# Source paper archive - Source title: DNA demethylases constrain multigenerational methylome and transcriptome divergence - Author(s): Dmitrieva, A. D., Christensen, K. I., Williams, B. P. - Source landing URL: https://www.biorxiv.org/content/10.64898/2026.10.29.762914v1 - PDF URL: https://www.biorxiv.org/content/10.64898/2026.10.29.762914v1.full.pdf - PDF SHA-256: e8782b838b1f5d9788d110c3955cb952f168705d91abe3e7f57c62dab3e27022 - Converter: pymupdf4llm 0.0.27 - Limits: Automated PDF text extraction; reading order, equations, tables and formatting may be incomplete or incorrect. Images are omitted, and no OCR was performed. Markup and links are neutralized for safe downloading. Consult the original PDF for the authoritative version. --- [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 1 **TITLE:** 2 **DNA demethylases constrain multigenerational methylome and transcriptome** 3 **divergence** 4 **Authors:** [1] Anna D. Dmitrieva, [1] Kate I. Christensen, [1,2] Ben P. Williams* 5 **Affiliations:** [1] Department of Plant & Microbial Biology, University of California, Berkeley, CA 6 94720, USA. [2] Innovative Genomics Institute, University of California, Berkeley, CA 94720, 7 USA. 8 9 *Correspondence: [benwilliams@berkeley.edu](mailto:benwilliams@berkeley.edu) 10 **Abstract** 11 In plants, animals and fungi, DNA methylation is important for the silencing of repetitive DNA 12 and impacts the expression of a subset of genes. DNA methylation landscapes exhibit a high 13 degree of stability over generations, yet this stability is the product of multiple dynamic 14 methylation and demethylation pathways. How these dynamic pathways function over longer 15 multigenerational timescales is unclear. To address this, we inbred quadruple DNA demethylase 16 Arabidopsis mutants and closely related wild type lines for nine generations by single-seed 17 descent. In inbred mutant lines, methylation gains accumulated across generations, with regions 18 of preexisting methylation becoming progressively reinforced or spreading to proximal 19 sequences. Additionally, each inbred line possessed hundreds of de novo methylation gains in 20 unmethylated regions. Two-thirds of all methylation gains and 85% of de novo methylated loci 21 were unique to only one inbred line, suggesting a high degree of stochasticity in 22 multigenerational methylation dynamics. Individual inbred lines also exhibited substantial 23 divergence in both their transcriptomes and small RNA profiles. Our data lead to a model in 24 which DNA demethylases constrain the epigenome and transcriptome space accessible over 25 multigenerational timescales. Long-term loss of DNA demethylase activity causes lineages to 26 drift further apart to more diverged epigenetic landscapes. 27 28 **Author summary** 29 DNA methylation is a chemical mark added to DNA that can impact gene expression. The 30 patterns that DNA methylation marks form across DNA are maintained by enzymes that add 31 and remove them. We wanted to understand the role that DNA methylation removal plays in 32 keeping methylation patterns stable over generations. We followed several genetically identical 33 lineages of Arabidopsis, a small flowering plant, over nine generations. In plants lacking four 34 enzymes that remove DNA methylation marks, methylation gradually accumulated over time. 35 Some existing patches of methylation spread into neighboring regions, and new patches 36 appeared in places that normally lack DNA methylation altogether. Different lineages 1 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 37 accumulated mostly different methylation patterns, suggesting that where DNA methylation 38 appears over time is mostly down to chance. Independent lineages also differed in their gene 39 expression patterns, with thousands of genes showing altered expression levels. Our findings 40 suggest that methylation removal limits the range of methylation patterns that can build up over 41 generations in plants. Preserving inherited methylation patterns therefore requires a mechanism 42 of methylation removal, which prevents new marks from taking hold in the wrong places. 43 **Introduction** 44 Across diverse eukaryotes, DNA methylation provides a critical layer of epigenetic regulation. 45 DNA polymerase cannot incorporate methylated bases into newly synthesized strands. DNA 46 [methylation patterns therefore require continual maintenance by DNA methyltransferases [1].](https://www.zotero.org/google-docs/?wAZrE2) 47 Despite this, DNA methylation landscapes can be stably propagated over 100 to thousands of 48 [years [2]. Mutation accumulation lines in](https://www.zotero.org/google-docs/?oL2PPR) _Arabidopsis thaliana_ (Arabidopsis), maintained for 30 49 generations, showed that over 99.998% of the methylated regions remained stably inherited 50 [between generations [3]. Such stability in the genomic distribution and levels of DNA](https://www.zotero.org/google-docs/?k13xIc) 51 methylation enables it to accurately perform its key functions: silencing transposons, regulating 52 [gene expression, and establishing imprinting [4]. However, this long-term stability does not](https://www.zotero.org/google-docs/?FF8AB9) 53 reflect biochemical inactivity or a lack of arising epimutations but rather results from continual 54 active maintenance of stable epigenetic states by specialized pathways, a concept called 55 [“epigenetic homeostasis” [1].](https://www.zotero.org/google-docs/?Wti2cf) 56 In plants, DNA methylation occurs at cytosines in CG, CHG, and CHH sequence contexts (H = 57 A, T, C). Methylation is established de novo in all three contexts by the RNA-directed DNA 58 methylation (RdDM) pathway, in which methylation is guided to specific genomic targets by 59 small RNAs (sRNAs) [[5]. Once established, methylation is maintained primarily by context-](https://www.zotero.org/google-docs/?5uHrk7) 60 specific methyltransferases, while CHH methylation must be continually re-established _de novo_ 61 either by CMT2 (primarily active in heterochromatin) or RdDM (primarily active in euchromatin) 62 [[4,6]. Methylation maintenance is aided by self-reinforcing interactions between DNA](https://www.zotero.org/google-docs/?9fQvg4) 63 methylation, the H3K9me2 histone modification, and small interfering RNAs (siRNAs) promoting 64 [repeated recruitment of methylation machinery to the same loci [7]. Although these feedback](https://www.zotero.org/google-docs/?WWdZ6a) 65 mechanisms are essential for continued silencing of transposons and other repetitive 66 sequences, they can also promote ectopic methylation or methylation spreading from the 67 [established targets into neighboring DNA, including genes and their regulatory regions [8].](https://www.zotero.org/google-docs/?jhiGMX) 68 Mechanisms that remove excessive methylation are therefore essential for enforcing 69 methylation boundaries and protecting gene regulation. 70 Active DNA demethylation contributes to the maintenance of epigenetic homeostasis by 71 counteracting the activities of DNA methyltransferases. In Arabidopsis, it is performed by four 72 DME-family 5-methylcytosine DNA glycosylases: DEMETER (DME) [ [9]](https://www.zotero.org/google-docs/?49JEqH), REPRESSOR OF 73 SILENCING 1 (ROS1) [[10]](https://www.zotero.org/google-docs/?p1dXM5), and DEMETER-LIKE 2 and 3 (DML2/3) [[11](https://www.zotero.org/google-docs/?1ptGfv) ], which prevent 74 genome-wide hypermethylation. The _ros1_ mutant and _rdd_ ( _ros1 dml2 dml3_ ) triple mutant show 75 thousands of hypermethylated loci, with many demethylase targets also targeted by RdDM [[11–](https://www.zotero.org/google-docs/?vLyZF9) 76 [13].](https://www.zotero.org/google-docs/?vLyZF9) To assess the contribution of DME to maintaining genomic methylation patterns in somatic 2 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 77 tissues, we previously generated a somatic quadruple _drdd_ mutant by rescuing _DME_ expression 78 specifically in the central cell, where its expression is required for reproductive development 79 [[14].](https://www.zotero.org/google-docs/?kVQjUP) Methylome analysis in the _drdd_ quadruple mutant showed exacerbated hypermethylation 80 compared to the _rdd_ triple mutant, including previously undetected hypermethylation inside gene 81 [bodies [14].](https://www.zotero.org/google-docs/?DC9ccL) 82 The trajectory, predictability, and effects of methylation change during prolonged loss of active 83 demethylation remain underexplored. Further, it is unknown whether hypermethylation arises 84 predominantly through reinforcement and expansion of previously methylated regions or 85 through _de novo_ methylation gains. A previously published multigenerational inbreeding 86 experiment profiled the _ros1_ methylome in generations 1 and 10 and found an increased rate of 87 epimutation accumulation compared to wild-type ( _DRDD_, hereafter referred to as “WT”) lines 88 [[15]. In this study, we assess the role of DNA demethylases in the long-term stability of DNA](https://www.zotero.org/google-docs/?KkJcyo) 89 methylation inheritance by propagating multiple _drdd_ and WT lineages for 9 generations by 90 single-seed descent. We performed methylation profiling at multiple generations, as well 91 as transcriptome and small RNA (sRNA) profiling at the final (9th) generation. Our results show 92 that independent lines accumulate lineage-specific gains of ectopic methylation, leading to 93 divergent methylomes and transcriptomes. New hypermethylation is almost entirely driven by 94 RdDM-pathway activity, occurring both proximally to existing RdDM targets and de novo in 95 unmethylated regions. Our study therefore uncovers long-term RdDM dynamics and the 96 epigenomic consequences of DNA demethylation over multiple generations of inheritance. 97 **Results** 98 **Inbred quadruple demethylase mutants gain methylation at thousands of intergenic** 99 **sequences** 100 In this study, three _drdd_ [and three WT segregant lines [14] were propagated by single-seed](https://www.zotero.org/google-docs/?PJElOT) 101 descent up to generation 9 (Fig. 1A, see Methods). We hypothesized that, without the 102 homeostatic activity of DNA demethylases, the _drdd_ lines would progressively gain methylation 103 over generations. To test this, we performed enzymatic methyl-seq (EM-seq) at the 5th, 7th and 104 9th inbred generations and identified differentially methylated regions (DMRs) relative to a 105 generation 2 ancestor (see Methods). One of the _drdd_ lines had to be terminated after 106 generation 7 as it gradually developed increasingly detrimental phenotypes, with few to no 107 seeds produced. Inbred _drdd_ lines showed a clear progressive increase in CG and CHG 108 hypermethylated regions (hyperMRs) over time, with the number of hyperMRs vs the 2nd 109 generation increasing in later generations (Fig. 1B). Conversely, WT inbred lines gained on 110 average ~6.6 times fewer CG and ~16.6 times fewer CHG hyperMRs than _drdd_ by the 9th 111 generation. Inbred _drdd_ lines also had some hypomethylated regions (hypoMRs). In the CG and 112 CHG contexts, there were 15.5-fold and 9.5-fold fewer hypoMRs than hyperMRs, respectively, 113 in the 9th generation (Fig. S1). 114 3 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 115 We next wanted to assess the chromatin landscapes where hyperMRs are most prevalent. We 116 [condensed previously characterized chromatin states [16](https://www.zotero.org/google-docs/?kYQLVn) ] into five categories and assessed the 117 proportion of hyperMRs overlapping each. In WT inbred lines, hyperMRs were most enriched in 118 gene bodies for CG methylation, consistent with the known patterns of epimutation described in 119 WT inbreeding experiments [[17] (Fig. S2). Additionally, CHG and CHH hyperMRs in WT were](https://www.zotero.org/google-docs/?yfNroK) 120 most enriched in heterochromatin (Fig. S2) . In contrast, hyperMRs in all three _drdd_ inbred lines 121 were more strongly enriched in non-heterochromatin intergenic regions and promoter regions 122 (Fig. 1C-D). This is substantially different from both the distribution of naturally occurring 123 epimutations that occur in WT, as well as the genomic distribution of chromatin states. 124 Together, these analyses demonstrate that inbreeding _drdd_ results in the accumulation of a 125 large number of methylation gains that are mostly distinct from the natural baseline processes of 126 epimutation. 127 Because several DNA methylation pathways rely on self-reinforcing loops, we wanted to assess 128 whether the methylation gained in _drdd_ over the course of inbreeding was largely spreading 129 from existing regions of methylation. We assigned hyperMRs to four classes based on their 130 location relative to methylated regions in a second-generation ancestor methylome (see 131 Methods). These classes are: “contained” for hyperMRs fully overlapping previously methylated 132 regions, “spreading” for hyperMRs located <= 600 bp immediately adjacent to ancestral 133 methylation, “de novo _”_ for hyperMRs >1200 bp away from previous methylation in any context, 134 and “ambiguous” for hyperMRs not matching any of the previous conditions (Fig. 2, A-B). 135 Between 50.8 and 74.3% o f all _drdd_ hyperMRs in generation 7 were classified as either 136 “contained” or “spreading,” suggesting methylation gains at existing methylated regions (e.g. a 137 higher percentage of methylated reads compared to the ancestor), or gains at cytosines 138 immediately adjacent to ancestrally methylated regions (Fig. 2C). However, we also identified 139 14.1 – 31.2% of all generation 7 _drdd_ hyperMRs as de novo (Fig. 2C), occurring in r egions 140 lacking DNA methylatio n in the ancestor. Notably, de novo hyperMRs were very rare in inbred 141 WT control lines, suggesting their presence is dependent on the loss of demethylation by 142 DRDD. These de novo hyperMRs in _drdd_ are completely absent in heterochromatin and have a 143 higher enrichment in non-heterochromatin intergenic regions compared to the genomic 144 distribution of all hyperMRs (Fig. S3). 145 To understand the stability and fate of acquired hyperMRs, we assessed methylation profile 146 dynamics at the hyperMRs over generational time. We examined the methylation profiles of 147 _drdd_ lines at each generation over generation 9 hyperMRs to trace how their methylation 148 changed over time (Fig. 2D). Both de novo and spreading hyperMRs exhibited progressively 149 increasing methylation levels from generation 2 onwards through inbreeding (Fig. 2D, S4). 150 Consistently, de novo hyperMRs exhibited an increased average percentage of methylation 151 over generations, whereas spreading hyperMRs exhibited a clear increase in the spreading 152 distance as generations progressed (Fig. 2D). This suggests that once gained, these newly 153 methylated regions were inherited and/or reinforced in subsequent generations. 4 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 154 **De novo methylation gains are RNA-directed** 155 Methylation spreading outwards from previously methylated regions in DNA demethylation 156 [mutants is a previously described phenomenon attributed to RdDM activity [11,18,13]. To](https://www.zotero.org/google-docs/?vXQXWX) 157 examine if RdDM could also be the mechanism underpinning the formation of spontaneous de 158 novo hyperMRs, we assessed the sequence contexts of methylation gains in these regions. We 159 found that on average 73.7% of all _drdd_ de novo hyperMRs in generation 7 had a methylation 160 gain of 5% or more in at least one other context (Fig. 3A). As multi-context methylation gains 161 are a hallmark of RdDM activity, we hypothesized that sRNAs are likely being produced to 162 target methylation to these regions. To test this, we performed sRNA-seq on the 9th inbred 163 generation of _drdd_ and WT control lines. We found that most frequently, _drdd_ de novo 164 hyperMRs were associated with _drdd_ -specific sRNA clusters which are absent in WT lines, 165 while other hyperMR classes had much higher overlap with WT sRNA clusters as well (Fig. 3, 166 B-C). This relationship was especially strong for CHH _drdd_ de novo hyperMRs where 81.4-91% 167 were associated with new sRNA clusters. Canonical RdDM relies primarily on 24-nucleotide (nt) 168 [siRNAs as guides and also generates 23 nt siRNAs [19,20]. Indeed, on average 92.8% of the](https://www.zotero.org/google-docs/?AKYwLt) 169 sRNA reads underlying _drdd_ de novo hyperMRs are 23–24 nt long, with around 74.7% being 24 170 nt (Fig. 3D). These data therefore strongly suggest that the RdDM pathway is the primary 171 mechanism contributing to de novo methylation gains in inbred _drdd_ . 172 173 **De novo hyperMRs occur in regions associated with Pol V activity in wild type Col-0** 174 In addition to siRNAs, RdDM relies on long noncoding RNA scaffolds that enable recruitment of 175 complementary AGO4-bound siRNAs to loci targeted for DNA methylation. These RNA 176 [scaffolds are synthesized by a specialized plant RNA polymerase, polymerase V (Pol V) [5].](https://www.zotero.org/google-docs/?vgUnCW) 177 Recently, low levels of Pol V activity have been detected at a large portion (potentially as high 178 [as 30%) of the wild type Col-0 genome [21], including many “non-RdDM” regions that are not](https://www.zotero.org/google-docs/?AURzTP) 179 associated with sRNAs or non-CG methylation. We therefore reasoned that these regions of the 180 genome exhibiting Pol V activity may have a higher probability of nucleating new RdDM sites in 181 inbred _drdd_ . We intersected each hyperMR class with previously published regions predicted to 182 [have Pol V transcription in Col-0 [21]. More than two thirds of de novo hyperMRs overlapped](https://www.zotero.org/google-docs/?ib6Fjn) 183 non-RdDM Pol V-transcribed regions (Fig. 4A-B), supporting the hypothesis that pre-existing 184 wild type Pol V activity in these regions may prime them for de novo methylation in _drdd_ 185 mutants. Removal of methylation by DRDD may explain the lack of RdDM occurring at these 186 sites in wild type Col-0. As expected, contained hyperMRs (which directly overlap previously 187 methylated regions) and spreading hyperMRs (which are adjacent to methylated regions) were 188 associated with regions of Pol V activity that are identified as RdDM targets in Col-0 (Fig. 4A). 189 These data suggest that Pol V activity may be one of the key factors that predisposes a region 190 to gain of new RdDM activity in _drdd_ mutants, and that regions with Pol V activity already 191 detectable in wild type Col-0 are likely locations for hypermethylated regions during _drdd_ 192 inbreeding. 5 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 193 **Loss of DRDD leads to long-term methylome and transcriptome divergence** 194 Given that multiple independent _drdd_ lines had a similar distribution of hyperMR classes (Fig. 195 2C), similar underlying RNA distributions (Fig. 3C) and Pol V activity in WT (Fig. 4A), we 196 predicted that independent _drdd_ lines may share many common hyperMRs and exhibit 197 predictable gains of methylation at the same loci. To test this hypothesis, we plotted average 198 methylation over a superset of all hyperMRs ever called, between all lines and generations. In 199 contrast to our initial hypothesis, methylation levels at all hyperMRs across different _drdd_ lines 200 showed distinct patterns (Fig. 5A, S 5), suggesting that methylation was initiated at unique sites 201 between the lines and further amplified to form progressively divergent methylomes. 202 Consistently, 65.7–85.4% of all hyperMRs were unique to one _drdd_ line, with de novo hyperMRs 203 being the most line-specific (85.0–85.4% unique) (Fig. 5B, S5). Additionally, most _drdd_ -specific 204 sRNA clusters underlying hyperMRs were unique to one of the two _drdd_ lines profiled (Fig. S7). 205 In contrast, WT lines retained stable average methylation at the _drdd_ hyperMRs through the 206 course of inbreeding (Fig. S6). 207 Because hyperMRs in the inbred _drdd_ were enriched in non-heterochromatic intergenic regions 208 and promoters (Fig. 1C), we hypothesized that they may be significantly affecting gene 209 expression. We performed RNA-seq on inbred _drdd_ and WT lines in the 9th generation, to 210 assess the effects of the highest final numbers of accumulated hyperMRs in _drdd_ on 211 transcription, as well as the impacts of multigenerational methylome divergence. We performed 212 RNA-seq using six biological replicates of each WT and _drdd_ line and called differentially 213 expressed genes (DEGs) separately for each of the two _drdd_ lines. First, we compared the 214 numbers of DEGs to the generation 2 _drdd_ [[14] and generation 4](https://www.zotero.org/google-docs/?xWiMyC) _drdd_ [[22]. The](https://www.zotero.org/google-docs/?yTN6AO) _drdd_ lines in 215 generation 9 showed a substantial increase in the numbers of DEGs, with a total of 4,057 and 216 2,539 DEGs in _drdd_ lines A and B, respectively, ~6-9.6 times higher than in generation 2 and 217 ~3.3-5.3 times higher than in generation 4 (Fig. 6A). Identifying DEGs with a reduced randomly 218 subsampled number of WT replicates did not impact this result (S3 Table). 219 220 The two independently profiled 9th generation _drdd_ lines shared 1,308 DEGs, suggesting they 221 do not share a closely related set of transcriptome changes (Fig. 6B). Principal component 222 analysis (PCA) showed that the two _drdd_ lines occupied largely distinct regions of transcriptional 223 space from one another, as well as from the WT lines, which clustered more closely together 224 (Fig. 6C). Consistent with this pattern, a heatmap of normalized expression across the union of 225 all _drdd_ DEGs showed broadly similar expression profiles among the WT lines but distinct 226 expression patterns in the two _drdd_ lines (Fig. 6D). These data suggest that the divergence in 227 methylation patterns between the two lines is associated with divergent transcriptome changes. 228 Consistently, direct comparison of the two _drdd_ lines identified 2,641 DEGs between them, 229 further supporting their substantial transcriptional divergence. To examine the biological 230 processes associated with these divergent transcriptional profiles, we performed Gene Ontology 231 (GO) enrichment analysis separately for the DEGs identified in each _drdd_ line. Both lines 232 showed enrichment for defense- and specialized metabolism-related processes among the top 6 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 233 enriched terms. Additionally, line A had line-specific enrichment for translation, ribosome 234 biogenesis, and cell-cycle-related processes (Fig. S8). 235 236 To examine if the methylation gains caused by multigenerational loss of DRDD activity may 237 underpin some of these transcriptome changes, we examined the proximity between hyperMRs 238 and DEGs called in each independent _drdd_ line. We observed that a small but nevertheless 239 significant subset of hyperMRs are in close genomic proximity (≤ 1 kb) to DEGs (Fig. 6E), 240 [similar to the association between DMRs and DEGs previously described [14]. Furthermore, this](https://www.zotero.org/google-docs/?1S8HbG) 241 analysis showed that out of all hyperMR classes, de novo hyperMRs are more frequently 242 associated with DEGs, and therefore may be the primary driver of transcriptome changes, with 243 17–22% of de novo hyperMRs being located within 1 kb of a DEG (Fig. 6E). Together, this 244 multigenerational multi-omic profiling of inbred lines leads to a model in which the loss of DRDD 245 leads to a shift in the accessible epigenetic landscape over time. In WT plants, DNA 246 demethylation is restricting the extent to which newly methylated regions can form and 247 consequently constraining the extent to which methylomes can diverge from one another within 248 related individuals. When the DRDD pathway is knocked out, independent lines are able to drift 249 further apart across epigenetic space, leading to divergent methylation patterns and 250 substantially different gene expression profiles. 251 **Discussion** 252 In this study, we demonstrate that the DNA demethylases in Arabidopsis stabilize the 253 methylome over multiple generations of inheritance by suppressing de novo methylation across 254 the genome. Our study also shows that in the absence of DNA demethylase activity, de novo 255 methylation events are stochastic and specific to individual lineages, causing independent 256 lineages to diverge in epigenetic space along distinct trajectories. While hypermethylated loci 257 are distinct between individual _drdd_ mutant lines, they appear to be drawn from a constrained 258 subset of the genome that is susceptible to methylation gain. The creation of distinct epigenomic 259 states in individual inbred lines is accompanied by substantial transcriptional divergence, 260 suggesting that the methylation gains prevented by DNA demethylation activity can broadly 261 impact gene regulation programs. Stable epigenetic inheritance therefore depends not only on 262 faithful propagation of existing methylation patterns, but also on the continual prevention of 263 ectopic methylation from becoming established and amplified by strong positively reinforcing 264 methylation pathways. 265 Our data strongly suggest that RdDM is responsible for the majority of methylation gains 266 occurring in inbred _drdd_ mutants. While we observe the hallmarks of RdDM maintaining 267 methylation at these sites, such as the presence of 24 nt sRNAs, these snapshots do not fully 268 explain how such de novo methylation events are initiated in an unmethylated parent genome. 269 While methylation at RdDM targets is reinforced through the canonical Pol IV-dependent 270 pathway, de novo methylation can be initiated through several noncanonical RdDM pathways in 271 which 21–22 nt sRNAs are generated from a variety of other sources, including aberrant 272 [transcripts [23,24]. Notably, most of the de novo sRNA-targeted hyperMRs we observed were in](https://www.zotero.org/google-docs/?7KVpct) 7 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 273 typically non-transcribed (by RNA polymerase II) intergenic regions. HyperMRs are therefore 274 either initiated from non-polymerase II sources, such as RNA polymerase IV, or are the product 275 of low-frequency polymerase II activity outside of typically transcribed regions. 276 Our data suggest that some genomic regions are more susceptible to de novo methylation than 277 others. We observe de novo methylation events more frequently in certain chromatin states 278 (Fig. S3), and a subset of loci became methylated in multiple independent lines (Fig. 5). A 279 central question therefore is – why are some loci more susceptible to de novo methylation gain 280 than others? The lineage specificity of many _drdd_ de novo hyperMRs and their frequent overlap 281 [with regions transcribed by Pol V in WT [21] support a “partial randomness” model in which de](https://www.zotero.org/google-docs/?5x49BP) 282 novo methylation is established stochastically within a constrained subset of susceptible loci. 283 This association is consistent with the proposed genome-surveillance model in which low-level 284 Pol V transcription makes genomic regions competent for RdDM if complementary siRNAs 285 [become available [21]. Some evidence suggests that siRNA-guided ARGONAUTE recruitment](https://www.zotero.org/google-docs/?BBDYjK) 286 precedes and directs Pol V recruitment, in which case Pol V activity may not reflect a genomic 287 surveillance function but rather a signature of low-frequency RdDM activity at certain genomic 288 [regions [25]. Once initial methylation became established at the new RdDM targets in](https://www.zotero.org/google-docs/?JKtJkp) _drdd_, 289 SUVH2 and SUVH9 could facilitate continued Pol V recruitment, creating a positive feedback 290 mechanism and contributing to RdDM reinforcement and the formation of newly methylated 291 [regions [26].](https://www.zotero.org/google-docs/?BVn9PS) 292 In addition to the gain of many newly methylated regions, inbreeding also causes substantial 293 transcriptomic divergence over time. Inbred _drdd_ mutant lines exhibit thousands of DEGs, 294 substantially more than the number identified in non-inbred _drdd_ mutants. In both mammals and 295 [plants, only a subset of genes are sensitive to proximal DNA methylation changes [27,28] and](https://www.zotero.org/google-docs/?Lw8Zbn) 296 genome-wide correlations between methylation and expression changes are typically weak. Our 297 data are consistent with this, with a minority of DEGs correlating with proximal hyperMRs that 298 occurred during inbreeding, typically within promoter regions. While our data cannot distinguish 299 between direct effects of proximal methylation changes and indirect effects from broader 300 transcriptional networks, the fact that transcriptomes were substantially different in independent 301 lines with divergent methylomes suggests that the transcriptomic changes are a consequence of 302 methylation changes in gene regulatory regions. Notably, at the organismal level, several _drdd_ 303 lineages became non-viable and could not be maintained through the full experiment, 304 suggesting that different methylation trajectories potentially produce different fitness 305 consequences. The lineages analyzed at generation 9 therefore represent a higher-fitness 306 subset of the possible outcomes of prolonged demethylation loss, potentially causing us to 307 underestimate its most severe epigenomic and transcriptional consequences. 308 Our study demonstrates that the activity of DNA demethylases is essential for stable epigenetic 309 inheritance over multiple generations. This mirrors the finding that elevated expression of a DNA 310 [demethylase also causes unstable inheritance of methylation over generations [29]. DRDD](https://www.zotero.org/google-docs/?RxeyLM) 311 activity must therefore be neither too high nor too low in order to enable optimal propagation of 312 epigenetic states and constrain epigenome divergence over generations. Mechanisms such as 8 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 313 the DNA methylation rheostat (methylstat) at the promoter of _ROS1_ [[30,31] are key to balancing](https://www.zotero.org/google-docs/?ToQBX7) 314 demethylation activity with the need for long-term stable epigenetic inheritance. 315 316 **Materials and Methods** 317 **Plant inbreeding and growth conditions** 318 The seeds from second generation individual _drdd_ and WT segregant progenitor plants were 319 [obtained from Williams et al. [14]. Two of the four](https://www.zotero.org/google-docs/?FVyOFc) _drdd_ lines already maintained to the third 320 generation showed detrimental phenotypes, with plants that became sterile within 1–2 additional 321 generations, and these lines were excluded from the study. To maintain three _drdd_ lines, we 322 derived two sub-lines from siblings of one parent line at generation 3 (Fig. 1A). The resulting 6 323 lines ( _DRDD_ lines A, B, and C; and _drdd_ lines A, B, and C) were subsequently inbred through 324 single-seed descent through 9 generations total, with the exception of _drdd_ line C which became 325 sterile and had to be terminated after generation 7. At each generation, plants that were 326 selected for genomic profiling were genotyped to confirm the correct genotype for each sample 327 (primers provided in S1 Table). Seeds were stratified for ~2 days at 4 °C and grown under long328 day conditions (LD, 16 h light / 8 h dark, ~21 °C, ~60% humidity) in a Percival growth chamber 329 (AR100L3). Most sample collection happened at 21 days after germination (DAG), described in 330 more detail in subsequent sections. 331 **Enzymatic methyl sequencing** 332 Whole-genome single-cytosine-resolution DNA methylation was profiled with enzymatic methyl 333 [sequencing (EM-seq) [32]. For each line at each generation assessed, a single 5th true leaf was](https://www.zotero.org/google-docs/?v9AVFo) 334 sequenced. Whole-leaf tissues were collected, flash-frozen in liquid nitrogen, and stored at -80 335 °C until use. In the 5th generation, 5th true leaves were collected when the plants were visually 336 at a similar developmental stage between 21–28 DAG. Frozen leaf tissue was processed using 337 a Qiagen TissueLyser II and incubated in CTAB buffer at 65 °C for about 30 minutes, followed 338 by the addition of 3 μl of RNase and incubation at 37 °C for 20 minutes. DNA was then purified 339 with chloroform, precipitated in 100% ethanol at -20 °C, washed with 70% ethanol, and eluted 340 with nuclease-free water. DNA was sheared to an average fragment size of ~300 bp with a 341 Covaris S220 Focused Ultrasonicator, using 99-second sonication. EM-seq libraries were 342 constructed according to the NEBNext® Enzymatic Methylseq Kit protocol with either 25 or 50 343 ng starting DNA, depending on the sample DNA concentration. Library quality was assessed 344 using a KAPA Library Quantification Kit, and average fragment size was verified using a 345 fragment analyzer. Libraries were sequenced with Illumina NovaSeq X, with 150 bp paired-end 346 reads. 347 [EM-seq data were mapped as described in Williams et al. [14]. Briefly, adapters were removed](https://www.zotero.org/google-docs/?WUGaK6) 348 using Trim Galore 0.6.6 (Babraham Bioinformatics). Filtered reads were mapped to the 349 _Arabidopsis thaliana_ TAIR10 genome using Bismark 0.22.3 (Babraham Bioinformatics), allowing 350 for 0 mismatched reads in the seed region and removing PCR duplicates. Percent methylation 351 for each cytosine was calculated using the Bismark methylation extractor function. The EM-seq 352 conversion efficiency was calculated based on the percentage of methylation of reads mapped 9 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 353 to the chloroplast, which is unmethylated. Conversion rates were > 98.7% across all samples 354 sequenced. Genome browser snapshots for EM-seq data and all other genomic data were 355 [taken in Integrative Genomics Viewer (IGV) v. 2.16.0 [33]. Sequencing and mapping statistics](https://www.zotero.org/google-docs/?e7UXSo) 356 for all EM-seq libraries are provided in S2 Table. 357 **DMR analysis** 358 [Differentially methylated regions (DMRs) were called using DMRcaller [34]. DMRs were called](https://www.zotero.org/google-docs/?18CbWX) 359 between pairs of samples: a generation 2 ancestor and an inbred line sample. Lists of 360 hyperMRs and hypoMRs are provided in S1–S4 Data. The generation 2 ancestors used are a 361 _drdd_ plant and a _DRDD_ plant, which are siblings of the generation 2 line progenitor plants 362 [profiled in Williams et al. [14]. Before using DMRcaller, read counts were first corrected based](https://www.zotero.org/google-docs/?9n2US6) 363 [on conversion rate as previously described [35]. DMRcaller was run using the bins method, with](https://www.zotero.org/google-docs/?T50QEj) 364 the following settings: bin size of 300 bp; p-value threshold of 0.01; minimum cytosine count of 365 10; min proportion difference of 0.35 for CG, 0.25 for CHG, and 0.1 for CHH; min gap between 366 DMRs of 500 bp, and min reads per cytosine of 5. The resulting DMRs were filtered for 367 hyperMRs, as these are most likely to be a direct impact of the mutants losing demethylase 368 activity. To analyze chromatin states, the resulting hyperMRs were overlapped with the nine 369 [chromatin states previously defined by Sequeira-Mendes et al. [16]. The nine states were](https://www.zotero.org/google-docs/?J0FtgP) 370 reduced into five chromatin functional groups that were most enriched in the original 9 states: 371 transcription start site (TSS) - state 1; proximal promoter - state 2; gene body - states 3, 6, and 372 7; heterochromatin - states 8 and 9; other intergenic - states 4 and 5. To generate control 373 chromatin state distributions, the _drdd_ and WT generation 2 ancestor genomes were filtered for 374 300 bp bins that had average methylation below 65% in the CG context, 75% in the CHG 375 context, and 90% in the CHH context to retain only regions which were eligible for being 376 classified as a hyperMR based on DMRcaller cutoffs. These control distributions are shown as 377 “filtered genome” in chromatin state distribution plots, with the _drdd_ control distribution used in 378 Fig. 1C and S3 and the WT control distribution in Fig. S2. 379 To classify hyperMRs based on their location relative to previously methylated regions in 380 generation 2, a custom R script was written. Briefly, the mapped EM-seq sequencing data from 381 generation 2 were broken into 300 bp bins, and average methylation was calculated over each 382 bin. Resulting bins were filtered to retain those with at least 10 cytosines and an average of 5 383 reads per cytosine. Regions were classified as methylated if the average methylation was at 384 least 50% for CG, 20% for CHG, or 5% for CHH. Methylated bins that were up to two bins (600 385 bp) apart were merged, unless the merged bin no longer met the cytosine and read 386 requirements, in which case they were kept separate. The output of these steps was a set of 387 “generation 2 methylated” regions. Next, regions that retained methylation were identified in 388 each inbred line at each generation, and regions that either lost methylation, or did not pass the 389 same cytosine count, average reads, and average methylation threshold filters were filtered out, 390 resulting in a final set of “retained generation 2 methylated” regions. Finally, hyperMRs found for 391 each inbred line were classified based on their location relative to retained generation 2 392 methylated regions: “contained” for hyperMRs fully contained in the original region, “spreading” 393 if the hyperMR overlapped the two closest bins (600 bp upstream or downstream), “de novo” for 10 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 394 hyperMRs beyond 1200 bp, and “ambiguous” for any hyperMRs not matching any of the 395 previous conditions, such as falling between the two distance cutoffs. The distances were 396 chosen based on a previous finding that methylation tends to spread up to ~500 bp away from 397 [heavily-methylated siRNA targets [8] and 600 bp being the closest distance that is a multiple of](https://www.zotero.org/google-docs/?3I4ypJ) 398 the 300 bp bin length for both DMRcaller and our script. 399 The heatmap of average methylation across a superset of all _drdd_ hyperMRs was generated in 400 R using the package ComplexHeatmap [[36], using hierarchical clustering to sort hyperMRs](https://www.zotero.org/google-docs/?D8Z2iy) 401 within each class. Within each hyperMR class and methylation context, hyperMRs identified 402 across all _drdd_ lineages and generations were combined, and overlapping intervals were 403 merged to generate a nonredundant superset of regions which were used for the heatmap. 404 Intersect analysis for methylation and all other data was performed using bedtools intersect, and 405 [proximity analysis was performed using bedtools closest [37]. Out of figures depicting hyperMRs](https://www.zotero.org/google-docs/?c0Okoh) 406 in a single generation or relying on single-generation hyperMRs for the underlying analysis, 407 figures 1C, 2B, 2C, 3A, S2, and S3 are based on hyperMRs in the 7th generation, so as to 408 compare three independent inbred lines. Figures 2D, 3B, 3C, 3D, 4A, 4B, 6E, S4, and S7 are 409 based on hyperMRs in the 9th generation, so as to analyze methylation gains in the final inbred 410 generation or to correlate methylation data with same-generation RNA-seq or sRNA-seq data. 411 **RNA-seq** 412 The 5th true leaf of six independent replicate plants was collected at 21 DAG for each line 413 analyzed at generation 9. Whole-leaf tissues were collected, flash-frozen in liquid nitrogen, and 414 stored at -80 °C until use. Total RNA was isolated using a QIAGEN RNeasy Plant Mini Kit, 415 following the manufacturer’s protocol. Briefly, frozen leaf tissue was ground to a powder using 416 glass beads using Qiagen TissueLyser II and resuspended in Buffer RLC with β417 mercaptoethanol. Plant debris was removed through centrifugation in a lilac spin column, 418 followed by RNA precipitation with 100% ethanol, washes with buffer RW1 and buffer RPE, and 419 elution in RNase-free water performed through centrifugation in a pink spin column. 420 Library preparation and sequencing were performed by Novogene, with sequencing done on a 421 NovaSeq 6000 using 150-bp paired-end reads. All samples were sequenced in one batch. The 422 [reads were mapped as previously described [38]. Briefly, reads were first processed with Trim](https://www.zotero.org/google-docs/?XVdSgs) 423 [Galore v0.6.6 (Babraham Bioinformatics) and cutadapt [39]. During this step, adapter](https://www.zotero.org/google-docs/?KMKC1M) 424 sequences were trimmed, the first 9 bp were clipped from the 5’ end of each read, low-quality 425 bases were trimmed from the 3’ end using a Phred quality threshold of 25, and reads shorter 426 than 20 bp after trimming were discarded. Reads were mapped to the Araport11 genome using 427 STAR 2.7.1a [[40], and alignments were filtered to retain uniquely mapped reads, allowing](https://www.zotero.org/google-docs/?ck4Ico) 428 mismatches up to 5% of the total read length. Sequencing and mapping statistics for all RNA429 seq libraries are provided in S2 Table. Gene-level read counts were generated with htseq-count 430 [from HTSeq v2.0 [41] and are provided in S5 Data. Normalization and differential expression](https://www.zotero.org/google-docs/?vSOoq5) 431 [analysis were performed using DESeq2 v1.40.1 [42]. Principal component analysis (PCA) was](https://www.zotero.org/google-docs/?jIOByo) 432 performed on regularized-logarithm (rlog)-transformed gene counts using DESeq2 (S6 Data), 433 and 95% multivariate _t_ [-distribution ellipses were plotted for each line using ggplot2 [43].](https://www.zotero.org/google-docs/?XExW2P) 434 Differential expression was called separately for each of the two _drdd_ lines, against a set of all 11 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 435 18 WT replicates (S7 and S8 Data). We also repeated the DEG analysis with 6 WT replicates 436 selected by random subsampling and achieved equivalent numbers of DEGs (S3 Table; S9 and 437 S10 Data). DEGs were defined as genes with log2 fold-change >1 and a Benjamini–Hochberg438 adjusted P value < 0.05. 439 GO enrichment analysis was performed separately for DEGs called in each _drdd_ line using the 440 [PANTHER 19.0 Statistical Overrepresentation Test [44]. For each analysis, the reference list](https://www.zotero.org/google-docs/?cfv0HL) 441 consisted of all genes with non-NA adjusted P-values in the corresponding DESeq2 comparison 442 (24,003 genes for each line). Enrichment was tested using the GO Biological Process Complete 443 annotation dataset, Fisher’s exact test, and false discovery rate (FDR) correction for multiple 444 testing. Overrepresented terms with FDR < 0.05 were considered. To reduce redundancy 445 among GO terms, one representative term was selected from each hierarchical group displayed 446 in PANTHER, choosing the most specific child term with at least five associated DEGs. 447 Subsequently, terms that were highly redundant were excluded. The resulting GO terms were 448 ranked by fold enrichment, and the top 20 terms were plotted for each line. 449 **sRNA-seq** 450 For each of the five lines at generation 9, 5th true leaves were collected at 21 DAG from six 451 individual plants and pooled together. The leaf tissue was flash-frozen in liquid nitrogen and 452 stored at -80 °C until use. Total RNA (without depletion of RNAs) was isolated using the 453 mirVana miRNA Isolation Kit. Briefly, frozen tissue was ground to a powder using glass beads 454 using Qiagen TissueLyser II and resuspended in 300 μl of Lysis/Binding Buffer, followed by the 455 addition of 30 μl miRNA Homogenate Additive. RNA was purified with acid-phenol:chloroform 456 and washed in a filter cartridge with 100% ethanol, followed by the mirVana Wash Solution 1 457 and Wash Solution 2/3. RNA was then eluted with the mirVana Elution Solution, pre-heated to 458 95 [o] C. 1 μl of Qiagen spike-ins were added to each sample. Libraries were prepared by 459 Novogene and sequenced with NovaSeq 6000. For mapping, sRNA reads were first aligned to 460 [Qiagen spike-ins using Bowtie2 [45], and reads that did not align to the spike-ins were retained](https://www.zotero.org/google-docs/?IrJn0j) 461 for genomic mapping. The retained reads were then aligned to the TAIR10 genome using 462 [ShortStack v4.1.2 [46]. Known Arabidopsis mature miRNA sequences from miRBase were](https://www.zotero.org/google-docs/?vQNbEk) 463 provided to ShortStack to guide miRNA locus identification. Sequencing and mapping statistics 464 for sRNA-seq libraries are provided in S2 Table. 465 **Pol V transcript analysis** 466 For Pol V transcription analysis in Figure 4, previously published genomic annotations reflecting 467 [Pol V and RdDM activity were used [21]. The whole genome was previously classified into four](https://www.zotero.org/google-docs/?2gaq7a) 468 states using a Hidden Markov Model (HMM), based on the likelihood of Pol V transcription and 469 [RdDM activity in each bin [21]. Here, two of the HMM states were combined, and hyperMRs of](https://www.zotero.org/google-docs/?EPGqLJ) 470 different classes were overlapped with the resulting classified clusters to obtain Figure 4A. The 471 random control regions were randomly sampled from a set of 300 bp bins which were lowly 472 methylated in generation 2 _drdd_ (average mCG < 50%, mCHG < 20%, and mCHH < 5%), with 473 five independent random samplings per context. These controls were selected because they 474 had potential to be classified as a de novo hyperMR by our DMR classifying script. The 12 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 475 numbers of control regions were 500, 200 and 200 for CG, CHG and CHH, respectively, 476 approximating average numbers of de novo hyperMRs called in each sequence context. 477 For read alignments in Figure 4B, IPARE-seq reads from Col-0 samples were obtained from 478 [GEO accession GSE146913 [21]. For each read pair, the cDNA-derived sequence was](https://www.zotero.org/google-docs/?5mu2yP) 479 contained in the R1 read, while the unique molecular identifier (UMI) was extracted from the 480 [corresponding R2 read using UMI-tools [47] and appended to the matching R1 read name.](https://www.zotero.org/google-docs/?LUllZA) 481 [Poly(A) sequences were trimmed using cutadapt [39], and reads shorter than 20 bp after](https://www.zotero.org/google-docs/?FFum7S) 482 trimming were discarded. Reads were then aligned as single-end reads to the TAIR10 genome 483 [using Bowtie2 [45], allowing one mismatch per seed alignment. Reads with mapping quality ≥ 2](https://www.zotero.org/google-docs/?JixzoG) 484 [were retained using SAMtools [48], and PCR duplicates were removed using UMI-tools](https://www.zotero.org/google-docs/?6weIsT) 485 deduplication. For genome browser visualization, deduplicated BAM files from all six Col-0 486 [replicates were merged and indexed using SAMtools [48].](https://www.zotero.org/google-docs/?kytwKZ) 487 488 **Acknowledgements** 489 We are most grateful to Andrzej Wierzbicki and Shriya Sethuraman for openly sharing their data 490 and analysis on RNA polymerase V transcripts. We are grateful to Dawei Dai for assistance with 491 EM-seq library preparation and to Ken Chen, Dawei Dai, and Nicole K. Smoot for assistance 492 with bioinformatics analysis. 493 **AI use disclosure** 494 AI was used in the final proofreading of the manuscript, as well as in the construction of analysis 495 pipelines. The accuracy of all AI output was rigorously checked and verified by the authors. 496 **Competing interests** 497 The authors have no conflicts of interest to declare. 498 **References** [1. Williams BP, Gehring M. Principles of Epigenetic Homeostasis Shared Between Flowering](https://www.zotero.org/google-docs/?l2mhyA) [Plants and Mammals. Trends Genet. 2020;36: 751–763. doi:10.1016/j.tig.2020.06.019](https://www.zotero.org/google-docs/?l2mhyA) [2. Hagmann J, Becker C, Müller J, Stegle O, Meyer RC, Wang G, et al. Century-scale](https://www.zotero.org/google-docs/?l2mhyA) [Methylome Stability in a Recently Diverged Arabidopsis thaliana Lineage. PLOS Genet.](https://www.zotero.org/google-docs/?l2mhyA) [2015;11: e1004920. doi:10.1371/journal.pgen.1004920](https://www.zotero.org/google-docs/?l2mhyA) [3. Hofmeister BT, Lee K, Rohr NA, Hall DW, Schmitz RJ. Stable inheritance of DNA methylation](https://www.zotero.org/google-docs/?l2mhyA) [allows creation of epigenotype maps and the study of epiallele inheritance patterns in the](https://www.zotero.org/google-docs/?l2mhyA) [absence of genetic variation. Genome Biol. 2017;18: 155. doi:10.1186/s13059-017-1288-x](https://www.zotero.org/google-docs/?l2mhyA) [4. Law JA, Jacobsen SE. Establishing, maintaining and modifying DNA methylation patterns in](https://www.zotero.org/google-docs/?l2mhyA) [plants and animals. Nat Rev Genet. 2010;11: 204–220. doi:10.1038/nrg2719](https://www.zotero.org/google-docs/?l2mhyA) 13 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [5. Matzke MA, Mosher RA. RNA-directed DNA methylation: an epigenetic pathway of increasing](https://www.zotero.org/google-docs/?l2mhyA) [complexity. Nat Rev Genet. 2014;15: 394–408. doi:10.1038/nrg3683](https://www.zotero.org/google-docs/?l2mhyA) [6. Stroud H, Do T, Du J, Zhong X, Feng S, Johnson L, et al. Non-CG methylation patterns shape](https://www.zotero.org/google-docs/?l2mhyA) [the epigenetic landscape in Arabidopsis. Nat Struct Mol Biol. 2014;21: 64–72.](https://www.zotero.org/google-docs/?l2mhyA) doi:10.1038/nsmb.2735 [7. Xu G, Law JA. Loops, crosstalk, and compartmentalization: it takes many layers to regulate](https://www.zotero.org/google-docs/?l2mhyA) [DNA methylation. Curr Opin Genet Dev. 2024;84: 102147. doi:10.1016/j.gde.2023.102147](https://www.zotero.org/google-docs/?l2mhyA) [8. Ahmed I, Sarazin A, Bowler C, Colot V, Quesneville H. Genome-wide evidence for local DNA](https://www.zotero.org/google-docs/?l2mhyA) [methylation spreading from small RNA-targeted sequences in Arabidopsis. Nucleic Acids Res.](https://www.zotero.org/google-docs/?l2mhyA) [2011;39: 6919–6931. doi:10.1093/nar/gkr324](https://www.zotero.org/google-docs/?l2mhyA) [9. Choi Y, Gehring M, Johnson L, Hannon M, Harada JJ, Goldberg RB, et al. DEMETER, a DNA](https://www.zotero.org/google-docs/?l2mhyA) [Glycosylase Domain Protein, Is Required for Endosperm Gene Imprinting and Seed Viability in](https://www.zotero.org/google-docs/?l2mhyA) [Arabidopsis. Cell. 2002;110: 33–42. doi:10.1016/S0092-8674(02)00807-3](https://www.zotero.org/google-docs/?l2mhyA) [10. Gong Z, Morales-Ruiz T, Ariza RR, Roldán-Arjona T, David L, Zhu J-K. ROS1, a Repressor of](https://www.zotero.org/google-docs/?l2mhyA) [Transcriptional Gene Silencing in Arabidopsis, Encodes a DNA Glycosylase/Lyase. Cell.](https://www.zotero.org/google-docs/?l2mhyA) [2002;111: 803–814. doi:10.1016/S0092-8674(02)01133-9](https://www.zotero.org/google-docs/?l2mhyA) [11. Penterman J, Zilberman D, Huh JH, Ballinger T, Henikoff S, Fischer RL. DNA demethylation in](https://www.zotero.org/google-docs/?l2mhyA) the _Arabidopsis_ [genome. Proc Natl Acad Sci. 2007;104: 6752–6757.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1073/pnas.0701861104](https://www.zotero.org/google-docs/?l2mhyA) [12. Lister R, O’Malley RC, Tonti-Filippini J, Gregory BD, Berry CC, Millar AH, et al. Highly](https://www.zotero.org/google-docs/?l2mhyA) [Integrated Single-Base Resolution Maps of the Epigenome in Arabidopsis. Cell. 2008;133:](https://www.zotero.org/google-docs/?l2mhyA) [523–536. doi:10.1016/j.cell.2008.03.029](https://www.zotero.org/google-docs/?l2mhyA) [13. Tang K, Lang Z, Zhang H, Zhu J-K. The DNA demethylase ROS1 targets genomic regions](https://www.zotero.org/google-docs/?l2mhyA) [with distinct chromatin modifications. Nat Plants. 2016;2: 16169.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1038/nplants.2016.169](https://www.zotero.org/google-docs/?l2mhyA) [14. Williams BP, Bechen LL, Pohlmann DA, Gehring M. Somatic DNA demethylation generates](https://www.zotero.org/google-docs/?l2mhyA) [tissue-specific methylation states and impacts flowering time. Plant Cell. 2022;34: 1189–1206.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1093/plcell/koab319](https://www.zotero.org/google-docs/?l2mhyA) [15. Tang K, Zhu X, Xie S, Lang Z, Zhu J-K. Transgenerational increases in DNA methylation in](https://www.zotero.org/google-docs/?l2mhyA) _Arabidopsis_ [plants defective in active DNA demethylation. Proc Natl Acad Sci. 2024;121:](https://www.zotero.org/google-docs/?l2mhyA) [e2320468121. doi:10.1073/pnas.2320468121](https://www.zotero.org/google-docs/?l2mhyA) [16. Sequeira-Mendes J, Aragüez I, Peiró R, Mendez-Giraldez R, Zhang X, Jacobsen SE, et al.](https://www.zotero.org/google-docs/?l2mhyA) The Functional Topography of the _Arabidopsis_ [Genome Is Organized in a Reduced Number of](https://www.zotero.org/google-docs/?l2mhyA) [Linear Motifs of Chromatin States. Plant Cell. 2014;26: 2351–2366.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1105/tpc.114.124578](https://www.zotero.org/google-docs/?l2mhyA) [17. Hazarika RR, Serra M, Zhang Z, Zhang Y, Schmitz RJ, Johannes F. Molecular properties of](https://www.zotero.org/google-docs/?l2mhyA) [epimutation hotspots. Nat Plants. 2022;8: 146–156. doi:10.1038/s41477-021-01086-7](https://www.zotero.org/google-docs/?l2mhyA) 14 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [18. Yamamuro C, Miki D, Zheng Z, Ma J, Wang J, Yang Z, et al. Overproduction of stomatal](https://www.zotero.org/google-docs/?l2mhyA) [lineage cells in Arabidopsis mutants defective in active DNA demethylation. Nat Commun.](https://www.zotero.org/google-docs/?l2mhyA) 2014;5: 4062. doi:10.1038/ncomms5062 [19. Erdmann RM, Picard CL. RNA-directed DNA Methylation. PLOS Genet. 2020;16: e1009034.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1371/journal.pgen.1009034](https://www.zotero.org/google-docs/?l2mhyA) [20. Loffer A, Singh J, Fukudome A, Mishra V, Wang F, Pikaard CS. A DCL3 dicing code within Pol](https://www.zotero.org/google-docs/?l2mhyA) [IV-RDR2 transcripts diversifies the siRNA pool guiding RNA-directed DNA methylation. eLife.](https://www.zotero.org/google-docs/?l2mhyA) 2022;11: e73260. doi:10.7554/eLife.73260 [21. Tsuzuki M, Sethuraman S, Coke AN, Rothi MH, Boyle AP, Wierzbicki AT. Broad noncoding](https://www.zotero.org/google-docs/?l2mhyA) [transcription suggests genome surveillance by RNA polymerase V. Proc Natl Acad Sci.](https://www.zotero.org/google-docs/?l2mhyA) [2020;117: 30799–30804. doi:10.1073/pnas.2014419117](https://www.zotero.org/google-docs/?l2mhyA) [22. Smoot NK, Zeng Y, Hochman RM, Williams BP. Complete vegetative propagation via an](https://www.zotero.org/google-docs/?l2mhyA) [epigenetic state change. bioRxiv; 2026. doi:10.64898/2026.03.27.714943](https://www.zotero.org/google-docs/?l2mhyA) [23. Cuerda-Gil D, Slotkin RK. Non-canonical RNA-directed DNA methylation. Nat Plants. 2016;2:](https://www.zotero.org/google-docs/?l2mhyA) [16163. doi:10.1038/nplants.2016.163](https://www.zotero.org/google-docs/?l2mhyA) [24. Kramer MC, Ratnayake TS, Edwards SA, Lowrey HL, Klaas G, Sidorenko L, et al.](https://www.zotero.org/google-docs/?l2mhyA) [Identification of a cleaved aberrant RNA associated with the initiation of transgene silencing.](https://www.zotero.org/google-docs/?l2mhyA) [Plant Cell. 2025;37: koaf219. doi:10.1093/plcell/koaf219](https://www.zotero.org/google-docs/?l2mhyA) [25. Sigman MJ, Panda K, Kirchner R, McLain LL, Payne H, Peasari JR, et al. An siRNA-guided](https://www.zotero.org/google-docs/?l2mhyA) [ARGONAUTE protein directs RNA polymerase V to initiate DNA methylation. Nat Plants.](https://www.zotero.org/google-docs/?l2mhyA) 2021;7: 1461–1474. doi:10.1038/s41477-021-01008-7 [26. Liu Z-W, Shao C-R, Zhang C-J, Zhou J-X, Zhang S-W, Li L, et al. The SET Domain Proteins](https://www.zotero.org/google-docs/?l2mhyA) [SUVH2 and SUVH9 Are Required for Pol V Occupancy at RNA-Directed DNA Methylation](https://www.zotero.org/google-docs/?l2mhyA) [Loci. PLOS Genet. 2014;10: e1003948. doi:10.1371/journal.pgen.1003948](https://www.zotero.org/google-docs/?l2mhyA) [27. Bock C, Beerman I, Lien W-H, Smith ZD, Gu H, Boyle P, et al. DNA methylation dynamics](https://www.zotero.org/google-docs/?l2mhyA) [during in vivo differentiation of blood and skin stem cells. Mol Cell. 2012;47: 633–647.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1016/j.molcel.2012.06.019](https://www.zotero.org/google-docs/?l2mhyA) [28. Meng D, Dubin M, Zhang P, Osborne EJ, Stegle O, Clark RM, et al. Limited Contribution of](https://www.zotero.org/google-docs/?l2mhyA) [DNA Methylation Variation to Expression Regulation in Arabidopsis thaliana. Gaut BS, editor.](https://www.zotero.org/google-docs/?l2mhyA) [PLOS Genet. 2016;12: e1006141. doi:10.1371/journal.pgen.1006141](https://www.zotero.org/google-docs/?l2mhyA) [29. Williams BP, Gehring M. Stable transgenerational epigenetic inheritance requires a DNA](https://www.zotero.org/google-docs/?l2mhyA) [methylation-sensing circuit. Nat Commun. 2017;8: 2124. doi:10.1038/s41467-017-02219-3](https://www.zotero.org/google-docs/?l2mhyA) [30. Lei M, Zhang H, Julian R, Tang K, Xie S, Zhu J-K. Regulatory link between DNA methylation](https://www.zotero.org/google-docs/?l2mhyA) and active demethylation in _Arabidopsis_ [. Proc Natl Acad Sci. 2015;112: 3553–3557.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1073/pnas.1502279112](https://www.zotero.org/google-docs/?l2mhyA) [31. Williams BP, Pignatta D, Henikoff S, Gehring M. Methylation-Sensitive Expression of a DNA](https://www.zotero.org/google-docs/?l2mhyA) [Demethylase Gene Serves As an Epigenetic Rheostat. Mittelsten Scheid O, editor. PLOS](https://www.zotero.org/google-docs/?l2mhyA) [Genet. 2015;11: e1005142. doi:10.1371/journal.pgen.1005142](https://www.zotero.org/google-docs/?l2mhyA) 15 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [32. Feng S, Zhong Z, Wang M, Jacobsen SE. Efficient and accurate determination of genome-](https://www.zotero.org/google-docs/?l2mhyA) [wide DNA methylation patterns in Arabidopsis thaliana with enzymatic methyl sequencing.](https://www.zotero.org/google-docs/?l2mhyA) [Epigenetics Chromatin. 2020;13: 42. doi:10.1186/s13072-020-00361-9](https://www.zotero.org/google-docs/?l2mhyA) [33. Thorvaldsdottir H, Robinson JT, Mesirov JP. Integrative Genomics Viewer (IGV): high-](https://www.zotero.org/google-docs/?l2mhyA) [performance genomics data visualization and exploration. Brief Bioinform. 2013;14: 178–192.](https://www.zotero.org/google-docs/?l2mhyA) doi:10.1093/bib/bbs017 [34. Catoni M, Tsang JM, Greco AP, Zabet NR. DMRcaller: a versatile R/Bioconductor package for](https://www.zotero.org/google-docs/?l2mhyA) [detection and visualization of differentially methylated regions in CpG and non-CpG contexts.](https://www.zotero.org/google-docs/?l2mhyA) [Nucleic Acids Res. 2018;46: e114. doi:10.1093/nar/gky602](https://www.zotero.org/google-docs/?l2mhyA) [35. Catoni M, Zabet NR. Analysis of Plant DNA Methylation Profiles Using R. In: Cho J, editor.](https://www.zotero.org/google-docs/?l2mhyA) [Plant Transposable Elements: Methods and Protocols. New York, NY: Springer US; 2021. pp.](https://www.zotero.org/google-docs/?l2mhyA) [219–238. doi:10.1007/978-1-0716-1134-0_21](https://www.zotero.org/google-docs/?l2mhyA) [36. Gu Z, Eils R, Schlesner M. Complex heatmaps reveal patterns and correlations in](https://www.zotero.org/google-docs/?l2mhyA) [multidimensional genomic data. Bioinformatics. 2016;32: 2847–2849.](https://www.zotero.org/google-docs/?l2mhyA) doi:10.1093/bioinformatics/btw313 [37. Quinlan AR, Hall IM. BEDTools: a flexible suite of utilities for comparing genomic features.](https://www.zotero.org/google-docs/?l2mhyA) [Bioinformatics. 2010;26: 841–842. doi:10.1093/bioinformatics/btq033](https://www.zotero.org/google-docs/?l2mhyA) [38. Williams CJ, Dai D, Tran KA, Monroe JG, Williams BP. Dynamic DNA methylation turnover in](https://www.zotero.org/google-docs/?l2mhyA) [gene bodies is associated with enhanced gene expression plasticity in plants. Genome Biol.](https://www.zotero.org/google-docs/?l2mhyA) 2023;24: 227. doi:10.1186/s13059-023-03059-9 [39. Martin M. Cutadapt removes adapter sequences from high-throughput sequencing reads.](https://www.zotero.org/google-docs/?l2mhyA) [EMBnet.journal. 2011;17: 10–12. doi:10.14806/ej.17.1.200](https://www.zotero.org/google-docs/?l2mhyA) [40. Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, et al. STAR: ultrafast](https://www.zotero.org/google-docs/?l2mhyA) [universal RNA-seq aligner. Bioinformatics. 2013;29: 15–21. doi:10.1093/bioinformatics/bts635](https://www.zotero.org/google-docs/?l2mhyA) [41. Putri GH, Anders S, Pyl PT, Pimanda JE, Zanini F. Analysing high-throughput sequencing](https://www.zotero.org/google-docs/?l2mhyA) [data in Python with HTSeq 2.0. Bioinformatics. 2022;38: 2943–2945.](https://www.zotero.org/google-docs/?l2mhyA) doi:10.1093/bioinformatics/btac166 [42. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-](https://www.zotero.org/google-docs/?l2mhyA) [seq data with DESeq2. Genome Biol. 2014;15: 550. doi:10.1186/s13059-014-0550-8](https://www.zotero.org/google-docs/?l2mhyA) [43. Wickham H. ggplot2: Elegant Graphics for Data Analysis. 2nd ed. Springer International](https://www.zotero.org/google-docs/?l2mhyA) [Publishing; 2016. doi:10.1007/978-3-319-24277-4](https://www.zotero.org/google-docs/?l2mhyA) [44. Thomas PD, Ebert D, Muruganujan A, Mushayahama T, Albou LP, Mi H. PANTHER: Making](https://www.zotero.org/google-docs/?l2mhyA) [genome-scale phylogenetics accessible to all. Protein Sci. 2022;31: 8–22.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1002/pro.4218](https://www.zotero.org/google-docs/?l2mhyA) [45. Langmead B, Salzberg SL. Fast gapped-read alignment with Bowtie 2. Nat Methods. 2012;9:](https://www.zotero.org/google-docs/?l2mhyA) 357–359. doi:10.1038/nmeth.1923 16 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [46. Axtell MJ. ShortStack: Comprehensive annotation and quantification of small RNA genes.](https://www.zotero.org/google-docs/?l2mhyA) RNA. 2013;19: 740–751. doi:10.1261/rna.035279.112 [47. Smith T, Heger A, Sudbery I. UMI-tools: modeling sequencing errors in Unique Molecular](https://www.zotero.org/google-docs/?l2mhyA) [Identifiers to improve quantification accuracy. Genome Res. 2017;27: 491–499.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1101/gr.209601.116](https://www.zotero.org/google-docs/?l2mhyA) [48. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, et al. The Sequence](https://www.zotero.org/google-docs/?l2mhyA) [Alignment/Map format and SAMtools. Bioinformatics. 2009;25: 2078–2079.](https://www.zotero.org/google-docs/?l2mhyA) [doi:10.1093/bioinformatics/btp352](https://www.zotero.org/google-docs/?l2mhyA) **Supporting information** S1 Table. Genotyping primers for _drdd_ . S2 Table. Sequencing and mapping statistics for all libraries. S3 Table. Differentially expressed gene numbers for all comparisons. S1 Data. _drdd_ hyperMRs. S2 Data. WT hyperMRs. S3 Data. _drdd_ hypoMRs. S4 Data. WT hypoMRs. S5 Data. Raw RNA-seq gene counts. S6 Data. Rlog-transformed RNA-seq gene counts. S7 Data. DESeq2 differential expression results for _drdd_ A vs. 18 WT samples. S8 Data. DESeq2 differential expression results for _drdd_ B vs. 18 WT samples. S9 Data. DESeq2 differential expression results for _drdd_ A vs. 6 WT samples. S10 Data. DESeq2 differential expression results for _drdd_ B vs. 6 WT samples. 17 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 18 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) **Figure 2. De novo and spreading methylation gains happen progressively over the** **course of inbreeding. A.** The distance thresholds used for classifying hyperMRs. Distances are measured in base pairs from methylated regions which have conserved methylation between the generation 2 ancestor and an inbred line. **B.** Representative “contained”, “spreading”, and “de novo” methylation gains in _drdd_ lines in generation 7, in all sequence contexts. For the “contained” class, mCHG is elevated in generation 7 relative to generation 2 _drdd_ . Generation 2 WT is shown as a control. Dashed pink outlines show called hyperMRs. The vertical scale for each track ranges from 0 to 100% methylation. Thin dark gray bars beneath each track indicate sites with EM-seq coverage but no detected methylation (0%). **C.** Numbers of hyperMRs in generation 7, showing their distribution across the four classes. **D.** Ends analysis plots showing average CG methylation levels over de novo and spreading hyperMRs called in generation 9 in _drdd_ line A, relative to generation 2 _drdd._ Generation 2 WT is shown as a control. 150 bp are shown on either side inside the dashed lines, and methylation is averaged over 25 bp bins. Left: de novo hyperMRs are between the vertical dashed lines. Right: spreading hyperMRs are outside of the vertical dashed lines, and the ancestral methylated regions adjacent to them are between the dashed lines. [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) hyperMRs by overlap with sRNA clusters in the corresponding _drdd_ line and in any WT line. **D.** Distribution of sRNA read lengths for sRNAs underlying de novo hyperMRs in 9th generation [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 22 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 26 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) **S4.** Ends analysis plots showing average methylation levels over de novo and spreading hyperMRs called in generation 9, relative to generation 2 _drdd._ Plots are shown for CHG and CHH contexts in _drdd_ line A and all contexts in _drdd_ line B. Generation 2 WT is shown for comparison. 150 bp are shown on either side inside the dashed lines, and methylation is averaged over 25 bp bins. Left: de novo hyperMRs are between the vertical dashed lines. Right: spreading hyperMRs are outside of the vertical dashed lines, and the ancestral methylated regions adjacent to them are between the dashed lines. 27 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 28 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) **S5.** Heatmaps of average methylation in _drdd_ lines over class-specific supersets of all _drdd_ CHG (above) and CHH (below) hyperMRs ever called, between all lines and generations. Generation 2 WT is shown for comparison. HyperMRs are sorted by hierarchical clustering within each class. Total numbers of hyperMRs in each reduced superset are included in parentheses underneath each hyperMR class name. In the hyperMR status within each class, “Unique (1 line)” means a hyperMR was only called in one line throughout the experiment. “Shared (2+ lines)” means a hyperMR ever called in one line was called in at least one other line in at least one of the generations. Underneath the heatmaps are distributions of _drdd_ hyperMRs present in 1 line or shared between 2 or 3 lines in each hyperMR class. 29 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 30 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) **S6.** Heatmaps of average methylation in WT lines over class-specific supersets of all _drdd_ hyperMRs ever called, between all lines and generations. Heatmaps for all contexts are shown. HyperMRs are sorted by hierarchical clustering within each class. Total numbers of hyperMRs in each reduced superset are included in parentheses underneath each hyperMR class name. [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) 32 [bioRxiv preprint doi: https://doi.org/10.64898/2026.10.29.762914; this version posted October 30, 2026. The copyright holder for this preprint](https://doi.org/10.64898/2026.10.29.762914) (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made [available under a CC-BY-NC 4.0 International license.](http://creativecommons.org/licenses/by-nc/4.0/) **S8.** Gene Ontology (GO) enrichment analysis of differentially expressed genes (DEGs) in generation 9 _drdd_ lines A and B. For each line, the top 20 GO terms ranked by fold enrichment were selected after reducing redundancy among terms and requiring a minimum of five associated DEGs (see Methods). Bubble size represents the number of DEGs associated with each term, and color indicates −log10(FDR). Some GO term names were slightly abbreviated for clarity where necessary. 33