Genomic approaches for understanding the genetics of complex disease
- 1Division of Endocrinology, Metabolism and Molecular Medicine, Department of Medicine, Northwestern University Feinberg School of Medicine, Chicago, Illinois 60611, USA;
- 2Department of Biostatistics and Bioinformatics, Duke University Medical School, Durham, North Carolina 27708, USA;
- 3Center for Genomic and Computational Biology, Duke University Medical School, Durham, North Carolina 27708, USA
- Corresponding author: tim.reddy{at}duke.edu
Abstract
There are thousands of known associations between genetic variants and complex human phenotypes, and the rate of novel discoveries is rapidly increasing. Translating those associations into knowledge of disease mechanisms remains a fundamental challenge because the associated variants are overwhelmingly in noncoding regions of the genome where we have few guiding principles to predict their function. Intersecting the compendium of identified genetic associations with maps of regulatory activity across the human genome has revealed that phenotype-associated variants are highly enriched in candidate regulatory elements. Allele-specific analyses of gene regulation can further prioritize variants that likely have a functional effect on disease mechanisms; and emerging high-throughput assays to quantify the activity of candidate regulatory elements are a promising next step in that direction. Together, these technologies have created the ability to systematically and empirically test hypotheses about the function of noncoding variants and haplotypes at the scale needed for comprehensive and systematic follow-up of genetic association studies. Major coordinated efforts to quantify regulatory mechanisms across genetically diverse populations in increasingly realistic cell models would be highly beneficial to realize that potential.
The ultimate goal of genetic association studies is both to define the genetic architecture of complex traits and diseases and also to provide new insights into normal physiology and disease pathophysiology. Accomplishing that goal will require defining the causal variants that account for the observed associations, their mechanism of action, and their target genes. Success would have both near- and long-term benefits to health and science. In terms of health benefits, causal relationships between noncoding genetic variants and disease risk can be used to improve the prediction of disease onset and the design of prevention and early detection strategies. Subsequently determining the effects of causal variants on gene expression can prioritize downstream efforts to characterize causal genes and their role in disease etiology. That prioritization is particularly valuable when the target genes have an unknown function. This discovery pathway can ultimately lead to novel and potentially patient-specific therapeutic targets. In terms of scientific benefits, expanding the catalog of noncoding variants that are known to contribute to human traits is needed to determine general and transferrable principles about the genetic basis of complex human diseases. Recent conceptual and technical advances in genetics and genomics together have the potential to greatly improve our understanding of the noncoding genetic contributions to human traits. Although there are a wide variety of ways in which noncoding variants may affect phenotypes, we will focus specifically on variants that alter the activity of gene regulatory elements and, subsequently, the expression of target genes.
The plummeting cost of DNA sequencing has enabled parallel paradigm shifts in human genetics and genomics. For genetic studies, the major benefit has been access to all variants in an individual for association testing. That benefit has been predominantly realized by using whole-genome sequences of related populations to impute the alleles of variants that have not been directly genotyped (The 1000 Genomes Project Consortium 2012; Delaneau and Marchini 2014; Gudbjartsson et al. 2015; Horikoshi et al. 2015; Kuchenbaecker et al. 2015; Surakka et al. 2015) and by whole-exome sequencing (for examples and reviews, see Bamshad et al. 2011; Chong et al. 2015; de Bruin and Dauber 2015). Meanwhile, the first association studies that replace targeted genotyping with whole-genome sequencing are now starting to appear (Gaulton et al. 2013; Morrison et al. 2013; Taylor et al. 2015).
Even with perfect genotype information, there will remain a need for downstream functional studies to identify causal variants that contribute to human phenotypes. One major reason is that the resolution of genetic association is limited by patterns of recombination in the study population: Without recombination between a causal mutation and a nearby noncausal mutation, there is no ability to unambiguously determine which of the two contributes to phenotype with association alone. The ability to discriminate causal effects within those regions requires alternative strategies that effectively separate out the effects of variants that are close to one another on the same chromosome.
Concurrent with the first human population sequencing projects, large and coordinated genomics efforts completed the first comprehensive maps of the molecular state of the human genome and epigenome (The ENCODE Project Consortium 2012; Roadmap Epigenomics Consortium et al. 2015), and hundreds of similar studies have been completed in other biological systems (Fig. 1). The resulting data sets provide researchers with extensive catalogs of transcription factor binding and chromatin states across noncoding genomic regions in a wide diversity of cell types and environmental conditions. Integrating results from studies of genomic regulatory activity with genetic associations has shown initial promise for resolving causal variants of human phenotypes after genetic association, as demonstrated both by overall trends (Wang et al. 2010; Boyle et al. 2012; Ward and Kellis 2012; Zhou et al. 2015) and specific examples (Zhang et al. 2012; Corradin et al. 2014; Huang et al. 2014; Guo et al. 2015).
Number of publications in the NCBI database matching the search queries for RNA-seq (circles) and ChIP-seq (squares). Queries were performed via the NCBI PubMed website (http://www.ncbi.nlm.nih.gov/pubmed/) on July 14, 2015. For RNA-seq and ChIP-seq, the exact query used was “RNA-seq OR RNAseq” and “ChIP-seq OR ChIPseq,” respectively.
Notwithstanding those initial successes, predicting the effect of noncoding genetic variation remains a foremost challenge for several reasons. First, regulatory activity across the genome varies dramatically between cell types and conditions (Thurman et al. 2012). Second, recent evidence suggests that few of the candidate regulatory elements defined by chromatin state and transcription factor binding have strong regulatory activity (Kwasnieski et al. 2014). Even with data supporting regulatory activity of an element, predicting the effects of genetic variants therein is complex and typically requires further empirical investigation. In this Perspective, we will describe recent advances at the interface between genetics and genomics that have improved the ability to identify regulatory mechanisms of disease. We particularly focus on emerging technologies that overcome some of the most eminent challenges and emphasize the need for collaborative studies between genetics and genomics investigators to realize the potential of those technologies.
The landscape of genetic association signals
As of February 2015, genome-wide association studies (GWAS) and other studies had demonstrated the association of more than 15,000 SNPs with a complex disease or trait (Welter et al. 2014). However, the mechanisms underlying these associations remain largely undefined. More generally, the underlying architecture of complex diseases and traits remains poorly defined. The common disease–common variant hypothesis (Gibson 2011) initially predicted that common variants present in all populations underlie phenotypic variation or disease risk and that, together, these variants have an additive or multiplicative effect on disease risk or trait variation. As initial genome-wide association studies failed to account for the observed narrow sense heritability of diseases and traits, alternative explanations have been proposed for the architecture of complex diseases and traits, including (1) a large number of small-effect common variants across the spectrum of allele frequency account for disease risk and quantitative trait variation; (2) a large number of large-effect rare variants underlie observed associations; or (3) a combination of genotypic, environmental, and epigenetic interactions account for the associations (Gibson 2011). It is likely that some combination of those different potential mechanisms accounts for the underlying architecture of complex diseases and traits as common, low frequency, and rare variants have all now been shown to be associated with complex diseases and traits (Fu et al. 2013; Morrison et al. 2013; Ratnapriya et al. 2014; Surakka et al. 2015).
The fundamental problem now faced by geneticists is that variants identified through genetic association studies are typically common SNPs that mark an associated locus rather than the variant that mechanistically contributes to the association. The reason is that alleles of variants that are close together in the genome are likely to be inherited together, a phenomenon known as linkage disequilibrium (LD). The small number of recombination events per human generation, the preferential occurrence of recombination events in certain genomic regions, the history of the population, and other influences all contribute to patterns of LD (for review, see Stumpf and McVean 2003). In the human genome, regions 10–100 kb in size within which causality cannot be inferred are typical (The 1000 Genomes Project Consortium 2012). For that reason, a small number of common variants can represent a large fraction of the genetic variation. Genetic association studies have taken advantage of that LD structure with great success by genotyping common variants rather than those most likely to cause the trait or disease.
The LD patterns that have made genome-wide association studies successful are also a major limiting factor in the identification of causal variants using statistical association alone. For example, a standard approach to identify a causal variant within a locus of genetic association is to first take advantage of different patterns of LD in different ancestry groups to narrow the boundaries of the association locus (for review, see Rosenberg et al. 2010; Edwards et al. 2013). Sequencing the narrowed locus in appropriate populations to identify all the genetic variation across the locus follows. At that point, the number of variants that could contribute to phenotype may number into the hundreds or thousands. Computational strategies to prioritize among the remaining variants based on genomics data and other features may help but, as described above, accurately predicting the functional impact of specific variants on the regulation of gene expression remains a largely unsolved problem.
Further complicating causal variant identification is the possibility that multiple as opposed to a single variant within an LD block may be functional and contribute to the observed association. Analogous to examples in which multiple coding variants in the same gene independently contribute to disease risk (Kotowski et al. 2006; Nejentsev et al. 2009; Rivas et al. 2011), Corradin and colleagues recently suggested a “multiple enhancer variant” (MEV) hypothesis (Corradin et al. 2014) based on investigation into six different autoimmune diseases. In that study, they provide evidence that multiple variants within an LD block impact the activity of multiple different enhancers, and those effects coordinately alter target gene expression. The MEV hypothesis is supported by case studies. In one example, we provided empirical evidence that regulatory variants spanning multiple enhancers within an LD block associated with maternal glucose levels during pregnancy have a coordinated allelic effect on expression of HKDC1 (Guo et al. 2015). Similar patterns were reported previously for the SOX9 region associated with prostate cancer risk (Zhang et al. 2012). The observation that multiple variants within an LD block can affect regulatory element activity and gene expression argues that testing single variants in isolation will be both an inefficient and potentially misleading approach for identifying causal variants. Instead, high-throughput strategies to systematically and comprehensively evaluate the function of variants and haplotypes present in a phenotype-associated locus are needed.
Genetic associations with gene expression reveal target genes
Studies to associate genetic variants with gene expression have generated extensive catalogs of expression quantitative trait loci (eQTLs) in diverse cell types and conditions (e.g., Gamazon et al. 2010; Lappalainen et al. 2013; Liang et al. 2013; The GTEx Consortium 2015). Known eQTLs are highly enriched in variants associated with traits and diseases (e.g., Nica et al. 2010; Nicolae et al. 2010; Torres et al. 2014), and those associations can mark candidate target genes for downstream mechanistic investigation (Cookson et al. 2009). That feature is especially useful when the target gene is not an obvious choice for follow-up because the gene is not in LD with the phenotype-associated variant or because there is not a clear biologic rationale for the association. For example, one of the most robust genetic associations is between genetic variants on Chromosome 16 and body mass index (Dina et al. 2007; Frayling et al. 2007; Scuteri et al. 2007). Despite localization of the associated variants in the first intron of the FTO gene and demonstration of a role for FTO in body weight regulation and fat mass in mouse models (Fawcett and Barroso 2010), recent eQTL analyses have suggested that the Iroquois-related homeobox 3 (IRX3) gene, located more than a megabase away from the most highly associated variant, may also be a causal gene (Smemo et al. 2014; Ronkainen et al. 2015). Mice deficient in Irx3 have obesity and diabetes-related traits, increasing confidence in this mechanistic connection (Smemo et al. 2014). Similar approaches have used eQTL analyses to reveal target genes in other studies (e.g., Teslovich et al. 2010; Innocenti et al. 2011; Hernandez et al. 2012; Farh et al. 2015), supporting the broad utility of the approach.
Expression QTLs have now been mapped for several different tissue types (Schadt et al. 2008; Dimas et al. 2009; Gibbs et al. 2010; Innocenti et al. 2011; Grundberg et al. 2012, 2013) and hormone responses (Maranville et al. 2011). Further studies have investigated the distribution of eQTLs across different tissues from the same individuals (Dimas et al. 2009; Nica et al. 2011; The GTEx Consortium 2015). Although those studies have revealed a substantial degree of shared eQTLs between tissues, the degree of sharing varies across tissues, and certain tissues such as brain appear to have an especially high degree of tissue-specific gene regulation (Hernandez et al. 2012; The GTEx Consortium 2015). The importance of tissue-specific eQTLs is supported by studies showing that GWAS results are specifically enriched for eQTLs in tissues that are relevant to the phenotype (Emilsson et al. 2008; Nica et al. 2010; Below et al. 2011; Brown et al. 2013; Torres et al. 2014). On the other hand, tissue-general eQTLs may be enriched for variants that have a function throughout the body and thus as a class may have a disproportionate effect on phenotypes. To the best of our knowledge, however, the relative contribution of tissue-general eQTLs to phenotypes has yet to be estimated. There is also growing evidence for substantial allelic heterogeneity in gene expression levels (Brown et al. 2013), in agreement with the previously described observations of multiple coordinated regulatory variants in disease loci (Zhang et al. 2012; Corradin et al. 2014; Guo et al. 2015). Taken together, increasing the diversity of primary tissues, cell types, and environments for which eQTLs have been mapped is likely to be highly valuable for identifying variants, genes, and tissues that contribute to phenotypes.
Genomic regulatory elements are highly enriched in phenotype-associated variants
Although eQTLs have demonstrated value in identifying target genes for genetic association studies, they too suffer from the same limitation that the most strongly associated variant may not be causal of the regulatory event. Comprehensive genomic measurements of the epigenetic and regulatory state of the genome, made possible by high-throughput sequencing, can overcome that limitation because LD does not limit their resolution. Many different assays have been developed to measure genomic components of gene regulation, and we will focus on two widely used approaches, chromatin accessibility mapping and ChIP-seq. The accessibility of chromatin to various enzymes such as DNase I is a well-established indicator of genomic regulatory activity. High-throughput sequencing-based assays such as DNase-seq and ATAC-seq exploit that principle to reveal comprehensive maps of chromatin accessibility across the human genome (Song and Crawford 2010; Thurman et al. 2012; Buenrostro et al. 2013). Similarly, the high-throughput sequencing version of chromatin immunoprecipitation, ChIP-seq (Johnson et al. 2007; Mikkelsen et al. 2007; Robertson et al. 2007), is now commonly used to identify binding sites for transcription factors and histone modifications associated with regulatory states of the human genome. ChIP-seq can localize a binding event or a modified histone to within 50 bp, and DNase-seq has a similar resolution.
There is now strong evidence that genetic variation within candidate regulatory elements identified with DNase-seq or with ChIP-seq contributes to human phenotypes. For example, several studies have found that phenotype-associated variants are enriched DNase- or ChIP-positive regions in a tissue-specific manner, and that tissue specificity can be used to implicate unexpected tissues in disease etiologies (Ernst et al. 2011; Maurano et al. 2012; Schaub et al. 2012; The ENCODE Project Consortium 2012; Parker et al. 2013; Pickrell 2014). Genetic variation in the same regions also accounts for a substantial and significantly enriched fraction of the heritability of complex human diseases (Gusev et al. 2014). Moreover, using an association approach similar to that used to identify eQTLs, several studies have identified genetic variants that are correlated with changes in chromatin accessibility, histone modifications, and DNA methylation; and they have shown that the identified variants explain a large fraction of eQTL associations (Degner et al. 2012; McVicker et al. 2013; Banovich et al. 2014). Together, those results suggest that variation in regulatory elements is a primary contributor to expression phenotypes. Those overall enrichments motivate a deeper investigation into the genetic architecture of gene regulation, with a particular focus on determining the specific variants that alter regulatory element activity.
Allele-specific genomic activity reveals candidate mechanisms of disease
The use of high-throughput sequencing as readout for DNase and ChIP assays not only improves detection of regulatory elements but also allows for the simultaneous observation of the genetic sequence of the identified elements. If one or more positions within the identified regulatory element are heterozygous in an individual, that feature can be leveraged to estimate the abundance of each allele in the assayed DNA. A significant deviation from the expected ratio based on that individual's genome indicates an allele-specific difference in the activity of that element. Allele-specific analyses were first used to investigate gene expression using a variety of approaches, including targeted sequencing, RT-qPCR, and microarrays (e.g., Singer-Sam et al. 1992; Yan et al. 2002; Bray et al. 2003; Lo et al. 2003; Pastinen et al. 2004; Gimelbrant et al. 2007; Ge et al. 2009; Adoue et al. 2014; for review, see Knight 2004; Pastinen 2010). Those studies typically found evidence for allele-specific expression of ∼10% of human genes. High-throughput sequencing of RNA advanced the field by making it possible to measure allele-specific gene expression genome-wide and agnostic of reference gene annotations (e.g., Degner et al. 2009; McManus et al. 2010; Pickrell et al. 2010; Reddy et al. 2012). Allele-specific analyses of DNase-seq and ChIP-seq data use similar strategies to reveal variants associated with chromatin state or transcription factor binding (Kasowski et al. 2010; McDaniell et al. 2010; Reddy et al. 2012). If the observations reflect direct local effects of genetic variation on gene regulation (Fig. 2A), then such allele-specific analyses can be used to identify individual causal variants within large LD blocks identified via association studies. On the other hand, long-range regulatory interactions may limit the ability to pinpoint individual genetic causes.
Mechanisms of allele-specific transcription factor occupancy. (A) Local effects occur when a genetic variant directly impacts the ability of a transcription factor to bind DNA. In this example, only the A allele is bound by the transcription factor and recovered by ChIP. (B) Genetic variants may lead to allele-specific binding of entire regulatory complexes. In this example, transcription factor TF1 binds the A but not the G allele. Because TF1 also recruits TF2 to the same regulatory complex, ChIP-seq for TF2 preferentially isolates the A and C alleles even though TF2 does not directly bind either variant. (C) Long-range interactions may also drive distal allele-specific effects. One potential mechanism is that TF1 and TF2 form a regulatory complex via DNA looping. Because occupancy of TF1 influences that of TF2, variants that impact TF1 binding lead to an allele-specific signal for TF2 occupancy.
Several findings indicate that allele-specific effects are indeed local events. The initial allele-specific ChIP-seq studies found that genetic variants with allele-specific transcription factor binding are enriched near the specific nucleotides bound by the transcription factor (Reddy et al. 2012). The inverse was also true: Variants without allele-specific binding were depleted near transcription factor binding sequences (Reddy et al. 2012). Moreover, the variants with the strongest effects typically altered the DNA sequences bound by transcription factors (Kasowski et al. 2010; Reddy et al. 2012). Allele-specific transcription factor binding and, to a lesser extent, chromatin state, are both heritable, indicating a clear genetic contribution (McDaniell et al. 2010; Reddy et al. 2012; Kasowski et al. 2013; Kilpinen et al. 2013). Although changes in the DNA sequence bound by the transcription factor explain the largest effects, most allele-specific effects are modest and cannot be explained by changes in the transcription factor binding sequence (Reddy et al. 2012). One potential explanation is that transcription factors often bind the human genome in complexes often referred to as cis-regulatory modules (for review, see Hardison and Taylor 2012). Observed allele-specific transcription factor binding may result from genetic variants that disrupt binding of other transcription factors in the same module (Fig. 2B). That model is supported by a strong degree of allele-specific coordination between multiple transcription factors and chromatin state at the same genomic locus (Reddy et al. 2012; The ENCODE Project Consortium 2012; McVicker et al. 2013; Soccio et al. 2015). On the other hand, observations of long-range coordination in allele-specific chromatin indicate that local effects of regulatory variants may affect distal sites on the same chromosome (Fig. 2C; Kilpinen et al. 2013). In that scenario, LD would still impair resolution of allele-specific analyses for identifying causal variants. Taken together, it is likely that allele-specific measurements of the regulatory state reflect a mixture of local and distal effects. Although the relative proportion of local and distal signals is not yet known, the contributions of local effects to the overall signal likely provide some ability to identify causal variants within regions of high LD.
There are several additional advantages of allele-specific analyses over association-based studies that motivate their increased use. Because the two alleles compete for regulatory factors in the same nucleus and in the same environment and because both alleles undergo the same sample processing steps, variation due to sample history or handling is unlikely to contribute to false positives. Allele-specific analyses also have a practical advantage in that, unlike for association studies, a large cohort of individuals is not needed to detect an allele-specific effect at an individual variant. In cases in which samples are rare or difficult to obtain, an allele-specific approach may therefore be the only viable path forward for identifying genetic associations with regulatory element activity. Finally, because comparisons are made between the two alleles present in the same individual, the power to detect an allele-specific effect of a heterozygous variant in that individual does not depend on the population frequency of the variant.
The corresponding limitations are that only heterozygous sites in an individual are informative, and there are additional analytical challenges over genetic association studies. The limitation to heterozygous sites means that homozygous individuals do not contribute to the power to detect an effect, and also that large cohorts are still needed to observe a rare variant. Ideally, it would be possible to combine allele-specific analyses with standard genetic association, and newly developed approaches to do so are a promising advance (van de Geijn et al. 2014). The primary additional analytical challenge is aligning short-read sequences in a manner that is not biased to the reference genome. Such alignment biases arise from numerous sources, including unobserved genetic variation and repetitive sequences (Degner et al. 2009; Stevenson et al. 2013). Alignment biases have been overcome previously by aligning sequences to personal genome sequences (McDaniell et al. 2010; Rozowsky et al. 2011; Reddy et al. 2012), prefiltering genomic regions prone to bias (Degner et al. 2009), and the development of variant-tolerant alignment algorithms (Wu and Nacu 2010). Once alignments are made, detecting a bias toward one allele requires statistical approaches to handle overdispersion in read-count data. Software packages for allele-specific alignment and statistical analysis have now been established, making those advances available to a wide diversity of researchers (Wu and Nacu 2010; Rozowsky et al. 2011; Skelly et al. 2011; van de Geijn et al. 2014). As for genetic association studies, allele-specific analyses rely on dense genotyping for each individual. However, recent advances to perform allele-specific analyses without supporting genome sequence information may remove that limitation (Harvey et al. 2015; Romanel et al. 2015). Such advances are major breakthroughs because they greatly reduce cost and complexity. For all of the aforementioned reasons, expanding studies to include allele-specific analyses is a promising strategy to improve the identification of causal regulatory variants in a diversity of tissues and cell types.
High-throughput measurement of regulatory element activity
One major outstanding challenge subsequent to the widespread adoption of DNase-seq and ChIP-seq is to reconcile the abundance of candidate regulatory elements identified. One likely explanation is that a small fraction of candidate regulatory elements are highly active and those elements affect the majority of gene regulation. Reporter-gene expression assays have a distinct advantage over eQTL associations and allele-specific genomic analyses because they directly measure the regulatory activity of a genomic sequence. Briefly, in a reporter assay, a candidate regulatory element is introduced into a plasmid that contains an easily observable reporter such as a fluorescent or chemiluminescent protein (Fig. 3A). The plasmid is then introduced into cells of interest by any of a variety of approaches. Once the plasmid enters the nucleus, transcription factors and RNA polymerases bind the plasmid and control reporter gene expression. Because reporter assays isolate regulatory elements from the surrounding genomic context, results are independent of adjacent elements that may be in LD. Critical for genetic studies, reporter assays can be used to estimate the effect of genetic variants on regulatory activity by comparing the activity of different alleles of the same regulatory element. For those reasons, reporter assays have been valuable for identifying individual regulatory variants that contribute to phenotype (Musunuru et al. 2010; Feng et al. 2013; Fogarty et al. 2014; Stadhouders et al. 2014; Guo et al. 2015).
High-throughput reporter assays. (A) In a standard reporter assay, a candidate regulatory element is placed upstream of a reporter gene that is expressed from a constitutively active promoter. (B) In a high-throughput version of the same system, a random DNA sequence known as a molecular barcode is inserted into the 3′ UTR of the reporter gene, and a library of candidate regulatory elements are placed upstream of the promoter. Each individual candidate regulatory element is physically linked to a unique molecular barcode. Measuring the expression of each molecular barcode can then be used to estimate the activity of the associated regulatory element. (C) An alternative strategy is to clone the library of candidate regulatory elements directly into the 3′ UTR of the reporter gene. By that construction, the regulatory element controls its own expression, which can be measured with paired-end high-throughput sequencing. (D) The preceding strategy can be modified for genetic studies by cloning genetically diverse regulatory elements captured from donor genomes into the reporter gene. By that construction, each allele is expressed at a level that is directly related to its regulatory activity.
The major drawback to standard reporter assay systems for genetic screens is the throughput. Because readout is limited to a single reporter gene, assays must be individually constructed and assayed. Multiwell plates and automated liquid handling increase throughput substantially (e.g., Landolin et al. 2010; Whitfield et al. 2012), but not to the extent required to routinely comprehensively assay regulatory variants in an entire LD region identified by a genetic association study. To address the need to increase the scale of reporter assays, high-throughput versions have been developed in which regulatory activity is measured using high-throughput sequencing rather than by observing a fluorescent protein. One strategy is to construct a library of regulatory elements that are uniquely associated with DNA barcode sequences embedded in an otherwise ignored reporter gene (Fig. 3B). High-throughput sequencing of the expressed barcodes can then be used to estimate activity of the associated element (Kwasnieski et al. 2012; Melnikov et al. 2012; Patwardhan et al. 2012). Each of the initially published mammalian examples focused on evaluating the effects of genetic variants within a small set of previously defined regulatory elements. The CRE-seq assay developed by Kwasnieski et al. (2012) used high-throughput DNA synthesis to generate and assay more than 1000 genetic variants of a 52-bp rhodopsin promoter. Similarly, Melnikov et al. (2012) used DNA synthesis to generate and assay more than 27,000 variants of two 87-bp inducible enhancers. In both cases, the length of the regulatory elements assayed was limited by DNA synthesis. Patwardhan et al. (2012) used degenerate PCR rather than DNA synthesis to generate random mutants of three known liver enhancers. That approach enabled generation of more unique versions (more than 100,000) of longer regulatory elements (258–619 bp). Finally, a related strategy known as functional identification of regulatory elements within accessible chromatin (FIREWACh) assayed captured DNase hypersensitive regions rather then predefined regulatory elements. By combining DNase-seq with a reporter assay, Murtha and colleagues were able to agnostically quantify the activity of approximately 80,000 open chromatin sites in a single assay (Murtha et al. 2014).
For barcode-based approaches, each regulatory element to be assayed must be linked with a unique barcode in the assay. The STARR-seq approach uses a different library construction strategy that obviates that step. In STARR-seq, the reporter element itself is cloned into the 3′ untranslated region (UTR) of the reporter gene and serves as its own barcode (Fig. 3C; Arnold et al. 2013). The major advantage of STARR-seq is that the greatly simplified library construction makes the approach particularly amenable to assaying highly diverse libraries of randomly fragmented DNA. Specifically, in the initial demonstration of STARR-seq, libraries of more than 10 million unique regulatory fragments were assayed, and the median size of the regulatory fragments was ∼600 bp. That level of diversity was sufficient to agnostically assay the entire Drosophila melanogaster genome in multiple cell lines and six human bacterial artificial chromosomes ranging in size from 150 to 185 kb. As in the FIREWACh assay, combining capture of regulatory elements with STARR-seq has enabled focused investigation of genomic regions that are of interest because they are likely to be functional or because they are associated with disease (Vanhille et al. 2015; Vockley et al. 2015). The STARR-seq assay has now been used to investigate changes in regulatory activity across species (Arnold et al. 2014), in response to hormones (Shlyueva et al. 2014), and in combination with different promoters (Zabidi et al. 2015), highlighting the flexibility of the approach.
Comprehensive evaluation of human and mouse candidate regulatory elements identified using ChIP-seq, DNase-seq, and integrative techniques has revealed that only a small fraction of elements typically have a strong effect on gene expression (Kwasnieski et al. 2014; Murtha et al. 2014). Interestingly, with growing evidence that transcription factors tend to bind the genome in heterotypic and homotypic clusters, Smith and colleagues used a massively parallel reporter approach to show that heterotypic clusters of transcription factors are especially potent regulators of gene expression (Smith et al. 2013). Similarly, mutagenesis studies have shown that mutations in distal regulatory elements typically have modest effects on regulatory element activity (Melnikov et al. 2012; Patwardhan et al. 2012). Together, these results indicate that allele-specific DNase-seq and ChIP-seq will be useful to reduce the search space for causal variants, but that additional functional assays will be needed to identify individual causal regulatory variants.
Using high-throughput assays to measure the effects of noncoding variants in GWAS cohorts is one possible strategy that uses existing technology to identify causal variants underlying a genetic association result. As an initial example, we recently used STARR-seq to measure the activity of candidate regulatory elements captured from the genomes of 95 individuals from a recent genetic association study (Fig. 3D; Urbanek et al. 2013; Vockley et al. 2015). That population-scale approach allowed identification of functional regulatory variants within a genetically linked region of association. Because the capture was performed from donor genomes, all variants and haplotypes tested were found in the study population, including a substantial fraction of variants not found in existing databases. We expect that continued development and application of such high-throughput reporter assays to expanded populations is a promising strategy to connect genetics and genomics and thereby reveal causal variation within large genomic regions associated with disease.
As with any approach, reporter assays have limitations, some of which can be mitigated with improved study designs. Regulatory element activity may require additional contexts such as a specific promoter, genomic integration, or cellular environment. Those concerns can be largely addressed with experimental designs that include custom promoters (Zabidi et al. 2015), genomic integration (Dickel et al. 2014; Murtha et al. 2014), and strategies to assay libraries in vivo (Kwasnieski et al. 2012; Patwardhan et al. 2012; Smith et al. 2013). Genome and epigenome editing strategies are also emerging as complementary strategies to investigate regulatory element activity in vivo (Mendenhall et al. 2013; Yin et al. 2014; Hilton et al. 2015). However, there are many contexts for which tractable culture models do not yet exist, and continued development of more realistic models will remain invaluable to determine contributions of regulatory variation to disease.
Finally, although reporter assays are a promising strategy to identify causal regulatory variants, integration with results from other approaches, such as ChIP-seq and eQTL studies, will be needed to identify the responsible transcription factors and the causal genes, respectively. An additional exciting possibility is the integration with chromatin conformation assays such as ChIA-PET (Li et al. 2014), Hi-C (Belton et al. 2012), and variants thereof (Jäger et al. 2015) that can reveal physical interactions between causal variants and causal genes.
Informing future association studies by improving models of regulatory variants
Prioritizing variants that are most likely to have a phenotypic effect is a promising strategy for improving resolution within an associated locus (for review, see Cooper and Shendure 2011). Briefly, that strategy is commonly applied to genetic variants in coding regions, where gene annotations, codon sequences, and protein structure can guide analyses (Sunyaev et al. 2001; Ng and Henikoff 2003; Adzhubei et al. 2010, 2013; Price et al. 2010; Schwarz et al. 2010; Hu et al. 2013; Ionita-Laza et al. 2013). A current need is to improve understanding of the basic characteristics of the types of variants that most impact gene expression to support the development of analogous methods for noncoding genomic regions. Approaches to computationally predict the effects of genetic variants in noncoding regions have largely relied on evolutionary conservation in closely related species (e.g., Cooper et al. 2005; Siepel et al. 2005; Pollard et al. 2010). With a dramatic increase in empirical data describing the epigenetic and regulatory state of the genome, integrative strategies have also recently emerged that combine both conservation and empirical data to identify and predict the effects of noncoding variants (Lee et al. 2011; Ernst and Kellis 2012; Hoffman et al. 2012; Ward and Kellis 2012; Khurana et al. 2013; Kircher et al. 2014; Ritchie et al. 2014; Shihab et al. 2015). Such integrative approaches are limited by the empirical data available. As the high-throughput empirical approaches described above are further developed and applied, integrative strategies to predict the effects of noncoding variants are likely to immediately benefit. Meanwhile, the catalog of regulatory variants that are known to contribute to human phenotypes will be greatly expanded. That expansion is critical to support the development of additional guiding principles for the interpretation of noncoding regulatory variation. For example, certain classes of variants, such as rare variants or variants in regions with specific histone modifications, may be more likely to alter regulatory element activity. Another possibility is that the three-dimensional structure of the genome will help to model genetic effects on target gene expression. Early indications suggest that prioritization based on such models will be helpful, and the full extent of possibilities remains to be determined.
Discussion
The aforementioned approaches are only a few examples of the ways in which the integration of genomic and genetic analyses can inform our understanding of noncoding mechanisms of human phenotypes. Although by no means comprehensive, the preceding examples represent both the basic principles and common challenges of using genomic assays to inform a positive genetic association. Specifically, as LD is disrupted in experimental systems, it becomes easier to finely map causal variants. The sacrifice is that context-dependent regulation and the identity of target genes is typically lost. For that reason, we expect that the greatest value for mechanistic interpretation will be achieved when integration across multiple levels of resolution reveals both candidate causal variants and one or more target genes for downstream study.
This Perspective has focused on associations with complex phenotypes in which causal variants are difficult to ascertain. Similar strategies will likely benefit the investigation of rare Mendelian diseases, specifically in cases in which a causative coding mutation has not been found. Especially in cases of intermediate phenotypes, regulatory mutations are a plausible explanation for the missing diagnosis (Weedon et al. 2014). Identifying such cases will likely improve diagnosis and could reveal patients who may be candidates for novel treatments.
Much of the work described here involved efforts of major consortia focused specifically on genetic associations or on highly coordinated genomic studies. Moving forward, we believe that the greatest benefit for human health will be obtained through joint studies that integrate both genetic and genomic principles in their design. We have found that such highly interactive studies yield substantial mutual benefits. For example, performing high-throughput reporter assays on DNA from genetic association cohorts required access to unique DNA samples that had been collected over several years and expert knowledge about the cell models and conditions that are most relevant to the phenotype (Guo et al. 2015). Then, once regulatory variants are identified via genomic strategies, follow-up genotyping in an independent cohort is needed to confirm the effects and to establish genetic risk scores. Meanwhile, informed genetic association based on genomic evidence of activity will require the generation of new functional genomic data in relevant cell models to support that development. Joint research efforts by genetics and genomics teams dramatically lowers the bar for such cross-cutting activities and, for that reason, we believe that such approaches will be well-positioned to realize translational benefits of biomedical research in both the short and long term.
Acknowledgments
This work was funded by National Institutes of Health R01 grants DK099820 and DK097534.
Footnotes
-
Article and publication date are at http://www.genome.org/cgi/doi/10.1101/gr.190603.115.
-
Freely available online through the Genome Research Open Access option.
This article, published in Genome Research, is available under a Creative Commons License (Attribution-NonCommercial 4.0 International), as described at http://creativecommons.org/licenses/by-nc/4.0/.
