Extreme haplotype variation in the desiccation-tolerant clubmoss Selaginella lepidophylla

Plant genome size varies by four orders of magnitude, and most of this variation stems from dynamic changes in repetitive DNA content. Here we report the small 109 Mb genome of Selaginella lepidophylla, a clubmoss with extreme desiccation tolerance. Single-molecule sequencing enables accurate haplotype assembly of a single heterozygous S. lepidophylla plant, revealing extensive structural variation. We observe numerous haplotype-specific deletions consisting of largely repetitive and heavily methylated sequences, with enrichment in young Gypsy LTR retrotransposons. Such elements are active but rapidly deleted, suggesting “bloat and purge” to maintain a small genome size. Unlike all other land plant lineages, Selaginella has no evidence of a whole-genome duplication event in its evolutionary history, but instead shows unique tandem gene duplication patterns reflecting adaptation to extreme drying. Gene expression changes during desiccation in S. lepidophylla mirror patterns observed across angiosperm resurrection plants.

T he ability to survive periodic drying was essential for the aquatic ancestors of terrestrial plants during their transition from sea to land. These early protective mechanisms are thought to be ancestral and conserved in most land plants 1 , forming the basis of modern seed and pollen desiccation pathways 2 . A small group of plants have maintained this ancestral desiccation tolerance, surviving near complete anhydrobiosis (<10% relative water content) during prolonged drought events. Desiccation tolerance is widespread in non-seed plants such as mosses, ferns, and lycophytes 3,4 , but is uncommon in angiosperms. High-quality genomes for several angiosperm resurrection plants have shed light on the molecular basis of desiccation tolerance 2,5,6 , but the sparse genomic resources for non-seed plants has hindered progress toward understanding the origin of this trait.
Here we report the small genome of Selaginella lepidophylla, a lycophyte that can survive the loss of >95% cellular water for decades in a quiescent state. Selaginella are among the earliest diverging lineages of vascular plants, and extant species bear striking morphological similarities to fossils from the upper Carboniferous (333-350 Myr ago) 7,8 . Genome size variation within this ancient group is minimal and Selagienlla are the only clade of land plants with no shared whole-genome duplication event 9 . Detailed comparative genomics across angiosperms and with the S. moellendorffii genome 10 shed light the origin of desiccation tolerance and unique properties of Selaginella.

Results
Haplotype-level genome assembly and annotation. Extensive within genome heterozygosity exists for most plant species, and this variation enhances fitness and serves as a reservoir for adaptive evolution. Heterozygosity is a major issue in plant genome assembly, and most heterozygous plant genomes are highly fragmented with collapsed haplotype regions 11 . S. lepidophylla has an estimated genome size of 109 Mb based on kmer frequency analyses, which is consistent with the small genomes observed across the genus 9 . Despite its compact size, the S. lepidophylla genome is highly heterozygous, with a bimodal distribution of kmers corresponding to distinct heterozygous and collapsed homozygous regions ( Supplementary Fig. 1). Initial sequencing attempts using an Illumina only approach yielded a highly collapsed and fragmented assembly of 21,282 scaffolds, totaling 54 Mb, with an N50 of 3 kb (Supplementary Table 1). To overcome assembly issues related to heterozygosity, we generated 175× coverage (19.3 Gb) of single-molecule real-time (SMRT) sequences from PacBio with a filtered subread N50 length of 18 kb ( Supplementary Fig. 2). The long read lengths of PacBio sequencing facilitate accurate separation and assembly of heterozygotic haplotypes 12 . Raw PacBio reads were error corrected and assembled using the Canu assembler 13 , which utilizes optimized overlapping and assembly algorithms to avoid collapsing repetitive regions and haplotypes. The resulting contigs were polished using the raw PacBio data with Quiver 14 , followed by a second round of polishing with Pilon 15 using the high-coverage Illumina data.
The final S. lepidophylla assembly spans 122 Mb across 1149 contigs with an N50 length of 163 kb (Supplementary Table 2). These assembly metrics are similar to the S. moellendorffii genome (contig N50 length = 120 kb), which was sequenced using a reiterative Sanger approach 10 . Our assembly includes 15 of the 20 terminal telomere tracks (2n = 20) and 128 rRNA arrays, which are typically unassembled in most next-generation sequencing (NGS) based genomes 11 . The S. lepidophylla assembly graph is highly complex with numerous interconnected edges, reflecting bubbles and ambiguities caused by heterozygous regions (Fig. 1). The long PacBio reads assembled 930 haplotype blocks spanning 33.7 Mb of separated haplotype sequences (Methods section; Fig. 2a; Supplementary Fig. 3). Other haplotype aware assemblers were tested but Canu provided the best separation of haplotype blocks as well as the best contiguity. We annotated 27,204 genes across a deduced 89 Mb haploid assembly, considerably more than S. moellendorffii (22,285 genes) 10 . The majority (96%) of core eukaryotic BUSCOs 16 were identified in the gene set, supporting the completeness of the genome assembly and quality of the annotation (Supplementary Table 3). Roughly 45% of genes (11,847) are found in duplicated pairs across the full 122 Mb assembly with an average protein homology of 99.1%. Gene density in S. lepidophylla is 4.0 kb per gene, which is higher than S. moellendorffi 10 and Arabidopsis 17 , but lower than the smallest sequenced plant genome (82 Mb) of Utricularia gibba 18 .
Genome evolution and haplotype-specific structural variation. Genome size varies by four orders of magnitude across land plants, ranging from 60 Mb in the carnivorous Genlisea to 152,000 Mb in Paris japonica. Lycophytes have a narrow variation in genome size compared to angiosperms 19 , and Selaginella sizes range from 90 to 182 Mb 9 . Rapid changes in genome size are driven by bursts of long terminal repeat (LTR) retrotransposon proliferation followed by aggressive deletion through unequal homologous recombination and double-strand break repair 20 . LTR elements are distributed non-randomly across the genome with Gypsy-like elements predominantly found in gene poor and repeat rich blocks while Copia-like elements are dispersed among genes (Supplementary Table 4). Most of the LTR elements are clustered in a relatively small number of LTR families (Supplementary Figure 4). We identified 744 full-length LTR retrotransposons (LTR-RTs), including 572 Gypsy and 17 Copia elements (Supplementary Table 5). The majority of intact LTRs are relatively young, with average insertion times of 0.1 and 0.4 MYA for Copia and Gypsy elements, respectively, and most intact LTR elements are younger than 0.5 Ma ( Fig. 2b and Supplementary Fig. 5). The proportion of young LTR elements is higher in S. lepidophylla than other high-quality plant genomes (Supplementary Figure 5). The recent LTR amplification suggests most elements have been active recently in Selaginella but they are rapidly fragmented and deleted. S. lepidophylla has a much higher ratio (4.06, Wilcoxon rank-sum, P < 0.05, same statistical test hereafter unless specified) of solo-intact LTR elements comparing to the genomes of rice and Arabidopsis (2.71 and 1.80, respectively) ( Supplementary Fig. 6), indicating more effective removal of LTR-RTs via unequal homologous recombination. Moreover, the average age of LTR families in S. lepidophylla is much younger (0.37 MY vs. 0.92 and 0.94 MY, respectively; P < 0.05) under the mutation rate of μ = 1.3 × 10-8 (per bp per year), suggesting rapid removal of LTR elements. The intensive and rapid removal of LTR-RTs is likely one of the causes of haplotype variation and enrichment of LTR-RTs (Fig. 2).
Haplotype blocks within the S. lepidophylla genome are largely collinear, with conserved gene content and 98.2% sequence homology ( Fig. 2a; Supplementary Fig. 3). However, since our analysis is based on both the number of contiguous protein and sequence homology, it is possible that we have missed large highly diverged haplotype blocks. A detailed comparison uncovered 104 large-scale haplotype-specific insertions or deletions (indels) across the genome. These high-confidence indels collectively span 2.5 Mb among 13.5 Mb of collinear sequences (18.5%) with individual sizes ranging from 4.8 kb to 85 kb with an average size of 24 kb (Supplementary Table 6). This is likely an underestimation as separate haplotypes were assembled for 35% of the genome and only high-confidence collinear regions spanning >75 kb were surveyed. Haplotype-specific regions are largely repetitive with 58% of bases consisting of LTR retrotransposons and only 9% genes (Fig. 2c). Flanking collinear regions show the opposite trend, with 86% genic and only 11% LTR retrotransposon sequences. We annotated 335 protein-coding genes in the gene poor, haplotype-specific regions, but most (74%; 248 genes) have little to no detectable expression (TPM <0.5). Young LTR retrotransposons (insertion times <0.1 Ma) are overrepresented in haplotype-specific regions, supporting a recent origin of these indels ( Fig. 2d; P < 0.01). DNA methylation is associated with suppressing repetitive element proliferation and larger, repeat-dense plant genomes tend to have higher levels of methylation 21 . We surveyed genome-wide DNA methylation in S. lepidophylla for patterns related to repetitive element silencing. Similar to other non-seed vascular plants 22,23 , S. lepidophylla lacks evidence of any gene body methylation (Fig. 3a). In contrast, intact LTR retrotransposons have high levels of CpG and CHG methylation with comparatively little CHH methylation (Fig. 3b). Methylation is highest in the body of LTR retrotransposons with levels dropping at the 5′-and 3′-flanking regions. Gene rich regions of the S. lepidophylla genome are highly compact with an average of 1.1 kb intergenic sequences with interspersed, repeat-dense blocks (Fig. 3c). This creates large tracks in the genome that are essentially devoid of methylation flanked by densely methylated, likely heterochromatic regions.
Comparative genomics across Selaginella and angiosperms. Selaginella arose in the early Carboniferous period 333-350 MYA 8 , and the ancestors of S. lepidophylla and S. moellendorffii diverged about 250 MYA 9 , well before the radiance of modern angiosperms 24 . Despite this separation, both lineages have maintained a similar genome size (~100 Mb) and neither has experienced a recent whole-genome duplication event, contrasting patterns observed in all other land plants 10,25 . S. lepidophylla and S. moellendorffii have maintained a surprising degree of collinearity with 36% (7849) of S. moellendorffi genes showing conserved order in S. lepidophylla. Genic regions are generally more compact in S. lepidophylla, correlating with the higher TE content (37.5%) in S. moellendorffii ( Supplementary Fig. 7). By contrast, only~15% of rice and core eudicot genes are collinear 26 , reflecting the role of WGD events in shaping gene order and genome architecture. In the absence of WGD, any new traits and adaptations associated with duplicate genes in Selaginella must arise through local or segmental gene duplications 27 . S. moellendorffii has significantly more tandem gene duplications (TD) than S. lepidophylla with 6693 compared to 4170, but both share similar enrichment pattern (Supplementary Tables 7-10). TDs in Selaginella are enriched in molecular function GO terms related to kinase activity, ion and other small molecule binding among others (Supplementary Tables 7 and 9). Few TDs were annotated as transcription factors or highly connected metabolic genes, which is consistent with expectations under the gene balance hypothesis 28,29 . Both Selaginella species have overlapping KEGG pathway enrichments across TDs including phenylpropanoid, cutin, and secondary metabolite metabolism (Supplementary Tables 8 and 10). S. lepidophylla has TDs enriched in phagosome pathways, which may be related to controlled autophagy during desiccation 30 .
Lycophytes are sister to all other vascular plants including angiosperms, the most diverse extant plant group on the planet. We surveyed changes in orthogroup composition across angiosperms and the two Selaginella species to identify patterns associated with WGD and life history traits. We identified 5891   Table 11). Angiosperm-dominated orthogroups have diverse GO terms and enriched KEGG pathways for alkaloid biosynthesis, proteolysis, and plant-pathogen interactions (Supplementary  Tables 12 and 13). The angiosperm biased orthogroups are more highly connected than expected by chance (~2.7 edges per gene; P ≤ 10 −20 , Fisher's exact test), fitting expectations for gene duplicates retained from WGD 28 . The 55 Selaginella-biased orthogroups have GO enrichment in vesicle related processes and KEGG pathway enrichment in phenylpropanoid biosynthesis (Supplementary Tables 14 and 15). Most Selaginella-biased orthogroups are unique and completely uncharacterized, reflecting the need for more functional studies in other major plant lineages.
Origins and evolution of desiccation tolerance. The transition from sea to land posed unique challenges for the ancestors of terrestrial plants. Early land plants needed to evolve protective mechanisms to survive periodic drying and low atmospheric water potential. Desiccation tolerance in angiosperms likely evolved from rewiring seed desiccation pathways 2 , but the origins of this trait in non-seed plants is unknown. We surveyed gene expression changes during the desiccation and rehydration process in S. lepidophylla and compared patterns to angiosperm resurrection plants. S. lepidophylla can survive for years or even decades in a desiccated state without loss in viability. Samples were taken from S. lepidophylla plants that were desiccated (0 h), 1, 6, and 24 h post rehydration (recovery), 120 h post rehydration (fully recovered), and after a 24 h dehydration (Fig. 4a). Despite prolonged quiescence, 77% of genes (21,062) have detectable expression in desiccated leaf tissue, implicating strong protective mechanisms to preserve the mRNA pool. We performed gene-co-expression analysis across the rehydration and subsequent dehydration RNAseq timecourse data. The 38 co-expression clusters fall under the following three general patterns with (1) high expression in desiccated tissues followed by decreases during rehydration, (2) increased expression during rehydration, and (3) steady expression across all samples (Fig. 4b,  Supplementary Fig. 8). Transcript levels are relatively consistent across the 0, 1, and 6 h post-rehydration time-points, with only 123 and 567 differentially expressed genes, respectively (0 h vs. 1 h and 1 h vs. 6 h). Clusters with increased expression during rehydration are enriched with functions related to DNA repair, oxidative and water stress response among others. Clusters with high expression in desiccated tissues and low expression during rehydration have functions related to cell protection and osmoprotectant accumulation as described in more detail below.
Gene duplications drive evolutionary innovation and tandem gene duplications (TDs) are important for adaptive evolution in dynamic environments 31 . Several clusters with peak expression during desiccation are enriched in TDs, including many with previously characterized roles in desiccation tolerance (Cluster 5, fdr = 1.26 e −16 ; Supplementary Fig. 9). Early light-induced proteins (ELIPs) protect photosynthetic apparatus against photo-oxidative damage under high light and other abiotic stresses 32 . Land plants typically have 1-2 ELIPs but S. lepidophylla has undergone a major expansion to 23 ELIPs. Most ELIPs are located in three tandem arrays with 8, 7, and 2 members, and the largest array is syntenic to the single ELIP found in S. moellendorffii (Fig. 4c). ELIPs are universally highly expressed during desiccation and show co-expression with other desiccation-related genes including heat-shock proteins, DNA    6 , suggesting these proteins are likely involved in the convergent evolution of desiccation tolerance. LEA proteins function in membrane stability and macromolecular protection with roles in seed development, response to abiotic stresses, and desiccation tolerance in angiosperm resurrection plants 2, 33, 34 . S. lepidophylla has an expansion in LEA proteins compared to S. moellendorffii with 65 vs 36 across six LEA subfamilies (Supplementary Table 17). The LEA2 and seed maturation protein (SMP) subfamilies show the most dramatic expansion with 31 and 11 members, respectively. LEAs are dispersed across the genome and few are found in tandem arrays. LEAs are among the most highly expressed transcripts in the S. lepidophylla desiccation/rehydration timecourse with 48 showing significant expression changes ( Supplementary  Fig. 10). Interestingly, LEAs are also highly expressed in wellhydrated tissues, suggesting they may be involved in constitutive protective mechanisms, similar to several osmoprotectants.
Desiccation-related genes are found in large tandem arrays in the S. lepidophylla genome with individual genes having similar expression patterns. This reflects the importance of TD in the evolution of novel traits and the advantages of gene proximity for coordinated expression. Patterns of gene expression during desiccation in S. lepidophylla mirror the pathways activated in angiosperm resurrection plants. These pathways were either independently co-opted through convergent evolution in both lineages or present in the common ancestor of early land plants. The latter explanation is more likely as dehydration tolerance was acquired during the colonization of land. This hypothesis can be further refined with additional non-seed resurrection plant genomes.
The extreme haplotype-specific variation in Selaginella has not been observed in other sequenced plant genomes. This dynamic flux of accumulating and purging sequences likely serves two purposes. Strong selective pressures for removal of excess LTR retrotransposons act to preserve proper gene and cis-element spacing, which in turn maintains the small genome size. Unequal crossing over between divergent haplotypes and LTR-mediated gene duplications create a reservoir of new genes to facilitate adaptive evolution. In this context, haplotype-specific gene content may supplement the diversity lost in the absence of WGD.

Methods
DNA extraction and sequencing. The Selaginella lepidophylla plants used in this study were collected in the Chihuahuan Desert in the Southwestern US. These plants were confirmed to be S. lepidophylla based on sequence homology to published vouchered specimens with the following nucleotide identities based on BLAST: rbcL (99.6%; GenBank: AF419051.1) ITS2 (99.8%; GenBank: AF419002.1) 28S (99.8%; GenBank: AJ507594.1). Desiccated plants were stored at 20°C in dark conditions for roughly 3 years prior to this experiment. Rehydrated plants were maintained in growth chambers under 24°C day, 20°C night, 12-h photoperiod, and 200 μmol m −2 s −1 light. High-molecular-weight genomic DNA was extracted from young leaf tissue of a single S. lepidophylla plant using a modified nuclei preparation 35 followed by phenol chloroform purification to remove residual contaminants. DNA quality was verified on a low-concentration agarose gel (0.5% W/V) prior to library construction. gDNA was sheared to a 30 kb target size followed by end repair and adapter ligation. The final library was size selected using a BluePippen size selection system (Sage Science) followed by purification using AMPure XP beads (Beckman Coulter). High-molecular-weight libraries were sequenced using the PacBio RsII instrument with P6C4 chemistry. A total of 21 SMRT cells were sequenced. A PCR-free Illumina DNAseq library was constructed using the same high-molecular weight DNA following the manufactories instructions. Paired end 300 bp reads were sequenced on an Illumina MiSeq platform yielding a total of 80× Illumina sequence coverage for error correction.
RNA extraction and sequencing. S. lepidophylla plants were maintained in growth chambers under 16-h-light/8-h-night regime at 22 ± 2°C for desiccation/ rehydration experiments. Sampling at the 0 h (desiccated), 24 and 120 h post rehydration and 24 h post dehydration were taken at T = 0 to avoid any variance associated with natural circadian oscillation. Sampling was done in triplicate with three biological replicates sequenced for each timepoint. The relative water content (RWC) for the S. lepidophylla rehydration time-course was calculated using the following equation: RWC = [(FW − DW)/(SW − DW)], where FW, DW, and SW indicate fresh weight, dry weight, and saturated weight, respectively. DW was obtained after drying plant leaf tissue at 103°C for 24 h and SW was obtained after submerging the leaf tissue in water for 36 h. RWC for each sample can be found in Supplementary Table 18 DNA extraction and methylation sequencing. Genomic DNA for methylation profiling was extracted from 500 mg of ground tissues using the CTAB method, followed by RNase digestion and column purification using Zymo Research genomic DNA clean & concentrator™ (Zymo Research). In total, 500 ng of genomic DNA was first sheared to an average size of 400-800 bp using Covaris M200. Libraries were constructed using KAPA Hyper Prep Kits (KapaBiosystem, #KK8504) according to the manufacturer's protocol. The sheared DNA fragments were end-repaired and A-tailed followed by adapter ligation. Adapter-ligated fragments were subjected to bisulfite conversion using EZ DNA Methylation-Lightning™ Kit (Zymo Research). Bisulfite-converted DNA was amplified using KAPA HiFi HotStart Uracil + Ready Mix (KapaBiosystem) with 8 PCR cycles. PCR products were then size-selected by gel electrophoresis for the range of 300-600 bp and recovered by Zymoclean ® Gel DNA Recovery Kit (Zymo Research). Multiplexed pooled libraries were sequenced on the Illumina HiSeq4000 platform under paired-end 150 nt mode.
Genome assembly. S. lepidophylla PacBio reads collectively span 19.3 Gb representing 175× genome coverage with a filtered subread N50 length of 18 kb. Raw reads were error corrected and assembled using the Canu (V1.4) 13 assembler with the following parameters: minReadLength = 3000, GenomeSize = 110 Mb, and minOverlapLength = 2000. Other Canu parameters were left as default. The Falcon assembler (V0. 3.0) was also tested but the overall contiguity was much lower (N50-110 kb) and a higher proportion of haplotypes were collapsed. Contigs were polished using a reiterative approach with two rounds of Quiver (V2.3.0) 14 with a minimum subread length = 5000 bp, minimum polymerase read quality = 0.8, maximum divergence percentage = 30 and minimum anchor size = 15. The polished contigs underwent a final round of error correction using high-coverage 300 bp PE Illumina data using Pilon (V1.21) 15 . Quality trimmed Illumina reads were aligned to the Quiver polished contigs using bowtie2 (V2.3.0) 36 with default parameters. Reads were locally re-aligned around insertions/deletions (indels) using the IndelRealigner from the genome analysis tool kit (GATK; V3.7) 37 . Pilon parameters were as follows: --flank 7, --K 51, and --mindepth 30. Illumina reads were re-aligned to the final polished genome to verify no residual errors remained using GATK.
Identification of haplotype-specific regions. Assembled haplotypes were identified using a two-step approach based on gene level microsynteny followed by whole-genome alignment. Gene synteny avoided spurious hits and misclassification around repetitive elements with whole-genome alignment facilitating base-pair resolution of overlapping haplotype regions. Syntenic gene pairs were identified with BLAST followed by filtering for only collinear pairs using the MCSCAN toolkit (V1.1) 38 . This approach flagged 11,847 duplicated gene pairs collectively representing~45% of genes across 28 Mb of the genome. The exact boundaries of haplotype overlaps and haplotype-specific structural variants were identified by whole-genome alignment using MUMmer (V3.0) 39 . The unmasked genome was aligned against itself using NUCmer with default parameters and a minimum overlap size of 2 kb. Only hits in contigs with collinear gene pairs were considered and overlaps in these regions were manually filtered removing misalignments in repetitive regions. In total, 33.7 Mb of assembled haplotype sequences were identified including 105 high-confidence haplotype-specific insertions/deletions (indel). The boundaries of each indel were manually refined using the GEvo module in CoGe (https://genomevolution.org) 40 . Haplotype-specific indels were verified through alignment of raw PacBio reads spanning the boundary regions in both haplotypes.
Transposable element identification. Non-autonomous DNA transposons in S. lepidophylla were identified by MITE-Hunter 41 and were classified according to the sequence of termini and length of target site duplication (TSD) provided in the output from MITE-Hunter based on the guide of Zhao et al. 42 . Candidate long terminal repeat retrotransposons (LTR-RTs) in S. lepidophylla were identified using LTR_Finder (V1.02) 43 and LTRharvest 44 . LTR_retriever 45 was used to filter out false LTR retrotransposons using structural and sequence features of target site duplications, terminal motifs, and LTR-RT Pfam domains. Whole-genome annotation of LTR retrotransposons was done by RepeatMasker (http://www. repeatmasker.org/) 46 using the non-redundant LTR-RT library constructed by LTR_retriever. DNA transposons that were nest-inserted in LTR retrotransposons were identified and removed in the LTR library using a Perl scrip "Purger.pl" from the LTR_retriever package. Non-LTR retrotransposons and other interspersed repeats were identified using RepeatModeler (V1.0.8) (http://www.repeatmasker. org). While RepeatModeler also reports LTR retrotransposons, their sequences were examined using TE-specific Pfam domains by HMMER (V3.1b2) 47 . Among 34 RepeatModeler-predicted LTR retrotransposons, none of them were exclusively matched with LTR-specific Pfam domains, thus all these sequences were labeled as unknown repeats instead of LTR retrotransposons. Candidate tandem repeats were identified by Tandem Repeat Finder (trf409.linux64) 48 and were masked by the TEs identified in previous steps. Unmasked candidates were labeled as tandem repeats, and the sequence redundancy were reduced using Cd-hit (V4.6.6) 49 with parameters -c 0.8, -G 0.8, -s 0.9, -aL 0.9, -aS 0.9, -M 0, and -T 5. The Perl script "Purger.pl" from LTR_retriever and the plant protein database "alluniR-efprexp082813" from MAKER (V2) 50 were used to remove any protein-coding sequences that were accidentally identified in these TEs. The top 10 most abundant repeat families (based on genome mass) were manually curated through alignment of multiple family members so to determine their boundary, termini, and TSD, which were used to classify the relevant repeat.
Solo LTRs were identified from the whole-genome annotation of LTR-RT using the Perl script "solo_finder.pl" in the LTR_retriever package. In brief, independent LTR sequences (no other LTR-related sequence presented in 300 bp distance) that cover at least 80% of the library LTR entry with alignment score >300 and at least 100 bp in length were identified as solo LTR. The solo-intact LTR ratio was calculated based on the number of solo LTR over the number of intact LTR-RT for each family, which was estimated using the Perl script "solo_intact_ratio.pl" in the LTR_retriever package. Since LTR length is positively correlated with the formation of solo LTRs, we only compared the S/I ratio using elements with comparable LTR length. This was achieved by using the 95 percentile of LTR length in S. lepidophylla (90-1540 bp) to screen for LTR-RT families within into this range in the rice and Arabidopsis genomes.
Estimation of insertion time for each intact LTR element was provided by LTR_retriever using the formula of T = K/2μ, where K is the divergence rate approximated by percent identity and adjusted using the Jukes-Cantor model for non-coding sequences. The neutral mutation rate of μ = 1.3 × 10 -8 mutations per bp per year was used 51 .
Genome annotation. The S. lepidophylla genome was annotated using MAKER (V 2.31.8) 52 . A set of high-quality transcripts for gene prediction were assembled with StringTie (V1.3.1) 53 using the stranded, paired end Illumina RNAseq data described above. Quality filtered RNAseq reads were aligned to the unmasked S. lepidophylla genome using TopHat (V2.1.0) 54 and the accepted_hits.bam file was used as input for StringTie. Default parameters were used for StringTie with the --merge flag to produce a set of non-redundant transcripts to feed into MAKER. These transcripts were treated as expressed sequence tag evidence and protein sequences from Arabidopsis 17 , Selaginella moellendorffii 10 , and UniprotKB plant databases 55 were used as protein homology evidence. Intact S. lepidophylla LTR retrotransposons were input as a custom repeat database along with default repeat libraries in MAKER. Ab initio gene prediction was done using the gene predictors SNAP and Augustus with reiterative training in S. lepidophylla. The initial MAKER Max gene set was filtered for gene models with Pfam domain evidence and annotation edit distance <1.0. Putative transposon-derived models were removed. This produced a set of 39,051 filtered gene models. A further 11,847 duplicated gene models corresponding to separate haplotypes were removed to produce a final set of 27,204 non-redundant genes.
Gene expression analysis. Adapter sequences and low-quality bases were filtered from the raw, paired end Illumina RNAseq reads using TRIMMOMATIC (v0.33) 61 and reads shorter than 36 bp were discarded. Expression levels (in Transcript Per Million, TPM) were quantified using Kallisto 62 with the 27,204 monoploid S. lepidophylla gene models provided as input. An average TPM for each time-point was calculated from the mean TPM of the three biological replicates. We preformed weighted gene co-expression analysis of the dehydration and subsequent rehydration time course RNA-seq data with the R package WGCNA 63 . The expression data were pre-filtered to remove genes with excessive missing values or invariant expression using the built-in quality control function. A signed coexpression network was constructed using a soft-thresholding power of 10 and default parameters. The only exception was the mergeCutHeight parameter, controlling the minimum distance between co-expression clusters, which was set to 0.15. The co-expression network resulted in 38 clusters, with 13 genes not placed in any cluster.
MethylC-seq analysis. Quality filtered MethylC-seq reads were filtered as described above. Three replicates for 0 h, 24 h, and 120 h post rehydration and unconverted control DNA were sequenced. The average genome coverage of each library is 12.3×. Reads were mapped to the in silico converted, unmasked S. lepidophylla genome for the forward and reverse strands using Bismark (v0.12.0) 64 . Genome-wide cytosine methylation was obtained using the following parameters in the Bismark methylation extractor step: -comprehensive -CX_context --cytosi-ne_report. The methylation level of each CHH, CHG, and CpG position was calculated as the number of methylated count/total of methylated and unmethylated count. Any position with coverage of less than three reads was removed. For each cytosine position, the average methylation level was calculated from the mean methylation level among biological replicates.
Methylation level of the upstream (1 kb), downstream (1 kb) and body (from transcription start site to transcription stop site) of all genic regions and intact LTR retrotransposons were calculated using BEDTools (v2.24.0) 65 and in-house scripts.
To study differential methylated regions between 0 h, 24 h, and 120 h post rehydration samples, the whole-genome methylated count of each library was obtained from the Bismark output. Differential methylated regions were analyzed using the R package methylKit (v1.2.0) 66 in 100 bp window and minimum coverage of 10 reads per site.
Data availability. The genome assembly, raw PacBio, Illumina DNAseq, and RNAseq data are available from the National Center for Biotechnology Information (NCBI) Short Read Archive (SRA). The RNAseq reads were deposited to NCBI SRA under BioProject PRJNA420971 (BioSample SAMN08128719-08128734) and MethylCseq data were deposited under BioProject PRJNA421023 (BioSample SAMN08132846-08132853). The genome assembly has been deposited under BioSample SAMN07071123 and BioProject ID PRJNA386571.