Nonlinear causal effects of estimated glomerular filtration rate on myocardial infarction risks: Mendelian randomization study

The study was performed in accordance with the Declaration of Helsinki and approved by the Institutional Review Boards of Seoul National University Hospital (No. E-2006-043-1131). The usage of the UK Biobank data was approved by the UK Biobank consortium (application No. 53799). Acquisition of informed consent was not required, as the study investigated anonymous public databases and genetic summary statistics.

The study was an MR analysis of the major findings from the CKDGen and UK Biobank data (Fig. 1). The CKDGen genome-wide association study (GWAS) meta-analysis provides the largest-scale information to date on single-nucleotide polymorphisms (SNPs) associated with kidney function traits (URL: https://​ckdgen.​imbi.​uni-freiburg.​de/​) [12, 13], thus, it was utilized to develop the genetic instruments for log-transformed eGFR. The UK Biobank is a population-scale prospective cohort that included > 500,000 participants aged 40–69 from diverse regions in the UK from 2006 to 2010 (URL: https://​www.​ukbiobank.​ac.​uk/​) [14]. The database includes a variety of clinicodemographic information and deep genotyping data and thus has been widely used for genetic studies, including MR analysis.

The main analysis was a two-sample MR analysis; genetic instruments were developed for log-transformed eGFR values calculated by the CKD-EPI equation using creatinine values from the phase 4 CKDGen GWAS meta-analysis [12], and MI outcomes in the UK Biobank with no sample overlap between the two data. Avoiding sample overlap in MR analysis has strength in a conservative sense, as a potential bias, particularly in the case of weak instruments, is toward false negative findings; thus, the robustness of a positive finding by two-sample MR analysis can be supported [15].

We also performed a secondary analysis using genetic instruments for log-transformed eGFR values based on serum cystatin C levels [13], as creatinine-based eGFR values are more likely to be biased by dietary factors or body shapes than cystatin C-based levels. Genetic instruments were developed from a recent GWAS meta-analysis that included both CKDGen studies and UK Biobank data. As the UK Biobank provided > 90% of the samples for the meta-analysis, the MR analysis is nearly a one-sample setting. Despite the sample overlap-related issues, the analysis has strength for replicative purposes, including an alternate kidney function parameter which is less affected from external factors.

Three core assumptions should be attained to demonstrate causal estimates by an MR analysis [16]. First, the relevance assumption is that the genetic instrument should be closely associated with the exposure phenotype, and as the instruments were associated with eGFR with genome-wide significance, this assumption was considered attained. We further tested the association strength by calculating the explained variance by R2 values. The other two assumptions, the independence and the exclusion-restriction assumptions, are regarded as an absence of a pleiotropic pathway. The independence assumption means that an instrument should not be associated with a confounder, and the exclusion-restriction assumption means that the causal effect should occur through the exposure of interest. In MR analysis under the linearity assumption, some pleiotropy-robust MR analyses are usually performed to address and test the pleiotropic effects (e.g., MR-Egger regression). Alternatively, in the current study, non-linear MR analysis directly adjusting the potential confounding covariates was performed to conduct a MR analysis even controlling some measured pleiotropic effects. In addition, we trimmed the genetic instruments to exclude the variants that are suspected to be weakly associated with kidney function traits because of potential confounding associations.

For the genetic instruments in the main two-sample MR analysis, we used the results of individuals of European ancestry (N=567,460) from the CKDGen data [12] to restrict our analysis to individuals of a single ancestry, as in our previous studies [17‐19]. The European participants had a median age of 54 years old, 91.4 mL/min/1.73 m² median eGFR values, and 9% prevalence of CKD determined by eGFR < 60 mL/min/1.73 m², and 50% of them were male.

The CKDGen GWAS meta-analysis reported 256 index SNPs at least 1 Mbp apart with a genome-wide significant association (P < 5 × 10⁻⁸) with log-transformed eGFR values based on creatinine levels. Additional trimming of the 256 SNPs was necessary, as in our previous MR analyses [17‐19], to remove genetic variants likely related to creatinine metabolism instead of kidney function. We performed an association analysis of eGFR based on cystatin C levels with data from 337,138 individuals of white British ancestry in the UK Biobank who passed genetic quality control with the exclusion of those outliers for heterozygosity, missing rate, or with sex chromosome aneuploidy. The variables for genetic quality control were predefined by the UK Biobank consortium, and those who were included in the principal component analysis and those who reported white British ancestry were included. The linear regression analysis of the 256 SNPs for the cystatin C eGFR level was adjusted for age, sex, age × sex, age², and the first 10 genetic principal components by PLINK 2.0 [20]. Other details of the genetic data structure and quality control process are available in the resources provided by the UK Biobank consortium [14]. We disregarded 115 SNPs identified from the association analysis in the UK Biobank data that showed different directions of regressed betas or those not reaching the Bonferroni adjusted significance level (P < 0.05/256) association with cystatin C-based eGFR values. Finally, the remaining 141 SNPs and their summary statistics in the CKDGen data were used as the genetic instrument for kidney function, and the combined allele score explained 2.69% of the variance in creatinine-based eGFR values in the UK Biobank data. When we calculated the F statistic, which should be over 10 to avoid weak instrument bias, by the equation [(n – k − 1)/(k)]*[R2/(1 − R2)], where n represents sample size, k represents the number of instruments, and R2 represents explained variance of the exposure phenotype, the F statistic was 66.1 [17]. The summary statistics for the instrumented SNPs are presented in Supplemental Table 1.

As serum creatinine levels are affected by muscle mass or diet, the cystatin C-based eGFR value has certain benefits when assessing kidney function [21]. Cystatin C-based eGFR value was a superior biomarker in regard to its power to predict cardiovascular diseases in the UK Biobank data [22].

A recent GWAS meta-analysis incorporating CKDGen studies and UK Biobank data reported 424 lead SNPs with genome-wide significant association with log-transformed eGFR based on creatinine levels. The study included a GWAS meta-analysis for log-transformed eGFR values based on cystatin C levels of samples from individuals of European ancestry (N=460,826) [13]. To include the SNPs consistently associated with kidney function-related biomarker, a trimming was performed similar to the aforementioned method and 348 SNPs were validated to be significantly (P < 0.05) associated with eGFR based on creatinine and blood urea nitrogen levels with consistent direction. We used the information of these 348 SNPs and the statistics of their association with eGFR based on cystatin C levels as the genetic instruments for the secondary analysis (Supplemental Table 2). Combined allele scores of the 348 SNPs explained 3.47% of the variance in cystatin C-based eGFR values in the data from individuals of white British ancestry in the UK Biobank, yielding an F statistic of 34.8 [23].

We used the UK Biobank data as the source of MI outcome, as individual-level large-scale genetic data, including information on both eGFR values and MI events, are necessary for a nonlinear MR analysis [10, 17]. The UK Biobank data defined MI events based on self-reports, hospital admission records, and death registries throughout the UK. Among the 337,138 individuals of white British ancestry in the UK Biobank data used in this study, 321,024 individuals had available creatinine- and cystatin C-based eGFR values, including 12,205 people with MI.

Conventional MR analysis (e.g., inverse variance weighted method) assumes a linear exposure-outcome relationship. As average causal estimates are calculated for the total ranges of an exposure in such an analysis, the causal estimates can be falsely attenuated if the true exposure-outcome relationship is nonlinear [10]. In such conditions, nonlinear MR analysis can be used.

In nonlinear MR analysis, the stratification of the population is performed by instrument-free exposure, the residual variation in the exposure conditioned for the instruments [10]. This is because directly dividing the study population according to exposure phenotype would bias the results by invalidating the MR assumptions; thus, the nongenetic component of the exposure is used to stratify the population. Next, localized averaged causal estimates are calculated as the association between the outcome and genetically predicted exposure divided by the association between the exposure and genetically predicted exposure. Finally, meta-regression of the localized causal estimates can be performed in nonlinear MR analysis to estimate the exposure-outcome relationship.

First, we plotted restricted cubic spline curves from the instrument-free exposure on MI risks to present interpretable MR estimates. The cubic spline curves based on logistic regression analysis for MI outcome was plotted with 10 knots determined based on decile values.

For non-linear MR analysis, we mainly used the fractional polynomial model [10], one of the methods that have been commonly used in recent non-linear MR studies [24, 25]. The degree 1 or degree 2 model is commonly used for nonlinear MR analysis, and whether the degree 2 model fits better, particularly when the exposure-outcome association is complex, can be tested. In the current study, we used the flexible degree 2 model, as the model fit was better (P = 0.004) than that of the degree 1 model, for the main two-sample MR analysis with 100 strata [10]. Allele scores for genetically predicted eGFR were calculated with PLINK 2.0 by multiplying the gene dosage matrix with the regressed betas from the GWAS summary statistics, which provided the genetic instruments [20]. Whether the scores followed a normal distribution was assessed by histograms, as a normal distribution supports the random allocation of genotypes, which is necessary for MR analysis [26]. The main nonlinear MR analysis included adjustments for the covariates age, sex, and the first 10 genetic principal components. The risks of MI according to creatinine-based eGFR or cystatin C-based eGFR, calculated by the CKD-EPI equation [27, 28], were investigated by nonlinear MR analysis. To robustly control the effects from clinical covariables, we additionally adjusted for body mass index, systolic blood pressure, hypertension medication history, hemoglobin A1c, history of diabetes diagnosis, levels of triglycerides, high-density lipoprotein, low-density lipoprotein, dyslipidemia medication history, and urine microalbumin levels in a sensitivity analysis (Supplemental Methods). The sensitivity analysis was performed on 245,398 individuals (9128 MI patients) with complete information on the covariates.

We additionally presented the results by piecewise linear method from the same models constructed in the above analysis by the fractional polynomial method.

The nonlinear MR analysis was performed by the “nlmr” package in R [10], and a two-sided P value < 0.05 was considered a significant finding. The reference point of the phenotypical eGFR value for the analysis was designated as 90 mL/min/1.73 m², which was suggested by the clinical guideline and was reported to be associated with minimal cardiovascular risks in previous observational studies [7].

We performed supplemental summary-level MR analysis by the inverse variance weighted method, weighted median method [29], and MR-Egger regression [30] to inspect causal estimates under the linearity assumption [31]. The analysis was first performed against the outcome data from individuals of white British ancestry in the UK Biobank, and the summary statistics for MI risk were generated by a GWAS adjusted for age, sex, age × sex, age², and the first 10 genetic principal components by PLINK 2.0 [20]. A replicative analysis was performed on the summary statistics provided by the CARDIoGRAMplusC4D consortium, which was from a GWAS meta-analysis including 43,676 MI cases and 128,199 controls of predominantly European ancestry samples who were not included in the UK Biobank data [32]. The other details for the summary-level MR analysis are presented in the Supplemental Methods.

At the baseline visits, the median age of the 321,024 individuals of white British ancestry was 58 years, and 46% of them were male (Table 1). The median creatinine-based eGFR and cystatin C-based eGFR values were 92.50 (2.3% with < 60) and 88.89 (4.7% with < 60) mL/min/1.73 m², respectively (Supplemental Fig. 1). Four percent (13,205 cases) had prevalent/incident MI events, and the proportion was higher in males (7%) than in females (2%).

The distributions of the allele scores for eGFR values followed a normal distribution (Supplemental Fig. 1).

We calculated localized averaged causal estimates by stratifying the population according to instrument-free exposure variables (Supplemental Table 3). The instrument-free variable showed U-shaped association with the risk of MI when we plotted cubic splines (Fig. 2). When genetically predicted creatinine-based eGFR was the exposure variable (Fig. 3 and Table 2), nonlinear MR analysis by fractional polynomial method demonstrated a quadratic, or a U-shaped, association (quadratic P value < 0.001) with MI risk, and the β1 (decreasing slope in low eGFR ranges) and β2 (increasing slope in high eGFR ranges) estimates were both significant. The results were similar even after clinical covariates were adjusted, and the slope was steeper in the low eGFR ranges where a higher genetically predicted eGFR was associated with a lower risk of MI.

When the allele score for cystatin C-based eGFR was the exposure variable, a similar quadratic relation between genetically predicted eGFR and MI risk was identified, with both directions of causal estimates again being statistically significant. The results were similar when additional clinical covariates were adjusted for the model.

The results by the piecewise linear method also demonstrated a U-shaped association for the causal estimates by eGFR on risks of MI (Fig. 4).

When the conventional inverse variance weighted method under the linearity assumption was used to yield causal estimates by summary-level MR, the causal estimates remained null for both the creatinine- and cystatin C-based eGFR exposures on MI risk in the UK Biobank data (Table 3). Although no significant directional pleiotropy was suspected by MR-Egger intercept P values, the pleiotropy-robust summary-level MR sensitivity analyses also provided null causal estimates. The results were similar when the independent summary statistics from the CARDIoGRAMplusC4D consortium were used as the outcome data.