Drosophila and genome-wide association studies: a review and resource for the functional dissection of human complex traits

Over the last two decades, human genome-wide association studies (GWAS) have begun to reveal the genetic risk factors for countless common disorders with complex genetic etiologies (Hardy and Singleton, 2009), including most of the major causes of morbidity and mortality in the developed world. Despite the impressive success rate for discovering disease susceptibility loci, few, if any, results from GWAS have yet to be successfully translated for delivery of new therapies. This is partly explained by challenges implicit in the experimental design; GWAS provide a list of implicated genomic loci from which causal genes must first be identified in order to confidently draw conclusions about biological mechanisms. Another important barrier stems from insufficient communication between the human geneticists who lead GWAS discovery efforts and the researchers best equipped with the tools that are pivotal for further investigations, including simple animal models. Here, we focus on the outstanding and largely untapped potential of the fruit fly, Drosophila melanogaster, for functional follow-up of human GWAS. We first provide a basic overview of the methodology behind GWAS, including addressing common misconceptions and highlighting challenges for identifying causal genes. Second, we introduce the Drosophila experimental model, covering some key contributions to biomedical science and the powerful genetic tools available for follow-up of GWAS. Third, we present results of cross-species bioinformatic analyses intended as a resource for both human and fly geneticists who are interested in working together to elucidate the genetic mechanisms that underlie complex human traits. Thus, we hope to promote more widespread experimental follow-up of human GWAS in fly models, and thereby accelerate insights leading to novel and much needed therapies.

The National Human Genome Research Institute Catalog of Published GWAS (Welter et al., 2014; http://www.ebi.ac.uk/gwas/) currently reports 33,004 variant associations with more than 1601 distinct human traits, based on aggregated results from 2224 published GWAS. Each of these human genetic findings is supported by robust statistical evidence. Compared with Mendelian disorders, which are caused by highly penetrant, single-gene mutations, complex genetic diseases are characterized by (1) common polymorphisms with more modest effect sizes; (2) a greater role for polygenicity (additive effects resulting from multiple risk alleles); (3) genetic heterogeneity, in which disease risk is influenced by at least partially non-overlapping loci; and (4) a more prominent contribution of non-genetic factors, including age and environmental exposures. For decades, such disorders evaded genetic dissection until the implementation of GWAS, which have now successfully revealed the presence of risk alleles for innumerable human traits, including coronary artery disease (Lu et al., 2012; Schunkert et al., 2011), multiple sclerosis (De Jager et al., 2009; Sawcer et al., 2011), Alzheimer's disease (Lambert et al., 2013; Naj et al., 2011), and schizophrenia (Ripke et al., 2013; Shi et al., 2011). In fact, as our understanding of human genetics has advanced, many of the distinctions between simple and complex genetic disorders have become less clear-cut. For example, GWAS have now identified common variant modifiers for rare Mendelian disorders (Lee et al., 2015), and reciprocally, next-generation sequencing approaches are beginning to reveal important contributions of rare risk alleles for common diseases (Goldstein et al., 2013). Additional examples have highlighted convergence between studies of rare and common alleles of the same gene, including SNCA in Parkinson's disease (Xu et al., 2015), PCSK9 in dyslipidemia (Teslovich et al., 2010), and TBX6 in congenital scoliosis (Wu et al., 2015).

In statistical genetics, the association study design simply compares the frequency of an allelic variant between a case and control sample (Balding, 2006) (Fig. 1A). Having controlled for potential confounders between the samples (e.g. ethnic composition, relatedness, or similar factors that might influence genetic makeup), a significant difference in variant frequencies between cases and controls signals the presence of a potential risk or protective allele. An improved understanding of human genomic variation coupled with advances in genotyping technology and statistical methods have enabled association testing on an unbiased, genome-wide scale. In the resulting GWAS, association tests are conducted iteratively for variants sampled across the entire genome. Owing to the block-like structure of correlated genetic variation within genomes (Frazer et al., 2007), current imputation methods allow estimation of millions of genotypes from a much smaller number of directly typed variants. Importantly, given the limitations imposed by commonly employed genotyping arrays and sample sizes, most GWAS conducted to date have been powered for detection of relatively common genomic variants (>1% minor allele frequency). More recently, however, GWAS are also being deployed successfully for analysis of exome-wide genotyping or next-generation sequencing data, highlighting less common or rare variant alleles in population-based case and control cohorts (Francioli et al., 2014; Gudbjartsson et al., 2015; Walter et al., 2015). Besides their application for human genetic investigation, GWAS have further proven to be a powerful method for discovery of genomic variants responsible for complex traits in other species, including Drosophila. Indeed, GWAS in flies have successfully identified susceptibility loci for sleep (Harbison et al., 2013), aggression (Shorter et al., 2015), brain size (Zwarts et al., 2015), courtship patterns (Gaertner et al., 2015), sexual characteristics (Takahara and Takahashi, 2015), longevity (Ivanov et al., 2015) and pigmentation (Dembeck et al., 2015), among other phenotypes.

As noted earlier, thousands of susceptibility loci have now been reported for a large diversity of human traits, including many diseases without effective treatments. A central justification for the substantial investment in GWAS has been the potential for advancement of our understanding of disease mechanisms, including insights that could promote development of new therapies. However, the successful discovery of a robust susceptibility signal via GWAS rarely amounts to definitive identification of the responsible gene(s), which is essential for moving from a genetic finding to therapeutic development. Instead, GWAS typically identify a marker single nucleotide polymorphism (SNP) that is correlated with an unknown causal genetic variant (Fig. 1B) (Cantor et al., 2010). Therefore, associated SNPs usually identify ‘haplotype blocks’ encompassing up to several hundred kilobases (Wall and Pritchard, 2003) and further fine mapping of such signals to define the responsible genes and variants has proven challenging in most cases (Edwards et al., 2013; Ioannidis et al., 2009). Thus, association peaks from GWAS typically identify a handful of gene candidates, and additional studies are required to determine whether one or more of these are likely to be causal.

Further complicating the picture, it now seems that most GWAS signals are probably the result of regulatory variation – alleles that fall within enhancer elements and therefore impact gene expression – rather than amino acid changes (GTExConsortium, 2015; Nicolae et al., 2010). Because enhancers often act over very long genomic distances, a large number of candidate genes might need to be considered for each associated polymorphism. Nevertheless, ongoing large consortium projects, such as ENCODE (EncodeProjectConsortium, 2004) and GTEx (GTExConsortium, 2015), are enhancing our understanding of non-coding genomic regulatory sequences, allowing improved estimation of the local and more distant changes in gene expression triggered by common variants. Other bioinformatic approaches have been developed to integrate knowledge of genetic pathways with GWAS to test the association between a pathway and a disorder (Cantor et al., 2010; Hindorff et al., 2009). Such efforts promise to simplify the prioritization of candidate genes that could be responsible for associated variants. However, as discussed further below, even once a convincing candidate gene is identified, substantial experimental work is needed to confirm its link to disease susceptibility, including elucidation of the underlying molecular mechanisms.

A frequent criticism of GWAS is that the effect sizes identified for most variants are quite modest. For example, among the ∼22 susceptibility loci identified by the largest Alzheimer's disease GWAS (Lambert et al., 2013), odds ratio estimates for the implicated risk alleles range from 1.1-1.4, whereas the epsilon 4 allele of the apolipoprotein E gene – discovered by linkage analyses completed before the GWAS era – is associated with a threefold increased risk of disease (Corder et al., 1993; Pericak-Vance et al., 1991). However, there are several caveats to consider for interpretation of effect sizes of GWAS. First, as the identified variants are nearly always proxies, they likely yield underestimates for the effect size of the true, but unknown, causal variant. Second, although susceptibility variant effect sizes from GWAS can be underwhelming, their commonality translates to a large contribution to disease risk on a population level. Additionally, individuals frequently harbor multiple such alleles, and corresponding aggregate genetic risk models reveal much stronger combined effects. Third, findings from GWAS hold enormous promise for novel mechanistic insights and potential breakthroughs in treatment, despite the few meaningful advances in clinical risk prediction to date (Manolio, 2013). Notably, evolutionary selective pressures might constrain the frequency of variants with strongly damaging or other functional consequences in human populations, thereby limiting the observed effect sizes of common genomic variants on human traits. By contrast, pharmacological manipulation is not subject to this potential ceiling effect, allowing for more potent therapeutic outcomes. For example, in GWAS of dyslipidemia, common polymorphisms in the HMG-CoA reductase gene have quite modest effects on low-density lipoprotein (LDL) cholesterol levels (Teslovich et al., 2010), whereas treatment with statins, designed to inhibit the encoded enzyme, are powerful LDL-lowering agents taken by millions worldwide for heart disease prevention.

In sum, GWAS have successfully identified thousands of common genomic variants responsible for countless human disease traits. Although such results are supported by robust statistical evidence and represent an enormous opportunity for novel biological insights relevant to disease pathophysiology, the majority of susceptibility loci await functional follow-up, and such work will be essential to leverage GWAS findings for therapeutic advances. Having introduced GWAS, we turn next to exploring the potential for Drosophila to accelerate the urgently needed follow-up studies.

The vast array of resources and tools that recommend Drosophila for functional genomic investigations, including follow-up of human GWAS, has been extensively reviewed (Matthews et al., 2005; Mohr et al., 2014; Ugur et al., 2016; Venken et al., 2011). One distinct advantage is the immediate availability of several large collections of reagents for gene manipulation (Fig. 2). In addition to alleles generated by chemical mutagenesis, transposable element insertions are available for the majority of fly genes (Fig. 2A), including homologs of candidate susceptibility genes from GWAS. Specifically, the Drosophila Gene Disruption Project has generated transposon insertion alleles for over two thirds of the organism's protein-coding genes (Bellen et al., 2011). These strains facilitate further genomic manipulations, including the generation of deletion alleles via imprecise excision of transposable elements (Fig. 2A). Collections of deficiency strains tiling nearly the entire Drosophila genome provide another valuable resource for reverse genetic studies (Cook et al., 2012; Ryder et al., 2007). Binary expression systems (Fig. 2B) in which a transcriptional activator binds to specific cis-enhancer elements, leading to activation of the adjacent gene, are enormously flexible, popular and powerful research tools. GAL4/UAS, which was co-opted from yeast, is the most widely used binary expression system, and thousands of GAL4 driver lines available from individual labs and public stock collections allow expression of desired target genes, typically cDNA transgenes under control of upstream activating sequence (UAS) sites, in precise spatial and temporal patterns (Brand and Perrimon, 1993). Extensive collections of transgenic RNA-interference (RNAi) lines are also available (Dietzl et al., 2007; Perkins et al., 2015). Under control of the GAL4-UAS system, these RNAi strains permit tissue-specific and/or conditional knockdown of up to 88% of all protein-coding genes (Fig. 2C), facilitating efficient analysis of loss-of-function phenotypes. Information on these and many other useful genetic reagents is available through FlyBase, a central, online annotated resource for Drosophila genetics (http://flybase.org; dos Santos et al., 2015; Millburn et al., 2016).

Leveraging substantial genomic conservation and powerful genetic tools outlined above, studies in Drosophila have touched nearly all branches of human disease biology (Wangler et al., 2015). This includes insights into mechanisms of congenital renal disease (Weavers et al., 2009), cardiotoxicity from a high-fat diet (Diop et al., 2015), sterol absorption in the gut (Voght et al., 2007) and neuromuscular dysfunction in mucolipidosis (Venkatachalam et al., 2008). Drosophila models have enabled powerful mechanistic insights into numerous neurodegenerative disorders (Jaiswal et al., 2012; Lessing and Bonini, 2009; Shulman et al., 2003) including Alzheimer's disease, Parkinson's disease, amyotrophic lateral sclerosis, Huntington's disease and spinocerebellar ataxias (Lessing and Bonini, 2009; Romero et al., 2008; Tsuda et al., 2008, 2005; Vos et al., 2012). Unbiased genetic screens leveraging RNAi reagents in the fly have been deployed to identify conserved regulators of cardiac function (Neely et al., 2010b), pain perception (Neely et al., 2010a) and adipocyte differentiation (Pospisilik et al., 2010). Other screens have highlighted Mendelian disease genes (Yamamoto et al., 2014). For such disorders, Drosophila readily facilitates confirmation of variant pathogenicity and elucidation of disease mechanisms (Bellen and Yamamoto, 2015; Shulman, 2015). In the broad scope of investigations using fly models for functional genomics, follow-up of results from human GWAS is still comparatively new. Nevertheless, there is enormous potential for such work, and we highlight below several noteworthy and pioneering examples.

Regardless of the experimental model selected, GWAS present unique challenges for functional follow-up investigations, and one immediate obstacle is how to prioritize specific candidate genes for further study based on associated variants. Because most SNPs detected by GWAS are not likely causal variants for disease risk but rather informative markers, it is often not productive to study their direct functional consequences. Furthermore, the non-coding sequences that usually harbor such changes show less conservation than exonic regions, especially in evolutionarily distant species. Instead, implicated SNPs must be mapped to the most promising gene candidates and usually, there are multiple prospects (Fig. 1B, Fig. 3A). One strategy is to leverage medium- to high-throughput screening assays in flies to help refine and prioritize from among several, equally good, candidate genes in order to identify one (or more) worthy of more detailed study. For example, Pendse et al. (2013) studied 38 human genomic regions based on SNPs linked to type 2 diabetes mellitus or related metabolic traits. A 100 kb genomic window centered around each associated SNP identified 130 human gene candidates, of which 71 were sufficiently conserved for follow-up in flies. Orthologous genes were serially targeted using available RNAi transgenic lines and examined for genetic interactions in a fly model relevant to diabetes, based on sucrose-induced toxicity. As illustrated in Fig. 2C, short hairpin RNA sequences homologous to each target gene were activated using a ubiquitously expressed tubulin-GAL4 driver line. Using this strategy, 34 human genes were highlighted based on enhancement or suppression of sucrose-induced lethality, following knockdown of their respective gene orthologs. The majority of genomic regions studied had multiple human candidate genes, and in six cases, a single gene was implicated based on the screening assay, allowing refinement of the initial list. Interestingly, in nine other cases, more than one gene in a region showed independent interactions with sucrose toxicity, potentially compatible with the contribution of multiple genes to each human susceptibility signal. Based on more recent GWAS (Bonnefond and Froguel, 2015), more than 90 susceptibility loci have now been implicated in type 2 diabetes, increasing the need for further follow-up studies. Indeed, the vast majority of susceptibility signals identified by human GWAS to date await further fine-mapping efforts to prioritize and confirm the responsible genes.

Where one or more well-conserved gene candidates are strongly implicated, Drosophila is an ideal model system for further functional elucidation. The prevailing strategy relies on the key assumption that beyond evolutionary conservation at the sequence level, homologous genes will have conserved functional requirements in humans and in Drosophila, leading to similar traits or phenotypes when subject to genetic manipulation. In a GWAS for alcohol consumption that included more than 40,000 individuals, the top-associated SNP fell within the intron of the autism-linked AUTS2 gene, which encodes a neuronal nuclear protein of uncertain function (Schumann et al., 2011). The implicated variant was further related to AUTS2 expression levels in postmortem human brain samples, consistent with AUTS2 being a candidate causal gene for the association. In order to complement the largely correlative data with functional evidence, the authors next turned to Drosophila, which has a single gene ortholog, tay. Ethanol tolerance was enhanced by inactivation of tay using transposable element insertional alleles or pan-neuronal knockdown of the gene by RNAi (Schumann et al., 2011). Specifically, tay loss-of-function flies were observed to maintain consciousness longer than control animals when exposed to ethanol vapor. In another study focused on identifying risk factors for idiopathic azoospermia in a cohort of more than 9000 men, a similar strategy was employed for follow-up of a strongly suggestive association signal proximal to the gene candidate, CDC42BPA. Indeed, knockdown of the fly homolog, gek, via RNAi transgene expression in the supporting somatic cells of the fly testis led to impaired sperm maturation and resulting infertility (Hu et al., 2014). Lastly, in a more targeted Drosophila follow-up study of a top variant associated with restless legs syndrome (Stefansson et al., 2007), mutant alleles of a conserved homolog of the leading candidate gene, BTBD9, were generated via transposon-mediated excision (Fig. 2A), causing motor restlessness and sleep fragmentation that was remarkably reminiscent of the human disorder (Freeman et al., 2012).

Given the success of using Drosophila to model neurodegenerative disorders, it is not surprising that such systems have been applied to follow up findings of GWAS in this field. In one early example of this strategy, we used fly transgenic models to test the hypothesis that candidate Alzheimer's disease susceptibility genes modulate neurodegenerative phenotypes induced by expression of the human MAPT protein, responsible for the characteristic neurofibrillary tangle pathology (Shulman et al., 2011). Specifically, MAPT-induced retinal degeneration causes a visible eye phenotype in flies that is modified by knockdown or activation of susceptibility gene homologs. In subsequent work (Chapuis et al., 2013; Moreau et al., 2014; Shulman et al., 2014), we and others discovered that fly homologs of several genes at loci implicated by GWAS, including BIN1, PICALM, CD2AP, CELF1 and FERMT2, can interact with MAPT toxicity in vivo, providing mechanistic insight into their potential links with Alzheimer's disease risk in humans. Similarly, GWAS in Parkinson's disease have recently (Nalls et al., 2014) expanded to ∼28 the number of susceptibility loci identified by common genomic variants, and studies in Drosophila have shown promise for mechanistic follow-up. Interestingly, results from two recent studies (Ivatt et al., 2014; MacLeod et al., 2013) highlight connections between candidate genes identified by GWAS and causes of Mendelian Parkinson's disease. Consistent with findings in neuronal cell cultures, Macleod et al. found that overexpression of the fly homolog of RAB7L1 (also known as RAB29), a candidate gene in the PARK16 susceptibility locus, was capable of rescuing dopaminergic neuronal loss and reduced survival induced by LRRK2^G2019S, which is associated with autosomal dominant familial Parkinson's disease. LRRK2 toxicity was also suppressed by overexpression of the conserved Drosophila homolog of VPS35, rare variants that also cause dominantly inherited Parkinson's disease. Prior studies in numerous models (including flies) demonstrated that VPS35 is required for retrograde transport of proteins within the endosomal-lysosomal pathway (Wang et al., 2014). In the second study, Ivatt et al. (2014) performed a genome-wide RNAi screen in Drosophila cells to identify genes required for translocation of the Parkinson's disease-associated protein parkin upon mitochondrial damage. Mutation of the parkin (PARK2) gene is a common cause of early-onset, autosomal recessive parkinsonism, and studies in Drosophila have contributed substantially to our understanding of its putative role in mitochondrial quality control (Haelterman et al., 2014). Interestingly, the cell-based parkin interaction screen identified multiple mediators of lipogenesis, including SREBF1, a Parkinson's disease susceptibility candidate gene identified via GWAS. Thus, these studies exemplify how investigation in Drosophila can not only link findings from GWAS to informative, disease-relevant biology, but can additionally reveal connections between simple Mendelian and more complex genetic forms of disease.

Despite the opportunities for advancing our mechanistic understanding of many common human diseases with complex genetic etiologies, Drosophila certainly has its limitations for follow-up of GWAS. As discussed above, the majority of susceptibility signals identified by GWAS likely point to regulatory variants. However, compared with protein-coding sequences, cross-species evolutionary conservation of genomic regulatory sequence is less well defined, especially between humans and flies. Thus, it is rarely feasible to directly examine the functional consequences of presumed regulatory variants. Instead, Drosophila is more appropriate for gene-centric strategies that investigate consequences of directed experimental manipulation of genes on relevant phenotypes. The potential application of this approach depends on gene conservation. Even where conservation of genes is strong, a related question is whether encoded proteins will subserve conserved functions in such evolutionary distant species. Thus, as considered systematically below, in order to understand the potential applicability of Drosophila for follow-up of human susceptibility loci, we need to understand not only the conservation of candidate susceptibility genes, but whether available evidence supports conserved functional requirements.

In order to assess the potential of Drosophila for follow-up of GWAS – and to directly facilitate future studies – we have undertaken systematic cross-species analyses based on the comprehensive results reported in the National Human Genome Research Institute–European Molecular Biology Laboratory-European Bioinformatics Institute (NHGRI-EBI) GWAS catalog (http://www.ebi.ac.uk/gwas). Selected findings and examples are highlighted below, and the full results are available in supplemental Tables S1-S4. As a starting point, we applied consistent criteria to define candidate human susceptibility genes based on reported SNP associations for each trait. The overall approach is illustrated in Fig. 3A, taking as an example a chromosome 6 SNP, rs9349407, discovered in a large GWAS of Alzheimer’s disease risk (Naj et al., 2011). This SNP falls within an intron of the CD2AP gene, which encodes an actin-binding and SH3 domain adaptor protein, but additional candidate genes can be defined based on genomic intervals centered around the index variant, the size of which determines the number of implicated genes. In our analyses, we define four nested categories that progressively consider an increasing number of candidate genes at each implicated locus: (1) ‘intragenic’, referring to genes in which the associated SNP falls within an intron or exon (n=0-1 genes/SNP); (2) ‘nearest neighbors’, including the genes immediately proximal and distal to the SNP (n=2-3 genes/SNP); (3) ‘within 250 kb’, including all genes within a genomic window 125 kb proximal and 125 kb distal to the SNP (n≈0-20 genes/SNP); and (4) ‘within 500 kb’, in which the genomic window is extended a further 250 kb proximal and distal to the SNP (n≈0-37 genes/SNP). Therefore in our example (Fig. 3A), the rs9349407 SNP identifies CD2AP (intragenic), TNFRSF21 and ADGRF2 (nearest neighbors), and lastly ADGRF4 (within 500 kb). These four levels of criteria were applied to all 15,825 reported SNP associations within the NHGRI-EBI GWAS catalog (September 2015 data freeze), resulting in between 4009 (intragenic) and 16,544 (within 500 kb) total human candidate susceptibility genes, which we consider for the analyses described below. In Table S2, we also make available the comprehensive list of candidates.

Having defined a candidate gene list, we next asked which loci are conserved in Drosophila, a key prerequisite for following up a candidate susceptibility gene from GWAS in flies. We took advantage of the published Drosophila Integrative Ortholog Prediction Tool (DIOPT) (Hu et al., 2011), which integrates 10 bioinformatic algorithms, to evaluate putative human-fly ortholog pairs. We required that at least two distinct algorithms agree (DIOPT score≥2) for determination of whether a human candidate susceptibility gene is conserved. Although these rather liberal criteria are potentially liable to false-positive calls of homology, they are adequate for our goals to (1) assess potential conservation among large groups of genes and (2) inform selection of genes for follow-up experimental validation, which ultimately is essential to confirm any bioinformatic predictions. Applying these criteria genome-wide, 11,122 out of 20,950 (53%) protein-coding genes in the human genome have a fly ortholog. Based on prior work documenting increased conservation of human genes linked to Mendelian disorders (Fortini et al., 2000; Hu et al., 2011) we hypothesized that on average, susceptibility genes for complex human diseases would also show increased conservation. Indeed, we discovered a modest enrichment (1.2-fold) of cross-species conservation for candidate genes mapped to SNPs discovered in human GWAS. Interestingly, the degree of conservation is strongest (64%, 2573 of 4009 genes) for intragenic genes, whereas enrichment is attenuated somewhat when the criteria are liberalized (58% for genes within 500 kb, 9660 of 16,544 genes) (Fig. 3B). If genes with roles in disease susceptibility (such as those implicated by GWAS) are indeed more likely to be evolutionarily conserved, than this observation might suggest that the genes harboring intragenic associated variants are more likely to be truly causal than candidates mapped at greater distances from SNPs; in other words, cross-species conservation could help to guide fine mapping of causal genes. For all subsequent analyses, we restricted our consideration to candidate susceptibility genes identified by intragenic SNPs.

The human GWAS catalog reports findings from studies of a diverse array of human traits and disease phenotypes, ranging from risk of autism to economic and political preference. We wondered whether all such traits are equally translatable for studies in fruit flies. In order to facilitate comparisons, we categorized each of the 1252 traits based on two separate criteria: (1) target tissue (Table 1) or (2) disease mechanism (Table 2) (full results in Table S1). For target tissue, we considered the cell type and/or organ system that is predominantly impacted by the trait or disease (e.g. heart failure primarily affects the cardiovascular system). However, diseases that affect the same organ systems often have widely divergent genetic mechanisms. Although ischemic stroke, multiple sclerosis, and Parkinson's disease similarly affect the central nervous system, distinct mechanisms are implicated (e.g. vascular, immunological and/or inflammatory, and neurodegenerative etiologies, respectively), implying that the underlying genes – and resulting extent of evolutionary conservation – might differ. The principle of aggregating traits based on common mechanism has been leveraged for the discovery of shared genetic risk factors, formally known as pleiotropy (Solovieff et al., 2013). For example, a GWAS integrating data on more than 30,000 subjects with autism, attention deficit-hyperactivity disorder, bipolar disorder, major depressive disorder and schizophrenia identified 38 candidate genes (Smoller et al., 2013), of which 27 (73.0%) are conserved in the Drosophila genome.

We therefore examined whether particular human phenotypes or categories are more or less amenable to study in Drosophila, again using conservation of implicated genes from GWAS as a benchmark (Tables 1, 2; Table S2). One potential caveat for interpretation of these analyses is that not all human traits (or trait categories) have been interrogated with the same intensity. Variation in the number of studies and sample sizes employed impact the statistical power for discovery of susceptibility loci. Nevertheless, based on either target tissues or mechanisms, all human trait categories showed evidence of increased conservation compared with the genome-wide average (53%, above). However, as shown in Tables 1 and 2, we observed substantial variation (range: 57-73%) in the degree of conservation for implicated genes depending on the specific category. For example, when compared with the composite set of all genes nominated by GWAS, several target tissue categories showed increased conservation (e.g. pulmonary, eye, cardiovascular, nervous system) whereas others showed reduced conservation (e.g. blood, urogenital, skin, gastrointestinal). Interestingly, the alternative classification scheme based on disease mechanisms generally increased conservation within and across categories; gastrointestinal disorders remained the lowest ranked group. As expected, even within categories there can be significant variation in conservation based on specific traits. For example, among all nervous system traits with at least 20 gene candidates identified, Alzheimer's disease cognitive decline was the trait with the greatest conservation (80%, 32 out of 40 gene candidates).

In sum, candidate susceptibility genes nominated by human GWAS generally show increased conservation in Drosophila when compared with the average level of conservation observed for all human genes. In addition, the degree of conservation depends on the specific trait, and is influenced by both the target tissue and underlying putative disease mechanisms. Our comprehensive analytic results (Table S2) will allow human and fly geneticists to infer the extent to which Drosophila is amenable for follow-up of specific disease traits and/or categories, and provide an accessible catalogue of the fly genes that are homologous to human susceptibility gene candidates identified by published GWAS and thereby represent a high priority for functional studies.

In addition to gene or protein sequence conservation, the tissue-specificity of expression patterns (or lack thereof) is highly relevant when considering investigation of a candidate human disease gene in an experimental model organism. For example, when confronted with a novel brain-expressed candidate gene for Parkinson's disease, well-conserved in flies, it might be important to ask whether the gene is consistently expressed in the Drosophila nervous system. As well as supporting further functional studies in flies, a positive answer could also increase confidence that the selected candidate gene is truly causal. In order to address this systematically, we leveraged publicly available, high-throughput Drosophila gene expression data (Brown et al., 2014) to determine the levels and potential tissue-specificity for Drosophila homologs of candidate susceptibility genes from human GWAS. Similar analyses can be readily performed using the publicly available Drosophila Gene Expression Tool (DGET, http://fgr.hms.harvard.edu/dget). We initially focused on nervous system traits, and found that 77% of fly homologs for human neurological disorder susceptibility genes were expressed in the adult fly head. Furthermore, expression of these genes seemed somewhat more likely to be detected in the fly head than in the fly digestive system (69%), or whole animal (72%) (Fig. 4A), consistent with the hypothesis that homologs of genes associated with human nervous system disorders, show a relative, albeit modest, tissue specificity in flies. In addition, the fly homologs of candidate genes expressed in the human nervous system are somewhat more likely to be expressed in the adult fly head than are homologs of candidate genes nominated by other human target tissue categories (Fig. 4B; full results in Table S3). One potential caveat comes from the finding that the fly homologs of human candidate susceptibility genes from GWAS tend to be expressed at higher levels not only in the fly head, but also across a broad range of other Drosophila tissues for which data are available (Fig. 4A; Table S3). Indeed, strongly conserved Drosophila genes, including the homologs of most candidate susceptibility genes from GWAS, seem to be expressed more widely and robustly, consistent with conclusions from our recent work (Y.H., unpublished observation). Given that most signals in GWAS are believed to represent the impact of genomic regulatory variants that cause modest transcriptional changes (GTExConsortium, 2015; Nicolae et al., 2010), it is intriguing that homologs of these genes are broadly expressed in Drosophila. Thus, future studies in flies could reveal the existence of conserved gene regulation and/or dose-sensitive requirements, consistent with the implicated mechanisms of susceptibility variants in human disease. Based on our results, however, we urge caution when drawing conclusions about cross-species functional conservation based solely on tissue expression patterns. Before deciding on the potential feasibility of pursuing follow-up studies in flies, it will be essential to integrate observations of expression with the results from directed experimental manipulations that yield more reliable evidence of conserved functional requirements, particularly within specific tissue contexts.

Documenting candidate susceptibility gene conservation and expression in relevant tissues might be important foundations for follow-up of GWAS; however, the ultimate goal is to deploy powerful genetic models, including flies, for functional dissection. FlyBase provides an excellent starting point by not only documenting existing strains for experimental manipulation, but also making accessible the data on previously reported phenotypes (dos Santos et al., 2015). Quick lookups can often provide clues to the evolutionary conservation of functions, including those that are of potential relevance to human disease traits (Millburn et al., 2016). To illustrate this, we examined promising results from two selected GWAS that interrogated risk factors for social communication impairment in children, a trait related to autism spectrum disorders (St Pourcain et al., 2014, 2013). Together, these studies identified seven independent signals, of which five are conserved in Drosophila. Using FlyBase as a guide, along with targeted literature mining, we discovered that the reported loss-of-function phenotypes suggest conserved gene functions that might be relevant to their associations with the complex human phenotype of social communication. For example, bru3, the fly ortholog of human CELF4, encoding an RNA-binding protein and regulator of transcript splicing and translation, was identified in a Drosophila genetic screen for gender-specific social responsiveness (Ellis and Carney, 2011). Another gene, Tmhs, homologous to LHFPL3, encoding a tetraspan membrane protein, is required for fly auditory perception (Cosetti et al., 2008; Coop et al., 2008). Lastly, CG4328, a homolog of LMX1B, encoding a transcription factor, was linked to sensory neuron dendritic arborization (Parrish et al., 2006). Though speculative in the absence of further experimental validation, these published loss-of-function phenotypes in Drosophila suggest that the homologous human susceptibility genes might subserve similar functions in the genesis of social communication disorders. Thus, targeted data mining allows results of human GWAS to be rapidly linked to relevant model organism phenotypes, informing mechanistic hypotheses for further testing.

These examples provide potential support for the functional conservation of genes underlying complex human traits; nonetheless, it is difficult to generalize based on anecdotal evidence alone. In order to test the hypothesis more systematically, we again considered all human candidate susceptibility genes for neurological traits, asking whether conserved Drosophila homologs are similarly associated with nervous system phenotypes in flies. For this analysis, we took advantage of the FlyBase ‘controlled vocabulary (CV)’ terms used to standardize phenotype reporting (Millburn et al., 2016). Of the homologs for human neurological susceptibility genes, 33% (306 out of 914) are established to cause neuronal or nervous system phenotypes in Drosophila, based on 41 FlyBase CV terms, compared with 25% for all fly genes. Although this represents only a modest enrichment over the 29% of all GWAS candidate homologs causing fly neuronal phenotypes, it is potentially consistent with our finding (Fig. 4A), that such genes are frequently expressed in the Drosophila nervous system.

Although the neuronal functions of Drosophila genes are more widely annotated than in humans or mammalian models, one important caveat is that many genes still remain incompletely studied. Even for genes with well-characterized loss-of-function phenotypes, semantic barriers can complicate the precise matching of fly to human traits, because Drosophila and human geneticists frequently use widely differing terminologies to describe phenotypes (Pospisilik et al., 2010). Given the growing recognition of this challenge, phenotype curation in both humans and model organisms and associated bioinformatic tools for data integration are likely to improve in the near future (Deans et al., 2015; Mungall et al., 2010, 2015). Community resources, such as the Monarch Initiative (https://monarchinitiative.org), promise to facilitate mapping of human to model organism traits, thereby enhancing the functional follow-up of implicated susceptibility genes.

In sum, as our systematic understanding of the functional requirements of all conserved Drosophila genes improves, so will the analytical power to pinpoint human traits amenable to cross-species functional dissection. However, this conclusion depends on the likely flawed assumption that conserved genes and associated genetic pathways will cause similar loss-of-function phenotypes in distantly related organisms, such as humans and Drosophila. In fact, many strongly conserved molecular systems, such as signal transduction pathways or gene coexpression networks, are ‘repurposed’ over evolutionary timescales for heterogeneous cellular, developmental and organismal functions. For example, mutations in the human sonic hedgehog gene (SHH) result in nervous system pattering defects and subsequent developmental malformations (e.g. holoprosencephaly). By contrast, mutations in the Drosophila ortholog, hedgehog (hh), disrupt embryonic segmentation and lead to altered appearance of larval denticle belts. Nevertheless, studies of this larval phenotype have facilitated successful identification and mechanistic dissection of hh and numerous genes encoding highly conserved downstream signaling components. This and numerous similar examples illustrate that the mere superficial equivalence of distal phenotypic outcomes is likely a poor predictor for the potential value of flies in the mechanistic follow-up of human complex traits; rather, it is the fundamental conservation and coherence of the more proximal genetic regulatory networks that are consequential. To put this more simply, many potentially valuable Drosophila ‘disease models’ might have little or no resemblance to their cognate human traits. Embracing this insight, Marcotte and colleagues (McGary et al., 2010; Woods et al., 2013) deployed an unbiased bioinformatic approach to define homologous phenotypes, or ‘phenologs’ (http://www.phenologs.org), between distantly related species, including humans and several experimental animal models (but unfortunately not Drosophila). Phenolog assignment was based on overlap between groups of conserved gene sets that cause similar phenotypes within each species. For example, this successful strategy led to testable predictions of new genes involved in human breast cancer and neural crest defects based on non-obvious homologous phenotypes in C. elegans (hermaphroditism) and Arabidopsis (gravitropism), respectively (McGary et al., 2010).

Full implementation of the phenolog strategy in Drosophila is outside the scope of this Review; however, we did examine the proportion of fly homologs of human susceptibility gene candidates from GWAS that are essential (Table 3; full results in Table S4), i.e. result in embryonic lethality when genetically disrupted. As a group, human GWAS identify orthologous fly genes that are strongly enriched for lethal phenotypes (43% versus 25% for all Drosophila genes). Interestingly, among disease mechanism categories, candidate susceptibility genes for human developmental disorders were even more likely to be required for embryonic viability in flies (50% lethal phenotypes). More broadly, 66% of all homologs for human susceptibility gene candidates from GWAS currently have phenotypic annotations in FlyBase, providing an immediate entry point for functional study in flies, such as rapid tests of phenotypic rescue by the human homologs and/or evaluation of genetic interactions with other candidate susceptibility genes.

Recently, the pre-eminence of GWAS as a tool for susceptibility locus discovery in human complex genetic disorders is being supplanted by next-generation sequencing approaches and the complementary search for rare variant risk alleles. This comes as the GWAS approach reaches maturity, with many common diseases having now been interrogated by meta-analyses that involve the largest feasible sample sizes. Similar to the arrival of GWAS about 10 years ago, this shift is further fueled by the tumbling cost of next-generation sequencing technology coupled with advances in statistical and analytical methods. Whereas GWAS usually implicates tag SNPs and genomic loci encompassing multiple potential causal genes, sequencing instead promises to pinpoint specific gene variants with putative functional consequences. Nevertheless, the large number of implicated variants emerging from most sequencing studies leads to a related set of challenges as well as opportunities for experimental follow-up (Shulman, 2015). While embracing the opportunities for sequencing-based discovery, it is essential that we not prematurely abandon GWAS, given the potential of these data to illuminate the yet largely unknown biological mechanisms underlying common diseases with complex genetic etiologies. Indeed, as we have emphasized throughout this Review, although the human genetic work might be winding down, a significant challenge remains to confirm the responsible genes and understand the relevant mechanisms, a task well-suited for the model organism research community.

Although we have focused on the powerful tools and approaches that Drosophila models can bring to bear on GWAS follow-up, other experimental systems also have important roles to play. Integrated studies in both simple (e.g. yeast, nematode, fly) and more complex (e.g. zebrafish, mouse) animal models can create powerful synergy for the follow-up of candidates arising from human genetic studies, highlighting the evolutionary conservation of disease mechanisms. Notably, recent advances in human induced pluripotent stem cell (iPSC) methods offer powerful, complementary experimental systems for functional validation of GWAS findings (Zhu et al., 2011). Further, the application of high-throughput technologies for profiling the epigenome, transcriptome, proteome, and/or metabolome in human tissues, coupled with robust systems biology approaches, show increasing promise for mechanistic dissection. In fact, some have argued that such advances could render model organisms obsolete for the study of complex human traits (Visscher, 2016). In our opinion, such predictions represent a grave misreading of the current research landscape. Despite offering the distinct advantage of a human genomic context for variant validation and more faithful recapitulation of the species-specific cellular milieu, iPSC-based approaches nevertheless fall short in several important areas. The mechanisms underlying complex human traits can rarely be reduced to single cells but rather play out at the tissue or even multi-system level, requiring in vivo, organismal models for mechanistic dissection. Other important factors contributing to complex traits that are challenging, if not impossible, to approximate in cell culture include the impact of developmental biology and aging. Moreover, bioinformatics using large-scale, comprehensive ’omic data can be extraordinarily powerful for hypothesis generation. However, results from these analyses are usually correlative in nature, crucially requiring experimental validation to establish causation (Chakravarti et al., 2013). All experimental systems, including both Drosophila and human subject investigations, have their limitations. As alluded to earlier, many, but not all, genes are conserved, and there are certainly many facets of human disease that are likely not amenable to modeling in Drosophila. Nevertheless, our analyses demonstrate how a majority of human susceptibility loci are highly conserved in flies, creating myriad opportunities for functional follow-up, even where such genes might be operating in different contexts. Given the remarkable scope of the challenges currently encountered in functional genomics, we must leverage all available tools to address these important problems, and it is incumbent upon all stakeholders to embrace such efforts, including investigators, funding bodies and publishers.

In sum, understanding the mechanisms of susceptibility for common and complex genetic diseases is an urgent public health priority. These disorders – including heart/lung disease, cancer, stroke, Alzheimer's disease, diabetes, and many others – account for the overwhelming population burden of morbidity and mortality in the developed world. The successful identification of risk loci by GWAS provides an enormous opportunity for translational research aimed at discovering completely novel drug targets. As highlighted here, susceptibility locus discovery is only the first step, with the crucial challenge remaining to define the relevant mechanisms. Only then can the therapeutic potential of GWAS findings be fully unleashed. We hope that this Review makes a compelling case that Drosophila models offer one important path forward.

The authors are grateful to Juan Botas and Neil Hanchard for feedback on the manuscript.