As the amount of experimental data grows and becomes more complex, the power of computational models to identify mechanisms of biological processes by integrating diverse sets of experimental data is being realized, particularly in organisms where high throughput perturbation analyses have been developed, including Saccharomyces cerevisiae, Caenorhabditis elegans, and Strongylocentrotus purpuratus [43–47]. However, finding and quantifying mammalian networks has been challenging due to the data requirements of current methods. In the present analysis, we have demonstrated the utility of bioinformatics approaches for elucidation of a GRN describing proneural bHLH (Ngn2 and Mash1) transcription factor regulation of murine telencephalon specification.
To develop a robust GRN, we first measured global gene expression patterns in GOF and LOF analyses to identify novel putative target genes for Ngn2 and Mash1, several of which were validated by in situ hybridization in slice preparations from mutant and electroporated mouse embryos. Using Bayesian network analysis, we have corroborated 25 literature-based GRN hypotheses by quantifying connections based on the compiled microarray datasets.
As highly conserved long-range enhancers are known to be particularly important for developmentally regulated genes [48–50], we utilized phylogenetic footprinting analyses to identify putative long-range enhancers in several predicted Ngn2 and Mash1 target genes. This approach suggests that no putative co-factor binding sites are specifically associated with conserved Mash1 binding sites in both Mash1 and common target genes, consistent with an instructive and relatively context-independent role for Mash1 in ventral cell fate determination . In addition, the analysis identifies several conserved TFBSs in close proximity to Ngn2 binding sites surrounding the Ngn2 target genes, consistent with previous findings supporting a more cell context-dependent role for Ngn2 in dorsal telencephalic fate specification . In addition, we hypothesized that important regulatory modules that bind transcription factors other than Ngn2 and Mash1 may be present surrounding the identified target genes. Therefore, our second comparative genomics analysis was designed to search for conserved TFBS enriched in target genes, but not necessarily in close proximity to Mash1 or Ngn2 binding sites. Using a novel Bayesian network analysis approach with an informative prior structure, we were able to synthesize the knowledge gained from each of the above experimental and computational approaches to predict connectivity between the novel target genes, co-factors, and co-regulators.
Our resultant GRN predicts that Creb1 and Crebbp are the most likely candidates for a dorsally expressed Ngn2 co-factor regulating cortical targets such as Neurod6, Eomes and NeuroD2. Interestingly, previous analyses have indicated interactions between bHLH transcription factors, Creb, and the Creb binding protein (CBP/P300 complex) in the differentiation of several cell types [52, 53], as well as in neurotrophin-mediated expression of vgf . Most notably, a conserved Neurod6 promoter has recently been shown to contain both Creb and E-box binding sites and is activated via cAMP exposure . Yy1 is also a potential Ngn2 co-factor, yet our microarray data suggest that Yy1 mRNA expression is confined to the ventral telencephalon. This finding suggests an inhibitory role for Yy1 in Ngn2-regulated transcription, which is supported by evidence showing Yy1 inhibition of both BMP induced cell differentiation  and Notch transcriptional activity .
Our GRN also supports a synergistic interaction between Pou-domain containing transcription factors and bHLH proneural proteins in the activation of several target genes. Interestingly, both Pou6f1 and Mash1 have been identified as important mediators of oligodendrocyte development [58, 59], and Pou6f1 is thought to synergistically interact with members of the Sox family [60, 61]. Moreover, other members of the Pou domain, class 3 family of transcription factors, Brn1 or Brn2, which exhibit similar affinity with Pou6f1 for certain binding sites [62, 63] have been experimentally shown to co-regulate several Mash1 targets in the ventral telencephalon .
Finally, our analysis predicts connectivity between Sox9, Fzd1, and Wnt7b, which is consistent with a recent report suggesting Sox9 regulation through Wnt signaling . Sox9 may serve as an inhibitory factor for several Ngn2 target genes including Gng2, Npdc1, Eomes, Fzd1, and Coro2b, consistent with the opposing roles of Sox9 and Ngn2 in the specification of glial fate [35, 36]. We also show that Mef2a, which has been shown previously to have complex regulatory functions in neuronal differentiation and plasticity [38–40], may co-activate several of the predicted Ngn2 targets, such as Neurod, Ngn1, Wnt7b, Mgst3, Gng2, Acpl2, and Dusp14. In addition, our findings suggest that transcriptional regulators downstream of Wnt signaling (Tcf4/Lef1) may bind to regulatory modules that also bind Ngn2, which is consistent with the role of Wnt signaling in the specification of the dorsal forebrain [65–67], and offers a hypothesis in which coordinated Wnt activation and Ngn2 expression act in concert to transcriptionally activate target genes. Interestingly, a recent report suggests that Wnt pathway-initiated neural differentiation, but not proliferation, requires specific interactions between Tcf/β-catenin and Crebbp . Another intriguing aspect of our analysis is the prediction of Elavl4 as an important regulator of Ngn2 targets, including Bhlhb5, Robo1, Nhlh1, Coro2b, Sox11, and Dll1, possibly through stabilization of mRNA . It is of note that two related mRNA stabilization genes, Elavl2 and Elavl3, are also predicted targets of Ngn2.
Our current research identifies important research steps for further refinement of the GRN. The cyclic nature of delayed negative feedback in the Notch pathway is thought to act as a molecular clock regulating the timing of several developmental processes . Future production of robust time-series datasets will allow for application of Bayesian methods that are not limited to the discovery of acyclic networks and linear relationships . Several recent reports have suggested connectivity between Notch and Wnt pathways [27, 68, 72, 73]. For example, one way in which the Notch pathway may regulate Wnt signaling is through Lef protein stabilization by Nrarp activation . However, because no data currently exist that would allow quantitative prediction via direct protein-protein interactions, we are unable to predict this relationship in the current network analysis. Nonetheless, our network analyses of transcriptional regulation offers other hypotheses such as a Tcf/Lef regulation of Lfng and Mfng, which is consistent with evidence during somitogenesis suggesting a Wnt mediated regulation of Lfng expression .
One important application of the network analysis is the prediction of the most useful perturbation or chromatin immunoprecipitation experiments for resolving the overall network structure . This is particularly relevant to studies in mammalian species, in which perturbation analyses are much more time and resource intensive. Our network analysis would suggest Centg3 and Elavl4 are important candidates for perturbation and subsequent global gene expression analysis for further network resolution, as they are both highly connected nodes and little is known about their function in the developing telencephalon. Interestingly, Centg3 may protect against neurodegeneration in Polyglutamine diseases . The central role of Elavl4, which regulates through mRNA stabilization , highlights the importance of moving beyond cis regulatory binding to elucidate network relationships.