Gene silencing by double-stranded RNA from C. elegans neurons reveals functional mosaicism of RNA interference

Abstract Delivery of double-stranded RNA (dsRNA) into animals can silence genes of matching sequence in diverse cell types through mechanisms that have been collectively called RNA interference. In the nematode Caenorhabditis elegans, dsRNA from multiple sources can trigger the amplification of silencing signals. Amplification occurs through the production of small RNAs by two RNA-dependent RNA polymerases (RdRPs) that are thought to be tissue-specific - EGO-1 in the germline and RRF-1 in somatic cells. Here we demonstrate that EGO-1 can compensate for the lack of RRF-1 when dsRNA from neurons is used to silence genes in intestinal cells. However, the lineal origins of cells that can use EGO-1 varies. This variability could be because random sets of cells can either receive different amounts of dsRNA from the same source or use different RdRPs to perform the same function. Variability is masked in wild-type animals, which show extensive silencing by neuronal dsRNA. As a result, cells appear similarly functional despite underlying differences that vary from animal to animal. This functional mosaicism cautions against inferring uniformity of mechanism based on uniformity of outcome. We speculate that functional mosaicism could contribute to escape from targeted therapies and could allow developmental systems to drift over evolutionary time.


INTRODUCTION
Animals have diverse cell types that perform specialized functions while retaining the ability to perform common functions. Such common functions could rely on the same molecular machinery in all cells or on different machinery in different cells. As a result, an apparently uniform organismal response could obscure differences in the mechanisms used by different cells. A common response to viral infection is the silencing of viral genes facilitated by the recognition of viral double-stranded RNA (dsRNA) (reviewed in (1)). The experimental addition of dsRNA triggers similar mechanisms that can silence any matching sequence (2). This process of RNA interference (RNAi) is a powerful approach for gene silencing applications in a variety of organisms (reviewed in (3)). In the nematode Caenorhabditis elegans, exposure to different sources of dsRNA can silence matching genes in many somatic cell types and in the germline (4)(5)(6). Studies in C. elegans have therefore been informative in piecing together the organismal response to RNAi in an animal. While similar silencing responses occur in diverse cell types, it is unclear whether dsRNA from every source engages the same molecular machinery in each cell.
Export of RNA from cells can result in diverse populations of RNA in the extracellular space (reviewed in (7)). In C. elegans, systematic expression of dsRNA in multiple tissues suggests that extracellular dsRNA from neurons reliably causes silencing in distant cells (8). The export of dsRNA could be the result of non-specific processes such as cellular damage or could require secretion mechanisms and specific processing of dsRNA by enzymes that splice (9,10), edit (reviewed in (11)), or modify (12) RNA. Collectively, these processes could potentially produce many forms of extracellular dsRNA.
Entry of extracellular dsRNA into the cytosol and subsequent silencing relies on the conserved dsRNA importer SID-1 (13)(14)(15)(16). SID-1-dependent silencing is observed in many tissues even when dsRNA is expressed within a single tissue, suggesting that form(s) of dsRNA move between cells. In particular, dsRNA expressed in neurons can silence a target gene in somatic tissues such as the intestine, muscle, and hypodermis (8,17,18) and in the germline (19). Silencing in these diverse target cells requires the dsRNAbinding protein RDE-4 (20,21) and the endonuclease DCR-1, which together process dsRNA into small-interfering RNAs (siRNAs) (22,23), and the Argonaute RDE-1, which binds siRNAs (24). Upon recognition of a matching mRNA by RDE-1-bound siRNAs, RNA-dependent RNA Polymerases (RdRPs) are recruited, resulting in the production of numerous secondary siRNAs (25,26). Testing multiple target genes suggests that two different RdRPs are used for silencing: RRF-1 for genes expressed in somatic cells (25)(26)(27) and EGO-1 for genes expressed in the germline (28,29). Secondary siRNAs can bind the Argonaute NRDE-3 in somatic cells (30) or the Argonaute HRDE-1 in the germline (31)(32)(33) and subsequently accumulate within the nuclei of cells that express the target gene. Through these events, extracellular dsRNA can reduce the levels of mRNA and/or pre-mRNA of a target gene.
While silencing by all extracellular dsRNA requires SID-1, DCR-1, and RDE-1, the requirement for other components can vary. For example, some genes expressed in somatic cells can be silenced by ingested dsRNA in the absence of RRF-1 (34). While many genes do not require NRDE-3 for silencing, the bli-1 gene requires NRDE-3 for silencing by ingested dsRNA or neuronal dsRNA (18). Finally, a strict requirement for NRDE-3 but not for RRF-1 is seen for the silencing of repetitive DNA that occurs in an enhanced RNAi background upon growth at lower temperatures (35). These observations suggest that a mix of mechanisms could underlie RNAi in C. elegans. Experiments that control one variable at a time are needed to elucidate features that dictate the choice of mechanism used for silencing.
Here we reveal that silencing by neuronal dsRNA can differ from silencing by other sources of dsRNA in its requirement for EGO-1 in the absence of RRF-1. We provide a single-cell resolution view of silencing by neuronal dsRNA and find that each animal has a different set of intestinal cells that can rely on EGO-1 for gene silencing.

Transgenesis
Caenorhabditis elegans was transformed with plasmids and/or PCR products using microinjection (37) to generate extrachromosomal or integrated arrays. pHC337 was used to express an inverted repeat of gfp in neurons (8), which is expected to generate a hairpin RNA (gfp-dsRNA). Generation of the array that expresses unc-22-dsRNA in neurons (qtEx136) was described earlier (17). To rescue silencing defects in rde-1(jam1) and rrf-1(jam3) animals (Supplementary Figure S2), genomic DNA from wild-type animals (N2 gDNA) was used as a template to generate fused promoter/gene products through overlap extension PCR using Expand Long Template polymerase (Roche) and PCR products were purified using QIAquick PCR Purification Kit (Qiagen). The plasmid pHC448 for DsRed2 expression in the pharynx or a PCR product, Prgef-1::DsRed2::unc-54 3 UTR, for DsRed2 expression in neurons was used as a co-injection marker (17). Additional details are provided in Supplementary Materials and Methods.

Genome editing
Synthetic CRISPR RNA (crRNA) and trans-activating cr-RNA (tracrRNA) (IDT) or single guide RNAs (sgRNA) transcribed in vitro were combined with Cas9 protein (PNA Bio Inc. or IDT) to generate complexes used for genome editing. To transcribe guide RNAs, the scaffold DNA sequence was amplified from pDD162 (Peft-3::Cas9 + dpy-10 sgRNA -Addgene plasmid # 47549, a gift from Bob Goldstein) (38) using a common reverse primer (P16) and targetspecific forward primers (see Supplementary Table S2), purified (PCR Purification Kit, Qiagen), and used for in vitro transcription (SP6 RNA polymerase, NEB). Deletions were made using two guide RNAs and a single-stranded DNA oligonucleotide repair template with a co-conversion strategy (39). Insertions of gfp were performed using a single guide RNA and a double-stranded repair template amplified using PCR (40). Punc-22::unc-22::gfp resulted in GFP fluorescence within the pharynx as reported earlier (41). Additional details are provided in Supplementary Materials and Methods.

Feeding RNAi
One generation of feeding RNAi was performed as described earlier (15) and the numbers of brightly fluorescent intestinal nuclei in animals subject to RNAi were counted for Figure 1D.
Genetic screen and whole genome sequencing AMJ1 animals were mutagenized with 25 mM N-ethyl N-nitrosourea (ENU, Toronto Research Chemicals) and ∼600,000 of their F2 progeny were screened for recovery of GFP expression in intestinal cells (performed by A.M.J. in Craig Hunter's lab, Harvard University). For 23 mutants that showed different degrees of fluorescence, we prepared genomic DNA from ∼1-2 ml of worms (200-800 ng/l of DNA per mutant, NanoVue Plus (GE)). Libraries for Illumina sequencing were prepared at the IBBR sequencing core as per manufacturer's instructions and sequenced using a HiSeq1000 (Illumina).

Bioinformatic analysis
All bioinformatic analyses were done using the web-based Galaxy tool collection (https://usegalaxy.org) (42)(43)(44). For each of the 23 mutant strains, we obtained ∼40 million 101 base fastq reads on average (Supplementary Table S3). One 5 -end base and three 3 -end bases were of lower quality and were trimmed from all reads before alignment to ce6/WS190 using Bowtie (∼36 million mapped reads per mutant on average). Sequence variants were filtered to call Nucleic Acids Research, 2019, Vol. 47, No. 19 10061 mutations (Phred33 ≥20, ≥2 aligned reads, and same variant call in ≥66% of reads). We intended to rescue any sid-1 mutations that might arise in the screen to avoid isolating many alleles of sid-1 (∼100 alleles of sid-1 were isolated in the original sid screen (13)). However, our sequencing data revealed that we had instead inadvertently introduced a non-functional copy with 12 missense mutations as part of the qtIs50 transgene (Supplementary Figure S1C). Therefore, the threshold for calling a mutation was reduced from 66% to 15% for sid-1 sequences. For all mutants, nonsynonymous changes, changes in splice junctions, and deletions (characterized by lower than average coverage) were analyzed further. Identical changes detected in two or more mutants were eliminated as potential background mutations that were likely present before mutagenesis. Pairwise comparisons were carried out between all mutants to identify cases of different mutations in the same gene (i.e. in silico complementation (45)). Because this process entails 253 pairwise comparisons, we expect that one or two such shared genes will be identified for some mutant pairs at random. For example, for mutant pairs with 30 mutated genes each, the P-value for one shared gene (0.044) and that for two shared genes (0.0009) are both larger than the Bonferroni corrected P-value of 0.0002 for 253 comparisons at ␣ = 0.05 (46).

Western blotting
Mixed stage animals were washed off three to five 100 mm plates and used for western blot analysis. Samples were sonicated four times (40% amplitude with 45 s pauses between 15 s pulses) using a probe sonicator with a microtip (Branson Sonifier). Proteins were separated on a 14% SDS-PAGE and then blotted onto nitrocellulose paper (TransBlot™ Turbo Midi transfer pack). The blot was probed for GFP first, stripped (incubated in 0.2% sodium dodecyl sulfate, 0.1 M Tris, pH 8.0 and 1.4% ␤-mercaptoethanol for 1 h at 65 • C), and then probed for Tubulin. The following primary antibodies were used: mouse anti-␣-Tubulin (Sigma: T5168; 1:4000 dilution) and mouse anti-GFP (Thermo Fisher Scientific: MA5-15256; 1:2000 dilution). The following corresponding secondary antibodies were used: Rabbit antimouse IgG1 HRP (Sigma: SAB3701171, 1:250 dilution) and goat anti-Mouse IgG(H+L) HRP (Thermo Fisher Scientific: 32430, 1:750 dilution). Blots were developed using chemiluminescence detection reagents (Thermo Fisher Scientific: SuperSignal™ West Pico PLUS) and imaged using a ChemiDoc (Bio-Rad). The western blots in Supplementary Figure S3D are representative of three technical replicates. Signal of the band of interest was quantified using FIJI (NIH, (49)) and is reported as median of ratios with respect to ␣-tubulin.

Microscopy
For Figures 2A and 3A and Supplementary Figures S3B,  S5, S6B, S6D and S7, animals were immobilized in 5 l of 3 mM levamisole (Sigma-Aldrich; catalog no. 196142), mounted on slides, and imaged using an AZ100 microscope (Nikon) at a fixed magnification under non-saturating conditions of the tissue being quantified for silencing. A C-HGFI Intensilight Hg Illuminator was used to excite GFP (filter cube: 450-490 nm excitation, 495 nm dichroic, 500-550 nm emission), which also resulted in some bleed through from the DsRed fluorescence (e.g. Figure 3A). For Figure 4A-D and Supplementary Figure S8, L4-staged worms were mounted onto a slide with a 3.5% agarose pad after incubating the worm for 10 minutes in 7l of 1mM freshly made levamisole. Extended exposure to levamisole was necessary for reliable immobilization of the worm for the ∼100 min of imaging that was required to obtain 512 × 512 images of entire L4-staged sur-5::gfp worms using a 63× lens in a Leica SP5X confocal microscope (average of 3 measurements per line, 319 slices per section, 5 sections, and 0.125 m between slices). A 488 nm laser was used to excite GFP (emission: 498-550 nm, NA = 1.4). For Figure 4E and F, DAPI, GFP and Quasar 670 fluorescence in intestinal cells anterior to the germline and posterior to the pharynx was acquired as 1024 × 1024 images (six slices, 0.5 m between slices) using a 63× lens and 2× digital zoom in a Leica SP5X confocal microscope. GFP was excited as described above, a 405 diode laser was used to excite DAPI (emission: 422-481 nm, 9% power) and a 633nm laser was used to excite Quasar 670 (emission: 650-715 nm, 50% power).

Image processing
All images being compared in a figure were adjusted identically using Adobe Photoshop and/or FIJI (NIH). Images taken on Nikon AZ100 were inverted (GFP = black), lookup tables were changed using Photoshop (190 = white and 255 = black for gtbp-1::gfp, eft-3::gfp, gfp::unc-22 and unc-22::gfp; 212 = white and 255 = black for sur-5::gfp), and cropped for display. When imaging using the SP5X confocal microscope ( Figure 4A-D and Supplementary Figure S8), our immobilization conditions resulted in the worm lying on the coverslip such that the middle of the worm (vulva region) was tightly sandwiched between the coverslip and the agarose pad but the rest of the worm (head and tail in particular) was free to assume different positions. To partially account for this variability and the observed loss in sensitivity with depth of imaging, stacks close to the coverslip that lacked any signal were removed (0-30 stacks, median 7 stacks) and an equivalent number of empty stacks were added beyond the worm for a consistent total of 319 stacks in all cases. For Figure 4A-D, Z-projections of the five stacks for each worm were stitched together using a combination of a pairwise stitching plugin (50) and manual alignment (Adobe Illustrator). For Figure 4A, each Z-stack was depth-coded using the 'temporal-color code' function in FIJI (6 colors with 53 stacks/color). For Supplementary Figure S8A, Z projections of maximum intensity were created using all 319 stacks (head and tail) or a subset of stacks (seam, uterus and vulva). For Figure 4E and F, Z projections of maximum intensity were created using five slices, inverted (GFP = black), cropped for display (full anterior region or zoomed-in region between two nuclei) and look-up tables were changed using Photoshop (160 = white and 255 = black). Composites of GFP, DAPI and Quasar 670 were created on FIJI (NIH) and look-up tables were changed to magenta, blue, or green.

Quantification of silencing
Silencing in response to unc-22-dsRNA was scored by calculating the percentage of L4-staged animals that twitched within 3 min in 3 mM levamisole. The silencing of GFP expressed from nrIs20 (sur-5::gfp) was determined by counting the number of intestinal nuclei that showed bright GFP fluorescence in L4-staged animals at a fixed magnification and zoom using a MVX10 stereomicroscope (Olympus). Average number of intestinal nuclei were determined by counting HC195 and was relatively constant in most genetic backgrounds with the exception of strains that lacked rrf-1 (e.g. 32.8 ± 0.6 nuclei in rrf-1(-); nrIs20 animals and 32.3 ± 0.8 nuclei in rrf-1(-); eri-1(-) nrIs20 animals, compared to 29.9 ± 1.2 nuclei in nrIs20 animals, errors indicate 95% CI). For images acquired using Nikon AZ100, silencing was quantified using FIJI (NIH) by measuring the fluorescence posterior to the pharynx in a region of interest (ROI) that included either a fixed area anterior to the germline ( Figure  2 and Supplementary Figure S3) or body-wall muscles all along the worm ( Figure 3 and Supplementary Figure S6), using the formula ROI fluorescence (arbitrary units) = intensity of ROI -(area of ROI × mean intensity of background). For images acquired using the SP5X confocal microscope, a combination of thresholding using the 3D object counter plugin (51) on FIJI (NIH) and manual verification was used to count various nuclei. To score nuclei as 'on' or 'off', different thresholds were used for intestinal nuclei located at different depths (70 for stacks 1-160; 20 for stacks 161-319) and a constant threshold was used for all other nuclei (20 for all stacks). For Figure 4E and F, the number of mRNA foci was counted using the 3D object counter on FIJI (NIH). A threshold of 50 was selected, objects <0.015 m 3 were eliminated as background, and objects >0.2 m 3 were eliminated as miscounts due to merging of multiple objects. For Figure 4G, the identity of each intestinal nucleus was inferred using its expected location and using the position of the vulva, anus, and the twisting rows of hypodermal cells (twist induced by the rol-6 co-injection marker for gfp-dsRNA [qtIs49]) as guideposts (52)(53)(54)(55)(56)(57).

Statistics
Significance of differences in silencing (P-value < 0.05, unless otherwise stated) were calculated using Student's t-test (two tailed) or a two-way analysis of variation (ANOVA) with replication. Error bars in Figure 3B, Right and Supplementary Figure S6A and S6C, Right indicate 95% confidence intervals for single proportions calculated using Wilson's estimates with a continuity correction (Method 4 in (58)) and significance of differences between strains was determined using pooled Wilson's estimates.

Silencing by neuronal dsRNA can be distinct from silencing by ingested or target-derived dsRNA
Double-stranded RNA can be introduced into C. elegans cells through the transcription of complementary sequences within the target cell, in a distant cell, or in ingested bacteria. While all these sources of dsRNA trigger RDE-1dependent gene silencing (59), each source could produce different pools of dsRNA and/or dsRNA-derivatives that are trafficked differently to the cytosol of the target cell where silencing occurs. Here we present evidence that different sources of dsRNA can differ in their requirement for RRF-1 to silence the same target gene.
This bypass of RRF-1 could be a feature of silencing by target-derived dsRNA, neuronal dsRNA, or a general feature of silencing by all sources of dsRNA. To determine RRF-1 requirements for silencing by different sources of dsRNA, we examined silencing by target-derived dsRNA using an eri-1(-) background, silencing by neuronal dsRNA in an eri-1(+) background, and silencing by ingested dsRNA in an eri-1(+) background. All three sources of dsRNA strictly required RDE-11, a dosagesensitive RNAi factor (61,62). In contrast, the requirement for RRF-1 varied depending on the source of dsRNA. The weak silencing by target-derived dsRNA was completely abolished in rrf-1 null mutants (Figure 1D orange). Equally weak silencing by neuronal dsRNA was not significantly altered in rrf-1 null mutants (Figure 1D blue). Yet, robust silencing by ingested dsRNA was strongly dependent on RRF-1 (Figure 1D black). These source-dependent differ-ences in extents of silencing could be caused by differences in the routes taken by dsRNA to reach the silencing machinery, the forms of dsRNA and/or the dosages of dsRNA. However, because weak silencing by neuronal dsRNA was partially independent of RRF-1, while strong silencing by ingested dsRNA was primarily dependent on RRF-1, a high dose of dsRNA from neurons cannot be the sole explanation for the observed RRF-1 independence. Therefore, these observations suggest that mechanisms engaged by ingested or target-derived dsRNA can differ from those engaged by neuronal dsRNA.

EGO-1 can compensate for lack of RRF-1
To determine if other targets could show silencing by neuronal dsRNA in the absence of RRF-1, we used the same source of neuronal dsRNA and examined silencing of GFP expression under the control of a different promoter introduced into different genomic loci. Silencing of gfp expressed under the control of the eft-3 promoter (Peft-3::gfp) from a single-copy transgene was partially independent of RRF-1 (Figure 2A). In the absence of RRF-1, a significant reduction in GFP fluorescence was detectable ( Figure 2B).
A similar extent of silencing in rrf-1(-) animals was observed using Peft-3::gfp transgenes located on three different chromosomes (Supplementary Figure S3A) and for a Cterminal gfp fusion of a ubiquitously expressed gene (Supplementary Figure S3B and Supplementary Figure S3C) generated using Cas9-based genome editing (Supplementary Figure S4). Thus, a measurable amount of silencing by neuronal dsRNA can occur in the absence of RRF-1 when gfp is expressed under the control of different promoters and from different chromosomes.
Although it is formally possible that neuronal dsRNA engages novel processing pathways that are not used by other sources of dsRNA, we found that additional components of canonical RNAi were required for silencing ( Figure 2B and Supplementary Figure S5). RDE-11, thought to facilitate the production of secondary siRNA (61,62), was required for most silencing ( Figure 2B). MUT-16, a poly-Q/N protein (63) and MUT-2/RDE-3, a putative nucleotidyltransferase (64), that together localize to perinuclear foci thought to be sites of secondary siRNA production (65,66), were both required for all observed silencing ( Figure 2B). Consistently, GFP protein levels in mut-16(-) animals were greater than that in rrf-1(-) animals ( Supplementary Figure S3D). Removal of MUT-16 in the rrf-1(-) background (Supplementary Figure S4) resulted in weaker silencing of this target (see persistent nuclear fluorescence in Supplementary Figure S5) and complete loss of silencing for another target (see below). These results suggest that silencing by neuronal dsRNA in the absence of RRF-1 either occurs through the action of primary siRNAs along with canonical factors such as RDE-11, MUT-16, and MUT-2/RDE-3, or through the production of secondary siRNAs using an alternative RdRP.
The C. elegans genome has four genes that encode proteins with RdRP domains, three of which have demonstrated roles in the production of RNA using RNA templates. RRF-3 is thought to act as a processive RdRP in an endogenous pathway (67, Supplementary Figure S9 in (68)) that competes with experimental RNAi for shared components (69) and therefore loss of rrf-3 enhances RNAi (70). RRF-1 and EGO-1 are thought to act as non-processive RdRPs that make siRNAs in the soma (25,26,69) and the germline (71), respectively. Preventing germline proliferation in rrf-1(-) animals was found to greatly reduce the levels of secondary siRNAs but not eliminate them (72), leaving open the possibility that the residual secondary siRNAs may be generated by an alternative RdRP. The fourth putative RdRP, RRF-2, was found to be not required for silencing by ingested dsRNA (27). To test if the silencing observed in the absence of RRF-1 depends on any of these other RdRPs, we generated mutants lacking RRF-2, RRF-3, or EGO-1 using Cas9-based genome editing (Supplementary Figure S4). In an rrf-1(-) background, loss of rrf-2 did not eliminate silencing and loss of rrf-3 resulted in enhancement of silencing ( Figure 2C and Supplementary Figure  S5). Evaluation of the loss of ego-1 is complicated by the sterility of ego-1(-) animals, reflecting the role of EGO-1 in germline development (28,29). However, ego-1(-) progeny of heterozygous animals lacked all silencing in the absence of rrf-1 despite the potential for parental rescue of ego-1 ( Figure 2C and Supplementary Figure S5), suggesting that EGO-1 made in progeny compensates for the absence of RRF-1. Hereafter, we shall refer to silencing in the absence of RRF-1 as silencing using EGO-1.
Taken together, these results reveal instances of silencing in somatic cells by a source of neuronal dsRNA through the use of two different RdRPs. silencing. Nevertheless, targeting gfp sequences using neuronal dsRNA resulted in silencing using EGO-1 in animals that lack rrf-1 (in Figures 1, 2, and Supplementary Figure  S3 using an integrated gfp-dsRNA source, and in 6/6 rrf-1(-); gtbp-1::gfp animals using an extrachromosomal gfp-dsRNA source). These results suggest that silencing in somatic cells using EGO-1 is not a generic property of all neuronal dsRNA and raise two possibilities: (1) sources that do not strictly require RRF-1 (e.g. neuronal gfp-dsRNA) differ from sources that require RRF-1 (e.g. neuronal unc-22-dsRNA); or (2) target sequences that do not strictly require RRF-1 (e.g. gfp) differ from target sequences that require RRF-1 (e.g. unc -22).

Context of target sequence can dictate RdRP usage
To examine silencing of a single target sequence by either source of dsRNA, we generated two chimeric genes (gfp::unc-22 or unc-22::gfp) that could both be silenced by either gfp-dsRNA or unc-22-dsRNA expressed in neurons ( Figure 3A, top left). Both chimeric genes express unc-22 and gfp sequences as a single transcript under the control of endogenous unc-22 regulatory sequences (Supplementary Figure S4) and were functional as evidenced by lack of twitching ( Figure 3B, right), which is a sensitive readout of reduction in unc-22 function (2). With either source of dsRNA, all measurable silencing required RRF-1 ( Figure  3, Supplementary Figures S6B and S6C). This complete dependence on RRF-1 was more evident when twitching was measured in response to the expression of either gfp-dsRNA or unc-22-dsRNA in neurons ( Figure 3B, right).
These results suggest that changing the context of a target sequence can change its need for RRF-1 versus the alternative use of EGO-1 for silencing by neuronal dsRNA. Specifically, silencing of the single-copy gfp target by neuronal gfp-dsRNA could use EGO-1 when gfp is present as part of eft-3::gfp or gtbp-1::gfp but not as part of unc-22::gfp. These differences in genomic location, associated regulatory elements, or site of expression could be responsible for the observed differential use of EGO-1.

Somatic cells that can use EGO-1 for silencing vary from animal to animal
To examine the use of EGO-1 for silencing in all somatic cells while keeping the genomic location and associated regulatory elements of the target gene constant, we generated a chimeric gene with gfp sequence fused to the endogenous sur-5 gene (sur-5::gfp, Supplementary Figure S4). This strain resulted in the expression of a nuclear-localized SUR-5::GFP fusion protein, enabling simultaneous visualization of every somatic nucleus using confocal microscopy ( Figure  4A). Expression of gfp-dsRNA in neurons resulted in silencing throughout the length of the animal that was entirely dependent on SID-1, consistent with silencing by neuronal dsRNA (Figure 4B and Figure 4C) and was not subject to silencing upon eri-1 loss by target-derived dsRNA (Supplementary Figure S7A-C) as is expected for a single-copy target (35). Silencing was easily detected in intestinal cells, hypodermal cells, body-wall muscle cells, and the excretory canal cell ( Figure 4B and Supplementary Figure S8A). Silencing was not detectable in some cells in the head, the vul-   Figure S8A). Interestingly, even neighboring, lineal sister cells sometimes showed very different extents of silencing (e.g. intestinal cells near the tail in Supplementary Figure S8A, top row). Nevertheless, the overall silencing observed was much more than that observed when the same source of dsRNA was used to silence a multi-copy sur-5::gfp transgene (1.0±0.4 visible intestinal nuclei for single-copy sur-5::gfp (Supplementary Figure S7C) versus 24.0±1.9 visible intestinal nuclei for multi-copy sur-5::gfp ( Figure 1A), P-value <10 −21 and errors indicate 95% CI). A simple explanation for this difference could be that silencing higher numbers of target transcripts requires higher amounts of dsRNA (see Discussion for additional possibilities). Thus, the single-copy sur-5::gfp gene is a sensitive target for evaluating the use of EGO-1 for silencing by neuronal dsRNA in somatic cells throughout the animal.
Silencing of single-copy sur-5::gfp by neuronal dsRNA was detectable in rrf-1(-) animals ( Figure 4D), but the extent of silencing and the locations of cells that showed silencing varied dramatically from animal to animal (Supplementary Figure S7D). To obtain a high-resolution view of silencing, we quantified silencing in multiple tissues by counting the number of nuclei that show fluorescence (Supplementary Movie S1). For quantifying silencing in hypodermal and body-wall muscle cells, we divided the body into three regions (Supplementary Figure S8B, Left): head (anterior to the posterior bulb of the pharynx), anterior body (anterior to the vulva), and posterior body (posterior to the vulva). In the head and anterior body, the average numbers of detectable nuclei in rrf-1(-) animals were not very different from the average numbers detectable in sid-1(-) animals (Supplementary Figure S8B, Right). The posterior body, however, showed marginal silencing of hypodermal and/or body-wall muscle cells in rrf-1(-) animals (50.0±7.6 nuclei versus 58.7±4.6 nuclei in sid-1(-), P-value = 0.08 and errors indicate 95%CI), suggestive of some use of EGO-1 for silencing. The intestine, however, showed obvious silencing in the absence of RRF-1. This silencing was associated with a reduction in mRNA levels ( Figure 4E and F) and required MUT-16 and EGO-1 ( Supplementary Figure S7E). Notably, loss of EGO-1 alone does not result in a detectable defect in silencing by neuronal dsRNA (Supplementary Figure S7F), suggesting that EGO-1 is not re-quired for silencing in any intestinal cell but rather can compensate for loss of RRF-1.
Because each of the 20 intestinal cells has an invariable lineal origin and position after morphogenesis ( Figure 4G, (52)(53)(54)(55)(56)(57)), we were able to examine whether silencing occurs in any discernible pattern correlated with lineage or position. Each tested worm had a different complement of cells with respect to RdRP use for silencing ( Figure 4G and Supplementary Figure S7D) such that no cell relied on only RRF-1 in every animal and no cell could use EGO-1 in every animal ( Figure 4G).
Together, these results show that neuronal dsRNA can cause robust silencing, but the particular cells that require RRF-1 for such silencing vary from animal to animal.

DISCUSSION
We examined RNA interference in the somatic cells of C. elegans and found that the source of extracellular dsRNA, the context of target sequences, and the identity of the tested cell can all dictate whether the RNA-dependent RNA polymerase RRF-1 is required for silencing. We discovered that silencing by neuronal dsRNA can be extensive and, when examined at single-cell resolution, different sets of cells rely on only RRF-1 or could also use EGO-1 in the absence of RRF-1 for silencing in each animal.

Silencing by neuronal dsRNA
Expression of dsRNA in all neuronal cells resulted in SID-1-dependent silencing in a variety of cell types throughout the animal (hypodermal cells, body-wall muscle cells, seam cells, intestinal cells, and excretory canal cell; Figure  4 and Supplementary Figure S8), suggesting that dsRNA molecules exported from neurons are widely available. Subsequent import depends on the levels of SID-1 in importing cells because cells that overexpress SID-1 can act as sinks for dsRNA and presumably reduce entry of dsRNA into other cells (Supplementary Figure S2 in (8,73)). The observed widespread silencing ( Figure 4) therefore suggests that no single tissue acts as a sink and that sufficient dsRNA is exported from neurons to reach cells throughout the animal.
Yet, silencing by neuronal dsRNA is not always detectable in all cells, which could reflect either inefficient im- port of dsRNA or inefficient silencing. For example, most intestinal cells were not silenced when neuronal dsRNA was used to silence a multi-copy sur-5::gfp transgene (Figure 1A). However, silencing of this multi-copy target was greatly enhanced upon loss of eri-1 ( Figure 1A), which releases shared factors used for endogenous RNAi-related processes (74). Therefore, this case of limited silencing by neuronal dsRNA likely reflects limited availability of such RNAi factors (e.g. RDE-4, DCR-1, etc.) and not poor access to dsRNA or poor import of dsRNA. Similarly, the lack of silencing of single-copy sur-5::gfp in the cells of some tissues (pharynx, vulva, and uterus, Supplementary Figure  S8A) could reflect inefficient silencing that could potentially be enhanced by providing limiting factors.
Genes required for the biogenesis and/or export of dsRNA from cells are currently unknown. Genetic screens that could have isolated mutations in such genes targeted fluorescent proteins expressed from repetitive transgenes for silencing by extracellular dsRNA ( (13,61), and this study). Because repetitive transgenes are themselves sources of dsRNA that can result in self-silencing (e.g. (8)) or in inhibition of silencing by extracellular dsRNA (18), the mechanism(s) of silencing disrupted in mutants from these screens are unclear. Nevertheless, these screens isolated genes required for import of dsRNA (sid-1 (13), sid-2 (75), sid-3 (76), sid-4 (77), sid-5 (78)), or for silencing within target cells (rde-10 (61,62), rde-11 (61,62)). We have constructed a screenable worm that could be used to isolate genes required for the biogenesis and/or export of dsRNA without confounding effects from repetitive transgenes expressed in target cells (Figure 4). A repetitive source of dsRNA from neurons, on the other hand, could be necessary for robust silencing of sur-5::gfp (Figure 4), although it does not guarantee robust silencing as evidenced by the weak silencing of bli-1 (Supplemental Figure 6A). Whether a single copy source of dsRNA can result in the export of dsRNA from neurons and efficient silencing of any target gene in distant cells is unclear.

Cellular origins of small RNAs
A wide range of endogenous small RNAs (miRNAs, siR-NAs, piRNAs etc.) is being analyzed by sequencing RNA from whole worms. Where any particular small RNA is made and where it acts are both obscured when worms are homogenized for extracting RNA. Base-paired RNAs such as long dsRNA (79), precursors of miRNAs (22,23) or precursors of 26G RNAs (68,69) could be transported through SID-1 such that they are made in one cell and cause effects in other cells. However, tests for such non-autonomous effects of the lin-4 miRNA suggest cell-autonomous action of this miRNA (80). Examination of some of the numerous anti-sense RNAs called 22G RNAs suggested that they are made by RRF-1 in somatic cells and both RRF-1 and EGO-1 in the germline (81). Our results suggest the possibility that some 22G RNAs could be made in the intestine in the absence of RRF-1 potentially using EGO-1 in the intestine or through indirect effects of EGO-1 function in the germline. Resolving the origin and the site of action of such an endogenous small RNA requires controlled exper-iments that consider both non-cell autonomy of the RNA and functional mosaicism of its biogenesis.

Functional mosaicism of RNAi in an animal
The identities of the intestinal cells that strictly require RRF-1 for silencing by neuronal dsRNA varied from animal to animal ( Figure 4G and Supplementary Figure S7). This variation observed in rrf-1(-) animals could be because of unequal and random availability of compensatory EGO-1 despite equal availability of neuronal dsRNA or because of unequal and random availability of neuronal dsRNA despite equal availability of compensatory EGO-1 (Supplementary Figure S9). Such functional mosaicism is masked in wild-type animals, where the amplification of silencing signals by RRF-1 results in uniform silencing. Thus, RRF-1 promotes silencing by extracellular dsRNA to ensure uniform silencing -a role that is reminiscent of the role for ERI-1 in opposing silencing by transgene-derived dsRNA to ensure uniform expression (35).
RNAi is an antiviral mechanism in many organisms (see (82) for a recent evolutionary perspective) and wild strains of C. elegans that are defective in RNAi can harbor viruses (83). Viral infection of C. elegans in the lab results in proliferation of the virus in some but not all intestinal cells (84). It would be interesting to determine whether mosaicism of specific components of the RNAi machinery underlies patterns of viral infection observed in the intestine of Caenorhabditis nematodes (83,84). We speculate that functional mosaicism and its control could be common in multicellular organisms because of the need to balance diversification of cell types with preservation of fundamental functions in all cells.

Functional mosaicism could enable escape from targeted therapies
Current examples of escape from therapeutic interventions could reflect unanticipated functional mosaicism -especially when such escape occurs in the absence of genetic mutations or overt differences. For example, bacterial cells can persist after treatments with antibiotics and the presence of such persister cells does not reflect genetic heterogeneity (85), but rather could reflect differences in underlying mechanisms among similar cells. Furthermore, while the genetic variation in cancers is well appreciated as a cause of resistance and relapse (86), the possible role of functional mosaicism as an additional contributor merits exploration.

Functional mosaicism could allow developmental systems to drift over evolutionary time
Analyses of variation in intact animals where organismal regulatory mechanisms are preserved, as described here using C. elegans, are an effective complement to analyses in single cells, which have begun to reveal heterogeneity in many processes (e.g. in gene expression (87), in membrane trafficking (88), and in subcellular organization (89)). This variation can be modified by the presence of maternal/zygotic factors (e.g. exonuclease ERI-1 (35)) or secreted factors (e.g. extracellular dsRNA, this study) Nucleic Acids Research, 2019, Vol. 47, No. 19 10069 that can act during development. Such modifiers of variation could allow diversification of underlying mechanisms in response to selection for the same function. As a result, functional mosaicism could persist without differences in phenotype. This hypothetical sequence of events supports the plausibility of mosaicism in a process existing in the ancestors of organisms with divergent developmental systems that nevertheless perform the same function (90). Consistently, evolutionary comparisons in nematodes and in arthropods suggest that transposons are silenced using a plurality of mechanisms that could have diverged from ancestors with multiple mechanisms (67,91). For example, efficient silencing can occur in the absence of RRF-1-like RdRPs (using RRF-3-like processive RdRPs (67)) or without any RdRPs (67,91). Evaluation of this hypothesis for any process requires analyses in closely related species at single-cell resolution.