Cells respond dynamically to pulsatile cytokine stimulation. Here we report that single, or well-spaced pulses of TNFα (>100 min apart) give a high probability of NF-κB activation. However, fewer cells respond to shorter pulse intervals (<100 min) suggesting a heterogeneous refractory state. This refractory state is established in the signal transduction network downstream of TNFR and upstream of IKK, and depends on the level of the NF-κB system negative feedback protein A20. If a second pulse within the refractory phase is IL-1β instead of TNFα, all of the cells respond. This suggests a mechanism by which two cytokines can synergistically activate an inflammatory response. Gene expression analyses show strong correlation between the cellular dynamic response and NF-κB-dependent target gene activation. These data suggest that refractory states in the NF-κB system constitute an inherent design motif of the inflammatory response and we suggest that this may avoid harmful homogenous cellular activation.
In biological systems, timing is critical in the precise order of events required to produce a functional signalling molecule, to the accurate interpretation of temporally encoded signals that determine cell fate. Cellular fate decisions may vary from fully committed binary outcomes, for example, live or die1, to graded responses that are fine-tuned according to the changing amplitude, duration and intensity of the signal2. Surprisingly, growing evidence suggests these responses might in fact be random and subject to changes over time3. This has been attributed to intrinsic noise in gene expression4, heterogeneous dynamics of key transcriptional networks5 as well as the existence of multiple cellular states in genetically identical populations6,7. Cells must reproducibly discriminate varying environmental signals over time; however, how these apparently heterogeneous responses may be coordinated in single cells and cellular populations is not fully understood.
The nuclear factor kappa B (NF-κB) transcription factor is among the best characterized mammalian signalling systems involved in an immune response8, and its deregulation is associated with inflammatory disease and cancer9. NF-κB p65 exhibits heterogeneous nuclear-to-cytoplasmic oscillations in its cellular localization in response to tumour necrosis factor α (TNFα)10,11,12,13, a principal inflammatory signalling molecule. These dynamics are in part due to NF-κB-dependent transcription of inhibitory kappa B protein family (mainly IκBα and IκBɛ), which regulate intracellular localization of the NF-κB (refs 10, 14). Changes in oscillation frequency were associated in part with differential gene expression15, suggesting that the NF-κB system, like calcium Ca2+ (ref. 16) and other biological oscillators5, may be capable of decoding extracellular signals by frequency. The activation of the NF-κB system is also encoded digitally, as the decrease of the TNFα concentration over four orders of magnitude (or the level of antigen stimulation in lymphocytes17) resulted in fewer responding cells in the population2,18. Additional analogue parameters, including the amplitude of NF-κB nuclear translocation, among others, also contributed to the downstream gene expression patterns2,15,19. A long-term pulsed cytokine input resulted in more synchronous NF-κB translocations and increased downstream gene expression, compared with a continuous treatment, suggesting that the NF-κB system may be capable of encoding rapidly changing environmental signals20.
The regulation of the IκB kinase (IKK) has been proposed to be particularly relevant for the temporal control of NF-κB responses21. IKK integrates different signals ranging from stress, bacterial endotoxin or cytokine stimulation, such as TNFα and interleukin 1β (IL-1β)22,23. Stimulus-dependent activation of IKK, a multi-protein complex composed of IKKα, IKKβ and a catalytic subunit NEMO, leads to degradation of IκB inhibitors and release of NF-κB into the nucleus8. IKK activity is temporally controlled via conformational and phosphorylation cycles24, which are dictated by a range of mechanisms. These involve a network of complex and not fully resolved interactions including over 20 molecular species, for example, TRAFs and RNF11 adaptors, RIP and TAK1 kinases as well as IRAK1-4, TAK1, Lubac, Cezanne, ABIN, Tpl2 and Itch among others8,25,26. These proteins play a key role in transduction of different signals; for example, TNFα and IL-1β act via their cognate receptors and have been shown to converge on IKK via TRAF2/RIP and TRAF6/IRAK-mediated pathways, respectively27,28,29,30. The responsiveness of cells may be limited by their ability to regulate IKK via A20 (a dual-function deubiquitinase and E3 ligase enzyme) in a negative NF-κB-dependent transcriptional feedback loop15,31. Indeed, A20-deficient cells are unable to resolve NF-κB response due to disruption of ubiquitin-controlled IKK activity29,32.
During inflammation, cells are often exposed to a rapidly changing tissue-level environment. For example, TNFα may be produced by a number of immune cells, at different times and doses, creating steep local concentration gradients over time33,34,35. IL-1β may be released in bursts associated with inflammasome-mediated activation of myeloid cells36. In this work, we investigated how dynamic cytokine signals are encoded by the NF-κB system in individual cells. We employ time-lapse microscopy to track p65 and IκBα fluorescent fusions in human S-type neuroblastoma (SK-N-AS) cells in response to pulses of TNFα and 1L-1β at different time intervals ranging from 50 to 100 min. We discovered that the observed single-cell NF-κB dynamics are controlled by a stimulus-specific refractory state, which varied between cells, resulting in cellular heterogeneity. The refractory state is not stochastic, because it was conserved in individual cells for several hours. Mathematical modelling predicts that it was due to pseudo-stable cellular states or levels of protein in the signal transduction pathway. We hypothesize that refractory states might be a part of an inherent design motif in the inflammatory response, which enables precise control of the tissue-level inflammatory response by the timing of cytokine stimulation and cellular heterogeneity.
NF-κB p65 and IκBα oscillate out-of-phase in single cells
To investigate cellular responses to temporal cytokine input, we visualized NF-κB system dynamics using time-lapse fluorescent microscopy10,15,18. We produced SK-N-AS cells (referred to here as C9 cells) stably expressing 3′UTR eGFP-tagged full-length human IκBα gene from a bacterial artificial chromosome (BAC) construct (Fig. 1a, see also Supplementary Note 1 for cell line development). Additionally, we also introduced a lentiviral construct expressing a NF-κB p65-mCherry fusion protein37 (Fig. 1a, referred to as C9L cells). Resting cells showed a cytoplasmic localization of both IκBα-eGFP and p65-mCherry. Treatment with 10 ng ml−1 of TNFα (or IL-1β) led to a rapid and synchronous degradation of IκBα-eGFP, which coincided with a nuclear p65-mCherry translocation (Fig. 1b–d, see Supplementary Table 1 for the number of cells analysed in specific experiments). The corresponding NF-κB p65 translocation amplitudes were very similar for both cytokine treatments (Fig. 1e, represented by a ratio of p65-mCherry nuclear to total intensity with the total concentration of p65 fusion per cell constant in time10,15,37). The timing of degradation and re-synthesis of IκBα-eGFP matched those exhibited by endogenous IκBα after TNFα stimulation (see Supplementary Figs 1 and 2 for immunoblotting analysis).
Over a long-time course (up to 800 min), single cells exhibited persistent oscillations in the level of IκBα-eGFP corresponding to multiple rounds of IκBα re-synthesis and degradation (Fig. 1c, see Supplementary Fig. 3 for the conserved NF-κB dynamics in different model systems). The IκBα-eGFP oscillations were out-of-phase with respect to nuclear translocations of p65-mCherry, as shown by the frequency analysis (Supplementary Fig. 4). In agreement with previous data in SK-N-AS cells10,15,18, NF-κB oscillations had an average period of ∼100 min, but were heterogeneous between cells, as depicted by individual peak-to-peak timings, Fig. 1f. Frequency analysis demonstrated that many individual cells exhibited distinct periods ranging from 80 to 140 min (see Fig. 1g and Supplementary Fig. 5 for power spectrum analysis, Fig. 1h for distribution of single-cell periods). This highlighted that some key aspects of NF-κB responses might be differentially controlled by parameters inherent to individual cells13.
Cells exhibit refractory periods to pulsatile stimulation
To understand how cells may encode changing inflammatory signals, first, C9 cells were stimulated with a single 5-min pulse of 10 ng ml−1 TNFα (by manually changing media on cells). This resulted in a single synchronous degradation of IκBα-eGFP with a trough at 20 min in all cells (as shown by a population average of 63 cells, Fig. 2a, mirroring previously described NF-κB p65 nuclear translocation15). Subsequently, IκBα-eGFP levels peaked at 125 min and returned to the pre-stimulation levels between 200 and 300 min. However, when two 5-min pulses of TNFα were applied at a 60 min interval (first pulse at time 0 min followed by a second pulse at 60 min), the responses to a second pulse were more asynchronous in comparison with responses to a first pulse, or single-pulse stimulation (Fig. 2a versus b). Clustering analysis of individual traces showed that while all the cells responded to the first pulse, only ∼30% (out of 113) cells responded to a second pulse of TNFα, based on degradation of IκBα-eGFP (Fig. 2c,d for images of representative cells). In the remaining 70% of cells IκBα-eGFP continued to accumulate after treatment, giving a profile similar to the single-pulse stimulation response (Supplementary Fig. 6). In comparison, when two pulses of IL-1β were applied at a 60-min interval, 75% of cells responded to the second pulse, suggesting that cell’s responsiveness is stimulus-specific (Supplementary Fig. 9).
Pulsatile TNFα stimulation was then applied at time intervals ranging from 50 to 100 min (as shown in Fig. 2e). We found that while all the cells responded to the first TNFα pulse, at shorter pulse intervals (<100 min) cells increasingly failed to respond to the second pulse (Fig. 2f, see Supplementary Note 2, Supplementary Fig. 7 for single-cell traces, and Supplementary Fig. 8 for the clustering analysis of ‘responding’ and ‘non-responding’ cells). At a 100-min pulse interval, 93% of cells responded to both pulses; however, the fraction of responding cells decreased to 70 and 30% at 70 and 60 min pulse intervals, respectively. With a 50-min interval, only 5% of cells showed NF-κB activation to a second pulse. These data suggest the existence of a refractory state induced by the initial stimulation, in which cells become temporally unresponsive to TNFα. The refractory state is heterogeneous, as the time period where cells are unresponsive to the second TNFα pulse (defined here as the refractory period) varied across the population (Fig. 2g, calculated based on the data in Fig. 2f, see Materials and methods section) with ∼40% of cells characterized by a refractory period between 60 and 70 min.
Previous analyses suggested that the timing of TNFα pulsing regulated the amplitude of NF-κB nuclear translocation15. In IκBα-eGFP p65-mCherry C9L cells pulsed at 100 min interval (see Fig. 2h for single-cell traces, Fig. 2i for fraction of responding cells), the average amplitude of p65-mCherry nuclear translocation was similar to that of cells stimulated with continuous TNFα (80% of the first-peak amplitude, Fig. 2j). However, when pulsed at 60 min, the nuclear p65-mCherry amplitude was significantly lower and equated to 60% of the first-peak amplitude.
Altogether, these data showed that individual cells exhibited stimulus-specific heterogeneous activation (as exhibited by IκBα-eGFP degradation) in response to pulsatile cytokine stimulation, with fewer cells responding at shorter TNFα pulsing intervals. The amplitude of NF-κB p65 nuclear translocation in cells that responded depended on the length of the pulsing interval, suggesting both digital and analogue signal-encoding elements. These responses were apparently stochastic, with most of the variability being captured by the fraction of responding cells (coefficient of variation of 1.1 at 60 min), rather than the amplitude of activation (coefficient of variation of 0.3 for the peaks 2 to 1 NF-κB p65 nuclear intensity ratio in responding cells).
Refractory period is encoded in the IKK network
Stimulus-induced activation of IKK is regulated by temporally coordinated conformational and phosphorylation cycles24. Existing models suggest that, on stimulation, IKK undergoes a rapid activation and before it can be reactivated must return to the unstimulated state2,14,15,23,28,31. We therefore compared responses to two TNFα pulses (TT) to stimulation with pulses of TNFα and IL-1β (TI). While only 30% of cells responded to two pulses of TNFα at a 60-min interval (Fig. 2f), 95% of cells responded when TNFα-treated cells were stimulated with IL-1β after 60 min (see Fig. 3a for an average of 95 C9 cells and Fig. 3b for the fraction of responding cells). Also, the amplitude of the second NF-κB p65-mCherry nuclear translocation in responding cells was higher when treated with IL-1β rather than TNFα (Fig. 3c) suggesting parallel signal transduction to IKK (refs 27, 28, 29, 30). Immunoblotting for IκBα protein level and serine 536-phosphorylated p65 (ref. 10) in WT SK-N-AS cells, confirmed NF-κB system activation in response to a second pulse of IL-1β, but not TNFα (Fig. 3d). This experiment suggested that the refractory period following TNFα stimulation is controlled upstream of IKK in the TNFα transduction pathway. 54% of cells responded to a TNFα pulse 60 min after a pulse of IL-1β (Supplementary Fig. 9), indicative of some cross-talk between IL-1β and TNFα signal transduction pathways.
We then investigated whether the TNFα refractoriness depended on the receptor availability2. First, cells were stimulated with 5 min pulses of fluorescein isothiocyanate (FITC) or Tx-Red fluorescently labelled TNFα at 60 min intervals. In agreement with previous reports38, a 5-min pulse produced rapid cell surface binding and internalization of FITC-labelled TNFα (see Fig. 3e and Supplementary Fig. 10 for representative cells and Fig. 3f for population level). When re-stimulated after 60 min with Tx-Red-labelled TNFα, all cells again exhibited rapid binding and internalization of the labelled cytokine indicative of TNFα receptor expression (Fig. 3e,f, and Supplementary Note 3 and Supplementary Figs 11 and 12 with reverse labelling). Importantly, when the availability of the tumour necrosis factor receptor-1 (TNFR1), a cognate TNFα receptor38, was measured directly by flow cytometry, no change of cell surface expression was observed between different pulsing protocols (in comparison to untreated cells, Fig. 3g). This suggests a model where the TNFR1 cell surface availability is constantly maintained via efficient trafficking to the cell membrane39, with the NF-κB response being independent of endocytosis (Supplementary Fig. 13). Overall, these data show that the TNFα signal was decoded in a transduction pathway upstream of the IKK but downstream the TNFα receptor.
Signal transduction controls single-cell responses
To theoretically investigate the encoding of pulsed TNFα stimulation in the NF-κB system, the structure2 and cell-system specific parameters14,15,18 of previously developed models were considered (see Supplementary Note 5 for model development and validation, and Supplementary Figs 14–20; refs 21, 28, 32, 40). The mathematical model comprised two interconnected modules (Fig. 4a). The base module described the interaction between NF-κB and its IκBα negative feedback15, whilst the IKK module depicted the transduction pathway downstream of TNFα and IL-1β. In particular, the model incorporated an upstream generic IKK kinase (IKKK) and A20 feedback protein, which together regulate IKK activity29,32,41. In this description, IKKK represents a large network of complex and not fully characterized molecular interactions (for example, TRAF adaptors, RIP and TAK1 kinases)8,30, while the A20 also includes other A20-like feedbacks (for example, cIAP or Cezanne)42 and proteins that control their activity (for example, ABIN, Lubac and TPL2)25,26,29. Based on the data in Fig. 3 and Supplementary Fig 9, we assumed that the TNFα and IL-1β signals converge on the IKK via signal-specific IKKK kinases (see Supplementary Fig. 16 for the model fit to TNFα and IL-1β pulsing)27. The total level of IKKK was assumed to be constant per cell2,14, while A20 limited IKK activity by inhibition of the ‘active’ IKKK.
The generic IKKK and A20 (but not IκBα, Supplementary Fig. 21) levels effectively divided the solution space into ‘responding’ and ‘non-responding’ regimes (Fig. 4b). Typical unstimulated cells are located in a ‘responding’ regime, characterized by a high ‘neutral’ IKKK (IKKKn, a form that could be activated by a cytokine signal) and low A20 protein levels (as depicted at t=0 for two representative cell trajectories, shown with broken lines). A pulse of TNFα leads to a rapid decrease of IKKKn (converted to its ‘active’ form), subsequent IKK activation, NF-κB translocation and A20 synthesis. This coincided with single-cell trajectories crossing over to the ‘non-responding’ regime, characterized by a low IKKKn and high A20, where they become refractory to a second pulse of TNFα (Fig. 4b). Over time equivalent to the refractory period, cells returned into the ‘responsive’ regime, as the IKKKn levels recovered from the ‘inactive’ state, while the A20 protein degraded.
The distribution of refractory period was recapitulated in the model by IKKK levels normally distributed in the population but fixed per cell (Fig. 4c,d, however, other distributions could also fit the data, Supplementary Fig. 22). For a fixed level of A20 protein, higher IKKK levels resulted in faster IKKKn recovery thus enabling subsequent activation (Supplementary Fig. 15). Therefore, when pulsed at longer intervals, responding cells exhibited higher amplitude of NF-κB p65 translocation as observed in the data (Fig. 4e). We found that changes in parameters of the IKKK distribution directly affected the mode and shape of the refractory period distribution. Increasing the mean IKKK level resulted in a shorter refractory period on average (Supplementary Fig. 23), while increasing the s.d. resulted in a more uniform distribution, with 40% of cells responding at the 50-min pulse interval (Fig. 4f).
We next used sensitivity analyses to understand which biochemical parameters may control the NF-κB responses. Latin hypercube sampling43 (and dynamic sensitivity analyses44, Supplementary Figs 26 and 27) showed that cell responsiveness to TNFα pulses (as defined by amplitude of NF-κB translocation) was controlled by parameters associated with the IKKK and A20 signalling (for example, A20 mRNA transcription, protein translation rates and half-lives, c1, c2, c3, c4; IKKK level, IKKKtott; A20-induced IKKK inhibition, kA20; half-maximal IKK activation, sIKKK, Fig. 4g shown in blue), but not the core NF-κB-IκBα or receptor signalling (see also Supplementary Fig. 28). In contrast, IκBα- and IKK-related parameters (transcription and translation rates, c1a, c2a; mRNA half-life, c3a; total NF-κB level, nfkbtot; and IKK inactivation rate, k1, Fig. 4g shown in brown) contributed (together with IKKK- and A20-related parameters) to the cumulative amount of nuclear NF-κB. The IκBα-related parameters also controlled the period of NF-κB oscillations, while A20- and IKKK-related parameters contributed to their amplitude (Supplementary Fig. 24). In agreement with the hypothesis that the depletion (or inhibition) of IKKK regulates the refractory period, all cells responded to a 10 ng ml−1 pulse 60 min after a low dose 0.1 ng ml−1 TNFα pulse. The mathematical model suggested that despite induced A20 protein levels in cells that responded to the first pulse, simulated IKKKn levels were sufficiently high to enable subsequent NF-κB activation, Supplementary Fig. 18a. Additionally, only 40% of cells responded to a 50 ng ml−1 pulse when the refractory period was set with a 10 ng ml−1 dose (in comparison to 30% of cells responding to two 10 ng ml−1 pulses), Supplementary Fig. 18b. This suggested that refractory state could not be efficiently reversed with a high-dose stimulation, supporting mathematical model predictions. Finally, when short interfering RNA (siRNA) was used to reduce A20 expression, a substantial increase in cells responding to TNFα pulses was observed (80% versus 25% scrambled siRNA control, Fig. 4h). In agreement, distributed individual parameters associated with A20/IKKK signalling (such as IKKKn recovery rate, m3, or total IKKK, IKKKt), but not IκBα (for example, protein half-life, c4a) were able to fit the measured refractory period distribution (Fig. 4i). This together showed the critical role of the TNFα signal transduction and A20 protein in the regulation of the refractory period.
Heterogeneous single-cell responses are imprinted
The mathematical model (Fig. 4) suggests a paradigm where heterogeneous responses to pulsatile TNFα stimulation are generated by ‘extrinsic’ differences between cells (modelled by distributed network parameters). In contrast, an alternative hypothesis suggests that this behaviour might arise from ‘intrinsic’ noise due to low numbers of interacting molecules, for example gene copies4 (Fig. 5a for schematic representation of noise in the NF-κB system). A model assuming a stochastic regulation of the A20 feedback, but a fixed level of IKKK, could explain digital NF-κB activation (Fig. 5b) and would be consistent with previous analyses14,20,45. To distinguish between these alternatives we stimulated cells with two pairs of TNFα pulses at 70 min intervals separated by a 4 h equilibration period (Supplementary Fig. 29). Model simulations predicted that under the ‘extrinsic’ noise hypothesis individual cells would show reproducible NF-κB activation when re-stimulated (Fig. 5c). However, under the ‘intrinsic’ noise hypothesis, the responses would be stochastic, that is, a cell that responded to a first set of pulses might not respond to the second pulsing protocol. We tested this prediction experimentally and found that the ‘extrinsic’ noise hypothesis was supported (P value<0.00001, for the Fisher’s exact test between obtained distributions); among 79 C9 cells (78% responding) only three cells behaved differently after the equilibration period (Fig. 5d,e, see also Supplementary Movie 1 and Fig. 5f for principal component analysis of IκBα-eGFP trajectories).
Altogether, these data suggested that NF-κB responses to pulsed stimulation (and thus refractory periods) were innate to individual cells and represented a non-genetic state. To test the stability of this state, the responses of daughter cells stimulated with two pulses of TNFα at a 70 min interval were compared (Fig. 5g). We found that ∼85% of daughter cells (out of 56 pairs) exhibited homogenous responses as depicted by single-cell IκBα-eGFP traces (Fig. 5h–j and Supplementary Fig. 30). In addition, no correlation between the behaviour of daughter cells and the time from cell division at stimulation was found (Supplementary Fig. 31). These data therefore suggested that the refractory states associated with NF-κB system were subject to change at a timescale longer than the cell cycle.
Refractory states integrate different cytokine inputs
Cells are often exposed to complex cytokine inputs. When stimulated with three pulses of TNFα (TTT) at 50 min intervals, most single cells were refractory to the second pulse, but responded to the third pulse (Fig. 6a). In contrast, stimulation with IL-1β in the second pulse instead of TNFα (TIT) induced a nuclear p65-mCherry translocation in 80% of cells (Supplementary Fig. 32). Although these responses were apparently noisy, responses to TIT stimulation were elevated in comparison to TTT (Fig. 6b), especially when stratified into time intervals between pulses (Supplementary Fig. 34a). This confirms that while refractory to TNFα, other cytokine inputs, for example IL-1β, can be processed by cells. To test if refractory states might enable functional discrimination between different cytokine inputs, we assayed expression of 70 TNFα and IL-1β-regulated genes46,47 with a Nanostring nCounter system48. Cells were stimulated with alternate 50 min TNFα and IL-1β pulses (TT, TIT and T_T), with additional single pulse and untreated controls (Fig. 6c). The comparison between the stimulation with two (T_T) and three (TTT) pulses of TNFα showed only eight differentially regulated genes (out of 62 expressed genes; including 18 constitutively expressed in addition to 5 housekeeping genes, see Supplementary Data 1 for details of the Nanostring analysis including raw and normalized data, and pair-wise comparisons between conditions). This confirmed that the cells were functionally refractory to the second pulse of TNFα at 50 min, in close agreement with imaging data (Fig. 6a,b). Differentially regulated genes included members of the NF-κB system, NF-κB1, NF-κB2 and IκBɛ; signalling molecules, CCL2 and CSF2; and A20-interacting TRAF1 and TNIP1 (ref. 27). The change in expression level compared with unstimulated control samples was relatively low and did not exceed 50%. In contrast, the comparison between TIT and TTT stimulation, showed 26 genes that were differentially regulated by IL-1β (Supplementary Fig. 33). This set included members of the NF-κB signalling system (Rel, RelB, NF-κB1, NF-κB2, IκBα, IκBɛ and A20) and a number of pro-inflammatory signalling molecules including cytokines and chemokines (IL8, CSF2, CCL2, CCL5, CXCL1, CXCL2, LIF, TNFα, TNFAIP6 and IL-1β). The first group showed only modest increase (<50%) in response to TIT compared with TTT stimulation (with A20 and IκBα highest with two-fold and 1.4-fold change, respectively). Notably, all of the signalling molecules exhibited changes over two-fold with CCL1, CXCL2, IL8 and TNFAIP6 higher than five-fold (Fig. 6f). In contrast, comparison between single TNFα and IL-1β pulses did not show a difference in the expression of those genes (see Supplementary Fig. 34e and Supplementary Data 1). Integrated single-cell nuclear NF-κB levels were very well correlated with the measured gene expression patterns across different pulsing conditions (Supplementary Fig. 34). In particular, we found the expression of all differentially regulated genes (including NF-κB-system and cytokine genes, Fig. 6f) showed a very high correlation (close to 1) irrespectively of their response amplitude. This analysis suggests that refractory states might enable cells to functionally discriminate between different closely timed cytokine inputs.
In this study, we investigated how rapidly changing cytokine inputs are encoded in the dynamics of the NF-κB signalling system. We employed time-lapse microscopy to quantitatively measure activation of the NF-κB p65 and its negative feedback IκBα in response to a pair of 5 min pulses of TNFα (and IL-1β), which were applied at different time intervals, ranging from 50 to 100 min. Single, or well-spaced pulses of TNFα (>100 min apart) gave a high probability of NF-κB activation. However, we found that at shorter pulse intervals (<100 min) responses were heterogeneous, with progressively more cells failing to respond to the second pulse (Fig. 2). This identified a heterogeneous refractory state in the NF-κB system. We asked how the variability between individual cells was generated (Fig. 5). We used a closely timed pair of TNFα pulses at 70 min interval as these discriminated cells into two pools that showed either a second response in IκBα degradation, or no response. We interpret this to mean that in one group cells have a refractory period of more than 70 min (non-responders), while in the other group cells have a refractory period of <70 min (responders). When the pair of pulses was repeated on the same cells several hours later, the presence or absence of a response was maintained. Daughter cell responses were also maintained in 85% of siblings. This implies that the refractory period was pseudo-stable, as characterized by a low switching probability over the timescale of the cell cycle. This mechanism enabled robust and reproducible digital responses in individual cells, with the timing of stimulation encoded in the fraction of responding cells. We hypothesize that this combination of digital- and analogue-encoding may induce a coordinated population-level response, enabling acute responses to temporal stimuli.
We predicted that the heterogeneous refractory period might be associated with cellular states encoded by the levels of protein in the TNFα transduction pathway, and mediated through a process downstream of TNFR and upstream of IKK. This mechanism has been represented in the mathematical model (in agreement with other NF-κB models49) by a simple nonlinear interaction between a generic IκB kinase kinase representing a complex TNFα transduction network (IKKK) and previously characterized NF-κB-dependent A20 negative feedback50 (Figs 3 and 4). In agreement with the model, siRNA knockdown of A20 protein increased the number of responding cells. A number of other proteins and interactions previously reported in the literature might also be involved, for example: TRAF adaptors51, RIP and TAK1 kinases8 as well as proteins involved in regulation of A20 enzymatic activity such as ABIN, RNF11, TAX1BP1 adaptors or the E3 ubiquitin-ligase Itch25,26,52. Response to TNFα (and IL-1β) might also involve other transcriptional networks, for example, mitogen activated protein (MAP) kinases or interferon regulatory factors (IRFs) (ref. 30), which also could constitute regulatory feedback and contribute to the observed behaviour. Experiments with alternate cytokine pulsing suggested that the TNFα stimulation had no effect on the IL-1β response; however, some dependency of the former on the IL-1β transduction pathway was found (Fig. 3 and Supplementary Fig. 9). To provide the simplest explanation for these data, we assumed that the signal cross-talk depended on respective levels of signal-specific IKKK and A20 inhibition; however, other mechanisms, for example, partial IKKK utilization by both pathways, cannot be excluded. It is possible that the cellular states within the IKK module might be controlled, for example, through a temporally stable epigenetic mechanism6 and this deserves further investigation. For example, the observed states are consistent with the reported long-term memory of protein levels53.
Refractory states in signalling have been previously associated with inhibition of cellular responses54,55. For example, endotoxin resistance was functionally linked with inhibition of over exuberant pro-inflammatory responses56. In contrast, we suggest that the refractory states in the NF-κB system (not only to TNFα and IL-1β, but also possibly to other input signals) might enable robust discrimination between different temporal cues. In accordance with this hypothesis, we showed that while in a refractory state to TNFα, cells responded to IL-1β stimulation (Fig. 6). Analysis of repeated pulses of TNFα, or alternating pulses of TNFα and IL-1β, suggested a strong correlation between single-cell NF-κB responses and the target gene expression (Supplementary Fig. 34). However, other downstream signalling systems including MAP kinases or IRFs30, stimulus-specific post-translational modification of the NF-κB (ref. 57) or timing of transcription and mRNA stability47 might also contribute to patterns observed for specific genes. At low doses, TNFα and IL-1β are the principal mediators of local inflammatory responses, however, at higher doses (>1 ng ml−1 in serum), TNFα (and IL-1β) becomes a danger signal that can lead to septic shock and ultimate death of the organism58. The level and timing of the signalling downstream of TNFα and IL-1β are therefore key parameters that dynamically control the activation and resolution of the immune response. We hypothesize that the refractory state might represent a mechanism in which responses to a single cytokine input are effectively dampened to avoid out-of-control activation, while multiple closely timed signals of more than one cytokine could synergistically activate an inflammatory response. This differential encoding of the temporal inputs is in part facilitated by the bow-tie topology of the NF-κB network, with the IKK effectively able to integrate multiple parallel upstream signals22,23. Other important mammalian signalling systems share a similar topology59, including epidermal growth factor60, G protein-coupled receptor61 and metabolic networks62. Hence, refractory states may be used in rapidly changing environments to fine tune patterns of differential gene expression. This may be a fundamental mechanism common to other cellular response systems.
Over the past decade, cellular heterogeneity has emerged as a common functional trait in many signalling systems with a number of studies suggesting a key involvement of stochastic gene regulation4. Intrinsic noise was thought to be particularly important for generation of heterogeneous NF-κB dynamics2,14,45. We previously argued that stochastic activation of IκB feedback regulated heterogeneity of NF-κB oscillation timing between individual cells, thus dampening potentially harmful fluctuations of NF-κB-dependent tissue-level cytokine secretion14,63. Intrinsic noise was also involved in the entrainment and amplification of NF-κB response to long-term periodic TNFα inputs20. In the present study, we identified an alternative source of cellular heterogeneity associated with pseudo-stable states embedded in the TNFα transduction network. We expect that many (if not all) network parameters may in fact exist in different states between individual cells (as suggested by measurements of protein1,53 and transcript levels64) and thus would contribute to the emergent behaviour. This behaviour is consistent with observed heritability of NF-κB responses13 and has been previously reported in MAP kinase and hypoxia-inducible factor signalling7. In agreement with our analyses, long-term repeat stimulation with TNFα pulses at 75 min intervals and lower resulted in a reduced system entrainment, possibly due to the refractory period20. This was exhibited by a fraction of cells responding with a frequency lower than that of the input (that is, effectively doubling the period of the NF-κB response). This also suggests that pulsing at intervals longer than the refractory period would result in better entrainment with most cells exhibiting the input frequency. Overall, we believe that the observed heterogeneous refractory states might be a part of the inherent design within the inflammatory response, to avoid out-of-control homogenous cell activation. This may be important for the balance between amplification and resolution of inflammation and its regulation could play a role in pathological inflammatory conditions.
Reagents and cell culture
SK-N-AS cells (ECACC 94092302 (ref. 10)) were cultured in minimum essential medium supplemented with 10% foetal calf serum (Gibco) and 1% NEAA (Sigma-Aldrich), and sub-cultured at densities between 80 and 90%. Stimulations were performed with 10 ng ml−1 (unless otherwise stated) of recombinant human TNFα or IL-1β (Calbiochem). Cells were stimulated with TNFα, washed three times with PBS and warmed culture medium was added. IL-1β neutralizing antibody (R&D Systems) was added (0.5 μg ml−1) to cells following IL-1β pulses and washed off with PBS before subsequent stimulation. Cells were routinely tested for mycoplasma contamination.
Derivation of IκBα-eGFP SK-N-AS cell line
To preserve physiological control of the IκBα reporter construct, a BAC system was utilized. Using seamless recombineering techniques65 we replaced the stop codon of the IκBα gene within the BAC CTD-3214F11 with the eGFP coding sequence (see Supplementary Note 1 for detailed description). Therefore, the resulting IκBα-eGFP fusion was expressed under the control of large regions of flanking sequences (63 kb upstream and 92 kb downstream), and maintained the UTR/Exon/Intron gene structure (Fig. 1a). The BAC was retrofitted with a Neomycin selection marker. SK-N-AS cells were transfected with IκBα-eGFP BAC, single-cell sorted and a clonal line (C9) carrying the construct was selected. Subsequently, the cell line was transformed with a p65-mCherry lentiviral plasmid37. In pulsing experiments, a cell was classified as a ‘responder’ if the gradient of the corresponding IκBα-eGFP trajectory at the time of stimulation was not positive; otherwise it was called a ‘non-responder’. This classification was independently verified by unsupervised clustering analysis (Supplementary Fig. 8). All experimental conditions and corresponding cell numbers are detailed in Supplementary Table 1.
Development of the p65-mCherry expression vector
Human NF-κB p65 (accession number NM_021975) was amplified by PCR using primers flanked by gateway recombination sequences (forward primer: 5′-ATGGACGATCTGTTTCCCCTCATCT-3′, reverse primer: 5′-CGAGCTGATCTGACTCAGCAGGGGCT-3′). A third generation ‘pLNT’ lentiviral transfer vector was used to enable Ubiquitin-ligase C promoter-mediated constitutive expression of N-terminally mCherry tagged amplified p65 gene37. Lentivirus production and transduction was performed as described fully in ref. 37.
Cells were lysed in hot lysis buffer (1% (w/v) SDS, 10% (v/v) glycerol, 10% (v/v) β-ME, 40 mM Tris pH 6.8, 0.01% (w/v) bromophenol blue). Samples and ladder (NEB #P7712S, USA) were resolved on polyacrylamide gels. Proteins were transferred to nitrocellulose membranes (Protran BA-83, GE Healthcare), incubated at room temperature in blocking buffer (5% (w/v) non-fat milk powder in TBS-T), washed in TBS-T and incubated overnight with primary human antibody (p65 Cell Signaling Technology #8242, phospho-S536 p65 CST #3039, IκBα CST #9242, Gapdh CST #2118 and α-Tubulin CST #2144) at 1:1,000 dilution in blocking buffer. Membranes were washed (TBS-T × 3), and incubated with 1:1,000 or 1:2,000 HRP-conjugated immunoglobulin-G (CST #7074) for 2 h. Membranes were washed (TBS-T × 3) then incubated with Luminata Crescendo Western HRP Substrate (EMD Millipore Corp.) and signal was detected by exposure to Carestream Kodak BioMax MR film (Sigma-Aldrich). Membranes were stripped using Restore Western Blot Stripping Buffer (Thermo Scientific) before re-probing with another primary antibody where necessary. See Supplementary Fig. 35 for uncropped scans.
C9 cells were plated into 35 mm culture dishes 24 h before transfection. Transfection mix was prepared using the SK-N-AS transfection kit (Altogen Biosystems, Las Vegas, USA) according to the manufacturer’s protocol, including complex condenser and transfection enhancer steps. Each dish was transfected with 100 nM of human A20 On-Target Plus siRNA or On-Target Plus non-targeting pool siRNA (both GE Dharmacon) 48 h before imaging. Cells were transferred to glass-bottomed dishes (Greiner Bio-One) 24 h before imaging.
Total RNA was extracted from wild-type SK-N-AS cells using the Roche High Pure RNA Isolation Kit. The nCounter Gene Expression assay (Nanostring Technologies, Seattle, USA) was performed according to the manufacturer’s instructions. Transcript counts were normalized to the relevant housekeeping genes using the NanoStringNorm package within Bioconductor66. The protocol was followed where the geometric mean was used to summarize the positive (CodeCount) and housekeeping controls, with a stringent background correction applied (mean±2 s.d.). Differential expression of genes was assessed on log2-normalized data with hierarchical regression models, using the eBayes function within limma67. A 0.01 false discovery rate cutoff was used to determine statistical significance. All genes were clustered with respect to median log2 fold changes across replicates (treated/untreated, see Supplementary Data 1 for detailed analyses) and visualized as a heat map.
Cells were plated onto 35 mm-glass-bottomed dishes (Greiner Bio-One) and incubated on the microscope stage at 37 °C in humidified 5% CO2. Several Zeiss confocal microscopes were used (LSM Pascal, Exciter, 510meta, 710 or 780) with fluar × 40 numerical aperture (NA) 1.3 or plan-apochromat × 63 NA 1.4 objectives and appropriate excitation and emission wavelengths for the two fluorophores. Image capture was performed using the Zeiss software, either ‘Aim version 4.2 utilizing the Autofocus macro68’ on the 5-series microscopes or ‘Zen 2010b SP1’ on the 7-series microscopes. Quantification of IκBα-eGFP fluorescent signal of whole cells was performed using region of interest analysis in ‘Zen 2010b SP1’. The data were exported as mean fluorescence intensity. For quantification of p65-mCherry fluorescence, Cell Tracker (version 0.6)69,70 was used to estimate mean nuclear and whole-cell fluorescence level, which was expressed as a nuclear to total ratio.
Evaluation of TNFα internalization
SK-N-AS cells were plated onto 4-compartment glass-bottomed imaging dishes (Greiner Bio-One) in culture medium and incubated at 37 °C in humidified 5% CO2 on the microscope stage. A Zeiss 780 confocal microscope with a plan-apochromat × 63 NA 1.4 oil objective was used with appropriate excitation and emission signal detection. Image capture was performed using Zeiss software ‘Zen 2010b SP1’ to take Z stacks using a stack separation of 0.8–1.2 μM. Maximum intensity projections were used for image analysis. Human recombinant TNFα biotin conjugate (1 μg ml−1, Fluorokine, R&D Systems, Wiesbaden) was diluted to 25 ng ml−1 in either 20 μl of avidin-FITC (10 μg ml−1) or 2 μl avidin-Texas-Red (2 mg ml−1, Life Technologies) and made up to 50 μl with minimum essential medium. Cells were washed with PBS before stimulation. Cells were pretreated with 80 μM Dynasore hydrate (Sigma) for 1 h where applicable. For acid wash treatment, cells were cooled to 4 °C and incubated with acid wash buffer (150 mM NaCl, 100 mM glycine pH 2.5) for 3 × 2 min. Cells were fixed with 3.7% formaldehyde in PBS for 15 min at room temperature, then washed with PBS. Fixed samples were imaged on a Zeiss 780 confocal microscope as above, with a plan-apochromat × 40NA 1.3 oil objective.
FACS analysis of TNFR1 level
SK-N-AS cells were scraped and fixed in 4% paraformaldehyde solution and then incubated on ice for 1 h with phycoerythrin conjugated TNFR1 antibody (Santa Cruz) according to the manufacturer’s protocol. Specificity of the antibody was confirmed using interferon γ stimulated cells, which exhibited higher TNFR1 expression comparing to untreated cells. Samples were analysed with a FACSVerse Flow cytometer. To eliminate cell debris or aggregated cells, events with low or high side and forward scatter were excluded. Subsequent data analysis was performed with FlowJo Software.
In this work, we considered the structure2 and parameters14,15,18 of previously developed models of the NF-κB system to recapitulate responses to pulsatile TNFα and IL-1β stimulation (see Supplementary Note 5 for model development and validation). Additional single-cell imaging data10,15 including responses to low TNFα doses2,18 were also recapitulated. The model also fitted population-level experimental data (nuclear NF-κB, total IκBα and A20 protein and mRNA, and IKK activity levels) in response to continuous and pulsatile stimulation of TNFα (of varying duration and concentration) in WT and A20 knockout cells21,28,32,40. Population-level heterogeneity was modelled by (1) normal distributions of network parameters, , with mean μ equal to nominal values of biochemical parameters simulated, and s.d. σ=0.3 μ, or (2) stochastic transcription of IκBα and A20 feedback genes in single cells14. Each model simulation was directly comparable to single-cell time-lapse microscopy data. A cell was classified as a ‘responder’ if the net peak amplitude of the nuclear NF-κB (calculated with respect to the nuclear NF-κB level at the time of stimulation) was >15% of the total level per cell; otherwise it was called a ‘non-responder’.
Simulations were performed in Matlab R2013a environment (see Supplementary Table 7 for all simulated conditions and Supplementary Software for computer codes). Power spectrum analysis was performed in Matlab using DFFT function.
Distribution of the refractory period
The refractory period is defined as the time period for which cells are unresponsive to TNFα (following initial stimulation). The distribution of refractory periods (Fig. 2g) was calculated based on fractions of responding cells measured at consecutive time points (based on data in Fig. 2f). For example, the fraction of responding cells decreased from 70 to 30% at 70 and 60 min pulse intervals, suggesting that ∼40% of cells had a refractory period between 60 and 70 min.
Global sensitivity analysis
Latin hypercube sampling was performed to calculate global parameter sensitivity with respect to model outputs43. For each model parameter, the range of ±50% around the nominal value was uniformly divided into equal k=1,000 slices. Values from each parameter range were then sampled randomly without replacement to ensure that the entire parameter range was explored. The model output was then simulated for each of the randomized k combinations of parameter values. A Spearman correlation coefficient was then calculated between each parameter and model output of interest (for example, fraction of responding cells, amplitude of NF-κB activation, period of oscillations and so on).
Statistical analyses were performed in GraphPad Prism V 6.0. For normally distributed samples (as assessed with D’Agostino and Pearson test) two-sided Student t-test was performed, otherwise nonparametric tests were applied (as indicated in the text). Goodness-of-fit comparison was performed using Fisher’s exact test. Sample sizes were selected based on the number of replicates (up to three) performed for each experiment. Statistical analyses are detailed in Supplementary Data 1.
The data that support the findings of this study are available from the corresponding authors on request.
How to cite this article: Adamson, A. et al. Signal transduction controls heterogeneous NF-κB dynamics and target gene expression through cytokine-specific refractory states. Nat. Commun. 7:12057 doi: 10.1038/ncomms12057 (2016).
We thank James Boyd and other members of Systems Microscopy Centre in Manchester for fruitful arguments. We also thank Mark Muldoon, Werner Müller and Kevin Couper for discussion, and the Genomic Technologies Core Facility in the Faculty of Life Sciences, Manchester for assistance. P.P. holds a BBSRC David Phillips Research Fellowship (BB/I017976/1). C.B. is supported by BBSRC PhD studentship. This work was also supported by MRC (MR/K015885/1), BBSRC (BBF0059381, BB/K003097/1) and Wellcome Trust (097820/Z/11/B). The work leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2012-2017) under grant agreement no. 305564.
Matlab scripts for computer simulations of NF-kB mathematical models.