Transcriptional control in the segmentation gene network of Drosophila

Abstract

The segmentation gene network of Drosophila consists of maternal and zygotic factors that generate, by transcriptional (cross-) regulation, expression patterns of increasing complexity along the anterior-posterior axis of the embryo. Using known binding site information for maternal and zygotic gap transcription factors, the computer algorithm Ahab recovers known segmentation control elements (modules) with excellent success and predicts many novel modules within the network and genome-wide. We show that novel module predictions are highly enriched in the network and typically clustered proximal to the promoter, not only upstream, but also in intronic space and downstream. When placed upstream of a reporter gene, they consistently drive patterned blastoderm expression, in most cases faithfully producing one or more pattern elements of the endogenous gene. Moreover, we demonstrate for the entire set of known and newly validated modules that Ahab's prediction of binding sites correlates well with the expression patterns produced by the modules, revealing basic rules governing their composition. Specifically, we show that maternal factors consistently act as activators and that gap factors act as repressors, except for the bimodal factor Hunchback. Our data suggest a simple context-dependent rule for its switch from repressive to activating function. Overall, the composition of modules appears well fitted to the spatiotemporal distribution of their positive and negative input factors. Finally, by comparing Ahab predictions with different categories of transcription factor input, we confirm the global regulatory structure of the segmentation gene network, but find odd skipped behaving like a primary pair-rule gene. The study expands our knowledge of the segmentation gene network by increasing the number of experimentally tested modules by 50%. For the first time, the entire set of validated modules is analyzed for binding site composition under a uniform set of criteria, permitting the definition of basic composition rules. The study demonstrates that computational methods are a powerful complement to experimental approaches in the analysis of transcription networks.

Figure 6. Module Predictions within the Segmentation Gene Network

(A) Schematic depiction of the regulatory relationships within the segmentation gene network. (B) Ahab-predicted modules in the control regions of segmentation genes were classified based on their composition into pair-rule driven (pr, red), maternal/gap driven (mg, green), and mixed but predominantly pair-rule (pr(mg), light red) or predominantly maternal/gap driven (mg(pr), light green); see text for details. For each gene, the number and type of modules in the control region is shown; grouping of genes is indicated by brackets and follows the hierarchy as depicted in (A). The type of regulatory input a gene receives is indicative of its position within the gene network.

Figure 1. Ahab Predictions and Recovery of Known Modules

(A) Histogram of genome-wide window scores for the Ahab mg run (maternal/gap input, window size 500 bp, window shift 50 bp, background model 2). As free energy cutoff we chose 15, which is approximately four standard deviations above the genome-wide mean (indicated by light blue line). (B) Pie chart summarizing results of Ahab predictions for gap and pair-rule genes, including recovery of known modules and testing of novel predictions. (C–F) For the genomic regions of selected gap and pair-rule genes, the free energy profiles of two Ahab runs (mg and mgpr) are shown. The free energy cutoffs are marked by dotted lines; statistically significant predictions for the mg run are marked by black arrow heads (cf. Figure 4). In the header above, the blastoderm expression pattern of the locus is depicted schematically, anterior to the left, posterior to the right. The position of experimentally validated modules within the control region is delineated by colored bars; the aspect of the endogenous pattern they drive is highlighted in matching color. Overall, the control regions of the gap genes hb and Kr and of the primary pair-rule genes eve and h are computationally well delineated with maternal/gap input. References: (1) Schroder et al. (1988), (2) Margolis et al. (1995), (3) Hoch et al. (1990), (4) Goto et al. (1989), (5) Fujioka et al. (1999), (6) Riddihough and Ish-Horowicz (1991), (7) Howard and Struhl (1990), and (8) Langeland and Carroll (1993).

Figure 4. Correlation of Expression Patterns with Module Composition

Based on the expression pattern they give rise to, known and newly validated modules are sorted into anterior, posterior, and terminal (if expression bridges the 50% EL line, the module is labeled ant/post), and their binding site composition is evaluated using Ahab output from the mg run. The expression pattern of a module is depicted schematically (anterior = 100% EL, left; posterior = 0% EL, right), followed by name of gene, name of module, recovery as significant prediction (marked by X) or as subthreshold peak (marked by (X)) in D. melanogaster and D. pseudoobscura, distance to the gene's transcription start site (negative values denote upstream location), and binding site composition. For references see Dataset S1. Expression patterns of previously known modules are in black, those of newly validated modules are in dark pink, and modules with unfaithful/unstable patterns are in light pink. Binding site composition is given in the form of integrated profile values for individual input factors (see Materials and Methods); higher color intensity emphasizes higher values. Diagnostic features are emphasized by black trim: In anterior modules Bcd sites are overrepresented and Cad sites are underrepresented, while in posterior modules Cad sites are overrepresented and Bcd sites underrepresented. Terminal modules are enriched in TorRE sites.

Figure 2. Expression Patterns Driven by Ahab-Predicted Modules I

Ahab-predicted modules in the control region of gap and pair-rule genes were tested by fusing putative modules to a basal promoter driving lacZ (module-basal promoter-lacZ; Thummel and Pirrotta 1991). The genomic regions, with free energy profiles, for two Ahab runs (mg and mgpr) are shown on the right. The free energy cutoffs are marked by dotted lines; mg run predictions with scores greater than 15 are marked by black arrowheads, tested subthreshold peaks with scores below 15 by open arrowheads. The transcribed region of the locus is marked in blue, the experimentally tested genomic regions are marked by pink bars and named according to distance from transcription start site to middle of the enhancer, and previously known modules are marked by orange bars. The endogenous gene expression is shown on the left (blue frame), the expression pattern driven by the module(s) in the center (pink frame). Embryos are oriented anterior to left, dorsal up. In a few cases, the patterns driven by Ahab-predicted modules are unfaithful to the endogenous gene expression; we distinguish “unfaithful” and insertion-dependent “unstable” patterns. For further description see text. (A) gt, (B) cnc, (C) oc, (D) D, (E) cad, (F) fkh, and (G) slp2. References: (1) Berman et al. (2002), (2) Gao and Finkelstein (1998), (3) Lee and Frasch (2000), and (4) Pankratz et al. (1992) and Rivera-Pomar et al. (1995).

Figure 3. Expression Patterns Driven by Ahab-Predicted Modules II

See legend for Figure 2. (A) kni, (B) knrl, (C) pdm2, (D) nub, and (E) odd.

Figure 5. Ap Distribution of Binding Sites and Cognate Input Factors

(A) Plots depict distribution of input factors (black) along the ap axis (anterior tip = 100, posterior tip = 0) (based on Myasnikova et al. [2001]) and the average number of binding sites (as measured by integrated profile values; Figure 4) found in all modules driving expression at a given percent EL (red) (see Materials and Methods). For TorRE, Bcd, and Cad, the distributions of binding sites and input factors are positively correlated. For Hb, Gt, and Kr, the distributions are negatively correlated; note that the number of binding sites is particularly high in modules expressed adjacent to the expression domain of these factors. In the case of Hb, modules with more Hb sites than Bcd sites (blue) show negative correlation with input factor distribution, and modules with fewer Hb sites than Bcd sites (green) show positive correlation, indicating bimodal function of Hb. For Kni and Tll, no clear correlations are found, possibly because of the unspecificity of their weight matrices. (B) Information scores of the Kr, Kni, and Tll weight matrices.

Figure 7. Genomic Position of Modules

Position of modules predicted by the Ahab mg run relative to the transcription start site of the cognate loci; predictions for the homeotic genes are excluded. The number of modules found at a given position is shown in blue. The black line indicates the probability of a module occurring at a given position (calculated by dividing the number of modules at a given position by the number of control regions extending to that position). The stippled black line shows that probability if modules were randomly distributed. Modules with maternal/gap input are clustered within the first 6 kb upstream, in the first 2 kb of intronic space, and around 2 kb downstream (measured from the end of the gene).