Actin networks modulate heterogenous NF-κB dynamics in response to TNFα

The canonical NF-κB transcription factor RELA is a master regulator of immune and stress responses and is upregulated in PDAC tumours. Here, we characterised previously unknown endogenous RELA-GFP dynamics in PDAC cell lines by live single cell imaging, which revealed rapid, sustained and non-oscillatory nuclear RELA following TNFα stimulation. Using Bayesian analysis of single cell datasets with variation in nuclear RELA, we computationally predicted that RELA heterogeneity in PDAC cell lines is dependent on F-actin dynamics. By RNA-seq, we identified the actin regulators NUAK2 and ARHGAP31 as transcriptionally regulated by RELA. In turn, NUAK2 or ARHGAP31 siRNA depletion downregulates TNFα-stimulated RELA nuclear localisation in PDAC cells, establishing a novel negative feedback loop regulating RELA activation by TNFα. We identify an additional actin-independent feedback loop involving RELB, which suppresses TNFα-mediated RELA nuclear localisation following RELA mediated upregulation of RELB. Taken together, we provide computational and experimental support for interdependence between the F-actin network and RELA translocation dynamics in PDAC.


Introduction 24
The NF-κB transcription factor RELA is an essential mediator of the inflammatory and 25 immune responses in all mammals (Hayden et al., 2006) and is central to the canonical NF-26 κB signalling pathway (Ghosh et al., 1998). As a transcription factor, RELA activation is 27 controlled in large part through its localisation. Inactive RELA is sequestered in the cytoplasm 28 by IκB proteins and IκB degradation by upstream cues, such as the potent inflammatory hyperphysiological TNFα doses (e.g. 10 ng/ml) and exogenous RELA reporters. 116 To identify whether cell cycle progression contributes to heterogeneity in RELA responses 117 to TNFα in PDAC cells, we categorised each tracked cell by cell cycle stage at the time of 118 TNFα addition, using changes in the appearance and intensity of endogenous PCNA-Scarlet 119 to mark cell cycle transitions (Figure 1 -Supplement 1A). For each cell, we calculated the 120 amplitude and timing of peak nuclear RELA, in addition to the mean nuclear RELA intensity 121 across all timepoints. Largely, there were no differences between cells by cell cycle stage in    Distributions of nuclear RELA in the five PDAC cell lines by immunofluorescence revealed 157 highly heterogeneous RELA responses within and across the PDAC cell lines (Figure 2A). 158 To identify features that predict RELA localisation differences, we collated and incorporated 159 normalised single cell measurements across all PDAC cell lines and TNFα treatments into 160 Bayesian networks, harnessing the observed variation in RELA ( Figure 2B). Bayesian 161 network models apply statistical inference to heterogeneous experimental data to predict the 162 conditional dependence of components on each other. Bayesian network models appear as 163 influence diagrams consisting of nodes, each representing a measured feature, and arcs 164 that depict predicted dependencies between the nodes. These dependencies represent 165 linear and non-linear relationships, direct and indirect interactions, and illustrate multiple 166 interacting nodes simultaneously (Sachs et al., 2005). We employed a hybrid class of Bayesian algorithm ('rsmax2') that generates models with unidirectional arcs using a 168 combination of constraint-based and score-based approaches (Scutari et al., 2018). 169 In order to provide the greatest heterogeneity in nuclear RELA and cytoskeletal/cell shape 170 features for Bayesian model generation, we collated data from all five PDAC lines and TNFα 171 doses (0, 0.01, 0.1 and 10 ng/ml), shown in Figure 2C. This model indicated that nuclear 172 RELA measurements are correlated to and predicted to be dependent on cytoplasmic actin 173 and tubulin intensity, suggesting that cytoskeletal dynamics influence heterogeneity in 174 nuclear RELA translocation with TNFα. Nuclear RELA is also predicted to be dependent on 175 cell area, although with a lower strength of the probabilistic relationship compared to 176 actin/tubulin, while nucleus roundness is predicted to be dependent on nuclear RELA.

177
Interestingly, both actin texture and the ratio of membrane to cytoplasm actin are also 178 predicted to be dependent on RELA, suggesting that RELA and actin dynamics are 179 interdependent. Altogether, our data suggest that the influence of cell shape on RELA 180 translocation we have previously described in breast cancer cells (Sero et al., 2015;Sailem 181 and Bakal, 2017) is likely mediated through cytoskeletal changes.

182
To understand how inter-line differences in cytoskeletal organisation may influence RELA 183 translocation dynamics, we additionally generated Bayesian network models by subsetting 184 data by cell line and TNFα concentration. We summarised dependencies involving nuclear 185 RELA in Figure 2D. In contrast to our prior findings using Bayesian modelling of RELA and  We generated a single cell dataset for use in Bayesian modelling by treating MIA PaCa2 and 215 PANC1 cells with selected drug doses for 2 hr then simultaneously with TNFα (0, 0.01, 0.1 216 and 10 ng/ml) for 1 hr and input these data into the same Bayesian algorithm as above 217 (rsmax2) ( Figure 2F). Interestingly, the Bayesian network following perturbations revealed 218 that 5/6 of the variables observed to correlate with RELA (whether they influenced RELA or 219 were influenced by RELA) in untreated cells were independent of RELA in the drug-treated 220 network. Only 'Cytoplasmic Actin Mean' remained as a influencing variable following drug-221 treatment, suggesting that actin abundance is a key regulator of RELA nuclear localisation.

222
While other variables (i.e. Tubulin and cell shape) can indirectly influence RELA, they do so 223 via regulating actin network organisation.

224
TNFα dose and duration determine profiles of RELA-dependent gene expression in 225 PDAC cells 226 As RELA is known to be involved in feedback loops with transcriptional targets, with the best 227 studied example being IκBα, we sought to identify what the transcriptional targets of RELA 228 are in PDAC and which influence RELA in actin dependent and independent ways.

229
To this end, we carried out RNA-seq analysis of PDAC cells at an early (1 hr) and late (5 hr)     We also considered how expression of NF-κB and IκB genes are affected by the treatment  To identify whether siRNAs affected the early/peak RELA and sustained RELA response to 321 TNFα, we calculated the fold change of mean nuclear RELA with each siRNA to NT siRNA 322 at 1 hr or 12 hr TNFα stimulation, and compared fold changes using t-test with multiple 323 comparison correction ( Figure 4B and 4C). Considering TNFα/NF-κB pathway components, 324 we found that several siRNAs enhanced nuclear RELA in PANC1 but not in MIA PaCa2 in 325 basal conditions, including NFKBIA, NFKBIB, RELB and ARHGAP31, suggesting that RELA 326 activation is actively and constitutively suppressed by several mechanisms in PANC1 cells.

327
Focusing on the early RELA response (1 hr) to TNFα, we found that NFKBIE and NFKB2 328 depletion reduced nuclear RELA in PANC1 in control conditions, suggesting that these 329 TNFα/NF-κB signalling components enhance nuclear RELA in basal conditions. Meanwhile,     Of note, we present the first evidence that ARHGAP31 is a transcriptional target of TNFα or 421 RELA. Moreover, ARHGAP31 is not previously associated with pancreatic cancer, but has 422 well studied roles in actin modulation, since ARHGAP31 is the human orthologue for the We thank Theodoros I Roumeliotis and Jyoti Choudhary for performing protein mass 453 spectrometry and analysis. We thank Andrea Brundin for assistance with single cell tracking. 454 We gratefully acknowledge funding for this work by Cancer Research UK, awarded to F.B.

457
The authors declare no competing interests.