Integration of summary data from GWAS and eQTL studies identified novel causal BMD genes with functional predictions

Abstract

Purpose:

Osteoporosis is a common global health problem characterized by low bone mineral density (BMD) and increased risk of fracture. Genome-wide association studies (GWAS) have identified > 100 genetic loci associated with BMD. However, the functional genes responsible for most associations remain largely unknown. We conducted an innovative summary statistic data-based Mendelian randomization (SMR) analysis to identify novel causal genes associated with BMD and explored their potential functional significance.

Methods:

After quality control of the largest GWAS meta-analysis data of BMD and the largest expression quantitative trait loci (eQTL) meta-analysis data from peripheral blood samples, 5967 genes were tested using the SMR method. Another eQTL data was used to verify the results. Next we performed a fine-mapping association analysis to investigate the functional SNP in the identified loci. Weighted gene co-expression network analysis (WGCNA) was used to explore functional relationships for the identified novel genes with known putative osteoporosis genes. Further, we assessed functions of the identified genes through in vitro cellular study or previous functional studies.

Results:

We identified two potentially causal genes (ASB16-AS1 and SYN2) associated with BMD. SYN2 was a novel osteoporosis candidate gene and ASB16-AS1 locus was known to be associated with BMD but was not the nearest gene to the top GWAS SNP. Fine-mapping association analysis showed that rs184478 and rs795000 was predicted to be possible causal SNPs in ASB16-AS1 and SYN2, respectively. ASB16-AS1 co-expressed with several known putative osteoporosis risk genes. In vitro cellular study showed that over-expressed ASB16-AS1 increased the expression of osteoblastogenesis related genes (BMP2 and ALPL), indicating its functional significance.

Conclusion:

Our findings support that ASB16-AS1 and SYN2 may represent two novel functional genes underlying BMD variation. The findings provide a basis for further functional mechanistic studies.

1. Introduction

The widespread application of genome-wide association studies (GWAS) has contributed to a revolution in the research of human genetics and the genetic determinants that underlie complex disease. A recent study suggested that selecting genetically supported therapeutic targets implicated by GWAS could potentially double the success rate in the clinical development of new pharmaceutical treatments [1]. While GWAS have identified thousands of genetic variants associated with human complex traits and diseases, there are still several important limitations to these prevailing types of genetic association studies [2,3]. A longstanding challenge of GWAS lies in exploring the mechanisms by which these genetic loci affect the variation of complex traits and the pathophysiology of complex diseases.

Osteoporosis is a common skeletal disease characterized by reduced bone mineral density (BMD) and increased risk of low trauma fractures [4]. In the United States, it has been estimated that the prevalence rate of osteoporosis in older adults was about 10.3%, while low bone mass prevalence rate was 43.9% [5]. In China, the prevalence rate of osteoporosis in older adults was estimated to be 15.7%, and it will increase rapidly with the increasing age of the total population [6]. Despite the significant impact on human health, there is still a lack of highly effective osteoporosis treatments that are free of negative side effects [7]. Hence, identification of additional therapeutic molecular targets for effective and efficient prevention and treatment of osteoporosis are needed. Heritability for BMD is estimated to be > 50% [8,9] and more than one hundred loci have been found to be associated with BMD or osteoporosis by previous studies [10–16]. However, although many associated genetic variants have been identified by GWAS, the causal variants truly with biological effects have remained largely unknown. Furthermore, only ~10% of the total BMD heritability has been explained by the current GWAS findings [14]. Additional genes and biological mechanisms underlying osteoporosis could be identified from existing GWAS data by using novel biostatistic and bioinformatic methods [17].

A genetic variant that influences a particular gene-expression level is known as expression quantitative trait loci (eQTL). Two studies using eQTL analysis in primary bone cell cultures discovered numerous eQTLs associated with BMD [18,19]. Kwan et al. found an eQTL (rs136564) regulated the expression of a novel transcript of FAM118A, and this eQTL was also found to be associated with BMD by GWAS analysis [19]. There are many other studies that have attempted to assess whether a SNP detected by GWAS was also an eQTL [20]; however most of them have analyzed the GWAS data and eQTL data in two sequential and separate steps rather than in an integrative manner [21]. Zhu et al. recently proposed an integrative method called summary data–based Mendelian randomization (SMR) that integrates independent GWAS summary statistics data with eQTL data from whole blood tissue to identify potential functionally relevant genes at the loci identified in GWAS and to identify novel trait-associated genes for five complex traits [22], and further applied the method to the analyses of another 28 traits [23]. It has been shown that eQTL effect in blood can be a proxy for eQTL effect in most relevant tissues for various traits or diseases [22,24]. Mendelian randomization is an instrumental variable analysis approach that uses genetic variants as instrumental variables (for example, eQTLs) to test whether an exposure (for example, the expression level of a gene) has a causal effect on an outcome (for example, trait value or disease risk) [25,26]. In this study, the phenotypic trait is the outcome (Y), gene expression is the exposure (X), and the top cis-eQTL that is strongly associated with gene expression is used as the instrumental variable (Z). The previous study [22] suggested that there are three models consistent with a significant association from the SMR test using only a single genetic variant. These three models include causality (Z → X → Y), pleiotropy (Z → X and Z → Y) and linkage (Z1 → X, Z2 → Y, and Z1 and Z2 are two variants in linkage disequilibrium (LD) in the cis-eQTL region). The purpose of this study was to identify genes whose expression levels had causal effect (Z → X → Y) on BMD. In Mendelian randomization studies, multiple uncorrelated instrumental variables (IVs) [22] (for example, the uncorrelated trans-eQTLs and/or cis-eQTLs) or multiple correlated IVs [27] were needed to identify the causality. Due to multiple uncorrelated IVs were not available in the Westra eQTL study [24], we applied weighted generalized linear regression method using correlated IVs to distinguish causality from pleiotropy.

It is clear that complex bone traits, like BMD, are not only the results of cumulative effects of individual genetic factor, but also the effects from interactive biological networks [28]. The weighted gene co-expression network analysis (WGCNA) is a useful method to construct the co-expression network [29]. Co-expression networks are modular, and each module represents a group of co-expressed genes. These modules tend to contain genes involved in similar biological processes [7]. Hence, under WGCNA, we explored the functional relevance of the identified novel genes to other known putative osteoporosis genes in order to explore the potential functional mechanisms of the identified novel genes.

In this study, we will identify genes whose expression levels have causal effects on BMD by using the SMR method and the weighted generalized linear regression method, followed by assessing functions of the identified genes through in vitro cellular study or previous functional studies.

2. Materials and methods

2.1. Data used in this study

2.1.1. GWAS summary data

The GEnetic Factors for OSteoporosis (GEFOS) Consortium used meta-analysis of whole genome sequencing, whole exome sequencing and deep imputation of genotype data (in reference to the UK10K and 1000Genomes data) to identify low-frequency and rare variants associated with risk of osteoporosis in 53,236 Caucasians [16]. The detailed description about genotype and imputation of GWAS data can be found in the previous study [16]. Each SNP with a minor allele frequency (MAF) > 0.5% was tested for association with an additive effect on femoral neck (FN), lumbar spine (LS) and forearm (FA) BMD, adjusting for sex, age, age² and weight [16]. The summary statistic data are available online (http://www.gefos.org/).

2.1.2. eQTL summary data

Westra et al. performed the largest eQTL meta-analysis so far in non-transformed peripheral blood samples of 5311 European healthy individuals with replication in 2775 European individuals [24]. Another eQTL study [30] which were performed to investigate the genetic architecture of gene expression (GAGE) in peripheral blood in 2765 European individuals was used to verify the results. It is widely accepted that eQTL effect in blood tissue can be a proxy for eQTL effects in most relevant tissues for various traits or diseases [22,24]. Particularly, there are several types of cells such as peripheral blood monocytes (PBMs) and B and T lymphocytes that are related to bone metabolism [31]. For example, PBMs have been well established as a working cell model for studying gene expression patterns in relation to osteoporosis risk in vivo in humans [32]. PBMs may act as precursors of osteoclasts since they can differentiate into osteoclasts [33], and they express different cytokines which are important for osteoclast differentiation, activation, and apoptosis [32,34]. B lymphocytes, an important cell type of the immune system, express/secrete factors involved in osteoclastogenesis, such as receptor tumor necrosis factor superfamily member 11 and osteoprotegerin [35]. The gene expression data were quantile-normalized to the median distribution, and subsequently log₂ transformed [24]. The eQTL summary data in SMR binary format can be downloaded from http://cnsgenomics.com/software/smr/download.html.

2.1.3. Gene expression data

We used previously published gene expression profile generated from PBMs [36] to generate gene co-expression networks to assess the potential interactions and thus potential mechanisms for the identified novel gene. PBMs were from 73 Caucasians females, which were stratified by hip BMD and menopausal status. In the high BMD group there were 42 subjects with 16 premenopausal and 26 postmenopausal women. In the low BMD group there were 31 subjects with 15 premenopausal and 16 postmenopausal women. Details of the samples’ information can be found in the previous study [32]. We downloaded the raw data under the accession number GSE56814 from Gene Expression Omnibus (GEO) website.

2.2. Statistical analysis methods

2.2.1. Summary data–based Mendelian randomization (SMR) analysis

The SMR method was detailed in the previous paper [22]. In brief, SMR applies the principles of Mendelian randomization (MR) [37,38] to jointly analyze GWAS and eQTL summary statistics in order to test for association between gene expression and a trait due to a shared variant at a locus. In this study, the phenotypic trait is the outcome (Y), gene expression is the exposure (X), and the top cis-eQTL that is strongly associated with gene expression is used as the instrumental variable (Z). Equivalently, it is an analysis to test whether the effect of Z on Y is mediated by X (a model of Z → X → Y). The SMR software was downloaded from http://cnsgenomics.com/software/smr/.

Since the SMR analysis assumes that the instrument (top cis-eQTL) has a strong effect on the exposure (gene’s expression level), only probes with at least one cis-eQTL at P_eQTL (a p value from the eQTL study indicating the significance of the eQTL associated with the gene expression) smaller than 5 × 10⁻⁸ in the cis-eQTL region were included in the eQTL summary data (hg19). We excluded cis-eQTL with MAF < 0.01 and cis-eQTL in the MHC region because of the complexity of LD patterns in this region [22]. After data processing, there were 5967 probes left. So the genome-wide significance level for SMR test was P_smr < 8.4 × 10⁻⁶ (0.05/5967, Bonferroni Correction).

Since the significant SMR results could also reflect linkage model that was of less biological interest, we used the heterogeneity in dependent instruments (HEIDI) test to distinguish pleiotropy (or causality) model from linkage model. The HEIDI test considers the pattern of associations using all the SNPs that are significantly associated with gene expression (eQTLs) in the cis-eQTL region (± 250 kb from the center of the gene probe). Under Hardy-Weinberg equilibrium and LD, b_XY estimated at the top associated cis-eQTL (b_XY(top)) will be equal to that estimated at any of the cis-SNPs in LD that is associated with gene expression. In the HEIDI test, we excluded the SNPs in strong LD with top cis-eQTL at r² > 0.9 which was calculated by using individual level data from the HapMap2 CEU. We also removed SNPs in the cis-eQTL region with a P_eQTL > 1.6 × 10⁻³ (equivalent to χ² < 10) to avoid weak instrumental variables according to the original paper [22]. The null hypothesis of the HEIDI test was that there was a single causal variant which meant that there was no heterogeneity in the b_XY values. A p value threshold of P_HEIDI > 0.05 was conservative for a gene having “pleiotropy” effects (no heterogeneity) [22].

Next, causal estimate was conducted by applying weighted generalized linear regression method using correlated IVs [27] to distinguish causality model from pleiotropy model. First, we assume the estimate of association for IV k = 1, …, K with the gene’s expression (X) is β_kX with standard error σ_kX. The estimate of association for IV k with BMD can be expressed as β_kY with standard error σ_kY. The correlation between IVs k₁ and k₂ can be defined as ρ_k1k2. Then we can perform a weighted generalized linear regression of the β_kY parameters on the β_kX parameters using the σ_kY⁻² parameters as inverse-variance weights and taking into account the correlation between the IVs. If we define Ω_k1k2 = σ_k1Yσ_k2Yρ_k1k2, b_XY can be estimated from a weighted generalized linear regression as b_XY = (β_kX^TΩβ_kX)⁻¹β_kX^T Ω⁻¹β_kY [39], withβ_kX^T being the transposed vector of β_kX. The standard error of the estimate is $\text{se}\left({\text{b}}_{\text{XY}}\right)=\sqrt{{\left({\mathrm{\beta }}_{\text{kX}}^{\text{T}}{\mathrm{\Omega }}^{-1}{\mathrm{\beta }}_{\text{kX}}\right)}^{-1}}$ [27]. After pruning for SNPs with linkage disequilibrium r² > 0.8, all variants with a P_eQTL < 5 × 10⁻⁸ in the cis-eQTL region of each probe available in both BMD and eQTL data were included in the analysis. The correlation between IVs was calculated in Plink software (version 1.0.7) by using HapMap 2 CEU as a reference panel. Cochran’s Q statistic was performed to test heterogeneity using the meta R packages [27], a p value > 0.05 can be interpreted as no more heterogeneity between causal effects estimated using the variants individually than would be expected by chance.

2.2.2. The regional association plot

SNAP (for SNP Annotation and Proxy Search, https://archive.broadinstitute.org/mpg/snap/ldsearch.php) was used to characterize the regional association plot for the interesting genes [40]. The LD between the putative causal SNP and other SNPs were calculated by using HapMap 2 CEU as a reference panel.

2.2.3. Fine-mapping

It is well documented that a SNP overlapping with a functional region is more likely to be a functional SNP [41]. To identify the causal SNPs in each identified gene, we used the annotation data from RegulomeDB [42], and performed a fine-mapping analysis called Probabilistic Identification of Causal SNPs (PICS) [43] to estimate the possibility of each SNP in the locus to be a causal SNP. By using information derived from the permutations, PICS score was calculated as the posterior probability of each SNP being the causal variant, given the observed pattern of association at the locus [43].

2.2.4. WGCNA

To identify functional connections for the identified novel genes with known genes, we performed WGCNA in gene expression profiles generated from PBMs. We used RMA (robust multiarray average) algorithm to correct for the background noise and normalize the expression data with the Expression Console software. The software can be downloaded from the Affymetrix website: http://www.affymetrix.com. The co-expression networks were generated using the WGCNA R package [44]. We extended gene boundaries by 20 kb upstream and downstream of the gene [45]. Any gene containing a SNP in the extended region with a p value < 0.01 for at least one of the two BMD trait (FN-BMD and LS-BMD) and 234 genes which were identified to be associated with BMD by previous studies [12,46,47] are here referred to as the nominally significant GWAS geneset (4796 genes). After excluding the non-expressed genes, we identified 3593 probes representing 3593 genes to construct the co-expression network. Finally, we exported the network of the interesting genes with Topology Overlap Matrix (TOM) value > 0.15. We visualized the networks by Cytoscape [48]. Due to that the TOM value could not exhibit the positive or negative correlation between co-expressed genes, we defined the signed expression similarity as the value of the correlation coefficient between the profiles of genes i and j: s_ij = cor(x_i,x_j). Person correlation coefficient was calculated by using R software.

2.3. Gene overexpression experiment for the identified ASB16-AS1

To validate the functional implication of the ASB16-AS1 gene in osteoblast, we performed gene overexpression experiment. Total cellular RNA was isolated from the human fetal osteoblastic 1.19 cell line (hFOB1.19, ATCC, Cat CRL-11372). Since the ASB16-AS1 transcription variant 2 uses an alternate splice site in the terminal exon compared to variant 1, the transcription variant 2 is shorter than variant 1. The same primers were used to amplify both the variants, only variant 2 was successfully subcloned into the pCEP4 vector. The primers for amplification of ASB16-AS1 were 5′-ggggtaccgtggcttcgcgactgcggaaggt-3′ (forward KpnI) and 5′- cgggatcctttttttttttttttttttatttttttttggtgcatactgtttaattttct-3′ (reverse BamHI). The procedure of cell culture was described previously [49]. Cells were seeded at 6 × 10⁵ cells/well in 6-well plates. After 24 h, cells were transfected with Lipofectamine 3000, expression vector pCEP4-ASB16-AS1 using pCEP4 as a control. Real-time Polymerase Chain Reaction (RT-PCR) was used to detect the relative mRNA level of osteoblastic genes (BMP2 and ALPL) 48 h later after transfection. The primers for RT-PCR were shown in Table S1. Student’s t-test was applied to determine statistical significance.

3. Results

3.1. SMR tests

SMR tests identified two genes (ASB16-AS1 and TMUB2) that were significantly associated with FN-BMD, two genes (SYN2 and LRP3) were associated with LS-BMD and no genes were identified to be associated with FA-BMD (Table 1). Of the four genes, only ASB16-AS1 and SYN2 passed HEIDI tests (P_HEIDI > 0.05), suggesting that there was no heterogeneity and gene expression of ASB16-AS1 (or SYN2) and the BMD variation were affected by the same variant. Using the GAGE eQTL data, we re-conducted the SMR tests. ASB16-AS1 was successfully verified to be associated with FN-BMD with P_SMR = 1.98 × 10⁻⁸ and P_HEIDI = 0.13. SYN2 was successfully verified to be associated with LS-BMD with P_SMR = 2.51 × 10⁻⁵ and P_HEIDI = 0.31. In the cis-eQTL region of SYN2 probe, there was no SNP with a GWAS p value < 5 × 10⁻⁸ (Fig. 1B). Therefore, SYN2 was a new gene associated with BMD discovered by the SMR test. The details of the results were shown in Table 1. The genes TMUB2 and LRP3 failed to pass the HEIDI tests suggesting there were two genetic variants, one affecting the gene expression and the other affecting the BMD variation. Subsequently, our study only focused on the two genes (ASB16-AS1 and SYN2) with potential functionally relevant variants in the downstream further analysis.

Table 1.

Four genes identified by the SMR analyses.

Trait	Probe ID	Chr	Gene	cis-eQTL	PGWAS	PeQTL	bXY	PSMR	PHEIDI	PICS
FN-BMD	ILMN_1676731	17	ASB16-AS1	rs227580	7.06 × 10−8	4.66 × 10−75	−0.13	1.30 × 10−7	0.11	0.055
FN-BMD	ILMN_2311779	17	TMUB2	rs228754	4.20 × 10−6	1.03 × 10−143	−0.077	3.70 × 10−6	0.007	NA
LS-BMD	ILMN_1781060	3	SYN2	rs310758	8.73 × 10−7	8.18 × 10−37	−0.18	2.83 × 10−6	0.34	0.035
LS-BMD	ILMN_2127605	19	LRP3	rs11084710	6.99 × 10−6	3.23 × 10−82	0.11	7.55 × 10−6	0.011	NA
FN-BMD	ILMN_1676731	17	ASB16-AS1	rs184478	9.34 × 10−9	8.37 × 10−72	−0.13	2.39 × 10−8	0.11	0.99
LS-BMD	ILMN_1781060	3	SYN2	rs795000	3.58 × 10−7	2.05 × 10−36	−0.28	1.44 × 10−6	0.34	1

Note: The first four lines were calculated with top cis-eQTLs which were the top associated cis-eQTL of the probes in the Westra et al. eQTL data. The last two lines were calculated with the possible causal SNPs in the region.

b_XY was the gene expression effect on BMD.

PICS was Probabilistic Identification of Causal SNPs.

Fig. 1.

The SMR results at gene loci for BMD. (A) The SMR result at ASB16-AS1 locus for FN-BMD. (B) The SMR result at SYN2 locus for LS-BMD. In the top plot, black dots represent the p values for the SNPs from the latest GWAS meta-analysis for BMD (Y-axis), diamonds represent the p values for probes from the SMR test. In the bottom plot, the eQTL p values of the SNPs were from the eQTL study (Y-axis) for the ILMN_1676731 probe (or ILMN_1781060 probe) tagging ASB16-AS1 (or SYN2).

In the cis-eQTL region of ASB16-AS1 probe, there were 73 SNPs with a P_eQTL < 5 × 10⁻⁸. After pruning for SNPs with LD r² > 0.8, a total of 16 SNPs in the cis-eQTL region of ASB16-AS1 probe were included to detect whether the expression of ASB16-AS1 had a causal effect on BMD. The estimated effect size of the expression of ASB16-AS1 on BMD (b_XY) was −0.14 (95% CI: −0.18, −0.10). The p value from Cochran’s Q statistic of heterogeneity in causal estimates for each genetic variant calculated individually was 0.45, suggesting that no heterogeneity existed. There were 160 eQTLs with a p value < 5 × 10⁻⁸, in the cis-eQTL region of SYN2 probe. After pruning, 15 SNPs were included to give the causal estimate with a p value from Cochran’s Q statistic 0.16. The estimated effect size of the expression of SYN2 on BMD (b_XY) was −0.20 (95% CI: −0.27, −0.13). Association results of ASB16-AS1 or SYN2 with BMD were displayed graphically in Fig. 2.

Fig. 2.

Estimated genetic associations and 95% confidence intervals with effect sizes in eQTL and GWAS studies for 16 genetic variants in the ASB16-AS1 gene region (A) and 15 genetic variants in the SYN2 gene region (B).

3.2. Fine-mapping

With the functional annotation data from RegulomeDB, in the ASB16-AS1 locus we found that rs184478, which was in perfect LD with rs227580 (Fig. 3), was an eQTL overlapping with many functional regions such as conserved motifs, transcription factor binding site, DNase I hypersensitive peak, Chip-Seq peaks and DNase I footprint region, while the top cis-eQTL rs227580 only overlapped with transcription factor binding site and DNase I hypersensitive peak. The PICS score of rs184478 was 0.987, while the PICS score of rs2275580 was only 0.055. When using rs184478 instead to perform the SMR test, we found the P_smr value (P_smr = 2.39 × 10⁻⁸) was smaller (thus more significant) than using the top cis-eQTL (P_smr = 1.30 × 10⁻⁷). In the SYN2 locus, rs795000 achieved the highest PICS score. When using rs795000 instead to perform the SMR test, the result was more significant than using the top cis-eQTL rs310758 (Table 1). Therefore, rs184478 and rs795000 had the largest possibility to be causal SNPs for ASB16-AS1 and SYN2 genes, respectively.

Fig. 3.

Regional association plot for ASB16-AS1 on chromosome 17. SNPs which were in this region were selected with their p values from the GWAS data of FN-BMD. r² of pairwise LD is calculated between rs227580 and other SNPs. C17ORF65 is also known as ASB16-AS1.

3.3. WGCNA

The biological roles of ASB16-AS1 in bone metabolism are not well known. WGCNA is a useful system genomics method to construct gene co-expression network and predict functional connections (thus potential functional mechanisms) between novel genes and known genes. After excluding non-expressed genes, we selected 3593 probes representing 3593 genes (~75% of the nominally significant GWAS geneset which contains 4796 genes). After hierarchical clustering and dynamic tree cutting, we chose the gene modules which contained ASB16-AS1 gene. Using Cytoscape, we visualized the gene centered co-expression network, respectively (Fig. 4). The network only showed genes which were shown to be associated with BMD in the previous studies and connected to the centered gene with a TOM > 0.15. With this criterion, there were 45 known genes co-expressed with ASB16-AS1 (Fig. 4). CTNNB1 positively co-expressed with ASB16-AS1, while TNFSF11 (cor = −0.43) was negatively co-expressed with ASB16-AS1.

Fig. 4.

The gene co-expression relationships for ASB16-AS1. ASB16-AS1 centered network provides a view of all edges and their corresponding nodes connected to ASB16-AS1 with a TOM > 0.15. We only selected those nodes proved to be associated with BMD before. Genes are color coded based on their correlation with ASB16-AS1, white (cor > 0) and grey (cor < 0).

3.4. Over-expression experiment of ASB16-AS1

The RT-PCR results showed that the expression of ASB16-AS1 transcript variant 1 was higher than the expression of AB16-AS1 transcript variant 2 (Fig. 5A). However, using the same primers, we failed to get a clone of variant 1. The variant 2 was successfully subcloned into the pCEP4 vector. After transfection of pCEP4-ASB16-AS1, the expression of ASB16-AS1 were highly enhanced (Fig. 5B). The expression of BMP2 and ALPL was significantly increased in the ASB16-AS1 over-expressed osteoblast compared with the control (Fig. 5C and D). The expression of RUNX2 was increased but the alteration was not significant (Fig. 5E).

Fig. 5.

ASB16-AS1 can promote the expression of osteoblastic genes. (A) The relative expressions of two variants of ASB16-AS1 were detected in hFOB1.19 cells. The relative expressions of ASB16-AS1 (B), ALPL (C), BMP2 (D) and RUNX2 (E) were detected after transfection with the pCEP4-ASB16-AS1 using the pCEP4 as control. Bars represented S.D. *p < 0.05, **p < 0.01.

4. Discussion

In the present study, by using the summary data-based Mendelian randomization analysis, we identified ASB16-AS1 associated with FN-BMD and SYN2 associated with LS-BMD and successfully verified by using another eQTL study. Under the WGCNA analysis, we found that ASB16-AS1 was co-expressed with dozens of known putative osteoporosis genes including CTNNB1, TNFSF11 and RUNX2. Over-expressed ASB16-AS1 would increase the expression of osteoblastic genes (BMP2 and ALPL).

The causal estimate of b_XY for ASB16-AS1 was −0.14 which meant that a genetic increase in log₂ transformed expression of ASB16-AS1 by one standard deviation decreased BMD by 0.14 standard deviation. Similarly, the causal estimate of b_XY for SYN2 was −0.20 which meant that a genetic increase in log₂ transformed expression of SYN2 by one standard deviation decreased BMD by 0.20 standard deviation.

SMR was an efficient method to identify associations between gene expression and complex traits using summary data from GWAS and eQTL studies. One advantage of SMR was that it was useful to prioritize functionally relevant genes in a trait/disease associated locus. A major challenge of GWAS in interpreting which gene is functionally associated with the trait comes from the fact that the significant SNP represents a large region of LD. Regions of strong LD can be very large, and the significant SNPs have been found in perfect LD with the causal SNP hundreds of kilobases away [41]. The SMR analysis (including a SMR test and a HEIDI test) was useful to prioritize the functional gene in the associated region. The top associated GWAS SNP rs228769 was located in HDAC5 gene region, 60 kb downstream from ASB16-AS1 (Fig. 3). The gene HDAC5 may contribute to osteoporosis etiology by controlling sclerostin expression in osteocytes and regulating osteoblast differentiation [50]. Since there was no probe for the HDAC5 gene in both eQTL studies, the present study cannot exclude the causal role of HDAC5 gene for BMD variation. However, our study clearly demonstrated that ASB16-AS1 was a putative novel functional gene in this known locus.

Complex traits, like BMD, may be not only the results of cumulatively additive effects of individual genetic factors, but also the results of gene interaction via biological pathways/networks [28]. Under WGCNA, we found ASB16-AS1 co-expressed with more than forty genes which were proven to be associated with BMD by previous studies. The result thus suggested that ASB16-AS1 may play an important role in bone biology. In the ASB16-AS1 centered network, gene CTNNB1 encodes β-catenin protein which can aggregate and move into the nucleus and then can active wnt/β-catenin signaling pathway. The target gene of β-catenin protein, RUNX2, is active to promote osteoblast proliferation and differentiation [51–53]. Interestingly, these genes which mainly play roles in osteoblasts, were found to be co-expressed with ASB16-AS1 in PBMs which are precursors of osteoclasts [33]. The reason why these genes express in PBMs was worth of further exploration. TNFSF11 was negatively correlated with ASB16-AS1. RANKL encoded by TNFSF11 is a key factor for osteoclast differentiation and activation [54]. Taken together, we predicted that ASB16-AS1 may play important roles in osteoblast and osteoclast proliferation and differentiation. BMP2 has been shown to potently induce osteoblast differentiation [55]. ALPL, coding for the liver/bone/kidney isozyme of alkaline phosphatase, are known to be regulated during osteoblastic differentiation [56]. The expression of BMP2 and ALPL were increased after transfection of pCEP4-ASB16-AS1, suggesting that ASB16-AS1 may promote the differentiation of osteoblast.

SMR was demonstrated that it was useful to prioritize novel genes associated with BMD. There was no GWAS signal within 0.5 Mb of the probe in SYN2 (Fig. 1B). However, by the SMR tests, our study identified significant association signal for SYN2. SMR tests reduced the multiple hypothesis burdens by testing tens of thousands of genes instead of millions of SNPs [17]. The synapsins are a family of 4 synaptic vesicle-associated proteins which are products from alternative splicing of two genes, SYN1 and SYN2. The previous study which showed that the release of glutamate via synapsin can directly promote osteoblast differentiation [57], suggesting that SYN2 played an important role in osteoblast differentiation.

Additionally, there are several other methods for detecting the association between genes and traits using the GWAS and eQTL data [58–60]. The COLOC method [58] is a very useful tool to detect co-localization of GWAS and eQTL signals at known GWAS risk loci with a Bayesian analysis approach; however it does not provide thresholds for the posterior probabilities to control for genome wide false positive rate. The PrediXican method [59] directly tests the molecular mechanism through which genetic variation affects a study trait; however it requires individual-level genotype and gene expression data in the training data set, and individual-level genotype and phenotype data in the target data set. Therefore, this approach may currently have limited power and feasibility because of the limited availability and small sample sizes of such required data sets at present. Gusev et al. [60] proposed a method to overcome this problem by performing a polygenic prediction analysis using summary-level statistic data. However, unlike the SMR method, neither PrediXcan nor the Gusev’s method distinguishes between pleiotropy and linkage.

The present study may have some limitations. First, in the Westra et al. eQTL data [24], there were only 5967 probes so that it had incomplete and sparse genomic coverage of relevant genes, which may lead to the fact that only four positive findings using the SMR test. Some potential genes may thus be missed. Second, in both eQTL data, there is no probe for HDAC5. So SMR test did not yield any signal of HDAC5 but we cannot exclude the importance of the gene. Despite this, we found ASB16-AS1 was another putative novel functional gene for BMD at the locus. Third, the Gene Tissue Expression (GTEx) project which was an NIH funded effort planned to generate RNA-seq expression profiles from > 40 tissues in a large genotyped human cohort [61], did not collect bone tissue and primary bone cells. We thus cannot get eQTL data from bone related cells/tissue from GTEx for our analyses at present. Fourth, SMR may lose power compared to standard GWAS when the true biological mechanism is independent of gene expression [17].

In conclusion, the present study identified that ASB16-AS1 and SYN2 may causally affect BMD through their gene expression. ASB16-AS1 may play an important role in osteoblast proliferation and differentiation, at least by interacting with known genes in these processes.

Supplementary data to this article can be found online at https://doi.org/10.1016/j.bone.2018.05.012.

Abbreviations:

BMD

bone mineral density

GWAS

genome-wide association studies

eQTL

expression quantitative trait loci

SMR

summary data-based Mendelian randomization

HEIDI

heterogeneity in dependent instruments

LD

linkage disequilibrium

WGCNA

weighted gene co-expression network analysis

IVs

instrumental variables

PICS

Probabilistic Identification of Causal SNPs

TOM

Topology Overlap Matrix