Maternal Pol IV inhibits, whereas paternal Pol IV promotes, interploidy seed abortion

We tested if the molecular data supporting distinct functions for Pol IV in the mother and the father [11] could be supported by genetic analyses. We reanalyzed previously published data [14] to specifically test the effects of the loss of maternal Pol IV versus paternal Pol IV in the context of interploidy, paternal excess crosses (diploid mother pollinated by tetraploid father). Inheritance of a mutant nrpd1 allele from diploid mothers resulted in 4% normal seed in a cross to tetraploid fathers, which was significantly different than 7.1% normal seed observed when WT diploid mothers are crossed to WT tetraploid fathers. Crosses between WT diploid mothers and tetraploid nrpd1 fathers resulted in 64.8% normal seed. Paternal rescue by nrpd1 was diminished when the diploid mother was also mutant, resulting in 37.5% normal seed. Thus, we conclude that maternal NRPD1 promotes interploidy seed viability and paternal NRPD1 represses seed viability (S1 Fig). This is consistent with observations that paternal excess seed viability was promoted by the maternal activity of DCL3 and repressed by the paternal activity of DCL3 (DCL3 functions downstream of NRPD1) [10]. Interploidy crosses are a sensitive genetic assay to detect endosperm phenotypic effects. However, paternal excess endosperm displays widespread transcriptomic changes [10], which make it a poor system to understand the specific role of RNA Pol IV in endosperm development. Therefore, for all subsequent experiments, we examined endosperm molecular phenotypes in the context of balanced crosses (diploid × diploid) where either one or both parents were homozygous mutant for the nrpd1a-4 allele.

Loss of maternal or paternal Pol IV activity impacts sRNAs at distinct sites

To determine the role of Pol IV in sRNA production in the endosperm, we first identified Pol IV–dependent endosperm sRNAs. Previously, we showed that 24-nt sRNAs were the predominant sRNA species in endosperm and exhibited a broader distribution over genes and transposable elements (TEs) than in other tissues [14]. We profiled sRNA populations in 3 replicates of endosperm derived from crosses of Ler nrpd1 females pollinated by Col-0 nrpd1 males (7 days after pollination) and compared them with our previously published sRNA libraries from Ler × Col-0 WT F₁ endosperm (female parent in cross written first) [14] (S1 Table).

We identified 21,131 sRNA peaks in WT endosperm using ShortStack [23]. A total of 76.9% of these were predominantly populated by 24-nt sRNAs, with another 1.1%, 0.2%, and 2.2% of peaks predominated by 23-, 22-, or 21-nt sRNAs, respectively. An additional 19.7% of peaks were either predominated by a noncanonical sRNA size or had no predominant size class (S2A Fig). The majority of sRNAs were genetically dependent on NRPD1, with 99% of 24-nt sRNA peaks, 94.87% of 22-nt sRNA peaks, and 70.1% of 21-nt sRNA peaks absent in nrpd1−/− endosperm (S2B Fig). To enable downstream comparisons to expression, we binned sRNAs by size (21 to 24 nt) and calculated read counts overlapping TEs and genes encoding proteins, miRNAs, and other ncRNAs. We used DESeq2 [24] to separately identify genes and TEs with significant differences in Pol IV–dependent sRNA populations. Consistent with the peak-based analysis, loss of RNA Pol IV abolished 21- to 24-nt sRNAs at most TEs and genes, while most miRNAs were not impacted (S2 and S3 Tables, S3 Fig). Moreover, 21- to 23-nt sRNAs were often lost at the same loci as 24-nt sRNAs (S2C Fig), suggesting that sRNAs of differing sizes arose from the same Pol IV transcript in the WT but were likely processed into RNAs shorter than 24 nt by different downstream DICERs or by the exosome components Atrimmer1 and 2 [25,26]. Pol IV–dependent 21- to 23-nt sRNAs have been identified in other tissues, indicating that this finding is not specific to endosperm [12,27,28].

After identifying Pol IV–dependent sRNAs, we asked whether loss of one parent’s Pol IV influenced the abundance of Pol IV–dependent sRNAs in nrpd1 heterozygous endosperm. We sequenced sRNAs from 2 replicates of Ler female × Col-0 nrpd1−/− male (referred to as pat nrpd1+/−) F₁ endosperm and 3 replicates of Ler nrpd1−/− female × Col-0 male (referred to as mat nrpd1+/−) F₁ endosperm. Because the endosperm is triploid, in these comparisons, there are 3 (WT), 2 (pat nrpd1+/−), 1 (mat nrpd1+/−), and 0 (nrpd1−/−) functional NRPD1 alleles in the endosperm. However, expression of NRPD1 is paternally biased in WT Ler × Col endosperm [21]. Consistent with paternal allele bias, mRNA sequencing (mRNA-seq) data show that NRPD1 is expressed at 42% of WT levels in pat nrpd1+/− and at 91% of WT levels in mat nrpd1+/− endosperm (S10 Table).

We found that the presence of functional NRPD1 inherited from either parent is sufficient for the biogenesis of nearly WT levels of 21- to 24-nt sRNAs in endosperm (Fig 1A, S3 Fig). However, although the overall sRNA population in the heterozygotes was similar to the WT (Fig 1A), loss of maternal and paternal NRPD1 had distinct impacts on sRNA at individual loci (Fig 1B–1F, S3 Fig, S4–S7 Tables). We identified genes and TE insertions that displayed at least a 2-fold change in the abundance of sRNAs in mat or pat nrpd1+/− compared to the WT (Fig 1B–1F, S3 Fig, S4–S7 Tables). Loss of paternal NRPD1 caused relatively small fold change reductions in 21- to 24-nt Pol IV sRNAs at a handful of loci, while loss of maternal NRPD1 had slightly greater yet limited impact (Fig 1B–1F, S3 Fig, S4–S7 Tables). For genic loci with NRPD1-dependent 24-nt sRNAs, 2% (327 genes) had significantly lower sRNA abundance in mat nrpd1+/− compared to WT; in contrast, 0.3% (60 genes) had significantly lower sRNAs in pat nrpd1+/− (Fig 1B). For TE loci with NRPD1-dependent 24-nt sRNAs, 2.8% (545 TE insertions) and 1.35% (261 TE insertions) exhibited significantly lower sRNA abundance in mat and pat nrpd1+/−, respectively (Fig 1B). Few of the loci with reduced sRNAs were shared between the reciprocal heterozygote—of 327 24 nt sRNA-expressing genic loci that were reduced by more than 2-fold in mat nrpd1+/−, only 22 were also reduced by 2-fold in pat nrpd1+/− (Fig 1F). Moreover, there was no quantitative or correlative relationship between loci affected in mat nrpd1+/− and pat nprd1+/− (Fig 1F). Thus, the vast majority of sRNA-producing loci in endosperm only require at least 1 functional copy of NRPD1 after fertilization.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Impact of loss of maternal, paternal, or both copies of NRPD1 on endosperm sRNAs.

(A) Loss of maternal or paternal NRPD1 does not substantially alter the endosperm sRNA pool. Fraction of aligned sRNA reads in each size class in the indicated genotypes. (B) Examination of 21- to 24-nt sRNAs over genes or TEs shows that inheriting a mutant maternal nrpd1 allele has a greater impact than inheriting a mutant paternal nrpd1 allele. Percent of loci showing at least a 2-fold reduction in sRNA abundance and padj < 0.05 according to DESeq2 are indicated in red (mat nrpd1+/−) or blue (pat nrpd1+/−). Genes and TEs included in this tally have a normalized WT read count of 5 or higher. (C) Snapshots of loci with Pol IV–dependent 24-nt sRNAs that show a specific loss of sRNAs in mat (left) or pat (right) nrpd1+/− endosperm. (D–F) Comparisons of genic 24-nt sRNAs upon loss of maternal, paternal, or both copies of NRPD1. Fold change as calculated by DESeq2. Only significant changes (padj < 0.05) are plotted. Underlying data for Fig 1 A, B, and D–F can be found in S1 Data. Pol IV, polymerase IV; sRNA, small RNA; TE, transposable element; WT, wild-type.

https://doi.org/10.1371/journal.pbio.3001602.g001

Evaluating memory of parental Pol IV activity and endosperm sRNA production

The absence of dramatic differences in sRNAs in heterozygotes could indicate that the alleles inherited from both the WT and the nrpd1−/− parent produce a WT level of sRNAs after fertilization. This result would be expected for a recessive mutation without parental effects. However, it is known that Pol IV activity at some loci requires prior Pol IV activity [29]. Under such a scenario, Pol IV activity in the parents before fertilization might be necessary for sRNA production from that parent’s allele in the endosperm after fertilization. Thus, the observed lack of differences in sRNA production at most loci in heterozygous nrpd1 endosperm (Fig 1) might be explained by an up-regulation of sRNA production from the alleles inherited from the WT parent (i.e., sRNAs are up-regulated from paternal alleles in mat nrpd1+/−, and maternal allele sRNAs are up-regulated in pat nrpd1+/− endosperm). To distinguish between these possibilities, we used the SNPs between Col-0 and Ler to identify the allelic origins of sRNAs in WT and heterozygous endosperm. We first confirmed prior observations that Pol IV sRNAs are biallelically expressed at most loci in endosperm and predominantly expressed from one parental allele, or imprinted, at several hundred others [14]. By examining sRNAs at genes and TEs, we found that both biallelically expressed 21- and 24-nt sRNA loci (defined as between 20% and 80% of sRNAs from maternally inherited alleles) and those predominantly expressed from one parental allele (>80% or <20% maternal) were Pol IV-dependent (i.e., their accumulation was significantly reduced in nrpd1−/− endosperm) (S4 Fig).

To test if sRNA production from alleles inherited from WT parents compensated for alleles inherited from an nrpd1−/− parent, we first assessed several thousand loci that were not significantly misregulated in nrpd1+/− endosperm. Overall, there were similar contributions from maternal and paternal alleles in mat and pat nrpd1 heterozygotes compared to WT endosperm (Fig 2A). This suggests that by 7 DAP (days after pollination), at most loci in the endosperm, sRNAs are produced from both maternal and paternal alleles regardless of whether the alleles were inherited from a WT parent or an nrpd1−/− parent. However, we found that imprinted sRNA regions (ISRs; 113.1-kb maternally imprinted and 1,215.6-kb paternally imprinted regions overlapping both genic and TE loci) [14] were impacted by loss of parental Pol IV (Fig 2C–2F). ISR loci have been filtered to remove regions that are also enriched for seed coat sRNAs [14] and thus preclude analytical artifacts that may arise due to maternal tissue contamination or due to any potential sRNA movement. A total of 179 of 206 ISRs where expression was maternally biased in the WT showed reduced 24-nt sRNAs in mat nrpd1+/− (Fig 2C). On the other hand, only a small subset (74 of 2,405 ISRs) of paternally biased ISRs produced fewer sRNAs in pat nrpd1+/− (Fig 2F) and slightly more sRNAs in mat nrpd1+/− (Fig 2D). We also note that maternally biased regions in WT showed slightly elevated production of sRNAs in pat nrpd1+/− endosperm (Fig 2E). Examination of the allelic origins of sRNAs at genes and TEs are also consistent with the ISR analysis. We found that sRNA loci showing dramatic reductions in abundance in mat nrpd1+/− tended to be maternally biased in WT endosperm (>80% of sRNAs from the maternally inherited alleles) (Fig 2B, leftmost column). Similarly, in pat nrpd1+/−, paternally biased sRNA (<20% sRNAs from the maternally inherited alleles) loci were more impacted (Fig 2B, rightmost column).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Effects of loss of maternal or paternal Pol IV on the allelic origin of sRNAs.

(A) Tukey plot shows no difference in allelic origin of genic and TE sRNAs between nrpd1 heterozygotes and WT. Loci plotted here showed similar sRNA abundances in WT and heterozygotes and have a sum of at least 10 allele-specific reads in 3 WT replicates and in heterozygotes. (B) Loci with reduced sRNAs in mat nrpd1+/− or pat nrpd1+/− exhibit maternally or paternally biased sRNAs in WT. Genes and TEs showing differential abundance of 24-nt sRNAs in nrpd1 heterozygotes were grouped into bins by the % of sRNAs produced from the maternal alleles of that locus in WT. Fold change was calculated by DESeq2. Tukey plot represents fold change in each group; circles show fold change at individual loci. Numbers below and above plot are total number of loci having significantly lower and higher abundance of 24-nt sRNAs in nrpd1+/− relative to the WT. (C–F) Loss of maternal NRPD1 leads to a reduction in the abundance of sRNAs from maternally biased ISRs (C) and gain of sRNAs from a subset of paternally biased ISRs (D). Loss of paternal NRPD1 has a negligible impact on maternally biased ISRs (E) and a relatively minor impact on paternally biased ISRs (F). The Col-Ler ISRs used here were defined by Erdmann and colleagues (2017). To identify regions with changes in sRNA abundance, read counts were calculated over sliding windows of 300 bp with 200-bp overlap. Windows with differential abundance were identified using DESeq2. Windows overlapping an ISR were identified using bedtools intersect. Overlapping windows were merged using bedtools merge and the median read count for each set of merged windows was plotted. Windows with and without significant differences in abundance are represented by black and gray circles. Data represented in this figure can be found in S3 Data. ISR, imprinted sRNA region; Pol IV, polymerase IV; sRNA, small RNA; TE, transposable element; WT, wild-type.

https://doi.org/10.1371/journal.pbio.3001602.g002

In summary, these results indicate that most maternally and some paternally biased imprinted sRNA loci in endosperm are dependent on Pol IV activity in the parents and are not established de novo postfertilization. Notably, these sites of Pol IV action are by definition distinct between maternal and paternal parents.

Maternal and paternal RNA Pol IV have antagonistic impacts on gene expression

We previously identified several hundred genes misexpressed in nrpd1−/− endosperm [14]. To test for maternal or paternal effects on endosperm gene expression, we performed mRNA-seq in 3 replicates each of mat nrpd1+/− and pat nrpd1+/−, along with appropriate WT controls and homozygous mutant nrpd1 endosperm (S1 Table). Evaluation of these datasets using a tissue enrichment test showed no indication of contamination with maternal seed coat tissue (S5 Fig). Differential expression analyses identified 1,791 genes whose transcripts were more abundant and 1,455 that were less abundant in nrpd1−/− compared to WT endosperm (Fig 3, S10 Table). Almost 50% of these genes (1,599) were similarly misregulated in mat nrpd1+/− (Fig 3A and 3B), along with 2,998 additional genes. In contrast, very few genes (90) changed in expression in pat nrpd1+/− compared to the WT (Fig 3A and 3B). In addition to the difference in the size of the effect, loss of maternal or paternal Pol IV altered the expression of different classes of genes. Panther overrepresentation tests [30] indicated that in mat nrpd1+/−, down-regulated genes were enriched for functions in the cell cycle, whereas up-regulated genes were enriched for functions in photosynthesis, stress response, and abscisic acid signaling (S11 Table). In pat nrpd1+/−, up-regulated genes were enriched for functions in heat stress response, while down-regulated genes were enriched for functions in responses to fungi (S11 Table). The expression of imprinted genes is known to be regulated epigenetically in endosperm. In mat nrpd1+/− endosperm, imprinted genes were more likely to be misregulated than expected by chance (hypergeometric test p < 10⁻¹⁵)—15 out of 43 paternally expressed and 45 out of 128 maternally expressed imprinted genes were misregulated in mat nrpd1+/−, whereas 2 maternally expressed imprinted genes but no paternally expressed imprinted genes were misregulated in pat nrpd1+/− (S6 Fig, S10 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Maternally and paternally acting Pol IV have antagonistic effects on endosperm gene expression.

(A) Venn diagrams showing overlap of genes with increased and decreased expression in comparison to WT endosperm for the indicated genotypes. (B) Scatter plots of genes that are significantly different (q ≤ 0.05, Log₂(Fold Change) ≥1 or ≤ −1) between WT and the indicated mutants. Fold change calculated using Cuffdiff. (C) Examples of genes that are antagonistically regulated by Pol IV. Gray bars represent mathematical sum of effects observed in mat and pat nrpd1+/−. (D–F) Inverse relationship between changes in gene expression in mat and pat nrpd1+/− relative to WT. Genes that are antagonistically influenced by maternal and paternal NRPD1 are colored purple, while genes not antagonistically regulated are colored gray. In D, genes up-regulated at least 2-fold in both nrpd1−/− and mat nrpd1+/− do not exhibit misregulation in pat nrpd1+/−, while genes up-regulated only in mat nrpd1+/− but not nrpd1−/− have decreased expression in pat nrpd1+/−. In E, genes down-regulated in both nrpd1−/− and mat nrpd1+/− are not misregulated in pat nrpd1+/−, while genes down-regulated only in mat nrpd1+/− but not in nrpd1−/− are overall slightly increased in expression in pat nrpd1+/−. In F, genes that are significantly (2-fold, q ≤ 0.05) misregulated in both mat nrpd1+/− and nrpd1−/− show little to no inverse relationship between mat and pat nrpd1+/− (slope = −0.07), while genes that are only misregulated in mat nrpd1+/− but not nrpd1−/− are inversely affected in pat nrpd1+/− (slope = 0.25). Plots D and E show median and interquartile range for log₂(fold change) in mutant/WT. Fold change were calculated by Cuffdiff. Data represented in this figure can be found in S4 Data. Pol IV, polymerase IV; WT, wild type.

https://doi.org/10.1371/journal.pbio.3001602.g003

Differential expression of a gene between WT and nrpd1−/− could represent (1) maternal and paternal effects arising from the loss of NRPD1 in parents; (2) zygotic effects arising from epistatic interactions between mat nrpd1− and pat nrpd1−; (3) effects from the loss of all NRPD1 in the endosperm; or (4) the sum of all 3 effects. As this study does not examine the effect of knocking down NRPD1 specifically in the endosperm, we can only detect parental effects. Curiously, 2,988 genes misregulated in mat nrpd1+/− were not misregulated in nrpd1−/− endosperm (Fig 3A). We hypothesized that genic misregulation found exclusively in mat nrpd1+/− (but not nrpd1−/−) was caused by separate transcriptional effects of maternal and paternal nrpd1 that were obscured in null mutants. To test this hypothesis, we compared gene expression between mat and pat nrpd1+/− (Fig 3C–3F). We found that 51/90 genes misregulated in pat nrpd1+/− endosperm were also misregulated in mat nrpd1+/− endosperm. However, 36 of these 51 genes changed expression in the opposite direction (hypergeometric test for enrichment, p < 10⁻¹⁰). For example, expression of the gene SUC2 decreased about 4-fold in pat nrpd1 +/− endosperm and increased about 8-fold in mat nrpd1+/− endosperm (Fig 3C). If NRPD1 loss has no endospermic (zygotic) effect on the expression of these genes, then the misregulation observed in nrpd1−/− endosperm would be the sum of the parental effects. Indeed, the change in abundance of these genes in nrpd1−/− endosperm is close to that predicted by an additive, antagonistic parental effect (compare gray and green bars in Fig 3C). SUC2 transcript abundance in nrpd1−/− changes by 2.7-fold compared to the predicted 2.18-fold change, and other genes show similar effects (Fig 3C). While the expression of these particular genes showed large effects in both heterozygotes, most genes misregulated in mat nrpd1+/− did not show a significant change (defined as >2-fold difference in transcript abundance) in pat nrpd1+/− endosperm. We therefore hypothesized that misregulation of genes in mat nrpd1+/− but not nrpd1−/− endosperm was due to a small antagonistic effect arising from the loss of pat NRPD1 in nrpd1−/−. To further test this hypothesis, we evaluated the expression of genes in pat nrpd1+/− endosperm that were either misregulated in both mat nrpd1+/− and nrpd1−/− or only in mat nrpd1 +/− (Fig 3D–3F). Transcripts that significantly increased exclusively in mat nrpd1+/− had slightly decreased expression in pat nrpd1+/− endosperm (Fig 3D, purple). In contrast, genes that were significantly up-regulated in both mat nrpd1+/− and nrpd1−/− were not affected in pat nrpd1+/− (Fig 3D, gray). Similarly, genes that showed a significant reduction in abundance only in mat nrpd1+/− were slightly higher expressed in pat nrpd1+/− endosperm (Fig 3E, purple), while genes with reduced abundance in both mat nrpd1+/− and nrpd1−/− were not affected in pat nrpd1+/− (Fig 3E, gray). Finally, the antagonistic relationship could also be observed when directly comparing changes in mRNA abundance at individual genes, genome-wide, and upon loss of maternal and paternal Pol IV. Genes that were similarly misregulated in mat nrpd1+/− and nrpd1−/− showed a limited relationship (slope = −0.07). By contrast, genes that were misregulated in mat nrpd1+/− but not nrpd1−/− exhibited a clear inverse relationship (slope = −0.25) These results are consistent with an antagonistic parent-of-origin effect model for the impact of Pol IV on endosperm transcriptomes. Although the antagonistic effect at most genes is less than the commonly used 2-fold threshold difference for a significant change in gene expression, it is similar in magnitude to dosage compensation effects in other systems, such as that observed for the fourth chromosome in Drosophila melanogaster and for genes mediating genetic compensation in zebrafish [31,32].

Evaluating possible mechanisms of Pol IV’s impact on gene expression

How does Pol IV have parent-of-origin effects on gene expression in the endosperm after fertilization? Pol IV effects could be direct or indirect at the affected loci. One possibility is that Pol IV modulates gene expression via the proposed posttranscriptional gene silencing (mRNA cleavage) or translational inhibition by 21- to 22-nt Pol IV–dependent sRNAs [27,33]. Or, Pol IV–dependent RdDM over genic sequences or associated gene regulatory elements might repress transcription in the WT. Alternatively, Pol IV could impact many genes in trans by regulating the expression of chromatin proteins, like the known target ROS1 (a DNA demethylase) (Fig 3C) [34], transcription factors [11,35], or by broadly influencing genome organization, which, in turn, affects gene expression [36,37]. To estimate the contribution of cis- and trans-effects and identify potential cis-regulatory targets of Pol IV that could drive widespread trans-effects, we analyzed the congruence of sRNAs, DNA methylation, mRNA cleavage patterns, and allele-specific changes driving gene expression changes in WT and mutant endosperm.

Assessing potential mRNA cleavage by Pol IV–dependent sRNAs in endosperm

Pol IV–dependent genic sRNAs are proposed to regulate gene expression by cleaving mRNA [27], and we previously demonstrated that endosperm has greater accumulation of genic sRNAs than other tissues [14]. To test if such cleavage events contribute to endosperm Pol IV–dependent transcript abundance, we first identified candidate genes that exhibited significantly increased (≥2-fold) mRNA abundance and significantly decreased 21-, 22-, or 24-nt sRNA abundance (≥2-fold) in nrpd1−/− endosperm (S7A–S7C Fig). This analysis suggested that at least 305 genes, or 16% of the genes that increase in expression in nrpd1−/− endosperm, were associated with Pol IV–dependent sRNAs. To directly assay if these genic sRNAs drive mRNA cleavage at levels sufficient to alter transcript abundance at specific loci, we mapped the 5′ ends of mRNA from WT and nrpd1−/− endosperm mRNA using NanoPARE sequencing [38]. NanoPARE maps both the 5′ ends of primary transcripts and those that result from mRNA cleavage. We confirmed that NanoPARE sequencing was working for us by identifying transcriptional start sites as well as internal cleavage sites for known miRNA targets (S7E Fig). We found that almost all genes exhibiting increased mRNA abundance also exhibited increased 5′ ends at transcriptional start sites, but did not have reduced 5′ ends internal to the gene (S7D Fig). This was confirmed by visual observation of individual loci (S7E Fig). This suggests that the increase in the transcript abundance of these genes in nrpd1−/− was not caused by reduced mRNA cleavage compared to the WT. Only 5 genes—PERK8, GLP2A, ETTIN, AAD3, and AT2G45245—exhibited reduced cleavage at a few sites in nrpd1−/− endosperm. This minimal effect is contrary to the expectation that candidate genes described above should have reduced cleavage and suggests that sRNA-mediated posttranscriptional gene silencing is not a key mechanism for Pol IV to control endosperm gene expression. We therefore did not test if mRNA cleavage is impaired in nrpd1+/− heterozygotes.

Assessing correspondence between sRNA and mRNA changes

Only a minority of the genes that have altered expression in nrpd1 endosperm have associated changes in sRNAs within those same genes (S7A–S7C Fig). However, Pol IV sRNAs may also act at sites proximal to a gene to regulate it. We assessed the distance between misexpressed genes and altered sRNAs in homozygous and heterozygous nrpd1 mutant endosperm. We found that 9.2%, 11.7%, and 3.3% of misregulated genes are within 1 kb of a site that loses sRNAs in nrpd1−/−, mat nrpd1+/−, and pat nrpd1+/− endosperm, respectively (S12 Table). To obtain a genome-wide perspective not focused on arbitrary distance cutoffs, we used the relative distance metric to test if genomic regions losing 24-nt sRNAs were associated with misregulated genes. The relative distance metric describes the spatial correlation between sRNA intervals and misregulated genes, compared to misregulated genes and random intervals [39]. This analysis found no enrichment in the association between Pol IV–dependent sRNAs and misregulated genes in any of the genotypes (S7F Fig).

Assessing correspondence between DNA methylation and mRNA changes

The relevant molecular function of RNA Pol IV with regard to gene expression is typically assumed to be its role in RdDM. To identify potential examples of DNA methylation mediating Pol IV’s impact on genes, we performed bisulfite sequencing of WT, mat nrpd1+/−, and pat nrpd1+/− endosperm DNA. We evaluated WT DNA methylation at individual cytosines within Pol IV sRNA-producing genes that were misregulated in nrpd1−/−, misregulated genes that showed increased sRNAs in nrpd1−/− (Pol IV-independent), and 5 control sets of genes that showed no change in sRNA abundance upon loss of Pol IV. We found that most cytosines were not methylated (median is near zero) in the genes we examined (S8A Fig). This suggests that Pol IV sRNAs do not generally target DNA methylation at genes. However, some misregulated genes with Pol IV–dependent sRNAs had higher CG methylation in WT endosperm (S8A Fig). These misregulated genes with Pol IV–dependent sRNAs were also likely to be longer (S8B Fig). Longer genes have higher sRNA read counts (S8C Fig), and, thus, differences in these genes are more likely be called as statistically significant by DESeq2 [40]. Longer genes also tend to have higher CG methylation [41,42]. We therefore argue that this increased CG methylation is likely an analytical artifact. Our results suggest that Pol IV–dependent genic sRNAs do not regulate endosperm gene expression by directing genic DNA methylation, consistent with our previous findings [14].

We also tested if changes in DNA methylation brought about by loss of parental Pol IV could explain changes in gene expression. Overall, loss of parental Pol IV had only minor effects on DNA methylation (S9 Table). Loss of Pol IV activity primarily reduces asymmetric CHH methylation [43]. Comparison of mat nrpd1+/− and pat nrpd1+/− CHH methylation with WT endosperm identified 2,234 and 2,056 DMRs (covering 812.7 kb and 759.9 kb, respectively) with 50% hypomethylated in mat nrpd1+/− and 54.8% hypomethylated in pat nrpd1+/− (S9 Table). Consistent with the parent-of-origin effects described for mRNAs and sRNAs, we made 3 observations that suggest that mat and pat Pol IV activity have distinct impacts on the endosperm methylome. First, only 50% of CHH DMRs are shared between the 2 heterozygous genotypes. Second, regions where sRNA accumulation is dependent on paternal inheritance of a WT NRPD1 allele have higher CHH methylation in WT endosperm than regions where sRNAs are dependent on maternal NRPD1 (S8D Fig). This pattern is consistent with our previous finding that maternally biased sRNAs are often not associated with methylated DNA in WT endosperm [14]. Third, an examination of regions with at least 10% CHH methylation in WT endosperm shows that loss of paternal NRPD1 had a more substantial impact on endosperm CHH methylation than loss of maternal NRPD1 (S8E Fig).

Symmetric CG and CHG methylation are typically less affected by loss of NRPD1 because other mechanisms exist to maintain this type of methylation. Comparison of CHG methylation between WT and either heterozygote identified fewer than 100 DMRs and was not investigated further. Both mat and pat nrpd1+/− endosperm exhibited changes in CG methylation compared to the WT (S9 Table). In mat nrpd1+/− endosperm, 48.5% of CG DMRs (of a total 600 DMRs spanning 207 kb) were hypomethylated relative to WT while in pat nrpd1+/−, 60% of CG DMRs (of a total 707 DMRs spanning 258 kb) were hypomethylated relative to WT. Further, we found that few of the sites hypo- or hypermethylated in the CG context in mat nrpd1+/− were shared with those changing methylation state in pat nrpd1+/− (S9 Table).

We used the DMRs in the analyses described above to assess their impact on gene expression. In mat nrpd1+/− endosperm, 2.6% and 3.4% of misregulated genes are within 1 kb of assayable regions with altered CHH methylation in mat nrpd1+/− endosperm (S12 Table). In addition, 2 genes and 1 gene are within 1 kb of a region that has higher and lower CHH methylation in pat nrpd1+/−. One gene is associated with increased CG methylation in pat nrpd1+/−. We also used relative distance analysis to see if mat nrpd1+/− misregulated genes are more likely to be associated with DNA methylation changes (there are too few genes associated with DNA methylation changes in pat nrpd1+/− to perform this analysis). Consistent with previous analyses, we find no clear relationship between DNA methylation changes and gene expression changes in the mat nrpd1+/− endosperm (S8F Fig).

Allelic analysis of misregulated genes to identify cis- or trans-effects of Pol IV

One method to assess whether Pol IV’s impacts on gene expression are predominantly cis- or trans-acting is to compare the allelic origins of mRNA in WT and nrpd1+/− endosperm. If a gene’s mRNA abundance in the endosperm is determined by the activity of Pol IV in cis either in the gametophyte or sporophyte, then the gene would be primarily misregulated from the allele inherited from a parent lacking Pol IV. Thus, in mat nrpd1+/− endosperm, misregulation of such genes would be driven predominantly by changes in expression from maternal alleles, whereas gene expression differences in pat nrpd1+/− endosperm would be driven by changes in expression from paternal alleles. In contrast, the predominance of trans-effects would be indicated by both parental alleles contributing to the changes in the total transcript abundance at most genes. We utilized SNPs between Col-0 (paternal) and Ler (maternal) genomes to identify allele-specific mRNA-seq reads. We evaluated the contributions of each parent’s alleles in the endosperm for 2372 misregulated genes that had at least 10 allele-specific reads in WT and mat nrpd1+/−. For the majority of genes, misregulation in mat nrpd1+/− was driven by effects on expression of both maternal and paternal alleles, with some notable exceptions (S9A Fig). For example, increased expression of DOG1 in mat nrpd1+/− was primarily due to increased expression from maternal alleles (S9B Fig). The transcript from AT4G12870 was repressed in mat nrpd1+/− primarily due to a loss of maternal allele expression (S9B Fig). In contrast, expression of SAC2 was primarily repressed in mat nrpd1+/− because of decreased expression from the paternal allele (S9B Fig). Overall, both maternal and paternal alleles made equal contributions to genic misregulation in the mat nrpd1+/− endosperm. Moreover, 4.7% of down-regulated genes and 5.3% of up-regulated genes showed at least a 20% increase or decrease in maternal allele contribution. This was roughly similar to the contribution of paternal alleles to misregulation in mat nrpd1+/−: 4.4% of down-regulated genes and 5.3% of up-regulated showed at least a 20% change in paternal allele contribution (S9A Fig). In pat nrpd1+/−, only 8% of down-regulated genes had lower contribution of paternal alleles, while both alleles contributed to up-regulation (S9A Fig). Overall, these results suggest that parental Pol IV’s impact on gene expression is largely due to trans-effects.

In summary, our analyses test and dismiss several cis-regulatory mechanisms for how Pol IV may mediate parent-of-origin gene expression effects on the endosperm. We also individually examined DNA methylation and sRNAs at genes showing antagonistic regulation by maternally and paternally acting Pol IV and found no evidence for a role for sRNAs and DNA methylation in their regulation. These results lead us to conclude that the parent-of-origin and antagonistic effects that we observe are likely the result of trans-acting effects of parental Pol IV activity.

Pol IV, conflict, and the genetic architecture of parental control

Parental conflict theory predicts that in viviparous and polyandrous species, mothers and fathers have antagonistic effects on regulating resource allocation and associated gene expression in offspring [45]. In practice, such effects are difficult to detect and have been infrequently described [46]. Analogous to observations for cryptic meiotic drive suppression systems [47], antagonistic parental effects are likely to be balanced in the individuals within an inbred population (like Arabidopsis) and are thus unobservable except in mutants or in hybrids where maternal and paternal effects are out of balance. When homozygous mutants are examined, these effects may be missed because they do not cause dramatic developmental phenotypes or because simultaneous loss of antagonistic maternal and paternal effects effectively cancels one another out. Thus, reciprocal heterozygotes need to be examined to detect antagonistic parent-of-origin effects. A close examination of our data provides insights into the genetic architecture mediating parental control of offspring development.

A key feature of the regulatory infrastructure that mediates parent-of-origin specific effects on zygotic gene expression is that maternal and paternal alleles need to be distinguished from each other in the zygote (in this case, endosperm is the relevant zygote). In A. thaliana endosperm, at many loci, maternally inherited alleles are DNA demethylated and marked with H3K27 methylation by polycomb repressive complex2 (PRC2), while paternally inherited alleles remain DNA methylated and have reduced H3K27me3 [21,48,49]. Maternal inheritance of mutations in the PRC2 subunits MEA, FIE, FIS2, and MSI1 leads to endosperm defects and seed abortion [50–53]. Similarly, inheritance of maternal mutations in the DNA demethylase DME increases DNA methylation on endosperm maternal alleles and causes seed abortion [54]. Paternal inheritance of mutations in these genes have no reported effect on endosperm development or gene expression. These results thus argued that the solution to the problem of distinguishing parental alleles from one another after fertilization was to mark maternal and paternal chromosomes with distinct epigenetic modifications. However, this model may not explain all parent-of-origin effects on gene expression, particularly outside of imprinted genes. Our study provides evidence for a distinct model in which the same epigenetic regulator—Pol IV—can mediate both maternal and paternal effects. The only other example of a gene with seemingly antagonistic effects on seeds is the maintenance methyltransferase MET1, whose mutation has opposing effects on seed size when inherited maternally or paternally, although the molecular basis of this phenotype is unknown [55].

How does Pol IV in the mother and the father have distinct impacts after fertilization? Pol IV targets can be tissue or developmental stage specific [56], and, thus, Pol IV may target different genomic regions during male and female gametogenesis. Pol IV could act pre- or postmeiotically in the parental sporophyte (diploid phase of the life cycle), in the gametophyte (haploid phase of life cycle), or postfertilization in the maternal sporophyte. Reverse transcription PCR (RT-PCR)-based examination of dissected synergids and central cells did not detect NRPD1 transcripts [44]. This suggests that on the maternal side, Pol IV influences endosperm gene expression by acting in the maternal sporophyte or in the female gametophyte prior to central cell formation. Alternatively, Pol IV could act in the maternal sporophytic integuments/seed coat after fertilization, when the endosperm is developing. One potential mechanism for this would be through Pol IV–dependent sRNAs moving from the seed coat to the endosperm [11,56]. However, examination of the levels of total Pol IV–dependent sRNAs, allele-specific data, and ISRs suggests that the potential influence of seed coat Pol IV function on endosperm expression would likely be independent of sRNA transfer. This conclusion is consistent with previous observations that endosperm and seed coat have distinct sRNA profiles [14].

We have shown that parental Pol IV activity is dispensable for guiding endosperm sRNA production at most loci, with the exception of ISRs, but that parental Pol IV activity plays an important role in guiding endosperm gene expression. The molecular nature of this memory is unknown, and at present, we can only speculate. Data from paternal excess interploidy crosses suggest that the molecular identity of Pol IV memory may differ between the maternal and paternal parents. In the father, the genes required for sRNA production (NRPD1, RDR2, and DCL3) and the genes required for downstream DNA methylation (NRPE1/Pol V and DRM2) are both essential to promote paternal excess seed abortion [10]. In contrast, in the mother, genes required for sRNA production but not for DNA methylation promote paternal excess seed viability [10]. This suggests that DNA methylation or another downstream chromatin mark directed by Pol IV–dependent sRNAs could be the identity of paternally inherited memory, but is unlikely to be the molecular identity of maternally inherited memory. What would be the nature of maternal DNA methylation–independent memory? Pol IV, like other RNA polymerases [57], could act as a chromatin remodeler. Or, Pol IV could direct a chromatin modification, produce sRNAs that posttranscriptionally control genes, or control the expression of genes whose products are deposited in the gametes, which, in turn, sets up a memory to direct gene expression programs in the endosperm after fertilization.

How might we interpret Pol IV’s parent-of-origin effects in terms of conflicts between parents? The WISO or “weak inbreeder/strong outbreeder” model [58] emerges from the dynamics of parental conflict and parent-of-origin effects. Under this model, a parent from populations with higher levels of outcrossing is exposed to higher levels of conflict and can thus dominate the programming of maternal resource allocation in a cross with an individual from a population with lower levels of outcrossing. Such a phenomenon has been observed in numerous clades including Dalechampia, Arabidopsis, Capsella, and Leavenworthia [59–62]. Intriguingly, loss-of-function phenotypes in the RdDM pathway are more severe in recently outcrossing species than in A. thaliana [12,13] and suggests that RNA Pol IV functions are more elaborate in these species. This raises the possibility that the role for RNA Pol IV and RdDM in parental conflict that we describe here in A. thaliana is likely heightened in and mediates the elevated level of parental conflict in species that are currently or have been recently outcrossing.

Studies on how resource allocation conflicts between parents impact gene expression have thus far been focused on imprinted genes. However, a handful of studies show the importance of nonimprinted genes in parent-of-origin effects [63,64]. For example, quantitative trait locus (QTL) analyses of a heterogeneous mouse stock showed that nonimprinted genes mediate parent-of-origin effects on the offspring’s immune system [64]. Our study describes for the first time a system in which an epigenetic regulator acts in the mother and the father to antagonistically regulate the same nonimprinted genes in the zygote. While the magnitude of effects at many genes may be small, it should be noted that small changes in gene expression can be associated with very different phenotypes [65]. Our allele-specific mRNA-seq data show that loss of Pol IV from one parent can impact alleles inherited from both parents in the endosperm. This suggests that Pol IV does not act directly at antagonistic loci but acts instead by regulating other modifiers of gene expression. Yet, this antagonistic regulation can also be viewed from another perspective. Parental conflict can be resolved or paused if both parents can modulate the expression level of a gene or the activity of a pathway to an optimum that is tolerable to each. Pol IV’s role in mediating the antagonistic effects of both parents makes it an ideal system to negotiate optimal gene expression levels. Thus, Pol IV may not be solely an agent of conflict, but also a means to resolving it. Overall, these data suggest that Pol IV is part of a gene regulatory network that is evolving under parental conflict.

Arabidopsis growth conditions, strains, and tissue collection

Plants used in this experiment were grown at 22°C in a Conviron chamber in 22-hour light (120 μM). The A. thaliana mutant used in this study was nrpd1a-4 (SALK_083051 obtained from ABRC, The Ohio State University) [66] in the Col-0 background. We also utilized nrpd1a-4 introgressed 4 times into Ler [14]. Endosperm from approximately 100 seeds (7 days after pollination) from at least 3 siliques was dissected free of embryos and seed coats and pooled for each biological replicate as previously described [22]. Each biological replicate was collected from crosses that used different individuals as parents. The number of replicates for each experiment was decided based on currently accepted practices in genomic studies. For sRNA experiments, we planned to sample 3 biological replicates for each genotype. However, we had to discard one of the 3 pat nrpd1+/− sRNA libraries because that library had too few reads.

mRNA, sRNA, and DNA isolation and library construction

Large- and small-sized RNAs were isolated using the RNAqueous-Micro RNA isolation kit (Thermo Fisher Scientific, Waltham, MA). Briefly, endosperm dissected from seeds was collected in lysis buffer and then homogenized with an RNAse-free pellet pestle driven by a Kimble motor. Large and sRNA species were isolated and separated using the manufacturer’s protocol. The RNA concentration of the larger fraction was measured by Qubit. sRNA libraries were constructed using the NEXTFLEX Small RNA-seq kit v3 (Bioo Scientific, Austin, TX). Final library amplification was carried out for 25 cycles, and the libraries were size selected (135 to 160 bp) using a Pippin Prep (Sage Science, Beverly, MA). mRNA-seq libraries were constructed using a Smart-Seq2 protocol [67]. NanoPARE libraries were built as described in [38]. All libraries were sequenced on the Illumina Hi-Seq 2500. Seed coat contamination in our samples was ruled out by examining transcriptome data using a previously published tool [68].

DNA for bisulfite sequencing was isolated from dissected endosperm at 7 days after pollination using the QIAamp DNA Micro kit (QIAGEN 56304 Hilden, Germany). Dissected tissue was obtained for 2 biological replicates for each genotype and incubated overnight in a shaker at 56°C in ATL buffer with Proteinase K. Between 70 and 100 ng of endosperm DNA was subjected to bisulfite treatment using the MethylCode Bisulfite Conversion kit (Invitrogen, Waltham, MA). Analysis of cytosines from chloroplasts with at least 10 sequenced reads showed a conversion rate of greater than 98% for all libraries. Bisulfite converted DNA was used to build libraries with the Pico Methyl-Seq library kit (Zymo Research, D5455, Orange, CA). A total of 7 cycles of amplification were used for library construction. All libraries were sequenced on the Illumina Hi-Seq 2500 (60-bp paired-end reads).

sRNA analysis

sRNA reads were trimmed with fastq_quality_trimmer (fastq_quality_trimmer -v -t 20 -l 25). Cutadapt [69] was used to identify adapter bearing reads of suitable length (cutadapt -a TGGAATTCTCGGGTGCCAAGG—trimmed-only—quality-base 64 -m 24 -M 40—max-n 0.5—too-long-output). Taking advantage of the random nucleotides on the adapters in NEXTFLEX kits, we used Prinseq (prinseq-lite-0.20.4) (prinseq-lite.pl -fastq <infile> -out_format 3 -out_good <filename> -derep 1 -log) to remove PCR duplicates [70]. Filtered reads were aligned to a genome consisting of concatenated Col-0 TAIR10 and Ler pseudo-genome (Col-0 genome substituted with Ler SNPs) using Bowtie (v 1.2.2) bowtie -v 2—best -p 8–5 4–3 4—sam <index file> <infile.fq> (2 mismatches, report best alignment, ignore 4 bases on 5′ and 3′ ends) [71]. Reads mapping to Ler were lifted over to Col-0 using custom scripts [14]. A custom script, assign-to-allele, was used to identify reads arising from Col-0 or Ler alleles (https://github.com/clp90/imprinting_analysis/tree/master/helper_scripts). Aligned reads between 21- and 24-nt in length were binned based on size. Bedtools was used to count reads in 300-bp windows with 20-bp overlaps and over annotated genes and TEs from Araport 11. DESeq2 [24] was used to identify features showing differences in sRNA abundance with an adjusted p-value of 0.05 or less. One complication with using DESeq2 is that the loss of Pol IV–dependent sRNAs at most loci in nrpd1−/− leads to an underestimation of WT library size by DESeq2, which increases the proportion of false negatives and undercounts the number of Pol IV–dependent sRNA loci. To allay this effect while analyzing genes, we excluded TEs and applied differential expression analysis to just genic and miRNA loci. These non-TE loci also included Pol IV-independent sRNA loci, which provide an estimate of library size. We separately examined TEs using genic sRNA counts to provide an estimate of library size. ShortStack version 3.8.5 [23] was also used as an orthogonal approach to identify sRNA peaks from bam alignment file output from Bowtie. Parameters chosen for ShortStack included dicermin = 20, dicermax = 25, and a mincov of 0.5 rpm. Weightage for multimapping reads was guided by uniquely mapping reads (option = u).

mRNA-seq and NanoPARE analysis

The reads from mRNA-seq and NanoPARE were trimmed for quality with “trim_galore -q 25—phred64—fastqc—length 20—stringency 5” and aligned to the TAIR10 genome using Tophat (v2.1.1) [72] using the command tophat -i 30 -I 3000—segment-mismatches 1—segment-length 18 —b2-very-sensitive. Cuffdiff (v2.1.1) [73] was used to identify differentially expressed genes for mRNA-seq data. Aligned NanoPARE read counts at each nucleotide in the genome were counted using bedtools. Sites with statistical differences in NanoPARE read counts were identified by DESeq2.

DNA methylation analysis

Reads from Bisulfite sequencing were trimmed for quality using Trim Galore (https://github.com/FelixKrueger/TrimGalore). Trimmed reads were aligned to the TAIR10 genome using Bismark [74] with parameters set to -N 1 -L 20—non_directional. For this alignment, paired-end reads were treated as single reads. Previously described Bismark methylation extractor and custom scripts [21,75] were used to determine DNA methylation/base, and then methylation was calculated for 300-bp windows that overlapped by 200 bp. Data from the 2 biological replicates for each genotype were pooled together for comparison between genotypes. To be included in analysis, windows needed to have at least 3 common informational cytosines and a depth of 6 reads/cytosine. Windows that differed between genotypes by 10% CHH, 20% CHG, or 30% CG DNA methylation were identified as differentially methylated. Overlapping windows with differential methylation between genotypes were merged into differentially methylated regions. To increase the robustness of our conclusions, we added 2 data filtering steps. DNA methylation in the endosperm varies between maternal and paternal alleles and bisulfite sequencing is known to potentially enrich for methylated DNA [76]. Since we were examining the consequences of loss of NRPD1 in either parent, we could preferentially lose DNA methylation information from one set of alleles. This could lead to lower coverage of one set of parental alleles and lead to faulty measurements of DNA methylation. We therefore limited our analyses to genomic regions in which reads arising from the maternally inherited genome accounted for 67%+/− 15% of total DNA reads (based on the fact that 2/3 of the DNA in endosperm is maternally inherited). Next, we identified DMRs between the 2 replicates for each genotype to mark regions where DNA methylation was variable within the same genotype. These regions were excluded from further analysis.