A Genetic Predictive Model Estimating the Risk of Developing Adolescent Idiopathic Scoliosis.

BACKGROUND: Previous GWASs have revealed several susceptible variants associated with adolescent idiopathic scoliosis (AIS). Risk prediction based on these variants can potentially improve disease prognosis. We aimed to evaluate the combined effects of genetic factors on the development of AIS and to further develop a genetic predictive model.
METHODS: A total of 914 AIS patients and 1441 normal controls were included in the discovery stage, which was followed by the replication stage composed of 871 patients and 1239 controls. Genotyping assay was performed to analyze 10 previously reported susceptible variants, including rs678741 of LBX1, rs241215 of AJAP1, rs13398147 of PAX3, rs16934784 of BNC2, rs2050157 of GPR126, rs2180439 of PAX1, rs4940576 of BCL2, rs7593846 of MEIS1, rs7633294 of MAGI1 and rs9810566 of TNIK. Logistic regression analysis was performed to generate a risk predictive model. The predicted risk score was calculated for each participant in the replication stage.
RESULTS: The association of the 10 variants with AIS was successfully validated. The established model could explain approximately 7.9% of the overall variance. In the replication stage, patients were found to have a remarkably higher risk score as compared to the controls (44.2 ± 14.4 vs. 33.9 ± 12.5, p <0.001). There was a remarkably higher proportion of the risk score i.e. >40 in the patients than in the controls (59% vs. 28.9%, p <0.001).
CONCLUSION: Risk predictive model based on the previously reported genetic variants has a remarkable discriminative power. More clinical and genetic factors need to be studied, to further improve the proba-bility to predict the onset of AIS.

INTRODUCTION

As a complex disease, the genetic background of adolescent idiopathic scoliosis (AIS) has been extensively investigated [1-7]. Many variants associated with AIS were reported through genetic association studies [8-14]. The combined effects of genetic variants may have a better prediction of the complex disease. Herein, early detection and risk prediction based on a panel of genetic markers could potentially improve the prognosis of AIS by allowing gene-based treatment or spurring them to modify lifestyle habits.

Ward et al. [15] developed the first genetic diagnostic kit for AIS that was registered as “ScoliScore”. It consisted of 53 genetic markers, and a numeric risk can be calculated to determine the probability of curve progression. Recently, a large-sale replication study of ScoliScore was performed in the Chinese population to validate its reliability [16]. However, none of the SNPs were replicated, indicating that ScoliScore might not be applicable to the Chinese AIS patients [16]. Similarly, lack of validation of ScoliScore was also reported in 3 other studies performed in the Japanese, the Caucasian and the Canadian population [17-19]. Apparently, to ensure the reliability of the diagnostic kit, the association of genetic markers with AIS should be well-refined before being embedded in the predictive model.

Previous GWASs have revealed several susceptible variants associated with AIS [8-11, 14, 20]. Zhu et al. [12, 14] reported 6 novel genes associated with AIS in the Chinese population, including PAX3, AJAP1, TNIK, MAGI1, MEIS1 and BCL2. GWAS performed in Japanese population revealed that LBX1, GPR126 and BNC2 were the susceptible genes of AIS [9, 10, 20]. Sharma et al. [11] reported that PAX1 was involved in the development of AIS in both Caucasians and Japanese. It is noteworthy that all the 4 genes have been successfully replicated in the Chinese population [12, 21-23]. It is thus plausible to predict the risk of AIS based on a combination of these 10 genetic variants mentioned above. This study aimed to evaluate the independent and combined effects of previously reported GWAS variants on the development of AIS and to develop a genetic predictive model.

METHODS

Subjects

Under the approval of a local Institutional Review Board, the current two-stage case-control study comprised of 1785 female patients and 2680 healthy controls from the Chinese Han population. Scoliosis patients who visited our clinic between April 2012 and January 2017 were retrospectively reviewed. The inclusion criteria were as follows: 1. Diagnosed with AIS through clinical and radiological examinations; 2. Aged between 10 and 16 yrs. Patients having scoliosis secondary to known etiology, including congenital scoliosis, neuromuscular scoliosis, scoliosis secondary to skeletal dysplasia, or connective tissue abnormalities were excluded from this study. Healthy subjects undergoing a physical examination prior to admission to high school were recruited as normal controls. A total of 914 patients and 1441 normal controls were assigned to the discovery stage, which was followed by the replication stage comprising of 871 patients and 1239 controls. The age was recorded for each subject. The curve severity was recorded as the Cobb angle measured at the latest visit. Guardians of all the participants signed an informed consent.

Genotyping of the Target SNP

The extraction of blood and DNA was approved by the local Institutional Review Board. Informed consent was obtained from the subjects or their guardians. Genomic DNA was extracted under standard protocols (Qiagen K.K., Tokyo, Japan). Genotyping assay was performed to analyze 10 previously reported susceptible variants, including rs678741 of LBX1, rs241215 of AJAP1, rs13398147 of PAX3, rs16934784 of BNC2, rs2050157 of GPR126, rs2180439 of PAX1, rs4940576 of BCL2, rs7593846 of MEIS1, rs7633294 of MAGI1 and rs9810566 of TNIK. The results were read with an ABI PRISM 7900HT sequence detection system (Applied Biosystems, Foster City, CA) as previously described [16]. Twenty percent of the samples were randomly selected to validate the reproducibility of the genotyping results.

Development and Replication of the Predictive Model

A predictive regression model was developed according to the independent contribution of genetic variants to the risk of AIS. The model was replicated in a different cohort of patients and controls. A predicted risk score was calculated for each participant, theoretically distributed between 0 and 1. For the convenience of understanding, the risk score was subsequently transformed by multiplying with 100.

Statistical Analysis

The Hardy-Weinberg Equilibrium (HWE) test was performed for all the subjects. The student t-test was used to compare the demographic data between the cases and the controls. The differences of genotype and allele distributions between the cases and the controls were compared using the Chi-square test. Odds ratio (OR) was calculated using the minor allele as a reference. Logistic regression analysis was performed to generate a risk predictive model. Sensitivity and specificity were calculated to maximize the sum of the 2 values using receiver operating characteristic (ROC) curves. The average risk score was compared between the cases and the controls with the student t-test. The SPSS software (version 16.0, Chicago, IL) was used for statistical analyses. Statistical significance was set at a p-value of less than 0.05.

RESULTS

Demographic Data

The two groups were matched in terms of age. The mean Cobb angle of the patients was 35.4 ± 18.3 degrees (range 25 - 73 degrees) and the mean age was 12.3 ± 1.9 years (range 10.3 - 16 years).

Genetic Association Analysis

All the 10 SNPs were successfully genotyped for all the participants. HWE test showed no significant difference in genotype frequencies in both patients and normal controls. All the SNPs were found to have significant differences in terms of allele frequency and genotype frequency between the two groups (Table ).

Establishment of the Predictive Model

The results of the logistic regression analysis are shown in Table . All the 10 SNPs were enrolled in the final model with statistical significance. The model was therefore established using the following formula:

p = 1/[1 + e^(−(−2.186 − 0.239 ∗ rs241215 + 0.178 ∗ rs7593846 + 0.405 ∗ rs13398147 + 0.222 ∗ rs7633294 + 0.311 ∗ rs9810566 + 0.199 ∗ rs2050157 + 0.310 ∗ rs16934784 + 0.388 ∗ rs678741 + 0.326 ∗ rs4940576 + 0.251 ∗ rs2180439) ) ].

The Hosmer-Lemeshow test showed no significant deviation between the observed and predicted values (p = 0.26). The value of Cox & Snell R Square was 0.079, indicating that the predictive model can explain approximately 7.9% of the overall variance. As shown in Table , with a cut-off set at 0.5, the predictive model yielded 85.9% sensitivity and 31.3% specificity.

Replication of the Predictive Model

For the replication population composed of 871 patients and 1239 controls, the mean risk score was averaged 44.2 ± 14.4 (range, 8.2 - 79.1) for the patients and 33.9 ± 12.5 (range, 11.0 - 83.7) for the controls. As shown in (Fig. ), patients were found to have a remarkably higher risk score as compared to the controls (p <0.001).

As summarized in Table , there were significantly different distributions of the risk score between the patients and the controls. There was a remarkably higher proportion of the risk score i.e. >40 in the patients than in the controls (59% vs. 28.9%, p <0.001). As shown in (Fig. ), the ROC analysis indicated that a score of 43 was the most discriminative cut-off point. The area under the curve (AUC) was 0.705 for the replication population.

DISCUSSION

In the current study, a predictive model for AIS was constructed on the basis of 10 susceptible genes discovered through GWAS. Our results suggested that the mathematic formula may serve as a helpful tool to identify the individuals at high risk of AIS prior to the onset of the spinal curve. To the best of our knowledge, this is the first predictive model developed for AIS in the Chinese population. As expected, we demonstrated independent significant associations between AIS and known genetic polymorphisms as reported by previous GWASs. The results of allele frequencies and the ORs for each variant were comparable with those of the previously published papers [12, 14, 21-23].

Targeting high-risk individuals could facilitate more frequent surveillance and active clinical conservative interventions. Moreover, patients would benefit from a more targeted therapy for a healthy lifestyle. Overall, our predictive model can explain 7.9% of the whole variance of AIS. With a cut-off value set at 0.5, the predictive model yielded 85.9% sensitivity and 31.3% specificity. As shown in the replication stage, most of the patients had a risk score of more than 40, with 15% of them exceeding 60. By contrast, only 2.8% of the controls had a risk score of more than 60. Apparently, a risk score of more than 60 was indicative of an extremely high risk of developing AIS. For children with such a high risk score, especially those with a family history of AIS, a close observation during growth spurt until skeletal maturity seems necessary.

Commonly, the AUC is recognized as the measure of the discriminatory power of the diagnostic model. Predictive models with an AUC above 0.75 are considered as clinically useful, and models with an AUC above 0.90 are considered as excellent [24]. Herein, it appears that the current risk model has a remarkable discrimination power but is still insufficient for clinical diagnostic use. We speculated that insufficient discrimination power was mostly attributed to the relatively small number of genetic markers and the complexity of the disease. Similar to the current study, a regression model was created in previous research to evaluate the risk of bracing failure in AIS patients [25]. With two genetic markers and one clinical parameter included in the model, the sensitivity and specificity were 92.3% and 41.7%, respectively. To further improve the prediction model of AIS, more genetic factors need to be studied and included in the model.

In this study, we described a formula facilitating the calculation of individual risk score of AIS, which was successfully replicated in an independent cohort. To be noted, however, the risk prediction resulted from this model is currently applicable only to the Chinese population. Also, this formula needs to be modified continuously with new predictors included in the model. Therefore, it should be carefully interpreted when extended to other populations in the current form.

The present study has several limitations. First, further validation of our model in different population is required to validate its reliability. Second, as only 10 SNPs were eligible to be included, identification of more AIS-associated loci in the Chinese population is required in order to strengthen the accuracy of the predictive model. Third, some confounding factors were not taken into consideration. We predicted the risk of AIS onset using genetic variants only. Other factors such as family history of idiopathic scoliosis, abnormal bone mineral density, abnormal body mass index, rapid growth velocity and unknown environmental factors can be associated with the development of AIS. In future studies, more factors need to be included in the predictive model to improve its sensitivity and specificity.

CONCLUSION

The present study established a novel predictive model concerning the risk of AIS onset in the Chinese population. Larger sample size, more influential factors and validation of the model in other populations will allow for the development of a more powerful prediction tool for the risk of AIS.

Table 1

Association of the 10 variants with AIS in Chinese.

SNPs	MA	Genotypea		P	MAF		P	Odds Ratio (95% CI) b
		Patients	Controls		Patients	Controls
rs241215	T	62/351/498	133/584/613	0.0001	0.261	0.320	2.2 x 10-5	0.75 (0.66 - 0.86)
rs7593846	G	112/432/353	107/636/589	0.0009	0.366	0.319	0.001	1.23 (1.09 - 1.39)
rs13398147	T	66/353/483	53/428/849	7.7 x 10-7	0.269	0.201	1.0 x 10-7	1.46 (1.27 - 1.49)
rs7633294	G	127/429/348	141/560/603	0.0007	0.378	0.323	0.0001	1.27 (1.12 - 1.44)
rs9810566	A	176/411/321	176/592/551	0.0001	0.420	0.358	2.6 x 10-5	1.30 (1.15 - 1.47)
rs2050157	T	139/442/325	173/576/581	0.001	0.397	0.347	0.0006	1.24 (1.10 - 1.41)
rs16934784	G	60/375/476	70/436/824	2.8 x 10-5	0.272	0.217	2.1 x 10-5	1.35 (1.18 - 1.55)
rs678741	T	255/452/201	248/684/400	3.4 x 10-8	0.529	0.443	1.1 x 10-8	1.42 (1.26 - 1.60)
rs4940576	A	209/447/255	206/636/487	6.3 x 10-7	0.475	0.394	8.6 x 10-8	1.39 (1.23 - 1.57)
rs2180439	G	188/429/293	197/639/497	0.0005	0.442	0.388	0.0002	1.25 (1.11 - 1.42)

aThe three values in the ‘genotype’ column indicate the numbers of homozygotes with respect to the minor allele, heterozygotes and homozygotes with respect to the major allele, respectively.

bCalculated with the minor allele as reference.

MA, Minor Allele; MAF, Minor Allele Frequency; CI, Confidential Interval.

Table 2

Results of logistic regression analysis.

SNPs a	Regression Coefficient	p	Odds Ratio (95% CI) b
rs241215	-0.239	0.0005	0.78 (0.69 - 0.90)
rs7593846	0.178	0.008	1.19 (1.05 - 1.36)
rs13398147	0.405	3.6 x 10-8	1.49 (1.30 - 1.73)
rs7633294	0.222	0.0006	1.25 (1.10 - 1.42)
rs9810566	0.311	7.1 x 10-7	1.37 (1.21 - 1.54)
rs2050157	0.199	0.001	1.22 (1.08 - 1.39)
rs16934784	0.310	2.3 x 10-5	1.36 (1.18 - 1.57)
rs678741	0.388	1.1 x 10-9	1.47 (1.30 - 1.67)
rs4940576	0.326	2.3 x 10-7	1.39 (1.22 - 1.57)
rs2180439	0.251	7.3 x 10-5	1.29 (1.14 - 1.45)

aGenotypes of homozygotes with respect to the minor allele, heterozygotes and homozygotes with respect to the major allele were coded as 2, 1 and 0, respectively.

bCI=Confidence Interval.

Table 3

The sensitivity, specificity and accuracy of the regression model.

Cut-off Point	Sensitivity	Specificity	Accuracy
0.1	1%	100%	38.9%
0.2	10%	97.3%	42.5%
0.3	37.7%	81.5%	54.7%
0.4	64.1%	57.8%	61.6%
0.5	85.9%	31.3%	64.7%
0.6	96.7%	15%	65.1%
0.7	99.3%	3.7%	62.2%
0.8	100%	0%	61.4%

Table 4

The distribution of individual risk score in the replication cohort.

Risk Score	Replication Population		p a
	Patients (n = 871)	Controls (n = 1239)
0 - 20	24 (2.8%)	158 (12.8%)	<0.001
20 - 40	333 (38.2%)	723 (58.3%)	-
40 - 60	373 (42.8%)	323 (26.1%)	-
60 - 80	136 (15.6%)	35 (2.8%)	-
80 - 100	5 (0.6%)	0 (0%)	-

a calculated by Chi-square test.