Methylome differences among barley inbreds and their association with genomic, transcriptomic, and phenotypic variation

Abstract

Epigenetic variation can play a crucial role in explaining the missing heritability of complex traits. To investigate genome-wide methylation in spring barley (Hordeum vulgare subsp. vulgare), we performed whole genome bisulfite sequencing on 23 parental inbreds from a community resource for genetic mapping. Our objectives were to characterize methylation variation, explore its association with single nucleotide polymorphisms (SNPs), and examine links to gene expression. The barley genome showed high average methylation levels of 88.6%, 58.1%, and 1.4% in the CpG, CHG, and CHH contexts, respectively. We identified nearly 500 000 differentially methylated regions (DMRs), with 64%, 64%, and 83% of DMRs in CpG, CHG, and CHH contexts, respectively, not associated with sequence variation. Around 6% of all DMRs showed significant associations with gene expression, with the direction of the correlations varying based on the DMR’s location relative to the gene with a recognizable pattern. We exemplified this association between DNA methylation and gene expression on the known flowering promoting gene VRN-H1, identifying a highly methylated epiallele linked to earlier flowering. Lastly, methylation improved the prediction abilities of genomic prediction models for various traits, outperforming models based solely on SNPs and gene expression. These findings emphasize the independent role of DNA methylation to sequence variation.

Methylome differences between spring barley inbred lines are associated with transcriptional changes, which can lead to phenotypic variation, independent of genetics.

Introduction

Barley (Hordeum vulgare subsp. vulgare) is one of the world’s oldest and most widely grown crops with worldwide importance for animal feed and alcohol production (Newton et al., 2011). In addition, it is not only important for global food security, but also serves as an important model organism for other cereals due to its smaller diploid genome and simple inbreeding genetics (Giraldo et al., 2019). This raises the need for a greater understanding of its genetics, genomics, and physiology to increase the yield and meet global demands. However, in any species, genetic variation alone cannot explain all of the phenotypic variation of any macroscopic or microscopic phenotype, a phenomenon known as missing heritability (Trerotola et al., 2015). Epigenetic variation was proposed as one of the possible molecular reasons for this (Cortijo et al., 2014).

Many definitions of epigenetics have been proposed over the years (Holliday, 2006; Deans and Maggert, 2015), but today it is commonly accepted that epigenetics summarizes those variations that are not necessarily associated with changes in the DNA sequence and includes chromatin and histone modifications, non-coding RNAs, and DNA methylation (Holliday, 2006; Bird, 2007; Dupont et al., 2009; Gallusci et al., 2017). One of the most researched aspects of epigenetics is DNA methylation, since evidence for its heritability is well-established in comparison with other epigenetic variations (Bird, 2007; Verhoeven et al., 2010; Stelpflug et al., 2014; Atta, 2015). DNA methylation is the addition of a methyl or hydroxymethyl group to the C5 position of cytosine forming 5-methylcytosine or 5-hydroxymethylcytosine, respectively (Kurdyukov and Bullock, 2016). In plants, methylation occurs in three sequence contexts: CpG, CHG (where H refers to any base but G), and CHH (Henderson and Jacobsen, 2007). De novo methylation of cytosines is catalysed by DOMAINS REARRANGED METHYLTRANSFERASE 2 (DRM2) (Law and Jacobsen, 2011). CpG methylation is then maintained by DNA METHYLTRANSFERASE 1, CHG by CHROMOMETHYLASE 3, and the asymmetric CHH methylation is maintained by continuous de novo methylation by DRM2 and CHROMOMETHYLASE 2. DNA methylation in plants largely targets repetitive elements like transposons (Law and Jacobsen, 2011; Eichten et al., 2016; Zhang et al., 2018). Genes are usually CpG methylated only, but not at transcription start sites (TSS) and transcription termination sites (TTS) where methylation is absent.

Investigation of the methylation landscape is important due to its commonly assumed regulatory functions. DNA methylation is considered to play an important role in plant development through zygotic gene expression regulation and imprinting, resulting in the differential expression of maternal or paternal alleles (Finnegan et al., 2000; Zhou et al., 2021). A conserved epigenetic modification can alter the state of chromatin and therefore influence chromosome interactions. For example, hypermethylation is associated with a decreased frequency of crossovers (Boideau et al., 2022). Methylation is associated with the silencing of transposable elements (TEs) and consequently contributes to the overall genome stability (Slotkin and Martienssen, 2007). Furthermore, earlier studies reported that promoter methylation is associated with the down-regulation of gene expression by inhibiting transcription activators, promoting repressors, or even influencing histone modifications (Bilichak et al., 2012; Zhang et al., 2018). On the other hand, gene body methylation is associated with an increase of gene expression, which is strongly conserved between orthologs and therefore a long-term property of evolutionary consequence (Takuno and Gaut, 2013; Zhang et al., 2018). For example, DNA methylation is associated with many abiotic stress responses in plants and is able to serve as a short- and long-term memory without sequence changes (Sun et al., 2022). Furthermore, such epialleles may be inherited independently from the DNA sequence (Eichten et al., 2014). It was hypothesized that barley may also regulate its drought and moisture stress responses by DNA methylation (Drosou et al., 2021; Falahi et al., 2021). However, the common assumption of gene expression changes being caused by DNA methylation changes is ultimately unproven (Secco et al., 2015). In addition, there are reports in the literature that imply that the causal relationship between DNA methylation and gene expression is in fact inverted. For example, a study of DNA methylation changes in rice suggested that environmentally induced gene expression changes are able to cause DNA methylation changes in nearby TEs at a later time point (Secco et al., 2015). In our study, however, we assessed environment-independent methylation differences as a heritable characteristic of inbred lines.

Finding differences in methylation between individuals is usually done by first testing for differentially methylated cytosines (DMCs) and then collapsing adjacent DMCs into differentially methylated regions (DMRs). However, DMRs can also be identified directly (Shafi et al., 2018). Interestingly, about half of the common DMRs in a maize study were not associated with local genetic state (Eichten et al., 2013). In a second study in maize (Xu et al., 2019), even 60% of the DMRs showed no association with sequence variation. Furthermore, these DMRs were strongly correlated with gene expression and even associated with phenotypic traits that could not be explained by single nucleotide polymorphisms (SNPs). Human genetic studies revealed that especially the combination of genetic and epigenetic information improves the phenotypic prediction score for complex traits (Shah et al., 2015). Studies with plants also indicate a huge benefit in the addition of epigenetics to phenotypic prediction for many important traits, especially seed quality, yield components, energy-use efficiency, and respiration, which are of high interest to plant breeders (Hauben et al., 2009; Shi et al., 2009 ; Long et al., 2011; Chen and Zhou, 2013). For example, in a study aiming to predict plant height of Arabidopsis, epigenetic variation explained 65% of the phenotypic variance (Hu et al., 2015). However, such information is not yet available for any small grain cereal.

Several studies of DNA methylation in barley are available. However, most of these studies focused on finding differences among two groups, typically a treatment and a control group. For example, the mediation of heat or drought stress through methylation has been frequently assessed in different barley varieties (Cai et al., 2021; Drosou et al., 2021; Sakai et al., 2022; Jiabu et al., 2023). While these studies provide insight into methylation changes in barley upon exposure to stress, they fail to characterize the methylome variation among inbred lines, as well as its association with genomic, transcriptomic, and phenotypic variation on a broad scale.

Initial efforts were made by Malinowska et al. (2020) to characterize methylome variation in barley, but the study was based on reduced representation bisulfite sequencing (RRBS) only covering 0.7% of the barley genome. The same limitation is also true for the study of Hansen et al. (2022). Therefore, in order to understand the genome-wide methylation variation and support quantitative trait locus cloning projects that rely on the alleles segregating among the 23 spring barley inbreds, the objectives of this study were to: (i) gain fundamental insight into the spring barley methylome through whole genome bisulfite sequencing (WGBS), (ii) characterize its variation among 23 parental inbreds of a community resource for mapping phenotypic traits, and (iii) assess the association of DMRs with SNPs and gene expression variation.

Material and methods

Plant material

From a total of 224 spring barley (Hordeum vulgare subsp. vulgare) accessions (Haseneyer et al., 2010), 23 were selected for maximal combined phenotypic and genotypic richness as previously described by Weisweiler et al. (2019).

These inbreds are the parental inbreds of the barley double round robin population (HvDRR; Casale et al., 2022). Seeds of the 23 inbreds were sown in a greenhouse in mid-July. Seedlings were placed in a vernalization chamber with 16 h of light (20 µmol m^–2 s^–1) at 4 °C for 5 weeks and subsequently repotted to ensure homogeneous growth. Tissue samples of 3×1 cm were collected from the youngest leaf 1 week after the transfer to the greenhouse. Apex samples were collected when stage 47 of the Zadoks scale was reached (Zadoks et al., 1974). Seedling samples were grown under the same conditions in Petri dishes and collected 5 d after germination.

Mixed tissue samples were used for HOR7985, HOR8160, Sissy, SprattArcher, Unumli-Arpa, and W23829/803911. The remaining inbred lines samples consisted only of leaf tissue (Supplementary Table S1). For Sissy, three additional samples were collected: one for leaf, apex, and seedling. The data set comprised biological replicates for the inbreds Sissy, K10693, HOR12830, and Unumli-Arpa.

Single nucleotide polymorphism and gene expression data

A set of 79 348 211 SNPs from whole genome DNA sequencing for the 23 barley inbred lines was available from Weisweiler et al. (2022). In order to reduce the covariance due to environmental factors between methylation and gene expression, we used two gene expression datasets that were available from Weisweiler et al. (2019), which were generated for the same genotypes but different experiments. This had the advantage that observed correlations are genotypic correlations as they are not caused by common environmental effects and, thus, avoid joint error covariance (Falconer and Mackay, 1983). The first dataset consisted of leaf expression data for 21 inbred lines, not including Kombyne and Sanalta. The second set consisted of seedling expression data for 21 inbred lines, not including IG128216 and Sanalta. Raw data were remapped against the Morex v3 reference and subsequently normalized using DESEq2 (Love et al., 2014). Finally, the gene expression for the missing inbreds was imputed with the mean gene expression of the 21 present inbreds per gene in both datasets individually, as is common practice in multi-omics prediction studies (Song et al., 2020).

Phenotypic data

Six phenotypic traits were assessed in seven environments (Cologne, 2017–2019; Mechernich and Quedlinburg, 2018–2019) in Germany. The 23 inbreds were sown as replicated checks in an augmented row–column design of other genetic material. The heading time trait was measured as days after planting. Plant height (cm) was measured after heading in Cologne and Mechernich. Seed area (mm²), seed length (mm), seed width (mm), and thousand grain weight were measured using grains from Cologne, 2017–2019, and Quedlinburg, 2018, with a MARVIN seed analyser (GTA Sensorik, Neubrandenburg, Germany) (Weisweiler et al., 2022).

Sequencing

Genomic DNA was extracted using the DNeasy Mini Kit Plant (Qiagen). Genomic DNA of 1 µg was submitted to mechanical shearing using a Covaris instrument with target fragment length set to 300–500 bp. Library preparation was performed with the NEBNext Ultra II DNA Library Prep Kit for Illumina kit (NEB). Methylated adapters were used to prevent conversion of adapters (NEB single methylated index, cat. no. e7535). Prior to PCR amplification, adapter ligated libraries were bisulfite treated using the EZ DNA-Methylation Lightning Kit (Zymo) following manufacturer’s instructions. The resulting paired end libraries were sequenced with the Illumina Hiseq 2000 and NovaSeq systems. For the second biological replicates of Sissy, K10693, HOR12830, and Unumli-Arpa, enzymatic methyl-seq tissues were homogenized in liquid nitrogen, and DNA was isolated with the DNeasy Plant Mini Kit (Qiagen). Libraries were prepared using the NEBNext Enzymatic Methyl-seq Kit (NEB) following the protocol for large DNA inserts. Therefore, 200 ng genomic DNA was combined with 0.002 ng CpG-methylated pUC19 DNA and 0.04 ng unmethylated lambda DNA. Fragmentation was done by using the Diagenode Bioruptor NGS in three rounds, 30 s on, 90 s off. The Agilent Technologies 4200 Tape Station was used to determine the size distribution and concentration of the libraries.

Bisulfite mapping

Raw reads were adapter and quality trimmed with Trimgalore (Krueger, 2015) and mapped with Bismark (Krueger and Andrews, 2011) using Bowtie 2 (Langmead and Salzberg, 2012), which is recommended especially for large genomes with many repetitive sequences (Grehl et al., 2020).

To increase the mapping efficiency, the reference sequence Morex v3 (Mascher, 2021) was SNP corrected with known SNPs for each inbred, respectively. Insertion and deletion correction was also evaluated but the increase in mapping efficiency was less than 0.5% compared with solely SNP correction (Supplementary Table S2). This was not enough to justify the added complexity of shifted genomic positions among the genotypes and was therefore not further evaluated.

The Bismark options --score_min L,0,−0.6 −X 1200 were selected to optimize the mapping efficiency without sacrificing quality (Supplementary Table S3). This decreased the minimum alignment score required and increased the maximum insert size of paired end reads (Espinas et al., 2020; Sarpan et al., 2020). The option -N 1, which sets the number of allowed mismatches to 1, was also evaluated but resulted in a worse mapping efficiency.

PCR duplicates were subsequently removed and DNA strands joined using the --comprehensive option.

Quality control

Identification of sample identity

To ensure that none of the samples had been swapped during library preparation, the identity of each sample was verified by comparing SNP calls in the bisulfite data with known SNPs. Firstly, SNPs were identified with Bis-SNP (Liu et al., 2012) using the uncorrected Morex reference sequence. SNPs were then filtered for shared genomic positions between both datasets. For each shared SNP, the allele calls were compared between the samples. If two samples have the same genotype, they should have the highest number of matching allele calls.

Conversion rate

Bisulfite conversion rates were assessed using standard protocols utilizing unmethylated chloroplast DNA and counting the number of methylated reads among all reads for each sample (Schmitz et al., 2022).

Reproducibility

In order to assess the technical error of our procedure, Pearson correlations of the proportion of methylated reads between the mixed Sissy inbred line sample and the weighted average across the three separate tissue samples of Sissy were calculated. Only sites with a minimum coverage of 5 in all samples were included.

The same procedure was chosen to assess the reproducibility with the biological replicates in a pairwise manner. Additionally, the below-describe DMR identification step was repeated one time considering the biological replicates and in another analysis ignoring them. DMRs of both runs were compared based on physical overlap. Furthermore, epiallele calls of the same inbred in both datasets were correlated as a measure of the reproducibility of the DMR identification step.

Differentially methylated region identification and characterization

DMRs were called using Methylscore at default settings (Hüther et al., 2022). Since methylscore does not natively support large chromosomes, a customized version was used. For the final analysis reported in the Results section, only one replicate was considered for each inbred in order to avoid bias in the results due to differences in the statistical power to detected features because of differences in the number of sequencing reads per inbred.

DMRs were classified as TE or genic when they overlapped to at least 50% of their physical length with TEs or annotated genes, respectively. The proportional overlap was calculated relative to the shorter feature so that the shorter length was the dividend. If DMRs intersected both genomic features simultaneously, they were classified as gene+TE. The remaining DMRs were classified as intergenic. The experiment was repeated a second time assessing overlaps of DMRs with specific TE classes present in the barley genome. The number of overlaps was normalized by the respective TE class frequency.

Population structure analysis

The population structure of the 23 spring barley inbreds was analysed with principal coordinates analyses (PCoAs) based on Euclidean distances of DMRs, SNPs, and both expression datasets separately. To quantify the similarities between the population structures derived from the different data sets, a generalized procrustes analysis (GPA) was performed with FactoMineR (Lê et al., 2008). Subsequently, 1−the procrustes similarity indexes were used as dissimilarity measurements in a final PCoA (Wu et al., 2022). DMR population structures were calculated separately for each sequence context for the GPA.

Local association

Local association among differentially methylated regions and single nucleotide polymorphisms

To identify DMRs with a significant association to SNPs, the approach of Eichten et al. (2013) was modified to save computation time. This procedure was applied to all DMRs for the three sequence contexts separately.

Firstly, DMRs were filtered for a maximum of 50% missing data. SNPs were filtered for a maximum of 20% missing data, maximum 20% heterozygosity, minimum minor allele frequency of 5%, and to include only biallelic SNPs. Each SNP in the region of any DMR ±10 kb was tested individually, in contrast to the study of Eichten et al. (2013), with a Wilcoxon rank-sum test, since the residuals were not normally distributed. To define a significance threshold, each DMR that had at least one SNP in the vicinity was tested with 10 regions of 100 random SNPs, which are not in proximity to any DMR. The significance threshold was set to the 5% quantile of all control P-values across the three methylation contexts. DMRs were filtered again for at least three significantly associated SNPs. For only those DMRs, 90 additional SNP regions were tested. The significance threshold was revised to the 1% quantile of all control P-values. For each DMR that still has at least three significantly associated SNPs, a ranking was conducted between the proportion of associated SNPs in its proximity and the proportion of associated SNPs in each control region. Only when the SNPs in the proximity of the DMR were in the top 5% of the ranking was the DMR classified as SNP associated.

Local association among differentially methylated regions and gene expression

Spearman’s rank correlation coefficients were calculated between the methylation levels of each DMR and the expression of the closest gene for those DMRs, where the variation of gene expression in inbred lines with the same epiallele was ≥0 for at least one epiallele. Association between DMRs and their closest gene expression state were performed for both expression datasets and the three sequence contexts separately. Benjamini–Hochberg false discovery rate control (FDR) was applied to find significant associations. Additionally, 1000 random DMRs were correlated with 1000 randomly selected genes for each context for comparison.

Genomic prediction

A linear mixed model was used to analyse each phenotypic trait across all seven environments and estimate adjusted entry means for all barley inbred lines:

$$y_{ijk}=\mu +E_j+G_i+(G\times E{)}_{ij}+{\mathrm{\epsilon }}_{ijk},$$

where y_ijk corresponds to the observed phenotypic value of inbred i in environment j of replicate k. The general mean is denoted as μ. G_i defines the genetic effect of inbred i, E_j the effect of the jth environment, and (G×E)_ij the interaction effect of the given inbred and environment. The random error is represented by ɛ_ijk. A variety of predictors was evaluated with respect to their performance to predict the adjusted entry means of all inbreds for each trait measured as the prediction ability (r), a Pearson correlation between the observed y and the predicted $\hat{y}$. The predictors for the genomic best linear unbiased prediction (GBLUP) (VanRaden, 2008) model were: an Illumina 50K barley SNP array (Bayer et al., 2017; Weisweiler et al., 2022), DNA sequencing SNPs (Weisweiler et al., 2022), seedling gene expression data (Weisweiler et al., 2019), and DNA methylation data from this study. Both the SNP array and DNA sequencing SNPs were filtered to remove markers which were monomorphic, had >20% missing data, and had a minor allele frequency ≤5%. Only biallelic SNPs were retained. Missing data were mean imputed. The DNA methylation data was filtered to remove sites with >20% missing data, monomorphic cytosine sites and was mean imputed.

W for predictor m had the dimensions of the number of barley inbreds (n=23) × the number of features of the given predictor (m_SNParray=38 285, m_SNPs=38 725 848, m_Expression= 43 769, m_{MethylationCpG}=238 842 649, m_{MethylationCHG}= 204 877 046, m_{MethylationCHH}=789 553 379). All W matrices were column centered, standardized to unit variance and denoted as W*. Additive relationship matrices were defined as:

$$G={\displaystyle \frac{W^{*}{W}^{*T}}{m},}$$

where W*^T was the transposed W* matrix. To combine the methylation information of the three sequence contexts into one G matrix, a joined weighted relationship matrix (Schrag et al., 2018) was formed using the number of cytosines as weights. As the focus of our study lay on the question of whether methylation can improve genomic prediction using multi-omics predictors, joined weighted relationship matrices of DNA sequencing SNPs and expression; DNA sequencing SNPs and methylation; and DNA sequencing SNPs, expression, and methylation were formed with all possible weight combinations ranging from 0 to 1 in steps of 0.1, where the summation of all weights needed to equal 1 for each respective combination.

For the investigation of the prediction ability, 200 five-fold cross-validation runs were used. The median correlation of the 5-folds was determined and the median of the median correlation across the 200 replicates was calculated as the prediction ability (Weisweiler et al., 2022).

Results

The methylation dataset consisted of >18.3 billion read pairs which were aligned to a SNP corrected Morex v3 reference sequence with customized parameters to increase the alignment rate (Supplementary Tables S2, S3). This resulted in >10.1 billion unique alignments with a mapping efficiency of 71.8%. The number of uniquely aligned read pairs per sample ranged from 104 118 586 to 619 615 095. A negative correlation of −0.39 between the mapping efficiency and the genetic distance of each inbred line and Morex was observed. The average conversion rate across all samples was ≥99%. The coverages of cytosine sites were on average across all samples 10.8, 11.0, and 11.7 in the CpG, CHG, and CHH sequence context, respectively. Since the main objective of our study was to assess methylation variation as a characteristic of genotypes instead of different tissues, Pearson correlation coefficients between a pool of three Sissy tissues and the average of the three individual tissues were calculated. The correlation coefficients were 0.86, 0.89, and 0.66 for the CpG, CHG, and CHH sequence context, respectively (Supplementary Fig. S1). To assess the reproducibility of the methylation levels in our study, a similar approach was chosen to correlate the genome wide methylation levels between the replicates of the same inbred. On average across the four inbreds for which biological replicates were available, correlation coefficients for the proportion of methylated reads of 0.91, 0.84, and 0.53 were observed in the CpG, CHG, and CHH context, respectively.

The barley methylome

The average methylation levels, defined as the proportion of methylated reads among all reads, were 89.3%, 57.9%, and 1.4% in the CpG, CHG, and CHH sequence contexts, respectively, across the 23 barley inbred lines. Mean methylation levels calculated in bins of 5 Mbp across 23 inbred lines tended to be lower at the distal regions of the chromosomes in the CpG and CHG context where the repeat content was also lower (Fig. 1). For CHH methylation, the opposite trend was observed. CpG methylation levels showed local minima in the centromeric region, while CHH methylation showed a minor peak in that region. The overall shape of the curves was largely uniform across the chromosomes with a mean standard deviation of 8.71%, 10.71%, and 0.28% in the CpG, CHG, and CHH context, respectively (Supplementary Fig. S2).

Fig. 1.

Mean methylation levels of the three sequence contexts calculated in bins of 5 Mbp across 23 inbred lines. The dashed gray lines indicate the centromere with the pericentromeric regions highlighted in gray (Casale et al., 2022). Proportions of genes and repeats were extracted from the Morex v3 GFF (Mascher, 2021).

In order to assess DNA methylation on a local level around genes, these were averaged in 100 bp bins between 10 kb up- and downstream of all TSSs and TTSs and then averaged across the inbreds (Fig. 2). In all contexts, DNA methylation minima were observed at TSSs and TTSs. The slope of methylation was generally steeper on the side before a TSS and after a TTS compared with after a TSS and before a TTS. The trend of the curves at TSSs and TTSs was largely symmetric. CpG and CHG methylation levels varied strongly between the genotypes around TSSs and TTSs, but the overall shape of the curves showed the same trend. CpG methylation displayed, in comparison with the other two sequence contexts, the steepest decline and increase right before and after a TSS and the steepest increase after a TTS. CHG methylation showed a less steep increase over a long genomic sequence after a TSS in comparison with CpG methylation. Interestingly, CHH methylation appeared to have a peak right before a TSS, followed by a minimum before returning to the normal level. The opposite trend was observed at a TTS. However, the peak in CHH methylation was more pronounced before a TSS than after a TTS.

Fig. 2.

Mean methylation levels of 10 kbp up- and downstream regions of transcription start sites (left) and transcription termination sites (right) calculated in bins of 100 bp across 23 inbred lines. Variability among the inbreds is illustrated by an approximation of the 95% confidence interval for each sequence context as the colored area.

Gene methylation and expression

To first investigate the association between the methylation level and the expression of genes in a single inbred, we categorized genes of the leaf expression dataset of Sissy based on their expression level into five groups (100% quantile≥High>75% quantile; 75% quantile≥Medium High>50% quantile; 50% quantile≥Medium Low>25% quantile; 25% quantile≥Low>0% quantile; None = 0) and divided each of these and also their corresponding 2 kb up- and downstream regions into 200 bins across their physical length. Subsequently, average methylation levels were calculated for each bin in the Sissy leaf tissue bisulfite data (Fig. 3). Highly expressed genes were strongly methylated in the CpG context and moderately methylated in the CHG context. They were also only slightly more highly methylated in the CHH context compared with less expressed or silenced genes. On the other hand, CpG and CHG methylation in the up- or downstream regions of the gene body were associated with a low expression. This was in contrast to CHH methylation, which was positively associated with an increase of expression. Interestingly, silenced genes showed the highest CHG methylation levels in the gene body region, while highly expressed genes were only slightly less methylated. The same results were obtained using the Sissy seedling gene expression and methylation data (Supplementary Fig. S3).

Fig. 3.

Average methylation levels of genes in Sissy leaf tissue categorized based on their expression levels in the same tissue (100% quantile≥High>75% quantile; 75% quantile≥Medium High>50% quantile; 50% quantile≥Medium Low>25% quantile; 25% quantile≥Low>0% quantile; None = 0).

Differential methylation

Among the 23 barley inbred lines, 244 689, 151 992, and 103 115 DMRs with a minimum of five Cs per DMR in the CpG, CHG, and CHH context, respectively, were discovered. Even though CHH sites were about 2.1 times more common than CpG and CHG sites together, they showed the least amount of differential methylation with a proportion of 20.63% of all DMRs. DMRs tended to be located at the ends of the chromosomes with up to 22.78-fold more DMRs compared with the pericentromeric regions (Fig. 4A). This was especially noticeable towards the 3′ end of the chromosomes. Nevertheless, a local maximum in the number of DMRs was found right at the centromeric sequence in all chromosomes when considering all three sequence contexts. DMRs of the CHG and CHH context were distributed similarly to that of the CpG contexts. However, CHH DMRs not only tended to be fewer, but were also generally shorter than DMRs of the other two contexts (Fig. 4C).

Fig. 4.

Distribution of differentially methylated regions. (A) Distribution of differentially methylated regions (DMRs) in bins of 5 Mbp across the chromosomes. The dashed gray lines indicate the centromeres with the pericentromeric regions highlighted in gray. (B) Proportional overlap of physical positions between DMRs of all three methylation contexts. (C) DMR length distribution on a logarithmic scale.

In addition to the correlation of methylation levels between the physically bulked Sissy tissues and the mean of the separate tissues, DMRs between the three tissue samples of Sissy were called. In comparison with the mean of 100 replications of three randomly selected unique combinations of barley inbred lines, the number of DMRs between the three tissues was about 224, 24 193, and 2 times lower in the CpG, CHG, and CHH context, respectively (Table 1). Additionally, the DMR calling step was repeated when including the biological replicates. In this case, about 88% of the DMRs in the CpG and CHG context overlapped with the DMRs detected when ignoring the replicates. A match of physical positions between the two DMR datasets with a tolerance of 100 bp was observed in about 81% of the cases in the CpG and CHG context. Furthermore, for each of the four inbreds we correlated the methylation level at the epiallele between the DMRs detected at the same physical position with consideration of the biological replicates and without. On average across the inbreds, a correlation coefficient of 0.99 was observed for both contexts. For CHH, the observed values were much lower implying that CHH methylation is mainly driven by environmental or error effects.

Table 1.

Summary of the number of differentially methylated regions (DMRs) between three tissues of inbred Sissy and the average of 100 random unique combinations of three inbred lines

Sequence context	DMRs
CpG
Tissues	303
Inbreds	67 790
CHG
Tissues	1
Inbreds	24 193
CHH
Tissues	11 830
Inbreds	27 425

DMRs were classified as TE, genic, gene+TE, or intergenic based on positional intersections (Fig. 5A) to investigate the amount of differential methylation in the genic and intergenic space for downstream analyses. The majority of all DMRs were classified as TE in the CpG and CHH context with a share of 45% and 70%, respectively. Intergenic was the second most common category with 40% and 25% for the CpG and CHH context, respectively. In contrast, CHG DMRs were most often classified as intergenic with a share of 48%, and TE was the second most common class with 40%. For CpG, the highest proportion of genic DMRs across the three sequence contexts was observed with 13%. Repeating the experiment with no minimum required overlap-size did not change the results considerably. Since the majority of DMRs were classified as TE, a more detailed analysis was conducted to determine which TE classes most frequently overlapped with DMRs. In total as well as in the CpG context, intersections with retrotransposons of the RXX class were the most common (Fig. 5B). In the CHG sequence context, intersections with transposons of the DHH class were more frequent, while DTX transposon intersections were the most common in the CHH context. Additionally, DMRs were found to be over-represented in miniature inverted-repeat transposable elements (MITEs) assessed with a binomial test (P<0.01) in all three contexts.

Fig. 5.

Transposable element annotation of differentially methylated regions. (A) Annotation of differentially methylated regions (DMRs) intersecting with genic, transposable element (TE), intergenic, or both TE and genic features for each sequence context. (B) Number of DMRs overlapping with specific TE classes normalized by the frequency of the given TE class in the barley genome.

Population structure analysis

To visualize and compare the information content of DMRs with the other omics data we calculated Euclidean distances from DMRs and subjected them to a PCoA (Fig. 6A). The axes explained 11.48% and 7.54% of the total variance. The resulting population structure formed three clusters. One cluster consisted of two-row landraces and cultivars, while six-row inbred lines were grouped in a separate cluster. The cluster in the center of the PCoA consisted of two- and six-row inbred lines. Inbred lines from the same country of origin mostly fell into the same cluster like the Syrian IG31424 and HOR12830, the Turkish HOR7985 and HOR8160, or the Indian Lakhan and Kharsila.

Fig. 6.

Population structure comparison. (A) Principal coordinate analysis (PCoA) of 20 inbred lines based on Euclidean distances determined by differentially methylated regions (DMRs). The percentage values on the axes refer to the proportion of variance explained by the respective axis. (B) PCoA comparing the information content of CpG DMRs, CHG DMRs, CHH DMRs, single nucleotide polymorphisms, and expression based on a generalized procrustes analysis.

Comparing the population structures derived from SNPs, the two gene expression datasets and DMRs, separate for each context, using GPA, revealed considerable differences (Fig. 6B). The DMR-based population structures clustered closely together, while the SNP and expression population structures were spread out. Interestingly, the distance between the SNP and expression population structures was smaller than the distance between each of the SNP and expression population structures and the DMR population structure. Furthermore, we observed a correlation of −0.07 between the distance matrices calculated from SNPs and DMRs using a Mantel test with 99 permutations.

Local association of differentially methylated regions

DMRs were subjected to local association analysis with SNPs and gene expression to investigate the association between genetic and epigenetic variation and transcriptomic and epigenetic variation. In the CpG, CHG, and CHH sequence context, 36.4%, 36.2%, and 16.6% of the DMRs, respectively, were significantly (P<0.01) associated with at least three local SNPs. SNP-associated DMRs largely accumulated at the ends of the chromosomes in all three sequence contexts, notably towards the 3′ end (Supplementary Fig. S4). This trend was similar to the distribution of all DMRs with an average Pearson’s correlation coefficient of 0.91 across all chromosomes and sequence contexts (Figs 4A, 7). In addition, we observed that the Kolmogorov–Smirnov test of uniformity of the number of SNP-associated DMRs corrected for the number of DMRs per bin was not significant (P<0.05) for all chromosomes and contexts. This shows, that when corrected for the total number of DMRs, the SNP-associated DMRs were distributed more towards the far chromosome ends than the total number of DMRs. Most notably, some chromosomes showed local maxima in the proportion of SNP-associated DMRs among all DMRs in the pericentromeric region, while others showed local minima even when corrected for the total number of DMRs (Fig. 7).

Fig. 7.

Percentage of single nucleotide polymorphism (SNP)-associated differentially methylated regions (DMRs) among all DMRs in bins of 5 Mbp across the genome. The dashed gray lines show the middle of the centromere with the pericentromeric regions highlighted in gray.

The distribution of the Spearman correlation coefficients between the methylation level at DMRs and the gene expression of the closest gene illustrated that only a low proportion of all genes were highly positively or negatively correlated with methylation across the inbreds, while most genes showed low to medium correlations in all sequence contexts (Supplementary Fig. S5). The same was also the case for 1000 random DMR correlations with 1000 randomly selected genes. Except for CpG and CHG, where a negative trend of correlation between gene expression and methylation at DMRs was found in close proximity to the TSS, all distributions were symmetric around 0 across the various distance groupings and did not differ from the random correlations. Only if significant correlations (FDR<0.05) were considered, the distributions showed a negative correlation trend in the CpG and CHG contexts upstream of, and most notably at, the TSS (Fig. 8). Correlations downstream of the TSS tended to be positive in the CpG context, but strongly negative in the CHG context up to 2 kb from the TSS. In the CHH context, all correlations between the DMRs and gene expression were strongly positive regardless of the distance to the TSS. However, a slight pattern can be recognized in stronger positive correlations closer to the TSS. Spearman’s correlation coefficient distributions of all contexts were similar between the leaf and seedling expression datasets.

Fig. 8.

Spearman’s rank correlation coefficient distributions of differentially methylated regions (DMRs) with gene expression filtered using the Benjamini–Hochberg false discovery rate control (FDR<0.05) and grouped by the DMR’s distance to the respective transcription start site (TSS) ±10 kb in intervals of 5 kb to 500 bp for the leaf and seedling expression data separately. Additionally, 1000 DMRs were randomly correlated with 1000 random genes, presented in the last category of each subplot. The numbers above the categories reflect the number of correlations, as well as the number of unique genes in the respective category.

Association of differentially methylated regions and phenotypic variation

We explored the physical overlap of known genes contributing to phenotypic variation of barley and the detected DMRs. One of the CHG DMRs with a high correlation of its methylation level with gene expression of +0.53 with one gene in the seedling expression dataset was indentified on chromosome 5H of the Morex v3 reference spanning the physical region of 528 153 135 to 528 153 251 in intron 1 of VERNALIZATION1 (Fig. 9). VRN-H1 encodes a MADS-box transcription factor that promotes flowering and is part of the main genes regulating vernalization response (Yan et al., 2003). We observed two epialleles at this DMR: epiallele 1 with a mean CHG methylation level of 73% and epiallele 2 with a mean CHG methylation level of 52% (Fig. 9A, B). The inbred lines carrying epiallele 1 showed an increased expression of VRN-H1 in both expression datasets. The DMR was not associated with SNPs or structural variants in the regulatory or coding region of VRN-H1 (Fig. 9D). Associations of SNPs and SVs within a ±5 kb region of VRN-H1 with the expression of this gene were not significant (P>0.05).

Fig. 9.

Differentially methylated region at VRN-H1. (A) Differentially methylated region (DMR) in an intronic region of VRN-H1. The DMR is highlighted with a colored background, where the inbred lines carrying the same epiallele share the same color. The red background corresponds to epiallele 1 with a mean methylation level of 73% and the blue background corresponds to epiallele 2 with a mean methylation level of 52%. For inbred lines with a white background color no epiallele could be assigned by Methylscore. The dots represent CHG sites colored according to their methylation level. The inbred lines are sorted by their expression of VRN-H1 in ascending order. (B) Average methylation level per CHG site for both epialleles. (C) Box plots of the CHG methylation level at the DMR, gene expression of VRN-H1 in the seedling dataset, gene expression of VRN-H1 in the leaf dataset, and adjusted entry means (AEM) for flowering time for inbred lines carrying either epiallele 1 or epiallele 2. (D) Histogram of P-values from Spearman correlations between single nucleotide polymorphisms (SNP) as well as structural variants (SV) within 5 kb of VRN-H1 and the DMR. The vertical red line indicates the significance threshold from the permutation test.

Methylation as a predictor of phenotypic variation

A SNP array, DNA sequencing SNPs, gene expression, and methylation data were used as single predictors to assess the median prediction ability for six traits in a GLUP framework. In addition, for the DNA sequencing SNPs, gene expression and methylation data, joined weighted relationship matrices were created with all possible weight combinations between 0 and 1 in steps of 0.1 to select the one with the highest median prediction ability (Fig. 10). The prediction abilities across all six traits ranged from 0.17 to 0.85. The joined relationship matrices of DNA sequencing SNPs, gene expression, and methylation (S+M+E) and the ones of DNA sequencing SNPs with methylation (S+M) had the highest prediction ability across the six traits followed by the SNP array and the DNA sequencing SNPs alone. The mean optimal weight across all traits in the S+M scenario was the highest for methylation with 0.63. In the S+E scenario, DNA sequencing SNPs had the highest mean optimal weight of 0.83. The highest mean optimal weight in the S+E+M scenario was observed for methylation with 0.52.

Fig. 10.

Prediction abilities for six traits from 200 five-fold cross-validation runs using either a single nucleotide polymorphism (SNP) Array, DNA sequencing SNP, combined SNP and methylation (S+M), combined SNP and gene expression (S+E), or combined SNP, methylation and gene expression (S+E+M) to establish relationship matrices among 23 barley inbred lines. The numbers below the combined datasets show the weights for the joined weighted relationship matrix with the highest prediction ability.

For the traits plant height and seed length, methylation outperformed any other predictor, which was indicated by a weight of 1 in the optimal joined relationship matrices. The prediction ability of plant height and seed length was improved by 0.1 and 0.16, respectively, using methylation data as a predictor compared with the SNP array.

For the traits seed area, seed width, and thousand grain weight, the combination of S+M outperformed the combination of expression data and DNA sequencing SNPs (S+E) and DNA sequencing SNPs alone with improvements ranging from 0.01 to 0.03 compared with the DNA sequencing SNPs. Instead, the prediction ability for the traits seed area and thousand grain weight was the highest using the SNP array. The prediction ability could not be improved by adding methylation data for the prediction of the trait heading time in comparison to S+E.

Discussion

From the observed high conversion rates of >99%, which are comparable to those reported in literature (Khodaeiaminjan et al., 2024; Wu et al., 2024), it can be concluded that the methylation data for all inbred lines is of high quality and that there is no bias due to the bisulfite treatment. In addition, we observed high and significant correlation coefficients among the biological and technical replicates which indicates a good reproducibility of this experiment in the CpG and CHG context. For the CHH context, the reproducibility was low, implying that CHH methylation is mainly driven by environmental or error effects (Supplementary Fig. S1). Furthermore, this observation of a high correlation between the methylation rates of a bulk sample and an artificial bulk sample created from individual tissue sequence data suggested that the bulking of tissue samples is a reasonable approach to assess a genotype specific methylation profile at moderate costs. This conclusion is also supported by the observation of several orders of magnitude less differential methylation among the tissues, especially in the CpG and CHG context, compared with the amount of differential methylation among the inbreds (Table 1). This illustrates that it is possible to investigate methylation differences of inbred lines even if the samples consisted of different or mixed tissues.

In our study, no full genome sequence was available for all 23 inbreds. Instead, we have used the SNPs of whole genome sequencing of the 23 inbreds to create a SNP-corrected reference sequence from the typically used barley reference sequence Morex (Mascher, 2021). This approach, however, did not fully remove the mapping bias. We observed a negative correlation between the mapping efficiency and the genetic distance of each inbred line and Morex of −0.39. However, as only sufficiently covered cytosine sites are considered for differential methylation analyses and the lowered mapping efficiency of distant inbreds affects both methylated reads and unmethylated reads equally, we expect that no methylation level bias is introduced by the differing relatedness of the characterized inbreds to Morex.

The barley methylome

The average methylation level of barley was exceptionally high when comparing it to Arabidopsis or rice (Cokus et al., 2008; Li et al., 2012). Compared with maize, the average methylation levels of barley in the CpG and CHG context were about 10% higher and in the CHH context 1% lower (Xu et al., 2020). The higher methylation levels in the CpG and CHG context can be explained by barley’s large genome with many repetitive elements (Fig. 1). However, these inter-species comparisons should be treated with caution as the library preparation as well as the analysis methods in our study are different from the studies mentioned above.

The results for the CpG and CHG context are in line with a previous RRBS study in barley (Malinowska et al., 2020). However, the average CHH methylation level observed in our study was about 48% lower compared with the study using the RRBS approach. This might be due to sampling bias as the previously mentioned study employing RRBS covered just 0.7% of the barley genome, whereas we used the gold standard approach of WGBS (cf. Yong et al., 2016).

Differential methylation

The approximation of the 95% confidence interval of methylation across inbreds indicates that the course of the methylation level around genes is the same among the inbred lines (Fig. 2). However, this analysis also suggested the presence of large differences in the extent of methylation among inbred lines, which will be discussed later. Explaining the variation of plant phenotypes is one of the key aspects of plant genetics (Acquaah, 2009). However, relying on sequence variation alone has resulted in missing heritability in previous research (Trerotola et al., 2015). As suggested by the tight relationship between methylation and gene expression within one genotype (Fig. 3; Supplementary Fig. S3), investigating the differences in DNA methylation may be one of the key aspects to unravelling more of the missing heritability (Cortijo et al., 2014; Trerotola et al., 2015).

More CpG DMRs were identified than CHG DMRs (Fig. 4A). The CHH context shows the fewest DMRs despite being the most frequent site (Fig. 4A). This can probably be accounted to CHH sites being largely unmethylated as well as being strongly driven by environmental and error effects (Fig. 1). The accumulation of DMRs towards the chromosome arms that was observed in our study (Fig. 1) might be explained by methylation being less conserved in the euchromatic chromosome arms between the inbred lines; for example, epimutation rates are more prevalent in euchromatic regions with many genic and intergenic sequences and depleted in pericentromeric regions (Johannes and Schmitz, 2019). A potential explanation might be that errors during meiosis lead not only to structural variants (Pokrovac and Pezer, 2022) but also differences in methylation. This hypothesis is supported by a very high average correlation between the number of DMRs and the recombination rate of 0.76 across all chromosomes and contexts (Casale et al., 2024, Preprint). Alternatively, accumulation of DMRs towards the chromosome ends could be a consequence of the selective activity of the RNA-directed DNA methylation pathway or potentially a mutual exclusiveness of DNA methylation and antagonistic histone variants like H2A.Z (Zilberman et al., 2008). However, why the positional bias of DMRs towards the chromosome arms cannot be observed in soybean and is either not present in maize or only to a far lower extent requires further investigation (Li et al., 2015 ; Shen et al., 2018; Xu et al., 2019; Fig. 4A).

Additionally, we characterized the DMRs based on their overlaps with genic, intergenic, or TE features. CpG DMRs were located more frequently at annotated genes than DMRs of the other contexts. This is likely because genes are predominantly CpG methylated (Zhang et al., 2018). DMRs of all three sequence contexts largely targeted TEs, especially of the RXX class, which is unsurprising given their high methylation ratio and the overall high repeat content in the barley genome (Fig. 5A, B). This association of DMRs and TEs was also observed in other crops, as DNA methylation is a key factor regulating their expression (Bestor, 1990; Li et al., 2015; Shen et al., 2018). Additionally, the over-representation of DMRs at MITEs, a known TE superfamily involved in post transcriptional regulation of gene expression (Kuang et al., 2009; Yan et al., 2011; Pegler et al., 2023), of all methylation contexts may hint that DMRs serve as an additional layer of gene expression regulation that activates or deactivates MITE transcription.

We observed local maxima of DMRs at the centromeres that might be explained by sequence variation among the studied inbred lines. In rice, functional centromeres were either hyper- or hypomethylated compared with the pericentromeric regions, which may be the result of sequence variation as hypothesized by Yan et al. (2010). This means, depending on the sequence composition of the inbred line, different centromeres can either have a higher or lower methylation level compared with the pericentromeric regions in the same cultivar. As local maxima of sequence variation between the same barley inbred lines were previously reported at centromeres (Weisweiler et al., 2022), it may be possible that in our study the sequence variation among the inbred lines is a cause for the local maxima of DMRs at the same centromeres. This hypothesis is supported by the high amount of differential methylation that targeted repeat rich genome regions (Fig. 5A), but this still requires further research.

Linkage disequilibrium of differentially methylated regions and single nucleotide polymorphisms

Investigating the linkage disequilibrium of DMRs and SNPs is crucial to understand the independence of epigenetic variation from sequence variation as only such variation is able to explain the missing heritability. In barley, 32% of all DMRs across all three contexts were associated with SNPs (Fig. 7; Supplementary Fig. S4). Similar results were observed in other crops like maize with less than 40% SNP-associated DMRs and soybean with about 23% (Shen et al., 2018; Xu et al., 2019). We observed that SNP-associated DMRs distributed more towards the far chromosome arms than the total amount of all DMRs (Fig. 7). However, unraveling the reasons for the local minima of SNP-associated DMRs in the pericentromeric regions of some chromosomes, while others show local maxima, requires further research.

The low amount of SNP-associated DMRs highlights that there is a considerable difference in information content derived from DNA methylation compared with sequence variation. This hypothesis is also supported by the GPA of population structures derived from DMRs and SNPs (Fig. 6B), where these differences in information content can be observed. This is further supported by the observation of a correlation of −0.07 between the distance matrices calculated from SNPs and DMRs. A low level of linkage disequilibrium between SNPs and DMRs emphasizes that DMRs reveal new levels of genotypic information not accessible by other means. As we observed a tight relationship between the gene expression and the extent of methylation within one genotype, as previously discussed, we were interested in understanding the predictive power of methylation differences among inbreds on the respective gene expression.

Association of differentially methylated regions with gene expression

The 23 inbred lines showed large differences in their methylation level around TSSs and TTSs (Fig. 2). We observed that most DMRs at genic loci were of the CpG context (Fig. 5A), as described previously. However, when looking at the number of significant associations of DMRs with the expression of the closest gene, most of them were in the CHG context (Fig. 8). This highlights that the relevance of CHG methylation for the regulation of gene expression, at least in barley, might be higher than previously described in literature (Bewick and Schmitz, 2017; Muyle et al., 2022).

When considering the association of DMRs with the expression levels of adjacent genes, it can be concluded that these differences have only minor effects on gene expression (Supplementary Fig. S5). Only when considering significant associations were strong correlations between DMRs and gene expression observed (Fig. 8). However, significant associations between DMRs and the gene expression of the corresponding gene were only observed for 5.69% out of all annotated genes across the seven chromosomes, which is of the same order of magnitude as the 3.38% observed gene–DMR associations in maize (Xu et al., 2019). The percentages of DMRs with gene expression associations were 1.3% and 1.7% for the leaf and seedling expression dataset, respectively. This illustrates, that the effect of methylation on gene expression across inbred lines is limited and suggests that other cis- (Engelhorn et al., 2025) or possibly trans-effects (West et al., 2007) are responsible for the differential expression of the majority of genes. However, this requires further research.

As described above for barley (Fig. 3) but also earlier in literature (Zhang et al., 2018; Cai et al., 2021), a pattern of promoter methylation in the CpG and CHG context leading to down-regulation of expression emerges when considering different genes in a single genotype, while CHH methylation results in the contrary. When observing the methylation differences and gene expression variation of the same genes across a diverse set of inbreds instead, a similar, but more specific and spatially confined, pattern was recognized. A slight trend towards negative correlations was observed between the extent of CpG methylation between −10 kb and −500 bp upstream of the TSS. Right at the TSS, a strong negative correlation was observed. In contrast, in the downstream regions of the TSS from 500 bp to 5 kb, a slight trend towards positive correlations can be observed. These observations are in contrast to the observations made across maize inbreds, where methylation at CpG DMRs associated with repressed expression in gene body regions (Xu et al., 2019).

The trend of association between methylation at CHG DMRs was largely consistent with that of CpG DMRs in the upstream region of the TSS from −5 kb to −500 bp and at the TSS from −500 to 500 kb. However, the trend of association between methylation at CHG DMRs in the downstream regions from 500 bp to 5 kb was strongly negative and consistent with the findings in maize (Xu et al., 2019). This may be explained by the association of H3K9me2 with CHG methylation, though this is a subject for further investigation due to conflicting studies in the past (Bernatavichute et al., 2008; Bewick and Schmitz, 2017).

Methylation at DMRs of the CHH context were positively correlated with gene expression throughout all distance groupings. This trend towards positive correlation was less prominent in the distal regions from the TSS and consistent with the observations in one barley inbred line and with previous observations in maize (Xu et al., 2019).

In conclusion, differences in gene expression are associated with more specific spatially confined differences in methylation across inbred lines than across different genes in one inbred line, especially in the CpG and CHG context.

Association of differentially methylated regions and phenotypic variation

Assessing the association of DMRs as a characteristic of genotypes with phenotypic variation is an important step to understanding the origin of the missing heritability better. In the frame of this study, we want to consider VRN-H1 as an example for this association. It is well known that VRN-H1 is a key regulator of flowering time in cereal crops and one of the main actors in the vernalization response (Yan et al., 2003). The vernalization-induced transcription of VRN-H1 is mediated by epigenetic regulation involving changes in chromatin state, through particular modifications in the pattern of histone methylation (Oliver et al., 2009). Previous research illustrated that a large deletion of around 5 kb in an intronic region of VRN-H1 in spring barley compared with winter barley is one of the reasons for an increased basal gene expression of VRN-H1 in non-vernalized seedlings (Oliver et al., 2009). However, the study of Hemming et al. (2009) illustrated that at least 10 alleles of VRN-H1 exist that contain different deletions or insertions in the first intron and that these alleles were associated with differing levels of VRN-H1 expression in non-vernalized plants as well as with different flowering behavior. Our observation of a positive correlation between the methylation level at a DMR in intron 1 of VRN-H1 with the gene expression suggests that not only a deletion in the first intron determines the expression of VRN-H1 among inbred lines, but also DNA methylation variation might contribute (Fig. 9). We identified within spring barley two epialleles where an increased CHG methylation was associated with a significantly increased VRN-H1 expression and therefore earlier flowering time (Fig. 9C). It is noteworthy that this DMR was not associated with local sequence variation (Fig. 9D). The regulation of flowering time by non-CG methylation has already been revealed in winter wheat as a consequence of vernalization treatment (Khan et al., 2013). Therefore, we speculate that non-CG DNA methylation may regulate the basal expression of VRN-H1 as a characteristic of spring barley genotypes contributing to earlier flowering times of inbred lines carrying epiallele 1. Previous research with the inbred lines of our study and the derived segregating populations has already reported a flowering time quantitative trait locus whose confidence interval included VRN-H1 in the progeny of the monomorphic parents IG31424, carrying epiallele 2, and Kharsila, carrying epiallele 1 at VRN-H1 (Cosenza et al., 2024). Both parents were monomorphic for the intron allele HvVRN-H1-1 of VRN-H1.

Methylation as a genomic predictor

Incorporating methylation information in genomic prediction has the potential to explain more of the missing heritability, and thereby provide a wider understanding of the causes of phenotypic variation, as methylation is largely independent of sequence variation (Fig. 7) and also partially independent of gene expression variation (Fig. 8). The prediction ability of DNA methylation data in genomic prediction models in barley was largely dependent on the predicted trait (Fig. 10). About half of the examined phenotypic traits showed noteworthy improvements by using methylation information as a predictor. These differences are presumably such that the traits differ in the importance of polymorphisms in the coding sequence versus gene expression effects. Nevertheless, our results highlight that especially the combination of DNA sequencing SNPs and methylation is valuable for many traits and performs generally better than the combination of DNA sequencing SNPs and gene expression. These observations in barley are in line with previous studies (Hu et al., 2015; Xu et al., 2019; Amiri Roudbar et al., 2020; Malinowska et al., 2022).

Supplementary data

The following supplementary data are available at JXB online.

Fig. S1. Correlation of the average methylation levels across the Sissy tissue samples and the bulked Sissy sample.

Fig. S2. Average chromosome methylation of each sequence context.

Fig. S3. Average methylation levels of genes in Sissy’s seedling tissue categorized based on their expression levels.

Fig. S4. Distribution of SNP-associated DMRs across the genome.

Fig. S5. Correlation coefficient distributions of DMRs with gene expression grouped by distance to the TSS.

Table S1. Barley inbred lines that were part of this study.

Table S2. Average mapping efficiency of the barley inbreds against the uncorrected, SNP corrected, and SNP and Indel corrected Morex reference.

Table S3. Average mapping efficiency of the 23 barley inbreds under varying parameters.

Author contributions

BS: conceptualization; MK, BS, and PW: methodology; MK, BS, and AS: investigation; JE, SM, and TH: data generation; MK: writing—original draft; MK, PY, AS, JE, SM, TH, and BS: writing—review & editing; MK: visualization; BS: supervision; BS: funding acquisition.

Funding

This research was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy—EXC 2048/1—Project ID: 39068611 and core funding of the Julius Kühn Institute.

Data availability

The data that support the findings of this study are publicly available from NCBI with the identifier PRJNA1100572.