Epigenetic and Conventional Regulation Is Distributed among Activators of FLO11 Allowing Tuning of Population-Level Heterogeneity in Its Expression

Epigenetic switches encode their state information either locally, often via covalent modification of DNA or histones, or globally, usually in the level of a trans-regulatory factor. Here we examine how the regulation of cis-encoded epigenetic switches controls the extent of heterogeneity in gene expression, which is ultimately tied to phenotypic diversity in a population. We show that two copies of the FLO11 locus in Saccharomyces cerevisiae switch between a silenced and competent promoter state in a random and independent fashion, implying that the molecular event leading to the transition occurs locally at the promoter, in cis. We further quantify the effect of trans regulators both on the slow epigenetic transitions between a silenced and competent promoter state and on the fast promoter transitions associated with conventional regulation of FLO11. We find different classes of regulators affect epigenetic, conventional, or both forms of regulation. Distributing kinetic control of epigenetic silencing and conventional gene activation offers cells flexibility in shaping the distribution of gene expression and phenotype within a population.


Introduction
Microbial cell populations employ a number of strategies to rapidly generate phenotypic diversity on relatively short time scales [1,2]. In some microbes, genes known as contingency loci contain tandem repeats of DNA whose recombination results in turning expression ON or OFF [3]. Other genetic strategies include the directed recombination of silent alleles into a particular active locus, as is the case for mating type switching in yeasts and surface antigen expression in T. brucei [4], the causative agent of African sleeping sickness. Another widely used strategy that generates phenotypic heterogeneity in clonal microbial cell populations is epigenetic gene regulation. In contrast to genetic strategies, this refers to the heritable change in a gene's expression that is not caused by changes in the underlying gene sequence. For example, the parasite P. falciparum (malaria) and the model organisms S. cerevisiae and E. coli use epigenetic mechanisms to variably express antigenic cell-surface proteins [2] and possibly escape immune surveillance and/or survive in an unpredictably changing environment.
Many epigenetically regulated genes can be considered switches as they have two heritable expression states, ''ON'' and ''OFF.'' A stable epigenetic marker maintains each state and can be encoded in cis or in trans. The molecular basis of local, cis markers involve covalent modifications of DNA or DNA-associated proteins. These include DNA methylation [5] and histone modifications that define silenced heterochromatin or active euchromatin in eukaryotes [6]. Global, trans markers are often transcription factor activity; the mechanism for stable, slow switching of these levels is positive or double negative feedback loops that generate heritable bistable gene expression states associated with high or low levels of transcription factor activity [7][8][9]. Switches using either scheme respond to environmental factors, but heterogeneity is observed even with constant environmental conditions, suggesting that the switch can rarely and randomly be toggled due to fluctuations in the intracellular environment. The two schemes can be also combined. For example, in uropathogenic E. coli the expression of pyelonephritis-associated pili is regulated by an epigenetic switch that maintains its state through both DNA methylation and a positive feedback loop [10].
The control of phenotypic heterogeneity is arguably as important as its rapid generation. Heterogeneity, or noise, in conventionally regulated gene expression has been well-studied in recent years. Single cell and single molecule studies have revealed that gene activation occurs in random, intermittent transcriptional bursts [11][12][13][14] due to fast promoter fluctuations (.once per cell cycle) between an inactive (but competent) and active promoter state. Mechanistically, this is an oversimplification as the promoter likely adopts a series of different states involving binding of various gene-specific and general transcriptional machinery that lead to productive transcription. Here, the active promoter state can be thought of as one where rapid initiation and reinitiation is possible.
For example, for regulable RNA Pol II-dependent promoters, transcriptional initiation is often rate-limiting and hence the active promoter state corresponds to pre-initiation complex formation. Expression heterogeneity caused by even these fast fluctuations can have consequences on phenotype and population-level fitness [15].
Noise in gene expression can be partitioned depending on whether its source is intrinsic or extrinsic to the process of gene expression. Intrinsic noise is due to the random nature of chemical transformations, including transcription and translation events. However, the random bursts of transcription thought to be associated with fast promoter fluctuations occurring in cis appear to be the dominant source of intrinsic noise in eukaryotes [16,17]. Extrinsic noise is due to cell-to-cell variation in trans factors affecting gene expression: for example, general and gene-specific transcriptional machinery, ribosome number and tRNA availability, or even cell morphology. The two sources can be experimentally distinguished using a dual-reporter assay, where two copies of the same promoter are used to drive distinguishable fluorescent protein variants [18]. Extrinsic noise is variation in protein levels between different cells; intrinsic noise is variation in protein levels within the same cell.
How regulators control the kinetics of intrinsic promoter fluctuations dictates the resulting expression heterogeneity. Stochastic models can be used to directly quantify this relationship [19]. Most transcriptional regulators appear to function by modulating the frequency of these bursts [19,20], probably in large part by increasing the rate of transcriptional initiation. Therefore, regulators do not control expression heterogeneity independently of expression level. In fact, heterogeneity is under genetic control as noisy promoters tend to have particular characteristics: strong TATA boxes, highly regulable, and dependent on chromatin remodeling activities [19][20][21].
While conventional gene regulation involves fast fluctuations between inactive (competent) and active promoter states, epigenetic silencing of gene expression involves slow fluctuations (,once per cell cycle) between a silenced and competent state. The kinetics of these fluctuations in trans-encoded switches involving feedback loops and associated with bistable gene expression have been studied in detail [7,22,23]. Both theory and experiment suggest that extrinsic fluctuations in the trans factor that overcome the stability of the two epigenetic states lead to switching [7]. However, much less is known of the precise role of regulators in modulating fluctuations of cis-encoded switches which must involve changes in the local promoter state. For example, activators could increase population-averaged expression by either stabilizing the competent state or destabilizing the silenced state. The resulting heterogeneity in expression is dictated by the specific kinetic role of the activator.
In a diploid organism, an epigenetically regulated gene might exhibit four different expression states if each copy switches independently. With global encoding, both copies respond to the same global factor and must switch in a correlated manner. However, with local encoding, each copy may respond independently if the fluctuation that trips the switch is a molecular event that occurs locally at one copy. In fact, a recent study demonstrated the random and independent switching of two copies of a reporter gene inserted within the canonically silenced mating type loci, HMR and HML, in S. cerevisiae. Four distinct expression states were observed in a sir1 background, where SIRprotein dependent silencing of these loci is partially impaired [24].
Multiple cis-encoded epigenetic switches that toggle slowly and randomly could lead a combinatorial explosion of expression states and represent a powerful strategy to generate phenotypic diversity. Is independent switching employed in nature and how are slow fluctuations regulated? The S. cerevisiae Flo11p is a cell-wall adhesin protein and member of the FLO gene family important in mediating cell-to-cell and hydrophobic cell-surface interactions [25]. In addition to traditional regulation via the MAPK and PKA pathways [26,27], at least three mechanisms are known to generate variation in cell-surface adhesins: ploidy regulation [28], frequent recombination of tandem repeats within adhesin genes [29] and epigenetic silencing [30]. Silencing at FLO11 occurs in a SIR-protein independent manner and is both promoter and position-specific [30]. Given the importance of phenotypic diversity in the adhesive phenotype and the epigenetic silencing at FLO11, independent switching could represent a fourth mechanism for generating variation.
At 3.5 kb, the FLO11 promoter is one of the largest in S. cerevisiae and regulated by many factors (Figure 1) whose kinetic roles are unknown. Silencing of FLO11 is thought to occur through the recruitment of the histone deacetylase Hda1p via the repressor Sfl1p through a yet to be defined mechanism [30]. The Sfl1p repressor binding site overlaps the Flo8p activator binding site [27]. Activation of FLO11 through the protein kinase A (PKA) pathway results in phosphorylation of both Sfl1p and Flo8p. While phosphorylation disables Sfl1p binding, it enables Flo8p binding [26,27]. Additional transcription factors bind directly to this promoter [26,27,31] including the MAPK regulated Ste12p/ Tec1p and Phd1p. These three activators require Flo8p for activation and play a significant role in determining the overall level of expression [32]. Two activators, Msn1p and Mss11p, do not require Flo8p for activation and operate through poorly understood mechanisms that do not seem to require DNA binding [33]. Msn1p acts at longer distances to destabilize chromatin [34]; Mss11p has glutamine rich activation domains and may associate weakly with Flo8p [32]. All these activators modulate plasmidborne FLO11 expression, a context where silencing does not occur [26,27]. However, their varied biochemical roles might imply distinct kinetic and functional roles in epigenetic regulation of FLO11.
Here, we provide evidence that FLO11 is indeed a cis-encoded epigenetic switch and identify the kinetic roles of trans factors in the epigenetic and conventional regulation of FLO11. Within a diploid yeast, each locus switches in a slow, random, and independent

Author Summary
In an uncertain and changing world, microbial populations with a diverse range of phenotypes may outperform a monolithic population. Over many generations, mutations can lead to genetic diversity in a population. However, microbes have strategies to generate such diversity quickly. For example, if multiple genes switch ON and OFF slowly, randomly, and independently of each other, then a large combination of gene expression states, and hence phenotypes, are possible. The different gene expression states do not involve changes in DNA sequence and are therefore epigenetically inherited. We show that the two copies of the FLO11 gene in S. cerevisiae can switch ON and OFF slowly and independently. In addition, we reveal a simple regulatory strategy by which cells can control the proportion of cells in different gene expression states. Because FLO11 encodes a cell-wall protein responsible for mediating cell-cell and cell-surface interactions, this control might literally allow natural populations to have a controllable fraction of cells ''stick around'' while the other fraction is easily washed away.
manner, with switching rates dependent on environmental conditions. Using a stochastic kinetic model, we infer the kinetic role that different regulators have on the slow promoter fluctuations associated with epigenetic transitions between a silenced and competent promoter state and the fast promoter fluctuations associated with conventional gene activation. We find three classes of FLO11 regulators: those that affect the stability of the competent state, affecting slow promoter fluctuations; those that regulate the burst frequency of transcription due to fast promoter fluctuations; and those that have both functions. Moreover, a single synthetic activator can mimic each of these three classes based on the location of its DNA binding site. Because the kinetic role of each regulator defines its impact on expression heterogeneity, this can be controlled by the choice of regulator class. Finally, ethanol controls the extent of gene silencing nearly independently of transcriptional activation through Flo8p, thereby dictating whether FLO11 expression responds in a graded or heterogeneous manner to other signals.

FLO11 switches between silenced and competent states independently at each locus
Under poor nutritional conditions, FLO11 is partially silenced in haploid cells and heterogeneous in expression. Members of this population are capable of reversibly transitioning between the OFF (silenced) and ON (competent) state [30]. To determine whether the transitions were due to a cis or trans fluctuation, we employed the dual-reporter assay, replacing the two copies of a FLO11 ORF in diploid yeast with a distinct fluorescent protein variant (Venus YFP and Cerulean CFP) ( Figure 2A). Importantly, we verified the independence and equivalence of the two reporters with respect to the presence of the other reporter ( Figure S2 and Figure S3). When grown in media with poor carbon sources, including ethanol, glycerol, galactose, and raffinose, we observed all four possible expression states ( Figure 2B, data not shown). Because endogenous Flo11p is not present in the dual-reporter strain, we verified that Flo11p did not affect expression at the FLO11 promoter in two ways. First, we added a plasmid constitutively expressing FLO11 and found no significant effect on fluorescent protein expression (data not shown). Second, we compared fluorescent protein expression in the dual-reporter strain to strains where only one FLO11 allele had been replaced with a fluorescent protein. There was no difference in expression levels ( Figure S2 and Figure S3).
If each allele switches independently, then at steady-state the proportion of cells in each expression state is given by p 2 (both ON), (12p) 2 (both OFF), or 2p(12p) (mixed ON/OFF and OFF/ON), where p is the proportion of cells with a particular allele ON. Note that p is identical for both YFP and CFP expression because the alleles are equivalent. We were able to verify the population's expression profile had reached steady-state ( Figure 3A and Figure S4). However, a naïve classification of expression state based on comparing a cell's fluorescence level to background is incorrect because it does not consider the long lifetime of the fluorescent proteins which obscures the true expression state of the promoter. Therefore, we directly measured the eight transition rates by real-time monitoring of FLO11 expression in single cells grown in a microfluidic chamber at constant conditions (Text S1 for details). All four ON to OFF and OFF to ON transition rates ( Figure 2C-2E) were found to be indistinguishable, demonstrating that each allele was switching independently. Furthermore, the fraction of cells turning ON or OFF are well-fit by a single exponential, confirming each transition is appropriately lumped as a pseudo-first order reaction. Switching was uncorrelated with cell-cycle stage ( Figure S7).

Using a stochastic kinetic model, static distributions can reveal kinetic information
Time lapse microscopy provides an accurate determination of the slow epigenetic transition rates and proportion of each expression  [37]. Binding sites are approximate and based on literature but most sites have not been confirmed directly. The three locations where a tetO sequence was inserted are also shown. See main text, Text S1, and Figure S1 for further details. doi:10.1371/journal.pgen.1000673.g001 state, but it is experimentally challenging and low throughput. Therefore, after determining that transition rates were accurately described as first order, we devised a way to infer these rates directly from static snapshots, accounting for the long lifetimes of the fluorescent reporters. Two-state models have been widely employed to model faster promoter fluctuations associated with conventional gene regulation [16,21,35]. In such models, the promoter can transition between an inactive but competent state and an active state that leads to transcription. Many eukaryotic genes appear to reside in the competent state, with rare transitions to the short-lived active state that result in a ''burst'' of transcription.
The observed variation in the FLO11 promoter can be divided into an intrinsic and extrinsic component. The FLO11 promoter is subject to both fast intrinsic fluctuations and a slow epigenetic transition, as depicted by the augmented three-state model in Figure 3B. Extrinsic noise also contributes to cell-to-cell variation in FLO11 expression levels when the promoter is not silenced. However, when the promoter is (partially) silenced, the predominant source of variation in FLO11 or reporter expression arises from the slow epigenetic transition between silenced to competent states because of (1) the smaller magnitude of the fast intrinsic and extrinsic fluctuations and the fact that (2) the faster fluctuations (,1 cell generation) are more completely time-averaged by the longlived reporter compared to the slow transition (.1 cell generation). Therefore, we can lump the fast transition rates (l9, c9, m9) into an overall transcription rate m and ignore extrinsic fluctuations. Gene expression can now be described using the commonly employed two-state model: , where x is the amount of reporter protein, m is the (lumped) protein production rate, d is a protein degradation rate (here the cell growth rate), and f(t) is a ''random telegraph process'' that takes values of 0 or 1 corresponding to a silenced or active promoter state, with exponentially distributed times between switching events ( Figure 3C). This stochastic equation has been solved analytically to yield a Beta distribution for protein number x at steady-state [35]. The slow epigenetic transition rates, l and c, correspond to those measured in the time lapse experiment. To infer these rates we assume our measured distribution of protein x is steady ( Figure S4) and fit it to the Beta distribution using a value of m based on the expression level of the ON population in a bimodal condition (the parameter d, the cell growth rate, is measured directly -see Text S1).
We tested this method in two different ways. First, we used the steady-state protein distribution of the time lapse experiments to estimate transition rates and found tight agreement between the inferred rates and those directly measured in time lapse ( Figure 3A and 3C). Second, this model allows proper estimation of the fraction of cells that appear ON in static distributions that are actually OFF because of the long lifetime of the fluorescent reporter (details in Text S1). We applied this correction to static snapshots of cells grown in different conditions. Although the fraction of cells in each expression state varied, the overall statistics were always consistent with independent switching at each promoter ( Figure 3D). Therefore, the upstream signaling network can map environmental inputs to a particular mixture of expression states through the modulation of transition rates.

A strategy for determining how regulators affect transition rates
Ultimately, environmental signals modulate epigenetic regulation of FLO11 through downstream regulators. The effect of these . Each marker represents the fraction of cells observed to switch at the particular time, and the pink curve is the same as the fit curve in Figure 3A. (E) As in (D) but for ON to OFF transition rates: c 1 (X), c 2 (X), c 3 (N), c 4 (N). The blue curve is the same as the fit curve in Figure 3A. (D) and (E) demonstrate that transition rates at one allele are independent of the state of the other allele. Even the null hypothesis that c 2 and c 4 are equivalent cannot be rejected at the 5% significance level (two-way T-test, p = 0.28) nor can the null hypothesis that their distributions are identical (two-way KS test, p = 0.47). doi:10.1371/journal.pgen.1000673.g002 regulators on both the mean level of expression and expression heterogeneity is succinctly and quantitatively described by their effect on the transition rates in the three-state model ( Figure 3B). Therefore, we decided to titrate trans factors and measure the quantitative response of the FLO11 promoter at the single cell level in hundreds of cells by fluorescence microscopy using the dualreporter assay. For each condition and strain, we always grew cells for .10 doublings, serially diluting them as needed to maintain low density and ensure a steady-state had been reached (further details in Text S1).
To obtain transition rates, we fit the measured fluorescence distributions arising from each titration to the Beta distribution, the solution to the simplified two-state model. As described previously, a two-state model which lumps the fast transitions is only strictly applicable when slow epigenetic transitions related to silencing dominate. The four quadrant plot in Figure  If epigenetic regulation is lost, expression levels can still change due to faster promoter fluctuations. A two-state model accurately describes intrinsic (but not extrinsic) fluctuations caused by transitions between the competent and active promoter states. For example, a conventionally regulated (but repressed) gene can be OFF and lie in the lower right quadrant. Activation leads to an increased burst frequency (l9 increases) and the graded, unimodal distribution of the upper right quadrant. Importantly, it is impossible to distinguish between conventional repression and epigenetic silencing in any population in the lower right quadrant  Figure S5 and Figure S6). (B) Three-state model of FLO11 activation showing separation of timescales between epigenetic (silencing) and conventional regulation. When slow transitions associated with silencing are present, the fast transitions of transcriptional bursting can be lumped into a single rate m. The model then collapses into the two-state model in (C). (C) (Left) Static distribution of YFP fluorescence of dual-reporter strain grown in identical media conditions as A but in deep well plates rather than the microfluidic device. Transition rates inferred from this snapshot using a stochastic kinetic model (right) agree closely with those obtained by timelapse microscopy. (D) Modulation of switching rates. The stochastic kinetic model's prediction of the fraction of cells in the mixed expression state corresponds to independent switching (given by 2p(12p), corresponding to the gray line) for a range of conditions. Error bars (x-axis) are from 95% confidence intervals from MLE fit of switching rate to estimate true fraction of ON cells; error bars (y-axis) are due to errors in the estimation of threshold fluorescence value for autofluorescence (see Text S1). doi:10.1371/journal.pgen.1000673.g003 where expression is completely OFF. Furthermore, fitting fluorescence expression distributions generated from a single FLO11 promoter driven reporter that is conventionally regulated does not yield the fast promoter transition rates l9 and c9 because here extrinsic fluctuations are significant. The extrinsic noise is due to cell-to-cell variation in factors like morphology, ribosome number, and/or upstream components in the FLO11 regulatory pathway and affects both promoters within the same cell in a correlated fashion. To properly measure the fast transition rates associated with conventional regulation, the intrinsic noise should be analyzed to determine the burst frequency (l9) and burst size (m9/c9) (see Text S1).

Three classes of regulators of FLO11 expression
To decouple complex upstream signaling events occurring at the promoter (Figure 1), individual trans factors were expressed heterologously under the control of a doxycycline-inducible promoter [36]. Because Sfl1p and Flo8p are post-translationally regulated, we needed a way to tune their relative strength. We chose ethanol, since the addition of ethanol activates FLO11 expression in a Flo8p-dependent manner (see below). All the trans factor titrations were performed in either SD ura-or SD ura-with ethanol. Titrations of Sfl1p, Flo8p, and Mss11p were done in a sfl1D, flo8D or mss11D background, respectively.
For each titration point, we inferred the transition rates l and c using the two-state model. To summarize the effect of various activators and Sfl1p on the stability of the silenced and competent states, we plot the series of (c,l) values determined on the four quadrant plot of Figure 4A. In SD ura-, FLO11 is OFF, corresponding to the lower right quadrant. Based on the response of the FLO11 promoter ( Figure 4B and 4C), we grouped the activators into 3 classes. Addition of three Class I activators, Tec1p, Ste12p, and Phd1p, appears to weakly stabilize the competent state, but expression remains extremely low. The Class II activators, Msn1p and Mss11p, stabilized the competent state by decreasing c and entering the heterogeneous region where slow promoter fluctuations dominate. At some critical c value, the silenced state is rapidly destabilized and the entire population turns ON. Flo8p constitutes a special class and was titrated in ethanol conditions where presumably some fraction of Flo8p is now phosphorylated and active. The population response is intermediate between Class I and Class II activators, but closer to Class II. Sfl1p has the exact opposite effect under the same ethanol conditions, consistent with the antagonistic role Sfl1p and Flo8p have through their overlapping binding sites and as being negatively and positively regulated by PKA, respectively.
To determine if activator class was correlated to binding site position or accessibility, we applied both an in silico nucleosomal occupancy model [37] and performed micrococcal nuclease mapping (Figure 1 and Figure S8) of the FLO11 promoter. Both techniques suggested the 21200 region containing overlapping binding sites for Sfl1p and Flo8p is nucleosome free. In contrast, binding sites of Class I activators occur in nucleosomally occluded regions. Class II activators are not known to bind DNA but are potent activators, with Msn1p having a known ability to recruit chromatin remodeling machinery [34].

A synthetic activator mimics each activator class depending on its binding site position
Are binding site position, accessibility and/or competition with Sfl1p sufficient to determine activator class? If so, a synthetic activator could have qualitatively different regulatory profiles depending on binding site position. We engineered dual-reporter yeast strains with the 19 bp tetO sequence inserted at 3 different locations within the FLO11 promoter. The first location was at 2350 in a nucleosomally occluded region close to the TATA box and transcriptional start site. The second location was at 21470, on the outer-edge of the nucleosome upstream of the Sfl1p binding site and far from the core promoter. The third location was at 21160, within a nucleosome-free region directly overlapping the Sfl1p binding site. Sfl1p binds as a dimer at two sites [27,38], so we replaced 19 bp of promoter sequence between the two sites with the tetO sequence to preserve the distance spanned by the two sites. The tetO is bound by rtTA, a synthetic activator that contains a strong acidic activation domain, VP16 [36], known to recruit the SAGA complex in yeast [39]. We titrated rtTA in these three strains grown in SD ura-(all initially OFF). Each location functionally mimics the response of the respective class of activators ( Figure 4B and 4D). Importantly, the silenced state stability is reduced for the third tetO location compared to the second tetO location. This occurred even in the absence of ethanol, suggesting that the difference in silenced state stability between Class II activators and Flo8p is not due to an alternative ethanol-specific effect.

Two modes of Sfl1p repression correspond to a graded or heterogeneous response
When performing the Sfl1p titration above ( Figure 4C and Figure 5A), we did so in a sfl1D background in ethanol. FLO11 was highly expressed (upper left quadrant) in the sfl1D background, as has been shown previously [26]. Surprisingly, in the Sfl1p titration in SD ura-media without ethanol, both promoters turned off in a graded fashion ( Figure 5B). This was in contrast to the heterogeneous population response observed for titrations of Class II activators (Figure 4) including Sfl1p titrations performed in ethanol ( Figure 5A), a condition where Flo8p is presumably more active. To explain this result, we hypothesized that a critical Sfl1p level is required to silence the promoter, and below this level Sfl1p still repressed transcription but in a conventional manner associated with faster promoter fluctuations. The model requires that Sfl1p is able to repress the FLO11 promoter as a silencer or as a conventional repressor; evidence exists for both modes [38,40].
To test this model, we generated a dual-reporter strain in an hda1D background and added back Sfl1p heterologously. When we titrated Sfl1p in this background in SD ura-media, we observed a graded response ( Figure 5C). This establishes that Sfl1p is capable of repressing expression in an Hda1p-independent manner. The graded response suggests Sfl1p is working as a conventional repressor rather than affecting the slower fluctuations between the epigenetically silenced and competent promoter states. To further demonstrate that Hda1p is necessary to silence FLO11, we measured the average H3 and H4 histone acetylation state at the FLO11 promoter by chromatin-IP in the dual-reporter strain grown in SD ura-( Figure 5D). Only the wildtype FLO11 promoter exhibited a hypoacetylated state, indicative of silenced chromatin [41] and similar to a silenced telomeric region. In both the hda1D and sfl1D backgrounds, the promoter was hyperacetylated, a chromatin state associated with lack of silencing. Together, this demonstrates that Sfl1p silences the promoter in an Hda1pdependent manner and the silenced state at FLO11 is correlated with hypoacetylation in at least one region (,21600 bp) of the promoter.
If silencing is eliminated in an hda1D background, titration of activators in the presence of high levels of Sfl1p will result in a graded response, since Sfl1p is now functioning as a conventional repressor. In addition, the threshold level of activator required to turn on FLO11 will be lower. We performed these titrations (shown for Tec1p in Figure 5E and 5F and other activators in Figure S8) and confirmed this prediction. In addition, the intrinsic noise of these strains was proportional to the inverse square root of the mean expression level ( Figure 5F inset and Figure S8), indicating that without slow promoter fluctuations, the activators regulate the burst frequency, l9 [19].

Ethanol controls the importance of silencing in a Flo8pdependent manner
To further understand the role of ethanol in FLO11 signaling, we grew the dual-reporter strain in SD ura-media in a range of ethanol concentrations. FLO11 expression exhibited a graded response to increasing ethanol levels, but the average expression level remained low even at the highest (3%) ethanol concentrations ( Figure 6A). The graded response suggested a lack of silencing possibly due to increased Flo8p activity. This led to the hypothesis that at low levels of ethanol (,1%), Flo8p activity eliminates Sfl1pmediated silencing, but has little effect on expression. Therefore, although FLO11 expression is OFF (lower right quadrant), the promoter is actually in the competent state. To test this idea, we titrated Class I and II activators in 1% ethanol. Both classes were capable of increasing FLO11 expression to high levels in a graded manner ( Figure 6B), implying that silencing no longer occurred in these conditions. The corresponding synthetic activators had a similar effect. In contrast, titration of the synthetic activator mimicking Flo8p resulted in a response similar to ethanol ( Figure 6C). Finally, Class I and II activator titrations in a flo8D in SD ura-with 1% ethanol ( Figure 6D) reverted to the behavior seen in SD ura-conditions. Therefore, ethanol controls the extent of silencing at FLO11 in a Flo8p-dependent manner. Both Flo8p and its synthetic analog affect slow promoter transitions, but neither is a strong conventional activator.

Discussion
The main results of our work are perhaps best understood with reference to the 3-state model in Figure 7, a simple, but useful paradigm for describing the kinetics of epigenetic gene regulation. Regulators can affect either the slow transition rates associated with epigenetic silencing or the fast transition rates associated with conventional gene activation. While a good deal is known about how regulators affect the fast transition rates, less is known about how regulators affect the slow transition rates, and whether effects on slow and fast transition rates are coupled. Our work demonstrates that slow transitions at the two copies of the FLO11 promoter in diploid yeast occur randomly and independently. Furthermore, we identify the role of various trans regulators of FLO11 in controlling both slow and fast transition rates; it appears that this control is distributed among various ''classes'' of regulators. Importantly, distributed control enables the cell to shape the diversity of FLO11 expression within an isogenic population by utilizing different combinations of regulators.
Our demonstration that two copies of the FLO11 promoter switch slowly and independently builds on previous work demonstrating (1) the FLO11 promoter is epigenetically regulated [30], and (2) two copies of a partially silenced URA3 promoter inserted at the mating type loci in S. cerevisiae switch independently in a sir1 background [24]. We now provide evidence that independent switching occurs in a natural gene whose epigenetic regulation is not SIR protein dependent. Given the rich diversity of the FLO11 gene pool [29] independent switching may be an additional mechanism for generating variation in adhesive phenotype.
To understand how expression heterogeneity at FLO11 is controlled, we took a functional approach and determined the kinetic role of different regulators on both slow and fast promoter fluctuations. Class I regulates fast promoter fluctuations exclusively. Intrinsic noise measurements confirmed these regulators destabilize the competent state and increase the burst frequency, a common theme for regulators in yeast [19,20]. Interestingly, we find that the Class II and Flo8p regulate slow promoter fluctuations primarily by stabilizing the competent state. The previous study at the mating type loci found an activator, Ppr1p, could challenge the silenced (D) ChIP assay probing for acetylated H3 and H4 histones at FLO11 promoter (strains grown in SD complete or SD leu-). Probes amplified the 21.7 to 21.5 kb region of the promoter. Signal (y-axis) represents anti-acetylated histone/anti-histone ratio, or an effective average acetylation per histone in the region. Deletion of both sfl1 and hda1 result in hyperacetylation of the FLO11 promoter which is associated with the abrogation of silencing. Therefore SFL1-dependent silencing at FLO11 requires HDA1. Error bars are standard error of triplicate quantitative PCR samples. (E) When the activator Tec1p is titrated in an hda1D background, the response is also graded (SD ura-). (F) Mean levels of expression during Sfl1p and Tec1p titrations in wildtype (blue curve) and hda1D (pink curve) backgrounds. Elimination of silencing because of lack of Hda1p lowers the threshold level at which activators function. Furthermore, the population response is graded (C and E, see Figure S8 for other activator titrations). Error bars represent 3 s.d. around mean calculated from bootstrap analysis. Inset: The square of intrinsic noise of Sfl1p (left) and Tec1p (right) titrated in hda1D is proportional to the reciprocal of protein abundance (here shown as the mean fluorescence level). This indicates that without silencing, regulators modulate expression by controlling burst frequency (l'). doi:10.1371/journal.pgen.1000673.g005 state and affected both slow transition rates [24], but this was only measured at one level of Ppr1p. It is only by titrating regulators that we were able to clearly discern their functional roles. Whether activators generally affect one or both of the transition rates of silenced genes remains an open question, but likely depends on the mechanism of silencing.
While the regulator titrations indicate that ethanol desilences the promoter via Flo8p, the pathway(s) by which ethanol activates Flo8p is unknown. The simplest mechanism is that long term growth in ethanol activates PKA (specifically Tpk2p), which then activates Flo8p and deactivates Sfl1p. In fact, glucose is not required as the promoter response when Flo8p was titrated in synthetic media with ethanol or glycerol as the sole carbon source was similar to that in SD 1% ethanol (data not shown). However, activation of the PKA pathway via ethanol has not been reported. It is also possible that Flo8p is activated by ethanol via another pathway independent of PKA, although to our knowledge, it is not known to be the target of any other kinase. A third possibility is that ethanol may enable synergistic interactions between Flo8p and other regulators that leads to desilencing, although there is no evidence of Flo8p interacting with any Class I activators. These possibilities could be distinguished by monitoring cyclic AMP levels and the phosphorylation status of Flo8p and Sfl1p in ethanol.
Our results lend support to the idea that binding site position within the FLO11 promoter can largely determine the kinetic role of the transcriptional regulator. However, the mechanistic description of how binding to particular sites affects slow epigenetic regulation and fast conventional regulation and the molecular nature of the silenced and competent promoter states is still unclear. A mechanistic explanation for the dual roles of Sfl1p is likely the clearest. Binding in the 21200 nucleosome-free region governs epigenetic silencing, possibly by recruiting Hda1p via Tup1p/Ssn6p corepressor [38]. Conventional regulation most likely occurs via Sfl1p binding to the 2200 region which contains a putative Sfl1p binding site and has been shown to bind Sfl1p in vitro. Indeed, preliminary ChIP experiments suggest Sfl1p is bound to this region in vivo (Octavio and Maheshri, unpublished results).
Among the activators, the role of Flo8p and its synthetic analog is perhaps the clearest. Flo8p binding and Sfl1p binding at the 21200 region are likely mutually exclusive because of overlapping binding sites, and so Flo8p can prevent the Sfl1p-mediated establishment of silencing but probably not directly affect conventional Sfl1p repression. This would explain the ability of Flo8p and its corresponding synthetic analog to affect the slow epigenetic transition independently of the conventional activation. In fact, binding of tetR, which lacks the VP16 activation domain of rtTA, to the 21160 tetO site is sufficient to abrograte silencing (data not shown), implying steric hindrance is the major mode of action. Furthermore, any weak activation via Flo8p might be through its known role in binding to the promoters of other Class I activators (including the ones tested here) and presumably upregulating their expression. In this manner, Flo8p activation can put the FLO11 promoter in a competent, ''poised'' state whose expression can be controlled by Class I -like regulators. There is evidence, though, that Flo8p can bind to other regions of the promoter [31].
Several possibilities exist for the inability of the Class I activators to challenge the silenced state, yet still regulate the burst frequency of the competent FLO11 promoter. For example, binding site proximity to the transcriptional start site could play a dominant role. Canonical yeast promoters are typically 150-400 bp with transcription factor binding sites are clustered 100-200 bp from the transcriptional start site [42]; this proximity allows direct interaction with general transcriptional machinery. Therefore, it may be that Class I activators bind in the core region of the FLO11 promoter, a region that may be inaccessible to transcription factors and/or transcriptional machinery in the silenced state. However, while the Class I synthetic analog binding site is at 2350 in a nucleosome occluded region, not all Class I activators have binding sites in this region [31].
A second possibility not mutually exclusive with the first is that Class I activators need not bind in the downstream region but can influence transcription rates via long range (but fast) mechanisms including DNA looping, cryptic transcription, or long range chromatin remodeling. This would provide an explanation for the presence of Class I activator binding sites in these regions that are known to be bound in vivo in activating conditions [31]. In addition, it might explain why even at high levels of expression in the absence of silencing the intrinsic noise at the FLO11 promoter is 10 times higher than that of a similarly highly expressed PHO84 promoter (data not shown). However, with either explanation, the inability of Class I activators to challenge the silenced state is not clear. Altered chromatin structure or reduced binding site accessibility could be invoked as Class I activator sites tend to be under nucleosomes. However, other than some nucleosome depletion in the core promoter and the 21300 region, no gross nucleosomal rearrangements seem to occur upon silencing ( Figure  S9) although higher resolution mapping may reveal finer differences.
While we do not know the biochemical intermediates during the slow promoter transitions, the pseudo-first order rates suggest a single slow step, rather than a distributed control mechanism. This is similar to both the partially silenced mating type loci [24] and the epigenetically regulated agn43 gene in E. coli [43]. Possibilities for the slow epigenetic step governing ON to OFF might include Sfl1p binding or Sfl1p-mediated recruitment of silencing factors, among others. Both Class II activators Msn1p and Mss11p are capable of stabilizing the active state, but their localization and activity with respect to the FLO11 promoter remains unclear [32,34]. The ability of the Class II synthetic analog to have a similar stabilizing effect on the ON state by binding to the 21470 region as well as the differential acetylation state of that region strongly suggests chromatin remodeling in the upstream region affects accessibility of Class I activators and the transition to the competent state. High resolution mapping of the chromatin state of the entire FLO11 promoter under various conditions should point toward the biochemical mechanism of the slow promoter transition and will be the focus of future work.
Our findings have implications for the regulation of various subtelomerically encoded gene families known to be epigenetically regulated. This includes the FLO gene family [30] and other closely related yeast adhesins [25] such as the EPA gene family in the pathogenic yeast C. glabrata [44]. Phenotypic variability in EPA gene expression might allow C. glabrata to rapidly colonize new host tissues and evade immune surveillance. Do such genes turn ON and OFF independently, does it depend on the mechanisms of their silencing (SIR-dependent, etc.), their relative chromosomal locations, or the presence of boundary elements? What promoter transitions do trans factors regulate? This understanding will allow the engineering of strains with well-defined levels of phenotypic heterogeneity. Such strains are a prerequisite to quantify what role if any phenotypic heterogeneity has on organismal fitness in natural environments.

Yeast strains and media
To use the dual-reporter assay to study switching of the FLO11 promoter, we replaced the FLO11 ORF in a haploid S1278b from the Heitman laboratory [45] with YFP-KanMX6 or CFP-KanMX6 cassettes by PCR integration, and then mated to create diploids. All strains and plasmids used are provided in Table S1 and Table  S2.
SD is synthetic defined media with 2% glucose. SD ura-media lacks uracil. SD ura-media with ethanol contains both 2% glucose and a specified amount of ethanol.

Single cell measurements and analysis
Cells from overnight cultures grown in SD ura-or SD ura-+ ethanol were inoculated at an initial OD 600 between 0.005 and 0.01, and grown for 15-20 hours in the same media. For the titration experiments, these cultures were treated with serial dilutions of doxycyline (0 to 10000 ng/ml) at 30uC in well-agitated deep well 96-well plates. Cells were harvested in mid-late log phase (OD 600 between 0.5 and 1.5), and placed on ice while other samples were being processed. Expression was measured using a Zeiss AxioObserver microscope with filters optimized for yECitrine, mCherry, and Cerulean (Chroma). Metamorph software (Molecular Devices) was used to analyze images and quantify single cell YFP and CFP fluorescence. Between 500 and 1500 cells were imaged for each sample. Fluorescence levels in the RFP channel was used to discard dead cells (usually ,5% of population). Details of data preprocessing and estimation of l and c are given in the Text S1.
In timelapse microscopy experiments, cells were loaded onto the ONIX Microfluidic Platform (CellASIC) with initially ,10 cells trapped in individual chambers. Media in the ,10 25 ml chambers was constantly replenished at a rate of 10 ml/hr. YFP, CFP, RFP fluorescence and bright field images were obtained every 15 minutes for 20 hours. An example is provided in Video S1. Image stacks were segmented using custom Metamorph journals. Single cell tracking and fluorescence was determined using custom MATLAB routines.

Chromatin immunoprecipitation assay
Chromatin IP's were done based on the method of [46]. Briefly, ,40 ml of cells were grown at 30uC in either SD complete or SD leu-to an OD 600 = 0.8. Lysates from the fixed cells were sonicated to shear the chromatin to an average length of 500 bp, and isolated chromatin was incubated with 2 ml antibodies (Upstate/ Millipore) against either histone H3 (Cat. No. 05-928), histone H4 (Cat. No. 05-858), acetylated histone H3 (Cat. No. 07-593) and acetylated histone H4 (Cat. No. 06-866). A sample with no antibodies was also prepared as a control. After reversal of crosslinks, DNA from immunoprecipated chromatin was purified and analyzed using quantitative PCR (Applied Biosystems). Primers amplifying the 21.7 to 21.5 kb region of the FLO11 promoter and primers amplifying a telomeric region in the right arm of chromosome VI [47] as a control for hypoacetylated histone signals were used. Applied Biosystems 7300 software was used to obtain cycle threshold values. All signals from experimental samples were quantified relative to signal from a known amount of genomic DNA from an unmodified, cogenic S1287b strain (MLY43, Table S1) that served as a positive control. The ratio of anti-histone/anti-acetylated histone signals was used as a measure of average H3 and H4 acetylation in the region.

Micrococcal nuclease assay
Micrococcal nuclease assays were performed as in [48]. All cells were grown at 30uC in either SD complete or SD leu-to OD 600 = 0.5. Figure S1 Nucleosome occupancy data from Figure S9 is plotted along with predicted nucleosome positions from studies predicting nucleosome position genome-wide prediction. The Kaplan et al [17] study provides a computational prediction of nucleosome positioning on any sequence. Only a few nucleosome poor regions bordered by well-defined nucleosomes are predicted, including the 21200 region. The Mavrich et al [18] study used a statistical model to analyze their experimental genome-wide nucleosomal occupancy data. The x-error bars denote how fuzzy the position is. The relative occupancy scale is arbitrary, and absolute number cannot be compared between datasets. However, some general trends in positioned and fuzzy nucleosomes are apparent and this was the basis of Figure 1 in the main text. Found at: doi:10.1371/journal.pgen.1000673.s001 (0.46 MB TIF) Figure S2 FLO11 expression on solid media -single reporter strains. 10 mL of a mid-log phase culture of various diploid strains with single reporters were spotted on a fresh YPD plate. Labels on the left indicate whether and at which locus the FLO11 ORF was replaced with a particular fluorescent protein variant. Plates were left at room temperature. Cells from all regions of the spot were sampled (see Text S1) and CFP and YFP expression of these samples was monitored every 2 days by fluorescence microscopy. Density plots for each sample are given, where the x-axis is log CFP fluorescence levels and the y-axis is log YFP fluorescence levels. Cellular autofluorescence can be estimated based on the (first) control strain. Both fluorescence reporters respond equivalently, whether integrated at the A or a locus. Found at: doi:10.1371/journal.pgen.1000673.s002 (1.26 MB TIF) Figure S3 FLO11 expression on solid media -dual reporter strains. As in panel Figure S2. Four different dual reporter strains were constructed, two with CFP at the a locus and YFP at the A locus, and two in the opposite configuration. Their response is similar, verifying that both reporters are equivalent. Furthermore, the expression distribution of each individual fluorescent reporter is equivalent to the corresponding single reporter strain in Figure  S2, verifying independence and the fact that FLO11 expression doesn't feedback and affect its own expression. Found at: doi:10.1371/journal.pgen.1000673.s003 (1.28 MB TIF) Figure S4 Static snapshots of Y45 cells in YP 1% Ethanol, 2% Glycerol maintained in exponential phase by dilution. The null hypothesis that distributions of YFP fluorescence at each time point are equivalent cannot be rejected (two-way Kolmogorov-Smirnov test, p = 0.60, 0.25, 0.85 for day 1 vs. day 2, day 2 vs. day 3 and day 1 vs. day 3 respectively). Similarly, the null hypothesis that CFP fluorescence distributions at each time point are equivalent cannot be rejected (two-way Kolmogorov-Smirnov test, p = 0.66, 0.36, 0.88 for day 1 vs. day 2, day 2 vs. day 3 and day 1 vs. day 3 respectively).  Figure 4B and Figure 5E in the main text where Tec1p was titrated in a wildtype (Y45) and hda1D background respectively, the other activators Phd1p and Mss11p also exhibit a graded response in an hda1D background (A and B). In the wildtype background, Phd1p (C) like Tec1p is unable to stabilize the ON state enough to enter the bimodal regime, whereas Mss11p (D), like Msn1p (shown in Figure 4B), is able to do so. As in Figure 5F, elimination of silencing in the hda1D background lowers the threshold level at which Mss11p and Phd1p function (E and F). Error bars represent 3 standard deviations around the mean from bootstrap analysis. Like Tec1p in hda1D in Figure 5F, both Mss11p and Phd1p control burst frequency (l9) in the absence of silencing, as the square of the intrinsic noise of Phd1p titrated in hda1D (G) and Mss11p titrated in hda1D (H) scale with the reciprocal of protein abundance. All titrations were done in SD ura-. Found at: doi:10.1371/journal.pgen.1000673.s008 (3.07 MB TIF) Figure S9 Micrococcal nuclease mapping of FLO11 was performed on cells grown in conditions where the promoter was completely silenced (growth of wildtype Y45 in SD complete) in (A) or completely active (growth of an sfl1D strain in SD complete plus 2% glucose) in (B). The overall structure agrees well with the in silico predictions in Figure 1. Although nucleosomal occupancy in the 21300 bp region and the 2150 bp region appears to be further depleted in the active state, there is surprisingly no gross rearrangement of nucleosomal structure between the silenced and active state. Error bars are standard error from triplicate quantitative PCR samples.   Video S1 Movie of cell growth within the microfluidic device. Red and green represent CFP and YFP expression respectively. Both ON to OFF and OFF to ON transitions for each copy of the FLO11 promoter are readily seen in this video. Each frame represents 15 minutes elapsed and the total movie represents 20 hours of growth. Found at: doi:10.1371/journal.pgen.1000673.s013 (3.83 MB AVI)