Development of a polygenic risk score to improve screening for fracture risk: A genetic risk prediction study.

BACKGROUND: Since screening programs identify only a small proportion of the population as eligible for an intervention, genomic prediction of heritable risk factors could decrease the number needing to be screened by removing individuals at low genetic risk. We therefore tested whether a polygenic risk score for heel quantitative ultrasound speed of sound (SOS)-a heritable risk factor for osteoporotic fracture-can identify low-risk individuals who can safely be excluded from a fracture risk screening program. METHODS AND
FINDINGS: A polygenic risk score for SOS was trained and selected in 2 separate subsets of UK Biobank (comprising 341,449 and 5,335 individuals). The top-performing prediction model was termed "gSOS", and its utility in fracture risk screening was tested in 5 validation cohorts using the National Osteoporosis Guideline Group clinical guidelines (N = 10,522 eligible participants). All individuals were genome-wide genotyped and had measured fracture risk factors. Across the 5 cohorts, the average age ranged from 57 to 75 years, and 54% of studied individuals were women. The main outcomes were the sensitivity and specificity to correctly identify individuals requiring treatment with and without genetic prescreening. The reference standard was a bone mineral density (BMD)-based Fracture Risk Assessment Tool (FRAX) score. The secondary outcomes were the proportions of the screened population requiring clinical-risk-factor-based FRAX (CRF-FRAX) screening and BMD-based FRAX (BMD-FRAX) screening. gSOS was strongly correlated with measured SOS (r2 = 23.2%, 95% CI 22.7% to 23.7%). Without genetic prescreening, guideline recommendations achieved a sensitivity and specificity for correct treatment assignment of 99.6% and 97.1%, respectively, in the validation cohorts. However, 81% of the population required CRF-FRAX tests, and 37% required BMD-FRAX tests to achieve this accuracy. Using gSOS in prescreening and limiting further assessment to those with a low gSOS resulted in small changes to the sensitivity and specificity (93.4% and 98.5%, respectively), but the proportions of individuals requiring CRF-FRAX tests and BMD-FRAX tests were reduced by 37% and 41%, respectively. Study limitations include a reliance on cohorts of predominantly European ethnicity and use of a proxy of fracture risk.
CONCLUSIONS: Our results suggest that the use of a polygenic risk score in fracture risk screening could decrease the number of individuals requiring screening tests, including BMD measurement, while maintaining a high sensitivity and specificity to identify individuals who should be recommended an intervention.

Introduction

Screening programs are generally designed to identify a proportion of the screened population whose risk of a clinically relevant outcome is high enough to merit an intervention. However, usually only a small proportion of individuals who undergo screening is found to be at high risk, indicating that much of the screening expenditure is spent on individuals who will not qualify for intervention.

Osteoporosis is a common and costly disease that results in an increased predisposition to fractures [1]. Many guidelines [2-6] aimed at the prevention of osteoporosis-related fractures incorporate the Fracture Risk Assessment Tool (FRAX) [7,8], a validated method to risk stratify individuals for treatment by assessing their 10-year probability of major osteoporotic fracture. Guidelines vary widely, but often recommend a staged process where individuals are first assessed with a clinical-risk-factor-based FRAX (CRF-FRAX), and those at increased risk of fracture are then additionally characterized with a more expensive bone mineral density (BMD)–based FRAX (BMD-FRAX) score. Such approaches are usually recommended in the setting of enhanced case-finding strategies, but recently, a large randomized controlled trial (SCOOP) demonstrated the potential benefit of community-based fracture risk assessment in reducing rates of hip fractures in elderly women [9]. This trial used a strategy based on the National Osteoporosis Guideline Group (NOGG) screening strategy [3], which implements fracture risk stratification through the use of FRAX scores. In this trial, the entire screened population underwent FRAX assessment using clinical risk factors, and almost half (49%) had a sufficiently high probability of fracture to warrant further testing using a BMD-FRAX test. Yet, only 14% of the screened population had a resultant probability of fracture high enough to warrant intervention. This suggests that a method that improves screening efficiency and decreases the number of persons undergoing risk stratification, particularly BMD-FRAX assessments, would be a welcome addition to the screening strategy.

Skeletal measures that predict fracture risk are highly heritable (50%–85%) and include BMD and quantitative ultrasound speed of sound (SOS) measurements, which are highly correlated [10-13]. Recently, large cohort resources have enabled the genomic prediction of such heritable clinical risk factors from genotypes through polygenic risk scores [14-20], which capture information from many single nucleotide polymorphisms assayed from genome-wide genotyping. These assays assess common genetic variation at millions of single nucleotide polymorphisms and cost approximately $40 in a research context. However, the clinical utility of such polygenic risk scores is unclear, widespread replication of polygenic risk scores is currently lacking, and it is unknown whether they can aid in screening programs. Defining their clinical relevance may be particularly relevant in a British context, where the National Health Service aims to sequence 5 million individuals within 5 years [21].

Very large cohorts are required to train polygenic risk scores, and current cohorts lack sufficient sample size to generate useful BMD polygenic risk scores. However, since BMD is strongly correlated with SOS [22] and SOS has been measured in 341,449 individuals in UK Biobank, we developed a polygenic risk score for SOS termed “gSOS” (for “genetically predicted SOS”) that could be used to identify individuals unlikely to have low enough BMD to warrant a clinical intervention. To improve screening efficiency, such individuals could be removed from an osteoporosis screening program prior to measurement of BMD. We then tested the generalizability and potential benefit of incorporating gSOS into the NOGG guidelines using 5 cohorts, comprising 10,522 eligible individuals. Last, we tested if gSOS could decrease the number of people requiring more detailed assessments, such as BMD measurement, while still identifying those who require interventions to decrease their risk of fracture.

Methods

Overall study design and cohorts

The purpose of this study was not to predict fractures. Rather, the purpose of this study was to understand if genetic prescreening could reduce the number of screening tests needed to identify individuals at risk of osteoporotic fractures. This study included 3 phases (Fig 1). The first 2 phases were conducted in 2 distinct subsets of the UK Biobank study cohort, and the final phase in a further subset of UK Biobank combined with 4 other cohorts. Characteristics of the cohorts are shown in Table 1, with the cohorts described in detail in Table A in S1 Tables.

Fig 1

Overall study design.

BMD, bone mineral density; CLSA, Canadian Longitudinal Study on Aging; GWAS, genome-wide association study; NOGG, National Osteoporosis Guideline Group; PRS, polygenic risk score; QC, quality control; SOF, Study of Osteoporotic Fractures; SOS, speed of sound; UKB, UK Biobank.

Table 1

Participant characteristics by dataset.

Participant characteristics	Model development cohorts		gSOS-based screening test cohorts
	UK Biobank Training Set	UK Biobank Model Selection Set	UK Biobank Test Set	CLSA	SOF	Mr OS US	Mr OS Sweden
Sample size	341,449	5,335	4,741	6,704	3,426	4,657	1,880
Individuals eligible for screening, N (%)	—	—	2,445 (51.6)	2,931 (43.7)	2,094 (61.1)	2,026 (43.5)	1,026 (54.6)
Age, mean (SD)	56.8 (8.0)	56.6 (8.1)	55.8 (7.6)	62.6 (9.9)	71.5 (5.3)	74.0 (6.0)	75.4 (3.2)
Women, N (%)	186,569 (55.6)	2,863 (53.7)	2,489 (52.5)	3,396 (50.7)	3,426 (100)	0 (0)	0 (0)
Smoker, N (%)	27,181 (8.0)	397 (7.4)	966 (20.4)	581 (8.7)	270 (7.9)	145 (3.1)	178 (9.5)
Previous fracture, N (%)	34,917 (10.2)		386 (8.1)	1,032 (15.4)	1,210 (35.3)	1,084 (23.3)	637 (33.9)
Use of glucocorticoids, N (%)	3,330 (1.0)	51 (0.8)	79 (1.7)	258 (3.9)	363 (10.6)	98 (2.1)	34 (1.8)
Alcohol user, N (%)	—	—	—	1,189 (17.7)	98 (2.9)	182 (3.9)	52 (2.8)
Fall within last 12 months, N (%)	69,057 (20.2)	1,052 (20.0)	1,500 (31.6)	699 (10.4)	1,021 (28.2)	984 (21.1)	298 (15.9)
Rheumatoid arthritis, N (%)	3,312 (1.0)	41 (0.8)	28 (0.6)	191 (2.9)	252 (7.0)	226 (4.9)	27 (1.4)
Secondary osteoporosis, N (%)	14,541 (4.3)	215 (4.0)	192 (4.1)	313 (4.7)	—	—	—
Parental history of fracture, N (%)	—	—	—	820 (12.2)	404 (14.4)	599 (16.8)	164 (8.7)
Baseline CRF-FRAX score for MOF, mean (SD)	5.1 (3.1)	5.0 (3.1)	4.8 (2.7)	8.1 (6.8)	18.7 (9.5)	9.5 (4.7)	11.1 (6.3)
Baseline BMD-FRAX score for MOF, mean (SD)	—	—	4.9 (2.6)	7.5 (5.8)	17.1 (9.5)	8.1 (4.4)	13.1 (5.6)
gSOS, mean (SD)	—	−0.002 (1.00)	0.043 (0.98)	−0.005 (1.00)	0 (0.99)	−0.033 (0.98)	−0.708 (0.46)

BMD-FRAX, bone-mineral-density-based Fracture Risk Assessment Tool; CLSA, Canadian Longitudinal Study on Aging; CRF-FRAX, clinical-risk-factor-based Fracture Risk Assessment Tool; MOF, major osteoporotic fracture; SOF, Study of Osteoporotic Fractures.

The first phase used least absolute shrinkage and selection operator (LASSO) regression [23] to train a set of polygenic risk score models to predict SOS in the UK Biobank Training Set (N = 341,449). In phase 2, the polygenic risk score model explaining the most variance in measured SOS in the UK Biobank Model Selection Set (N = 5,335) was selected and named gSOS. The ability of gSOS to explain variance in measured SOS was then tested in the UK Biobank Test Set (N = 84,768). In phase 3, gSOS was tested for its performance in a screening strategy, based on NOGG guideline thresholds of fracture risk, applied to a population of 10,522 individuals derived from 5 separate cohorts. Inclusion in the screening program required these individuals to be ≥50 years with at least 1 risk factor and available measurement of femoral neck BMD. This population comprised a further distinct subset of the UK Biobank Test Set (N = 2,445), as well as individuals from the Canadian Longitudinal Study on Aging (CLSA) (N = 2,931), the Study of Osteoporotic Fractures (SOF) (N = 2,094), Mr OS US (2,026), and Mr OS Sweden (N = 1,026). Together these 5 cohorts in phase 3 are referred to as the validation cohorts. Next, to test the effect of gSOS on fracture screening by age, we stratified the CLSA cohort by age, dividing the population into 3 age groups: 50–59 years, 60–69 years, and ≥70 years. The CLSA cohort was chosen for this age-stratified analysis, because it was the largest validation cohort and had the widest age range. To assess the performance of gSOS in ancestries other than White British, we tested it in individuals in the UK Biobank Test Set who were eligible for screening and were of non–White British ancestry, as defined by genotypes (see S1 Text for further details of definition of ancestry; Table B in S1 Tables shows the demographic and risk factor characteristics of the sub-population).

This study adheres to the GRIPS statement (see S1 Checklist) and did not have a pre-specified analysis plan [24]. Specific ethics approval was not required for this study.

SOS and BMD measurement

We decided to use polygenic risk scores to predict SOS, rather than BMD, because polygenic risk scores require a large number of individuals with proper phenotyping and genome-wide genotyping. The largest dataset for SOS is approximately 10-fold larger than that for BMD [10,25]. SOS also predicts fracture, with similar performance characteristics compared to BMD, and the 2 measures are correlated (r = 0.4–0.6) [22]. However, since femoral neck BMD is required for FRAX calculations used in screening programs [26], we required that all individuals in the phase 3 analysis have femoral neck BMD measure available. Details of SOS and BMD measurement are available in S1 Text. All analyses used SOS standardized to a mean of 0 and standard deviation of 1.

Development of machine learning model to predict SOS

Training, model selection, and test datasets

To develop and test gSOS, we followed best practices in clinical prediction to ensure unbiased estimates of model performance by developing the models in datasets distinct from the datasets that were used to test model performance [27]. Participants in the UK Biobank with White British ancestry (see S1 Text), measured SOS, and genotyping information (N = 426,811) were randomly assigned to the UK Biobank Training Set (80% of participants), the UK Biobank Model Selection Set (1.25% of participants), or the UK Biobank Test Set (18.75% of participants) (Fig 1; Table 1). Since BMD was measured in only 4,741 individuals in all of UK Biobank [28], these individuals were assigned to the UK Biobank Test Set to enable them to be used in phase 3 of the study.

Genome-wide association study (GWAS)

Using methods from our previous GWAS of estimated BMD in UK Biobank [25], but using a different sample size and SOS as the outcome, we undertook a GWAS for SOS in the UK Biobank Training Set (N = 341,449 individuals with White British ancestry). We tested the additive allelic effects of each of the 13.9 million SNPs passing quality control, separately, on SOS using a linear mixed model to adjust for cryptic relatedness and population stratification [29], as well as adjusting for age, sex, assessment center, and genotyping array (S1 Text). Linkage-disequilibrium-independent associations where obtained using PLINK by clumping SNPs in linkage equilibrium at a r2 > 0.05 and selecting a single most significant SNP from within each clumped set. To reduce potential bias due to population stratification, the UK Biobank Training, Model Selection, and Test Sets included only White British participants, while all other cohorts included only people of general European ancestry (as defined in S1 Text). Further, as stated above, the performance of gSOS-based screening was also tested in non–White British participants in UK Biobank.

Polygenic risk scores using LASSO

Using the UK Biobank Training Set, we fitted 6 LASSO models [23] to predict SOS using only SNPs with p-values smaller than a chosen set of thresholds (Table C in S1 Tables). The UK Biobank Model Selection Set was then used to identify the p-value threshold and regularization parameter (λ) that resulted in the lowest root mean square error for the prediction of SOS. This p-value threshold and regularization parameter were then taken forward for testing in the UK Biobank Test Set. Hence, we ensured that the performance of only 1 final polygenic risk score was evaluated in the UK Biobank Test Set. We refer to this final predictor as gSOS, in which SOS is predicted only from genotype.

Traditional polygenic risk scores

Traditional polygenic risk scores [15] were derived from the GWAS for SOS performed in the UK Biobank Training Set, without the use of LASSO, by including different sets of SNPs, selected by p-value threshold and linkage disequilibrium clumping as described in S1 Text (Table C in S1 Tables).

Generation of FRAX scores

FRAX risk scores for major osteoporotic fracture (hip, clinical vertebra, proximal humerus, or wrist) can be generated with or without BMD, referred to in this paper as BMD-FRAX and CRF-FRAX, respectively [26]. Therefore CRF-FRAX and BMD-FRAX were calculated for all participants in each validation cohort [26]. FRAX clinical risk factors were assessed at the baseline visit for each cohort and included age, sex, body mass index (BMI), previous fracture, smoking, glucocorticoid use, rheumatoid arthritis, and secondary causes of osteoporosis. Measures of more than 2 daily units of alcohol and parental history of hip fracture were not available in UK Biobank and were set to “no” for this cohort, as suggested by FRAX guidelines. Not all secondary causes of osteoporosis were available for the SOF, Mr OS US, and Mr OS Sweden cohorts, and these variables were also set to “no” for these cohorts, as recommended by FRAX. Age was recorded at baseline visit. Sex was self-reported and verified by genotype. Individuals with discordant sex between self-report and genotype were excluded. CRF-FRAX and BMD-FRAX were calculated for all participants in each of the clinical cohorts, using country-specific FRAX models [26].

Genomic screening in fracture risk screening

In the absence of an international consensus on fracture risk screening [2,4,5,30], we chose to use the assessment and management clinical algorithm developed by NOGG [3], since a screening program similar to the NOGG screening strategy is supported by randomized controlled trial evidence [9]. The NOGG screening strategy uses 10-year absolute probability of fracture as calculated by FRAX and suggests treatment or reassurance based on thresholds of risk, which are age dependent and consider competing risks. The NOGG guidelines (Fig 2) also aim to identify individuals at risk for fracture in a cost-efficient manner by reserving clinical visits and more costly BMD testing for those at intermediate risk, i.e., where the FRAX score lies close to an intervention threshold. This intervention threshold is equivalent to the age-specific FRAX 10-year probability in women with a prior fragility fracture, since nearly all such women would be recommended an intervention [3]. Individuals without any risk factors are excluded from the CRF-FRAX assessment. By applying CRF-FRAX, individuals can be recommended for either an intervention (high risk), a BMD-FRAX assessment (intermediate risk), or reassurance and no further participation in the screening program (low risk). Those having a BMD-FRAX assessment can then be recommended an intervention if their resulting 10-year probability of major osteoporotic fracture exceeds the age-specific threshold, or they can be reassured (see Fig 2).

Fig 2

NOGG screening strategy.

Both CRF-FRAX and BMD-FRAX generate a 10-year probability of major osteoporotic fracture, which is used to designate risk of fracture. BMD-FRAX, bone-mineral-density-based Fracture Risk Assessment Tool; CRF-FRAX, clinical-risk-factor-based Fracture Risk Assessment Tool; NOGG, National Osteoporosis Guideline Group.

Despite the efficiencies gained by using this stepwise approach [31], false negatives can occur when interventions are not recommended to individuals who have a low CRF-FRAX-based probability and are discharged from subsequent screening, whereas if they had undergone BMD-FRAX, would have qualified for intervention. Likewise, false positives can arise when an individual is recommended for an intervention based on the CRF-FRAX score but would not have qualified for an intervention with BMD-FRAX.

To try to reduce the number of individuals undergoing testing, particularly more costly BMD testing, who would subsequently not require intervention, we introduced a gSOS-based screening step in the NOGG algorithm, where individuals were reassured if their gSOS was above a threshold (Fig 3). This is because individuals with a high SOS are likely to have a high BMD and are thus less likely to be recommended for an intervention. The trade-off of this strategy is that it could result in reassurance of individuals who, if their BMD was measured, would have been recommended an intervention. This would result in a decrease in sensitivity to identify individuals requiring an intervention. To calculate the sensitivity and specificity of the gSOS-modified NOGG algorithm, we used BMD-FRAX as a reference standard within the NOGG screening strategy (Fig 4). According to NOGG guidelines, women ≥50 years with a prior fragility fracture are recommended treatment without further FRAX testing. As a result, these individuals were assigned an intervention recommendation when calculating the sensitivity and specificity of correct treatment assignment (Fig 4).

Fig 3

NOGG screening strategy with a gSOS screening step.

Both CRF-based FRAX and BMD-based FRAX generate a 10-year probability of major osteoporotic fracture, which is used to designate risk of fracture. gSOS is standardized to have a mean of 0 and standard deviation of 1. BMD-FRAX, bone-mineral-density-based Fracture Risk Assessment Tool; CRF-FRAX, clinical-risk-factor-based Fracture Risk Assessment Tool; NOGG, National Osteoporosis Guideline Group.

Fig 4

Calculation of sensitivity and specificity of correct treatment assignment.

BMD-FRAX, bone-mineral-density-based Fracture Risk Assessment Tool; NOGG, National Osteoporosis Guideline Group.

Since resources are often expended to measure BMD-FRAX in individuals whose final probability of fracture is too low to warrant intervention, we also estimated the number of CRF-FRAX and BMD-FRAX tests that were performed but led to the individual being reassured without a recommended intervention.

We chose the sex-specific thresholds of gSOS that reduced CRF-FRAX and BMD-FRAX testing but minimized the loss of sensitivity to identify individuals who would be recommended for treatment. This threshold was chosen using data from the UK Biobank Test Set (S4 Fig). The generalizability of the selected gSOS threshold was then tested in the remaining 4 validation cohorts (CLSA, SOF, Mr OS US, and Mr OS Sweden). The number of CRF-FRAX and BMD-FRAX tests performed but not leading to an intervention were counted. These analyses were conducted in each validation cohort, men and women separately, and in all groups combined. We also tested individuals of non–White British ancestry in UK Biobank (N = 350), i.e., the individuals who remain subsequent to filtering out the White British subset and who have available measurements of femoral neck BMD. The characteristics are provided in Table B in S1 Tables.

Results

Cohort characteristics

Table 1 describes the FRAX risk factors for all of the cohorts. There were few clinically relevant differences in any of the osteoporosis-related risk factors in the UK Biobank Training, Model Selection, and Test Sets, as expected, since these sets were generated randomly. As planned, all individuals from UK Biobank with BMD measures were included in the UK Biobank Test Set, to ensure availability of BMD-FRAX scores as the reference standard. There were few differences in demographics or clinical risk factors between individuals with and without BMD measured. The validation cohorts (CLSA, SOF, Mr OS US, and Mr OS Sweden) provided a range of characteristics, allowing for a better assessment of the generalizability of results (Table 1).

GWAS

After quality control (see S1 Text), 13,958,249 SNPs were included in the GWAS. The GWAS in the training set identified 1,404 independent (r ≤ 0.05) genome-wide significant loci at a p-value threshold of <5 × 10−8. S1 Fig shows the Manhattan and QQ plots for this GWAS.

Variance explained in SOS in the UK Biobank Model Selection Set

The polygenic risk score models trained with LASSO explained at most 25.0% (95% CI 23.0%–27.0%) of the variance in SOS in the UK Biobank Model Selection Set (Table C in S1 Tables). S2 Fig provides detailed information on the optimal algorithm tuning parameters. None of the traditional polygenic risk scores performed better than the polygenic risk score derived from the LASSO regression. S3 Fig demonstrates that, as expected, the estimated effects of the activated SNPs from the LASSO algorithm were attenuated compared to the effects estimated from the GWAS.

Variance explained in SOS in the UK Biobank Test Set

Age, sex, and BMI explained 4.0% (95% CI 3.7%–4.2%) of the variance in SOS. Adding all available FRAX clinical risk factors increased the variance explained to 5.3% (95% CI 5.0%–5.6%). The polygenic risk score from the UK Biobank Model Selection Set explaining the most variance in measured SOS was designated as “gSOS” and was then tested for its correlation with SOS in the UK Biobank Test Set. This model explained 23.2% (95% CI 22.7%–23.7%) of the variance in measured SOS and included 21,717 SNPs activated from a total of 345,111 SNPs that had p-values for association with SOS of ≤5 × 10−4 (Table C in S1 Tables; Fig 5).

Fig 5

Variance explained in SOS by clinical risk factors and gSOS in the UK Biobank Test Set.

Available FRAX clinical risk factors included age, sex, BMI, smoking, previous fracture, use of glucocorticoids, rheumatoid arthritis, and secondary osteoporosis. BMI, body mass index; FRAX, Fracture Risk Assessment Tool; SOS, speed of sound.

Screening by NOGG guidelines in validation cohorts

The validation cohorts comprised 10,522 individuals eligible for fracture risk screening (Table 1). Both the sensitivity and specificity of the NOGG screening strategy to identify individuals at high enough risk to merit an intervention, compared to the reference standard, BMD-FRAX, were high (99.6% and 97.1%, respectively; Fig 6; Table D in S1 Tables). This high sensitivity and specificity required CRF-FRAX tests to be undertaken in 81% of the population eligible for screening, with BMD-FRAX tests subsequently recommended in 37% of the population. In total, 74% of those requiring CRF-FRAX tests were classified for reassurance, i.e., without a recommendation for an intervention. As well, just over one-third of all individuals who received a BMD-FRAX test were classified for reassurance without intervention (Fig 6; Table D in S1 Tables).

Fig 6

Performance characteristics of screening with and without a gSOS screening step.

BMD-FRAX, bone-mineral-density-based Fracture Risk Assessment Tool; CRF-FRAX, clinical-risk-factor-based Fracture Risk Assessment Tool.

Screening incorporating a gSOS-based screening step

Using the UK Biobank Test Set, we selected the threshold of gSOS that would minimize the number of BMD tests done in persons who would ultimately be reassured rather than receiving an intervention, but also would minimize the number of false negatives (S3 Fig). Applying this threshold separately in men and women, we found that a threshold of standardized gSOS set to 0.5 and 0 for men and women, respectively, balanced these goals in the UK Biobank Test Set, and subsequently individuals above these thresholds were excluded from further screening in the validation cohorts, prior to receiving a CRF-FRAX or BMD-FRAX test (Fig 3). The utility of this threshold was then tested in all validation cohorts.

Fig 6 demonstrates that applying a gSOS screening step in the validation cohorts resulted in a small decrease in sensitivity to identify eligible participants for therapy, to 93.4%, but that the specificity increased slightly, to 98.5%. However, the proportion of screened individuals requiring CRF-FRAX testing decreased from 81% to 51% (representing a relative decrease of 37%) compared to NOGG-based screening without a gSOS screening step. Additionally, the proportion of screened individuals requiring BMD-FRAX testing decreased from 37% to 22% (representing a relative decrease of 41%) (Fig 6; Table D in S1 Tables).

The proportion of CRF-FRAX and BMD-FRAX tests that resulted in an individual being excluded from the screening program without a recommendation for an intervention also decreased from 74% to 46% and from 34% to 20%, respectively (Fig 6; Table D in S1 Tables). Cohort-specific results are shown in Tables E–I in S1 Tables.

The positive predictive value for correct treatment assignment in all validation cohorts was 91.8% without a gSOS screening step and increased to 95.4% with the gSOS screening step (Table D in S1 Tables; cohort-level results and subgroup results are available in Tables D–P in S1 Tables).

Women and men separately

The SOF cohort was composed of only women, while Mr OS US and Mr OS Sweden were composed of only men, providing the opportunity to explore performance characteristics by sex. Further, we divided the UK Biobank Test Set and CLSA into sex-specific datasets (Tables J–M in S1 Tables). Amongst 4,859 women who were eligible for screening in the cohorts (SOF, UK Biobank Test Set, and CLSA), the sensitivity and specificity for correct treatment assignment were high (99.9% and 95%, respectively). Nevertheless, 58% of the population required CRF-FRAX tests, and 43% required BMD-FRAX tests (Table N in S1 Tables).

When applying a gSOS screening step, the sensitivity decreased marginally, to 94.6%, and the specificity increased marginally, to 98.2%. The proportion of the population requiring a CRF-FRAX test reduced from 58% to 27% (representing a relative decrease of 55%), while the proportion requiring a BMD-FRAX test reduced from 43% to 20% (representing a relative decrease of 55%) (Table N in S1 Tables).

Amongst the 5,668 men eligible for screening, the sensitivity and specificity were 96.9% and 98.2%, respectively, using CRF-FRAX alone as the screening step. In order to achieve this performance, 100% of men had a CRF-FRAX test, and 31% required a BMD-FRAX test. The yield of high-risk individuals from these tests was low, such that 94% of men receiving a CRF-FRAX test were reassured, and 29% of those receiving a BMD-FRAX test were reassured (Table O in S1 Tables). Applying a gSOS screening step to these men reduced the sensitivity to 82% while maintaining a similar specificity at 99%. However, the proportion of men requiring a CRF-FRAX test reduced to 72% (representing a relative decrease of 28%), and the proportion undergoing BMD-FRAX reduced to 23% (representing a relative decrease of 25%).

Stratification by age

We next tested the performance of gSOS in different age strata to understand if the screening efficiency improved more for one age group than another. Using the largest cohort, with the largest variation in age (CLSA, N = 6,704), we found that gSOS had the highest performance in individuals aged ≥70 years. Specifically, the sensitivity and specificity to identify individuals who require an intervention remained high, at 99.6% and 94.9%, respectively. The proportion of screened individuals requiring CRF-FRAX testing decreased from 73% to 37% (representing relative decrease of 49%) compared to the NOGG screening strategy without a gSOS screening step. Additionally, the proportion of screened individuals requiring BMD-FRAX testing decreased from 24% to 12% (representing a relative decrease of 50%) (Table F in S1 Tables). In contrast, in individuals aged 50–59 years, sensitivity reduced to 86%, but specificity was 99.6%. The percent of individuals requiring CRF-FRAX and BMD-FRAX testing reduced by 51% and 50%, respectively. This demonstrates that gSOS pre-screening improves the efficiency of screening, but that the sensitivity to correctly identify individuals requiring therapy is maximized in older age groups.

Non–White British individuals

We then assessed the effect of a gSOS pre-screening in individuals from UK Biobank with dual-energy X-ray absorptiometry BMD measures who were of non–White British ancestry (Table B in S1 Tables). We found that the results were generally consistent with those in individuals of White British ancestry. Specifically, the proportion of screened individuals requiring CRF-FRAX testing decreased from 94% to 48% (representing a relative decrease of 49%) compared to NOGG-based screening without a gSOS screening step. Additionally, the proportion of screened individuals requiring BMD-FRAX testing decreased from 39% to 17% (representing a relative decrease of 57%) (Table P in S1 Tables).

The proportion of CRF-FRAX and BMD-FRAX tests that resulted in an individual being excluded from the screening program without a recommendation for an intervention also decreased from 92% to 47% and from 38% to 16%, respectively (Table P in S1 Tables).

Discussion

By building a polygenic risk score using 341,449 individuals and validating its utility in fracture risk screening in 5 separate cohorts totaling 10,522 individuals, we determined that genomics-enabled fracture risk screening could reduce the proportion of people who require BMD-based testing by 41%, while maintaining a high overall sensitivity and specificity for correct treatment assignment. While these findings are not meant to be prescriptive, they indicate the possible utility of polygenic risk scores in screening programs that are dependent on heritable risk factors.

Fracture risk assessment is expensive, with estimates of approximately US$50,000 per quality-adjusted life year gained [32], but is less expensive, or even cost-saving, using NOGG assessment strategies [33,34], because NOGG decreases the number of individuals who require CRF-FRAX and BMD-FRAX testing. Current guidelines suggest testing a large proportion of the population [2,3,5], yet most screened individuals are not identified as having a clinically actionable level of fracture risk [9,35]. This circumstance provides an opportunity for genetically derived measures of risk to increase cost-efficiencies in healthcare systems where investments have been made in genome-wide genotyping. Already at least 7 large healthcare systems have invested in genome-wide genotyping of a large proportion of their population, among whom electronic health record data are available [36,37]. Since the costs associated with genome-wide genotyping have now dropped below those of several routine clinical tests, the use of polygenic risk scores could be particularly helpful in these environments since a one-time genotyping cost could be used to generate several polygenic risk scores. However, there is a clear need to study the translation of such polygenic risk scores to clinical applications [38]—and the work presented here provides one example of how such scores could be translated to the clinic.

Previous attempts to predict osteoporosis from genomic data did not substantially increase discrimination compared to standard clinical measures alone, likely because the GWAS that underpinned these attempts was derived from 32,961 individuals and explained only 5.8% of the variance in BMD [39,40]. The improvement in variance explained in this study was attributable to the increase in sample size afforded by UK Biobank and to the LASSO regression’s ability learn SNP associations with SOS jointly, as opposed to summing over independently measured effects on BMD. Other attempts to predict BMD have been based on several dozen genome-wide significant SNPs [39], whereas our approach used machine learning to jointly consider the effects of 642,127 SNPs (Table C in S1 Tables). LASSO regression has recently been used to predict estimated BMD, but from a GWAS sample size that was one-third of that used here, explaining only 17.2% of the BMD variance, and it was not used in a screening program [14]. Our work has improved the genomic prediction of BMD and demonstrated its potential clinical relevance.

We observed similar predictive performance across all LASSO models in the model selection step (Table C in S1 Tables); therefore, it remains possible that a more parsimonious model containing fewer SNPs would perform as well. As a result, further exploration of these LASSO models is warranted in a future technical study. However, should a more complex model with more SNPs prove to be ideal, the hinderance to clinical translation should be minimal, as the computational burden is limited to the training of the models, and is not in the prediction of an individual’s genetic risk.

The sensitivity and specificity to correctly assign intervention was maximized in individuals ≥70 years of age. This could be clinically relevant because this is the age range for which the SCOOP trial demonstrated that a community-based screening program could be effective in reducing hip fractures [9].

We acknowledge that for many practicing physicians, such as those in the UK, who have access to an automatically generated electronic-health-record-based CRF-FRAX test, the result of interest would be the reduction in BMD-FRAX tests. However, we observed no appreciable difference in the sensitivity and specificity to correctly identify individuals requiring therapy if the gSOS screening step was placed prior to the CRF-FRAX test or immediately after the CRF-FRAX test. Tables E–O in S1 Tables show the results for a reduction in BMD-FRAX tests by cohort and sex.

Limitations

We have generated a polygenic risk score for SOS, rather than BMD, since there are insufficient data resources to generate such a score for BMD. Nevertheless, the correlation between SOS and BMD has enabled the identification of individuals unlikely to have a BMD low enough to warrant an intervention. Further refinement could improve the efficiencies presented here, including a polygenic risk score for BMD, when sample sizes are large enough to enable this. While nearly all FRAX risk factors were available for study, alcohol intake and parental history of fracture were not available from the UK Biobank cohorts. However, these were available in the other validation cohorts. Secondary causes of osteoporosis were not uniformly available in SOF, Mr OS US, and Mr OS Sweden. Nevertheless, CLSA provided similar results to other cohorts and had all required information. Like participants in most cohort studies, the participants used in these studies are, on average, healthier than the general population [41]. Thus, external validation in a truly population-based study may provide helpful estimates of the real-world performance of genomics-enabled fracture risk screening. While we have tested the utility of gSOS in individuals of non–White British ancestry, the sample size available for study was relatively small, and thus results should be replicated in additional cohorts of different ancestry, underlining the need for large-scale GWAS datasets in individuals of non-European ancestry [42]. We recognize that different approaches could be taken to incorporate polygenic risk scores into fracture risk screening, but here we offer a simple approach that could be readily implemented in a genotyped population with required FRAX risk factors using the NOGG strategy [9].

Conclusions

In summary, we have developed and tested gSOS, a polygenic risk score for SOS, which when introduced into a fracture risk screening program decreased the number of people requiring CRF-FRAX and BMD-FRAX assessments, while still maintaining a high sensitivity and specificity to identify individuals in whom an intervention should be recommended. These findings highlight the role that genetic prediction could play in screening programs that rely upon heritable risk factors.

GRIPS Checklist.

(DOCX)

Click here for additional data file.

GWAS of SOS.

(A) Manhattan plot from GWAS of SOS. (B) QQ plot from GWAS of SOS.

(TIFF)

Click here for additional data file.

Performance of each SNP set using LASSO regression in the model selection set.

Each feature set consists of a set of SNPs associated with SOS at a specified p-value threshold (sub-panel titles). For each feature set, we fit a regularized model to the training set over a range of regularization constants (λ) (top left), with each λ resulting in a variable subset of activated features (bottom left). The model with the minimal root mean square error in the model selection set (top right) was selected to compare the variance explained (r2, bottom right among all feature sets.

(TIFF)

Click here for additional data file.

Correlation of effect estimates from the GWAS and the coefficients from the LASSO regression for activated SNPs.

Activated SNPs are those SNPs chosen by the machine learning algorithm to be in gSOS, the final selected model.

(TIFF)

Click here for additional data file.

Effects of gSOS threshold on treatment assignment.

Results stratified by women (top) and men (bottom).

(TIFF)

Click here for additional data file.

Supplemental tables.

(XLSX)

Click here for additional data file.

Supplemental text.

(DOCX)

Click here for additional data file.

4 Feb 2020

Dear Dr. Richards,

Thank you very much for submitting your manuscript "A Polygenic Risk Score to Improve Screening for Fracture Risk" (PMEDICINE-D-19-04448) for consideration at PLOS Medicine.

Your paper was discussed among the editorial team, evaluated by an academic editor with relevant expertise, and sent to independent reviewers, including a statistical reviewer. The reviews are appended at the bottom of this email and any accompanying reviewer attachments can be seen via the link below:

[LINK]

In light of these reviews, we will not be able to accept the manuscript for publication in the journal in its current form, but we would like to invite you to submit a revised version that fully addresses the reviewers' and editors' comments. You will appreciate that we cannot make a decision about publication until we have seen the revised manuscript and your response, and we expect to seek re-review by one or more of the reviewers.

In revising the manuscript for further consideration, your revisions should address the specific points made by each reviewer and the editors. Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. In your rebuttal letter you should indicate your response to the reviewers' and editors' comments, the changes you have made in the manuscript, and include either an excerpt of the revised text or the location (eg: page and line number) where each change can be found. Please submit a clean version of the paper as the main article file; a version with changes marked should be uploaded as a marked up manuscript.

In addition, we request that you upload any figures associated with your paper as individual TIF or EPS files with 300dpi resolution at resubmission; please read our figure guidelines for more information on our requirements: http://journals.plos.org/plosmedicine/s/figures. While revising your submission, please upload your figure files to the PACE digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at PLOSMedicine@plos.org.

We hope to receive your revised manuscript by Feb 25 2020 11:59PM. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

***Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.***

We ask every co-author listed on the manuscript to fill in a contributing author statement, making sure to declare all competing interests. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. If new competing interests are declared later in the revision process, this may also hold up the submission. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. YOU MUST NOT ADD OR REMOVE AUTHORS UNLESS YOU HAVE ALERTED THE EDITOR HANDLING THE MANUSCRIPT TO THE CHANGE AND THEY SPECIFICALLY HAVE AGREED TO IT. You can see our competing interests policy here: http://journals.plos.org/plosmedicine/s/competing-interests.

Please use the following link to submit the revised manuscript:

https://www.editorialmanager.com/pmedicine/

Your article can be found in the "Submissions Needing Revision" folder.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see http://journals.plos.org/plosmedicine/s/submission-guidelines#loc-methods.

Please ensure that the paper adheres to the PLOS Data Availability Policy (see http://journals.plos.org/plosmedicine/s/data-availability), which requires that all data underlying the study's findings be provided in a repository or as Supporting Information. For data residing with a third party, authors are required to provide instructions with contact information for obtaining the data. PLOS journals do not allow statements supported by "data not shown" or "unpublished results." For such statements, authors must provide supporting data or cite public sources that include it.

Please let me know if you have any questions. Otherwise, we look forward to receiving your revised manuscript.

Sincerely,

Richard Turner PhD, for Louise Gaynor-Brook, MBBS PhD

Associate Editor, PLOS Medicine

rturner@plos.org

-----------------------------------------------------------

Requests from the editors:

Throughout your text, please adapt the wording to reflect the experimental design rather than anticipations of possible use dependent on further clinical studies. For example, "The use of a polygenic risk score ... can reduce ..." seems too stridently worded.

Please add a study descriptor to the title. We suggest: "Development of a polygenic risk score to improve screening for fracture risk: a cross-sectional study".

Please convert the abstract to PLOS Medicine style, i.e., with three subsections. The final sentence of the "methods and findings" subsection should summarize the study's main limitations.

Please begin the "Conclusions" subsection of your abstract with "In this study, we found that ..." or similar.

Please quote summary demographic and clinical details for study participants in your abstract.

In your abstract and elsewhere in the paper, please add p values alongside 95% CI where available.

After your abstract, we will need to ask you to add a new and accessible "author summary" section in non-identical prose. You may find it helpful to consult one or two recently published research papers in PLOS Medicine to get a sense of the preferred style.

Early in your methods section, please state whether the study had a protocol or prespecified analysis plan and if so attach the relevant document(s) as a supplementary file (referred to in the text). Please highlight analyses that were not prespecified.

Please remove trademarks from the manuscript.

Throughout the text, please format reference call-outs as follows: "... that for BMD [10,24]."

Please ensure that citations in your reference list meet journal style. All boldface and italic elements should be converted into plain text. Where appropriate, 6 author names should be listed, followed by "et al.".

Please add a journal name to reference 37 as appropriate.

If references 41 and 43 are preprints, please indicate this.

Please add a completed checklist for the most appropriate reporting guideline, which may be STROBE, referred to in your methods section. In the checklist, individual items should be referred to by section (e.g., "Methods") and paragraph number rather than by page or line numbers, as the latter generally change in the event of publication.

Comments from the reviewers:

*** Reviewer #1:

The authors present a well-written and interesting study on developing a polygenic risk score for heal quantitative ultrasound speed of sound (SOS) for identify low risk individuals who can excluded from fracture risk screening. The strengths of the study include, developing and testing the genetic risk score (GRS) in a large population cohort (UK Biobank), interestingly using both a LASSO-based data-driven approach, and then validating the performance of the GRS in 5 other cohorts with FRAX-score risk factors available , including Bone Mineral Density. The key finding being that there were substantial reductions in proportion of individuals requiring FRAX-based tests (with or without BMD).

There are quite a few analyses in this and some of the results are difficult to follow as they are given in excel workbooks instead of word format. However, the statistical methods appear sound - with appropriate QC of genotyping data and SNP selection. The introduction of the LASSO model compared to simple polygenic risk score is also a nice addition and showed that using a data-driven approach can improve the variance in the model (up to 25%).

Much of the literature on polygenic risk scores has been done in the context of "how much added values does GRS have above and beyond a clinical prediction model" with a focus on the overall change in AUC. The authors have framed their research question differently, in particular about how a GRS could preclude people from undertaking FRAX risk assessment (which have some level of costs). Whilst the modelling is convincing comparing these strategies - I am still left wondering about the clinical utility and feasibility of such an approach given that genotyping will need to be done on ALL individuals. Albeit the authors mentioned genotyping is more available now but that is really limited still to research based studies - it not a strategy which is indeed offered as routine clinical care.

Specific comments:

(1) The project clearly shows the potential of using polygenic risk score but it would be useful if the authors could comment about the clinical utility of their approach. There is a growing drive for genomic testing in routine practice but given that their LASSO model included 21,717 activated SNPs mean that's genotyping arrays of this magnitude will need to be utilised - which will indeed a barrier for implementation. The authors themselves had to use Amazon cloud computing to conduct the analyses. I'm wondering if a reduced set of SNPs would confer similar set of results. In Table S2, the LASSO approach with 6,823 SNPS explains about 23.4% which is not far off the 24.99% achieved by having 21,717 SNPs. There is in fact an argument for using more simplistic GRS scores with fewer SNPS.

(2) Please check text results section (Lines 344) as it says the SNP array with 21,717 activated SNPS has a r^2 of 23.2% but in Table S2 the r^2 for the 21,717 SNPS

(3) CRF-FRAX is significantly cheaper to apply without BMD - in fact many EHRs have this embedded within medical record. In the UK for instance, TPP offers CRF-FRAX and EMIS offers QFracture using data which already collected in patient's medical records. There is logical mismatch in argument made by the authors in the discussion on "requiring all individuals to undergo a CRF-FRAX test" as the scores are CRF-FRAX or QFractures scores are estimated instantaneously from the medical record. The reductions in individuals requiring CRF-FRAX is actually not a huge burden with algorithms being integrated into EHRs. The more compelling argument here is the reduction of BMD-based FRAX tests.

(4) I found it interesting the authors looked at placing the genomic screening step after the CRF-FRAX and similar to other studies which have looked at the added value of the GRS on top of the clinical prediction model, they found no further improvement in sensitivity and specificity. As a population screening model the key comparison should be the CRF-FRAX model as this has been embedded in primary care EHRs at very minimal costs. The current comparison is BMD-FRAX but the more commonly and cheaper alternate which is adopted in primary care is CRF-FRAX due to BMD not being routinely available.

*** Reviewer #2:

Forgetta et al., 2019

Forgetta et al. analyze data from UK Biobank to derive a polygenic risk score (gSOS) for heel ultrasound speed of sound. They show that the polygenic risk score can be used to reduce the number of patients that require DXA scans for bone mineral density in the assessment of fracture risk.

This paper tests gSOS in clinical cohorts. The authors show that gSOS can be used to eliminate the need for DXA scans for some patients, thereby decreasing the cost of screening a general population. The authors set a threshold for gSOS such that patients above the threshold are "reassured" and do not require further screening. The key result can be seen in Fig. 6, which shows that incorporating gSOS retains sensitivity and specificity but decreases the number of patients that require CRF-FRAX and BMD-FRX. The figure clearly shows the cost savings of genetic testing in diagnosing osteoporosis.

I reviewed a previous version of this paper, and the current submission has already responded to many of my comments.

Minor comments

Line 441 "Recently, LASSO regression was used to predict estimated BMD, but from a GWAS sample size that was one third of that used here, captured only 20% of the its variance and was not used in a screening program.41" This statement should refer to reference 14, which developed gSOS previously based on the identical UK Biobank cohort and with nearly identical performance (i.e. correlation to SOS) as in this submitted work.

In Table 1, some of the demographics differ between groups and should have comments:

1. The percent smoker in the Biobank test set is much different than the training or selection set. Smoker is a key factor for FRAX, and this difference could impact performance of gSOS.

2. Age in the four clinical cohorts is much higher than the Biobank study.

3. The percent of previous fracture in the four clinical cohorts is higher than in the UK Biobank. Previous fracture is a very strong risk factor for osteoporotic fractures. This difference likely impacts the results.

4. The percent of falls in the Biobank test set was higher than in the training or marker selection set. This is very worrisome as it suggests that the test set was at higher risk for fragility fracture than the training set.

Table 1 should list 95% CI for each of the percentages to show which of the difference are statistically significant.

*** Reviewer #3:

Forgetta et al. have developed and tested a polygenic prediction tool for bone fracture, based on the large UK Biobank sample (341,449 individuals) and a polygenic risk score for heel quantitative ultrasound speed of sound (SOS). The paper is timely and interesting, and shows that genetic testing may reduce the need for additional screening tests. Some issues should be addressed:

1) The use case can be improved. The risk for an event is less clinically relevant than prediction the time of the event. It would be of interest to see the results with for example the polygenic hazard score in this context (published in PLOS Medicine in 2017).

2) Further, in what age groups are the gSOS most predictive? This should be added to the model.

3) What is the positive predictive value of gSOS? These result should be presented.

4) Outcome: the sensitivity and specificity did not improve, while the proportion of individuals requiring CRF-FRAX testing decreased from 81% to 51% and the proportion of screened individuals requiring BMD-FRAX testing decreased from 37% to 22%. However, this seems to be age dependent, and for which age group is this relevant?

5) The sensitivity and specificity of gSOS is not better than the for the other measures (a small decrease in sensitivity to 93.4%, while the specificity increased slightly to 98.5%). This result should be discussed, and the reason for lack of improvement should be presented.

6) Earlier intervention is often better for prevention. Does the gSOS improve the predictive value of BMD-FRAX or CRF-FRAX if applied to younger age groups? This is a highly relevant analysis.

7) UK Biobank - the GWAS did not include age as variable? Is it possible to provide result of age dependent GWAS, at least for different age groups?

8) There should be non-European samples available that the authors can use for testing the performance in other ancestries.

***

Any attachments provided with reviews can be seen via the following link:

[LINK]

27 Mar 2020

Submitted filename: r1-PLOSMED-gSOS-rebuttal.docx

Click here for additional data file.

19 Apr 2020

Dear Dr. Richards,

Thank you very much for submitting your revised manuscript 'Development of a polygenic risk score to improve screening for fracture risk' (PMEDICINE-D-19-04448R1) for consideration at PLOS Medicine.

Your paper was discussed with our academic editor and was also seen by one peer reviewer, whose comments you can read at the bottom of this email. Any accompanying reviewer attachments can be seen via the link below:

[LINK]

We would now like to invite you to submit a further revised version that fully addresses the reviewer's and editors' comments. Please note that we cannot make a decision about publication until we have seen the revised manuscript and your response, and we may seek re-review by one or more reviewers.

In revising the manuscript for further consideration here, your revisions should address the specific points made by each reviewer and the editors. Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. In your rebuttal letter you should indicate your response to any comments from reviewers or editors and the changes you have made in the manuscript. Please submit a clean version of the paper as the main article file; a version with changes marked should be uploaded as a marked up manuscript.

In addition, we request that you upload any figures associated with your paper as individual TIF or EPS files with 300dpi resolution at resubmission; please read our figure guidelines for more information on our requirements: http://journals.plos.org/plosmedicine/s/figures.

While revising your submission, please upload your figure files to the PACE digital diagnostic tool, https://pace.apexcovantage.com/ PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at PLOSMedicine@plos.org.

We hope to receive your revised manuscript by May 08 2020 11:59PM. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

We ask every co-author listed on the manuscript to fill in a contributing author statement, making sure to declare all competing interests. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. If new competing interests are declared later in the revision process, this may also hold up the submission. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. YOU MUST NOT ADD OR REMOVE AUTHORS UNLESS YOU HAVE ALERTED THE EDITOR HANDLING THE MANUSCRIPT TO THE CHANGE AND THEY SPECIFICALLY HAVE AGREED TO IT. You can see our competing interests policy here: http://journals.plos.org/plosmedicine/s/competing-interests.

Please use the following link to submit the revised manuscript:

https://www.editorialmanager.com/pmedicine/

Your article can be found in the 'Submissions Needing Revision' folder.

Please let me know if you have any questions. Otherwise, we look forward to receiving your revised manuscript shortly.

Sincerely,

Richard Turner PhD

Senior editor, PLOS Medicine

rturner@plos.org

-----------------------------------------------------------

Requests from the editors:

To your data statement, please add the UK Biobank contact details for those wishing to inquire about access to data.

Please add a study descriptor to your title, following a colon, e.g. "...: a genetic risk prediction study".

Please add a new concluding sentence to the "methods and findings" subsection of your abstract, quoting 2-3 of the study's main limitations.

If not done, please add a sentence to the methods section to note that the study did not have a pre-specified analysis plan.

Throughout the paper, please format reference call-outs as follows, preceding punctuation: "... [1].".

Please add p values alongside 95% CI, where available.

Please remove all trade marks, e.g., at line 67.

Please present the reporting checklist in a separate supplementary document, referred to in your methods section.

Comments from the reviewers:

*** Reviewer #3:

The authors have addressed my comments adequately, except for the points about prediction which seems to be based on a misunderstanding.

I understand they are not predicting fracture. However, they are predicting an event: who requires an intervention ("be eligible for an intervention", "identify individuals requiring treatment".)

They are estimating the sensitivity and specificity for this event.

Thus, they should be able to provide

1) Estimate the time to the event (requiring treatment), or this is a limitation

2) The age group where the sensitivity and specificity is highest for the event (to identify individuals requiring treatment),

3) The positive predictive value for the event (requiring treatment)

They should provide standard clinical measures for prediction performance to identify individuals who require an intervention based on their quantitative measure. Otherwise, the approach has no clinical interest

***

Any attachments provided with reviews can be seen via the following link:

[LINK]

29 Apr 2020

Submitted filename: r2-PLOSMED-gSOS-rebuttal.docx

Click here for additional data file.

8 May 2020

Dear Dr. Richards,

Thank you very much for re-submitting your manuscript "Development of a polygenic risk score to improve screening for fracture risk: a genetic risk prediction study" (PMEDICINE-D-19-04448R2) for consideration at PLOS Medicine.

We have discussed the paper with our academic editor and it was also seen again by one reviewer. I am pleased to tell you that, provided the remaining editorial and production issues are fully dealt with, we expect to be able to accept the paper for publication in the journal.

The remaining issues that need to be addressed are listed at the end of this email. Any accompanying reviewer attachments can be seen via the link below. Please take these into account before resubmitting your manuscript:

[LINK]

Our publications team (plosmedicine@plos.org) will be in touch shortly about the production requirements for your paper, and the link and deadline for resubmission. DO NOT RESUBMIT BEFORE YOU'VE RECEIVED THE PRODUCTION REQUIREMENTS.

***Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.***

In revising the manuscript for further consideration here, please ensure you address the specific points made by each reviewer and the editors. In your rebuttal letter you should indicate your response to the reviewers' and editors' comments and the changes you have made in the manuscript. Please submit a clean version of the paper as the main article file. A version with changes marked must also be uploaded as a marked up manuscript file.

Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. If you haven't already, we ask that you provide a short, non-technical Author Summary of your research to make findings accessible to a wide audience that includes both scientists and non-scientists. The Author Summary should immediately follow the Abstract in your revised manuscript. This text is subject to editorial change and should be distinct from the scientific abstract.

We hope to receive your revised manuscript within 1 week. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

We ask every co-author listed on the manuscript to fill in a contributing author statement. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. YOU MUST NOT ADD OR REMOVE AUTHORS UNLESS YOU HAVE ALERTED THE EDITOR HANDLING THE MANUSCRIPT TO THE CHANGE AND THEY SPECIFICALLY HAVE AGREED TO IT.

Please ensure that the paper adheres to the PLOS Data Availability Policy (see http://journals.plos.org/plosmedicine/s/data-availability), which requires that all data underlying the study's findings be provided in a repository or as Supporting Information. For data residing with a third party, authors are required to provide instructions with contact information for obtaining the data. PLOS journals do not allow statements supported by "data not shown" or "unpublished results." For such statements, authors must provide supporting data or cite public sources that include it.

Please let me know if you have any questions. Otherwise we look forward to receiving the revised manuscript shortly.

Sincerely,

Richard Turner PhD

Senior editor, PLOS Medicine

rturner@plos.org

------------------------------------------------------------

Requests from Editors:

Please truncate the short title.

Thank you for adding an additional sentence at the end of the "methods and findings" subsection of your abstract. However, journal style calls for a sentence that summarizes limitations of the present study rather than detailing future research plans. Please substitute a sentence such as "Study limitations include a reliance on cohorts of predominantly European ethnicity, and use of a proxy of fracture risk.".

Our academic editor commented that, in view of the in silico research design, the language used regarding implied clinical utility would need to be amended prior to publication. For example, at line 34 rather than "... we found that the use of ... decreased ...", more apt wording would be "... our results suggest that the use of ... could decrease the number of individuals ...". Similarly, at line 50, the wording should be changed to "... we estimate that reassuring individuals ... could reduce ...".

Please review the entire manuscript and make similar changes (e.g., at lines 386 the text could be amended to "... screening could reduce the proportion ..."; and at line 388 "... they indicate the possible utility of polygenic risk scores...").

Around line 145, please add "Specific ethics approval was not required for this study" or similar.

Please make that "characteristics are" at line 264.

Please adapt the formatting of the reference call-outs throughout the text ("... are available [37,38]. Since the ...").

Please add full author and access details to reference 10.

Please break the completed GRIPS checklist out into a separate supplementary file, referred to in your methods section as "See S1_GRIPS_Checklist" or similar.

Comments from Reviewers:

*** Reviewer #3:

The remaining issues have now been addressed

***

Any attachments provided with reviews can be seen via the following link:

[LINK]

3 Jun 2020

Dear Dr. Richards,

On behalf of my colleagues and the academic editor, Dr. Christelle Nguyen, I am delighted to inform you that your manuscript entitled "Development of a polygenic risk score to improve screening for fracture risk: a genetic risk prediction study" (PMEDICINE-D-19-04448R3) has been accepted for publication in PLOS Medicine.

PRODUCTION PROCESS

Before publication you will see the copyedited word document (in around 1-2 weeks from now) and a PDF galley proof shortly after that. The copyeditor will be in touch shortly before sending you the copyedited Word document. We will make some revisions at the copyediting stage to conform to our general style, and for clarification. When you receive this version you should check and revise it very carefully, including figures, tables, references, and supporting information, because corrections at the next stage (proofs) will be strictly limited to (1) errors in author names or affiliations, (2) errors of scientific fact that would cause misunderstandings to readers, and (3) printer's (introduced) errors.

If you are likely to be away when either this document or the proof is sent, please ensure we have contact information of a second person, as we will need you to respond quickly at each point.

PRESS

A selection of our articles each week are press released by the journal. You will be contacted nearer the time if we are press releasing your article in order to approve the content and check the contact information for journalists is correct. If your institution or institutions have a press office, please notify them about your upcoming paper at this point, to enable them to help maximize its impact.

PROFILE INFORMATION

Now that your manuscript has been accepted, please log into EM and update your profile. Go to https://www.editorialmanager.com/pmedicine, log in, and click on the "Update My Information" link at the top of the page. Please update your user information to ensure an efficient production and billing process.

Thank you again for submitting the manuscript to PLOS Medicine. We look forward to publishing it.

Best wishes,

Richard Turner, PhD

Senior Editor

PLOS Medicine

plosmedicine.org