Principles of dimer-specific gene regulation revealed by a comprehensive characterization of NF-κB family DNA binding

Designing an NF-κB-specific protein-binding microarray

To examine the DNA-binding specificities of NF-κB dimers in a systematic and unbiased manner, we utilized PBM technology. PBMs are double-stranded DNA microarrays that allow the in vitro characterization of protein binding to tens of thousands of unique DNA sequences in a single experiment^24,26,27. The universal PBM (uPBM) developed previously²³ allows a comprehensive, unbiased assessment of protein-DNA binding to all ungapped and gapped 8-bp sequences. We carried out uPBM experiments with six human and mouse NF-κB dimers (c-Rel:c-Rel, RelA:RelA, p52:p52, p50:p50, c-Rel:p50, RelB:p52) to perform an initial, comprehensive survey of potential κB site sequences. DNA binding site motifs derived from the uPBM experiments were in excellent agreement with published SELEX data on cRel:cRel, RelA:RelA and p50:p50 homodimers¹³ (Supplementary Fig. 1), demonstrating highly specific binding in our assay.

The uPBM platform assesses binding to 8-bp sequences; however, the canonical NF-κB binding site (κB site) is 10-bp long^6,7. Therefore, we created a custom NF-κB PBM containing 10-bp sequences prioritized according to the uPBM 8-bp data (see Supplementary Methods). We compiled the 1,000 top-scoring 10-bp sequences determined for the six NF-κB dimers into a list of 3,285 non-redundant sequences that represent the top-scoring set of potential κB site sequences. These 10-bp κB sites were incorporated into a custom NF-κB PBM with each site situated within constant flanking sequence (Fig. 1a, see Methods).

We initially examined the binding of RelA:p50 to our custom NF-κB PBM. To assess significance of the results, the natural log of the median PBM signal intensity of each 10-bp site was transformed into a z-score using the scores from 1,200 randomly chosen 10-bp sites as a background distribution (Fig. 1b; see Methods). Many potential κB sites, including a set of validated κB sites, scored significantly higher than the background distribution (z-score > 4), demonstrating that the custom PBMs reveal the specific DNA binding sites of NF-κB dimers. Thus, our custom NF-κB PBM provides a unique platform to assess the DNA-binding specificities of different NF-κB dimers for a large set of potential κB site sequences.

Three distinct DNA binding classes

To examine the DNA-binding preferences of different NF-κB dimers, we performed custom NF-κB PBM experiments for ten dimers from mouse or human. We compared the DNA-binding specificities of different dimers by correlating their κB site z-scores. Hierarchical clustering revealed that the NF-κB dimers separated into three distinct classes: p50 or p52 homodimers; heterodimers; and c-Rel or RelA homodimers (Fig. 1c). This subdivision is similar to the basic division of the NF-κB family members into two subclasses based on protein sequence of the Rel-homology domains: p50 and p52 (subclass 1); c-Rel, RelB and RelA (subclass 2). The common binding specificity observed for the heterodimers suggests a canonical DNA-binding contribution from members of each NF-κB subclass.

To highlight the differences between these three NF-κB classes, we constructed a representative DNA binding site motif for each class (Fig. 1d; Supplementary Fig. 2 for all individual motifs). We identified different DNA binding site motif lengths for each class: 9-bp for c-Rel and RelA homodimers; 10-bp for heterodimers; 11- or 12-bp for p50 and p52 homodimers. We note that while our custom NF-κB PBM was designed to assay binding to a large collection of 10-bp sequences, our de novo motif finding approach arrived at a longer motif for p50 and p52 homodimers; length preferences are examined more directly below. The 10-bp motif for the heterodimer class is in excellent agreement with the known NF-κB consensus sequence 5'-GGGRNWYYCC-3', demonstrating that we correctly identified the known high-affinity binding sites. The variant 9-bp and 11-bp motifs for the c-Rel,RelA and p50,p52 homodimer classes, respectively, also agree with the reported DNA-binding preferences of these different homodimers⁶. A comprehensive overview of the binding landscape for all dimers to the 10-bp κB sites is provided (Supplementary Fig. 3a), along with example genomic regions in which our PBM data are used to annotate dimer preferences for putative κB sites (Supplementary Fig. 3b,c).

In our pair-wise comparisons, all heterodimers exhibit a common DNA-binding specificity. Of particular interest was our observation that RelB-containing heterodimers exhibit DNA-binding specificity similar to that of c-Rel- and RelA-containing heterodimers. The biological significance of this finding is discussed below (see Discussion).

Dimer preferences for traditional and non-traditional κB sites

Dimer-specific DNA-binding preferences provide a mechanism for NF-κB dimers to target distinct binding sites, and thus to regulate distinct target genes. We observed the most distinct DNA-binding preferences (the lowest z-score correlation) between members of the two homodimer classes (Fig. 1c). To investigate these differences further, we compared the binding specificities of the most dissimilar dimers: p50:p50 and c-Rel:c-Rel (z-score correlation = −0.13) (Fig. 2a). We observed many off-diagonal features that correspond to sites bound preferentially by one of the dimers (‘dimer-preferred’ κB sites).

To identify sequence features that could explain the relative dimer preferences, we examined the c-Rel:c-Rel-preferred κB sites (κB sites with p50:p50 z-score < 2 and c-Rel:c-Rel z-score > 4). We found that a majority had a strong 5'-HGGAA-3' half-site (Fig. 2a, red dots). Many of these sequences conform to the canonical c-Rel,RelA-preferred 9-bp binding site with two 5'-HGGAA-3' half-sites separated by a 1-bp spacer^12,28. However, we also found many non-traditional κB site sequences with only one 5'-HGGAA-3' half-site. We use the term “non-traditional” in lieu of “non-canonical” to avoid potential confusion with variant κB sites (referred to as non-canonical) reported to be downstream of the non-canonical NF-κB signaling pathway¹⁶. Non-traditional sites are defined as sites that score poorly by the widely used NF-κB PWMs^13,29 (see Supplementary Methods); examples include CD28RE from the Il2 and Csf2 gene promoters (5'-GGAATTTCT-3', c-Rel:c-Rel z-score = 8.5) and the κB site from the murine Plau gene promoter (5'-GGAAAGTAC-3', c-Rel:c-Rel z-score = 12.9)⁵. We also found that a motif constructed from the c-Rel:c-Rel-preferred κB sequences exhibited a strongly degenerate half-site (Fig. 2b). Therefore, while the highest scoring c-Rel:c-Rel-preferred sites are pseudo-symmetric (Fig. 1d), a large number of non-traditional, c-Rel:c-Rel-preferred sites (and RelA:RelA-preferred, Supplementary Fig. 4) scored significantly above background yet have only a single canonical 5'-HGGAA-3' half-site.

Examining the p50:p50-preferred κB sites (p50:p50 z-score > 4, c-Rel:c-Rel z-score < 2), we found a number of κB sites with a G-rich 5' half-site. Highlighting the κB sites that conform to the pattern 5'-GGGGGNNNNN-3' (Fig. 2a, yellow dots), we observed a strong p50:p50 preference, although a subset of the sites is also bound well by c-Rel:c-Rel (discussed further below). A motif constructed from the G-rich sites bound well by p50:p50 (z-score > 4) exhibits a 5'-GGGGG-3' half-site and a degenerate 3' half-site (Fig. 2b), although a moderate preference for adenine and thymine bases 3' to the G-run was observed. Therefore, similar to the c-Rel:c-Rel-preferred sites, we observed statistically significant binding to a large group of κB sites defined by a single half-site sequence.

To ensure that the observed dimer-specific binding to non-traditional κB sites is not an artifact of our PBM approach, we examined binding to a set of traditional and non-traditional κB sites using SPR (Figs. 2b–d, Supplementary Fig. 5). Due to the very fast on-rates (K_on) of some dimers, we were unable to obtain reliable K_on measurements. However, we were able to obtain reliable off-rate (K_off) values and found excellent agreement between the SPR-determined K_off values and our PBM-determined z-scores (Figs. 2d,e). These results are consistent with previous reports²⁵ showing differential off-rates as the major contributor to binding affinity differences between κB sites. Our data demonstrate that our PBM-determined z-scores reflect equilibrium binding measurements and lend further support for the potential regulatory significance of the numerous non-traditional κB sites in our dataset.

Dimer preferences for κB sites of different lengths

DNA binding studies¹³ and X-ray crystal structures^6,12 have revealed different κB site lengths for c-Rel,RelA homodimers (9 bp), heterodimers (10 bp), and p50,p52 homodimers (11 bp). These length preferences are consistent with the DNA binding site motifs we determined for each dimer class (Fig. 1a). However, in light of the number of non-traditional binding sites in our dataset, we sought to determine whether binding site length preferences depend on the binding site sequence itself.

We examined how the DNA bases flanking 10-bp κB sites affect the binding to different dimers. Since p50:p50 binds an 11-bp site, we expected to observe a strong effect due to flanking base identity. We measured binding by PBM experiment to all 16 κB site variants in which the bases immediately 5' and 3' to the 10-bp site were exhaustively sampled (Fig. 3a). Examining the binding of p50:p50, RelA:p50 and c-Rel:c-Rel to traditional κB site sequences, we observed improved binding by p50:p50 and RelA:p50 with the addition of 5' guanine to one strand (Fig. 3a, columns 1 and 2). These differences are readily understood in terms of the known half-site preferences, with high-affinity binding occurring on 11-bp sites with symmetrically opposed, optimal 5-bp half-sites: 5'-GGGGA(A)TCCCC-3' and 5'-GGGAA(A)TTCCC-3'. However, for p50:p50 the highest-affinity binding occurred with 5' guanines flanking both half-sites, demonstrating a preference beyond the 5-bp half-site and showing that a 12-bp site can be differentiated from an 11-bp site.

In contrast, we observed that binding of all three dimers was unaffected by the identity of the bases flanking non-traditional κB sites (Fig. 3a, columns 3 and 4). This suggested a different mode of protein-DNA interaction for non-traditional κB sites. To delimit their lengths, we used our PBM dataset to examine binding to shorter κB sites. For example, to interrogate binding to a 9-bp sub-sequence of the 5'-GGGGAATTTT-3' site, we examined binding to the four κB sites in our dataset of the form 5'-NGGGAATTTT-3'. Binding of p50:p50 and RelA:p50 was insensitive to base identity at position 5 (positions numbered as 5'-G₋₅G₋₄G₋₃A₋₂A₋₁T₊₁T₊₂T₊₃T₊₄T₊₅-3') and only moderately sensitive to the base identity at position −5 (Fig. 3b). Therefore, the length of this non-traditional site is 8 to 9 bp (5'-gGGGAATTT-3'), which differs considerably from the 9- to 11-bp traditional κB site sequences (Fig. 3a). In contrast, c-Rel:c-Rel binding is insensitive to positions −5 and −4, but sensitive to base identity at position 5, demonstrating a similarly short 8-bp length but to a different sub-sequence (5'-GGAATTTT-3'). The same preferences were observed for the 5'-GGGGGTTTTT-3' site (Supplementary Fig. 6). Our data suggest that a fundamentally different mode of binding mediates the recognition of traditional versus non-traditional sites and that this translates into κB sites of different lengths. Furthermore, we observed that the length of the binding site is dimer-specific.

We analyzed the role of flanking bases to 17 additional κB sites and similarly found that 5' guanine bases were the most predictive of increased binding affinity for all NF-κB dimers, and that the effect of a 5' guanine was dependent on the G-content in adjacent bases. We generated a linear model to predict the PBM scores for all 12-bp κB sites based on our set of 10-bp κB sites (see Supplementary Methods). We demonstrate the efficacy of this extended dataset below in our comparison of PBM data with chromatin immunoprecipitation (ChIP) data, and make the full dataset available for use in genomic analysis (see Supplementary Data).

Affinity versus specificity of c-Rel and RelA homodimers

The Rel homology regions (RHRs) of c-Rel and RelA are more similar to each other than are the RHRs of any other pair of NF-κB family members³⁰. In vitro DNA-binding studies have demonstrated highly similar binding specificities, although c-Rel homodimers appear to bind a broader range of κB site sequences than RelA¹³. We observed highly correlated binding of c-Rel and RelA homodimers over our large set of ~3,300 κB sites (Figs. 1,3 and Supplementary Fig. 2). Despite these similarities, c-Rel and RelA can elicit distinct biological functions in vivo^6,17. c-Rel homodimers can preferentially activate the mouse Il12b gene by binding non-traditional NF-κB sites with much higher affinity than RelA homodimers¹⁷. Forty-six amino acids within the c-Rel RHR were found to be responsible for the enhanced binding affinity, and a chimeric RelA protein containing these 46 residues rescued Il12b expression in cRel^−/− macrophages.

To determine the relationship between our PBM profiles and binding affinity, SPR was performed with six different DNA sequences. For all sequences tested, except for one that bound poorly to both dimers, we observed much slower dissociation rates (K_off) for c-Rel homodimers than for RelA homodimers (Table 1). Importantly, swapping these 46 residues of the c-Rel RHR domain into RelA (protein RelA/N3,4) led to substantially slower dissociation rates (Table 1). RelA/N3,4 homodimers also exhibited a binding specificity that correlated more closely with c-Rel homodimers (Pearson r = 0.87) than with RelA homodimers (Pearson r = 0.82) (Fig. 4). However, these specificity differences are subtle in comparison to the global difference in binding affinity distinguishing c-Rel from RelA homodimers (median fold difference of 8.7 for c-Rel, RelA K_off values). Thus, while the DNA-binding specificities of RelA and c-Rel homodimers are highly correlated, c-Rel homodimers have much slower off-rates than RelA homodimers, resulting in a higher overall affinity and contributing to the selective regulation of c-Rel-dependent genes.

These results raised the question of whether binding affinities might discriminate other NF-κB dimers with correlated binding specificities. We performed SPR experiments with the six different DNA sequences using mouse p50:p50 homodimers and c-Rel:p50, RelA:p50, RelB:p50 and RelB:p52 heterodimers (Supplementary Table 1). The results failed to reveal differences of the same magnitude as those found for c-Rel and RelA homodimers. The most notable difference was that c-Rel:p50 heterodimers exhibited slower off-rates with some DNA sequences than the other heterodimers. Although the magnitudes of these differences were smaller than those observed with c-Rel and RelA homodimers (median K_off fold difference of 8.7 for the c-Rel, RelA homodimer, versus pairwise heterodimer differences ranging from 1.1 to 5.0), the results raise the possibility that enhanced binding affinity allows c-Rel:p50 heterodimers to selectively regulate some genes.

Comparison with in vivo binding data

We examined the relationship between our PBM-derived binding data and available genome-scale ChIP datasets on in vivo occupancy of RelA and p50^31–33. We found highly significant enrichment for high-scoring PBM-determined κB site sequences within ChIP-enriched (i.e., dimer-bound) regions (Supplementary Fig. 7). Moreover, we found significant enrichment when traditional κB sites were masked from the genomic sequence. These results demonstrate that both traditional and non-traditional κB sites in our PBM dataset represent binding sequences utilized in vivo.

To determine whether PBM-determined dimer-specific differences correlate with dimer-specific binding differences in vivo, we examined an NF-κB ChIP dataset³³ in which ChIP-chip was performed on LPS-stimulated human macrophages for all five NF-κB proteins. Focusing on p50 binding, which had the largest number of bound regions, we separated regions into those bound by p50 only (regions bound by p50:p50 homodimers, Fig. 5a) and those also bound by at least one of RelA, c-Rel or RelB (regions bound by p50 heterodimers or multiple dimers, Fig. 5a). This analysis allowed us to examine whether particular κB sequences distinguish regions bound only by p50:p50 homodimers and whether our PBM data for different dimers capture these sequence differences.

The ~8,000 human promoter regions in the ChIP dataset³³ were scanned with our PBM-determined 12-bp κB site sequences and assigned the z-score of the top-scoring κB site (see Supplementary Methods). Receiver-operating-characteristic (ROC) curve analyses were performed to quantify whether the p50-bound regions (true positives) scored more highly than the unbound regions (true negatives). We observed strong area under the ROC curve (AUC) enrichment scores for all three dimers tested (Fig. 5b–d). However, strikingly, we observed that p50:p50 PBM data yielded significantly higher enrichment scores for the regions bound only by p50 than for the regions bound by additional NF-κB members (AUC=0.84 versus 0.67, Fig. 5b). This was not the case for the RelA PBM data, which showed no discrimination between the two types of regions (Fig. 5d). This demonstrates that κB sequence features can discriminate the p50-specifically bound regions, and that these features correlate best with p50:p50 homodimer PBM data. The same analysis performed with PBM-determined 10-bp or 11-bp κB sites did not show the same discriminatory capacity for p50:p50 (data not shown) suggesting that the p50-bound sites are discriminated primarily by p50:p50 preferences for 12-bp long κB sites. These results demonstrate that the PBM-derived, dimer-specific binding differences relate directly to dimer-specific binding differences in vivo.

Preparation of protein samples

Mouse sequences for RelA, c-Rel, p50, p52, and RelB were cloned into a modified pET11a expression vector for purification. Constructs contained the Rel-homology region (RHR) of each subunit: RelA (1–314 a.a.), c-Rel (1–282 a.a.), p50 (1–429 a.a.), RelB (1–400 a.a.). The p50 subunit had a C-terminal FLAG tag. Proteins were expressed in BL21 (DE3) Escherichia coli cells (0.1 mM isopropyl β-D-1-thiogalactopyranoside (IPTG) induction) for 16 h at 25°C. Heterodimer subunits were co-expressed using a bicistronic expression plasmid⁴³. Protein purification was performed on a Q-Sepharose High Performance anion exchange column (GE Healthcare) and a SP Sepharose High Performance cation exchange column (GE Healthcare) and analyzed by SDS-PAGE and EMSA. The final purified protein samples were then frozen in aliquots in a storage buffer of 50 mM Tris-HCl, pH 8.0 150 mM NaCl, 5 mM DTT, and 10% glycerol.

Expression constructs for the human NF-κB dimers were created as established previously⁴⁴. Briefly, histidine-tagged recombinant proteins were produced using pET vectors in BL21 (DE3) E. coli (Merck). Constructs contained the Rel-homology region (RHR) of each subunit: RelA (1–307 a.a.), c-Rel (1–285 a.a.), p50 (7–356 a.a.), p52 (4–332 a.a.) RelB (120–401 a.a.). Proteins expressed was induced with 0.2 mM isopropyl β-D-1-thiogalactopyranoside (IPTG) at 30 °C for 5 h. Cell pellets were harvested in “Ni-NTA Binding” buffer with added EDTA-free Protease Inhibitor (Roche), pulse-sonicated for 2 min and debris removed via centrifugation at 16,000 g. A two-step purification procedure was then employed, first with the “Ni-NTA His-Bind Resin” system (Merck #70666) and then a subsequent purification based on DNA-affinity isolation of functional, DNA-binding protein. Ni-NTA purification was carried according to manufacturer’s guidelines while for DNA-affinity isolation, the processing of a sample derived from 250 ml of bacterial culture required 0.128 µM of the oligonucleotides “TNF-promoter” (biotinylated) and “TNF-promoter complementary”. Oligonucleotides were annealed via incubation in NEB Buffer 3 (New England Biolabs) at 94°C for 1 min, followed by 69 cycles of 1 min with step-wise decrease of 1°C. 712.5 µl of pre-annealed oligonucleotide mixture was conjugated with streptavidin-agarose (Sigma).

Surface Plasmon Resonance (SPR) experiments

Sensorgrams were recorded on a Biacore T100 (GE Healthcare) using streptavidin chips (Sensor Chip SA). Biotinylated oligonucleotide probes were immobilized on the surface of the streptavidin sensor chip in running buffer (10 mM HEPES, pH 7.5 150 mM NaCl, 3 mM EDTA, 0.005% Tween 20). Protein samples were applied to the sensor chip at 50 µL/min, at 10°C and referenced to an unmodified surface. Binding data was collected in the running buffer described above; sensor chip surface was regenerated with a 90 second pulse of 2 M NaCl followed by 180 second pulse of the running buffer. Dissociation rates were obtained by global fitting of the real-time kinetic data using the Scrubber2 software (BioLogic Software) and a simple 1:1 binding model. Six concentrations of each NF-κB protein were used, ranging from 1 nM to 1 µM.

Comparison and clustering of NF-κB PBM data

NF-κB dimer binding specificities was compared using the Pearson correlation coefficient of κB site z-scores.. Only κB sites with a z-score > 1 in at least one experiment were included in these calculations. Further, possibly redundant κB sites were ignored in the calculation if they could be explained by a higher-scoring κB site (i.e., if a higher-scoring κB 10-bp site matched its probe sequence); ~500 κB sites met this criteria. Calculations were performed using the R statistical software package. Hierarchical clustering and visualization of the comparison matrix (Fig. 1c) were performed using the heatmap function in R, with a ‘euclidean’ distance function and a ‘complete’ clustering function.

DNA binding site motif analysis

Binding motifs for universal PBM experiments were derived using the Seed-and-Wobble algorithm^22,26. DNA binding site motifs from top-scoring κB sites identified by custom NF-κB PBM experiments were determined by running the Priority 2.1.0 motif finding algorithm⁴⁵ on the 10-bp sequences. Graphical sequence logos were generated using enoLOGOS⁴⁶.