Identification of Novel Potentially Pleiotropic Variants Associated With Osteoporosis and Obesity Using the cFDR Method

Abstract

Context

Genome-wide association studies (GWASs) have been successful in identifying loci associated with osteoporosis and obesity. However, the findings explain only a small fraction of the total genetic variance.

Objective

The aim of this study was to identify novel pleiotropic genes important in osteoporosis and obesity.

Design and Setting

A pleiotropic conditional false discovery rate method was applied to three independent GWAS summary statistics of femoral neck bone mineral density, body mass index, and waist-to-hip ratio. Next, differential expression analysis was performed for the potentially pleiotropic genes, and weighted genes coexpression network analysis (WGCNA) was conducted to identify functional connections between the suggested pleiotropic genes and known osteoporosis/obesity genes using transcriptomic expression data sets in osteoporosis/obesity-related cells.

Results

We identified seven potentially pleiotropic loci—rs3759579 (MARK3), rs2178950 (TRPS1), rs1473 (PUM1), rs9825174 (XXYLT1), rs2047937 (ZNF423), rs17277372 (DNM3), and rs335170 (PRDM6)—associated with osteoporosis and obesity. Of these loci, the PUM1 gene was differentially expressed in osteoporosis-related cells (B lymphocytes) and obesity-related cells (adipocytes). WGCNA showed that PUM1 positively interacted with several known osteoporosis genes (AKAP11, JAG1, and SPTBN1). ZNF423 was the highly connected intramodular hub gene and interconnected with 21 known osteoporosis-related genes, including JAG1, EN1, and FAM3C.

Conclusions

Our study identified seven potentially pleiotropic genes associated with osteoporosis and obesity. The findings may provide new insights into a potential genetic determination and codetermination mechanism of osteoporosis and obesity.

Pleiotropic analysis of osteoporosis and obesity was performed by cFDR. Strong pleiotropic enrichment and seven potentially pleiotropic SNPs were found for osteoporosis and obesity.

Recently, genome-wide association studies (GWASs) have identified thousands of variants for many human complex diseases, but these single-nucleotide polymorphisms (SNPs) failed to explain a substantial proportion of the heritability of most complex diseases studied, such as osteoporosis and obesity. Because of polygenic architecture, a large number of SNPs with small to modest effect sizes (but still important) have not been identified (1, 2). Currently, an important task and challenge is to take full advantage of existing GWAS data to explore more novel genetic variants underlying complex diseases, especially for the SNPs that do not meet the conservative Bonferroni-corrected significance level. Previous studies estimated that about 4.6% of SNPs and 16.9% of genes are involved in pleiotropic effects (3). The presence of pleiotropy in the genome indicates that correlated traits may have overlapping genetic determinants.

Osteoporosis is the most common metabolic bone disease that is characterized by reduced bone mineral density (BMD), deficiencies in the structure of bone tissue, and increased risk of low-trauma fractures (4). The most widely accepted measurement for predicting the risk of osteoporosis is BMD (4). BMD is a highly heritable trait, and its heritability ranges from 0.5 to 0.9 (5). To date, previous GWASs had identified ∼200 genes associated with osteoporosis-related traits (6–9), which explained <10% of the total genetic variance in any single study population. Therefore, novel statistical approaches to identify novel genes/variants important for osteoporosis-related traits are needed.

Obesity is a disease of excessive storage of body fat because of chronic imbalance of energy intake and consumption (10). It becomes a common public health problem because it can increase the risk of developing other serious diseases, such as type 2 diabetes, hypertension, cardiovascular disease, and coronary heart diseases (11, 12). A widely used measurement for obesity research is body mass index (BMI), defined as body weight divided by the square of height, which has a strong genetic determination, with heritability of 0.4 to 0.7 (13–15). Body fat distribution is also a heritable trait, with heritability ranging from 0.3 to 0.6 (16, 17).

Osteoporosis and obesity are closely related diseases (18). Obesity has long been viewed as a protective factor to osteoporosis (19, 20), whereas other studies suggested that fat accumulation had a negative effect on bone mass (18, 21, 22). Adipocyte tissue secretes the estrogen synthesis enzyme, leptin, adiponectin, and various proinflammatory cytokines, which are important to bone remodeling (23, 24). Adipocytes and osteoblasts originate from a common progenitor, bone marrow mesenchymal stem cells (25). Some studies suggested that bone-derived factors, such as osteopontin and osteocalcin, exert an endocrine regulation on body weight and glucose homeostasis (26). In addition, many reports have proved overlapping genetic susceptibility in osteoporosis and obesity. The genetic correlation between BMD and BMI is about 0.4 in Chinese Han ethnicity, suggesting that the two traits share pleiotropic genes (27). For the currently reported osteoporosis and obesity GWAS SNPs/genes, 12 genes or loci are shared between osteoporosis and obesity, such as DNM3, CDKAL1, and MPP7. Several obesity and osteoporosis shared genomic regions have been found through a bivariate whole-genome linkage scan, which include interleukin 6 and tumor necrosis factor α that influence both osteoporosis and obesity (28). A bivariate genome-wide association analysis discovered that SOX6 played a pleiotropic role in both bone and fat, providing further support for the existence of pleiotropic genes for osteoporosis and obesity (29). However, despite these interesting yet limited findings of the pleiotropic loci underlying osteoporosis and obesity, there are many other shared genetic loci influencing these two diseases awaiting identification.

As far as we know, Andreassen et al. (30) recently proposed a novel genetic pleiotropy-informed conditional false discovery rate (cFDR) method to identify shared genetic variants using existing GWAS summary statistics data. This method exploits the idea that a variant with statistically significant effects on two related traits is more likely to be a true effect and therefore has a higher probability of being detected in multiple independent studies. This method incorporates the summary statistics from two independent GWASs to test variants for association with the principal trait (the trait that is being tested for the association) conditional on different strengths of association with the second trait (conditional trait). The cFDR method allows us to focus specifically on the subset of variants with a given strength of association in the conditional trait (30). Andreassen et al. (30) have applied this method and successfully discovered novel pleiotropic loci that are associated with schizophrenia and bipolar disorder/cardiovascular disease risk factors (30, 31). Our group also has applied this method in a number of other pairs of correlated traits/diseases (32–38).

In this study, we applied the cFDR method with GWAS summary statistics data from three large independent studies to identify novel variants with pleiotropic effects on the two related traits, femoral neck (FN) BMD and BMI or FN BMD and waist-to-hip ratio (WHR). With this approach, we identified seven pleiotropic variants implicating a shared genetic mechanism for FN BMD and BMI or FN BMD and WHR. The results of this study enable us to better characterize the potentially genetic mechanisms underlying the correlations of FN BMD and BMI or FN BMD and WHR, as well as better understand the potentially mechanistic relationships of osteoporosis and obesity.

Materials and Methods

GWAS data sets

The GWAS data sets we applied in this study were all downloaded from publicly available online database. These data sets contained summary statistics, including the P values of association and direction of effect for each variant. To our knowledge, the data sets we used are among the largest and/or the latest meta-analysis results in the osteoporosis and obesity fields, respectively.

The data set containing the summary statistics results for association with FN BMD was derived from the study of the Genetic Factors for Osteoporosis Consortium in 2015 (http://www.gefos.org/?q=content/data-release-2015) (9). This GWAS meta-analysis used whole-genome sequencing, whole-exome sequencing, and deep imputation of previous genotype data to identify variants associated with FN BMD, lumbar spine BMD, and forearm BMD. The GWAS meta-analysis was performed on >10 million genotyped or imputed SNPs in 33 individual studies consisting of 53,236 Caucasian participants. Because hip fractures are the most severe type of osteoporotic fractures and directly associated with high morbidity and mortality (39), FN and total hip BMD are the most important risk traits for the study of osteoporosis.

The data set with summary statistics results of BMI was taken from a GWAS meta-analysis of 339,224 (322,154 European descent individuals and 17,070 non-European descent individuals) individuals performed by the Genetic Investigation of Anthropometric Traits Consortium (http://www.broadinstitute.org/collaboration/giant/index.php/GIANT_consortium_data_files) (40). The meta-analysis that combined 125 studies and contained >2 million genotyped or imputed SNPs was to identify the variants association with BMI. The WHR association data set was downloaded from a GWAS meta-analysis in up to 224,459 individuals (210,088 European ancestry individuals and 14,371 non-European ancestry individuals) analyzed by the Genetic Investigation of Anthropometric Traits Consortium (41). The WHR association meta-analysis combined 57 GWAS studies and 52 Metabochip studies, which included ∼2.5 million genotyped and imputed SNPs. An easily accessible and widely used measurement of body fat distribution is WHR, a ratio of waist and hip circumferences. WHR is a better index of abdominal fat and fat distribution than waist circumference, where a ratio of waist and hip circumferences can better distinguish abdominal or gluteal-femoral adiposity as waist circumference divided by hip circumferences.

Statistical analysis

SNP pruning and annotation

The original GWAS results were adjusted by using genomic control to ensure that the variance estimates for each SNP are not inflated due to potential population structure. First, we removed the SNPs that passed the Bonferroni-corrected significance threshold in the original GWAS and annotated SNPs to genes in the GWAS studies (FN BMD, BMI, and WHR). Next, we removed variants with large correlations between pairs of variants based on linkage disequilibrium. The pruning algorithm started at a window of 50 SNPs, where linkage disequilibrium was calculated between each pair of SNPs by using the HapMap III genotypes. If the pairs have an r² value >0.2, the SNP with the smaller minor allele frequency of that pair will be removed. After the initial removal of SNPs, the window slid five SNPs forward and repeated the process until there were no pairs of SNPs with r² > 0.2. After the pruning process, 122,152 variants for FN BMD and BMI and 120,573 variants for FN BMD and WHR were used in the cFDR analysis.

Conditional Q-Q plots for assessing pleiotropic enrichment

To assess the enrichment of association with the principal trait, we constructed conditional Q-Q plots based on varying levels of significance with the conditional trait. The Q-Q plots present a graph of the observed distribution of P values plotted against the expected distribution of P values under the null hypothesis. We plotted the Q-Q curve for the empirical quantiles of nominal –log₁₀(P) values for SNPs’ association with the principal trait for the subset of SNPs with significance levels below the varying thresholds with the conditional trait. The nominal P values are plotted on the y-axis, and q values are plotted on the x-axis. The q values are the quantiles of the nominal P values for SNPs’ association with the principal phenotype for the subset of SNPs with significance levels below the varying thresholds with the conditional phenotype. The q values can be calculated as the number of SNPs with P values less than or equal to the P value divided by the total number of SNPs in the group. Pleiotropic enrichment can be observed when the Q-Q curve is leftward deflected from the expected null line. Pleiotropic enrichment means an increase in the number of loci that are associated with the principal trait when gradually restricting the subset of SNPs at stronger significance levels of association with the conditional trait. Larger spacing between conditional Q-Q plots stands for a greater extent of pleiotropic genes shared between traits, and the earlier departure from the null line suggests a greater proportion of true associations for the nominal P value of the given phenotype.

Conditional statistics for one trait

The unconditional false discovery rate for a set of variants is defined as the probability of a false-positive association for a random SNP. For a set of observed association P values, the unconditional false discovery rate is estimated as the ratio of the expected quantile of the P value under the null hypothesis with the observed quantile.

The cFDR extends this idea to the two-phenotype case, which is characterized as the probability that a random SNP is not associated with the principal phenotype given that the observed P values for each phenotype are smaller than predefined disease-specific significance thresholds. The cFDR is given by cFDR(p_i|p_j), where $H_0^{\left(i\right)}$represents the null hypothesis that the particular SNP is not associated with the principal trait i, p_i represents the association for the particular variant with the principal trait, and p_j represents the association of the same SNP with the conditional trait.

$$\text{cFDR}\left(p_i\text{|}{p}_j\right)=\text{Pr}(H_0^{\left(i\right)}|P_i\le p_i,P_j\le p_j)$$

We computed the cFDR of each variant for the BMI or WHR conditioned on the association with FN BMD (BMI|FN BMD or WHR|FN BMD) and vice versa (FN BMD|BMI or FN BMD|WHR). To estimate whether the cFDR method results in the enrichment of associated SNPs, we controlled the subset of SNPs being tested using the following criteria for P_j ≤ p_j; P_j ≤ 1 (all SNPs), P_j ≤ 0.1, P_j ≤ 0.01, P_j ≤ 0.001, and P_j ≤ 0.0001, based on significance for the association with the conditional trait. SNPs with a cFDR value <0.05 were deemed to be statistically significantly associated with the principal phenotype.

Conjunction statistics for both traits

Next, we computed the conjunction cFDR (ccFDR), which is defined as the probability that a given SNP has a false-positive association with both the principal and conditional traits. The maximum cFDR value of the two traits was deemed equal to the ccFDR value of each variant.

$$\text{ccFDR}=\text{max}({\text{cFDR}}_{i|j},{\text{cFDR}}_{j|i})$$

All SNPs were considered to be statistically significantly associated with both traits when the ccFDR was <0.05.

Manhattan plots for conditional statistics and conjunction statistics

To illustrate the localization of the independent loci associated with BMI or WHR conditional on FN BMD and vice versa, as well as independent loci with a pleiotropic effect on FN BMD and BMI or FN BMD and WHR, we present conditional or conjunction Manhattan plots, which display all the SNPs analyzed in the study and their chromosomal location in the genome. Any variant with a –log₁₀(cFDR)/–log₁₀(ccFDR) value >1.3 (equivalent to cFDR or ccFDR <0.05) was determined to be statistically significantly associated with the principal trait or both traits.

Functional prediction for SNPs

For the potentially pleiotropic SNPs identified in FN BMD and BMI or FN BMD and WHR, we applied the F-SNP program (http://compbio.cs.queensu.ca/F-SNP/) to analyze and predict their potential functions. The F-SNP program provides the functional effects at splicing, transcriptional, translational, and posttranslational levels (42).

Differential expression analysis

We performed differential expression analysis in two expression data sets for the identified potentially pleiotropic genes to explore their functions in relationship to the related phenotypes (individually or pleiotropically). We downloaded two normalized publicly available expression data sets from GEO (https://www.ncbi.nlm.nih.gov/gds/). An experiment for the data set GSE7429 was performed to identify genes associated with bone using a design of high vs low BMD in circulating B lymphocytes. B lymphocytes, an important cell type of the immune system that is related to bone metabolism, express/secrete factors involved in osteoclastogenesis, such as receptor tumor necrosis factor superfamily member 11 and osteoprotegerin (43). In addition, B cells may serve as precursors of osteoclasts, and the precursors have the ability to differentiate into osteoclasts after estrogen deficiency (44, 45). GSE2508, which contains data on transcriptomic expression of 20 lean children and 19 obese children, was designed to detect obesity-related genes in adipocytes. t Tests were performed to identify differential expression genes through comparing means of the gene expression signals in two groups. The nominal significance threshold was set as P values <0.05.

Weighted gene coexpression network analysis

To explore potential cofunctions of putative pleiotropic genes identified in our study, we used weighted genes coexpression network analysis (WGCNA) to construct interaction networks of pleiotropic genes with known osteoporosis or obesity genes, respectively (46). First, we searched osteoporosis or obesity GWAS references with keywords BMD, WHR, BMI, and GWAS in PubMed and selected osteoporosis/obesity-associated SNPs with a P value ≤5.0 × 10⁻⁸. A total of 203 osteoporosis and 379 obesity association genes were selected. Next, combining osteoporosis/obesity GWAS association genes and cFDR suggested genes, we identified probes representing 152 osteoporosis-related genes (14 novel and 138 known genes) and 303 obesity-related genes (25 novel and 278 known genes) and applied WGCNA (R package) to generate networks in gene expression data sets E-MEXP-1618 (http://www.ebi.ac.uk/arrayexpress/experiments/E-MEXP-1618/) and GSE2508, respectively. Expression data set E-MEXP1618 was generated from transiliac bone biopsy specimens from 84 postmenopausal white women in which most cells were osteocytes, with some very small proportions being osteoclasts and osteoblasts, with 41 low FN BMD women (T score ≤–1.0) and 43 normal or high FN BMD women (T score >–1.0). Finally, the module network was visualized by the software Cytoscape (47), including nodes and module edges with topological overlap measurement (TOM) larger than 0.10 with pleiotropic genes. TOM is a parameter ranging from 0 to 1, where a TOM of 0 means that two genes are not connected at all and two genes with a high TOM are considered to highly interconnect with the same sets of genes (48).

Results

Assessment of pleiotropic enrichment

Conditional Q-Q plots for FN BMD given nominal P values of association with BMI (Fig. 1A) or WHR (Fig. 1B) showed general polygenic pleiotropic enrichment across different significance thresholds for BMI or WHR. The gradual leftward shift when restricting the P value thresholds of associations with BMI or WHR indicated an increase in the number of true associations with FN BMD conditioned on BMI or WHR. A stronger enrichment can be observed for BMI or WHR conditional on FN BMD association nominal P values (Fig. 1C and 1D), as there appears to be a greater amount of separation between the different curves.

Figure 1.

Stratified Q-Q plots of nominal vs empirical –log₁₀P values in principal trait below the standard GWAS threshold of P < 5.0 × 10⁻⁸ as a function of significance of the association with conditional trait at the level of P < 1, P < 0.1, P < 0.01, P < 0.001, and P < 0.0001, respectively. (A) FN BMD|BMI (FN BMD conditional on BMI), (B) FN BMD|WHR, (C) BMI|FN BMD, and (D) WHR|FN BMD. The diagonal indicates the null hypothesis.

FN BMD–associated gene loci identified with cFDR

We identified a total of 27 SNPs associated with FN BMD when conditional on BMI and 32 loci given the association with WHR with a significance threshold of cFDR <0.05 (Supplemental Tables 2 and 4). Of the 27 SNPs for FN BMD|WHR, several SNPs are located at or near BMD-related genes, such as MARK3, ZBTB40, OPG, SOX6, and GALNT3 (Supplemental Table 2). Some SNPs are located at or near noncoding RNA, such as rs944082, rs11209223, and rs11952384 (GNG12-AS1 and MEF2C-AS1). Some SNPs found by FN BMD|WHR are located at known BMD-related genes, such as MARK3, SMAD3, and DNM3, and four SNPs are located at a noncoding RNA (GNG12-AS1) (Supplemental Table 4). Of these identified SNPs associated with FN BMD conditional on BMI or WHR, 20 SNPs overlapped; therefore, a total of 39 SNPs (44 genes) were associated with FN BMD, in which 19 genes were not identified in any previous GWASs.

BMI- or WHR-associated gene loci identified with cFDR

We detected a total of 147 SNPs statistically significantly (cFDR <0.05) associated with BMI and 19 SNPs statistically significantly associated with WHR given their association with FN BMD (Supplemental Tables 1 and 3). There were 87 SNPs with P values smaller than 10⁻⁵, which were not detected in the original meta-analysis. Of these SNPs, we found six SNPs located at FTO, five at MC4R, five at SEC16B, and four at TFAP2B, which were known candidate genes related to BMI and obesity. Our analysis also confirmed several loci associated with WHR, such as TBX15 and DNM3, which were associated with WHR in a previous meta-analysis (41). Of these SNPs, we identified 72 novel statistically significant loci, such as DNMT3A and POC5, which need further experimental and statistical validation. There were three SNPs (rs11079813, rs6723710, rs751008) located at SKAP1, LINC01122, and DOCK1 identified in the analyses of BMI|FN BMD and WHR|FN BMD.

Pleiotropic gene loci for FN BMD and BMI or FN BMD and WHR

We detected seven independent pleiotropic variants—rs3759579 (MARK3), rs2178950 (TRPS1), rs1473 (PUM1), rs9828174 (XXYLT1), rs2047937 (ZNF423), rs17277372 (DNM3), and rs335170 (PRDM6)—which were statistically significantly associated with FN BMD and BMI or FN BMD and WHR (ccFDR <0.05). Of these loci, rs3759579 (MARK3), rs2178950 (TRPS1), rs1473 (PUM1), and rs9828174 (XXYLT1) were associated with FN BMD and BMI and rs2047937 (ZNF423), rs17277372 (DNM3), and rs335170 (PRDM6) with FN BMD and WHR (Fig. 2 and Table 1). Conforming to the positive relationship between BMD and BMI found in most epidemiological studies, three (MARK3, PUM1, and XXYLT1) of four potentially pleiotropic SNPs of FN BMD and BMI have the same effect direction on both FN BMD and BMI (Table 1). Interestingly, all of FN BMD and WHR association loci (ZNF423, DNM3, and PRDM6) have an opposite influence on both FN BMD and WHR, which is relatively consistent with a negative relationship of osteoporosis and body fat distribution in previous studies (49, 50). Taken together, the relationship and mechanism between obesity and osteoporosis may be complex. Four potentially pleiotropic SNPs (rs3759579, rs2178950, rs17277372, and rs2047937) located at MARK3, TRPS1, DNM3, and ZNF423 genes, respectively, were known to be related with bone or fat metabolism. Three SNPs (rs1473, rs9828174, rs335170) located at or close to genes (PUM1, XXYLT1 and PRDM6) were novel.

Figure 2.

Conjunction Manhattan plot of conjunction –log₁₀ ccFDR values for FN BMD and BMI or FN BMD and WHR, and the solid line marking the –log₁₀ ccFDR value of 1.3 corresponds to a ccFDR of 0.05. (A) FN BMD and BMI and (B) FN BMD and WHR. The figure shows the genomic locations of pleiotropic loci.

Table 1.

Pleiotropic Loci Associated With Both FN BMD and BMI/WHR With cFDR <0.05

Characteristic	rs3759579	rs2178950	rs1473	rs9825174	rs2047937	rs17277372	rs335170
Gene	MARK3	TRPS1	PUM1	XXYLT1	ZNF423	DNM3	PRDM6
Chromosome	14q32.32	8q24.12	1p35.2	3q29	16q12	1q24.3	5q23.2
A1	A	C	A	C	C	A	A
A2	G	G	G	T	T	G	C
P value BMI	2.18E-05	8.86E-04	4.89E-04	7.13E-04	/	/	/
P value WHR	/	/	/	/	3.80E-04	1.60E-03	2.20E-03
P value FN BMD	3.31E-05	3.95E-04	2.67E-04	2.50E-04	2.10E-07	7.80E-07	1.80E-05
Effect for BMI	−	−	+	−	/	/	/
Effect for WHR	/	/	/	/	−	+	+
Effect for FN BMD	−	+	+	−	+	−	−
cFDR BMI|FN BMD	7.43E-04	3.06E-02	2.66E-02	3.64E-02	/	/	/
cFDR FN BMD|BMI	3.54E-03	4.08E-02	4.33E-02	4.78E-02	/	/	/
cFDR WHR|FN BMD	/	/	/	/	1.14E-03	6.40E-03	1.91E-02
cFDR FN BMD|WHR	/	/	/	/	2.21E-05	1.01E-04	1.96E-03
ccFDR	3.54E-03	4.08E-02	4.33E-02	4.78E-02	1.14E-03	6.40E-03	1.91E-02

Independent gene loci (r² < 0.2) with SNPs that have ccFDR <0.05 in FN BMD and BMI/FN BMD and WHR. All SNPs are listed with their nearest gene(s), chromosomal location, effect allele A1, noneffect allele A2, raw P values for univariate GWAS for each trait, cFDR values for each trait, and ccFDR values. The effect direction of the potentially pleiotropic SNP on FN BMD and BMI/WHR was obtained from the original GWAS data. The effect represents the summary of effect directions (“+” indicates positive effect of A1 allele, “−” indicates negative effect of the A1 allele, and “/” indicates there is no value for the column).

Using the F-SNP program, we investigated the potential functions of these seven SNPs. The SNP rs3759579, within the intron of MARK3, is located at potential transcription factor–binding sites and may have a role in transcriptional regulation. A G-A change at rs3759579 may produce binding sites of CP2 and p300. Previous studies suggested the protein encoded by CP2 may be involved in osteoblast differentiation (51). P300, a histone acetyltransferase, has been demonstrated to be involved in adipogenesis through regulation of peroxisome proliferator activated receptor γ (PPARγ) and C/EBPβ expression (52) and osteoclastogenesis (53). The XXYLT1 SNP rs9825174 is located at a transcription factor–binding site, and the C-T transition may generate a binding site of C/EBPβ, which plays important roles in inducing expression of PPARγ (54). The G-C change in TRPS1 rs2178950 was found to produce a new exonic splicing enhancer motif—ACTGAA, which may regulate splicing of its pre-messenger RNA (mRNA). The other four SNPs did not show any known transcriptional regulation functions or splicing regulation functions according to the F-SNP program.

By performing t tests for differential expression analysis for the pleiotropic genes in the transcriptomic expression data of GSE2508, we found that TRPS1, MARK3, PUM1, and PRDM6 were differentially expressed between the lean and obese participants (TRPS1, P = 0.002; MARK3, P = 0.004; PUM1, P = 0.033; and PRDM6, P = 0.018). Furthermore, t tests of the transcriptomic expression data of GSE7249 showed that two genes (PUM1 and DNM3) were differentially expressed between the high BMD and low BMD participants with P = 0.042 and P = 0.014, respectively (Tables 2 and 3). The results of differential expression analysis for the seven potentially pleiotropic genes in the E-MEXP1618 were not statistically significant. A possible reason is that bone biopsy specimens were collected from several bone cells, potentially leading to the differential expression of these genes confounded by heterogeneous cell proportions (thus with larger noise) of these cells in different participants.

Table 2.

Basic Characteristics of the Studied GEO Data Sets

Characteristic	GSE7429	E-MEXP-1618	GSE2508
Design	High BMD vs low BMD	Normal or high FN BMD vs low BMD	Obese vs lean
Disease	Osteoporosis	Osteoporosis	Obesity
Target cells	Circulating B lymphocytes	Transiliac bone biopsies	Adipocytes
Sample size	(10 high vs 10 low)	(43 normal or high vs 41 low)	(19 obese vs 20 lean)
Platform	Affymetrix Human Genome U133A Array	Affymetrix GeneChip Human Genome U133 Plus 2.0	Affymetrix Human Genome U95 Array

Table 3.

Differential Expression Analysis for the Pleiotropic Genes in Osteoporosis- or Obesity-Related Cells

Gene Symbol	Probe ID	t Test P Value
MARK3	202568_s_at	4.02E-01
MARK3	232537_x_at	5.00E-02
MARK3	49762_at	4.00E-03
TRPS1	218502_s_at	5.22E-01
TRPS1	234351_x_at	1.05E-01
TRPS1	55443_at	2.00E-03
PUM1	201166_s_at	4.20E-02
PUM1	201164_s_at	5.50E-02
PUM1	40048_at	3.30E-02
XXYLT1	NA	NA
XXYLT1	226891_at	2.49E-01
XXYLT1	78532_at	8.53E-01
ZNF423	214761_at	5.43E-01
ZNF423	214761_at	2.61E-01
ZNF423	83639_at	6.57E-01
DNM3	209839_at	1.40E-02
DNM3	1558502_s_at	4.07E-01
DNM3	51761_at	5.40E-02
PRDM6	NA	NA
PRDM6	230311_s_at	8.00E-02
PRDM6	71140_at	1.81E-02

We only listed the most statistically significant expression results of probes with one gene with multiple detected probes. Bold values indicate nominally significant (P ≤ 0.05) of t-test.

Abbreviation: NA, not available.

Using WGCNA for the 152 osteoporosis genes (138 known and 14 novel genes) in bone biopsy specimens, 137 genes were parsed into two gene modules. Potentially pleiotropic genes MARK3, TRPS1, PRDM6, and XXYLT1 were in the blue module, whereas potentially pleiotropic genes DNM3, PUM1, and ZNF423 were in the turquoise module. In the turquoise module, ZNF423 is one of three top network hub genes (JAG1, ZNF423, and HOXC6). With the TOM ≥0.10, there were 21 known osteoporosis-related genes interconnected with ZNF423 (Fig. 3A), and ZNF423 was positively correlated with all of the 21 known osteoporosis-associated genes, such as JAG1, EN1, and SMAD3. PUM1 was interconnected with AKAP11, JAG1, and SPTBN1 in the turquoise module with TOM ≥0.10 (Fig. 3A). Furthermore, PRDM6 interacted with FAM155A and WNT4 in the blue module (Fig. 3B). There were three pleiotropic genes identified in two modules, MARK3 and TPRS1 in the turquoise module and ZNF423 in the blue module, when applying network analysis for 303 obesity-related genes (278 known and 25 novel genes) in adipocytes. However, the interconnection between ZNF423, MARK3, TRPS1, and obesity-related genes was relatively weak.

Figure 3.

(A) ZNF423 and PUM1 centered network view, which characterizes the coexpression relationships for ZNF423 and PUM1 with BMD-associated genes in the osteoporosis turquoise module. This network view includes all edges and their corresponding nodes connected to ZNF423 and PUM1 with a TOM ≥0.10. The width of edges presents the value of TOM; the larger the TOM, the larger the width of the edge. All of the correlations between nodes are positive. (B) PRDM6 centered network view in the osteoporosis blue module, which presents the edges and nodes connected with PRDM6 with TOM ≥0.10.

Discussion

In the current study, we applied the cFDR method that combined the summary statistics from independent GWAS meta-analyses to identify the variants that are associated with FN BMD and BMI or FN BMD and WHR, as well as successfully identified seven pleiotropic variants (rs3759579, rs2178950, rs1473, rs9825174, rs2047937, rs17277372, rs335170) that were associated with FN BMD-BMI or FN BMD-WHR.

Although previous studies have found the four genes (MARK3, TRPS1, DNM3, and ZNF423) playing an important role in osteoporosis or obesity individually, our study for the first time discovered the pleiotropic effects of these genes in both osteoporosis and obesity. MARK3 was previously found to associate with BMD (55). Lennerz et al. (56) reported that the mice loss of Par-1a/MARK3/C-TAK1 kinase showed increased energy expenditure and reduced adiposity. DNM3 was associated with BMD (55) or WHR (57) in previous GWASs. Our results, together with previous studies, suggested that MARK3 and DNM3 play pleiotropic roles in osteoporosis and obesity.

The SNP rs2047937, located at the intron of the ZNF423 (also called ZFP423) gene, encodes a nuclear protein that is a DNA-binding transcription factor by using distinct zinc fingers in different signaling pathways. Multipotent mesenchymal precursor cells can differentiate into several cell types, such as osteoblasts and adipocytes (58). Numerous studies have demonstrated that the bone morphogenetic protein (BMP) and small mothers against decapentaplegia (SMAD) signal pathway not only induced the recruitment and differentiation of mesenchymal precursors into osteoblasts (59) but also controlled the preadipocyte commitment (60) and the differentiation of committed preadipocytes to adipocytes (61). ZNF423 was identified as a DNA-binding cofactor associated with SMADs in response to BMPs, leading to activate transcription of BMP target genes (62). Gupta et al. (60) reported that ZNF423 acted as a prominent preadipocyte commitment factor and regulated the expression of PPARγ through amplification of the BMP signaling pathway, and they discovered that Znf423 knockout mice exhibited noticeable defects in adipose mass. In the osteoporosis turquoise module, ZNF423 was found to highly interact with reported osteoporosis-associated genes, such as JAG1 and EN1. JAG1 participated in normal trabecular bone formation and inhibited periosteal expansion through affecting osteoprogenitor cells and their progeny but not mature osteoblasts (63). Recently, Zfp423 was found to act as a hinge regulated by Zfp521 coordinating regulation of adipocyte and osteoblast differentiation (64). Statistical evidence obtained in our study, together with previously reported biological functions, suggested that ZNF423 may play a role in both bone and fat metabolism through progenitor cells regulating cell differentiation between osteoblasts and adipocytes.

The TRPS1 gene encodes a zinc finger protein that has several different zinc finger motifs, including a GATA-type and IKAROS-like zinc fingers (65). Mutation of the human TRPS1 gene leads to trichorhinophalangeal syndrome, which exhibits abnormal development of various organs, including craniofacial and skeletal malformation due to abnormal differentiation of cartilage and bone (66). Numerous studies demonstrated that TRPS1 not only modulated chondrocyte proliferation, differentiation, and ossification but also controlled mitotic progression in chondrocytes (67). In addition, TRPS1 can directly bind the osteocalcin promoter in the presence or absence of RUNX2, mediating the osteocalcin transcription to regulate the osteoblast differentiation and deposition of mineralized matrix (68). TRPS1 mRNA was downregulated in obese adipocytes compared with nonobese adipocytes, suggesting that it might contribute to obesity pathology. Combining with biological functions, our study implicates that TRPS1 may play a pleiotropic role in bone and obesity metabolism.

Importantly, three novel SNPs located at 1p35.2 (PUM1), 5q23.2 (PRDM6), and 3q29 (XXYLT1) have not been reported in previous osteoporosis- or obesity-related research. PUM1 encodes a member of the PUF family that is implicated in various physiological processes and serves as a translational regulator of specific mRNAs by binding to their 3′ untranslated regions (69). Many studies indicated that PUM1 represses gene expression by binding to an eight-nucleotide RNA consensus sequence (the pumilio response element, 5-UGUANAUA-3), resulting in mRNA decay and translation inhibition (70, 71). By using the RNA-binding protein immunoprecipitation chip method, >1000 PUM1 target mRNAs have been identified (71, 72). In the list of mRNA targets, we found several genes related to osteoporosis or obesity, such as LRP5, WNT5A, BMP2, LMX1B, RQCD1, and ETS2 (40, 55, 73). PUM1 can recruit CCR4-NOT deadenylase subunits to cause translational inhibition and mRNA degradation (74). Several studies found that subunits of CCR4-NOT complex were involved in bone mass regulation (75) and adipocyte differentiation (76), suggesting that PUM1 can recruit CCR4-NOT deadenylase subunits and subunits of CCR4-NOT complex were involved in bone mass regulation. Therefore, we speculated that PUM1 may recruit CCR4–NOT complex controlling gene expression and involved in osteoporosis and obesity pathology. PUM1, meanwhile, was differentially expressed in two groups of B lymphocytes and adipocytes, suggesting its pleiotropic effect in the two phenotypes. PUM1 serves as a translational regulator and interacts with several osteoporosis GWAS-associated genes (AKAP11, JAG1, and SPTBN1), suggesting that it might regulate osteoporosis association genes and participate in osteoporosis pathology. Taken together, our study suggests that PUM1 probably plays an important role in regulating posttranscriptional translation of osteoporosis- and obesity-related genes. However, its functions in osteoporosis and obesity pathology need further functional studies.

Compared with a standard single phenotype analysis, cFDR simultaneously analyzing multiple related traits allows for an increased discovery of trait-associated variants using GWAS summary data (30). In addition, it allows us to take advantage of polygenic pleiotropy for further exploring common loci or genes that may influence multiple related phenotypes. Furthermore, through restricting the subset of SNPs at a different significance level, it can greatly lessen the burden of multiple testing and subsequently improve the detection of trait-associated variants. It is worth noting that the study samples for the cFDR method should be independent under the assumption so that maximum power can be achieved, because the overlap in the GWAS samples may reduce the information used for detection of potentially pleiotropic loci. In our study, a small proportion of cohorts overlapped in each pair of the summary statistics (FN BMD and BMI or FN BMD and WHR). Although the power of cFDR may be reduced owing to the overlapping samples, the potentially genetic pleiotropic variants identified in our study were robust and reliable.

In conclusion, we detected seven pleiotropic loci (MARK3, DNM3, ZNF423, TRPS1, PUM1, PRDM6, and XXYLT1) for FN BMD and BMI or FN BMD and WHR by using the cFDR method with existing individual GWAS summary statistics data. Of the seven genes, four of them (MARK3, DNM3, ZNF423, and TRPS1) were reported as potential candidate genes in osteoporosis or obesity in individual GWAS, and our study for the first time suggested their potentially pleiotropic roles. The other three novel potentially pleiotropic genes (PUM1, PRDM6, and XXYLT1) were not found in previous bone or obesity research. The findings in the current study enhance our knowledge of the shared genetic influences between osteoporosis and obesity and suggest a possible research direction for subsequent functional studies of these potentially implicated genes in the joint pathophysiology of both disorders.

Abbreviations:

BMD

bone mineral density

BMI

body mass index

BMP

bone morphogenetic protein

ccFDR

conjunction conditional false discovery rate

cFDR

conditional false discovery rate

FN

femoral neck

GWAS

genome-wide association study

mRNA

messenger RNA

PPARγ

peroxisome proliferator activated receptor γ

SNP

single-nucleotide polymorphism

TOM

topological overlap measurement

WGCNA

weighted genes coexpression network analysis

WHR

waist-to-hip ratio.