Portal de Periódicos da CAPES

Structure of silent transcription intervals and noise characteristics of mammalian genes

2015; Springer Nature; Volume: 11; Issue: 7 Linguagem: Inglês

10.15252/msb.20156257

1744-4292

Benjamin Zoller, Damien Nicolas, Nacho Molina, Félix Naef,

Genomics and Chromatin Dynamics

Article27 July 2015Open Access Structure of silent transcription intervals and noise characteristics of mammalian genes Benjamin Zoller Benjamin Zoller The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Damien Nicolas Damien Nicolas The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Nacho Molina Nacho Molina The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Felix Naef Corresponding Author Felix Naef The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Benjamin Zoller Benjamin Zoller The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Damien Nicolas Damien Nicolas The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Nacho Molina Nacho Molina The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Felix Naef Corresponding Author Felix Naef The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Search for more papers by this author Author Information Benjamin Zoller1, Damien Nicolas1, Nacho Molina1 and Felix Naef 1 1The Institute of Bioengineering, School of Life Sciences, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland *Corresponding author. Tel: +41 21 693 16 21; E-mail: [email protected] Molecular Systems Biology (2015)11:823https://doi.org/10.15252/msb.20156257 PDFDownload PDF of article text and main figures. Peer ReviewDownload a summary of the editorial decision process including editorial decision letters, reviewer comments and author responses to feedback. ToolsAdd to favoritesDownload CitationsTrack CitationsPermissions Figures & Info Abstract Mammalian transcription occurs stochastically in short bursts interspersed by silent intervals showing a refractory period. However, the underlying processes and consequences on fluctuations in gene products are poorly understood. Here, we use single allele time-lapse recordings in mouse cells to identify minimal models of promoter cycles, which inform on the number and durations of rate-limiting steps responsible for refractory periods. The structure of promoter cycles is gene specific and independent of genomic location. Typically, five rate-limiting steps underlie the silent periods of endogenous promoters, while minimal synthetic promoters exhibit only one. Strikingly, endogenous or synthetic promoters with TATA boxes show simplified two-state promoter cycles. Since transcriptional bursting constrains intrinsic noise depending on the number of promoter steps, this explains why TATA box genes display increased intrinsic noise genome-wide in mammals, as revealed by single-cell RNA-seq. These findings have implications for basic transcription biology and shed light on interpreting single-cell RNA-counting experiments. Synopsis Analysis of transcriptional bursting from time-lapse imaging of single alleles in mammalian cells identifies the kinetic structure of promoter cycles underlying refractoriness, and explains noise in mRNA abundance. Quantitative modeling of single allele time-lapse recordings in mouse cells identifies minimal models of promoter cycles, which inform on the rate-limiting steps responsible for refractory periods. The structure of promoter cycles is gene specific and independent of genomic location. Typically, five rate-limiting steps underlie the silent periods of endogenous promoters, while minimal synthetic promoters exhibit only one. Promoter architecture constrains intrinsic noise depending on the structure of the promoter cycles, notably, TATA box genes display increased intrinsic noise in mammals, as confirmed in single-cell RNA-seq. Introduction Gene expression is intrinsically dynamic and varies greatly from cell to cell (Raj & van Oudenaarden, 2008). In isogenic cell populations, such variability arises naturally from randomness in the processes governing gene expression. Typically, low numbers of molecules are involved in transcription, leading to unavoidable stochasticity in both mRNA and protein levels (Elowitz et al, 2002; Paulsson, 2004). In fact, fluctuations in mRNA numbers can significantly exceed what constitutive expression predicts (Poisson statistics) (Blake et al, 2003; Raser & O'Shea, 2004), and it was proposed that this originates in short and intermittent activations of the genes called transcriptional bursts. Transcriptional bursting was formalized as a telegraph model (Peccoud & Ycart, 1995), in which a promoter toggles between transcriptionally active (on) and inactive (off) states. The size of the bursts (b) represents the average number of transcripts produced during the active period. Recent assays in single cells confirmed transcriptional bursting in many organisms (Golding et al, 2005; Chubb et al, 2006; Raj et al, 2006; Zenklusen et al, 2008). Although not all genes are transcribed in bursts (Zenklusen et al, 2008), bursting appears predominant in mammals (Suter et al, 2011; Dar et al, 2012; Bahar Halpern et al, 2015). The mechanisms causing bursts in eukaryotes are still elusive but most likely involve the interplay between transcription factors (Larson et al, 2013; Senecal et al, 2014), chromatin remodelers (Coulon et al, 2013; Voss & Hager, 2013), the formation of gene loops and pre-initiation complexes (Blake et al, 2003; Zenklusen et al, 2008), and transcription initiation and elongation (Jonkers et al, 2014; Stasevich et al, 2014). Recent time-lapse imaging to monitor bursting of endogenous mammalian genes (Harper et al, 2011; Suter et al, 2011) reported peaked silent transcriptional intervals, suggesting a refractory period lasting about 1 h preceding transcription reactivation. Similarly, promoter refractoriness to reactivation was reported in Neurospora, indicating a form of molecular memory (Cesbron et al, 2015). Refractory periods support a model of promoter progression (Hager et al, 2006; Métivier et al, 2006) in which sequential metastable changes in the local chromatin template underlie a multi-step progression toward transcription activation (Coulon et al, 2013). In first approximation, this promoter progression can be considered as an irreversible cycle (Zhang et al, 2012), whose rate-limiting steps need to be estimated, which we address here. Detailed knowledge on the transcriptional kinetics also allows better understanding of noise in gene expression (Ozbudak et al, 2002; Swain et al, 2002; Paulsson, 2004; Sanchez & Kondev, 2008), which is relevant notably in the context of RNA-counting experiments in developmental (Little et al, 2013; Bothma et al, 2014) or cell differentiation systems (Chang et al, 2008; Abranches et al, 2014; Ochiai et al, 2014). Importantly, the structure and kinetics of the promoter cycles will also impact the noise in gene expression since it determines the statistics of the off intervals (Pedraza & Paulsson, 2008). Indeed, in addition to the standard transcriptional parameters (burst size, activation frequency), the number of rate-limiting steps also tunes noise levels (Zhang et al, 2012). Here, we combined temporal single-cell measurements of short-lived and highly sensitive luciferase reporters with mathematical modeling to characterize silent transcriptional intervals. In particular, by modeling promoters as an irreversible cycle, we estimated the number and durations of the rate-limiting steps responsible for refractory periods in mammalian gene reactivation. We found gene-specific structure and kinetics of the promoter cycle. Typically, endogenous promoters showed five sequential inactive steps, while minimal synthetic promoters exhibited only one. Two groups of promoter architecture showed distinct transcriptional kinetics; notably, TATA box promoters had only few inactive steps, independently of their genomic location. Moreover, intrinsic noise in our clones was constrained due to transcriptional bursting, and buffered by additional inactive promoter steps. Finally, we analyzed single-cell RNA-seq in mouse embryonic stem cells (mESCs) to validate genome-wide the prediction that TATA box promoters, owing to their reduced number of promoter steps, showed increased intrinsic noise. Results Refractory period in gene reactivation modeled by a promoter cycle The two-state promoter cycle (telegraph model) predicts exponentially distributed transcriptionally silent periods, yet evidence points toward peaked (non-exponential) durations (Harper et al, 2011; Suter et al, 2011), which implies out of equilibrium dynamics and irreversibility in the underlying processes (Tu, 2008). A simple yet still sufficiently general model compatible with this constraint is an irreversible N + 1-state promoter cycle (Zhang et al, 2012), consisting of one transcriptionally active state (on) and N sequential inactive states (off), modeling the scenario of promoter progression (Hager et al, 2006; Métivier et al, 2006). Both the number of states N and their durations (not necessarily equal) are not known and will be estimated from data. The resulting stochastic gene expression model (Appendix Supplementary Methods) consists in a two-layered cascade of birth and death processes, describing the production and degradation of mRNAs and proteins (Fig 1A). Although this model is a coarse-grained description of gene expression, it accommodates for the observed refractory periods while remaining sufficiently parsimonious to allow inference. Figure 1. The promoter cycle as a generic stochastic gene expression model to analyze time-lapse imaging data in single mammalian cells The stochastic model describes gene activation, transcription, translation, and degradation of mRNA and proteins. The promoter state follows an irreversible cycle composed of one transcriptionally active state and multiple (N) sequential inactive states describing the promoter progression toward activation. Gene-specific rates for the different processes are indicated. Stochastic simulation of protein numbers, mRNA numbers, and gene activity with N = 1 inactive state. Here, the duration of silent intervals is exponentially distributed (left) and the gene expression traces are irregular. With N = 6 states, the duration of silent intervals is now peaked and the expression pattern more regular. Parameters reflect a realistic situation, that is, the average duration of the total silent period T is identical in both simulations and set to T = 90 min, τa = 8 min and km = 5 mRNA per minutes. Download figure Download PowerPoint To illustrate the behavior of the model, we compared two realistic simulations differing only in the partitioning of the silent period T. A unique step (N = 1) yielded exponentially distributed off-times (Fig 1B), while partitioning T in six subintervals of equal average duration (N = 6) followed a peaked (Gamma) distribution. For N = 1, we observed large variability in the silent periods. Due to the short active periods, the mRNA and protein time traces were irregular (Fig 1B). By contrast, the profiles for N = 6 were more regular and the fluctuations in mRNA and protein numbers were reduced (Fig 1C), which follows from the more evenly spaced activation events (Pedraza & Paulsson, 2008). Identification of optimal promoter cycles To characterize the promoter cycles in a set of NIH3T3 cell lines expressing a single allele of a short-lived luciferase reporter driven by different promoters, we extended our computational approach for estimating transcriptional parameters from time-lapse recordings (the transcription rate km, the active period τa, and the total silent period T) (Suter et al, 2011; Molina et al, 2013) to identify the number N and durations τi of transcriptionally inactive states. The translation rate kp and the degradation rates of both the protein γp and mRNA γm were measured (Table EV1) and therefore did not need to be inferred. Briefly, we followed a Bayesian approach to estimate the joint posterior probabilities on N and the kinetic rates using a reversible jump Markov chain Monte Carlo (RJ-MCMC) algorithm (Green & Hastie, 2009) (Materials and Methods). RJ-MCMC is a model selection method in which more complex models (larger cycles) are naturally penalized, thus avoiding over-fitting. Implementing this scheme requires computing the likelihood of each bioluminescence time trace under a model (specified by N and all kinetic rates). For the likelihood, we used calibrated luminescence signals (Suter et al, 2011; Molina et al, 2013) (Appendix Supplementary Methods) and the transition probabilities between promoter states, mRNA and protein numbers over the 5-min sampling interval, as dictated by the master equation for the promoter cycle. For the RJ-MCMC sampling, we implemented model-crossing jumps by adding or removing inactive states while keeping T constant (Fig 2A). Figure 2. Model selection and parameters estimation based on reversible jump Markov chain Monte Carlo (RJ-MCMC) sampling The number of inactive states in the promoter cycle defines a class of nested models. To sample the different models, we implemented moves (jumps) between models differing by one inactive state. Typical MCMC run, here on simulated data (64 individual traces of 48 h each) generated with N = 6 inactive steps. Kinetic parameters and number of inactive states N are sampled invariably. Left: MCMC traces, note the short burn in period. Right: Histograms reflecting the estimated posterior distributions, these are centered on the mean values (dashed line), and N is between 6 and 9 (most probable is N = 7). Posterior distribution for N, inferred from synthetic data as in (B) (with 48 cells per condition), with N = 1, 2, and 4, respectively, keeping identical mean silent period T. Performance of the inference on individual transcriptional parameters in function of the simulated mean mRNA numbers (48 cells per mean mRNA). The dashed lines represent the expected values. To vary the mean, either only the transcription rate is increased (blue) or both the on-time and the transcription rate (green) are increased. Crosses show the posterior mean and error bars the 5th and 95th percentiles. Download figure Download PowerPoint To validate the method, we simulated bioluminescence time traces that mimicked our experiments in terms of the number of cells, length of time traces, measurement noise, and sampling rate (Appendix Supplementary Methods), and tested whether N, km, τa, and could be recovered. For simplicity, we assumed that the kinetic rates were constant and equal for all cells from the same clone. We estimated posterior distributions of the parameters from four populations sharing identical parameters except T, which was partitioned into N = 1, 2, 4 and 6 intervals. As exemplified for N = 6, we recovered these parameters with good accuracy, albeit with small biases (< 8%) (Fig 2B and Table EV2). Similarly, the posterior probability on N bracketed the true number, with a tendency to overestimate the most likely value by one (Fig 2B and C). To test whether low expression would deteriorate performance, we explored how the mean number of mRNAs and active period τa affect the inference. We generated synthetic populations spanning a realistic range in mRNA expressions (Fig 2D) and varied the expression either by changing the transcription rate km or by changing both km and τa (Appendix Supplementary Methods). Remarkably, the recovered parameters were close to the input values even for τa smaller than the 5-min sampling interval and for the lowest expressions (Fig 2D and Table EV2). Finally, we tested whether heterogeneous kinetic parameters would affect our estimates. Although inter-cell variability may shorten τa and increase km, the burst sizes b, N, and T were not subject to similar biases (Appendix Fig S1). Thus, considering that we used a limited amount of data to mimic the bioluminescence signal and that some parameters describe processes that are filtered at the level of the measured protein expression, we concluded that the inference method performs remarkably well. Two groups of gene-specific promoter cycles We then applied the method to characterize promoter cycles in 16 mouse fibroblasts cell lines (NIH3T3 cells) stably driving a short-lived luciferase reporter from a single allele (Suter et al, 2011). These included reporter lines driven by two distinct insertions of the Bmal1 promoter (B clones); seven clones obtained by lentiviral trapping (gene trap, GT) of endogenous promoters (gene names in Table EV1); and five clones that used the FRT/Flp system to insert into a common location single copies of either the Dbp gene (including its promoter) or minimal synthetic promoters combining a TATA box and one (H1) or two (H2) CCAAT boxes with multiple mutations. Additionally, we generated two more H1 clones that used new FRT sites in different genomic locations (Appendix Supplementary Methods). Importantly, to minimize transcriptional disturbances during the cell cycle, non-dividing (highly confluent) cells were continuously recorded over approximately 2 days (Chassot et al, 2008). We then estimated the transcriptional kinetics from temporal traces in single cells for each clone. The clones spanned a wide range of burst sizes b (from 1 to 80), independent of the fraction of time spent in the active state, which remained under 10% (Fig 3A). The infrequent promoter activations clearly indicated that transcription occurs in bursts. Moreover, b depended predominantly on the promoter and, to a lesser extent, on the genomic locus, as exemplified by multiple Bmal1 and H1 clones. The average duration of the silent period T exhibited a smaller dynamic range (from 30 min to 3 h) than the burst sizes, which was the most varying kinetic parameter among the clones (Fig 3B). Notably, b and T appeared largely uncorrelated among the clones. Overall, the extended model yielded kinetic parameters that were largely consistent with previous estimates (Suter et al, 2011) (Table EV3). Clearly, the short activation times and large burst sizes implied that transcription in this set of clones is highly discontinuous. Figure 3. Structure and kinetics of the promoter cycles for the NIH3T3 clones Burst size vs. the fraction of time the gene is active. Each clone is represented by a 95% confidence ellipse from the posterior distribution. All the analyzed clones burst, characterized by small activity fractions. Burst sizes show a large dynamic range across clones (˜80-fold). Inset: Magnification of the lower left corner. Burst size vs. the total silent period T. Elongated confidence ellipses reflect the dependence between those two quantities and the mean mRNA. Although the dynamic range of the silent period (˜6-fold) is smaller than for the burst size, it is also gene specific. The synthetic (warm colors) and endogenous (cold colors) promoters cluster in distinct regions. Number of inactive states vs. T, crosses indicate mean and error bars stand for the 5th and 95th percentiles of the posterior. Endogenous promoters tend to show more inactive steps and shorter cycle times (cluster around N˜6 and T˜60 min) compared to synthetic promoters (cluster around N˜1–2 and T˜130 min). Partitioning of the silent period for the optimal models. The light and dark bars show the mean durations of each sub-step. Partitions in endogenous promoters tend to be more uniform compared to the synthetic promoters. Average inactive times for endogenous promoter are around 10 min, whereas synthetic promoters have average inactive times close to 100 min (˜115 min for the first and ˜25 min for the subsequent intervals). Download figure Download PowerPoint Examining the structure of the promoter cycles (Fig 3C and D), we found that the number of inactive steps N differed between the clones (N = 1–7). Although it is difficult to gain further insights on the nature of these rate-limiting steps, their timescales of 10 min were more consistent with the dynamics of histone modifications than the interactions of transcription factor with DNA (Discussion). Supporting this, for the Bmal1 promoter treated with the histone deacetylation inhibitor (TSA), which renders the chromatin more permissive for transcription, N reduced from 7 to 3 and T reduced from 60 to 40 min (Fig 3C and D). The durations of the sub-intervals τi in the endogenous promoters were fairly homogenous, with intervals between 6 and 14 min, whereas synthetic promoters showed one dominating interval. This implied that the silent periods of endogenous promoters should display peaked distributions, whereas the silent periods of synthetic promoters should approximate exponential distributions. To assess the consistency of the inferred promoter cycles with the data, we compared the distributions of silent and active periods from the optimal model with the one obtained using Gibbs sampling (Appendix Figs S2 and S3). Gibbs sampling reconstructs mRNA and gene activity trajectories conditioned on the data in each individual cell using the optimal model as a prior (Appendix Supplementary Methods). It appeared that, for most genes, both the modeled and Gibbs distributions matched closely, confirming the previously observed peaked silent distributions, as well as the aforementioned difference between endogenous and synthetic promoters (Suter et al, 2011). Moreover, we did not observe refractory active periods on the scale of the sampling times (Appendix Fig S3). Intriguingly, the relationship between N and T suggested two groups, namely promoter cycles with few steps (Group I: N ~ 1–2) and ones with markedly more steps (Group II: N ~ 6) (Fig 3C). In addition, in the first group, all synthetic promoters (six) as well as Dbp had long cycles (130 min), while the endogenous promoters (Ctgf, Prl2C2) had shorter cycles (50 min). Moreover, all promoters with large N were endogenous. As shown for representative cells for the H1 (Group I synthetic), Prl2C2 (Group I endogenous), and Gls promoters (Group II), the distinct kinetics are visible in individual cells, based on the raw signals as well as the mRNA counts and gene activities (Fig 4A–C). Figure 4. The kinetic structure of the silent intervals reveals different classes of promotersThree classes emerge from the characterization of the promoter cycle: synthetic promoters with a single long silent interval (Group I, synthetic), endogenous promoters with a shorter single interval (Group I, endogenous), and endogenous promoters displaying a refractory period (Group II). A–C. Deconvolution of individual cell traces illustrating the three groups of promoter cycles for three promoters (H1b, Prl2C2, Gls). Measured single-cell bioluminescent time traces (red), deconvolved mRNA (green), and gene activity (gray shading indicates probability of the ‘on’ state, black is highest). D. Scheme showing the three groups: Top, distributions of silent intervals. Bottom, Simulations of gene activity patterns for the three groups show qualitative differences. Download figure Download PowerPoint In summary, the analyzed promoter cycles suggested two distinct groups, simple promoter cycles and complex promoter cycles (Fig 4D). Simple promoter cycles (Group I) caused nearly refractory-less and irregular activations, although the irregularity in the endogenous promoters (Ctgf and Prl2C2) was alleviated by more frequent activations. Complex promoter cycles (Group II) involved several transitions and short silent periods, thus leading to more regular activation patterns constrained by a refractory period. Promoter architecture influences the promoter cycles Since all the synthetic promoters from the original library (H1a, H2, H2 1M, H2 2M) were inserted into the same genomic location, the low number of states and long promoter cycle observed might reflect a property of the insertion site, for example, the chromatin state, rather than the promoter architecture. We therefore generated additional clones (H1b, H1c) by integrating the minimal promoter H1 at distinct genomic locations. Remarkably, the three H1 insertions retained very similar promoter cycles (Fig 3C and D). While endogenous promoters with similar cycles (Group II) were inherently located in different genomic loci, the two Bmal1 clones (Bmal1a, Bmal1b) in two distinct locations also showed very similar cycles (Fig 3D), further supporting that the structure of the cycles is primarily a property of the promoters. Interestingly, the synthetic (Group I) and the two endogenous promoters with small N (Group I with the exception of Dbp) contained a canonical TATA box element (Dreos et al, 2013), which was absent from other endogenous promoters (Group II) with larger N. Although the numbers were low, the presence of TATA boxes in promoters with small N (in mouse, only < 15% of promoters contain TATA boxes) was non-random (P < 0.01, binomial sampling). Promoter architecture and, in particular, the presence of TATA boxes seemed to influence the promoter cycles. Further evidence that this holds genome-wide is presented below. Intrinsic transcriptional noise dominates in non-dividing mammalian cells We next studied the implications of promoter cycles and transcriptional kinetics on population noise in mRNA numbers, defined as the variance over the mean squared (total noise). Since a fraction of the total noise is expectedly due to extrinsic variability, we split the total noise as . Although this separation can be subtle (Swain et al, 2002; Hilfinger & Paulsson, 2011) (Materials and Methods), η2 (intrinsic noise) arises from gene-specific fluctuations whereas (extrinsic noise) reflects other sources of heterogeneity. To estimate both components for each clone, we used Gibbs sampling to reconstruct the empirical distributions of mRNA numbers in each individual cell (Materials and Methods and Appendix Fig S4). Simulations with heterogeneous cell populations showed that Gibbs sampling accurately recovered the simulated mRNA distributions in each individual cell (and also in the cell population), providing an excellent proxy for (Appendix Fig S5). In the clones, the empirical population distributions tended to be more dispersed than the fitted model (Fig 5A). Indeed for some clones, for example, NcKap1 or Ctgf (Appendix Fig S4B), the model did not capture enrichment at low transcript numbers or longer tails in the empirical distributions. These deviations likely originated from extrinsic noise, such as kinetic parameters differing between cells or over time. Consistent with this interpretation, the circadianly transcribed Bmal1 promoter showed the largest deviation (Appendix Fig S4B). Figure 5. Separating intrinsic and extrinsic noise mRNA distribution (Gibbs, light green) and steady-state distribution of the optimal promoter cycle (dark green). Differences, also reflected in the noise values, originate from extrinsic variability. Separation of the total noise in intrinsic and extrinsic components. For the majority of clones, intrinsic noise dominates. The modeled noise corresponds to 83% of the estimated intrinsic noise on average. Data information: In (B, C), the error bars stand for the 5th and 95th percentiles of the estimate (parametric bootstrap). Download figure Download PowerPoint The mRNA distributions in individual cells allow splitting of the total variance into the mean variance (proxy for intrinsic variance) plus the variance in the means across cells (extrinsic variance) (Swain et al, 2002; Hilfinger & Paulsson, 2011). As verified by simulations (Appendix Fig S5), this split captures η2 and for static cellular heterogeneity (i.e., parameters in each cell remain constant during the recording). Importantly, the recordings were performed in non-dividing cells, removing one important source of temporal heterogeneity (Zopf et al, 2013). Since the Bmal1 and Dbp clones are sensitive to circadian oscillations, we restricted our noise analysis to the other clones, except for Bmal1 treated with TSA, which abolishes circadian oscillations while maintaining transcriptional bursting (Suter et al, 2011). In most clones, η2 exceeded (Fig 5B), and η (coefficient of variation CV) was on the order of 100% (η2 between 0.3 and 2.1), independent of expression levels. As shown below, this a direct consequence of transcriptional bursting. In comparison, ηe was in the range of 70% ( between 0.1 and 1.2) for a majority of clones. Among the few clones dominated by extrinsic noise, Ctgf is known to be highly sensitive to stimulations (Molina et al, 2013). Importantly, in both the clones and the simulations (Fig 5C and Appendix Fig S5D), a high portion of the estimated intrinsic noise (83% on average) was captured by the optimal model (Fig 5C), which allows us to study how the noise depends on transcriptional parameters (Fig 6). Figure 6. Relationship between mRNA noise and promoter cycles Intrinsic noise η2 (modeled noise) for the different clones in function of the mean mRNA expression (number of copies). The Poisson component sets a lower bound on intrinsic noise (lower dashed line). Thus, the promoter noise dominates for most genes, as reflected by Fano factors much larger than 2