Background

One of the most direct approaches to elucidating the role of any particular gene is to characterize its loss-of-function phenotype. Loss-of-function phenotypes have now been analyzed for almost all of the predicted genes of Saccharomyces cerevisiae [1], Caenorhabditis elegans [2], and Drosophila melanogaster [3], and there are ongoing efforts to make comprehensive collections of mouse knockouts. In all, this gives us an unprecedented level of insight into eukaryotic gene function. However, the loss-of-function phenotype of any individual gene is highly dependent on the genetic context; specifically, variations in the activities of other genes will affect this phenotype (for review [4]). If changes in the activity of one gene affect the loss-of-function phenotype of a second gene, then these two genes are said to interact genetically. Genetic interactions can be used to identify novel components of molecular pathways and can reveal the redundancy that underlies the robustness of genetic networks. Thus, although analyzing the loss-of-function phenotypes of all genes in a wild-type animal is a major advance, an understanding of how each phenotype is modulated by the activities of other genes will prove to be just as critical.

Recently, genetic interactions in S. cerevisiae were investigated in a systematic manner using matings within a comprehensive collection of mutant strains. Pair-wise matings have identified over 4500 genetic interactions, demonstrating the extensive degree of redundancy in yeast [5, 6]. However, this approach is not currently feasible in any animal. No complete collection of mutant strains exists, and even if such strains were all available, large-scale matings are far more laborious in animals than in yeast, and so alternative strategies are needed.

One underlying cause of genetic redundancy may be gene duplication. Duplicated genes that retain at least partially overlapping functions can confer robustness to mutation in the other copy [7, 8]. However, there is still much debate about whether redundancy of duplicated genes can be evolutionary selected [9–11]. Theoretical models have been proposed to explain the evolutionary stability of redundancy [12, 13], and indirect experimental evidence for the redundant functions of duplicated genes comes from the analysis of loss-of-function phenotypes of single genes; in both yeast and worms, inactivation of a duplicated gene is less likely to result in a nonviable phenotype than inactivation of a single copy gene [2, 14, 15]. However, there are strong biases in the types of genes that are duplicated in genomes, which complicates the interpretation of these results [16], and no attempt has yet been made to examine the extent of redundancy between duplicated genes in vivo directly and systematically.

RNA-mediated interference (RNAi) is a powerful tool for studying the loss-of-function phenotypes of genes. In particular, in C. elegans, RNAi by bacterial feeding has been used for genome-wide screens because it allows high-throughput (HTP) and low-cost analysis of the loss-of-function phenotypes of genes in vivo [2]. However, RNAi has only been used extensively to target single genes. To study genetic redundancy systematically and to identify genetic interactions using RNAi, it is critical to establish and validate robust methods for simultaneously targeting multiple genes by RNAi using bacterial feeding ('combinatorial RNAi'). In the present report we show that by using combinatorial RNAi by bacterial feeding we can identify the majority of a testset of previously described genetic interactions. We used this technique to provide the first large-scale analysis of the redundant functions of duplicated genes in any organism, and we found that many duplicate gene pairs can retain redundant functions for more than 80 million years of evolution.

Results

Effectiveness of combinatorial RNA-mediated interference

We sought to establish HTP methods for simultaneously targeting multiple genes in C. elegans using RNAi by bacterial feeding ('combinatorial RNAi') on a large scale. We recently developed HTP methods for using RNAi by feeding to target single genes (see Materials and methods, below); these assays allow us to identify the vast majority (>85%) of previously published nonviable RNAi phenotypes with high reproducibility (>90%) [17, 18]. We wished to determine whether we could adapt these methods, which are efficient for analyzing the RNAi phenotypes of single genes, to targeting multiple genes by combinatorial RNAi.

To investigate whether we could target effectively more than one gene in a single animal using bacterial-mediated RNAi, we used three tests. First, we assessed whether we could simultaneously target two independent genes, each with a known loss-of-function phenotype, and generate phenotypes for both genes in the same animal. For example, targeting lin-31 by RNAi generates multivulval worms, targeting sma-4 generates small worms, and targeting both would be expected to generate small worms with multiple vulvae if combinatorial RNAi is effective. We chose well characterized genes with non-overlapping phenotypes (Table 1) to ensure that we could investigate each phenotype independently. We examined all possible pair-wise combinations of our four test genes either in wild-type animals or in the RNAi-hypersensitive strain rrf-3 [19], and scored for the known RNAi phenotypes. We found that we could detect five of the five possible additive phenotypes in both wild-type and rrf-3 worms (Table 1; see Figure 1 for an example), demonstrating that it is feasible to target two genes in the same animal by bacterial-mediated RNAi. In addition to generating additive phenotypes, we found that the simultaneous targeting of sma-4 and lon-2 produced only small worms (the phenotype of sma-4 alone). Thus, we can use combinatorial RNAi to recapitulate a previously demonstrated epistatic relationship between SMADs and lon-2 [20]. Finally, although we could detect additive RNAi phenotypes in wild-type worms, we noted that the penetrance was often higher in the rrf-3 RNAi-hypersensitive strain, suggesting that this background might be more suitable for combinatorial RNAi; we examine this in more detail below.

Table 1 Combinatorial RNAi effectively generates additive phenotypes
Full size table
Figure 1
figure 1

Combinatorial RNA interference (RNAi) can target two genes in the same animal. Exposing worms to a mixture of two double-stranded (ds)RNA-expressing bacterial clones, one targeting lin-31 and the other one targeting sma-4, resulted in small worms with multiple vulvae along their ventral side. Shown are RNAi-hypersensitive rrf-3 animals [19] fed on bacteria expressing (a) a nontargeting dsRNA (control) and (b) combined bacterial clones expressing dsRNA against lin-31 and sma-4 (magnified in (c)). Pseudovulvae are indicated by white arrowheads.

Full size image

We next tested a set of known synthetic lethal interactions compiled from literature [21–25] (Table 2 and Figure 2). In rrf-3 animals, we were able to detect reproducibly all seven tested genetic interactions (Table 2 and Figure 2). However, in wild-type animals only five of these interactions could be recapitulated (Table 2). Not only did we fail to detect two out of seven interactions in wild-type worms, the five detected interactions were also weaker than in rrf-3, demonstrating that for effective combinatorial RNAi it is often essential to use RNAi-hypersensitive strains.

Table 2 Combinatorial RNAi can identify known synthetic lethal interactions
Full size table
Figure 2
figure 2

Combinatorial RNA interference (RNAi) can recapitulate known synthetic lethal interactions. To test whether combinatorial RNAi could recapitulate seven synthetic lethal interactions that were identified from literature (see Table 2 for references), brood size and embryonic survival measurements following co-targeting of both genes of a synthetic lethal pair (Observed Gene1 + 2) were compared with that following the targeting of each single gene alone (Gene1 or Gene2) and with the calculated product of the single gene brood sizes and embryonic survival measurements (Expected Gene1 + 2); this product represents the predicted outcome if the genetic interaction is purely additive. Values plotted represent the percentage of average wild-type brood size and embryonic survival rates, and are the arithmetic mean of two independent experiments performed in the RNAi-hypersensitive strain rrf-3 [19]. ***p < 1.0 × 10-02; *p < 5.0 × 10-02, by Student's t-test.

Full size image

Finally, we investigated whether we could use combinatorial RNAi to recapitulate known genetic interactions that result in post-embryonic phenotypes. To do this we focused on the well characterized synthetic multivulval (synMuv) genes [26–28]. The synMuv genes are organized into two redundant genetic pathways that are required for normal development of the hermaphrodite vulva. Inactivation of either a synMuv A pathway gene or a synMuv B pathway gene alone results in no vulval defect, but inactivation of both a synMuv A and a synMuv B gene in combination results in the multivulva (Muv) phenotype. Using combinatorial RNAi, we co-targeted three synMuv A genes with the canonical class B gene lin-15B, and co-targeted 12 synMuv B genes with the canonical synMuv A gene lin-15A in either wild-type or rrf-3 animals. In each experiment, we scored progeny for the multivulva phenotype; we expected to see this phenotype only if combinatorial RNAi targets both genes effectively in the same animal. We observed Muv worms for 13 out of 15 test cases in the RNAi-hypersensitive rrf-3 background, and for 8 out of 15 possible viable combinations in wild-type animals (Table 3).

Table 3 Genetic interactions of synthetic multivulval genes can be recapitulated by combinatorial RNAi
Full size table

Taken together these results demonstrate that combinatorial RNAi by feeding using our HTP platform works efficiently in rrf-3 animals; we were able to generate additive phenotypes and to detect the great majority of previously described genetic interactions.

Effect of dilution on phenotype strength

In analyzing the phenotypes produced through combinatorial RNAi, we and others [29, 30] observed that some of the single gene phenotypes were qualitatively weaker when two genes were targeted together than when each gene was targeted alone. Because such dilution effects will affect both the false negative rate in large-scale screens and the possible number of genes that can be co-targeted effectively, we wished to investigate the extent to which combining double-stranded (ds)RNA-expressing bacteria leads to reduced strength of RNAi phenotypes. To do this, we selected 282 genes from chromosome III that have a nonviable (embryonic lethal or sterile) RNAi phenotype [2] (Additional data file 1) and examined whether their phenotypes change as the targeting bacteria are diluted with increasing amounts of unrelated dsRNA-expressing bacteria (Figure 3).

Figure 3
figure 3

Effect of dilution on strength of RNA interference (RNAi) phenotype. The RNAi phenotype of each nonviable gene on chromosome III [2] was assessed following dilution with increasing amounts of bacteria expressing a nontargeting double-stranded (ds)RNA. The percentage of genes with phenotypes that are either identical to that observed when targeted alone (red) or weaker than when targeted alone (blue) is shown for each dilution. This was examined for three phenotypes: (a) all nonviable phenotypes, (b) complete sterility (no progeny), and (c) partial sterility (some progeny). (d) False negative rate (in percentage) of combinatorial RNAi at a given dilution. Data shown are representative of two independent experiments performed in the RNAi-hypersensitive rrf-3 background [19].

Full size image

We found that the strength of RNAi phenotypes for many genes was indeed reduced with increasing dilution of control bacteria (Figure 3). For example, we were able to detect phenotypes for about 90% of genes with nonviable RNAi phenotypes (Figure 3a) when the targeting strains were diluted with equal amounts of a bacterial strain expressing a control nontargeting dsRNA. This detection rate dropped further to about 70% at threefold and to about 60% at fourfold dilution (Additional data file 1). We found essentially identical results when we diluted with a dsRNA-expressing bacterial strain targeting lin-31 (data not shown), showing that the observed dilution effect appears not to be specific to the diluting dsRNA-expressing strain.

We next considered whether the effect of dilution on the observed phenotype was related to phenotypic strength. To this end, we determined the dilution behavior for genes that have different strengths of brood size defects when targeted alone (Figure 3b,c). We found that genes with weak RNAi phenotypes were indeed more likely to appear wild-type following dilution - and thus to be missed in screens - than were genes with strong, highly penetrant phenotypes. For example, we could still detect phenotypes for about 80% of genes that normally have a completely sterile phenotype at a fourfold dilution; however, only about 20% of genes conferring partial sterility (a reduction in brood size) had a detectable phenotype at this dilution. Although this indicates that genes with weaker phenotypes are more likely to appear wild-type when targeted in combination with other genes, we conclude that on average about 90% of genes with a detectable RNAi phenotype still have sufficient knockdown when diluted with equal amounts of a second dsRNA-expressing bacterial strain.

Overall, these experiments allow us to estimate the false-negative rates induced by dilution effects in combinatorial RNAi (Figure 3d; see Materials and methods for calculation). Assuming that each gene behaves independently, we expect that about 80% of bigenic interactions yielding visible RNAi phenotypes will be detectable by combinatorial RNAi. Because RNAi in rrf-3 recapitulates null phenotypes for about 70% of known genetic nulls, we thus estimate that combinatorial RNAi can detect about 50% of all bigenic interactions yielding nonviable phenotypes.

Investigating the redundancy of duplicated genes in C. elegans

Having validated combinatorial RNAi by using bacterial feeding as a method to inhibit simultaneously the expression of any pair-wise combination of genes, we wished to use this approach to investigate functional redundancy in the genome of C. elegans. One obvious possible cause of genetic redundancy is through gene duplication. Duplicated genes that have retained at least partially overlapping functions can confer robustness to mutation in the other copy [7, 8], and genome-wide loss-of-function screens provide indirect evidence that duplicated genes may often share redundant functions [2, 14, 15]. However, this hypothesis has never been directly tested with systematic experimental approaches.

We used the InParanoid algorithm [31] to identify 239 pairs of C. elegans genes that correspond to single orthologs in S. cerevisiae or D. melanogaster genomes (see Materials and methods, below). These genes have thus been duplicated in the genome of C. elegans since the divergence from either species. To determine whether there is functional redundancy between the duplicated genes, we compared the phenotype resulting from targeting both duplicated genes simultaneously by RNAi with the RNAi phenotype of each gene alone. We interpret a synthetic genetic interaction - that is, where the combined phenotype is greater than the product of the individual phenotypes [32] - as indicating redundancy. Of 143 duplicate gene pairs amenable to analysis by combinatorial RNAi (see Materials and methods, below; Additional data file 2), we found 16 pairs of duplicated genes to show reproducible synthetic RNAi phenotypes by quantitation (Table 4 and Figure 4), indicating that they are, at least in part, functionally redundant. Of these pairs only two have previously been identified as having redundant functions [33, 34]. The pairs of genes that when co-targeted give synthetic phenotypes encode diverse molecular functions, ranging from structural constituents of the ribosome (for example, rpa-2 + C37A2.7, rpl-25.1 + rpl-25.2), signaling proteins (for example, lin-12 + glp-1, C13G3.3 + W08G11.4), and transcription factors (for example, elt-6 + egl-18) to polyadenylate-binding proteins (for example, pab-1 + pab-2; Table 5).

Table 4 C. elegans duplicate gene pairs with at least partially redundant functions
Full size table
Table 5 Molecular functions of C. elegans duplicate gene pairs with synthetic phenotypes
Full size table
Figure 4
figure 4

Quantitative analysis of synthetic phenotypes following the simultaneous targeting of both genes of a duplicate pair. For duplicate gene pairs that yielded reproducible synthetic effects, phenotypes produced by combinatorial RNA interference (RNAi) were quantitated. For each gene pair, brood size and embryonic survival following co-targeting of both duplicates (Observed Gene1 + 2) were compared with that following the targeting of each single gene alone (Gene1 or Gene2) and with the calculated product of the single gene brood sizes and embryonic survival measurements (Expected Gene1 + 2). Values plotted represent the percentage of average wild-type brood size and embryonic survival rates, respectively, and are the arithmetic mean of two independent experiments performed in the RNAi-hypersensitive strain rrf-3 [19]. ***p < 1.0 × 10-02, *p < 5.0 × 10-02, by Student's t-test. Note that combinatorial RNAi against the gene pair ptr-2 + ptr-10 resulted in a significantly increased number (p = 7.4 × 10-09, by Student's t-test) of first-generation larval growth arrested worms, rather than a brood size defect, hence these data are not shown.

Full size image

The duplicated genes that we focused on in the worm correspond to single genes in either S. cerevisiae or D. melanogaster genomes. We wished to investigate whether the known function of the single yeast or fly gene was a good predictor of the RNAi phenotype identified by co-targeting the duplicated worm genes with redundant functions. If this were the case, then it is most likely that the redundancy that we observed is due to both duplicates retaining the ancestral function. Based on the gene deletion phenotypes of the single copy orthologs in yeast, we split our set of C. elegans duplicated genes into those corresponding to essential and to nonessential S. cerevisiae genes (Additional data file 2). We found that five out of 18 worm duplicates (28%) that are orthologous to yeast essential genes exhibited synthetic phenotypic effects by combinatorial RNAi. In contrast, only five out of 55 C. elegans duplicated genes (9%) that are orthologous to S. cerevisiae nonessential genes were found to produce a synthetic phenotype when co-targeted. We conclude that duplicated genes in C. elegans that are related to an essential gene in yeast are about three times more likely to have an essential redundant function than those related to a nonessential yeast gene. Strikingly, this is the same enrichment for nonviable RNAi phenotypes as for nonduplicated genes; 61% of C. elegans single copy orthologs of S. cerevisiae essential genes have nonviable RNAi phenotypes, as compared with 20% of orthologs of yeast nonessential genes (Additional data file 3). Thus, our finding is entirely consistent with a simple model of redundancy, suggesting that the function of a single gene identified in one organism is a good predictor of the redundant function covered by a pair of duplicated genes in a second organism.

Duplicated genes can maintain redundant functions for more than 80 million years of evolution

By using combinatorial RNAi we found that 11% of C. elegans duplicate gene pairs corresponding to single yeast or fly genes had synthetic phenotypes. These data clearly demonstrate that duplicated genes in metazoans often have at least partially redundant functions, but they do not address the underlying causes for this redundancy. Two simple models might explain why some duplicated genes appear to have redundant functions. First, the redundancy may represent a transient state resulting from a recent duplication event. In this model, the pairs of genes we found to be redundant are likely to be more recent duplicates than those for which we found no functional overlap. Alternatively, several groups have established population-genetic frameworks suggesting that redundant functions can be maintained by natural selection over substantial evolutionary times [12, 13]. In this case, we would expect no difference in age between the sets of duplicated genes for which we observed redundant phenotypes and gene pairs with no apparent redundant functions. Instead, we anticipated that there would be evidence that the redundant duplicated genes have been maintained relative to their ancestral sequence, thus retaining their overlapping, redundant functions.

Remarkably, 14 out of the 16 pairs of duplicated genes that we identified as having redundant essential functions in C. elegans were duplicated before the divergence from the related nematode C. briggsae (see Materials and methods, below; Additional data file 4). C. elegans and C. briggsae, despite being morphologically very similar, last shared a common ancestor 80-110 million years ago [35]. It is extremely unlikely that the redundancy between these 14 genes has been maintained for more than 80 million years of evolution merely as a consequence of the rate of neutral evolution, that is, that there has been insufficient evolutionary time for the duplicates to drift. To place this time period in the context of the rate of change of coding genes, C. elegans and C. briggsae only share about 60% of their genes as 1:1 orthologs, and a full 10% of genes encoded in either genome has no identifiable match in the other genome [35]. We thus considered the possibility that these 14 gene pairs retained redundant functions simply as a result of neutral evolution to be very unlikely; instead, these data suggest that the redundancy between these duplicated genes has been maintained over an extensive evolutionary period.

If there has been selection for the maintenance of redundancy between two duplicated genes, then we would expect these duplicates to encode more similar proteins than non-redundant duplicates. Indeed, we found that pairs of redundant duplicated genes are more similar to each other at the amino acid level (p = 1.6 × 10-02, by Wilcoxon rank sum test), have a greater similarity in alignable protein length (p = 2.2 × 10-02), and also exhibit a lower rate of nonsynonymous nucleotide substitution per nonsynonymous site (mean Ka for redundant duplicates = 0.34; mean Ka for non-redundant duplicates = 0.50; p = 3.8 × 10-02) than non-redundant duplicates (Additional data file 4). Using the rate of synonymous nucleotide substitutions (Ks) as a measure of the evolutionary age of gene duplicates, we found no evidence that the redundant genes represent more recent gene duplications (mean Ks = 13.41 for redundant duplicates, mean Ks = 9.48 for non-redundant duplicates; Additional data file 4). Thus, we believe that it is unlikely that this greater similarity is a trivial consequence of their having duplicated more recently. Rather, we suggest that the protein sequences of redundant gene pairs have been maintained relative to each other since duplication as the result of selective pressure to maintain their redundant functions.

Discussion

RNAi has emerged as a key technique for the analysis of the in vivo function of single genes in C. elegans. For the systematic identification of genetic interactions by RNAi, we have established and validated methods that allow us to study the loss-of-function RNAi phenotypes of any pair-wise combination of C. elegans genes in a high-throughput manner. We found that we can use this methodology to identify the great majority of a testset of previously known synthetic lethal and post-embryonic genetic interactions. This approach should therefore allow researchers to explore genetic interactions in the worm in a far more systematic manner than has been possible in the past.

We used our method to examine systematically the potentially redundant functions of duplicated genes in the genome of C. elegans, focusing on genes that correspond to single orthologs in S. cerevisiae or D. melanogaster. These genes have thus duplicated in the C. elegans genome since the divergence from either species. Of the 143 pairs of duplicate genes amenable to analysis by combinatorial RNAi, we identified 16 gene pairs that exhibited unambiguous synthetic RNAi phenotypes, demonstrating that they are at least partially functionally redundant. We found that just as single copy worm genes are more likely to have a nonviable RNAi phenotype if they are orthologous to an essential gene in yeast, duplicated worm genes are more likely to have a redundant essential function if they are co-orthologous to an essential yeast gene. It should therefore be possible to predict the redundant functions of many duplicated genes in higher organisms based on the functions of single copy orthologs in lower organisms.

Most intriguingly, the redundancy we observed between duplicated genes cannot simply be explained by a very recent duplication event; 14 of the 16 redundant gene pairs were duplicated before the divergence of C. elegans and C. briggsae 80-110 million years ago [35]. The redundancy between these 14 gene pairs has therefore been maintained for more than 80 million years of evolution. We believe that it is extremely unlikely that the functional overlap between these 14 duplicated genes is present merely due to the lack of evolutionary time since duplication. Not only is the average half-life of a gene duplicate in eukaryotes typically about 4 million years [11] but also, over this time period, the C. elegans and C. briggsae genomes have diverged greatly; they only share about 60% of their genes as 1:1 orthologs, and a further 10% of genes are present exclusively in one or other genome [35]. Rather, our findings are consistent with population genetic simulations that demonstrate that under appropriate (but realistic) conditions it is possible to select, directly or indirectly, for redundancy between duplicates to be maintained [12].

Conclusion

Our data provide the first systematic investigation into the redundancy of duplicated genes in any organism and strongly support models of gene evolution, which suggest that redundancy is not just a transient side effect of recent gene duplication but is instead a phenomenon that can be maintained over substantial periods of evolutionary time.

Materials and methods

Ninety-six-well liquid feeding assay

Selected bacterial strains of the C. elegans RNAi feeding library [2] were grown to saturation at 37°C in 96-well deep plates in 400 μl 2 × TY containing 100 μg/ml ampicillin. To induce dsRNA expression, 4 mmol/l IPTG (isopropyl-beta-D-thiogalactopyranoside) was added for 1 hour at 37°C before cultures were spun down at 3500 rpm for 5 min and finally resuspended in 400 μl of NGM (nematode growth medium) with 100 μg/ml ampicillin and 4 mmol/l IPTG. Finally, 10 (for wild-type N2) or 15 (for NL2099 rrf-3 [pk1426] II) L1-stage worms in 15 μl M9 buffer were aliquoted into each well of a 96-well flat-bottom plate and 40 μl of the resuspended bacterial cultures were added. For combinatorial RNAi feeding experiments, resuspended saturated cultures of different bacterial strains were mixed to give a final volume of 40 μl. Plates were incubated shaking at 150 rpm, 20°C, for 96 hours. Worms were scored for embryonic lethality, sterility, and growth defects using a dissecting microscope.

Testing additive RNAi phenotypes and known synthetic genetic interactions

To score post-embryonic phenotypes (Table 1 and Table 3), L1 larvae from the 96-well liquid feeding assay were collected after 96 hours and allowed to develop further on 12-well NGM plates. Cultures were filtered through a 11 μm nylon mesh (MultiScreen™ Nylon Mesh, Millipore Corporation, Bedford, MA, USA) and L1 larvae were spotted onto 12-well NGM plates containing 100 μg/ml ampicillin and 1 mmol/l IPTG, seeded with bacteria expressing a nontargeting dsRNA (Ahringer library clone Y95B8A_84.g). Adult worms were scored after further incubation at 20°C for 72 hours. Because we were assessing second generation (post-embryonic) phenotypes, we had to exclude genes that resulted in sterility, embryonic lethality, or larval growth arrest after RNAi. Only genes that were (according to the above criteria) amenable to analysis in both wild-type worms and the RNAi-hypersensitive rrf-3 background could be included in the study.

Estimation of false-negative rate of combinatorial RNAi

Assuming each gene is an independent targeting event in combinatorial RNAi, and having evaluated the average failure rate for the successful generation of a phenotypically detectable knockdown for single genes at a given dilution, we were able to estimate the false-negative rate of combinatorial RNAi for multigenic interactions. We calculated the detection rate of n-genic interactions to be xn, where x is the detection rate of single gene phenotypes at n-fold dilution.

Identification of duplicated genes and synthetic phenotypes

We used the InParanoid algorithm (version 4.0) to identify C. elegans orthologs of S. cerevisiae and D. melanogaster genes [31]. All genes that are targeted by bacterial clones from the C. elegans whole-genome RNAi library [2] with inserts having more than 80% nucleotide identity over 200 base pairs with multiple predicted genes were excluded from the analysis. This is the threshold for cross-reaction used by Kamath and coworkers [2]. Furthermore, genes that resulted in first-generation larval growth arrest after RNAi were not included in the study for synthetic interactions, because this strong phenotype cannot be enhanced any further.

When screening for phenotypic differences between single gene and combinatorial RNAi knockdowns, single gene phenotypes (as references) were compared with combinatorial phenotypes side by side. To account for dilution effects arising from combining two dsRNA-expressing bacteria, equal amounts of nontargeting dsRNA-expressing bacteria were added to bacteria expressing dsRNA targeting the reference genes. Screens for synthetic phenotypic effects were performed at least twice in triplicates within independent assays. For synthetic interactions to be scored positive, synthetic phenotypes had to be unambiguous and reproducible in at least two independent RNAi experiments.

Statistical analysis

For statistical analysis of the data, measurements of brood size and embryonic viability following RNAi were normalized to measurements obtained after RNAi against control genes that give no detectable phenotypes ('wild-type brood size' and 'wild-type embryonic survival'). To examine whether the combinatorial phenotypes were synthetic or merely additive, we compared the quantitative phenotypes following combinatorial RNAi with the calculated products of measurements for both individual genes of a pair. Duplicate brood size and embryonic survival measurements for two individual genes were multiplied to generate an array of 16 values that was compared with six measurements obtained for synthetic phenotypes, using a Student's t-test (two-tailed distribution, two-sample equal variance). In cases in which measurements for brood size and embryonic viability exceeded 100% of wild-type brood and viability, values were set to 100% of wild-type values.

Evolutionary analysis

We used InParanoid (version 4.0) to identify C. elegans orthologs of S. cerevisiae, D. melanogaster, and C. briggsae genes [31]. If both C. elegans duplicates had a single identifiable ortholog in C. briggsae, then this implies that the duplication predates the divergence of C. elegans from C. briggsae. Protein sequences were aligned using the CLUSTAL W program to determine the percentage of identity between gene duplicates [36].

Additional data files

The following additional data are included with the online version of this article: A Word document listing C. elegans chromosome III genes with a previously assigned nonviable (embryonic lethal or sterile) RNAi phenotype [2] and the effect of dilution following combinatorial RNAi (Additional data file 1); a Word document listing C. elegans pairs of duplicated genes that have been screened for synthetic RNAi phenotypes (Additional data file 2); a Word document listing C. elegans 1:1 orthologs of S. cerevisiae genes and their RNAi phenotypes (Additional data file 3); and a Word document presenting C. elegans duplicate gene pairs included in this study, their orthologous genes in C. briggsae, and levels of protein identity between C. elegans gene duplicates, as well as Ka and Ks values between duplicate gene pairs (Additional data file 4).