A strategy to study pathway cross-talks of cells under repetitive exposure to stimuli
© Fu et al.; licensee BioMed Central Ltd. 2012
Published: 17 December 2012
Skip to main content
© Fu et al.; licensee BioMed Central Ltd. 2012
Published: 17 December 2012
Cells are subject to fluctuating and multiple stimuli in their natural environment. The signaling pathways often crosstalk to each other and give rise to complex nonlinear dynamics. Specifically repetitive exposure of a cell to a same stimulus sometime leads to augmented cellular responses. Examples are amplified proinflammatory responses of innate immune cells pretreated with a sub-threshold then a high dose of endotoxin or cytokine stimulation. This phenomenon, called priming effect in the literature, has important pathological and clinical significances.
In a previous study, we enumerated possible mechanisms for priming using a three-node network model. The analysis uncovered three mechanisms. Based on the results, in this work we developed a straightforward procedure to identify molecular candidates contributing to the priming effect and the corresponding mechanisms. The procedure involves time course measurements, e.g., gene expression levels, or protein activities under low, high, and low + high dose of stimulant, then computational analysis of the dynamics patterns, and identification of functional roles in the context of the regulatory network. We applied the procedure to a set of published microarray data on interferon-γ-mediated priming effect of human macrophages. The analysis identified a number of network motifs possibly contributing to Interferon-γ priming. A further detailed mathematical model analysis further reveals how combination of different mechanisms leads to the priming effect.
One may perform systematic screening using the proposed procedure combining with high throughput measurements, at both transcriptome and proteome levels. It is applicable to various priming phenomena.
A cell needs to constantly sense and response to various signals from both external and internal environments. The requirement on generating appropriate response to specific signals forces cells to develop a complex signaling network that often involves multiple highly intertwined signaling pathways [1–3]. It becomes increasingly clear that pathway cross-talks play critical roles in cellular signaling and decision making process . For example, cross-talks may increase the nonlinearity in the signaling network, resulting in various synergistic and antagonistic effects in cellular responses [5–8]. A nonlinear response refers to the cellular response to multiple different stimuli, or repetitive stimulus that is not simply the sum of responses to each individual stimulus. Cells in vivo are constantly exposed to a variety of stimulus with fluctuating concentration. Therefore it is of great importance to study how cells utilize complex pathway cross-talks to generate appropriate response or make correct decision to multiple or repetitive stimulus. Pharmaceutically, it is also a common treatment strategy to use combinations of multiple drugs simultaneously in order to generate synergistic effect [8, 9]. Therefore, the nonlinear phenomena due to pathway cross-talks have important physiological and clinical significances.
In the previous study, we applied a computational analysis to enumerate all possible network motifs that are able to induce priming effect in a generic three-node regulatory network. Strikingly, we found that the in silico discovered priming motifs naturally fall into three priming mechanisms. Based on the finding, the main purpose of this study is to design and apply a general combined experiment and computation strategy to search for molecular candidates contributing to the priming effect for a given stimulus. The remaining part of the paper is organized as follows. First we summarize the main results of our first study, and outline the strategy. Then we demonstrate how to apply the strategy to analyze a set of published microarray data on IFN-γ-mediated priming effect. Next we show further analysis on a detailed ordinary differential equation based model.
In the first paper , we enumerated all possible network structures and kinetics that are able to induce priming effect with a generic three-node model (Figure 1B). The three-node model represents the minimal abstraction of the two cross-talking pathways (e.g., MyD88-dependent and -independent branches of Toll-like receptor 4 (TLR4) signaling pathway). Each node in the model can either positively or negatively regulate the activity of the other nodes or itself. We simulated the dynamics with a set of nonlinear ordinary differential equations with 14 variable parameters. Through a two-stage Metropolis algorithm, we analyzed the dynamical behavior of over 1.5 × 105 different networks that can generate priming effect. Here we refer to priming effect as a set of dose-response behaviors: (1) A single low dose stimulant (LD) cannot activate the readout x3 (< 0.1 in a reduced unit with 1 the maximum induction). (2) A single high dose stimulant (HD) can activate x3. (3) Sequential stimulation with LD first followed by HD (LD+HD) can activate x3 to a maximum level that is at least 50% higher than that under HD alone.
As shown in Figure 1C, the parameter sets leading to priming effect clearly cluster into two regions, in terms of the change in the two regulators, x1 and x2, at the end of LD pretreatment ( , i = 1,2). Data in the left region locate approximately along the negative side of x-axis, that is, a LD pretreatment decreases x1 in this region (i.e., , with an arbitrarily chosen cutoff to account for possible experimental resolution). Notice x2 in this region spread out vertically, that is, x2 can either increase or decrease to some extent under LD pretreatment. Based on this observation, we want to find out any possible constraint on x2 in this region. To do this, we plotted the distribution of the difference between the maximum response of x2 under LD+HD and that under HD alone. We found that x2 from this region can be either HD-responsive or LD-responsive, but with a constraint that the maximum expression under LD+HD makes no difference with that under HD alone (i.e., ) [see Additional file 1]. On the other hand, the data in the right region demonstrate a significant increase in x2, but not x1, after LD pretreatment (Figure 1C) (i.e., ). The maximum expression of x1 under LD+HD makes no difference with that under HD alone (i.e., ) [see Additional file 1]. However, this overlapped region can be further separated into two sub-groups, pathway synergy (PS) and activator induction (AI), if plotted against another experimentally measurable quantity: the difference in the maximum level of x2 under LD+HD vs under HD (Figure 1D). It is obvious that the data from the red group, but not the green group, shows a significant increase in the maximum level of x2 under LD+HD compared to that under HD alone (i.e., ) (Figure 1D).
The physics underlying the three priming mechanisms turns out to be simple and beyond the current three-node model . For Pathway Synergy, both of the two pathways activate the priming readout x3, but one has a fast time scale and a high activation threshold while another one has a slow time scale and a low activation threshold. When given a single HD stimulation, the regulation on x3 from the two pathways is temporally separated. A LD pretreatment brings forward the slow pathway so that the two pathways can achieve a transient synergy to boost the production of x3 (Figure 2). Similarly, for Activator Induction and Suppressor Deactivation, a LD pretreatment separates the two originally temporally overlapping but antagonistic pathways by either advancing the activator or delaying the suppressor (Figure 2).
Record the time course of the cellular response under single LD, single HD, and LD+HD, respectively.
Identify the priming readout genes as those with higher response to LD+HD than HD, but with no significant response to LD.
Identify the genes induced or reduced by LD (LD-responsive genes), and those responding to HD only (HD-responsive genes).
Pathway Synergy: (1) LD-responsive genes (with the expression under LD+HD higher than that under HD alone) and (2) HD-responsive genes; (3) both activate a downstream readout gene.
Activator Induction: (1) LD-responsive genes (with the expression under LD+HD similar to that under HD alone) and (2) HD-responsive genes; (3) the LD-responsive gene activates while the HD-responsive gene inhibits a downstream readout gene.
Suppressor Activation: (1) LD-reduced genes and (2) LD/HD-responsive genes (with the expression under LD+HD similar to that under HD alone); (3) the LD-reduced gene inhibits while the LD/HD-responsive gene activates a downstream readout gene.
Furthermore, five genes (SLC2A3, ST3GAL5, DNAJB1, STAT1, UBE2S) are identified as possible priming readout genes (x3), which show negligible expression under LD, but considerable higher expression under LD+HD than under HD alone. However, among the five genes, only UBE2S shows a significant change between LD+HD and HD (by ≥ 2 fold) that passes t-test with p < 0.05. Considering microarray data are usually noisy, one needs more quantitative measurements, e.g., real time PCR to confirm these results. Here we used the experimentally confirmed molecular species, such as phosphorylated STAT1 dimmer, IRF-1 and IP-10 as the priming readout . After selecting and grouping genes based on the guideline in Figure 3, we then placed them in the context of regulatory networks in order to identify possible priming mechanism on the molecular interaction level. The regulatory network associated with these selected genes is constructed in IPA ® database (see Methods for details).
In our proposed strategy it is essential to examine the genes identified from the high throughput data in the context of the regulatory network. In many cases gene activities are correlated, e.g., due to a common upper stream regulator. As an illustrative example, suppose the activities of genes A and B are correlated and are both up-regulated by the low dose stimulant, but only A regulates the downstream readout gene C. Based on the absence of regulation from B to C in the regulatory network, one can only conclude that the existing experimental result suggests A, but not B, as a potential contributor to the priming of C. In another situation, if a molecular species (e.g., a transcription factor) shows priming effect, the priming effect may be transmitted to its downstream targets. The detailed model discussed later gives such an example.
To investigate how LD priming affects macrophage cellular functions, we conducted the ontology analysis of the genes that show significant fold change (≥ 2 fold, p < 0.05) after LD priming. Additional file 2 shows the clustering result, and lists the top 10 significantly enriched molecular functions found for LD IFN-γ induced and reduced genes, respectively. We found that in general, genes that are significantly increased by LD priming are associated to inflammatory response and immune system process; genes that are significantly decreased are associated to negative regulation of T cell mediated cytotoxicity and immunity. This result suggests that LD priming prepares macrophages for a stronger inflammatory response by elevating a number of proinflammatory genes and inhibiting some negative regulators, reflecting a cellular adaptivity of innate immune cells.
We want to make it clear that the generic procedure shown in Figure 3 is not restricted to microarray data analysis. Microarray is a high throughput technique but less quantitative. One can only detect genes with significant fold change (usually by ≥ 2 fold). For many priming effects, the fold change is less than 2 [10, 13]. Often more quantitative methods such as real time PCR are needed to confirm the microarray findings. Furthermore information on posttranslational and epigenetic modifications requires other techniques. In many applications, it is advantageous to combine time course data under LD, HD, and LD+HD stimulant obtained with different techniques. Here we use one example to illustrate this point.
Our microarray analysis suggested that STAT1 and SOCS1 may participate in a potential priming motif activated by IFN-γ (Figure 6), which is in consistence with the experimental investigations by Hu et al., . Hu et al., reported that a pretreatment of a sub-threshold of IFN-γ sensitized the Janus kinase (JAK)-signal transducer and activator of transcription (STAT) signaling for a second dose of IFN-γ . They found that a low dose IFN-γ exposure is able to switch on the transcription of STAT1. However, LD IFN-γ can only weakly activate the inhibitor SOCS1 in a transient manner . Since STAT1 protein is more stable than SOCS1 protein, the elevated expression of STAT1 actually increased the pool for STAT1 docking and phosphorylation in response to the second dose of IFN-γ, thereby contributing to the induction of priming effect.
Notice that in this model we only considered the coupling between IFN-γ induced STAT1 gene expression and the canonical Jak/STAT pathway. Figure 6 suggests a number of parallel pathways that may contribute to the observed IFN-γ priming effect. These pathways function together to make the temporal profile and amplitude of the priming phenomenon more complex.
Molecules within a cell interact with each other and form a large interconnected network. Consequently cellular information seldom propagates linearly through a single pathway. The priming effect, which widely studied using immune cells, is such an example. Based on our previous in silico studies , in this work we proposed a generic procedure to identify possible molecular candidates contributing to the priming effect through combined experimental time course measurement, subsequent data analysis and computational modeling. We demonstrated the procedure with high throughput microarray and other data on interferon-γ induced priming effects. This procedure is generally applicable to other similar problems. Especially it is of great significance to examine the generality and the specificity of the observed priming effects, in terms of stimulant and cell types. One may perform systematic screening using the proposed procedure combining with high throughput measurements, at both transcriptome and proteome levels.
The microarray data were downloaded from Gene Expression Omnibus (GEO, accession number: GDS1365). The data record the expression profile of approximately 12,000 gene probes with 3 independent pools. This is the only dataset we could find from GEO that include systematic time course measurement under either single dose or sequential stimulations (Control, HD 3hr, HD 24hr, LD Control, LD+HD 3hr, LD+HD 24hr. LD: 0.15 µg/L IFN-γ, HD: 5 µg/L IFN-γ).
In order to analyze the gene expression pattern, we first filtered out genes that contain no "Present Call" in all three independent pools. Genes without differential expression (by fold change < 2) under all of the following conditions were also filtered out: LD vs Control, HD (3hr) vs Control, HD (24hr) vs Control, LD+HD (3hr) vs Control and LD+HD (24hr) vs Control. All Differential expression was statistically analyzed by Welch's t-test with FDR correction. The threshold of p-value is set to be 0.05.
We used the commercial database IPA® (@Ingenuity) to query the molecular interactions among interested genes and products. IPA® assembles the signaling/regulatory network on a literature basis. Database query was restricted to immune cells and immune cell lines in Mus musculus or Homo sapiens. Interaction type was chosen to be either direct or indirect (i.e., interaction with intermediates). Prediction on potential priming candidates was made by comparing the priming motifs shown in Figure 2 and the signaling/regulatory networks constructed by IPA®.
We used a mathematic model adapted from Yamada et al.  to simulate the dynamics of Jak/STAT pathway in macrophages under different stimulation scenarios. Hu et al. have reported that increased expression of STAT1 induced by the first dose of IFN-γ treatment was responsible for sensitization of Jak/STAT1 pathway , we therefore added two additional reactions to the original model: STAT1 transcription triggered by IFN-γ and STAT1 translation. In addition, we introduced two reactions describing IRF-1 transcription and translation. Adding these two reactions allows us to exam the expression behavior of downstream gene IRF-1 for priming effects. As it is unclear how IFN-γ affects STAT1 expression, we proposed that an unknown intermediate × transduces the signal from IFN-γ to STAT1 gene.
As shown in additional file 3 and 4, our model includes 36 variables and 50 parameters. Most of the rate equations are presented using Mass-action kinetics. Several equations presenting gene transcription are denoted using Michaelis-Menten kinetics. We employed the same initial conditions for Jak, IFN-γ receptor, PPX, PPN and SHP-2 as in the work of Yamada et al. Other initial conditions are set to be the steady-state values achieved given zero IFN-γ signal. These ODEs are solved using standard ODE solver in Matlab. In our simulation, macrophages were primed with 0.15 µg/L of IFN-γ for 3 days, after which cells were washed for 10 minutes with fresh medium and re-stimulated with 5 µg/L IFN-γ for 2 days . The total STAT1 and SOCS1 proteins under repetitive two stimulations and single high dose of IFN-γ treatment were analyzed. In addition, phosphorylated STAT1 dimer and IRF-1 were examined as readouts to quantify the level of priming effect .
signal transducer and activator of transcription 1
interferon regulatory factor 1
interferon gamma-induced protein 10
Toll-like receptor 4
suppressor of cytokine signaling 1
tumor necrosis factor-alpha
Src homology 2
phosphorylated STAT1 dimer
unidentified phosphatase in the cytoplasm
SH2 domain-containing tyrosine phosphatase 2
IFN-γ-IFNR-Jak complex dimer
IFN-γ-IFNR-Jak complex phosphorylated dimer
phosphorylated cytoplasmic STAT1
phosphorylated nuclear STAT1
phosphorylated nuclear STAT1 dimer
phosphorylated cytoplasmic STAT1 dimer.
We thank Drs Xiaoyu Hu and Roderick Jensen for helpful discussions. This work was supported in part by the National Science Foundation (DMS-0969417), and by the National Institute of Allergy and Infectious Diseases (AI099120-01).
This article has been published as part of BMC Systems Biology Volume 6 Supplement 3, 2012: Proceedings of The International Conference on Intelligent Biology and Medicine (ICIBM) - Systems Biology. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcsystbiol/supplements/6/S3.
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.