Reconstructing dynamic regulatory maps

Jason Ernst¹, Oded Vainas², Christopher T Harbison^3,a, Itamar Simon² & Ziv Bar-Joseph^1,4

^aPresent address: Bristol-Myers Squibb Pharmaceutical Research Institute, Princeton, NJ 08543, USA

Article highlights

Synopsis

Understanding the dynamic programs that a cell utilizes in response to internal or external stimuli is an important challenge. These programs activate regulatory networks controlled by several transcription factors (TFs) (Harbison et al, 2004) and can involve a large number of genes (Natarajan et al, 2001). Direct information about this process has been obtained from genome-wide chromatin immunoprecipitation (ChIP-chip) experiments and comparative motif studies that have been carried out to identify some of the regulators involved (Hahn et al, 2004; Harbison et al, 2004; Xie et al, 2005; Workman et al, 2006). Time-series microarray expression experiments identified hundreds of genes that are activated or repressed following stress (Gasch et al, 2000). There has been a lot of recent interest in combining these and other data sources for inferring aspects of static regulatory networks (Figure 2C), but limited previous work on dynamic aspects. Here, we present a computational method for inferring the temporal activation mechanism that drives the observed expression response. Our method uses an input–output hidden Markov model (Bengio and Frasconi, 1995) to represent a temporal mapping from ChIP-chip and motif data to observe temporal expression profiles. This is achieved by focusing on bifurcation events. These events occur when sets of genes that have roughly the same expression level up until some time point diverge. The goal of the method is to detect these bifurcation events and explain them in terms of regulation by TFs. Based on these events and the explanation assigned to them, our method produces a map of the dynamic regulation that is taking place during the activation of a cell's response process. Genes are assigned to paths in the maps based on their expression profiles and the TFs that control them. This process results in temporal treemaps where each gene is assigned to a specific path in the tree (Figure 1). Following this assignment, we compute association scores for TFs with bifurcation events. In addition, the maps allow us to determine which TFs are controlling the response observed, when they are activating the genes they regulate, and the cascade of regulatory events that are responsible for the observed expression profiles.

Figure 1

Model overview. (A) Plots of time series expression profiles generated to illustrate the model. (B) Static TF-DNA binding data—DREM integrates TF-gene regulatory relationships derived from ChIP-chip or motif data with the time series expression data. For this example a majority of the pink genes in (A) are regulated by TF A, the blue genes by TF B and the red genes by TF C and D. (C) The model structure inferred by DREM for the data in (A) and (B). After the model is derived genes are assigned to their most likely paths based on their expression profile as well as on the set of TFs that regulate them. TF labels appear on some of the paths out of splits. (D) IOHMM model—each state has a Gaussian emission distribution for the expression values and the transition probabilities for a gene depend on the set of TFs that regulates it. A logistic regression classifier (Krishnapuram et al, 2005) maps the set of regulating TFs to transition probabilities. The classifiers are denoted by question marks in the figure. Example transition probabilities are given for a gene which is regulated by TF B. These probabilities are greater for the states with distributions similar to those of TF B regulated genes. The TF information also affects the structure of the resulting IOHMM model. Based on this information some splits can be added and some splits are removed from the model.

Full figure and legend (200K)Figures & Tables index

We initially focused on amino acid (AA) starvation response. Figure 2A presents a map of the dynamic AA starvation response inferred by our method from time-series expression data (Gasch et al, 2000) and 34 TFs profiled with ChIP-chip experiments in condition-specific binding data (Harbison et al, 2004). The map correctly identified known aspects of the response pathway including activation of genes by master regulators such as Gcn4 and Cbf1 and secondary activators such as Met32 and Arg81, among others. A repressed pathway associated with ribosomal TFs Fhl1, Rap1, and Sfp1 was found in AA starvation and was common across multiple conditions analyzed. We next augmented the set of TFs with 75 TFs that were profiled with ChIP-chip experiments in YPD media, but not AA starvation. As Figure 2B shows, using these additional TFs, we were able to explain a number of additional regulatory events that could not be explained with the original set of TFs. This indicates that a number of other TFs not known to be involved in regulating this process are taking part in the activation of the AA response pathway. For example, the new map predicted that Ino4, which was not profiled under AA starvation conditions, has a key role in regulating genes activated 4 h after the response. To test this, we performed a new genome-wide binding experiment for Ino4 under 4 h after AA starvation and observed a significant increase in the set of genes it binds compared to normal conditions. We have also experimentally validated another prediction for the role of Gcn4 in controlling yeast response to methyl-methanesulfonate stress.

Figure 2

Dynamic regulatory map and static network for yeast response to AA starvation. (A) Dynamic map of yeast response to AA starvation using static input from condition-specific binding experiments and time-series expression data. TFs with split score below 0.001 appear next to the split they regulate, in ranked order of scores. Nodes in the graph represent hidden states. The area of a node is proportional to the s.d. of the expression of the genes assigned to that node. Green nodes represent split nodes. Many of the TFs were correctly assigned to the time points they are known to regulate. For example, Gcn4, which is a known master regulator of AA starvation response, is correctly assigned to the first split. Many of the TFs assigned to the second split regulate specific AA biosynthesis pathways. (B) Dynamic map of yeast response to AA starvation using input from both condition- and non-condition-specific ChIP-chip experiments. Several additional TFs not profiled with a condition-specific ChIP-chip experiment under the AA condition were determined to be participating in the response and recovery processes. These included Abf1, Swi4, Mbp1, and Ino4. In addition to identifying these TFs as potential participants in the response, DREM also identifies their time of influence. (C) Static regulatory graph for AA starvation. Nodes correspond to genes or TFs. An edge implies that the TF binds the gene with a P-value <0.005 in an AA starvation ChIP-chip experiment. Blue edges represent interactions between TFs. Whereas some properties of the networks can be derived from the static representation, many of the dynamic aspects of the system are lost when not using the time-series data.

Full figure and legend (352K)Figures & Tables index

Next, we applied our method to study several stress responses in yeast. Similar to our results for AA starvation, the method was able to correctly reconstruct maps containing many known aspects of the temporal responses for these systems. Comparing the different stress experiments, we identified a number of common pathways and control mechanisms. Some of these common pathways were experimentally validated, leading to better understanding of the mechanisms controlling cell cycle in response to stress. By using the time of activation that our method assigned to TFs, we were able to identify cascades of activators. Analysis of these cascades provides insights into the utilization of network motifs and condition-specific regulation in response to stress.

MORE ARTICLES LIKE THIS

These links to content published by NPG are automatically generated.