Stochastic Expression of the Interferon-β Gene

The analysis of stochastic interferon-beta gene expression in virus-infected mammalian cells reveals that the levels of components required for virtually every step in the virus induction pathway are limiting.


Introduction
Eukaryotic cells respond to extracellular signals and environmental stresses by coordinately activating specific sets of genes. Signals from the cell surface or cytoplasm trigger signaling pathways that culminate in the binding of distinct combinations of coordinately activated transcription factors to promoter and enhancer elements that regulate gene expression. A wellcharacterized example of this is the activation of type I interferon (IFN) gene expression in response to virus infection or doublestranded RNA (dsRNA) treatment [1,2]. After infection, viral RNA is detected in the cytoplasm by one of two RNA helicases, retinoic acid-inducible gene I (RIG-I) or melanoma differentiation-associated gene 5 (MDA5), which respond to different types of viruses [3]. RIG-I recognizes short dsRNA or panhandle RNA bearing a 59 triphosphate group [3], and its activity is positively regulated by the ubiquitin E3 ligase tripartite motif 25 (Trim25) [4]. When RIG-I or MDA5 bind to RNA, they form heterodimers, undergo a conformational change, and expose a critical Nterminal caspase-recruiting domain (CARD) [5,6]. This domain interacts with the CARD domain of the downstream adaptor protein mitochondrial antiviral signaling (MAVS) (also known as IPS-1/Cardif/VISA) on the mitochondrial membrane [7]. The association of RIG-I with MAVS initiates the recruitment of adaptor proteins and leads to the activation of the transcription factors IFN regulatory factors 3 and 7 (IRF3 and IRF7) and NF-kB by the TANK-binding kinase 1 (TBK1) [8][9][10] and IKKa and IKKb, respectively [7,11]. Activated IRF3/IRF7 and NF-kB translocate into the nucleus and, along with the transcription factors ATF2/cJun, bind the IFN-b gene enhancer and recruit additional transcription components to form an enhanceosome [12]. This complex signaling and promoter recognition mechanism functions to coordinately activate a specific set of transcription factors that recognize the unique enhancer sequence of the IFNb gene and thus specifically activate IFN gene expression.
Early in situ hybridization (ISH) studies revealed that induction of IFNb expression by virus infection or dsRNA treatment in both human and mouse cells is stochastic [13,14]. That is, only a fraction of the infected cells express IFNb. This ''noisy'' expression is not due to genetic variation within the cell population, as multiple subclones of individual cells display the same low percentage of cells expressing IFNb [14]. In addition, different mouse and human cell lines display different percentages of expressed cells, and the levels of IFNb gene expression can be increased in low expressing cell lines by fusing them with high expressing lines, or by treating low expressing lines with IFNb [13,14]. These studies suggest that stochastic expression of the IFNb gene is a consequence of cell-to-cell differences in limiting cellular components required for IFN induction, and that one or more of the limiting factors are inducible by IFNb [13].
Stochastic expression has been observed with a number of other cytokine genes, including IL-2 [15], IL-4 [16,17], IL-10 [18], IL-5, and IL-13 [19]. In many of these cases, expression is both stochastic and monoallelic. Recent studies of IFNb gene expression revealed that stochastic expression in human cells is initially monoallelic, and becomes biallelic later in the induction [20,21]. In one study the stochastic expression of the IFNb gene was proposed to be a consequence of intrinsic noise due to stochastic enhanceosome assembly [21]. Subsequently, an analysis of human HeLa cells identified a specific set of Alu-repetitive DNA sequences bearing NF-kB binding sites that associate with the IFNb gene through interchromosomal interactions, and in so doing are thought to increase the local concentration of NF-kB. Initially, only one of the two chromosomes associates with the specialized NF-kB binding sequence, resulting in early monoallelic expression. Secretion of IFN leads to an increased expression of limiting factors (most likely IRF7, which is inducible by IFN), obviating the need for interchromosomal interactions, and leading to the activation of the second IFNb allele [20]. More recently, heterogeneity in the infecting viruses, rather than cell cycle differences, has been proposed to be the primary source of IFN stochastic expression [22]. Many functions have been proposed for biological noise, ranging from cell fate decisions during development to survival in fluctuating environments [23]. In the case of the IFN genes, neither the mechanisms nor functions of biological noise are well understood.
Here we report a detailed analysis of stochastic IFNb gene expression in mouse cells. We make use of an IFN-IRES-YFP reporter mouse [24] to perform a detailed analysis of differences between virus-infected cells that either express or do not express IFNb. Our results reveal a complex picture of stochastic expression of the IFNb gene, in which the levels of components required for virtually every step in the virus induction pathway are limiting. This includes components required for viral replication and expression, for sensing the presence of viral RNA by the host, and for the virus induction signaling pathway, and the transcription factors required of IFNb gene expression. Remarkably, in spite of this complexity the percentage of expressing cells remains constant through recloning and cell division, indicating that the stochasm of clonal cells is genetically programmed.

Stochastic Expression of Mouse and Human IFNb Genes
Sendai virus (SeV) infection of either mouse or human cells leads to the expression of IFNb mRNA in only a fraction of the infected cells ( Figures 1A, 1B, and S1A), and the percentage of expressing cells differs between different cell lines. The time course of mouse IFNb expression determined by ISH ( Figure 1B) is consistent with that from the quantitative PCR (qPCR) analysis ( Figure S1B and S1C). Remarkably, the percentage of cells expressing IFN did not exceed 20%, even at the latest time point ( Figure 1B). The absence of IFNb signal in the majority of cells is not an artifact of hybridization, as b-actin mRNA was detected in all cells ( Figure S1D). IFNb mRNA is specifically detected with an antisense IFNb RNA probe, while no signal is detected with a sense RNA probe ( Figure S1E). In addition, similar percentages of IFNb-expressing cells were detected by immunofluorescent staining using an IFNb antibody ( Figure S1F), strongly supporting the reproducibility and specificity of the IFNb ISH.
As mentioned above, enhanceosome assembly and limiting amounts of NF-kB have been proposed to be the primary limiting steps in stochastic expression of the human IFNb gene [20,21]. To determine whether this stochastic expression is unique to the IFNb gene because of the complexity of the IFNb enhanceosome, or is more general, we examined the expression of the IFNa genes, which are coinduced with IFNb, but have simple enhancer/ promoters, and do not require NF-kB [25,26]. Using either a mouse IFNa4 or human IFNa8 probe, we found that IFNa genes are also stochastically expressed in both mouse and human cells, respectively ( Figures 1C and S1G). Although NF-kB has been shown to be a limiting factor in the activation of the human IFNb gene [20], it is not required for IFNb expression in mouse cells [27]. Thus, in spite of this difference both the mouse and human IFNb genes are stochastically expressed. We also examined other virus-inducible genes, and found that they too are stochastically expressed (see below). Each of these virus-inducible genes requires different levels and combinations of transcription factors, yet they are all stochastic. In all of these cases (mouse and human IFNb and IFNa and the other virus-inducible genes), the common requirement is the RIG-I virus-inducible signaling pathway. We therefore carried out experiments to determine whether limiting components in this pathway contribute to the observed stochastic expression.

Separation and Characterization of IFN-Expressing and Non-Expressing Cells
To investigate the mechanism of stochastic IFNb gene expression, we made use of an IFNb reporter-knock-in mouse, in which YFP expression allows tracking of IFNb expression at a single-cell level [24]. Using IFNb/YFP homozygous mouse embryonic fibroblasts (MEFs) and fluorescence-activated cell sorting (FACS), we obtained pure populations of IFNb-producing and IFNb-negative cells upon SeV infection. As expected, IFNb mRNA is high in the YFP-positive cells, and very low in the YFPnegative cells ( Figure S2A). As expected, the IFNa2 and IFNa4 genes are also highly expressed in the YFP-positive cells, and not in the YFP-negative cells ( Figure S2A). These observations indicate that replication of the infecting virus and/or components in the RIG-I pathway are the limiting steps in the uninduced cells, rather than intrinsic differences in the IFNb and a promoters.

Author Summary
Eukaryotic cells can respond to extracellular signals by triggering the activation of specific genes. Viral infection of mammalian cells, for example, induces a high level of expression of type I interferons (IFNa and b), proteins required for antiviral immunity that protects cells from the infection. Previous studies have shown that the expression of the IFNb gene is stochastic, and under optimal conditions only a fraction of the infected cells express the IFNb gene. At present neither the mechanisms nor functions of this interesting phenomenon are well understood. We have addressed this question by analyzing IFN-expressing and non-expressing mouse cells that were infected with the highly transmissible Sendai virus. We show that stochastic IFNb gene expression is a consequence of cell-to-cell differences in limiting levels and/or activities of virus components at every level of the virus induction process, from viral replication to expression. These differences include the sensing of viral RNA by host factors, the activation of the signaling pathway, and the levels of activated transcription factors. Our findings reveal the complexity of the regulatory mechanisms controlling stochastic IFNb gene expression. We propose that the stochastic expression of IFN allows for an even distribution of IFN, thus avoiding over-expression of IFN in infected cells.
We also detected the relative mRNA abundance of other virusinducible genes in IFNb-expressing and non-expressing cells. As shown in Figure 1D, transcription levels of all tested inflammatory cytokine or chemokine genes are much higher in IFNb-producing cells compared to nonproducers. Considering the fact that IFNbproducing cells account for only 10% of the total cell population, we conclude that expression of all these virus-inducible genes is also stochastic and that these genes are coordinately activated with the type I IFN genes. Activation of these virus-inducible genes is known to require the RIG-I signaling pathway [28][29][30][31]. Thus, our results indicate that stochastic gene expression is due primarily to limiting components in the signaling pathway and not to gene-togene variation in the mechanism of gene activation.
In the case of human cells, stochastic expression of the IFNb gene is randomly monoallelic early and biallelic late in infection, and the activation of the second IFNb allele is inducible by IFN [20,21]. However, the nature of allelic expression of the IFNb gene has not been addressed in mouse cells. By using IFNb/YFP heterozygous MEFs, we showed that early after infection ( Previous studies have shown that the levels of IFNb gene expression can be increased by priming the cells with IFNb [13]. Using both mouse and human primary fibroblasts, we showed that IFNb pretreatment also increases the percentages of IFNbexpressing cells ( Figure S3), indicating that the limiting factor(s) contributing to stochastic IFNb gene expression are, indeed, inducible by IFNb. One of these IFN-inducible factors is IRF7 ( [20] and see below).

Viral Replication Is More Efficient in IFNb-Producing Cells
To examine the role of the infecting virus in stochastic IFNb gene expression, we infected primary MEFs with SeV followed by immunofluorescent staining using a SeV antibody. As shown in Figure S4A, most, if not all, of the cells are uniformly infected by SeV, far more than could explain the small percentage of cells expressing IFNb gene. When we used increasing multiplicities of SeV (as defined by hemagglutination units [HAU]) to infect primary MEFs, we found that the percentage of IFNb-producing cells increased as the HAU was increased, reaching a maximum of approximately 18% at the peak ( Figure S4B). However, as more virus was added (.200 HAU), the percentage of IFNb-producing cells decreased. Thus, the viral titer is not a limiting factor in the observed stochastic IFNb gene expression. Next, we determined viral transcript levels in both IFNb-producing and nonproducing cells. We found that the nucleoprotein (NP), matrix protein, and L polymerase protein mRNA transcripts were present at significantly higher levels in IFNb-producing cells compared to the nonproducers (Figures 2A and S4C). In addition, higher levels of SeV NP protein were detected in IFNb-producing cells ( Figure 2B).
The RNA helicase RIG-I detects viral genomic RNA and defective interfering (DI) genomes [32,33]. We therefore examined the levels of viral and DI genomes in both IFNb-producing and nonproducing cells. As shown in Figure 2C (upper panel), more SeV DI genomes were detected in IFNb-producing cells compared to IFNb-nonproducing cells at 8 and 12 h.p.i. Using a primer pair that specifically detects viral genomic RNA, we also detected more viral genomes in IFNb-producing MEFs 8 and 12 h.p.i. ( Figure 2C, lower panel). These results are consistent with the observed viral NP mRNA levels ( Figure S4C), and indicate that viral replication is more efficient in the IFN-producing cells. We also investigated the induction activities of total RNA extracted from both IFNb-producing and nonproducing cells. As shown in Figure S4D, total RNA from IFNb-producing cells infected for 8 or 12 h induced more IFNb expression compared to total RNA from IFNb-nonproducers at the same time points. We conclude that viral mRNA, DI genomes, and viral genomes are present at higher levels in IFNb-producing cells than in nonproducers. Thus, differences in the efficiency of viral replication/transcription contribute to the stochastic expression of the IFNb gene.
Previous studies led to the conclusion that the stochastic expression of the IFNb gene is a feature of the infecting virus, and not of the host cell [22]. To address this possibility, we determined the number of cells that have high levels of viral RNA and produce IFNb at 8 h.p.i. As shown in Figure 2D, after 8 h of virus infection, approximately 38% SeV-high cells (upper left and upper right) were detected, and about 9% YFP-positive cells (upper right and lower right). Although a higher percentage of IFNb-expressing cells was observed within the SeV-high cell population (6.56% versus 2.42%), only 17% (6.56% out of 38%) of SeV-high cells produce IFN. Thus, although cell-to-cell differences in viral replication contribute to the stochastic expression of IFN, these differences are not sufficient to explain the extent of stochastic IFN gene expression.

The RIG-I Signaling Pathway Is Activated and More Potent in IFNb-Producing Cells
To further investigate the mechanism of stochastic IFNb gene expression, we determined the localization of various components of the signaling pathway required for IFN production using nuclear and cytoplasmic fractions separated from both expressing and non-expressing cells. Consistent with the limiting component hypothesis, we detected phosphorylation and translocation of IRF3 in the YFP-positive cells, but not in the YFP-negative cells ( Figure 3A). Previous studies have shown that IRF3, like IRF7, is phosphorylated by the TBK1 kinase, and translocates from the cytoplasm to the nucleus. As both IRF3 and IRF7 are activated via the RIG-I pathway, our results suggest that one or more components of the RIG-I signaling pathway are limiting in the cells that fail to express IFN. A similar result was obtained with sorted cells at 12 h.p.i. ( Figure 3B).
In human cells both NF-kB and IRF3/IRF7 are required for virus induction of the IFNb gene [12,34]. The human and mouse IFNb enhancers differ in only two nucleotides out of 45 bases. However, in mouse cells NF-kB is required only for early antiviral activity, when the level of active IRF3 is low, but is not required for maximum levels of IFNb expression late in induction [27,35]. Consistent with this finding, we show that only a small fraction of the p65 subunit of NF-kB translocates to the nucleus 8 h.p.i., and little difference is observed in NF-kB localization between the YFP-positive and YFP-negative cells ( Figure 3A).
The observation that IRF3 activation and translocation occurs in only a fraction of virus-infected cells suggests that upstream components in the RIG-I signaling pathway differ in IFNbproducing and nonproducing cells. Western blotting results ( Figure 3C) showed that IFNb-producing cells have higher levels of both RIG-I and MDA5 than the nonproducing population. Trim25, an E3 ligase required for RIG-I activation [4], is also present at a higher level in the IFNb-producing cells ( Figure 3C). The increase in protein levels appears to be a consequence of differential transcription of the tested genes, as mRNA levels of all three genes are higher in IFNb-producing cells ( Figure 3D). We conclude that the IFNb-producing cells have higher levels of essential RIG-I signaling pathway components than the IFNbnonproducing cells. Thus, at least part of the observed stochastic expression is due to limiting RIG-I pathway components in the cells that do not express IFN.
By contrast to the RNA detectors, the protein levels for both MAVS and TBK1, two essential components of the RIG-I signaling pathway [7,9], were lower in the IFNb-producing cells ( Figure 3C). However, this is likely due to the degradation and/or cleavage of the MAVS protein in infected cells [36][37][38]. The data of Figure 3C suggest that TBK1 is also targeted for degradation during virus infection, consistent with the observation that TBK1 is subject to proteasome-dependent degradation [39]. Thus the turnover of both MAVS and TBK1 may be required for the postinduction turn-off of IFNb gene expression [38].

Over-Expression of Individual Components of the RIG-I Signaling Pathway Increases the Percentage of Cells Expressing IFNb
We have shown that the RIG-I signaling pathway is selectively activated in IFNb-expressing cells, and this is due only in part to the cell-to-cell differences in virus infection/replication. Our results also suggest that IFNb-producing cells have a more potent signaling pathway than IFNb-non-expressing cells. To further explore this possibility, we established a series of L929 stable cell lines that express RIG-I, MDA5, or Trim25 under the control of a tetracycline-inducible promoter ( Figure S5A). As shown in Figure  S5B and S5C, high levels of exogenous RIG-I only slightly increased the percentage of IFNb-producing cells. A larger increase was observed with MDA5 and Trim25, but the final percentage in both cases was still under 30%. Thus, these upstream components appear to be among several limiting factors in the cell population.
Additional components in the RIG-I signaling pathway were tested using the same approach, and high percentages of IFNbproducing cells were observed ( Figure 4A and 4B). While a large difference between tetracycline-negative and tetracycline-positive cells was observed with the TBK1 line, only a small difference was observed between the corresponding MAVS lines. However, a large difference was observed between the non-transformed and transformed MAVS lines, suggesting that a low level of leaky transcription in the MAVS line is sufficient to dramatically increase the number of IFNb-expressing cells. These data clearly indicate that both MAVS and TBK1 are limiting components in the RIG-I pathway and therefore contribute significantly to stochastic IFNb expression.
We have shown that over-expression of RIG-I or Trim25 alone only slightly increases the percentage of IFNb-producing cells, but it is possible that both must be expressed to achieve maximum levels of IFNb production. We therefore transfected RIG-I stable transfectants with a Trim25 expression plasmid, and the other way around. The cells were then induced with tetracycline, infected with SeV, and examined for IFNb mRNA expression. Control experiments using a GFP reporter indicated that under our experimental conditions approximately 70% of cells can be transfected with the second plasmid ( Figure S5D). As shown in Figure 4C and 4D, a dramatic increase was observed only 6 h.p.i. when either the RIG-I or Trim25 lines were transfected with Trim25 or RIG-I, respectively. This observation was confirmed by carrying out intracellular staining and flow cytometry experiments using IFNb/YFP homozygous MEFs ( Figure S6). We conclude that the combination of RIG-I and Trim25 is limiting in the RIG-I pathway.
We note that the increase of IFNb-expressing cells was not observed in uninfected cells, with the only exception being MAVS.
Thus, over-expression of these signaling components did not bypass the requirement for signaling pathway activation.

IRF7 Is a Primary Limiting Factor in Stochastic IFN Gene Expression
Expression of the IFNb gene requires an active RIG-I signaling pathway and assembly of the enhanceosome complex on the IFNb promoter. To investigate whether individual enhanceosome components are limiting factors, we established a series of tetracycline-inducible L929 stable lines that express IRF3, IRF7, or p65 genes. Figure 5A and 5B show that, without tetracycline induction, only 10%-15% of the cells produce detectable levels of IFNb mRNA in response to virus infection. Remarkably, the percentage of IFNb-producing cells upon SeV infection increased to 85% when IRF7 expression was induced by tetracycline in every cell ( Figure S7A). A smaller increase (55%) was observed when IRF3 was over-expressed, whereas increasing the concentration of NF-kB had little effect, consistent with the data in Figure 3A, and previously published studies [27]. Interestingly, IRF7 over-expression also significantly increased the percentage of IFNa-producing cells after virus infection ( Figure S7B and S7C). It is known that IRF7 is required for maximum induction of type I IFN genes [25], and its basal protein level is very low in most cell types except for plasmacytoid dendritic cells [26,40]. We conclude that IRF7 is a critical limiting factor that is a major contributor to stochastic expression of mouse IFNa and b genes. This conclusion is also supported by our ISH results from 4E-BP1/4E-BP2 doubleknockout MEFs ( Figure 5C and 5D). Previous studies have indentified 4E-BPs as negative regulators of type I IFN production via translational repression of IRF7 mRNA [41]. As shown in Figure 5C and 5D, we observed a 4-fold increase of the percentage of IFNb-expressing cells in 4E-BP1/4E-BP2 double-knockout MEFs compared to wild-type MEFs, consistent with the conclusion that a limiting amount of IRF7 is a major contributor to the stochastic expression of IFNb.
We also found that type I IFN induction was exceptionally high, with much faster kinetics in cells expressing exogenous IRF7 than in control cells ( Figure S7D). In the absence of tetracycline induction, low levels of IFNb, IFNa4, and IFNa2 mRNA were first detected 6 h, 9 h, and 12 h.p.i., respectively. When the cells were treated with tetracycline, the kinetics of IFN gene transcription changed significantly. IFNb, IFNa4, and IFNa2 transcripts could be detected as early as 4 h after virus infection. Even at 24 h.p.i., steady and robust transcription of these genes could still be detected. These observations are consistent with a model in which IRF3 is normally activated early for IFN gene induction. Later, higher levels of IRF7 are produced by IFN and are required for both IFNb and IFNa gene expression, but IRF7 is rapidly turned over, leading to the cessation of both IFNb and IFNa gene expression [1,25,26]. By contrast, in the presence of excess IRF7 in the tetracyline-activated cells, both IFNb and IFNa are activated earlier, and continue to be expressed because of the continuous presence of IRF7.

IRF7 Positively Regulates the RIG-I Signaling Pathway
We have shown that over-expression of IRF7 or both RIG-I and Trim25 almost completely eliminates stochastic IFNb expression ( Figures 4C, 4D, and 5). To investigate the connection between these observations, we carried out microarray analysis to compare genome-wide expression profiles of L929-IRF7 stable transfectants treated with or without tetracycline. Interestingly, upon IRF7 over-expression, only two up-regulated signaling pathways were identified from the KEGG Pathway Database, and the RIG-I-like receptor signaling pathway is the most upregulated (p = 3.6E-06) ( Figure S8A and S8B) [42]. We did not identify signaling pathways that were similarly enriched among the down-regulated genes. Using qPCR, we confirmed that the mRNA levels of both RIG-I and Trim25 were higher in IRF7 over-expressing cells ( Figure S8C). Considering the low basal expression level of IRF7, we conclude that a high level of IRF7 protein increases the percentage of IFNb-expressing cells not only by increasing its own abundance, but also by up-regulating the RIG-I signaling pathway to increase the potency of activation of the IFNb gene.
Stochastic Expression of IFNb Induced by dsRNA-poly I:C Is Due to Limiting Amounts of MDA5 and IRF7 IFNb gene expression can also be induced by transfection of the synthetic dsRNA polyriboinosinic polyribocytidylic acid (poly I:C), and this induction occurs mainly through the MDA5 signaling pathway [43]. Early studies revealed that induction of IFNb expression by dsRNA treatment is also stochastic [13,14]. We therefore asked whether stochastic IFNb gene expression induced by dsRNA is due to cell-to-cell variation in the levels of MDA5 and IRF7. Using FACS analysis, we found that poly I:C-induced IFNb expression is also stochastic ( Figure 6A). When IFNb/YFP homozygous MEFs were electroporated with Cy5-labeled poly I:C, only 9% of the cells produced IFNb as detected by the presence of YFP. However, the electroporation efficiency was over 99% ( Figure 6A, left panel). Interestingly, based on the Cy5 intensity, there were two populations of cells, which contained different amounts of poly I:C. When we gated these two populations out as ''poly I:C-high'' and ''poly I:C-low'', we observed that the ''poly I:C-high'' population included more cells producing IFNb ( Figure 6A, right panel), indicating that the amount of inducer does affect the extent of stochastic IFNb expression. However, only a small percentage of ''poly I:C-high'' cells expressed the IFNb gene, clearly indicating that other limiting factor(s) dominate the stochastic IFNb expression induced by poly I:C transfection. We therefore carried out experiments to identify these limiting components.
L929-MDA5 and L929-RIG-I stable transfectants were transfected with poly I:C followed by ISH to detect IFNb expression. As shown in Figure 6B and 6C, over-expression of RIG-I only slightly increased the percentage of IFNb-producing cells. By contrast over-expression of MDA5, the major cytoplasmic receptor for poly I:C, led to a substantial increase in the percentage of IFNbproducing cells (from 15% to 65%). Considering that the transfection efficiency is approximately 75% (data not shown), over-expression of MDA5 basically eliminates stochastic expression of the IFNb gene in response to poly I:C transfection. Furthermore, the results of the flow cytometry experiment also supported this conclusion. As shown in Figure 6D, after 8 h of poly I:C stimulation, we observed approximately 2.6% YFP-positive cells. Within this population, about 70% of the YFP-positive cells had higher levels of MDA5 protein (1.86% out of 2.67%). We note that the percentage of YFP-positive cells is much lower than that observed with virus infection (Figures 2D and S6).
Over-expression of IRF3 or IRF7 also increased the percentage of IFNb-producing cells in response to poly I:C ( Figure S9A and S9B). As shown in Figure S8, over-expression of the IRF7 gene upregulates MDA5 gene expression. Considering its low basal expression level, IRF7 is also an important limiting factor in stochastic IFNb expression induced by poly I:C transfection. Taken together, these data show that poly I:C-induced stochastic IFNb expression depends on the abundance of both poly I:C and signaling pathway protein MDA5 as well as IRF3/IRF7, which is similar to what was found in the case of virus infection.

Variation in the Levels of RIG-I Signaling Pathway Components
We also asked whether the concentrations of proteins regulating IFNb expression are sufficiently different from cell to cell to account for the stochastic IFNb expression. Using flow cytometry, we measured the distributions of six components in the RIG-I signaling pathway for which specific antibodies are available. As shown in Figure 7A and 7B, all six proteins were log-normally distributed across the population. Quantitative immunofluorescence data for individual components show similar distributions of each factor at the single-cell level ( Figure S10). Combined with our previous data, these observations suggest that naturally occurring differences in the protein levels of signaling pathway components are the primary cause of cell-to-cell variability in IFNb expression upon virus infection.

IFN-Inducible Antiviral Genes Are Not Stochastically Expressed
When IFN is secreted from virus-infected cells in vivo, it binds to type I IFN receptors on surrounding cells and activates a large set of genes encoding antiviral proteins (interferon-stimulated genes [ISGs]) via the Jak/STAT signal transduction pathway. We therefore carried out experiments to determine whether the induction of antiviral ISGs is also stochastic. As shown in Figure 7C, ISG15 is expressed in all cells upon treatment with IFNb. Thus, when IFN is secreted, all of the surrounding cells produce antiviral proteins. This result is also consistent with previous observations showing that the antiviral response induced by IFN is a robust feature common to all cells, and is independent of the stochastic expression of IFN receptor IFNAR [44].

Discussion
Regulation of type I IFN production is essential for the innate immune response to viral infections [45,46]. However, high levels of IFNb can be toxic [47,48]. Thus, IFNb production must be tightly regulated. This regulation appears to be both temporal and stochastic. Type I IFN genes are tightly repressed prior to virus infection, activated upon infection, and then rapidly turned off several hours later ( Figure S1B and S1C). Previous studies of several cytokine genes suggest that this stochastic gene expression provides an additional mechanism of regulation whereby optimal levels of cytokine production are determined by the frequency of expressing cells rather than by protein levels per cell [18,19,49]. Thus, it is possible that stochastic expression is a primary mechanism for controlling the optimal level of IFNb production in vivo. In particular, we have shown that while IFN production is stochastic, the activation of the antiviral gene program by secreted IFN is not. Thus, stochastic expression of IFN would allow the regional distribution of the cytokine and activation of the surrounding cells, without producing toxic levels of IFN.
Previous studies have implicated as limiting steps enhanceosome assembly [20,21] and the assembly of an interchromosomal transcriptional hub formed through interactions between Alu elements bearing NF-kB sites [20]. More recently, the infecting virus, rather than intrinsic properties of the infected cell, has been implicated in this stochasm [22]. The data presented here reveal a far more complex mechanism in which cell-to-cell variations in limiting components required to support viral replication, to detect and signal the presence of viral RNA, and to activate transcription factors all contribute to the observed stochastic expression ( Figure 7D). It seems likely that the key limiting factor varies between cell types, cell lines, and organisms.
The earliest step in the virus induction signaling pathway is entry of virus or dsRNA into the cell. We have shown that both inducers elicit stochastic expression, but in neither case is this due to limiting inducer (Figures S4B and 6A). We showed that both IFNb-producing and nonproducing cells were infected by SeV ( Figure 2B). However, the IFNb-producing cells contained significantly higher levels of the products of viral replication and transcription. Thus, it appears that there are cell-to-cell differences in the ability to support efficient viral replication, and these differences influence the probability of IFNb gene expression. Presumably, high levels of RNA inducer in the IFN-producing cells overcome limiting amounts of RIG-I or MDA5. However, differences in viral replication alone cannot explain the observed stochasm in IFNb production. A previous study, using a cell line transfected with an IFNb-GFP reporter, concluded that stochastic IFNb expression is due entirely to heterogeneity in the infecting virus [22]. However, in that study the IFNb-GFP cell line was preselected to minimize stochastic expression of the reporter. In addition, that study involved a stably transfected gene, while the present study made use of the endogenous gene. The results presented here strongly indicate that heterogeneity of both the virus and host cells together are responsible for the stochastic expression of IFNb.
We have identified multiple limiting steps in the activation of IFNb gene expression, ranging from initial steps in virus infection and replication, to the signaling pathway, to the activation and binding of transcriptional activator proteins to the IFNb promoter. For example, over-expression of individual components in the RIG-I signaling pathway increases the percentage of IFNexpressing cells. The largest increase was observed with IRF7, which lies at the endpoint of the RIG-I pathway, and also positively controls the expression of components in the RIG-I signaling pathway. Taken together, these data are consistent with a model in which the probability of expression of the IFNb gene in individual cells depends primarily on the activation of the RIG-I signaling pathway and the presence of sufficient numbers of IRF7 molecules to activate transcription ( Figure 7D). This conclusion is consistent with the observation that both IFNb and IFNa are stochastically expressed in response to virus infection ( Figure 1A and 1C). The expression of both genes requires activation of the RIG-I pathway and active IRF7 [50].
We find that limiting amounts of other RIG-I pathway components also contribute to stochastic expression of the IFNb gene, as we observed higher levels of RIG-I/Trim25 and MDA5 mRNA and protein levels in the IFNb-producing cells than in the nonproducers (Figure 3). In addition, over-expression of RIG-I and Trim25 together leads to a dramatic increase in the percentage of cells that express IFNb ( Figure 4C and 4D). Similar results were obtained with high levels of expression of the RIG-I signaling components MAVS and TBK1 and the transcription factors IRF3 and IRF7 (Figures 4A, 4B, and 5). Thus, it appears that many, if not all, of the components in the RIG-I signaling pathway, from the sensors of viral RNA to the essential transcription factors, can be limiting components in the virus induction pathway.
The largest increase in the percentage of IFN-producing cells was observed when IRF7 was over-expressed. IRF7 is the master regulator of type I IFN gene expression [25], and is present at low levels in all cell types except plasmacytoid dendritic cells, where it is constitutively abundant [26,40]. Our over-expression experiments show that high levels of IRF7 promote the transcription of type I IFN genes ( Figure S7D), and essentially eliminate the stochastic expression of both the IFNb and a genes (Figures 5 and  S7). In a previous study in human cells, both NF-kB and IRF7 over-expression was shown to partially suppress stochastic IFNb expression [20]. Our results are consistent with this observation. However, there are two differences. First, based, at least in part, on the lack of requirement of NF-kB in murine cells, we observed a relatively small effect of increasing NF-kB expression. Second, we saw a greater effect of IRF7 expression in murine cells than was observed in human cells. Over-expression of IRF7 in L929 cells almost completely eliminated stochastic expression of both IFNb and a genes, while in human HeLa cells high levels of IRF7 increase the percentage of IFNb-producing cells to almost 55% [20]. Deleting the IRF7 translational repressors, 4E-BPs, also increased the IFNb-expressing MEFs by 4-fold ( Figure 5C and 5D). We also showed that the RIG-I signaling pathway, and in particular RIG-I and Trim25, are up-regulated in IRF7 overexpressing cells ( Figure S8). We conclude that limiting amounts of active IRF7 appear to be overcome by two mechanisms: positive auto-regulation of IRF7 expression, and IRF7-dependent upregulation of the RIG-I signaling pathway.
We note that in addition to IFNb, several other virus-inducible genes, including TNFa, IL-6, CCL4, and CCL5, are highly expressed in the IFNb-producing cells compared to nonproducers, suggesting that many, if not all, of the virus-inducible genes are stochastically expressed. The common feature of the activation of all of these genes is that they all require the RIG-I signaling pathway [28][29][30][31]. Thus, we conclude that stochastic gene expression is primarily due to limiting components in the signaling pathway but not gene-to-gene variation in the mechanism of gene activation.
We showed that although the IFNb gene is stochastically expressed upon virus infection, the antiviral ISGs, e.g., ISG15, were equally induced in all cells ( Figure 7C). However, we note that RIG-I, Trim25, and MDA5, which are also antiviral ISGs, are highly expressed in IFNb-producing cells compared to nonproducing cells ( Figure 3C and 3D). We believe that the differences we observed here reflect naturally occurring cell-to-cell variability in the levels of expression of these genes prior to virus infection, and that this variability is the primary source of stochastic IFNb gene expression. However, at later times after virus infection, we expect that the differences in the mRNA or protein levels of these genes between the YFP-positive and YFPnegative populations will be much smaller compared to those at earlier stages (8 h.p.i.). As shown in Figure S11A and S11B, our qPCR data and Western blot data support this expectation. The IFNb gene is also stochastically expressed in IFNAR-deficient MEFs, which suggests that the IFNAR levels or an IFNb feedback loop are not major factors responsible for stochastic IFNb gene expression ( Figure S11C). We further measured the distributions of six components in the RIG-I signaling pathway. As shown in Figures 7A, 7B, and S10, all six proteins were log-normally distributed across the cell population, an observation that is consistent with data on other proteins [51,52]. Thus, naturally occurring differences in the protein levels and activities of individual signaling pathway components and transcription factors account for stochastic IFNb expression induced by both poly I:C induction and virus infection.
Previous studies have shown that naturally occurring differences in the levels of proteins in the apoptotic signaling pathway are the primary reasons for cell-to-cell variability in the probability of cell death [52]. Thus, the results presented here not only reveal the complexity of the regulatory mechanisms controlling stochastic IFNb gene expression, but also suggest a general mechanism used in different biological processes to establish and control stochastic gene expression. A remarkable feature of stochastic expression is that it appears to be an intrinsic property of different clonal populations of cells. For example, if a particular cell line displays a certain percentage of activated cells, that percentage differs from other cell lines, and is retained when the cells are recloned [14]. Thus, the extent of stochasm appears to be a genetic and epigenetic feature of clonal cell populations.

Cells, Reagents, and Plasmids
All cell lines, including L929, RAW 264.7, MG63, and 293T, were from the American Type Culture Center; primary MEFs were isolated using standard protocols from IFNb/YFP mice [24]. Primary human foreskin fibroblast cells were purchased from PromoCell. All cells were cultured in DMEM (Gibco) supplemented with 10% FBS (Gibco) in a 5% CO 2 incubator. Cycloheximide was purchased from Sigma-Aldrich. Human and mouse recombinant IFN proteins were purchased from PBL Interferonsource. Brefeldin A solution was purchased from eBioscience. Poly I:C was purchased from InvivoGen. Cy5-labeled poly I:C was generated using Label IT Nucleic Acid Labeling Kit (Mirus). The different expression constructs were generated by cloning the coding sequences of each gene by PCR and inserting them into the vector pt-REX-DEST30, which has the tetracyclineinducible promoter (Invitrogen).

RNA Preparation and PCR
Total RNA was extracted with Trizol reagent (Invitrogen). Real-time quantitative reverse transcription PCR (qRT-PCR) was conducted according to standard protocols.

Antibodies and Western Blot
Antibody against YFP was from Chemicon (Millipore) or Abcam. RIG-I, MAVS, and GAPDH antibodies were from Cell Signaling. Antibodies against p65, HDAC1, and Trim25 were from Santa Cruz Biotechnology. MDA5 and TBK1 antibodies were from Abcam and Imgenex, respectively. IFNb antibody used for FACS was from Millipore. SeV antibodies were kindly provided by Dr. Atsushi Kato (National Institute of Infectious Diseases, Japan). Nuclear/cytosol fractionation was performed using Nuclear/Cytosol Fractionation Kit (BioVision). Western blots were carried out using standard protocols.

In Situ Hybridization
Antisense RNA probes recognizing mouse IFNb or b-actin were synthesized using T7 or SP6 polymerase and digoxigenin-labeled nucleotides (Roche Applied Science). Cells were cultured on poly-D-lysine-coated 24-well plates (Fisher) and either mock-or virusinfected for the times indicated. Cells were then washed twice with PBS and fixed with 4% paraformaldehyde. Hybridization, washes, and staining were carried out as previously described [53].

Flow Cytometry
MEF cells were fixed with IC Fixation Buffer and permeabilized with Permeabilization Buffer (both from eBioscience). After incubation with appropriate antibodies, flow cytometry was done with a FACSCalibur, and data were analyzed with CellQuest software (both from Becton Dickinson).

Microarray Analysis and KEGG Pathway Enrichment Analysis
Total RNA from untreated and tetracycline-induced L929-IRF7 cells were prepared using Trizol reagent (Invitrogen) followed by purification using MEGAclear (Ambion). Biotinylated RNA probes were synthesized by two rounds of amplification using the MessageAmp II aRNA Amplification kit (Ambion). The probes were hybridized with Affymetrix Mouse Genome 430A_2.0 array chips. Affymetrix DAT files were processed using the Affymetrix Gene Chip Operating System to create CEL files. Normalized expression values were analyzed with the Bioconductor Limma package, an approach for implementing empirical Bayes linear modeling [42]. For all comparison tests, genes with an absolute fold change in transcript level exceeding 1.5 and p,0.05 were selected for further analyses. The likelihood of overrepresentation of KEGG signaling pathways in the up-or downregulated gene list relative to a background of all array genes was calculated by Fisher's exact test for statistical analysis.