Fold change of nuclear NF-κB determines TNF-induced transcription in single cells

Abstract

In response to tumor necrosis factor (TNF), NF-κB enters the nucleus and promotes inflammatory and stress-responsive gene transcription. Because NF-κB deregulation is associated with disease, one might expect strict control of NF-κB localization. However, nuclear NF-κB levels exhibit considerable cell-to-cell variability, even in unstimulated cells. To resolve this paradox and determine how transcription-inducing signals are encoded, we quantified single-cell NF-κB translocation dynamics and transcription in the same cells. We show that TNF-induced transcription correlates best with fold change in nuclear NF-κB, not absolute nuclear NF-κB abundance. Using computational modeling, we find that an incoherent feedforward loop, from competition for binding to κB motifs, could provide memory of the preligand state necessary for fold-change detection. Experimentally, we observed three gene-specific transcriptional patterns that our model recapitulates by modulating competition strength alone. Fold-change detection buffers against stochastic variation in signaling molecules and explains how cells tolerate variability in NF-κB abundance and localization.

Figure 1. TNF-induced NF-κB sub-cellular localization is variable

(A) Fixed-cell RelA immunofluorescence images of HeLa cells treated with 10 ng/mL TNF for indicated times; scale bar is 10 μm. (B&C) Frequency histograms for total nuclear (B) and nuclear density (C) of endogenous RelA for HeLa cells treated (red) or not (blue) with 10 ng/mL TNF (t = 30 min, dose and time when nuclear translocation was maximal; Figure S1); n = 800–1000 cells.

Figure 2. TNF-induced NF-κB translocation varies in live cells

(A) Time-lapse images of stable HeLa FP-RelA treated with 10 ng/mL TNF. Arrow and asterisk indicate nuclei of cells with different FP-RelA translocation dynamics; scale bar is 10 μm. (B) Single-cell FP-RelA nuclear density time courses quantified from time-lapse images of HeLa FP-RelA treated with 10 ng/mL TNF. To reduce the influence of high frequency noise in mean fluorescence intensity, nuclear time courses were represented as 3-frame running averages (Figure S2). Inset defines time course descriptors. (C) Bar graph of the coefficients of variation (CV) for select time course descriptor of FP-RelA nuclear density (descriptors are defined in Figure S2). Error bars represent the standard deviation of the mean (S.E.M.) for triplicate experiments.

Figure 3. Variability of TNF-induced NF-κB-dependent transcription is transcript specific

(A) Transmitted light, (B) Hoechst channel and (C) maximum intensity projection images of a representative TNFAIP3 smFISH fluorescence z-stack. The perimeters of nuclei are marked (dashed yellow line) and side view projections of z-stack are depicted. (D) Linescans from image in (C) demonstrate the high signal-to-noise ratio for typical mRNAs (blue line) and active transcription sites (ATS, red line) where nascent transcripts accumulate on the gene locus (Raj et al., 2008). (E) Images merging Hoechst and fluorescent channels for control samples that were not exposed to smFISH probes but were otherwise treated identically to all other samples. (F) Transmitted light and IL8, TNFAIP3 and NFKBIA smFISH images of untreated and TNF-treated HeLa. Nuclei were counterstained with Hoechst (blue). (G) Boxplots of the IL8, TNFAIP3 and NFKBIA mRNAs/cell distributions for parental (P) and FP-RelA HeLa lines and indicated treatments. Red bars and notches indicate median and 95% confidence interval; statistical significance of differences were assessed by two-sample Kolmogorov-Smirnov test (n > 45, **p ≪0.01; Figure S3); scale bars are 10μm for all.

Figure 4. Transcriptional responses to TNF are determined by fold-change of nuclear NF-κB

(A) Workflow connecting live-cell imaging of FP-RelA nuclear translocation to same-cell smFISH (Movie S2). (B) Assessment of correlations between descriptor and mRNA number. Example weak and strong correlations are shown, corresponding to mean nuclear fluorescence at t=30′ (F_t=30′) and maximum nuclear fold-change (F_max/F_i) respectively. (C) Bar graph of the coefficient of determination (R²) of each nuclear FP-RelA descriptor for three NF-κB-dependent transcripts (see also Figure S4). (D) Plots showing the R² of listed descriptors at each time point for IL8, TNFAIP3 and NFKBIA. The straight red line indicates the R² for maximum fold-change of nuclear NF-κB, the strongest single predictor of the transcriptional response for all three genes ((C) and Figure S4). The summary of cell numbers for same-cell FP-RelA translocation and transcription experiments is in Table S3.

Figure 5. An I1-FFL model of NF-κB-mediated transcription recapitulate experimental transcriptional patterns

(A) Schematic diagram of an I1-FFL network motif. (B&C) Diagram of the NF-κB-induced transcriptional network showing the pulse-generator (I) and transcriptional (II) modules for direct (B) and I1-FFL-like (C) transcriptional models. (D&E) Scatter plots of transcript numbers vs. total FP-p65 in untreated cells (left), and vs. maximal nuclear FP-RelA fold-change for TNF-treated cells (center, data from cells treated with 0.1, 1 and 10 ng/mL TNF are all plotted on the same graph). Bar graphs show the relative variance for all three genes at different fold-change levels (right; Figure S7). Results are shown for simulations with high (cyan), moderate (red) and low (yellow) affinity competition (D) and for experiments for IL8 (cyan), TNFAIP3 (red), and NFKBIA (yellow) (E).

Figure 6. Individual genes show different sensitivity to knockdown of candidate competitors

(A) Graphs of the predicted change in transcript abundance (expressed as fold induction over the no-knockdown condition) as a function of competitor knockdown efficiency for baseline conditions (left) and TNF-treatment conditions (right). Simulations of the D2FC model were run using initial conditions mimicking parental HeLa cells (see Figure S6C and its legend for details). As affinity of the competitor for the target gene promoter was increased, the predicted change in abundance also increased (arrow). Regions of high and low competitor affinity, used to model baseline and TNF-induced transcription corresponding to IL8 and NFKBIA, are shown as yellow and blue regions respectively. (B) Bar graphs showing fold induction in transcript abundance over control siRNA (si-ct) condition as measured by quantitative PCR for parental HeLa cells transfected with siRNA targeting three candidate competitors (p50, p52 and BCL3). IL8 (yellow bars) and NFKBIA (blue bars) mRNA levels were quantified in baseline condition (untreated cells; left) and in cells treated with 10 ng/ml TNF for 60 min (right). mRNA abundances that are significantly greater than in control siRNA conditions are marked with their p-values (grey lines, two-tailed t-test). Asterisks mark values where IL8 siRNA abundance changes significantly more than the abundance of NFKBIA mRNA (p < 0.05 in a one-tailed t-test). Error bars represent the standard deviation from at least four independent siRNA transfection experiments (C&D) Scatter plots on the left show correlation between nuclear densities of RelA vs. p50 (C) or BCL3 (D). Histograms of total nuclear Hoechst intensity (middle) were used to categorize cells as G1 (smaller nuclei, less DNA content) or G2 (larger nuclei, more DNA content) and scatter plots on the right compare nuclear densities for cells with similar DNA content.

Figure 7. The model explains how transcription patterns are tuned by changes to competitor affinity and abundance

(A) Hypothetical plots mapping the multifactorial system that regulates gene expression. ‘Hard-wired factors’ can be described using plots showing hypothetical affinity of competitor complex vs. RelA dimer affinity for a series of competitor complexes (left); shaded regions represent a hypothetical space occupied if plotting values for all κB sites and squares show that the binding affinity for different competitor complexes can be different for two hypothetical κB sites. Competitor affinity, RelA dimer affinity and competitor identity are only three axes in a larger multidimensional space representing all the factors that affect NF-κB driven gene expression (right). (B) Schematics for repressed, inducible and constitutive patterns of transcription (also see Figure S7). (C) Matrix of scatter plots showing transcript number vs. fold-change in nuclear RelA for simulations with increasing competitor abundance (from left to right) and increasing competitor affinity for target promoter (from bottom to top). In the model, competitor expression and affinity are lumped variables representing the aggregate abundance of all competitor complexes that regulate a gene, and their combined affinity for the κB sites in the promoter. Patterns were classified as ‘constitutive’, ‘inducible’ or ‘repressed’ (see Parameter Sweep section in the Supplementary Information for additional discussion).