A comprehensive GWAS dataset, featuring 179 lipid species across 13 classes and 4 major categories—glycerolipids, glycerophospholipids, sphingolipids, and sterols—is depicted in Figure 1 ( Supplementary Data Sheet 1 ) (17). These lipid species’ summary statistics are cataloged in the GWAS Catalog (https://www.ebi.ac.uk/gwas/) with accession numbers ranging from GCST90277238 to GCST90277416 (18). The dataset originates from the GeneRISK study, which includes 7,174 European participants (4,579 female and 2,595 male), recruited between 2015 and 2017, and aged 45–66. The primary goal of GeneRISK is to evaluate the effects of genetic risk information on cardiovascular disease (19). The original publication details the inclusion and exclusion criteria for participants, along with the population’s characteristics (19). Before blood sample collection for plasma, serum, and DNA extraction, participants observed a 10-hour fasting period. Collected biological and demographic data, alongside health, genetic, and lipidomic profiles, are stored at the THL Biobank (https://thl.fi/en/research-and-development/thl-biobank). Additionally, the summary-level statistics for diabetic neuropathy were retrieved from the R9 release of the FinnGen GWAS results, under the phenotypic classification “DM_NEUROPATHY”, with small overlap (<5%) with GWAS of lipid species, reducing the risk of bias (https://r9.finngen.fi) (20). The determination of diabetic neuropathy cases was anchored in the standardized International Classification of Diseases (ICD) coding system. The dataset included a total of 2,843 cases and 271,817 control individuals of European descent, with adjustments made for variables such as age, sex, genetic relatedness, genotyping batch, and the first 10 principal components. On the other hand, we also obtained the GWAS summary statistics for body mass index (BMI) and glycated hemoglobin (HbA1c) from the European cohorts, sourced respectively from the Genetic Investigation of ANthropometric Traits (GIANT) Consortium (https://portals.broadinstitute.org/collaboration/giant/index.php/GIANT_consortium) and the Meta-Analyses of Glucose and Insulin-related Traits Consortium (MAGIC) (https://magicinvestigators.org/) (21, 22).

The MR analysis relied on three principal assumptions: (1) The IVs used in the study were significantly associated with lipidomes. (2) There was independence between the selected IVs and any confounding factors that could affect both lipidomes and diabetic neuropathy. (3) The IVs exerted an effect on diabetic neuropathy solely through their influence on lipidomes (23). To enable MR analysis across all lipid species, we adopted a significance threshold of P < 5e-6 for selecting genetic variants. This threshold is considered appropriate for MR studies when SNPs available for exposure are scarce (24, 25). To curtail the influence of correlated SNP associations, linkage disequilibrium (LD) scrutiny was performed using the European 1000 Genomes Project Phase 3 as a reference, with an r2 threshold of <0.001 and a clumping distance set at 10,000 kb. Palindromic SNPs were omitted to avoid inconsistencies in allelic interpretations that could skew causal inferences. SNPs with strong associations to the outcome were also discarded. The MR Steiger filter was applied to eliminate SNPs with an incorrect direction of effect. The IVs for lipid species were evaluated using the variance explained (R²) and the F-statistic, with those yielding an F-statistic below 10 being eliminated. The F-statistic is calculated using the formula R²(N-K-1)/[K(1-R²)], where R² is the variance of the exposure explained by the IVs, N is the effective sample size, and K is the number of variants in the IV model. The PhenoScanner, an online resource, was utilized to identify and exclude SNPs linked with potential confounders (26). Finally, to evaluate the statistical power, we utilized the online tool available at https://shiny.cnsgenomics.com/mRnd/ (27).

The MR analysis was conducted using the R software (version 4.3.1), with specialized packages including “TwoSampleMR” (version 0.5.7), “MR-PRESSO” (version 1.0), and “MendelianRandomization” (version 0.9.0). A two-sample MR analysis was employed using five MR methods: inverse variance weighted (IVW), weighted median, simple mode, weighted mode, and MR-Egger regression. The primary method used was the fixed-effects IVW, which combined the Wald ratios from individual SNPs to provide a summary estimate. Additional methods served to validate the findings. Multiple testing was accounted for using False Discovery Rate (FDR) correction, with a significant association defined by an FDR of less than 0.05. Sensitivity analyses were performed to confirm the robustness of the MR results. Cochrane’s Q test assessed heterogeneity among IVs, with a P value below 0.05 indicating significant heterogeneity and necessitating the use of a random-effects IVW model instead of a fixed-effects model. The MR-Egger regression intercept was employed to detect potential horizontal pleiotropy, with significance set at a P value below 0.05. The MR-PRESSO global test was also used to assess horizontal pleiotropy, with outlier correction to mitigate its influence. The stability of the results was further examined using a leave-one-out strategy, where each SNP was sequentially excluded from the analysis, and the IVW method recalculated the effect. Finally, to explore the potential vertical pleiotropic pathways, we also performed multivariable MR (MVMR) analysis, including MVMR-IVW, MVMR-Egger, and MVMR-Median, to assess the direct causal impacts of these lipid species on diabetic neuropathy after adjusting for BMI and HbA1c levels. The parameter settings were consistent with univariable MR analysis.

After thorough quality control measures, we have pinpointed 2,722 SNPs to serve as IVs in examining the causal relationship between lipid species and diabetic neuropathy. Each of these SNPs has an F-statistic value exceeding 10, which suggests that the potential for bias due to weak instruments is minimal. Detailed information for each SNP, encompassing the effect allele, other allele, β value, standard error (SE), and P value, is provided in Supplementary Data Sheet 2 .

A thorough MR analysis results of 179 lipid species and diabetic neuropathy was presented in Supplementary Data Sheet 3 . Through the IVW analysis, we have identified 17 lipid species with potential causal relationship. Among these, 5 lipid species are identified as potential factors that may contribute to an elevated risk of diabetic neuropathy, whereas 12 lipid species seem to confer some protection against diabetic neuropathy, as illustrated in Figure 2 . It is worth noting that phosphatidylcholine constitutes the majority.

Following the application of FDR correction, two lipid species have been identified that show a significant causal link with the risk of diabetic neuropathy, as illustrated in Figure 3 . Individuals with genetically higher levels of phosphatidylcholine (16:0_20:2) had a 18% lower risk of developing diabetic neuropathy (OR = 0.82, 95%CI: 0.73-0.91; P < 0.001, FDR = 0.033). Similarly, higher genetically inferred levels of phosphatidylcholine (16:1_18:1) were associated with a 23% reduced risk of diabetic neuropathy (OR = 0.77, 95%CI: 0.67-0.88; P < 0.001, FDR = 0.019). Additional analyses using methods such as MR-Egger, weighted median, simple mode, and weighted mode have also consistently indicated a trend in these causal associations. Supplementary Data Sheet 4 presents scatterplots that clearly depict the causal connections between these lipid species and the risk of diabetic neuropathy.

Sensitivity analyses presented in Table 1 , including the Cochrane’s Q test, showed no heterogeneity among the genetic variants associated with two lipid species in predicting diabetic neuropathy. The MR-Egger regression intercept, used to assess the risk of bias from unbalanced horizontal pleiotropy, suggested no significant impact on our findings concerning diabetic neuropathy. The robustness of our MR results was further confirmed by the MR-PRESSO test, reinforcing the reliability of our conclusions. Supplementary Data Sheet 4 demonstrates that removing any single SNP in the leave-one-out analysis did not significantly change the MR estimates, affirming the IVW method as the preferred analytical strategy, given the lack of significant heterogeneity or unbalanced pleiotropy in explaining the variability in the risk of diabetic neuropathy (28–30).

We further applied a MVMR analysis to assess the direct impacts of these two lipids on diabetic neuropathy ( Figure 4 ). According to the MVMR-IVW findings, the effect of genetically predicted phosphatidylcholine (16:1_18:1) (OR = 0.70, 95% CI: 0.51-0.98; P = 0.036) on diabetic neuropathy remained significant after adjusting for BMI and HbA1c. The consistent direction and magnitude of results from the MVMR-Egger and MVMR-Median models further support the causal inference ( Supplementary Data Sheet 3 ).

After refining our criteria based on a genome-wide significance threshold of P <5e-8, we selected IVs for these two lipid species ( Figure 5 ). Phosphatidylcholine (16:0_20:2) had six SNPs as IVs; phosphatidylcholine (16:1_18:1) had five SNPs as IVs. Further IVW analysis indicated that the increased levels of phosphatidylcholine (16:1_18:1) remain as the protective factors for diabetic neuropathy (OR = 0.80, 95% CI: 0.66-0.98; P = 0.032).