Abstract
Summary
By adopting the extension approaches of Mendelian randomization, we successfully detected and prioritized the potential causal risk factors for BMD traits, which might provide us novel insights for treatment and intervention into bone-related complex traits and diseases.
Introduction
Osteoporosis (OP) is a common metabolic skeletal disease characterized by reduced bone mineral density (BMD). The identified SNPs for BMD can only explain approximately 10% of the variability, and very few causal factors have been identified so far.
Methods
The Mendelian randomization (MR) approach enables us to assess the potential causal effect of a risk factor on the outcome by using genetic IVs. By using extension methods of MR—multivariable MR (mvMR) and MR based on Bayesian model averaging (MR-BMA)—we intend to estimate the causal relationship between fifteen metabolic risk factors for BMD and try to prioritize the most potential causal risk factors for BMD.
Results
Our analysis identified three risk factors T2D, FG, and HCadjBMI for FN BMD; four risk factors FI, T2D, HCadjBMI, and WCadjBMI for FA BMD; and three risk factors FI, T2D, and HDL cholesterol for LS BMD, and all risk factors were causally associated with heel BMD except for triglycerides and WCadjBMI. Consistent with the mvMR results, MR-BMA confirmed those risk factors as top risk factors for each BMD trait individually.
Conclusions
By combining MR approaches, we identified the potential causal risk factors for FN, FA, LS, and heel BMD individually and we also prioritized and ranked the potential causal risk factors for BMD, which might provide us novel insights for treatment and intervention into bone-related complex traits and diseases.
Keywords: Causal relationship, Mendelian randomization, Multivariable MR, Osteoporosis
Introduction
Osteoporosis (OP), a common metabolic skeletal disease, is mainly characterized by reduced bone mineral density (BMD) that results in increased risk for bone fragility and eventually bone fractures. BMD, a highly heritable trait and an important index of bone strength, is frequently used in for OP clinical diagnosis and genetic research. Previous studies estimated that nearly 75% of the variance in BMD at the site of the femoral neck (FN) and 83% in the lumbar spine (LS) may be explained by genetic determinants [1, 2]. Although genome-wide association studies (GWASs) have identified > 100 BMD-associated loci, the causal mechanisms implicated in the onset of OP remain largely unknown [3].
The Mendelian randomization (MR) approach [4] enables us to assess the potential causal effect of a risk factor on the outcome by using genetic instrumental variables (IVs). As an extension approach of MR, multivariable MR (mvMR) [5] is able to incorporate the different effect sizes for the pleiotropic SNPs with their related traits to simultaneously assess the causal effect of multiple related risk factors on the outcome. In comparison with two-sample MR, mvMR assumes the genetic IVs are associated with a set of risk factors, although not necessarily with every risk factor. Additionally, under the circumstances of directional pleiotropy [6], causal effect can be assessed even if none of the variants shows specific associations with any individual risk factor [5].
Recently, Zuber V et al. developed a novel approach [7] which combined multivariable MR with Bayesian model averaging (MR-BMA) that scales to high dimensional settings and can select biomarkers as potential causal risk factors for disease of interest. The authors demonstrated in their realistic simulation study that the method can detect and prioritize true risk factors even when the multiple risk factors are highly correlated [7]. The MR-BMA approach was applied on publicly available summary data of metabolites to successfully prioritize the most likely causal biomarkers for age-related macular degeneration [7].
Studies with established evidence showed that metabolic traits and related traits shared consistent association with OP [8-10]. However, traditional study results may be biased by potential confounding factors and reverse causality. Furthermore, among those risk factors that are known to be associated with OP, it remains unclear which are causal, and MR approaches have been tested for some of the factors like body mass index (BMI), type 2 diabetes (T2D), and lipids profile; however, fasting glucose (FG), fasting insulin (FI), hip circumference (HC), waist circumference (WC), and waist-to-hip ratio (WHR) had never been tested. Besides, we do not know which risk factor may play a more pivotal role in disease susceptibility. In the current study, we intend to apply the mvMR and MR-BMA approach to identify and prioritize the most likely potential causal metabolic risk factors related to BMD and OP fracture including BMI, T2D, HC, WC, WHR, high-density lipoprotein cholesterol (HDL cholesterol), low-density lipoprotein cholesterol (LDL cholesterol), and triglycerides (TG), FG, FI. To exclude the potential mediation effect of BMI, several traits were adjusted for BMI, HC adjusted for BMI (HCadjBMI), WCadjBMI, and WHRadjBMI, to further estimate their role in BMD or OP.
Material and methods
Data sources and SNP selection
Summary statistics for risk factor-associated SNPs were extracted from the largest available GWAS datasets to date performed by the corresponding Consortia in European populations (Table S1). For the implementation of mvMR, we selected SNPs that achieved genome-wide significance (P < 5 × 10−8) in the GWAS datasets for each risk factor as IVs. Effect estimates of these risk factor-associated SNPs on the risk of OP were assessed using the summary statistics of European individuals for FN, forearm (FA), and LS BMD from the GEFOS Consortium (Ntotal = 32,965) [11] and heel BMD (Ntotal = 426,824) [12]. The European samples from the 1000 Genomes Project reference panel were adopted to estimate linkage disequilibrium (LD) between chosen SNPs. When target SNPs were not available in the outcome study, we used proxy SNPs that were in high LD (r2 > 0.8) with the SNPs of interest. Summary statistics from these consortia can be downloaded at the public websites; for detailed information, please see Table S1.
IV selection and validation
There are several important assumptions for mvMR. First, genetic IV is associated with one or more of the risk factors. Second, genetic IV is not associated with a confounder of any of the risk factor–outcome associations. Third, the genetic IV is conditionally independent of the outcome given the risk factors and confounders. To ensure the SNPs used as IVs for risk factors are not in LD with each other, a vital assumption of MR, we calculated pairwise-LD between all our selected SNPs in the 1000 Genomes European reference sample using PLINK 1.90 [13]. For all pairs of SNPs determined to violate the independence assumption with r2 > 0.001, we retained only the SNP with the smaller association P value. To ensure the effect of a SNP on the exposure and the effect of that SNP on the outcome correspond to the same allele, we harmonized the effect of these instrumental SNPs by using a function that ensures all corresponding risk factors and outcome (FN, FA, LS, and heel BMD) alleles are on the same strand where possible. If they are not, then the function will flip alleles and use allele frequency to infer the strand of palindromic SNPs.
mvMR and univariate estimates
In the current study, standard inverse-variance-weighted (IVW) [14] analysis was used to estimate the causal effect of the multiple related risk factors on the BMD traits. After obtaining the selected instruments for each exposure, unadjusted BMI exposures and adjusted BMI exposures for those SNPs were then regressed against the outcome separately, weighting for the inverse variance of the outcome to ensure the genetic instruments with more precise association receive more weight in our analysis. MR Steiger test [15] was also performed to infer the causal direction between exposures and BMD traits. It calculates the variance explained in the exposures and outcome by the instrumenting SNPs, and tests if the variance in the outcome is less than the exposures. A Bonferroni-corrected threshold of P = 0.005 (0.05/10) was considered to be significant for each BMD trait individually, and 0.005 < P < 0.05 was considered suggestive of evidence for a potential association. To clarify if the effects are mediated by the other risk factors, we performed the univariate estimates for each risk factor (including FIadjFG—snps associated with FI but not with FG) individually to compare with mvMR estimates. A P < 0.05 was considered to be significant for each BMD trait individually. To highlight the genes/loci that were driving a link between fasting insulin and BMD measurements, we also performed pathway analysis using PANTHER Classification System (http://pantherdb.org/).
Sensitivity analysis
Furthermore, the weighted median estimator and maximum likelihood method (MLM) [16, 17] were also utilized to complement the IVW to provide more robust MR estimates. Weighted median is believed to generate unbiased estimates of the MR causal effect provided that > 50% of the weight comes from valid SNPs. Previous studies have confirmed that the weighted median estimator affords some distinct superiorities for its improved power of causal effect detection, lower type I error, and robustness to the MR assumptions. And MLM has also been encouraged in weak instruments situation, as the median of the distribution of the its estimator is close to unbiased even with weak instruments.
To assess whether there are horizontal pleiotropic effects where IVs affect BMD via more than one biological pathway, we employed MR-Egger regression [18], which is often used in meta-analysis to examine whether there is evidence of publication bias. When applying the MR-Egger method, the SNP’s effect estimate for exposures is plotted against its effect estimate on the outcome, and an intercept that deviates from the origin may provide evidence for potential pleiotropic effects across the genetic IVs.
MR-BMA estimates
MR-BMA was then applied to prioritize the most causally related risk factors for BMD. MR-BMA assumes that the true potential causal risk factors are very few and it considers the risk factor selection as a variable selection problem in the linear regression model. The approach considers all possible combinations of the risk factors and generates posterior probability (PP) for each specific model, where PP means the probability of including a specific risk factor in the model. Furthermore, MR-BMA adopts BMA which computes a marginal inclusion probability (MIP) for each risk factor, where MIP refers to the sum of the PP over all possible models where the risk factor is present. Then, MR-BMA will compute the model-averaged causal estimate (MACE) for each risk factor by ranking all the risk factors according to the corresponding MIP. Finally, MR-BMA will prioritize the best model by the PP value for each individual model. We presented the MIP values for all the risk factors and the best ten individual models for FN, FA, LS, and heel BMD according to the PP values of the corresponding risk factors being included in the specific models.
Two-sample MR estimates of T2D on fracture
To further validate the role of T2D in fracture, we applied the principles of two-sample MR to assess the role of T2D (62,892 T2D cases and 596,424 controls) [19] in the susceptibility of fracture. Summary statistics for fracture was derived from GEFOS 2018 release (37,857 cases and 227,116 controls) [20]. Same methods described previously were used to assess the effect estimates of T2D on fracture. All the analyses were implemented in R software environment.
Results
SNP selection and validation
Overall, we obtained 973 and 1028 LD-independent SNPs (Table 1) that achieved genome-wide significance for all the BMI-unadjusted and BMI-adjusted risk factors after implementing the pruning strategy previously described. Then, those SNPs were extracted from each BMD trait, respectively. After harmonizing the exposure and outcome datasets, there were 792/835, 830/872, 792/835, and 857/904 SNPs remained to perform the MR analysis for outcome (FN, FA, LS, and heel BMD). Detailed information for the number of SNPs used for each risk factor is shown in supplementary Tables S2-S9.
Table 1.
SNPs used for each risk factor in the MR analysis
| Exposure | SNPs |
|---|---|
| Type 2 diabetes | 118 |
| HDL cholesterol | 89 |
| LDL cholesterol | 81 |
| Triglycerides | 55 |
| Hip circumference | 52 |
| HCadjBMI | 75 |
| Waist circumference | 42 |
| WCadjBMI | 65 |
| Waist-to-hip ratio | 29 |
| WHRadjBMI | 38 |
| Fasting glucose | 35 |
| Fasting insulin | 14 |
| Body mass index (BMI) | 458 |
Multivariable MR estimates
There were suggestive positive associations between genetically determined T2D (beta = 0.033, standard error (se) = 0.270, P = 0.010), genetically FG (beta = 0.155, se = 0.073, P = 0.032), and FN BMD (Fig. S1). However, HC showed suggestive inverse association with FN BMD after adjusted for BMI (Fig. S1).
For FA BMD trait, we found that genetically predicted higher FI was significantly associated with increased FA BMD (b = 0.863, se = 0.280, P = 0.002, Fig. 1), and genetically determined T2D showed suggestively positive association with FA BMD (b = 0.070, se = 0.026, P = 0.008, Fig. S2). After adjusted for BMI, we observed suggestive and significant negative associations between WCadjBMI (Fig. S2), HCadjBMI, and FA BMD (Fig. 1), respectively.
Fig. 1.
Multivariable MR analysis forest plot: effect of multiple risk factors on FA BMD
We observed a significantly positive association between genetically determined T2D and LS BMD (b = 0.061, se = 0.015, P = 4.24E–05, Fig. 2), and genetically predicted higher FI was suggestively associated with elevated LS BMD (Fig. S3). However, genetically predicted HDL cholesterol was found inversely associated with LS BMD (Fig. S3).
Fig. 2.
Multivariable MR analysis forest plot: effect of multiple risk factors on LS BMD
For heel BMD trait, genetically determined BMI, FI, HC, T2D, WC, WHR, and WHRadjBMI all showed significant positive associations with heel BMD (Fig. 3), and genetically increased level of FG showed a suggestive association with heel BMD (Fig. S4). However, genetically determined HDL and LDL cholesterol showed significant inverse associations with heel BMD (Fig. 3). After adjusted for BMI, genetically HC showed negative association with heel BMD (Fig. 3).
Fig. 3.
Multivariable MR analysis forest plot: effect of multiple risk factors on heel BMD
Univariate MR estimates
For FN BMD, our univariate MR analysis suggested T2D (P = 0.017), WCadjBMI (P = 0.036), and FIadjFG (P = 0.048) were potential causal risk factors (Table S10). While T2D, FG, and HCadjBMI become potential causal risk factors after performing mvMR analysis. In univariate analysis, FI (P = 0.009) and HCadjBMI (P = 0.002) showed significant association with FA BMD (Table S11), except for those two, T2D and WCaDjBMI also showed association with FA BMD in the mvMR estimates. For LS BMD, univariate analysis demonstrated FI (P = 0.035) and WHR (P = 0.048) were associated with increased LS BMD, while mvMR analysis showed FI and T2D were associated with LS BMD, and HDL cholesterol was associated with decreased LS BMD (Table S12). While for heel BMD, the association signals for the rest factors in the univariate MR were mostly the same as the mvMR results, except that FG and HDL cholesterol becomes potential risk factors in mvMR estimates but not in the univariate MR estimates (Table S13).
In the pathway analysis, ten genes (GRB14, GCKR, RSPO3, LYPLAL1-AS, ARL15, LOC105369944, TET2, FTO, MAP3K19, and TCF7L2) for FInotFG were subjected to the pathway analysis, and the results demonstrated that genes (TCF7L2 and GRB14) were enriched in the “Wnt signaling pathway (P00057),” “Angiogenesis (P00005),” “Alzheimer disease-presenilin pathway (P00004),” and “Cadherin signaling pathway (P00012)” (Table S14).
Sensitivity analysis
For FN BMD trait, MLM and weighted median estimator both confirmed suggestive positive association between genetically predicted FG and FN BMD, and MLM also confirmed that suggestive positive association between genetically predicted T2D and FN BMD; however, MLM suggested inverse association between HCadjBMI and FN BMD, which is partially consistent with the main MR results (Fig. S1).
For FA BMD trait, MLM and weighted median both replicated the positive association between genetically predicted FI and FA BMD, and MLM also showed genetically increased risk of T2D was associated with increased FA BMD; however, we observed inverse association between genetically determined higher HCadjBMI, WCadjBMI, and FABMD, which is similar to the IVW results (Fig. S2).
For LS BMD trait, consistent with main MR approach, MLM and weighted median both detected positive association between FI and LS BMD, and MLM found that genetically increased risk of T2D was significantly associated with increased LS BMD; however, MLM illustrated inverse association between genetically predicted HDL cholesterol and LS BMD (Fig. S3).
For heel BMD trait, MLM and weighted median estimator both supported positive association between FI, BMI, HC, T2D, WC, WHR, and heel BMD; MLM showed WHRadjBMI and FG were positively associated with heel BMD. However, MLM and weighted median estimator demonstrated negative associations between HCadjBMI, LDL cholesterol, and heel BMD, and only the MLM sensitivity approach supports the IVW result that LDL cholesterol was negatively associated with heel BMD (Fig. S4).
Additionally, MR-Egger regression intercept estimate demonstrated that no horizontal pleiotropic exists in our selected IVs (Table S15-S18). And our Steiger test showed the inferred causal direction between exposures and each BMD trait was “TRUE.”
MR-BMA estimates
All the risk factors were then prioritized and ranked by their MIP. The top four risk factors for FN BMD were FI, T2D, FG, and HCadjBMI, which were confirmed to be included in the best four individual models (Table 2). For FA BMD, FI, HCadjBMI, LDL cholesterol, and WCadjBMI became the first four risk factors and were further included in the best five individual models (Table 3). The top five risk factors for LS BMD included FI, T2D, and HDL cholesterol, and they also achieved to be included in the best six individual models (Table 4). For heel BMD, all the significant risk factors in mvMR analysis achieved top 10 risk factors and were further included in the best ten individual models (Table 5). All those results from MR-BMA were consistent with that from mvMR results.
Table 2.
Ranking of risk factors for FN BMD. (A) According to their marginal inclusion probability (MIP). (B) The best ten individual models according to their posterior probability (PP)
| Risk factors | MIP | Model-averaged causal estimate | |
|---|---|---|---|
| (A) According to their MIP | |||
| 1 | FI | 0.32 | 0.08 |
| 2 | T2D | 0.265 | 0.012 |
| 3 | FG | 0.14 | 0.018 |
| 4 | HCadjBMI | 0.098 | − 0.009 |
| 5 | WCadjBMI | 0.067 | − 0.006 |
| 6 | BMI | 0.053 | 0.004 |
| 7 | WHR | 0.053 | 0.004 |
| 8 | WMRadjBMI | 0.038 | 0 |
| 9 | WC | 0.033 | 0 |
| 10 | LDL cholesterol | 0.032 | − 0.001 |
| Individual models | Risk factors combination | PP | Causal estimate |
| (B) The best 10 individual models according to their PP | |||
| 11 | FI | 0.251 | 0.253 |
| 1 | T2D | 0.21 | 0.047 |
| 10 | FG | 0.101 | 0.134 |
| 6 | HCadjBMI | 0.067 | − 0.096 |
| 8 | WCadjBMI | 0.036 | − 0.075 |
| 13 | BMI | 0.035 | 0.07 |
| 9 | WHR | 0.034 | 0.071 |
| 12 | WMRadjBMI | 0.023 | 0.023 |
| 2 | HDL cholesterol | 0.02 | − 0.039 |
| 3 | LDL cholesterol | 0.02 | − 0.042 |
Table 3.
Ranking of risk factors for FA BMD. (A) According to their MIP. (B) The best ten individual models according to their PP
| Risk factors combination | MIP | Model-averaged causal estimate | |
|---|---|---|---|
| (A) According to their MIP | |||
| 1 | FI | 0.571 | 0.308 |
| 2 | HCadjBMI | 0.239 | − 0.056 |
| 3 | LDL cholesterol | 0.095 | − 0.013 |
| 4 | WCadjBMI | 0.094 | − 0.017 |
| 5 | BMI | 0.079 | 0.015 |
| 6 | WHR | 0.06 | 0.007 |
| 7 | FG | 0.055 | 0.007 |
| 8 | HC | 0.055 | − 0.006 |
| 9 | WHRadjBMI | 0.049 | 0.001 |
| 10 | WC | 0.047 | − 0.003 |
| Individual models | Risk factors combination | PP | Causal estimate |
| (B) The best 10 individual models according to their PP | |||
| 11 | FI | 0.345 | 0.544 |
| 6 | HCadjBMI | 0.115 | − 0.239 |
| 6, 11 | HCadjBMI, FI | 0.05, − 0.215 | 0.496 |
| 3 | LDL cholesterol | 0.035 | − 0.133 |
| 8, 11 | WCadjBMI, FI | 0.03, − 0.205 | 0.59 |
| 8 | WCadjBMI | 0.029 | − 0.164 |
| 13 | BMI | 0.029 | 0.153 |
| 9 | WHR | 0.023 | 0.137 |
| 10 | FG | 0.021 | 0.124 |
| 3, 11 | LDL cholesterol, FI | 0.019, − 0.123 | 0.52 |
Table 4.
Ranking of risk factors for LS BMD. (A) According to their MIP. (B) The best ten individual models according to their PP
| Risk factors combination | MIP | Model-averaged causal estimate | |
|---|---|---|---|
| (A) According to their MIP | |||
| 1 | FI | 0.77 | 0.336 |
| 2 | FG | 0.1 | 0.017 |
| 3 | T2D | 0.083 | 0.004 |
| 4 | LDL cholesterol | 0.051 | − 0.004 |
| 5 | HDL cholesterol | 0.038 | − 0.002 |
| 6 | BMI | 0.036 | 0.003 |
| 7 | WHRadjBMI | 0.031 | − 0.002 |
| 8 | WHR | 0.029 | 0 |
| 9 | HC | 0.026 | 0.002 |
| 10 | WC | 0.025 | 0.001 |
| Individual models | Risk factors combination | PP | Causal estimate |
| (B) The best 10 individual models according to their PP | |||
| 11 | FI | 0.593 | 0.44 |
| 1 | T2D | 0.051 | 0.055 |
| 10 | FG | 0.041 | 0.18 |
| 10, 11 | FG, FI | 0.038, 0.154 | 0.413 |
| 3, 11 | LDL cholesterol, FI | 0.02, − 0.079 | 0.425 |
| 2 | HDL cholesterol | 0.018 | − 0.073 |
| 13 | BMI | 0.018 | 0.112 |
| 3 | LDL cholesterol | 0.017 | − 0.087 |
| 11, 12 | FI, WHRadjBMI | 0.015, 0.475 | − 0.088 |
| 1, 11 | T2D, FI | 0.014, 0.041 | 0.342 |
Table 5.
Ranking of risk factors for heel BMD. (A) According to their MIP. (B) The best ten individual models according to their PP
| Risk factors combination | MIP | Model-averaged causal estimate | |
|---|---|---|---|
| (A) According to their MIP | |||
| 1 | WHR | 0.811 | 0.248 |
| 2 | FI | 0.247 | 0.096 |
| 3 | HCadjBMI | 0.167 | − 0.027 |
| 4 | WCadjBMI | 0.056 | − 0.008 |
| 5 | WC | 0.046 | 0.012 |
| 6 | HC | 0.029 | − 0.009 |
| 7 | WHRadjBMI | 0.021 | − 0.001 |
| 8 | BMI | 0.011 | 0.001 |
| 9 | T2D | 0.005 | 0 |
| 10 | FG | 0.003 | 0 |
| Individual models | Risk factors combination | PP | Causal estimate |
| (B) The best 10 individual models according to their PP | |||
| 9 | WHR | 0.585 | 0.311 |
| 9, 11 | WHR, FI | 0.08, 0.234 | 0.294 |
| 6, 11 | HCadjBMI, FI | 0.075, − 0.166 | 0.439 |
| 11 | FI | 0.055 | 0.486 |
| 6, 9 | HCadjBMI, WHR | 0.05, − 0.128 | 0.26 |
| 8, 9 | WCadjBMI, WHR | 0.046, − 0.156 | 0.393 |
| 6, 7 | HCadjBMI, WC | 0.014, − 0.237 | 0.198 |
| 5, 7 | HC, WC | 0.012, − 0.444 | 0.556 |
| 6, 9, 11 | HCadjBMI, WHR, FI | 0.009, − 0.131, 0.18 | 0.301 |
| 11, 12 | FI, WHRadjBMI | 0.006, 0.412 | 0.162 |
Two-sample MR estimates of T2D on fracture
We obtained 118 LD-independent SNPs that achieved genome-wide significance for T2D. Then, those SNPs were extracted from the datasets for fracture. After data harmonization, there were 104 SNPs remained to perform the MR analysis for fracture, as shown in Table S19. Standard IVW MR results showed no association between T2D and fracture risk, and the estimates of weighted median and MR Egger were all consistent with the IVW result (Table S20). The intercept of MR Egger demonstrated there was no directional pleiotropy that exists in the genetic variables (P = 0.179). Because of the existence of heterogeneity among the selected SNPs (P = 0.0007), we present the IVW (random effect) results which still demonstrated no causal association.
Discussion
In the present study, by performing mvMR and MR-BMA analysis together using summary statistics for FN, FA, LS, and heel BMD and multiple risk factors, we successfully identified potential causal risk factors T2D, FG, and HCadjBMI for FN BMD; four risk factors FI, T2D, HCadjBMI, and WCadjBMI for FA BMD; and three risk factors FI, T2D, and HDL cholesterol for LS BMD, and all risk factors were causally associated with heel BMD except for triglycerides. Furthermore, we also prioritized and ranked those several risk factors for BMDs of different regions, which might provide us novel insights to determine the potential causal risk factors for complex traits and diseases.
Our results are consistent with previous two-sample MR results that suggest T2D and FG are causally related to FN BMD [21]; however, our current study did not implicate any causal effect role of T2D on fracture risk. Traditional observational studies reported poorly controlled T2D subjects tend to have higher risk for fracture, and the possible mechanisms underlying this relationship might be the poor bone quality in T2D patients because of decreased bone turnover, altered bone material properties, and bone microstructure [22]. Shanbhogue et al. stated that insulin resistance [23] was associated with greater volumetric BMD; our current study demonstrated the causally relationship of FI for FA, LS, and heel BMD. Previous studies detected the significantly correlation between HC and osteoporosis [24], and our mvMR results also suggested HCadjBMI were causally associated with FN and FA BMD. Recent MR analysis suggested that LDL cholesterol has a potential causal role on heel BMD but not HDL cholesterol and TG [25], and our univariate MR analysis also demonstrated the same result for LDL cholesterol and TG; however, the mvMR did identify HDL cholesterol as another potential causal risk factor for LS and heel BMD, which might be the mediated effects of other included risk factors.
As for the relationship between obesity/BMI and BMD, our analysis found that elevated BMI was associated with increasing heel BMD. The conclusions from previous studies are controversial. Some studies indicate that obesity/BMI may have a beneficial role in BMD [26, 27]; however, emerging contrasting studies suggest obesity/BMI may not have a protective effect on BMD [28-31]. Although, one two-sample MR study suggested that BMI/adiposity has been found to be causally related to increased BMD [32]; this conflicting study [32] was conducted in 5221 children with mean age of 9.9 years and was also performed on individual-level dataset rather than by using the summary statistic information from large datasets.
Lian-Hua Cui et al. [33] revealed that WC is independently and inversely associated with BMD (FN and LS) even after adjusting for age, weight, height, regular exercise, and percent body fat. Our mvMR estimates demonstrated that WCadjBMI was causally associated with FA and heel BMD. There might be several reasons why our analysis did not identify WC/WCadjBMI as potential causal risk factors for FN and LS BMD traits. First, they used traditional multiple linear regression analysis which could be biased by potential cofounding factors or reverse causality. Additionally, their analysis was performed on Korean population while our analysis was conducted in Europeans.
This study highlighted two genes (TCF7L2 and GRB14) were driving a link between FI and BMD phenotypes, and those two genes were enriched into “Wnt signaling pathway (P00057)” and “Cadherin signaling pathway.” Previous studies have frequently reported the role of Wnt signaling in bone metabolism and skeletal disorders [34, 35]. Furthermore, mice studies illustrated that N-cadherin regulation of bone growth and homeostasis is osteolineage stage-specific, and may therefore widen the therapeutic window of osteoanabolic agents [36].
There are several important strengths in the current study. Our MR analysis results may provide evidence of the causal role of T2D, FI, FG, HCadjBMI, WCadjBMI, and HDL cholesterol in the development of OP since the influence of traditional confounding factors in observational studies is minimized/eliminated. By leveraging the summary statistics from the largest available GWAS studies for multiple risk factors and BMD, we were able to increase our discovery power. Furthermore, previous studies have demonstrated that performing the MR analysis by using summary statistics data and by using individual-level data has similar efficiency [37]. Finally, the MR-BMA model [7] could select and prioritize the potential risk factors from a set of related candidate risk factors. Compared with the IVW approach, MR-BMA have the ability of detection of true risk factors and could produce reduced variance.
There may also be some limitations in our study. First, the datasets we used in our study only contain European population, and therefore the results may not be generalizable to other ancestries. Additionally, the BMD GWAS dataset released in 2015 [11] was adjusted for weight while the GWASs for the risk factors were not, which might introduce bias to our MR estimates; however, with the improved statistical power of mvMR, we assumed that this would not drive away the association signals.
Conclusions
In conclusion, by combining MR approaches, we identified potential causal risk factors FI, T2D, FG, HCadjBMI, WCadjBMI, and HDL cholesterol for BMD and we also prioritized and ranked the risk factors for BMDs of different regions, which might provide us novel insights into the causal mechanisms of OP. Our current study might provide us novel insights for treatment and intervention into bone-related complex traits and diseases.
Supplementary Material
Acknowledgments
QZ as the first author performed data analysis and wrote the manuscript. JG contributed suggestions for manuscript revision and revised the manuscript. HS, LJZ, WDZ, and CQS provided advice and suggestions while we met some problems during the data analysis process. HWD conceived and initiated this project, provided advice on experimental design, oversaw the implementation of the statistical method, and revised/finalized the manuscript.
Funding This research was partially supported by Key Science and Technology Development of Henan Province (Grant No.: 192102310191). We appreciate the support from Zhengzhou University in providing necessary support for this collaborative project. HWD was partially supported by grants from the NIH (R01-AR069055, U19-AG055373, R01-MH104680, R01-AR059781, and P20-GM109036) and Edward G. Schlieder Endowment fund at Tulane University.
Footnotes
Conflicts of interest None.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material The online version of this article (https://doi.org/10.1007/s00198-020-05640-5) contains supplementary material, which is available to authorized users.
Data availability
No original, unprocessed data was used in the present study. The summary datasets used in our study were derived from the following resources available in the public domain (detailed in Supplemental Data).
References
- 1.Xu XH, Dong SS, Guo Y, Yang TL, Lei SF, Papasian CJ, Zhao M, Deng HW (2010) Molecular genetic studies of gene identification for osteoporosis: the 2009 update. Endocr Rev 31(4):447–505 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Liu YZ, Liu YJ, Recker RR, Deng HW (2003) Molecular studies of identification of genes for osteoporosis: the 2002 update. J Endocrinol 177(2):147–196 [DOI] [PubMed] [Google Scholar]
- 3.Liu YJ, Zhang L, Papasian CJ, Deng HW (2014) Genome-wide association studies for osteoporosis: a 2013 update. J Bone Metab 21(2):99–116 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Smith GD, Ebrahim S (2003) ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol 32(1):1–22 [DOI] [PubMed] [Google Scholar]
- 5.Burgess S, Thompson SG (2015) Multivariable Mendelian randomization: the use of pleiotropic genetic variants to estimate causal effects. Am J Epidemiol 181(4):251–260 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Hemani G, Bowden J, Davey Smith G (2018) Evaluating the potential role of pleiotropy in Mendelian randomization studies. Hum Mol Genet 27(R2):R195–R208 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Zuber V, Colijn JM, Klaver C, Burgess S (2020) Selecting likely causal risk factors from high-throughput experiments using multivariable Mendelian randomization. Nat Commun 11 (1):29. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Lin H-H, Huang C-Y, Hwang L-C (2018) Association between metabolic syndrome and osteoporosis in Taiwanese middle-aged and elderly participants. Arch Osteoporos 13(1):48–48 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Wong SK, Chin K-Y, Suhaimi FH, Ahmad F, Ima-Nirwana S (2016) The relationship between metabolic syndrome and osteoporosis: a review. Nutrients. 8(6):347. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Zhou J, Zhang Q, Yuan X et al. (2013) Association between metabolic syndrome and osteoporosis: a meta-analysis. Bone 57(1):30–35 [DOI] [PubMed] [Google Scholar]
- 11.Zheng HF, Forgetta V, Hsu YH et al. (2015) Whole-genome sequencing identifies EN1 as a determinant of bone density and fracture. Nature 526(7571):112–117 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Morris JA, Kemp JP, Youlten SE et al. (2019) An atlas of genetic influences on osteoporosis in humans and mice. Nat Genet 51(2):258–266 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MAR, Bender D, Maller J, Sklar P, de Bakker PIW, Daly MJ, Sham PC (2007) PLINK: a tool set for whole-genome association and population-based linkage analyses. Am J Hum Genet 81(3):559–575 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Bowden J, Davey Smith G, Haycock PC, Burgess S (2016) Consistent estimation in Mendelian randomization with some invalid instruments using a weighted median estimator. Genet Epidemiol 40(4):304–314 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Hemani G, Tilling K, Davey SG (2017) Orienting the causal relationship between imprecisely measured traits using GWAS summary data. PLoS Genet 13(11):e1007081. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Bowden J, Davey Smith G, Haycock PC, Burgess S (2016) Consistent estimation in Mendelian randomization with some invalid instruments using a weighted median estimator. Genet Epidemiol 40(4):304–314 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Burgess S, Butterworth A, Thompson SG (2013) Mendelian randomization analysis with multiple genetic variants using summarized data. Genet Epidemiol 37(7):658–665 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Bowden J, Davey Smith G, Burgess S (2015) Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. Int J Epidemiol 44(2):512–525 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Xue A, Wu Y, Zhu Z et al. (2018) Genome-wide association analyses identify 143 risk variants and putative regulatory mechanisms for type 2 diabetes. Nat Commun 9(1):2941. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Trajanoska K, Morris JA, Oei L et al. (2018) Assessment of the genetic and clinical determinants of fracture risk: genome wide association and Mendelian randomisation study. BMJ 362:k3225. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Ahmad OS, Leong A, Miller JA, Morris JA, Forgetta V, Mujammami M, Richards JB (May 2017) A Mendelian randomization study of the effect of type-2 diabetes and glycemic traits on bone mineral density. J Bone Miner Res 32(5):1072–1081 [DOI] [PubMed] [Google Scholar]
- 22.Napoli N, Chandran M, Pierroz DD et al. (2016) Mechanisms of diabetes mellitus-induced bone fragility. Nat Rev Endocrinol 13:208. [DOI] [PubMed] [Google Scholar]
- 23.Shanbhogue VV, Finkelstein JS, Bouxsein ML, Yu EW (2016) Association between insulin resistance and bone structure in non-diabetic postmenopausal women. J Clin Endocrinol Metab 101(8):3114–3122 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Aghaei Meybodi H, Hemmat-Abadi M, Heshmat R, Rezaei Homami M, Madani S, Ebrahimi M, Adibi H, Larijani B (2011) Association between anthropometric measures and bone mineral density: population-based study. Iran J Public Health 40(2):18–24 [PMC free article] [PubMed] [Google Scholar]
- 25.Zheng J, Brion MJ, Kemp JP, Warrington NM, Borges MC, Hemani G, Richardson TG, Rasheed H, Qiao Z, Haycock P, Ala-Korpela M, Davey Smith G, Tobias JH, Evans DM (2020) The effect of plasma lipids and lipid-lowering interventions on bone mineral density: a Mendelian randomization study. J Bone Miner Res 35:1224–1235 [DOI] [PubMed] [Google Scholar]
- 26.Felson DT, Zhang YQ, Hannan MT, Anderson JJ (1993) Effects of weight and body-mass index on bone-mineral density in men and women - the Framingham study. J Bone Miner Res 8(5):567–573 [DOI] [PubMed] [Google Scholar]
- 27.Edelstein SL, Barrett-Connor E (1993) Relation between body size and bone mineral density in elderly men and women. Am J Epidemiol 138(3):160–169 [DOI] [PubMed] [Google Scholar]
- 28.Zhao L-J, Liu Y-J, Liu P-Y, Hamilton J, Recker RR, Deng H-W (2007) Relationship of obesity with osteoporosis. J Clin Endocrinol Metab 92(5):1640–1646 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Zhao LJ, Jiang H, Papasian CJ, Maulik D, Drees B, Hamilton J, Deng HW (2008) Correlation of obesity and osteoporosis: effect of fat mass on the determination of osteoporosis. J Bone Miner Res 23(1):17–29 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Sharma S, Tandon VR, Mahajan S, Mahajan V, Mahajan A (2014) Obesity: friend or foe for osteoporosis. J Mid-Life Health 5(1):6–9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Janicka A, Wren TA, Sanchez MM et al. (2007) Fat mass is not beneficial to bone in adolescents and young adults. J Clin Endocrinol Metab 92(1):143–147 [DOI] [PubMed] [Google Scholar]
- 32.Kemp JP, Sayers A, Smith GD, Tobias JH, Evans DM (2016) Using Mendelian randomization to investigate a possible causal relationship between adiposity and increased bone mineral density at different skeletal sites in children. Int J Epidemiol 45(5):1560–1572 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Cui L-H, Shin M-H, Kweon S-S, Choi JS, Rhee JA, Lee YH, Nam HS, Jeong SK, Park KS, Ryu SY, Choi SW (2014) Sex-related differences in the association between waist circumference and bone mineral density in a Korean population. BMC Musculoskelet Disord 15:326–326 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Maeda K, Kobayashi Y, Koide M, Uehara S, Okamoto M, Ishihara A, Kayama T, Saito M, Marumo K (2019) The regulation of bone metabolism and disorders by Wnt signaling. Int J Mol Sci 20(22):5525. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Baron R, Gori F (2018) Targeting WNT signaling in the treatment of osteoporosis. Curr Opin Pharmacol 40:134–141 [DOI] [PubMed] [Google Scholar]
- 36.Fontana F, Hickman-Brecks CL, Salazar VS, Revollo L, Abou-Ezzi G, Grimston SK, Jeong SY, Watkins M, Fortunato M, Alippe Y, Link DC, Mbalaviele G, Civitelli R (2017) N-cadherin regulation of bone growth and homeostasis is osteolineage stage-specific. J Bone Miner Res 32(6):1332–1342 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Burgess S, Butterworth A, Thompson SG (2013) Mendelian randomization analysis with multiple genetic variants using summarized data. Genet Epidemiol 37(7):658–665 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
No original, unprocessed data was used in the present study. The summary datasets used in our study were derived from the following resources available in the public domain (detailed in Supplemental Data).