studies can be invaluable for detailed mechanistic analysis of promoters. However, they do not prove that an interaction occurs at physiological concentrations of the factor in the native chromatin environment of the target gene. Moreover, they cannot address in which cells the interaction occurs in vivo. In this respect, chromatin immunoprecipitation (ChIP) has emerged as a method for demonstrating that a transcription factor binds to a gene segment in a specific cell type (Fig. 1) [15]. Other useful in vivo tools are transgenics using wild-type and mutated regulatory DNA segments linked to a reporter gene, or targeted mutations of cis elements to determine their role in the transcription of a gene at distinct developmental stages [16, 17, 18].

Fig. 1
figure 1

Analysis of transcription factor promoter interactions using the chromatin immunoprecipitation (ChIP) assay. The ChIP assay can show that a native transcription factor binds to an endogenous gene in a specific cell type [15]. Living cells are exposed to formaldehyde to crosslink proteins bound to DNA. The cells are lysed, chromatin is sheared by sonication, and an antibody against a DNA-binding protein is used to immunoprecipitate the chromatin fragments to which the protein is bound. After DNA purification, specific DNA segments are assayed for enrichment in the immunoprecipitate. This is achieved by PCR amplification of the candidate target gene segment and a negative control DNA, which is not expected to be enriched (A and B respectively). The amplification pattern is compared with that obtained in unenriched (“input”) genomic DNA [167]. Alternatively, ChIP-enriched DNA can be hybridised to microarrays containing a large number of promoter or intergenic fragments, a technique known as ChIP-on-chip [168, 169]. Afor, A forward; Arev, A reverse oligonucleotide primer

A key issue in ChIP and expression profiling experiments is that relationships between transcription factors and target genes are highly cell-specific, and thus need to be demonstrated in specified cell types or developmental stages [19, 20, 21]. Throughout this review, therefore, the emphasis is on direct regulatory interactions that have been shown to occur during defined cellular stages of the developing pancreas.

Transcriptional networks in very early pancreatic bud multipotent cells

In the mouse embryo, the first visible signs of pancreatic development appear at approximately embryonic day (E) 9 (Fig. 2) [3, 22]. Two pancreatic buds emerge independently from the ventral and dorsal regions of the foregut endoderm. The ventral bud is formed from cells that escape the inhibitory effect of fibroblast growth factor and bone morphogenetic protein signalling from the cardiac and septum transversum mesoderm [23, 24]. In contrast, signals from the notochord promote dorsal pancreas development through a process involving the repression of Sonic hedgehog expression in the dorsal endoderm [25, 26]. It is likely therefore that important aspects of the transcriptional programs orchestrating the specification and development of ventral and dorsal buds are distinct.

Fig. 2
figure 2

Schematic representation of three stages of pancreatic development in the mouse embryo. At E9-E10 (left) dorsal and then ventral pancreatic buds arise from foregut endoderm. At this stage most epithelial cells are multipotent precursors expressing markers such as Pdx1 and Nkx2.2 (see also Fig. 3). The representative immunofluorescence image (E10.5) shows dorsal (dp) and ventral (vp) buds marked by anti-Nkx2.2 (green), while intense reactivity for anti-Hnf1α (red) identifies primitive liver (li). At E14-E17 (middle), the pre-differentiated epithelium has formed a well-defined ductal network distinguished from acinar and endocrine cells by the expression of Hnf1β. Numerous Ngn3+ cells arise from the Hnf1β+ ductal epithelium at this time, later giving rise to endocrine cells. The immunostaining image (E14.5) depicts ducts formed by Hnf1β+ cells (red), Ngn3+ cells (blue) and endocrine cells expressing exclusively Nkx2.2 (green) (some cells co-express these markers, see main text). On the right is a schematic representation of the mature pancreas. Immunofluorescent stainings were provided by M. A. Maestro

Fate-mapping and immunolocalisation studies have shown that most epithelial cells of the early buds are pre-differentiated precursors of exocrine and endocrine cells [27, 28, 29]. These cells express a combination of transcription factors, including Pdx1, Ptf1a (p48), Nkx6.1, Nkx2.2, Hb9, Hex, hepatocyte nuclear factor (Hnf) 6 and Foxa2 (Hnf3β) [29, 30, 31, 32, 33, 34, 35].

Knock-out studies show that many of these factors play essential roles in early buds (Fig. 3). Thus Pdx1 and Ptf1a are indispensable for the formation or progression of both buds [29, 36, 37, 38]. On the other hand, Hb9 and Isl1 are only required for the specification or differentiation of the dorsal pancreatic bud, while Hex is selectively needed for specification of the ventral counterpart, which is consistent with the different mechanisms underlying the formation of the two buds [33, 35, 39, 40]. Interestingly, at this stage Isl1 is predominantly present in mesenchymal cells, and its mutant phenotype is due to lack of dorsal pancreas mesenchyme [39]. Hex contributes to the growth of ventral foregut endoderm so that part of it is at a site that is permissive for pancreatic organogenesis [35]. The findings for Isl1 and Hex show that the critical roles of transcription factors in pancreatic development can be mediated by completely indirect mechanisms that do not involve cell-autonomous activation of pro-differentiation genes in precursor epithelial cells.

Fig. 3
figure 3

Essential roles and expression patterns of transcription factors during different embryonic stages of the beta cell lineage. At the top, arrows point to the developmental steps at which the indicated transcription factors have been shown to be indispensable in genetic knock-out studies of mice, irrespective of when these genes are expressed. At the bottom the relative expression levels of different transcription factors throughout pancreatic development are indicated schematically

There is thus considerable information on which transcription factors are expressed in pancreatic precursors and which ones are essential for early organogenesis. However, very little information is available at present on how transcriptional networks are organised within these cells. For example, nothing is known about the transcription factors that directly activate Ptf1a, Nkx6.1, Nkx2.2 and Hlxb9 (Hb9) genes in the pancreatic primordium. There are also no data about their direct targets at this stage.

Information on the regulators and targets of Pdx1 in early bud networks is at a more advanced stage. The importance of understanding the network linked to Pdx1 in pancreas organogenesis cannot be overstated. Pdx1 is not only essential for early pancreas formation [36], but can also activate pancreatic developmental programs if misexpressed in the right cellular context. Two laboratories have shown that Pdx1 can induce beta cell properties when expressed in adult or fetal liver cells [41, 42]. Another study demonstrated that a fusion protein containing the Xenopus orthologue of Pdx1 and the VP16 transcriptional activation domain can induce the formation of pancreatic tissue in the liver when misexpressed using the transthyretin promoter [43]. In this experiment VP16 provided a very strong activation domain, which probably acts as a surrogate for other activators that need to act in concert with Pdx1 at common genomic targets.

So which transcription factors control the activation of Pdx1 in early bud cells? It was recently shown that the activation of retinoic acid receptors during gastrulation specifies the location of Pdx1 expression along the anterior–posterior boundary of the endoderm in zebrafish [44]. Moreover, Hb9 (in keeping with its role in the specification of dorsal pancreas) activates Pdx1 in the dorsal region of the foregut by an unknown mechanism, while Hex1 is indirectly required for Pdx1 expression in the ventral foregut [33, 35, 40]. Another recent study reported that the onecut homeodomain protein Hnf6 regulates Pdx1 in the primordium [45]. Hnf6 is expressed in the foregut endoderm prior to Pdx1. Embryos lacking Hnf6 exhibit reduced expression of Pdx1 at E10 and delayed development of both ventral and dorsal pancreas, resulting in pancreatic hypoplasia (as well as other abnormalities discussed later) [45]. In vitro binding and cotransfection studies indicate that Hnf6 binds and transactivates mouse Pdx1 5′ flanking DNA [45], suggesting that Hnf6 is the first known direct regulator of Pdx1 expression during early bud formation (Fig. 4).

Fig. 4
figure 4

Genetic networks during defined cellular stages of pancreatic development. The four compartments represent consecutive embryonic cellular stages analogous to those shown in Figure 3. Gene promoters are represented by a rectangle and a thick arrow. Binding of a transcription factor to a promoter is represented by a thin line. A perpendicular line replaces an arrow when an interaction is inhibitory. Only interactions shown to occur in vivo and for which there is biochemical evidence of a direct effect are depicted

The role of two other candidate regulators of Pdx1 remains less well established. High-affinity binding of Pdx1 to the Pdx1 gene has been reported, consistent with an autoregulatory mechanism that operates once the gene is turned on [46, 47]. However Pdx1 −/− mice exhibit apparently normal expression of a LacZ reporter knocked into the Pdx1 locus in early bud cells [38]. Similarly, although Foxa2 (Hnf3β) is expressed in the foregut and early pancreatic cells, and binds the Pdx1 gene [47, 48], LacZ transgenics containing a mutation of a Pdx1 gene–Foxa binding site affect expression in beta cells rather than in pancreatic rudimentary epithelial cells [49]. Cultured Foxa2 −/− embryoid bodies exhibit reduced expression of Pdx1 mRNA [50], but evidence of a similar role in the early embryonic pancreas is still lacking.

A different but equally crucial question is that of how Pdx1 controls pancreas organogenesis. The identification of the genomic targets of Pdx1 in early pancreatic buds should provide important clues in this regard. So far, the only direct target of Pdx1 reported at this stage is the nuclear receptor named liver receptor homologue 1 (Lrh1, NR5A2) [51]. This receptor is expressed in early bud epithelium, differentiated ducts and acinar cells [52]. Pdx1 was shown to bind to the Lrh1 gene in vitro and in vivo in embryonic pancreas, and to transactivate episomal and endogenous Lrh1 in pancreatic carcinoma cells [51]. Furthermore, Lrh1 mRNA is reduced in pancreatic buds of Pdx1 −/− embryos at E9.5 (although, surprisingly, this study reported reduced expression in liver) [51]. Unfortunately Lrh1 −/− embryos die before the pancreas is formed [51], so further experiments will be needed to define the possible role of Lrh1 as a downstream Pdx1 target. Other Pdx1-dependent genes are Hlxb9 (Hb9), which is induced by ectopic expression of Pdx1 in chick embryo endoderm, and Sonic hedgehog, a negative regulator of pancreatic specification that is suppressed by Pdx1. However, it is not known if these effects are direct [53].

The regulatory network controlling the formation of neurogenin 3+ cells from embryonic duct precursor cells

The embryonic mouse pancreas undergoes major changes by E13.5 to E17.5. The two buds fuse, and signs of acinar and endocrine differentiation become evident, while the remaining pre-differentiated cellular epithelium is organised into prominent branching ductal structures (Fig. 2) [3]. In addition to these morphological changes, the pre-differentiated duct epithelium differs from earlier Pdx1+ bud epithelial cells at the level of gene expression. Pdx1, Nkx2.2 and Nkx6.1 are visibly down-regulated (though still expressed), while Hnf6 and Hnf1β expression is maintained [34] (Fig. 2, Fig. 3, bottom panel). In fact, Hnf1β is slightly induced and serves as a specific marker for this ductal epithelial domain because it is not visibly expressed in differentiated acinar or endocrine cells [34].

The predominant wave of pancreatic endocrine cell generation takes place precisely at this embryonic stage [54, 55]. Endocrine cells originate from lineage-committed progenitors that are marked by the helix–loop–helix transcription factor neurogenin 3 (Ngn3) [56]. Ngn3+ cells are present intercalated with Hnf1β+ cells in the epithelium that lines the lumen of embryonic ducts, and in their immediate vicinity (Fig. 2) [34]. Ngn3+ progenitors are short-lived cells lacking a significant self-renewal capability [27]. This raises the question of which cells provide the precursor pool for Ngn3+ cells. Several observations, including the existence of frequent Hnf1β+/Ngn3+ transition cells and the paucity of Hnf1β/Ngn3 cells in embryonic duct epithelium, support the idea that Hnf1β+ cells are the direct precursors of Ngn3+ cells [34]. The embryonic Hnf1β+ duct cell domain probably also provides the precursors of differentiated pancreatic duct cells, as pancreatic ducts are essentially composed of Hnf1β+ epithelial cells from E13.5 up to birth [34]. Thus, Hnf1β+ ductal cells at E13.5 to E17.5 can be regarded as a distinct stage of embryonic precursors that immediately precedes the initiation of lineage-specific differentiation (Fig. 2, Fig. 3).

What are the transcriptional mechanisms controlling the generation of Ngn3+ cells within the Hnf1β+ epithelial domain? The current evidence points to the presence of positive and negative mechanisms regulating mutually exclusive pathways either to differentiation or to self-renewal of the Hnf1β+ epithelium. Hes1, a basic helix loop helix transcriptional repressor that is an effector of the Notch receptor signalling pathway, is a major negative regulator of endocrine cell commitment [57]. Thus, Hes1 homozygous null mutant embryos exhibit accelerated differentiation of Ngn3+ and endocrine cells, resulting in pancreatic hypoplasia due to early depletion of precursors [57]. This effect may be mediated through direct repression of the Ngn3 gene by Hes1 [58]. Hes1 is activated by Notch receptor signalling, and accordingly the knock-outs for the Notch pathway genes delta-like ligand 1 (dll1) or RBP-Jκ/CBF-1 exhibit phenotypes closely related to that of the Hes1 knock-out [59]. In contrast, overexpression of activated Notch in embryonic precursors using Pdx1 promoter-driven transgenics suppresses both endocrine and exocrine differentiation [60, 61].

The inactivation of Notch/Hes1 in pancreatic duct precursors is therefore a decisive event for endocrine (and exocrine) differentiation. Accordingly, Hes1 is absent in endocrine cells [62]. However, the mechanisms whereby Notch/Hes1 is turned off in some precursor cells are unknown. One speculative scenario is that Ngn3 participates in this process, forming a negative cross-regulatory feedback loop. The activation of Ngn3 beyond a threshold in a given cell could thus result in the inhibition of its own repressor, hence consolidating commitment to the endocrine lineage. On the other hand, because Ngn3 turns on Notch ligand genes such as delta-like 1, activation of Ngn3 in a precursor cell is thought to suppress differentiation in the surrounding cells as a mechanism to modulate the number of endocrine cells being generated [59, 63, 64].

The most prominent known positive regulator of Ngn3+ cell formation is Hnf6 [65]. Hnf6 is normally expressed in Hnf1β+ duct epithelium and acinar cells, but not in endocrine cells [30, 34, 65]. The Hnf6 knock-out develops a pancreas but almost completely fails to generate Ngn3+ progenitors and endocrine cells during embryogenesis [65]. Moreover, Hnf6 interacts in vitro with a 5′ flanking cis element of the Ngn3 promoter and transactivates it in transient transfections [65]. It thus forms part of the program that induces the formation of Ngn3+ cells in concert with the inactivation of Notch signalling.

Another candidate key component of the Hnf1β+ precursor cell gene program is Hnf1β. Embryos deficient in Hnf6 have markedly reduced Hnf1β in primitive pancreatic ducts, demonstrating that Hnf6 is an important regulator of Hnf1β in these cells [34]. Interestingly, Hnf6 also controls Hnf1β in the bile duct epithelium, and both Hnf6-deficient mice and liver-specific Hnf1β null mutants exhibit arrested bile duct differentiation, indicating that the Hnf6–Hnf1β hierarchy regulates bile duct differentiation [66, 67]. However, the exact role of Hnf1β in pancreatic ducts is still unclear. Whereas homozygous germ-line Hnf1β mutations cause embryonic death before the pancreas is formed [68, 69], heterozygous mice have not been reported to exhibit a phenotype. Yet humans with diabetes owing to heterozygous Hnf1β mutations do exhibit pancreatic hypoplasia, which would be the expected phenotype, if Hnf1β acts as a regulator of pluripotent pancreatic precursor cells [70]. The Hnf6–Hnf1β hierarchy thus links two transcription factors with important pancreatic developmental phenotypes, suggesting that it is an integral component of an as yet incompletely understood genetic program that regulates the function of ductal precursor cells.

Intriguingly, the embryonic bile duct epithelium shares many of the transcriptional components with its pancreatic equivalent, including the Hnf6–Hnf1β hierarchy, expression of Hes1, and even Pdx1 [34, 38, 66, 71]. A striking recent report showed that Hes1-deficient embryos exhibit conversion of bile duct epithelium into a pancreas-like structure containing both acinar and islet endocrine cells, in a process that involves activation of Ngn3 [71]. Thus, embryonic pancreas and bile duct epithelium share a common intrinsic program capable of driving pancreatic differentiation. Unlike pancreatic duct cells, bile ductal cells normally suppress this program due to their inability to switch off Hes1 expression. An understanding of the reasons for this difference could provide important clues on how to sort out the mechanisms underlying pancreatic differentiation. This finding also offers an extremely attractive potential route to induce the formation of beta cells in Type 1 diabetes.

Neurogenin 3 orchestrates a network that drives endocrine differentiation

Mice lacking Ngn3 have no pancreatic endocrine cells [56], indicating that Ngn3 is not just a marker of progenitors, but also an essential effector. In fact, Ngn3 is sufficient to induce endocrine differentiation in appropriate cellular contexts [53, 59, 63, 72]. Thus, transgenic overexpression of Ngn3 in early Pdx1+ cells results in massive endocrine differentiation, mostly glucagon cells [59, 72]. Similarly, electroporation of chick embryo endoderm with Ngn3 and Pdx1 gives rise to endocrine cells [53]. Again, glucagon but not insulin cells are generated, even if beta-cell-promoting factors such as Pax4 are added, indicating that the refinement of the endocrine subtype fate requires additional unidentified ingredients. In a different model, adenovirus was used to express Ngn3 in purified adult human ductal cells, resulting in expression of endocrine differentiation markers such as PC1, and even low levels of insulin [63].

The pro-endocrine role of Ngn3 results from the activation of several genes encoding endocrine differentiation regulators. Thus, in the human duct study described above, forced expression of Ngn3 induced expression of endocrine transcription factors NeuroD1, Pax4, Nkx2.2 and Pax6 [63], while in the Ngn3 knock-out [56] a similar set of genes became inactive.

Many of the Ngn3-dependent, pro-endocrine regulators are direct genomic targets. In vitro Ngn3 binds and transactivates the 5′ flanking region of NeuroD1 (Beta2), another cell-restricted basic helix–loop–helix transcription factor gene that regulates the survival and terminal differentiation of beta cells [73]. NeuroD1 is likely to be directly responsible for part of the pro-endocrine effects of Ngn3, given that its misexpression in precursor cells of transgenic embryos or in human ducts partially mimics the effects of Ngn3 [63, 72].

Neurogenin 3 also binds and transactivates a cis element in the Pax4 promoter [74], while Ngn3 −/− mice fail to express Pax4 mRNA [56], suggesting that Pax4 is also a direct target of Ngn3. During embryogenesis, Pax4 is only transiently expressed in progenitors (Fig. 3). Its function is to sub-specify the beta cell type within the endocrine progenitor pool [75]. Interestingly, one role of Pax4 is to directly or indirectly suppress the expression of another transcription factor, Arx, whose function is to promote the alpha rather than beta cell fate of Ngn3+ progenitors [76].

Another Ngn3 target, Nkx2.2, also drives endocrine differentiation. The regulation of Nkx2.2 is somewhat complex. Nkx2.2 is expressed at low levels in pre-differentiated epithelial cells, then exhibits a sharp induction in Ngn3+ cells and remains highly expressed in endocrine cells [32, 34, 64, 77] (Fig. 3). Nkx2.2 is controlled by alternative promoters at different cellular stages [78]. Immediately after the Ngn3+ stage, promoter 1a drives Nkx2.2 expression in endocrine cells [78]. It has been suggested that Ngn3 and NeuroD1 consecutively activate Nkx2.2 through this promoter in progenitor and endocrine cell stages, a suggestion based on the presence of E boxes that are bound and transactivated by these factors [78]. In contrast, promoter 1b is activated in Ngn3+ progenitor cells and probably requires a separate activation mechanism parallel to the induction of Ngn3, while the cis regulatory elements involved in the earlier bud cells are less well characterised [78].

Even though Nkx2.2 is present in pancreatic epithelium prior to the activation of Ngn3, its first essential role is not until the immediate post-Ngn3 stage (Fig. 3). The inactivation of Nkx2.2 gives rise to endocrine-like cells lacking insulin or Glut2 but expressing other endocrine markers such as amylin and Isl1, which suggests that the endocrine differentiation process is blocked or diverted [32]. In contrast, inactivation of another homeodomain transcription factor, Nkx6.1, results in a severe reduction of insulin-producing cells [31]. Remarkably, Nkx2.2/Nkx6.1 double knock-outs entirely mimic the Nkx2.2 mutant pancreatic phenotype [31], suggesting that the essential role of Nkx6.1 takes place after the step controlled by Nkx2.2.

Interestingly, Ngn3 has also been shown to negatively regulate its own promoter in transient transfection studies, providing a potential mechanism for self-inactivation that is consistent with its transient expression during pancreatic development [79].

Pdx1, a critical regulator of the differentiated beta cell phenotype

Once beta cells are formed during pancreatic development, genetic programs are needed to carry out their specialised functions. One obvious function of such programs is to support high levels of insulin gene transcription. Another equally important task is to activate a vast number of genes involved in (i) the sensing of extracellular signals such as nutrients, hormones and neurotransmitters; (ii) the coupling of these signals to the process of secretion; and (iii) the regulation of beta cell mass or the ability to respond to immune-mediated damage [80, 81, 82].

In addition to its role in early organogenesis, Pdx1 is also a key regulator of differentiated beta cells. Shortly after the Ngn3+ cell stage, Pdx1 is sharply induced in insulin-producing cells and remains highly expressed throughout adult life (Fig. 3). Targeted disruption of the Pdx1 gene in mature beta cells leads to a diabetic phenotype due to impaired beta cell function [6, 83, 84]. This is associated with reduced activity of islet-specific genes regulated by Pdx1 such as insulin, IAPP, Nkx6.1 and Glut2 [6, 84]. A similar expression profile results from dominant negative suppression of Pdx1 in INS1 beta cells [85]. In vitro binding and transient transfection studies suggest that Pdx1 interacts directly with the 5′ flanking regions of these four genes [86, 87, 88, 89], while ChIP experiments have confirmed that Pdx1 binds in vivo to the insulin and IAPP genes in cultured beta cell lines [90, 91].

The key role of Pdx1 in beta cells is further emphasised by mouse and human genetic studies showing that beta cells are sensitive to Pdx1 gene dosage. Thus heterozygous mutant mice exhibit beta cell dysfunction and diabetes associated with defective expression of Glut2, IAPP and Nkx6.1 [92, 93], as well as increased apoptosis [94], whereas in humans heterozygous IPF1 (Pdx1) mutations lead to Maturity onset diabetes of the young (MODY) 4 [95].

How is cell-specific, high-level Pdx1 expression achieved in beta cells?

Given the pre-eminent role of Pdx1 in differentiated beta cell function, it is not surprising that the regulation of the Pdx1 gene in these cells has attracted considerable interest. Numerous transgenic and cell culture studies have by now established that the cell-specific transcription of the Pdx1 gene is mediated by several cis regulatory areas located at more than 1.6 kb from the transcription site, most of which are evolutionarily conserved [48, 96, 97].

Several regulators of Pdx1 in beta cells have been identified, one of which is Foxa2. As mentioned earlier, in vitro binding studies and transient transfection studies have identified multiple Foxa2-binding sites in the distant 5′ flanking regions of human, mouse and rat Pdx1 genes [47, 48, 50, 97, 98], while ChIP analysis has shown that Foxa2 binds to the mouse Pdx1 distant enhancer region in beta cell lines [49]. Transgenic experiments have shown that the so-called Area II of the Pdx1 distant 5′ flanking region, which contains one of the high-affinity Foxa2-binding sites, can drive beta-cell-specific expression, whereas mutation of the Foxa2-binding site reduced beta cell expression penetrance of another transgene [49]. Furthermore, beta-cell-specific ablation of Foxa2 with the Cre/loxP system results in down-regulation of islet Pdx1 mRNA and protein levels [99].

Recently, the beta-cell-enriched MafA basic-leucine zipper protein was found to bind to a conserved Pdx1 Area II sequence in vitro and in vivo in a beta cell line [100]. Transient transfection studies with substitution mutants of the Pdx1 gene Maf-binding site suggest that MafA exerts positive regulation of this gene in cultured beta cell lines [100]. The same region is bound by Pax6 in beta cell lines, both in vitro and in vivo, and mutational analysis of the Pax6-binding site suggests that it too is a positive regulator of the Pdx1 gene [49].

As mentioned earlier, some data point to autoregulation of Pdx1, a finding based on the in vitro [47] and in vivo binding of Pdx1 to a conserved distant enhancer region of Pdx1 in cultured beta cells [46, 90]. The physiological consequences of this loop in beta cells are still unclear, as Pdx1+/− islets exhibit, as expected, roughly 50% expression of Pdx1 mRNA, whereas some existing studies have proposed either positive or negative regulatory effects [46, 47, 83, 93].

Hnf1α, a POU-homeodomain protein involved in MODY3, binds in vitro to the mouse and human Pdx1 genes [46, 98]. The genetic data on the functional consequences of this interaction are somewhat controversial. Whereas one study showed a 2.9-fold reduction of Pdx1 mRNA in islets of Hnf1α −/− mice [21], two others failed to show this in young Hnf1α −/− mice [19, 101] (the latter [101] did find a reduction after 6 weeks of age). Moreover, dominant negative suppression of Hnf1α in INS1 cells fails to inhibit Pdx1 mRNA [102]. Potential explanations of these discrepancies include (i) the effects of diverse genetic backgrounds, and (ii) the suggestion that Hnf1α is required after a certain postnatal stage and/or that the observed effects are secondary to hyperglycaemia.

So far, in short, no single factor has been found to be absolutely indispensable for Pdx1 expression in beta cells. On the other hand, recent studies have revealed a combination of cell-specific transcriptional activators that bind to the Pdx1 gene, many of which are involved in ensuring high levels of expression in beta cells.

A signalling cascade acting through Pdx1

Recent studies have provided interesting insights into how signalling cascades can mediate their effects through the modulation of beta cell transcription networks. A revealing example is the regulation of Pdx1 by Irs2 and Foxo1, two transducers of key extracellular signals such as insulin and IGF1 [103].

Several genetic experiments support the proposal that Irs2 is required for Akt to phosphorylate the forkhead transcription factor Foxo1, which is then excluded from the nucleus and prevented from inhibiting the Pdx1 gene. Irs2 −/− mice thus exhibit reduced phosphorylation of Foxo1 and diminished expression of Pdx1 [104, 105]. Haplo-insufficiency for Foxo1 partially restores the expression of Pdx1 and reverses beta cell failure caused by Irs2 deficiency. Moreover, a constitutively active Foxo1 mutant inhibits Pdx1 expression in beta cells [106].

The mechanism whereby Foxo1 exerts this negative effect on the Pdx1 gene is not known, although one suggestion is that it inhibits Foxa2-dependent regulation of Pdx1 [104]. On the other hand, the extent to which the IRS2 and Foxo1 findings relate to insulin and IGF1-dependent signalling has yet to be established. Beta-cell-specific knock-outs for insulin- and IGF1-receptor genes have reduced beta cell function, but this has not been reported to result in reduced Pdx1 expression [107, 108, 109, 110]. Future studies will thus need to further define the precise signals and regulatory contexts that drive Irs2/Foxo1 regulation of the Pdx1 gene.

Multi-input motifs control the transcription of beta-cell-specific genes

Much of what we know about the transcriptional control of beta cells stems from a wealth of studies that analysed the cis and trans elements involved in insulin gene transcription [111, 112]. Three major pancreatic transcription factors were discovered precisely because of their ability to bind to several evolutionarily conserved sites in the insulin gene promoter, namely Pdx1 [113], MafA [114, 115] and NeuroD1/Beta2 [116], although these same factors were also independently cloned through other routes [117, 118, 119, 120, 121]. Subsequent in vivo studies have verified that all three factors occupy insulin gene chromatin in cultured beta cell lines [90, 91, 115]. As discussed earlier, Pdx1 is essential for normal insulin gene expression in mouse beta cells [6, 83, 84]. The MafA knock-out has not been reported on, although cell culture studies indicate that one major role may be related to glucose induction of insulin gene transcription [120], while transgenic studies analysing a DNA rat insulin promoter 3 element, to which MafA binds, suggest that it contributes to cell-specific expression [122]. Surprisingly, homozygous NeuroD1 mutant mice showed that this gene is not indispensable for insulin gene expression, despite numerous functional studies in cultured cell lines indicating that it plays a regulatory role in rodent and human insulin genes [116, 123, 124, 125]. To date, no detailed analysis of these mice has determined whether insulin mRNA content is partially reduced or whether the physiological regulation of insulin gene expression is impaired.

Two other factors that bind the insulin gene in vitro and in vivo in cultured beta cell lines are Nkx2.2 and Pax6 [91]. As discussed earlier, Nkx2.2 is essential for insulin gene activity in mice [32], and in one of the two Pax6 mutants marked reductions of insulin gene expression were observed [126]. Thus, analogously to the situation described for the Pdx1 gene, beta-cell-specific activation of the insulin gene is dependent on a unique combination of nuclear factors occupying its 5′ flanking region in beta cells.

Remarkably, a nearly identical collection of transcription factors (i.e. Pdx1, Pax6, NeuroD1, Nkx2.2) has been reported to occupy a set of beta cell promoters that includes glucokinase, IAPP and Pax4 [91, 127] (Fig. 4). This pattern, whereby multiple factors control a common set of genes, is a characteristic motif encountered in transcription networks and referred to as a multi-input motif [128, 129] (Fig. 5). It has been suggested that such multi-input motifs confer the ability to provide coordinated gene responses to diverse regulatory signals [128], but they can also more simply underlie the combinatorial code required to activate a particular set of cell-specific genes. On the other hand, MafA and Pax6 both regulate the insulin gene as well as Pdx1 [49, 91, 100, 115], which itself is a regulator of the insulin gene. Feed-forward motifs of this sort are thought to serve as a switch, capable of responding to sustained rather than transient inputs (the distal target gene is activated only if a previous input has enabled expression of the intermediate gene) [128, 129, 130, 131]. It could therefore be speculated that such motifs modulate insulin transcription as a function of long-term metabolic demands.

Fig. 5
figure 5

Basic regulatory motifs in genetic networks. Certain patterns of regulation between transcription factor genes are over-represented in gene networks. Circles represent transcription factor genes, arrows show that a gene product exerts direct control of another gene or its own gene. This figure is largely based on data presented in references [128, 129]

Collectively these results show that the transcription of the insulin gene results from the concerted action of a complex non-lineal beta-cell-specific network. While several major components of this network have been identified, the regulatory logic of its structure is not yet understood.

Several MODY genes play key roles in a common differentiated beta cell network

The genetic analysis of MODY has provided a mine of previously unsuspected information on beta cell transcription networks. We now know that MODY is caused by heterozygous mutations in at least five genes encoding transcription factors: HNF4α/HNF4A (MODY1), HNF1α/TCF1 (MODY3), PDX1/IPF1 (MODY4), HNf1β/TCF2 (MODY5) and NEUROD1/BETA2 (MODY6) [95, 124, 132, 133, 134] (reviewed in references [14] and [13]). MODY2, which is so far the only subtype not related to a transcription factor [135], is caused by mutations in the glucokinase gene.

MODY is essentially a haplo-insufficient phenotype, based on functional studies showing that most mutations are loss-of-function defects and that disease-causing mutations exist in promoter regions [7, 136, 137, 138, 139, 140]. Mutation carriers develop a severe impairment of insulin secretion and diabetes after at least 10 years of age [141, 142, 143]. Interestingly, beta cell dysfunction does not result from the sheer lack of insulin, as elegantly shown in a recent human therapeutic trial. This trial [144] revealed that MODY3 patients retain a considerable ability to secrete insulin in response to sulphonylureas. This suggests that Hnf1α controls specific differentiated functions of beta cells, rather than solely the beta cell mass or insulin gene transcription.

Hnf1α −/− mice are viable, and have thus turned out to be a very fruitful model to understand the function of MODY transcription factors in beta cells [19, 20, 21, 145, 146, 147]. These mice have abnormal glucose-induced insulin release and develop diabetes shortly after birth, in addition to other metabolic abnormalities in the liver and kidney [145, 146]. Reduced nutrient-induced insulin release in Hnf1α-deficiency has been linked to impaired islet glucose metabolism [102, 147, 148]. A number of beta cell targets of Hnf1α have been identified, e.g. Glut2 and L-pyruvate kinase (Pklr), but the role of specific gene defects in the stimulus–secretion coupling phenotype of Hnf1α-deficient islets is unclear [19, 20, 102, 147, 148]. Interestingly, mRNA levels of several beta cell targets such as Glut2 and Pklr are not reduced in hepatocytes or pre-differentiated pancreatic cells from Hnf1α −/− mice, suggesting that the role of Hnf1α is highly dependent on the developmental and cell-type-specific context [19].

Islets from Hnf1α −/− mice exhibit decreased expression of several transcription factor mRNAs, including Hnf4α, Hnf4γ, Shp and Foxa3, pointing to the existence of a broad Hnf1α-dependent transcriptional network [20, 21]. Being also a MODY gene, Hnf4α is a particularly interesting transcription factor target. Importantly, Hnf4α mRNA is dependent on Hnf1α specifically in differentiated pancreatic cells, but not in the liver [20]. This is because in the pancreas the Hnf4α gene is almost exclusively driven by the P2 promoter, which is bound and controlled by Hnf1α, whereas in the liver and most other tissues Hnf4α is predominantly driven by the P1 promoter, which does not require Hnf1α [20, 149, 150].

The tissue specificity of Hnf1α-dependent expression of Hnf4α is concordant with the common beta cell phenotype in MODY due to mutations in these two genes. This suggests that Hnf1α control of Hnf4α P2 promoter could be a critical regulatory link in beta cells. This is strongly reinforced by the identification of a natural human mutation in the HNF4α P2 promoter that selectively impairs binding by Hnf1α and causes diabetes [138]. What is remarkable about this mutation is that although Hnf1α regulates multiple genes in beta cells, the selective disruption of its regulation of Hnf4α has similar consequences to mutations that completely disrupt Hnf1α function [138]. On the other hand, this Hnf1 site mutation is one of three reported genetic defects that disrupt P2 promoter function and cause MODY [149, 151], and together they provide compelling confirmation that this is the biologically significant Hnf4α promoter in the pancreas, despite the fact that P1-driven transcripts can be detected by PCR in human islet RNA [152] (S.F. Boj and J. Ferrer, unpublished).

Regulation in the opposite direction, i.e. the regulation of Hnf1α by Hnf4α, also occurs. However, in this case there is no beta cell selectivity. The germ-line Hnf4α knock-out is lethal at a very early embryonic stage [153], but Hnf1α expression is reduced in Hnf4α −/− embryoid bodies [136], in Hnf4α-deficient hepatocytes from chimeric mutant embryos [154], and in a Cre/loxP-based Hnf4α liver-specific knock-out (Iannis Taliandidis, unpublished observations). In vitro and in vivo transgenic studies have shown that regulation of the Hnf1α gene by Hnf4α is dependent on an evolutionarily conserved cis sequence element in the Hnf1α promoter [17, 155, 156]. Support for the notion that Hnf4α controls Hnf1α expression both in liver and in pancreatic cells comes from a human mutation in this same conserved cis element that co-segregates with MODY in a large pedigree [140]. Furthermore, dominant negative suppression of Hnf4α in INS1 cells decreases expression of Hnf1α, resulting in inhibition of the same mRNAs as dominant negative suppression of Hnf1 [4]. Thus, Hnf1α is dependent on Hnf4α in numerous cell types, but specifically in pancreatic cells the two genes are interdependent.

The Hnf1α–Hnf4α positive cross-regulatory loop could act as a cellular memory mechanism that maintains the expression of the two genes and their targets in differentiated pancreatic cells. Other self-sustaining cross-regulatory circuits connecting MODY genes are also likely to operate in beta cells. For example, Hnf1α mRNA is reduced in Pdx1 +/− mice [93], while in some studies discussed earlier Pdx1 and NeuroD1 mRNA was reduced in Hnf1α −/− mice [21]. A regulatory interaction of some sort between Pdx1 and Hnf1α is further supported by the finding that Pdx1 +/− Hnf1α +/− double heterozygous mice exhibit a more severe beta cell defect than expected from the sum of the single heterozygous defects [93]. The regulation of Pdx1 by Hnf4α has not yet been analysed, but Pdx1 interacts in vitro with a TAAT element in the Hnf4α P2 promoter and transactivates it in transient transfections [149]. A mutation in this cis element causes MODY in a large pedigree, indicating that the sequence is essential for Hnf4α promoter activity [149].

MODY as a paradigm of a haplo-insufficient defect resulting from the breakdown of cell-specific networks

One of the outstanding questions in the genetics of MODY is why heterozygous mutations in genes like HNF1α or HNF4α cause a phenotype essentially restricted to beta cells, whereas homozygous mutations give rise to phenotypes affecting numerous other cell types [7, 153, 157, 158]. The resolution of this paradox may reside in the fact that these transcription factors form part of different networks in distinct cell types. The network operating in beta cells may be sensitive to haplo-insufficiency, whereas other networks need not be affected.

It is thus worth reflecting on the consequences of reduced gene dosage in the context of the pancreas-selective Hnf1α–Hnf4α circuit configuration [7]. Positive cross-regulatory loops like this one (gene A is dependent on B, gene B is dependent on A) typically exhibit bistable behaviour [9, 159, 160]. By this we mean that reciprocal activation between two genes (such as Hnf1α and Hnf4α) can serve to maintain the activity of both genes in a very stable state. However, because they are interdependent, a transient reduction of either gene that is severe enough to fall below a certain threshold can lead to a severe inhibition of the other gene. This results in an alternate stable state, in which the two genes are expressed at a level below that required to activate each other [160]. Reduced gene dosage of Hnf1α or Hnf4α could therefore increase the probability that the transition from the active to the inactive stable state occurs due to the diminished ability of a one-allele system to prevent severe transient reductions of a gene product [7, 161, 162]. In such a scenario, a germ-line mutation that disrupts a single Hnf1α or Hnf4α allele can lead to the stable functional knockdown of all four alleles that compose the circuit. This is expected to occur selectively in cells where the two genes are interdependent, in keeping with the cellular-specificity of MODY.

MODY thus apparently results from the collapse of cell-specific transcriptional networks due to the haplo-insufficiency of key network genes. In fact, the genetic analysis of MODY may have helped to reveal not only unsuspected beta cell transcription factors, but also the keys to understand the functional logic of underlying regulatory networks.

Pancreatic transcription networks as complex systems

As more knowledge is gained about pancreatic transcription networks, it will become theoretically possible to interpret general aspects of their behaviour, and not solely their individual components. Studies in much more manageable models than the mammalian pancreas, such as Escherichia coli and Saccharomyces cerevisiae, have provided enough data to indicate that transcription factors that control most cellular functions form complex networks [128, 129, 130, 163]. It is thus worth considering some overarching principles governing the behaviour of complex networks.

One interesting notion is that networks as diverse as the Internet, ecological webs or transcription and metabolic networks are generally arranged in a highly non-random manner [130, 164, 165]. A very common non-random configuration is known as the scale-free network. In such a network a few nodes are disproportionately highly connected to other nodes, whereas most nodes have very few links [164]. Such networks are extremely robust inasfar as the individual loss or change in behaviour of most nodes have few consequences. However, selective inactivation of the few highly interconnected nodes can be catastrophic [164]. It will be interesting to see to what extent this concept applies to pancreatic networks as well as potentially to the fact that some transcription factor mutations in mice and humans with pancreatic transcription factor loss of function lead to particularly severe pancreatic phenotypes, whereas others appear to be inconsequential.

Another interesting non-random feature of most complex networks is the over-representation of connectivity patterns or “motifs” [128, 129, 130]. Experimental studies have shown that transcription factor networks in particular possess a disproportionately high representation of certain motifs such as feedforward loops, autoregulation, multi-component feedback loops or multi-input motifs [128, 129, 130] (Fig. 5). Of particular interest is a landmark study that used ChIP-on-chip to identify the in vivo binding sites in the S. cerevisiae genome for 106 of the 141 known S. cerevisiae transcription factors [128] (schematic description of ChIP-on-chip, see Fig. 1). This study provided a bird’s eye view of the topology of a transcription network in a eukaryotic cell and showed that transcription factor genes are highly interconnected, with frequent well-defined regulatory motifs such as those shown in Fig. 5, as well as highly connected hub-like nodes [128].

Interestingly, similar motifs have also been encountered in pancreatic networks, as discussed in this review. Experimental and mathematical modelling data suggest that such over-represented motifs fulfil specific regulatory strategies, some of which have been discussed earlier in this review [9, 129, 130, 131, 159]. In some cases, motifs or combinations of motifs form modules whose function can be analysed independently of the remaining network [166]. In fact, it is possible that an understanding of the function of large networks can, to a large extent, be achieved through careful hypothesis-driven analysis of key modules. Resources such as ChIP-on-chip screens could thus provide the “road maps” that describe the global network topology and help identify unsuspected subnetworks.

Concluding remarks

The development of the pancreas is controlled by transient networks that operate during consecutive precursor cell stages. Specific gene networks are also in place to provide long-term maintenance of the differentiated beta cell phenotype. The past few years have seen unprecedented progress in our understanding of pancreatic networks, largely through the testing of focused hypotheses. Further advances are expected in the field as a consequence of large-scale gene expression, ChIP, proteomics and mutation screens. The significance of these studies will be determined by whether they can be applied to the major developing and adult pancreatic cell types, as opposed to transformed cell lines or heterogeneous cellular populations.

From what is already known, it is apparent that most pancreatic transcription networks are not simple lineal hierarchies, but complex networks analogous to those recently described in more simple cellular models. Further complexity is to be expected from the integration of gene networks with other regulatory networks, including those formed by signal transduction pathways and protein–protein interaction webs. We believe that an approach combining the availability of large-scale network maps with more focused hypothesis-driven experimentation will eventually provide an in-depth understanding of the behaviour of pancreatic networks that is concordant with the complex biological reality. This in turn should provide foundations for the development of new therapies for beta cell loss and dysfunction in human diabetes.