Transcriptional burst frequency and burst size are equally modulated across the human genome

Abstract

Gene expression occurs either as an episodic process, characterized by pulsatile bursts, or as a constitutive process, characterized by a Poisson-like accumulation of gene products. It is not clear which mode of gene expression (constitutive versus bursty) predominates across a genome or how transcriptional dynamics are influenced by genomic position and promoter sequence. Here, we use time-lapse fluorescence microscopy to analyze 8,000 individual human genomic loci and find that at virtually all loci, episodic bursting—as opposed to constitutive expression—is the predominant mode of expression. Quantitative analysis of the expression dynamics at these 8,000 loci indicates that both the frequency and size of the transcriptional bursts varies equally across the human genome, independent of promoter sequence. Strikingly, weaker expression loci modulate burst frequency to increase activity, whereas stronger expression loci modulate burst size to increase activity. Transcriptional activators such as trichostatin A (TSA) and tumor necrosis factor α (TNF) only modulate burst size and frequency along a constrained trend line governed by the promoter. In summary, transcriptional bursting dominates across the human genome, both burst frequency and burst size vary by chromosomal location, and transcriptional activators alter burst frequency and burst size, depending on the expression level of the locus.

There exists conflicting evidence over the predominant mode of gene expression in both prokaryotes and eukaryotes. The classical view of gene expression as a constitutive, Poisson-like accumulation of gene products (Fig. 1A) is supported by a comprehensive large-scale study in bacteria, demonstrating that > 400 genes appear to follow constitutive (or Poisson-like) gene expression (1). Constitutive expression has also been reported for subsets of human genes (2). Conversely, several elegant studies showed that specific promoters in bacteria and yeast express gene products in an episodic process (Fig. 1B), characterized by pulsatile bursts in transcription (3–9). Given this conflicting evidence, it remains unclear if episodic bursting is the predominant mode of gene expression across a genome or just a highlighted exception. If bursting is predominant, it is not clear if or how it depends on genomic location.

Fig. 1.

Fluctuations in gene expression to differentiate between alternate models of transcription across the genome. (A and B) Schematics of the constitutive, Poisson-expression model and the episodic, bursty gene-expression model, together with three expression trajectories from hypothetical genomic loci. Sites that exhibit constitutive (i.e., Poisson) expression exhibit small and relatively fast fluctuations in gene products over time. Alternatively, loci that exhibit episodic expression bursts generate large, slow fluctuations in gene expression. (C) The principle of noise space. The three-dimensional noise space consists of noise magnitude, noise autocorrelation, and mean expression level. Small, fast fluctuations have a small noise magnitude and short autocorrelation times and thus cluster (after normalization) at the origin of the noise magnitude-autocorrelation plane (gray region, Lower Left). Large, slow (i.e., bursty) fluctuations have expanded noise magnitude and extended autocorrelation times (red ovals). The three-dimensional space can be decomposed into two additional two-dimensional projections of noise magnitude and noise autocorrelation versus mean expression level (Lower Center and Lower Right). For episodic-bursty expression, a trajectory’s noise-space coordinates are invariably shifted away from the constitutive model into the burst model space depending on changes to their transcriptional parameters (10).

To globally determine if constitutive Poisson-like expression or episodic bursty expression dominates throughout the human genome, we capitalize on a recently proposed theoretical framework (10) for extracting the details of gene regulation from the time-resolved structure of fluctuations (i.e., noise) in gene expression. This analysis quantifies time-lapse expression trajectories to obtain three orthogonal measures of expression: the average expression level, the magnitude of expression fluctuations (as measured by the coefficient of variation squared, CV²), and the autocorrelation time of expression fluctuations (as measured by the noise autocorrelation time at half of its initial value, τ_1/2) (11, 12) (Fig. 1C). Although this three-dimensional noise space is impractical to analyze directly, different two-dimensional projections of noise space allow the quantification of rate parameters in gene-regulatory models and provide a convenient method to differentiate between underlying gene-expression mechanisms, such as constitutive versus bursty transcription (Fig. 1C). For example, transcriptional bursting increases both noise magnitude and noise autocorrelation time and shifts points in the CV² versus τ_1/2 plane to the upper right quadrant relative to a constitutive expression model (Fig. 1C, Bottom Left). Conversely, translational bursting shifts noise magnitude, but not the autocorrelation time (10, 13).

Importantly, analysis of the τ_1/2 axis is critical to fully parameterize two-state transcriptional bursting models (Fig. 1C), which always include at least three unknown parameters: the rate of transition to a transcriptionally active state (k_on), the rate of transitioning to a transcriptionally inactive state (k_off), and the rate of transcription once in the active state (k_m) (13, 14). Analyses of a single two-dimensional plane (e.g., CV² versus expression level) cannot fully determine these three rate parameters. Conversely, analyses of CV² versus expression level and τ_1/2 versus expression level allow the determination of these three parameters, and analysis of CV²-versus-τ_1/2 facilitates direct comparisons of data containing widely varying expression levels, because it removes the reciprocal dependence of noise magnitude on expression level (10).

The ability to accurately quantify these transcriptional rate parameters is essential for answering basic questions about the mechanisms that regulate transcription. Previous studies elegantly applied flow cytometry (6, 7) and time-lapse microscopy (1, 15, 16) to analyze gene-expression noise in large subsets of genes. However, a tedious experimental bottleneck of subcloning and expansion of isogenic populations necessarily limits the throughput of these noise-analysis approaches. Here, we circumvent this subcloning requirement to globally apply the analytical framework of noise space across the human genome and quantify these transcriptional rate parameters. The analysis addresses two specific questions regarding transcriptional regulation in human cells. (i) Does constitutive (Poisson-like) expression or episodic (bursty) expression dominate throughout the human genome? (ii) Does genomic location influence either mode of expression? For example, if bursting is operant, does genomic location influence burst size or burst frequency, and which is predominantly influenced? Recent studies (17–19) have tackled these questions for specific genes, but no clear and broad consensus has yet emerged.

Results

To globally apply the analytical framework of noise space to screen for constitutive versus bursty expression across the human genome, we capitalized on the semirandom pattern of integration exhibited by the HIV-1 lentivirus, where the majority of integrations (approximately 69%) occur within transcriptionally active regions (20, 21). Jurkat T lymphocytes were infected with HIV-based lentiviral vectors encoding a short-lived, two-hour half-life version of green fluorescent protein (referred to as d₂GFP), to generate a library of cells in which the vector is integrated at a different genomic position in each individual cell (i.e., a polyclonal library) (Fig. 2A). To focus on measuring the intrinsic fluctuation dynamics of the genomic region surrounding the vector-integration site, we utilized a vector encoding the HIV-1 LTR promoter, which is relatively weak and heavily influenced by the expression dynamics of the local chromatin environment (22).

Fig. 2.

Extracting transcriptional parameters from the noise space. In individual isoclones, burst dynamics vary with genomic location. (A) Cells are infected with a lentiviral vector expressing a 2-h half-life GFP reporter (d₂GFP) at a low multiplicity of infection (moi) to ensure a single semirandom integration in each cell. Individual single cells are isolated, grown (creating isoclone populations), and imaged by time-lapse fluorescence microscopy. (B and C) Single-cells are tracked for 12–18 h, and an individual cell’s mean expression level, variance (σ²), and autocorrelation time (τ_1/2) are extracted from the time trace (e.g., the green circle represents a single cell’s noise space coordinate). A constitutive model of gene expression that displays abundance dependence (bold red arrows from black model lines) was used to normalize each cell’s noise magnitude (CV²) and autocorrelation (τ_1/2). The normalized noise magnitudes and autocorrelations are plotted in a Δ log CV² - Δ log τ_1/2 noise map (Left). (D) Consistent shifts to the Upper Right quadrant in Δ log CV² - Δ log τ_1/2 space observed for three LTR isoclones (clones 3, 4, and 5), are indicative of transcriptional bursting relative to the least bursty isoclones (clones 1 and 2). Bursting dynamics varies between different clones as evidenced by shifts in both noise autocorrelation and magnitude. The isoclonal signature is taken from 18 h trajectories of 400 cells.

Initially, the analytical framework for distinguishing bursty expression was applied to five isoclonal populations: Each population was grown from a single parent cell, and all daughter cells, therefore, share the same LTR genomic integration site (Fig. 2 A–D). Cell fluorescence was then imaged for 18 h, and the resulting fluorescence intensity trajectories were used to construct a three-dimensional noise space with three axes: noise magnitude as measured by the coefficient of variation squared (CV²), noise autocorrelation represented by the half-autocorrelation time (τ_1/2), and mean expression level () (Fig. 2 B–C and SI Appendix). Analysis of these trajectories in the noise space allows comparisons between isoclonal populations and differentiation between isoclones that exhibit constitutive transcription versus episodic bursts of transcription (Fig. 2D). To generate an initial baseline origin for noise-space analysis of the single-cell data, we identified isoclones that were the most Poissonian in their expression fluctuation profiles from a library of isoclonal populations (17) (SI Appendix). Two isoclonal populations exhibiting the fastest fluctuation autocorrelation decays (i.e., shortest autocorrelation times) were selected as the most Poissonian and used to establish an origin of the CV²-versus-τ_1/2 noise map (Fig. 2D, clones 1 and 2). The isoclone heat map represents the probability of where a randomly chosen single cell would land on the noise map (Fig. 2D).

Importantly, nontranscriptional phenomena that influence noise behavior (e.g., protein and mRNA lifetimes, GFP maturation, extrinsic noise) are already embedded in the noise of these isoclonal populations, specifically the most Poissonian reference clones. Thus, comparison between these necessarily precludes the possibility that noise-map shifts are due to these nontranscriptional phenomena. In addition, the high-frequency noise-processing technique minimizes extrinsic noise effects (SI Appendix) (12). Because it is likely that these two isoclones are somewhat bursty in their expression (SI Appendix), comparisons to these isoclones represent a highly conservative assay for bursty expression. Nevertheless, the results show that even in a small panel of five clones, there are marked changes in both noise magnitude and autocorrelation time of approximately 1.5-fold in normalized space (Fig. 2D). Clones 4 and 5 show significant changes in autocorrelation time with smaller differences in noise magnitude. These differences in noise magnitude were validated against conventional flow cytometry measurements (SI Appendix, Fig. S1). The agreement between data from the 12–18 h microscopy experiments and flow cytometry demonstrates that this CV²-versus-τ_1/2 analysis has the fidelity to differentiate transcriptional dynamics between different isoclones.

We next extended the analysis to image polyclonal populations—consisting of thousands of integration sites (Fig. 3A)—to globally apply the analytical framework of noise space to screen for constitutive versus bursty expression across the human genome. We analyzed more than 8,000 distinct genomic loci with three different promoters integrated throughout the genome by imaging cells for 12–18 h (Fig. 3B). The constitutive expression origin of the CV²-versus-τ_1/2 noise map previously determined for the most-Poissonian isoclones (Fig. 2D and SI Appendix) was compared to the 8,000 loci. To control for LTR-specific or vector-specific artifacts, we also tested self-inactivating lentiviral vectors that encode either promoters for human elongation factor 1α (EF1A) or human ubiquitin C (UBC) that in turn drive d₂GFP. UBC and EF1A are promoters that drive essential cellular housekeeping genes: UBC promotes the ubiquitinization cascade by marking proteins for proteosomal degradation, and EF1A promotes the GTP-dependent binding of an aminoacyl-tRNA to ribosomes. UBC and EF1A are among the most abundant proteins in eukaryotic cells, and their promoters exhibit robust high-level expression across integration sites in different cell types (23, 24). Virtually all examined genomic loci for all promoters exhibit noise-map shifts to the upper right (Fig. 3B), indicating significant bursting in gene expression at virtually all genomic loci. Similar widespread bursting is also observed in the THP-1 human monocyte cell line and compared to a constitutive simulation (SI Appendix, Fig. S2). Both synchronized and unsynchronized cells exhibit similar shifts in noise space (SI Appendix, Fig. S3), indicating that transcriptional bursts appear to be common throughout the cell cycle, as reported (25).

Fig. 3.

Episodic-bursty expression dominates across the human genome. (A) To create the polyclonal population, cells are infected with a lentiviral vector expressing d₂GFP so that each cell represents a unique clone harboring a single semirandom integration of reporter. (B) Resultant noise maps for over 8,000 individual cell trajectories for the HIV-1 LTR promoter, EF1A promoter, and UBC promoter. The constitutive origin is derived from Fig. 2D (18 h).

The UBC and EF1A promoters showed markedly lower CV² values than the LTR promoter (Fig. 3B and SI Appendix, Fig. S4), which is consistent with our previous study (17) showing that the LTR promoter displays relatively higher levels of noise than other eukaryotic promoters in yeast (6, 7). This shift in noise magnitude is consistent with the well-known transcriptional elongation stall that characterizes LTR expression (26). This stall results in delayed switching to the transcriptional ON state in a two-state transcription model (17) and predicts that noise frequency (not only noise magnitude) (27) is modulated in different genomic or chromatin environments (SI Appendix). In agreement with this prediction, the distribution of points in the LTR noise map indicates significant differences from the UBC and EF1A noise maps. In further support, an alternate representation of the noise-map distribution that converts the distributions to centroids (with error bars) can be used to conveniently visualize these differences between promoters (SI Appendix, Fig. S4) (11).

Given conflicting reports on whether burst frequency varies with genomic location (17–19, 28), we next determined if transcriptional burst size, burst frequency, or both changed across the genome (Fig. 4A). As mentioned above, transcriptional bursting can be quantified by a two-state model of transcription (13, 14) in which switching between the two states occurs at rates k_on and k_off, and transcription only occurs in the ON state with a rate k_m (Eqs. 1–3 and Materials and Methods). The burst size, or number of mRNAs generated per activity pulse, is typically defined as k_m/k_off and, in the limit of k_off≫k_on, the burst frequency is defined as k_on (Fig. 4A) (17). To directly test if transcriptional burst frequency changes across genome location, we analyzed the polyclonal three-dimensional noise-space data to fit values for k_m, k_off, and k_on. LTR polyclonal trajectories are subclustered into groupings of approximately 60 cells, so that each subcluster represents cells in a specific range of gene-expression levels (SI Appendix, Fig. S5), and average noise autocorrelation is calculated for each subcluster by autocorrelation analysis (12). To validate this subclustering approach, we verified that CV² values from subclustered polyclonal trajectories agree with CV² from conventional flow cytometry data from isoclonal populations (Compare Fig. 4C to SI Appendix, Figs. S1 and S6) . Thus, each polyclonal subcluster corresponds to an isoclonal population in terms of average expression level (SI Appendix, Fig. S6). Strikingly, this genome-wide data demonstrate that autocorrelation time first increases with increasing expression and once an expression threshold is reached (gray line, Fig. 4B), autocorrelation time decreases as expression level increases (Fig. 4B). This pattern of concavity is inconsistent with constant burst frequency (i.e., constant k_on) across genomic locations and provides a genome-wide measurement of k_on and k_off changes that is independent of k_m (10).

Fig. 4.

Transcriptional burst frequency and burst size vary equally across the genome and are strongly dependent on expression level. (A) Schematic of the two-state model of transcriptional bursting, where the promoter switches between ON and OFF states at rates k_on and k_off and transcribes at rate k_m in the ON state. Transcriptional dynamics are modulated through changes in burst size, burst frequency, or both. (B) Noise autocorrelation, noise magnitude (C), burst frequency (D), and burst size (E) versus abundance for polyclonal subclusters of 2,000 12-h Ld2G single-cell trajectories. Low and high abundance domains are separated by a solid gray threshold line which indicates the changes in the trends of noise autocorrelation, noise magnitude, and hence burst size and burst frequency is observed. (F and G) As a function of , fold changes in burst size and frequency are comparable, with an initial increase of frequency in all promoters investigated.

Conventional approaches to quantify transcriptional burst kinetics analyze noise magnitude (CV²-versus-) (6, 7) and the polyclonal data from Fig. 3 can also be analyzed in terms of noise magnitude on the CV²-versus- plane of noise space (Fig. 4C). This noise magnitude analysis shows a strong initial decrease in CV² at low expression levels (Fig. 4C). Then when an expression threshold is reached (gray line, Fig. 4C) a leveling off of CV² is observed for higher expression levels (Fig. 4C). However, noise magnitude is insufficient to uniquely parameterize the two-state model. Because burst size couples k_m and k_off, it is only through the τ_1/2 measurement, which is not influenced by k_m, that the two parameters can be differentiated from one another. Note, that the gray line in Fig. 4 B and C correspond to the same expression threshold.

Fitting of the two-state model in the polyclonal three-dimensional noise map space shows a strong initial increase in burst frequency at low expression levels, whereas burst size remains almost constant (Fig. 4 D and E). Upon reaching a threshold expression level (gray vertical line in Fig. 4D), a switch in burst dynamics occurs, and burst size increases, whereas burst frequency remains constant (Fig. 4 D and E). The fold change in transcriptional burst size and burst frequency values reveals that both vary equally across genomic loci (Fig. 4F and SI Appendix, Fig. S6). In addition, the measured burst-size range predicts an average mRNA level of 110 molecules per cell, which is consistent with previous measurements that used single-molecule mRNA fluorescent in situ hybridization (29, 30). The success of fitting the three-dimensional noise space is reflected in the close agreement between a simulated autocorrelation curve and the experimental trend (Fig. 4B) with the fit model parameters. This fit shows that the assumed two-state model is sufficient to describe the measured system (SI Appendix). Much like LTR, UBC, and EF1A display similar fold changes in burst size and frequency and exhibit a similar pattern of increasing burst frequency, followed by increasing burst size (Fig. 4G). These data indicate that integration site influences burst kinetics, irrespective of promoter type (i.e., cis sequence). However, UBC and EF1A exhibit almost constant τ_1/2 at the highest expression levels indicating increases in only k_m at these levels (SI Appendix, Fig. S7). Interestingly, these two strong promoters individually span the range of burst frequencies recently reported for a variety of mammalian genes (18) (SI Appendix, Fig. S7), whereas the LTR functions at much lower burst frequencies (Fig. 4).

To test how transcriptional activators influence burst dynamics (Fig. 4), transcription was perturbed with transcriptional activators, including the histone deacetylase inhibitor (HDAC) trichostatin A (TSA), and the cell-signaling molecule tumor necrosis factor α (TNF). TNF enhances expression by stimulating recruitment of a p50-RelA heterodimer to nuclear factor κB (NFκB) binding sites (31) and the HIV-1 LTR encodes multiple NFκB binding sites and is potently activated by TNF (22). We previously reported that TNF only changes burst frequency of the LTR while conserving burst size in a limited number of isoclones (17), and were interested to see how widespread this phenomena was across the genome. The τ_1/2-versus- analysis of TNF stimulation shows a significant decrease in τ_1/2 with increasing expression level (Fig. 5A). Here, τ_1/2 decrease with increasing abundance is a direct indication of kinetic changes and demonstrates that increasing expression level cannot be explained solely by modulations of k_m (SI Appendix, Eq. S1) (10, 13).

Fig. 5.

Transcriptional burst size and frequency are altered by transcriptional activators. (A–D) TNF addition (filled red circles) shifts the measured integration sites to the higher abundance and burst dynamic domain along the nondrug curve (empty circles). Large autocorrelation shifts implicate changes in burst kinetics. (E) Estimated residence times in the active (ON) and inactive (OFF) states.

Fitting of the different two-dimensional planes of three-dimensional noise space upon TNF induction (Fig. 5 A and B) demonstrates that both burst frequency and size significantly increase as expression levels increase, with burst frequency increasing at low expression levels and burst size increasing at higher expression levels (Fig. 5 C and D). Interestingly, there appears to be a threshold in expression level, above which k_on plateaus to values observed before adding TNF, and k_off appears to decrease. These data and analysis are corroborated by conventional flow cytometry measurements of 35 isoclonal populations (SI Appendix, Fig. S6). Overall, these results suggest that TNF induces expression from the LTR along existing burst trends (Fig. 5C and D), and the use of TSA, which induces expression through a different mechanism than TNF (32), corroborates this observation (SI Appendix). This observed decrease in k_off with TNF induction leads to extended duration of bursts and is consistent with the reported inhibition of p50-HDAC1 repressive-complex formation at LTR NFκB sites by p50/RelA heterodimers (33)—the successful formation of HDAC1 leads to weakened recruitment of RNA polymerase II and weakened transcriptional initiation (34). The observed increases in k_on are also consistent with increased recruitment of RNA polymerase II to the LTR promoter NFκB sites induced by TNF (35, 36). Fitted parameter estimates of LTR residency time in the presence of TNF were used to represent an average over the first 12 h of stimulation given the dynamic nonlinear nature of the NFκB response (37, 38). Collectively, these results enable estimation of LTR residency time in the transcriptional ON and OFF states and show that TNF extends duration in the ON state up to eightfold (Fig. 5E).

Discussion

The analysis of noise space presented here provides a high-throughput method to dynamically profile gene-regulatory mechanisms and the effects of perturbations on gene expression. A significant methodological advantage of analyzing three dimensions of noise space is the ability to more accurately constrain two-state transcriptional burst models and the polyclonal nature of the approach enables shotgun mapping of gene regulation dynamics on a genome-wide scale.

The resulting genome-wide data demonstrate that constitutive transcription is rare across the human genome. Instead, the overwhelming majority of human genomic loci appear to stochastically fire in episodic bursts. Analysis of noise space demonstrated that both transcriptional burst frequency and burst size vary in roughly equal degree across the human genome (Fig. 4 D–G). Intriguingly, there appears to be a threshold expression level below which integrations modulate only transcriptional burst frequency and above which only burst size is modulated (Fig. 4 B–G and SI Appendix, Figs. S6 and S7). This transition could result from recently reported refractory periods inherent to bursting kinetics (18, 39). Burst frequency can be increased at loci where transcriptional bursts are infrequent, but as frequency increases, the refractory period temporally precludes further increases in frequency. Therefore, once this frequency ceiling is reached, the only way to increase expression is to increase the transcription rate or extend the duration of each burst.

As proposed (40), widespread episodic bursting may allow limited transcriptional resources within the cell to be efficiently allocated to achieve high-level transcription across large numbers of loci. This efficient allocation of resources may be the biological analog of time-domain multiplexing approaches used to efficiently transmit data in signal processing applications.

Materials and Methods

Lentiviral Vectors.

Lentiviral vectors were cloned as described (41) and used to infect 5 × 10⁵ Jurkat cells at a multiplicity of infection < 0.1, resulting in 25,000–50,000 infected cells each with a unique integration site. Cells were then sorted by FACS and fluorescently imaged on glass-bottom dishes in RPMI medium 1640 with 10% fetal calf serum and 1% penicillin-streptomycin.

Imaging.

Imaging was performed in humidified conditions at 37 °C and 5% CO₂ for 12–24 h with a 40X (1.2 N.A.) oil-immersion objective on an Olympus DSU microscope equipped with an automated linear-encoded X-Y stage, as described in refs. (12) and (42). Image processing and cell tracking were performed in Matlab with an in-house algorithm (12) and a single 12-h experiment could generate up to 1,000 trajectories for analysis.

Calculations.

For each trajectory, noise autocorrelation (τ_1/2) and noise magnitude (CV²) were calculated using an established noise-processing algorithm (11, 12). A reported theory (10, 13) of the two-state transcriptional bursting model yields analytical expressions for both the autocorrelation of the noise, τ_1/2, and the noise magnitude, CV²(see SI Appendix, Materials and Methods). Detailed discussion of the noise-mapping approach, including analytical arguments and stochastic simulations (of Poissonian gene expression) by the Gillespie algorithm (43, 44), are described in detail in SI Appendix.

Transcriptional burst dynamics are quantified by deriving analytical expressions for burst size and burst frequency with formulations from previous analyses (13, 14, 17) and low promoter activity assumptions where k_off≫k_on, k_off≫k_m, k_off≫γ_p and k_m≫(γ_m + γ_p)):

[1]

[2]

[3]

where BS is the burst size, BF is the burst frequency, k_m is the transcription rate, k_p is the translation rate, γ_m and γ_p are the mRNA and protein decay rates, respectively, is the mean protein abundance, and b is the translational burst rate. , or the mean number of GFP molecules in the measurements, is assumed to be directly proportional to , the mean fluorescence intensity. Eqs. 2 and 3 reveal that measurements of CV² and are sufficient to quantify burst size and burst frequency within a constant, which is only dependent on the translation rate and decay rates of mRNA and protein. Assuming these remain constant, while varying integration site or promoter sequence, an abundance-dependent burst size and frequency trend can be directly resolved.

Note.

Full decomposition of k_off, k_on, and k_mcan only be performed through the use of the full 3D noise space.