Understanding the genetic architecture of polygenic traits requires investigating how complex networks of interacting molecules mediate the effect of genetic variation on organismal phenotypes. We used a combination of P-element mutagenesis and analysis of natural variation in gene expression to predict transcriptional networks that underlie alcohol sensitivity in Drosophila melanogaster. We identified 139 unique P-element mutations (124 in genes) that affect sensitivity or resistance to alcohol exposure. Further analyses of nine of the lines showed that the P-elements affected expression levels of the tagged genes, and P-element excision resulted in phenotypic reversion. The majority of the mutations were in computationally predicted genes or genes with unexpected effects on alcohol phenotypes. Therefore we sought to understand the biological relationships among 21 of these genes by leveraging genetic correlations among genetically variable transcripts in wild-derived inbred lines to predict coregulated transcriptional networks. A total of 32 “hub” genes were common to two or more networks associated with the focal genes. We used RNAi-mediated inhibition of expression of focal genes and of hub genes connected to them in the network to confirm their effects on alcohol-related phenotypes. We then expanded the computational networks using the hub genes as foci and again validated network predictions. Iteration of this approach allows a stepwise expansion of the network with simultaneous functional validation. Although coregulated transcriptional networks do not provide information about causal relationships among their constituent transcripts, they provide a framework for subsequent functional studies on the genetic basis of alcohol sensitivity.
KNOWLEDGE of the genetic basis of phenotypic variation is critical for predicting disease risk and individual therapeutic treatments in human populations, understanding processes maintaining genetic variation in natural populations, predicting adaptive evolution, and selective improvement of agriculturally important species. Most phenotypic variation is genetically complex, attributable to multiple segregating quantitative trait loci (QTL) with small, environmentally sensitive, and often context-dependent allelic effects (Falconer and Mackay 1996; Lynch and Walsh 1998). Many recent large genome-wide studies mapping QTL affecting diseases and complex traits in human populations have identified novel loci and led to new biological insights about the biology of disease, but taken together these variants only explain a small fraction of the total segregating variation (Altshuler et al. 2008; Donnelly 2008; Frazer et al. 2009; Manolio 2010). Genetic dissection of complex traits in model genetic organisms, in which mutagenesis as well as analysis of segregating polymorphisms can be applied, also reveal that many novel, pleiotropic alleles affect most traits; these alleles often have sex-, environment- and genetic background-specific effects (Falconer and Mackay 1996; Flint and Mackay 2009).
The current challenge facing the genetic analysis of complex traits is to understand how allelic variants at hundreds of loci act together to affect organismal phenotypes. The key to solving this challenge is the realization that new mutations and segregating variants do not affect organismal phenotypes directly, but do so via complex interacting networks of transcriptional, protein, metabolic, and other molecular endophenotypes (Anholt et al. 2003; Sieberts and Schadt 2007; Chen et al. 2008; Emilsson et al. 2008; Keller et al. 2008; Ayroles et al. 2009; Cookson et al. 2009; Dobrin et al. 2009; Edwards et al. 2009a; Harbison et al. 2009; Morozova et al. 2009; Schadt 2009; Schadt et al. 2009; Winrow et al. 2009). If we know the network elements (genes) and connections (cis- and trans-regulation) associated with any complex phenotype, we can begin to make predictions about the effects of genetic perturbations on the organismal trait and whole network responses to such perturbations.
The wealth of genetic and genomic resources, diversity of complex phenotypes, and ability to rear large numbers of genetically identical individuals under controlled environmental conditions makes Drosophila melanogaster an attractive model organism for elucidating general principles of systems genetics of complex traits. Two complementary approaches have been used to begin to define coregulated transcriptional networks associated with complex traits. The first involves transcriptional profiling of mutations affecting the trait that have been generated in a common isogenic background; genes encoding coregulated transcripts are themselves candidate genes affecting the trait (Anholt et al. 2003; Dubnau et al. 2003; Keegan et al. 2007; Bauer et al. 2008; Rollmann et al. 2008; Seugnet et al. 2009; Magwire et al. 2010; Slack et al. 2010). However, the expense of whole-genome expression analysis limits the number of mutations studied to only a few of the large number of mutations affecting each trait, and the pleiotropic effects of each mutation on multiple traits confounds our ability to infer which of many coregulated transcripts are associated with the focal trait. The second approach utilizes populations of lines suitable for linkage or association mapping of complex traits in which genome-wide expression and organismal phenotypes have been jointly assessed (Passador-Gurgel et al. 2007; Ayroles et al. 2009; Harbison et al. 2009; Morozova et al. 2009). Quantitative trait transcripts associated with the trait can then be grouped into correlated transcriptional networks associated with the trait. The limitation of this approach is that only coregulated transcripts associated with allelic variation in the study population can be assessed.
Here, we combine mutagenesis with transcriptional variation in wild-derived Drosophila lines to derive computationally predicted networks that affect alcohol sensitivity. We chose alcohol sensitivity as an example of a complex trait because alcohol is a well-defined, ecologically relevant chemical cue for localization of decomposing fruit on which flies feed, and understanding genetic networks associated with alcohol sensitivity in flies is likely to show evolutionary conservation with relevance to human alcohol consumption and addiction (Park et al. 2000; Rothenfluh and Heberlein 2002; Guarnieri and Heberlein 2003; Mackay and Anholt 2006; Li et al. 2008; Morozova et al. 2009). Furthermore, alcohol knockdown time can be readily and accurately measured in an “inebriometer” (Weber 1988). When exposed to alcohol, flies initially become hyperactive, but soon lose postural control and fall through the inebriometer at a rate that is correlated with their sensitivity. Following recovery, a second exposure reveals the development of tolerance reflected in a longer knockdown time (Moore et al. 1998; Scholz et al. 2000; Singh and Heberlein 2000; Cowmeadow et al. 2005, 2006; Scholz et al. 2005; Wen et al. 2005; Morozova et al. 2006, 2009; Kong et al. 2009).
We screened a collection of 653 co-isogenic P[GT1]-element insertional mutations (Bellen et al. 2004) and identified 162 mutations that affect alcohol sensitivity. We then constructed networks of genetically correlated transcripts centered on a subset of these loci as focal genes, using whole-genome expression data from 40 inbred wild-derived lines (Ayroles et al. 2009). We validated hub genes in these networks by RNAi-mediated suppression of expression of target genes and identified interconnected genes that showed altered regulation of gene expression. We then assessed the effects of those genes on alcohol sensitivity and constructed new networks around the new focal genes. This iterative approach enables the gradual buildup of genetic networks for alcohol sensitivity and resistance with simultaneous functional validation. This approach is, in principle, applicable to any complex trait that can be analyzed with mutations and for which genetically correlated transcriptional networks can be constructed.
MATERIALS and METHODS
Homozygous lines containing P[GT1] elements in or near candidate genes in a co-isogenic Canton-S (A–F) background were generated as part of the Berkeley Drosophila Gene Disruption Project (Bellen et al. 2004). Revertant alleles were generated using crossing schemes that preserved the co-isogenic background of the revertant lines. For example, we constructed chromosome 2 revertant lines by crossing w1118; P; iso3 females to w1118; CyO/Sp; SbΔ2–3/TM6,Tb males. We mated male offspring of genotype w1118; CyO/P; SbΔ2–3/iso3 to w1118; CyO/Sp; iso3 females, and crossed single male offspring of genotype w1118; CyO/P−; iso3 in which the P element had excised to w1118; CyO/Sp; iso3 females. In the following generation, we mated males and females of genotype w1118; CyO/P−; iso3 inter se to make a homozygous revertant stock of genotype w1118; P−; iso3. w1118 and iso3 are the two isogenic chromosomes of the appropriate Canton-S strain; P refers to the chromosome with the P-element insertion; and P− indicates a P-element excision allele. Similar crosses were used to construct X and chromosome 3 revertant alleles (Rollmann et al. 2006). We verified excision of the P[GT1] element by PCR, using primers flanking the insertion site, followed by sequencing of the amplification products.
Seven RNAi transgenic fly lines (51977fng, 100739Thor, 104005Cyp1, 104016Men, 106641Pdk, 106875H15, and 107177Mlc-c) and the progenitor control line (60000) were obtained from the Vienna Drosophila RNAi Center (VDRC) from the φC31 RNAi library stocks (Dietzl et al. 2007). An ubiquitous tub-Gal475138 driver line was obtained from the Bloomington Drosophila Stock Center. This line was supplemented with a UAS-GAL4 enhancer. We crossed males from each of the transgenic UAS-RNAi lines to virgin females from the tub-Gal4 driver line to suppress the expression of the gene of interest in hybrid F1 offspring.
Fly stocks were reared under standard culture conditions on cornmeal–molasses–agar medium at 25°, 60–75% relative humidity, and a 12 hr light/dark cycle. Flies were not exposed to CO2 anesthesia for at least 24 hr prior to assays for alcohol sensitivity.
Quantitative assay for alcohol sensitivity and tolerance:
We assessed ethanol sensitivity for males only from 653 lines with single P[GT1] transposon insertions and their co-isogenic control line with two replicate measurements (N = 70 per replicate) per line. The replicates for each line were assessed on different days together with the appropriate co-isogenic control. We placed the flies in an inebriometer, a glass column with slanted mesh partitions (Weber 1988) preequilibrated with ethanol vapor, and collected them as they fell through the column at 1-min intervals. The mean elution time (MET) is a measure of alcohol sensitivity. Following the initial screen, we retested males from 210 P[GT1]-element insertion lines with significantly different mean elution times from their controls and females from a subset of these lines (four to five replicates per line, N = 60–70 per replicate). We determined alcohol sensitivity of revertant alleles simultaneously with the P-element mutant lines and the Canton-S control (five replicates per line, N = 70 per replicate).
We assessed both ethanol sensitivity and induction of tolerance for tub-Gal4;UAS-RNAi and progenitor F1 offspring by recording elution times upon initial exposure to ethanol (E1) and a second exposure (E2) of the same flies 2 hr later (five replicates per line per sex, N = 70 per replicate). The scaled difference of METs between the second and first exposures is a measure of the induction of tolerance, i.e., (E2 − E1)/(0.5(E1 + E2)) (Morozova et al. 2009).
Quantitative genetic and statistical analyses:
We used a one-way ANOVA model with post hoc Dunnett's test (Y = μ + L + ε) to assess differences in ethanol sensitivity between replicate measurements of each individual P[GT1]-element insertion line and its co-isogenic control at P < 0.05, where μ is the overall mean, L is the random main effect of line (P-element insertion vs. control lines), and ε is the environmental variance between replicates. The measurements for replicates were randomized over time along with appropriate contemporaneous P-element free Canton-S controls. We did not find significant differences between control replicates done on different days. We estimated the total mutational variance among lines as σL2 (the among-line variance component). The total phenotypic variance is σP2 = σL2 + σE2, where σE2 is the environmental variance component. We estimated broad sense mutational heritabilities as H2 = σL2/σP2. We quantified the standardized mutational effects of the P-element mutations as a/σP, where a is one-half of the difference between the homozygous mutant and control line, and σP is the phenotypic standard deviation of the control (Falconer and Mackay 1996).
Functional gene annotations are based on FlyBase (Drysdale and Crosby 2005; Tweedie et al. 2009) and Affymetrix (Netaffx Analysis Center, http://www.affymetrix.com) compilations. Gene ontology enrichment analysis used the DAVID program (Huang da et al. 2009).
To identify networks of genetically correlated transcripts, we computed the sex-centered correlation rij between all pairs of significant transcripts i and j among 40 wild-derived inbred lines (Ayroles et al. 2009). To enable visualization and to retain the most highly correlated transcripts, we used the absolute correlations |rij| to construct transcriptional networks around focal genes at an arbitrary threshold level determined for each gene individually.
We quantified mRNA levels by quantitative RT–PCR with the SYBR Green detection method according to the protocol from Applied Biosystems (Foster City, CA), using the ABI PRISM 7900 sequence detection system (Applied Biosystems). We used the glyceraldehyde-3-phosphate dehydrogenase gene as an internal standard. Three independent replicates of total RNA were isolated from P[GT1]-element insertion lines or tub-GAL4;UAS-RNAi and progenitor line F1 offspring using the Trizol reagent (GIBCO-BRL, Gaithersburg, MD), and cDNA was generated from 100 ng of total RNA by reverse transcription using the high capacity cDNA reverse transcription kit (Applied Biosystems). Primer Express software (Applied Biosystems) was used to design transcript-specific primers to amplify 100- to 150-bp regions of genes of interest. Primers were designed to encompass common regions of alternative transcripts. Negative controls without reverse transcriptase were used for all genes to exclude potential genomic DNA contamination. Samples in each run were normalized relative to a control sample, using 2–ΔΔCt values (Abi 2001). Statistical significance for differences in gene expression levels was determined by two-tailed Student's t-tests on ΔCt values.
To examine developmental stage-specific gene expression levels in P[GT1]-element insertion lines and revertant alleles, we analyzed relative levels of expression as described above after extraction of triplicate RNA samples from embryos between 5 and 8 hr after oviposition, third instar larvae, pupae, adult heads and bodies, and whole adult flies.
To identify transcriptional networks for increased sensitivity or resistance to ethanol exposure we first used mutational analysis to identify genes that directly impact alcohol sensitivity in an isogenic background. We then capitalized on natural variation in wild-derived inbred lines to integrate these genes into coregulated networks. Critical hub genes in these networks were then verified by quantitative RT–PCR.
Identification of co-isogenic P[GT1]-transposon insertion lines with altered alcohol sensitivities:
We measured the METs of 653 lines with single P[GT1]-transposon insertional mutations in common Canton-S (A–F) genetic backgrounds (Lukacsovich et al. 2001; Bellen et al. 2004). We found significant variation in alcohol sensitivity among the P[GT1]-insertion lines (Figure 1 and supporting information, Table S1). The average deviation of METs from the control for all 653 lines was 0.5 ± 0.2 min, with a range from −3.7 min to 7.7 min. The average mutational effect (Falconer and Mackay 1996) was 0.45 min, with a range of −0.7 to 1.7 min. The broad sense mutational heritability for alcohol sensitivity was HM2 = 0.17. The range of variation in alcohol sensitivity among the P-element insertion lines is comparable to that previously observed among 40 wild-derived inbred lines after a single ethanol exposure with broad sense heritability of 0.244 (Morozova et al. 2009).
We identified 210 P[GT1] lines with significant deviations from the controls from the initial screen, and retested these lines with five additional replicates per assay. The second screen confirmed that 162 mutant lines were significantly different from the control (77% of the lines detected in the initial screen and 28.7% of all 653 lines tested), with 27 sensitive and 135 resistant lines (Table S2). These 162 lines included 139 unique P[GT1]-insertion sites with 124 insertions in or near known genes (Table S2). Seventy-two lines are significantly different from the control even after Bonferroni correction for multiple testing (P < 0.0003). Only a few sensitive lines were confirmed after retesting, because the inebriometer assay is biased toward detecting alcohol resistance rather than hypersensitivity—the time it takes for flies to pass through the inebriometer sets the low threshold for the assay. In addition, Canton-S (E) flies had an elution time of 3.9 ± 0.1 min, precluding the detection of hypersensitivity in this genetic background.
The insertion site of the P[GT1]-element could be determined unambiguously in all but 5 of the 162 lines. A total of 10 insertions were in regions with no annotated gene within 2 kb of the P[GT1]-element insertion site. Several genes were tagged by multiple P[GT1]-element insertions (e.g., Beadex, capricious, Calreticulin, High mobility group protein D, Malic enzyme, Mnt, Thor, wing blister, Table S2). In addition, 51 out of 162 lines with altered alcohol sensitivity encode predicted transcripts with unknown functions.
Gene ontology analysis and correlated phenotypes:
Gene ontology (GO) analysis reveals diverse categories of genes that confer increased sensitivity or resistance to ethanol (Figure 2). Two GO categories are prominently associated with increased sensitivity to ethanol: olfactory behavior and nervous system development. The group of lines with increased resistance to ethanol exposure also showed enrichment for the “nervous system development” category, but in addition was enriched for genes associated with imaginal disc, organ, tissue and sensory organ development, as well as cell motility and neurogenesis. Most of the mutations that affect alcohol sensitivity have pleiotropic effects on other complex traits, such as startle-induced locomotion (Yamamoto et al. 2008), aggression (Edwards et al. 2009b), sleep (Harbison and Sehgal 2008), and starvation stress resistance (Harbison et al. 2005) (Table S3). We found significant positive correlations with aggression (N = 43, r = 0.35, P = 0.012) (Edwards et al. 2009b), lifespan (N = 142, r = 0.32, P < 0.0001) (Magwire et al. 2010), and 24 hr sleep (N = 22, r = 0.52, P = 0.008; Figure S1) (Harbison and Sehgal 2008).
To assess the correlation for alcohol sensitivity between the sexes, we examined differences in responses to a single ethanol exposure between males and females from 11 sensitive and 16 resistant P[GT1]-element insertion lines. We did not observe significant sex differences (r = 0.98, P < 0.0001; Figure S2). This lack of sexual dimorphism agrees with our previous study on alcohol sensitivity in 40 inbred wild-derived lines (Morozova et al. 2009), but contrasts with other traits examined in these lines, including sleep (Harbison and Sehgal 2008), startle behavior (Yamamoto et al. 2008), and olfactory avoidance behavior (Sambandan et al. 2006), where significant differences in mutational effects between the sexes have been reported.
Analysis of P[GT1]-element revertant alleles:
To demonstrate that the effects on alcohol sensitivity were due to the P[GT1] insertions rather than unrelated mutations in the genome, we excised the P[GT1] elements in a subset of the mutants. We chose two sensitive (Sip1 and psq) and five resistant (Fs, CG14430, CG14591/SCAP, Men, and Osi9) P[GT1]-element mutations and created revertant alleles for these lines. In addition, we used a previously generated revertant for BG01214 [sugarless (sgl)] (Edwards et al. 2009b) and BG01272 (Tre1/Gr5a) (Figure S3; Rollmann et al. 2006). We sequenced the revertant alleles and identified at least one precise revertant for each line except CG14430 and Tre1/Gr5a, where excision was imprecise.
We showed that the MET of the excision lines reverted to the control phenotype, including the imprecise excisions of CG14430 and Tre1/Gr5a (Figure 3A). The Osi9 revertant allele showed a MET significantly lower than both the mutant and control lines, possibly due to a mutation created during mobilization of the P[GT1] element. We characterized gene expression of the mutations and revertant alleles using qRT–PCR analysis at four developmental stages: embryos, larvae, pupae, and adults, as well as adult heads and bodies, separately (Figure 3B). All mutant lines showed changes in levels of expression in one or more developmental stages. Expression of psq is higher in the P[GT1]-element insertion line at all developmental stages, except embryos, as observed previously (Sambandan et al. 2006). Transcript levels in Sip1 and Fs mutants were lower in pupae but increased in adults. Men and sgl mutants were associated with decreased transcript levels throughout development. Osi9 and Tre1 mutants had reduced levels of gene expression in embryos but increased levels at all other stages except pupae. Transcript levels in the SCAP mutant were increased in adults, but decreased in larvae. Finally, the CG14430 mutant had reduced levels of gene expression in embryos and pupae, and increased levels in adult bodies. These results are consistent with the notion that insertions of transposons in regulatory regions of genes can alter their temporal expression during development, as reported previously (Sambandan et al. 2006).
To characterize revertant alleles further, we estimated transcript levels of the candidate genes at those developmental stages at which P[GT1]-element insertions significantly affected expression levels (Figure 3C). Transcript levels of revertant alleles were not different from controls for all candidate genes with the exception of Men and CG14430, where transcript levels were significantly different from both mutant and control. This cannot be readily explained, as the P[GT1]-element excision for Men was precise. For CG14430, partial restoration of transcript level can be explained by imprecise excision of the P element. Alcohol sensitivity for both of these revertant alleles, however, was not distinguishable from the control (Figure 3A).
Systems genetics analysis:
We took advantage of a previous study, which quantified genetic variation for 10,096 transcripts among 40 wild-derived inbred lines (Ayroles et al. 2009). These data have been used previously to regress phenotypic values of traits of interest on transcript abundance to identify candidate genes affecting these phenotypes, including alcohol sensitivity and tolerance (Morozova et al. 2009), aggression (Edwards et al. 2009a), and sleep (Harbison et al. 2009). Here, we assessed the extent to which the 10,096 transcripts were genetically correlated with genes affecting alcohol sensitivity identified in this study. This approach is limited by two assumptions, which require verification. First, we are assuming that genome-wide effects of mutational variation on the transcriptome will be the same as segregating variation. Second, we assume that genetic correlation is the result of coregulation of gene expression. If these assumptions are correct (as we will argue below), then we are justified in utilizing previously characterized transcriptional genetic networks from a natural population to infer coregulated networks in the mutant backgrounds.
We built transcriptional genetic correlation networks using 9 genes with mutations associated with increased alcohol sensitivity and 12 genes with mutations associated with alcohol resistance as focal genes and the transcriptional profiling data from the 40 wild-derived inbred lines (Ayroles et al. 2009). This analysis reveals highly interconnected transcriptional networks for each focal gene (Figure 4). Thirty-two genes occur in two or more networks. Thirteen of these 32 genes encode predicted transcripts of unknown functions. Correlated transcripts interconnect not only within alcohol-related phenotypic groups (i.e., networks around focal genes associated with enhanced sensitivity or resistance to alcohol), but also between phenotypic groups, indicating the complex genetic architecture that underlies alcohol sensitivity.
Gene ontology analyses for each predicted coregulated transcriptional network separately reveals that the major biological function categories that overlap between sensitive and resistant gene networks include regulation of primary metabolic process, nervous system and organ development, oogenesis, cell cycle phase, and positive and negative regulation of cellular process (Figure 5). At the same time transcripts affecting transcription, embryonic axis specification, and immune system development are enriched in transcriptional networks associated with increased sensitivity to alcohol. Transcripts affecting RNA processing are only associated with the CG10778 network. Synaptic transmission, response to light, and oxidative phosphorylation appear as unique gene ontology categories for transcriptional networks associated with increased resistance to alcohol. Genes annotated as being associated with translation are only apparent in the CG14430 network, where fatty acid and oxoacid metabolic processes are significantly interconnected in the transcriptional network around the Men focal gene. Finally, gene products implicated in signal transduction are enriched in the Tre1 module.
Previously, we identified 1,133 transcripts associated with one or more alcohol phenotypes by regressing phenotypic values on transcript abundance among the 40 wild-derived inbred lines (Morozova et al. 2009) as well as 2,615 transcripts associated with correlated transcriptional responses to divergent artificial selection for alcohol sensitivity and resistance (Morozova et al. 2007). Here, we combined identification of candidate genes by P[GT1]-element insertional mutagenesis with whole-genome transcriptional data to build networks of highly interconnected genes on the basis of genetic correlations among transcripts in a natural population. We predicted to find more overlap between our previously documented transcriptional networks and the networks shown in Figure 4 than expected by chance. Indeed, we found 78 transcripts and 95 transcripts that were in common (χ12 = 23, P < 0.0001 and χ12 = 26, P < 0.0001) between the present study and Morozova et al. (2009) and Morozova et al. (2007), respectively, including 11 focal candidate genes (Bx, CG14430, CG14591, Doa, Men, Pdk, Prosap, psq, Sdc, Sip1, and Tl; Figure 4). Fifty-five of these transcripts encode gene products of unknown functions (Table S4). Imaginal disc, nervous system, organ development, and olfactory behavior were enriched among the major biological function gene ontology categories for the common transcripts.
The complex interconnectivity of the networks can be broken down into subsets of focal genes and associated networks that contribute to resistance to alcohol (Figure 6A). This reveals an underlying network structure of hub genes associated with alcohol resistance, including Cyp1, fng, H15, and Mlc-c.
Validation of candidate genes and interconnected networks:
We used RNAi lines from the Vienna Drosophila Research Center collection (Dietzl et al. 2007) to validate candidate genes found by the P[GT1]-element screen and to test our computationally predicted networks. We focused on a subset of candidate genes in which P[GT1]-element insertions lead to increased alcohol resistance, including Men, Pdk, and Thor (Figure 6A).
To validate the connectivity of the predicted networks, we chose four genes, Cyclophilin 1 (Cyp1), fringe (fng), H15, and Myosin light chain cytoplasmic (Mlc-c) that were common for two or more predicted networks (Figure 6A). We crossed lines containing UAS-RNAi constructs for each target gene and the progenitor control to a ubiquitous tubulin-Gal4; driver line to suppress the expression of the target genes. F1 tub-GAL4;UAS-RNAi hybrids were used for both molecular and behavioral experiments. The efficiency of RNAi-mediated suppression of individual target genes in tub-GAL4;UAS-RNAi F1s was assessed by qRT–PCR measured on whole flies (Figure 6B). Transcript levels of the target genes were suppressed in all cases compared to the progenitor control line and varied from 20% (H15 females) to 90% (Cyp1 males) reduction in gene expression. Next, we measured METs after an initial exposure to ethanol (E1) and after a second exposure 2 hr later (E2). As illustrated in Figure 6C, RNAi-mediated suppression of fng, Mlc-c, and Pdk resulted in greater sensitivity to a single ethanol exposure than the control, while RNAi-mediated suppression of Cyp1, H15, Men, and Thor resulted in enhanced resistance. A second exposure showed development of tolerance through a shift in MET, except for Mlc-c and Pdk males. The MET averaged over all RNAi hybrids was greater after the second (11.1 min) than the first exposure (7.8 min), again indicating the development of tolerance. We also quantified tolerance (T) as the shift in MET between the first and second exposures, scaled by their average values (T = (E2 − E1)/0.5(E1 + E2)). Overall, we did not find significant changes in tolerance with the exception of fng, which showed a significant increase in tolerance in both sexes (Figure 6C).
In summary, using UAS-RNAi lines we independently confirmed that the three candidate genes tested (Men, Pdk, and Thor) contribute to alcohol sensitivity. In addition we showed that genes in the same transcriptional genetic network as Men, Pdk, and Thor (Cyp1, fng, H15, and Mlc-c) also affect alcohol sensitivity.
Predictive value of transcriptional connectivity:
Previous studies showed that networks of correlated transcripts can be used to identify new candidate genes affecting a phenotype (Edwards et al. 2009a; Harbison et al. 2009; Morozova et al. 2009). We tested whether we could use an iterative process to expand our networks by constructing new networks of correlated transcripts. We used Cyp1, fng, H15, and Mlc-c as focal genes together with transcriptional profiling data from the 40 wild-derived inbred lines (Ayroles et al. 2009) to assess the genome-wide connectivity of these four transcripts. This resulted in a highly connected ensemble of 85 genes, with 65, 17, and 3 genes shared by two, three, and four modules around the focal genes, respectively (Figure 7A). The strength of absolute correlation of the transcripts for these networks was surprisingly high: ≥0.65 for H15 and fng, and ≥0.75 for Mlc-c and Cyp1. Biological function GO categories of the 85 transcripts showed enrichment for protein metabolic processes, microtubule cytoskeleton organization, vesicle-mediated transport, and translation. Thirty nine of the 85 transcripts from computationally predicted networks were common between differentially expressed transcripts from at least one of the previous studies (Morozova et al. 2006, 2007, 2009) (Table S5). It should be noted that 82% of these Drosophila transcripts have human orthologs.
Since expression levels of the focal genes in corresponding tub-GAL4;UAS-RNAi lines were suppressed (Figure 6B), we predicted that transcripts genetically correlated with the focal genes would also show altered transcript levels in the Cyp1, fng, H15, and Mlc-c RNAi backgrounds. Indeed, expression of the most highly interconnected transcripts, such as Histone H4 replacement (His4r), Tctp, and Synaptobrevin (Syb) were altered in the RNAi-hybrid lines in both sexes (Figure 7B). Since these transcripts were among the correlated transcripts in networks containing Pdk and Thor, we tested their expression levels in tub-GAL4;UAS-RNAi hybrids with reduced expression of Pdk and Thor. Here, we also found reduced expression levels of the His4r, Tctp, and Syb transcripts (Figure 7B).
We have applied a combination of mutational and computational approaches using P-element insertional mutagenesis while capitalizing on naturally occurring variation in genome-wide transcript abundance levels to predict transcriptional networks that determine alcohol sensitivity in D. melanogaster. This approach enables a systematic walk through the transcriptome to identify interconnected hub genes that can serve as landmarks for iterative expansion of the network. The hub genes also serve as cornerstones for subsequent comparative genomic approaches in which orthologous networks for alcohol-related phenotypes in other species, including humans, can be superimposed on the transcriptional network for alcohol sensitivity in Drosophila. The feasibility of such a translational approach has been demonstrated previously by implicating cytoplasmic malic enzyme 1 in human drinking behavior on the basis of the identification of Men as a gene associated with alcohol sensitivity in Drosophila (Morozova et al. 2009).
P[GT1]-element mutations affecting alcohol sensitivity:
Approximately 29% of the P-element mutations tested for alcohol sensitivity and resistance differ significantly from the co-isogenic P-element free controls. This percentage demonstrates a large mutational target for alcohol sensitivity and indicates that a large fraction of the genome contributes to alcohol sensitivity, implicating extensive pleiotropy (Table S3). Indeed, most of the mutations that affect alcohol sensitivity have pleiotropic effects on other complex traits, including aggression (Edwards et al. 2009b), life span (Magwire et al. 2010), and sleep (Harbison and Sehgal 2008) (Table S3).
The large mutational target appears to be a characteristic of quantitative traits. For example, ∼37% of P-element insertion lines tested for startle-induced locomotion differed significantly from the P-element–free, isogenic Canton-S control line (Yamamoto et al. 2008). About 6% of P-element insertion lines had aberrant olfactory avoidance behavior (Sambandan et al. 2006). A total of 35% of the P-element insertion lines tested for aggression exhibited levels of aggression that differed significantly from the control (Edwards et al. 2009b). Fifteen percent of P-element insertions with known effects on physiology, development, and behavior affect 24-hr sleep time (Harbison and Sehgal 2008). The numbers of abdominal and sternopleural bristles were affected by 38.3% of co-isogenic P-element insertions (Norga et al. 2003).
We found a significant positive correlation between alcohol sensitivity and the number of aggressive encounters under conditions of competition for a limited food source (Figure S1). Thus, mutations associated with ethanol resistance are also associated with increased aggressive behavior. We also found a positive correlation with 24-h sleep and lifespan (Figure S1). Thus, mutations associated with resistance to ethanol exposure sleep longer and live longer.
Several of the genes implicated in alcohol sensitivity in Drosophila in this study have been implicated previously, either as genes that are differentially expressed after repeated exposures to ethanol in a Canton-S (B) background (Morozova et al. 2006), genes associated with correlated transcriptional responses to divergent artificial selection for alcohol sensitivity and resistance (Morozova et al. 2007), or genes computationally implicated in transcriptional networks for alcohol sensitivity and tolerance in 40 inbred wild-derived lines (Morozova et al. 2009) (Table S6). Genes that were consistently identified in all studies include CG12505, lola, Men, and Pdk. CG12505 is also involved in starvation resistance (Mattaliano et al. 2007). lola encodes a transcription factor required for axon growth and guidance in Drosophila (Giniger et al. 1994), and plays a role in aggression (Edwards et al. 2009b) and startle behavior (Yamamoto et al. 2008). Men and Pdk encode malic enzyme and pyruvate dehydrogenase kinase, respectively, and function in energy metabolism (Morozova et al. 2006).
Systems genetics analysis:
P-element insertional mutations can cause widespread alterations in transcript levels throughout the genome (Anholt et al. 2003; Harbison et al. 2005; Rollmann et al. 2006). It is reasonable to hypothesize that genes that are coregulated in mutant backgrounds are themselves candidate genes affecting the trait. Here, we used a combination of P-element insertional mutagenesis and analysis of natural variation to derive computational networks that underlie alcohol sensitivity in D. melanogaster. An iterative approach, which uses RNAi-mediated inhibition of expression of focal genes to expand computational networks, while at the same time confirming effects of hub genes on the phenotype, presents a paradigm for stepwise expansion of genetic networks associated with a specific phenotype. Gene ontology analysis of networks built around hub genes associated with high or low sensitivity to alcohol exposure that contain genes associated with specific biological processes (e.g., RNA processing, synaptic transmission, signal transduction), whereas other networks are composed of more diverse covariant transcripts. Despite this complexity, P-element mutagenesis allows the identification of gene ensembles associated with high sensitivity or resistance to alcohol exposure. The boundaries between these networks are, however, fluid, as is evident from our observation that P-element insertions in genes that affect alcohol sensitivity can either be recruited in networks associated with low or high sensitivity depending on the precise insertion site and consequently the precise effect of the P element on transcription.
We used 9 P-element–tagged alleles associated with increased alcohol sensitivity, and 12 alleles associated with alcohol resistance to build correlated transcriptional networks associated with each hub gene (Figure 4). We can validate the predicted transcriptional networks by measuring transcript levels of genes connected to a P-element–tagged focal gene. For example, a previous study showed that expression levels of CG31063, CG6503, CG1443, CG14277, CG12262, CG32032, CG11200, CG4786, CG5399, and Obp18a were altered in the mutant MenBG02365 P-element insertion background (Morozova et al. 2009). It should be noted that CG31063, CG5399, CG6503, CG14277, and CG32032 have previously unknown biological functions; thus, their connectivity within a network associated with a phenotype, i.e., alcohol sensitivity, provides a functional annotation for these genes.
The complexity of interconnections between clusters and the number of hub genes could be simplified by considering only a subset of genes with P-element–tagged alleles, including CG14430, Men, Pdk, Sdc, and Thor, as focal genes. All five mutant lines have increased ethanol resistance relative to the control. CG14430, Men, and Pdk have been associated previously with changes in transcript abundance after a second exposure to ethanol (Morozova et al. 2009); Men and Pdk also showed differential expression after artificial selection for alcohol sensitivity (Morozova et al. 2007) and after repeated exposures to ethanol in the Canton-S (B) genetic background (Morozova et al. 2006). Repeated exposures to ethanol also resulted in changes in Thor transcript abundance (Morozova et al. 2006). We built transcriptional networks around these genes (Figure 6) and showed that RNAi targeted against Men, Pdk, and Thor, Mlc-c, H15, Cyp1, and fng decreases their expression. We confirmed not only that expression of these genes was down-regulated, but also that alcohol sensitivity in RNAi lines targeting these genes was altered upon repeated exposure to ethanol (Figure 6). We then chose Mlc-c, H15, Cyp1, and fng for further validation, as these genes emerged as hub genes.
Next, we constructed new predicted networks of genes using Cyp1, fng, H15, and Mlc-c as focal genes. Unexpectedly, we found an even more highly connected network of 85 genes shared by two and more gene clusters (Figure 7). We chose His4r, Syb, and Tctp for validation because they showed the highest interconnectivity level and were correlated with one of the focal genes in Figures 4 and 6. In addition, His4r and Tctp showed differential expression after a second ethanol exposure in 40 wild-derived lines (Morozova et al. 2009) (Table S5), and Syb was differentially expressed among selection lines for alcohol sensitivity (Morozova et al. 2007) and after repeated ethanol exposures in a Canton-S (B) line (Morozova et al. 2006). An alcohol-related phenotype has also been reported for the murine ortholog of Syb (Vamp2) (Mulligan et al. 2006). Using RT–PCR analysis we confirmed that expression levels of His4r, Syb, and Tctp were altered in all of the RNAi lines targeting Cyp1, H15, fng, Mlc-c, and Pdk in both sexes.
The error rate underlying the predicted networks is unknown. Thus, it is possible that some of the transcripts interconnected with focal genes may not be associated with alcohol sensitivity. Determination of the error rate in building iterative networks can only be done empirically. On the basis of Morozova et al. (2009), the error rate for genes coregulated with Malic enzyme and associated with a second exposure to ethanol, estimated by qRT–PCR, was 25% at r > 0.3. Similarly, the error rate for genes coregulated with pipsqueak and associated with an initial exposure to ethanol, estimated by qRT–PCR, was 26% at r > 0.3. Harbison et al. (2009) calculated an error rate of 19% at r > 0.7 for covariant transcriptional networks associated with sleep phenotypes. In building these networks, the number of interconnected hub genes can be increased or decreased by adjusting the criterion for strength of correlation. However, reduction of the threshold for correlation will introduce more false-positive associations.
To reduce the false-positive rate, we used two procedures. First, we focused only on genes with the strongest correlations with focal genes known to affect alcohol sensitivity. Second, we performed validation of several interconnected genes both by qRT–PCR and, wherever possible, by assessing the effect on the phenotype. These safeguards protect the iterative networking approach we describe from accumulating incremental errors. The observation that 39 of the 85 transcripts from computationally predicted networks were common between differentially expressed transcripts from at least one or more previous studies (Morozova et al. 2006, 2007, 2009) (Table S5) provides further validation for the networks despite the lack of a precise false-positive error estimate.
Previous studies have shown associations between polymorphisms in Adh and Aldh with ethanol sensitivity in Drosophila (Fry and Saweikis 2006; Fry et al. 2008). Unfortunately, our collection of P-element insertion lines did not contain an insertion affecting Adh or Aldh. Furthermore, the strength of correlations of Adh or Aldh transcripts with any of our focal genes did not reach our minimal arbitrary threshold for significance (r > 0.45).
Our results demonstrate that sensitivity to alcohol exposure is determined by the composition and dynamics of its genome-wide transcriptional underpinnings. Although coregulated transcriptional networks do not provide information about causal relationships among their constituent transcripts, they provide a useful context to guide subsequent functional studies. Our studies also emphasize the importance of assessing the genetic context when evaluating the effects of single genes on alcohol sensitivity and to note that selection for low or high sensitivity acts on genetic networks rather than on single alleles.
We thank Elaine Smith and Cynthia Rouf for technical assistance, and Julien Ayroles for help with statistical analyses. This work was supported by grant AA016569 (T.F.C.M.) from the National Institutes of Health.
Supporting information is available online at http://www.genetics.org/cgi/content/full/genetics.110.125229/DC1.
Available freely online through the author-supported open access option.
Communicating editor: M. Johnston
- Received November 15, 2010.
- Accepted January 22, 2011.
- Copyright © 2011 by the Genetics Society of America
Available freely online through the author-supported open access option.