Strong epistatic and additive effects of linked candidate SNPs for Drosophila pigmentation have implications for analysis of genome-wide association studies results

Over the past years, genome-wide association studies (GWAS) have become the method of choice to identify functionally relevant variation for a wide set of traits [1, 2]. Despite the increased resolution of GWAS, compared with experimental crosses, the mapping resolution is frequently not sufficiently high to map the causative variants. The haplotype blocks identified by GWAS often carry multiple variants and it is not clear how many of them and which ones are causative. Many approaches have been proposed to prioritize candidate single nucleotide polymorphisms (SNPs). While rather advanced approaches are available for protein-coding variants (e.g. SIFT, PolyPhen-2, MutationAssessor [3–5]), the characterization of regulatory SNPs suffers from low sequence conservation and incomplete knowledge about functionally important motives. The large number of candidate SNPs also precludes experimental inference of epistasis since the number of possible haplotypes scales exponentially with the number of loci. Indeed, only a few studies analyzed the fitness consequences of all combinations of small sets of mutations (reviewed in [6]). Consequently, the experimental validation of GWAS results is often confined to cell culture or knockdown of candidate genes.

Drosophila pigmentation has a long tradition as model to study the genetic basis of phenotypic variation within and between species [7]. Pool-GWAS in three natural D. melanogaster populations identified three closely linked SNPs as major contributors to female abdominal pigmentation [8, 9]. These three SNPs are located within a 208-bp window of the t-MSE enhancer, which controls the expression of the pigmentation gene tan (t) in the abdominal epidermis [10]: X-9121129 (SNP1), X-9121094 (SNP2), and X-9120922 (SNP3) (Additional files 1 and 2: Table S1 and Figure S1). An independent study, which does not rely on Pool-Seq [11] of extreme phenotypes but analyzed individual isofemale lines, confirmed the importance of this regulatory region for natural variation in female abdominal pigmentation [12]. Using the D. melanogaster Genetic Reference Panel, which was established from a North American population, two of the SNPs identified by Pool-GWAS (SNP1, SNP2) were also detected [12].

Building on the previously identified GWAS signal in the regulatory region of the pigmentation gene tan (t), we performed the first in vivo characterization of all GWAS candidate SNPs in a quantitative trait locus (QTL) region. It could have been expected that a single causative variant was linked to two neutral SNPs. However, we find that all three candidate SNPs have strong additive and complex epistatic effects on pigmentation. Interestingly, only two agree with the effect predicted by GWAS; the third one goes in the opposite direction.

We used transgenic lines to evaluate the functional impact of SNP1, SNP2, and SNP3. We constructed rescue transgenes, each containing the t-MSE with one of the eight combinations of the three SNPs fused to the hsp70 minimal promoter and the t complementary DNA (cDNA). Each transgene was inserted at the same genomic position using phiC31 integrase-based transgenesis [13]. The effect of the different SNP combinations on abdominal pigmentation was assessed in females with an otherwise identical genetic background, mutant for t (allele t ^d07784 [14]). The different genotypes showed pronounced pigmentation differences (Fig. 1, left). For a quantitative analysis of pigmentation differences, we analyzed 15 individuals for each genotype and quantified the pigmentation in segments A5, A6, and A7 on mounted abdominal segments (Additional file 3: Raw Data; Fig. 1, right). We detected clear differences between the genotypes, covering a similar phenotypic space as wild-type (WT) flies (Fig. 1). We determined the main effects of all SNPs and their interactions using ANOVA (Additional file 4: Table S2). All SNPs had very significant main effects in all segments (p < 0.001), explaining up to 71% of the total variance in pigmentation (Table 1). SNP1 had the strongest main effect, explaining 29–53% of the variation. Epistatic effects were considerably weaker accounting for 6–11% of the total variation. For all segments, a model with epistatic effects provided a significantly better fit to the data than a model with main effects only (p ≤ 0.001). The presence of epistatic effects in a regulatory element is in line with results for transcription factor binding sites [15].

	A5		A6		A7
	p value	Eta Squared	p value	Eta Squared	p value	Eta Squared
1	<0.001	0.422	<0.001	0.533	<0.001	0.295
2	<0.001	0.059	<0.001	0.097	<0.001	0.132
3	<0.001	0.130	<0.001	0.081	<0.001	0.180
1 × 2	0.206	0.004	<0.001	0.025	0.394	0.002
1 × 3	0.260	0.003	<0.001	0.032	0.074	0.009
2 × 3	<0.001	0.058	0.013	0.010	<0.001	0.047
1 × 2 × 3	<0.001	0.044	<0.001	0.042	0.025	0.015

1, 2, and 3 in the first column correspond to SNP1, SNP2, and SNP3. P values for the different effects are indicated as well as Eta Squared, a measure of the effect size

The effects of the three SNPs on the pigmentation were largely consistent across segments. Nevertheless, some differences between segments were noticed (Fig. 1, Table 1, Additional file 4: Table S2). SNP1 had the strongest main effect in A6 and the weakest main effect in A7, whereas the opposite was observed for SNP3. SNP2 had the strongest main effect in A7 and the weakest main effect in A5. The two-way interactions involving SNP1 were only significant in A6, while all other interactions were significant across all three segments (Table 1). We attribute this heterogeneity among segments to differences in expression of genes regulating tan. The transcription factors Abdominal-B and bric-à-brac (BAB1 and BAB2), which regulate abdominal pigmentation and are expressed at an increasing level along the anteroposterior axis in pupal epidermis [16, 17], are good candidates, although the polymorphic SNPs are not located in characterized binding sites.

Importantly, not all Dark alleles predicted by GWAS increased pigmentation in our functional assays. While the Dark alleles of SNP1 and SNP3 increase pigmentation, the Dark allele of SNP2 has the opposite effect. Indeed, the darkest genotype corresponds to the combination D.L.D, whereas the lightest genotype corresponds to the opposite combination L.D.L. Thus, opposite to the predictions of three GWAS studies, SNP2 does not affect pigmentation in the same direction as the linked SNPs. A closer inspection of the underlying haplotype structure shows that despite the high recombination rate and large population size of D. melanogaster, the absence of the haplotypes with extreme pigmentation prevented the correct inference of the effects of the three causative SNPs [9].

The functional impact of natural variants on a given trait is one of the most pressing questions in genetics. GWAS and QTL mapping studies substantially advanced our understanding of the contribution of genes at different, well-separated locations in the genome. Nevertheless, the analysis of linked variants has been severely limited by linkage disequilibrium. In this report, we use, to our knowledge for the first time, transgenesis to break up linkage disequilibrium among closely linked sites to determine their contribution to the trait of interest (pigmentation) and to study the underlying genetic architecture. Our result of all candidate SNPs of the tan locus contributing to the pigmentation demonstrates that multiple SNPs could be contributing to a single QTL. The current literature already indicates that this phenomenon may be not uncommon: several multiparent advanced intercrosses noted that a given QTL explained a significant fraction of the variation, but variants segregating in the corresponding region did not [18–21]. The authors interpreted this discrepancy as evidence for multiple functional alleles contributing to the trait [18–20]. Similarly, multiple alleles have been suggested to contribute to a given eQTL [22, 23]. The strongest independent support for multiple functional variants comes from a study on the Alcohol dehydrogenase (Adh) in D. melanogaster, where the authors divided the gene into three parts with multiple SNPs and showed that each part contributed to Adh expression [24].

Analyzing all eight possible combinations of the three adjacent functional intra-specific variants provided the first in vivo analysis of the genetic architecture of a regulatory module in a higher eukaryote. Although additive effects explain most of the trait variation, up to 11% of the variation arises from epistatic interactions between the three sites. The regulatory architecture of pigmentation is similar in the three abdominal segments, but some noticeable differences are present. The about tenfold difference in the variance, explained by epistatistic two-way interactions involving SNP1, results in a significant effect only in A6. In addition to cuticle pigmentation, tan is also involved in vision [14]. The expression of tan in photoreceptors is responsible for the hydrolysis of carcinine, a different substrate than Beta-alanyl-dopamine which is responsible for cuticle pigmentation [25, 26]. We anticipate that such other roles of tan could have their own regulatory architecture, which differs from the one of abdominal pigmentation.

Our observations have direct implications for other GWAS studies: (1) consistent with a previous report concluding that complex traits are mainly governed by additive effects [27], we find that the main effects are much stronger than epistasis; (2) linkage disequilibrium between causal SNPs could result in predicting effects in the wrong direction. Very importantly, this is not an artifact limited to this specific case in Drosophila, it will apply to all GWAS studies [28]. The confounding effect of multiple causative alleles at the same locus has also been described for several other systems (e.g. [28, 29]).

We propose that the possibility of linked causal variation needs to receive more attention and emphasize the importance of experimental systems, such as fruit flies, which allow functional testing of candidate SNPs in vivo as well as the inference of the underlying genetic architecture.