Extrachromosomal circular DNA and structural variants highlight genome instability in Arabidopsis epigenetic mutants

Abundant extrachromosomal circular DNA (eccDNA) is associated with transposable element (TE) activity. However, how the eccDNA compartment is controlled by epigenetic regulations and what is its impact on the genome is understudied. Here, using long reads, we sequence both the eccDNA compartment and the genome of Arabidopsis thaliana mutant plants affected in DNA methylation and post-transcriptional gene silencing. We detect a high load of TE-derived eccDNA with truncated and chimeric forms. On the genomic side, on top of truncated and full length TE neo-insertions, we detect complex structural variations (SVs) notably at a disease resistance cluster being a natural hotspot of SV. Finally, we serendipitously identify large tandem duplications in hypomethylated plants, suggesting that SVs could have been overlooked in epigenetic mutants. We propose that a high eccDNA load may alter DNA repair pathways leading to genome instability and the accumulation of SVs, at least in plants.

Abundant extrachromosomal circular DNA (eccDNA) is associated with transposable element (TE) activity. However, how the eccDNA compartment is controlled by epigenetic regulations and what is its impact on the genome is understudied. Here, using long reads, we sequence both the eccDNA compartment and the genome of Arabidopsis thaliana mutant plants affected in DNA methylation and post-transcriptional gene silencing. We detect a high load of TE-derived eccDNA with truncated and chimeric forms. On the genomic side, on top of truncated and full length TE neo-insertions, we detect complex structural variations (SVs) notably at a disease resistance cluster being a natural hotspot of SV. Finally, we serendipitously identify large tandem duplications in hypomethylated plants, suggesting that SVs could have been overlooked in epigenetic mutants. We propose that a high eccDNA load may alter DNA repair pathways leading to genome instability and the accumulation of SVs, at least in plants.
Extrachromosomal circular DNA (eccDNA) has been described for decades in eukaryote cells including fungi, Drosophila, nematodes, plants and mammals [1][2][3][4] . Only recently this fraction of the cell genetic material has gained attention thanks to the development of specific eccDNA-sequencing approaches allowing its characterization, such as mobilome-seq 5,6 , CIDER-seq 7 and circleseq 8 . EccDNA can originate from spurious homologous recombination between tandem copies (for example ribosomal DNA or telomeric DNA) or micro-homologies 9 . EccDNA can also be formed from circularization of extrachromosomal linear DNA of active transposable elements (TEs) 5,10 through homologous recombination (HR) or non-homologous end-joining (NHEJ). The roles of eccDNA are just being unraveled. For example the presence of eccDNA is associated with senescence in yeast and with apoptosis in mammal cells 11,12 . In stressful environmental conditions, eccDNA can provide the cells with an advantage. For instance in yeast, eccDNA arise in nitrogen limiting conditions and contributes to stress adaptation 13 . In cancer cells, abundant eccDNAs originating from chromothripsis (a massive event of genome rearrangement) can lead to oncogene amplification and contribute to tumor evolution [14][15][16] . In weeds, eccDNA encoded genes can contribute to herbicide resistance 17 , demonstrating the adaptive role of the eccDNA compartment in plants exposed to strong selective pressure 18 . Recently, PopRice, an abundant TEderived eccDNA in rice endosperm 5 was shown to act as a sponge for a transcription factor during germination 19 . The role of eccDNA upon their potential re-integration in the genome has been characterized in cancer cells 16 but the interplay between eccDNA and genome stability is poorly described in other biological systems. In yeast and cattle, early works suggested that  29 . Mutant plants affected in one or a combination of these pathways have a high level of TE transcription 30,31 . DNA methylation is maintained directly by methyltransferases and by the RNA-directed DNA methylation pathway (RdDM). It is also indirectly maintained by the chromatin remodeler DDM1 (Decrease DNA Methylation 1) involved in the deposition of the heterochromatic histone variant H2A.W at full-length TEs 32 . DDM1 belongs to DHL chromatin remodelers, an acronym for a family also comprising human HELLS (Human helicase lymphoid specific) and mice Lsh (Lymphoid specific helicase) 33 . Recent studies suggest a role of histone variants in DNA repair 34 . For instance, HELLS plays a role in HR repair of heterochromatic breaks 35 and Lsh promotes DNA repair in mice 36 . HELLS and Lsh are expressed in testis and in lymphoid cells, where they participate in V(D)J recombination. In Arabidopsis thaliana, DDM1 ensures HR repair of double strand breaks (DSBs) 37 . Loss of DDM1 is associated with a complete change in the epigenome, notably through changes in DNA methylation, histone methylation at H3K9, and H2A.W 33 . Because of its multiple roles, we wondered whether loss of DDM1 could have an impact on the accumulation of eccDNA in plants.
Here, using a reverse genetics approach, we can perturb the regulation of eccDNA and directly address the role of eccDNA on genome stability in A. thaliana mutant plants. In order to further the load of TEderived eccDNA in A. thaliana (hereafter TE-eccDNA), we use a combination of mutant plants affected in DDM1, RDR6 (involved in PTGS) and NRPD1 (the largest catalytic subunit of the POL IV holoenzyme that participates in RdDM). We perform mobilome-seq and long read genome resequencing to characterize both the eccDNA and genome contents and analyze the impact of eccDNA on genomic stability. Thanks to long read mobilome-seq, we show that TE-eccDNA can be truncated or chimeric. By analyzing the genomic content of these mutants plants, we uncover examples of SVs suggesting a high level of genome instability in these genetic mutant backgrounds.

Detection of eccDNAs and TE-mediated SVs in epigenetic mutants
To characterize the eccDNA repertoire of the wild type and epigenetic mutants (ddm1, polIV, rdr6, ddm1 polIV, ddm1 rdr6, and ddm1 polIV rdr6), a series of 21 mobilome-seq experiments were conducted (Fig. 1a). Briefly, after removal of genomic linear DNA with exonuclease V, the circular DNA molecules were amplified by random rolling circle and sequenced using Illumina or ONT platforms ( Supplementary Fig. 1). Additionally, thirteen whole genome resequencing datasets using ONT long reads were produced from these samples, to investigate SVs (Fig. 1a). The most abundant class of TE-eccDNA was identified as corresponding to the long terminal repeat (LTR) retrotransposon EVD, a TE previously observed as eccDNA in hypomethylated mutants 38 . EVD-eccDNA were present in all libraries from ddm1 mutant combinations mutants, but not in polIV or rdr6 single mutants (Fig. 1b). The sequencing coverage at the EVD/AT5G17125 locus reached up to 37,260X in the triple mutant, compared to 6X in the WT. The second enriched TE-eccDNA corresponds to the DNA transposon of the Mutator family VANDAL21 (Supplementary Fig. 2). The sequencing coverage at the VANDAL21/AT2TE42810 locus reached up to 2,080X in the triple mutant, compared to 6X in the WT. At both EVD (Fig. 1c) and VANDAL21 ( Supplementary Fig. 2) loci the reads cover the entire elements, confirming that these loci lead to eccDNA formation. In order to capture the full picture of the circular structures, we performed mobilome-seq using ONT reads on the WT and ddm1 polIV rdr6 triple mutant plants. Thanks to eccDNA amplification with random rolling circle, the ONT sequencing reads contain tandem repeated units of the eccDNA template ( Supplementary Fig. 1), allowing us to decipher the real structure of eccDNAs. Compared to Illumina mobilome-seq, ONT mobilome-seq data yielded a higher and more homogenous sequencing depth across both EVD (Fig. 1c) and VANDAL21 (Supplementary Fig. 2), showing clearer boundaries. The accumulation of reads corresponding to both TE families was specific to the triple mutant and consistent in the two biological replicates ( Supplementary Fig. 3). In contrast, reads corresponding to chloroplast and mitochondrial circular genomes and to endogenous rDNA eccDNAs (positive controls also amplified in our procedure) were detected in all samples ( Supplementary Fig. 3). Altogether these data confirm the formation of bona fida eccDNAs from both EVD and VANDAL21 TE families. Among the 318 TE families that we analyzed, these two were the only ones enriched in both biological replicates of the triple mutant compared to the WT (Fig. 1). For example we found evidence of eccDNA for ATCOPIA51 and ATCOPIA52/SYSIPHUS only in one biological replicate.
To investigate the activity of corresponding TEs, we analyzed genomic SVs mediated by EVD and VANDAL21, and detected up to 73 and 6 new insertions for EVD and VANDAL21, respectively, in a pool of 3 ddm1 polIV rdr6 mutant plants (Fig. 1d). In the single ddm1 individual plant we detected 26 EVD insertions, whereas in individual ddm1 polIV we detected 56 and 1 insertions for EVD and VANDAL21, respectively (Fig. 1d). Intriguingly in ddm1 rdr6 we did not detect any EVD insertion. Additionally, in ddm1 polIV and ddm1 polIV rdr6 mutants we detected insertions of two DNA transposons (ATENSPM3 and ATMU5) and the retrotransposon ATCOPIA21, not detected as eccDNA and that could be active in earlier generations or at a different developmental stage. Copia retrotransposons (EVD and ATCOPIA21) inserted preferentially within exons, VANDAL21 mainly targeted the 5′ UTRs of active genes, and ATENSPM3 insertion sites were more widely distributed next to genes (Fig. 1e).

Truncated TEs are present as both eccDNA and new insertions loci
Ty1/Copia retrotransposons comprise two LTRs surrounding an open reading frame that encodes a GAG (Group-specific Antigen) and a POL (polyprotein), which is subsequently cleaved into four active functional domains, namely AP (aspartic protease), IN (integrase), RT (reverse transcriptase) and RH (RNase H) (Fig. 2a). In addition to eccDNAs corresponding to expected full-length EVD with one or 2 LTRs (Fig. 2b), we identified eccDNAs with partial structures: loss of GAG domain (13%), loss of IN and RT domains (19%), and loss of LTR (20%). These truncated EVD-eccDNAs account for 52% of all EVD-eccDNAs, while full-length EVD-eccD-NAs represent 48% of all EVD-eccDNAs (Fig. 2b). No hotspot for truncation could be detected ( Supplementary Fig. 4), despite a very high coverage with ONT reads (up to 34,695X, Fig. 1c).
Given the high frequency of truncated eccDNAs we analyzed their potential impact on genomic SVs. In this aim we selected EVD containing reads in our ONT genome resequencing datasets for WT, ddm1, ddm1 polIV, and ddm1 polIV rdr6 mutants. We excluded single polIV and rdr6 mutants and ddm1 rdr6 double mutants as there was no new detected EVD insertion in these mutant backgrounds. New insertions corresponding to truncated retrotransposon structures were detected at distinct loci in mutants (Fig. 2c, d). We also observed a truncated insertion for ATCOPIA21 ( Supplementary Fig. 5). Two of these truncated insertions contain clear target site duplications (TSDs). Of note, all truncated insertions possess two LTRs (Fig. 2c, d) suggesting that only a fraction of truncated TEs are linked to new insertion events.

Both eccDNAs and genomic SVs contain chimeric TEs
All eccDNAs described so far consisted exclusively of TE-derived DNA. In contrast, in ddm1 polIV rdr6 plants we observed a striking example of chimeric eccDNA, defined as a circle originating from different genomic loci. In this eccDNA, supported by 7 copies in a single long read (Fig. 3a), a partial EVD fragment and a portion of the AT5G66440 gene (encoding a tRNA methyltransferase) had fused. EVD containing reads were extracted from ONT mobilomeseq datasets for the triple mutant plants and mapped to the reference genome. The mapping profile confirmed the presence of chimeric reads at the AT5G66440 locus (Fig. 3b). Additionally, we detected an EVD insertion at the corresponding locus (AT5G66440) in the genome of the same plants (Fig. 3c), flanked by a 5 bp TSD, a signature of integrase-mediated insertion (Fig. 3d). This finding suggests that the chimeric eccDNA might originate from the new insertion of EVD at this locus.
To investigate whether this chimerism was specific to the EVD family or could be detected for other TE families, we extracted ATCOPIA21 long reads from two replicates of ddm1 rdr6 polIV ONT genomic data. We noticed that these ATCOPIA21-containing reads displayed complex re-arrangements when mapped to the reference genome. They were divided into three segments mapping to distinct genomic regions, and here-after referred to as "3-hit" reads. One "3-hit" read started from the AT4G16950 gene encoding the pathogen response gene RPP5 (for RECOGNITION of PERONOSPORA PARASITICA 5), then spanned ATCOPIA21, and ended at the AT4G16970 gene encoding a CRK19 (CYSTEIN-RICH RECEPTOR-LIKE PROTEIN KINASE 19) located 3 Mb away from the RPP5 cluster on the reference genome ( Supplementary Fig. 6). This was confirmed by mapping the ATCOPIA21 containing ONT reads on the reference genome at both RPP5 and CRK19 loci. The insertion of ATCOPIA21 at the CRK19 gene created a 9 bp TSD (TATAGTAGC). This TSD seems unusually long 39 . but suggests a proper integration event. This structural remodeling resulted in RPP5 gene duplication close to the CRK19 gene in the ddm1 rdr6 polIV triple mutant ( Supplementary Fig. 6). The RPP5 locus is highly repetitive (Supplementary Fig. 7). The detection of chimeric duplicated RPP5 locus in ddm1 rdr6 polIV prompted us to analyze the stability at the endogenous RPP5 locus in this mutant background. While we did not detect any SV neither in Col-0 nor in ddm1 at this locus, we detected a 13.3 kb deletion in the triple mutant, in one biological replicate (Supplementary Fig. 7). This replicate comes from a pool of 3 plants, suggesting that one plant could be heterozygous for the deletion. ddm1 mutants have hypomethylated TEs. In order to test a different line showing hypomethylated and active TEs, we analyzed met1-derived epiRILs genomes 40,41 . When compared to the Col-0 control, 9 out of 21 lines analyzed displayed a lack of coverage at the RPP5 locus, compatible with a recombination event between the two flanking genes AT4G16940 and AT4G16960 (Supplementary Fig. 8). This result suggests that, in independent conditions of hypomethylation, this locus displays genomic instability.

Large SVs in hypomethylated plants beyond eccDNA
The prevalence of detected SVs in the epigenetic mutants prompted us to characterize all SVs in these mutants, beyond TEs producing TE-eccDNAs. We detected a 55 kb (Chr1:5,548,395-5,603,615) and a 56 kb duplication (Chr2:231,518-287,731) in two ddm1 siblings, respectively (Fig. 4a). The first duplication starts at the 5'UTR of AT1G16220 and ends within the AT1G16390 exon (Fig. 4b). We analyzed long reads crossing the junction of the two tandem copies and confirmed the 55 kb duplication by dot plots (Fig. 4c). We did not detect SNPs in the two tandem copies, but the second copy is slightly hypomethylated at the left breakpoint compared to the first one ( Supplementary Fig. 9). This suggests that the two tandem copies are recent, differing only in their epigenetic patterns. The second duplication is 56 kb long, starting in the exon 2 of the AT2G01510 gene and ending between two DNA transposons (AT2TE01165 and AT2TE01180) (Fig. 4d). A third 34 kb long duplication was identified in ddm1 polIV in a repeated region ( Supplementary Fig. 10).
We did not identify eccDNAs corresponding to these large segmental duplications suggesting that these eccDNA forms could be transient or that these SVs could be initiated though other pathways. We reasoned that the accumulation of eccDNA might affect the DNA repair pathway. Large tandem duplications have been observed in Caenorhabditis elegans mutant affected in DNA repair, together with small deletions 42 . To evaluate whether ddm1 mutant phenocopies this DNA repair mutant, we analyzed small deletions in single plants of ddm1 and ddm1 polIV and found 85 and 51 deletions absent in the WT, respectively (Fig. 4).

Discussion
We investigated both eccDNA content and SVs in A. thaliana epigenetic mutants in order to characterize the eccDNA repertoire and document whether it can impact genomic SVs (Fig. 5). In WT plants the formation of TE-eccDNAs is prevented by DNA methylation pathways and PTGS. In ddm1 rdr6 polIV mutant plants, we detected a heavy load of eccDNAs, most of them originating from EVD, an LTR retrotransposon active in hypomethylated plants 38   b eccDNA-seq ONT reads containing EVD were selected and mapped to the reference genome and shown at the AT5G66440 locus. c New TE insertions detected in the ddm1 rdr6 polIV mutant genome. A new copy of EVD is detected at the AT5G66440 locus (red triangle). d The EVD insertion at AT5G66440 created a 5 bp target site duplication (TSD, black line), visible thanks to read mapping locations ending or starting at the same position on the reference.
EVD-eccDNAs were composed of truncated EVD in ddm1 rdr6 polIV. Truncated insertions of EVD were detected in the corresponding ddm1 rdr6 polIV genome, but also in single ddm1 mutant plants, suggesting that the ddm1 mutant background itself is sufficient to promote these truncated TE insertions. This points to a mechanism of TE internal deletion and formation of non-autonomous elements. Truncated insertions have been described earlier 44-47 however it is not yet clear how these insertions occur. Future investigations will be needed to characterize whether this phenomenon occurs (i) early during TE lifecycle in the VLP, as suggested by reshuffling for Onsen 48,49 or ATCO-PIA52/SISYPHUS 50 , (ii) post-integration in the genome or (iii) if truncated eccDNA would be able to re-insert into the genome. TE-eccDNAs are generally considered as a dead-end, not being able to reintegrate into the genome 51 . However, in yeast, cattle and cancer cells, integration of eccDNAs has been suggested 20,21,52-54 . While mobilome-seq data suggested that EVD and VANDAL21 were active in the ddm1 mutant combinations, genome resequencing using long reads revealed insertions of TEs from additional families: ATCO-PIA21, ATENSPM3 and ATMU5. Several hypotheses could explain why eccDNA corresponding to these 3 families could not be detected. These eccDNAs could be expressed in a tissue specific manner, or with a very limited short life, for yet unknown reasons, limiting our power to detect them. For instance, concerning class II elements, Mu and Ds elements preferentially insert in genes expressed in meristematic enriched tissues 55 . Future experiments of targeted mobilome-seq could help addressing this issue. The activity of TE families might also differ between plants or lines of the same genotype, as shown before for VANDAL21, CACTA, ATCOPIA13 or ATGP3 for instance 43 . This could also explain why we did not detect new insertions of EVD in ddm1 rdr6, contrary to a previous report 50 , and despite the presence of EVD-eccDNA.
On top of expected TE neo-insertions, we uncovered complex SVs in ddm1 mutant combinations. For example we detected chimeric genic reads consisting of DNA sequences originating from ATCOPIA21 and two genes fragments: RPP5 and CRK19, revealing a partial duplication of the RPP5 locus. This locus is located within a resistance (R) gene cluster that plays an important role in the innate immune response to pathogens in A. thaliana 56 . Clusters of R genes are critical for plant resistance in the evolutionary arms race with pathogens, and these regions are enriched for TEs 57 . Intriguingly we observed a large deletion at the RPP5 locus both in one biological replicate of ddm1 rdr6 polIV mutants and in met1-derived epiRIL population suggesting that genetic instability could be a direct or indirect consequence of DNA hypomethylation at this locus. This observation is consistent with the lack of RPP5 expression previously reported in met1 epiRILs 58 . Some years ago a 55 kb duplication at this locus (named bal) had been reported and associated with a dwarf plant phenotype 59 . A parsimonious explanation could be that TE instability can lead to genomic SVs at this locus. It is also possible that TEs preferentially insert in regions of instability while SVs are caused by changes in DNA methylation regardless of TE insertion. The role of TEs in remodeling the genome is reminiscent of the recent discovery of LINE1-induced SVs in human cancer genomes 60 . More systematic analyses of individual plants with TE bursts and associated eccDNAs will allow addressing the extent of TE-driven structural remodeling. Beyond chimeric or truncated TE insertions linked to eccDNA we have identified large tandem duplications and small deletions in single plants of ddm1 and ddm1 polIV, not associated with eccDNA, suggesting an indirect link between the presence of eccDNAs and these SVs. Large tandem duplications have also been reported in lines derived from a fas2 mutant 61 . Both loss of DDM1 or FAS2 correlate with higher rates of meiotic 62 or homologous recombination 63 , respectively, showing that DNA repair pathways are affected in these mutants. Interestingly we found that the pattern of SVs in ddm1 and ddm1 polIV (large tandem duplications and small deletions) phenocopies the BRCA1-associated SVs found in a C. elegans model 42 . During their life cycle, proliferating TEs produce linear extrachromosomal DNA that presents double strand breaks (DSBs) and eventually get circularized presumably though homologous recombination or non-homologous end-joining to produce eccDNA. In hypomethylated plants, the high load of eccDNA coupled with active integration of TEs (through integrase or transposase) could result in an excess of DSBs in the cell, overloading the repair machinery (Fig. 5). We propose that this could trigger alternative pathways for DNA repair, such as the recently described TMEJ involving the polymerase theta 64 , leading to genomic SVs. Of note, while this study was in press, a publication reported the role of DNA repair pathways in eccDNA formation in Drosophila 65 .
By sequencing both the mobilomes and genomes of ddm1, ddm1 polIV and ddm1 rdr6 polIV plants we detected a high eccDNA load associated with genome instability. We speculate that the abundance of eccDNA in ddm1 mutant combinations triggers alternative DNA repair pathways leading to SVs and mimicking mutants in DNA repair. During evolution, TE-derived eccDNAs could contribute to hotspots of SVs and participate to the rapid evolution of disease resistance clusters in natural plant populations. Analyses in other biological models would allow testing the hypothesis that TEs and/or TE-eccDNA contribute to a « two speed » genome evolution mechanism by contributing to instability at genomic hotspots.

Plant material and growth conditions
Seeds from A. thaliana WT ecotype Columbia-0 and mutants were sown on soil, stratified for 2 days at 4°C and grown in a growth chamber at 21°C under long day conditions (16 h light). The mutant alleles used in this study were ddm1-2 66 first and six generation mutants, rdr6-15 (SAIL_617) and polIV/nrpd1-3 67 . The mutant combinations have been previously described 68 . For eccDNA analyses, leaf material was harvested from a pool of 12 plants after 3 weeks. For ONT genome resequencing, aerial part from individual plants was harvested after 3 weeks, except for the dwarf ddm1 polIV rdr6 mutant plants for which 3 siblings had to be pooled for DNA extraction. We performed at least two biological replicates for all DNA extractions, except for the polIV and rdr6 single mutants where one replicate was used, as these single mutants have no reported evidence of active TEs.

DNA extraction
Seedlings were collected into one tube immediately snap-frozen in liquid nitrogen and stored at −80°C until DNA extraction, in duplicate or triplicate. Total DNA was extracted using the CTAB approach 5 . Total DNA quantity was measured with a Qubit Fluorometer (Thermo Fisher Scientific). Genomic DNA was quality controlled using Nanodrop measures, Qubit quantifications and visualization on a 1% agarose gel electrophoresis.
eccDNA enrichment eccDNA enriched was performed using the Phi29 enzyme 5 . Briefly, genomic DNA (2 μg) was first cleaned up using the PCR Cleanup kit (Qiagen) to remove large genomic fragments. The DNA on the column was eluted with 30 μL elution buffer. DNA concentration was measured with a Qubit fluorimeter. In order to eliminate linear DNA fragments, 25 μL of purified DNA was digested with the Plasmid-Safe ATP-Dependent DNase (Epicenter) according to the manufacturer's instructions. Following digestion, eccDNA was precipitated with 0.1 volume of 3 M sodium acetate (pH 5.2), 2.5 volumes of ethanol, and 1 μL of GlycoBlueTM Coprecipitant (Ambion) overnight at −20°C. After centrifugation at 4°C for 1 h and washing with 70% ethanol, eccDNA was directly resuspended in the Illustra TempliPhi Sample Buffer and then amplified by random rolling circle amplification (RCA) using the Illustra TempliPhi Amplification Kit (GE Healthcare) according to the manufacturer's instructions, except that the reaction was incubated for 65 h at 28°C. The enzyme was inactivated for 30 min at 70°C. Out of 10 μL TempliPhi reaction, 1 μL was used to perform an EcoRI digestion (New England Biolabs) overnight at 37°C. Quality control was performed by loading the EcoRI digestion product on a 1% agarose gel. The presence of a smear on the gel indicated a proper amplification step.

Mobilome-seq with Illumina
For sequencing, 5 ng of amplified DNA (typically using a few μL of a 1/20 dilution of the TempliPhi amplification product) were used to prepare libraries using the Nextera XT kit, following the manufacturer's instructions. Twelve libraries were pooled per run. Sequencing was performed in house (UPVD Bio-Environment sequencing platform) using a Miseq with a 2 × 300 bp sequencing cassette. Quality control was performed on the resulting FASTQ files.

Mobilome-seq with Nanopore
For ONT sequencing, 500 ng of amplified eccDNA from the TempliPhi product were used to prepare an ONT library using the Nanopore Rapid Sequencing Kit (SQK-RAD004) following the manufacturer's instructions. The libraries were sequenced in house (UPVD Bio-Environment sequencing platform) on ONT MinIONs using R9.4.1 flow cells. Basecalling was done on a GPU machine using Guppy with high-accuracy mode.

Detection of eccDNAs from Illumina and ONT eccDNA-seq
The eccDNA producing loci from each Arabidopsis epigenetic mutant were detected using ecc_finder 6 with default parameters of short-readmapping and long-read-mapping mode (for Illumina and ONT data, respectively). eccDNAs originating from organelle DNA fragments mapping to their nucleic copy (such as NUPTs for nucleoplasmic DNA and NUMTs for nuclear mitochondrial DNA), from ribosomal DNA repeats (rDNA) and from centromeric repeats were removed. The remaining eccDNAs were grouped into TE families by mapping them with BWA to the reference genome and normalized with FPKM (fragments per kilo base per million mapped reads, paired-end). For ONT eccDNA data, an additional counting of the unique reads was done. Reads from FASTQ files were mapped on the TAIR10 reference genome using minimap2 (version 2.24) with the following command $ mini-map2 map-ont -ax -t 8 -MD reference.fa file.fastq | samtools view -Sb -u | samtools sort > $output.file.sort.bam. The bam file was converted into bed to collect the reads positions using bedtools (version 2.30.0). Unique reads were collected and the number of reads per TE or gene was calculated using bedtools intersect with the TE/gene annotation file and an homemade script. For each TE family the total number of unique reads was compiled using R for the 318 annotated TE families.

EVD functional annotation
The two LTR sequences were identified by self-to-self alignment using BLAST. The sh-GAG structure of EVD was identified from long read RNA-seq 30 . RH sequence was identified using VLP data from the same ddm1 mutant 50 .
Detection of truncated and chimeric eccDNAs from ONT eccDNA-seq EccDNA-seq data were filtered for reads containing EVD. These EVD reads were mapped on the annotated structural domains of the EVD sequence using minimap2. Different profiles, such as the loss of different structural domains, were visualized using dot plot and then systematically grouped using bedtools groupby. In the next step, EVD reads were remapped to the reference genome using minimap2 to check for unmapped sequences. Chimeric eccDNAs supported by at least 5 reads were retained. The same approach was used for ATCO-PIA21 eccDNAs.
Genomic DNA sequencing with Nanopore Genomic DNA was quality controlled using Nanodrop and Qubit quantifications. The libraries were sequenced in house (UPVD Bio-Environment sequencing platform) on ONT MinIONs using either R9.4.1 or R10 flow cells. Basecalling was done on a GPU machine using Guppy with high-accuracy mode.

Detection of TE insertion polymorphisms from ONT genomic reads
Reads spanning the entire insertion and deletion sequence do not cause alignment breakpoints, but are flagged in the CIGAR. We developed a script to filter the CIGAR output. This pipeline can be found on GitHub (https://github.com/njaupan/CIGAR_SV). Briefly, in order to extract all reads containing insertions and deletions, we generated a PAF file from minimap2 (-cs -cx map-ont) 69 . The CIGAR metrics were then indexed until the position of the SV on the reference genome. Breakpoints with more than 4 supported reads at the same position were selected. The start and end positions of the breakpoints were extracted from the PAF file and grouped to generate common breakpoints displayed in BED format. The breakpoint locations were finally filtered for the presence of 5-20 bp TSDs supporting bona fide TE insertion polymorphisms.

Identification of SVs
SVs from minimap2 alignments were detected using Sniffles v2.0 70 , which generated a VCF file for each A. thaliana genotype separately. For duplications the VCF files were furthered filtered for duplications larger than 1 kb. Read depth was calculated with samtools depth and visualized with samplot. Reads spanning the junctions of the tandem duplications were extracted to manually validate the segmental duplications. For deletions the following parameter was used: -minsvlength 20 in order to recover small deletions for all samples. Deletions specific to the mutants were detected comparing the VCF files from the mutants versus the WT using bedtools intersect with -f = 1E-9 (1 base overlap).

Detection of DNA methylation from ONT genomic reads
Cytosine methylation patterns were detected from ONT reads using Nanopolish. The methylation patterns were parsed and plotted using methylartist (https://github.com/adamewing/methylartist).

Re-analysis of DNAseq from met1-derived epiRILs
Raw data from DNAseq of met1-epiRILs was obtained from previously published work 41 (GEO ID: GSE120571) and re-analyzed. Briefly, after alignment with Bowtie2, peak calling was done for each epiRIL mapped bam file in comparison to the Col-0 control condition with MACS2 (https://github.com/taoliu/MACS/), setting the fragment size to 75 and extended size to 400 bp, with the options -B, -SPMR and -no model and Columbia-0 as control. The Log based background subtraction was performed with MACS2 bdgcmp with "logLR" as model and a P-value threshold of 0.00001.

Plots and visualization
For the visualization samplot was used. For the scheme we used bioicons (https://bioicons.com/).

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
All high-throughput sequencing data generated in this study have been deposited at NCBI under Bioproject accession PRJNA956454. Source data are provided with this paper.