2.1 Study Overview

This study estimated heritability and genetic correlations of the phenotype by analyzing GWAS data. Subsequently, it identified disease-associated genetic instruments and assessed the bidirectional causal relationship between RLS and diabetes, including its complications, using a two-sample MR approach. The analysis employed the latest GWAS data for RLS, as well as for Type 1 and Type 2 diabetes (T2D), incorporating additional GWAS data for diabetes-related complications, such as diabetic nephropathy, retinopathy, and neuropathy. The discovery phase utilized data, primarily from the UK Biobank cohort, whereas validation was performed with data from a FinnGen cohort. In the forward MR analysis, the causal impact of RLS on diabetes and its complications was evaluated across five phenotypes. In contrast, the reverse MR analysis used diabetes and its complications as exposure variables to assess their potential genetic causal effects on RLS. Heterogeneity and pleiotropy analyses were conducted for quality control to ensure the robustness and reliability of the findings. The methodological workflow, encompassing GWAS summary statistics, MR assumptions, linkage disequilibrium score regression (LDSC), and bidirectional MR analysis, is illustrated in Figure 1, whereas a detailed, step-by-step flowchart of the entire analytical strategy is provided in Figure S1.

2.2 Data Source

RLS GWAS data were obtained from the international EU-RLS-GENE consortium, encompassing a discovery meta-analysis of 7248 cases (2479 males and 4769 females) and 19,802 controls (10,422 males and 9380 females). RLS cases were recruited from multiple European countries, along with Canada (Quebec) and the United States. Diagnoses were conducted by expert neurologists or sleep specialists through face-to-face interviews, based on the diagnostic criteria of the International Restless Legs Syndrome Study Group (IRLSSG). This dataset is accessible from the HGRI-EBI Catalog GWAS under ID GCST90399568 (https://www.ebi.ac.uk/gwas/search?query=GCST90399568), with additional details available in Table 1. For comprehensive information on trait definitions and covariate adjustments, please refer to the original studies (Schormair et al. 2024).

For the discovery study, Type 1 diabetes (T1D) dataset (ID: GCST90014023) comprises 18,942 patients and 501,638 controls, whereas T2D dataset (ID: GCST90018926) includes 38,841 patients and 451,248 controls. The diabetic nephropathy (D-NEP) dataset (ID: GCST90018832) contains 1032 patients and 451,248 controls, whereas the diabetic retinopathy (D-RET) dataset (ID: GCST90043640) includes 308 patients and 456,040 controls. The diabetic neuropathy (D-NEU) dataset (ID: GCST90129436) consists of 772 patients and 435,199 controls, all of European ancestry. All datasets are available in the HGRI-EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/home). Validation study GWAS data were obtained from the FinnGen database; further details can be found in Table 1.

2.3 Instrument Selection

MR relies on assumptions of correlation, independence, and exclusion. The correlation assumption requires that genetic variants, specifically single nucleotide polymorphisms (SNPs) linked to a particular phenotype, be selected as IVs under stringent criteria. For both forward and reverse MR analyses, IVs for RLS, diabetes mellitus, and diabetic complications were chosen using a significance threshold of p < 5E − 8. This widely accepted threshold represents a Bonferroni correction for the approximately one million independent common variants in genomes of European ancestry, effectively controlling the family-wise error rate at 5% (Uffelmann et al. 2021; Chen et al. 2021). The independence assumption was met by applying a low linkage disequilibrium (LD) threshold (10,000 kb, r² = 0.001). To ensure adequate instrument strength, an F value was calculated for each SNP to confirm it exceeded 10, following the approach of Pierce and Burgess (2013). Sensitivity analyses were conducted to verify that causal inference was unaffected by pleiotropy, thereby meeting the exclusion assumption.

2.4 Statistical Analysis

LDSC, implemented via the ldscR R package, was employed to investigate genetic associations and interactions among RLS GWAS, diabetic phenotypes, and SNPs. LDSC quantifies polygenic signals and mitigates bias from population stratification, enabling unbiased genetic correlation assessment even with overlapping samples (Bulik-Sullivan et al. 2015). Furthermore, LDSC was also used to estimate associations between GWAS phenotypes and SNPs. The method also assesses LD analysis to estimate heritability (h²), with a significance threshold of p < 0.05 indicating significant genetic correlations.

MR analysis was conducted utilizing the TwoSampleMR software package. Within the framework of bidirectional MR analysis, the causal relationship between the exposure and outcome was estimated using the inverse-variance weighted (IVW) or Wald ratio method. Additionally, we employed the IVW radial approach and the robust adjusted profile score (RAPS) method for causal effect estimation. The IVW method applied a p value threshold of <0.05 as the main threshold for evaluating causal effects, without applying Bonferroni adjustments to maintain the exploratory nature of the study. The robustness of MR findings was established if at least one of the alternative methods yielded a similar significant result.

2.5 Sensitivity Analyses

A comprehensive suite of sensitivity analyses was conducted to test for potential violations of MR assumptions, particularly heterogeneity and horizontal pleiotropy. Our multi-pronged strategy included the following: First, we used the MR-Egger intercept test to specifically detect directional pleiotropy, with an intercept p value < 0.05 indicating its presence. Second, we applied Cochran's Q-test to assess heterogeneity among the IVs, as significant heterogeneity can indicate underlying pleiotropy. Third, to further scrutinize and mitigate the influence of individual variants, we employed radial MR methods (implemented via the RadialMR R package) to identify and remove outlying SNPs that might have pleiotropic effects. Finally, a leave-one-out sensitivity analysis was performed to ensure that the overall causal estimate was not driven by any single, disproportionately influential SNP (Greco et al. 2015; Burgess and Thompson 2017).

2.6 Ethical Approval

Regarding the ethical approvals and informed consent for our study, I would like to clarify that the analysis conducted in this manuscript is based on publicly available GWAS data that had already received the necessary ethical approvals and consent at the time of the original data collection and publication. As no new data collection or direct interaction with participants was performed, additional Institutional Review Board (IRB) approval or informed consent is not required for this analysis.

3.1 Heritability Estimation

Heritability for RLS, T1D, T2D, D-NEP, D-RET, and D-NEU was assessed for both the discovery and validation studies using LDSC regression analysis. The LDSC results revealed significant heritability in 9 out of 11 phenotypes analyzed, including RLS, T1D, T2D, D-NEP, and D-NEU (GCST90129436), as well as D-RET (finn-b-DM_RETINOPATHY). Common genetic variants explained between 1.9% and 16.0% of phenotypic variance, with T2D (finn-b-E4_DM2) showing the highest heritability (h² = 0.160, p = 1.28E − 31), followed by RLS (h² = 0.145, p = 5.79E − 08). Detailed results are provided in Figure 2A.

3.2 Forward MR and Sensitive Analyses

The forward MR analysis investigated the causal relationship between RLS and diabetes mellitus (T1D and T2D), as well as diabetic complications, including D-NEP, D-RET, and D-NEU. The analysis utilized 11 IVs for RLS as the exposure, with F-statistic values ranging from 32.33 to 880.53, effectively mitigating concerns about weak instrument bias (Table S1).

In the discovery phase, the forward MR analysis indicated a genetic causal association between RLS and diabetic nephropathy (estimated ORs: IVW = 1.049, p = 0.0238; IVW radial = 1.049, p = 0.0238; RAPS = 1.049, p = 0.2851), as detailed in Figure 2D and Table 2. The IVW method, which assumes all IVs are valid, is a weighted regression model that regresses the effect sizes of each genetic instrument on exposure and outcome, providing results with high statistical power. IVW radial, on the basis of the original IVW model, refines the analysis of residuals and outliers to better detect and exclude invalid instruments, thus reducing the impact of these outliers. The RAPS method, which is more stringent by incorporating penalty terms to minimize the influence of invalid instruments, such as those with pleiotropy or bias, yielded consistent effect estimates, though no significant association was found (p = 0.2851). This indicates potential pleiotropy may have affected the results. Further sensitivity analysis using MR-Egger indicated intercepts close to zero (p = 0.6094), suggesting no horizontal pleiotropy bias (Figure 2B). The consistency in effect estimates across IVW, IVW radial, and RAPS analyses indicated robustness of the findings, despite no significant causal association with other diabetic complications.

In the validation phase using the FinnGen GWAS database, the results were consistent with those of the discovery study, indicating a strong genetic association between RLS and an increased risk of diabetic nephropathy (estimated ORs: IVW = 1.067, p = 0.0028; IVW radial = 1.067, p = 0.0028; RAPS = 1.068, p = 0.0140) (Figure 2C,D; Table 2). All three methods produced significant results, further confirming the robustness of the findings. No significant causal relationships were found for other forms of diabetes or diabetic complications. Thus, both the discovery and validation studies concluded that RLS is a risk factor for diabetic nephropathy. More detailed results are presented in Table S2.

In the validation phase using data from the FinnGen GWAS database, results were consistent with those of the discovery study, indicating a strong genetic association between RLS and an increased risk of diabetic nephropathy (estimated ORs: IVW = 1.067, p = 0.0028; IVW radial = 1.067, p = 0.0028; RAPS = 1.068, p = 0.0140) (Figure 2C,D; Table 2). Each method yielded significant results, further supporting the robustness of these findings. No significant genetic associations were observed for other diabetes types or diabetic complications. Consequently, both the discovery and validation phases suggest that RLS predisposes individuals to diabetic nephropathy. Detailed results are provided in Table S2.

In sensitivity analyses for exposures and outcomes demonstrating significant causal associations, Cochran's Q-test detected no heterogeneity in either the IVW or MR-Egger models for the specific exposure-outcome pairs, with p values ranging from 0.6300 to 0.9982 (Table 3, Table S3), as also illustrated in the funnel plots (Figure 3A,B). Furthermore, the MR-Egger intercept test indicated no evidence of directional pleiotropy, with p values between 0.6094 and 0.7937 (Table 3). Additionally, the leave-one-out analysis confirmed that the overall IVW estimates were not influenced by any single SNP, supporting the robustness of the findings (Figure 3C,D).

3.3 Reverse MR and Sensitive Analyses

A reverse MR analysis was performed to investigate whether diabetes (T1D and T2D) and its complications (D-NEP, D-RET, and D-NEU) are genetically linked to a higher likelihood of developing RLS. Appropriate IVs were selected for each exposure. In the discovery study, the qualifying SNPs for T1D, T2D, D-NEP, D-RET, and D-NEU were 89, 183, 1, 1, and 1, respectively, with the lowest F-statistic observed at 29.60 (F > 10). For the validation study, the eligible SNPs were 16, 61, 2, 13, and 4, respectively, with the lowest F-statistic observed at 29.73, as shown in Table S4.

The IVW, IVW radial, and RAPS reverse MR analyses indicated no significant causal relationship between diabetes, its complications, and RLS in both the discovery and validation phases, except for D-NEU (GCST90129436), for which no suitable IVs were available under stringent criteria (Table 4, Figure 4). More detailed results of the reverse MR analyses can be found in Tables S5 and S6. However, significant heterogeneity or pleiotropy (p < 0.05) was detected in T1D and T2D during heterogeneity and horizontal pleiotropy testing (Tables S7 and S8). To ensure robustness, outliers were identified and excluded using the radial IVW and MR-Egger methods, after which the analyses were repeated. The corrected analyses included T1D datasets (GCST90014023 and finn-b-E4_DM1) with adjusted SNP counts of 78 and 14, respectively, and T2D datasets (GCST90018926 and finn-b-E4_DM1) with adjusted SNP counts of 158 and 53, respectively. For specific outlier information, refer to Table S9. The revised MR analyses continued to show no significant causal relationship, and no significant heterogeneity or horizontal pleiotropy was detected (Table 5, Tables S10 and S11).