Enhanced RNAi stability through imperfect inverted repeats: nucleotide mismatches prevent intrinsic self-silencing of hpRNA transgenes in plants

Hairpin RNA (hpRNA) transgenes, with a perfect inverted-repeat (IR) DNA, have been the most successful RNA interference (RNAi) method in plants. Here we show that hpRNA transgenes were invariably methylated in the IR DNA and the adjacent promoter, causing transcriptional self-silencing and preventing the full potential of RNAi. Nucleotide substitutions in the sense sequence, which disrupts the perfect IR DNA structure, were sucient to prevent the intrinsic DNA methylation resulting in more uniform and persistent RNAi. Substituting all cytosine (C) with thymine (T) nucleotides, in a G:U hpRNA design, prevented DNA methylation and self-silencing but still allowed for the formation of perfect hpRNA due to G:U wobble base-pairing. The G:U design induces effective RNAi in 90–96% of transgenic lines, compared to 57–65% for the traditional hpRNA design. Furthermore, while a traditional hpRNA transgene showed increasing DNA methylation and self-silencing from cotyledons to true leaves, the G:U transgenes avoided this developmental progression of self-silencing and induced RNAi throughout plant growth. The G:U and traditional hpRNA transgenes generated small interfering RNA (siRNA) with different 5’ phosphorylation, which resembled the endogenous tasiRNA and miRNA, respectively. Furthermore, our results suggest that siRNAs from the two transgene designs function differently to induce target DNA methylation, one (from traditional hpRNA) through the canonical RdDM pathway and the other (G:U hpRNA) a non-canonical pathway. Our study not only revealed a methylation-resistant RNAi transgene design but also provided new mechanistic insights into small RNA biogenesis and function in plants.


Introduction
RNA silencing is an evolutionarily conserved gene silencing mechanism in eukaryotes, where long dsRNA is processed by Dicer or Dicer-like (DCL) proteins into 20-30 nucleotide (nt) small RNA (sRNA) that induces RNA degradation via sequence complementarity [1][2][3] . In plants, multiple RNA silencing pathways exist, including microRNA (miRNA), trans-acting small interfering RNA (tasiRNA), repeat-associated siRNA (rasiRNA) and exogenic (virus and transgene) siRNA (exosiRNA) pathways 4 . miRNAs are 20-24 nt sRNAs processed in the nucleus by DCL1 from short self-folding RNAs transcribed from MIR genes 2 . tasiRNAs are 21 nt secondary siRNAs derived from DCL4 processing of dsRNA synthesized by RNA-dependent RNA polymerase 6 (RDR6) from miRNA-cleaved TAS RNA fragment 4 . The 24-nt rasiRNAs are generated from repetitive DNA in the genome by the combined function of DNA-dependent RNA polymerase IV (Pol IV), RDR2 and DCL3 5 . The exosiRNA pathway overlaps with the tasiRNA and rasiRNA pathways and both DCL4 and DCL3 are involved in exosiRNA processing. In addition to DCL1, DCL3 and DCL4, plant genomes encode DCL2 or equivalent, which generates 22-nt siRNAs including 22 nt exosiRNAs, and plays a key role in systemic and transitive gene silencing in plants 6 . All of these plant sRNAs are methylated at the 2'-hydroxyl group of the 3' terminal nucleotide by HUA Enhancer 1 (HEN1), which is thought to stabilize the sRNAs 7 . miRNAs, tasiRNAs and exosiRNAs are functionally similar to sRNAs in animals, and involved in post-transcriptional RNA degradation. The rasiRNAs, however, are unique to plants and function to direct de novo cytosine methylation at the cognate DNA, a transcriptional gene silencing mechanism known as RNA-directed DNA methylation (RdDM) 5 . The post-transcriptional RNA silencing mechanism has been extensively exploited as a gene knockdown technology in various eukaryotic systems, generally referred to as gene silencing or RNA interference (RNAi) technologies. In plants, the different RNA silencing pathways have led to different technical approaches, such as arti cial miRNA, arti cial tasiRNA and virus-induced gene silencing technologies 4 . However, long hpRNA transgenes, designed to express long hairpin-structured dsRNA, are the most widely used RNAi technology in plants, and a variety of successful applications of this technology have been demonstrated in plant biotechnology 4 . It can be anticipated that this RNAi approach will continue to be a powerful tool in many areas of crop improvements such as host-induced RNAi against pests and pathogens and metabolic engineering of novel traits through spatial and temporal gene knockdown, which is di cult to achieve using gene knockout technologies such as the CRISPR/Cas9 approach.
An hpRNA construct typically consists of a perfect inverted repeat (IR) of a target gene sequence (forming the dsRNA stem of hpRNA) separated by a spacer sequence (forming the loop). Tandem DNA repeats, particularly the IR DNA structures, are widely observed to attract strong DNA methylation causing transcriptional silencing 8,9 . Beside the IR structure, siRNAs derived from hpRNA transgenes can potentially direct DNA methylation to their own sequence via the RdDM pathway 10,11 . hpRNA transgenes therefore differ from normal transgenes and are potentially subject to self-induced transcriptional silencing. Indeed, a previous study showed that hpRNA transgene-induced RNAi in Arabidopsis was enhanced in an RdDM mutant, and that this enhanced RNAi effect correlated with reduced DNA methylation spanning from the IR DNA to the upstream promoter sequence 12 . An RNAi design that can prevent self-induced silencing would therefore be desirable for achieving durable and potent RNAi in plants.
In this study we investigated the effect of nucleotide mismatches in hpRNA-induced RNAi in plants.
Introducing nucleotide mismatches to disrupt IR DNA structure resulted in uniform and persistent RNAi against a reporter gene and two endogenous genes. We discovered that the traditional hpRNA transgenes with a perfect IR structure are generally prone to self-induced methylation and transcriptional silencing (referred to as self-silencing hereafter) causing large variability in RNAi e cacy, whereas the enhanced RNAi effect of mismatched hpRNA constructs was due to the prevention of methylation in both the IR and the promoters preventing self-silencing. Additionally, we generated evidence that the IR-associated DNA methylation and self-silencing is independent of the RdDM pathway, and that siRNAs from G:U hpRNA transgenes are processed and function differently from traditional hpRNA transgenes, providing novel insights into IR-induced gene silencing and siRNA biogenesis in plants.

Results
Evenly mismatched and G:U basepaired hpRNA constructs induce uniform RNAi We rst tested three mismatched constructs in Nicotiana tabacum using the β-glucuronidase (GUS) reporter gene as the RNAi target ( Figure 1A). These constructs contained the same 200 bp antisense wildtype (WT) GUS sequence as the traditional hpRNA construct (hpGUS[WT]) to ensure perfect sequence complementarity between antisense siRNAs and target GUS mRNA. The mismatched construct hpGUS [1:4] had one nucleotide substitution in every 4 nucleotides of the 200 bp sense sequence; hpGUS[2:10] contained 2 consecutive nucleotide substitutions in every 10 nucleotides; and hpGUS[G:U] had all 52 cytosine (C) nucleotides changed to thymine (T) nucleotides ( Figure 1A, S1). The C to T changes in hpGUS[G:U] disrupted the perfect IR DNA structure but did not prevent the formation of perfect hpRNA due to G:U wobble base-pairing.
The hpGUS[WT] transgenic population showed a wide range of RNAi e ciency, with 35 of the 59 independent lines analysed (59.3%) showing strong RNAi (GUS activity ≤10% of the untransformed GUS plants), 9 showing weak RNAi (GUS activity 10-30% of the untransformed), and 15 almost no silencing (Table 1 and Figure 1B), which was typical for traditional hpRNA constructs 13 [1:4] coincided with its dsRNA stem having the lowest predicted thermodynamic stability ( Figure 1C). Consistently, there appeared to be a good correlation between the extent of GUS RNAi and the predicted dsRNA stability of the four hpRNAs ( Figure  1C).
As the G:U hpRNA construct induced strong and uniform RNAi against GUS, we tested this design against two endogenous genes in Arabidopsis, the ethylene insensitive 2 (EIN2) and phytoene desaturase (PDS) genes, silencing of which can be scored based on hypocotyl length of dark-germinated seedlings on 1aminocyclopropane-1-carboxylic acid medium 14 and photo-bleaching 15 Figure   3B; 7 days). Thus, both constructs induced effective PDS RNAi in cotyledons. However, the two populations showed a clear difference when true leaves emerged (14 days and beyond), with a much larger number of hpPDS[WT] plants giving green leaves that indicated a loss of strong RNAi ( Figure 3B).
As summarised in Figure 3C, the hpPDS[G:U] population contained much higher proportions of the strongly and moderately silenced lines (63% and 30% respectively) than the hpPDS[WT] population (34% and 23%). In addition, most of the weakly silenced hpPDS[G:U] plants still showed mild mottling on true leaves, in contrast to the weakly silenced hpPDS[WT] plants that mostly had fully green leaves.
The GUS, EIN2 and PDS RNAi results collectively con rmed that the G:U hpRNA construct induces more uniform RNAi than the traditional hpRNA construct. Importantly, the PDS RNAi result indicated a developmental stage variability of RNAi by the traditional hpRNA transgene, being more effective in cotyledons than leaves, and suggested that the G:U hpRNA transgenes are developmentally more stable.  [1:4] lines with representative GUS RNAi levels (indicated by asterisks in Figure 4A). Consistent with the McrBC-digestion PCR result, all four hpGUS[G:U] lines had very low levels of DNA methylation at the 35S promoter based on bisul te sequencing ( Figure 4C). The four randomly selected hpGUS [1:4] lines all showed low to moderate levels of DNA methylation ( Figure 4C), with their average promoter methylation levels correlating inversely with the extent of GUS RNAi ( Figure S4). The two hpGUS[WT] lines, despite strong RNAi, both showed moderate levels of DNA methylation in the upstream region of the 35S promoter and high levels (60-100%) of DNA methylation near the 35S-GUS junction ( Figure 4C). Thus, the hpGUS[G:U] transgene, and to a lesser degree the hpGUS[1:4] transgene, had reduced promoter methylation across the transgenic population. Taken together, the methylation analyses indicated that the relatively uniform RNAi of the mismatched hpRNA lines was due to diminished promoter methylation and that the traditional hpRNA transgenes are inherently prone to promoter methylation with all lines having some levels of promoter methylation. The result also suggested that promoter methylation of traditional hpRNA transgenes is developmental stage dependent.
The intrinsic methylation of traditional hpRNA transgenes is not affected in RdDM mutants It was thought that the methylation in the IR region of a traditional hpRNA transgene is induced by hpRNA-derived siRNAs via the RdDM pathway. Consequently, it was expected that the traditional hpRNA transgenes would lose the methylation in a RdDM mutant resulting in uniform RNAi across transgenic populations. It was also expected that the traditional hpRNA transgenes would induce more effective RNAi than the G:U hpRNA transgenes in RdDM mutants due to stronger dsRNA stability. We investigated these using two Arabidopsis RdDM mutants, nrpd1a-3 (a T-DNA insertion mutant of the upstream siRNA biogenesis factor Pol IV) and ocp11 (a dominant-negative mutant of the downstream effector AGO4).
The traditional hpRNA constructs, targeting PDS or EIN2, indeed induced uniform RNAi in the two RdDM mutants, with over 84~100% of transgenic lines showing RNAi ( Table 2). The white cotyledon-to-green leaf-type of PDS RNAi phenotype of the Col-0 background ( Figure 3) also largely disappeared in the RdDM mutants, with most of the hpPDS[WT] plants showing relatively uniform photo-bleaching from cotyledons to leaves ( Figure S6A; Figure S7A). However, to our surprise, the traditional hpRNA transgenes induced weaker RNAi than the G:U transgenes in both mutant backgrounds ( Figure S6; Figure S7A). In particular, the hpPDS[G:U] construct induced extreme photo-bleaching in 100% of the transgenic lines in ocp11, compared with moderate to high levels of photo-bleaching in 95% of the hpPDS[WT] lines ( Figure S6A; Figure S7A). The RNAi of ocp11/hpEIN2[G:U] lines could not be properly assayed using hypocotyl length because many lines showed poor seed germination in the dark on ACC medium (likely due to strong EIN2 RNAi based on our previous observation). Nevertheless Figure S7B). Thus, strong DNA methylation inside the perfect inverted-repeat DNA, as well as its spread to the upstream promoter, was not affected in the RdDM mutants.
The EIN2 and PDS genomic targets also showed DNA methylation but at a much lower level than the IR region of the hpEIN2[WT] and hpPDS [WT] transgenes, particularly at the CHG and CHH sites ( Taken together, experiments with the RdDM mutants con rmed that DNA methylation of traditional hpRNA transgenes is intrinsic to the IR DNA structure and independent of the RdDM pathway. This intrinsic methylation prevented the traditional hpRNA transgenes from reaching their full RNAi e cacy, even in the RdDM mutants. However, the increased cross-line uniformity of PDS and EIN2 RNAi in the RdDM mutants suggested that the RdDM pathway contributes to genomic position or copy numberdependent silencing of hpRNA transgenes. siRNAs from traditional and G:U hpRNA are differently processed One obvious question was whether G:U base-paired hpRNAs were e ciently processed by Dicer into siRNAs. Northern blot analysis detected abundant siRNAs from the hpEIN2[G:U] ( Figure 7A Figure S9B). This result suggested that siRNAs from the traditional and G:U hpRNAs possess different chemical modi cations at the termini.
Dicer-processed sRNAs were assumed to have 5' monophosphate but in C. elegans many siRNAs are found to possess di-or tri-phosphate which increases gel mobility 18 . Alkaline phosphatase treatment homogenised the gel mobility of hpRNA[WT] and hpRNA[G:U]-derived siRNAs ( Figure 7C), indicating that the two siRNA populations have a similar size pro le and that the differential gel mobility of untreated siRNAs was due to different 5' phosphorylation. The siRNA bands of hpEIN2[WT] plants aligned well with the 21 and 24 nt sRNA size markers that were monophosphorylated with radioactive 32 P ( Figure 7A; Figure S10A), suggesting that these siRNAs are largely monophosphorylated. The G:U hpRNA-derived siRNAs, with faster mobility, were therefore likely to possess 5' di-or multi-phosphate. This possibility was further suggested by the under-representation of hpEIN2[G:U]-derived antisense sRNAs compared to hpEIN2[WT]-derived antisense siRNAs in the sRNA-seq data, despite the similar or even stronger northern blot bands of hpEIN2[G:U] siRNAs ( Figure 8, Figure S10B, Figure S11); a standard sRNA-seq protocol involves adaptor ligation with 5' monophosphorylated sRNAs but not di or multi-phosphorylated sRNAs 18,19 . Northern blot hybridisation detected high amounts of long dsRNA species in the hpEIN2[G:U] lines but not in the hpEIN2[WT] plants ( Figure 7B), suggesting that the two types of hpRNA are processed differently, which could account for the differential 5' phosphorylation of the siRNAs. The Arabidopsis microRNA miR168 resembled hpRNA[WT]-derived siRNAs in gel mobility, whereas the trans-acting siRNA tasiR255 were similar to hpRNA[G:U]-derived siRNAs (Figure 7c, Figure S10a), suggesting that plant endogenous sRNAs also possess different 5' phosphorylation.
Deep sequencing of hpGUS [1:4] lines also detected siRNAs but with a much lower abundance ( Figure  S9c

Discussion
In this study we showed that the traditional hpRNA transgenes are invariably methylated at the IR DNA structure and the adjacent promoter sequences compromising RNAi e ciency. This widespread intrinsic DNA methylation and self-silencing of hpRNA transgenes were not reported before but is nevertheless unsurprising. IR DNA structures have long been reported to attract DNA methylation that can extend short distance to upstream promoters in plants, and the methylated IR locus can induce homology-dependent trans-methylation of single-copy loci in the genome [20][21][22] . The best studied IR DNA is the naturally occurring PAI1-PAI4 locus in Arabidopsis ecotype Wassilewskija, which always carries dense DNA methylation independently of its transcriptional activity or RdDM factors 23 . Evidence exists that supports DNA:DNA pairing in IR-induced methylation, but dsRNA and sRNA signals are also suggested to contribute the methylation particularly at the homologous trans-methylated non-IR loci 21,24 . Our results showed that strong DNA methylation in the hpRNA transgenes was largely retained in the RdDM mutants of both the upstream siRNA biogenesis factor Pol IV and the downstream effector AGO4, which seems to support a RdDM-independent DNA:DNA pairing model in IR methylation. However, the increased crossplant uniformity of RNAi in the RdDM mutants by the traditional as well as G:U hpRNA transgenes suggest that hpRNA transgenes, like any type of transgenes, are subject to insertion pattern or positiondependent transcriptional silencing, and that RdDM plays a key role in this type of transgene silencing.
It is interesting to note that RNAi potency was generally reduced in the Pol IV mutant compared to wildtype Col-0 and the AGO4 mutant, as indicated by the uniform but weak photo-bleaching phenotypes of hpPDS lines and the low amount of hpEIN2-derived siRNAs in the nrpd1 background. It has been proposed previously that Pol IV may use either methylated DNA and/or dsRNA as template to generate dsRNA and siRNAs 25 . A more direct evidence for the dsRNA-templated model came from a study showing that RNA virus-derived siRNAs, without RNAi transgenes, are strongly reduced in a Pol IV mutant 26 . The hpEIN2[G:U] plants accumulated high amounts of long dsRNA species, and the bulk of sense siRNAs had the C to U-modi ed sequence, indicating that siRNAs were mostly derived from direct Dicer processing of the primary G:U hpRNA transcript independent of Pol IV. For the hpEIN2[WT] transgenes, however, long dsRNA was almost undetectable and there was a strong reduction in siRNA accumulation in the nrpd1a-3 background ( Figure 7B). This raises the possibility that Pol IV may contribute speci cally to siRNA production from the traditional hpRNA transgenes using the low amounts of the primary perfect dsRNA as template. Interestingly, siRNA bands of hpRNA[WT] looked more scattered on the gel blot than those of hpRNA[G:U] (Figure 7), which implies that hpRNA[WT]-derived siRNAs are a mixture of different biogenesis processes with different 5' phosphorylation hence gel mobility (e.g. direct Dicer processing of primary hpRNA plus Pol IV-mediated ampli cation), unlike the G:U hpRNA-derived siRNAs that are largely derived from the primary hpRNA transcript.
The key nding of this study is that C to T substitutions or around 25% nucleotide modi cations in the sense DNA sequence prevented the intrinsic methylation of the hpRNA transgenes, resulting in uniform RNAi across independent transgenic lines. The C to T substitutions also prevented the cotyledon to true leaf progression of methylation and self-silencing observed for the hpPDS[WT] transgene, a phenomenon that has not been reported before but has important implications in studying developmental stagedependent RNAi and transcriptional gene silencing. Thus, disruption of perfect IR DNA structures is su cient to block IR methylation and self-silencing of hpRNA transgenes. The promoter methylation level in the C-to-T modi ed transgenes (hpGUS[G:U] and hpEIN2[G:U]) was more reduced than in the 1-in-4 mismatched transgene (hpGUS [1:4]), suggesting that depletion of cytosines, the target of DNA methylation, in the sense sequence, further reduces promoter methylation. It is interesting to note that microRNA precursors in plants usually contain mismatches or G:U base pairs in the duplex regions. Considering the results from our study, this structural feature may have evolved to disrupt IR DNA structure preventing transcriptional self-silencing of miRNA genes.
As illustrated by the different GUS RNAi e cacy by the four hpGUS constructs, reduced dsRNA stability due to nucleotide modi cations in the sense strand reduces RNAi e ciency presumably because of ine cient Dicer processing. Weak to moderate RNAi can have speci c applications, particularly when the target genes are required for plant viability. The potential drawback of reduced RNAi, however, is largely overcome with the G:U hpRNA constructs, where the C-to-T changes in the sense sequence disrupt the IR DNA structure but still allow the formation of perfect hpRNA structure due to G:U wobble basepairing. Consequently, all three G:U hpRNA constructs tested induced strong and uniform RNAi. hpRNAs containing multiple G:U base-pairs (up to 17.5%) has been previously shown to induce RNAi in animals and confer virus resistance in plants 27,28 . In our study all cytosines in the sense sequence, constituting 18~26% of the target sequences, were substituted in the G:U hpRNA constructs. Future studies should examine the number of C-to-T substitutions that are required for reducing self-silencing while maximizing RNAi e ciency.
Our study indicated that G:U hpRNA-derived siRNAs have different 5' phosphorylation to those of traditional hpRNA. This was unexpected, as Dicer processed small RNAs in plants were assumed to carry 5' monophosphate and standard sRNA cloning strategies have been based on this 5' chemistry. This nding has important implications in the interpretation of current sRNA-seq data and development of sRNA sequencing protocols, as sRNAs with different 5' phosphorylation requires different cloning methods 18 . Methylation analysis of hpRNA transgenes in the RdDM mutants suggested that the G:U hpRNA-derived siRNAs, unlike those of the traditional hpRNA, induce RdDM through a Pol IV-independent pathway. Thus, G:U hpRNA-derived siRNAs may have distinct functional properties from the traditional hpRNA-derived siRNAs, possibly due to different biogenesis or 5' modi cation. How the traditional and G:U hpRNA-derived siRNAs are differentially phosphorylated is beyond the current study. The similar 5' phosphorylation between the two groups of siRNAs and miR168 (a nuclear processed sRNA) or tasiR255 (a likely cytoplasmic sRNA) raises the possibility for siRNA processing in the nuclei for perfect hpRNA and cytoplasms for G:U hpRNAs.
In conclusion, our study uncovered a new RNAi construct design that overcomes transcriptional selfsilencing to induce more uniform and persistent  Table 1), containing XhoI and BamHI sites or HindIII and KpnI sites, respectively. PCR fragment was inserted into pGEM-T Easy (Promega), the correct nucleotide sequence con rmed by sequencing, and inserted as a BamHI/HindIII fragment into pKannibal 30 forming the 35S-P::PDK intron::antisense GUS::Ocs-T cassette (pMBW606). This plasmid was used as the base vector for assembling the four GUS hpRNA constructs as follows.
hpGUS followed the following rule: C is changed to G, G to C, A to T and T to A. The PCR fragments were ligated into the pGEM-T Easy vector, the correct nucleotide sequences con rmed by sequencing, and then inserted as a XhoI/KpnI fragment into pMBW606. The resulting 35S promoter::hpRNA::OCS terminator cassette was excised with Not1 and inserted into the NotI site of pART27, forming the three mismatched constructs.
Perfect and G:U base-paired EIN2 and PDS hpRNA constructs: DNA fragments spanning the 200 bp regions of the wild-type EIN2 cDNAs were PCR-ampli ed from Arabidopsis thaliana Col-0 cDNA using the oligonucleotide primer pairs EIN2wt-F and EIN2wt-R (Supplementary Table 1) and cloned into pGEM-T Easy. The 200 bp C to T modi ed sense sequence (EIN2[G:U]) was assembled by annealing the overlapping oligonucleotides EIN2-GU-F and EIN2-GU-R (Supplementary Table 1), followed by PCR extension of 3' ends using LongAmp Taq polymerase, and also cloned into pGEM-T Easy and sequenced. DNA fragments of 450 bp wild-type and C-to-T modi ed sequences of PDS cDNA (Supplementary Table   2) were synthesized by GeneArt TM .
The 35S-P::sense fragment::PDK intron::antisense fragment::OCS-T cassettes were prepared in the same way as for the hpGUS constructs. Essentially, the wild-type sequences were excised from the respective pGEM-T Easy plasmids by digestion with HindIII and BamHI, and inserted into pKannibal between the BamHI and HindIII sites so they would be in the antisense orientation relative to the 35S promoter. The wild-type or C to T modi ed fragments were then excised from the respective plasmids using XhoI and KpnI and inserted into the same sites of the respective antisense-containing clone. All of the cassettes in the pKannibal vector were then excised with NotI and inserted into pART27 to form the nal binary vectors for plant transformation.\

Stable transformation and identi cation of transgenic lines
All four GUS hpRNA constructs were transformed into the GUS-expressing tobacco lines PPGH11 and PPGH22 using the Agrobacterium-mediated leaf-disk method 32 . EIN2 and PDS hpRNA constructs were transformed into A. thaliana by the " oral dipping" method 33 . To select for transgenic Arabidopsis lines, mature seeds were sterilized 34 and spread on MS plates containing 50 µg/mL kanamycin plus 150 µg/mL timentin to inhibit Agrobacterium growth. The phenotype of PDS silencing was recorded for the primary (T1) transformants. The surviving T1 lines of PDS hpRNA constructs, and those of EIN2 hpRNA construct, were transferred to soil, self-fertilised and grown to maturity. Seed collected from these plants (T2 seed) was used to establish T2 plants that were used for further gene silencing, DNA methylation, and transgene segregation analyses.
Analysis of GUS and EIN2 silencing GUS activity was quantitatively determined using uorimetric 4-methylumbelliferyl-β-D-glucuronide (MUG) assay 34  DNA and RNA analysis DNA, small RNA and large RNA from all transgenic tobacco lines were prepared following the phenol extraction method as previously described 10  Bisul te conversion and puri cation were performed using the EpiTect Bisul te kit (QIAGEN) following the procedures recommended by the manufacturer. Bisul te PCR was performed as a nested PCR (two PCR reactions). The primers used in the rst and second round PCR was listed in the Supplementary Table 3.
The PCR cycles was the same as described previously 36 . The PCR products from the second PCR were puri ed using an UltraClean DNA Puri cation Kit (MO BIO) following the manufacturer's instructions.
Approximately 50-200ng of puri ed bisul te PCR product was sequenced with BigDye Terminator V3.1 premix (Applied Biosystems) using one of the nested primers. Cytosine methylation levels were determined from the sequencing trace les using the same procedure as described in Le et al. 37 .
Declarations Tables   The plants were classi ed into three groups based on strong (strong photo-bleaching in the whole plant), moderate (pale green or mottled leaves) and weak (fully green or weakly mottled leaves) PDS RNAi.

Figure 4
Page 23/30 a. GUS expression patterns of the independent transgenic lines analysed by McrBC-digestion PCR. The red asterisks indicate the lines analysed by bisul te sequencing in (c Bisul te sequencing of the 35S promoter regions as indicated by the arrows. Bisul te treatment of genomic DNA converts unmethylated cytosine bases to uracil (U) (shown as thymine in PCR product) but methylated cytosines are not affected. PCR ampli cation of bisul te-treated DNA followed by sequencing therefore detects methylated cytosines at a single-nucleotide resolution. PCR primers were designed to speci cally amplify only the 35S promoter sequences of the hpGUS transgenes but not the one driving HPT expression in the target GUS gene ( Figure 1A; 35S').   blot was stripped and re-hybridized with antisense tasiR255 oligonucleotide probe. Note that the APcaused gel mobility shift of miR168 and tasiR255 bands resembles that of traditional hpRNA-derived siRNAs and G:U hpRNA-derived siRNAs, respectively, which is more clearly shown in Figure S10A.