Adolescent idiopathic scoliosis (AIS) is a disorder with a three-dimensional spinal deformity. AIS is defined as scoliosis that develops during adolescence with a Cobb angle (CA) of 10 degrees or more on plain radiographs in the standing position (1, 2). AIS is a common disease affecting 1-5% of adolescents (3), progressing rapidly during periods of rapid growth (4), and is associated with physical (5), mental (6), and respiratory problems (7). To prevent these problems, identification of the cause of the onset of AIS is demanded. In Japan, AIS has been adopted as one of the screening factors for adolescents and is considered an important adolescent health problem.

AIS is also known as a complex disease involved in environmental and genetic factors. In twin’s studies, he concordance of AIS between twins in monozygotic pair is as high as 0.73 to 0.92 and its heritability is estimated as 87.5% (8–11), suggesting that genetic factors account for a large fraction of AIS onset. Research groups including us have identified 20 GWAS significant loci in AIS (12).

BMI is also known as a highly polygenic trait with 670 GWAS significant loci identified so far (13). A relationship between AIS and body mass index (BMI) has been epidemiologically (12) suggested. A clinical study has reported that AIS patients had a significantly lower BMI in comparison with those healthy adolescents (14, 15). We showed a negative genetic correlation between AIS and BMI, which indicates that AIS and BMI share genetic backgrounds or polygenicity with opposing directions of effects (12, 16). This is in line with the epidemiological association between the onset of AIS and low BMI. However, the causal relationship between AIS and BMI remains to be elucidated.

Using genetic data is becoming popular to resolve those phenotypic relationships (17). Mendelian randomization (MR) is an approach to summarize causal inference between phenotypes using the genome-wide association study (GWAS) results (18, 19). The genetically determined phenotypic profiles are robust for lifelong confounding factors, which can be interpreted as ideal randomization of the subject. Because of the development of MR analytic methods (20) and the achievement of large-scale GWASs of various human phenotypes with publicly available data, MR is one of the best methods to infer causality. Here we conducted MR analysis to investigate the causal inference between BMI and the onset of AIS ( Figure 1 ).

The ethical committees at all collaborating institutions and RIKEN approved the study. Informed consent was obtained from subjects and/or parents. The details about collecting samples and inclusion criteria have been shown in the previous studies (12, 21–24). In the Japanese AIS cohort, the age of the case (93% of female ratio) was 10 -18 years, and the age of the controls (46% of female ratio) was 18 -100 years. To evaluate the causal relationship in a trans-ethnic manner, we also used another GWAS for AIS summary statistic derived from European ancestry. The age of the case (86% of female ratio) was 12.2 -17 years and the age of the control (55% of female ratio) was 44 -88 years for control.

We used AIS GWAS summary statistics derived from two US cohorts of European ancestry: Missouri (MO1) described in a previous study (25) and Texas (TX). The cases in the TX cohort (n=1,358) were recruited at Texas Scottish Rite Hospital for Children as approved by the Institute Review Board of the University Texas Southwestern Medical Center as previously described (26) and genotyped on the Illumina HumanCoreExome BeadChip. For controls in the TX cohort, we utilized 12,507 non-AMD GRU (non-age related macular degeneration general research use) subjects of European ancestry, downloaded from dbGaP website (https://www.ncbi.nlm.nih.gov/gap) from the International Age-Related Macular Degeneration Genomics Consortium study (IAMDGC: phs001039.v1.p1.c1). Cases and controls genotypes were merged and imputed as described in a previous study (25). Then only bi-allelic common (MAF>0.01) SNPs with imputation quality Rsq >=0.3 were included for further analysis. Genetic association for the imputed allele dosages was performed in Mach2dat (27) using logistic regression with gender and 10 principal components as covariates. The major histocompatibility complex (MHC) region was excluded form analysis.

For AIS, we used published summary statistics of Japanese AIS (12). For summary statistics of European AIS, a new GWAS was performed, the details of which are described above.

For BMI, we used published summary statistics derived from Japanese populations based on the Biobank Japan project (range of 18 -100 years, 46% of female ratio) (16). For European ancestry, we used two published summary statistics derived from European populations; a meta-analysis of GWAS on BMI in UK biobank (range of 40 -69 years, 54% of female ratio) and genetic investigation of anthropometric traits (GIANT) (range of 40 -75 years, 55% of female ratio) (13), and a result of GWAS for BMI of European children (range of 2 -10 years, 51% of female ratio) (28). In addition, summary statistics from the population architecture using genomics and epidemiology (PAGE) consortium (range of 18 -79 years, 50% of female ratio), including African American, Asian, Hispanic, Native American, and Native Hawaiian populations quantities were also used (29). In the above three BMI GWASs, age and sex were added as covariates for the analysis.

The overlapped significant SNPs between BMI and AIS GWAS were used for the analysis ( Table 1 ). The Analyses were performed for each population. In the above three BMI GWASs, age and sex were added as covariates for the analysis.

We also evaluated the F-statistics metric to include the strength of genetic instruments for all instruments in different populations. Based on the previous studies, the F-statistic metric was calculated by using the formula: F = (R2/k)/([1 − R2]/[n − k – 1]). R2 represents the proportion of variance explained by the exposure SNP-instruments, k represents the number of instruments, and N represents the exposure GWAS sample size (30, 31).

MR is a method for evaluating the causal effect of a risk factor on an outcome from observational data using genetic variants (32, 33). We adopted two-sample MR which is one of the MR analysis approaches. Two sample MR in which SNP exposure and SNP outcome associations are estimated handling GWAS summary statistics derived from independent studies was applied to the analysis. In this study, we focused on the causality of BMI to AIS using this method implemented in the R software “TwoSampleMR” which is composed of the R language (19). The instruction of this software also provides source codes for analysis. The association between BMI and AIS was evaluated using the inverse-variance weighted (IVW) method, a typical and conventional method in MR, to obtain a weighted average of the effect estimates.

If SNPs act through a pleiotropic pathway, it would violate the MR assumption that the instrumental variable affects the outcome only via exposure and bias the causal estimate (that is, the effect of a genetic variant on the outcome acts entirely through the exposure of interest.). Because testing the validity of these assumptions is difficult in indeed, therefore, we conducted additional two MR methods in addition to IVW, namely Egger regression method (MR-Egger) and the weighted median estimator method (Weighted median), to check for the robustness of the estimates for IVW. MR-Egger is similar to IVW other than that the intercept is not limited to passing through the origin, and the nonlinear intercept indicates the possibility of directional pleiotropy (20, 34). MR-Egger is statistically less powerful but can adjust the pleiotropy. The Weighted median method works even when at least half of the subset of genetic variants of the instrumental variables are valid and provides a valid causal estimate (35). The number of SNPs used in each analysis is described in Table 1 . In Japanese, we also evaluated potential bias in the results of MR analyses by sensitivity analyses including leave-one-out analysis, intercept test for MR-Egger followed in the R software “TwoSampleMR” which also provides source codes for these analyses and global test followed in the R software “MR-PRESSO” (36).

We performed the reverse-direction MR analysis inferring the causality of AIS on BMI. For Japanese overlapped significant SNPs reported in Japanese GWAS for AIS were used as instrumental variables ( Table 2 ) (12). Unfortunately, the number of independent SNPs passing the genome-wide significant threshold (p<5×10^-8) was too small to be analyzed using US AIS summary statistics.

To estimate the genetic correlation of GWAS SNPs from different populations, the python package, popcorn (version 0.9.9; https://github.com/brielin/popcorn), was used (37). East Asian and European data of the 1000 Genomes Project (1KG) were used to compute cross-population scores considering each linkage disequilibrium (LD) structure (38). Consequently, a high genetic correlation of AIS between Japanese and Europeans (rho:0.87 (SE:0.11)) was observed.

We performed power calculations for MR regarding the previous study (39). Using the GWAS data of the Japanese cohort as a reference, we set input parameters including sample size (79,211), the proportion of cases in the study (0.067), type I error rate (0.05), and true odds ratio of the outcome variable per standard deviation of the exposure variable calculated from the estimate of MR analysis with Japanese BMI and AIS GWAS result.

Since MR analysis using summary statistics with different populations is not recommended, we assessed genetic backgrounds between BMI and AIS. The correlation of the odds ratios of the genome-wide SNPs, excluding the BMI and AIS-associated loci, were estimated for sets of SNPs stratified by p-values of GWAS thresholds in each trait. For the shared SNPs between the two traits, LD pruning of the SNPs (r2<0.3) was adopted at first. When stratifying the SNPs based on the p-values in each GWAS, the correlation of odds ratios of the SNPs between BMI and AIS was evaluated by Pearson’s correlation coefficient. This approach has been established to assess the direction of genetic risk between traits or populations as a previous study reported (40).

The primary analyses included 83 SNPs in the Japanese GWAS as instruments for BMI (16). We found significant causality of genetically decreased BMI on the risk of AIS (estimate=-0.56, P=1.110^-3 in IVW) ( Figure 2 ; Table 2 ). The significant causality was also supported in the weighted median method (estimate=-0.56, P=5.110^-3) and the MR-Egger method (estimate=-1.50, P=22.8×10^-3) ( Table 2 ). The direction of estimates was consistent across all results, indicating that the risk SNPs of increased BMI had suppressive effects on the onset of AIS. These results are consistent with that of an epidemiological study (14). We conducted sensitivity analyses, intercept tests of MR-Egger and MR-PRESSO to evaluate potential bias in the results of MR analysis. According to the MR-Egger intercept test and MR-PRESSO, we did not observe obvious horizontal pleiotropy ( Table 3 ).

Next, we addressed whether the causal inference could be generalizable regardless of the selection of instruments. BMI GWASs have been mainly conducted in European populations. In US GWAS, we observed the analysis was conducted without any problems (λ_GC=1.03) and three genome-wide significant signals were observed. Since US GWAS has a small sample size, we conducted a trans-ethnic genetic correlation of AIS between Japanese and Europeans using the Popcorn program (37) and observed a high genetic correlation of AIS (rho:0.87 (SE:0.11)). Therefore, to address the generalizability of this inference on AIS and to confirm this trend in different populations, we conducted MR analysis using summary statistics of GWAS for AIS from the US population (European ancestry) and analyzed in three methods (IVW, weighted median method, and MR-Egger) as in the Japanese AIS analysis. We repeated the analyses, using 537 of the 640 BMI-associated SNPs identified by the meta-analysis of the UK biobank and GIANT consortium, 31 of 43 BMI-associated SNPs identified by GWAS for BMI derived from European children, and 61 of the 63 BMI-associated SNPs identified by PAGE consortium ( Table 1 ).

Before the analysis, the power calculation was performed to evaluate the results more conservatively. This calculation is based on each variant, not summary statistics. Therefore, we selected two variants, rs11642015 and rs633715, showing the highest association in the Japanese analysis. The power for MR analysis was 0.63 for rs11642015 and 0.36 for rs633715. On the other hand, when we tried to obtain the same results with the US cohort as those of the Japan cohort, the power was 0.23 for rs11642015 and 0.17 for rs633715. There is no significant difference in the case ratios between the Japanese and US cohorts, indicating that the sample size has a pure effect. Additionally, the total sample size was calculated to be 80,918 for rs11642015 and 80,515 for rs633715 to obtain the same power as in Japan with the same cohort ratio. These results indicate that the results of this MR analysis are mainly due to the small sample size of the US cohort.

According to the result of power calculation for MR analysis, a small sample size of the US GWAS lowered statistical power. However, even though most of the results were not significant, but showed consistent trends with those using summary statistics of Japanese GWAS for AIS were observed ( Table 2 ). We also repeated the sensitivity analyses, intercept test and MR-PRESSO for each BMI instrument using both of AIS instruments ( Table 3 ). We did not find apparent bias by sensitivity analyses, and similar trends like MR analyses were observed between Japanese and US instruments. Finally, we applied the reverse MR analysis inferring the causality of AIS on BMI ( Table 1 ). We did not observe evidence of the causality of AIS on BMI in any MR methods ( Table 2 ). Since the number of independent SNPs passing genome-wide significant threshold (p<5×10^-8) were too small (3 SNPs) for reverse MR to be analyzed using US AIS summary statistic.

We further tried to provide evidence to support the causal relationship between low BMI on the development of AIS (and not vice versa). We conducted polygenic assessment for common variants shared between GWAS summary statistics for AIS and BMI to compare genetic backgrounds between BMI and AIS. We stratified the SNPs based on the p-values of GWAS (P) in each summary statistic and evaluated the correlation of odds ratios of the SNPs between the two traits. We observed significant negative correlations of odds ratios for the SNPs related to BMI, even for those showing modest association (P<0.0001 in the GWAS for BMI of Biobank Japan; r=-0.65~-0.18 for each P value range, p-value<0.005 for each correlation test) ( Figure 3 ), confirming agreement in direction of genetic risk in a polygenic manner (40, 41). Observed correlations (r) of odds ratios indicate substantial overlap of the genetic risk (with opposing directions of effects) between the two traits, not only in the identified BMI susceptibility loci exceeding the GWAS significant level but also at the loci showing nonsignificant associations. These findings are consistent with the negative relationship between AIS with BMI reported in the previous studies (12, 16). When repeating these analyses using other BMI summary statistics from other populations, similar trends of correlation of odds ratios were observed in the analyses using the BMI summary statistics of transethnic analysis (PAGE) and European child. On the other hand, we did not observe modest correlations of odds ratios for the SNPs in p-value based on GWAS for AIS ( Figure 3 ).