Buffering of Genetic Regulatory Networks in Drosophila melanogaster

Abstract

Regulatory variation in gene expression can be described by cis- and trans-genetic components. Here we used RNA-seq data from a population panel of Drosophila melanogaster test crosses to compare allelic imbalance (AI) in female head tissue between mated and virgin flies, an environmental change known to affect transcription. Indeed, 3048 exons (1610 genes) are differentially expressed in this study. A Bayesian model for AI, with an intersection test, controls type I error. There are ∼200 genes with AI exclusively in mated or virgin flies, indicating an environmental component of expression regulation. On average 34% of genes within a cross and 54% of all genes show evidence for genetic regulation of transcription. Nearly all differentially regulated genes are affected in cis, with an average of 63% of expression variation explained by the cis-effects. Trans-effects explain 8% of the variance in AI on average and the interaction between cis and trans explains an average of 11% of the total variance in AI. In both environments cis- and trans-effects are compensatory in their overall effect, with a negative association between cis- and trans-effects in 85% of the exons examined. We hypothesize that the gene expression level perturbed by cis-regulatory mutations is compensated through trans-regulatory mechanisms, e.g., trans and cis by trans-factors buffering cis-mutations. In addition, when AI is detected in both environments, cis-mated, cis-virgin, and trans-mated–trans-virgin estimates are highly concordant with 99% of all exons positively correlated with a median correlation of 0.83 for cis and 0.95 for trans. We conclude that the gene regulatory networks (GRNs) are robust and that trans-buffering explains robustness.

DIFFERENCES in the level of expression between two alleles are widespread and play an important role in complex traits. They have been implicated in human disease phenotypes such as cancer and are an important component of variation in Drosophila, as well as yeast, plants, animals, and humans (Mendell and Dietz 2001; Brem et al. 2002; Cowles et al. 2002; Yan et al. 2002; Wittkopp et al. 2004, 2006, 2008b; Springer and Stupar 2007; Hutter et al. 2008; Smith and Kruglyak 2008; Tirosh et al. 2009; Emerson et al. 2010; McManus et al. 2010; Crowley et al. 2015). In Drosophila, variation in gene expression is heritable (Wayne et al. 2004, 2007) and evidence for both cis- and trans-regulatory polymorphisms is widespread (Wittkopp et al. 2004, 2008b; Hughes et al. 2006; Genissel et al. 2008; Wang et al. 2008; Graze et al. 2009, 2012, 2014) and evidence for cis- by trans-interactions has been reported (Wittkopp et al. 2004, 2008b; Wang et al. 2008; Graze et al. 2014). It has been argued that these interactions may also present evidence for coevolution of cis- and trans-regulation between species (Wittkopp et al. 2004).

A variety of genetic designs have been used to identify cis- and trans-effects, including chromosome substitution (Hughes et al. 2006; Lemos et al. 2008; Wang et al. 2008; Wittkopp et al. 2008a; Graze et al. 2014), expression QTL (eQTL) (reviewed by Brem et al. 2002; Kirst et al. 2005; Genissel et al. 2008; Gilad et al. 2008; Mackay et al. 2009; King et al. 2012, 2014; Massouras et al. 2012), and allelic imbalance (reviewed by Yan et al. 2002; Lo et al. 2003; Wittkopp et al. 2004, 2008a,b; Ronald et al. 2005; Guo et al. 2008a; Graze et al. 2009, 2012; Tirosh et al. 2009; Zhang and Borevitz 2009; McManus et al. 2010; Pastinen 2010). While eQTL studies often refer to cis- and trans-effects, eQTL cis-effects are local effects and trans-effects are distal effects (Rockman and Kruglyak 2006). In contrast, the cis- and trans-effects in allelic imbalance (AI) studies refer to molecular effects in the regulatory region of the locus itself (cis) or variation in genes molecularly interacting with common regulatory elements (trans) (Figure 1) (Yan et al. 2002; Wittkopp et al. 2004; Graze et al. 2009, 2012; Wittkopp and Kalay 2012).

Figure 1.

Cartoon of cis- and trans-regulation. A1 and A2 are alleles from two different testcrosses of a parental line (A1, red; A2, cross-hatched red) with a tester allele (At, blue). Alleles are tracked by SNPs within the coding sequence. Note that these SNPs are not necessarily causal. For testcross 1, there is no allelic imbalance; allele-specific expression is equal between the A1 allele and the At allele. For testcross 2, the A2 allele is different from the At allele, indicating cis-regulation. In this cartoon, the expression of the exon is regulated by a regulatory SNP within the gene’s promoter. Trans-effects are identified by comparing the expression of the At allele in testcross 1 to the expression of the At allele in testcross 2.

Previous studies of AI in Drosophila examining interspecies hybrids (Michalak and Noor 2003, 2004; Ranz et al. 2004; Wittkopp et al. 2004; Ortiz-Barrientos et al. 2007; Graze et al. 2009; McManus et al. 2010, 2014; Graze et al. 2012) identified cis-effects in 20–90% of genes (Wittkopp et al. 2004; Graze et al. 2009, 2012; McManus et al. 2014; Meiklejohn et al. 2014). Most intraspecific studies have used chromosomal substitutions and these studies have found contributions of trans-effects ranging from 0.01% (McManus et al. 2014) to 40–50% of genes (Wittkopp et al. 2004; Meiklejohn et al. 2014) but sometimes reaching 90% of genes (Wang et al. 2008).

In a single hybrid, cis-regulatory variation is identified by comparing gene expression of two alleles at the same loci within that hybrid genotype (reviewed by Pastinen 2010) while trans-estimates are derived from comparing the same allele between hybrid genotypes (Figure 1). For a single hybrid, the estimates of cis are confounded with cis- by trans-interactions, and interactions are usually assumed to be negligible relative to the main effect of cis. When multiple genetically variable lines are crossed to a common tester line, cis- and trans-effects for each allele in the population can be estimated relative to the effects of the common tester allele (Nuzhdin et al. 2012). By modeling AI as a function of cis and trans and their interaction in the population, the contribution of these effects to AI can be estimated.

Identifying genes where AI occurs is a critically important biological problem, yet successfully identifying AI is technically challenging. Whereas mapping to a common reference strain is convenient, it introduces bias in the assessment of AI due to the difference in polymorphisms between the alleles interrogated and the reference (Degner et al. 2009; Graze et al. 2012). Masking SNPs and extensive filtering may ameliorate some sources of potential bias (Stevenson et al. 2013; van de Geijn et al. 2014), but bias may come from multiple sources and not all of these can be effectively filtered (Degner et al. 2009; Leon-Novelo et al. 2014; van de Geijn et al. 2014). There is mounting evidence that using personal genomes is more effective at reducing bias, particularly from structural variants (Degner et al. 2009; Rozowsky et al. 2011; Graze et al. 2012; Yuan and Qin 2012; Munger et al. 2014; Zou et al. 2014). Further, previously unidentified structural variation can be identified and the bias corrected using appropriate DNA controls (reviewed by Wittkopp et al. 2004; Pastinen 2010; Graze et al. 2012) and it is impossible to detect and filter such bias without DNA controls. This is often ignored, yet small amounts of unaccounted for bias lead to gross inflation of type I error rates (Leon-Novelo et al. 2014). Bias correction remains an important, and mostly ignored, feature of testing for AI.

The extent to which regulation of expression is sensitive to the environment is an interesting and not fully resolved issue. A frequent lack of concordance between eQTL across environments or genetic backgrounds has been previously interpreted as strong evidence for differential gene regulation in Drosophila (Mackay et al. 2009; Massouras et al. 2012; Huang et al. 2015). However, eQTL are reproduced between homozygous and heterozygous flies (Massouras et al. 2012), and regulatory variation measured using allelic imbalance appeared robust to environmental perturbations in Arabidopsis (Cubillos et al. 2014). Here, we have assayed the transcriptome in virgin and mated flies to test whether AI variation responds to mating.

Following mating in Drosophila melanogaster the fly experiences changes in metabolism (McGraw et al. 2004, 2008; Dalton et al. 2010), immune defense (reviewed by Lawniczak and Begun 2004; Peng et al. 2005; Lawniczak et al. 2007; Morrow and Innocenti 2012), life span (Chapman et al. 1995), detoxification products (McGraw et al. 2004), and gene expression (Smith et al. 2013; Zhou et al. 2014). Metabolic changes are found throughout the whole body (McGraw et al. 2004, 2008) as well as specific changes in the female head fat body (Dalton et al. 2010), including effects on nutritional state postmating.

This study explores whether an allelic imbalance in an interspecific population of D. melanogaster differs by environment and whether regulatory variation equally involves cis- and trans-effects. Consistent with the literature, this study finds 54% of all genes have evidence for AI. Coevolutionary processes between species result in cis- by trans-interactions, which have been clearly demonstrated in melanogaster/simulans hybrids (Wittkopp et al. 2004, 2008b; Landry et al. 2005; Graze et al. 2009, 2012; Fontanillas et al. 2010; Stevenson et al. 2013; McManus et al. 2014). However, coevolution is not a plausible explanation within species. We propose an alternative hypothesis for apparent cis–trans compensation, where the gene expression level perturbed by cis-regulatory mutations is partially compensated through trans-regulatory mechanisms. This is supported by recent experimental evidence of widespread molecular interactions as a result of cryptic genetic variation (Paaby and Rockman 2014).

Methods

All original data are attached to this article in File S1 and described in the Supplemental Material. All programs used for analysis, plus documentation, are at http://github.com/McIntyre-Lab/papers/tree/master/fear_ase_2016.

Sequencing data

RNA-seq data from a panel of 68 D. melanogaster F₁ hybrids (Kurmangaliyev et al. 2015) were used to measure allele-specific expression. Briefly, females from two populations of naturally derived D. melanogaster strains [54 from the Drosophila Genetic Reference Panel (DGRP) (Mackay et al. 2012) and 14 from Winters, California] were crossed to males of a single D. melanogaster laboratory strain (w¹¹¹⁸). Kurmangaliyev et al. (2015) isolated RNA from pools of ∼50 adult female heads from the F₁ hybrid progeny. At least 3 independent biological replicates of each cross were evaluated. Libraries were constructed using standard protocols and pooled at equal molar concentrations. They were sequenced in eight or more lanes of Illumina HiSeq 2000. FlyBase v5.51 was used and exons were identified from this build. Exons that overlapped within a gene were collapsed as in Graze et al. (2012).

Genome ambiguity: simulation 1 random SNPs

Genome ambiguity is a major source of mapping bias (Degner et al. 2009), but it can be estimated via simulation (Degner et al. 2009; Leon-Novelo et al. 2014). To identify regions of genome ambiguity we simulated data similar to the intraspecific hybrid population studied here. On average, parental genotypes have ∼160,000 exonic SNPs compared to the reference (FlyBase v5.51). Parental genotypes for 101 lines were simulated by randomly incorporating 160,000 SNPs into 63,181 exons of the D. melanogaster reference (FlyBase v5.51). All possible 95-bp reads were created from each reference, using a sliding window. An F₁ panel (n = 100) was constructed by taking simulated reads from each of 100 simulated genotypes (line_i where i = 1–100) and mixing them with a single common simulated genotype (line_i: tester). Reads were aligned to each “parental” reference (line₁₀₁: tester), using Bowtie (v0.12.9, -m1, -v3) (Langmead et al. 2009) and Last (v247, -l 25) (Frith et al. 2010). Genotype-specific alignments were compared as in Graze et al. (2012) and reads were categorized as aligning best to the “line” or the “tester’ or aligning equally well to “both.” For each exon, bias (q) was estimated by taking the proportion of reads that aligned to the tester over the sum of the allele-specific reads (line_i + tester). Because the simulated F₁ panel was created by equally mixing reads from parental strains, q is expected to be 0.5. When q deviates from 0.5, the exons must contain sequence ambiguities that affect the mappability of this region. A total of 55,647 exons showed no sequence ambiguity, while 7534 have some evidence of ambiguity (Table S1). Of these there were 1372 that were ambiguous in >50% of simulated lines, and 807 were ambiguous in all 100 simulated F₁ hybrids (Figure S1A). We removed the 807 exons that were ambiguous in all lines from the remainder of analyses.

Genotype-specific references

Genotype-specific references have been used to reduce map bias (Degner et al. 2009; Rozowsky et al. 2011; Graze et al. 2012; Yuan and Qin 2012; Stevenson et al. 2013; Leon-Novelo et al. 2014; Munger et al. 2014) by incorporating a filtered set of SNPs and indels that were identified in each parental strain. The GATK UnifiedGenotyper tool (ver. 2.1-8) was used to identify SNPs and indels in 68 genotypes (DGRP and Winters; line) and the laboratory strain (w¹¹¹⁸; tester). To minimize biases from miscalled variants (Leon-Novelo et al. 2014), SNPs and indels were strictly filtered as follows. All variants were considered as a line–tester pair, and if a variant was heterozygous in either the line or the tester, it was removed. The remaining SNP locations were masked in the genomic reference (FlyBase r5.51) for each line–tester pair (n = 68), by replacing the current nucleotide with an “N”. RNA-seq data from each F₁ hybrid were aligned to the masked reference, using Bowtie (v0.12.9, -a, -v3) (Langmead et al. 2009) followed by Last (v247, -l 25) (Frith et al. 2010). A custom python script summarized masked alignment files by counting the number of RNA-seq reads that supported each variant call (SNP or indel). A variant (SNP or indel) was retained if it had at least one supporting RNA-seq read or if the DNA coverage for the original variant call was five or more. Filtered variants (File S1) were used to create a genotype-specific exome reference by updating the D. melanogaster genomic reference (FlyBase r5.51) with SNPs and indels and removing redundant exons.

Genotype-specific bias estimation: simulation 2 (qSIM)

Previous studies have used DNA read counts to estimate inherent bias and then corrected these biases by including a random effect parameter in their allelic imbalance models (Wittkopp et al. 2004; Graze et al. 2012; Leon-Novelo et al. 2014). Here we do not have DNA read counts; however, we use read simulation to identify exons that contain ambiguity and to create a bias correction parameter (q = qSIM; see Leon-Novelo et al. 2014). To mimic the information provided by the DNA, we simulated all possible 95-bp DNA reads from each parental genotype-specific exome reference (DGRP, Winters, and w¹¹¹⁸) (see Methods: Genotype-specific references). Similar to genome ambiguity simulation, a hybrid was created by mixing reads from each line (n = 49) with reads from the common laboratory strain (w¹¹¹⁸). Simulated hybrid reads were aligned separately to the line and tester genotype-specific references, using Bowtie (v0.12.9, -m1, -v3) (Langmead et al. 2009) and Last (v247, -l 25) (Frith et al. 2010). Genotype-specific alignments were compared as in Graze et al. (2012) and reads were categorized as aligning best to the line or the tester or aligning equally well to both. For each exon, the bias term (qSIM) was calculated by taking the proportion of reads that aligned to the tester over the sum of the allele-specific reads (line_i + tester). After removing 807 exons that are always biased in genome ambiguity simulations (see Methods: Genome ambiguity: simulation 1 random SNPs ), 53,923 exons showed no bias (qSIM = 0.5), while 8451 regions showed bias in at least 1 line–tester combination (Figure S1B and Table S1). The bias with the biggest difference from 0.5 was used as the estimate for bias for each exon.

Mapping

Genotype-specific references were created by updating the genome, using variants identified by global alignments of these and all other data for D. melanogaster lines simultaneously (Table S2). Data for the 68 lines were mapped to each parent and w¹¹¹⁸ separately to the line and tester genotype-specific references, using Bowtie (v0.12.9, -m1, -v3) (Langmead et al. 2009) and Last (v247, -l 25) (Frith et al. 2010). Reads in exons were counted and classified as aligning to the line, to the tester, or to both. Several filtering steps were performed. Exons that always showed bias in the genome ambiguity simulation were removed. Exons with very low coverage (average per nucleotide coverage <25) were removed (n = 31,791). Exons that were not present in at least 10% of genotypes were removed (n = 11,382). Exons that were not present in both environments (virgin, mated) were removed (n = 1253; 5391 remaining). After filtering exons, genotypes whose median value across all lines (for the ratio of tester-specific to total allele-specific read counts) was extreme (≤0.4 or ≥0.6) were removed (n = 9) as these likely reflect poor variant detection and systematic bias. Finally, genotypes with <500 remaining exons were removed (n = 13). A total of 49 genotypes and 5391 exons were analyzed for allelic imbalance (File S2).

An intersection test for AI

We use the Poisson Gamma (PG) model developed by Leon-Novelo et al. (2014). Briefly, under the null hypothesis of no allelic imbalance ${\mathrm{H}}_0:\theta =0.5$ vs. the alternative of allelic imbalance${\mathrm{H}}_1:\theta \ne \mathrm{0.5.}$ Let $x_i$ and $y_i$ be the line and tester RNA allele-specific read counts in biological replicate $i(i=1,\dots I):$

$$y_i|\mu ,\alpha ,{\beta }_i,q\sim \text{Poisson}\left(\mu \alpha {\beta }_{i}q\right),$$

and

$$x_i| \mu ,{\beta }_i,q\sim \text{Poisson}\left(\mu {\beta }_i\left(1-q\right)\right).$$

Here μ is the overall mean, ${\beta }_i$ is the variation of biological replicates ($i=1,\dots ,I$), α is the effect of a read having AI, and q is a constant to incorporate bias information, where values >0.5 indicate bias toward the tester allele and values <0.5 are bias toward the line allele. If θ is the real proportion of reads from the tester allele, then

$$\theta = \frac{\mu \alpha {\beta }_i}{\mu {\beta }_i+\mu \alpha {\beta }_i}= \frac{\alpha }{1+\alpha }.$$

When there is no AI ($\theta =0.5$), therefore$\alpha =1.$ The bias correction parameter q can either be a fixed constant (e.g., 0.5) or be a random variable. The PG model requires that exons have at least three biological replicates for a given genotype × mating status and that the exon is expressed in at least one of these replicates (i.e., average per nucleotide coverage >0). For those exons where qSIM captures the potential bias this value is used in the Bayesian model. The simulated value of qSIM, when different from 0.5, has been shown to reflect bias from DNA controls and to be close to the value of bias estimated from DNA (Leon-Novelo et al. 2014). Note that the input to this model is raw read counts and that no additional normalization is performed.

The type I error rate for this test is <5% only when there is modest misspecification of bias; however, error rate increases as the percentage of misspecification increases (Leon-Novelo et al. 2014). For the commonly used binomial procedure in the test for AI type I error rates are large (Degner et al. 2009; Fontanillas et al. 2010; Nothnagel et al. 2011; Yuan and Qin 2012; Leon-Novelo et al. 2014; van de Geijn et al. 2014; Zou et al. 2014; Castel et al. 2015). Using DNA controls to estimate bias reveals that most bias (95%) is <20% (Graze et al. 2012). This translates to values of the parameter q between 0.4 and 0.6. Evaluating the posterior at three values of q, 0.4, 0.5, and 0.6, provides an estimate of the behavior of the posterior for the most likely set of conditions. Using an intersection test approach (Stell et al. 1980; Berger 1997; Berger and Boos 1999; Coffman et al. 2003), AI is declared if the credible interval excludes 0.5 for all three values of q.

The type I error behavior of the proposed intersection test was evaluated using simulation. Read counts for 10,000 genes were simulated using a Poisson distribution. The Poisson model was parameterized similarly to the above AI PG model: with mean read counts estimated from the population panel for three classes of coverage (low, medium, and high), the bias varied from 0.35 to 0.65 (19 values altogether), and all simulations were the null case (no AI). For each level of bias simulated, the simulated bias was used in the model and several deviations from the simulated bias, misspecification, were evaluated (1%, 2%, 3%, 5%, 10%, and 20%). For each combination of bias and misspecification (19 × 6) the posterior distribution of θ was estimated using the above AI PG model and intersection test (q = [0.4, 0.5, 0.6]). Note that the simulation is performed in relative bias amounts and not absolute. For example, 10% of 0.4 is 0.04 and 10% of 0.6 is 0.06. This was done for demonstration purposes, as the direction of the AI is arbitrary and this reveals more about the behavior of the bias.

Estimation of cis- and trans-effects

With the tester-cross design, allele-specific read counts can be used to separate cis- and trans-effects (Nuzhdin et al. 2012). For the sake of clarity in this article the reasoning in Nuzhdin et al. (2012) is reproduced here. In the testcross, the expression of the population allele i and the tester allele t in genotype i can be expressed as

$$E_{ii}=\mu +C_i+\frac{T_i+T_t}{2}\hspace{0.17em}\text{Line}$$

$$E_{ti}=\mu +C_t+\frac{T_i+T_t}{2}\hspace{0.17em}\text{Tester},$$

where$\mu$is the population mean,$C_i$is the cis-effects of allele i,$T_i$is the trans-effects of the allele i, and $T_t$ is the trans-effects of the tester allele. Several constraints are necessary to solve this system. For simplicity assume ${\displaystyle {\sum }_{i=1}^n{C}_i=0}$ and ${\displaystyle {\sum }_{i=1}^n{T}_i=0}.$ The expected difference between the alleles over the population is ${\displaystyle {\sum }_{i=1}^n\left(\left(E_{ti}-E_{ii}\right)/n\right)},$ which is equivalent to ${\displaystyle {\sum }_{i=1}^n\left(\left(C_t-C_i\right)/n\right)}=C_t.$ The tester trans-effects are estimated by $T_t=2\times \left({{\displaystyle \sum }}_{i=1}^n{E}_{ti}/n-\mu -C_t\right),$ which is $-C_t.$ The cis-effects for each allele are estimated as $C_i = E_{ii}-E_{ti}+C_t.$ The line trans-effects are estimated by $T_i=2\times \left(E_{ti}-\mu -C_t\right)-T_t.$ This formulation assumes equal coverage across lines. To account for the varying number of reads and the varying specificity of each cross, allelic expression was scaled to the number of non-allele-specific reads in the same region. That is, E_ii is estimated as the number of reads mapping uniquely to allele i divided by the number of reads mapping to the region but not assignable to allele i or the tester allele. Without scaling, the estimation is reflective only of the relative number of reads for that cross. In addition, this is a system of four unknowns and two equations, meaning there is not a unique solution for all four effects. The effects of C_t and T_t in this design are confounded with each other and will vary, depending on the exact estimation strategy for the population mean. However, the estimation of the effects C_i and T_i is not sensitive to the strategy for estimation of the population mean.

Data availability

The authors state that all data necessary for confirming the conclusions presented in the article are represented fully within the article.

Results

Allele-specific expression was compared between 49 testcrosses of females that were either mated or virgin. F₁ hybrids were produced between a w¹¹¹⁸ tester strain and genotypes from the DGRP and a Winters collection (Campo et al. 2013). We use a Bayesian test for AI that accounts for bias and controls type I errors (Figure S2). The ability to detect AI is related to coverage. Exons with allelic imbalance were underrepresented in genotypes with low levels of coverage (Table S2) of fewer than three reads per nucleotide. AI is concordant across the length of the transcript with very few (n = 82) exceptions (Table S3). These genes show AI in two or more exons in different directions, with at least one exon with more expression in the line and at least one other exon with more expression from the tester allele. The few differences across the gene may represent a frequency difference in isoform usage between alleles (Kurmangaliyev et al. 2015) and are consistent with data observed in other species (Skelly et al. 2011; Anders et al. 2012; Trapnell et al. 2013).

As with previous studies (Lawniczak and Begun 2004; McGraw et al. 2004, 2008; Mack et al. 2006; Kocher et al. 2008; Dalton et al. 2010; Smith et al. 2013) there is substantial difference in expression between virgin and mated flies. We observed 3048 exons from 1610 genes exhibiting differential expression at a false discovery rate of 0.05 (Benjamini and Hochberg 1995). In addition, 7 of the top 10 genes were reported previously in an analysis of gene expression in the heads of virgin and mated flies, which provides support for the present results (Dalton et al. 2010).

Of the 5391 exons analyzed, 4090 (76%) showed evidence of allelic imbalance in at least 1 cross. Fourteen regions had AI in >40 crosses (of 49 total), and 31 regions had AI in 100% of the crosses measured. The few that are different in all crosses likely result from regulatory mutations unique to the w¹¹¹⁸ strain. There were 3004 exons that had AI in ≤5 crosses, and 570 exons showed AI in <5 crosses. The distribution of θ is centered on 0.5 for both mated and virgin environments (Figure 2A). An average of 19% of exons show evidence for AI in any one cross. The proportion of exons with significant AI ranged from 8% of exons tested to 39% (Figure 2B and Table S3). Of the exons with allelic imbalance, one-third are also differentially expressed between environments.

Figure 2.

Distribution of AI. (A) Distribution of $\widehat{\theta }$ for mated (blue) and virgin (red) environments for all data. The distribution is centered at 0.5, indicating that there is no overall bias toward the line or the tester allele. (B) The proportion (y-axis) of exons significant for AI, determined by the intersection test, for each line (x-axis) in mated (blue) and virgin (red) environments.

Environmental effects on allelic imbalance

We looked to see whether there were any differences in AI between mated and virgin flies. The exons that show differences in estimates of allelic imbalance between mated and virgin flies are listed in Table S4. There were 93 genes that were significant in AI in mated flies only. Several genes associated with reproduction were significant for AI in mated flies only, including Arpc1, CG10433, CG17919, fry, heph, Sec5, Cg25C, and Shark. Mating has been shown to reduce life span (Chapman et al. 1995; Flatt 2011; Landis et al. 2015) and several genes reported to affect life span were also significant for AI in mated flies only (α-Man-Ia, bmm, mt:ATPase6, sNPF, and sm). Postmating immunity is also compromised in female fruit flies (reviewed by Peng et al. 2005; Lawniczak et al. 2007; McGraw et al. 2008) and the genes CG10433, CG17652, CG17919, Gs1l, α-Man-Ia, mtd, Sec5, and Shark are involved in processes such as apoptosis, phagocytosis, and general immune responses. In addition, Cyp4ac1, bmm, Gs1l, and sNPF are associated with metabolism, which is also altered by mating in fruit flies (McGraw et al. 2004, 2008; Dalton et al. 2010; Zhou et al. 2014).

Interestingly, there were two genes that have been shown to be directly involved in regulatory regions of the genome that were significant for AI in mated flies. cha is involved in cholinergic neuronal regulation and cis-regulation in the brain (Yasuyama et al. 1995). cnc encodes the Drosophila homolog of the Nrf2 transcription factor that is directly linked to metabolic regulation, stress response (Misra et al. 2011; Sykiotis et al. 2011; Pickering et al. 2013), and life span (Obata and Miura 2015). cnc has a direct influence on the oocyte nucleus and gurken, signaling (Guichet et al. 2001) innate immunity and lipid metabolism (Karim et al. 2015), and its overexpression can restore movement in a Drosophila model of Parkinson’s disease (Barone et al. 2011).

In virgin flies, there were 99 genes that were significant for AI. Reproductive-related activities were also represented in virgins as they were in mated flies (Cdc42, Cys, dnc, fl(2)d, mei-P26, qm, and zip). In addition, Tim10, a negative regulator of innate immune response, was also significant in AI in virgin females only, suggesting that upon mating this gene is differentially regulated in virgin but not in mated flies.

There were several genes involved in regulatory regions that were significant for AI in virgin flies only. Act57B contributes to developmental processes and cytoskeleton dynamics but has also been shown to play a role in courtship conditioning (Winbush et al. 2012). Kank, a gene that affects muscle tendon development, has been shown to be differentially expressed downstream of dsx (Lebo et al. 2009), suggesting AI regulation in the sex determination pathway in virgins only. p24-1 has sex-specific expression (Boltz et al. 2007) and has been shown to play a role in the neural signaling involved in female egg-laying defects and reproductive behavior (Saleem et al. 2012). In addition, ζCOP is involved in lipid storage and regulation of lipid droplets (Guo et al. 2008b) consistent with regulatory changes in lipid metabolism after mating in the fruit fly.

Cis- and trans-effects

We compared all estimates of AI in both environments (Figure 3). AI is largely concordant between environments (Figure 3 and Figure 5), a somewhat surprising result given the magnitude of expression differences reported for these conditions previously (Dalton et al. 2010) and in these data (see above) and the presence of ∼200 differentially regulated genes. We hypothesize that the large degree of concordance in expression regulation between environments is due to a robust gene regulatory network (Cubillos et al. 2014). To explicitly test this hypothesis, we can estimate cis- and trans-effects for each exon within the population based on the F₁-hybrid tester design (Nuzhdin et al. 2012). For exons significant for AI in at least 10 lines (n = 879) cis- and trans-effects were estimated for all lines in mated and virgin conditions separately (Figure 4 and Table S3).

Figure 3.

Estimates of AI in the mated vs. virgin environment. Three lines (r324, r365, and w47) were selected at random. $\widehat{\theta }$ for the mated environment (x-axis) and the virgin environment (y-axis) for all exons was evaluated. Each exon is a point on the graph. Estimates of AI are remarkably similar between mated and virgin environments. The red line in the panels is a regression while the blue line has an intercept of (0, 0).

Figure 5.

His3.3B: cis- and trans-estimates. This gene was chosen at random as an example to demonstrate how cis- and trans-effects were compared. Each circle represents one testcross/F₁-hybrid line. The top row compares cis- and trans-estimates within environment. The x-axis is the estimated trans-effect and the y-axis is the cis-effect. Cis- and trans-effects are negatively correlated for this exon (85% of exons show negative associations in both environments). The bottom row compares cis- and trans-effects across environments. The x-axis is the mated environment and the y-axis is the virgin environment. In this example, trans-effects are positively associated between mated and virgin flies. Cis effects are less correlated in this example but the correlation is positive (99% of all exons are positively correlated for both cis- and trans-effects across environments with a median correlation of 0.83 for cis and 0.95 for trans). The red line is the regression. Figure S4 shows the same relationships between cis- and trans-effects across all crosses and exons.

Figure 4.

Distribution of cis and trans for exons with significant AI. (A) Distribution of estimates of cis and trans for the mated environment. Trans-estimates appear to have a larger variance than the cis-estimates. Cis and trans are centered at zero, as expected. (B) Distribution of estimates of cis and trans for the virgin environment. Trans-estimates appear to have a larger variance than the cis-estimates. Cis and trans are centered at zero, as expected. (C) Distribution of the variance for cis- and trans-estimates for the mated environment. For each exon, the variance of the estimate is calculated. The variance in trans-estimates can be much larger than the variance in cis-estimates. (D) Distribution of the variance for cis- and trans-estimates for the virgin environment. For each exon, the variance of the estimate is calculated. The variance in trans-estimates can be much larger than the variance in cis-estimates.

The distribution of estimates (Figure 4, A and B) shows a larger range for trans-estimates than for cis-estimates. For each exon in mated and virgin environments we calculated the variance of cis- and trans-estimates (Figure 4, C and D). The variance of trans-estimates is much larger than the variance of cis-estimates in both environments, consistent with previous studies (Genissel et al. 2008). The direction of the estimates for cis and trans for a particular cross/environment is often different (63%, Table 1) and the association between estimates for cis and trans for a particular exon across crosses is negative 85% of the time (Figure 5, gene His3.3B; and Figure S4). In contrast, when comparing trans-effects for the same exon across environments the results are strikingly concordant (Figure 5 and Figure S4). Cis-effects are likewise similar across environments (Figure 5 and Figure S4).

Table 1. Direction of cis- and trans-effects.

	Direction trans
Direction cis	+	−	Total
+	9,090	22,300	31,390
−	11,869	10,621	22,490
Total	20,959	32,921	53,880

If the value of the estimate is positive, the direction is recorded as (+); similarly negative values are scored as (−). The direction of cis-effects is compared to the direction of trans-effects. There are (9090 + 10,621) cis- and trans-estimates that are concordant and (22,300 + 11,869) cis- and trans-estimates that are compensatory.

AI is a result of both cis and trans and the interaction between them (Figure 1). For each environment we modeled the effect of AI as a function of cis- and trans-effects and their interaction, using a linear model: Y_i = μ + C_i +T_i + C_i × T_i, where Y is AI, C_i is the estimate of cis, and T_i is trans for line i (i = 1–n_j). The association between cis-effects and AI was nearly always significant. For 75% of the exons tested, cis-effects explained at least 45% of the variation in expression (Figure 6 and Table S5). For 25% of exons, cis-effects explained >85% of the variance. Trans-effects added on average 8% to the proportion of variance explained. We found that 25% of exons had at least 10% of their variance explained by trans. Even after accounting for both cis and trans, the interaction between cis and trans explains an additional 11% on average. The full model explained at least 40% of the variance for 95% of the exons modeled and 90% of the variance for 50% of the exons (Figure 6, Figure 7, and Table S5).

Figure 6.

The distribution of R² for the regression Y_ij = μ + C_i + T_i + C_i × T_i, where Y is $\widehat{\theta },$ C_i is the estimate of cis, and T_i is trans for line i for each exon j. R² is an estimate of the contribution of cis-interactions, trans-interactions, and cis- by trans-interactions to AI. Distributions of R² are similar in the mated and virgin environments. Cis-effects explain at least 45% of the variation in expression for 75% of exons tested. Trans-effects explain 8% of the variance on average. The cis- by trans-interaction explains an additional 11% of the variance on average.

Figure 7.

Three exons displaying different patterns of association between cis-interactions, trans-interactions, and cis- by trans-interactions and $\widehat{\theta }.$ Each point is a testcross/F₁ hybrid. On the y-axis is the estimated AI ($\widehat{\theta }$) and on the x-axis in left, center, and right columns are cis, trans, and cis by trans. The top row represents the gene Men. This gene has ∼97% of the variance explained by the cis-effect and a negative association between cis and AI. The middle row represents the gene CHOp24. This gene has ∼97% of the variance explained by trans-effects. The bottom row represents the gene CG3036. This gene has ∼53% of variance explained by the cis- by trans-interaction. Figure S3 depicts lack of association between cis- and trans-effects with $\widehat{\theta }$ when all crosses and all exons are considered simultaneously.

We identified genes that are regulated completely by cis (R² > 0.90, n = 170), have a large trans-effect (R² due to trans >0.40, n = 50), or have a large cis- by trans-interaction effect (R² due to the interaction >0.40). Genes regulated primarily by cis were enriched for biological Gene Ontology processes involved in centrosome organization (P = 0.001439, 12 genes), centrosome cycle (P < 0.0001, 11 genes), and centrosome duplication (P = 0.001839, 10 genes). In addition, phototransduction (P = 0.004142, 8 genes), detection of light stimulus (P = 0.00949, 8 genes), response to light stimulus (P = 0.017, 11 genes), and detection of visible light (P = 0.046, 6 genes) were also enriched. Phototransduction genes are known to be altered by mating in Drosophila (Gioti et al. 2012; Landis et al. 2015).

Of the genes with a large trans-effect on AI, Chmp1 is involved in the negative regulation of the epidermal growth signaling pathway as well as involved in chromatin structure and cell cycle progression (Stauffer et al. 2001). HPS4 is also involved with chromatin, but in the silencing and negative regulation of gene silencing (Lee et al. 2009). jim is also involved in chromatin silencing (Mugat et al. 2015) and PIWI-interacting RNA processes in follicle cells (Saito et al. 2009). Finally, ps is spliced sex specifically (Telonis-Scott et al. 2009) and is a splicing factor (Seshaiah et al. 2001) and Vmat is involved in transmembrane transport as a splice variant that is involved in dopamine, serotonin, and octopamine transport (Greer et al. 2005). Both dopamine signaling and octopamine signaling are implicated in mediating the postmating response (Rezaval et al. 2014; Landis et al. 2015).

Discussion

This study supports previous findings of differences in expression between mated and virgin flies and goes further in identifying ∼200 genes with AI differences between environments. Yet, when AI is detected in both environments, it is concordant and cis- and trans-effects are concordant across the environments. When AI is detected, cis is almost always significant and in this first population-scale study of natural variation within a species, trans-effects and cis- by trans-interactions are sizable. Trans-effects are evidence of molecular connections across loci and cis- by trans-interactions are evidence of genetic variation across loci influencing the phenotype of gene expression. In stark contrast to the cis_M–cis_V and trans_M–trans_V comparison across environments the cis- and trans-effects within an environment (cis_M–trans_M; cis_V–trans_V) are markedly not concordant in this study (Table S6). The cis/trans combinations +/+ and −/− are concordant while +/− and −/+ are compensatory. Here 63% of the effects within a line/exon are compensatory and across lines 85% of the associations between cis- and trans-effects have a negative slope, leaving us with an unambiguous finding of large-scale compensatory effects in expression regulation. Large-scale compensatory effects within the same species are also supported by a recent elegant experimental study in Caenorhabditis elegans (Paaby and Rockman 2014).

There is a strong consensus emerging from the analyses of a variety of model organisms: in yeast (Kvitek et al. 2008; Artieri and Fraser 2014), fly (Wang et al. 2008; Coolon et al. 2014, 2015; Graze et al. 2014), and mouse (Goncalves et al. 2012; Crowley et al. 2015; Pinter et al. 2015) that cis- and trans-effects are compensatory. In intraspecific D. melanogaster F₁’s, compensatory interactions were observed in 79% of cases, and in interspecific D. simulans/D. sechelia and D. melanogaster/D. simulans crosses, they corresponded to 73% and 87% of cases (see summary in Coolon et al. 2014). How does one explain this strong and phylogenetically ubiquitous pattern?

One frequent explanation for compensatory effects is that trans- and cis-factors coevolve, as there is a stabilizing selection for overall transcript level (Wittkopp et al. 2004, 2008b; Landry et al. 2005; Graze et al. 2009, 2012; Fontanillas et al. 2010; Stevenson et al. 2013; McManus et al. 2014). A neutral trans-mutation is envisioned that spreads jointly with a slightly deleterious cis-mutation (see Takahasi et al. 2011 for a helpful summary). Takahasi et al. (2011, p. 15,279) demonstrate via extensive simulation that “compensatory cis-trans interactions gradually accumulated over time.” Also of note, in their simulations (Takahasi et al. 2011, p. 15,279) “background dependency of relative allelic expression was not observed within species.” Within species, cis- and trans-effects cannot coevolve as they are not cotransmitted. In this experiment in Drosophila the likelihood of long-range LD in flies accounting for ∼85% compensatory vs. ∼15% concordant cis–trans combinations is vanishingly small as the scale of this effect is known to be very small (Pool et al. 2012). What is then the evolutionary explanation of such compensation within species?

Gene expression evolves under a house-of-cards model of stabilizing selection (Hodgins-Davis et al. 2015), maintaining population genetic variation (Jin et al. 2001; Gibson and Dworkin 2004; Nuzhdin et al. 2004). This variation contributes to phenotypic diversity and is widespread. If instead of thinking about individual loci we focus on gene regulatory networks (GRNs), then logically there must be variation among loci. How then would that variation be measured in the context of cis- and trans-effects? If each locus contained variation that affected expression at that locus, we would expect a large number of cis-effects, as observed here. Yet, GRNs are inherently models of trans-effects, as the connections among genes in the pathway are trans. For these data a pathway analysis of the virgin environment modeled regulation of the sex hierarchy, using structural equations (Fear et al. 2015). This analysis suggested that there were 754 genes that were linked to the core sex hierarchy. We conducted a similar path analysis of the mated environment and found 380 (50%) were in common, suggesting a large overlap in the trans-regulation of genes upstream and downstream of the sex hierarchy in both environments.

In thinking about cis- and trans-effects in the context of a GRN the relative directions of the cis- and trans-effects must be considered carefully. For example: If an allele 1 in genotype 1 is upregulated due to cis-regulatory mutation (+ effect) and there is also a significant trans-effect estimated for the same allele, what direction—upregulating or downregulating (− effect)—would it more frequently have? An alternate effect of the allele in cis and trans suggests that the overall regulation of the GRN is constrained. Compensatory effects of the allele within a GRN may explain widespread observations of GRN robustness.

GRN robustness has been explicitly formulated to account for the prevalence of compensatory cis–trans interactions (Denby et al. 2012; Bader et al. 2015). Denby et al. (2012) have proposed that negative feedback controlling the level of RNA expression could be a common mechanism to buffer effects of regulatory variants in yeast. Screening for autoregulated transcription factors in yeast, Denby et al. (2012) found ROX1 to be under strong negative feedback. Mutant experiments showed that this negative feedback confers robustness to the expression of ROX1 in the face of naturally occurring allelic variants present in a set of divergent yeast strains. This study demonstrated for a single gene that negative feedback could act as a buffering mechanism for regulatory variants. Bader et al. (2015) quantified buffering by feedback against naturally occurring regulatory variants genome-wide to be ∼15%. Note that when buffering arises from feedback, it is not necessary to invoke additional genetic variation across the genome to explain the trans-effects.

If GRNs are generally robust, they have implications in human health research. Analysis of the structure of GRNs in the K562 cancer cell line suggests poor buffering compared to the noncancer cell line GM12878 (Albergante et al. 2014). Most of the genome-wide association study associations are between noncoding regions and disease (reviewed in Zhang and Lupski 2015). One potential explanation is that genes are misregulated in disease (reviewed in Lee and Young 2013). As GRN robustness should buffer against misregulation, it implies that disease may be a result of lack of effective buffering. This may make the hunt for genes underlying disease a little more straightforward, as rather than looking for direct links with phenotype, lack of robust regulation can be used to identify potential candidate pathways.