Femoral neck width genetic risk score is a novel independent risk factor for hip fractures

Abstract Femoral neck width (FNW) derived from DXA scans may provide a useful adjunct to hip fracture prediction. Therefore, we investigated whether FNW is related to hip fracture risk independently of femoral neck bone mineral density (FN-BMD), using a genetic approach. FNW was derived from points automatically placed on the proximal femur using hip DXA scans from 38 150 individuals (mean age 63.8 yr, 48.0% males) in UK Biobank (UKB). Genome-wide association study (GWAS) identified 71 independent genome-wide significant FNW SNPs, comprising genes involved in cartilage differentiation, hedgehog, skeletal development, in contrast to SNPs identified by FN-BMD GWAS which primarily comprised runx1/Wnt signaling genes (MAGMA gene set analyses). FNW and FN-BMD SNPs were used to generate genetic instruments for multivariable Mendelian randomization. Greater genetically determined FNW increased risk of all hip fractures (odds ratio [OR] 1.53; 95% CI, 1.29–1.82 per SD increase) and femoral neck fractures (OR 1.58;1.30–1.92), but not trochanteric or forearm fractures. In contrast, greater genetically determined FN-BMD decreased fracture risk at all 4 sites. FNW and FN-BMD SNPs were also used to generate genetic risk scores (GRSs), which were examined in relation to incident hip fracture in UKB (excluding the FNW GWAS population; n = 338 742, 3222 cases) using a Cox proportional hazards model. FNW GRS was associated with increased risk of all incident hip fractures (HR 1.08;1.05–1.12) and femoral neck fractures (hazard ratio [HR] 1.10;1.06–1.15), but not trochanteric fractures, whereas FN-BMD GRS was associated with reduced risk of all hip fracture types. We conclude that the underlying biology regulating FNW and FN-BMD differs, and that DXA-derived FNW is causally related to hip fractures independently of FN-BMD, adding information beyond FN-BMD for hip fracture prediction. Hence, FNW derived from DXA analyses or a FNW GRS may contribute clinically useful information beyond FN-BMD for hip fracture prediction.


Introduction
Hip fractures account for the greatest impact of osteoporosis in terms of mortality, morbidity, and health economic impact. 1 DXA-derived BMD is widely used to evaluate hip fracture risk, for which femoral neck (FN) BMD has greater predictive value compared with measurements at other sites. 2X-ray attenuation, used by DXA to estimate BMD, reflects both bone density and depth of bone.Given the lack of correction for depth, DXA-derived BMD is expressed as g/cm 2 and referred to as an "areal" bone density.Alternative methods for fully correcting BMD for bone size have been proposed.For example, recalibrating lumbar spine BMD to estimate true "volumetric" bone density by dividing lumbar spine BMC by bone area (BA) raised to the power of 1.5 (based on the assumption that a vertebra represents a cuboid) corrected ethnic differences in lumbar spine BMD due to size differences. 3owever, the same geometric model does not apply to the hip.
Current methods for deriving BMD are well established, and there is a case for retaining these but combining with a separate measure of bone size.Since the height of the FN region of interest (ROI) on DXA scans is fixed, FN-BA solely depends on average femoral neck width (FNW).As the cross section of the FN approximates to a circle, FNW may provide a reasonable estimate of depth within the FN ROI.Hip structural analysis (HSA), developed over 25 yr ago, provides an automated means of deriving FNW from hip DXA scans, as well as other geometric indices and estimates of hip strength. 4Using this method, Rivadeneira et al. found that as well as a lower FN-BMD, hip fracture cases had greater FNW, 5 consistent with the expected reciprocal relationship between FNW and "volumetric" FN-BMD.This raises the possibility that prediction of hip fracture by "areal" FN-BMD might be preferentially enhanced by the addition of FNW, through the provision of missing depth information at the FN ROI.Since HSA or equivalent software is widely available, if confirmed, such understanding could be readily applied to improve hip fracture prediction by DXA through combination of FN-BMD and FNW results.
In the present study, we examined whether FNW contributes to hip fracture prediction independently of FN-BMD, using a genetic approach.First, we aimed to perform a genome wide association study (GWAS) of FNW, derived from automated annotation of hip DXA scans obtained in 38 150 individuals in UK Biobank (UKB) from which minimum FNW could be calculated (Supplementary Figure S1).Results were then used to provide genetic instruments in Mendelian randomization (MR) analyses to determine if FNW is causally related to hip fracture risk, including multivariable MR (MVMR) to establish if any such effect is independent of FN-BMD.Finally, given the recent development of genetic risk scores (GRSs) for BMD to predict fractures, 6 we examined whether combining a FN-BMD GRS with a FNW GRS provides greater prediction of hip fractures compared with the use of a single GRS alone.

Study population
The UKB is a prospective cohort study which recruited 500 000 adults from the United Kingdom, aged between 37 and 73 yr at a baseline visit which took place between 2006 and 2010. 7The participants have undergone comprehensive genetic and physical phenotyping (see website for comprehensive catalogue of variables available http://bio bank.ctsu.ox.ac.uk/crystal/).The study is overseen by the Ethics Advisory Committee and received approval from the National Information Governance Board for Health and Social Care and Northwest Multi-Centre Research Ethics Committee (11/NW/0382), all participants provided informed consent for this study.As part of the UKB extended imaging study, which commenced in 2014, a number of imaging modalities, including DXA, are being collected. 8Information on hip fracture was obtained from linkage to the hospital episode statistics (HES) database.All UKB participants were linked, both prospectively and retrospectively at baseline.HES records began on April 01, 1997, and end date for data capture for the present study was September 30, 2021.The maximum duration of follow-up for hip fracture from the baseline visit was 15.5 yr.

FNW measurement
We used left hip DXA scans to train an 85-point Statistical Shape Model based machine-learning system (Supplementary Figure S1A) to outline the proximal femur and acetabulum in all available images as of April 2021. 9A custom Python 3.0 script was developed to calculate the minimum FNW in the FN region accessible online (Faber B. Geometric Parameters Python 3.0 Code.2022.https://zenodo.org/badge/latestdoi/518486087).The pixel dimension data stored in DXA DICOM images were converted to millimeters (mm).FNW was defined as the shortest distance measured between the superior and inferior FN.The inferior side of the FN was mapped with points 6-12, and the superior side with points 32-38.A line-segment approach was used to automatically calculate the narrowest distance between these points (Supplementary Figure S1A and B).A description of this approach has been published previously. 10

Genetic analyses
Preparation and quality control of genetic data in UKB Genotyping, imputation, and quality control were performed by UKB as previously described 7 (see Supplementary supplementary methods).

Femoral neck width genome-wide association study
To test the association between genetic variants and FNW, FNW was stratified by sex and adjusted for age, genotyping chip, and the first 20 ancestry principal components.Residuals resulting from female and male analyses were standardized (mean = 0, SD = 1), and then combined into a single outcome for GWAS.We used a linear mixed model assuming an additive allelic effect implemented in BOLT-LMM (v2.3) to account for cryptic population structure and relatedness.The GWAS involved high-quality genome-wide imputed v3 genetic data (∼10 million SNPs, INFO >0.3, MAF > = 1%).SNPs reaching genome-wide significance (5 × 10 -8 ) were taken forward for conditional association analysis.The same methods were also used to generate a GWAS for hip axis length (HAL), derived from hip outline points as previously described. 10

Genome wide complex trait conditional and joint analysis (GCTA-COJO)
To detect multiple independent association signals at each of the genome-wide significant (GWS) loci, we applied approximate conditional and joint genome-wide association analysis 11 in conjunction with a UKB reference panel using GCTA v1.93 software.Conditionally independent variants at GWAS significance level were annotated with the closest gene using bedtools 12 v2.3.0 and the Hg19 Gene list available from https://www.cog-genomics.org/link2/.
Look ups, MAGMA gene-set analysis, and Gene-set enrichment analysis in FUMA Independent FNW signals were looked up in publicly available FN-BMD summary statistics 13 (Tables 1A and 1B and Supplementary Figure S3A).To gain an overview of which biological pathways are involved, FNW and FN-BMD 13 GWAS summary statistics were uploaded to FUMA webbased platform 14 to perform gene-set analysis with MAGMA v1.06 (Table 2, Supplementary Tables S4 and Supplementary S5).Gene sets were obtained from MsigDB v7.0.A total of 15 496 gene sets, including curated gene sets (5500) and gene ontology (GO) terms (9996), were available for testing.Curated gene sets consist of 9 data resources including Kyoto encyclopedia of genes and genomes (KEGG), Reactome, and BioCarta (http://software.broadinstitute.org/gsea/msigdb/collection_details.jsp#C2for details).GO terms consists of 3 categories, biological processes, cellular components, and molecular functions.All parameters were set as default.The output of gene-set analysis contains all genes in each significant gene set (Table 2, Supplementary Tables S4 and S5).
Based on the genes at the identified loci, we also performed a gene-set enrichment analysis as implemented by the FUMA SNP2GENE function (Supplementary Tables S6 and Supplementary S7).We also looked up the independent FNW signals in the GWAS catalogue (Supplementary Table S8).

Genetic correlation
To estimate the genetic correlation between FNW and related traits, we used (cross-trait) linkage disequilibrium score regression (LDSR) 15 as implemented in the LD score tool LDSC available on github. 16This method uses the crossproducts of summary test statistics from 2 GWASs and regresses them against a measure of how much variation each SNP tags (its LD score).The LDSR analyses were restricted to HapMap3 SNPs with minor allele frequency (MAF) > 5% in the 1000 Genomes European reference population.We used precalculated LD scores from the same reference population (https://data.broadinstitute.org/alkesgroup/LDSCORE/). We estimated the genetic correlation between FNW and 3 related traits: hip fracture, 17 fracture at any bone site, 18 and FN-BMD, 13 using public available GWAS summary statistics.
We accounted for multiple testing by using a conservative Bonferroni correction for 3 tests (P = .05/3= .017).

Mendelian randomization
To assess the effects of FNW on the risk of fractures at different bone sites, we performed 2-sample Mendelian randomization (MR) analyses.We used a MVMR approach to estimate the independent causal associations for genetically determined FNW and FN-BMD with risk of fractures at different bone sites.Genetic instrument variables for the FNW exposure were derived from the present GWAS on FNW, while genetic instruments for FN-BMD were derived from previous GWAS on FN-BMD. 13,19The highest available number of FN-BMD signals (n = 49) was identified in the FN-BMD GWAS by Estrada et al. 19 However, HapMap imputation was used in that early GWAS meta-analysis, hindering the ability to examine FN-BMD associations for FNW signals identified in the more detailed imputation panel used by UKB.We therefore selected the GWS FN-BMD signals from Estrada et al. as instruments for FN-BMD but used effect estimates from the later FN-BMD GWAS study by Zheng et al. 13 (the imputation panel in the latter included the majority of FNW GWAS SNPs identified here, making MVMR feasible).We only used variants that were available in both the present FNW GWAS and the FN-BMD GWAS by Zheng et al.The variants were required to have a MAF > 1% and to be associated with either FNW or FN-BMD at a GWS level (P < 5 × 10 -8 ).We selected instruments with a pairwise r 2 < 0.01 (based on the European populations in LDlink) 20 to ensure that there was low correlation between instruments.One SNP associated with FNW was strongly correlated with one SNP associated with FN-BMD (r 2 = 0.73).For this pair, we first selected the SNP that is most strongly associated with FN-BMD in the GWAS.In sensitivity analyses, we selected the other SNP that is most strongly associated with FNW, revealing virtually identical results. 21Two palindromic FNW SNPs with non-clear strand were removed from the analyses.After quality filtering, 40 FNW SNPs and 40 FN-BMD SNPs remained.Using Steiger filtering, no more FNW SNPs were suggested to be removed.We estimated the F-statistic as a measure of instrument strength. 22he outcome fracture associations (logistic regression adjusted for age, and sex) used in the 2-sample MR were derived from summary statistics from a previous GWAS on hip fractures (including 11.515 hip fracture cases 17 ) or newly performed association analyses in UKB, excluding the FNW GWAS data set (Supplementary Table S1).As the primary MR analyses, we used combined weighted estimates by IVW using fixed or random effects depending on Cochran's Q statistic test of heterogeneity.We then used weighted median MR as a sensitivity analysis and MR-Egger regression to test for possible directional horizontal pleiotropy.To reduce the possible impact of heterogeneity, we also performed sensitivity analyses excluding outliers of genetic instruments using MR-LASSO.The MR analyses were conducted using the R-package MR.

Weighted GRSs
We defined a weighted GRS for FNW (FNW GRS) based on the 71 conditionally independent (COJO) significant SNPs identified in the present study.We also defined a GRS for FN-BMD (FN-BMD GRS) based on 49 SNPs previously identified to be associated with FN-BMD at GWS level. 19For each individual, the GRSs were defined as the weighted sum of SNP dosages, where SNP effects from the corresponding BMD GWAS were used as weights.The GRSs were standardized to have a mean of and SD of 1.The separate and combined associations for FNW GRS and FN-BMD GRS with incident fracture were calculated using Cox-regression in UKB samples excluding the samples used for the FNW GWAS.The effects are given as hazard ratios (HRs) per SD increase in GRS.The base model included sex and baseline age as covariates.

Genome-wide association study
The discovery set comprised 38 150 UKB participants with available FNW analyses, mean 63.8 yr, of whom 48% were male (Table 3).We identified 71 independent signals at 61 loci passing GWS level (P < 5 × 10 −8 ; Tables 1A and 1B; Supplementary Table S2 and Supplementary Figure S2), of which 8 were low-frequency (MAF ≤ 5% but >1%) and 63 were common (MAF > 5%).These 71 signals explained 7.6% of the variance of FNW (Supplementary Table S2) and showed limited associations with FN-BMD as evaluated in a previous FN-BMD GWAS data set 13 (Supplementary Table S3, Supplementary Figure S3A).Conversely, the previously identified 49 GWS FN-BMD signals showed limited associations with FNW in the present GWAS dataset (Supplementary Figure S3B).MAGMA gene set analyses of the FNW GWAS data identified different gene sets (representing cartilage differentiation, hedgehog signaling, skeletal development) compared with FN-BMD GWAS (runx1/Wnt signaling), demonstrating distinct underlying biology for the regulation of FNW and FN-BMD (Table 2, Supplementary Tables S4 and S5).This notion was further supported by clear differences in FUMA gene set enrichment for FNW and FN-BMD signals (Supplementary Tables S6 and S7), with the FNW GWAS comprising signals for hip dimensions, fat distribution, and height, while the FN-BMD GWAS included signals for fracture risk and BMD at other sites.
Look-up in the GWAS catalogue of the 71 conditionally independent FNW signals including linked signals (r 2 > 0.8) identified multiple signals for osteoarthritis, hip circumference, and hip bone size (Supplementary Table S7).We observed a strong genetic correlation for FNW with hip fractures (r g = 0.48) but not with fractures at any bone site (r g = 0.07) (Supplementary Table S9).A modest inverse genetic correlation was observed for FNW with FN-BMD (r g = -0.20).A second parameter of hip geometry, HAL, could be derived from our automated annotation of hip shape, for which GWAS summary statistics were also obtained.Although HAL was strongly correlated genetically with FNW (r g = 0.61), this showed a considerably weaker genetic correlation with hip fracture (r g = 0.24).

Mendelian randomization
SNPs selected from the current FNW GWAS, and previous FN-BMD GWAS, 19 provided well-powered genetic instruments for subsequent MR analyses (F-statistic 40.6 and 31.0,respectively, Supplementary Table S10).We evaluated the causal associations for genetically determined FNW and FN-BMD with hip fractures using a hip fracture GWAS meta-analysis for the outcome analyses 17 (Supplementary Table S11 and Supplementary Figure S4).Univariate MR revealed a significant effect of both genetically determined FNW (OR = 1.55, 95% CI,  S4) as exposures.
As an alternative exposure in the MR analyses, we used a weighted GRS for FNW (FNW GRS, 71 SNPs).Using this FNW GRS as exposure, MR confirmed that FNW is causally associated with all hip (OR = 1.64, 95% CI, 1.34-2.02,per SD increase) and femur neck (OR = 1.85, 95% CI, 1.45-2.36)but not trochanteric (OR = 1.22,95% CI, 0.81-1.83)fractures.In sex stratified analyses, the causal effects of FNW on all hip fractures and FN fractures were slightly greater in women compared to men (Supplementary Table S13).Collectively, these data demonstrate that genetically determined FNW is causally associated with the risk of hip fractures, specifically that of FN fractures.

FNW GRS and FN-BMD GRS add independent information for prediction of incident hip fractures
We next determined whether the FNW GRS and/or FN-BMD GRS predict incident hip fractures.In separate models, a high FN-BMD GRS was associated with reduced risk of incident hip fracture at any site (HR 0.83; 0.80-0.86 per SD increase), whereas a high FNW GRS was associated with increased hip fracture risk (HR 1.09; 1.05-1.13;Figure 2, Supplementary Table S14).The associations between the FNW GRS and hip fracture risk were more pronounced for FN (HR 1.11; 1.06-1.16)compared with trochanteric (HR 1.03; 0.97-1.11)fractures.Similar results were observed in combined analyses (including both FNW GRS and FN-BMD GRS) and in models additionally adjusted for BMI; however, the association between FNW GRS and hip fracture risk was attenuated by approximately 25% following separate adjustment for height and weight to account for effects of body size (Supplementary Table S14).An age interaction (P = 1.9 × 10 -3 for the age/FNW GRS interaction term) was observed for the association between the FNW GRS and FN fracture risk, reflected by a higher HR for younger (age ≤ 71; 1.14; 1.08-1.21)compared with older (age > 71; 1.06; 1.00-1.13)individuals (Figure 2, Supplementary Table S15).Similar associations for the FNW GRS with incident hip fractures were observed in both men and women (Figure 2, Supplementary Table S15).Additionally, we observed an interaction between the FN-BMD GRS and the FNW GRS for the prediction of hip fracture risk (P = .04for the interaction term), where individuals with genetically determined low FN-BMD had a more pronounced increased risk from genetically determined high FNW.
Finally, we examined additive associations for binarized high risk FNW GRS and binarized high risk FN-BMD GRS with hip fractures risk.Participants in UKB were classified as high risk (yes/no) for high FNW based on their FNW GRS and at high risk (yes/no) for low FN-BMD based on their FN-BMD GRS.Participants were divided into 4 different groups (group 1 = no/no; group 2 = yes/no; group 3 = no/yes; and group 4 = yes/yes) based on their binarized FNW GRS and binarized FN-BMD GRS.We used the lowest risk group (group 1 = no/no) as reference to study the possible additive associations for the 2 different GRSs.We used 3 different cut-off limits as definitions of high risk for the 2 GRS (50%, 25%, and 10%; Figure 3).A high binarized FNW GRS was associated with high hip fracture risk, while a low binarized FN-BMD GRS was associated with high hip  Associations for FNW GRS and FN-BMD GRS with incident fractures in UKB The effects are given as HRs per SD increase in GRS.All models were adjusted for sex and baseline age, except the sex stratified models where sex was not included as a covariate.Age stratification was based on the median age at hip fracture (71.7 yr).* Significant age interaction (P = 1.3 × 10 -4 ) for FNW GRS.* * Significant age interaction (P = 1.9 × 10 -3 ) for FNW GRS.fracture risk.Using the 10% cut-off for high risk, subjects in the high-risk GRS group for both FNW GRS and FN-BMD GRS (ie, group 4 = yes/yes) had a more than 2-fold increased risk of hip fractures compared with those in the low-risk group for both binarized GRSs (ie, group 1 = no/no).
Binarized FNW GRS and FN-BMD GRS contributed independently to hip fracture prediction.For example, on comparing groups 3 and 4, the high risk FNW GRS added information beyond the high risk FN-BMD GRS (Figure 3, Supplementary Table S16).

Discussion
We investigated whether FNW contributes to hip fracture risk independently of FN-BMD, using a genetic approach.Having performed a GWAS of FNW derived from hip DXA scans in over 38 000 individuals in UKB, we identified 71 conditionally independent signals in 61 different loci explaining 7.6% of the variance of FNW, of which 70 signals represented novel genetic associations with FN bone size (the FN-area signal at the HHP locus was previously reported by Styrkarsdottir et al. 23 ).The genetic architecture of FNW appeared to be distinct to that of FN-BMD, suggesting these 2 traits are in large part independent.Less than 20% of FNW genetic signals were nominally associated with FN-BMD.FNW showed a relatively weak genetic correlation with FN-BMD, and MAGMA gene set analysis revealed involvement of FNW and FN-BMD SNPs in distinct biological pathways.
Although FNW only showed weak genetic correlation with FN-BMD, it was correlated relatively strongly with hip fractures.Given the suggestion that FNW is largely independent of FN-BMD, we investigated whether FNW is causally related to the risk of hip fracture, independently of FN-BMD.MVMR revealed that greater FNW, or a highly correlated hip shape parameter, increases the risk of any hip fractures, and that of FN fractures specifically, despite adjustment for FN-BMD.In contrast to FNW which was only related to risk of hip fracture/FN fracture, FN-BMD was also related to risk of trochanteric and forearm fractures.
A GRS based on genetic associations with FN-BMD has previously been proposed as an adjunct in clinical fracture prediction, either in isolation or in combination with clinical risk factors such as those included in FRAX. 6,24herefore, we examined whether a GRS based on FNW might also have utility in hip fracture prediction.We found that both FNW and FN-BMD GRSs were independently related to risk of hip/FN fracture, and that these exerted additive effects on hip fracture risk.For example, an individual with a GRS in the highest 10% risk category for both FNW and FN-BMD has an approximately 2-fold increased risk of hip/FN fracture, compared with a 25%-35% increase based on either parameter alone.In addition to improving the use of GRSs to predict fractures by combining 2 independent scores, an FNW GRS may also prove useful due to its presumed independence from DXA BMD.This would represent an important advantage over a BMD GRS, which provides little additional predictive value if BMD is already known. 25hough the present study focused on additive effects of GRSs for FNW and BMD on risk of hip fracture, it may be possible to use an equivalent approach based on measured parameters.In the present study, FNW was derived using a bespoke method based on points annotated as part of a separate study on hip shape. 9However, an equivalent measure, obtained using HSA software, 4 is widely available (these were strongly correlated [r 2 = 0.97] in a subset of 1744 DXA images where FNW was obtained using both methods).An alternative method would be to combine BMD with FN BA, which is provided routinely during hip DXA measurements, and also correlated strongly with FNW (r 2 = 0.93).
When used alone, the FNW GRS had a clear independent relationship with hip fracture risk, in the opposite direction to that of FN-BMD and when combining both GRSs, a marked improvement in predictive ability was observed.One explanation for these findings is that the input of additional information about bone size, provided by FNW, enables a more accurate estimate of "volumetric" BMD than that provided by "areal" BMD, by providing missing information about depth.As discussed in the introduction, several approaches have been attempted to more fully account for bone size when evaluating BMD by DXA.Our results suggest that this concept can also be applied to GRSs used for fracture prediction.Rather than providing missing depth information, it may be that greater FNW per se has a negative effect on bone strength and fracture risk.For example, for a given cortical thickness, greater FNW is inversely related to resistance to buckling as reflected by buckling ratio. 5On the other hand, FNW is positively related to bending strength as reflected by cross-sectional moment of inertia 26 ; to the extent that both types of forces contribute to risk of hip fracture, whether FNW has any net direct effect on hip fracture risk is currently unclear.A further explanation is that, rather than a direct association, FNW is related to hip fracture risk through co-association with height, which is also positively related to risk of hip fracture, 27 possibly because height is a proxy for leg length. 28hough the overall genetic correlation between FNW and FN-BMD was relatively weak, for those SNPs related to both traits, an inverse relationship between these 2 traits was generally observed.However, the LRP5 locus was an exception, since this was positively related to both FN-BMD and FNW, suggesting that in contrast to other loci, the LRP5 locus has a positive effect on both bone size and BMD.Inkeeping with this suggestion, individuals with high bone mass as a consequence of a mutation in LRP5 have been found to have both increased BMD and bone size, as reflected by tibial and radial cortical area and thickness. 29Although the LRP5 locus appears to affect bone size of the skeleton as a whole, genetically determined FNW only influenced risk of FN/hip fracture, suggesting fractures at the latter site are particularly influenced by bone geometry.This contrasts with the genetically determined FN-BMD GRS, which influenced fracture risk at multiple sites, suggesting that FN-BMD signals influence BMD throughout the skeleton.In addition to being restricted to the prediction of FN/hip fractures, the FNW-GRS was more strongly related to risk of hip fracture in younger individuals.One possible explanation for this finding is the greater contribution to hip fracture of risk factors unrelated to skeletal fragility in older individuals, such as factors related to fall risk.
This paper reports the first FNW GWAS, which provided the basis for a novel FNW GRS which may have clinical utility as a BMD independent risk factor for hip fracture.In terms of limitations, though many novel loci related to FNW were identified, functional genomic studies intended to characterize genetic mechanisms contributing FNW were not undertaken as part of the current study.That said, MAGMA gene set analysis was used to characterize the biological pathways identified by our GWAS.These primarily comprised genes representing cartilage differentiation, hedgehog signaling, and skeletal development, consistent with the determination of overall skeletal size.This contrasted sharply with findings for FN-BMD, which primarily identified Wnt signaling genes, consistent with the important role of Wnt signaling in the regulation of bone mass. 30n terms of other limitations, UKB on which this GWAS was based is primarily comprised of Caucasians, and further studies are required to investigate whether the GRS has equivalent predictive value in other ethnic groups.To establish the clinical utility of our FNW GRS, further studies are required to confirm that this GRS predicts hip fracture independently of BMD, as well as established clinical risk factors.Furthermore, it should be acknowledged that although the FNW GRS may have clinical utility for predicting fractures, unlike BMD, this appears to be limited to hip fractures.Finally, although we have described relationships between genetically predicted FNW and BMD, and hip fracture, associations between hip fracture and directly measured FNW and BMD were not presented.The latter analyses are restricted to the subgroup with DXA data, and given this smaller sample and the shorter follow-up period, there were relatively few hip fractures on which to base analyses, limiting statistical power.The number of participants undergoing DXA scans in UKB, as well as the duration of follow up, is increasing substantially with time, and we plan to reexamine these relationships once more hip fracture cases are available.
In conclusion, our FNW GWAS demonstrates that the biology underlying this trait differs from that of FN-BMD.Consequently, although FNW or a highly correlated hip shape parameter is causally related to hip fractures, this is independent of FN-BMD, and DXA-derived FNW adds information beyond FN-BMD for hip fracture prediction.Based on the genetic evidence presented herein, we propose that FNW derived from DXA analyses or a FNW GRS may add clinically useful information beyond FN-BMD for hip fracture prediction.

Figure 1 .
Figure 1.Independent effects of genetically determined FNW and FN-BMD on fracture risk in UKB: MVMR analyses to estimate the effect of genetically determined FNW and femoral neck bone mineral density (FN-BMD) on the risk of hip fracture, FN fractures, trochanteric fractures, and forearm fractures in UKB.Both prevalent and incident fractures were included.OR and 95% CIs are given.

Figure 2 .
Figure 2.Associations for FNW GRS and FN-BMD GRS with incident fractures in UKB The effects are given as HRs per SD increase in GRS.All models were adjusted for sex and baseline age, except the sex stratified models where sex was not included as a covariate.Age stratification was based on the median age at hip fracture (71.7 yr).* Significant age interaction (P = 1.3 × 10 -4 ) for FNW GRS.* * Significant age interaction (P = 1.9 × 10 -3 ) for FNW GRS.

Figure 3 .
Figure 3. Additive associations for binarized high risk FNW GRS and, binarized high risk FN-BMD GRS with fractures risk.Participants in UKB were divided into 4 different groups (no/no, yes/no, no/yes and yes/yes) based on their binarized FNW GRS and binarized FN-BMD GRS, using 3 different cut-off limits as definitions of high risk for the 2 GRS (50%, 25%, and 10 %).
Conditionally independent significant signals within chromosomes 10-21 associated with femur neck width (FNW) in 38 150 UKB participants.Results are presented as estimated association (beta) and standard error (SE) expressed per effect allele (EA).Beta, SE, and P are from the conditional (COJO) analysis.OA, other allele; EAF, effect allele frequency.

Table 2 .
MAGMA gene set analysis.
Significant gene sets from the MAGMA gene-set analysis of the FNW GWAS and the FN-BMD GWAS (PMID: 26367794): SET and VARIABLE: name of the gene set, NGENES: number of genes in the gene set.In total 15 496 gene sets were available for testing; 15 488 of them were represented in the FNW GWAS and 15 485 were represented in the FN-BMD GWAS.A gene set was considered to be significant if P < .05/15496= 3.2 x 10 -6 .

Table 3 .
Characteristics of UKB Study participants in the FNW GWAS.Population characteristics of the UKB participants in the FNW GWAS with complete FNW and covariate data.