NF-κB-Chromatin Interactions Drive Diverse Phenotypes by Modulating Transcriptional Noise

SUMMARY

Noisy gene expression generates diverse phenotypes, but little is known about mechanisms that modulate noise. Combining experiments and modeling, we studied how tumor necrosis factor (TNF) initiates noisy expression of latent HIV via the transcription factor nuclear factor κB (NF-κB) and how the HIV genomic integration site modulates noise to generate divergent (low-versus-high) phenotypes of viral activation. We show that TNF-induced transcriptional noise varies more than mean transcript number and that amplification of this noise explains low-versus-high viral activation. For a given integration site, live-cell imaging shows that NF-κB activation correlates with viral activation, but across integration sites, NF-κB activation cannot account for differences in transcriptional noise and phenotypes. Instead, differences in transcriptional noise are associated with differences in chromatin state and RNA polymerase II regulation. We conclude that, whereas NF-κB regulates transcript abundance in each cell, the chromatin environment modulates noise in the population to support diverse HIV activation in response to TNF.

Graphical abstract

INTRODUCTION

Gene expression noise arises in part because transcripts are produced in episodic bursts during infrequent transitions from inactive to active promoter states, a process referred to as transcriptional bursting (Bahar Halpern et al., 2015; Raj et al., 2006; Singh et al., 2010; Skupsky et al., 2010; Suter et al., 2011). In many systems, regulatory networks amplify gene expression noise to drive phenotypic diversity (Acar et al., 2008; Chang et al., 2008; Weinberger et al., 2005). Thus, modulating transcriptional bursting offers a strategy to alter clinically relevant phenotypic outcomes.

To effectively tune transcriptional bursting, we must better understand its molecular regulatory mechanisms (Sanchez and Golding, 2013; Symmons and Raj, 2016). Evidence supports roles for nucleosome positioning and occupancy (Dey et al., 2015; Raser and O’Shea, 2004), chromatin modifications (Suter et al., 2011; Viñuelas et al., 2013; Weinberger et al., 2012), and transcription factor concentration (Octavio et al., 2009; Senecal et al., 2014), but how these factors interact to regulate bursting and alter phenotypes remains unclear.

HIV is a model system used to study both the sources and phenotypic consequences of gene expression noise and transcriptional bursting. HIV integrates into host genomes, and the chromatin environment at the integration site regulates the population-average expression supported by the HIV long terminal repeat (LTR) promoter (Jordan et al., 2001; Miller-Jensen et al., 2012). HIV LTR-driven gene expression is bursty, and bursting kinetics vary with integration site (Singh et al., 2010; Skupsky et al., 2010). Furthermore, HIV encodes its own positive transcriptional transactivator, Tat (Nabel and Baltimore, 1987), and there is evidence that Tat-mediated amplification of basal HIV transcriptional noise establishes phenotypic diversity in which some viruses replicate while others remain latent (Miller-Jensen et al., 2013; Weinberger et al., 2005).

The persistence of latent HIV reservoirs in T lymphocytes is a major barrier to curing HIV infections. One promising eradication strategy is to use small molecules or proteins to activate HIV transcription within these reservoirs to render viruses susceptible to existing antiretroviral therapy (ART; Margolis et al., 2016). However, this approach is limited, because latent viral activation appears stochastic in patients (Ho et al., 2013). Recent studies suggest that modulating transcriptional bursting, and thus noise, at latent HIV promoters may increase their likelihood of response to exogenous stimulation (Dar et al., 2014). Still, the ability to systematically alter viral phenotypes via modifying bursting will depend on identifying its molecular regulators.

The HIV LTR is regulated by nuclear factor κB (NF-κB), a family of transcription factors that activate or repress transcription (Hayden and Ghosh, 2008). Selective induction of NF-κB-target genes ensures proper regulation of many processes, including inflammation and development (Smale, 2011). NF-κB-target promoters have varied chromatin environments that affect selectivity (Ramirez-Carrozzi et al., 2009; Saccani et al., 2001), but how NF-κB together with chromatin modulates transcriptional bursting is unknown. In basal conditions, the NF-κB heterodimer RelA:p50 is preferentially shuttled to the cytoplasm by inhibitor of NF-κB-alpha (IκBα). Stimulation by tumor necrosis factor (TNF) leads to IκBα degradation, allowing RelA:p50 to accumulate in the nucleus, bind to target gene promoters, and stimulate transcription (Hayden and Ghosh, 2012). Cell-to-cell variability in RelA activity has been correlated to variability in target gene outputs (Lee et al., 2014; Sung et al., 2014). Thus, we hypothesized that RelA-chromatin interactions may contribute to the regulation of transcriptional bursting and phenotypic diversity in latent HIV activation.

Here, we demonstrate that NF-κB and chromatin act to modulate transcriptional bursting at latent HIV LTR promoters to drive divergent TNF-induced viral activation. We observed distinct HIV transcriptional noise patterns after TNF treatment, indicating that TNF can increase transcription by changing either burst size or burst frequency, and we show that these differences are sufficient to drive divergent viral phenotypes when amplified by HIV-Tat positive feedback. We find that cell-to-cell variability in RelA activity in part controls strength of viral activation, while differences in RNA polymerase II (RNPII) regulation underlie transcriptional noise and bursting. Altering LTR-proximal chromatin can change transcriptional bursting and viral phenotypes. We expect that our results showing how RelA and chromatin each act to tune transcriptional bursting and noise will apply more broadly to endogenous RelA-target promoters.

RESULTS

Latent HIV Integrations Exhibit Phenotypic Diversity in Response to TNF Treatment

To study TNF-induced phenotypic diversity of latent viral activation and the underlying sources of regulation, we established clonal populations of Jurkat T lymphocytes with unique HIV integrations that are phenotypically silent in the basal state but are activated by TNF (“latent-but-TNF-inducible”; Figure 1A). Cells were infected with HIV lentiviral constructs driving expression of GFP and HIV Tat to retain the Tat-mediated positive feedback that contributes to viral phenotypic diversity (Figure S1A; Weinberger et al., 2005). Two of these clonal populations (J-Lat 8.4 and 10.6) have been studied extensively to determine molecular mechanisms of HIV latency (Blazkova et al., 2009; Jordan et al., 2003; Williams et al., 2006). An additional pair was isolated from a Jurkat cell line stably expressing NF-κB mCherry-RelA (Ch-RelA) (J65c 4.4 and 6.6; Figures S1A–S1D) so that we could directly explore how NF-κB signaling dynamics contribute to HIV phenotypic diversity (see below).

Figure 1. TNF Induces Diverse Viral Activation Phenotypes Associated with Different Chromatin Environments.

(A) Diagrams of HIV-LTR integration types (left) and potential sources of regulation of HIV-LTR activation investigated in this study (right). Integrations can be latent (silent with or without TNF), active (expressing even with no TNF), or latent but TNF inducible (focus of this study, with a spectrum of low to high activating phenotypes with TNF). Latent-but-TNF-inducible HIV-LTR activation may be modulated by chromatin state, NF-kB signaling, RNPII regulation, and transcription noise, which itself may be amplified by Tat positive feedback. We investigate both early signaling and transcription response (<4 hr after TNF addition) and resulting phenotypic outcomes (12–24 hr) as measured by onset time, maximal expression, and fraction of activating cells.

(B) Histograms of HIV transcript number per cell in the basal state (no TNF, no CHX), as measured by smFISH for indicated clonal single-HIV-integrant Jurkat populations. J-Lat 8.4 basal transcript distribution is significantly different from the others (p < 0.05 by Kolmogorov-Smirnov [K-S] test); all others are similar [p > 0.05]; n = 63–210 cells per clone).

(C) Bar graph of mean number of HIV transcripts from distributions in (B). Error bars represent 95% confidence intervals (CIs) of the basal mean (bootstrapping for combined smFISH experiments). Independent replicates quantifying basal J65c transcription distributions were not significantly different (Figure S1E).

(D) Single-cell time courses of HIV-GFP signal after TNF addition (no CHX), measured by fluorescence time-lapse imaging. Cells with measurable activation are shown from ~40 to 80 cells imaged per clone (J65c 6.6, n = 54 cells; J65c 4.4, n = 25 cells; J-Lat 8.4, n = 16 cells; and J-Lat 10.6, n = 34 cells).

(E and F) Bar graphs of mean HIV-GFP onset time (E) and maximum GFP fluorescence intensity (F) by 12 hr after TNF addition (no CHX) for LA (J-Lat 8.4 and J65c 4.4, orange) and HA (J65c 6.6 and J-Lat 10.6, blue) clones. Error bars represent bootstrapped 95% CIs.

(G) Bar graphs of the percentage of HIV-GFP+ cells 24 hr after TNF addition (no CHX) for LA (orange) and HA (blue) clones. We show mean ± SD of independent biological triplicates.

(H and I) Bar graphs of basal enrichment (% input by ChIP) of histone H3 (H) and AcH3 (I) at the LTR and GAPDH for LA (orange) and HA (blue) clones. Data from cells pretreated with 160 ng/mL CHX for 1 hr (mean ± SD of independent biological duplicate). *p < 0.05, **p < 0.005 by ANOVA. No significant differences in H3 and AcH3 were measured for GAPDH across clones.

(J) Time courses of enrichment (percent input by ChIP) of RelA at the LTR for LA J65c 4.4 (orange) and HAJ65c 6.6 (blue) after TNF addition. We show mean ± SD of independent biological duplicates; we found no significant differences in time course as calculated by two-way ANOVA.

See also Figure S1.

To confirm that phenotypically silent (GFP–) cells have equivalently low levels of Tat activity across all clonal populations, we measured HIV-Tat transcripts using single-molecule RNA fluorescent in situ hybridization (smFISH). The transcript count distributions across all four populations showed near-silent transcription in the basal state (<2 mean transcripts per cell; Figures 1B and 1C), with J-Lat 8.4 having the lowest counts or the most transcriptionally repressed behavior. A small fraction of cells from each clonal population exhibited viral activation in the basal state (0.1 to 6% GFP+; Figure S1D), and these were excluded from the study. Spontaneous activation of GFP− cells in the absence of TNF occurs on a slow timescale of days to weeks (Miller-Jensen et al., 2013), and thus by focusing our analysis on initially GFP− cells, we could specifically investigate TNF-induced activation (Figure 1A).

TNF activates latent HIV by initiating transcription that is amplified by Tat-mediated positive feedback (Figure 1A), as Tat binds to nascent HIV RNA and promotes its elongation (Barboric et al., 2001). Despite similar latent HIV transcriptional profiles across these clonal populations, TNF induced divergent phenotypic behavior with differences in both dynamics and extent of activation (Figure 1D). J-Lat 8.4 and J65c 4.4 showed slow activation with low maximum GFP in only a small fraction (J-Lat 8.4) or a minority (J65c 4.4) of cells, while J-Lat 10.6 and J65c 6.6 both increased GFP earlier and to higher levels in a majority of cells (Figures 1E–1G). In all four clones, some cells showed activation, and some did not. Thus, divergent phenotypic behavior occurred both between cells with the same viral integration (cell-to-cell variability) and between populations with different viral integrations (clone-to-clone variability). Hereafter, we refer to divergent clonal behaviors as low activating if activation is slow, low, and only takes place in a minority of cells (LA; J-Lat 8.4 and J65c 4.4) and high activating if activation occurs relatively early and robustly in a majority of cells (HA; J-Lat 10.6 and J65c 6.6). These clonal populations are representative of the range of phenotypes observed when activating HIV latency experimental models with TNF (Spina et al., 2013).

Differences in inducible latent viral activation have been linked to differences in the basal chromatin environment at the LTR promoter (Miller-Jensen et al., 2012), and therefore, we measured the relative levels of histone 3 (H3) and acetylated H3 (AcH3) at these LTRs by chromatin immunoprecipitation (ChIP). Promoter-bound H3 is linked to nucleosome occupancy (Kim and O’Shea, 2008; Rafati et al., 2011), while AcH3 is linked to promoter accessibility and transcriptional competence (Lusic et al., 2003; Van Lint et al., 1996). Overall, we found that LA LTRs had more H3 and less AcH3 compared to HA LTRs (Figures 1H and 1I). Of note, these differences did not alter transcription factor accessibility, because we observed equivalent RelA binding at LA and HA LTRs at the peak of TNF-induced RelA translocation (Figures 1J and S1F). Thus, differential sensitivity to TNF between LA and HA clones is linked to differences in basal H3 occupancy and H3 acetylation, but not to differences in RelA binding at the LTR.

Viral Integrations with Distinct Activation Phenotypes Exhibit Different TNF-Induced Transcriptional Noise

We next sought to identify clone-to-clone differences in early, TNF-dependent transcription that could be amplified by Tat to achieve low versus high viral activation. To separate the effects of TNF and Tat on HIV transcription in the first 4 hr after TNF addition, we blocked the TNF-induced production of Tat with a low dose of the translation inhibitor cycloheximide (CHX). The low CHX dose was sufficient to block most of the new TNF-induced Tat-mediated GFP production, even for a 24-hr TNF treatment (Figure S2A), as previously described (Blazkova et al., 2009), but did not affect early TNF-induced RelA translocation dynamics (Figure S2B).

We quantified HIV transcription in single cells by smFISH (Figures 2A and 2B) and found that 2 hr after TNF addition in the presence of CHX, the fraction of cells with at least three HIV transcripts increased for all four clones (Figure 2C). Furthermore, mean transcript number increased similarly in LA J65c 4.4 and HA clones (Figure 2D), despite large differences in dynamics and extent of HIV-GFP expression in the 12 hr after TNF addition in the presence of Tat positive feedback (Figures 1D–1F). Even the repressed LA J-Lat 8.4 clone increased its mean transcripts per cell, although less so than other clones. Based on similar average TNF-dependent early transcription and RelA binding at the LTR across these different clones (Figure 1J), we conclude that low versus high GFP expression cannot be explained by different thresholds of RelA activation set by chromatin.

Figure 2. HIV Transcription Increases to a Similar Mean Level but with Distinct Noise in LA versus HA Clones after TNF Treatment without Tat Positive Feedback.

(A) Representative maximum intensity projection images of an HIV-GFP smFISH fluorescence z stack in LA J65c 4.4 (left) or HA J65c 6.6 (right) cells before (top) and 2 hr after (bottom) 20 ng/mL TNF treatment in the presence of CHX. Blue dashed lines outline cells. Initially GFP+ cells were excluded from additional analyses. Scale bar, 5 μm.

(B) Histogram of HIV-GFP transcript number per cell before (gray) and after (red) a 2-hr (J65c 4.4, J65c 6.6 and J-Lat 10.6) or 4-hr (J-Lat 8.4) TNF treatment with 160 ng/mL CHX. Each distribution has 30 bins; n = 196 - 650 cells per condition.

(C–F) Bar graphs of fraction of cells with at least 3 transcripts (C), mean transcript numbers per cell (D), CVs (E), and Fano factors (F) before and after a 2-hr or 4-hr 20 ng/mL TNF treatment with 160 ng/mL CHX for LA clones (J-Lat 8.4 and J65c 4.4; orange) and HA clones (J-Lat 10.6 and J65c 6.6; blue). Error bars represent bootstrapped 95% CIs, combining the data from independent smFISH experiments (n > 196 cells for each condition) (Supplemental Experimental Procedures). The difference between two moments was inferred to be significant (p < 0.05) if the 95% CIs did not overlap.

See also Figure S2.

Despite similarities in average early TNF-dependent transcription across LA and HA clones, their transcript number distributions were distinct. HA clones had a skewed distribution, with a tail of cells producing many transcripts (30+), while most cells produced very few; in contrast, distributions for LA clones were broader and less skewed (Figure 2B). To quantify the differences in noise across HIV transcript number distributions, we compared coefficient of variation (CV), the relative variance of transcript numbers, and Fano factor, a measure of how much a distribution deviates from Poissonian behavior. For LA clones, CV decreased ~2-fold with TNF treatment, and the Fano factor remained the same (Figures 2E and 2F). By contrast, CV did not change for HA clones, while the Fano factor significantly increased (Figures 2E and 2F). When Tat production was not blocked, TNF-induced transcription increased similarly in all clones except J-Lat 8.4 (Figures S2C and S2D), but the trends in CV and, most strikingly, Fano factor were still distinct between LA and HA clones (Figures S2E and S2F), showing that Tat activity does not alter these differences. Notably, transcript number distributions and noise trends for NFKBIA (encoding IκBα), an early and abundantly transcribed endogenous RelA-target gene, were similar for both LA and HA clones, confirming that RelA signaling does not differ clone to clone and that not all RelA-target promoters display clone-to-clone differences in TNF-induced transcription (Figures S2G–S2J). Together, these observations reveal that TNF differentially activates transcription at latent HIV integrations with low versus high activation phenotypes, even as it stimulates similar RelA signaling.

Simulations Show that Differences in TNF-Induced Transcriptional Bursting Can Cause Divergent Viral Activation in LA versus HA Clones

Previous studies have shown that differences in transcription factor mechanism of action can be inferred by how they affect transcriptional noise at inducible promoters (Neuert et al., 2013; Senecal et al., 2014) (Box 1). In the two-state model of promoter activity (Figure 3A, black), activation of transcription occurs by increasing either burst frequency (rate of transition to active state) or burst size (mean number of transcripts produced per burst in active state). We fit transcript number distributions for LA and HA clones pre- and post-TNF addition to the predicted transcript number probability density function of the two-state model (Dey et al., 2015; Raj et al., 2006). We found that the two-state model closely fit all distributions (Figure S3). In untreated cells, model-inferred transcriptional burst frequencies were similar for all clones, but the burst size for J-Lat 8.4 was lower, consistent with the lower basal mean (Figures 3B and 3C). Model fits were consistent with TNF treatment markedly increasing burst size without changing burst frequency for HA clones (Figures 3B and 3C). By contrast, fits for LA clones were consistent with TNF treatment significantly increasing burst frequency with no increase (J-Lat 8.4) or only a modest increase (J65c 4.4) in burst size (Figures 3B and 3C). Our results confirm that TNF-stimulated RelA activates transcription differently at LA versus HA latent viral integrations.

Box 1. Transcriptional Noise and the Two-state Model of Promoter Activity.

If a gene promoter is constitutively active and transcription occurs stochastically, but at a constant rate, then the transcript number per cell becomes a Poisson process and transcript number distribution should follow a Poisson distribution (Peccoud and Ycart, 1995). However, transcript number distributions for many genes, including for HIV, are better fit by hypergeometric distributions (or, in continuous form, gamma distributions; Friedman et al., 2006), which are replicated by the two-state model of promoter activity (Peccoud and Ycart, 1995; Raj et al., 2006).

In the two-state model, a promoter is assumed to switch between an “on” state, producing transcripts at a constant rate, and an “off” state. Therefore, transcription occurs in sporadic bursts (see Box 1 diagram above) and transcription activation is assumed to result from increasing either average burst frequency (the rate of transition to on state, k_a), or average burst size (the number of transcripts produced in a burst, b = k_m/k). Burst size increase could occur via an increased transcript production rate (k_m) and/or a decreased rate of transition to the off state (k_i).

The two-state model is surely an oversimplification of the transcription process, and models with three states (e.g. on, off-repressed, and off- poised; Bintu et al., 2016; Chavali et al., 2015) and even a continuum of states (Corrigan et al., 2016) have been proposed. Still, the two-state model is useful because of the clear mathematical relationships between the model parameters of burst frequency and burst size and the moments of an expected distribution of transcript numbers, which are mathematically related to distribution noise (Skupsky et al., 2010). When an increase in gene expression is caused by a burst frequency increase, noise (CV = σ/μ) decreases monotonically with expression (σ²/μ₂ < σ₂/μ₂). In contrast, when increased gene expression is caused by a burst size increase, there is a corresponding increase in noise strength (Fano factor, F = σ²/μ, thus ${\mathrm{\sigma }}_2^2/{\mathrm{\mu }}_2>{\mathrm{\sigma }}_1^2/{\mathrm{\mu }}_1$) and no change in noise (Carey et al., 2013; To and Maheshri, 2010). Therefore, if measured transcript number distributions can be fit by a two-state model, then the model can infer specific values of burst frequency and burst size, providing interpretable quantities to describe a transcription process.

Figure 3. Two-State Model Fits Are Consistent with TNF Treatment Increasing Transcription Burst Frequency in LA Clones and Burst Size in HA Clones, Accounting for Differences in Viral Activation.

(A) Schema of the two-state promoter model (black) and transcription amplification by Tat positive feedback (gray). Burst frequency (k_a) and burst size (b = k_m/k_i) were fit based on measured transcript number distributions; Tat positive feedback was simulated computationally.

(B and C) Bar graphs of burst frequency (B) and burst size (C) parameter fits for transcript number distributions measured before and after 20 ng/mL TNF treatment (with 160 ng/mL CHX; no Tat feedback) for LA (J-Lat 8.4 and J65c 4.4; orange) and HA clones (J-Lat 10.6 and J65c 6.6; blue). Error bars show bootstrapped 95% CIs. Differences between parameters were inferred to be significant (p < 0.05) if 95% CIs did not overlap.

(D) Simulations of Tat protein abundance (model equivalent of measured HIV-GFP) after TNF treatment of HA and LA clones when amplified by Tat positive feedback. For each clone, basal steady state was simulated with burst sizes and burst frequency values similar to fitted values derived from experimental data. To simulate TNF treatment, at time t = 0, burst frequency (LA clones, left) or burst size (HA clones, right) was increased to the fitted value, whereas all other parameters remained the same. Tat protein levels were simulated for n = 500 individual cells for 12 hr (Supplemental Experimental Procedures).

See also Figures S3 and S4.

Our results suggest that model-inferred burst size increases in HA clones are efficiently amplified by Tat positive feedback, but burst frequency increases in LA clones are not, leading to the observed differences in viral phenotypes. To test this hypothesis, we simulated viral activation in HA and LA clones using a mathematical model of a two-state LTR promoter with Tat positive feedback (Figure 3A, gray) (Chavali et al., 2015). First, we identified parameter values describing Tat feedback that reproduced the low transcript counts observed in the basal state (Figure S4A). To simulate TNF activation, we changed burst frequency or burst size parameters at time t = 0 to match the model-inferred parameters for LA and HA populations after 2 hr of TNF treatment and simulated viral activation with Tat positive feedback over 12 hr (Supplemental Experimental Procedures). We found that our simulations qualitatively recapitulated the observed differences in rate and level of viral expression for LA versus HA clones (Figure 3D).

Simulating model behavior over a range of burst size and burst frequency values (Figure S4), we found qualitative evidence of HA behavior arising when burst size is high (>10), while LA behavior is maintained when burst size is low (<2), even for high values of burst frequency (e.g., J-Lat 8.4). These results suggest that a certain burst size may be required to generate enough Tat molecules to initiate Tat-driven positive feedback. Because of the instability of transcripts and/or Tat protein, more frequent but smaller transcription bursts may be less likely to generate the required number of Tat molecules. Importantly, for intermediate burst size increases, as inferred for LA J65c 4.4, LA versus HA behaviors depend on the burst frequency increase (Figure S4). Thus, we conclude that the differences in transcriptional bursting after TNF addition are sufficient to support substantial phenotypic diversity when coupled to Tat positive feedback.

RelA Signaling Quantitatively Determines Relative Viral Activation Independent of the Transcriptional Noise Profile

In addition to clone-to-clone variability, we observed cell-to-cell variability in viral transcription and activation within each clonal population after TNF addition (Figure 1D). This raises the question: is cell-to-cell variability in viral activation mainly driven by stochastic events like transcriptional bursting (i.e., intrinsic noise), as has been shown for latent HIV in unstimulated cells (Weinberger et al., 2005)? Alternatively, does cell-to-cell heterogeneity in TNF-induced RelA nuclear translocation, an extrinsic noise source observed in Jurkat (Figure S5A) and other cell types (Lee et al., 2014; Nelson et al., 2004), contribute to variability in viral activation? Intrinsic and extrinsic noise contributions are not mutually exclusive, but their relative importance will affect phenotype distributions. If extrinsic noise dominates, we expect that within each clone, RelA signaling will strongly correlate with the strength of Tat-mediated viral activation. If intrinsic noise dominates, we expect little to no correlation with RelA.

To distinguish between these possibilities, we tracked nuclear translocation of stably integrated Ch-RelA and HIV-GFP activation dynamics in the same single cells by time-lapse imaging for both LA J65c 4.4 and HA J65c 6.6 (Figures 4A and 4B). Prior to TNF addition, nuclear Ch-RelA fluctuated only slightly (Figure S5B); after TNF addition, it increased transiently in both populations (Figure 4B). We analyzed RelA nuclear translocation dynamics in single cells by extracting quantifiable features from individual Ch-RelA nuclear intensity time courses (Cohen-Saidon et al., 2009; Lee et al., 2014) (Figure S5C). The average nuclear RelA translocation and CVs of features extracted from single-cell time courses were very similar for LA J65c 4.4 and HA J65c 6.6 cells (Figures S5D an S5E), yet their viral activation dynamics were significantly different (Figures 1D–1G and 4C).

Figure 4. RelA Signaling Accounts for Cell-to-Cell Variability in Viral Activation for Both LA and HA Clones.

(A) Time-lapse images Ch-RelA (top) and HIV-GFP (bottom) in HA J65c 6.6 cells after treatment with 20 ng/mL TNF (no CHX). Cells with strong (arrows) and weak (asterisks) nuclear translocation of Ch-RelA are indicated. Scale bar, 10 μm.

(B) Single-cell time courses of nuclear Ch-RelA intensity for LA J65c 4.4 (top) and HA J65c 6.6 (bottom) cells treated with 20 ng/mL TNF (no CHX). Time courses for HIV-GFP signal from the same cells are in Figure 1D (n = 68 cells [HA J65c 6.6] and n = 77 cells [LA J65c 4.4]).

(C) Time courses for pairs of LA J65c 4.4 (orange) and HA J65c 6.6 (blue) cells selected for similar Ch-RelA nuclear translocation (left) and their respective viral activation (right).

(D) Scatterplots of maximum nuclear RelA (F_max; top) and maximal nuclear RelA fold change (F_max/F_i; bottom) versus HIV-GFP area under the curve (AUC) for LA J65c 4.4 (left) and HA J65c 6.6 (right). Spearman correlation coefficients (r_s) and their p values are reported.

(E) Bar graph of Spearman correlation of each nuclear Ch-RelA dynamic feature (defined in Figure S5C) with HIV-GFPAUC (defined in Figure S5F) for HA J65c 6.6 (blue) and LA J65c 4.4 (orange). Error bars show bootstrapped 95% CIs; *p < 0.001.

See also Figure S5.

To determine whether RelA signaling could predict viral activation within each clonal population, we extracted quantitative descriptors of the GFP, or viral activation, time courses (Ramji et al., 2015; Figure S5F) and then calculated within-clone correlations between features of Ch-RelA translocation dynamics and HIV-GFP expression in the same cells (Figures 4D, 4E, and S5G). Maximum nuclear Ch-RelA intensity (F_max) after TNF addition was a moderate predictor of HIV-GFP expression for HA J65c 6.6 but a poor predictor for LA J65c 4.4 (Figure 4D, top). In contrast, maximum nuclear Ch-RelA fold change (F_max/F_i) was a moderate predictor of TNF-induced HIV-GFP gene expression for LA J65c 4.4 and strong predictor for TNF-treated HA J65c 6.6 cells (Figure 4D, bottom). Overall, although nuclear RelA fold-change signaling features were the strongest predictors of viral activation for both clones, measures of absolute nuclear Ch-RelA concentration were adequate predictors only in the HA population (Figure 4E and S5G). This LA versus HA difference may be due to differences in how RelA activates transcription in these clones. Nevertheless, our results suggest that cell-to-cell variability in RelA signaling partially explains within-clone phenotypic diversity, although other factors, perhaps including intrinsic noise, must also contribute.

RelA-Mediated Increases in Burst Size or Burst Frequency Are Associated with Different Mechanisms of RNPII Regulation

At its target promoters, RelA can directly recruit general transcription factors and other modulators of RNPII activity (van Essen et al., 2009), as well as chromatin-remodeling proteins that facilitate histone acetylation (Zhong et al., 2002). Therefore, we considered whether changes in burst size versus burst frequency were linked to differences in RelA-mediated molecular events at the LTR.

First, we measured changes in H3, AcH3, and RNPII at the LTR in HA J65c 6.6 and LA J65c 4.4 cells at 2 and 4 hr after TNF addition using ChIP. We found that after TNF addition, H3 decreased and AcH3 increased at both LTRs (Figures 5A and 5B). However, even though H3 dropped to similar occupancy at both LTRs by 2 hr, AcH3 levels, and thus the AcH3:H3 ratio, remained consistently higher at the HA LTR (Figure 5C). Moreover, significantly more RNPII was bound to the HA LTR versus LA LTR by 4 hr after TNF addition, despite similarly low basal RNPII binding at both LTRs (Figure 5D).

Figure 5. RelA and Chromatin Linked to Differential Regulation of RNPII Activity at the HIV LTR in LA versus HA Clones.

(A–C) Plots showing enrichment (% input by ChIP) of histone H3 (A) and AcH3 (B) at the LTR for HA J65c 6.6 (blue) and LA J65c 4.4 (orange) in the basal state (0 hr) and after treatment (2 and 4 hr) with 20 ng/mL TNF with 160 ng/mL CHX. Ratio of AcH3:H3 (C) was calculated as ratio of % inputs. All show mean ± SD of independent biological duplicates.

(D–G) Enrichment of RNPII (D), ser5-p RNPII (E), ser2-p RNPII (F), and (G) NELF-E for HA J65c 6.6 (blue) and LA J65c 4.4 (orange) as quantified by ChIP. The same conditions as in (A)-(C) were used, and data represent mean ± SD of independent biological duplicates (or triplicate for RNPII).

For (A)–(G), statistical significance between the HA J65c 6.6 and LA J65c 4.4 time courses are indicated (two-way ANOVA; *p < 0.05; **p < 0.005; ***p < 0.0001; n.s., not significant). The change between 0 hr and 4 hr was statistically significant for all HA J65c 6.6 time courses and for H3, AcH3, and the AcH3:H3 ratio LA J65c 4.4 time courses.

(H) Diagram of our working model of molecular events regulating transcriptional and phenotypic outcomes at latent-but-TNF-inducible HIV LTRs (gray, literature inferred; black, supported by our data). At HA LTRs (bottom), high AcH3 levels are linked to efficient RelA-mediated recruitment of initiating RNPII (ser5-p) and NELF in addition to elongating RNPII (ser2-p), yielding increased transcription burst size and high activation. At LA LTRs (top) with lower AcH3 levels, RelA- mediated histone acetyl transferase (HAT) binding and chromatin remodeling must occur before efficient recruitment of initiating RNPII (ser5-p) and only burst frequency is increased, yielding low activation. HDAC is recruited to HIV-LTRs by NF-κB p50:p50, and TSA and SAHA are two HDAC inhibitors used to increase AcH3 at HIV-LTRs.

See also Figure S6.

As a point of reference, we measured enrichment for the same factors at the NFKBIA promoter (Figure S6). We found that in the basal state, only the LA LTR had significantly less AcH3 than NFKBIA, but both LA and HA LTRs had more H3, and thus a lower AcH3:H3 ratio and also lower RNPII binding, than NFKBIA (Figures S6A–S6D). However, after 4 hr of TNF treatment, AcH3 and RNPII enrichment at the HA LTR, but not the LA LTR, approached that measured at the NFKBIA promoter (Figures S6A–S6D). Therefore, the model-inferred increase in burst size at the HA LTR was associated with consistently higher AcH3:H3 ratio and a TNF-induced increase in RNPII binding, both approaching levels comparable to those at the promoter of the highly transcribed NFKBIA.

Even though the TNF-induced increase in RNPII binding was significantly greater at the HA versus the LA LTR, both J65c LTRs supported similar mean transcription during the first 4 hr (Figure 2D). To explain this paradox, we looked for differences in RNPII post-translational modifications. While initiating RNPII is phosphorylated at serine 5 (ser5-p RNPII), elongating RNPII is phosphorylated at serine 2 (ser2-p RNPII); both forms of RNPII are required for production of transcripts, but accumulation of initiating RNPII can signal transcriptional pausing (Adelman and Lis, 2012). Using ChIP, we found that TNF induced agreater enrichment of ser5-p RNPII at the HA J65c 6.6 LTR than at the LA J65c 4.4 LTR after 2 and 4 hr (Figure 5E). In contrast, both LTRs showed similar enrichment in elongating ser2-p RNPII (Figure 5F), consistent with their similar average gene expression. This enrichment of ser2-p RNPII at both LTRs was notably lower than that at the NFKBIA promoter, consistent with their much lower transcription (Figures S6F, S2G, and S2H).

To understand what may prevent initiating RNPII from fully elongating transcripts at HA LTRs, we considered reports showing that the negative elongation factor (NELF) pauses initiating RNPII by forming stable RNPII-NELF complexes proximal to the HIV LTR and endogenous promoters (Gilchrist et al., 2010; Henriques et al., 2013). We measured enrichment of the NELF subunit NELF-E at HA and LA J65c LTRs and found that TNF increased NELF-E recruitment only to the HA J65c 6.6 LTR (Figure 5G). The buildup of both initiating ser5-p RNPII and NELF-E at the HA J65c 6.6 LTR within 4 hr of TNF addition indicates that RNPII is selectively initiated yet paused at HA LTR, but not at LA LTR, and this is associated with a change in transcriptional noise (Figure 5H).

Modifying Basal H3 Acetylation Is Sufficient to Change RNPII Regulation, Transcriptional Noise, and Activation Phenotype

LA LTRs have higher H3 occupancy and lower H3 acetylation than HA LTRs. Thus, we considered whether the HA versus LA differences in TNF-mediated transcription could be altered by changing the chromatin environment at the LTR (Figure 5H). We pretreated LA J65c 4.4 cells with the histone deacetylase (HDAC) inhibitor trichostatin A (TSA) for 4 hr in the presence of CHX to block Tat production. After TSA pretreatment, H3 acetylation and H3 occupancy at the LA J65c 4.4 LTR now resembled the basal state of the HA J65c 6.6 LTR (Figure 6A). TSA pretreatment modestly increased the basal-state mean transcripts per cell, but not the mean transcripts at 4 hr post-TNF addition (Figure 6B). When comparing the TNF-induced transcript number distributions, we found that TSA pretreatment significantly altered the distribution for LA J65c 4.4 (Figures S7A–S7C), changing its activation behavior to resemble that of HA clones. Fits of transcript number distributions to the two-state model were consistent with TSA-pretreated LA J65c 4.4 cells having a substantial TNF-induced increase in HIV LTR transcriptional burst size, but not in burst frequency (Figure 6B).

Figure 6. TSA Pretreatment Increases Acetylation at the LA J65c 4.4 LTR and Changes Its Transcriptional and Activation Response to One Similar to that of the HA J65 6.6 LTR.

(A) Bar graph showing basal state enrichment (% input by ChIP) of histone H3 (left), AcH3 (center), and their ratio (right) at the LTR for LA J65c 4.4 treated with 400 nM TSA (with 160 ng/mL CHX) for 4 hr (green bars), compared to basal state results for untreated LA J65c 4.4 (orange) and HA J65c 6.6 (blue) cells replotted from Figures 1H and 1I. Data represent means ± SD of independent biological duplicates.

(B) Bar graphs of mean mRNA per cell (left), burst size (middle), and burst frequency (right) for TSA-pretreated LA J65c 4.4 (green bars) before and after 2- or 4-hr treatment with 20 ng/mL TNF (with CHX), compared to no-TSA data for LA J65c 4.4 (orange dots) and HA J65c 6.6 (blue dots) replotted from Figures 2D, 3B, and 3C; error bars represent bootstrapped 95% CIs.

(C–E) Plots of ChIP-quantified enrichment of RNPII (C), ser5-p RNPII (D), and ser2-p RNPII (E) for LAJ65c 4.4 + TSA (green) before and after 2- or 4-hr treatment with 20 ng/mL TNF (with CHX). For comparison, no-TSA data for LA J65c 4.4 (orange) and HA J65c 6.6 (blue) cells are replotted from Figures 5D–5F; all shown as mean ± SD of biological duplicates. HA J65c 6.6 and LA J65c 4.4 + TSA are not statistically significantly different (two-way ANOVA, p > 0.05).

(F and I) Schema of the timeline of pretreatment for 400 nM TSA (F) and 20 ng/mL TNF (I).

(G and J) Plots of percentage of GFP+ cells for LA J65c 4.4 and HA J65c 6.6 measured by flow cytometry after 0, 12, and 24 hr of 20 ng/mL TNF treatment (no CHX), after indicated TSA (G) or TNF pretreatment (J). Means ± SD of independent biological duplicates are shown. Statistical significance between each pair of pretreated versus non-pretreated condition time courses is indicated (two-way ANOVA; *p < 0.05; **p < 0.01; ***p < 0.001; n.s., not significant).

(H and K) Plot of activation synergy from indicated TSA (H) or TNF pretreatment (K) quantified by the Bliss independence model for LA J65c 4.4 (orange markers) and HA J65c 6.6 (blue markers). Data represent mean ± SD of independent biological duplicates.

See also Figure S7.

To determine whether changing basal H3 acetylation also affected RelA-mediated RNPII regulation, we measured ChIP enrichment of RNPII, ser5-p RNPII, and ser2-p RNPII after TNF treatment. We found that TSA pretreatment significantly increased the TNF-induced enrichment of RNPII and ser5-p RNPII in LA J65c 4.4 to levels similar to those of HA J65c 6.6 (Figures 6C and 6D); in contrast, ser2-p RNPII did not change (Figure 6E). Overall, increasing basal H3 acetylation and decreasing H3 occupancy at the LA J65c 4.4 LTR is linked to an increase in initiated and paused RNPII in response to TNF. Moreover, RelA-mediated RNPII pausing is linked to increases in transcriptional burst size, but without RNPII pausing, transcription seems to proceed via burst frequency increases.

We next asked whether the TSA-induced changes in RNPII enrichment and transcriptional noise altered viral activation behavior. We found that with Tat positive feedback (i.e., no CHX), TSA pretreatment significantly increased the fraction of GFP+ cells in response to sustained TNF for LA J65c 4.4 and J-Lat 8.4, but not HA J65c 6.6 and J-Lat 10.6 (Figures 6F, 6G, and S7D). This is consistent with our conclusion that increased burst size drives Tat-mediated HIV activation more efficiently than increased burst frequency. TSA pretreatment induced viral activation synergistically with TNF for LA clones (J65c 4.4 and J-Lat 8.4), but not for HA clones (J65c 6.6 and J-Lat 10.6; Figures 6H and S7E). Pretreatment of LA J65c 4.4 with another HDAC inhibitor, suberoylanilide hydroxamic acid (SAHA), also changed TNF-induced transcriptional noise to be more HA-like. The effect of SAHA was weaker than that of TSA (Figures S7F–S7I) and the inferred “switch” from increasing burst frequency to burst size was more modest (Figure S7J), consistent with the more modest increase in viral activation for SAHA (Figure S7L). Overall, our data show that it is possible to increase viral activation in LA clones by modifying TNF-induced transcriptional regulation and noise via HDAC inhibitors.

Similar to HDAC inhibitors, TNF-induced RelA binding also lowers H3 and increases H3 acetylation at the LTR (Figures 5A and 5B). Thus, we asked whether TNF pretreatment of LA J65c 4.4 cells would also synergistically increase viral activation upon a second TNF treatment. Indeed, pretreating LA J65c 4.4 cells with a 30-min or 2-hr “pulse” of TNF, allowing the cells to rest, and then treating again with sustained TNF (Figure 6I) strikingly increased the fraction of LA J65c 4.4 cells activating HIV-GFP after 12 and 24 hr, even though the initial pulse itself did not activate these cells (Figures 6J and 6K). By contrast, the TNF pulse alone activated a fraction of HA J65c 6.6 cells, and the TNF pulse pretreatment did not increase the fraction of cells that were activated by sustained TNF treatment (Figures 6J and 6K), similar to our observations for TSA pretreatment of these cells. Taken together, our data strongly support a role of H3 acetylation in regulating transcriptional bursting and clone-to-clone variability in viral activation.

Latent Viruses Occupy Genomic Integrations that Support a Diversity of Inducible Phenotypes

In HIV-infected patients, latent-but-inducible HIV integrations activate to reseed high levels of viral infection in the absence of ART, and ongoing efforts aim to develop drugs that purge these latent viruses from patients (Margolis et al., 2016). Transcriptional noise profiles consistent with transcriptional bursting have been observed at all HIV integration positions studied, but latent viruses inducible by TNF or other drugs are rare (Figure 1A). Thus, we considered if the genomic locations occupied by these latent viruses confer specific noise characteristics that affect reactivation and treatment.

We compared our basal (no TNF) smFISH data for latent but inducible HIV integrations to smFISH transcript data collected for HIV LTR integrations with basal transcription in the absence of Tat (Dey et al., 2015) and analyzed the RNA mean versus noise relationship (noise defined as relative variance or CV²; Figure 7A). Our latent LTRs had much lower transcription than these LTRs, yet they still fell along the same inverse power-law noise-versus-mean trend line. Thus, HIV latency seems to be established at integrations with very low mean, but highly noisy, basal transcription.

Figure 7. Latent-but-Inducible HIV LTRs Occupy Genomic Locations with Noisy Basal Transcription and Divergent TNF Responses.

(A) Log-log graph of mean versus noise (CV²) of basal mRNA distributions for latent LTRs from this study (blue and orange) and LTRs measured in Dey et al. (2015) (gray). Linear regression (R²) is reported (p < 0.001).

(B) Schema showing theoretical lines of Poisson and bursty transcription across the genome for log-log mean versus noise (CV²) plots. Increasing transcription can occur by increasing burst size (moves to a new trend line) or frequency (moves along the same trend line).

(C) Log-log graph of mean versus noise (CV²) showing shifts in transcript distribution from before (filled circles) to after (open circles) a 2-hr 20 ng/mL TNF treatment (with CHX) for LA clones (orange; along the same trend line) and HA clones or TSA-pretreated LA J65c 4.4 (blue; shift to a new burst size trend line).

Information about how perturbations affect burst size and frequency can be inferred from the noise-versus-mean plots (Dar et al., 2012; Dey et al., 2015; Box 1). Increasing transcription solely by increasing burst frequency results in a shift along the negatively sloped line of constant burst size and decreasing noise. In contrast, increasing burst size results in a shift to a new burst size trend line with higher noise for the same mean (Figure 7B). Using this framework, we mapped how TNF affected transcriptional bursting across latent viral integrations (Figure 7C). TNF treatment of LA clones increased mean transcription along the existing burst size trend line, characteristic of increased burst frequency, while TNF treatment of HA clones transferred transcript distributions to a new burst size trend line, characteristic of increased burst size. These observations raise the possibility that latent but inducible HIV viruses have chromatin features that regulate a bifurcation in bursting trends in response to RelA signals. Together, our results suggest that latent viruses occupy integrations that support large differences in TNF-induced viral phenotypic outcomes. These differences might facilitate viral persistence and complicate purging strategies in the absence of chromatin perturbation.

DISCUSSION

Transcriptional bursting is a major source of gene expression noise, but few studies focus on the phenotypic consequences and molecular mechanisms associated with different modes of bursting. We used clonal T cell populations with unique latent HIV integrations to determine how NF-κB acts with chromatin to modulate bursting and drive strongly divergent viral activation, a phenomenon with clinical relevance. By connecting molecular events from signaling, promoter binding, and transcriptional noise to network amplification, we explain how the same NF-κB RelA signal at the same promoter drives these divergent phenotypes via chromatin-mediated modulation of transcriptional bursting.

We find that TNF induces distinct changes in transcriptional noise profiles that are sufficient to explain clone-to-clone phenotypic variability. Our observations strongly support a role for RNPII promoter-proximal pausing in regulating transcriptional burst size. RNPII promoter-proximal pausing occurs at many mammalian gene promoters and downstream of the HIV LTR transcription start site (Adelman and Lis, 2012). Pausing assists productive elongation by priming RNPII to rapidly respond to elongation signals (Henriques et al., 2013). We propose that assembly of paused RNPII complexes at HA LTRs prepares them to receive other elongation signals (here, likely the Tat-mediated recruitment of the human positive elongation factor b [p-TEFb]) (Barboric et al., 2001). By contrast, at LA LTRs, RelA leads to transcription initiation and elongation, but, without assembly of paused RNPII complexes, elongation enhancement by Tat is inefficient. We note that, although we suppressed new Tat production with CHX, it is hard to fully rule out the possibility that residual Tat in a few rare cells may enhance the inferred increase in burst size. However, because Tat does not promote RNPII recruitment (Mbonye and Karn, 2017), we conclude that RelA-chromatin interactions are needed to establish the different promoter configurations that ultimately modulate transcriptional noise, differential amplification by Tat, and viral phenotype.

Differential RNPII regulation at LA versus HA clones is correlated with differences in basal AcH3:H3 ratios, and HDAC inhibitor pretreatment experiments suggest that H3 acetylation has a role in regulating RelA-mediated events at the LTR. Other epigenetic modifications have also been implicated in HIV latency (Blazkova et al., 2009; Boehm et al., 2017; Pearson et al., 2008); it is likely that some of these also regulate viral phenotypic diversity. One possibility is that different chromatin modifications determine different timescales of provirus epigenetic memory (Bintu et al., 2016). For example, some modifications may impart long-term memory determining whether latency is reversible or not, while others (e.g., histone acetylation) may impart short-term memory determining the extent of latency reversal after stimulation.

Interestingly, RelA seems to quantitatively regulate cell-to-cell variability in viral activation independent of the mechanism for transcription increase. We find that, like the transcriptional output of endogenous genes in HeLa cells (Lee et al., 2014), viral activation is correlated to the maximum nuclear RelA fold change. Our finding that fold-change detection (FCD) of nuclear RelA holds for an exogenous viral promoter, as well as across chromatin environments and in cells of different origin (lymphocytes versus epithelial cells), shows that FCD is a general mechanism of signal encoding. FCD requires a mechanism to establish memory of baseline nuclear RelA at the promoter, and our previous work suggested that the NF-κB p50:p50 homodimer, which competes with RelA for binding at target promoters and is positively regulated by RelA activation (Ten et al., 1992), could provide this mechanism (Lee et al., 2014). Interestingly, NF-κB p50:p50 is bound at latent-but-inducible HIV LTR promoters (Williams et al., 2006), suggesting that p50 binding may be a shared mechanism mediating FCD at the HIV LTR and endogenous promoters. Moreover, p50:p50 recruits HDAC-1 to the LTR, leading to histone H3 deacetylation (Williams et al., 2006). In the future, characterizing in-depth the role of p50:p50 at the HIV LTR should elucidate whether it explains why FCD seems more critical for determining responses at LA versus HA promoters.

Previous studies have used HIV LTRs integrated throughout the genome to assess how genomic location alters transcriptional bursting (Dar et al., 2012, 2016; Dey et al., 2015; Singh et al., 2010; Skupsky et al., 2010). These studies examined HIV integrations supporting moderate-to-high basal activity (5 to ≥40 average mRNAs) in the absence of stimulation or Tat feedback. Interestingly, we find that the inverse scaling between noise and mean observed for such integrations (Dar et al., 2016) is maintained for the LTRs in our study which exhibit near-silent basal activity (average <2 mRNAs). Previous studies that measured how long-term TNF treatments (18 hr) modulate bursting found that TNF increases burst frequency at LTR integrations with moderate basal gene expression but increases burst size at LTR integrations with high basal gene expression (Dar et al., 2012, 2016). We also find that TNF can increase either burst size or burst frequency and additionally show that this distinction is not always associated with differences in basal expression. Further studies will be needed to directly relate the previously reported TNF effects on constitutively expressed HIV LTRs to our results examining how TNF affects inducible gene expression within 4 hr of TNF addition.

We focused on inducible transcription from LTRs with nearly silent basal activity, as they are representative of latent HIV integrations maintained long-term in host cells. Previous studies hypothesized that burst size increases may drive HIV activation more efficiently than burst frequency increases and that this may be important for HIV latency treatment (Dar et al., 2014; Rouzine et al., 2014). Our measurements of transcript distributions and simulations of Tat positive feedback confirm this hypothesis, showing that RelA-induced burst size increases at latent-but-TNF-inducible integrations can induce robust viral activation, while RelA-induced burst frequency increases cannot. Moreover, we add molecular underpinnings to this, finding that changes in bursting kinetics are linked to differences in histone acetylation and in the buildup of primed and paused RNPII upon TNF addition. Overall, our findings suggest that the chromatin features that maintain these latent viral transcript levels may uniquely support large differences in transcriptional bursting mode following treatment with TNF and other activators, which may complicate efficient latency reversal.

EXPERIMENTAL PROCEDURES

Further details and an outline of resources used in this work can be found in Supplemental Experimental Procedures.

Cell Lines, Retroviruses, and Lentiviruses

Jurkat T cell clone E6-1 and HEK293T cells were obtained from ATCC and J-Lat 8.4 and 10.6 (Jordan et al., 2003) from the NIH AIDS Research and Reference Reagent Program (NIAID). J65c 4.4 and 6.6 were created as described in Figure S1; pMSCV-mCherry-RelA was packaged into an amphotropic retrovirus in HEK293Ts via Ca₃(PO₄)₂ transfection. psLTR-Tat-GFP lentivirus was packaged, harvested, and titered as described previously (Miller-Jensen et al., 2013).

smFISH Microscopy and Image Analysis

After treatment, cells were fixed, hybridized with probes overnight as described previously (Lee et al., 2014) using conditions optimized for LTR probe sets, and then imaged on a spinning disk confocal microscope (100× oil objective). FISH-Quant software (Mueller et al., 2013) was used to quantify mRNA content.

Live-Cell Imaging and Analysis

Live-cell imaging was performed in an environmentally controlled chamber on a spinning disk confocal microscope (60× oil objective). After stimulation, images were captured at 3-min (mCherry) and 10-min (GFP) intervals. Data were extracted from background-corrected images in ImageJ; mean fluorescence intensity of nuclear Ch-RelA and total HIV-GFP was manually collected for each cell at each time point.

ChIP

EZ ChIP Kit (Upstate) reagents and protocols were used as described previously (Miller-Jensen et al., 2012). DNA isolated from ChIP was quantified by qPCR (Bio-Rad iCycler, iQ5) using SYBR Green Supermix (Bio-Rad).

Fitting Two-State Model

Maximum-likelihood estimation (MLE) was used to select burst frequency (k_a) and size (k_m/k_i) parameters that best fit the measured RNA distributions as described previously (Dey et al., 2015).

Mathematical Model and Computational Simulations of HIV Activation

We built a stochastic model according to our prior work (Chavali et al., 2015) with some modifications (Supplemental Experimental Procedures).

Statistical Analyses

Differences between transcript distributions were determined by Kolmogorov- Smirnov test (p < 0.05) and between two time courses by a two-way ANOVA (p < 0.05). 95% confidence intervals (CIs) on descriptive statistics of RNA distributions were estimated from the 2.5% and 97.5% quantiles of bootstrapped copy numbers per cell as described (Dey et al., 2015). 95% CIs on fit parameters (burst size and frequency) were estimated from the log-likelihood function assuming asymptotic normality of the estimates and using 1.92 log-likelihood ratio units as described previously (Dey et al., 2015). For CIs obtained empirically, the difference between two quantities was inferred to be significant if the 95% CIs did not overlap. All statistical test sand regression and correlation analyses were performed in Prism (GraphPad) unless otherwise indicated.

Highlights.

• TNF-induced transcriptional noise patterns predict diverse activation of latent HIV

• Fold change in nuclear RelA determines strength of viral activation in single cells

• Modifying transcriptional noise by targeting chromatin changes viral activation

• Changes in transcriptional noise are associated with altered regulation of RNAPII