Introduction RNA interference (RNAi) is targeted gene silencing via double-stranded RNA (dsRNA); a gene is inactivated by specific breakdown of the mRNA (Fire et al. 1998; Montgomery et al. 1998). It is an ideal method for rapid identification of in vivo gene function. Initial studies on RNAi used microinjection to deliver dsRNA (Fire et al. 1998), but it was subsequently shown that dsRNA can be introduced very easily by feeding worms with bacteria that express dsRNA (Timmons and Fire 1998). Using this technique on a global scale, an RNAi feeding library consisting of 16,757 bacterial clones that correspond to 87% of the predicted genes in Caenorhabditis elegans was constructed (Fraser et al. 2000; Kamath et al. 2003). Upon feeding to worms, these clones will give transient loss-of-function phenotypes for many genes by inactivating the target genes via RNAi. By feeding the clones in this library to wild-type Bristol N2 worms, loss-of-function phenotypes were assigned to about 10% of genes. However, RNAi phenotypes were missed for about 30% of essential genes and 60% of genes required for postembryonic development, probably because RNAi is not completely effective (Kamath et al. 2003). Other global RNAi screens have been recently performed in C. elegans using this RNAi library or other techniques (Gönczy et al. 2000; Maeda et al. 2001; Dillin et al. 2002; Piano et al. 2002; Ashrafi et al. 2003; Lee et al. 2003; Pothof et al. 2003). These screens were done using wild-type worms. We have already shown that mutation of rrf-3, a putative RNA-directed RNA polymerase (RdRP), resulted in increased sensitivity to RNAi (Sijen et al. 2001; Simmer et al. 2002). There are four RdRP-like genes in C. elegans. Two of these, ego-1 and rrf-1, are required for efficient RNAi, as apparent from the fact that these mutants are resistant to RNAi against germline or somatically expressed genes, respectively (Smardon et al. 2000; Sijen et al. 2001). A third gene, rrf-2, appears to have no role in RNAi. The rrf-3 strain, mutated in the fourth RdRP homolog, shows an opposite response to dsRNA; this mutant has increased sensitivity to RNAi (Sijen et al. 2001). A more detailed study of RNAi sensitivity of rrf-3 mutants using a set of 80 genes showed that rrf-3 is generally more sensitive to RNAi than wild-type worms (Simmer et al. 2002). RNAi phenotypes in rrf-3 animals are often stronger, and they more closely approximate a null phenotype, when compared to wild-type. In addition, loss-of-function RNAi phenotypes were detected for a number of genes using rrf-3 that were missed in a wild-type background. For example, known phenotypes were detected for many more neuronally expressed genes in the rrf-3 background. These features suggest that the rrf-3 strain could be used to improve and extend functional information associated with C. elegans genes. We have conducted a genome-wide RNAi screen using the rrf-3 strain. In total, we found reproducible RNAi phenotypes for 423 clones that previously did not induce a phenotype (corresponding to 393 additional genes). To explore the variability of global RNAi screens, we performed the rrf-3 screen twice for Chromosome I and carried out a Chromosome I screen with wild-type. These were cross-compared and also compared to the results of the wild-type screen of Fraser et al. (2000). From this, we find that rrf-3 consistently allowed detection of more phenotypes than wild-type. In addition, we found that there is a significant screen-to-screenvariability (10%–30%). Results Comparative Analysis of RNAi for Chromosome I with Wild-Type and rrf-3 We first conducted a pilot screen of Chromosome I using rrf-3 and found RNAi phenotypes for 456 bacterial clones. We compared these data to those obtained by Fraser et al. (2000) for a screen in the wild-type Bristol N2 strain. For 153 of these 456 clones, no phenotypes were reported by Fraser et al. (2000) and phenotypes were observed for 303 clones in both screens. The N2 screen done by Fraser et al. (2000) resulted in RNAi phenotypes for 40 clones for which no phenotypes were found using rrf-3 (Figure 1A). These results indicate that rrf-3 can be used in a global screen to identify loss-of-function phenotypes for additional genes. However, some phenotypes were missed in the rrf-3 screen. To explore the reproducibility and variability of RNAi screens, we next screened the clones of Chromosome I using N2 and rrf-3 side by side. We detected phenotypes for 447 clones: 140 were found only in rrf-3, 11 only in N2, and 296 in both strains (Figure 1B). These data confirm that rrf-3 is more sensitive to RNAi and, in addition, these data indicate that global RNAi screens with rrf-3 will result in more clones with a detectable phenotype. Variability of the RNAi Effect When we compared the RNAi results that we obtained using N2 with the Fraser et al. (2000) data, we were surprised to find significant differences: we only detected phenotypes for 75% of the clones that gave a phenotype in Fraser et al. (2000), and these researchers reported phenotypes for 84% of clones for which we found a phenotype (Figure 1C). The differences do not appear to be due to false positives. For example, Fraser et al. (2000) detected the predicted phenotype for goa-1 and unc-73, whereas we did not detect a mutant phenotype. Similarly, we detected the known mutant phenotype for egl-30 and cdc-25.1, which were not detected by Fraser et al. (2000). In addition, we found that the false-positive rate is negligible (see below). It is possible that different laboratories or investigators have slightly different results. However, when we compare the results that we obtained with two independent screens of Chromosome I using rrf-3 in our laboratory, we also see differences. For 394 clones we detected a phenotype in both experiments, 54 are specific for the first experiment, and 34 for the second (Figure 1D). Among the clones that only gave an RNAi phenotype in one of the experiments are again clones that induced the predicted phenotype based on the phenotypes of genetic mutants (unc-40, gpc-2, and sur-2). These data show that large-scale RNAi screens done within the same laboratory and by the same investigators also give variable results. A few examples of variable RNAi results are shown in Table 1. In conclusion, we find that RNAi results from different laboratories and from experiments done in the same laboratory vary from 10% to 30%. This appears to be due to a high frequency of false negatives in each RNAi screen, even when the same method is used in the same laboratory. The Genome-Wide RNAi Screen Based on the positive results of the Chromosome I screen using the rrf-3 strain, we next screened the complete RNAi library with rrf-3 mutant animals. We obtained results for 16,401 clones and detected phenotypes for 2,079 (12.7%). Of these, we identified phenotypes for 625 clones for which no phenotype was reported in the Fraser et al. (2000) or Kamath et al. (2003) screens using N2, with the remaining 1,454generating phenotypes in both screens (Table S1). In addition, there are 287 clones for which only Fraser et al. (2000) or Kamath et al. (2003) found phenotypes (23 of these were not done in our screen). The clones for which we only detected an RNAi phenotype once and that were specific for the rrf-3 screen were retested. Subsequently, the phenotypes of the clones corresponding to Chromosomes II to X that were not confirmed by this repetition were tested once more. In this way, the clones specific for the rrf-3 screen had two chances to be confirmed. Of the 625 clones for which no phenotype was found in the Fraser et al. (2000) and Kamath et al. (2003) N2 screens, the phenotypes of 423 clones were confirmed and 202 remained unconfirmed (Table 2; see Table S1). Combining the N2 screens and these 423 clones, the percentage of clones with a phenotype increases from 10.3% to 12.8%. Some of the RNAi phenotypes only found with rrf-3 that remained unconfirmed could be confirmed by RNAi phenotypes detected with other clones of the RNAi library corresponding to the same gene or by other laboratories using different RNAi methods. For example, for the clones corresponding to the predicted genes F56D1.1 (a member of the zinc finger C2H2-type protein family) and F27C8.6 (a member of the esterase-like protein family), we detected sterile progeny (Stp) and embryonic lethality (Emb), respectively; these were also found by Piano et al. (2002). In addition, some unconfirmed RNAi phenotypes are confirmed by comparing to phenotypes of genetic mutants such as gpc-2, hlh-8, and unc-84. This suggests that many of the unconfirmed phenotypes reflect true gene functions. Analysis of the rrf-3 Results To validate the results obtained using rrf-3, we first assayed the rate of false positives in the total dataset (all RNAi results obtained with rrf-3 for the 16,401 clones tested). In the assay used by Kamath et al. (2003), a set of genes for which it is known that genetic mutants display no lethality was selected. A false positive in the RNAi data is then defined as detecting a lethal RNAi phenotype for any of these genes. In the N2 screen, the false-positive rate was 0.4%. We find that the false-positive rate in the rrf-3 data is similarly low (0 of 152 genes). To further determine the effectiveness of the screen, we compared the RNAi phenotypes with loss-of-function phenotypes of genetic mutants. For all chromosomes except for Chromosome I, the rrf-3 data were confirmed by refeeding only if there was no phenotype detected in the N2 screens by Fraser et al. (2000) or Kamath et al. (2003). Therefore, to compare the difference in detection of known phenotypes between the rrf-3 and the N2 screens, we used the Chromosome I datasets, where phenotypes were confirmed independently for the two strains. Of 75 genetic loci on Chromosome I, Fraser et al. (2000) detected 48% of published phenotypes, compared to 59% for rrf-3 (Table S2). Using the genome-wide rrf-3 dataset (excluding the 202 unconfirmed phenotypes), we detected the published phenotype for 54% of 397 selected loci, compared to 52% for N2 (Table 3; see Table S2). We next asked whether using the rrf-3 strain improved general phenotype detection or whether certain types of phenotypes were particularly increased compared to the N2 screens by Fraser et al. (2000) and Kamath et al. (2003). To do this, we analysed the detection rate of different types of Chromosome I loci. First, we looked at a set of 23 loci with nonlethal postembryonic mutant phenotypes. Using rrf-3, we reproducibly detected the published phenotype for 11 of these compared to only two for N2. Of 50 loci required for viability (essential genes), we detected 31 using rrf-3, compared to 33 for N2. Thus, detection of essential genes was similar in the two strains, but detection of postembryonic phenotypes was improved with rrf-3. Finally, for the whole genome using rrf-3, we reproducibly detected the published phenotypes for 34 genetic mutants for which no RNAi phenotype was reported in the N2 screens (nine essential genes, 21 with postembryonic mutant phenotypes, and four with a slow-growth mutant phenotype). By comparison, published phenotypes were detected for 23 loci only with N2 (16 essential genes and seven with postembryonic mutant phenotypes) (see Table S2). We conclude that rrf-3 particularly improves detection of genes with postembryonic mutant phenotypes, a class that is poorly detected using wild-type N2. A striking feature of the rrf-3 dataset is the high number of clones where a slow or arrested growth (Gro/Lva) defect was induced, without associated embryonic lethality or sterility. Overall, 619 clones induced a Gro/Lva defect using rrf-3, compared to 276 for N2, whereas the number of essential genes detected was similar (1,040 versus 1,170, respectively). In addition, in the confirmed set of 423 clones with rrf-3-specific phenotypes, Gro/Lva defects are the largest category (42%), whereas this is only 18% for N2, with the largest category being essential genes (49%). These data suggest that rrf-3 might particularly enhance detection of genes that mutate to a slow-growth phenotype; we cannot easily test this hypothesis, as there are currently few known loci with this mutant phenotype. In some cases, a Gro/Lva phenotype was seen in rrf-3, whereas a different phenotype was seen in N2 (e.g., either lethality or a weak postembryonic phenotype). This suggests that some of the Gro/Lva phenotypes detected are due to incomplete RNAi of an essential gene (where lethality was seen in N2) or by a stronger RNAi effect (where no growth defect was seen in N2). In addition, it is possible that some of the Gro/Lva phenotypes detected are synthetic effects of using the rrf-3 mutant strain. To summarise, using the rrf-3 RNAi supersensitive strain in large-scale screens increases the percentage of clones for which it is possible to detect a phenotype. Detection of postembryonic phenotypes is particularly increased, whereas detection of essential genes is similar in rrf-3 and N2. In addition, using rrf-3, there is a high rate of induction of Gro/Lva defects. Positional Cloning of Genetic Mutants with Visible Phenotypes Despite the advantages of RNAi, genetic mutants remain indispensable for many experiments. In the past decades, forward genetic screens identified a large number of genetic mutants, many of which are not yet linked to the physical map. We used the RNAi phenotypes obtained with the genome-wide screens to test whether we could systematically clone genes that are mutated in existing genetic mutants. First, the genetic map positions of all uncloned genetic mutants with visible phenotypes were checked using WormBase (http://www.wormbase.org, the Internet site for the genetics, genomics, and biology of C. elegans). Second, we searched for clones near the defined map positions that, when fed to N2, rrf-3, or both, gave phenotypes corresponding to the phenotypes of the genetic mutants. For most genetic mutants, more than ten clones with a similar phenotype were found in the interval to which the genetic mutant was mapped. However, for 21 genetic mutants, only one or a few candidate clones were found. The genes corresponding to these clones were subsequently sequenced in the genetic mutant to determine whether a mutation was present. In total, we sequenced 42 predicted genes for the 21 genetic mutants (Table S3). For seven of these—bli-3, bli-5, dpy-4, dpy-6, dpy-9, rol-3, and unc-108—we found a mutation in one of the sequenced genes (Table 4). The mutated gene was confirmed by sequencing the same gene in a second or third allele (or both) of these genetic mutants (Table 4). The identification of mutations in unc-108 encoding the homolog of the small GTPase Rab2 is of particular interest. The RNAi phenotype of this gene gives a clue about the genetic property of the mutations in the mutants of unc-108. With rrf-3, we find that inactivation of Rab2 (F53F10.4) by RNAi causes uncoordinated movement (Table 4). Mutations in unc-108 were isolated in a screen for dominant effects on behaviour; heterozygous unc-108 mutants display dominant movement defects and are indistinguishable from homozygous mutants (Park and Horvitz 1986). RNAi phenocopies a loss-of-function phenotype, suggesting that the dominant movement defects of unc-108 mutants may be due to haplo-insufficiency. In eukaryotes, Rab2 is involved in regulating vesicular trafficking between the endoplasmic reticulum and Golgi. Based on the movement defects of unc-108 mutants, UNC-108 might be involved in vesicle transport in neurons that regulate locomotion. Thus, the RNAi data are a powerful tool to facilitate rapid cloning of the genes identified by genetic mutants and will provide important starting points for further studies of their function. Discussion With this genome-wide RNAi screen using the hypersensitive strain rrf-3, we have significantly increased the functional information on the C. elegans genome, and we confirmed many RNAi phenotypes observed previously. We have assigned RNAi phenotypes for 406 genes (corresponding to the 423 extra clones) using rrf-3. For 13 genes, Kamath et al. (2003) or Fraser et al. (2000) had already found a phenotype using a different clone from the RNAi library that targeted the same gene, and for at least 44 genes a genetic mutant exists (see Table S2). Other investigators have also found RNAi phenotypes for some of the genes using different methods. However, for most genes our result is to our knowledge the first hint about their biological function. Although we have identified new RNAi phenotypes for a substantial number of genes, others will have been missed in our screen for the following reasons. First, besides its increased sensitivity to RNAi, the rrf-3 strain has an increased incidence of males (Him) and displays slightly increased embryonic lethality and a reduced brood size (Simmer et al. 2002). In our rrf-3 experiments, we therefore made some minor adaptations to the original RNAi protocol described by Fraser et al. (2000). We did not score for the Him phenotype and had more stringent criteria for embryonic lethality and sterility. This may have reduced the number of extra clones identified with a phenotype. Moreover, the changes in the protocol can also account for some differences in the detection of RNAi phenotypes between rrf-3 and N2. Second, when an RNAi phenotype is detected with N2 and not with rrf-3, the lack of a detectable phenotype may be the result of variability in the efficiency of RNAi. This is consistent with the fact that we observe differences between experiments done with the same strain. When an RNAi phenotype is detected with rrf-3 and not with N2, this can be due to the increased sensitivity to RNAi of rrf-3. However, besides the higher sensitivity, we may also be observing synthetic effects with rrf-3 (e.g., embryonic lethality, sterility, or developmental delay). In particular, a large number of clones induced a developmental delay phenotype using rrf-3. Synthetic effects cannot be excluded without investigating genetic mutants. Again, variability in the efficiency of RNAi will also contribute to these differences, and a small portion may be false positives. In general, the few false positives that occur in the screen are most likely due to experimental errors, whereas the false negatives are due to reduced efficiency of the RNAi. Finally, differences between rrf-3 and N2 do not only involve the absence and presence of an RNAi phenotype, but also differences in the phenotypes for clones that did induce phenotypes in both screens (e.g., embryonic lethal in one screen and a postembryonic phenotype in the other). For example, we detected for unc-112 a 100% embryonic lethal (Emb) phenotype with rrf-3, whereas Kamath et al. (2003) detected an adult lethal (Adl), uncoordinated (Unc), and paralyzed (Prz) phenotype with N2. Conversely, Kamath et al. (2003) detected for gon-1 a 100% Emb phenotype and other phenotypes with N2, while we did not detect an Emb phenotype with rrf-3. What could be the source of the interexperimental variation of RNAi? Different phenotypes for the same gene can possibly occur owing to slight differences in the developmental stage at which the animals are exposed to dsRNA and owing to changes in temperature during the experiment. However, this probably does not account for the differences we see, as we always used animals of the same larval stage (L3/L4) and used incubators for constant temperature. It was shown previously that the level of induction of dsRNA production by isopropylthio-β-D-galactoside (IPTG) can modify the penetrance of the RNAi phenotype (Kamath et al. 2000). Therefore, differences in the induction of the dsRNA either by changes in the concentration of IPTG, temperature, timing, or the bacteria may be an important source of the variation in the outcome of RNAi. RNAi is starting to be used extensively in other systems experimentally, as well as therapeutically and agriculturally. The relative variability of the RNAi effect is an important fact to take in account also for the use of RNAi in other systems. The RNAi data can be a useful starting point for many new experiments, such as positional cloning of genetic mutants. By sequencing candidate genes based on the RNAi phenotypes, we identified the causal mutation in seven genetic mutants. Identification of these mutated genes gives insight into the biological process in which they are involved. In addition, cloning of these genes increases the resolution of the genetic map of C. elegans, since these mutants have been extensively used as visible markers in linkage studies. The complete set of RNAi phenotypes detected for the 2,079 clones using rrf-3 will be submitted to WormBase, annotated as confirmed or unconfirmed. There the data can be evaluated in the context of information on gene structure, expression profiles, and other RNAi results. Materials and Methods Nematode strains. We used the following C. elegans strains: Bristol N2, NL4256 rrf-3(pk1426), CB767 bli-3(e767), MT1141 bli-3(n259), CB518 bli-5(e518), BC649 bli-5(s277), CB1158 dpy-4(e1158), CB1166 dpy-4(e1166), CB14 dpy-6(e14), CB4452, dpy-6(e2762), F11 dpy-6(f11), CB12 dpy-9(e12), CB1164 dpy-9(e1164), BC119 dpy-24(s71), CB3497 dpy-25(e817), MT1222 egl-6(n592), MT1179 egl-14(n549), MT1067 egl-31(n472), MT151 egl-33(n151), MT171 egl-34(n171), egl-34(e1452), MQ210 mau-4(qm45), CB754 rol-3(e754), BC3134 srl-2(s2507dpy-18(e364); unc-46(e177)rol-3(s1040), CB713 unc-67(e713), CB950 unc-75 (e950), HE177 unc-94(su177), HE33 unc-95(su33), HE151 unc-96(su151), unc-96(r291), HE115 unc-100(su115), MT1093 unc-108(n501), and MT1656 unc-108(n777). RNAi by feeding. RNAi was performed as described elsewhere (Fraser et al. 2000; Kamath et al. 2000) with minor adaptations when the rrf-3 strain was used: after transferring L3- to L4-staged hermaphrodites onto the first plate, we left them for 48 h at 15°C instead of 72 h and then plated single adults onto other plates seeded with the same bacteria. Furthermore, we did not remove the mothers from the second plates. The phenotypes assayed are these: Emb (embryonic lethal), Ste (sterile), Stp (sterile progeny), Brd (low broodsize), Gro (slow postembryonic growth), Lva (larval arrest), Lvl (larval lethality), Adl (adult lethal), Bli (blistering of cuticle), Bmd (body morphological defects), Clr (clear), Dpy (dumpy), Egl (egg-laying defective), Lon (long), Mlt (molt defects), Muv (multivulva), Prz (paralyzed), Pvl (protruding vulva), Rol (roller), Rup (ruptured), Sck (sick), Unc (uncoordinated) Thin and Pale. Emb was defined as greater than 10% dead embryos for N2 and greater than 30% dead embryos for rrf-3. Ste required a brood size of fewer than ten among fed N2 worms and fewer than five among rrf-3. Each postembryonic phenotype was required to be present among at least 10% of the analysed worms. Sequencing of genetic mutants. The coding sequence and the 5′- and 3′-untranslated region (about 500 bp upstream and downstream of the coding sequence) of the predicted genes, as annotated in WormBase, was analysed for mutations by sequencing amplified genomic DNA of the genetic mutants (see Table S3). Nested primers were designed using a modification of the Primer3 program available on our website (http://primers.niob.knaw.nl/). Sequence reactions were done using the ABI PRISM Big Dye terminator sequencing kit (Applied Biosystems, Foster City, California, United States) and were analysed on the ABI 3700 DNA analyser. Sequences were compared to the genomic sequence of C. elegans using the BLAST program (http://www.sanger.ac.uk/Projects/C_elegans/blast_server.shtml) or analysed using the PolyPhred program (available from http://droog.mbt.washington.edu/PolyPhred.html). Supporting Information Table S1 RNAi Phenotypes for Bacterial Clones Using rrf-3 (482 KB PDF). Click here for additional data file. Table S2 Detailed Comparison of RNAi Phenotypes with Those of Known Loci (188 KB PDF). Click here for additional data file. Table S3 Summary of Genes Sequenced in Several Genetic Mutants (25 KB DOC). Click here for additional data file. Accession Numbers RNAi data from this study will be submitted to WormBase (http://www.wormbase.org).