Communication between levels of transcriptional control improves robustness and adaptivity

Alexander M Tsankov^1,3, Christopher R Brown², Michael C Yu^2,a, Moe Z Win³, Pamela A Silver² & Jason M Casolari^2,a

^aPresent address: Department of Biological Sciences, State University of New York at Buffalo, Buffalo, NY 14260, USA

^aPresent address: Department of Biochemistry, Stanford University, Stanford, CA 94305, USA

Article highlights

Synopsis

The nucleus is the distinguishing feature of the eukaryotic cell. It provides added mechanisms for regulating gene expression at the levels of chromatin organization, RNA processing, and selective export via the nuclear pore complex. Groups of proteins that mediate these processes have been extensively characterized. We propose that these functional groups of proteins exhibit extensive connectivity within and between groups in order to establish and maintain the transcriptional and nuclear architecture. These groups include transcription factors (TFs), RNA processing and nuclear proteins (RPs), nuclear transport (import/export) proteins (NTs), nucleosome remodelers (NRs), histone modification (e.g. acetylation) states (HSs), and histone modifying proteins (HMs).

Chromatin-immunoprecipitation experiments in combination with microarrays (termed ChIP-chip) have mapped the genomic occupancy of the aforementioned protein classes in living cells. Genome-wide identification of binding sites has allowed for the inference of which genes are regulated by such factors. For example, various TFs, HMs, and NRs have been shown to regulate specific gene expression programs (Lee et al, 2002; Ng et al, 2002; Robyr et al, 2002; Bar-Joseph et al, 2003; Robert et al, 2004; Gelbart et al, 2005). However, a unified model that incorporates all these different levels of transcriptional control remains undefined. Ideally, integrating ChIP-chip data from different labs can predict novel connections between individual regulators while providing a systems-level description of the greater transcriptional architecture. However, achieving such a model is hindered by the variability in microarray technologies and statistical analyses used by different labs.

Here, we combined and normalized ChIP-chip data for 317 regulators to gain an integrative view of the genome-wide interplay between different regulatory groups in Saccharomyces cerevisiae. As labs used disparate microarray technologies, we integrated the heterogeneous data sets using an assignment algorithm that maps each ChIP-chip measurement to its pertinent annotated gene. In addition, we developed a general method for standardizing the level of binding to P-values in order to normalize the data.

With the normalized data, we inferred biological relationships between any two regulators based on similar co-occupancy in the genome. Using communication theory, we identified all significant co-occupancy relationships between different protein groups and built a transcriptional network (Figure 2). In the process, we introduced mutual information, filtered correlation, and semi-supervised clustering approaches for analyzing genome-wide binding data. Our resulting transcriptional network identified over 340 known biological relationships, including associations within and between protein complexes studied in different labs (Figure 2). Moreover, previously detected protein–protein interactions also confirmed a significant portion of our predictions.

Figure 2

Resulting network. (A) Each circle (node) represents a regulator from a color-coded group and each link (edge) represents a significant synergistic (positive) binding relationship between two factors (see Supplementary Figure 1 for opposing (negative) links). Each node is labeled by the regulator's common name followed by an 'i' or 'o' if its genomic occupancy was measured at intergenic or ORF regions, respectively. The dotted box shows factors that have a preference for binding active genes. (B–G) Zoom in on several known interactions highlighted in solid boxes (see text). Network visualization was performed using Pajek (http://vlado.fmf.uni-lj.si/pub/netw
orks/pajek/doc/pajekman.htm).

Full figure and legend (987K)Figures & Tables index

Our integrative approach also uncovered novel biological phenomena, including unexpected connections between actively transcribed genes and silencing proteins Sir2 and Esc1 (Figure 2B). We validated these associations using ChIP-chip and quantitative PCR experiments. Sir2 and Esc1 are also known to localize to the nuclear periphery, suggesting a coupling between silencing and nuclear transport factors. The experiments demonstrate that our network predictions represent an in silico screen for discovering new biological processes.

We also analyzed the topology of the network to gain an integrative, systems-level description of the eukaryotic transcriptional architecture. We calculated several standard and new network statistics that describe the connectivity of each protein group. Our analysis formally shows that NTs, RPs, and NRs act as modular units that mediate general functions for large numbers of transcripts, whereas TFs and HMs are the specialists that provide gene target specificity.

We found that sequential in silico removal of network regulators, analogous to in vivo biological deletions, caused the connectivity between TFs to disintegrate more rapidly in the TF subnetwork than in the overall network (Figure 5B and C). A disconnected regulator cannot exchange information with the rest of the network, which may lead to a cellular malfunction; hence, interconnectedness between different regulatory groups made the overall network more robust to sequential deletions. Further, we found that proteins preferred to form modular subunits within their own class and communicate with other regulatory groups in a more selective manner. We conclude that modularity within a protein class helps localize the deleterious effect of a dysfunctional regulator to its class; for example, sequential removals of over 70 TFs had negligible effects on the connectivity within other protein groups (Figure 5B and C).

Figure 5

Network analysis. (A) Topology measures (first column, see text for definitions) for each color-coded level, active, and inactive regulators (first row). (B, C) Network robustness. Sequential attacks against TFs cause the characteristic path length to rise (B) as connectivity decreases (C) until the network reaches a breakdown point (peaks in panel B). TF attacks cause the connectivity of the TF subnetwork to disintegrate more rapidly than the overall transcriptional network, without affecting communication within other regulatory levels (flat lines). (D) Network adaptivity. In contrast to inactive genes, increased communication between regulators at active genes expedites the propagation of information (compare average neighboring levels and characteristic path length of last 2 columns in panel A) and may improve the speed and redundancy of the cell's response to dynamic environmental conditions. Our unified model connects several factors individually implicated in active gene expression (Lieb et al, 2001; Casolari et al, 2004; Robert et al, 2004).

Full figure and legend (191K)Figures & Tables index

Additional characterization of the network topology revealed that a significant proportion of the hubs that link levels of the transcriptional architecture are comprised of essential proteins. We found that 73% of the essential proteins linked two or more levels and 50% connected four or more levels, compared to 42 and 14% of non-essential proteins, respectively. For example, the essential TF Rap1 has been implicated in the recruitment of regulators from several functional groups to active genes (Lieb et al, 2001; Bernstein et al, 2004; Casolari et al, 2004).

Furthermore, regulators that preferentially bound to either active or inactive genes had opposing topological characteristics that significantly differed from the rest of the regulators in our network. We observed that control of inactive genes depends on disconnected, class-specific regulators (Figure 5A). Upon induction, highly interconnected regulators from all six groups converge onto genes that require activation (Figure 5A). These results suggest a model whereby increased communication between regulatory groups at active genes may improve the adaptivity and redundancy of the cell's response to changing conditions (Figure 5D).

Our work shows that integrating ChIP-chip data can provide new insight into the eukaryotic transcriptional architecture as a whole while also predicting novel associations between individual protein regulators. As more genome-wide localization data become available, we believe that the statistical methodology presented here can be extended to mammalian cells and other complex organisms.

MORE ARTICLES LIKE THIS

These links to content published by NPG are automatically generated.