Defining the sequence-recognition profile of DNA-binding molecules

Results

Array Design. The duplex DNA sequences are designed as self-complementary palindromes interrupted at the center by a TCCT sequence to facilitate the formation of DNA hairpins (Fig. 1). The 34-residue oligonucleotide is synthesized directly on the glass surface by using a maskless array synthesizer (25) that can readily create up to 786,000 spatially resolved features. After inducing hairpin formation, we found that 95% of the oligonucleotides in the array form duplexes (see Materials and Methods). In our hairpin design, we added three constant base pairs on either side of the 8 bp that were permuted (N¹–N⁸ in Fig. 1). Previous work shows that this addition is sufficient to buffer the core of the hairpin stem against thermal end-fraying of the duplex and against deviations from B-form DNA resulting from the presence of the loop (26). There is good evidence that the core of a hairpin stem interacts with proteins and small molecule ligands indistinguishably from DNA duplexes composed of two individual complementary strands (27, 28).

Array Validation Using an Engineered Small Molecule. To test the accuracy and fidelity of the CSI array, we used a polyamide engineered to target a specific DNA sequence (PA1, Fig. 2A). Polyamides are DNA-binding small molecules composed of N-methylpyrrole (Py) and N-methylimidazole (Im) heterocycle rings. The arrangement of the heterocycles (Im or Py) can be programmed to create polyamides that target most naturally occurring 6- to 8-bp DNA sequences (2). PA1, in particular, was designed to target the sequence 5′-WWGWWCWW-3′ (W = A or T) (Fig. 2) (29). A Cy3 fluorescent dye is conjugated to the N-methyl position of an internal pyrrole (Py*). Such conjugation does not meaningfully alter the DNA-binding properties of the polyamides (30). Previous solution-based footprinting (29) and dye displacement assays (28) have shown that polyamides discriminate very highly between their targeted cognate site and sites that differ by a single base pair. Thus, PA1, a well characterized DNA-binding molecule, serves as a stringent test for the ability of the CSI array to accurately identify its sequence recognition landscape.

PA1 was incubated with the array and a distinct pattern of fluorescent binding features was readily discernible, and the pattern did not change over a broad range of PA1 concentrations (0.5–500 nM). The array-to-array variability was very low, with an average correlation coefficient of 0.88 (Fig. 6, which is published as supporting information on the PNAS web site). A majority of the features showed low background fluorescence, and a small subset of the features were of high intensity (Fig. 2B and Fig. 7, which is published as supporting information on the PNAS web site). The duplicate features within an array and replicate features between arrays were averaged together to give finalized intensities. These averaged intensities were then converted into Z scores [Z = |signal – mean|/standard deviation] to reflect the signal-to-noise ratio (Fig. 2B).

Sequences in the highest Z score bin (≥25) were subjected to several motif-searching algorithms (31–33), which identified 5′-W¹W²G³T⁴W⁵C⁶W⁷W⁸-3′, a motif that is nearly identical to the predicted binding site for the polyamide 5′-WWGWWCWN-3′ (Fig. 2C). Parsing of the core sequences (N²–N⁷) showed that not all permutations of the consensus are bound equally well. In particular, all sequences that contained the sequence 5′-WWGATCWW-3′ had significantly lower intensities than other permutations of the consensus sequence (Fig. 2D). This observation is consistent with previous solution studies (34). Furthermore, the flanking sequence (N¹,N⁸) showed an equally strong preference for a W (A/T) in both positions (Fig. 8, which is published as supporting information on the PNAS web site). This observation is also in agreement with the preference of the polyamide γ-aminobutyric acid turn and Dp tail for A/T residues (35). Finally, we found that the cognate site preferences identified by the array were entirely consistent with reported solution binding studies of this polyamide for five different sequences (Fig. 2E and Table 1, which is published as supporting information on the PNAS web site). The high correlation (r² = 0.997) of feature intensity on the array with affinity for different cognate sites in solution provides significant confidence in the veracity of the cognate site preferences identified by the array. Taken together, these correlations demonstrate that the CSI array correctly identifies the cognate sites of a DNA-binding molecule, and that the CSI array accurately ranks each cognate site in the order of increasing affinity.

Comprehensive Mutational Analysis. In essence, the array performs a comprehensive “mutational” analysis as it queries the entire sequence space (within a defined size) to determine the contribution of every base pair in the cognate site for molecular recognition (Fig. 3). By examining the array data, it is apparent that substituting an S (G or C) at position 8 only subtly decreases binding by PA1. This finding is consistent with the ability of this symmetric polyamide to bind the sequence in only one orientation. Replacing one of the S residues at positions 6 (or 3′) with a W significantly attenuates, but does not abolish, binding. However, substituting any other position that prefers a W in the motif with an S residue nearly abolishes binding by PA1 (Fig. 3). The data also show that despite a double substitution at positions 3 and 6 to W, the resulting A/T stretch retains its ability to bind PA1. This observation is likely a result of the inherent affinity of polyamides for A/T-rich sequences (36).

Transcription Factor Binding and Cooperative Assembly. Having demonstrated the accuracy of the CSI arrays, we probed the sequence preferences of molecules that bind DNA cooperatively. We examined the cognate site preference of Exd, a transcription factor that plays an essential role in Drosophila development and is highly conserved across species, including humans (37). Exd binds DNA cooperatively with transcription factors from the Hox (homeobox) family (37, 38). Individual Hox proteins, as well as Exd, bind DNA with very low affinity and with poor specificity (38). Cooperative binding dramatically increases the affinity of Exd and Hox proteins for DNA and strongly influences DNA sequence specificity such that different Hox–Exd complexes target different genes (38). We have generated synthetic molecules (polyamide-peptide conjugates) that can mimic two key functions of the Hox family of transcription factors (39). First, they can bind sites targeted by specific Hox proteins, and second, they can cooperatively recruit Exd to an adjacent cognate site.

To determine the sequence specificity of Exd, we labeled it with Cy3 at a unique cysteine residue on an unstructured portion of the protein (Fig. 9A, which is published as supporting information on the PNAS web site.) (40). The modified protein does not differ in its ability to bind cooperatively with ultrabithorax (Ubx), a member of the Hox family, or with a synthetic Hox mimic to a known cognate DNA site. When tested on the CSI array, Exd alone, as expected, demonstrated little sequence-specific binding at concentrations ranging from 0.2 to 200 nM. It does, however, show an unexpected preference for stretches of consecutive G residues (Fig. 9B). Initial studies suggest that these sequences can form non-B-form, likely G-quadruplex (41), structures (Fig. 9D). The physiological importance of this binding interaction remains to be investigated.

When incubated with two different synthetic Hox mimics (PA2 and PA3), the Cy3-labeled Exd displayed an unambiguous pattern of feature binding in both sets of experiments (Fig. 4). PA2 and PA3 (ImImPy*Py-γ-ImPyPyPy-β-Dp) are designed to target the sequence 5′-WGWCCW-3′. Furthermore, instead of a Cy3 dye, PA2 and PA3 do not bear any dye but are conjugated to an Exd-binding peptide (N-FYPWMK-C). PA2 and PA3 differ solely by a single methylene in the linker connecting the Exd-binding peptide to the polyamide (Fig. 4A) (42). Because PA2 and PA3 are not fluorescently labeled, we detected cognate sites bound cooperatively by these synthetic Hox mimics and Exd, as well as sites bound by Exd alone.

The raw array data for the above experiments were treated as described for Fig. 2 (Fig. 10, which is published as supporting information on the PNAS web site). In addition to the G-stretches that Exd binds in the absence of any partner, three clear motifs emerged from the PA2–Exd data, whereas only two of those motifs were found in the PA3–Exd data (Fig. 4B). The Exd binding motif is 5′-NGAN-3′, which is consistent with the structural and genetic studies of Hox–Exd cognate sites (43). In other words, the 5′-GA-3′ dinucleotide is the only required sequence determinant for Exd binding to DNA. Remarkably, the array identified the differences in the arrangement of polyamide and Exd binding sites because of an ≈1.25-Å difference in the linker length between PA2 and PA3 (Fig. 4C). The other important result that emerged is that cooperative ternary assembly with Exd stabilizes binding of synthetic Hox mimics to truncated sites (5′-WGWC-3′). This stabilization is often seen in nature, where cooperative assembly of transcription factors utilizes suboptimal binding sites to ensure that only a higher order complex can efficiently bind to a regulatory element (44, 45).

Solution Binding and Molecular Modeling. To validate the unexpected differences in the motifs identified by each polyamide with Exd, we performed EMSAs. These studies with Exd and the two Hox mimics strongly support the cognate site preferences identified by the array (Fig. 5 A and C). Furthermore, molecular modeling (46, 47) analyses of PA2 and PA3 with Exd (with a docked Hox hexapeptide) agree well with the CSI array data. Both demonstrate that the linkers for PA2 and PA3 (9.98 and 11.25 Å, respectively) are able to deliver the hexapeptide to Exd at the composite consensus site (Fig. 5B). The array data indicate that both PA2 and PA3 reach Exd at the gapped composite site (consensus +1); however, simple geometric measurements with some energy minimization (48) suggest that the linker of PA3 should not span the distance. In the case of inverted binding sites, it is clear from modeling that the linker of PA3 is incapable of reaching Exd, and that the linker of PA2, even when fully extended, would be suboptimal, yielding an unstable ternary complex with Exd. These predictions are in good agreement with the observed CSI array binding data and EMSA results (Fig. 5C). However, the array data also demonstrate that a single base overlap (consensus –1) in the binding sites is not able to support binding of the complex, despite the fact that modeling indicates that the distance is similar to that of the consensus +1 site (Fig. 5B). The binding of either partner to overlapping sites may deform the DNA and prevent complex formation, even though modeling studies suggest that polyamide or Exd binding to the consensus –1 site should not disfavor complex formation (40, 49). Therefore, the ambiguities in molecular docking and energy minimization methods prevent precise prediction of the geometry of DNA grooves and distances between the interacting partners. In other words, the dramatic consequences on cognate site preference because of subtle, seemingly trivial, alterations in the linker length would not be readily apparent without the CSI array analysis. Therefore, this approach provides unexpected insight into molecular recognition properties of DNA-binding molecules when they bind individually or in cooperative pairs.

Discussion

In the CSI approach, the comprehensive sequence recognition landscape of DNA-binding molecules is determined in a rapid, unbiased, and unsupervised manner. Although we have directly labeled the DNA-binding molecules with a fluorescent dye, we anticipate that fluorescently labeled antibodies would serve equally well in detecting target proteins bound to their cognate sites. Because of the display of entire sequence space (within a certain size) on the array, there is no limitation on the use of proteins of a specific organism (as would be the case with ChIP-chip) or a specific class of small molecule ligands.

Because the CSI array analysis examines the entire sequence space at once, it also performs a comprehensive mutational analysis in a single experiment. Thus, one obtains information on the contribution of each nucleotide residue to the molecular recognition event between the DNA ligand and its cognate site(s). Moreover, DNA-binding preferences of the ligands (proteins or small molecules) are queried under identical conditions, yielding high-quality information. In the future, the accumulation of binding data from CSI analysis of different molecules will lead to the elucidation of the molecular recognition by a cluster of residues displayed on the surface of DNA-binding molecules. Such integrative perspective may be necessary to decipher the principles of molecular recognition displayed by DNA-binding molecules.

By determining the complete sequence recognition profile of DNA-binding molecules, the CSI array analysis also bridges the gap between the ChIP-chip approach and bioinformatic approaches of identifying regulatory DNA elements in genomes. For example, from a single CSI array experiment, one can validate (and order by affinity) all binding sites identified by ChIP-chip assays. Furthermore, the rank-order of the sequences can be used to computationally mine the genome for possible binding sites. The CSI array analysis would enable a coherent analysis of transcriptome studies by scanning for the presence of a range of possible binding sites in coregulated genes. Thus, including a CSI analysis in conjunction with other approaches will greatly aid in reducing the discrepancies between and absence of discernible binding sites in coregulated genes or the inability to detect protein binding at all biologically relevant sites in vivo by ChIP-chip analysis. The CSI array also provides a much-needed high-throughput approach for the design and development of novel classes of sequence-specific DNA-binding molecules.

As the ability to display more oligonucleotide features on a surface increases, the CSI approach can be easily scaled to represent larger sets of sequence variants. The most current technology (up to 10 positional variants), which is accessible to the entire scientific community, is sufficient to determine the binding preferences of nearly all metazoan DNA-binding proteins and engineered DNA-binding small molecules. Thus, this platform provides a powerful tool for tackling the important challenge of deciphering the DNA recognition code of DNA-binding molecules, individually or in cooperatively assembling complexes.

Materials and Methods

Duplex DNA Arrays. Microarrays were synthesized by using a Mask-less Array Synthesizer (NimbleGen Systems, Madison, WI) (25). Homopolymer (T₁₀) linkers were covalently attached to monohydroxysilane glass slides. Oligonucleotides were then synthesized on the homopolymers to create a high-density oligonucleotide microarray. The array surface was derivatized such that the density of oligonucleotides was sufficiently low within the same feature so that no one oligonucleotide would hybridize with its neighbors. Four copies of every sequence required a total of 131,584 features per array.

Hairpin Formation Percentage. In two distinct features on the array we present two sequences: one that forms a hairpin (5′-CGCTTAGTTCA-CGC-TCCT-GCG-TGAACTAA-GCG-3′) and one that does not (5′-CGC-TTAGTTCA-CGC-3′). By using a Cy3-labeled DNA probe that is complementary to the core sequence (5′-CGC-TTAGTTCA-CGC-3′) present in both oligonucleotides, we determined the ability of the complementary strand to bind the hairpin versus the single-stranded DNA molecules. The fluorescence intensity of the hairpin sequence was divided by the fluorescence intensity of the single-stranded sequence. The averaged background-subtracted intensity ratio of the double-stranded versus the single-stranded features indicated 95.6% hairpin formation.

Polyamide Synthesis. Polyamide-Cy3 conjugate PA1 was prepared by employing an orthogonally protected N-(phthalimidopropyl)pyrrole building block in standard Boc-based solid-phase synthesis (50). Cleavage of the polyamide from 100 mg of phenylacetamidomethyl (PAM) resin by treatment with 1 ml of dimethylaminopropylamine also removed the phthalimide protecting group to give the free base. The crude cleavage mixture was diluted with 0.1% trifluoroacetic acid (aq) and acetonitrile to a final volume of 5 ml and loaded onto a preconditioned solid-phase extraction column (C₁₈ bonded phase). After washing with a 4:1 (vol/vol) solution of 0.1% trifluoroacetic acid (aq) and acetonitrile, product was eluted with methanol, and solvents were removed by azeotropic distillation from toluene. The resulting aminopropyl precursor of PA1 was a slightly yellow solid. Analytical HPLC and MALDI-TOF MS verified the identity and purity of this intermediate, and it was used without further manipulation.

The intermediate free base (0.5 μmol) was dissolved in 0.45 ml of anhydrous dimethylformamide and 0.05 ml of diisopropylethylamine. An amine-reactive Cy3 fluorophore (1 mg) (Amersham Pharmacia) was added to this solution, and the resulting mixture was agitated for 4 h. Crude products were purified by preparative HPLC. The purity and identity of the product was confirmed by analytical HPLC and MALDI-TOF MS.

PA1. UV-Vis (H₂O) λ_max in nm (ε in M^–1·cm^–1): 313 (69,500), 555 (75,000). MALDI-TOF MS (monoisotopic) [M + H] 1,877.60 (1,877.81 calculated for C₉₁H₁₁₂N₂₄O₁₇S₂).

Binding Assay. Microarray slides were immersed in 1× PBS and placed in a 90°C water bath for 30 min to induce hairpin formation of the oligonucleotides. Slides were then transferred to a tube of nonstringent wash buffer (saline/sodium phosphate/EDTA buffer, pH 7.5/0.01% Tween 20) and scanned to check for low background (<200 intensity). Microarrays were scanned by using a ScanArray 5000 (GSI Lumonics, Billerica, MA), and the image files were extracted with genepix pro Version 3.0 (Axon Instruments, Foster City, CA).

Polyamide binding. Microarrays prepared as above were placed in the microarray hybridization chamber and washed twice with nonstringent wash buffer. Polyamide was diluted to 5 nM in Hyb buffer (100 mM Mes/1 M NaCl/20 mM EDTA, pH 7.5/0.01% Tween 20). Polyamide (5 nM) was then added to the hybridization chamber and incubated at room temperature overnight for 16 h. Finally, the microarrays were washed twice with nonstringent wash buffer and scanned.

Protein binding. The microarrays were washed with reaction buffer containing 150 mM potassium glutamate, 50 mM Hepes (pH 7.5), and 5% glycerol for 5 min. Cy3-labeled Exd (extradenticle) protein and polyamide were diluted in reaction buffer to a final concentration of 20 nM and 50 nM, respectively. This solution was added to the hybridization chamber and incubated for 30 min. Subsequently, the microarrays were washed with reaction buffer and scanned.

Data Processing. For each replicate, global mean normalization was used to ensure the mean intensity of each microarray was the same. Local mean normalization (51) was then used to ensure that the intensity was evenly distributed throughout each sector of the microarray surface. Outliers between replicate features were detected by using the Q test at 90% confidence and filtered out. The replicates were then quantile-normalized (52) to account for any possible nonlinearity between arrays. Duplicate features were then averaged together. The median of the averaged features was subtracted to account for background.

Z scores were calculated as |signal – median|/standard deviation. Because of the right-handed tail effect, standard deviation of the background signal was on the basis of the standard deviation from the median of all signals less than the median. The relationship of Z scores to P values can be found in Table 2, which is published as supporting information on the PNAS web site. Motifs were then found by running several motif-finding algorithms (31–33) on sequences in the highest Z score bin. Logos (53) of each motif were then created by using sequences from the highest Z score bin that contained the motif.

Molecular Modeling. Molecular models were created by aligning the coordinates of Exd crystallized with DNA (Protein Data Bank ID code 1B8I) with DNA of hairpin polyamide crystallized with DNA (Protein Data Bank ID code 1M18) (54). The DNA was aligned at four different positions by using structural alignment software (46), creating consensus, consensus +1, consensus –1, and inverse binding of the polyamide relative to Exd. The distance from the N-methyl group of the heterocycle ring, which is analogous to the N-methyl group of the ring of our polyamide that bears the hexapeptide, to the carboxyl carbon of the methionine of the recruitment peptide bound to Exd in the crystal structure was then calculated for each of the four alignments. This calculation demonstrated the distance that the linker in our polyamide (PA2 and PA3) would have to reach to recruit Exd to DNA (46). The alignments were visualized by using vmd (Visual Molecular Dynamics) software (47). The linkers for PA2 and PA3 were then drawn and energy was minimized to estimate how far each linker could likely reach (48).

Dye Conjugation to Exd. A pET3A vector containing the Exd sequence (residues 1–88) was mutated by using standard quick-change mutagenesis procedures to replace the cysteine with a serine (C41S), and an arginine was replaced with a cysteine (R2C) to generate Exd R2C (see Fig. 9A and Supporting Methods, which are published as supporting information on the PNAS web site). This Exd mutant was found to be stable, and the mutation had a minimal effect on DNA binding affinity. Exd R2C was then labeled with Cy3 by using a Cy3 maleimide Mono-Reactive dye pack (no. PA23031; Amersham Pharmacia Biosciences). The molar dye/protein ratio was determined to be 0.96 (quantified as follows): [Cy3] = (A₅₅₂ × dilution factor)/150,000 M^–1·cm^–1; [Exd-R2C] = [A₂₈₀ – (0.08 × A₅₅₂)]/12,090 M^–1·cm^–1; Dye/protein = [Cy3]/[Exd-R2C]; Exd sequence: ¹A(R→C)RKRRNFSK ¹¹QASEILNEYF ²¹YSHLSNPYPS ³¹EEAKEELARK ⁴¹(C→S)GITVSQVSN ⁵¹WFGNKRIRYK ⁶¹KNI.

EMSAs. Forty-mer DNA sequences labeled with ³²P (as per standard methods) were used in all reactions. Reactions were performed in a buffer containing 150 mM potassium glutamate, 50 mM Hepes (pH 7.5), 2 mM DTT, 100 ng/μl BSA, 10% DMSO, and 10% glycerol. For the binding tests, polyamide-peptide conjugates (50 nM final concentration) and [³²P]DNA were incubated together for 30 min at 4°C. Exd was then added to bring the reaction volume to 20 μl. Exd final concentrations were 0.033, 0.1, 0.33, 1, 3.3, 10, 33, and 100 nM. These reactions were incubated at 4°C for 1 h, and 15 μl was loaded onto a prerun 10% acrylamide/3% glycerol gel (1× TBE: 90 mM Tris/64.6 mM boric acid/2.5 mM EDTA, pH 8.3).

Supplementary Material