Investigating the effects of statins on ischemic heart disease allowing for effects on body mass index: a Mendelian randomization study

Despite effective lipid reduction and corresponding benefits for cardiovascular disease prevention and treatment, statins have pleiotropic effects potentially increasing the risk of ischemic heart disease (IHD), particularly by increasing body mass index (BMI). We assessed whether the effects of genetically mimicked statins on IHD were strengthened by adjusting for BMI in men and women. We also assessed if increasing BMI was specific to statins in comparison to other major lipid-lowering treatments in current use, i.e., proprotein convertase subtilisin/kexin type 9 (PCSK9) inhibitors and ezetimibe. Using univariable and multivariable Mendelian randomization (MR) we found genetically mimicked effects of statins increased BMI (0.33, 95% confidence interval (CI) 0.28 to 0.38), but genetically mimicked PCSK9 inhibitors and ezetimibe did not. Genetically mimicked effects of statins on IHD reduction in both sexes (odds ratio (OR) 0.55 per unit decrease in effect size of low-density lipoprotein cholesterol (LDL-c), 95% confidence interval (CI) 0.40 to 0.76), was largely similar after adjusting for BMI, in both men (OR 0.48, 95% CI 0.38 to 0.61) and women (OR 0.66, 95% CI 0.53 to 0.82). Compared with variations in PCSK9 and NPC1L1, only variation in HMGCR was associated with higher BMI. The effects on IHD of mimicking statins were similar after adjusting for BMI in both men and women. The BMI increase due to statins does not seem to be a concern as regards the protective effects of statins on IHD, however other factors driving BMI and the protective effects of statins could be.


Results
We used a single nucleotide polymorphisms (SNP) rs12916 to mimic statins. In sensitivity analysis, the six correlated SNPs from HMGCR (Supplementary Table S1-2, r 2 > 0.13) were used taking account of their correlations. Of the 7 SNPs mimicking effects of PCSK9 inhibitors from PCSK9 (Supplementary Table S1 and Supplementary  Table S3), the three independent (r 2 < 0.05) SNPs (rs11206510, rs2149041 and rs7552841) were used in the main analysis ( Fig. 1). Of the 5 correlated SNPs mimicking effects of ezetimibe from NPC1L1 (Supplementary Table S1  and Supplementary Table S4, r 2 > 0.05), rs10260606 (proxy of rs2073547, r 2 = 0.99) was used in the main analysis given it had the strongest association with LDL-c (Fig. 1). These associations with LDL-c for the SNPs mimicking statins, PCSK9 inhibitors and ezetimibe are given overall (Supplementary Table S5), in men (Supplementary  Table S6) and in women (Supplementary Table S7). Instrument strength. The F statistic for most SNPs used to genetically mimic the effects of statins, PCSK9 inhibitors and ezetimibe were > 10 for men, women and overall (Supplementary Table S5-7). The F statistics for the 264 SNPs (men), 314 SNPs (women) and 530 SNPs (overall) predicting BMI were all greater than 10, with median 49.8, 42.2 and 41.8 respectively.
Associations of genetically mimicked lipid modifiers with BMI. Genetically mimicked effects of statins increased BMI in men and women whereas genetically mimicked effects of PCSK9 inhibitors and of ezetimibe were not associated with BMI (Fig. 2). Further analysis for PCSK9 inhibitors and ezetimibe was not conducted, given their lack of association with BMI.
Associations of genetically mimicked effects of statins with BMI and IHD. Genetically mimicked effects of statins reduced the risk of IHD overall (odds ratio (OR) 0.55, 95% CI 0.40 to 0.76), and in men and   (Table 1). However, the I GX 2 was low (less than 90%; 89.8% overall, 85.0%, 84.9% for men and women, respectively), suggesting the violation of the NOME assumption for the MR-Egger estimates which were biased towards the null 12 , and that the MR-Egger estimates may not be reliable for one-sample MR 13 . The weighted median, consistent between one-sample and two-sample MR 13 , showed positive associations of BMI with IHD in both sexes (Table 1). In multivariable MR the conditional F-statistics for effects of genetically mimicked statins were 10.8 overall, 5.7 for men and 11.9 for women and for BMI were 61.8 overall, 31.6 for men and 56.6 for women. The multivariable MR-Egger intercept was significant overall and for men (P = 0.002, 2.33*10 -8 , respectively) and the corresponding modified Q statistic was also significant (877.5 and 435.3, respectively; P values < 0.001), so we focused on the MR-Egger estimates overall (OR 0.60, 95% CI 0.52 to 0.69) and for men (OR 0.48, 95% CI 0.38 to 0.61). The Q statistic (399.6, P < 0.001) also suggested pleiotropic effects for women, althought the MR-Egger intercept was not significant, where the MR-Egger estimate was also used (OR 0.66, 95% CI 0.53 to 0.82).
Compared with the univariable estimates for genetically mimicked effects of statins on IHD, the multivariable estimates (OR) adjusted for BMI were slightly smaller for men using MR-Egger, and were slightly larger for women and overall using MR-Egger (Table 2      www.nature.com/scientificreports/ Replication. We replicated the primary results using a different GWAS, i.e., GLGC for LDL-c, UK Biobank and GIANT 14 seperately for BMI and CARDIoGRAMplusC4D for IHD which had fewer SNPs for BMI. Similarly, genetically mimicked effects of statins increased BMI, whereas these effects were not evident for PCSK9 inhibitors or ezetimibe (Table 3). Genetically mimicked effects of statins were negatively associated with IHD, while BMI were positively associated with IHD (Table 4). Compared with the univariable estimates, the multivariable estimates (OR) for effects of statins on IHD adjusting for BMI were largely similar (Table 5). In the main analysis, the lead SNP, rs12916 was used for mimicking statins. The MR-Egger method was used for presenting the univariable MR estimates, given that the multivariable MR-Egger estimates were suggested for overall, men and women.

Discussion
Consistent with previous findings our study provides genetic evidence that statins raise BMI 3,4 , while PCSK9 inhibitors or ezetimibe do not affect BMI 3,15 . BMI increases the risk of IHD in both sexes 16,17 . Our study adds by showing that the BMI increase caused by statins may not detract from the protective effects of mimicking statins on IHD in both men and women. Although statins are known to increase BMI, the underlying mechanism remains unknown. Previous studies have shown dietary changes, particularly higher caloric intake for statin users 18 . Compared with statins, the other two lipid modifiers, PCSK9 inhibitors and ezetimibe however, did not affect BMI, which implies that higher BMI is not a consequence of lipid lowering, particularly of LDL-c lowering. Differences in effects on BMI for the three lipid modifiers may be related to their mechanisms of action. PCSK9 inhibitors enhance the clearance of LDL-c by increasing LDL receptors 19 , while ezetimibe reduces the cholesterol absorption 20 . In contrast, statins inhibit cholesterol synthesis 21 , an upstream bioprocess involving de novo synthesis in Leydig cells, which affects steroidogenesis 22,23 . Genetic scores for HMGCR , PCSK9 and NPC1L1 have been similarly associated with a slightly increased risk of diabetes 24 , and LDL-c lowering genetic variants located around genes targeted by lipid-lowering therapy have been positively associated with risk of type 2 diabetes 25 . Taken together these studies indicate detrimental pleiotropic effects of these variants on diabetes as a possible consequence of lipid lowering, which is different from our findings, suggesting statins specifically increase BMI, possibly indicating a different underlying pathway.
Although the estimates for BMI on IHD were not entirely consistent, it is likely that BMI increases the risk of IHD, with possibly more convincing estimates for men than women, which may also be partially explained by the hormone changes resulting from higher BMI. BMI raises plasma insulin in both sexes 26 , resulting in testosterone stimulation in vivo 27 , possibly via gonadotropin releasing hormone 28 . Testosterone stimulation is more relevant in men than in women given that men have higher testosterone than women 29 . Meanwhile, testosterone increases the risk of IHD more in men than women 9 . An MR study has also shown insulin raises the risk of CVD in men but not in women 30 . Conversely, BMI reduces testosterone in men 31 . As such, the overall effects of statins mediated by BMI in men may be relatively neutral due to these pleiotropic effects.
The effects of genetically mimicking statins on IHD were largely similar after adjusting for BMI in both men and women, suggesting that BMI may not necessarily mediate or undermine the effects of statins on IHD, despite the positive association of BMI with IHD. BMI is not a well-defined intervention reflecting measurable factors, such as diet and exercise, where different ways of changing BMI may have different effects 32 . Thus, a BMI increase due to statin therapy, but may not necessarily lead to IHD. However, although the association of statins with IHD does not seem to be mediated by BMI, the increase of BMI per 1 mmol/L LDL-c corresponds to an increased risk of IHD of OR 1.13 95% CI 1.07 to 1.19. Statins also increase the risk of type 2 diabetes, however, we did not adjust for type 2 diabetes because type 2 diabetes may be a downstream effect of BMI, whereby higher BMI increases the risk of type 2 diabetes 33 , likely resulting in higher risk of IHD 6 . Similarly, we did not adjust for HbA1c which may also partially mediate the effects of BMI on IHD 34 .
Despite providing sex-specific estimates and showing sex differences for the direct effects of statins on IHD allowing for BMI, this study has some limitations. First, instruments should satisfy three key assumptions, i.e., the instrument is strongly associated with the exposure, not related to factors confounding the exposureoutcome relation, and influences the outcome only via effects on the exposure. The genetic variants used are well Table 4. Univariable MR for genetically mimicking statins and BMI on IHD (CARDIoGRAMplusC4D for IHD, GLGC for LDL-c, GIANT from MR-base for BMI). *Excluded 3 palindromic SNPs (rs1558902, rs17001654, rs9641123) and rs12016871 that was not available in the CARDIoGRAMplusC4D.   24 . However, despite the high conditional F-statistics for BMI in the multivariable MR, the conditional F-statistics for genetically mimicked statins were low, and the Q statistic was high for overall estimate and men in the multivariable MR, suggesting pleiotropic effects for these two analyses, which were addressed by using multivariable MR-Egger. Pleiotropic effects for multivariable MR were not evident for women in this study. We cannot exclude the possibility that the SNPs mimicking effects of statins influence IHD through other pathways given the potential pleiotropic effects of these SNPs 35 .
Our findings, mainly based on populations of European descent, may not be transportable to other populations. Nevertheless, causes are consistent across populations, although relevance of the causal mechanism may vary by population 36 . Third, we replicated the results based on other large GWAS consortia for both exposure and outcome. However, due to lack of availability of sex-specific summary statistics for IHD in relevant populations, this study lacks sex-specific replication. Fourth, we assumed a linear relation for the association of lipid modifiers and BMI with IHD 37 . Fifth, linear regression was used for a binary outcome in the GWAS of IHD from Neale lab, which may generate false positives particularly when case numbers are small or minor allele frequency (MAF) is low (http:// www. neale lab. is/ blog/ 2017/9/ 11/ detai ls-and-consi derat ions-of-the-uk-bioba nk-gwas). However, the number of IHD cases was large and all SNP MAFs were over 1%. Sixth, our exposure statin use is similar to a dichotomized binary exposure, which does not have a clear interpretation 38 . However, statin use was instrumented by a number of SNPs and our estimates may correspond to different levels of statin use. Seventh, the estimates here do not necessarily reflect the exact risk reduction from pharmacologic treatments 39 , but the lifelong effects of endogenous exposures per unit decrease in LDL-c estimated by Mendelian Randomization 9 , which are usually greater than the short-term effects of pharmacologic intervention from an RCT 40 .

Methods
Genetic predictors mimicking effects of lipid modifiers. We used well-established genetic variants mimicking effects of statins (rs12916, rs10066707, rs17238484, rs2006760, rs2303152 and rs5909), PCSK9 inhibitors (rs11206510, rs2149041, rs7552841, rs2479409, rs2479394, rs10888897 and rs562556) and ezetimibe (rs10260606 (proxy of rs2073547, r 2 = 0.99), rs217386, rs7791240, rs10234070 and rs2300414) 24 , from the corresponding genes encoding the target protein for each lipid modifier (HMGCR , PCSK9 and NPC1L1, respectively). As the few SNPs mimicking each lipid modifier are correlated, independent (r 2 < 0.05) genetic predictors of the lipid modifiers strongly associated with LDL-c (p < 5*10 -8 ) were used in the primary analysis, i.e., rs12916 for statins, rs11206510, rs2149041, rs7552841 for PCSK9 inhibitors and rs10260606 for ezetimbe. For sensitivity analysis all correlated SNPs given above for each lipid modifier were used 24 with a correlation matrix between SNPs for Europeans obtained from the 1000 Genomes catalogue using LDlink (https:// ldlink. nci. nih. gov). MR estimates are presented in effect sizes of LDL-c reduction to mimic the effects of statins, PCSK9 inhibitors and ezetimibe, extracted from LDL-c genetic summary statistics from the UK Biobank where a sex-specific GWAS is available (http:// www. neale lab. is/ uk-bioba nk). The summary statistics are based on white British people and were adjusted for age, age 2 , sex, age*sex, age 2 *sex and the first 20 principal components for both sexes, and were adjusted for the first 20 principal components, age and age 2 for sex-specific GWAS. For replication, the associations of the relevant SNPs with LDL-c were also taken from the Global Lipids Genetics Consortium (GLGC), a meta-analysis of 188 577 participants mainly of European descent adjusted for age and sex, and excluding people known to be on lipid lowering medication 41 (Fig. 2).
Genetic predictors of BMI. SNPs that were strongly (p < 5*10 -8 ) and independently (r 2 < 0.001) associated with BMI were also obtained from the largest sex-specific GWAS of the meta-analysis 11 of the UK Biobank GWAS and the GIANT consortium ( Fig. 1), which included 806,834 participants of European ancestry, including 374,756 men and 434,794 women adjusted for age, sex, recruitment centre, genotyping batches and 10 principal components. GIANT 14 with 339,224 participants of European ancestry was adjusted for age, age 2 , and any study specific covariates, such as principal components. In GIANT estimates for unrelated individuals were stratified by sex and case/control status, while for family based studies sex was included as a covariate.
Genetic associations with IHD. Sex-specific associations of selected SNPs with IHD were also extracted from a UK Biobank GWAS provided by Neale lab, which concerned up to 361,194 participants with 20,857 cases of IHD, including 167,020 men with 15,056 cases of IHD and 194,174 women with 5,801 cases of IHD. Sex-combined associations with IHD were from the UK Biobank were taken from SAIGE genetic summary statistics 42 . Replication was based on CARDIoGRAMplusC4D (Coronary ARtery DIsease Genome wide Replication and Meta-analysis (CARDIoGRAM) plus The Coronary Artery Disease (C4D) Genetics) consortium, a meta-analysis of GWAS of IHD mainly based on people of European descent including 60,801 cases and 123,504 controls adjusted for study specific covariates and over-dispersion 43 (Fig. 2). Statistical analysis. Instrument strength was assessed by the F statistic using an established approximation, i.e., the square of beta of the SNP-exposure association divided by the square of its standard error 12 .
Genetic associations for each SNP were aligned on the same effect allele for exposure and outcome. Inversevariance weighting (IVW) with multiplicative random effects 44 combining SNP-specific Wald estimates (SNPoutcome divided by SNP-exposure) was used to obtain overall and sex-specific estimates of the associations of genetically predicted effects of lipid modifiers (i.e., statins, PCSK9 inhibitors and ezetimibe) with BMI and IHD, and estimates of the associations of genetically predicted BMI with IHD.
Multivariable MR was used to assess overall and sex-specific associations of genetically predicted effects of lipid modifiers with IHD adjusted for BMI. We combined the genetic variants for lipid modifiers and BMI in the multivariable MR, removed any BMI variants correlated with statins SNPs (r 2 < 0.001) and extracted their www.nature.com/scientificreports/ associations with LDL-c, BMI and IHD using corresponding GWAS. We used the conditional F-statistic to give an estimate of instrument strength conditional on the other exposure in multivariable MR 45 . We also used the modified Q statistic to assess horizontal pleiotropic effects using the WSpiller/MVMR package 45 , and used the MR-Egger estimate when a significant Q statistic 46,47 suggested violation of exclusion restriction assumption.

Sensitivity analysis.
In sensitivity analysis, we used all SNPs mimicking effects of statins, PCSK9 inhibitors and ezetimibe using correlation matrices obtained from LDlink (https:// ldlink. nci. nih. gov/). We used MR-Egger as sensitivity analysis for the univariable analysis given its less stringent assumptions assuming the instrument strength is independent of the direct effect 48 , while IVW assumes balanced pleiotropy. However, we did not use MR-Egger for correlated instruments due to concerns about interpretability 47 . We used the IGX2, an adaptation of the I 2 statistic, to assess the violation of the no measurement error (NOME) assumption, where a low IGX2 (< 90%) suggests a violation of the assumption for MR-Egger 12 . We also used the weighted median as it provides valid estimates when more than 50% of the information comes from valid instruments 49 . All statistical analyses were conducted using R (version 4.0.3; R Foundation for Statistical Computing, Vienna, Austria). The univariable and multivariable MR estimates were obtained using the MendelianRandomization R package and the WSpiller/MVMR package to obtain the conditional F-statistic and the modified Q statistic.
Ethics approval and consent for publication. Ethics approval was not required as only publicly available data was used in this study.

Conclusion
Compared with variations in PCSK9 and NPC1L1, HMGCR was uniquely associated with higher BMI for both sexes. Similar positive associations of BMI with IHD were found for men and women. The effects of mimicking statins on IHD per unit decrease in LDL-c were similar after adjusting for the harmful effects of BMI in both men and women.

Data availablity
The datasets generated and analysed during the current study are available in the UK Biobank repository, (http:// www. neale lab. is/ uk-bioba nk).