A quantitative model of the Hog1-Msn2/4 Network

To examine the interaction between Hog1 and Msn2/4 in detail, we used the mutant cycle approach (Fig. 2a) to determine the value of the three expression components in the system: H, M and Co. Expression was examined after 20 min of stress treatment (0.4 M KCl) since this is near the peak of the transient response¹⁰ but is early enough to avoid monitoring secondary effects in the mutant strains (Hog1 and Msn2/4 are inactive in prestress conditions; Table S1). We find that the influence and interaction of Hog1 and Msn2/4 varies dramatically from gene to gene (Fig. 2b); we observe a total of eight distinct regulatory modes based on the combination of statistically significant expression components at genes induced in osmotic stress (Fig. 2c). From these data it is clear that: (i) Hog1 and Msn2/4 interact, since 190 of the 273 genes in the network have a statistically significant Co component (Groups 1, 2, 5, 7, 8; Fig. 2c); and (ii) that both Hog1 and Msn2/4 are activated, and can act, separately since significant H and M components are found at 112 (Groups 4–8; Fig. 2c) and 64 genes (Groups 2, 3, 6–8; Fig. 2c), respectively.

It is not possible to translate these data directly into a mechanistic network wiring diagram since the cooperative interaction between Hog1 and Msn2/4 could be established at either the promoter (Hog1 and Msn2/4 interacting on the promoter) or signaling level (Msn2/4 activity being regulated by Hog1) (Cooperative Mechanism, Fig. 1a). We surmised that the interaction between Hog1 and Msn2/4 is likely to be established, at least in part, at the signaling level, since Hog1 is a protein kinase and is required for full expression of almost all Msn2/4-dependent genes (190/203; Groups 1, 2, 5–7; Fig. 2c). Therefore, to test for activation of Msn2/4 by Hog1, we monitored the stress-induced import of Msn2/4 into the nucleus in wild-type and hog1Δ mutant strains containing GFP-tagged Msn2 or Msn4 and a nuclear marker. We observe that Hog1 is activated in KCl stress (black points, Fig. 3a) and that it contributes to activation of Msn2/4 (compare nuclear Msn2 levels in the wt and hog1Δ strains, Fig. 3a). However, Msn2/4 must also be activated by another pathway since some Msn2 is imported into the nucleus (in response to stress) even in the absence of Hog1 (Fig. 3a, Msn2-GFP in hog1Δ).

Given these connections at the signaling level, the data from the Hog1-Msn2/4 mutant cycle (Fig. 2c) can be explained by a simple wiring diagram (Fig. 3b) in which the Co component is assigned to Hog1-dependent gene activation through Msn2/4 while the H and M components are due to direct activation by Hog1 and Msn2/4, respectively. Hog1 could activate Msn2/4 through phosphorylation at one or more of 10/11 MAPK consensus sites found in these proteins, or indirectly through the other kinases, phosphatases and 14-3-3 proteins that regulate Msn2/4 nuclear import and export^16–18.

Our Hog1-Msn2/4 network model defines only three classes of genes (Fig. 3b): (I) genes regulated by Hog1 alone; (II) genes regulated primarily by Hog1 through Msn2/4 (3 genes by Msn2/4 only); and (III) genes regulated by Hog1 both through Msn2/4 and independently of Msn2/4 (mixed regulation). However, the genes in Classes II (Groups 1–3) and III (Groups 5–8) show distinct behavior in deletion mutants, resulting in several groups in the expression component analysis (Fig. 2c). This can be explained if different groups of genes within a given class have different thresholds for gene activation by Msn2/4: high, low or intermediate. For example, genes in Groups 1 (Co) and 5 (H+Co) appear to have a high threshold for activation by Msn2/4 as they are insensitive to the low levels of nuclear Msn2/4 found in the absence of Hog1 (Fig. 2c; no M component). In contrast, genes in Groups 3 (M) and 6 (H+M) appear to have a low threshold for activation by Msn2/4 as they are fully activated at the low levels of nuclear Msn2/4 found in the absence of Hog1 (Fig. 2c; M but no Co component). Finally, genes in Groups 2 (M+Co) and 7 (H+M+Co) appear to have an intermediate threshold for activation as they are partially activated at the low nuclear level of Msn2/4 (Fig. 2c; M and Co component) (see Supplement for explanation of all groups).

Incorporation of Sko1 and Hot1 into the Network Model

To explain how Hog1 activates genes independently of Msn2/4 (112 genes with an H component, Groups I and III Fig. 3b) we used microarray analysis to test the role of all five TFs known or suspected to be activated by Hog1 (Sko1, Hot1, Msn1, Smp1 and Cin5^19–22). To our surprise, we find that only two TFs, Sko1 and Hot1, have a significant effect on osmotic stress dependent gene expression (Table S1), and that Sko1 activates many more genes (40 at >2-fold induction) than previously^23,24 appreciated (Fig. S4).

To incorporate these factors into the network model we used the mutant cycle approach to dissect the influence of, and interaction between, Sko1/Hot1 and Msn2/4 (Fig. 3c). We find a striking correlation between the Sko1/Hot1 component determined in this analysis and the H component determined in the Hog1-Msn2/4 mutant cycle analysis (R=0.90, Fig. 3d). Therefore, Msn2/4-independent gene induction by Hog1 occurs almost entirely through Sko1 and Hot1. In fact, by measuring the influence that Hog1 has on gene expression in the absence of Sko1, Hot1 and Msn2/4 (on a single array, Table S1), we find that Sko1, Hot1 and Msn2/4 are required for 88% of Hog1-dependent gene activation (calculated by comparing the sum of the fold induction by Hog1 in the absence of Sko1, Hot1 and Msn2/4 to that in the wild-type strain). Only 17 of the 273 genes regulated by the HOG pathway (red points, Fig. 3d) are activated >1.5-fold (p<0.05) by additional (unknown) Hog1-dependent TFs.

By analyzing the cooperative component from the Sko1/Hot1-Msn2/4 mutant cycle (Fig. 3c) we were also able to define the logic gates at individual promoters. We find that there are very few positively cooperative (AND) interactions between Sko1/Hot1 and Msn2/4 (i.e. few genes with a statistically significant positive cooperative component; 5 observed versus 2 false positives expected, at p<0.01, and 9 versus 9 at p<0.05), validating our assertion that positive Hog1-Msn2/4 cooperativity is established at the signaling level (i.e. Hog1 regulating Msn2/4 activity; Fig. 3b). Instead, we find that Sko1/Hot1 and Msn2/4 act redundantly (negative Co component, OR interactions) or through SUM gate logic (no Co component; the output is the log sum of the individual components) at the promoter level (see Supplement).

To complete our model, we dissected the influence of Sko1/Hot1 into individual expression components using two further mutant cycles (see Supplement and Table S2) and identified the Sko1 and Hot1 target genes using chromatin immunoprecipitation followed by microarray hybridization (ChIP-chip) analysis (Fig. 4a). These data revealed that 65–80% of the genes repressed by Sko1 (27 total), activated by Sko1 (52 total), or activated by Hot1 (15 total) are bound by the appropriate factor in the appropriate condition at p<0.05 (Fig. 4b and Supplement); these findings are further supported by our detailed analysis of regulatory motifs where we find that Sko1, Hot1 and Msn2/4 binding sites are highly enriched in the appropriate gene-sets (see Supplement). Finally, we find that over half of the Sko1 and Hot1 target sites identified through ChIP analysis are silent (<1.5-fold activation), and thus non-functional, in the conditions studied here (Fig. 4c, d and Supplement). These results highlight both the accuracy of our mutant cycle approach and the limitations of using ChIP-chip (alone) for identifying functional interactions within transcriptional networks.

Signal Integration in the Hog1 Network

Taken together, our data provide a detailed model of the Hog1 transcriptional network in KCl-induced osmotic stress (Fig. 5). Examination of this network reveals that the signals sent through Hog1 and the general stress (Msn2/4) pathways are integrated at two levels. At the signaling level, Hog1 and at least one additional pathway function together to activate Msn2/4 and trigger its nuclear import (Fig. 3b). At the promoter level, the signal transmitted through Hog1, via Sko1 and Hot1, combines with Msn2/4 at a subset of the general stress response genes (Fig. 5). Therefore, the Hog1-Msn2/4 transcriptional network appears to have evolved to create an osmotic stress response that is modulated by signals that regulate Msn2/4 (which could include the PKA, TOR, Snf1 and other pathways^16–18,25).

To test this prediction, we examined the Hog1-Msn2/4 network in an additional stress condition: hyper-osmotic stress caused by high glucose concentrations. Glucose is known to reduce Msn2/4 activity^16,25 and is biologically relevant, as high glucose levels (including levels similar to those tested here) are encountered by yeast when they grow on fruit²⁶. To simplify our analysis, the level of osmotic stress (total molar osmolarity) used in the glucose and KCl experiments was identical. Since the HOG pathway senses the level of osmotic stress (turgor pressure²⁷), we expected that Hog1 would be activated to a similar level in both KCl and glucose, but that Msn2/4 activation would be different in these two conditions.

We find that the HOG pathway activates fewer genes in glucose than in KCl (187 vs 367 at >1.5-fold). To identify the basis of this change, we applied the mutant cycle approach (Fig. 2a) to determine the value of the three expression components (H, M and Co) in glucose and compared the data to that from KCl stress for each gene (Figs. 6a–c). In agreement with our initial predictions, we find that in the absence of Msn2/4, Hog1 has a similar impact on gene expression in glucose and KCl stress (H component, Fig. 6a). By contrast, Msn2/4-dependent gene activation (M+Co components) is substantially decreased in glucose (Fig. 6b) and this decrease extends to Hog1-Msn2/4 dependent gene induction (Co component, Fig. 6c). In accord with these results, activation of Msn2/4 (monitored by nuclear localization) is decreased in glucose compared to KCl stress, while activation of Hog1 is identical in the two stress conditions (Fig. 6d). Thus, the osmotic stress response in high glucose is modulated, when compared to that in high salt, by inhibition of Msn2/4 activity (Fig. 6e, left vs. right panels). This leads to an overall decrease in the activation of the general stress response, and shifts the Hog1-dependent expression program towards genes regulated by Sko1 and Hot1.

S. cerevisiae Strains

The strains examined in this study were constructed in a W303 background, as described in the supplement, and are listed in Table S5.

Expression Microarrays

An overnight culture of yeast was used to inoculate a 1 L culture to an OD₆₀₀ of 0.05 in a 2 L conical flask shaking at 200 rpm at 30 °C. These cells were grown to an OD₆₀₀ between 0.55–0.60 and then 250 ml of cells were collected by filtration and frozen in liquid nitrogen. At this time 500 ml of YEPD containing 0.9375 M KCl (at 30 °C) was added to the culture, and then the cells were harvested (after 20min), again using filtration, and frozen. In each case, strains that were compared on a single two-color microarray were grown in parallel in the same batch of medium and treated with identical YEPD + KCl. RNA was then purified from the frozen cells, converted into cDNA using reverse transcription, labeled with Cy3 or Cy5 and examined using Agilent G4140A microarrays (see supplement for details).

Microscopy

Strains expressing a GFP-tagged protein and RFP-tagged Nhp6a (a nuclear marker), were grown in SD medium to an OD₆₀₀ of 0.1 at 30 °C. These tubes were then transferred to a roller-drum in the microscope room (23 °C) for approx 1 hr. 50 µl of cells were then added to a well of a 96 well glass bottomed plate and allowed to settle for 5–10 min. At this time 30 µl of 1.0 M KCl in SD medium (or SD medium alone for the background control) were added to the cells and DIC and fluorescence images (in the eGFP and RFP channels) were collected in five separate fields using a Zeiss Axiovert inverted microscope fitted with a Cascade 512B camera and an oil-immersion Zeiss 63x achromatic objective. The nuclear fluorescence of each cell was then determined in both the GFP and RFP channels using Metamorph (version 7). The nuclear region was identified using the signal in the RFP channel and overlaid onto the GFP image. The nuclear fluorescence within these regions was then calculated for each cell, and averaged. Data was only recorded for cells that were free from overlap in the DIC image and had their nuclei in the focal plane (based on a cutoff for low RFP signal intensity), usually 100–200 cells per time point. The values reported are the average and standard deviation from three separate experiments. Sample images are shown in Table S4.

Chromatin immunoprecipitation and read out on microarrays (ChIP-chip)

Cells with HA-tagged Sko1 or TAP-tagged Hot1 were grown to OD₆₀₀ of 0.6 in YEPD as described for the expression arrays, and then treated with YEPD + KCl (to 0.375 M final) or YEPD alone. Five minutes after the application of stress, cells were treated with 1% formaldehyde for 15 min at room temperature. Cross-linking was then stopped by the addition of 125 mM glycine to the culture and the cells were washed twice with PBS at 4 °C and harvested by centrifugation. Cells were lysed by bead beating as described in (33) and the chromatin sheared using a Missonex 3000 sonicator fitted with a microtip (5 × 15 sec, power setting 1.5). This led to an average fragment size of 500–1000 bp. The DNA crosslinked to the TF was then immunopreciptated using 12CA5 and protein G Fastflow Sephadex (Pharmacia) for Sko1-HA, or IgG magnetic beads (Dynal) for Hot1-TAP, and purified as described in (33). These samples were then amplified, in parallel with the original sonicated DNA from the same strains (as a genomic control), using random priming PCR³⁴ with amino allyl-UTP in the mix, as described in (35). Immunoprecipitated samples were then labeled with Cy5, and genomic DNA labeled with Cy3, as described for the expression arrays. 2 µg of a Cy5-labeled sample and 2 µg of the appropriate Cy3-labeled genomic control were then hybridized to a custom Agilent microarray with 44,000 60bp probes (see supplement for a description). These arrays were then washed and scanned using the procedure described for the expression arrays. Similar procedures were also carried out for Msn2 (tagged with HA or TAP), but here we were unable to detect significant binding by real-time PCR even at well-established target genes (including CTT1 and HSP12). Inspection of previous ChIP data for Msn2 and Msn4 reveals that only a small subset of the known target genes for these factors are enriched by ChIP⁴, suggesting that the problems arise from a property of the TFs themselves.

Expression Component Analysis

As described in detail in the supplement, the system of equations listed in Fig. 2a can be formulated as the following matrix multiplication:

or Y = X * β + ε, where Y are the measured values from each microarray, X is the design matrix, β is the contribution of the three expression components, and ε is the noise. For each gene, we wish to find a β which minimizes the errors, ε.

To solve this linear model, we applied a multiple linear regression algorithm which minimizes the least squares fit of X*β, assuming a zero-mean Normal distribution of the errors ε. Specifically, the equation above X * β = Y is multiplied (from the left) by X^T, to get: X^T * X * β = X^T * Y. In our case, the matrix X^T *X is non-singular, and so we invert X^T *X and use it to multiply the equation (from left), and obtain a unique solution for the vector of regression coefficient β = (X^T * X)⁻¹ * X^T * Y.

It is assumed that all the coefficients in β have a zero-centered normal distribution, and so we can estimate their variance and covariance values. Specifically, Cov(β) = σ² * (X^T * X)⁻¹, where σ² is the variance of the fit. As described in the supplement, these properties pave the way for testing hypotheses about the estimated values of regression coefficients β. A similar approach was used to analyze the other mutant cycles in this study (see supplement).