The mirage of DNA methylation in transcriptional regulation of plants

Abstract

“Cytosine methylation plays an important role in the regulation of gene expression in plants.” Some iteration of this statement can be found in most papers centered on plant epigenetics and has become a widely accepted textbook claim. However, our generalized understanding of how DNA methylation exerts control over transcription is now challenged by observations demonstrating that transcriptional levels of most genes are unresponsive to DNA methylation changes. On a genome‐wide scale, associations between DNA methylation and transcription are usually statistically weak. Even when correlations are found, the cause and effect can be difficult to identify, as methylation changes sometimes follow rather than precede transcriptional changes. While a growing number of studies explore a possible connection between differentially expressed genes (DEGs) and differentially methylated genes (DMGs), we demonstrate here that DEG‐DMG overlaps are often significantly smaller than what could be expected by chance. This indicates that, contrary to expectations, changes in DNA methylation and changes in transcription sometimes avoid one another. Here, we discuss such observations and their implications for the hypothesis of a widespread control of gene expression directly by DNA methylation. While there are well‐documented examples where DNA methylation regulates transcription, we argue that such cases represent a minority of genes, and we opine that approaches of reverse epigenetics are therefore unlikely to find broad application in breeding.

Core Ideas

• Links between DNA methylation and transcription remain difficult to generalize and are weak on genome‐wide scales.

• Multiple lines of evidence suggest DNA methylation plays only a minor role in the regulation of gene expression.

• Various studies show mutual avoidance of transcriptional and methylation changes, challenging the current paradigm.

Plain Language Summary

Regulation of gene activity—by increasing or suppressing transcription—is a key to understanding how cells work, organisms develop, respond to their environments, and differ from each other. Some of this regulation is believed to be epigenetic, that is, determined by malleable changes of genetic material—DNA methylation and histone modifications. However, clear associations between DNA methylation and gene expression are surprisingly difficult to find. Here we highlight that the relationship is statistically weak and inconsistent across plant species. By testing results from many studies involving various species and experimental conditions, we found that changes in gene methylation and expression are often unrelated. More surprisingly, multiple studies show that genes that change their DNA methylation tend to have stable transcription, or vice versa. These observations suggest that the role of DNA methylation in regulating gene expression is minor and still poorly understood.

Abbreviations

DEG

differentially expressed gene

DMG

differentially methylated gene

DMR

differentially methylated region

gbM

gene body DNA methylation

Pi

inorganic phosphate

RdDM

RNA‐directed DNA methylation

TE

transposable element

TF

transcription factor

TFBS

transcription factor binding site

TSS

transcription start site

UTR

untranslated region

ZF

zinc finger

1. MECHANISTIC MODELS AND REAL‐LIFE EXAMPLES

The level of DNA methylation can influence the level of gene expression, perhaps to a degree equivalent to an “on‐off” switch. Our understanding of the mechanistic basis of this effect presupposes two different processes. First, DNA methylation in cooperation with histone modifications affects chromatin states, turning large chromosomal regions into tightly packed heterochromatin with reduced accessibility for the transcriptional machinery. As a result, condensed heterochromatin is generally (though not always) associated with transcriptional silencing (Grewal & Jia, 2007). Second, there is evidence that DNA methylation of regulatory elements, such as transcription factor binding sites (TFBS), can prevent the binding of transcription factors (TFs). If the TF works as an activator of transcription, methylation of its TFBS leads to gene silencing; if the TF is a repressor, methylation of its TFBS leads to gene activation (for a recent review, see Zhang & Zhu, 2024). Conversely, demethylation of TFBS can cause gene activation or repression, respectively. In vitro experiments show that >75% of TFs in Arabidopsis thaliana are sensitive to the methylation of their TFBSs, suggesting a potential for methylation‐dependent regulation (O'Malley et al., 2016).

There are several notable examples from the plant kingdom demonstrating the phenomenon of DNA‐methylation‐driven control of gene expression in vivo. Epigenetic regulation is known to be crucial for fruit ripening in tomatoes, oranges, and strawberries (Zhang & Zhu, 2024). The role of DNA methylation was first revealed in tomatoes, thanks to a natural colorless nonripening (Cnr) epimutation. A spontaneous Cnr mutant with fruits that fail to ripen was described in 1999 (Thompson et al., 1999), and epiallelic variation at the causative locus was discovered after an analysis of the cloned region failed to identify nucleotide differences (Manning et al., 2006). It has been observed that demethylation of the Cnr promoter (and promoters of other ripening‐associated genes) correlates with binding of Ripening inhibitor (RIN), a MADS‐Box TF indispensable for fruit ripening (Z. Liu et al., 2022). In wild‐type plants, RIN binding and local demethylation correlate with transcriptional activation of ripening genes. Conversely, in the Cnr epimutant, RIN fails to bind, the Cnr promoter remains hyper‐methylated, and fruits do not ripen (Zhong et al., 2013). However, the Cnr promoter remains hyper‐methylated also in the rin mutant, suggesting that rather than being sensitive to the methylation of its binding sites, RIN assists in their demethylation (Z. Liu et al., 2022; Zhong et al., 2013). Moreover, it has been shown that DNA methylation does not prevent RIN binding to naked DNA in vitro and that RIN binding to the Cnr promoter is insensitive to DNA methylation in vivo, since it is maintained in the demethylase mutant dml2‐3 (Niu et al., 2025). Fruit ripening is a complex process involving additional TFs, histone modifications, and redundant pathways, and while DNA methylation clearly participates, we still do not know which gene(s) are directly regulated by it (Y. Gao et al., 2019). The hyper‐methylated Cnr promoter may therefore be an “epi‐marker” rather than the causal epiallele of tomato ripening.

Besides fruit ripening, other cases where a spontaneous or induced epimutation triggers a phenotype that deviates from the norm are known. These include an epimutation causing dwarf phenotype in rice (Miura et al., 2009); an epimutation changing flower symmetry in Linaria vulgaris (Ehrenhofer‐Murray, 2021); sup epialleles affecting floral organization in Arabidopsis (Bondada et al., 2020); and late flowering in Arabidopsis fwa mutants caused by demethylation and activation of an FWA homeobox gene that is silent in wild‐type plants (Kinoshita et al., 2006; Soppe et al., 2000). The fwa epiallele has recently proven useful as a reporter gene in experiments testing molecular tools for targeted (sequence‐specific) hypermethylation (Ghoshal et al., 2021; M. Wang, Zhong, et al., 2023). Such epigenome editing has also been reported for a crop cassava (Manihot esculenta), which is susceptible to Xanthomonas phaseoli pv. manihotis, the causal agent of bacterial blight. During a susceptible reaction, the pathogen delivers a TF (TAL20) that binds to the promoter of MeSWEET10a, activation of which promotes an infection. Veley et al. (2023) demonstrated that TAL20 does not efficiently bind to its DNA target in vitro when the target sequence is methylated. Subsequently, Veley et al. (2023) generated transgenic cassava plants expressing an artificial zinc finger (ZF) domain fused with DMS3, a component of the RNA‐directed DNA methylation (RdDM) pathway. The ZF domain directs the fused protein to the promoter of MeSWEET10a, where RdDM locally increases CG methylation. This leads to a significant lowering of MeSWEET10a transcript levels and reduced symptoms of bacterial blight. The fwa and sup epimutants, MeSWEET10a, and other examples with various strengths of evidence support the theory of gene regulation by DNA methylation.

The theoretical basis and the cases described above can lead to an expectation that most (if not all) plant genes respond transcriptionally to DNA methylation changes in their vicinity. After all, all transcriptional regulation involves TFs (most of which are methylation‐sensitive), and all genes reside in chromatin, be it hetero‐ or euchromatin. The methylation‐transcription relationship can then operate on two levels of complexity. The interaction is (i) “passive” or “static”, with transcriptional levels of genes predetermined by DNA methylation. In this scenario, DNA methylation is ontogenetically stable but can change spontaneously or artificially within a population. Most of the described epialleles fall in this category. A higher‐order interaction (ii) can be characterized as “active” or “dynamic,” where local DNA methylation levels are under the control of the genome and are utilized to alter transcriptional levels during development or in response to stress. The epigenetic regulation of fruit ripening likely falls under this type of methylation‐transcription relationship.

While the fractions of genes in the categories (i) and (ii) remain unknown even for model organisms, it is often assumed that methylation‐driven gene regulation is common. This assumption is implicit in the experimental design and result interpretation of many studies and nourishes expectations about the practical utilization of epigenetic diversity. However, multiple lines of evidence indicate that methylation‐responsive genes may be in the minority and that transcriptional regulation rarely employs dynamic changes in DNA methylation. The following two sections provide an overview of such evidence.

Core Ideas

• Links between DNA methylation and transcription remain difficult to generalize and are weak on genome‐wide scales.

• Multiple lines of evidence suggest DNA methylation plays only a minor role in the regulation of gene expression.

• Various studies show mutual avoidance of transcriptional and methylation changes, challenging the current paradigm.

2. UNIMPRESSIVE GENOME‐WIDE SIGNALS

The study of the methylation‐transcription relationship in plants is difficult. Given the poor characterization of plant TFBS (Jyoti et al., 2024) and the rarity of distinct CpG islands in plant promoters and plant genomes in general, there is typically no prior information on the position of elements whose methylation has consequences for the transcription of adjacent genes. DNA methylation is therefore usually measured in sliding windows across the entire genome or specifically in the transcribed gene partitions (usually referred to as “gene bodies”) and loosely defined promoters. DNA methylation of gene bodies often forms a pattern of a relatively high CpG methylation that dips around transcription start and termination sites, called “gene body DNA methylation” (gbM) (Bewick & Schmitz, 2017). However, gbM is persistently difficult to relate to transcription. For example, Bewick et al. (2016) analyzed gene expression and methylomes in Eutrema salsugineum, a species that lacks gbM, and in A. thaliana epigenetic recombinant inbred lines, and found no evidence supporting a regulatory role of gbM in transcription. Despite observations showing that transcription start sites (TSS) are unmethylated in most transcribed genes across multiple species (Niederhuth et al., 2016), the relationship between promoter methylation and transcription is similarly unclear. Promoter methylation and transcription in Brachypodium distachyon, maize, and other species are only weakly negatively correlated on a genome‐wide scale, with many highly expressed genes having highly methylated promoters (Figure 1a,b; with promoters defined as regions [−200, +50] bp around TSS). These genome‐wide assessments are not much different from those querying the methylation of gene bodies and transcription, where Pearson correlation coefficients are either insignificant or inconsistent across different species (Figure 1c,d). In contrast to the trends observed on these scatter plots for Brachypodium and maize, weak positive Pearson correlation between gbM and transcription was observed in poplar leaves (Liang et al., 2019). Extremely weak correlation (R ² < 0.01) was observed between gene expression and 3′UTR (untranslated region) methylation in Mikania micrantha (Sang et al., 2024). Nonetheless, scatter plots of DNA methylation and transcription are rarely shown in studies comparing these two variables, perhaps because it is difficult to draw firm conclusions from the lack of correlation. As mentioned above, TFBS methylation can impede the binding of both—transcriptional activators and repressors—possibly leading to conflicting signals. On top of that, ∼4.3% of Arabidopsis TFs preferentially bind to methylated DNA in vitro (O'Malley et al., 2016). This assumed complexity can be used to explain the absence of a genome‐wide correlation between DNA methylation and transcription, without rejecting their possible causal relationship. Other lines of evidence therefore need to be considered.

FIGURE 1.

Scatter plots showing a relationship (or lack thereof) between DNA methylation and gene expression in developing grain. Methylation and log‐transformed TPM (Transcripts Per Million) data described in Bellec et al. (2023) were used. Top panels show data for Brachypodium distachyon; bottom panels for maize. (a) methylation of promoters (−200, +50 bp around transcription start sites [TSSs]) in the CG context. (b) methylation of promoters in the CHG context. (c) methylation of exons in the CG context. (d) methylation of exons in the CHG contexts. Where correlation is significant (p < 0.01), a linear regression curve is shown in red. Correlations between transcript abundance and CHH methylation were very weak and inconsistent (r = 0.05 and −0.015 for promoters in maize and Brachypodium, respectively, and r = −0.113 and −0.043 for exons in maize and Brachypodium, respectively), and the corresponding scatter plots are not shown here.

In the model plant A. thaliana, mutant lines with defective methyltransferases or other components of DNA methylation pathways can be used to examine the importance of DNA methylation for gene expression. A complete removal of DNA methylation was achieved recently in a quintuple Arabidopsis mutant (mddcc) with all five known methyltransferases, MET1, DRM1, DRM2, CMT3, and CMT2, knocked out (He et al., 2022). The complete absence of DNA methylation was accompanied by 3738 differentially expressed genes (DEGs) (fold change >2; false discovery rate < 0.01) in aerial tissues, with 2273 upregulated and the remainder downregulated in respect to wild‐type plants (He et al., 2022). This may look like definitive proof of transcriptional control by DNA methylation. However, these DEGs represent only 13.5% of all annotated protein‐coding genes in the A. thaliana genome. In other words, transcript levels of 86.5% of genes did not significantly change in mddcc. Importantly, almost all of this effect can be ascribed to CpG methylation, since the selective removal of CHG and CHH methylation (ddcc mutant) led to only 310 DEGs (1.1% of annotated genes). Although we still do not know what fraction of genes would transcriptionally respond to hypermethylation, it is nonetheless striking that such a large majority remained transcriptionally unaffected by a complete loss of DNA methylation. Moreover, as the authors point out, the 3738 DEGs include transcriptional changes linked to methylation both directly and indirectly. What does that mean? In their subsequent paper (Zhao et al., 2022), the same team reported that the complete removal of DNA methylation causes extensive switches of chromatin states. The authors noted that 71% of the upregulated DEGs in mddcc were associated with gain/increase of active histone marks or loss/decrease of H3K27me3, while 64% of the downregulated DEGs were associated with loss/decrease of active marks or gain/increase of H3K27me3. Moreover, “most loci in which histone modifications remained stable in mddcc displayed similar expression levels in the wild type and mddcc, regardless of DNA methylation” (Zhao et al., 2022). The authors conclude that “the transcriptional reprogramming observed upon loss of DNA methylation correlates with the redistribution of the examined histone modifications or through the modulation of the levels of histone modifications.”

It is now recognized that DNA methylation and histone modifications are interconnected. In particular, DNA methylation and H3K9me2 methylation are in a self‐reinforcing loop, while CpG methylation restrains the binding of H3K27me3 (J. Wang, Meng, et al., 2016; Nakamura et al., 2025). Removal of DNA methylation can therefore lead to the expansion of H3K27me3 (a transcriptionally repressive mark associated with tissue‐specific and developmentally regulated genes) and lowering of H3K9me2 (a repressive mark), resulting in down‐ and up‐regulation of affected genes, respectively. Since H3K9me2 and H3K27me3 often cooperate with activating histone marks (H3K9ac and H3K4me3, respectively) (J. Wang, Meng, et al., 2016), the absence of DNA methylation can cascade into additional transcriptional effects. Due to this complexity of the crosstalk between DNA methylation and histone modifications, it is difficult to establish to what extent DNA methylation affects transcriptional levels directly and on its own or acts only indirectly by modulating the levels of histone marks, the real modulators of transcription. The absence of strong genome‐wide correlations between DNA methylation and transcription, together with the analyses of the quintuple mutant (He et al., 2022; Zhao et al., 2022), may suggest that the effect of DNA methylation is mostly indirect.

3. DEG‐DMG OVERLAPS

The studies on methyltransferase mutants, like the one mentioned above, explore the link between DNA methylation and transcription by comparing two groups of samples—the mutant and the wild‐type. But across‐sample comparisons of methylomes and transcriptomes can also include different genotypes, different tissues or developmental stages of the same genotype, or various environmental or experimental conditions (e.g., stress and control). This type of analysis yields a set of DEGs and a set of DMGs, both of which are smaller than the full set of annotated genes. These subsets (DEGs and DMGs) never overlap fully, as some DEGs are not differentially methylated, and some DMGs are not differentially expressed. Nonetheless, for the group of genes that are simultaneously DMGs and DEGs, it is usually concluded that the change of transcription resulted from the change in DNA methylation. However, two subsets of a larger set can overlap by chance (even when drawn independently), and unfortunately, most studies comparing DEGs and DMGs do not provide statistical tests of significance. Two types of tests are appropriate for the statistical assessment of the DEG‐DMG relationship. The first one is the 2 × 2 chi‐square contingency table, comparing the observed sizes of four gene categories (DEG‐DMG, DEG‐nonDMG, nonDEG‐DMG, and nonDEG‐nonDMG) to the counts expected if the two variables were independent. However, the result of the chi‐square test may not be reliable if any of the four categories have very low counts. The alternative test uses the hypergeometric distribution to measure the statistical significance of having drawn a sample consisting of a specific number of successes from a larger population containing a known fraction of successes. The hypergeometric test is used, for example, to assess the significance of gene ontology term enrichment (Rivals et al., 2007), and in the context of the DEG‐DMG comparisons can be understood as follows. Within a population of genes (i.e., all annotated genes in a genome), there is a known fraction that is differentially methylated (successes). If we draw a subset (i.e., all identified DEGs) from the full set of genes, we can test whether the fraction of DMGs in this subset corresponds to the DMG fraction of the full set or, alternatively, is over‐ or underrepresented. In other words, we can assess the over‐ or under‐representation of the DEG‐DMG overlap, with associated p‐values. If the DEG‐DMG overlap is not statistically different from a random overlap, the transcriptional and methylation changes occur independently. If the DEG‐DMG overlap is significantly overrepresented, then differential gene expression tends to co‐occur with DNA methylation changes. Conversely, if the DEG‐DMG overlap is significantly underrepresented, then transcriptional and methylation changes tend to avoid one another (Figure 2a–c).

FIGURE 2.

Possible outcomes of a DEG‐DMG comparison, verifiable by testing the sizes of intersections against the hypergeometric distribution. (a) Avoidance—expression changes (DEGs) and methylation changes (DMGs) occur preferentially on different sets of genes. (b) Independence—DEGs and DMGs are two completely independent gene sets that intersect randomly. (c) Co‐occurrence—expression changes and methylation changes occur preferentially on the same genes. Both avoidance and co‐occurrence indicate a possible causal relationship, although working in different directions. In the case of co‐occurrence, changes in one variable may cause changes in the other variable. In the case of the avoidance, changes in one variable may prevent changes in the other variable. (d) Overview of the DEG‐DMG intersections in published studies. The dotted line in the center indicates a random overlap, with underrepresentation to the left and overrepresentation to the right. Statistical significance (hypergeometric test) is indicated by the color of the intersection symbol. *An outlier data point (Feng et al., 2016) is not shown on this figure.

Here, we reanalyzed some of our own data (Sammarco et al., 2024; Table 1) and compiled other studies that quantify both DEGs and DMGs in corresponding samples in order to assess the DEG‐DMG overlaps by the hypergeometric test (Table 2). We conducted a search using Web of Science (Clarivate), querying terms (DEG* AND differential* AND express* AND methylat* AND (DMG* OR DMR*)) in the ‘topic’ of research papers in ‘Plant Sciences’ category. The search identified 33 research papers, but the statistics could not be calculated for most of them due to missing information. We supplemented the systematic search with additional papers we knew to contain all relevant numbers, yielding 24 papers altogether. Only four of these studies provide a statistical test for the DEG‐DMG overlap, two of them using the hypergeometric test (Lafon‐Placette et al., 2018; Y. Zhang et al., 2023), and the other two using the chi‐square test (Wang et al., 2025; Xue et al., 2022).

TABLE 1.

Fractions of 210 inter‐genotype comparisons of Fragaria vesca where significantly overrepresented DEG‐DMG overlaps were found.

Context	Test	Promoters	Gene bodies
CpG	Chi‐square	0	0.514
	Hypergeometric	0.014	0.567
CHG	Chi‐square	0.143	0.957
	Hypergeometric	0.167	0.948
CHH	Chi‐square	0.057	0.581
	Hypergeometric	0.067	0.576

Abbreviation: DEG‐DMG, differentially expressed gene‐differentially methylated gene.

TABLE 2.

Statistical assessment of DEG‐DMG overlaps in published studies.

			Detection threshold p | q		Hypergeometric test
Species	Comparison	DEG and DMG data sets	DEGs	DMRs	Representation factor for DEG‐DMG (excess genes above random expectation)	p‐value	Comments	Study
Fragaria vesca	Pairwise comparisons of 21 wild populations grown in a common garden	Promoters (all contexts)	– | 0.05	– | 0.05	1.06× (4)	0.318	Median values of the representation factor and FDR across 210 inter‐genotype comparisons; see also Table 1	Sammarco et al. (2024)
		Gene body (all contexts)			1.3× (15)	0.044
		CG (all partitions)			1.16× (10)	0.126
		CHG (all partitions)			1.61× (13)	0.015
		CHH (all partitions)			1.22× (9)	0.132
Hordeum vulgare	Immature embryos and derived callus	2 kb upstream (all contexts)	– | 0.05	?	0.86×	8.3e‐15	Similar results were obtained when gene partitions were analyzed for each DMG context separately	Huang et al. (2025)
		Gene body (all contexts)			0.97×	0.196
		2 kb downstream (all contexts)			0.87×	9.2e‐10
		CG (all partition)			0.91×	1.7e‐04
		CHG (all partition)			0.75×	9.1e‐08
		CHH (all partition)			0.89×	6.6e‐14
Glycine max	MET1 knockout mutants and wild‐type	CG 2 kb upstream	– | 0.05	– | 0.05	1.29× (59)	9.8e‐06	–	Wang et al. (2025)
		CG gene body			0.73×	1.0e‐25
		CG 2 kb downstream			0.99×	0.481
Cicer arietinum	Stages of embryo development; different cultivars at the same stage	Stage 1‐stage 2	– | 0.05	– | 0.01	0.90×	0.01	–	Rajkumar et al. (2020)
		Stage 2‐stage 3			1.30× (72)	5.0e‐07
		Stage 3‐stage 5			1.14× (105)	3.4e‐06
		Stage 5‐stage 7			1.02× (26)	0.176
		J‐H (stage 3)			1.02× (22)	0.195
		J‐H (stage 5)			1.13× (55)	0.001
Saccharum officinarum	Different tissues	Root‐leaf	– | 0.05	?	1.08× (411)	2.5e‐16	–	Xue et al. (2022)
		Rind‐leave			1.05× (321)	1.3e‐13
		Pith‐leave			1.04× (343)	1.3e‐15
		Rind‐root			1.06× (244)	1.7e‐12
		Pith‐root			1.05× (226)	1.3e‐11
		Pith‐rind			1.10× (65)	4.5e‐04
Arabidopsis thaliana	Chromosome inversion lines and wild‐type	CS1282‐wt	– | 0.01	– | 0.001	0.61×	1.5e‐63	DEG‐DMG overlaps obtained by personal communication with the authors	Khosravi et al. (2025)
		RW290‐wt			0.62×	1.3e‐66
		RW295‐wt			0.63×	6.8e‐64
Fragaria vesca	Heat stress effect followed over clonally‐propagated generations	M*			0.27×	1.8e‐441	No DEG‐DMR overlap was observed in the third‐generation St3	López et al. (2024)
		St1			0.56×	1.9e‐11
		St2			0.9×	0.398
		St3			–	0.573
Oryza sativa	Drought‐tolerant and drought‐sensitive cultivars; stress‐control	DK151c‐IR64c	– | 0.05	– | 0.1	0.76×	0.001	Microarray instead of RNA‐seq	W. Wang, Qin, et al. (2016)
		IR64s‐IR64c			0.74×	0.071
		DK151s‐DK151c			1.79× (5)	0.035
Arabidopsis thaliana	Stages of flower development (meristem and early flower)	Gene body	?	– | 0.001	1.22× (255)	1.1e‐25	–	Yang et al. (2015)
		1 kb upstream			0.57×	4.8e‐16
Malus domestica	Drought‐tolerant and drought‐sensitive cultivars; control and water‐deficit	Qc‐Hc (all)	– | 0.05	?	0.95×	9.7e‐05	–	Xu et al. (2018)
		Qc‐Hc (hyper‐down)			0.98×	0.281
		Qc‐Hc (hypo‐up)			0.94×	0.034
		Qc‐Qwd (all)			0.86×	1.5e‐15
		Qc‐Qwd (hyper‐down)			0.87×	4.0e‐04
		Qc‐Qwd (hypo‐up)			0.89×	0.005
		Hc‐Hwd (all)			0.88×	1.8e‐10
		Hc‐Hwd (hyper‐down)			0.89×	0.003
		Hc‐Hwd (hypo‐up)			0.89×	0.004
Oryza sativa	Cadmium stress and control	(‐Cd)‐(+Cd)	– | 0.05	– | 0.05	14.68× (1355)	0	–	Feng et al. (2016)
Brassica napus	Freezing‐resistant and sensitive cultivars under two freezing treatments	T1 (all) T2 (all)	0.001 | –	‐ | 0.05	0.53× 0.53×	4.8e‐86 1.3e‐107	DEGs and DMRs between the two cultivars under control conditions were excluded from the totals under treatments	Zheng et al. (2025)
Vitis vinifera	Fruit ripening in parent and derived early‐ripening mutant	Across stages parent (all)	0.05 | ‐	0.00005 | –	1.51× (157)	4.5e‐22	Only genes reoccurring in 2 or more stage comparisons were counted as DEGs and DMGs	Wei et al. (2025)
		Across stages mutant (all)			1.59× (100)	5.8e‐15
		Parent‐mutant (all)			1.37× (22)	0.003
Citrus junos, Poncius trifoliata	Diploids and derived autopolyploids	Citrus junos (CG)	0.05 | –	0.01 | –	2.21× (4)	0.038	–	H. Gao et al. (2024)
		Citrus junos (CHG)			2.87× (5)	0.006
		Citrus junos (CHH)			0.57×	0.217
		Poncius trifoliata (CG)			1.61× (14)	0.002
		Poncius trifoliata (CHG)			2.24× (17)	1.5e‐05
		Poncius trifoliata (CHH)			1.29× (4)	0.148
Hevea brasiliensis	Grafted plants and donors	Gene body (hypo‐up)	– | 0.05	0.05 | –	1.03× (17)	0.23	–	H.‐L. Li et al. (2024)
		Gene body (hypo‐down)			1.07× (28)	0.064
		Gene body (hyper‐up)			0.98×	0.46
		Gene body (hyper‐down)			0.85×	0.169
		Promoter (hypo‐up)			0.93×	0.099
		Promoter (hypo‐down)			1.03× (6)	0.341
		Promoter (hyper‐up)			0.97×	0.45
		Promoter (hyper‐down)			0.99×	0.461
Osmanthus fragrans	Low‐temperature treatment of floral buds	Gene body (hypo‐up)	– | 0.05	?	0.62×	1.3e‐16	Only the longest treatment (L72) is summarized here; the shorter treatments give similar results	Zhong et al. (2024)
		Gene body (hypo‐down)			0.76×	6.9e‐07
		Gene body (hyper‐up)			0.78×	0.079
		Gene body (hyper‐down)			1.17× (6)	0.156
		Promoter (hypo‐up)			0.87×	0.002
		Promoter (hypo‐down)			0.79×	4.7e‐07
		Promoter (hyper‐up)			0.96×	0.452
		Promoter (hyper‐down)			0.84×	0.22
Fragaria vesca	Effect of increased temperature on 4 ecotypes	ES12 (all)	0.05 | –	– | 0.003	1.19× (75)	1.5e‐05	The original study provides slightly different p‐values; we confirmed the numbers given here are correct (Y. Zhang, personal communication)	Y. Zhang et al. (2023)
		ICE2 (all)			0.97×	0.284
		IT4 (all)			0.84×	0.002
		NOR2 (all)			0.99×	0.46
Dendrobium officinale	Normal and malformed (hyperhydric) plants	Hypo‐up	?	?	0.82×	0.002	Two other comparisons with different patterns not shown here	H. Zhang et al. (2022)
		Hypo‐down			1.66× (130)	1.4e‐22
		Hyper‐up			0.75×	0.01
		Hyper‐down			1.74× (61)	5.2e‐11
Hibiscus cannabinus	Anthers after 5‐azacytidine treatment	5‐azaC0‐5‐azaC200 (all)	0.05 | –	0.05 | –	2.63× (45)	6.7e‐14	methylRAD instead of whole‐genome BS‐seq	Z. Li et al. (2021)
Prunus avium	Chilling treatment in different cultivars	Kordia‐443CH (hypo‐up)	– | 0.01	– | 0.01	0.94×	0.475	Four other comparisons with similar and mostly statistically random DEG‐DMG overlaps are not shown here	Rothkegel et al. (2020)
		Kordia‐443CH (hypo‐down)			1.25× (6)	0.125
		Kordia‐443CH (hyper‐up)			0.88×	0.287
		Kordia‐443CH (hyper‐down)			1.1× (5)	0.246
Prunus persica	Normal and mealy fruits at different stages	Normal E1‐E3 (all)	– | 0.01	0.05 | –	0.55×	0.168	–	Rothkegel et al. (2021)
		Mealy E1‐E3 (all)			1.31× (4)	0.135
		E1 mealy‐normal (all)			2.62× (7)	0.002
		E3 mealy‐normal (all)			1.67× (6)	0.034
Populus × euramericana	Different watering conditions	All	– | 0.05	– | 0.05	0.85×	.5e‐10	Microarray chips instead of whole genome analyses; only DMRs reappearing in at least two comparisons	Lafon‐Placette et al. (2018)
Pyrus pyrifolia	Covered (B) and sun‐exposed (D) fruits	D1‐B1 (all) D2‐B2 (all) D3‐B3 (all)	– | 0.05	– | 0.05	1.12× (66) 1.09× (49) 1.13× (81)	9.1e‐04 0.01 1.4e‐04	Counts are also available per sequence context and separately for hyper‐/hypo‐methylation and up‐/down‐regulation, but the test results are not statistically significant due to low numbers	H.‐N. Liu et al. (2023)
Oryza sativa	Two transgenic lines, their cross and parent	n‐ZH11 (gene body)			0.63×	3.0e‐08	Four other genotype combinations are available, all with similar statistical results	X. Wang, Niu, et al. (2023)
		En‐ZH11 (promoter)			0.41×	2.1e‐47

Abbreviation: DEG‐DMG, differentially expressed gene‐differentially methylated gene; DMR, differentially methylated region; FDR, false discovery rate.

Our transcriptomic and methylomic data collected from 21 wild populations of woodland strawberry (Fragaria vesca) grown in a common garden (Sammarco et al., 2024) offer a high number of comparisons (210 pairwise inter‐genotype combinations) that were assessed in respect to sequence context and gene partition (promoters 2 kb upstream of TSS and gene bodies). Across all the comparisons, the DEG‐DMG overlaps are either random or significantly overrepresented, but never significantly underrepresented. The chi‐square and the hypergeometric test gave very similar results (Table 1). Overrepresented DEG‐DMG overlaps were most common in the gene bodies and the CHG context (∼95% inter‐genotype comparisons with FDR < 0.05; Table 1), but the number of DEG‐DMG genes above random expectation was generally very low, averaging around 13 genes per comparison. In the vast majority of comparisons involving promoters, we observed statistically random DEG‐DMG overlaps, with a few significant cases seemingly resulting from an expansion of CHG methylation from the gene bodies (Table 1). When merging all sequence contexts and partitions, median inter‐genotype DEG‐DMG overlaps are random, that is, without statistical significance.

Across the additional reviewed studies, we detected DEG‐DMG overlaps of all three categories—random, significantly overrepresented and significantly underrepresented—without a recognizable pattern (Figure 2d). The patterns prevalent in the 210 pairwise inter‐genotype comparisons of F. vesca are not reflected in other studies that separately look into DNA methylation of promoters and gene bodies, where mostly significant underrepresentation of DEG‐DMG overlaps was found (Figure 2d; Table 2). With the exception of across‐tissue comparisons represented by a single study (Xue et al., 2022), statistical assessments of DEG‐DMG overlaps are inconsistent across studies and species. We should point out that most tests of the DEG‐DMG overlaps do not distinguish between up‐ and down‐regulated DEGs or between hyper‐ and hypo‐methylated DMGs. It may therefore be suggested that we see random or underrepresented DEG‐DMG overlaps because hyper‐methylated DMGs are very rarely up‐regulated, and hypo‐methylated DMGs are very rarely down‐regulated. When all categories are grouped together, the dearth of those “unexpected” categories counteracts the excess of the “expected” categories (i.e., hyper‐methylated/down‐regulated and hypo‐methylated/up‐regulated). However, studies that allow to calculate the hypergeometric test for each category show that even the “expected” categories are mostly random and can be significantly underrepresented (Table 2; Figure 2d).

While overrepresented and random DEG‐DMG overlaps are easy to interpret (concluding that transcriptional changes are driven by or unrelated to DNA methylation changes, respectively), the frequent observations of significant DEG‐DMG underrepresentation (avoidance) are perplexing and hard to explain within the current understanding of the role of DNA methylation. It needs to be emphasized that avoidance is not a lack of a significant overlap. Instead, it means that the observed overlap is significantly smaller than the one that can be expected by chance. The cases of underrepresented DEG‐DMG overlaps therefore suggest that either transcriptional changes preferentially occur in genes with stable DNA methylation or that DNA methylation profiles change preferentially in genes that do not change transcription. Either of these two options is difficult to rationalize, but we offer possible explanations in the next section.

4. THE MINOR ROLE OF DNA METHYLATION IN REGULATING GENE EXPRESSION

The sizes of DEG‐DMG overlaps across the reviewed studies suggest that the capacity of genomes to modulate transcription via DNA methylation is rather limited. As discussed above, in the extreme case of the total loss of DNA methylation in the Arabidopsis mddcc mutant, only 13.5% of genes significantly changed transcription, much of which could be a consequence of chromatin state switches rather than DNA methylation alone (He et al., 2022; Zhao et al., 2022). Under more natural circumstances, the range of DNA methylation changes is much smaller, suggesting an even lower potential for methylation‐driven regulation. Furthermore, the overrepresented DEG‐DMG overlaps observed across tissues, developmental stages, and genotypes confirm that the number of genes responding to DNA methylation changes is very low (dozens‐hundreds). For example, Xue et al. (2022) found that in sugarcane, 33,855 genes (40.4%) were differentially methylated between roots and leaves, while 12,441 genes (14.8%) were differentially expressed. If methylation and transcription were entirely independent, we would expect to see 5025 genes that are simultaneously DEGs and DMGs. The actual number observed was 5436, only 411 genes above the random expectation. We can therefore conclude that among genes differentiating leaves and roots of sugarcane, only ∼411 are (co‐)regulated by DNA methylation. That is only 3.3% of the DEGs and ∼0.5% of all annotated genes.

Moreover, the comparisons where statistically random overlaps were observed put the number of methylation‐regulated genes at zero. And the cases with significantly underrepresented DEG‐DMG overlaps suggest an even more counterintuitive scenario—genes that need to change their transcriptional level tend to stabilize their DNA methylation, at least in a significant majority of cases. DNA methylation changes still occur, but away from the DEGs. How can we explain such a pattern? Perhaps the cis‐regulatory elements that are sensitive to DNA methylation are more distant from their targets than typically assumed (i.e., >2 kb away), making them difficult to identify and associate with relevant genes. While the epialleles mentioned in the first section are located within genes and 2 kb upstream regions, this may be a subset biased by such difficulties rather than a guiding principle. The Cnr epiallele is located 1485–1771 bp upstream of Cnr TSS; the epiallele causing dwarfism in rice is 2 bp upstream of TSS; the FWA epiallele is located in a region (−100, +400) bp relative to TSS, and the Peloria and sup epialleles are hypermethylated within the 5′UTR and the coding sequence of the relevant genes.

To find a possible explanation for the mutual avoidance of DEGs and DMGs, perhaps we need to take a step back and look at DNA methylation in the context of non‐epigenetic mechanisms. A gene is usually transcribed or silenced when appropriate TFs (responding to biochemical, developmental, or environmental stimuli) are present and allow or prevent the initiation of transcription, respectively. This is arguably the dominant mechanism of transcriptional regulation. If the expression of a particular gene needs to be tightly regulated, it may be undesirable for DNA methylation, which can be quite erratic, to interfere with this process. For example, DNA methylation levels can fluctuate due to the availability of various metabolites, for example, S‐adenosyl‐l‐methionine, the methyl‐group donor required by methyltransferases (Huang & He, 2024). Metabolic changes throughout development or under particular environmental conditions can alter the availability of S‐adenosyl‐l‐methionine, with possible effects on the maintenance and dynamics of epigenetic marks (Mentch & Locasale, 2016). This could be the reason why DNA methylation alterations, primarily unrelated to gene regulation, are observed under various developmental or environmental circumstances. Such inherent instability of DNA methylation may be tolerated at some loci but actively mitigated at inducible genes in order to prevent their misregulation.

Another possible explanation of the mutual avoidance of DEGs and DMGs is related to the maintenance of regulatory integrity in the vicinity of strongly up‐regulated genes. Consider a stress‐responsive DEG strongly up‐regulated by TFs, independently of DNA methylation. This is accompanied by a locally open chromatin and a high concentration of the transcriptional machinery, possibly increasing unintended transcription of neighboring genes. If the expression of the surrounding genes needs to stay unchanged, DNA methylation may step in to stabilize it. As a consequence, the methylation level of the strongly up‐regulated gene remains unchanged, but the methylation level of neighboring genes is altered to rebalance their expression. A similar mechanism has been proposed by Secco et al. (2015), who observed that under inorganic phosphate (Pi) starvation in rice, transcriptional changes in some genes are accompanied by DNA methylation changes in proximal transposable elements (TEs) located upstream. A detailed time course analysis in Secco et al. (2015) allowed deciphering the temporal hierarchy between changes in transcription and changes in DNA methylation. For a set of Pi‐starvation‐induced genes associated with proximal differential methylation, the authors demonstrated that while the majority (55%) of these genes changed transcription by the seventh day of Pi‐starvation, only 9% of their DMRs (differentially methylated regions) appeared by this time point, with the remainder detectable at a later stage (24 days). This implies that methylation modifications associated with Pi‐starvation in rice follow rather than precede transcriptional changes and cannot therefore be the causal force. Since almost all of the Pi‐starvation‐induced, DEG‐associated hypermethylated regions overlapped with TEs, Secco et al. (2015) suggest that the induction of methylation may be involved in repressing the activity of specific TEs close to highly upregulated genes, via hypermethylation of the TEs in the CHH context.

Finally, the statistical incoherence of the DEG‐DMG overlaps, showing independence, over‐ and under‐representation with similar frequencies (Figure 2d), may be in part due to technical limitations, particularly in the case of DMG detection. Differentially methylated regions are often screened by calculating averages in fixed‐size windows (e.g., 200 bp) across the genome, and when significant DMRs overlap with genes (and ∼2 kb up‐/downstream), such genes are labeled DMGs. However, averaging DNA methylation in fixed windows could miss relevant regulatory regions that are much smaller than the window size (D. Sun et al., 2014) (or get split into two neighboring windows), while many of the DMRs may relate to the silencing of TEs rather than to the regulation of a gene 1–2 kb away. As a consequence, our current strategies of DMR detection may yield DMG sets that are largely random in respect to gene regulation.

Perhaps other reasons for the DEG‐DMG avoidance exist and may be related to functions of DNA methylation that remain unknown. The present article argues that variation in cytosine methylation is mostly neutral in the evolutionary sense (similarly to mostly neutral nucleotide substitutions) and plays only a minor role in the regulation of gene expression, perhaps limited to a small fraction of genes. The main function of DNA methylation then appears to revolve around TE silencing or, perhaps more generally, to ensure the maintenance and reinforcement of predetermined transcriptional states and prevent spurious and aberrant transcription, as described in mammals (Neri et al., 2017; Wu et al., 2023). These roles may seem underwhelming, considering the taxonomic omnipresence and evolutionary conservation of DNA methylation pathways, but perhaps we need to reappraise their significance without overestimating their involvement in dynamic transcriptional regulation. Such an outlook has implications for the exploitation of DNA methylation variants (natural or induced) in crop breeding (“epibreeding”), which is anticipated by some authors (Dalakouras & Vlachostergios, 2021; Kawakatsu & Ecker, 2019; Singh et al., 2021; C. Sun et al., 2021; Turcotte et al., 2022; Varotto et al., 2020). A promising example of such opportunities is offered by the cassava case (see above and in Veley et al., 2023), where TAL20 binding to MeSWEET10a (a gene of susceptibility to Xanthomonas phaseoli) is inhibited by DNA methylation. It should be noted that before Veley et al. (2023) generated transgenic lines, they tested TAL20 sensitivity to DNA methylation in vitro (because the sensitivity cannot be generally assumed), and despite confirming that TAL20 is a methylation‐sensitive TF, the targeted hypermethylation of the MeSWEET10a promoter did not lead to a complete silencing of MeSWEET10a and only reduced, not prevented, the disease symptoms. When aiming for epigenetic applications, it needs to be recognized that, unlike MeSWEET10a, many genes could be entirely insensitive to DNA methylation changes. The ones that are responsive may not show on‐off variability, or may depend on additional epigenetic factors, like histone marks and others. Such a complex situation suggests that approaches of reverse epigenetics, where DNA methylation of specific genes is artificially modified in the hope of eliciting desired phenotypes, may have success rates too uncertain to be applied on an ad hoc basis. Instead, practical applications of epigenetic variants in breeding will more likely rely on forward epigenetics screens, which start with a discovery of an interesting phenotype within the standing variation of a crop or a population of epigenetic mutants (ideally, epigenetic recombinant inbred lines), followed by a search for the genetic and epigenetic determinants of that phenotype and subsequent functional validation. This view is consistent with the current track record of achievements in the field of epigenetics. While the Cnr and other epialleles were discovered through the forward (epi)genetics approach, we are still waiting for the first report of a bioengineered epiphenotype successfully applied in the fields or even introduced into the breeding process. Harnessing the diversity of DNA methylation in plant breeding, if possible and practical at all, will require a level of understanding of this epigenetic phenomenon that we simply do not have at this moment.

AUTHOR CONTRIBUTIONS

Peter Civan: Conceptualization; formal analysis; funding acquisition; investigation; methodology; visualization; writing—original draft; writing—review and editing. Iris Sammarco: Data curation; formal analysis; writing—review and editing. Meriem Banouh: Formal analysis; writing—review and editing.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Civan, P. , Sammarco, I. , & Banouh, M. (2026). The mirage of DNA methylation in transcriptional regulation of plants. The Plant Genome, 19, e70208. 10.1002/tpg2.70208