All NGS data analyzed in this study were downloaded from GEO (https://www.ncbi.nlm.nih.gov/geo/) and all come from the Arabidopsis thaliana species. Since the SPT6L ChIP-seq dataset that was key to our study came from 10-day-old seedlings, when searching for additional datasets, we aimed to have them as close to this stage as possible and selected the one that best matched with the SPT6L ChIP-seq data. We do sample-to-sample comparisons only in cases when the samples came from the same experiment. Remaining analyses focus on comparing sets of loci within one sample. All details about the analyzed datasets are listed in ( Supplementary Table S1 ; Mi et al., 2008; Havecker et al., 2010; Luo et al., 2013; Johnson et al., 2014; Yelagandula et al., 2014; Zhang et al., 2016; Yang et al., 2017; Liu et al., 2018; Ma et al., 2018; Chen et al., 2019; Gallego-Bartolomé et al., 2019; Shu et al., 2019; He et al., 2021).

Downloaded raw reads ( Supplementary Table S1 ) were aligned to the TAIR10 reference genome using bowtie2 (v2.4.5; Langmead and Salzberg, 2012; https://github.com/BenLangmead/bowtie2; with parameters: -L 22 -i ‘S,1,1.15’ –local). Peaks were called using macs2 callpeak (v2.2.7.1; Zhang et al., 2008; https://github.com/macs3-project/MACS; with parameters: –gsize ‘135000000’ –broad –broad-cutoff=‘0.01’ –keep-dup ‘auto’ –qvalue ‘0.05’; with additional parameters: –nomodel –extsize #size –shift ‘0’ for single-end reads; the #size value was determined by macs2 predictd). In the case of more replicates, peaks were first called on pooled replicates to create a set of reference peaks and then on individual replicates. Peaks from pooled replicates were then retained only when present also in all individual replicates to keep only reproducible peaks (Stefflova et al., 2013). The final set of significant peaks was filtered for those with two-fold enrichment or higher. Alignment files were filtered by samtools (v1.15.1; Danecek et al., 2021; https://github.com/samtools/samtools) to remove duplicate reads and coverage tracks were generated using deepTools bamCoverage (v3.5.1; Ramírez et al., 2016; https://github.com/deeptools/deepTools; with parameters: -bs 20 –effectiveGenomeSize 135000000 –normalizeUsing RPGC –extendReads #size).

To obtain coverage over a set of loci, deepTools computeMatrix was used (with parameters: -m 500 -b 1000 -a 1000 -bs 20 –missingDataAsZero). The resulting data were then used to create metaplots (showing average and standard error) and heatmaps using deepTools, or the coverage over each locus was averaged and used to generate boxplots.

ATAC-seq data were downloaded as coverage tracks (Zhong et al., 2021) and onward processed the same way as ChIP-seq data.

Lists of DMRs were retrieved from Zhang et al., 2018b (260 datasets in total). To make the DMR data comparable with ChIP-seq peaks, the 100 bp long DMRs were merged to a single region if they were within 1 kbp from each other (such DMR regions overlap the ChIP-seq peaks in approximately 1:1 ratio). Bedtools intersect (v2.30.0; Quinlan and Hall, 2010; https://github.com/arq5x/bedtools2) was used to identify the DMRs/peaks with overlap. After the DMRs/peaks were counted for all the categories, the enrichment was calculated using the following formula: enrichment = log₂(observed/expected); where expected = (total # of all P5 peaks with DMR overlap)*(P5 peaks of interest/all P5 peaks). The statistical significance of overlap was evaluated using hypergeometric distribution and the resulting p-value was corrected for multiple testing using Benjamini-Hochberg procedure. Samples (datasets) were sorted into categories based on available literature (mostly original studies which produced the data). The categories are not mutually exclusive and were devised based on relevance to our study and samples available in the dataset list. All the samples and their categories are listed in Supplementary Table S3 .

The methylation tracks were retrieved from GEO (GSE39901; Stroud et al., 2013). Absolute values of methylation levels were calculated (ignoring the strandedness of methylation data). The procedure to obtain methylation level over a set of regions was similar as described in ChIP-seq analyses (with computeMatrix parameters: -m 500 -b 1000 -a 1000 -bs 20).

Downloaded raw reads ( Supplementary Table S1 ) were trimmed with gsat (v0.1; https://github.com/MikeAxtell/gsat) using default settings. The read alignment, sRNA cluster identification and read counting was done by ShortStack (v3.8.5; Axtell, 2013; https://github.com/MikeAxtell/ShortStack; with parameters: –genomefile TAIR10.fasta –dicermin 20 –dicermax 25 –nohp for discovering new loci and with an additional parameter: –locifile loci_of_interest.bed for counting reads over analyzed loci). The resulting readcounts were normalized to the dataset size as reads per million (RPM).

To analyze loci where SPT6L and sRNAs overlap, bedtools intersect (default settings) was used to identify the intersection between the sRNA loci and SPT6L loci (i.e. SPT6L∩sRNA loci are parts of the original loci with exact overlap; Supplementary Figure S1A ). With this definition of loci, the analysis was done only at the parts of SPT6L loci that are directly targeted by sRNAs, the non-overlapping parts of SPT6L (sRNA) loci were excluded because there, we could not expect that the sRNAs-SPT6L interaction has any effect. These loci were then randomly shuffled with bedtools shuffle to provide loci for plotting background.

To classify Pol V loci based on the overlap with either SPT6L or Pol II, bedtools intersect (parameters -wa or -v) was used (i.e. in this case we looked at whole Pol V loci, not just their overlapping parts; Supplementary Figure S2B ). This approach was used because the goal was to compare different categories of the Pol V loci.

Pol V loci were annotated with associated genomic features (genes, pomoters, TEs etc.) using uropa (v3.3.4; Kondili et al., 2017; https://github.com/loosolab/UROPA; with parameters: “feature.anchor”:[“start”, “center”, “end”], “distance”:[0, 0], “internals”:”0.5”). The loci were annotated with the features independently, so each Pol V locus could be classified with multiple features. Promoter was defined as region 1 kbp upstream of TSS of the gene. Terminator was defined as 3’UTR plus 500 bp downstream. Araport11 was used for gene annotations and TE Transcripts v1.0 (Panda et al., 2016; https://github.com/KaushikPanda1/AthalianaTETranscripts) was used to annotate TEs. This TE annotation was also a source for all presented TE characteristics (length, distance to nearest gene, type of silencing etc.). The enrichment of certain features at the loci of interest was calculated using formula: enrichment = log₂(observed/expected); where expected = (total # of all P5 peaks with genomic feature overlap)*(P5 peaks of interest/all P5 peaks). The statistical significance of enrichment was evaluated using hypergeometric distribution.

For statistical comparison of quantitative characteristics (e.g. ChIP-seq enrichment, loci length…) between two sets of loci Mann–Whitney U test was used.

Protein sequences were retrieved from UniProt (https://www.uniprot.org/), chain IDs are listed in Supplementary Table S2 . The protein complex predictions were done with AlphaFold-Multimer (v2.2; Evans et al., 2022; https://github.com/deepmind/alphafold; the model was run in full_dbs mode with default parameters). AlphaFold-Multimer produces ipTM score, for which in range between 0.7 and 0.8 over 60% of complexes exceed DocQ 0.49 (medium quality) and over 80% of complexes exceed DockQ 0.23 (acceptable quality; Basu and Wallner, 2016; Evans et al., 2022). The protein structures were visualized by ChimeraX (v1.5; Pettersen et al., 2021; https://www.rbvi.ucsf.edu/chimerax/).

The presence of conserved SH2-binding motifs in the protein sequences in question was checked by blastp (https://blast.ncbi.nlm.nih.gov/Blast.cgi).

Published sequences were retrieved from NCBI (https://www.ncbi.nlm.nih.gov/), Phytozome v12.0, (https://phytozome.jgi.doe.gov), and Phycocosm (https://phycocosm.jgi.doe.gov).

Arabidopsis thaliana protein sequences of AT1G65440 (SPT6L), AT5G04290 (SPT5L), AT1G05460 (SDE3), AT2G16485 (NERD), AT4G35800 (NRPB1 - Pol II), AT2G40030 (NRPE1 - Pol V), AT1G63020 (NRPD1 - Pol IV), AT5G49160 (MET1), AT5G14620 (DRM2), and AT1G69770 (CMT3) were used as queries for searches against genomes spanning all Viridiplantae with main focus on basal streptophyte algae. As DNMT3 is lost in angiosperms the protein sequence of DNMT3 from Ceratopteris richardii was identified and used instead. Manual checks against protein annotation were made where needed to identify correct proteins.

When needed FGENESH+ (Softberry Inc., New York, NY USA; www.softberry.com; Solovyev et al., 2006) was used to predict gene coding sequences from genomic sequences using the A. thaliana protein model when specific annotations were unavailable.

After obtaining sequences of desired paralogs, their AGO-hook was identified by using Wsearch (v1.0; Zielezinski and Karlowski, 2015; http://150.254.123.165/whub/) using a plant matrix. For a protein to be annotated as AGO-hook carrying we set the threshold for motifs score ≥ 6 and the number of such motifs must be ≥ 3 (Sheu-Gruttadauria and MacRae, 2018).

For marginal cases manual check was made to see if the composed domain is biologically relevant and for annotation of Klebsormidiophyceae SPT6L see Supplementary Figures S4H, I .

In case of detection of more transcript isoforms, only the longest variant was chosen and each candidate gene was also cross-checked in an annotation file of each species if available. Orthology was assessed through reciprocal BLAST searches and PFAM protein domain prediction. The multiple sequence alignment of protein sequences identified in the previous step was done by MAFFT v7.455 using the parameters E-INS-I or L-INS-I in Geneious Prime 2023.1.2 (https://www.geneious.com). We manually checked the alignments to correct the misalignment of homologous regions.

Maximum-likelihood phylogenetic trees were constructed by RAxML-HPC (Stamatakis, 2014) tools version 8.2.12 using a General Time Reversible (GTR) model with a gamma distribution of rate heterogeneity, and 100 bootstrap replicas were used to obtain support values.

To visualize the acquired trees FigTree v1.4.4 (http://tree.bio.ed.ac.uk/software/figtree/) was used.

In Arabidopsis thaliana genome 22 genes are predicted to contain the AGO-hook domain (Karlowski et al., 2010), yet only a few of them have been assigned to some sRNA-related pathway so far; NRPE1, SPT5L, NERD, SDE3 and SUO; Onodera et al., 2005; Bies-Etheve et al., 2009; Garcia et al., 2012; Pontier et al., 2012; Yang et al., 2012). Among these predicted genes with AGO-hook domain of unknown function is SPT6L (AT1G65440), which until now was known only as a transcription elongation factor of Pol II . Because SPT6L binds to chromatin and all other chromatin binding AGO-hook proteins discovered so far are involved in RdDM, we hypothesized that it could be involved in this pathway as well. To address this question, we made use of published ChIP-seq data on SPT6L in A. thaliana (Chen et al., 2019) and combined them with sRNA-seq data (Ma et al., 2018) to identify genomic loci where SPT6L and sRNAs overlap (SPT6L∩sRNA; Figure 1A ; Supplementary Figure S1A ). Thereafter, we analyzed DNA methylation at the overlapping parts of these loci ( Figure 1B ). Indeed, we saw there high cytosine methylation in all three sequence contexts as would be expected for the RdDM pathway, which contrasted with very low methylation level at SPT6L loci not overlapping with sRNAs. We looked also at other chromatin marks at these loci, but unlike DNA methylation, most of these had more or less intermediate levels at the overlapping loci compared to loci where SPT6L and sRNAs do not overlap ( Supplementary Figures S1B–D ). To find the reason for increased DNA methylation at the overlapping loci, we checked available ChIP-seq data for Pol II (the known interactor of SPT6L) and key RdDM components (Pol IV, Pol V and AGO4). All the RdDM proteins are strongly enriched at the overlapping loci ( Figure 1C ; Supplementary Figures S1E, F ), which brought up the possibility that besides Pol II, SPT6L also functions with Pol V (already known to be targeted by sRNAs). Indeed a cluster analysis of the SPT6L loci for SPT6L, Pol II and Pol V ChIP-seq data shows a clear formation of two clusters: in the first one, SPT6L colocalizes with Pol V and in the second with Pol II. Moreover, there is only a minor overlap between the Pol II and Pol V loci ( Figure 1D ). Since Pol IV is also present at most of Pol V loci, both Pol IV and Pol V were possible candidates for the colocalization with SPT6L. To differentiate between these two possibilities, we looked at the Pol V loci, which are not enriched for Pol IV and vice versa. SPT6L was strongly enriched only at loci where Pol V was present but not at loci where Pol IV was without Pol V ( Supplementary Figures S1G, H ). This together strongly suggests that SPT6L could be a component of the Pol V complex.

The interaction between SPT6/SPT6L (called SPT6(L) onwards) and Pol II is facilitated by the SH2 domain of SPT6(L) that binds to the phosphorylated serine, threonine and tyrosine residues at the linker to the C-terminal domain (CTD) of NRPB1, the largest subunit of the Pol II complex (Sdano et al., 2017; Vos et al., 2018; Chen et al., 2019). We checked for conserved SH2-binding motifs (residues) at the paralogous Pol V subunit NRPE1 but we did not find any. However, we were also unable to find these motifs in the largest subunit of yeast Pol I (data not shown), which has also been indicated to bind SPT6 (Engel et al., 2015). To further address the means of possible SPT6L binding to Pol V, we reanalyzed ChIP-seq data for A. thaliana treated with flavopiridol (an inhibitor of Pol II kinase P-TEFb, responsible for phosphorylation of the NRPB1 CTD linker; Vos et al., 2018; Chen et al., 2019) and an A. thaliana mutant lacking the SH2 domain of SPT6L (SPT6LΔSH2; Chen et al., 2019). As expected, we could see a decrease in SPT6L binding to loci occupied by Pol II when the samples were treated with the kinase inhibitor. In contrast, at loci occupied by Pol V, the treatment caused a slight increase in SPT6L binding ( Figure 1E ). Surprisingly, the missing SH2 domain reduced the binding of SPT6L to both Pol V and Pol II occupied loci ( Figure 1E ). This shows that the SH2 domain affects SPT6L binding even to loci where only Pol V is present but this binding is not dependent on the phosphorylation machinery that normally deposits phosphate groups recognized by the SH2 domain.

Although NRPB1 CTD linker is important for the binding of SPT6(L) to Pol II, a much larger interacting interface between SPT6(L) and Pol II is at NRPB4/7 subunits and also at its elongation factor SPT5 ( Figure 1F ; Vos et al., 2018; Chen et al., 2019). To assess whether SPT6L may bind these interfaces also at the Pol V complex we used the AlphaFold-Multimer predictions (Evans et al., 2022). Besides predicting the secondary structure of proteins or complexes, AlphaFold also calculates the accuracy of such structures. We used this score to estimate the likeliness of SPT6L interaction with Pol V subunits. Because assembling the whole polymerase complex using AlphaFold would be extremely demanding on computational resources due to its size, we resorted to testing pairwise interactions. We first ran the prediction for SPT6L and selected interacting subunits of the Pol II complex (NRPB1/4/6/7/SPT5) to see how AlphaFold will perform ( Supplementary Table S2 ). Of these, NRPB7 showed the best-supported interaction ( Figure 1F ). Therefore, we ran further predictions for all NRPB7 paralogues ( Supplementary Table S2 ). In this analysis NRPE7 reached the highest score, very similar to that of NRPB7 ( Figure 1F ; Supplementary Figure S1I ). The fact that NRPE7 probably retained the ability to interact with SPT6L throughout the evolution further supports our hypothesis that SPT6L is a component of the Pol V complex.

We also tried to search for any co-immunoprecipitation experiments (co-IP LC-MS/MS) that would indicate Pol V-SPT6L interaction. Unfortunately, from many co-IP studies involving Pol V only one presented the full dataset (Li et al., 2018). However, this study failed to detect SPT6L even in the Pol II complex, indicating that probably their co-IP conditions were not favorable to retain SPT6L in the immunoprecipitated complex.

SPT6L may participate in the Pol V complex as a transcription elongation factor with histone chaperone activity and besides that, it brings a third AGO-hook to the complex in addition to those already present on NRPE1 and SPT5L. SPT6L is significantly enriched only at some Pol V loci ( Figure 2A ), so determining the characteristic features of these loci would help to better understand the role of SPT6L in the RdDM pathway. For this purpose, we have defined Pol V loci as peaks from ChIP-seq data (Liu et al., 2018) and classified them based on their overlap with similarly defined SPT6L and also Pol II loci (Chen et al., 2019). In total, we have identified 8574 Pol V loci, of these 3472 loci (40%) are overlapped by SPT6L loci. We call these Pol V loci P5(+)SPT6L, in contrast to Pol V loci without SPT6L, which we call P5(-)SPT6L. Within the P5(+)SPT6L loci, 997 loci (29% ) also overlap with Pol II loci (these loci represent 11.6% of all Pol V loci; Figure 2A ; Supplementary Figures S2A, B ). At these loci, SPT6L can also be present as a component of the Pol II complex, so to filter this out in all the subsequent analyses, we will also separately provide results for P5(+)SPT6L loci without the overlap with Pol II, i.e. P5(+)SPT6L(-)P2 loci as Supplementary Data .

The Pol V loci with and without SPT6L were compared in four different aspects; i) chromatin state (components and modifications), ii) the mode of DNA methylation (proteins involved in its establishment/maintenance); iii) small RNAs mapping to the loci, and iv) functional classification of genomic elements overlapping with the loci.

Protein SPT6 has two paralogues in A. thaliana: SPT6L (AT1G65440) with the AGO-hook domain and SPT6 (AT1G63210) without this domain. The first one is likely part of the Pol V complex as shown in the previous sections and also of the Pol II complex (Chen et al., 2019), while the second paralogue is not very well described but it might be a part of the Pol II complex (Antosz et al., 2017). It somewhat resembles the situation with a pair of paralogues of SPT5(L), another RNA polymerase elongation factor (non-homologous to SPT6L). SPT5 without the AGO-hook functions in the Pol II complex (a role conserved to all eukaryotes), while SPT5L with the AGO-hook acts in the Pol V complex (a function specific to plants; Bies-Etheve et al., 2009). To gain a better understanding of the evolution of SPT6 paralogues, we compared detailed phylogenetic trees for SPT5(L) and SPT6(L) ( Figure 4 ; Supplementary Figures S4A, B ).

These trees indicate that SPT5 diverged in angiosperms ( Supplementary Figure S4B ; Matzke et al., 2015) into two separate clades characterized by the presence/absence of the AGO-hook, with both paralogues detected in all analyzed angiosperm species. On the contrary, there is no such deep dichotomy in the phylogenetic tree of SPT6(L). The only conserved paralogue in plants is SPT6L with the AGO-hook domain ( Supplementary Figure S4A ). Occasional occurrences of SPT6 without the AGO-hook reflect recent parallel duplications (in some Brassicaceae and a few other clades), likely followed by a secondary loss of the AGO-hook domain in one of the paralogues. SPT6L with the AGO-hook is missing only in Carica papaya, where surprisingly, also NERD, another AGO-hook protein related to the RdDM pathway, is present without the AGO-hook domain, and even SPT5L has its AGO-hook significantly truncated (only 13 motifs compared to usual 50+ motifs in other species; Supplementary Figures S4A–C ).

As already mentioned, SPT6L is just one of several AGO-hook-carrying proteins possibly affecting the RdDM pathway (NRPE1, SPT5L, NERD, SDE3). To better understand the RdDM evolution in the context of its newly identified component (SPT6L), we searched for the first emergence of all these proteins in the Viridiplantae clade ( Figure 4 ).

Protein SDE3 obtained the AGO-hook most recently and possesses this domain only in some angiosperm families, specifically in Asteraceae and Brassicaceae and its function is only partially connected with DNA methylation (Garcia et al., 2012; Supplementary Figure S4D ).

The previously mentioned SPT5L likely originated from SPT5 duplication in the ancestor of angiosperms ( Supplementary Figure S4B ). Protein NERD is older and likely emerged in the common ancestor of Zygnematophyceae (detected in Mesotaenium), mosses and vascular plants. Interestingly, in gymnosperms that do not have SPT5L, NERDs are characterized by longer AGO-hook domains compared with other clades ( Supplementary Figures 4B, C ).

The most ancient AGO-hook domain, which has been connected with RdDM, belongs to the largest subunit of Pol V, NRPE1 (more exactly the common ancestor of NRPD1/E1 of Pol IV and Pol V). It arose by duplication of the gene encoding the largest subunit of Pol II (NRPB1; Huang et al., 2015) and can first be found in the basal orders of streptophyte algae Klebsormidiophyceae and Charophyceae ( Supplementary Figure S4E ; for details see Trujillo, 2019). The emergence of NRP(D)E1 in these algae is accompanied by the emergence of DRM methyltransferase, which is considered necessary for the RdDM pathway ( Supplementary Figure S4F ). Although the duplication event giving rise to NRP(D)E1 likely happened shortly after the divergence of streptophyte algae from chlorophyte algae, the most basal clades of streptophyte plants represented by Mesostigma and Chlorokybus do not possess the NRP(D)E1 gene. Surprisingly, our phylogenetic analysis detected SPT6L with the AGO-hook motifs already in these two algae ( Supplementary Figure S4A ) in which the emergence of SPT6L coincides with the appearance of chromomethylases ( Supplementary Figure S4G ), so the acquisition of the SPT6L’s AGO hook into the Pol II complex preceded the emergence of Pol V. In chlorophyte algae, only the canonical SPT6 was detected, making SPT6L a synapomorphy of Streptophyta. In summary, the four AGO-hook binding proteins have appeared successively at different stages of plant evolution and were often accompanied with gains or loss of some cytosine methyltransferase genes ( Figure 4 ). SPT6L appeared as the most ancient component of any RNA polymerase complex that acquired the ability to bind AGOs in Viridiplantae.

The absence of SPT6 in most plant species indicates that the same SPT6L with the AGO-hook may be present in both Pol II and Pol V complexes. The fact that SPT6L did not diverge in evolution into specialized paralogues that would support these complexes separately suggests that the function of SPT6L in both complexes may be interconnected. SPT6L has been shown to have a role in chromatin opening, which is required for the initiation of Pol II transcription (Shu et al., 2022). The same may be true for Pol V at some of its loci. Pol V is predominantly attracted to its loci by SUVH2/9 that recognize methylated DNA (Johnson et al., 2014), likely primarily “hemimethylated” DNA after replication. In the case of unmethylated naïve loci, Sigman et al., 2021 suggested that Pol V could be guided by AGO4/6/9, which interact with Pol II transcripts. SPT6L, with its AGO-binding domain, histone-binding domain, and presumed ability to interact with both polymerases, seems like a good candidate to facilitate this process. Because Sigman et al., 2021 suggested that this AGO-dependent targeting should occur at loci where Pol V targeting is independent (or only partially dependent) on SUVH2/9, we took a closer look at these loci ( Figure 5A ). Here we could observe much higher levels of Pol II, but only small differences in SPT6L binding when compared to loci fully dependent on SUVH2/9. SPT6L present at such loci is mostly not a part of the Pol II elongation complex, as the presence of SPT6L there is not reduced by the flavopiridol treatment. As expected, suvh2/9 mutation causes depletion of NRPE1 (Pol V) from both types of loci, but the effect is much stronger in the case of SUVH2/9-dependent loci. Deletion of the SPT6L’s SH2 domain (SPT6LΔSH2) strongly reduced its binding at loci dependent on SUVH2/9. This was expected based on the previous analysis which showed that the SH2 domain is important for SPT6L binding to Pol V loci ( Figure 1E ). Surprisingly, SPT6L level was almost unaffected at loci where Pol V binding can be independent of SUVH2/9. It is unclear how SPT6L can bind to these loci and if its binding there involves Pol II (even in absence of the SH2 domain). But since this binding is likely independent of Pol V, it opens the possibility to switch the roles so that SPT6L guides Pol V to these loci ( Figure 5B ).

Our phylogenetic data indicate that SPT6L was the first component of any RNA Polymerase complex in the Viridiplantae clade that acquired an AGO-hook domain. We found SPT6L with this domain already in the most basal streptophyte algae Chlorokybus and Mesostigma. Mesostigma viride genome contains only about 1% of CHH methylation (Liang et al., 2020), which indicates no (or only negligible) RdDM activity. This corresponds to the fact that these algae contain neither an orthologue of DRM nor NRPD1/E1 (the Pol V subunit), which is considered essential for RdDM. Both these genes first appeared in Klebsormidiophyceae and Charophyceae (Luo and Hall, 2007; Trujillo, 2019; Yaari et al., 2019, this study). This indicates that the AGO-hook domain of SPT6L in the two most basal groups of streptophyte algae (Chlorokybophyceae and Mesostigmatophyceae) was unlikely engaged in some ancestral form of RdDM. However, it might be a prerequisite for evolution of the pathway. As the Pol V evolved, it likely contained two AGO-hook components (SPT6L and NRPD/E1) from the very beginning ( Figure 6 ).

Functionally specialized Pol IV and Pol V, as we know them from higher plants, likely evolved in the most recent common ancestor of land plants (Wang and Ma, 2015). Although SPT6L was probably well established at this evolutionary stage, our analysis of available ChIP data shows that SPT6L strongly colocalizes only with Pol V but not with Pol IV. It can be related to differences in the processivity of these polymerases. Pol V produces longer transcripts that can span one or more nucleosomes (Blevins et al., 2015; Böhmdorfer et al., 2016). In agreement with this, Pol V loci with SPT6L are generally longer than those without. Although the AGO hook of SPT6L is likely to have some function in the Pol V complex, the mere presence of SPT6L in this complex may also be required for its histone chaperone activity (regardless of the AGO hook).

Identification of another AGO-hook component of the Pol V complex raises the question of why the Pol V complex does need so many. When considering that two or three tryptophans are sufficient to bind one AGO protein (Sheu-Gruttadauria and MacRae, 2018) and tens of such motifs are present in the typical AGO-hook domain of NRPE1, SPT5L and also SPT6L ( Supplementary Figures S4A, B, E ), the Pol V complex with all these components may have the capacity to bind more than ten AGOs at once. Small RNAs that drive methylation are generated from a huge number of genomic loci, so their sequence variability can be extremely high. The high number of AGO-binding sites may be necessary to maximize the likelihood that a nascent transcript of Pol V is recognized by an AGO that is just carrying a complementary sRNA within the short time window of ongoing transcription. In our previous study, we observed that the likelihood of de novo methylation of a naïve locus depended on the level of the respective sRNAs (Čermák et al., 2020) which indicates that this interaction can indeed be the limiting step of the whole process. In conclusion, we suggest that increasing the number of AGO-hook domains in plant evolution may be needed to maintain sufficient sensitivity of the RdDM pathway in the face of increasing sRNA and TE complexity.

However, AGO-hook of SPT6L does not necessarily function in attracting AGOs to the Pol V complex. Lahmy et al., 2016 showed that AGO-hooks of NRPE1 and SPT5L are fully redundant and their deletion leads to loss of RdDM dependent DNA methylation. This would either mean that the AGO-hook of SPT6L is not sufficient to attract AGOs on its own or it has other functions. It could function in attracting other proteins than AGOs, like it has been shown for GW182 in attracting CCR4-NOT complex (Chekulaeva et al., 2011). AGO-hook domains are intrinsically disordered, so they likely tend to participate in (or even directly induce) the formation of biomolecular condensates in the nucleus via the liquid-liquid phase separation (LLPS). The overall protein composition of these condensates can be determined by specific combinations of participating intrinsically disordered domains (Laflamme and Mekhail, 2020). Thus various combinations of the three (or even more) AGO-hooks associated with the Pol V complex (NRPE1, SPT5L, SPT6L) might participate in defining specific LLPS condensates that attract/contain a specific wider set of chromatin-binding and chromatin-modifying proteins which directly and indirectly affect the epigenetic landscapes of individual loci.

A. thaliana is one of the few exceptional plant species that contains not only SPT6L but also an additional SPT6 paralogue without the AGO-hook domain. Although SPT6 was detected in the Pol II complex (Antosz et al., 2017), nothing is known about its function. The presence of both SPT6 and SPT6L in this plant model was likely the reason for overlooking the fact that the only conserved homolog in Viridiplantae is SPT6L that carries the AGO-hook with one exception found, which is Carica papaya. This contrasts with SPT5 and SPT5L. These paralogues are regarded to be specialized for separate roles in the Pol II and Pol V complexes, respectively (Bies-Etheve et al., 2009). Different evolutionary trajectories can result from the fact that SPT5L, unlike SPT6L, evolved in the ancestor of vascular plants with already well-defined RdDM machinery. However, it is also possible that there was some evolutionary constraint on the subfunctionalization of SPT6L. Such a reason could be if one protein molecule had to physically interact first with one and then with the other polymerase complex. Chen et al., 2019 documented that SPT6L without SH2 domain (necessary for binding to both Pol II and Pol V loci; Figure 1E ) is still able to bind in gene promoters which indicates its ability to bind to chromatin independently of the Pol II complex. In the follow-up work, SPT6L was shown to participate in attracting chromatin remodeling complexes that open chromatin for initiation of Pol II transcription (Shu et al., 2022). Based on these findings, we speculate that SPT6L might play a similar role for the Pol V complex in de novo DNA methylation of naïve loci, where Pol V cannot be tethered by SUVH2/9 that do this at already methylated loci (Johnson et al., 2014). Recently published data by Sigman et al., 2021 indicated that Pol II can recruit AGO4 (and possibly other AGOs) to naïve loci that are targeted by sRNAs and that it might participate in Pol V attraction to these loci. We hypothesize that AGO4 is present at these loci together with SPT6L. They both may be deposited there together by Pol II whose nascent transcripts was just recognised by sRNA loaded on AGO4 attached to SPT6L in the Pol II complex. Thereafter, SPT6L can stay bound to histones in the locus and mediate chromatin re-opening also for Pol V transcription ( Figure 5B ).