Using “-omics” Data to Inform Genome-wide Association Studies (GWASs) in the Osteoporosis Field

Abstract

Purpose of Review

Osteoporosis constitutes a major societal health problem. Genome-wide association studies (GWASs) have identified over 1100 loci influencing bone mineral density (BMD); however, few of the causal genes have been identified. Here, we review approaches that use “-omics” data and genetic- and systems genetics–based analytical strategies to facilitate causal gene discovery.

Recent Findings

The bone field is beginning to adopt approaches that are commonplace in other disease disciplines. The slower progress has been due in part to the lack of large-scale “omics” data on bone and bone cells. This is however changing, and approaches such as eQTL colocalization, transcriptome-wide association studies (TWASs), network, and integrative approaches are beginning to provide significant insight into the genes responsible for BMD GWAS associations.

Summary

The use of “-omics” data to inform BMD GWASs has increased in recent years, leading to the identification of novel regulators of BMD in humans. The ultimate goal will be to use this information to develop more effective therapies to treat and ultimately prevent osteoporosis.

Introduction

Osteoporosis is characterized by low bone mineral density (BMD) and deteriorated bone microarchitecture which leads to an increased risk of fracture [1, 2]. In the USA, over 12 million individuals have been diagnosed with osteoporosis, leading to over 2 million fractures per year, a number expected to nearly double by 2025 [3]. Importantly, of the ~300,000 people that suffer from a hip fracture annually, 1 in 5 will die in the subsequent 12 months [4]. Osteoporotic fractures are also costly accounting for approximately $26 billion in health-care expenditures [3].

BMD is one of the strongest predictors of fracture [5]. It is also a highly heritable quantitative trait (h² = 0.5–0.8) that can be measured in large cohorts of individuals [6-9]. Although genome-wide association studies (GWASs) have become the mainstay of investigations into the genetic basis of BMD, genetic studies of bone traits began before the GWAS era [10]. Prior to GWASs, genetic studies of BMD and osteoporosis involved linkage in families and candidate gene association studies [9]. Linkage studies identified several loci for BMD; however, with notable exceptions (see [11] as an example), the challenges of gene discovery in the context of linkage studies limited their utility for unraveling complex traits such as BMD. Additionally, replication of loci identified by linkage has been low [12]. Similarly, candidate gene studies identified several associations for BMD, few of which have been replicated in large cohorts [13, 14].

Most GWASs conducted for osteoporosis have focused on BMD. BMD can be measured using dual-energy X-ray absorptiometry (DEXA) or quantitative ultrasound (generates measures of estimated BMD (eBMD)). The largest GWAS for DEXA-derived lumbar spine and femoral neck BMD was performed on ~80K individuals and identified 56 loci [15, 16]. The largest eBMD GWAS performed to date used the UK Biobank (N~420K) and identified 501 loci harboring 1103 independent associations which explain 20.3% of the total variance of the trait [14].

Although GWASs have revolutionized the identification of BMD loci, few of the underlying causal genes have been identified. This is largely due to the fact that, unlike Mendelian disease, >90% of GWAS loci for common diseases are due to non-coding variants [17]. This suggests that most associations are caused by changes in gene regulation [18]. As a result, it is possible that variants in a GWAS locus may regulate a gene a considerable distance (100s of Kbps) up- or downstream. Additionally, extensive linkage disequilibrium (LD) adds to the difficulty in assigning target genes to loci and identifying the underlying causal variant(s) at each locus [19]. Together, these challenges have made it difficult to pinpoint causal genes highlighting the importance of developing novel approaches to inform BMD GWASs [20].

There are three primary reasons why causal gene discovery is important. First, the identification of new genes responsible for variation in BMD from GWAS [21, 22] has already shed light on important new processes impacting bone [23, 24]. This will only continue and increase in impact as the approaches we discuss below are utilized more widely to interrogate BMD GWAS. Second, the hopes of precision medicine for osteoporosis, which aims to tailor therapeutics based on individualized risk factors (i.e., an individual’s genotype), rely on a comprehensive understanding of the genes impacting bone. Third, and possibly the most important, GWAS is a powerful approach to identify antiosteoporotic therapeutic targets. Historically, many drug targets from traditional studies have failed in clinical trials [25]. There are many reasons for these failures including that targets of investigation are often not causally linked to a disease. Recently, it has been shown that drug targets with evidence from genetic studies (including GWAS) are twice as likely to succeed in clinical trials [25, 26]. Together, these factors are driving the focus on causal gene discovery.

In this review, we highlight how molecular “-omics” data and cutting-edge analytical approaches are being used to facilitate gene discovery from GWAS. Our aim is not to be comprehensive but to highlight specific studies that demonstrate how “-omics” data and analytical approaches can be used to “make sense” of GWAS. We also discuss resources that are needed in the bone field and novel approaches that will be used in the coming years.

Approaches for Causal Gene Discovery

Since the first GWAS for any disease in 2006 [27], several approaches have been developed with the goal of identifying causal genes. All of these approaches leverage the generation of molecular data and its analysis using genetic- and/or systems genetics–based analytical strategies. The concept is simple; genetic variants mediate their effects on a phenotype by altering molecular changes, such as differences in gene expression, alternative splicing, intron retention, protein levels, protein activities, and molecular interactions. As a result, the identification of disease-associated variants that influence molecular phenotypes allows one to identify causal genes and begin to unravel their mechanisms of action.

The advent of next-generation sequencing (NGS) has made profiling many of these changes straightforward and feasible to do in large human populations. This has led to a revolution in the use of molecular quantitative trait locus (QTL) data which are beginning to help us understand how BMD-associated variants impact molecular processes and, in turn, how these changes influence BMD and ultimately risk of fracture. The next logical step is to identify the causal genes in order to enhance our understanding of disease mechanisms. At the heart of this step are the approaches to correctly annotate causal genes and understand how they influence bone. This section will address current concepts related to establishing the cause-and-effect relationship between BMD GWAS reported genetic variants and their respective functional genes in bone.

Expression Quantitative Trait Loci (eQTL) Colocalization

One of the most widely used approaches to inform GWAS is through identification and colocalization of expression quantitative trait loci (eQTL) [28]. An eQTL is an association between a set of genetic variants and gene expression levels [28] (Figs. 1 and 2). eQTLs are divided into two categories based on their proximity to target genes: local (also referred to as cis) and distant (also referred to as trans). Local eQTLs are located in close proximity (typically defined as ± 1 Mbp) to the gene they regulate [28]. An example of a local eQTL would be a polymorphism in the promoter of a gene that leads to altered transcription factor binding and allele-specific expression. In contrast, distant eQTLs are located far from the genes they regulate and are often on different chromosomes [28]. A distant eQTL could manifest from a polymorphic transcription factor that influences its target genes differently based on its genotype. The first step in the identification of eQTLs consists of collecting and profiling the transcriptome of disease-relevant tissues or cell types using RNA-seq in a population of densely genotyped individuals. These data are then used to identify eQTLs by conducting association tests between millions of single-nucleotide polymorphisms (SNPs) and thousands of genes. The most direct and common way that local eQTLs are used to inform GWAS is through colocalization [30, 31]. Colocalization is a set of statistical approaches that test the hypothesis that an eQTL and GWAS association (or any two associations) are driven by the same shared variant (Table 1, Fig. 2). Essentially, colocalization is testing whether or not BMD-associated variants also influence gene expression. If so, then one can hypothesize that the BMD-associated variants influence gene expression and the change in gene expression alters BMD.

Fig. 1.

Examples of quantitative trait loci (QTL) for molecular phenotypes such as gene expression (eQTL), alternative splicing (sQTL), DNA methylation (meQTL), chromatin accessibility (caQTL), and protein expression (proQTL). QTL approaches can generally be applied to any molecular trait quantifiable in a population of individuals.

Fig. 2.

eQTL discovery and colocalization. A Two examples of an eQTL with box plots showing that gene expression is (left) or is not (right) correlated with the genotype of a single-nucleotide polymorphism (SNP). B An example of a colocalizing eQTL for MARK3 visualized using RACER [29]. Every circle represents a SNP. An eQTL for MARK3 is shown on the top panel of the mirror plot. The BMD GWAS association is plotted on the bottom panel. Note that the same SNPs associated with BMD are associated with the expression of MARK3. The colors signify r², a measure of linkage disequilibrium.

Table 1.

Defining concepts used in this review

Approach	Description	Reference
Genome-wide association study (GWAS)	Hypothesis-free statistical approach that identifies associations of genetic variants and diseases or disease-associated traits. GWASs are divided into two types: case-control and quantitative	[14, 16, 32]
Transcriptome-wide association study (TWAS)	Statistical approach that leverages gene expression imputation to identify significant gene-trait associations by estimating the genetic component of gene expression in a reference population where gene expression and genotype have been measured (GTEx is an example) and then imputing (predicting) gene expression in a much larger population (such as those used in a BMD GWAS). Once gene expression is imputed, genetically regulated gene expression is associated with a disease or disease phenotype	[33, 34]
Colocalization	Statistical test to determine whether a single variant is responsible for both GWAS and molecular QTL signal (i.e., eQTL) in a genomic region. One can draw a parallel to fluorescence microscopy where colocalization refers to the observation of overlap between different fluorescent labels that tag different “targets” located in the same area of the cell	[30, 34]
Network analysis	Approaches using “-omic” data (e.g., RNA-seq, proteomics, etc.) to partition genes into groups based on functional similarities in an unbiased manner. A growing number of studies have demonstrated the ability of networks to predict causal genes at GWAS loci	[35-37]

Several studies have demonstrated that many eQTLs are tissue or cell type specific, likely reflecting the cell type–specific nature of the epigenome [38, 39]. As a result, eQTLs identified in disease-relevant tissues or cell types are likely to be the most informative for use in GWAS colocalization [28, 38]. One of the largest projects for eQTL identification and analysis is the Genotype-Tissue Expression (GTEx) project [38]. GTEx is an ongoing effort to build a comprehensive public resource to study tissue-specific gene expression and regulation. The project has profiled ~50 tissues in hundreds of individuals using RNA-seq and identified thousands of eQTLs [38, 39]. In fact, GTEx has identified significant local eQTL for nearly 95% of all protein-coding genes in the human genome [39]. While GTEx has significantly increased our understanding of eQTLs and how they mediate the effects of GWAS, one of its limitations for the bone field is that bone or bone cells are not included in the 50 tissues investigated. Fortunately, studies are beginning to be conducted that identify eQTLs in bone and bone cells [40].

Bone tissue is primarily made up of three major cell types: osteoblasts, cells that form bone; osteoclasts, cells that break-down bone; and osteocytes, cells that coordinate the function of both osteoblasts and osteoclasts. A dynamic equilibrium between these cell types ensures that the skeleton is being properly maintained through bone remodeling. Ideally, we would have eQTL data (and other molecular data types) on all three important cell types.

The first and only osteoblast eQTL dataset was generated in 2009 from primary cultured human osteoblasts (HOb) derived from 95 unrelated donors of Swedish origin [41]. Global gene expression from these cells were profiled using microarray technology [41]. The authors then identified local eQTLs and used them to prioritize the genomic loci from one of the first BMD GWAS studies [42]. They identified serine racemase (SRR) as a novel BMD-associated gene [41]. Since its publication, this dataset has also been used by groups performing BMD GWAS to help identify causal genes [16, 32, 43]. However, its sample size, low-density genotyping, and use of microarray technology to profile expression have limited its effectiveness for gene discovery. The second population scale transcriptomic dataset was generated in 2010 on iliac crest bone biopsies from 84 postmenopausal women in Norway [44]. This study suffers the sample limitations as described for the study above, but it has been used by several groups performing GWAS to provide insight into potentially causal genes [16, 45].

In 2018, Mullin et al. [40] generated a RNA-seq–based eQTL dataset using osteoclast-like cells differentiated in vitro from peripheral blood mononuclear cells (PBMCs) obtained from 158 female patients. These data were used to identify genes with eQTLs that colocalized with loci from two eBMD GWASs [14, 32]. The authors used coloc [30], a widely used colocalization approach that tests whether association signals are driven by the same causal variant. In the first study [40], using 307 BMD GWAS significant SNPs, eight genes were reported to have a significantly colocalizing eQTL. In the second study [46], using 1103 significant GWAS SNPs, evidence of colocalization of GWAS and eQTL association signals was identified for 21 genes. The low percentage of GWAS loci with a colocalizing eQTL in osteoclasts may reflect the cross-sectional nature of the GWAS with differences in BMD being driven primarily by bone formation.

Studies support the use of eQTL data in aiding the interpretation of GWAS results in other disease fields such as Crohn’s disease [47], bipolar disease [48], and diabetes [49]. Here we highlight the transcription factor KLF14 and its role in type 2 diabetes (T2D) [50]. The genetic variants associated with T2D and other metabolic phenotypes map to a region of 3–48 kb upstream of KLF14 [50]. The GWAS SNPs associated with the KLF14 are colocalized with eQTLs only in adipose tissue despite KLF14 being expressed in multiple tissues. Small et al. [50] showed that these SNPs act in adipose tissue to reduce KLF14 expression and modulate, in trans, expression of 385 genes. The study also demonstrated the mechanism in which KLF14 expression increases pre-adipocyte proliferation but disrupts lipogenesis [50]. Additionally, in vivo knockout in adipose tissue in mice partially recapitulated the human phenotype of insulin resistance, dyslipidemia, and T2D [50]. This is an excellent example of how eQTL data can inform GWAS and how such findings could similarly be used in the bone field, especially once large-scale bone-relevant datasets have been generated.

Most eQTL studies focus on “total” gene expression that is transcript levels summed over the exons of a gene. However, genetic variation can impact all aspects of transcriptional and posttranscriptional regulation [51]. For example, recent studies have identified splicing QTL (sQTL), which are loci influencing mRNA splicing [38] (Fig. 1). No sQTL studies have been conducted in bone or bone cells; however, this approach has been used in other tissues [52]. For example, 8966 sQTL were identified using dorsolateral prefrontal cortex (DLPFC) RNA-seq data from >200 individuals [53]. When they compared sQTL SNPs and GWAS SNPs (an approach similar to colocalization, but statistically less stringent), a significant overlap was observed for schizophrenia and other diseases, suggesting that part of the genetic risk for complex diseases is due to sQTL. Therefore, to facilitate comprehensive gene discovery, future eQTL studies in bone should address how BMD-associated variants impact all levels of gene regulation.

Transcriptome-wide association studies (TWASs)

In recent years, transcriptome-wide association studies (TWASs) have become widely used approaches that utilize gene expression data to measure the association between genetically regulated gene expression and complex phenotypes [33] (Table 1). In an individual, gene expression (and as noted above, other aspects of gene regulation) is influenced by genetics and the environment [54, 55]. The genetic contribution to gene expression can be quantified using eQTL and used to predict or impute expression in an individual based on genotype. For example, if a local eQTL explains 100% (no environmental contribution in this hypothetical example) of the variance in gene x, then all we need to know is an individual’s eQTL genotype to know the expression of gene x in that individual. TWAS extends this example by estimating the genetic component of gene expression (using the advanced statistical analysis of eQTL data) across the genome in a reference population where gene expression and genotype have been measured (GTEx is an example) and then imputing (predicting) gene expression in a much larger population (such as those used in a BMD GWAS) [33]. Once gene expression is imputed, genetically regulated gene expression is associated with a disease or disease phenotype. Most genes identified by TWAS are located in GWAS associations for that disease (due to genetically regulated differences in gene expression being the basis of most GWAS associations). As a result, TWAS can pinpoint genes likely to be causal at GWAS loci.

TWASs for BMD are sparse but are starting to be performed. In one of the first studies, gene expression data from GTEx muscle and whole blood tissues in combination with the largest eBMD GWAS to date [56] identified 276 genes with significant gene-trait associations. To further pinpoint causal genes, the authors used colocalization to demonstrate that 142 of the 276 showed strong evidence for colocalization using GTEx data. Of the 142 genes, many were well-known regulators of BMD. Another study utilized 48 GTEx tissues and reported 88 significant genes, many of which were located in total body (TB) BMD GWAS [57]. Lastly, a recently published resource, PhenomeXcan [34], integrated TWAS gene-trait associations with colocalization to prioritize GWAS loci. A total of 675 genes were identified with both significant TWAS associations with BMD and colocalizing eQTLs.

Network Analysis

As mentioned above, GWASs have identified thousands of associations for BMD; however, the scarcity of population scale human RNA-seq datasets on bone or bone cells has hindered our ability to directly inform BMD GWAS. To address this limitation, it has recently been demonstrated that network approaches using transcriptomic (and in some cases other “-omic” data types) data can be used to provide information on which genes at a GWAS locus might be causal [35, 36]. The general idea is simple; genes responsible for GWAS associations likely function in pathways that impact bone, such as osteoblast-mediated bone formation or osteoclast-mediated bone resorption. In turn, biological network reconstruction approaches can take molecular data and group genes, in an unbiased way, into groups (or pathways) based loosely on function (Table 1). As a result, it is possible to use networks to identify “network modules” that are enriched in GWAS genes and likely represent key pathways or biological processes regulating BMD. In other words, biological networks provide us with cellular wiring diagrams and GWAS points to “circuits” that when disrupted (by genetic variation) lead to disease. We can then use the knowledge of key circuits to inform GWAS.

In a series of studies, our lab has used co-expression networks to predict causal BMD GWAS genes. In a co-expression network, genes are connected based on the correlation of their expression [58, 59]. Groups or “modules” of highly intercorrelated genes are identified by clustering [58, 59], and modules have been shown in a number of studies across species to loosely group genes based on functional similarities [60, 61]. Calabrese et al. [36] generated a co-expression network using transcriptomic data on marrow-free cortical bone from the Hybrid Mouse Diversity Panel (HMDP; a panel of 96 inbred mouse strains) [62]. We used mouse data to profile “pure” cortical bone since these data were not available in humans. To identify network modules that represented key “circuits,” we mapped the mouse homologues of genes implicated by (i.e., located in a GWAS locus) the largest BMD GWAS at the time [16] onto the mouse bone network. We identified two modules with a significant enrichment of GWAS genes, collectively named the Osteoblast Functional Module (OFM). Based on a detailed characterization, we hypothesized that many of the OFM genes were causal GWAS genes and that they likely influenced BMD via a role in modulating osteoblast activity. The OFM allowed us to predict and infer the function of causal genes for 30 of 64 reported BMD GWAS loci. We further investigated two BMD loci on chromosomes 2p16.2 and 14q32.32. Based on the network analysis, we predicted that Sptbn1 and Mark3 were responsible for the effects of the two loci, respectively. In support of these predictions, we used GTEx to identify that both genes were regulated by eQTLs in multiple tissues that colocalized with their respective GWAS associations. We also showed for both genes that BMD was altered in mouse knockout models in the same direction predicted by the GWAS/eQTL data.

There have been two follow-up studies [35, 37] that have refined the above approach. Sabik et al. [35] generated a cell type–specific (osteoblast) and time point-specific (mineralization) co-expression network using RNA-seq data on calvarial osteoblasts from a separate panel of inbred mouse strains. We identified a co-expression module enriched for genes implicated by BMD GWAS, correlated with in vitro osteoblast mineralization, and associated with skeletal phenotypes in human monogenic disease and mouse knockouts. We further investigated four loci and found that Cadm1, B4galnt3, Dock9, and Gpr133 all had human colocalizing eQTL and altered BMD in knockout mice.

Our next refinement was the use of Bayesian networks. Bayesian networks differ from co-expression networks in that they use advanced statistical approaches to add directions to the network that allow one to infer causal relationships between genes [63]. Al-Barghouthi et al. [37] demonstrated the utility of this approach by generating a co-expression network from cortical bone RNA-seq data collected from outbred mice. We then created Bayesian networks for each co-expression module and performed what is called a “key driver analysis” (KDA) [37, 64]. In a KDA, genes that are known to play important roles in bone (such as RUNX2, a transcription factor essential for osteoblastogenesis) are “seeded” onto the network. For each gene in the network, we then counted the number of “known” genes it was connected to and determined if this number was more than would be expected by chance. We then identified key drivers which were located in a GWAS locus and regulated by a colocalizing eQTL. This approach yielded 46 genes likely to be causal for human BMD GWAS associations. We further investigated two novel genes, Sertad4 and Glt8d2, and demonstrated that BMD was altered in knockouts, further suggesting they were causal for their respective GWAS association. These data supported the idea that Bayesian networks provided a new perspective and approach to identify causal BMD GWAS genes.

Another study informed GWAS using a co-expression network generated on macrophages from a cross between Wistar Kyoto (WKY) and LEW rats [65]. MMnet (the macrophage multinucleation network) was a module in this network. Importantly, macrophages can fuse to form osteoclasts in bone, and WKY rats experience spontaneous macrophage fusing events, while LEW rats do not [66]. MMnet contained 190 genes regulated by trans eQTLs that were driven by the Trem2 gene [66]. The authors reported that WKY rats show low bone mass, mineralization, and strength relative to LEW rats, suggesting that MMnet genes may control bone homeostasis. The authors show that the MMnet was enriched for genes located in BMD GWAS loci [65]. Given the evidence, the authors conducted in vivo knockout experiments of Bcat1, the most highly connected gene in the network, and showed that Bcat1 deficiency results in high bone mass. Next, readily available knockout mice of 12 MMnet individual genes were obtained, half of which showed skeletal phenotype abnormalities. As the authors established a strong association between MMnet and osteoclast activity, they investigated whether the human orthologues share the same activity as their rat counterparts. Knockdown of 11 MMnet genes (three of which are GWAS hits) in vitro (TRAP+ human osteoclasts) and knockout of the same genes in vivo (mouse models) were concordant and strongly correlated. This study identified a physiologically important network that is highly conserved in rats, mice, and humans and was enriched in GWAS-implicated genes. This is a great representation of how network analysis can inform GWAS and provide important information on the function (in this case osteoclast multinucleation) of potentially causal genes.

Epigenomics

Another “-omics” data type that has proven useful for informing GWASs is epigenetic data [67]. The majority (>90%) of BMD GWAS loci are found within noncoding regions [17], indicating that they perturb gene regulation. Thus, it is likely that causal GWAS variants reside in regulatory elements, such as promoters, enhancers, and CpG islands, which can be identified using epigenomic data.

In the same way that eQTL can be identified for gene expression, QTL can be identified using epigenomic data [68]. Examples are chromatin accessibility QTLs (caQTLs), which are loci that influence chromatin accessibility [69] (Fig. 1). Chromatin accessibility is a measure of the usage and activity of regulatory elements and is often assayed using ATAC-seq [70]. caQTLs can highlight potentially causal SNPs that may be driving genetically regulated changes in expression. In this way, they are most useful for identifying causal variants, not causal genes. However, they can also be integrated with eQTL information to link caQTLs to the genes they regulate, increasing confidence for a particular set of putatively causal genes [71].

To our knowledge, there are no studies using caQTLs to inform BMD GWAS. However, in T2D, a study profiled chromatin accessibility in pancreatic islet samples from 19 genotyped individuals and identified 2949 caQTLs [72]. The authors performed a functional follow-up on 13 of the reported caQTLs using luciferase reporter assays in MIN6 β-cells and showed that more than half exhibited effects on enhancer activity that were consistent with in vivo chromatin accessibility changes. Importantly, islet caQTL analysis nominated putative causal SNPs in 13 T2D-associated GWAS loci, linking seven and six T2D risk alleles, respectively, to gain or loss of in vivo chromatin accessibility.

Lastly, accumulating evidence suggests that genetic variants may impact a complex disease by modulating DNA methylation levels (meQTLs) (Fig. 1). To date, no meQTL analysis has been performed on bone samples. However, a study in depression cohorts has implicated gene targets by testing associations between SNPs and DNA methylation levels in whole blood [73]. Another study used meQTL to inform GWAS of asthma in exaggerated bronchoconstriction of airway smooth muscle cells (ASMCs) and showed that GWAS variants in asthma were significantly enriched for meQTLs [74].

Integrating “-omics” Data Types

So far, we have highlighted the use of single “-omic” data types. However, the use of multidimensional datasets (layering of various datasets) increases statistical power [75], potentially provides stronger evidence for causality, and captures biology that would not have been informed with any one modality. In a recent study, Qiu et al. [76] performed multi-“omics” analyses with expression, methylation, and metabolite QTLs to identify osteoporosis biomarkers. Their approach involved performing individual transcriptomic, methylomic, and metabolomic analysis in 119 European female subjects with high (n = 61) and low (n = 58) BMD. Using advanced statistical approaches, the authors identified gene-based biomarkers, some of which corresponded to genes located in BMD GWAS loci, suggesting they are causal.

Recently, Chesi et al. [77] took a different approach. Instead of generating population-level “-omics” data, the authors broadly profiled multiple “-omics” layers in human mesenchymal stem cell (MSC)–derived osteoblasts. Their approach was anchored on promoter Capture-C [78]. Promoter Capture-C is a technique that uses promoter “baits” to “fish-out” interactions between promoters and the rest of the genome. These data provide links between regulatory elements (e.g., enhancers) and the promoters of their target genes that are presumably important for gene expression. In osteoblasts, they identified interactions that were in close proximity of BMD-associated variants identified by GWAS. They then used RNA-seq data to confirm target gene expression in osteoblasts as well as ATAC-seq data to confirm the region interacting with the promoter was a region of open chromatin and presumably an active regulatory element. They fine-mapped 273 BMD GWAS loci in primary osteoblasts. The authors report observing consistent contacts between candidate causal variants and putative target gene promoters in open chromatin for ~ 17% of the 273 BMD loci investigated. Knockdown of two novel implicated genes, ING3 and EPDR1, inhibited osteoblastogenesis, while promoting adipogenesis.

Computational Prediction of Effector Genes at GWAS Loci

The advent of machine learning approaches to tackle biological problems has demonstrated potential in prioritizing GWAS genes. Evidence of this phenomenon is the proliferation of algorithms predicting effector genes at GWAS loci. Here, we highlight two recently published algorithms: Polygenic Priority Score (PoPS) [79] and Effector Index (Ei) [80].

PoPS is an ensemble approach that leverages GWAS data and gene expression, biological pathway, and predicted protein-protein interaction data to computationally prioritize GWAS-implicated genes. For BMD, PoPS correctly prioritized the gene CPE, which is regulated by the variant rs1550270. This variant co-localized with an eQTL for CPE in osteoclasts [46] and was predicted to have a role in regulating CPE as well as four other nearby genes. The knockout of Cpe in mice resulted in low bone mineral density and showed indications of increased bone turnover [81].

Ei aims to assign a probability of causality for each gene at a GWAS locus. The authors defined a set of “positive control genes” (68 genes for BMD, including 14 putative drug targets, and 54 genes whose perturbations directly affect BMD). Next, they assessed the genomic annotations at GWAS loci using the latest GWAS from the UK Biobank that was enriched for said “positive control genes.” Then, to predict the probability of causality for each gene at a GWAS locus, the authors used features, such as the number of genes at a locus and distance of a gene to the nearest associated SNP. The authors trained the model on 11 diseases including BMD data and applied the model on one disease (T2D). The model performance showed a predictive value of ~80%. The overall goal of Ei is to provide a rapid and readily understandable way to assess the probability that a gene at a GWAS locus is causal for a disease or trait.

Using Zebrafish to Inform BMD GWAS

In recent years, zebrafish have become a popular model for functional genomics experiments for a wide range of diseases including osteoporosis [82]. An example of this work is a study by Xiao et al. [83] who sought to validate a GWAS association implicating the MPP7 gene. Through the partial knockdown of mpp7, they showed that bone mineralization was affected and that MPP7 was a potential causal gene. Another group [84] followed up on an association identified by GWAS for adolescent idiopathic scoliosis (AIS) and found that overexpression of bnc2 (highly expressed in bone) in zebrafish embryos influenced spinal curvature. In conclusion, there is a strong rationale supporting the use of zebrafish for human GWAS follow-up because of the large number of conserved genetic and phenotypic features (reviewed elsewhere [85]).

Future Directions

In this review, we discuss the many ways that “-omics” data can be used to identify genes responsible for the effects of BMD GWAS loci. One limitation, as discussed above, is the paucity of “-omics” data on bone tissue and bone cells. While the number of such studies is growing, there is a need to generate population-scale transcriptomics data on the three main cell types in bone: osteoblasts, osteoclasts, and osteocytes. These data would significantly increase our ability to identify and characterize causal genes responsible for BMD GWAS associations. They would also be of significant use to the larger bone and human genetics communities to address other aspects of disease.

Another exciting approach that will impact our ability to use gene expression data to inform GWAS is single-cell RNA sequencing (scRNA-seq). scRNA-seq is emerging as a powerful tool to examine transcriptomes of individual cells. The clear benefit of this technology is its ability, when used in populations, to identify cell type–specific eQTLs, many of which are lost when using bulk methods that take into account average gene expression across all different cell subtypes [86]. In early 2020, the single-cell eQTLGen consortium (sceQTLGen) was founded, aimed at pinpointing the disease-causing genetic variants and their effect on gene expression [86]. The single-cell toolbox can be extended beyond the transcriptome to the epigenome with single-cell ATAC-seq (scATAC-seq). A recent study by Rai et al. [87] attempted to identify cell type–specific regulatory signatures underlying T2D in pancreatic islets. They reported that T2D GWAS SNPs were significantly enriched in the open chromatin of beta cells, but not in alpha or delta cell–specific open chromatin. In the bone field, scRNA-seq and scATAC-seq are just beginning to be performed, but have already demonstrated the extensive cellular heterogeneity of bone [88, 89]. Using post-GWAS approaches described by this review on data generated by these two approaches will lead to an increase in our ability to inform GWAS.

Proteomic analysis offers another type of data that could be integrated into systems genetics approaches. Protein quantification studies have shown that transcript abundance is not highly correlated with protein translation [90, 91]. Generally, proteomics technologies used in publications to study bone metabolism can be inherently divided into two categories: (i) expression screening and (ii) quantitative mass spectroscopy (MS) [91, 92]. The main challenge facing proteomic work in bone is the efficient extraction of proteins from bone cells [93]. Therefore, most studies have instead used blood serum/plasma or PBMCs to study cellular signaling, secretory proteins, and differential protein expression between conditions [92, 94]. However, the goal of GWAS follow-up at the proteome level is identifying genetic variants associated with protein concentrations (protein quantitative trait loci; proQTL) [95] (Fig. 1). Integrating proQTL with GWAS variants using approaches such as colocalization may inform GWAS beyond what can be accomplished with transcriptomics data and bridge the knowledge gap regarding SNP-disease associations [95]. For example, using GWAS data from Framingham participants (long-term cardiovascular study cohort) reported 13 proteins harboring proQTL variants that match coronary disease-risk variants from GWAS, not all of which also had colocalizing eQTLs [96].

It should be noted that all the approaches discussed in this review provide hypotheses that must be tested in order to confirm gene discovery. It is impossible to confirm these hypotheses in humans, thus highlighting the importance of using model organisms such as mice. Available resources such as the International Mouse Phenotyping Consortium (IMPC) [97] and Knockout Mouse Project (KOMP) [98] which aim to identify the function of every protein-coding gene in the mouse genome, the Origins of Bone and Cartilage Disease (OBCD) initiative [99, 100], and Bonebase project [100] that are providing detailed bone phenotyping of mice from these efforts are key components of this step. Additionally, in vitro (using human bone cell lines or human primary cells) or in vivo (using model organisms) testing of genes using gene editing technologies such as CRISPR/Cas9 [101] or its variations (CRISPRi, CRISPRa, etc. [102]) will be key in uncovering the biology identified by GWAS. Finally, the use of genome-scale CRISPR/Cas9 functional genetic screens could potentially uncover other genes beyond those directly implicated by GWAS.

Conclusions

Over the past 12 years, GWASs have identified a large number of variants influencing BMD. Post-GWAS efforts are attempting to identify the genes responsible for their effects. We now know that most of the variants identified by GWAS exert their impact on bone by altering gene regulation. We also know that the discovery of causal genes has the potential to provide insight into osteoporosis etiology and identify novel therapeutic targets. The approaches and findings highlighted in this review will only continue to improve given rapid advances in statistical approaches and technologies to profile molecular phenotypes. As the field progresses and we continue to unlock the secrets of the human genome, it is our hope that we will be able to use this information to develop more effective therapies to treat and ultimately prevent osteoporosis.