To identify potential positive modulators, approximately 14,500 mouse and human full-length genes were screened for their ability to activate an NFκB luciferase reporter. Individual genes were transfected into HEK-293T cells together with the NFκB reporter and a constitutively expressed DsRed fluorescent protein using high-throughput transfection methods (Figure 1). In the initial screen, 183 genes activated the NFκB reporter greater than three-fold when normalized to the fluorescent transfection control and had luciferase values greater than 5 standard deviations (SD) from the experimental mean. These genes were rearrayed and screened in triplicate to confirm activation of the NFκB reporter. Eighty-four percent (154 genes) of the original positive modulators were confirmed in these studies on the basis of an average three-fold activation and an adjusted p value of less than 0.05. Of these 154 genes, 45 have been previously shown to modulate NFκB signaling (Additional data file 1).

Among the known NFκB modulators, genes within both the canonical and atypical NFκB signaling branches were identified, including known receptor ligands (TNF, IL1B, and LTA), TNF receptors (TNFRSF1A, TNFRSF10A, TNFRSF11A, TNFRSF12A, TNFRSF25, and TNFRSF10B), B-cell receptor (CD40), adaptor proteins (TIRAP, MYD88, and FADD), TNF receptor-associated factors (TRAF2 and TRAF5), members of the NFκB transcriptional complex (RELA, RELB, and NFKB1), and others. The known positive modulators were mapped to the canonical TNF, IL1, B-cell receptor, and T-cell receptor signaling pathways (Additional data files 5 and 6). The functional breakdown of the positive modulators is shown in Figure 2a. The breakdown contains a diverse array of categories that reflect the wide range of stimuli capable of activating NFκB and potential cross-talk with other signaling pathways.

In order to identify inhibitors of NFκB activity, full-length cDNAs were introduced into HEK-293T cells as described in the preceding section and the cells were incubated in the presence of either TNF or IL1 (Figure 1). A total of 235 genes were identified that reduced NFκB reporter activity greater than three-fold when normalized to the fluorescent transfection control and had luciferase values less than 5 SD from the experimental mean. These genes were rearrayed and screened in replicate to confirm the reduction of the NFκB reporter activity when exposed to either TNF or IL1. Forty-four percent (104 genes) of the original negative modulators were confirmed in these studies based on an average three-fold reduction and an adjusted p value of less than 0.05 (Additional data file 2). To exclude nonspecific negative modulators, an identical assay was performed using the glucocorticoid response element as the reporter and dexamethasone as the stimulus. Sixteen genes that showed inhibition of reporter activity in both screens were labeled as 'general inhibitors' and were removed from subsequent analysis. For the remaining 88 negative modulators, the functional breakdown is shown in Figure 2b. The breakdown was less diverse than with the positive modulators and had a relatively high percentage of genes related to DNA binding and transcriptional activity, suggesting that a number of the negative modulators act downstream in the pathway.

Among the 88 negative modulators, 16 have been previously shown to negatively regulate the NFκB pathway; they include genes such as NFKBIB (IκBβ) 2, NFKBIE (IκBε) 2, PIAS4 13, and RHOB 14. The known negative modulators were also mapped to the canonical TNF, IL1, B-cell receptor, and T-cell receptor signaling pathways (see Additional data files 5 and 6). Several genes identified as negative modulators in our screen have been previously shown to positively regulate NFκB and three genes (GPD1, TRAF2, and TSPAN13) were identified as both positive and negative modulators in our screens (see Additional data files 1 and 2). For most of the genes previously shown to be positive regulators, it is unclear why they are negative modulators in our system. One possibility is simple methodological differences. For example, some genes previously shown to positively modulate NFκB were identified in a screen performed 24 hours after transfection 10, whereas genes in our study were identified 48 hours after transfection and in the presence of TNF or IL1. A more likely possibility is that they can either negatively or positively modulate NFκB, depending on the cellular context. There are numerous examples of differential gene function depending on cellular context 1516171819, and this behavior has also been observed for genes within the NFκB network 2021. For example, TRAF2 has been previously shown to play a positive role in CD40 signaling in B cells and a negative role in TNF signaling in macrophages 21. The mechanism underlying these different behaviors has not been conclusively identified, but it has been suggested that it can both activate and degrade proteins by attaching distinct types of polyubiquitin chains 22. In addition, activation of the IKK complex can have pro-apoptotic or anti-apoptotic effects, depending on timing and mechanism of activation 20. Therefore, it is possible that a cell-based, genomic screen would identify differential behaviors for true modulators of NFκB under different treatment conditions (that is, untreated cells to identify positive modulators and cells treated with TNF or IL1 to identify negative modulators).

To obtain a general understanding of how the positive and negative NFκB modulators were co-expressed across various tissues, gene-expression data from 79 different tissues was obtained from Symatlas 23. The mean expression level and a 99% confidence interval for each modulator were then calculated across the 79 human tissues. Using the lower 99% confidence limit as a cutoff, modulators that fell below the cutoff were considered absent in that tissue, whereas those that were expressed above the cutoff were considered present. As expected, tissues involved in the immune response (for example, peripheral blood BDCA4⁺ dendritic cells) had a higher average number of positive and negative modulators present in the tissue when compared with non-immune tissue (152 versus 108; p < 0.0001) (Figure 3). These results suggest that regulation by these modulators is not distributed uniformly across tissues and that one mechanism for controlling the number of branches in the NFκB network is through differential expression. This mechanism has been described previously for G-protein-coupled receptors 24. Notably, the number of positive and negative modulators that were present in a given tissue was highly correlated (r = 0.944; p < 0.001) suggesting that the degree of positive regulation is counterbalanced by negative regulators at the tissue level.

To map the positive modulators within the NFκB structural network, the effects of each modulator was examined in the presence of a series of dominant-negative mutants. These mutants were chosen due to their roles and relative locations within the NFκB network. Individual positive modulators were screened in triplicate with IKBKB, IKBKG, TRAF2, and MAP3K7 (TAK1) dominant-negative mutants (Additional data file 3). If the activation of the NFκB reporter was blocked by the dominant-negative mutant (≥ 70% average reduction), the positive modulator was considered upstream of the dominant-negative mutant. The results from this analysis organized the positive modulators into four distinct groups - upstream of TRAF2, upstream of MAP3K7, upstream of the IKK complex, and no inhibition by any dominant-negative mutant (Figure 4).

The group of positive modulators identified as upstream of TRAF2 included a variety of genes with a known dependence on TRAF2 for signal propagation. They included genes for several TNF receptors (for example, TNFRSF1A, TNFRSF1B, and TNFRSF12A), and for LTA, FADD, BIRC2, RIPK3, and CD40. Of these genes, the FADD protein is known be a part of a complex that includes TRAF2, RIPK1, and TRADD 25. For BIRC2, physical interaction with TRAF2 alters nuclear translocation of the protein 26 and a dominant-negative TRAF2 has been demonstrated to inhibit RIPK3 signaling 27. Other genes in this group have been associated with NFκB signaling, but have not previously been identified as TRAF2-dependent. These included PYCARD, TMEM9B, and TIRAP 102829. For TIRAP, previous reports have shown that it is involved in Toll-like receptor signaling 29, raising questions about its dependence on TRAF2. One explanation is that the overexpression of the TRAF2 dominant-negative mutant results in a nonspecific inhibition of TRAF6. However, neither MYD88 nor any of the Toll-like receptors that were also identified as positive modulators were inhibited by the overexpression of the TRAF2 mutant. Another explanation is that the physical association identified between TBK1 and TIRAP 30 is part of a larger complex that also includes TRAF2. TRAF2 has also been shown to associate with TBK1 31 and formation of the larger complex may be required for TIRAP signaling.

In the group of positive modulators identified as acting upstream of MAP3K7, several genes have a known dependence on MAP3K7 for signaling, including MAP3K7IP2 and CARD4 (Figure 4). Of the proteins encoded by these genes, MAP3K7IP2 activates MAP3K7 by forming a complex with TRAF6, MAP3K7 and TAB1 32. CARD4 has been shown to induce transcription of MAP3K7 and of the genes for other components of the IL1/TLR branch leading to activation of NFκB 33. Other genes in this group have been associated with NFκB signaling, but not previously identified as MAP3K7-dependent. These included PAK4, NUTF2, RRAS, EEF1A1, TRAF5, TBL1X, PYCARD, ATP2A2, TMEM9B, TRAF2, and TNFRSF12A. For TRAF2 and TRAF5, previous reports have linked their involvement with the TNF signaling branch 3435. However, other reports have shown that TRAF2 can activate MAP3K7 under certain conditions 1236 and that MAP3K7 still contributes to IκB phosphorylation under TNF stimulation 37, suggesting that signaling through each branch is not exclusive. This cross-talk may contribute to the number of genes that were found to be upstream of both TRAF2 and MAP3K7 (Figure 4, green text).

For the positive modulators inhibited by either the TRAF2 or MAP3K7 dominant-negative mutant, all genes were also inhibited by at least one of the IKK mutants. For the genes identified as upstream of TRAF2, MAP3K7, and the IKK complex, the majority was inhibited by both the IKBKB and IKBKG dominant-negative mutants, and all genes required IKBKB for activation. This suggests that most of the positive modulators were from what has historically been referred to as the canonical branch. Interestingly, five positive modulators were not inhibited by the IKBKG dominant-negative mutant: they comprise three of the genes upstream of MAP3K7 and two of the genes upstream of the IKK complex (Figure 4, red type). The IKBKG dominant-negative mutant used in the analysis is deficient for binding to CHUK and IKBKB 38, suggesting that IKBKG binding within the IKK complex is not required for NFκB activation by these genes.

The group of positive modulators identified as upstream of IKBKB and IKBKG was substantially larger than those identified as upstream of TRAF2 and MAP3K7. This primarily reflects its importance as point of convergence in the pathway, as it integrates signals from the various upstream branches 39. The size of the group may also have been inflated as a result of the relatively strict analysis criteria that may not have correctly identified some genes as TRAF2- or MAP3K7-dependent. These genes include those for the Toll-like receptor TLR9, three TNF receptors (TNFRSF10B, TNFRSF11A, and TNFRSF25), and genes such as TAB3 and IL1B. The signals from these genes were inhibited by the TRAF2 or MAP3K7 dominant-negative mutants, but not to the extent required by the cutoff. Apart from these genes, other positive modulators that are generally accepted to be upstream of TRAF2 or MAP3K7 may have slipped into the IKK group as a result of the branched structure of the network, which would allow the signal to be propagated around the dominant-negative mutant. For example, MYD88 plays a central role in IL1/TLR signaling and has been shown to activate NFκB through both MAP3K7-dependent and -independent mechanisms 9.

In the final group of positive modulators, six genes were not found to be upstream of any of the dominant-negative mutants used in these studies. This group included genes encoding the components of the NFκB transcription complex (RELA, RELB, and NFKB1), an NFκB-inducing kinase (MAP3K14), a member of the claudin superfamily (CLDN12), and a relatively uncharacterized gene (C20Orf102). The contextual placement of MAP3K14, CLDN12, and C20Orf102 was difficult to localize as the lack of inhibition by the dominant-negative mutants could mean that the modulators act through a separate, independent branch that converges at the transcriptional complex. For MAP3K14, previous reports are contradictory, with some investigators demonstrating MAP3K14 activation of the IKK complex 404142 whereas others have reported that MAP3K14 can activate NFκB independently of IKBKG through a p38 MAPK-dependent RELA phosphorylation pathway 43.

To assess the relative location of the negative modulators, each modulator was screened together with a constitutively active IKBKB mutant (Additional data file 4). If activation of the NFκB reporter was blocked by the negative modulator (≥ 70% reduction), the negative modulator was considered downstream of the constitutively active mutant. If activation of the NFκB reporter was not blocked, the negative modulator was considered unclassified as to its inhibitor influence. For the group of negative modulators downstream of IKBKB, several genes were identified that were previously demonstrated to inhibit NFκB signaling (Figure 5). These genes include NFKBIB, NFKBIE, PIAS3, PRKACG, PTGER2, and PTGER4. The downstream effects of the two IκB genes (NFKBIB, NFKBIE) have been well established 2, whereas PIAS3 functions downstream by binding to RELA and suppressing NFκB-dependent transcription 44. The gamma catalytic subunit of protein kinase A (PRKACG) has been shown to repress NFκB signaling by phosphorylating the p50 homodimer, which acts a transcriptional repressor by maintaining DNA binding in unstimulated cells 45. Finally, for the PTGER2 and PTGER4 genes, a clear mechanism has not been elucidated, but a recent report by Akaogi and colleagues 46 showed that the NFκB-dependent transcription of TNF was inhibited by 70% in response to activation of prostaglandin E2 and E4 receptors.

To combine the experimental data into a structural network representing NFκB signaling, an adjacency matrix was constructed based on the orientation flags (that is, upstream, downstream, or unclassified) from the dominant-negative and constitutively active screens used in the contextual organization analysis. The adjacency matrix containing the experimental data was then merged with a separate adjacency matrix containing members of the currently accepted NFκB signaling network. By using the known NFκB network as a scaffold, we were able to include additional signaling intermediates and create a more hierarchically consistent network. Based on the combined adjacency matrix, a network map was constructed as a rooted tree with the NFκB complex serving as the obligate root node. The longest path was calculated from each terminal node to the root node using a previously defined algorithm 47. The inferred NFκB structural network is provided as an additional data file (Additional data file 7).

Constructing the NFκB signaling map provides insight into the potential logic embedded within the network and allows additional hypotheses to be generated regarding influence and potential cross-talk from the positive and negative modulators identified in this investigation. The potential for positive cross-talk in our investigation was concentrated upstream of MAP3K7 and directly upstream of IKBKB and IKBKG. In contrast, a significant number of negative modulators entered the network between the IKK complex and the activated transcription factor, providing a large number of modes for negative regulation.

A set of bacterial stocks from the Mammalian Gene Collection (MGC) containing 30,720 cDNA clones in 80 384-well plates were ordered from the American Type Culture Collection (Manassas, VA). Putative full-length mouse and human cDNA clones in mammalian expression vectors were identified by cross-referencing the library with the 'Static Clone List' generated by MGC 53 and the 'IRAK/IRAL Cumulative File' generated by the IMAGE consortium 54 from which the MGC set is derived. Full-length status was determined by the MGC and IMAGE consortium through end sequencing, and full-length clones were given a GenBank identifier. The bacterial stocks for the cDNA clones matching the full-length criteria and cloned in mammalian expression vectors were rearrayed using a Genetix Q-bot (Hampshire, UK) into 96-well plates. A total of 8,700 full-length mouse and 5,800 full-length human genes were rearrayed. The rearrayed bacterial clones were amplified using a Hi-Gro incubator (Genomic Solutions, Ann Arbor, MI) in deep-well, 96-well plates containing 1.5 ml of Terrific Broth plus antibiotic (20 h; 500 rpm). Plasmid DNA was isolated from the bacterial clones and purified using a Roboprep 2500 (MWG, High Point, NC) and Macherey-Nagel NucleoSpin Robot-96 kits (Macherey-Nagel, Easton, PA). After purification, DNA concentrations were quantified on the basis of absorbance at 260 nm and normalized to a standard concentration of 2 ng/μl using a Theonyx liquid-handling robot (MWG). Individual clones were sequenced at random to verify gene inserts using universal M13 sequencing primers.

Human embryonic kidney (HEK)-293T cells were obtained from the American Type Culture Collection (ATCC No. CRL-11268, Rockville, MD) and were grown in Dulbecco's Modified Eagle Medium supplemented with 10% fetal bovine serum and 100 U/ml penicillin-streptomycin (Invitrogen, Carlsbad, CA). Cells were maintained at 37°C with 5% CO₂ in a humidified chamber. The plasmids used in this study were from the following sources: pNFκB-Luc (5X TGGGGACTTTCCGC) and pGRE-Luc (Stratagene, La Jolla, CA); pCMV-Sport6-Empty and pUC18-Empty (Invitrogen); pCS2-DsRed (kindly provided by R. Davis); pCMX-IKBKB SS-AA (Ser177A & Ser181A) and pCMX-IKBKB SS-EE (Ser177E & Ser181E) (kindly provided by S. Chanda); pCMV-MAP3K7 (K63W) (kindly provided by J. Ninomiya-Tsuji); pCMV-ΔTRAF2 (1-211) (kindly provided by G. Natoli); and pcDNA3-HA-IKBKG (134-419) (kindly provided by G. Nuñez).

High-throughput transfections for the gain-of-function screens were performed in 384-well plates with each well containing 20 μl OptiMEM with GlutaMax (no serum or antibiotic), 20 ng pNFκB-Luc, 15 ng pCS2-DsRed, and 150 nl Fugene6 (Roche Diagnostics, Indianapolis, IN). The transfection mix was dispensed using a Multidrop-384 (Titertek, Huntsville, AL). Purified plasmid DNA (15 ng) for each gene was added to an individual well of a 384-well plate using a Biomek FX robot (Beckman Coulter, Fullerton, CA) and the DNA:Fugene mixture was incubated for 30 min at room temperature. Each 384-well plate contained eight empty wells for the manual addition of positive and negative control plasmids. A pCMV-Sport6-Empty plasmid lacking a cDNA insert was included in each plate as the negative control. Following incubation, HEK-293T cells were added to each well in 20 μl serum-containing medium (20% fetal bovine serum, 200 U/ml penicillin-streptomycin) at a concentration of 400,000 cells/ml. The plates were then incubated for 48 h at 37°C with 5% CO₂. At the end of the incubation period, 40 μl BriteLite (Perkin Elmer, Wellesley, MA) was added to each well using the Multidrop-384 (Titertek). Within 5 min after substrate addition, luminescence and fluorescence were read using an Analyst HT plate reader (Molecular Devices, Sunnyvale, CA). For the loss-of-function screens, transfections were performed identically to the gain-of-function screens except that either purified TNF or IL1B (Sigma Aldrich, St. Louis, MO; Catalog Nos. T0157 and I9401) was added in 10 μl medium, 4 h post-transfection. The concentrations of TNF and IL1B were 20 ng/ml (final) and 0.5 ng/ml (final), respectively.

To contextually organize the positive NFκB modulators, each combinatorial screen was run using paired 384-well plates. In one of the paired plates, the positive modulators were co-transfected with either 15 or 20 ng of the dominant-negative expression plasmid. Fifteen ng of the dominant-negative expression plasmid was used for pCMV-ΔTRAF2 and pCMV-MAP3K7 (K63W) while 20 ng was used for pCMX-IKBKB SS-AA and pcDNA3-HA-IKBKG (134-419). In the second paired plate, the positive modulators were co-transfected with either 15 ng or 20 ng of pUC18-Empty to match the amount of the dominant-negative expression plasmid. The pUC18-Empty was used as a control for the dominant-negative mutant to keep the total amount of DNA constant. Both of the paired 384-well plates contained eight empty wells for manual addition of positive and negative control plasmids. A pCMV-Sport6-Empty plasmid was included on each paired plate as the negative control for the positive modulator. Assay conditions were identical to those described for the gain-of-function screen except that the amount of Fugene6 (Roche Diagnostics) was increased to either 195 nl (TRAF2, MAP3K7) or 210 nl (IKBKB, IKBKG). To contextually organize the negative NFκB modulators, paired 384-well plates were also used. In one of the paired plates, the negative modulator was co-transfected with 10 ng of the pCMX-IKBKB SS-EE plasmid expressing the IKBKB constitutively active mutant. In the second paired plate, the negative modulators were co-transfected with 10 ng pUC18-Empty. The pUC18-Empty was used as a control for the constitutively active mutant to keep the total amount of DNA constant. Both of the paired 384-well plates contained eight empty wells for manual addition of positive and negative control plasmids. A pCMV-Sport6-Empty plasmid was included on each paired plate as the negative control for the negative modulator. Assay conditions were identical to those described for the gain-of-function screen except that the amount of Fugene6 (Roche Diagnostics) was increased to 180 nl. Each paired combinatorial screen was performed at least three times.

Luciferase measurements for each gene in the gain-of-function and loss-of-function screens were adjusted for background luminescence using control wells containing only media and BriteLite (PerkinElmer). The background-subtracted values were then normalized for transfection efficiency using the fluorescent DsRed measurements. Normalized fold-change ratios were calculated by dividing the normalized luciferase values for each gene by the normalized luciferase values for the pCMV-Sport6-Empty negative control on each plate. In the case where multiple control wells on each plate were transfected with the pCMV-Sport6-Empty plasmid, the average normalized luminescence value was used to calculate the normalized fold-change ratio. In the initial gain-of-function screen, significant positive modulators were required to have a normalized fold-change ratio greater than three-fold and luminescence values greater than 5 standard deviations (SD) from the experimental mean. In the initial loss-of-function screen, significant negative modulators were required to have a normalized fold-change ratio less than 0.34 (that is, > 3-fold reduced) and luminescence values less than 5 SD from the experimental mean. Confirmation assays on the significant positive and negative modulators were performed at least three times and analyzed using a one-sample t-test for its statistical significance from a fold change of 1. Probability values were adjusted for multiple comparisons using a false-discovery rate correction 55. Confirmed hits were required to have an average normalized fold-change ratio greater than three-fold (positive modulators) or less than 0.34 (negative modulators) and an adjusted p value < 0.05.

Functional analysis of the positive and negative modulators was performed using NIH David 2.1 56 and based on molecular function ontologies at level 2. Mapping of positive and negative modulators to the known signaling pathways was performed using GenMAPP 2.1 57. The canonical TNF, IL1, B-cell receptor, and T-cell receptor signaling pathways were obtained from the Human Protein Reference Database 58.

Gene-expression data for 79 human tissues using the human Affymetrix U133A array were obtained from Symatlas 23. The raw data were preprocessed using GCRMA with a log₂ transformation. The mean expression level and a 99% confidence interval for each gene were then calculated across all 79 human tissues. Using the lower 99% confidence limit as a cutoff, modulators that fell below the cutoff were considered absent in that tissue, whereas those that were expressed above the cutoff were considered present. Statistical analysis of the average number of modulators present in immune tissue versus non-immune tissue was performed using a two-sample t-test.

Combinatorial screens for the positive modulators were analyzed by first subtracting background luminescence values and normalizing for transfection efficiency using the fluorescent DsRed measurements. Normalized fold-change ratios were calculated for each positive modulator on each of the paired 384-well plates by dividing the normalized luciferase values by the normalized luciferase values for the pCMV-Sport6-Empty negative control. A percent reduction was calculated based on the normalized fold-change ratio of the positive modulator that was co-transfected with the dominant-negative mutant and the normalized ratio of the positive modulator that was co-transfected with pUC18-Empty. The percentage reduction by the dominant-negative mutant was averaged for at least three experimental replicates and then subjected to a lower-bound cutoff that flagged each positive modulator as either upstream or unclassified relative to the dominant-negative mutant. A lower-bound cutoff of ≥ 70% was used to flag positive modulators as upstream. Percent reductions less than the cutoff were flagged as unclassified. Analysis of the combinatorial screens for the negative modulators was performed similarly, except that modulators showing ≥ 70% reduction were flagged as downstream relative to the constitutively active mutant and all others were flagged as unclassified.

An adjacency matrix was constructed based on the orientation flags (that is, upstream, downstream, or unclassified) from the dominant-negative and constitutively active screens used in the contextual organization analysis. The adjacency matrix containing the experimental data was then merged with a separate adjacency matrix containing the known NFκB signaling network. On the basis of the combined adjacency matrix a signaling network map was constructed as a rooted tree with the NFκB complex serving as the obligate root node. The longest path was calculated from each terminal node to the root node using a previously defined algorithm 47. For example, if A is upstream of B and C, and B is upstream of C but downstream of A, then the graph would be drawn as A → B → C and not as A → B, A → C and B → C. The signaling networks were visualized using Cytoscape software 59.