Comprehensive Identification of Human bZIP Interactions with Coiled-Coil Arrays

See allHide authors and affiliations

Science  27 Jun 2003:
Vol. 300, Issue 5628, pp. 2097-2101
DOI: 10.1126/science.1084648


In eukaryotes, the combinatorial association of sequence-specific DNA binding proteins is essential for transcription. We have used protein arrays to test 492 pairings of a nearly complete set of coiled-coil strands from human basic-region leucine zipper (bZIP) transcription factors. We find considerable partnering selectivity despite the bZIPs' homologous sequences. The interaction data are of high quality, as assessed by their reproducibility, reciprocity, and agreement with previous observations. Biophysical studies in solution support the relative binding strengths observed with the arrays. New associations provide insights into the circadian clock and the unfolded protein response.

In DNA binding transcription factors, autonomously folding domains are often responsible for the formation of homo- and heterodimeric protein complexes. Transcription factor dimerization can increase the selectivity of protein-DNA interactions and generate a large amount of DNA binding diversity from a relatively small number of proteins (1, 2). Dimerization also leads to the establishment of complex regulatory networks (3), and dimers can interact with other proteins required for transcription (4). The bZIP transcription factors constitute an important class of eukaryotic DNA binding proteins in which dimerization has been shown to occur by way of coiled-coil regions. Although previous studies have described bZIP pairings and the biological effects of these associations (3, 59), many substantive questions remain unanswered. First, little is known about the relative strength of bZIP interactions. Second, the use of complex reaction mixtures has made it difficult to assess whether interactions reported in earlier studies are direct or are assisted by bridging molecules. Third, the extent to which interaction specificity is inherent to the leucine-zipper region has not been fully established. Finally, the ability of many combinations of bZIP proteins to partner remains to be tested.

The emerging technique of printing and probing microscale protein arrays makes it possible to screen many interactions in parallel with small amounts of protein and uniform experimental conditions (10, 11). Protein arrays also permit the detection of direct interactions, something that is not guaranteed in two-hybrid or co-immunoprecipitation assays (1215). Further, by studying interactions at the domain level (16), undesirable effects due to oxidation, incorrect folding, and proteolysis can be reduced (when compared with the use of full-length sequences), resulting in fewer false-positives and false-negatives.

To prepare leucine-zipper arrays, we produced high-performance liquid chromatography (HPLC)–purified peptides corresponding to 49 (out of ∼55) human bZIP proteins (table S1). The sequences used span 16 different families, defined on the basis of sequence similarity in the coiled-coil region (17). Three peptides were prepared in duplicate to serve as internal controls. For comparative purposes, a further 10 bZIP domains from the budding yeast Saccharomyces cerevisiae were also expressed and purified. Each of these 62 constructs contained the entire predicted coiled-coil domain (18) and varying amounts of adjacent sequence to facilitate cloning.

The influence of valency on the strength of protein-protein interactions makes it important to present bZIP domains as monomers in association assays (16, 19). To this end, we printed reduced and guanidine hydrochloride–denatured bZIPs onto aldehyde-derivatized glass slides. Further, we rapidly diluted fluorescent peptides (hereafter, probes) from denaturant immediately before adding them to the processed array surfaces (17). In a series of pilot experiments, we determined that we could detect both strong (Fos/c-Jun, Kd ∼ 50 nM) (20) and weak (ATF-2/ATF-2, Kd ∼ 3 μM) (21) bZIP associations with probes at apparent concentrations about 1/100 that of the Kd.

To carry out a comprehensive analysis, we measured the interactions of all binary combinations of human and yeast bZIP domains with probe concentrations between 0.16 and 160 nM. Although the bZIP coiled coils share considerable sequence similarity (table S1), we observed very high interaction selectivity (Fig. 1, A and B). To quantify the results, we used data from five experiments performed at 40 or 160 nM probe concentration. Because surface peptides were spotted in quadruplicate, this yielded up to 20 measurements of each possible interaction (17). Each measurement was converted into a Z-score [(signal – mean)/estimated standard deviation] (17) reflecting the signal-to-noise ratio, normalized independently for each probe in different experiments. The Z-scores from all measurements were used to define each possible protein pair as “interacting” (Z > 2.5), “not interacting” (Z < 1.0), or “undetermined” (not meeting either criterion) (17) (Fig. 2). This quantitative analysis emphasizes the selectivity exhibited by the bZIPs: Out of 3481 measured pairs (i.e., 592 pairs of nonredundant peptides), only ∼14% resulted in interactions, and 5.8% were strong, with Z ≥ 10 (22). The ability to detect only one or a few partners for a given peptide, even among a large set of proteins with similar sequences, provides confidence in the reported associations.

Fig. 1.

bZIP interactions. (A) Interactions of fluorescent peptide Cy3-Bach1 (at ∼160 nM) with a bZIP array. Labels identify the surface features, spotted in duplicate, that bind Cy3-Bach1 most strongly. Immobilized peptides are located at the intersection of the tick marks and are printed successively by row, first from row 1, column 1 (R1C1) through R12C12, and then in a permuted order from R1C13 through R12C24. Scale bar: 1 mm. (B) Selectivity of binding. Mean fluorescence signals generated with 49 unique human bZIP peptides as probes. Each spot represents the average fluorescence signal for one of the surface peptides, spotted in quadruplicate, interacting with a given probe. Most surface peptides bind relatively little labeled peptide; only a few interact strongly. This is true even when the maximum signal is relatively low (e.g., columns 12, 15, and 16). The observed interaction selectivity is not due to an intrinsic inability of surface- or solution-phase peptides to associate with other sequences. All of the printed peptides interact with at least one solution-phase probe and vice versa (except ATF-4L1, column 30).

Fig. 2.

Consensus interaction matrix for 49 human and 10 yeast bZIP peptides. Fluorescent proteins (probes) are listed at the top, and surface-phase proteins are listed at the left. Control peptides spotted in duplicate are indicated name_1, name_2. Peptide pairs were assigned a Z-score (17) that corresponds to the highest value for which the probability of seeing the observed number of occurrences by chance is less than 104. Z > 20 (black); Z > 10 (dark blue); Z > 5 (medium blue); Z > 2.5 (light blue); Z > 1.5 (light green); at least 75% observations with Z < 1.0 (yellow); no assignment with confidence meeting P-value test (white); signal observed was not reciprocal (i.e., for the heterodimer XY, ZXY > 2.5, ZYX < 1) (gray). Comparisons of relative intensities are most accurate within columns, which share the same probe protein. The sparseness of strong human/yeast bZIP interactions (0.5% versus ∼5% for human/human or yeast/yeast) further points to the specificity of this motif. The background-corrected fluorescence data are available as part of the Supporting On-line Material.

The sequence-defined bZIP families exhibited characteristic interaction patterns (Fig. 2). As anticipated, peptides within families made similar interactions, whereas peptides in different families formed distinct sets of associations. Notably, the interaction array showed very high symmetry (more than 90% of observations were reciprocal), indicating that the assay was not sensitive to whether a protein was printed on the surface or used as a probe. The array data also showed high reproducibility, especially in comparisons to high-throughput studies of protein-protein interactions (23).

To compare interactions on the array with those observed in solution, we used circular dichroism (CD) spectroscopy. E4BP4, ATF-6, OASIS, and XBP-1 all formed strong homodimeric associations on the array, whereas c-Jun formed a weaker homodimer. In solution these peptides exhibited CD spectra typical of coiled coils, with double minima at 208 and 222 nm (Fig. 3A). In contrast, potential homodimers involving Fra1, Fra2, JunB, and JunD were each classified as noninteracting or undetermined on the array, and their CD spectra were not α-helical but instead showed random-coil character (Fig. 3B). We also measured the stability of these nine constructs as a function of temperature. Again, the solution data supported the observations made with the arrays. E4BP4, ATF-6, OASIS, XBP-1, and c-Jun unfolded cooperatively. c-Jun had a midpoint temperature for thermal unfolding (Tm) of 40°C; the other four peptides had Tms between ∼50° and 70°C. JunB, JunD, and the Fra peptides did not exhibit cooperative unfolding transitions but showed 50% loss of signal at 222 nm at temperatures between 10° and 15°C.

Fig. 3.

Correlation of array and solution studies. CD spectra were measured with a protein concentration of ∼5 μM at 25°C in phosphate buffered saline (pH 7.4). (A) Four proteins that form strong (Z > 10) homodimers (OASIS, ATF-6, E4BP4, and XBP-1) and one protein that forms a medium-strength (Z ∼ 5) homodimer (c-Jun) have CD spectra characteristic of folded coiled coils, as evidenced by minima at 208 and 222 nm. The mean residue ellipticity does not necessarily correspond directly to the helical content of the coiled-coil domains [e.g., basic regions can be helical in the absence of DNA (39)]. The symbols are, in order of increasingly negative signal at 222 nm: XBP-1 (▲), ATF-6 (◼), E4BP4 (♦), OASIS (⚫), and c-Jun (♦). (B) Four proteins that do not form homotypic interactions on the array surface have CD spectra characteristic of the random-coil configuration. The symbols are Fra2 (Δ), JunB (◯), Fra1 (▢), and JunD (♢). (C) The relation between the Tm values for a series of disulfide–cross-linked peptides studied by Kim and colleagues (25) and the fluorescence signal obtained using recombinant Cy3-GCN4 to probe bZIP arrays. Data generated by a single experiment share a common symbol [also true for (D) to (F)]. Data points at 36°, 64°, and 83°C correspond to GCN4/Fos, GCN4/c-Jun, and GCN4/GCN4 dimers, respectively. (D to F) The relation between the Gibbs free energy of unfolding for a series of disulfide-bonded synthetic peptides (17) and the fluorescence signal obtained using recombinant Cy3–c-Jun (D), Cy3–ATF-2 (E), and Cy3–ATF-3 (F) to probe bZIP arrays. Unfolding free energies were determined by denaturation with guanidine hydrochloride. In (D) the complexes at 4.0, 5.2, 5.6, and 6.1 kcal/mol occur between c-Jun and ATF-4, ATF-3, c-Jun, and ATF-2, respectively. In (E) the complexes at 3.7, 3.9, and 6.1 kcal/mol occur between ATF-2 and ATF-2, ATF-3, and c-Jun, respectively. In (F) the complexes at 1.5, 3.9, and 5.2 kcal/mol occur between ATF-3 and ATF-3, ATF-2, and c-Jun, respectively.

Information about relative binding strengths is critical for decoding the biophysical basis of interaction specificity and for formulating quantitative models of cellular networks. This type of information is not available from two-hybrid or immunoprecipitation studies (24). However, we observed a high correlation between fluorescence signals generated by Cy3-GCN4 interacting with surface proteins GCN4, c-Jun, and Fos, and the reported solution Tm's of disulfide–cross-linked peptides GCN4/GCN4, GCN4/c-Jun, and GCN4/Fos (25) (Fig. 3C). Good agreement was also found when comparing signals generated with probes Cy3–c-Jun, Cy3–ATF-2, or Cy3–ATF-3 with the Gibbs free energy of unfolding for a series of seven disulfide-linked heterodimers (Fig. 3, D to F) (26).

Taken together, the reproducibility and symmetry of the array data, the high purity of the reagents used, and the good agreement with solution studies all indicate that we have identified bona fide interactions. This finding is supported by excellent agreement with previous bZIP interaction studies, where such studies exist. Sixteen bZIP families (table S1 and Fig. 2) can potentially give rise to 136 [i.e., n(n + 1)/2] different family-family interactions. Out of these 136 possibilities, we found reports describing 58 inter- or intrafamily interactions. Our data agree with 50 of these previous studies (including five contested cases) and partially agree in another six instances (fig. S1). We do not observe two sets of interactions that others have detected. In a more detailed analysis, we compared the array data with associations reported to involve proteins in the well-studied Jun and Fos families. We find complete agreement with 41 of 51 interactions reported, and our data are consistent with seven other studies (fig. S2). We do not observe three reported interactions (27). Thus, we estimate our rate of false-negatives as less than ∼6%, with no evidence for false-positives. Collectively, these comparisons confirm that coiled-coil domains alone, and in the absence of DNA, are often sufficient for determining bZIP interaction specificity.

Charged and polar residues at the interface between coiled-coil strands are known to contribute to partnering specificity (9, 25, 28, 29). We found that simple empirical rules describing pairwise interactions of such residues can be used to detect unfavorable pairings. For example, using one rule (17), we identified 372 of 890 noninteracting pairs (42%) with only one false-negative out of 80 strongly interacting pairs (1.3%) (fig. S3). However, such rules do not fully separate strongly interacting pairs from noninteracting pairs, which confirms that other factors contribute to coiled-coil interaction specificity and should be considered when developing methods for prediction (30, 31).

We uncovered a number of interactions that, to our knowledge, have not been previously detected, including associations between proteins belonging to the following families: B-ATF and C/EBP, B-ATF and ATF-2, E4BP4 and CREB, E4BP4 and OASIS, and ATF-6 and XBP-1 (fig. S1). Because the domain-based array provides a reliable assay for direct interactions, these observations provide excellent candidates for functional studies. Particularly suggestive are the interactions of individual bZIP proteins E4BP4 with CREB and the interactions of XBP-1 with ATF-6 and with Zhangfei (ZF).

E4BP4 and CREB are present in the same tissues, at the same time, and are linked to the same processes (3234). Additionally, the consensus half sites for each homodimer differ by a single base pair, and both homodimers can bind DNA sequences comprising adjacent E4BP4 and CREB half sites (35), suggesting that E4BP4-CREB heterodimers will bind similar sites. Because E4BP4 is a component of the central circadian oscillation machinery, whereas CREB is known to target genes that respond to circadian rhythms (e.g., those involved in gluconeogenesis), we postulate that the E4BP4/CREB association joins the central circadian machinery to downstream processes.

The arrays also revealed candidate interactions involving proteins with a role in the intracellular signaling pathway known as the unfolded protein response (UPR). The UPR consists of at least three branches, specified by different protein sensors (i.e., ATF-6, IRE1, and PERK) and their targets. ATF-6 and XBP-1 (representing the ATF-6 and IRE1 UPR branches) interact directly in our assay, and the resulting heterodimer is a strong candidate to promote the transcription of a subset of UPR genes (36). XBP-1 also interacts with a single other transcription factor in our screen, the incompletely characterized protein ZF. ZF homodimerizes, and interacts with ATF-4, a downstream target of the UPR sensor PERK. Thus, ZF potentially links the IRE1 and PERK branches of the UPR.

One important future goal is to compare and contrast in vitro and in vivo interaction data to better understand how associations are affected by other cellular factors. In addition, although two-hybrid or co-immunoprecipitation assays are valuable for genome-wide analyses that aim to uncover new interactions (1215), these methods are susceptible to high false-positive and false-negative rates (37). A domain-based approach using protein microarrays can provide more reliable and quantitative information. Finally, the approach described here is not limited to coiled coils but potentially can be used to analyze interactions that occur between any of the computationally detectable interaction motifs that abound in eukaryotic genomes (38).

Supporting Online Material

Materials and Methods

Figs. S1 to S3

Tables S1 to S6


References and Notes

View Abstract

Stay Connected to Science

Navigate This Article