Open Access

Revealing a signaling role of phytosphingosine‐1‐phosphate in yeast

L Ashley Cowart, Matthew Shotwell, Mitchell L Worley, Adam J Richards, David J Montefusco, Yusuf A Hannun, Xinghua Lu

Author Affiliations

  1. L Ashley Cowart1,3,
  2. Matthew Shotwell2,
  3. Mitchell L Worley1,
  4. Adam J Richards1,
  5. David J Montefusco1,
  6. Yusuf A Hannun*,1 and
  7. Xinghua Lu*,1
  1. 1 Department of Biochemistry and Molecular Biology, Medical University of South Carolina, Charleston, SC, USA
  2. 2 Department of Medicine, Medical University of South Carolina, Charleston, SC, USA
  3. 3 Ralph H. Johnson Veteran's Affairs Medical Center, Charleston, SC, USA
  1. *Corresponding authors. Department of Biochemistry and Molecular Biology, 174 Ashley Avenue, Charleston, SC 29425, USA. Tel.: +1 843 876 1111; Fax: +1 843 876 1126; E-mail: lux{at} or Tel.: +1 843 792 9318; Fax: +1 843 792 4322; E-mail: hannun{at}


Sphingolipids including sphingosine‐1‐phosphate and ceramide participate in numerous cell programs through signaling mechanisms. This class of lipids has important functions in stress responses; however, determining which sphingolipid mediates specific events has remained encumbered by the numerous metabolic interconnections of sphingolipids, such that modulating a specific lipid of interest through manipulating metabolic enzymes causes ‘ripple effects’, which change levels of many other lipids. Here, we develop a method of integrative analysis for genomic, transcriptomic, and lipidomic data to address this previously intractable problem. This method revealed a specific signaling role for phytosphingosine‐1‐phosphate, a lipid with no previously defined specific function in yeast, in regulating genes required for mitochondrial respiration through the HAP complex transcription factor. This approach could be applied to extract meaningful biological information from a similar experimental design that produces multiple sets of high‐throughput data.

Visual Overview


In contemporary biomedical research, gene mutation remains the most powerful and commonly used tool in molecular and systems biology for perturbation and dissection of biological systems. However, as biological systems consist of highly connected networks, for example, metabolic networks or signal transduction networks, perturbing one portion could result in widely spread effects across the network. Such ‘ripple effects’ in systems pose a challenge to the paradigm of investigating the role of a metabolite through mutating enzymes required for its production. In this study, we have developed a systems biology approach that integrates different types of ‘‐omics’ data to identify signal transduction pathways involving spingolipids and gene expression. See Figure 1 for an overall scheme of our approaches.

Sphingolipids are a family of bioactive lipids that have important signaling functions in cells; in yeast, de novo synthesis is required to mediate the cell response to heat shock. We hypothesized that a specific sphingolipid, phyto‐sphingosine‐1‐phosphate (PHS1P), functions as a signaling molecule in the heat stress response (HSR) because, though its mammalian counterparts are known to have important signaling roles, the function of this metabolite in yeast remains unknown. To identify a putative role of PHS1P in the HSR, we deleted the genes involved in production (LCB4 and LCB5) and degradation (DPL1) of PHS1P to perturb its levels in cells. In wild‐type cells, heat shock induces a significant increase in PHS1P. Over the same course, expression of over a thousand genes was modulated.

While deleting the genes involved in PHS1P metabolism ‘clamped’ the PHS1P concentration as expected, these mutations also resulted in wide spread changes in many sphingolipids in addition to PHS1P. This ‘ripple effect’ prevented direct identification of signaling role of PHS1P in gene expression. We overcame this difficulty by using a set of systems approaches as follows: (1) identifying the information between levels of each individual sphingolipid species and gene expression through combining correlation analysis and clustering; (2) identifying the putative PHS1P‐sensitive subset of genes by analyzing the results from step 1; (3) identifying transcription factors (TFs) that potentially regulate these PHS1P‐sensitive genes thought promoter analysis; (4) modeling the activation states of the TFs by combining gene expression data and promoter sequence data; and finally, (5) modeling the relationship between sphingolipids and activation of TFs.

Our study showed that 441 genes were differentially expressed in the lcb4Δ/lcb5Δ strain in comparison to wild‐type strain; however, only 77 genes among them showed a significant correlation with respect to PHS1P, with 22 genes positively correlated and 54 genes negatively correlated. The results led to a hypothesis that the genes showing significant correlation were PHS1P sensitive whereas differential expression of other genes resulted from the compounding ‘ripple effects’ of the gene deletions. We tested this hypothesis by directly treating cells with PHS1P and monitoring the expression levels of the genes that were PHS1P sensitive and PHS1P insensitive, and the results showed that the expression of PHS1P‐sensitive genes indeed changed in response to the treatment whereas others did not. We developed a statistical model referred to as Bayesian transcription factor state model to infer activation states of TFs in cells under a specific condition based on the genomic information and gene expression data. We then used a Bayesian logistic regression to further model the relationship between the lipid concentrations and activation states of the TFs. Combined TF enrichment analysis and TF state modeling indicated that the HAP TF complex was likely responding to the signal from PHS1P and mediating the regulation of PHS1P‐sensitive genes. We tested this hypothesis by treating wild type and a strain of yeast with deletion of HAP4 gene (hap4Δ), a component of the HAP complex, with PHS1P and monitoring the expression of PHS1P‐sensitive genes. Indeed, the PHS1P induced the genes in the wild‐type strain but not in hap4Δ, thus indicating that induction of the PHS1P‐sensitive genes required a functioning HAP complex (see Figure 5 ).

In summary, our experiments demonstrated that, though gene mutation remains one of the most powerful tools to perturb biological systems, the high connectivity of biological systems poses a challenge for using this approach to identify signaling roles of bioactive metabolites. Here, we demonstrated combining the information from multiple types of ‘‐omics’ data using systems approaches, it is possible to circumvent these difficulties and reveal novel signal transduction pathways.

  • Perturbing metabolic systems of bioactive sphingolipids with genetic approach

  • Multiple types of “omics” data collected from the system

  • Systems approach for integrating multiple “omics” information

  • Predicting signal transduction information flow: lipid; TF activation; gene expression


Sphingolipids, a class of lipids found in all cell types across eukaryotic species, include bioactive molecules such as sphingosine‐1‐phosphate and ceramide (Zheng et al, 2006; Hannun and Obeid, 2008b). In budding yeast, the synthesis of sphingolipids increases acutely on heat stress and mediates cell survival at high temperature (Dickson et al, 1997; Jenkins et al, 1997). Several sphingolipid‐dependent subprograms of the cell response to heat stress have been identified including cell cycle arrest, regulation of protein synthesis and degradation, and transcriptional reprogramming (Chung et al, 2000; Jenkins and Hannun, 2001; Cowart et al, 2003; Cowart and Hannun, 2005; Meier et al, 2006); however, these findings derive largely from studies wherein all sphingolipid synthesis is blocked, and thus, biological roles for specific lipid species remain unknown. Although specific genetic manipulations can be designed to determine the effects of deletion or over‐expression of enzymes of sphingolipid metabolism, the interconnectedness of the sphingolipid metabolic network (as in many metabolic pathways) leads to widespread changes in lipid levels on single enzyme mutation, thus preventing attribution of the observed effects to specific lipid species. Here, we illustrate this problem in the yeast heat stress response. We then present a systems biology approach designed to reveal potential lipid‐specific signaling pathways by deconvoluting a body of heterogeneous biological information derived from genomic, transcriptomic, lipidomic, and functional annotation data (Figure 1). The analysis indicates that phytosphingosine‐1‐phosphate (PHS1P), a sphingolipid with no previously known biological function in Saccharomyces cerevisiae, regulates the expression of genes involved in cellular respiration in a manner that requires the HAP2/3/4/5 transcription factor (TF) complex (Bonander et al, 2008). Biological validation of these findings indicated the systems analysis successfully identified a biological pathway mediated by PHS1P. Importantly, the methods and approaches developed in this study may be applicable to other metabolomic and genomic studies that generate high‐throughput data sets across multiple platforms and where the interest is in the function of specific metabolites.

Figure 1.

Overview of the integrative systems approach. The lipidomics, transcriptomic, and genomic data were collected from experiments and databases; (A, B, C) example data points or data matrix. Integrating the matching lipidomic and transcriptomic data in a correlation analysis lead to a gene‐versuslipid correlation coefficient matrix shown as a heat map shown in (D). Genomic and transcriptomic data were combined to infer the activation states of TFs under each experiment, shown as a TF‐versus‐condition heat map representing the activation states in (E). The inferred activation sates of TFs from (E) were combined with lipidomic data (A) to model the relationship between lipid mass and activation of TFs, shown as a heat map representing the significant logistic parameters in (F). The results from (E) and (F) resulted in the hypothesis that PHS1P mediated regulation of a subset of genes through activation of the HAP complex, which was tested in a series of genetic and pharmacological experiments (G).


Sphingolipid metabolism

The initial metabolites resulting from sphingolipid biosynthesis include the sphingoid bases dihydrosphingosine (DHS) and phytosphingosine (PHS), which in turn serve as metabolic precursors for synthesis of an array of chemically diverse species including ceramides and sphingoid base phosphates (Figure 2). Though the homologous mammalian lipid sphingosine‐1‐phosphate activates a variety of signaling events through receptor‐mediated modulation of characterized signaling pathways (Alvarez et al, 2007), specific cell functions for sphingoid base phosphates in yeast remain unknown. PHS1P was shown earlier to transiently increase during heat stress (Skrzypek et al, 1999), suggesting a potential role for this lipid in the heat stress response. Moreover, mutant yeast strains that accumulate PHS1P exhibited poor growth and heat stress resistance (Skrzypek et al, 1999; Kim et al, 2000), though a specific role for this lipid in the heat stress response remains unidentified. We hypothesized that PHS1P may mediate subprograms of the heat stress response, and, therefore, we designed experiments to perturb PHS1P metabolism using a conventional gene deletion approach. Deletion of LCB4 and LCB5 genes, encoding the yeast sphingoid base kinases (Nagiec et al, 1998), attenuates PHS1P production whereas deletion of DPL1, encoding the sphingoid base phosphate lyase (Saba et al, 1997), blocks its degradation and results in accumulation of PHS1P (Figure 2). Thus, mutating the above genes allows a ‘clamp’ on PHS1P in cells at either low or high levels, respectively. Log‐phase cultures of the lcb4Δ/lcb5Δ, the dpl1Δ, and wt strains were subjected to heat stress, and samples were collected at 5‐min intervals over a time course of 30 min in two duplicate experiments. To collect multiple ‘‐omics’ data reflecting cellular signaling systems, each sample was divided into two portions, thus allowing mRNA and sphingolipid extraction from each sample for transcriptomic and lipidomic analyses (Bielawski et al, 2006), generating a total of 42 transcriptomics measurements paired with 42 corresponding lipidomics measurements.

Figure 2.

Summary of major sphingolipid biosynthetic pathways in Saccharomyces cerevisiae.

Transcriptomic responses to heat stress

As previously reported (Gasch et al, 2000; Gasch and Werner‐Washburne, 2002; Cowart et al, 2003), heat stress significantly induced or repressed the expression of over a thousand genes in the wt strain, and a significant portion of this program was conserved in both the lcb4Δ/lcb5Δ and the dpl1Δ strains. However, we further identified the genes that are differentially expressed among the strains under comparable conditions. This group included 687 probe sets (corresponding to 441 genes with gene names) that were differentially expressed after heat stress in the lcb4Δ/lcb5Δ mutant when compared to wt strains. Analysis of this group using GOStat (Beissbarth and Speed, 2004) revealed a broad range of functional categories including DNA repair, translational regulation, post‐translational modification of proteins, cell wall organization and biogenesis, and others (Figure 3A).

Figure 3.

Effects of mutations in specific sphingolipid metabolic genes on gene expression and total sphingolipid profiles. (A) Overrepresented Gene Ontology annotations for genes aberrantly regulated during heat stress in the lcb4Δ/lcb5Δ mutant strain. (B) Heat map depicting changes in sphingolipid profiles over a time course of heat stress. Data are shown as a pseudo‐colored heat map reflecting the logarithms of lipid mass measurements normalized to total phospholipid content of the sample. Log values of normalized measurements are color coded as indicated in the scale to the left of the heat map. (C) A double‐sided clustering map depicting relationships between specific lipid–gene pairs over heat stress. A statistically significant (P‐value ⩽0.05 and q‐value <0.1) positive correlation coefficient between a gene and a lipid is shown as a red bar; a significant negative one is shown as a green bar; the value of correlation coefficient is pseudo‐color coded. In the map, rows represent genes, and columns represent lipids. The clustering tree on the left side of the map indicate gene clusters; a block across rows in the map represents a group of genes sharing similar information with respect to lipids; lipids with similar information with respect to gene expression (columns with similar color pattern) are grouped close to each other.

Lipidomics responses to heat stress

Recognizing the connectedness of sphingolipid metabolism (Alvarez‐Vasquez et al, 2005; Hannun and Obeid, 2008a), we hypothesized that the mutations might cause broad changes in sphingolipid metabolism and that only a subset of these genes spanning broad functional categories actually resulted from the lack of PHS1P, whereas many of these genes’ aberrant regulation might result from modulation of other sphingolipids in this mutant. Therefore, we further evaluated the lipidomic data to determine the impact of heat stress on lipid species and to determine whether the effects of mutations generate a metabolic ‘ripple effect’ leading to changes in other metabolites. Indeed, out of 40 distinct sphingolipid species measured, 28 demonstrated changes at least at one time point of heat stress in wild‐type cells (Figure 3B). Heat stress induced an increase in C18 PHS1P from nearly undetectable levels to 0.02 pmol/nmol phosphate (see Supplementary Table 1), peaking around 20 min and returning to basal levels by 30 min. These kinetics were similar to previously published data that demonstrated an eight‐fold increase in PHS1P that peaked around 10 min and returned to basal levels around 20 min (Skrzypek et al, 1999). As expected, these changes did not occur in the lcb4Δ/lcb5Δ mutant strain. On the other hand, the dpl1Δ strain showed constitutively elevated PHS1P, which further increased during heat stress (Figure 3B, bottom row). More importantly, the data revealed widespread differences in lipid profiles between the three strains, both basal and under heat stress. For example, the lcb4Δ/lcb5Δ strain demonstrated significant elevation of PHS, increased α‐hydroxylated ceramides of 20 and 22‐carbon N‐acyl chain length, increased phytoceramides of 24 carbon chain length, and decreased dihydroceramide of 26 carbon chain length. In fact, out of the 28 lipid species showing changes in the wild type over the time course, at least 19 species showed differences between wt and the lcb4Δ/lcb5Δ strain at least one time point.

Lipidomics profiles of the dpl1Δ strain also demonstrated widespread differences as compared to the parental strain (Figure 3B, right panel). With the exception of the expected accumulation of PHS1P, lipid measurements in this strain revealed fewer and more subtle variation from the wild‐type strain at basal temperature; however, time course measurements indicated that deletion of DPL1 significantly altered the heat stress sphingolipid response in that at least 20 of the 28 measured lipids exhibited differences from the parental strain in at least one point of the time course. Moreover, changes in this mutant were partially distinct from changes observed in the lcb4Δ/lcb5Δ mutant strain.

In summary, these mutations not only had the expected effects on PHS and PHS1P (substrate and product), but also caused widespread changes in many sphingolipid species. Therefore, any changes in gene regulation observed in the lcb4Δ/lcb5Δ or dpl1Δ mutant strains could not be readily attributed to PHS1P (or any single lipid species). To circumvent the limitations of ‘gene‐centric’ approach, an alternative approach was devised to integrate information from lipidomic and transcriptomic data in a manner that allows inferring lipid‐mediated events.

Indentify potential sphingolipid‐regulated genes

To identify and quantify the information connecting lipidomic and transcriptomic changes, we performed covariance analysis (DeGroot and Schervish, 2002) between all lipid‐probe pairs, which led to a genes‐versus‐lipids matrix in which an element contained a correlation coefficient (ρ) for a lipid–gene pair if the ρ is statistically significant or a zero otherwise. It was noted that a row of the matrix constituted a correlation (information) pattern demonstrated by a gene with respect to all lipids, and similarly a column encoded an information pattern demonstrated by a lipid with respect to all genes. We sought to identify the shared information patterns among genes and lipids by applying a double‐sided hierarchical clustering analysis, which grouped genes (and lipids) sharing similar information patterns into clusters. Selected results focusing on key sphingolipids are shown in Figure 3C as a heat map (also see Supplementary Table 2). The results indicate that lipid species demonstrated distinct correlation patterns with respect to modules of genes, suggesting potential regulatory roles of the lipids on the modules. The double‐sided clustering revealed clusters of genes sharing similar information with respect to lipid (blocks across rows) and clusters of lipids sharing similar information with respect to gene expression data (columns with similar correlation coefficient patterns). For example, DHS and PHS, two closely related metabolites in the metabolic network, were grouped together because of their shared information with respect to clusters of genes at the top region Figure 3C, but they also showed distinct information with other gene clusters.

Importantly, this procedure identified subsets of genes that were significantly correlated to PHS1P, with 44 positively correlated probe sets (mapped to 23 named genes) from the microarrays and 61 negatively correlated ones (mapped to named 54 genes), which was significantly less than the genes identified in the microarray analysis of the lcb4Δ/lcb5Δ mutant (441 genes). Among these 77 PHS1P‐sensitive genes, 40 genes were also deemed differentially expressed according to differential expression analysis, whereas the other 33 genes were not. The results indicate a tentative statistical advantage of correlation analysis over the differential expression analysis in that the former can use all samples (42 samples for a lipid‐versus‐gene pair) whereas the latter can only use samples from specific conditions (12 lcb4Δ/lcb5Δ microarrays versus 12 wt microarrays after heat stress).

We applied a method referred to as GO Steiner Tree (GOSteiner, 2009) to analyze the functional coherence of the gene sets and to visualize their functional relationships. The method represents the genes and their associated Gene Ontology terms as a graph, finds a subgraph (a Steiner tree) connecting all genes and their annotations with a shortest total functional semantic distance, and finally evaluates the statistical properties of the tree as metrics of functional coherence. Supplementary Figure 1 shows the GO Steiner tree for genes that demonstrated failure to induce on heat stress in the lcb4Δ/lcb5Δ strain when compared to wt. Although the analysis showed that the lcb4Δ/lcb5Δ‐sensitive genes had diverse functions, the visualization of the GO Steiner tree revealed clusters of genes with coherently related functions, including genes involved in mitochondrial metabolism, for example, COX4, INH1, and ATP17, and vesicular transport, for example, APS2, SVP26, TPM2, and TVP23. The results indicate that the deletion of LCB4 and LCB5 led to aberrant regulation of several distinct groups of genes, among which genes showing high correlation to PHS1P represented only a subset of those genes (highlighted in Supplementary Figure 1). Importantly, this subset included genes in specific sub‐categories (e.g. the mitochondrial metabolism), but not others (e.g. vesicular transport) of the LCB4/LCB5‐regulated genes.

The above results led to the hypothesis that, although both sets of genes were lcb4Δ/lcb5Δ sensitive, only the genes showing strong correlation to PHS1P levels were truly PHS1P sensitive, whereas dysregulation of other genes in this mutant strain could result from confounding changes in other sphingolipids which had become apparent from the lipidomics analysis (Figure 3B). We tested this hypothesis by treating wild type cells (wt) with exogenous PHS1P in at non‐heat stress temperature and monitored the expression of sample genes from the putative PHS1P‐dependent set involved in the mitochondrial metabolism (COX4, INH1, and ATP17) and a putative non‐PHS1P‐dependent set involved in vesicular transport (APS2, SAR1, SVP26, TPM2, and TVP23). Indeed, expression of the genes involved in vesicular transport failed to be induced by PHS1P (Figure 4A), whereas the mitochondrial metabolism genes demonstrated ∼2.4–3‐fold increases on treatment (Figure 4B). To further test whether the effects were specific for PHS1P, wt and lcb4Δ/lcb5Δ mutant cells were treated with PHS, the metabolic precursor of PHS1P, which undergoes conversion to PHS1P in the wild type but not in the lcb4Δ/lcb5Δ mutant. Indeed, in wild‐type cells, PHS produced a significant upregulation of the genes in the respiration set, which was totally lacking in the lcb4Δ/lcb5Δ mutant (Figure 4C). Thus, the integrated analysis enabled the identification of the PHS1P‐dependent subset of genes within the larger set of genes that showed failed regulation in the lcb4Δ/lcb5Δ mutant.

Figure 4.

PHS1P‐mediated gene expression regulation. (A) Treatment with PHS1P in the absence of heat stress did not induce all genes aberrantly regulated in the lcb4Δ/lcb5Δ strain. (B) Treatment with PHS1P in the absence of heat stress induces gene expression of the putative PHS1P‐dependent genes identified by the integromics analysis. (C) The metabolic precursor of PHS1P, PHS, upregulated PHS1P‐dependent genes in the wild‐type strain, but not in the lcb4Δ/lcb5Δ mutant, which cannot phosphorylate PHS. Experiments were performed two to three times in triplicate and represented as mean±s.e.m.

Identifying candidate TFs

Defining specific lipid‐mediated responses, beyond the overall gene‐mediated responses manifested by mutations, is important in that it allows studying the distinct roles and mechanisms of specific lipids—the putative functional mediators—that contribute to the overall responses. Highly selective PHS1P‐mediated expression of specific genes led to the hypothesis that PHS1P regulates a specific pathway by activating some downstream TFs, which then mediate the regulation of these genes in response to PHS1P. To identify such putative TFs, a transcription factor binding site (TFBS) matrix was constructed, in which an element contains a binary variable indicating if a gene (g) can be bound by a TF (t). AN element, btg, of the matrix is set to 1 if the analysis of the chromatin immunoprecipitation experiments (Lee et al, 2002; MacIsaac et al, 2006) indicates that the TF t is capable of binding to the promoter sequence of gene g, or if there is documentations of the interaction between the TF and the gene according to the yeast TF database, YEASTRACT (Monteiro et al, 2008). The knowledge of TFBSs enabled us to evaluate whether any TFBS was significantly enriched in the promoters of a gene set. Using a hypergeometric‐distribution‐based model, we assessed enrichment of TFBSs in the promoters of the genes that demonstrated positive correlation to PHS1P and identified 22 TFBSs as significantly ‘enriched’ (see Supplementary information). This led to the hypothesis that the activation states of some of the candidate TFs were sensitive to changes in PHS1P, thus transmitting the signal from PHS1P to genes. To test the hypothesis, we further developed a two‐stage Bayesian information integration model to reveal information flow from bioactive lipids → TFs → gene expression.

Inferring activation states of TFs

We developed a novel Bayesian model to infer the activation states of the TFs under each experimental condition. This model extends our previous published method (Lu et al, 2004) such that the genomic data of TFBSs can be integrated with transcriptomic data to infer the activation state of TFs. In this model, the activation state of a TF (t) under a specific condition (a) is represented as a binary variable, sat, such that sat=1 indicates the TF is active and sat=0 otherwise. Representing activation states of TFs as binary variables has two advantages: (1) it provides a more intuitive representation of activation (active/inactive) state of a TF, in comparison to possible negative activation state allowed by some other models (Lee and Batzoglou, 2003; Liao et al, 2003; Gao et al, 2004; Ochs et al, 2004; Battle et al, 2005; Sun et al, 2006); (2) it renders the mathematical convenience to model the relationship between lipids and activation states of TFs using logistical regression, in which the sigmoid function mimics dose–response curves commonly observed in signal transduction pathways.

The model specifies that the expression value of a gene from a specific experiment (ega) is influenced by three factors: (1) TFs that bind to its promoter, indicated by bgtt∈{1,…..,T}; (2) the states of each of the TFs under this specific condition, represented by satt∈{1,…..,T}; (3) and the strength and direction (induction or repression) of an activated TF on its expression, represented by wgtt∈{1,…..T}. We define the probabilistic relationship between the above parent variables and the gene expression value as follows:

Embedded Image

where ε and τ represent the noise of the system and N stands for the Gaussian distribution. It is of interest to note that the product of two binary variables, bgtsat, encodes a logic AND relationship between the two variables in the equation, such that the equation can be interpreted as follows: TF t influences the expression value of gene g under the condition a if and only if it has a binding site in the promoter of the gene (bgt=1) AND it is activated under the condition (sat=1). The equation also reflects coordinated influences of multiple activated TFs on a gene's expression in a linear form (usually in logarithmic scales), an assumption widely used in modeling of expression systems (Liao et al, 2003; Gao et al, 2004; Lu et al, 2004; Battle et al, 2005). Given TFBS matrix and expression data, the model probabilistically infers the state of each TF under a specific condition using a variational Bayes technique (Lu et al, 2004) (see Supplementary information for detailed description). Figure 5A shows that many TFs switch states during heat stress, among which many are well‐documented stress‐responding TFs.

Figure 5.

Modeling information flow from lipid, to TFs, and to gene expression. (A) Inferred TF activation states through integrating genomic and transcriptomic data. Red color indicates activated state and black denotes inactivate states. The TFs were grouped according to their state across experiment; the yellow block indicates a group of TFs ‘turned off’ after heat stress; the purple box outlines the TFs ‘turned on’ after heat stress. (B) Logistic regression modeling of the relationship between sphingolipids and TF states. Statistically significant regression parameters are shown as a TF‐versus‐lipid heat map. The orange box indicates the significant parameters associated with PHS1P with respect to Hap2p and Hap4p. (C) The ability of PHS1P treatment to induce PHS1P‐dependent genes in the absence of HAP4 was determined. The experiment was performed three times in triplicate and represented as mean±s.e.m.

Modeling the role of lipids in TF activation

On the basis of the estimated binary TF states from the previous section, a Bayesian logistic regression model was applied to investigate the relationship between the levels of sphingolipids and the putative TF activation states. In this model, the states of a TF under each experimental condition were modeled as a sigmoid function of the concentrations of sphingolipids, where the probability that a TF (t) is active, conditioning on observed lipid profiles (l1,,L) under an experiment condition (a), is defined as follows: Embedded Image. In this model, a parameter βlt reflects the strength and direction of influence that the sphingolipid l has on the activation state of the TF t. Using a Gibb's sampling‐based Bayesian logistic regression, we identified all statistically significant parameters, and we interpret a significant parameter βlt with respect to a TF t as an indication that the TF is regulated by lipid l, and the selected results are shown in Figure 5B. A total of 13 TFs showed with significant coefficient with respect to PHS1P, with 11 positively influenced and 2 negatively influenced. Among the positive TFs, Hap2p, Hap4p, NRG1 and MGA1 were also among the 22 TFs deemed significantly enriched in the PHS1P postively correlated gene set. Hap2p and Hap4p are members of the HAP complex (Chodosh et al, 1988; Buschlen et al, 2003), whose binding sites were dominant in the PHS1P postively correlated gene set, particularly in the subset involved in the cellular respiration. Thus, modeling of the relationship between sphingolipids and TF states, in combination with the static information inferred from promoter analysis, provided evidence at a mechanistic level that gave rises to the hypothesis that PHS1P regulates expression of respiratory genes through modulating the activity of the HAP complex.

To test this hypothesis, we evaluated the ability of PHS1P to regulate PHS1P‐dependent genes in a mutant strain deleted for the gene encoding Hap4p (hap4Δ cells), an essential component of the HAP complex. Wt cells or hap4Δ cells were treated with exogenous PHS1P added into the culture media. RNA was isolated from each culture, and gene expression was determined by real time RT–PCR. Indeed, as compared with the parental background strain, the hap4Δ strain demonstrated complete loss of PHS1P‐mediated induction of the target genes (Figure 5C). The result indicated that regulation of these genes by PHS1P required a functional HAP complex.


Although there are publications (Fischer, 2005; Hirai et al, 2005; Ippolito et al, 2005) that simultaneously analyze metabolomic and transcriptomic data, these studies mainly concentrate on the relationships between the expression levels of the enzymes and their metabolites. To our best knowledge, this study represents a novel approach to integrating multiple ‘‐omics’ data to infer signal transduction pathways involving bioactive lipids at a mechanistic level. By integrating information from heterogeneous types of data in the principled probabilistic framework, the approaches developed in this study overcame the difficulties associated with conventional gene deletion/overexpression experiments commonly used in studies to delineate signaling pathways. The problem that single gene mutation leads to system‐wide perturbation is likely to be a general case in many biological systems, particularly in metabolic networks; the interconnectedness of sphingolipid metabolism demonstrated in this study is likely an example rather than an exception of such systems. From a systems biology point of view, genetic manipulation of genes is a powerful tool to perturb systems, which provides opportunities to study the systems at mechanistic level but it is not necessarily sufficient to derive a causal relationship between metabolites and their effectors. By collecting and assimilating information at systems level, our study transcends the ‘gene‐centric’ framework, leading to the identification of the signaling role of PHS1P in cellular stress responses in yeast. Progressive analyses generated specific hypotheses at different mechanistic levels: (1) PHS1P specifically regulates the expression of a set of genes involved in cellular respiration; and (2) this regulation requires the HAP TF complex; both findings received support from experimental validation. The study demonstrates the utility of integromic approaches in studying cellular signaling systems. This should be of great value in the study of bioactive lipids and other metabolic pathways, where the need arises to dissect functions of specific metabolites.

Materials and methods

Yeast culture and treatment

Strains of S.cerevisiae used are listed in Supplementary information. Yeasts were routinely cultured in Yeast Proteose Dextrose media at 30°C. Working cultures were seeded from overnight 5 ml cultures of a single colony and allowed to grow with 200–250 r.p.m. shaking to mid‐logarithmic phase (OD=0.4–0.8). For sample treatment, cells were shifted to 39°C over a time course of 5, 10, 15, 20, 25, and 30 min. Cells were collected by centrifugation at 3500 g for 3 min, and pellets were snap‐frozen in an ethanol/dry ice bath. Frozen pellets were stored at −80°C. For treatment with exogenous compounds, cells were maintained at 30°C and compounds were added as solutions in dimethylsulphoxide (DMSO), or DMSO alone was added as a negative control. Treatments with exogenous compounds were for 15 min.

RNA preparation and microarray hybridization

RNA was prepared from snap‐frozen pellets using the RNeasy kit from Qiagen according to manufacturer's directions. Preparation of target for microarray hybridization to the Affymetrix YG‐S98 chip was performed according to manufacturer's instructions. Microarray hybridization was performed at the Microarray Core Facility at the Medical University of South Carolina. Microarray analyses were performed using R packages of Bioconductor suite (Bioconductor, Gentleman et al, 2004). Microarrays were normalized using the robust microarray averaging package, and differential expression was assessed using the linear model for microarray data package from Bioconductor; false discovery rate was assess with Q‐value package (Storey and Tibshirani, 2003). The threshold for differential expression and significant correlation coefficient was set at P‐value <0.05 and q‐value ⩽0.1. The microarray data set is publicly available at the Gene Expression Omnibus database (Barrett et al, 2009), with an accession number of GSE18121.

Lipidomics analysis

Lipids were extracted from snap‐frozen yeast pellets as described (Bielawski et al, 2006) and subjected to high‐throughput LC/MS analysis as described earlier (Bielawski et al, 2006). Quantification was based on comparison of peak intensity to internal standards as described earlier.

Real Time RT–PCR

Cells were grown in YPD media to mid‐logarithmic phase and treated with compounds dissolved in DMSO as indicated or vehicle alone for 15 min in a 30°C water bath with 200–250 r.p.m. shaking. Cells were collected by centrifugation at 3500 g for 3 min, decanted, and immediately frozen at −80°C. RNA was extracted from frozen pellets using the RNeasy kit from Qiagen. cDNA was prepared from RNA using Superscript II or Superscript III (Invitrogen) according to manufacturer's directions. cDNA was diluted 10–20‐fold before real‐time PCR using the SybrGreen solution and protocols (Bio‐Rad) and primers as indicated in the Supplementary information. Gene intensity signals were normalized to levels of RNA for either ribosomal 18S subunit or actin. Reactions were conducted in a Bio‐Rad iCycler.

Two‐staged Bayesian information flow model

Detailed mathematical and computational descriptions of the Bayesian latent variable model and Bayesian logistic regression are available as online Supplementary information at Molecular Systems Biology.

Conflict of Interest

The authors declare that they have no conflict of interest.

Supplementary Information

Supplementary Material

Supplementary materials, methods and results. [msb20103-sup-0001.doc]

Supplementary Table 1

A comma separated value (CSV) file contains the lipid measurements. The file contains the numeric values of lipids corresponding to Figure 3B. [msb20103-sup-0002.xls]

Supplementary Table 2

A CSV file contains the numeric values for the heatmap of Figure 3C [msb20103-sup-0003.xls]


We acknowledge the following grant supports: P20 RR017677‐07 (to LAC and XL). 1R01LM10144 to XL, 5R01GM63265 to YAH, the Department of Veterans’ Affairs Merit award to LAC, 5R01LM009153, 3T15LM07438 to XL and AJR, and 5T32GM074934 to MS. We thank Alan Wilder and Jason Gandy for technical assistance. We also acknowledge the MUSC Lipidomics Core Facility of the COBRE in Lipidomics and Pathobiology for sample analysis, the MUSC Proteogenomics Core Facility for microarray assays, the MUSC department of Art Services for assistance with figure preparation, and Dr Hiroko Hama for critical reading of the paper.


This is an open‐access article distributed under the terms of the Creative Commons Attribution License, which permits distribution, and reproduction in any medium, provided the original author and source are credited. This license does not permit commercial exploitation without specific permission.