Clinical risk factors (CRFs), with or without bone mineral density (BMD), are used to determine the risk of osteoporotic fracture (OF), which has a heritable component. In this study we investigated whether genetic profiling can additionally improve the ability to predict OF. Using 1229 unrelated Korean postmenopausal women, 39 single-nucleotide polymorphisms (SNPs) in 30 human genomic loci were tested for association with osteoporosis-related traits, such as BMD, osteoporosis, vertebral fracture (VF), nonvertebral fracture (NVF), and any fracture. To estimate the effects of genetic profiling, the genetic risk score (GRS) was calculated using five prediction models: (Model I) GRSs only; (Model II) BMD only; (Model III) CRFs only; (Model IV) CRFs and BMD; and (Model V) CRFs, BMD, and GRS. A total of 21 SNPs within 19 genes associated with one or more osteoporosis-related traits and were included for GRS calculation. GRS associated with BMD before and after adjustment for CRFs (p ranging from <0.001 to 0.018). GRS associated with NVF before and after adjustment for CRFs and BMD (p ranging from 0.017 to 0.045), and with any fracture after adjustment for CRFs and femur neck BMD (p = 0.049). In terms of predicting NVF, the area under the receiver operating characteristic curve (AUC) for Model I was 0.55, which was lower than the AUCs of Models II (0.60), III (0.64), and IV (0.65). Adding GRS to Model IV (in Model V) increased the AUC to 0.67, and improved the accuracy of NVF classification by 11.5% (p = 0.014). In terms of predicting any fracture, the AUC of Model V (0.68) was similar to that of Model IV (0.68), and Model V did not significantly improve the accuracy of any fracture classification (p = 0.39). Thus, genetic profiling may enhance the accuracy of NVF predictions and help to delineate the intervention threshold. © 2013 American Society for Bone and Mineral Research.
Osteoporosis is a common disease characterized by low bone mass and defects in the microarchitecture of bone tissue, which impairs bone strength and leads to an increased risk of fracture. Osteoporotic fracture (OF) is one of the leading causes of significant morbidity and disability in old people and places a substantial economic burden on health care systems. In Korea, the residual lifetime probabilities of OF at the age of 50 years are 60% for women and 24% for men.
Bone mineral density (BMD) measured by dual-energy X-ray absorptiometry (DXA), which is often used to measure bone mass, is often employed to diagnose osteoporosis because bone mass accounts for approximately 70% of bone strength. Although the risk of OF increases as BMD values fall, about two-thirds of individuals who suffer a fracture do not have osteoporosis as defined on the basis of BMD values. Therefore, several fracture prediction models based on clinical risk factors (CRFs), with or without BMD data, have been developed to identify individuals who are at high risk of fracture.[5-7] However, because the prediction accuracies of these models are less than perfect,[6-8] their prediction accuracy needs to be improved further.
Although fracture risk is multifactorial, one of the most important CRFs is a positive family history. This indicates that genetic factors play a significant role in the pathogenesis of osteoporosis and/or OF. Moreover, genetic factors also account for a large proportion of the variance in OF risk factors such as BMD, bone loss, quantitative ultrasound properties of bone, bone turnover markers, and composite bone phenotypes derived from geometrical variables. Thus, it may be useful to include genetic factors when seeking to predict OF risk. However, several studies have suggested that single genes or single-nucleotide polymorphisms (SNPs) contribute only minimally to OF prognosis because their effect size is so small.[10, 11] This suggests that combining the effects of many SNPs might improve the ability to predict the risks of osteoporosis and OF. Although the clinical usefulness of genetic profiling remains controversial, a recent study in European populations showed that simulated integration of the effects of many genes into current prognostic models may significantly improve their ability to predict the fracture risk of a given individual.
However, variations in the underlying linkage disequilibrium (LD) structure and in the frequency of sequence variants are observed between European and Asian populations, and influence the results of association studies. A recent study showing the effects of genetic profiling was performed in European populations. Therefore, variations between populations of different ancestries might carry different effects of genetic profiling. In addition, the recent study was only a simulated design rather than a real genotyping study. The present study investigated whether genetic profiling improves the ability to predict the risk of OF in Koreans.
Subjects and Methods
The study population consisted of Korean postmenopausal women (n = 1229) who visited the Asan Medical Center (AMC) in Seoul. The study was approved by the AMC ethics review committee and written informed consent was obtained from all subjects. Menopause was defined as the absence of menstruation for at least 1 year and was confirmed by measuring the serum follicle-stimulating hormone levels. Information regarding the histories of smoking (current smoker), alcohol intake (≥3 units/d), medication, previous medical or surgical diseases, and reproduction, family history of fracture, and history of any fracture was obtained by providing a self-administered questionnaire.
Women exhibiting premature menopause (<40 years of age), and subjects who had taken drugs that could affect bone metabolism (e.g., glucocorticoid, sex hormone, bisphosphonate, or other treatments for osteoporosis) for more than 6 months or who had taken these drugs within the previous 12 months were excluded from the study. In addition, subjects were excluded from the study if they had experienced any disease that might affect bone metabolism, such as diabetes, cancer, hyperparathyroidism, or rheumatoid arthritis. Women who had suffered a stroke or who exhibited dementia were also excluded because of concerns relating to their limited physical activity. Women were also excluded if they had osteophyte formation above the fourth grade of the Nathan classification system and/or severe facet joint osteoarthritis in the lumbar spine, as determined by conventional spine radiographs.
BMD measurement and fracture assessment
In 784 women, areal BMD (g/cm) was measured at the lumbar spine (L2–L4) and femur neck by dual-energy X-ray absorptiometry using Lunar equipment (Expert XL; Lunar Corp., Madison, WI, USA). In the remaining 445 women, BMD was estimated using Hologic equipment (QDR 4500-A; Hologic, Waltham, MA, USA). As to precision of the Lunar and Hologic equipment, variation coefficients of their measurements were 0.82% and 0.85% for the lumbar spine, and 1.1% and 1.2% for the femur neck, respectively. These values were obtained by scanning 17 volunteers who were not part of the study. Each volunteer underwent five scans on the same day, getting on and off the table between examinations. To derive cross-calibration equations between the two systems, BMD values were measured in 109 Korean healthy women (mean ± SD, age 55 ± 11 years; range, 31–75 years) using the two machines, and cross-calibration equations were calculated as follows:
- LS BMD (g/cm2): Lunar = 1.1287 × Hologic – 0.0027
- FN BMD (g/cm2): Lunar = 1.1556 × Hologic – 0.0182
According to the World Health Organization (WHO) definition, osteoporosis was defined as the lowest T-score ≤ –2.5 SDs at the lumbar spine (LS) or femur neck (FN).
The prevalence of morphological vertebral fracture (VF) in all study subjects was determined by obtaining lateral thoracolumbar (T4–L4) radiographs. The assessment of VF was made in accordance with the recommendations of the Working Group on Vertebral Fractures. A VF was defined quantitatively as a more than 20% reduction in any of the measured vertebral heights (i.e., anterior, middle, or posterior). Nonvertebral fractures (NVFs), namely those at the wrist, hip, forearm, humerus, rib, and pelvis, were assessed by applying a self-administered questionnaire. Fractures clearly caused by major trauma (such as motor vehicle accidents or falls from higher than standing height) were excluded. Thus, fractures were only included if the report of fracture was definite and an interview confirmed that they had occurred with low trauma (e.g., fall from standing height or less) after the age of 50 years or after menopause.
Selection of markers for genotyping and association analyses
In total, 39 SNPs in 30 genomic loci listed in Supplementary Table S1 were tested: (1) group I: eight SNPS in six loci (CALCR, HSD11B1, IL15, ITGA1, SEMA7A, and TNFRSF11A) that were identified in our previous candidate gene association studies[21-26]; (2) group II: 13 SNPs in 11 loci (CTNNB1, ESR1, FLJ42280, GPR177, MEF2C, MEPE, SOST, SOX6, TNFSF11, TNFRSF11B, and ZBTB40) that associated with BMD in both European and East-Asian populations (including Koreans); (3) group III: nine SNPs in eight loci (ADAMTS18, JAG1, LRP4, LRP5, MARK3, MHC, SPTBN1, and TNFRSF11A) that associated with both BMD and fracture in genome-wide association studies[11, 27-31]; and (4) group IV: nine SNPs in eight loci (ARHGAP1, C6orf97, CRHR1, DCDC5, ESR1, SP7, TNFRSF11A, and TNFRSF11B) that associated with BMD in the genome-wide association studies.[11, 30] Because the 18 SNPs in 15 loci (ADAMTS18, ARHGAP1, C6orf97, CRHR1, DCDC5, ESR1, JAG1, LRP4, LRP5, MARK3, MHC, SP7, SPTBN1, TNFRSF11A, and TNFRSF11B) of group III and IV have not been previously evaluated for any association with osteoporosis-related traits in a Korean population, their associations were tested in this study.
For genotyping purposes, all DNA samples were diluted to a concentration between 2.5 and 10 ng/µL. Each DNA sample was genotyped for all 39 SNPs using the iPLEX assay on the Sequenom MassARRAY platform (Sequenom, San Diego, CA, USA), according to the manufacturer's instructions. All SNPs achieved call rates of >90%.
Selection of markers for generating a genetic risk score
The risk alleles that associated with one or more osteoporosis-related traits were selected. A risk allele was defined to associate with decreased BMD or increased risk of osteoporosis, VF, NVF, or any fracture with marginal significance, as the p values were lower than 0.05 but higher than a significance level of Bonferroni correction for multiple testing α = 0.05/39 SNPs/3 genetic models = 0.00043. For each individual, each gene was coded as 2 when two risk alleles were present, 1 with one risk allele, or 0 with no risk allele. To estimate the cumulative effect of the genetic markers, all scores were summed to yield the genetic risk score (GRS).
All data are presented as means ± SD, or as numbers and percentages. Multiple regression analyses of BMD at the LS and FN with GRS were performed using age, weight, height, current smoker, alcohol intake (≥3 units/d), and family history of fracture as covariates. The participants with and without osteoporosis were analyzed in terms of their GRS using logistic regression with adjustment for age, weight, height, current smoker, alcohol intake (≥3 units/d), and family history of fracture. The participants with and without fractures (nonvertebral, vertebral, or any) were analyzed in terms of the GRS using logistic regression with adjustment for age, weight, height, current smoker, alcohol intake (≥3 units/d), and family history of fracture.
The aim was to determine whether genetic profiling can improve the ability to predict the risk of OF. To address this, five prediction models were considered, as follows:
- Model I included only the GRS data.
- Model II included only the FN BMD data, which are included in most fracture risk prediction models.[5-7] LS BMD data were not included because of technical limitations in determining them such as incomplete exclusion of osteoarthritis, degenerative changes (bony spurs), or aortic calcification.
- Model III included only the CRFs that have been shown to associate with fracture risk. These CRFs included age, height, weight, current smoker, alcohol intake, and family history of fracture.
- Model IV included the CRFs and BMD data.
- Model V included the CRF, BMD, and GRS data.
The ability of the GRS to predict osteoporosis and fracture was quantified using the area under the receiver operating characteristic (ROC) curve (AUC). Because the AUC was found to be too insensitive to addition of markers in several recent studies, a reclassification analysis was proposed as a new way to measure the usefulness of adding markers in prediction models. Therefore, to assess how much genetic profiling would improve the ability to predict osteoporosis and fracture, a reclassification analysis was performed.
In this reclassification analysis, the probability of events (VF, NVF, or any fracture) was estimated for each individual using Model IV (CRFs + BMD) and Model V (CRFs + BMD + GRS). We chose the values of 10% and 15% of the probability of events as the cut-points for this analysis, the former as a diagnostic threshold and the latter as a treatment threshold, as previously suggested in Japan.[36, 37] The individuals were then classified into three risk groups: those with a less than 10% risk, those with a risk between 10% and 15%, and those with a risk exceeding 15%. The proportion of individuals who would be reclassified into the three risk groups for Model IV (CRFs + BMD) and Model V (CRFs + BMD + GRS) was calculated.
Thus, if GRS is useful for events prediction, the probability of events estimated by the model with GRS (Model V) would be increased for the event group and decreased for the no-event group compared to the model without GRS (Model IV). This predictive improvement was quantified as computing the net reclassification improvement (NRI) index. The NRI was computed as follows:
NRI = (probability of up in the event group) – (probability of down in the event group) – (probability of up in the no-event group) + (probability of down in the no-event group).
The statistical significance of NRI was estimated by Z-test, a simple asymptotic test for the null hypothesis of the NRI index = 0. If the net gain in reclassification proportion is significantly greater than zero, there is a statistical significance. All statistical analyses were performed using SPSS statistical software (SPSS Inc., Chicago, IL, USA), and p ≤ 0.05 was considered to be statistically significant.
The characteristics of the study subjects are shown in Table 1. Among the 1229 postmenopausal women, osteoporosis was noted in 383 and any fractures (i.e., VFs and NVFs combined) were noted in 188. Of the 188 subjects with OFs, 96 exhibited VFs and 105 exhibited NVFs. Many SNPs listed in Supplementary Tables S2 and S3 associated with one or more osteoporosis-related traits such as BMD (12 SNPs in 11 loci), osteoporosis (seven SNPs in six loci), VF (four SNPs in four loci), NVF (seven SNPs in six loci), and any fracture (six SNPs in six loci) in this study using the Korean postmenopausal women. In summary, a total of 22 SNPs in 19 genomic loci, ARHGAP1, C6orf97, CALCR, CTNNB1, DCDC5, GPR177, HSD11B1, IL15, ITGA1, LRP5, MEF2C, SEMA7A, SOST, SOX6, SPTBN1, TNFSF11 (alternatively known as RANKL), TNFRSF11A (RANK), TNFRSF11B (OPG), and ZBTB40 associated with one or more osteoporosis-related traits.
|Age (years)||58.2 ± 7.1|
|Height (cm)||155.1 ± 5.9|
|Weight (kg)||56.5 ± 7.0|
|Years since menopause (years)||8.6 ± 7.3|
|Current smoker, n (%)||23 (1.0)|
|Alcohol intake with ≥3 units/day, n (%)||6 (0.5)|
|BMD at lumbar spine (g/cm2)||0.881 ± 0.014|
|BMD at femur neck (g/cm2)||0.714 ± 0.014|
|Osteoporosis, n (%)||383 (31.2)|
|Family history of fracture, n (%)||159 (12.8)|
|Any fracture, n (%)||188 (15.3)|
|Vertebral fracture||96 (7.8)|
|Nonvertebral fracture||105 (8.5)|
Each locus association was represented by one SNP, except for HSD11B1, IL15, and TNFRSF11A, which were represented by two SNPs. The two SNPs in HSD11B1 were correlated little with each other (r2 = 0.020 between rs701950 and rs1000283) as were the two in TNFRSF11A (r2 = 0.050 between rs884205 and rs3018362). However, the two in IL15 were highly correlated with each other (r2 = 0.94 between rs1057972 and rs2857261), and so one of them (rs1057972) was excluded from calculation of the GRS to eliminate redundancy of IL15 risk alleles. Finally, only 21 SNPs in 19 loci listed in Supplementary Table S4 were included in the GRS calculation.
The distribution of simulated GRS had a median of 21 (range, 9–30) (Fig. 1). Odds ratios of the risk alleles for any fracture did not vary much (range, 1.00–1.31; Supplementary Table S4) with all 95% confidence intervals (CIs) overlapping each other, and so the contribution of each SNP to the GRS was regarded as equal. Thus, the GRS was not calculated using each SNP's weighted-score. Table 2 shows how the GRS associated with BMD and osteoporosis. A higher GRS associated significantly with decreased BMD at the lumbar spine (p = 0.018) and at the femur neck (p < 0.001). After adjustment for the CRFs, higher GRS still associated significantly with decreased BMD at the lumbar spine (p = 0.0010) and at the femur neck (p < 0.001). GRS associated significantly with the risk of osteoporosis before and after the adjustments for CRFs (p = 0.048 and 0.0054, respectively).
|Variable||Lumbar spine BMD||Femur neck BMD||Osteoporosis|
|Model I β (p)||Model III β (p)||Model I β (p)||Model III β (p)||Model I OR, 95% CI (p)||Model III OR, 95% CI (p)|
|GRS||−0.004 (0.018)||−0.005 (0.0010)||−0.005 (<0.001)||−0.006 (<0.001)||1.04, 1.00–1.09 (0.048)||1.07, 1.02–1.12 (0.0054)|
|Age (years)||−0.007 (<0.001)||−0.006 (<0.001)||1.10, 1.08–1.12 (<0.001)|
|Height (cm)||0.001 (0.40)||0.000 (0.53)||0.99, 0.97–1.02 (0.50)|
|Weight (kg)||0.005 (<0.001)||0.004 (<0.001)||0.94, 0.92–0.96 (<0.001)|
|Current smoker||−0.015 (0.64)||−0.042 (0.11)||1.12, 0.44–2.91 (0.81)|
|Alcohol intake||−0.099 (0.11)||−0.085 (0.086)||1.49, 0.24–9.44 (0.67)|
|Family history of fracture||−0.012 (0.44)||−0.013 (0.22)||1.33, 0.91–1.96 (0.15)|
Table 3 shows how GRS associated with VF, NVF, and any fracture. GRS did not associate with the risk of VF either before or after the adjustment but did associate with the risk of NVF before adjustment (p = 0.041). The adjustment for CRFs strengthened the statistical significance of this association (p = 0.045), and this was further improved by adjustment for both CRFs and BMD (p = 0.017). Although GRS did not associate with the risk of any fracture before adjustment, the adjustment for CRFs and BMD made this association statistically significant (p = 0.049).
|Variable||Model I||Model III||Model IV|
|Association with vertebral fracture|
|GRS||1.01, 0.94–1.09 (0.71)||1.03, 0.96–1.11 (0.41)||1.03, 0.95–1.11 (0.51)|
|Age (years)||1.10, 1.06–1.13 (<0.001)||1.12, 1.08–1.16 (<0.001)|
|Height (cm)||0.99, 0.96–1.03 (0.66)||0.99, 0.95–1.02 (0.47)|
|Weight (kg)||1.01, 0.97–1.04 (0.74)||1.00, 0.96–1.03 (0.83)|
|Family history of fracture||0.87, 0.42–1.79 (0.70)||0.92, 0.44–1.93 (1.0)|
|Femur neck BMD (g/cm2)||21.5, 2.9–192.9 (0.0060)|
|Association with nonvertebral fracture|
|GRS||1.06, 1.00–1.14 (0.041)||1.08, 1.00–1.16 (0.045)||1.09, 1.02–1.18 (0.017)|
|Age (years)||1.06, 1.03–1.09 (<0.001)||1.06, 1.03–1.10 (<0.001)|
|Height (cm)||1.04, 1.00–1.09 (0.067)||1.04, 1.00–1.09 (0.080)|
|Weight (kg)||1.00, 0.97–1.03 (0.94)||1.00, 0.97–1.03 (0.91)|
|Family history of fracture||1.65, 0.95–2.86 (0.076)||1.68, 0.96–2.93 (0.067)|
|Femur neck BMD (g/cm2)||1.09, 0.12–9.55 (0.94)|
|Association with any fracture|
|GRS||1.03, 0.98–1.09 (0.23)||1.05, 0.99–1.11 (0.082)||1.06, 1.00–1.11 (0.049)|
|Age (years)||1.09, 1.06–1.11 (<0.001)||1.10, 1.07–1.13 (<0.001)|
|Height (cm)||1.02, 0.99–1.05 (0.28)||1.02, 0.98–1.05 (0.31)|
|Weight (kg)||1.00, 0.98–1.03 (1.0)||0.99, 0.97–1.02 (0.67)|
|Family history of fracture||1.29, 0.80–2.07 (0.30)||1.33, 0.82–2.15 (0.25)|
|Femur neck BMD (g/cm2)||4.81, 0.90–25.68 (0.066)|
Table 4 shows the AUCs of the models for NVF and any fracture. The AUC of the model for NVF with GRS only was 0.550. The AUCs of the models for NVF with BMD data only (Model II) and CRFs only (Model III) were 0.597 and 0.643, respectively. The AUC of the model with both CRFs and BMD (Model IV) was 0.653. When the GRS was added to the latter model for NVF (Model V), the AUC increased to 0.673. Similarly, when the GRS was added to Model IV for any fracture (Model V), the AUC increased slightly to 0.683 relative to the AUC of Model IV (0.679).
|Model||Nonvertebral fracture||Any fracture|
|AUC||Improvement over Model III||AUC||Improvement over Model III|
|I. Genetic risk score (GRS)||0.550||0.528|
|II. Bone mineral density (BMD)||0.597||0.548|
|III. Clinical risk factors (CRFs)||0.643||Ref||0.678||Ref|
|IV. CRFs + BMD||0.653||0.010||0.679||0.001|
|V. CRFs + BMD + GRS||0.673||0.030||0.683||0.005|
Table 5 shows the reclassification analysis that revealed the effects of adding GRS in Model IV (Model V). For NVF, compared to the CRFs and BMD model (Model IV), the addition of GRS (Model V) reclassified 18 of 105 individuals with NVF (17.1%) into higher risk groups but only reclassified 5 of the 105 individuals with NVF (4.8%) into lower risk groups. Thus, there was a gain of 12.3% (17.1% minus 4.8%). For the no NVF group, Model V for NVF compared to Model IV for NVF reclassified 100 of the 1124 no NVF individuals (8.9%) into a higher risk group, while reclassifying 91 of the 1124 no NVF individuals (8.1%) into a lower risk group. Thus, there was a loss of 0.8% (8.1% minus 8.9%). In total, there was a net gain of 11.5% by Model V for NVF, and it was statistically significant (p = 0.014).
|Risk group in model IV||Risk group in model V||Up (model V > model IV)||Down (model V < model IV)||Reclassified (%)||Net reclassification improvement (%)|
|Subjects with nonvertebral fracture||18 (17.1%)||5 (4.8%)||12.3||11.5 (p = 0.014)|
|Subjects without nonvertebral fracture||100 (8.9%)||91 (8.1%)||−0.8|
|Subjects with any fracture||9 (4.8%)||9 (4.8%)||0.0||2.2 (p = 0.39)|
|Subjects without any fracture||69 (6.6%)||92 (8.8%)||2.2|
For any fracture, compared to the CRFs and BMD model (Model IV), the model with GRS added (Model V) reclassified 9 of the 188 individuals with fractures (4.8%) into higher risk groups but only reclassified 9 of the 188 individuals with fractures (4.8%) into lower risk groups, which meant no gain (0.0% = 4.8% minus 4.8%). For the not any fracture group, Model V for any fracture compared to Model IV for any fracture reclassified 92 of the 1041 no NVF individuals (8.8%) into lower risk groups, while 69 of the 1041 no NVF individuals (6.6%) were reclassified into higher risk groups. Thus, there was a gain of 2.2% (8.8% minus 6.6%). In total, there was a net gain of 2.2%, but it was not statistically significant (p = 0.39).
The present study showed that 21 risk alleles of 19 genomic loci associated with one or more osteoporosis-related traits. Genetic profiling was used to calculate the GRS to estimate the cumulative effect of the 21 risk alleles associated with BMD and the risk of osteoporosis in postmenopausal women. The GRS also associated with the risk of NVF and any fracture in these women independently of the CRFs and BMD. Furthermore, when GRS was added to current fracture risk assessment models that employed CRFs with or without BMD, their ability to predict the risk of NVF was indeed improved.
OF is the clinical endpoint and the single most important manifestation of osteoporosis. Therefore, risk factors for OF may help to identify individuals who are at high risk of OF. Their use may also heighten the clinical awareness of osteoporosis and aid the development of strategies to treat osteoporosis and prevent OF. BMD is an important predictor of future fracture risk. For every SD decrease in age-adjusted BMD, the overall fracture risk increases by approximately twofold (range, 1.6-fold to 2.6-fold).[38, 39] Therefore, the majority of the previous clinical guidelines that aim to prevent OF have been intervention recommendations that are based predominantly on the T-score. However, BMD tests alone are not optimal for the detection of individuals at high risk of OF because epidemiological studies and clinical experience have shown that OF can occur in patients with any given T-score, even in individuals with normal BMD values. Given these limitations in BMD, several models that use CRFs with or without BMD had been proposed to express the fracture risk for clinical assessment and to determine interventional thresholds.[5-7] The ability of these fracture risk assessment models to discriminate fracture from nonfracture cases has been modest.[6-8] Thus, the ability of these models to predict fractures needs to be improved further.
Several studies have shown that BMD and other determinants of OF risk (such as ultrasound properties of bone, skeletal geometry, and bone turnover) are highly heritable. In keeping with this, several investigators have reported that OF also has a heritable component.[40, 41] Therefore, genetic profiling might help to improve fracture risk assessment. The genetic architecture of OF is typical of a complex disease; namely, there are contributions from many genes, which individually show a small effect. This suggests that combining the effects of many SNPs could improve the ability to predict the risk of OF. The reclassification analysis in the present study showed that integrating genetic profiling in the form of GRS into the prediction models that use CRFs and BMD significantly improved the accuracy with which they predicted the risk of NVF for a given individual. These results are consistent with those of a previous study that showed simulated genetic profiling could enhance the predictive accuracy of current prognostic models in Europeans.
In addition, we included the family history of fracture as a CRF for fracture in the present study. The family history of fracture might already contain some information from genetic profiling. However, GRS associated with the risk of NVF and any fracture regardless of whether the family history of fracture was included or excluded (data not shown). This suggests that the family history may not be sufficient to fully reflect the genetic heritability for fracture, and that genetic profiling might provide more information beyond the family history of fracture.
The present study showed that the GRS associated with BMD at the lumbar spine and femur neck both before and after adjustment for age, height, weight, current smoker, alcohol intake (≥3 units/d), and family history of fracture. GRS also associated with an increased risk of osteoporosis before and after adjustment for CRFs. These results are consistent with those of a previous genome-wide meta-analysis showing that women in the highest risk score bin had 1.6-fold increased odds (95% CI, 1.1–2.2) of osteoporosis, compared to those in the middle risk bin. Furthermore, GRS itself associated with NVF risk and tended to associate with risk of any fracture after adjustment for CRFs. After the adjustment for CRFs and BMD, GRS associated with the risk of NVF and any fracture. These results were also consistent with those of previous studies, which showed that the heritable component of fracture seems to be independent of BMD, at least in part.[40, 41]
Despite the robust proof of the relationship between GRS and the risk of NVF and any fracture, ROC analysis showed that GRS alone had a relatively small discrimination ability (AUC = 0.55 for NVF prediction ability), although it did achieve statistical significance (p = 0.045). This is similar to the ability measured by a previous genome-wide meta-analysis (AUC = 0.57 for OF prediction ability). Adding GRS to a model with CRFs and BMD (AUC = 0.65) did not substantially increase its discrimination ability (AUC = 0.67). This was also observed in a previous genome-wide meta-analysis. Like our study and the previous genome-wide meta-analysis, several other studies have shown that ROC analysis is insensitive to addition of genes, as this approach did not appreciably increase the accuracy with which type 2 diabetes and cardiovascular diseases[43, 44] could be predicted. Therefore, to better quantify how much GRS addition improves the ability of a model to predict OF, we performed reclassification analysis. This reclassification analysis showed that adding GRS to a model with CRFs and BMD data improved the accuracy of NVF classification by 11.5%. This suggests that the integration of genetic profiling into the current prediction models could significantly improve the accuracy with which they predict the fracture risk of an individual. To our knowledge, this is the first study showing that genetic profiling with real genotyping results improves the ability to predict NVF.
We modeled femur neck BMD and GRS (including BMD-associated loci) together, although there could be interaction between them. In fact, weak correlation existed between the two variables (γ = − 0.11). When logistic regression analysis was performed using CRFs, femur neck BMD, GRS, and a BMD-GRS interaction term, however, the interaction was not significant for NVF risk (p = 0.31) and for any fracture risk (p = 0.88). Therefore, the interaction between BMD and GRS could be neglected in this dataset.
The present study had several limitations. First, not all of the SNPs identified by previous studies [10, 11, 14, 21-27, 29, 31] were genotyped. Instead, several were randomly selected. Second, neither the ROC analysis nor the reclassification analysis showed that adding GRS to a model with CRFs and BMD improved the accuracy with which VF or any fracture could be predicted. This may reflect the fact fewer risk alleles associated with BMD at the lumbar spine (n = 5) than BMD at the femur neck (n = 9). Also, fewer risk alleles associated with VF (n = 4) than with NVF (n = 6). Another explanation may be that too few risk alleles were tested: a previous study has suggested that a profile of up to 25 genes in the presence of CRFs, with or without BMD, was required to improve the accuracy of the model.
Third, this was a cross-sectional study but not a longitudinal study. A cross-sectional study without a detailed timeframe of fractures can only access the relative risk or the prevalence of the fractures rather than creating a real prediction model. Therefore, further prospective studies in an independent validation setting with a large number of patients are needed.
Fourth, the models that we considered here were based on the assumption that the effects of all genes are totally independent. Considering the complex phenotypes of OF, it is likely that there are gene-to-gene interactions. However, it is quite challenging to identify these interactions because the current linear statistical genetic methods used for analyzing and detecting gene-to-phenotype associations in human populations are not sensitive enough to detect the nonlinear interaction effects that are the result of the combinatorial complexity of gene-to-gene interactions.
Finally, histories of previous fracture, glucocorticoids use, rheumatoid arthritis, and secondary osteoporosis were not included as CRFs for fracture in this study. These might reduce the power of fracture prediction of CRFs.
In summary, this study demonstrated that genetic profiling could enhance the accuracy with which current models can predict NVF and could therefore help to identify high risk individuals for appropriate risk management or intervention.
All authors state that they have no conflicts of interest.
This work was supported by grants from the National Project for Personalized Genomic Medicine, Korean Ministry of Health and Welfare (A111218-GM03 to J-MK) and National Research Foundation of Korea (2011-0017999 and 2011-0020334 to CK). We thank Seon-Ok Kim at the Department of Clinical Epidemiology and Biostatistics, Asan Medical Center, Seoul, for advice and assistance in the statistical analysis.
Authors' roles: Study design: SHL, J-MK, and CK. Data collection: SWL, SHA, TK, K-HL, B-JK, E-HC, S-WK, T-HK, J-MK, GSK, and S-YK. Data analysis: SWL and SHL. Data interpretation: SHL, SWL, E-HC, TK, J-MK, and CK. Drafting of the manuscript: SHL, SWL, E-HC, J-MK, and CK. Finalizing of the manuscript: SHL, SWL, SHA, TK, K-HL, B-JK, E-HC, S-WK, T-HK, J-MK, GSK, S-YK, J-MK, and CK. J-MK and CK take responsibility for the integrity of the data analysis.