Genetic studies on human populations: trials and tribulations

Understanding the relationship between genotypic variation and variation in complex and quantitative trait phenotypes in human populations remains an ongoing challenge (Mackay 2014; Boyle et al. 2017). Yet, being able to make causal links between DNA variants and variation in organismal traits will result in a major leap forward to ultimately enable predicting both healthy phenotypes (e.g., lifespan, body weight) and susceptibility to disease.

Rare diseases that occur with high incidence within families or inbred populations and with high penetrance and that arise from single mutations with large effects can be subjected to classical linkage analysis in pedigrees to identify the causal DNA variant. However, most traits and common diseases have a complex genetic basis and their manifestation depends on multiple segregating genes, and interactions between them and environmental factors. Analyses of such traits require population-based association mapping approaches.

The classic measure that partitions phenotypic variation in genetic and environmental variance is the heritability (Falconer and Mackay 1996). The narrow sense heritability (h 2) is the portion of additive genetic variance that contributes to the total phenotypic variance that can be transmitted from parents to offspring. This measure of heritability is important for animal and plant breeding as it can be readily estimated from correlations among relatives and predicts response to artificial selection. The broad sense heritability (H 2) represents the fraction of phenotypic variance in a population that can be attributed to all sources of genetic variation, including both additive and non-additive (dominance, epistatic) genetic variance. High values of H 2 provide favorable scenarios for genome-wide association (GWA) analyses. However, even when H 2 is substantial, establishing causality of candidate polymorphisms to the phenotype in human GWA studies is challenging due to large blocks of linkage disequilibrium (LD)—the correlation in allele frequencies between polymorphic sites—in the human genome (Reich et al. 2001). Furthermore, genotype by environment interactions may result in lack of reproducibility between populations, when associations that are significant in a population in one environment cannot be resolved in a population under different environmental conditions. Gene–gene interactions (epistasis) may also result in failure to reproduce associations in different populations with different allele frequencies of causal loci (Mackay 2014, 2015). Moreover, when GWA analyses on individuals with vastly different genetic backgrounds are combined, spurious associations may be observed due to admixture (ethnic stratification; Pritchard and Rosenberg 1999). In addition, often the causal relationship between the trait and genes of unknown function or without biological context remains enigmatic.

The structure of the human genome, which contains regions of linkage disequilibrium, and differences in developmental history, lifestyle, and social and physical environments, requires phenotyping very large populations to resolve significant effects of candidate polymorphisms on quantitative traits (e.g., height; Wood et al. 2014) and susceptibility to common diseases (e.g., Zeggini et al. 2008). Such large population studies, often with tens of thousands of subjects, have identified many candidate genes that contain single nucleotide polymorphisms (SNPs) associated with common diseases, such as diabetes (Gaulton et al. 2015; Mohlke and Boehnke 2015) or cardiovascular disease (CARDIoGRAMplusC4D Consortium et al. 2013) as well as a vast number of copy number variants associated with schizophrenia (Stefansson et al. 2008, 2009). The concept of “common variants for common diseases” (Reich and Lander 2001), however, needs to be re-evaluated, as many rare variants with relatively large effects may cumulatively contribute to a significant fraction of disease risk within a population (Keinan and Clark 2012; Bomba et al. 2017).

Whereas certain human parameters, such as height, can be quantified precisely (Wood et al. 2014), many other phenotypes are difficult to quantify consistently. For example, evaluation of propensity to use alcohol or drugs often relies on self-reported questionnaires, and alcohol use and addiction may be confounded by psychiatric disorders or stress conditions. In addition, different investigators may use different criteria for alcohol use and addiction (Morozova et al. 2014). Furthermore, criminalization of illegal drug use makes it difficult to recruit subjects for GWA studies to explore genetic variants that may predispose to substance abuse. Genetic susceptibility to toxic exposure (e.g., heavy metals) is also hard to quantify as it is often confounded by exposure to multiple toxicants, and toxic effects may become evident long after the initial exposure. Similarly, cumulative effects of oxidative stress, for example, due to exposure to pesticides, which may lead to neurological disorders such as Parkinson’s disease (Dardiotis et al. 2013), are challenging to quantify. In addition, many human disease phenotypes, e.g., autism, are heterogeneous and present a spectrum of manifestations and severity, which may not necessarily arise from the same genetic risk factors (Chahrour et al. 2016). Finally, most human studies on genetic disease risk focus on extreme states, i.e., presence of disease, whereas clinical manifestation may result from surpassing a threshold of normal trait distribution within the population. For example, sociopathic aggression represents an extreme along the spectrum from assertiveness to shyness. Where to draw the threshold in this case is not unambiguous, but may be influenced by sociocultural boundaries.

Despite significant advances in studies on the genetic underpinnings of complex traits in human populations, both in health and disease, the challenges and impediments inherent in these studies illustrate a compelling need for comparative studies in model organisms. Studies on the genetic architectures of complex traits in Drosophila melanogaster have provided many insights with translational potential for human population genetics.

Drosophila: an advantageous model for the genetic dissection of complex traits

Many of the constraints encountered in human population studies can be overcome in the D. melanogaster genetic model system. We can inbreed flies, thus enabling strict control over the genetic background, and we can rear virtually unlimited numbers of genetically identical individuals. Drosophila has a 2-week generation interval under standard laboratory conditions and can be grown rapidly under well-controlled environmental conditions, without regulatory restrictions and at relatively low cost. We can readily substitute chromosomes, generate transgenic flies, and use CRISPR/Cas9 technology for gene editing, which allows us to create “designer” genotypes (Bassett and Liu 2014). About 75% of human disease-associated genes have a Drosophila ortholog (Reiter et al. 2001); and a wide range of morphological, physiological, behavioral, and life history traits, many with relevance to analogous human traits, can be quantified precisely. Hence, evolutionary conservation of fundamental biological principles empowers translational inferences across phyla, including humans.

The genetic architectures of complex traits can be explored by large-scale mutational screens, conventional quantitative trait locus (QTL) linkage mapping approaches, and GWA analyses. Large collections of Drosophila stocks with transposon insertions, such as P-element insertions, which often target promoter regions, or piggyBac inserts, which disrupt genes without promoter bias, are publicly available for mutational screens (Zhai et al. 2003; Thibault et al. 2004; Bellen et al. 2011). Transposon-based mutagenesis screens have shown that complex traits present large mutational targets, which are affected by mutations in many genes, often representing a substantial fraction of the genome, implicating extensive pleiotropy (Sambandan et al. 2006; Mensch et al. 2008; Magwire et al. 2010; Zwarts et al. 2011).

Whereas mutational screens identify genes that contribute to manifestation of the trait under study, linkage and association mapping analyses, respectively, identify a subset of those genes that harbor polymorphisms that contribute to variation in the phenotype. Genome-wide studies can identify candidate genes that contribute to phenotypic variation, while subsequent mutational analyses can provide evidence that disruption of the gene indeed affects the trait and provide a focus for detailed mechanistic studies. Thus, these approaches are complementary. In Drosophila, GWA analyses became possible with the development of the D. melanogaster Genetic Reference Panel (DGRP; Mackay et al. 2012; Huang et al. 2014).

The D. melanogaster Genetic Reference Panel (DGRP)

The DGRP is a population of 205 inbred wild-derived lines, which were generated by subjecting the offspring of individual gravid females collected from the Farmer’s Market in Raleigh, North Carolina, to 20 generations of full sib inbreeding (Mackay et al. 2012; Huang et al. 2014). Inbreeding minimizes genetic variation among individuals within each line, while genetic variation between the lines reflects the variation in the population from which they were derived (Fig. 1a). The lines were sequenced to high coverage (average ~ 27-fold) and a total of 4,565,215 naturally occurring molecular variants were identified, including 3,976,011 high-quality single or multiple nucleotide polymorphisms; 169,053 polymorphic insertions; 293,363 polymorphic deletions; and 125,788 polymorphic microsatellites. The vast majority of the DGRP lines are genetically unrelated, except for a few lines which may have resulted from random sampling of related individuals from the population (Fig. 1b). Residual heterozygosity in the lines is largely due to segregating polymorphic inversions, which represent islands of diversity within otherwise homozygous genomes. In addition, 53% of the DGRP lines are infected with the maternally transmitted endosymbiotic bacterium, Wolbachia pipientis (Huang et al. 2014).

Fig. 1
figure 1

Modified from Mackay et al. (2012)

The Drosophila melanogaster Genetic Reference Panel. a Diagram of the derivation of DGRP lines from a natural population. b Genomic relatedness among DGRP lines. The distribution of the relationship between all DGRP lines and the reference sequence is displayed as a box plot. From Huang et al. (2014). c Decay of LD (average R-squared) in bp as a function of physical distance.

Full size image

It should be noted that lines that survived inbreeding have been purged of highly deleterious alleles and selection may have occurred for epistatic interactions among variants that protect fitness during inbreeding. Also, long-term maintenance in the laboratory may give rise to adaptation to the laboratory environment. Despite these limitations, the DGRP provides a rich collection of natural variants that have survived the sieve of natural selection and are a treasure trove of genetic variation that can be harnessed for GWA studies. All traits measured in the DGRP to date show extensive phenotypic variation that far exceeds variation observed among standard laboratory inbred strains or recombinant inbred lines. Since individuals within each line are genetically virtually identical, the same genotypes can be measured repeatedly, which enables accurate quantification of phenotypes. Since phenotypic measurements can be spread over time, environmental noise can be randomized, further minimizing error in phenotypic estimates. Furthermore, since the DGRP is a publicly available resource, different laboratories can measure and correlate phenotypes on the same genotypes. Table 1 lists GWA analyses performed to date using the DGRP.

Table 1 Genome-wide association studies in the DGRP
Full size table

One example of the translational potential between flies and humans is a study which examined genome-wide variation in gene expression with phenotypic variation in alcohol sensitivity across 40 DGRP lines followed by validation of candidate genes through transposon-mediated mutagenesis (Morozova et al. 2009). One gene that emerged as a particularly interesting candidate gene was Men, which encodes malic enzyme. Malic enzyme represents a metabolic switch between energy production and lipid biosynthesis. Since excessive alcohol consumption in people results in fatty liver syndrome, a subsequent study in the Framingham Heart Study Offspring Cohort population focused on the cytoplasmic Malic Enzyme 1 (ME1) gene and identified polymorphisms associated with variation in alcohol consumption, which would not have been otherwise detected in a GWA study in this limited size population (Morozova et al. 2009).

Another example of how studies on the DGRP can guide investigations on human disease-associated genes is the recent identification of candidate modifier genes in a Drosophila model for retinitis pigmentosa (Chow et al. 2016).

Statistical considerations for GWA analyses

Large LD blocks in the human genome enable the use of tagging SNPs for GWA studies (Reich et al. 2001; Ke et al. 2004). In contrast to the human genome, LD in the D. melanogaster genome decays rapidly with physical distance, on average within a few hundred base pairs (Fig. 1c; Mackay et al. 2012). The advantage of low LD is that associated SNPs are likely causal or very near the causal variant. However, low LD prevents the use of tagging SNPs and mandates whole genome sequences for GWA analyses. Because the X chromosome experiences a lower effective population size, LD decays more slowly on the X chromosome than on the autosomes, and LD is notably increased in areas of reduced recombination, i.e., in the vicinity of autosomal telomeres and centromeres, and in polymorphic inversions, where recombination is suppressed (Mackay et al. 2012). Furthermore, due to the restricted population size of the DGRP, variants with minor allele frequencies of less than 5% may exhibit elevated long-range LD, since such rare variants will be in perfect LD with all variants that are private to a particular DGRP line. Thus, polymorphisms with minor allele frequencies of less than 5% should be excluded from GWA analyses.

The gold standard for genome-wide statistical significance in human GWA studies is a Bonferroni-adjusted threshold for multiple testing. GWA analyses in the DGRP are confined to only 205 lines with several million tests. Consequently, we seldom observe SNPs that show associations with P values that exceed a strict Bonferroni-corrected threshold. However, quantile–quantile plots (Q–Q plots) can indicate deviations from linearity between the distributions of observed and predicted values, indicating enrichment of true positive associations at P values which are orders of magnitude below the Bonferroni threshold. For most GWA analyses in the DGRP, we can identify approximately several hundred polymorphisms associated with phenotypic variation at an empirical threshold of P < 10− 5, supported by Q–Q plots (Weber et al. 2012; Swarup et al. 2013; Harbison et al. 2013; Morozova et al. 2015; Shorter et al. 2015; Garlapow et al. 2015; Carbone et al. 2016). Since all polymorphisms in the DGRP are known, those polymorphisms represent the top SNPs and/or insertions–deletions, which are associated with variation in the trait under study.

Mutational analyses or targeted RNAi experiments can be used to assess whether candidate genes that harbor such associated polymorphisms themselves affect the trait when their expression is disrupted. Such validation on a sample of candidate genes for which viable mutants are available establishes an empirical false discovery rate at the gene level, albeit not at the level of molecular variants. Previous studies have shown that we can validate about 60–80% of candidate genes using this approach (Weber et al. 2012; Swarup et al. 2013; Harbison et al. 2013; Morozova et al. 2015; Shorter et al. 2015; Garlapow et al. 2015; Carbone et al. 2016). Within pleiotropic genes different SNPs can be associated with variation in specific organismal traits (Carbone et al. 2006; Wang et al. 2010).

We can assess to what extent candidate genes with associated polymorphisms can be organized in interaction networks based on known genetic or physical interactions, and we can estimate the likelihood that a similar size network would emerge if the same number of genes were chosen at random. Here, the P value for significance of the network is not encumbered by multiple test considerations. Moreover, genes that form networks show very low empirical false discovery rates and validate at a high rate when subjected to mutational analysis (Fig. 2). However, only a restricted subset of genes with associated polymorphisms can be interconnected to form a network, although some of the remaining genes can sometimes be implicated in similar biological processes represented by the network through gene ontology analyses.

Fig. 2
figure 2

Modified from Swarup et al. (2013) and Arya et al. (2015)

Genetic networks for variation in olfactory behavior. a A network of interactions among candidate genes associated with variation in response to the odorant benzaldehyde among DGRP lines. Candidate genes are indicated by rectangles, missing genes (i.e., genes without significant associations) by triangles, and metabolites by circles. Components of the network associated with distinct interconnected cellular processes are highlighted by the colored backgrounds. b Validation of the connectivity of the predicted network. Transposon insertion in the Pkc53e locus results in a two-fold increase in gene expression accompanied by aberrant olfactory behavior. qRT-PCR of transcripts of genes connected in the network shown in panel a in a Pkc53e P{MiET1}-insertion mutant show increased expression levels compared to control, corroborating their functional connectivity. Error bars indicate SEM. *0.01 ≤ P ≤ 0.05; **0.001 ≤ P ≤ 0.01; ***0.0001 ≤ P ≤ 0.001. c An extended network of candidate genes associated with variation in responses to 14 structurally different odorants based on GWA and extreme QTL mapping analyses. Note the absence of missing genes in this network.

Full size image

Extreme QTL mapping

We can increase the power of GWA analyses by generating advanced intercross populations (AIP) from DGRP lines. There are several strategies for generating such AIPs. We can select a small number of lines with extreme phenotypes and cross them for many generations (Swarup et al. 2013; Shorter et al. 2015); we can select a random number of DGRP lines (Huang et al. 2012; Morozova et al. 2015; Carbone et al. 2016); or, we can select DGRP lines that are genetically unrelated with minimal residual heterozygosity, and free of inversions and infection of the symbiotic bacterium, W. pipientis (Garlapow et al. 2017). The lines can be crossed in a round-robin crossing design (Fig. 3) or a diallel crossing scheme (Griffing 1956) to generate a base population. Many generations of intercrossing in large population sizes to minimize loss of diversity due to drift will result through recombination in a virtually unlimited number of unique genotypes. We can now select and pool individuals with extreme phenotypes, subject the pools to bulk DNA sequencing, and identify alleles that differentially segregate among the phenotypic extremes (Fig. 3). This experimental approach, known as “extreme QTL mapping” (Ehrenreich et al. 2010), resembles case–control studies in human genetics, where allele frequencies associated with disease status are contrasted with those among unaffected age-matched individuals.

Fig. 3
figure 3

Diagrammatic representation of the generation of advanced intercross populations and their application for extreme QTL mapping. Note that replicate AIPs are always constructed from a base population derived by round-robin crossing of the same parental lines and that sexes are analyzed separately

Full size image

The advantages of the AIP design are that alleles that are present at low frequencies in the DGRP and may have large phenotypic effects, but cannot be analyzed by conventional GWA due to the risk of spurious LD, will be present at higher frequencies in the AIP base population, and hence their effects can be assessed. The diversity of genotypes generated through intercrossing increases statistical power, provided the pools for bulk DNA sequencing contain sufficiently large numbers of individuals. Indeed, extreme QTL mapping studies using DGRP-derived AIPs can resolve trait-associated variants that surpass a Bonferroni threshold for multiple tests. It is noted, however, that generating an AIP results in loss of polymorphisms that are not represented among the lines used to generate the base population. Nevertheless, AIPs derived from ~ 40 DGRP lines still captured a substantial amount of the genetic variation present in the DGRP (Huang et al. 2012).

From flies to human genetics: common concepts

The transcriptional niche

Genes do not act in isolation, but form part of functional expression networks. Introduction of a mutation that alters gene expression can result in altered transcript abundances of a suite of genes in addition to that of the target gene (Anholt et al. 2003). Analyses of correlations among transcripts of 40 DGRP lines reared under standard growth conditions showed a modular organization of the transcriptome, in which 10,096 genetically variable transcripts could be clustered in 241 modules, such that the genetic correlation of transcripts within each module was maximized and the genetic correlation of transcripts between modules minimized (Ayroles et al. 2009). This modular organization, however, is dynamic and changes under different environmental conditions (Zhou et al. 2012). Analyses of correlated transcript abundances with variation in gene expression of a focal gene define a group of correlated transcripts, the size of which depends on the statistical correlation threshold. We have designated such a modular array around a focal gene as the gene’s “transcriptional niche” (Fig. 4a; Arya et al. 2010). Gene ontology analyses of transcriptional niches can provide functional contexts to genes of unknown function or indicate pleiotropy of the focal gene (Fig. 4b, c).

Fig. 4
figure 4

The transcriptional niche. a Gene X gives rise to transcript X the abundance of which is correlated with that of other transcripts, indicated by circles of which the color shades indicate different statistical levels of correlation. b Different segments of the transcriptional niche of gene X can be recruited to give rise to different phenotypes, indicating that the transcriptional niche can serve as a conduit for pleiotropy. c The transcriptional niche is plastic and can change, for example, as a function of developmental stage, accompanied by a switch in its relationship to the organismal phenotype, as illustrated in this diagram, where purple circles indicate transcripts that are uniquely associated with the transcriptional niche of gene X in adulthood. d Transcriptional niches of different genes can overlap and interact to modify the organismal phenotypes associated with genes X and Y and give rise to different phenotypes, here illustrated as phenotype Z

Full size image

The concept of the transcriptional niche is of relevance to human genetics, but often underappreciated. Identification of a SNP associated with a disease phenotype does not necessarily mean that the gene harboring this SNP is directly causal to the phenotype, since its effect may be mediated through a shift in its transcriptional niche, such that modulation of expression of a different gene may actually be causal to the phenotype (Fig. 4d).

Epistasis, an inconvenient truth

Early QTL mapping studies on D. melanogaster recombinant inbred lines provided evidence for epistasis (Long et al. 1995; Leips and Mackay 2000; Dilda and Mackay 2002; Montooth et al. 2003). Subsequent studies on olfactory behavior, sleep, and waking activity, using chromosome substitution lines in which different chromosomes extracted from DGRP lines were introduced in a common genetic background with and without single P-element insertional mutations, provided evidence for suppressing epistasis, i.e., epistatic interactions tend to suppress the effects of new mutations (Yamamoto et al. 2009; Swarup et al. 2012). However, it remained unclear how prominent the contribution of epistasis is to the genetic architecture of complex traits.

When results from GWA analyses in the DGRP on startle behavior, recovery time from a chill-induced coma, and resistance to starvation stress were compared to the results from extreme QTL mapping experiments in DGRP-derived AIPs, SNPs identified by extreme QTL mapping were different from those identified by GWA analysis for all three traits (Huang et al. 2012). However, gene ontology and network analyses revealed that candidate genes implicated by both analyses converged on the same biological processes. Pairwise analyses of epistasis revealed extensive networks of epistatic interactions connecting results from the two analyses (Huang et al. 2012). A similar result emerged from studies on the genetic underpinnings of olfactory behavior, where combined GWA and extreme QTL mapping analyses implicated a network of genes associated with neural development and function as a substrate for variation in the behavioral phenotype (Swarup et al. 2013). This observation was extended and further confirmed by a GWA analysis that identified epistatic partners which interact with transposon-tagged mutants of two neurogenic genes, Sema-5c and neuralized, both of which have large effects on olfactory behavior (He et al. 2016). Similarly, transgenic expression of a misfolded mutant of human preproinsulin, proposed as a model for neonatal diabetes mellitus, causes morphological defects in adult flies; when crossed to different DGRP lines, F1 offspring displays a continuous range of morphological phenotypes, indicating genetic background-dependent modulation of the effect of the transgene (Park et al. 2014). Epistatic interactions between SNPs identified through GWA analysis and SNPs identified through extreme QTL mapping were also identified and validated in a study on variation in aggression (Shorter et al. 2015). These studies, along with others (Sanjuán and Elena 2006; Corbett-Detig et al. 2013; Chow et al. 2016), show convincingly that epistasis is a prevalent determinant of the genetic architecture of complex traits.

Lack of replication of associated polymorphisms between GWA studies in the DGRP and extreme QTL mapping analyses in the AIP is most likely due to the different frequencies of causal molecular variants in the two populations; note that this will almost always occur since the AIP populations are derived from a small subset of DGRP lines. A hallmark of epistatic interactions is that the effect of a focal locus on a quantitative trait depends on the allele frequencies at the interacting locus/loci (Fig. 5; Mackay 2015). This is not true for additive interactions among loci. Epistatic interactions are also sensitive to environmental conditions and are likely to affect the composition of the transcriptional niches of interacting partners, causing higher-order ripple effects. Thus, epistasis might be a confounding factor for the interpretation of GWA studies in human populations. For example, well-executed studies with sufficient statistical power to replicate associated variants in an independent population often fail to replicate the initial findings. However, lack of replication is actually expected in the presence of epistasis when populations have different allele frequencies at the genotyped loci (Greene et al. 2009; Moskvina et al. 2011; Mackay 2015). Disease-associated polymorphisms that are ignored because they do not replicate in different populations may nevertheless contribute important disease risk to members of the population in which they could be identified.

Fig. 5
figure 5

Epistasis depends on allele frequencies of interacting partners. a, b Absence of epistatic effects between alleles of locus B and alleles of locus A. c Differential epistatic effects of alleles B1 and B2 with alleles of locus A and d the dependence of the effect of locus A on the frequency of the B1 allele. e Antagonistic epistatic interactions between the B1 (enhancing epistasis) and B2 (suppressing epistasis) alleles with locus A and f the resulting dependence of the effect of locus A on the frequency of the B1 allele

Full size image

Although epistasis has been recognized in genetic studies on human populations (Webber 2017), it is often dismissed in analyses, because assessing pairwise epistatic interactions genome-wide is subject to a huge multiple testing penalty such that only interactions with very large effects could be detected. However, genetic variants uncovered in extensively studied human traits (a notable example is height; Wood et al. 2014) account for only a small fraction of the heritability (Manolio et al. 2009). Epistasis is one factor that can contribute to this “missing heritability” (Zuk et al. 2012).

Sexually dimorphic architecture of complex traits

Virtually every complex trait examined to date by mutational analyses, QTL analysis, GWA analysis, or extreme QTL mapping shows extensive genetic variation in sexual dimorphism, often with little overlap between trait-associated polymorphic markers of males and females (Fig. 6; Weber et al. 2012; Swarup et al. 2013; Harbison et al. 2013; Morozova et al. 2015; Garlapow et al. 2015; Zhou et al. 2016; Carbone et al. 2016). This is perhaps not surprising since extensive sexual dimorphism and genetic variation in sexual dimorphism is also apparent in genome-wide transcript abundances across the DGRP (Ayroles et al. 2009; Huang et al. 2015). In addition to inherent sex bias, environmental or genetic effects on gene expression may show sex specificity or sexually antagonistic modulation.

Fig. 6
figure 6

Adapted from Harbison et al. (2013)

Sexual dimorphism is prevalent among phenotypes measured in the DGRP. The example of sexual dimorphism shown here is phenotypic variation in sleep duration. Blue bars in the histograms in panels a and c denote male line means and purple bars denote female line means. The difference in line means between males and females (male–female) is shown in b and d. a Night sleep. b Male–female night sleep. c Day sleep. d Male–female day sleep.

Full size image

It is well known that disease susceptibility varies between men and women in human populations. It is, therefore, essential to include sex as a variable in analyses of GWA data (Golden and Voskuhl 2017). Failure to do so can reduce statistical power if a DNA variant affects a phenotype in one sex only or can lead to the erroneous conclusion that a particular variant poses a disease risk in both sexes. Considerations of sexual dimorphism are, therefore, essential for the development of genetic information-based precision medicine.

Genes and the environment

Phenotypic plasticity is the ability of a genotype to give rise to different phenotypes under different environmental conditions (Fig. 7). Studies in which an AIP derived from 40 DGRP lines was subjected to a wide range of different exposures showed that most of the Drosophila transcriptome is robust in the face of environmental changes, while the environmentally sensitive transcriptome includes genetically variable transcripts associated with detoxification, metabolism, proteolysis, heat shock proteins, and transcriptional regulation (Zhou et al. 2012).

Fig. 7
figure 7

Phenotypic plasticity and genotype by environment interaction. The diagram shows schematic reaction norms for two genotypes in two environments (Env). Phenotypic plasticity is evident when the reaction norms are parallel, i.e., the two genotypes respond similarly to a change in environment. Genotype by environment interaction is evident when the genotypes respond to different extents or in opposite directions to environmental change, or when one genotype responds to environmental change, whereas another is unaffected

Full size image

Genotype by environment interaction is due to genetic variation in phenotypic plasticity, i.e., different genotypes respond differently in response to environmental changes (Fig. 7). Some examples of genotype by environment interactions have been documented in studies of human populations. A classic study by Caspi et al. (2002) showed that a variable number tandem repeat polymorphism at the promoter of the MAOA gene was associated with violent behavior only in individuals who had experienced maltreatment as children. Furthermore, this effect was sex-specific and only observed in males. Similarly, polymorphisms in the promoter of the serotonin transporter gene have been associated with depression, dependent on stressful life experiences (Caspi et al. 2003; Rocha et al. 2015). These well-executed studies have not been universally corroborated (Munafò et al. 2009), potentially due to epistasis or differences in environmental sensitivity of allelic effects in different populations, as mentioned earlier. A well-established example of genotype by environment interactions is the observation that the effects of polymorphisms in immune response genes, including members of the interleukin family, associated with risk for asthma, are dependent on previous exposure to allergens (Sordillo et al. 2015; Bønnelykke and Ober 2016; Li et al. 2016).

Genotype by environment interaction is notably evident upon exposure to toxins. The Drosophila model is eminently suitable as a model system for population-based large-scale toxicogenomic studies. Several studies have focused on the genetic factors that underlie individual variation in susceptibility to heavy metals, such as cadmium, lead, and methylmercury (Akins et al. 1992; Hirsch et al. 2012). Metallothioneins are small cysteine-rich proteins that can bind ingested toxic heavy metals, notably cadmium, and contribute to their long persistence following exposure (Isani and Carpene 2014). In Drosophila, levels of metallothionein contribute to differences in sensitivity to cadmium between different strains (Gill et al. 1989), but cadmium sequestration by metallothionein is not the only factor that is responsible for cadmium resistance (Nguyen et al. 2014), indicating that other mechanisms contribute to cadmium toxicity.

Flies reared on low concentrations of lead acetate show changes in courtship, fecundity, and locomotor activity (Hirsch et al. 2003). A study on recombinant inbred lines, which used expression microarrays to identify cis-eQTL (i.e., where a local gene controls its own transcription) and trans-eQTL (i.e., where the activity of distant genes controls transcription) that were differentially expressed among control and lead-exposed flies, identified a co-regulated ensemble of 33 lead-induced genes, many of which are associated with neurodevelopment (Ruden et al. 2009). A study on lead sensitivity across the DGRP showed similar results. DGRP flies showed different effects on development time and adult locomotion when reared on medium supplemented with lead acetate, and gene ontology and network analyses showed enrichment of genes associated with early development and function of the nervous system (Zhou et al. 2016). Observations in the Drosophila model can guide studies of lead toxicity in children, which leads to neurological and cognitive deficits (Canfield et al. 2003; Jakubowski 2011; Liu and Lewis 2014; McDermott et al. 2014), by suggesting candidate pathways or cellular mechanisms.

A subsequent extreme QTL mapping study on a DGRP-derived AIP identified allelic variants associated with sensitivity to lead and cadmium exposure (Zhou et al. 2017). This study revealed genetic networks on which human counterparts of Drosophila genes could be superimposed. Human genes previously implicated in heavy metal toxicity could be placed in biological context along with identification of novel targets for heavy metal toxicity. This study showed that evolutionary conservation of fundamental biological processes enables Drosophila to serve as a translational model for toxicogenomics studies.

Methylmercury, a toxic heavy metal that can accumulate in seafood, has detrimental effects on the developing nervous system by interfering with the Notch receptor pathway (Bland and Rand 2006; Alattia et al. 2011; Engel et al. 2012). A GWA study on susceptibility to methylmercury exposure of DGRP lines identified candidate genes involved in muscle and neuromuscular development, and pupae exposed to methylmercury showed disrupted development of indirect flight muscle (Montgomery et al. 2014). In addition to effects on neural development, these observations implicate effects on muscle development as a consequence of methylmercury exposure. Glutamate cysteine ligase provides protection against methylmercury and overexpression of this enzyme in muscle-rescued eclosion of flies reared on methylmercury-supplemented medium. Mutations in kirre, a myogenic gene identified as a candidate gene associated with variation in methylmercury sensitivity, modulated eclosion rates upon exposure to methylmercury (Montgomery et al. 2014).

Drosophila has also proven to be a valuable model to study genetic variation in susceptibility to environmental oxidative stress agents, notably the highly toxic herbicide paraquat, which has been implicated as a causative agent for Parkinson’s disease, characterized by degeneration of dopaminergic neurons. Paraquat neurotoxicity in Drosophila also affects dopaminergic neurons (Martin et al. 2014) and expression of the dopamine receptor (Cassar et al. 2015). However, GWA studies on DGRP lines show that the genetic architecture that determines variation in susceptibility to paraquat and other oxidative stressors is highly polygenic. There is extensive variation in sensitivity to acute exposure to two oxidative stress-inducing agents, paraquat and menadione sodium bisulfite with little overlap of SNPs associated with variation in sensitivity to these compounds (Weber et al. 2012). GWA analyses of DGRP lines exposed to chronic oxidative stress induced by menadione sodium bisulfite revealed a network of candidate genes on which human orthologues could be superimposed. These candidate genes were associated with inositol triphosphate signaling and synaptic transmission, intermediary metabolism, signaling by NGF, EGFR, and Rho GTPases, and DNA replication (Jordan et al. 2012). These studies further underscore the translational potential of the Drosophila model as they highlight molecular processes that are conserved across phyla.

Orthologous networks

As mentioned earlier, we can assemble candidate genes identified through GWA analyses of the DGRP and/or extreme QTL mapping analyses of DGRP-derived AIPs into networks (Fig. 2). By superimposing human orthologues on these Drosophila genetic networks, we can build a translational blueprint to target candidate genes for subsequent focused studies in human populations (Fig. 8; Jordan et al. 2012; Zhou et al. 2016, 2017; Carbone et al. 2016). This approach is based on the principle of evolutionary conservation of fundamental biological processes and has several advantages. First, the Drosophila network provides a functional biological context for its human orthologue. Second, association of a gene with a trait in Drosophila, especially if it is a hub gene in the network, increases the chances for discovering an association of its orthologue with an analogous phenotype in a human population. Third, by identifying candidate genes with human orthologues through genome-wide screens in Drosophila, one can focus a subsequent association study on a single orthologue, which greatly reduces the multiple testing problem and, thus, increases statistical power.

Fig. 8
figure 8

A genetic network for susceptibility to lead exposure. The network was derived from candidate genes identified in GWA analyses for development time, viability, and activity of flies exposed to lead acetate. Yellow square boxes indicate candidate genes associated with any of these traits, while gray ovals represent computationally recruited intermediate genes. Blue font indicates genes with human orthologues. Note the extent by which human orthologues can be superimposed on their Drosophila counterparts to identify potential human candidate genes that may contribute to susceptibility to lead toxicity. From Zhou et al. (2016)

Full size image

Conclusion

GWA studies in Drosophila highlight the importance of sex-, environment-, and genetic background-dependent (epistatic) effects, modularity of gene expression, pleiotropy, and interactions among these parameters, which give rise to a dynamic integrative genetic architecture for complex traits. Understanding the genetic mechanisms that lead to the manifestation of genotype by environment interactions is critical for human health and developing genetics-informed precision medicine. Elucidating these mechanisms requires comprehensive empirical and computational approaches which integrate DNA sequence variation with variation at the levels of the transcriptome, proteome, and metabolome, while accounting for the complexity and dynamics of epistatic interactions, genome–environment interactions, and plasticity of transcriptional niches. D. melanogaster is the most favorable model organism for the pursuit of such a systems genetics endeavor. Principles derived from studies on this powerful genetic model system are universal and apply across phyla, from the DGRP to human populations.