Existence, Transition, and Propagation of Intermediate Silencing States in Ribosomal DNA

The MET3 promoter (MET3pr) inserted into the silenced chromosome in budding yeast can overcome Sir2-dependent silencing upon induction and activate transcription in every single cell among a population. Despite the fact that MET3pr is turned on in all the cells, its activity still shows very high cell-to-cell variability.

ABSTRACT

The MET3 promoter (MET3pr) inserted into the silenced chromosome in budding yeast can overcome Sir2-dependent silencing upon induction and activate transcription in every single cell among a population. Despite the fact that MET3pr is turned on in all the cells, its activity still shows very high cell-to-cell variability. To understand the nature of such “gene expression noise,” we followed the dynamics of the MET3pr-GFP expression inserted into ribosomal DNA (rDNA) using time-lapse microscopy. We found that the noisy “on” state is comprised of multiple substable states with discrete expression levels. These intermediate states stochastically transition between each other, with “up” transitions among different activated states occurring exclusively near the mitotic exit and “down” transitions occurring throughout the rest of the cell cycle. Such cell cycle dependence likely reflects the dynamic activity of the rDNA-specific RENT complex, as MET3pr-GFP expression in a telomeric locus does not have the same cell cycle dependence. The MET3pr-GFP expression in rDNA is highly correlated in mother and daughter cells after cell division, indicating that the silenced state in the mother cell is inherited in daughter cells. These states are disrupted by a brief repression and reset upon a second activation. Potential mechanisms behind these observations are further discussed.

INTRODUCTION

Gene expression is intrinsically variable from cell to cell due to the low copy number of DNAs and mRNAs in a single cell. Such variability, or “noise,” of gene expression can lead to significant physiological consequences (1–3). The noise level also reflects the underlying gene regulatory pathway (1–3). Elucidating the sources and the control mechanisms of gene expression noise is thus an important aspect of understanding gene regulation.

A well-characterized mechanism of gene repression in budding yeast is Sir2-dependent silencing. Transcriptional silencing occurs at HML/HMR, the telomere, and ribosomal DNA (rDNA) repeats through inhibitory chromatin configurations mediated by histone deacetylase Sir2 and associated complexes (the Sir2-4 complex at mating-type loci and telomere and the RENT complex at rDNA) (4, 5). The physical association and the chemical modifications generated by these complexes induce a repressive chromatin structure that blocks the assembly of DNA polymerase II (Pol II) and/or a downstream step and attenuate transcription (6, 7).

Besides reducing the average expression in a population of cells, silencing was also found to elevate the expression noise by generating bimodal, “on” or “off,” expression in individual cells (8–10). In “telomeric position effects,” for example, the expression of a reporter gene inserted into subtelomeric regions can switch stochastically between “on” and “off” states, and each state can propagate stably for multiple cell generations (8). As a result, cells containing the ADE2 gene near the telomere grow into white (ADE2 on) or red (ADE2 off) subpopulations; with URA3 at a similar locus, the cells are either uracil independent (URA3 on) or 5-fluoroorotic acid (FOA) resistant (URA3 off) (8). Besides these phenotypic assay, the bimodal URA3 expression can be directly measured through fluorescence microscopy or flow cytometry (11). URA3 inserted into HML or HMR was completely silenced in wild-type cells but also exhibited bimodal expression upon mutation of certain silencing factors, such as Sir1 (12). Interestingly, the “on” state is likely to be different from the “unsilenced” state: by isolating uniformly “on” or “off” cell populations and probing the chromosome configurations near the telomeric URA3, it was found that the “on” cells have distinct chromatin state that lies in between the fully silent and fully derepressed chromatin (11). The “on” telomeres are still associated with some silencing factors such as Sir3 and hypoacetylated H4, but they lack other features such as the depletion of H3K79 methylation. Correspondingly, the “on” state gene expression may have characteristics different from those of expression in regular euchromatin. Such expression is not well characterized, especially at the single-cell level.

In our previous work (13), we generated strains with MET3 promoter (MET3pr) driving the green fluorescent protein (GFP) gene inserted into all three silenced regions. MET3pr is an inducible promoter that is activated by Met4 when methionine is depleted. Consistent with previous findings that strong transcription activation overcomes silencing (14), MET3pr can be activated in these regions with an expression level lower than that in euchromatin. Our subsequent single-cell analysis revealed that the MET3pr-GFP is expressed in all cells. Despite the fact that the whole cell population is converted to the “on” state upon induction, the steady-state GFP level still shows unusually high noise. In particular, compared with transcriptional interference, another mechanism of gene repression, silencing can reduce the MET3pr-GFP expression to a similar extent but generate much higher cell-to-cell variability (13). The mechanism of such elevated noise is not understood.

We previously measured the MET3pr-GFP expression using flow cytometry by taking “snapshots” of a population of cells at a single time point. In this work, we followed the activation, repression, and steady state of the MET3pr-GFP expression inserted into rDNA using time-lapse fluorescence microscopy. We found that the noisy “on” state is in fact a collection of multiple substable, partially silenced states with discrete expression levels. GAL1Spr-GFP inserted into the same rDNA locus and MET3pr-GFP inserted into other Sir2-silenced regions also show multimodality in their expression. These intermediate states stochastically transition between each other, with “up” transitions occurring exclusively near the mitotic exit and “down” transitions occurring throughout the rest of the cell cycle. These states are likely to be inherited in daughter cells, as the GFP expression states in mother and daughter cells are highly correlated for a few hours after cell division. These states are disrupted by a brief repression and reset upon a second activation. Potential mechanisms behind these observations are further discussed.

RESULTS

MET3pr-GFP expression in rDNA shows higher variability among single cells.

We constructed two diploid strains containing a single copy of MET3pr-GFP inserted into either silenced (NTS1 in rDNA) or euchromatic (ATG36 open reading frame [ORF]) regions. We monitored the GFP expression during the MET3pr activation by combining time-lapse fluorescence microscopy with a microfluid device (see Materials and Methods) (15). As a control, these strains also contain a MET3pr-mCherry reporter at CDC20 in euchromatin. As shown in Fig. 1A to D, GFP is induced in both strains upon the depletion of methionine. However, while the expression of GFP at the ATG36 locus is roughly uniform across the cell population, GFP in rDNA shows very high cell-to-cell variability.

FIG 1.

MET3pr-GFP expression in rDNA shows higher variability among single cells. (A) Time-lapse images of a diploid strain containing a single copy of MET3pr-mCherry at the CDC20 locus and MET3pr-GFP at the ATG36 locus (both inside euchromatin). (B) Quantification of the mCherry (orange) and GFP (green) expression upon induction in the strain used for panel A. The cells were first incubated in 10× methionine medium before switching to 0× methionine medium at 0 h. Each trace represents the fluorescence measured in a single cell (total number of cells, 104). (C and D) The same as for panels A and B, respectively, except in a strain with mCherry at the same locus but GFP in rDNA. These cells show high cell-to-cell variability in the GFP expression (total number of cells, 105). (E and F) Correlation of mCherry and GFP expression levels in individual cells of the strain in panel A (E) and strain in panel C (F). Each dot represents the average steady-state GFP versus mCherry intensity (4 h after induction) of one cell. The correlation coefficients are 0.616 and 0.273, respectively. (G) The same as for panels C and D except both mCherry and GFP are in allelic loci in rDNA. (H) Correlation of mCherry and GFP expression levels in the strain used for panel G. The correlation coefficient is 0.034.

We quantified the GFP and mCherry fluorescence intensities as a function of time in each single cell (sample traces are plotted in Fig. 1B and D). GFP intensity increases sharply after ∼30 min of induction and reaches steady state after ∼3 h. The increase in the mCherry intensity slightly lags behind that of GFP due to a lower maturation rate (15). The mCherry expression levels in Fig. 1B and D are comparable, showing that the two strains have similar degrees of the methionine starvation response (e.g., similar concentrations of the transcriptional activator Met4). The average expression of GFP is ∼60% lower in the rDNA region with much higher cell-to-cell variability. This partial silencing completely relies on SIR2, as SIR2 deletion restores the MET3pr activity to the euchromatic level (see Fig. S1 in the supplemental material). Importantly, the expression of GFP in rDNA has low correlation with that of mCherry in the same cell (Fig. 1E and F), indicating that the variation in the GFP expression is not due to global variables such as cell size or fluctuations of transcription factor concentration (“extrinsic noise”) but is more related to local features such as variations in the rDNA chromatin structure (“intrinsic noise”) (16).

To further prove that the variability in GFP expression is indeed intrinsic, we generated another strain with MET3pr driving mCherry and GFP inserted into the allelic NTS1 loci (Fig. 1G). Similar to the case for GFP, mCherry expression is partially repressed with high variability. Again, the correlation between mCherry and GFP expression in the same cell is very low (Fig. 1H) (R = 0.034), confirming that the gene expression variation observed on these two alleles is caused by local and independent factors.

Comparison of the GFP activation rates in rDNA and in the euchromatic region.

GFP activation in the rDNA region involves competition between the silencing and activating factors. It was proposed that the silent chromatin is a relatively stable structure that resists activation to some degree (14). We therefore compared the MET3pr-GFP activation times in rDNA and euchromatin (see Materials and Methods). When induced by medium containing no methionine (maximum activation strength), the activation of MET3pr in rDNA is only slightly slower than that in euchromatin (Fig. 2A and B), showing that silencing causes only a minor delay in the activation. Another variable that may affect silencing strength is the cell cycle. Telomeric silencing, for example, is more likely to be antagonized after the end of S phase due to the disruption of silent chromatin during DNA replication (14, 17). If such a trend is also present in rDNA silencing, the activation time of MET3pr-GFP may show cell cycle dependence. We therefore constructed a new strain containing a cell cycle indicator, Myo1-mCherry (18). This protein forms a ring around the yeast bud neck from S to M phase, and its disappearance marks the time of cell division (18). After aligning the activation curves relative to the cell division time (only mother cycles were aligned, as mother and daughter have different G₁ lengths), we found that the activation can occur at any cell cycle stage, with the activation time uniformly distributed over the cell cycle (Fig. 2C and D).

FIG 2.

Comparison of the GFP activation rate in rDNA versus the euchromatic region. (A) Sample traces of the MET3pr-GFP expression in the early stage of induction by 0× methionine medium. MET3pr-GFP is located in the euchromatic (blue; n = 193) or rDNA (red; n = 187) region (the same color scheme applies for other panels). The crosses mark the starting point of the GFP activation. (B) Box plot of the GFP activation time in the rDNA versus the euchromatic region. GFP in both regions is activated within 1 h. (C) Sample traces of the MET3pr-GFP expression aligned by the cell cycle (0 marks the time of cytokinesis, as in other panels). The curves were deliberately separated along the y axis for visualization purpose. (D) Histogram of the GFP activation time across the cell cycle for both regions (n = 65 and 82 cycles for euchromatin and rDNA, respectively), focusing on mother cells. The activation can happen at any cell cycle stage, and there is no significant difference between the MET3pr located in euchromatin or rDNA. Error bars were calculated from $\sqrt{p(1-p)\hspace{0.17em}/\hspace{0.17em}N},$ where p is the probability and N is number of cell cycles. (E to H) The same as for panels A to D, respectively, except that cells were activated by medium containing 0.02× methionine (n = 79 and 89 for euchromatic and rDNA regions, respectively). The activation time in the euchromatic region is not affected, while that in rDNA region tends to be slower with broader distribution. The cell cycle distributions of the activation time in the two regions remain the same (n = 27 and 45 cycles for euchromatic and rDNA, respectively).

As activation needs to compete with silencing, the repression effect may be more severe with lowered activation strength. To test this idea, we dosed the induction medium with 0.02× methionine (0.4 mg/liter), which lowers the MET3pr activity in both euchromatic and rDNA regions (see Fig. S2 in the supplemental material). Under this condition, MET3pr shows a much-delayed activation in rDNA compared to euchromatin (P < 10⁻²²) (Fig. 2E and F). These results are consistent with the notion that silencing is more effective against weaker activation. Even with the lowered induction, the activation still occurs rather uniformly across all cell cycle stages (Fig. 2G and H), indicating that the variability of the silencing strength over the cell cycle, if any, is not large enough to cause a significant differential in the activation time.

Multiple intermediate silencing states exist with discrete expression levels.

To understand the MET3pr-GFP expression noise during steady state, we next carried out further analysis of the GFP traces after 4 h of induction. We smoothed the GFP traces and examined the distribution of the GFP intensity at different time points. When located in euchromatin, the histogram of the MET3pr-GFP intensity shows a single Gaussian distribution (Fig. 3A and B). In contrast, the GFP steady-state level in rDNA is better fitted by multiple Gaussian peaks (Fig. 3C and D; see Fig. S3 in the supplemental material). The multimodal distribution is particularly obvious at early time points in the steady state (Fig. 3D), and we suspect that this is because stochastic GFP expression builds up over time and causes mixing of different peaks. As different steady-state GFP levels reflect variable rates of GFP transcription, these data suggest that there are discrete intermediate silencing states.

FIG 3.

Multiple intermediate silencing states exist with discrete expression levels. (A and C) Sample traces of the smoothed GFP traces in the euchromatic (A) or rDNA (C) region induced by 0× methionine medium. (B and D) Histograms of GFP expression levels in panels A and C, respectively, at different time points (marked by the bars in panels A and C). The distributions are multimodal (*, peak location). (E and F) Same as for panel D except that GFP expression was induced by medium containing 0.005× (E) or 0.01× methionine (F). (G and H) Same as for panels C and D, respectively, except that GFP is driven by Gal1spr. (I and J) Same as for panels C and D, respectively, except that the MET3pr-GFP is integrated into a different rDNA locus (RDN37-1). (K and L) Same as for panels C and D, respectively, except that the MET3pr-GFP is integrated into a telomere locus.

To understand if the multimodality is a general phenomenon, we extended our analysis to more strains and conditions. First, we analyzed the steady-state expression of MET3pr-GFP at the rDNA NTS1 locus induced with 0.005× (0.1 mg/liter) or 0.01× (0.2 mg/liter) methionine. The histograms of GFP intensity again show multiple peaks (Fig. 3E and F). The peaks with lower GFP intensities become more populated as the methionine concentration increases and the activation strength decreases (Fig. 3E and F). Second, to test whether this phenomenon is specific to the MET3pr, we constructed a strain with Gal1spr-GFP reporter integrated into the same rDNA locus. Similar high noise and multimodality are found in the Gal1spr steady-state expression level, indicating that such a phenomenon is general among different promoters (Fig. 3G and H). Third, we constructed a strain with the same MET3pr-GFP reporter integrated into another rDNA locus (RDN37-1) or a telomere locus (see Materials and Methods). The GFP expression in RDN37-1 shows lower peaks with a long tail at high expression levels (Fig. 3I and J). GFP expression in the telomere shows multiple peaks with slightly higher levels than those in the rDNA (Fig. 3K and L). Given that Sir2 is recruited to NTS1, RDN37-1, and the telomere through different mechanisms (19, 20), these results indicate that the multimodality is a general property of Sir2-mediated silencing.

Cell cycle dependence of the transitions between intermediate states.

With multiple substable states, some cells may experience “transition” from one state to another, causing large changes in the GFP expression level. The rDNA GFP traces typically exhibit two types of fluctuations, cell cycle-dependent oscillations and unidirectional shifts in GFP intensity (examples of traces are highlighted in Fig. 3C). The former results from a DNA content change during the cell cycle (21–23), while the latter likely reflects the transition between different states. To selectively detect the transition events, we applied a low-pass filter to the data, which effectively eliminated the cell cycle oscillations (see Materials and Methods and Fig. S4 in the supplemental material). We then identified “up” and “down” transitions by applying thresholds on the amplitude and the rate of GFP intensity change (see Materials and Methods and Fig. S5 in the supplemental material). In total, we found 37 “up” and 40 “down” transitions among 733 h of steady-state GFP traces, resulting in a transition rate of 0.050 and 0.054 h⁻¹, respectively.

If the transitions involve changes of chromatin structure, they may preferably occur at certain cell cycle stages. To test this idea, we determined the transition time with respect to the cell cycle (see Materials and Methods). We again focused only on the mother cycles for the cell cycle alignment. The “up” transitions tend to happen near the beginning/end of a cell cycle (Fig. 4A and B), whereas the “down” transitions occur throughout the rest of the cell cycle (Fig. 4C and D). To confirm such cell cycle dependence, we generated a new diploid strain that can be arrested at metaphase by Cdc20 depletion (see Materials and Methods). Since Cdc20-depleted cells are arrested in metaphase prior to the end of the cell cycle (24), we expected to see a decreased “up” transition frequency in the arrested cells. Indeed, in the cycling cells with constitutive expression of Cdc20, GFP expression shows transition dynamics similar to those of the wild-type strain (Fig. 4E and G). However, in the arrested cells, “up,” but not “down,” transitions completely disappear (Fig. 4F and G). Overall, these observations strongly indicate that the cell cycle plays a major role in the transitions between the substable states.

FIG 4.

Cell cycle dependence of the transitions between intermediate states. (A) Sample traces of “on” transitions aligned with the cell cycle. The curves are deliberately separated along the y axis for visualization purpose. The crosses mark the starting points of the transitions. (B) Histogram of the “up” transition time relative to the cell cycle. Most of the “up” transitions occur near the end of the cell cycle (n = 33 cycles, focusing on mother cell cycles). (C) Sample traces of “down” transitions. (D) Histogram of the “down” transition times (n = 36 cycles, focusing on mother cell cycles). (E and F) Sample traces of GFP expression in the strain containing inducible CDC20 in the absence (E) or presence (F) of cell cycle arrest (total number of cells, 42 and 56, respectively). (G) Frequency of “up” and “down” transitions in wild-type cells (blue), Cdc20p⁺ cycling cells (yellow) and cdc20p-arrested cells (purple). The error bars show the SE among three biological replicates. In the arrested cells, the “up” transition is completely eliminated. The frequencies reported here are after 4 h of induction. (H) Histogram of the “up” (green) and “down” (orange) transition time relative to cell cycle for MET3pr-GFP in the telomeric region.

Since MET3pr-GFP in the telomere also showed multiple levels of GFP expression, we used the same method described above to identify transition events in these traces and examine their cell cycle dependence. The overall transition rate in these traces is smaller than that in the rDNA traces (0.010 h⁻¹ and 0.050 h⁻¹ for “up” transitions in telomere and rDNA, respectively; 0.022 h⁻¹ and 0.055 h⁻¹ for “down” transitions). More importantly, the transition time does not show a strong cell cycle dependence (Fig. 4H). The difference in the transition time in rDNA versus the telomere is likely related to different dynamic properties of the silencing complexes (see Discussion).

Simulation of the transition between intermediate silencing states.

Because it takes time for GFP to be produced and to mature, the transition time we observed in the GFP fluorescence may not reflect the transition time at the transcription level. To further pinpoint the transition time, we constructed a mathematical model to simulate the experimental data above (see Materials and Methods). In this model, GFP is produced, matured, and diluted with the rates of k_p, k_m, and k_d, respectively (Fig. 5A). The maturation constant k_m came from a previous study (25). The dilution constant k_d was fitted by the GFP exponential decay curve to be 1/145 min⁻¹ (see Fig. S6A in the supplemental material). In our model, k_p,eu goes through an initial spike upon induction before going down to the steady-state level, as derived from the GFP expression data (Fig. 5B and S6B). Noise was added to k_p,eu to account for the cell-to-cell variation of the production rate, and a sine wave was also added to simulate the cell cycle-dependent fluctuation of GFP intensity. The steady-state level of k_p,eu was fitted so that the simulated amplitude of the mature GFP (GFP*) matches the corresponding measurement (Fig. 5C and D).

FIG 5.

Simulation of the transition between intermediate silencing states. (A) Dynamics of the production, maturation, and dilution of GFP in the cell. [GFP] and [GFP*] represent the concentrations of immature and mature GFP. (B) Modeling of the production rates. k_p,_eu and k_p,_rDNA are the production rates in the euchromatic region and the rDNA region, respectively. The k_p,_rDNA transitions between two silenced states. (C and E) Examples of simulated traces in euchromatin (C) and rDNA (E). (D and F) Histograms of GFP expression levels in panels C and E, respectively, at different time points. (G and H) Histograms of the “up” (G) and “down” (H) transition times relative to the cell cycle from simulated GFP traces.

To model GFP expression in rDNA, we added another variable called the “silencing strength.” For simplicity, we assumed that this silencing strength can take on two discrete values, representing two silencing states. The net GFP production rate, k_p,rDNA, was calculated as k_p,eu minus the silencing strength (Fig. 5B). The GFP* level simulated with such k_p,rDNA shows a clear bimodal distribution (Fig. 5E and F). When we allowed the silencing strength to drop from the upper to the lower state exclusively during the last 10% of the cell cycle (corresponding to the “up” transitions in the GFP level) and the reverse change to occur randomly during the rest of the cell cycle, the GFP* transition kinetics agreed well with the experimental data (Fig. 5G and H). We therefore conclude that the “up” transitions at the transcription level happen near the mitotic exit.

Expressions in mother and daughter cells are correlated.

Silencing relies on histone modification, which can be partially inherited through DNA replication (26). The inherited histones may in turn facilitate the establishment of silencing. Through this mechanism, the silencing state from the mother cells can potentially be copied into the daughter cells. Even if the silencing state changes over time, similar starting points of the chromatin configuration in mother and daughter cells may allow them to change in a more synchronized fashion. Consistent with this idea, a previous study showed that mother and daughter cells often silence in concert upon the induction of Sir2 (27).

To understand whether the intermediate silencing states are correlated in mother and daughter cells, we tracked the change in GFP expression (ΔGFP) in mother/daughter pairs at different time points after their division (Fig. 6A). For comparison, we shuffled the mother/daughter cells, picked pairs of unrelated cells that have similar GFP levels upon division, and calculated the GFP level change at the same time points. Within 2 to 4 h, the correlation of ΔGFP between unrelated pairs is >0 (∼0.3) (Fig. 6B), which is expected because cells starting with the same GFP level tend to have similar GFP expression changes (e.g., cells with a very low GFP level are more likely to have “up” instead of “down” transitions). Importantly, such correlation is much higher in mother/daughter pairs (∼0.7) (Fig. 6B). These results indicate that the change of silencing states is more synchronized in mother and daughter cells, which may reflect similar chromatin structures in rDNA after DNA duplication.

FIG 6.

Expressions between mother in daughter cells are correlated. (A) Method to quantify the inheritance of GFP expression in rDNA. When the mother cell (m) gives birth to the daughter cells (d), they have the same GFP intensity. The changes of GFP expression at different time points after cytokinesis are calculated for mother (ΔGFP_m) and daughter (ΔGFP_d) cells. For comparison, we found another daughter cell (d′) that was born with a GFP intensity similar to that of m and d, and we calculated the ΔGFP_d′ at the same time points. (B) ΔGFP at 2, 3, 4, and 5 h after cell division of mother/daughter cells (red) in comparison to unrelated cells (black). At 2, 3, and 4 h, the former shows higher correlation in ΔGFP.

Intermediate states are not “memorized” after a brief repression.

Upon MET3pr repression, the activators will dissociate from the promoter, causing the partially activated local chromatin to revert back to its original configuration. We next asked whether the intermediate silencing state can be “memorized” through a brief repression. We first induced the MET3pr-GFP expression to the steady state, followed by 10× methionine repression for 1 to 3 h, and then induced again (Fig. 7A; see Materials and Methods). If there is a memory of the previous silencing state, we expect to see a positive correlation between the steady-state GFP levels during the first and second inductions. The MET3pr-GFP expression in euchromatin shows a small positive correlation between the two inductions, regardless of the repression time (Fig. 7B and C). We speculate that this correlation results from some global cellular properties that are maintained during the measurement, e.g., cell size. For MET3pr-GFP in the silenced locus, the correlation between the two expression levels turns out to be even smaller (Fig. 7B and C). We observed many cases where a low-expression state during the first induction jumped to a high-expression state after the brief repression and vice versa. These results indicate that the configuration of the silencing state is quickly lost once it is disrupted by repression, and the expression level during the subsequent induction needs to be “reset” by the competition between the activator and the local repressive chromatin.

FIG 7.

Intermediate states are not “memorized” after a brief repression. (A) Sample traces of GFP expression in the euchromatic (blue) or rDNA (red) region with 4 h of the initial induction, 2 h of repression, and 4 h of the second induction. The steady-state expression following the first and second inductions used the data in the box. (B) Average steady-state GFP expression level in the first induction compared with that of the second induction with 2 h of repression. Each dot represents one trace, and the x and y axes represent the average steady-state GFP expression during the first and second inductions. The correlation coefficient for the euchromatin (blue) and the rDNA region (red) are shown (total number of cells, 33 and 64 for euchromatin and rDNA, respectively). (C) Correlation of GFP levels with different repression durations.

DISCUSSION

Reporter genes inserted into silencing regions often exhibit “position effect variegation” that is characterized by an “all-or-none” expression pattern, leading to a large cell-to-cell variability in expression. Increasing the strength of activation biases the expression towards the “on” state. In this paper, we show that even when the “off” state is completely eliminated by strong activation, the reporter gene in silencing regions still shows high cell-to-cell variability in the “on” state. Further analyses suggest that the “on” state is an ensemble of multiple intermediate states, each showing a discrete level of expression that is in between the basal and fully expressed (unsilenced) level. The distribution among these substable states is modulated by the activation strength, which reflects the dynamic equilibrium between activation and silencing. Multiple intermediate states were also observed at the telomere, indicating that they are not due to some special properties of the RENT complex but are rather a general character of the Sir2-mediated silencing.

What could be the molecular nature of these intermediate silencing states? Because the MET3pr-GFP expression does not correlate with the MET3pr-mCherry expression in the same cell (Fig. 1F and H), it is unlikely that different silencing states reflect the activity of a global regulator, such as Met4 or Pol II. Instead, we suspect that the states are caused by variations in the local chromatin structure. Upon the induction of MET3pr in the rDNA, the histone-modifying enzymes and nucleosome-remodeling machineries recruited by the activator compete with the neighboring repressive chromatin to generate a local environment that is semipermissive of transcription. For example, some nucleosomes near the MET3pr may be more dynamic, carrying acetylation marks, and not associated with the Sir2 protein. The configuration of such partially opened chromatin may be variable from cell to cell. The discrete expression level of MET3pr-GFP raises the intriguing possibility that the intermediate states are determined by the number of permissive nucleosomes near the MET3pr or in the GFP gene body (Fig. 8A). Further experiments are needed to test this hypothesis, for example, by sorting the cells with different silencing states using fluorescence-activated cell sorting (FACS) and probing the chromosome configurations in each group.

FIG 8.

A model of multiple silencing states and the transitions in between. (A) A model of multiple silencing states. Higher or lower expression in the rDNA region (left and right panels, respectively) may be caused by different distributions of the RENT complex and histone modifications. Repressive histone marks may be carried by different numbers of nucleosomes, causing discrete levels of expression. (B) Maintenance or transition of the silencing states through mitotic exit. During mitotic exit, the RENT complex is transiently dissociated from rDNA. The remaining histone marks may be able to maintain the transcription level (pathway 1); in some cases, however, repressive histone marks may be replaced by more-activating ones, causing the transition to a higher-expression state (pathway 2).

We also found that the rDNA silencing states go through transitions, and the transition time is cell cycle dependent, with the “up” transitions occurring mostly at the end of the cell cycle. Such timing can be well explained by the dynamics of the RENT complex. Net1, a core subunit of RENT, is essential for rDNA silencing by tethering Sir2 to this region (28). During mitotic exit, Net1 is inactivated by phosphorylation, and a fraction of Sir2 leaves the nucleolus (28) (Fig. 8B). It was suggested that rDNA silencing may be weakened at this point of the cell cycle, and the timing of the “up” transitions provides strong support for this idea. The RENT complex then reassociates with rDNA during interphase, which may cause “down” transitions occurring at these stages of the cell cycle. Such a mechanism indicates that the transition time observed in Fig. 4B and D should be specific to the RENT complex and rDNA; indeed, transitions in the telomere do not have the same cell cycle dependence (Fig. 4G). Different from the “up” transitions between the intermediate states, the activation upon methionine depletion does not show cell cycle dependence (Fig. 2D and H). This is likely because the activation is stronger and can outcompete silencing even when the RENT complex is fully associated with rDNA.

It is important to point out that, although the RENT complex dynamically associates and dissociates from rDNA, we do not observe “up” and “down” transitions in every single cell cycle. In fact, these transitions occur approximately once every 7 cell cycles, and the GFP level is rather stable most of the time. This observation indicates that, in most cell cycles, the intermediate silencing state is not destroyed and reset during every mitotic exit; instead, the information of the previous state can be passed on despite the temporary disruption of the RENT complex. One potential candidate that can carry such heritable information is histone modifications. It is possible that certain histone modifications that contribute to rDNA silencing will be maintained during the mitotic exit and facilitate the rebinding of RENT during the next interphase, allowing stable propagation of the silencing states (Fig. 8B). The configuration of the histone modifications may be maintained during DNA replication, generating two cells with similar silencing states. We also found that the configuration is not robust against repression: once the activating components dissociate from the chromatin, the silencing complex likely reoccupies the reporter gene in a short amount of time. In the next round of activation, the activators need to compete with the inhibitory chromatin structure again to establish a new intermediate silencing state that has no correlation with the old state.

MATERIALS AND METHODS

Strain construction.

Strains and plasmids were constructed using standard methods. All strains used here are based on Y800 background strain. The two strains used for Fig. 1A to D and Fig. 3A to D contain a single copy of Saccharomyces kudriavzevii MET3pr-mCherry flanked by URA3 marker inserted into the CDC20 ORF and also a single copy of S. kudriavzevii MET3pr-GFP flanked by a KanMX marker inserted into the ATG36 ORF (chromosome X: 82713) or rDNA (NTS1 locus between the sequences TATTGCACTGGCTATTCATC and TTGCACTTTTCCTCTTTCTT) (13). The KanMX gene in the construct is driven by a strong promoter, TEF1pr. We suspect that silencing can only quantitatively reduce the activity of TEF1pr, and the remaining KanMX expression is sufficient for the survival of the strain in G418. The TEF1pr does not activate MET3pr directly, as the MET3pr-GFP expression is at an undetectable level without methionine depletion. The strain used for Fig. 1G and H contains the same GFP and mCherry reporters both flanked by KanMX marker and inserted into the same rDNA NTS1 locus. The strains used for Fig. 2, 4, 3E and F, 6, and 7 lack MET3pr-mCherry; instead, they have MYO1-mCherry at the endogenous MYO1 locus as a cell cycle marker. For the strain used for Fig. 3G and H, Gal1spr-GFP instead was inserted into the same rDNA NTS1 locus. For the strain used for Fig. 3I to L, MET3pr-GFP was inserted into RDN37-1 (between the sequences ACTACCACCAAGATCTGCAC and TAGAGGCCGTTCGACCCGAC) or a telomere locus (between the sequences TCTGGAACATCATCGCTATC and CAGCTCTTTGTGAACCGCTA). For the cell cycle arrest strain used for Fig. 4E and F, the endogenous CDC20pr (from −285 to −1 relative to the start of ORF) was replaced by a modified GAL1pr activated by the LexA-ER-VP16 fusion protein. CDC20 is constitutively expressed in the presence of 80 nM β-estradiol, allowing the cell cycle progression; washing out the β-estradiol causes cell cycle arrest.

Time-lapse fluorescence microscopy.

The time-lapse microscopy and the data acquisition procedure were described in our previous studies (29, 30). In short, cells were grown in CellASIC Onix microfluidic plates, which are controlled by CellASIC Onix2 microfluidic system. The software (CellASIC Onix2 software 1.0.4) controls the medium flow into the plates. For induction studies for Fig. 1 and 4, cells were first grown in synthetic complete medium with glucose (SCD)–10× methionine medium (repressive condition) for 2 to 4 h before switching to SCD–0× (0.5%×, 1%×) methionine for induction. For induction for Fig. 3G and H, cells were first grown in SC-raffinose medium (repressive condition) for 4 h before switching to SC-galactose for induction. For the cell cycle arrest experiment for Fig. 4E and F, yeast culture was first grown in SCD–10× methionine–80 nM β-estradiol, washed 4 or 5 times to remove the β-estradiol, loaded into the plates, and then induced with SCD–0× methionine medium in the absence of β-estradiol (we found that it is hard to wash out β-estradiol inside the microfluidic device, probably because it gets absorbed onto the plastic surface). For memory experiments for Fig. 7, cells were activated for 4 h with SCD–0× methionine medium, repressed for 1 to 3 h with SCD–10× methionine medium, and activated again for 4 h with SCD–0× methionine medium. Images were taken every 5 min.

Image analysis.

In-house MATLAB software was used to analyze the images. Cell contours were annotated and tracked from frame to frame, and the mean fluorescence intensities in each cell were extracted. The Myo1-mCherry fluorescence signals were also annotated and used to define the cell cycles. Because mother and daughter cells have significantly different G₁ phase durations, we focused only on mother cells. We also ignored very long cell cycles outside 3 standard deviations, which likely go through cell cycle arrest due to photo damage.

GFP data analysis.

The GFP intensity trace was treated by a median filter and then smoothed by a small moving window (three points moving average). To reduce the cell cycle-dependent fluctuations in the GFP intensity, a second-order Butterworth filter with cutoff frequency at 1/78 min⁻¹ (the frequency of the cell cycle) was applied to the GFP traces (see Fig. S4 in the supplemental material). To obtain the activation and transition times, zero crossings were detected from the first derivative of the low-pass-filtered fluorescence traces, which cut the curve into several monotonic intervals. An interval will be considered a transition only if the range and the rate of GFP change are both among the top 1/3 (see Fig. S5 in the supplemental material). After identifying the transition or activation events, the starting point was determined by the first point where the first derivative was raised from zero to 10% of its maximum value. By comparing the starting point with the cell division time, the starting point in the unit of cell cycle (from 0 to 1, where 1 is the end of cytokinesis) was calculated.

A Gaussian mixture model was used to fit the steady-state fluorescence distribution. The best fit was determined by the model with the lowest Akaike information criterion (AIC) score (see Fig. S3 in the supplemental material). In some cases, although the lowest AIC score suggests a model with multiple components, there is only one dominant peak (for example, GFP expression in the euchromatic region). To focus on the main components of the distribution, we ignored the peaks with probability smaller than 10%.

To search for the correlation between mother and daughter cells (Fig. 6), we calculated the change of GFP intensity in mother and daughter cells after cell division. As a control, we shuffled the relationship between cells. The daughter cells were randomly linked to unrelated mother cells that divided around the same time and had similar GFP expression at the end of division.

Simulations.

In the model, the initial expression occurs at 16.8 min (standard error [SE] = 0.6 min) after activation regardless of the time of the cell cycle or the silenced state. The production rate k_p first rises to 2 to 3 times (generated by random numbers) the steady-state level before it goes back down at around 80 min (Fig. 5B). k_p was modeled this way based on the experimental data (see Fig. S6B in the supplemental material). In the case of silent chromatin, the steady-state k_p adopts two values, with the lower one at 60% of the level of the upper one. To account for the cell cycle-dependent GFP fluctuations, a sine wave was added to k_p, oscillating around the mean steady-state level with a 30% amplitude. The autofluorescence background was set to a fixed number to mimic the background signal in real experiments.

The durations of cell cycles and the initial phase of cell cycle are randomly generated to mimic an asynchronized cell population. The G₁, S, and M phases take about 18.8%, 37.5%, and 43.7% of the mother cell cycle, respectively (31). In M phase, only metaphase and anaphase are considered, and they span 18.7% and 25% of the cell cycle, respectively (31). To simulate the cell cycle-dependent transitions, we allow the strong silencing state to switch to the weak one during the last 10% of the cell cycle with the probability calculated from the experiment, which leads to “up” transitions in GFP expression. We allow the opposite change to occur randomly during the rest of the cell cycle, leading to “down” transitions. With k_p generated by the model, we calculated the GFP intensity based on the differential equations in Fig. 5A using the finite-element method. The maturation constant k_m, ln(2)/18.5 min⁻¹ for GFP in yeast, was obtained from a previous study (25). The dilution constant k_d was fitted to be 1/145 min⁻¹ from exponential decay of GFP after MET3pr-GFP repression. To include the detection noise and other extrinsic noise, we added white noise to the final GFP (SE = 1.1). Subsequently, all simulation data were analyzed with the algorithm described in “GFP data analysis” above.