The specificity landscape of WRKY transcription factors reveals the bidirectional influence of non-CG methylation

Abstract

A distinct epigenetic feature of plants is the DNA methylation in non-CG contexts. Although the physiological roles of non-CG methylation have been elucidated, its direct impact on transcription factor (TF)–DNA interactions remains largely unexplored. Focusing on WRKY-family TFs, here we investigated how non-CG methylation influences their DNA binding specificity and genome-wide cis-regulatory elements (CREs). By generating 461 SELEX and DAP-seq libraries for 54 AtWRKYs, we show that DNA methylation alters both monomeric and dimeric binding specificities of WRKYs, leading to an overall increase in specificity divergence among family members. We curated 201 WRKY motifs and clustered them into 11 classes, 5 of which represent previously unreported specificities. Notably, the known WRKY cis-element PRE4 was found to be recognized only when methylated. The comprehensive dataset of accurate WRKY motifs also enabled the identification of the amino acid discriminants of W-box and WT-box. Expanding on prior knowledge, we demonstrate that methylation not only decreases but can also increase the affinity of WRKYs. This bidirectional effect has globally reshaped the genomic binding landscape of WRKYs upon methylation. Finally, we constructed the WRKY Regulatory Code Database (https://transysbio.cn/WRKYRCDB.php) to facilitate data access.

Graphical Abstract

Graphical Abstract.

Introduction

DNA methylation (5-methylcytosine, 5mC) is a fundamental epigenetic modification that regulates plant growth, development, and adaptation [1, 2]. Unlike in animals, where DNA methylation predominantly occurs in the CG context, plant genomes are also methylated in the CHG and CHH contexts [3, 4]. Non-CG methylation in plants is essential for silencing transposable elements (TEs) [1, 5] and also contributes to transcriptional regulation. In maize, many genes are flanked by regions of elevated CHH methylation and reduced CG/CHG methylation [6, 7]. Transcription factors (TFs) act as transducers that interpret CG methylation states into transcriptional outputs of the target genes [8]. This role of TFs also extends to non-CG methylation. In tomatoes, for example, the MADS-box ripening regulator RIN binds DNA in a methylation-sensitive manner [9], despite the fact that RIN recognizes cytosines in the CHH context [10]. Given that almost all TFs recognize sequences containing cytosines, non-CG methylation in plants is expected to fine-tune DNA binding for a broad spectrum of TFs.

To interrogate how DNA methylation impacts the binding of plant TFs, previous studies have compared TF affinities at methylated and unmethylated genomic cis-regulatory elements (CREs) [11–13]. The CRE sites of plant TFs were identified through DNA affinity purification sequencing (DAP-seq) or chromatin immunoprecipitation sequencing (ChIP-seq), and their methylation states were assessed using whole-genome bisulfite sequencing (WGBS). These approaches have revealed an overall repressive role of DNA methylation in the binding affinity of plant TFs [11, 12]. However, due to the scarcity of methylated (particularly non-CG methylated) CRE sites, these approaches have limited capacity to accurately infer the methylated models—TF binding models in the context of DNA methylation.

Methylated models for >500 animal TFs have been derived using EpiSELEX-seq [14], methyl-SELEX [8], and Methyl-Spec-seq [15]. Instead of genomic DNA, these approaches utilize a randomized DNA library as input to increase the complexity, and employ the methyltransferase M.SssI to introduce saturated CG methylations. However, non-CG methylations are also common in plants, whereby 5mCs are equally partitioned into two types of mutually exclusive regions: mCG-only regions and mC-all regions [3, 4, 11]. To examine TF specificities also in the mC-all regions, recent efforts have developed a modified methyl-SELEX protocol that incorporates 5mCs into all contexts by using 5-methyl-CTP in PCR [16, 17]. Similarly, this strategy has been extended to develop methyl-ampDAP-seq, which resolves genome-wide CREs of a TF in a fully methylated context and provides a “reference map” that complements ampDAP-seq [17]. Here, we establish high-throughput workflows of these methods and systematically examine how the mC-all context can modify the specificity and genome-wide CREs of the WRKY-family TFs.

WRKY-family TFs are named after the conserved WRKYGQK sequence within their DNA-binding domain (DBD), and represent one of the largest TF families in higher plants [18, 19] (see also Supplementary Fig. S1A). WRKYs are well known for their roles in plant immunity and stress responses [20, 21]. This includes responses to pathogens [22, 23], abiotic stresses [24], and stress-associated phytohormones [25]. Additionally, WRKYs are integral components of various signaling pathways [22, 26], and have been increasingly implicated in the regulation of plant development [27–29] and secondary metabolism [22]. WRKYs share a highly conserved specificity [30] and primarily recognize non-CG cytosines [11, 12, 31], making them ideal candidates for studying the effects of non-CG methylation.

WRKY–DNA interactions have been explored without DNA methylation or with sparse natural methylation. The genome-wide CREs of WRKYs were mapped using DAP-seq [11, 32–35] and ChIP-seq [36, 37], and the DNA-binding specificities of WRKYs were characterized using protein binding micrarray (PBM) [38, 39], KaScape [40, 41], and DNA–protein interaction enzyme-linked immunosorbent assay (DPI-ELISA) [42]. By combinatorial analyses of WRKY occupancy and methylation at genomic CREs, it was shown that cytosine methylation at all examined positions of WRKY CREs has decreased the affinity of WRKYs [12]. However, this conclusion is dependent on 10–20 CREs with high methylation rates, and assuming that WRKY motifs remain unchanged upon methylation.

To systematically investigate how non-CG methylation interferes with WRKY–DNA interactions, and to provide a high-quality, comprehensive dataset for WRKYs of Arabidopsis thaliana, this study employs high-throughput approaches based on SELEX and DAP-seq and generates 461 libraries for 54 AtWRKYs. Data analyses have revealed five classes of novel DNA-binding specificities of WRKYs, and demonstrated that instead of being monotonically repressive, DNA methylation can bidirectionally influence WRKY’s affinity. Additionally, the WRKY Regulatory Code Database (https://transysbio.cn/WRKYRCDB.php) was developed to facilitate data access.

Materials and methods

Plant materials and growth conditions

Seeds of A. thaliana used in this study were from the Columbia ecotype (Col-0). Seeds were stratified for 2–3 days at 4°C and grown in half-strength Murashige and Skoog (1/2 MS) (2.22 g/l MS medium, 10 g/l sucrose, 0.5 g/l MES, 8 g/l agar, pH 5.9) for ~1 week, after which the seedlings were transplanted into charcoal soil mixed with vermiculite in a 3:1 ratio in a greenhouse at 22°C day/night temperatures under long-day conditions (16 h light/8 h dark). Cotyledons and roots were separated from 7-day-old seedlings. Flowers were collected from 5-week-old plants. Siliques were collected from 8-week-old plants.

TF cloning and protein expression

The open reading frames (ORFs) of 54 full-length Arabidopsis WRKYs were cloned from the reported library [43] (Supplementary Table S1). The ORFs were recombined using LR Clonase (ThermoFisher, 11 791 100) into the pIX-HALO expression vector. The DNA-binding domain (DBD) and mutated DBD sequences of WRKY45/50/51/55 and the truncations of WRKY4 were cloned (Supplementary Table S2) and inserted into the pIX-HALO vector with the In-Fusion Kit (Vazyme, C112). These HALO-tagged WRKYs were expressed using the TnT® SP6 High-Yield Wheat Germ Protein Expression System (Promega, L3261) following the manufacturer’s instructions: ∼2 μg of plasmid DNA was added into a 50 μl reaction and incubated for 2 h at 25°C. The expressed proteins were then used in subsequent SELEX- and DAP-seq-based assays.

High-throughput SELEX and methyl-SELEX

High-throughput SELEX (HT-SELEX) was carried out as previously described [44, 45] with minor modifications. The SELEX ligands contain a 101 bp randomized region flanked by adaptors designed according to the TruSeq Illumina library (Supplementary Table S3). Briefly, 2 µl of HALO-tagged WRKY protein and 150–200 ng of SELEX ligand were mixed in 30 µl of TCAPT buffer (140 mM KCl, 5 mM NaCl, 1 mM MgCl₂, 3 µM ZnSO₄, 100 µM EGTA, 10 mM Tris, pH 8, 0.1% Tween). Next, 2 µl of Magne HaloTag Beads (Promega, G7282) were added and incubated for 1 h at 25°C. Beads were then washed by a HydroSpeed plate washer (Tecan, 30 190 101) and resuspended with 30 µl of elution buffer (0.1% Tween, 10 mM Tris pH 7.8, 1 mM MgCl₂), and then used as the template and PCR-amplified using 2 × Phanta Max Master Mix (Vazyme, P515) with the Amplification_primers (Supplementary Table S3). This process was repeated for a total of 3–5 cycles. The enriched libraries were further amplified with the indexed PE_PCR_primers (Supplementary Table S3), purified with VAHTS DNA Clean Beads (Vazyme, N411), and sequenced on Illumina NOVA Xplus.

In methyl-SELEX, PCR amplifications of the input library and between cycles were performed using the Taq DNA polymerase (Vazyme, P101) and with 5-methyl-dCTP (NEB, N0356S).

DAP-seq-based assays

Similar to as previously described [11], genomic DNA (gDNA) was extracted from Arabidopsis tissues using the Hi-DNAsecure Plant kit (TIANGEN, GDP350) and then fragmented with a Covaris M220 Sonicator into 150–200 bp fragments. The DNA ends were repaired and A-tailed with the End Preparation Module (Vazyme, N203), and subsequently ligated to the truncated Illumina Y-adapter (Supplementary Table S3) using the Adapter Ligation Module (Vazyme, N204). Next, 100–200 ng of gDNA ligands were incubated with 5 µl of HALO-tagged WRKY protein and 2 µl of Magne HaloTag Beads (Promega, G7282). After washing by the HydroSpeed washer (Tecan, 30190101), the resuspended bead solution was used as a template and amplified with the indexed PE_PCR_primers (Supplementary Table S3) using 2 × Phanta Max Master Mix (Vazyme, P515). The PCR products were purified with VAHTS DNA Clean Beads (Vazyme, N411), and sequenced on Illumina NOVA Xplus.

In ampDAP-seq and methyl-ampDAP-seq assays, the end-repaired and adaptor-ligated gDNA libraries were further PCR-amplified, respectively, with dCTP (Vazyme, P036) and 5-methyl-dCTP (NEB, N0356S).

Electrophoretic mobility shift assay (EMSA)

As larger amounts of proteins are required for EMSA, we used Escherichia coli cells to produce WRKY50 protein according to the previously described protocol [8]. The ORF was gateway-cloned into the pETG20A recipient vector and transformed into Rosetta 2 (DE3) pLysS strains. The transformed E. coli were grown in auto-induced ZYP5052 medium [44] at 37°C for 8 h and 17°C for 12 h. The bacteria were collected and lysed with lysis buffer [0.5 mg/ml lysozyme, 2 mg/ml DNase I, 1 mM phenylmethylsulfonyl fluoride (PMSF)], incubated with His-tag Ni Sepharose (Sangon, C600332) at 25°C for 1 h, and washed with a gradient of imidazole (10, 50, and 500 mM) in buffer A (50 mM Tris, 300 mM NaCl, pH 7.5) to purify the recombinant protein. The purified WRKY protein was then exchanged with the EMSA binding buffer (10 mM Tris, 50 mM KCl, pH 7.5) and diluted. Different concentrations of WRKY50 (57, 114, and 228 nM in the reaction) were incubated with 20 ng of DNA probes (Supplementary Table S2) in 20 µl of EMSA binding buffer at 25°C for 1 h. The reaction mixture was loaded onto a 6% native TBE gel and run (100 V) in 0.5× TBE buffer at 4°C for 100 min. The gel was stained with Gel Blue (UElandy, S2019L) and imaged with a Bio-Rad scanner (Bio-Rad, 1708195EDU). To prepare methylated DNA probes in EMSA, the original sequence was PCR-amplified using Taq DNA polymerase (Vazyme, P101) and with 5-methyl-dCTP (NEB, N0356S).

Data processing of SELEX-based assays

The raw sequencing data (paired-ended) were quality controlled using Fastp v.0.20.1 [46] to remove adapter sequences and low-quality reads with parameters: -q 10 -m -a –overlap_len_require = 5 -gx –length_required = 15 –n_base_limit = 10 -y –complexity_threshold = 5. Duplicate sequences from PCR amplification were also removed.

The SELEX reads were then cut into 40 bp fragments and used in motif discovery as described [47, 48]. Motif discovery utilizes Autoseed [47] with parameters: -40N <Background sequence> <Signal sequence> 1 8 10 0.35 - 50 100. The obtained motifs were curated and visualized using ggseqlogo [49].

To assess the quality of SELEX data, we employed Kullback–Leibler (KL) divergence as the measure. For all 6-mers with 0–10 bp gaps in the middle, we calculated the total KL divergence that quantifies the difference between the observed and expected 6-mer frequencies:

Where represents the expected frequency of a 6-mer (product of the frequencies of the two constituent 3-mers), and is the observed actual frequency of the 6-mer. Because the prevalence of TF-binding sequences will bias the 6-mer distribution, a high KL divergence is indicative of TF signals. Enriched-sequence-based mutual information (E-MI) was calculated and visualized as described before [50].

Data processing of DAP-seq-based assays

The raw sequencing data were quality controlled using Fastp v.0.20.1 [46] to remove adapter sequences and low-quality reads with parameters: -q 30 -y -Y 60 -W 5 -M 30 -5 -3 -l 50. The trimmed reads were mapped to the TAIR10 genome using BWA v.0.7.17 with default parameters. The mapped reads were further filtered for MAPQ > 30 using samtools v.1.9 [51] to reduce multiple mapping. The Bigwig files were created using bamCoverage in deeptools v3.5.1 [52] with the following parameters: –binSize 10 –normalizeUsing CPM. To define CREs from the DAP-seq-based libraries, peak calling was performed using MACS3 v.3.0.0a7 [53] with parameters: -f BAMPE –d-min 5 –min-length 15 –call-summits –keep-dup all –cutoff-analysis, using the mock sample (no WRKY protein) for background subtraction. A blacklist of peak regions (peaks in input and mock samples) was further excluded from consideration. The filtered peaks were then used in subsequent analyses, and calculated for the fraction of reads in peaks (FRiP) as the quality indicator.

Motif discovery of DAP-seq-based libraries also uses Autoseed [47] and follows the same commands as in SELEX, using the sequences extracted from the top 600 peaks. The 200 bp wide regions centered on the peaks were cut into 40 bp fragments, and then used as input for Autoseed.

Data processing of ATI, ATAC-seq, and RNA-seq

To examine whether the identified WRKY specificities are involved in transcriptional regulation, we collected the published Sea-ATI (sequential extraction assisted-active TF identification) and ATAC-seq (assay for transposase-accessible chromatin using sequencing) libraries [31] and generated the RNA-seq libraries of each tissue (constructed and sequenced by Novogene, Shanghai, China). The ATI reads are processed in the same way as described above for SELEX. ATAC-seq reads were first trimmed for adaptor sequences using Trim Galore with parameters: -q 30 –paired –stringency 5 –fastqc –gzip. The mapping and peak calling steps are the same as described above for DAP-seq. RNA-seq reads were mapped to the TAIR10 genome using HISAT2 [54]. Gene expression levels were quantified with featureCounts and normalized to transcripts per kilobase million (TPM). Genes with expression levels of the top and bottom 15% were identified based on TPM values.

Data processing of WGBS

For each Arabidopsis tissue, a 200 mg sample was collected and sent to Anoroad Genome (Beijing, China) for library construction and sequencing. The raw reads were quality-controlled and filtered with Trim Galore with parameters: -q 30 –paired –stringency 5 –fastqc –gzip. The remaining reads were mapped to the bisulfite-converted reference genome by bismark with parameters: bismark –bowtie2 -p 20 –bam –score_min L,0,-0.2 [55]. The reference genome was translated into a bisulfite-converted version (C to T and G to A) by bismark_genome_preparation. The mapped reads were then deduplicated by deduplicate_bismark and sorted by samtools v1.9 for further analysis. Methylated cytosines were identified, extracted, and counted by bismark_methylation_extractor. Undersampled cytosine sites with a < 5 coverage were excluded. The methylation rate of each cytosine was calculated as the ratio of methylated cytosines to the total coverage at that site. For visualization, the per-base methylation probabilities were further converted into the BigWig format.

Motif classification and enrichment analyses

The three most enriched monomeric motifs (primary, secondary, and tertiary, Supplementary Table S4) were curated from SELEX and methyl-SELEX libraries. These motif position frequency matrixes (PFMs) were then aligned, resized into a uniform width of 13 bp, and flattened into one-dimensional arrays for clustering. Spearman’s rank correlation coefficients were calculated between the PFMs to obtain the distance matrix and, based on the distances (distance measure: maximum), a hierarchical clustering tree was built with the complete linkage. By cutting the clustering tree at a height of 0.37, all monomeric motifs were categorized into 11 classes. To derive the representative motif for each class, we first calculated the arithmetic means respectively for high-information content (IC) motifs (mean IC ≥ 0.5) and low-IC motifs (mean IC < 0.5) in the class. The two obtained PFMs were combined into a weighted average (high-IC, 0.8; low-IC, 0.2) and used as the representative motif (PFMs in Supplementary Table S5). More weights are placed on the high-IC motifs because they are derived from SELEX-based libraries with stronger TF signals and better reflect the intrinsic specificity of the TFs.

The representative motifs were then used in enrichment analyses, which were performed with the R package motifmatchr [56] with P= 1e-5. Specifically, a motif is matched both to the SELEX-based library and the shuffled SELEX-based library. The numbers of motif hits were then compared to derive the enrichment ratio:

For each SELEX-based library, the enrichments of all representative motifs were then divided by the maximum enrichment for normalization. This facilitates comparison between different WRKYs because the signal strength (thus the absolute enrichments) can vary substantially across SELEX libraries. The enrichments of WRKY consensus sequences were calculated by directly counting their occurrences in the original and shuffled SELEX libraries, allowing 0 (W-, WT-, WK- boxes) or 3 (Control, PRE4, and SURE) mismatches, and then also normalized by dividing the maximum enrichment of all consensuses. For WRKY18, the enrichment of AAGTTTTC was used as the maximum enrichment in normalization.

The enrichments of motifs around transcription start sites (TSSs) and in various CREs (defined by ATI, ATAC-seq, and RNA-seq) were calculated as previously described [31]. Briefly, TSS enrichments visualize motif densities ±5 kb around the TSSs: motif hits at each position were divided by the total hits in the 10 kb region, and then a rolling average (20 bp window) was applied. The ATI CRE enrichment is the ratio of motif hits in the original and shuffled ATI libraries. The ATAC-seq CRE enrichment is the ratio of motif densities in the open (THS) and closed (non-THS) chromatin regions. The RNA-seq CRE enrichment refers to the ratio of motif densities in the promoters (±300 bp around the TSSs) of high-expression (top 15%) and low-expression (bottom 15%) genes. ComplexHeatmap [57] was used to visualize the enrichments.

Enrichment of dimeric configurations was analyzed by counting the occurrences of concatenated half-site strings (GTMAA), with the spacings ranging from 0 to 30 bp. The two concatenated GTMAA half-sites can assume any of the three relative orientations (DR, IR, or ER). The occurrences of the concatenated strings in the original library (either SELEX-based or DAP-based) and in the shuffled library were then counted. The ratio between the counts is defined as enrichment for each dimeric configuration. For each library, the enrichments of all dimeric configurations were then divided by the maximum enrichment to normalize. This facilitates comparison between different WRKYs because the signal strength can vary substantially across libraries. Sequences from the SELEX-based library or within the peaks of DAP-seq (200 bp at the center of the top 600 peaks) were used in the analyses.

Structural modeling and visualization

AlphaFold 3 [58] was employed to build the structural model of the WRKY50–DNA complex. The DBD sequence of WRKY50 (Supplementary Table S2) and the consensus of its WT-box motif (AAAAGTCAA) were used. Visualization of the structural model was achieved with PyMOL (https://pymol.org/). The contacts between WRKY50 and DNA were visualized with DNAproDB [59].

The bidirectional affinity effects of methylation

To examine the positionally dependent effects of cytosine methylation on the affinity of WRKYs (Fig. 5E), the WT-box motifs were first derived for the paired SELEX and methyl-SELEX libraries of each WRKY by using AAGTCAAC as the seed [47]. Next, the PFMs of the methylated and the unmethylated motifs were compared, to calculate the probability change and fold change of C at individual positions. To examine how WRKY35’s occupancy on CREs changes with the identity of the position 4 nucleotide in the WT-box motif (Fig. 5G), we substituted the fourth position of the WT-box motif into an even distribution (P= 0.25 for each nucleotide), and then used the substituted motif to locate CREs in the genome (using motifmatchr, P= 1e-5). The resulting CREs are then separated according to their nucleotide identity at the fourth position, and calculated for the occupancy of WRKY35 using ampDAP-seq and methyl-ampDAP-seq data. To profile the interpositional dependency of dinucleotides (Fig. 5H), the expected dinucleotide frequency (product of nucleotide frequencies at the two constituent positions) is subtracted from the actual dinucleotide frequency. For a pair of positions, only the largest deviation among all 16 dinucleotides is visualized in the heatmap.

Figure 5.

The bidirectional effects of methylation on the affinity of WRKYs. (A and B) Binding-mode-dependent bidirectional effects. (A) Bins with C/G-containing 8-mers are colored blue. The 8-mers containing only A/T bases (orange bins, medium value represented by the vertical line) whose affinities are not changed by methylation were used as a reference. Seeds are indicated on top of each motif. (B) Genomic CREs of m1 and m2 modes show higher occupancies when methylated, while CREs of the s1 mode show the reverse. (C–E) Position and TF-dependent bidirectional effects. (C) While cytosines at C4 always decrease in affinity upon methylation, methylation on C7 cytosines can both increase and decrease the affinity depending on WRKY identity. (D) Enrichments of 8-mers containing GTCAAC (blue line) and GTCAAA (green line) in methyl-SELEX and SELEX were compared. Note that the relative ordering of the lines is reversed between WRKY29 and WRKY65. (E) The methylation effects on cytosines of all positions of the WT-box motif were examined. WRKYs are sorted by the affinity effect of C7 methylation. (F–H) Bidirectional effects on the affinity of A/T. (F) Methylation changed the specificity at A/T positions of the WRKY40 motif (left, GTCAA to GTCTT). The A/T ratios at individual positions are also different for the SELEX and methyl-SELEX WT-box motifs of WRKY35 (right). (G) Occupancy at CREs of WRKY35’s WT-box motif (with nucleotide variations at position 4). Upon methylation, the occupancy at GTAAA sites has significantly increased, while it has only slightly decreased at GTCAA sites. (H) Dinucleotide dependency of WRKY70 increases upon methylation. Pixels in the heatmap are colored by the largest dinucleotide frequency deviation from the independent assumption. Magenta: dinucleotide frequency > estimated. Green: dinucleotide frequency < estimated. The colors of the two halves of the circle represent the bases in the most deviated dinucleotide.

Prediction and the precision–recall curve

The specificity of WRKYs was used for the prediction of their genome-wide CREs through a binary classification. First, from a DAP-seq-based library, we extracted 200 bp DNA sequences from the center of the top 1000 peaks; these sequences represent CRE regions (real positives, RPs). The genomic sequences outside of the DAP peaks represent non-CRE regions (real negatives, RNs); they were also cut into non-overlapping 200 bp segments. The 200 bp RP and RN sequences were then matched with the top 1000 most enriched 8-mers (with a 0–8 bp gap in the center) from the SELEX-based library. The fold enrichments of all matched 8-mers were summed for each 200 bp sequence and defined as its score. To plot the precision–recall curve, the scores of RP and RN sequences were then compared with a series of thresholds. Sequences with scores exceeding the threshold are predicted as CRE regions (predicted positives, PPs), and sequences with scores lower than the threshold are predicted as non-CRE regions (predicted negatives, PNs). At each threshold, PP and PN were compared with RP and RN to calculate the precision [(RP ∩ PP)/PP] and recall [(RP∩PP)/RP] values, and then plotted as a curve.

Construction of the WRKY Regulatory Code Database (WRCD)

Similar to as previously described [60], the WRCD was built on the LNMP architecture (Linux, Nginx v1.20.1, MySQL v8.0.41, PHP v5.4.16) and deployed on the Tencent Cloud (Tencent, China), with front-end pages developed using HTML5, CSS3, JavaScript, and AJAX for asynchronous data communication with the back end. JBrowse genome browser (v.1.16.11) was integrated for visualization of the genomic tracks. SELEX- and DAP-seq-based libraries were processed as described above to derive motifs, peaks, and coverages. The ChIP-seq data [36] were analyzed as described for DAP-seq. The published PBM motifs were directly collected and visualized. Regulatory networks were constructed using ampDAP-seq and methyl-ampDAP-seq libraries of the same WRKY. Peaks located around TSSs (−1000 to +500 bp), with signal values > 10, and not overlapping the background peaks (in input and mock) were used in network construction. Visualization of the networks was achieved using network3D. The Gene Ontology (GO) enrichment tool was realized with R scripts.

Results

A comprehensive dataset for WRKYs

First, a comprehensive dataset was generated for AtWRKYs to explore their specificity and genome-wide CREs, in either the presence or absence of DNA methylation (5mC). The SELEX-based methods are preferable for building specificity models due to their complex input library (∼10¹⁴ binding sites). On the other hand, the DAP-seq-based methods directly use gDNA as the input, and are thus preferable in the task of locating genomic CREs. Due to the abundance of methylation contexts (CG/CHG/CHH) in plants, in methyl-SELEX all cytosines on the DNA ligands were substituted with 5mCs, by using 5-methylcytosine instead of cytosine in the PCR [16, 17]. Similarly, methyl-ampDAP-seq was performed to identify WRKY CREs upon methylation. In parallel, SELEX [44] and ampDAP-seq [11] were carried out to generate binding profiles in the absence of 5mC. DAP-seq is also performed using gDNA from Arabidopsis cotyledons to map WRKY CREs under the natural state of DNA methylation.

In total, 140 SELEX-based libraries and 321 DAP-seq-based libraries (two replicates each) were generated for 54 AtWRKYs, offering currently the most comprehensive dataset for WRKY–DNA interactions. The generated dataset is compared with previous PBM, DAP-seq, and ampDAP-seq datasets [11, 38, 39] for their coverages (Fig. 1A, SELEX, ampDAP, and DAP). This study not only offered independent repeats for most AtWRKYs examined before (yellow area in Fig. 1A), but also uniquely covered ∼1/3 of AtWRKYs (red area in Fig. 1A). The previous datasets are either without 5mC (PBM and ampDAP-seq) or with only sparse 5mCs (DAP-seq); therefore, the libraries with 100% 5mC are not paralleled by the published data (Fig. 1A, methyl-SELEX and methyl-ampDAP). The coverage of all WRKY subgroups has been considerably improved. Moreover, we report only the high-quality data by applying stringent QC criteria—WRKY motifs need to be de novo discovered for all SELEX- and DAP-seq-based libraries, while previous datasets may contain 50–76% of libraries without motifs [11, 33]. Other parameters, such as KL divergence, E-MI [31, 50], and FRiP [61], were also employed to confirm the data quality (Supplementary Fig. S1C–E).

Figure 1.

Overview of the datasets. (A) Data coverage of AtWRKY subgroups. All cytosines are methylated in the methyl-SELEX and methyl-ampDAP-seq libraries. DAP-seq uses gDNA with natural methylation. Published PBM data (hatched area) are accounted for in the statistics of “SELEX”. (B) Annotation of additional WRKY CREs. The overlaps of the generated and previous DAP-seq datasets are shown. (C and D) Methylation diversifies the sequence specificity of WRKYs. Specificity became less similar for WRKY pairs upon methylation, as shown by the decreased correlation of 8-mer enrichments (C). The correlation before and after methylation is exemplified in (D) for the WRKY35–WRKY33 pair.

The DAP-seq-based methods have identified ∼130 000 CREs for all AtWRKYs, among which ∼100 000 CREs were not annotated by the previous data (Fig. 1B). The genuineness of the newly annotated CREs is supported by the enrichment of the typical WRKY motif (Supplementary Fig. S1F). Compared with ampDAP-seq, methyl-ampDAP-seq has identified more non-overlapping CREs with DAP-seq (Fig. 1B, right bottom). Given that DAP-seq directly uses gDNA as input, this suggests that most WRKY CREs are not methylated in the natural genome.

In all five types of datasets (Fig. 1A), SELEX has covered the largest fraction of the WRKY TFs. We therefore compared the specificity of all WRKYs by performing principal component analysis (PCA) for the SELEX libraries (Supplementary Fig. S1B). WRKYs within the same subgroup tend to cluster (although loosely) in the PCA plot, suggesting that similarity in their protein sequences has led to similarity in their DNA sequence specificity. In addition, two replicates of the same WRKY agree well with each other (Supplementary Fig. S1B, the pairs of points connected by lines), supporting the experimental reproducibility (Supplementary Data S1A).

Methylation diversifies the sequence specificity of WRKYs

WRKY is among the TF families with the most conserved specificities [30]. In the absence of DNA methylation, the sequence specificities of WRKYs are highly similar both within and across the plant species [30, 38]. However, surprisingly, in the presence of 5mC, the specificity has considerably diverged between different WRKYs (Fig. 1C; Supplementary Fig. S1I, J). For example, when DNA is unmethylated (SELEX), both WRKY33 and WRKY35 bind with the highest affinity to the 8-mer AGTCAACG (Fig. 1D, left), and the affinities of all 8-mers in the two SELEX libraries are also highly correlated, indicating an overall similarity of their specificity. In contrast, the specificity of WRKY33 and WRKY35 has diverged upon DNA methylation (methyl-SELEX), as evidenced by the bifurcated distribution of the 8-mer points (Fig. 1D, right). Upon methylation, the favorite 8-mer for WRKY33 is AAGTATGC, whereas the favorite 8-mer for WRKY35 is AAGTnAACC. Using these two 8-mers as the seeds, distinct motifs were derived for WRKY33 and WRKY35 (Fig. 1D, right). The specificity divergence upon DNA methylation is also confirmed by analyzing the genomic CREs: different WRKYs share more common CREs in ampDAP-seq than in methyl-ampDAP-seq (Supplementary Fig. S1G). Therefore, DNA methylation has enhanced the ability to distinguish WRKYs and helped resolve the “specificity paradox” of eukaryotic TFs—how similar TFs regulate different genes and execute distinct physiological functions [48, 62]. The methylation-induced specificity divergence is consistent with the observations that TF paralogs vary in their methylation sensitivity [8, 14]. In contrast to most WRKYs, the specificity for a few pairs of WRKYs (e.g. WRKY40–WRKY25) became more similar after DNA methylation (Fig. 1C).

Methylation changes the monomeric specificity of WRKYs

The monomeric motif models (PFMs in Supplementary Table S4) were de novo discovered from 101 SELEX libraries and 39 methyl-SELEX libraries. Altogether, 201 monomeric motifs were curated (Fig. 2). The SELEX and methyl-SELEX motifs, respectively, reveal the specificity of WRKYs in the absence and presence of DNA methylation. For comparison, monomeric motifs were also curated from DAP-seq, ampDAP-seq, and methyl-ampDAP-seq libraries (Fig. 2, right column). The input library of SELEX features a higher complexity and a lower sequence bias; accordingly, the SELEX motifs cover a larger specificity space compared with DAP-seq and ampDAP-seq motifs (Supplementary Fig. S1H). This can be explained by the fact that while SELEX motifs reflect only the biochemical specificity, the DAP-seq-based motifs are additionally affected by evolutionary drifts that deviate the abundance of CRE sequences from thermodynamic equilibrium [31]. The direction of deviation is the same for all WRKYs, thereby diminishing their specificity differences.

Figure 2.

Overview of WRKYs’ monomeric motifs. The de novo monomeric motifs were discovered from SELEX- and DAP-seq-based libraries. For WRKYs not covered by this study, the published PBM and DAP-seq motifs are presented (black boxes). See Supplementary Table S4 for motif matrices.

A single TF frequently exhibits multiple binding modes [63]. This is illustrated by the SELEX-based methods that systematically discover all binding modes for a TF, yielding not only the strongest (primary) motif but also those with relatively weaker affinities (secondary and tertiary motifs, Fig. 2). Through examining individual regulatory sequences, previous studies have identified five types of WRKY CREs: W-box (A/GGTCAA/G) and WT-box (AAAGTC) are recognized by most WRKYs, and a few specific WRKYs recognize WK-box (TTTTCCAC), PRE4 (TACTGCGCTTAGT), and SURE (TAAAGATTACTAATAGGAA) [64]. Here, the specificity profiles suggest that many WRKYs recognize more than one type of the reported CREs. For example, WRKY3 and WRKY44 bind to both the W-box and the WT-box motifs; WRKY50 and WRKY51 bind to both the WT-box and the WK-box motifs (Fig. 2, left column).

DNA methylation has changed the specificity of WRKYs, especially for the secondary and tertiary motifs (Fig. 2, middle column). WRKYs recognizing similar motifs in the absence of 5mC can bind to distinct motifs in the presence of 5mC (e.g. WRKY33 and WRKY35). This phenomenon is most obvious for WRKY18, while in SELEX it recognizes similar motifs (WT-box and W-box) as most other WRKYs, in methyl-SELEX the primary motif of WRKY18 is distinct from all other WRKYs. These observations are consistent with the overall specificity diversification upon methylation (Fig. 1C), suggesting that DNA methylation can evoke latent specificities of WRKYs.

Identification of novel specificities of WRKYs

We next clustered the monomeric WRKY motifs into 11 classes and derived the representative motif for each class (Fig. 3A; PFMs in Supplementary Table S5). The representative motifs of classes 1–5 are more enriched in the SELEX libraries, while those of classes 6–11 are more enriched in the methyl-SELEX libraries (Fig. 3B). Therefore, classes 1–5 represent WRKYs’ specificity in the absence of methylation, and classes 6–11 represent WRKYs’ specificity in the presence of methylation. The identified specificity classes also have different prevalence—classes 1, 2, 3, 9, and 10 are shared by many WRKYs, whereas classes 4, 5, 6, 7, 8, and 11 are specific for only a few WRKYs.

Figure 3.

The major classes of WRKY specificities. (A) The 11 classes of monomeric WRKY motifs. Monomeric motifs were clustered into 11 classes, with the representative motif of each class shown to the right. Classes of as yet unreported WRKY CREs are in bold and named with WS boxes (unmethylated specificities) and WmS boxes (methylated specificities). Two subclusters of class 9 (WmS9a and WmS9b) are also annotated because they represent strong binding modes with unreported specificities. Individual motifs are represented by colored rectangles showing their consensuses. Motif names are concatenations of SELEX Type, WRKY Identify, and Motif Rank. For example, m40_3 means the tertiary (third) motif of WRKY40 in methyl-SELEX. (B) Enrichments of representative motifs and CRE sequences in SELEX libraries. The representative motifs were derived for each class. Enrichment analyses were performed for each WRKY (heatmap) and summarized for all WRKYs (barplot). (C) WRKY50 recognizes both WT-box and WK-box CREs. EMSAs were performed with 0, 57, 114, and 228 nM (from left to right) of WRKY50. (D) The coverages of ampDAP-seq and methyl-ampDAP-seq are shown around genomic sites with 0–3 mismatches to the PRE4 consensus. (E) Alignment of the secondary and tertiary motifs of WRKY4 to the PRE4 sequence shows the potential recognition of PRE4 by WRKY dimers. (F) Enrichment of representative motifs around the TSSs. (G) Enrichment of representative motifs in known CREs. The enrichments of WRKY motifs were evaluated for active CREs that bind to TFs (ATI versus shuffled), within open chromatin (ATAC-seq peaks versus outside peaks), and promote gene expression (RNA-seq, promoters of the top 15% genes versus promoters of the bottom 15% genes). (H) Sequence features of WRKYs with W-box and WT-box specificities. The difference between W-box and WT-box motifs is shaded gray. The different high-IC amino acid positions that change directionally from W-box to WT-box are shaded blue. (I) Toggling between W-box and WT-box specificities by mutations. Mutating the four positions (H, blue shaded) has allowed WT-box WRKYs to recognize W-box (left), and vice versa (right). SELEX was performed with DBDs and their mutations; the primary motifs are shown. (J and K) Contacts between WRKY50 and WT-box consensus. The binding model is generated with AlphaFold 3 [58]. All contacts are visualized with DNAproDB [59] in (J). Contacts with the identified positions (15 and 16 in H) are colored red in (J) and enlarged in (K).

Among the 11 specificity classes, 6 of them correspond well with the reported WRKY CREs (classes labeled with non-bold text, Fig. 3A). W-box and WT-box are the dominant types of WRKY CREs; their specificities are respectively captured by class 1 and classes 2–3. Interestingly, after DNA methylation, only the WT-box but not the W-box type specificity is observed (classes 9–10, Fig. 3A). The WK-box type specificity is captured by class 5, which contains WRKY50 and WRKY51. Previous findings have suggested that the WK-box is recognized by WRKYs featuring a WRKYGKK sequence instead of WRKYGQK (most WRKYs) [65]. In agreement, both WRKY50 and WRKY51 contain WRKYGKK in their DBDs. Although without a SELEX library, the other WRKY that contains WRKYGKK, WRKY59, has also enriched the WK-box motif in its ampDAP-seq peaks (Supplementary Fig. S2). However, whereas it was proposed that WRKYGKK exclusively binds to the WK-box [65], we found that WRKY50 and WRKY51 still bind to WT-box motifs with high affinities (Figs 2 and 3B). This is also validated by the EMSA experiments (Fig. 3C).

The remaining five classes (classes 4, 6, 7, 8, and 11) represent novel WRKY specificities that were not previously reported (not found in JASPAR [66], PlantPAN [67], PlantTFDB [68], or MethMotif [13]). According to whether these classes are over-represented in SELEX or methyl-SELEX, they are termed WS-boxes and WmS-boxes (Fig. 3A, classes labeled with bold text). We identified four WmS-boxes but only one WS-box, despite the fact that far fewer methyl-SELEX (22 WRKYs) libraries are available than SELEX (53 WRKYs). This suggests that in the presence of DNA methylation, the specificity landscape of WRKYs is largely unexplored. The WS4-box (TTTTCAAC) resembles the WK-box (TTTTCCAC) but is recognized by typical WRKYGQK WRKYs, such as WRKYs 12, 13, 9, 31, and 42 (Fig. 2). The WmS6-box and WmS7-box are recognized by WRKYs 40, 18, and 55 (Fig. 3A, B). The WmS8-box is recognized by WRKY4 and WRKY33. The WmS11-box is recognized mainly by WRKY28, 71, 65, and 8. Note that although with fewer base changes, two subclasses of class 9 (WmS9a-box and WmS9b-box, Fig. 3A) also represent distinct specificities upon methylation. They are discussed in particular because they represent high-affinity binding modes (Fig. 2, middle column). For example, the WmS9a-box is the primary motif for WRKYs 4, 33, 25, and 32.

We also examined the enrichment of the consensus sequences of the reported WRKY CREs (W-box, WT-box, WK-box, PRE4, and SURE). Consistent with the motif analyses, the consensuses of the W-box and WT-box are enriched for most WRKYs (Fig. 3B, right panel); the WK-box consensus is enriched for WRKY50 and WRKY51. The consensus of SURE and a random control sequence (C) showed no enrichment. Interestingly, the PRE4 consensus bound by OsWRKY13 [69] is enriched in the methyl-SELEX but not the SELEX libraries (Fig. 3B, right panel), indicating that PRE4 is bound by WRKYs only upon DNA methylation. In agreement, at genomic sites with PRE4-like sequences, the occupancy of most WRKYs is higher in methyl-ampDAP-seq libraries than in ampDAP-seq libraries (Fig. 3D; Supplementary Fig. S3). OsWRKY13 binds PRE4 upon bacterial infection (Xanthomonas oryzae, Xoo) but not under other conditions [70]. Consistently, PRE4 is not methylated in uninfected rice tissues (Supplementary Fig. S3B). It would be of interest to further explore PRE4’s methylation state after Xoo inoculation and its physiological roles with methylation engineering. Alignment analysis suggests that PRE4 could have been recognized by WRKY dimers; for example, the secondary and tertiary motifs of WRKY4 (m4_2 and m4_3, respectively, corresponding to classes 9a and 8) all align with half of the PRE4 sequence (Fig. 3E).

We next examined whether the identified WRKY specificities are enriched in known cis-regulatory sequences. Functional CREs are typically enriched around TSSs. Although to a different extent, the representative motifs of the 11 classes are all enriched around the TSSs (Fig. 3F). ATI incubates random DNA ligands with the nuclear extract (containing TFs) and identifies CREs of all active TFs in the target tissue [31, 71]. Because ummethylated DNA ligands were utilized, the ATI libraries mainly enriched the unmethylated motifs of classes 1–5 (Fig. 3G, left). The methylated motifs of classes 9 and 10 are also detected due to their similarity to the unmethylated class 1 motif. The methylated motifs (classes 6–11) are not detected in the open chromatin (Fig. 3G, middle). This is probably because the open chromatin regions are associated with a low level of DNA methylation [72]. The methylated specificities (classes 6–11) are enriched in the promoters of actively expressed genes across all tissues (Fig. 3G, right). Given their absence in the open chromatin (Fig. 3G, middle), we hypothesize that the active promoters harboring methylated WRKY CREs may reside in chromatin regions with a lower accessibility.

The abundant and accurate monomeric motifs from SELEX have enabled the identification of the amino acid determinants of WRKYs’ monomeric specificity. The W-box and WT-box are the two representative binding modes of WRKYs (Fig. 3A, classes 1 and 2). The difference between them primarily resides in the IC of the 5′ A-stretch of the motifs (Fig. 3H, left, gray shaded). We classified all WRKYs into three groups according to the IC of the A-stretch in their motifs, and derived PWM models for the protein sequences of each group. Near the conserved WRKYGQK sequence, we observed four high-IC positions (15, 16, 22, and 25) that directionally changed from the W-box to the WT-box group (Fig. 3H, right, blue shaded). Mutations at these four positions were able to toggle between the W-box and the WT-box specificities (Fig. 3I). For example, the WT-box WRKYs 45, 50, and 51 were switched to recognize the W-box, and the reverse is achieved for WRKY55. Congruent with the mutational study, AlphaFold modeling has suggested that residues at positions 15 and 16 are in contact with the A-stretch when WRKY50 binds to the WT-box (Fig. 3J, K). These contacts require further structural validations. In addition to positions 15, 16, 22, and 25, the residues at positions 17, 24, and 27 also changed between the W-box and the intermediate groups (Fig. 3H), and may have contributed to WRKY’s recognition of the A-stretch.

Methylation changes the dimeric specificity of WRKYs

In contrast to the typical binding strength of a TF (K_d ≈ 10 nM), the dissociation constant of WRKYs from their monomeric sites is 600–700 nM [12]. The stable binding of WRKYs thus may require their dimerization. We analyzed the SELEX libraries to derive the configurational preferences for WRKY homodimers. The calculated SELEX enrichments of dimeric configurations are generally reliable, especially for the most preferred configurations of each WRKY (Supplementary Data S1B, D). While WRKYs share highly similar monomeric motifs (Fig. 2, left column), their dimeric configurations are diverse (Supplementary Fig. S4A; Supplementary Data S1B–D). Therefore, the spacing and orientation of dimeric WRKY CREs can determine to which WRKY they bind, and in turn dictate the spatiotemporal expression of the target genes. This is similar to the previous findings for ARFs [73–77], MYBs [48], and animal TFs [44], emphasizing the importance of the dimeric configuration of CREs in transcriptional control. Most WRKYs prefer the DR (direct repeat) dimeric configurations, especially DR1 and DR6. Only a few WRKYs bind to the IR (inverted repeat) and ER (everted repeat) configurations (Supplementary Fig. S4A). We also derived the configurational preferences from the ampDAP-seq libraries (Supplementary Fig. S4B). The dimeric profiles are in rough agreement with SELEX but contain more noise. We proposed that such noises could have originated from TE propagation, because the noise level is much higher when analyzing DAP-seq of wheat WRKYs [31], whereby the large genome contains more TEs.

Dimeric profiles from the methyl-SELEX libraries (Supplementary Fig. S4C) suggest that upon DNA methylation, the dimeric configuration remains unchanged for most WRKYs. However, for a few WRKYs (WRKYs 15, 25, 65, and 8), DNA methylation can either decrease or increase the affinity of a specific dimeric configuration (Fig. 4A). For example, WRKY65 binds to DR1 and IR4 in the absence of methylation, but binds to DR1, DR6, and IR7 in the presence of methylation. While protein-level contacts lead to well-defined dimeric/multimeric preferences (e.g. ARFs [78, 79], MADS-boxes [80, 81], and bZIPs [82]), most TF dimers do not involve large-area protein-level contacts between the two TFs [83, 84], i.e., they can be dynamic and stimuli responsive. This could have allowed WRKY dimers to sense the minor modifications in DNA and change their dimeric configurations.

Figure 4.

Methylation changes the dimeric specificity of WRKYs. (A) Methylation changes the preference of dimeric configuration. Enrichments of dimeric WRKY CREs with different spacings and three relative orientations are visualized. Two “GTMAA” strings are concatenated with the annotated spacing (x-axis) and assessed for enrichments. (B) Methylation changes nucleotide specificity for WRKY dimers. Dimeric motifs are named according to their relative orientations and spacings. Their enrichments in SELEX and methyl-SELEX libraries are also indicated to the right. (C) Methylation increases dimeric combinational diversity. WRKYs almost always recognize half-sites of the class 2 specificity in the absence of methylation (upper), while recognizing multiple half-site specificities in the presence of methylation (lower). (D) Dimeric preferences for WRKY4 truncations. The dimeric preference is assessed the same way as in (A). WRKY4 belongs to subgroup I. It contains two WRKY domains (C- and N-terminal). Note that the dimeric preference of WRKY4-C still resembles that of the full-length protein.

In addition to the dimeric configuration, DNA methylation also changes the nucleotide specificity for closely spaced WRKY dimers. For example, in the absence of DNA methylation, only the right half-site of the DR-1 mode of WRKY70 has a high IC (Fig. 4B), suggesting a weak contact between the left half-site and the TF protein. Upon methylation, the IC of the left half-site has significantly increased, and two types of specificities have developed (Fig. 4B, the two methylated DR-1 modes of WRKY70). An increased IC is also observed for the left half-site of WRKY33 after methylation. For WRKY25, its ER2 mode only appears upon DNA methylation, and the two nucleotides between the two half-sites (GACCGGTC) have developed strong preferences.

The monomeric specificity is versatile for WRKYs in both the absence and presence of DNA methylation (Fig. 3A). However, upon homodimeric binding, WRKYs almost always recognize half-sites of the class 2 specificity in the absence of methylation (Fig. 4C). In contrast, WRKYs can recognize multiple half-site specificities in the presence of methylation. The monomeric specificities of classes 7, 8, 10, and 11 can flexibly combine, thereby enriching the diversity of WRKY dimeric sites (Fig. 4C).

The subgroup I WRKYs have two WRKY domains. However, previous studies have proposed that subgroup I WRKYs bind DNA only with one WRKY domain (in the C-terminus) [85, 86]. Accordingly, the SELEX libraries of subgroup I WRKYs are most enriched with the monomeric motifs, and the specificities of their monomeric motifs are also similar to those of other WRKYs (Fig. 2). We next examined whether the other WRKY domain (in the N-terminus) could have contributed to the dimeric binding of subgroup I WRKYs. Unexpectedly, for the truncation retaining only the C-terminal WRKY domain of WRKY4, the dimeric configuration remained similar to the full-length WRKY4 (Fig. 4D). This is in contrast to most TFs for which the non-DBD regions drastically affect their dimeric configurations [48, 75, 87]. On the other hand, retaining only the N-terminal WRKY domain gives a distinct dimeric configuration (Fig. 4D, WRKY4-N). Therefore, these results supported the previously proposed dominance of the C-terminal WRKY domain in the function of subgroup I WRKYs.

Methylation bidirectionally affects the affinity of WRKYs

By analyzing 10³–10⁴ genomic CREs, DNA methylation was reported to play a monotonic inhibitory role on DNA binding of WRKYs [12]. This conclusion is derived from correlation analyses (Pearson’s r) between the methylation rate of individual cytosines and the relative WRKY occupancy (DAP/ampDAP). Due to the overall low level of CHH methylation (in A. thaliana) and the sensitivity of Pearson’s r to outliers, the evaluated methylation impacts were dominated by 10–20 highly methylated CRE sites. Here, we compared the SELEX and methyl-SELEX libraries that respectively contain ∼10⁶ TF-binding sites (TFBSs), and show that the methylation effect on WRKY–DNA binding is bidirectional—methylation can either increase or decrease WRKYs’ DNA affinity, depending on the binding mode, TF identity, and the position of cytosine in the motif (Fig. 5).

First, we found that the most enriched 8-mers in the methyl-SELEX libraries also contain C/G bases, and are different from those in the SELEX libraries (Fig. 5A; Supplementary Fig. S5A). Using the enriched 8-mers as seeds, motif discovery has yielded distinct binding modes for SELEX (s1, Fig. 5A) and methyl-SELEX (m1 and m2, Fig. 5A) of WRKY35. We next asked if the binding modes m1 and m2 have indeed increased their affinity upon methylation, or if they are just relatively enriched after an overall decrease of C/G affinity. For this purpose, the 8-mers containing only A/T bases (orange bins in Fig. 5A, the medium value represented by the vertical line) were used as a reference, because their affinities are not (or less) affected by 5mC. Compared wirh the A/T-only 8-mers, the consensus of m1/m2 modes and a large amount of other C/G-containing 8-mers are clearly enriched in the methyl-SELEX libraries, indicating their increased affinity upon DNA methylation. Consistent with the SELEX analysis, genomic CREs of the m1/m2 modes show higher signals in methyl-ampDAP-seq than in ampDAP-seq (Fig. 5B), while CREs of the s1 mode show the reverse. The EMSA analyses also confirmed that the m1 mode consensus (AAAAGTAAAC) has a higher affinity for WRKY50 upon methylation (Supplementary Fig. S5B).

Second, cytosines in different positions of the WT-box WRKY motif respond differently to DNA methylation. While cytosines at the C4 position always decrease in affinity upon methylation (Fig. 5C), the methylation effect on the C7 cytosine can be diverse, and dependent on the TF identity. For example, C7 methylation increases the DNA affinity for WRKY29 and WRKY35, decreases the affinity for WRKY46 and WRKY65, and has little effect on the affinity for WRKY14 and WRKY15 (Fig. 5C). The bidirectional effect of C7 methylation is also revealed by comparing the relative affinity between the GTCAAC and GTCAAA subsequences (Fig. 5D). We next examined the effect of 5mCs in all positions of the WT-box motif (Fig. 5E). The results again suggest that the direction of the methylation effect is dependent on position and TF identity. Overall, DNA methylation decreases the affinities of C-2, C-1, and C4 in the positive strand and C-1, C0, C1, and C8 in the negative strand, while increasing the affinities of C2, C3, and C8 in the positive strand and C3, C4, C5, and C6 in the negative strand (Fig. 5E). The direction of the affinity effect is TF dependent for C7 and C9 in the positive strand and for C2 in the negative strand.

Notably, DNA methylation bidirectionally affects not only the C/G bases but also the A/T bases. For example, methylation can change the consensus sequence GTCAA into GTCTT for both the W-box (Fig. 5F, left) and the WT-box (Supplementary Fig. S5C) modes of WRKY40. For WRKYs with less drastic motif changes (e.g. WRKY35), the A/T ratios at individual positions are still different for the SELEX and methyl-SELEX WT-box motifs (Fig. 5F, right). We also show that the presence of 5mC can affect the absolute affinity of the A/T bases. For example, position 4 of the WT-box motif of WRKY35 is C in the absence of methylation, and contains both C and A upon methylation (Fig. 5F, right). Such a change can either be due to the decreased absolute affinity of 5mC, or due to the increased absolute affinity of A. Analyses of the DAP-seq-based libraries have disentangled the two possibilities—methylation only slightly decreased WRKY35’s occupancy at CREs with a C at position 4 (Fig. 5G), but significantly increased WRKY35’s occupancy at CREs with an A at position 4. The results indicate that methylation has primarily increased the absolute affinity of A. The affinity shifts of neighboring A/T bases could have originated from the methylation-induced changes in the overall geometry of the local DNA segment [88–90].

In addition to changing the specificity of individual positions in TFBSs, methylation can also change the interpositional dependency. For WRKY70, the interpositional dependency of dinucleotides has overall increased upon methylation (Fig. 5H), including dinucleotides both containing and without C/G bases. The enhanced interpositional dependency is consistent with the previous MEDEMO modeling [91], whereby considering the interpositional dependency was critical for an accurate prediction of the genomic CREs.

Distinct genomic targets of WRKYs in the methylation context

The bidirectional effects of DNA methylation not only drastically changed WRKYs’ specificity (Fig. 5A; Supplementary Fig. S5), but also systematically redefined genomic CREs of WRKYs, leading to both diminishment and emergence of CRE peaks (Supplementary Fig. S5D). AmpDAP-seq and methyl-ampDAP-seq, respectively, define CREs on the unmethylated and the methylated genome. The two libraries for the same WRKY only share ∼14% CREs in common (Fig. 6A, ampDAP versus methyl-ampDAP). This value is much lower than the shared CREs (∼54.7%) between two experimental replicates (Fig. 6A, ampDAP replicates), and even lower than the shared CREs between two different WRKYs (∼19.3%) in the absence of methylation (Fig. 6A, blue dash). Therefore, DNA methylation has overall redirected WRKYs to a distinct set of genomic CREs. The fraction of promoter region CREs was also reduced upon methylation (Supplementary Data S3A). Genomic CREs in the absence and presence of methylation can respectively be predicted by the enriched 8-mers in SELEX and methyl-SELEX libraries (Fig. 6B, C), and also by the enriched motifs (Supplementary Data S3B), suggesting the importance of specificity data in understanding the genomic targets of TFs.

Figure 6.

Methylation redefines the genomic targets of WRKYs. (A) WRKYs bind to a distinct set of CREs upon methylation. While two ampDAP-seq replicates share a large fraction of common peaks (∼54.7%, black line), ampDAP-seq and methyl-ampDAP-seq of the same WRKY share very few peaks in common (∼14.0%, red line); this fraction is even lower than ampDAP-seq libraries between different WRKYs (19.3%, blue line). The top 600 peaks of each library were considered. (B and C) Sequence specificity predicts CREs of WRKYs. (B) Genomic CREs in the absence and presence of methylation are respectively predicted by the top 1000 enriched 8-mers in SELEX and methyl-SELEX (P-values are from t-tests). The precision–recall (PR) curves of WRKY70 are exemplified in (C). (D–F) Methylated WRKY CREs also regulate typical WRKY functions. For WRKY75, the naturally methylated WRKY CREs were first extracted by overlapping DAP and methyl-ampDAP libraries (D) and exemplified in (E). GO analysis was performed for the genes associated with naturally methylated CREs (F), and has enriched terms of typical WRKY functions (responses to stress and pathogen). (G) Statistics of the DAP-seq libraries of non-cotyledon tissues. To explore how differential methylation between Arabidopsis tissues affects WRKY binding, DAP-seq libraries were constructed with gDNA from additional non-cotyledon tissues. (H) Tissue-specific methylation leads to differential binding of AtWRKY75.

We next examined whether natural methylation on gDNA has contributed to WRKY’s occupancy. Because DAP-seq uses unamplified gDNA (from cotyledon) as input that keeps the natural methylation state, the CREs defined by DAP-seq are expected to include both methylated and unmethylated binding sites of WRKYs. Intersecting DAP-seq and ampDAP-seq peaks extracts the unmethylated WRKY CREs (Supplementary Fig. S6A), whereas intersecting DAP-seq and methyl-ampDAP-seq extracts the methylated WRKY CREs (Fig. 6D). For example, the naturally methylated CRE of WRKY75 is found in the promoter of AtMYB123 (Fig. 6E), a regulator of anthocyanin synthesis. It is noteworthy that the naturally methylated CREs of WRKY75 also regulate genes with typical functions of WRKYs. For example, their target genes have enriched the GO terms “response to salicylic acid”, “response to molecule of bacterial origin”, and “systemic acquired resistance” (Fig. 6F). These GO terms are similar to the terms enriched for the unmethylated CREs of WRKY75 (Supplementary Fig. S6B, C), and suggest the involvement of DNA methylation in wiring the downstream regulatory network of WRKYs [92]. While the GO terms are similar upon methylation, the targeted genes are distinct (Supplementary Fig. S6D). It would be of interest to further explore the physiological consequences of such methylation-induced rewiring.

Because DNA methylation of CREs can be dynamic across tissues, we also explored whether this dynamic has resulted in differential binding of WRKYs. To answer this question, we additionally generated 72 DAP-seq libraries for silique, flower, and root tissues of Arabidopsis (Fig. 6G). Indeed, we observed that differential methylation of WRKY CREs near the genes AtAEP1 and At5G24680 has changed the occupancy of WRKY75 thereof (Fig. 6H).

Database for TF–DNA interaction of WRKYs

A comprehensive database addressing the DNA binding of WRKYs is as yet unavailable [93, 94]. To facilitate access to WRKYs' TF–DNA interaction data, we developed WRCD (WRKY Regulatory Code Database, https://transysbio.cn/WRKYRCDB.php). This database is built on the LNMP architecture (Linux, Nginx, MySQL, and PHP) (Fig. 7A). WRCD has collected 571 WRKY–DNA interaction profiles (Fig. 7A), including the dataset generated in this study (101 SELEX, 39 methyl-SELEX, 162 DAP-seq, 94 ampDAP-seq, and 65 methyl-ampDAP-seq libraries), as well as the published datasets (28 PBM profiles, 16 ChIP-seq, 40 DAP-seq, and 26 ampDAP-seq) [11, 36, 38, 39, 95–97]. We also integrated several CRE-related analytic tools into WRCD, such as track browsing, regulatory network visualization, gene search, and GO enrichment.

Figure 7.

The WRKY Regulatory Code Database (WRCD) overview and key functionalities. (A) Graphical abstract of WRCD. WRCD collects 140 SELEX-based and 321 DAP-seq-based libraries of AtWRKYs generated in this study along with published datasets (28 PBM, 16 ChIP, 40 DAP, and 26 ampDAP), mainly providing source files and visualization of binding motifs and genomic CREs. Download links and tools for exploring the datasets (e.g. JBrowse, Gene Search, and Regulatory Network) were also integrated. (B) All functions can be accessed from the home module. (C) The TFBS Module displays information for DAP-seq/ChIP libraries. (D) The DNA-binding specificity landscape of WRKY TFs under unmethylated and methylated conditions. (E) The Gene Search tool visualizes CREs around a specified gene. The table summarizes information of the CRE peaks, and the JBrowse plug-in displays tracks from the libraries.

WRCD contains six major modules—Home, TFBS, Specificity, JBrowse, Tools, and Download. The Home page summarizes the datasets and provides navigation to other models (Fig. 7B). The TFBS module collects DAP-seq-based/ChIP-seq libraries and the associated information, such as gene ID, motif logo, and E-value of the motif, while also linking to the bed and bigwig files (Fig. 7C). The Specificity module collects SELEX-based libraries and visualizes both the monomeric and the dimeric motifs (Fig. 7D). The JBrowse module is pre-loaded with all DAP-seq datasets to visualize the genome-wide WRKY-binding sites. The Tools module enables comparison of methylated and unmethylated regulatory networks, allows analyses of GO enrichment, and offers the search tool that identifies CREs of all WRKYs neighboring a user-specified gene (Fig. 7E). Finally, the Download module contains links to the source files of the WRKY–DNA interaction libraries.

Discussion

Focusing on AtWRKYs, we demonstrated that DNA methylation in the mC-all context has drastically reshaped their specificity and genomic CREs. Upon DNA methylation, we identified four classes of unreported WRKY specificities (Fig. 3A), observed distinct dimeric configurations (Fig. 4), and proved that incorporation of 5mCs not only decreases but can also increase WRKY’s affinity (Fig. 5). All these specificity changes further defined a distinct set of genomic CREs upon methylation (Fig. 6).

The bidirectional effect of methylation reported here is in contrast to previous findings, which suggest that the mC-all context has overall repressed DNA binding of plant TFs [11], including AtWRKYs [12]. The previous analyses primarily identified the repressive effects probably due to the usage of the unmethylated motifs for locating CRE segments. In this study, we first profiled the specificity of WRKYs with SELEX-based methods, and discovered the methylated motifs. We found that the methylated motifs are distinct from the unmethylated ones, yet still contain C/G nucleotides (Figs 2, 3). Therefore, the observed bidirectional effect of methylation is not contradictory to previous findings, but suggests a more comprehensive picture—the mC-all genomic context will decrease WRKY occupancy on CREs of the unmethylated motifs, but increase WRKY occupancy on CREs of the methylated motifs (Fig. 5A). The bidirectional effect also implies that the function of pioneer factors can be sequence dependent. Pioneer factors are TFs that reprogram cell fates by activating genes in the heterochromatin. To achieve this function, pioneer TFs not only have to overcome the nucleosome barrier, but also need to access methylated DNA [13, 14, 98]. Because the methylated and non-methylated binding models of a TF can be dissimilar, pioneer TFs may have to recognize distinct CRE sequences in the methylated heterochromatin when executing their “pioneer” functions.

Non-CG methylation is rare in animals, the use of CG-methylated ligands in affinity purification assays was therefore sufficient to capture the major effects of methylation [8, 14, 15]. In contrast, this study utilized mC-all ligands in SELEX- and DAP-seq-based assays. This is due to the substantial presence of non-CG methylation in plants. For instance, in A. thaliana, 6.7% of CHG and 1.7% of CHH are methylated [3, 4], and the highest levels of non-CG methylation are observed in Beta vulgaris, with 81.2% of CHG and 18.8% of CHH sites methylated [99]. The level of non-CG methylation (especially CHH) also varies across cell types, with the columella being the most hypermethylated [100]. The regulatory significance of non-CG methylation is further underscored by its tendency to form clusters near the genic loci, which increases the likelihood of creating mC-all contexts in CREs. A well-defined mC-all context is the mCHH islands that are located at ∼400 bp upstream of TSSs [7, 101], which overlap the promoter region. The mCHH islands are maintained by the RNA-directed DNA methylation (RdDM) mechanism that methylates not only CHH but also the CHG and CG contexts [102, 103]. Co-localizing with the mCHH islands, we found that the mC-all methylated WRKY motifs are enriched upstream of TSSs (Fig. 3F). The analyses also reveal a potential involvement of CHH islands in the regulation of the PRE4 element of OsWRKY13, as it is bound only when methylated (Fig. 3B, D, E). The methylation homeostasis in Arabidopsis is maintained by a “methylstat” mCHH island that is located upstream of the DNA demethylase ROS1 [104, 105]. Hypermethylation at the “methylstat” facilitates the expression of ROS1, which in turn decreases the cellular methylation level (including the “methylstat”) and switches off ROS1. We examined the DAP-seq-based libraries and confirmed that this “methylstat” is not recognized by WRKYs.

This work has generated a comprehensive dataset describing TF–DNA interactions of WRKY-family TFs. This paralleled dataset can be of particular value for investigating the relationships among protein sequence identity, DNA binding specificity, regulatory selectivity, and methylation sensitivity of WRKYs. By dissecting the correlation between protein sequence identity and DNA binding specificity, we identified key amino acids responsible for differentiating the W-box and WT-box preferences of WRKYs (Fig. 3H, I). Furthermore, we also demonstrated that the binding specificity of WRKYs can predict their regulatory selectivity and capture the methylation sensitivity (Fig. 6B, C). Recent advances in deep generative modeling have integrated cis-regulatory grammar and enabled designs of functional CREs that are active across species [106, 107], species specifically [108, 109] or cell type specifically [110–112]. We anticipate that models trained on the dataset generated here will inspire the design of WRKY CREs responding exclusively to specific WRKYs, or with activities tunable by DNA methylation.

In this study, binding assays were performed separately with either methylated or unmethylated DNA ligands. We expect that mixing both types of ligands in a single binding assay, such as in Methyl-Spec-seq [15] and EpiSELEX [14], will offer more insights by facilitating comparisons between the methylated and unmethylated affinities of individual sequences.

Altogether, this work unraveled the bidirectional effect of non-CG methylation, and implicated the possibility of overcoming methylation-induced silencing by strategically engineering the CRE sequences. This holds promise for addressing several key challenges in synthetic biology, such as preventing transgene-induced silencing and maximizing protein yields in biological chassis.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

The authors declare no competing financial interests, and requests for materials should be addressed to F.Z. (fjzhu@fafu.edu.cn).

Funding

The National Natural Science Foundation of China [32370582 and 32170554 to F.Z.]; the National Key Research and Development Program of China [2024YFC3407200 to F.Z. ]); “Chu Ying” Young Talent Project of Fujian to F.Z.; Ministry of Human Resources and Social Security ["funding for high-level overseas personnel" to F.Z.]; and Fujian Agriculture and Forestry University [KFXH23027 to H.G.].

Ethics approval

No animals or humans were involved in this study.

Data availability

All sequencing data have been deposited to China National Genomics Data Center under accession PRJCA040216 (https://ngdc.cncb.ac.cn/bioproject/browse/PRJCA040216), with raw sequence data under GSA: CRA025757 (https://download.cncb.ac.cn/gsa5/CRA025757). The sample information, including library types and treatment conditions, is provided in Supplementary Table S6. The genome-wide data tracks can be accessed from https://transysbio.cn/WRCDjbrowse.html. The generated codes are deposited on GitHub (https://github.com/Jiang-Bio/WRKY_RCDB) and Zenodo (https://doi.org/10.5281/zenodo.17306247).