Specificity landscapes of DNA binding molecules elucidate biological function

CSI analysis of the natural and engineered DNA binding molecules yielded a comprehensive binding profile across the entire sequence space of a binding site (Fig. S1C). The highest intensity binding sites were compiled to identify consensus binding motifs (Fig. S1D) which are displayed as sequence Logos (33) in Fig. 1. This comparative analysis indicates that the consensus motif bound by polyamide PA-1 is similar to motifs recognized by natural proteins, both in length and information content.

While consensus motifs derived from PWMs summarize the best binding sequences, they overlook significant insights embedded within the complete sequence recognition profiles captured by CSI analysis (Fig. 2 A left panel) (34 –36). To surmount this limitation, we created the SSL as an adaptable tool to display and interpret the full recognition preferences of DNA binding molecules (Fig. 2). SSLs present the entire binding dataset through a series of concentric rings or a linear format (Fig. 2 A). In a circular SSL, the innermost ring displays sequences that contain a perfect match to a given seed motif (0-mismatch). The subsequent circles, going outward, represent increasing mismatches from the seed motif. The height of each color-coded peak on the individual rings corresponds to the CSI binding intensity data. In this way, the entire sequence space can be displayed in a comprehensive yet easily interpreted format. To generate an SSL, a seed motif is used as the starting input; this motif may be generated from a known PWM, although a motif from any source may be used to initialize the sequence alignment. Through iterative optimizations the seed motif is refined until one obtains an SSL with high affinity sequences restricted to the 0-mismatch ring, and moderate-to-low affinity sequences assigned to the appropriate outer mismatch rings. The sequences in the center ring are sorted first alphabetically by motif (or submotif) and then alphabetically by the flanking sequences. This provides a consistent ordering of sequences in the center ring. The sequences represented in the outer rings are organized first by the position of the mismatch and then alphabetically (e.g. in the motif G₁A₂T₃A₄A₅,G₁ is substituted by A₁,C₁,T₁, followed by A₂ to C₂,G₂,T₂, and onwards through all positions of the motif). In the linear format, the display permits vertical alignment of the sequences by mismatch for each position in the motif (Fig. 2 A and Fig. S2B). In essence, the SSL format reduces the high-dimensionality problem of displaying nonlocal sequence interdependencies of nearly a million different permutations into a readily interpretable graph that reports the entire specificity-spectrum of a DNA binding molecule.

A key advantage of the landscape display is that all data points are displayed without sequence compressions that are necessary to generate PWMs and consensus motifs. The absence of sequence compression in SSL facilitates the identification of optimal motif(s). If the seed motif, used to initiate an SSL, is too restrictive then several high-intensity sequences appear in the mismatch rings, and conversely if the motif is too inclusive then low-intensity peaks (i.e. valleys) would invade the innermost match ring (Fig. S2C). Importantly, if a chosen seed motif ignores the existence of alternative modes of binding to different sequences, then outer mismatch rings would display clustered peaks which would reveal the unique binding motifs (see below). Moreover, because the motifs identified by CSI (Fig. 1) are shorter than the permuted DNA duplexes on the array (5–7 bp motifs embedded within a 14–20 bp duplex), the landscapes identify the contributions of flanking sequences on the binding energetics of a given motif. In contrast to binding profiles predicted by PWMs (Fig. S2A), the flanking sequence influences binding properties to an unexpected degree, yielding highly textured binding profiles and rugged landscapes even in the 0-mismatch ring (Fig. 2). This additional contextual information greatly improves the accuracy of regulatory element annotation across genomes.

The SSLs of natural and engineered DNA binders immediately reveal several interesting results (Fig. 2 B). Hairpin polyamides display a high-degree of sequence specificity as compared to natural DNA binding proteins. For PA-1, as well as previously studied PA-2 (chemical structure of PA-2 shown in Fig. S1B) (31), the highest intensity peaks are clustered within the perfect match ring (innermost circle). While few strong binding sites are present in the 1-mismatch ring, virtually none are present in the 2-mismatch ring. By contrast, widely differing specificity spectra are evident among the protein-based DNA binding domains. Notably, the C₄-type ZF from Gata4 shows high-fidelity binding to sequences bearing the known motif. Although p53 prefers its consensus site, binding to other biologically relevant deviations of this sequence is also detected by CSI-SSL analysis (37, 38). TBP prefers an AT stretch but permits some deviation from this sequence. The SSL displays also shed light on the specificity of weak protein binders and resolve inconsistencies among reports in the literature. While Jumonji was shown to bind AT-rich sequences (24) others find that Jumonji prefers additional sequences (39). Our findings indicate that Jumonji binds with low specificity to A/G-rich sequences, thus resolving the two conflicting views.

We also analyzed ZFP-1, an engineered C₂H₂-type ZF protein. This protein was designed to target the sequence 5′-GGA-GTT-AAG-3′ using three linked ZF modules. Although the PWM-based Logo (from a CSI array with 10 permuted bp) suggests that the best binding sequences include 5′-RGGAGT-3′, the SSL analysis indicates that ZFP-1 recognizes 5′-GGAGTT-3′ (Fig. S2D). The 6-bp site indicates that only two of the three linked fingers are binding DNA specifically (6). The third “nonspecific” finger, when placed between flanking ZF modules, targets the 5′-AAG-3′ trinucleotide (40). However placing this finger at the N terminus leads to a significant diminution of its sequence specificity. This observation is consistent with previous observations of nonadditive behavior of these modules (34) and recent reports that the predictability of sequence recognition by engineered ZFPs is less than desired (41). Taken together, SSLs show that hairpin polyamides are among the most specific DNA binders.

The SSL of a homeodomain protein, Nkx-2.5, shows that the consensus motif 5^′-T₁T₂A₃A₄G₅T₆G₇-3^′ provides an incomplete view of the true binding specificity (Fig. 1 B). The submotifs T₁N₂AAGTG and N₁T₂AAGTG are strong binding sites with TTAAGTG being the best (29). These compressed submotifs are readily deconvoluted as high-intensity peaks in the 1-mismatch ring of an SSL generated with the TTAAGTG seed motif (compare Fig. 2 A with dual submotif landscape in Fig. 2 B). To explore the generality of this phenomenon, we examined motifs of murine homeodomain proteins that were recently reported (42). Evaluating all 18 classes of homeodomains using the SSL displays revealed that nearly 40% of the motifs could be optimized to better define the specificity of the proteins (Fig. S2E). For example, in the case of Barx1 the reported consensus motif compresses two submotifs and inaccurately includes poorly binding sequences within the motif. The SSLs of the same dataset yield a more accurate view of the cognate sites preferred by Barx1 as well as several other homeodomains (Fig. S2E).

CSI analysis combined with SSL display permits the rapid and rigorous evaluation of engineered DNA ligands with altered chemical composition and architecture. We synthesized a 6-ring hairpin polyamide (PA-3) rather than the larger 8-ring hairpin polyamide (Fig. 3 A and Fig. S3). The CSI-SSL analysis shows that this minimized molecule binds a shorter core sequence but its specificity profile is comparable to 8-ring hairpin polyamides. This has significant implications in engineering smaller molecules that retain specificity yet have improved cell permeability properties.

The SSL profiles also indicate that modular recognition between heterocycle rings of polyamides and individual base steps of DNA governs recognition in this class of molecules. The results strongly support the recognition rules described by Dervan and others (9, 11, 15). We also used the SSL format to display CSI data of a hairpin polyamide (PA-4) containing an atypical 3-chlorothiophene ring (Fig. 3 A and Fig. S3) (30). The SSL of this bioactive molecule validates the increased preference for Thymine by the thiophene ring. This result highlights the importance of the CSI-SSL approach in evaluating specificities of new chemical entities that target DNA. Finally, the SSL of a linear polyamide, PA-5, (30) that targets sites abnormally repeated in Friedrich’s ataxia patients shows binding to a larger 9 bp site (Fig. 3 A and Fig. S3) (17). In the landscape, significant binding is detected in the outer mismatch rings. Examination of the outer-ring peaks identified three related submotifs (Fig. 3 B) that were concealed in consensus motifs derived from PWMs (30). This example of relaxed specificity displayed by a linear polyamide (PA-5), along with the proteins described above, serve to highlight the ability of SSLs to identify multiple binding submotifs that are compressed into a single consensus by motif-finding algorithms. These results also reemphasize the utility of the CSI-SSL approach in creating and evaluating molecules that target DNA with the desired degree of specificity.

SSLs directly translate to binding energy landscapes for hairpin polyamides because their CSI intensities correlate strongly with equilibrium binding energies measured in solution (29, 30). To evaluate if this relationship holds for a natural DNA binding protein, we measured affinities of Nkx-2.5 for six DNA sequences that span a range of CSI intensities. As a control we refined the previously reported (29) affinity of PA-1 for a similar range of sites identified by CSI. The binding isotherms for these sequences were determined by nuclease protection assays for PA-1 and by EMSA for Nkx-2.5 (Fig. 4 and Fig. S4). The CSI intensities of both molecules were highly correlated with equilibrium binding energy. Not only do we observe a clear correlation at strong-to-moderate binding sites, the CSI intensities indicate differential binding at suboptimal sites that are too weak to be evaluated by EMSA analysis. The linear correlation observed, indicates that the fractional occupancy on the array is sufficiently low so as to still correlate with solution binding affinity. The low fractional occupancy is likely due to a combination of multiple factors. Thus, SSLs directly translate to energy landscapes and provide unprecedented information on binding energetics across broad sequence space (see Fig. 5).

A compelling need for developing CSI-SSL is to map the regulatory elements across the genome. Based on the CSI scores, we assigned binding probabilities across the entire human genome (Fig. 5). We designate these genome-wide binding landscapes, as “genomescapes.” Here, we focused on the CSI-SSL scores of a hairpin polyamide (PA-4) that targets the binding sites of the hypoxia induced transcription factor, HIF-1α. PA-4 binding to HIF-responsive-elements (HREs) blocks the activation of the hypoxia induced genes (18), including VEGF, a cytokine that is implicated in cancer and angiogenesis (43). Blocking HIF-1α dependent expression of VEGF suffices for averting tumorigenesis, thus marking this binding event as a target for therapeutic intervention. Previous genome-wide expression profiles indicated that genes bearing an HRE could be targeted in cells by the designed polyamide (PA-4) (Fig. 5 B) (18). However, in those studies it was not clear why the endothelin gene, ET-2, bearing a moderate HRE was robustly inhibited by the compound. To resolve this outstanding issue, we examined the genomic DNA for nonobvious binding sites that may exert additional inhibition. The CSI genomescape of ET-2 promoter regions identifies moderate PA-4 sites across the known HIF-1α binding site. More importantly, in contrast to other HIF-1α regulated genes, ET-2 also has several moderate PA-4 binding sites across the Transcription Start Site (TSS) and into the coding region (Fig. 5 C and Fig. S5). At 1 μM concentrations of the polyamide in the media, ET-2 gene is robustly down-regulated. CSI genomescapes suggest that this unusually high level of inhibition might be due to multiple synergistic inhibitory events where both the transcription factor and the transcriptional machinery are occluded from their binding sites. The binding energetics for this compound suggests that nanomolar binding affinities are required for effective action of the compound in living cells. We further anticipate that the CSI genomescapes will greatly aid in predicting how DNA binding molecules localize across the genome and in elucidating seemingly off-target events in living cells.

The fluorescently-labeled polyamide was diluted to a final concentration (250 nM for PA-1, PA-2, PA-5; 20 nM for PA-3; and 10 nM for PA-4) in hybridization buffer (1 M NaCl, 100 mM MES pH 7.5, 20 mM EDTA, 0.01% Tween-20). The polyamide was added to the hybridization chamber (Grace BioLabs) on the array and rotated for 1 h at 23 °C. The arrays were washed with nonstringent wash buffer (6X SSPE pH 7.5, 0.01% Tween-20), dried, and scanned using an Axon 4000B (Molecular Devices). Data was viewed using GenePix™ Pro 6.0 (Molecular Devices).

Arrays were blocked with 2.5% nonfat dried milk for 1.5 h. The protein was diluted to a final concentration in buffer and mixed with a directly-labeled fluorescent antibody to either the protein or a tag on the protein. The protein-antibody mixture was added to the hybridization chamber on the array and rotated for 1 h. The arrays were washed, dried, and scanned as above.

To order the binding data for presentation in the SSL, the binding site on each sequence is determined using the best match to the seed motif. Landscapes are optimized by maximizing the number of high-intensity sequences and minimizing the number of low-intensity sequences in the 0-mismatch ring. The data is parsed according to the number of mismatches from the motif being plotted. If the seed motif has multiple submotifs, we use the minimum number of mismatches to the best submotif. Each ring is sorted first by the position of the mismatch and then alphabetically. The intensities are then plotted using MatLab R2006a (The MathWorks, Inc,). The intensities are smoothed using 0.5% of the data on each ring.

Additional details are included in SI Text.