Mendelian randomization analysis reveals causal effects of blood lipidome on gestational diabetes mellitus.

This study was conducted and reported according to the STROBE-MR guideline [ 19 ]. We first conducted a two-sample MR analysis to evaluate the causal relationships between 295 lipid metabolites and GDM, and then an extended Phe-MR analysis of 1,210 non-GDM diseases to predict a wide range of potential side effects associated with targeting identified lipid metabolites (Fig.  1 ). Fig. 1 Conceptual framework of Mendelian randomization study. The study consists of two parts. a Part 1, we assessed the causality for the association between 295 lipid metabolites and the risk of GDM. b Part 2, we investigated a broad spectrum of side effects associated with targeting identified lipid metabolites in 1210 non-GDM diseases. GWAS = genome-wide association study, MR = Mendelian randomization, N in a  = number of lipid metabolites, N in b  = number of phenotypes Conceptual framework of Mendelian randomization study. The study consists of two parts. a Part 1, we assessed the causality for the association between 295 lipid metabolites and the risk of GDM. b Part 2, we investigated a broad spectrum of side effects associated with targeting identified lipid metabolites in 1210 non-GDM diseases. GWAS = genome-wide association study, MR = Mendelian randomization, N in a  = number of lipid metabolites, N in b  = number of phenotypes Summary statistics of single nucleotide polymorphisms (SNPs) associated with the blood lipidome were taken from three large-scale GWASs of European individuals [ 18 – 21 ] (Table  1 ). (1) Surendran et al. analyzed 271 lipid metabolic traits in 14,296 participants (8,455 from the INTERVAL study, and 5,841 from the EPIC-Norfolk study) with up to 87,696,888 imputed autosomal SNPs [ 18 ]. (2) Cadby et al. conducted a GWAS using data from the Busselton Health Study (BHS) involving 596 lipid species (539 of them found to have statistically significant associations) and 4,492 individuals of predominantly European ancestry, with 39,117,105 SNPs available for analysis after imputation [ 19 ]. (3) Karjalainen et al. conducted a meta-analysis of 213 lipid and lipoprotein parameters across 33 cohorts, with up to 13,389,637 imputed autosomal SNPs included in the meta-analysis [ 20 ], and our study specifically utilized data from 92,664 non-Finnish European participants from 21 of these cohorts, with detailed cohort information provided in Table S1. Table 1 Characteristics of blood lipid metabolites GWAS studies References Cohort(s) Population Sample size Mean Age (y) Female (%) No. SNP Platform Metabolites* Sample type Surendran et al. (2022) [ 18 ] INTERVAL European ancestor 8,455 44.0 48.8 87,696,888 MS 271 (98) Fasting plasma EPIC-Norfolk 5,841 59.8 53.3 Cadby et al. (2022) [ 19 ] BHS Australian 4,492 50.8 56.0 39,117,105 MS 539 (135) Fasting serum Karjalainen et al. (2024) [ 20 ] Meta-GWAS Non-Finnish European 92,664 (21 cohorts) 46.3 56.3 13,389,637 NMR 213 (62) Mostly fasting serum/plasma Total 1023 (295) BHS the Busselton Health Study, EPIC-Norfolk the EPIC-Norfolk study, GWAS genome-wide association study, INTERVAL the INTERVAL study, MS mass spectrometry, NMR  nuclear magnetic resonance *The numbers in brackets are the number of lipid metabolites included in the following analyses Characteristics of blood lipid metabolites GWAS studies BHS the Busselton Health Study, EPIC-Norfolk the EPIC-Norfolk study, GWAS genome-wide association study, INTERVAL the INTERVAL study, MS mass spectrometry, NMR  nuclear magnetic resonance *The numbers in brackets are the number of lipid metabolites included in the following analyses The GWAS summary statistics for GDM were obtained from the Release 10 results provided by the FinnGen consortium. This GWAS included data from 14,718 GDM cases (identified from registry data using ICD-9 and 10 codes O24.4) and 215,592 parous female controls. All participants in this study were of European (Finnish) ancestry [ 22 ]. Three key assumptions of MR must be satisfied in order to ensure the validity of the instrument variables (IVs): (1) IVs are strongly associated with lipid metabolites; (2) IVs are independent of any observed or unobserved confounders of lipid metabolites-GDM associations; (3) IVs affect GDM only through their effect in lipid metabolites without any alternative pathways. To meet the above assumptions, IVs were selected based on rigorous criteria. Specifically, (1) genetic variants were screened with genome-wide significance ( P  < 5e−8); (2) for each lipid metabolites, genetic variants were clumped to retain independent SNPs using a linkage disequilibrium threshold of r 2  = 0.001 and a clumping window size of 500 kb; (3) The strength of the selected IVs was evaluated using the F -statistic, calculated according to Eq.  1 , where r 2 represents the proportion of variance explained by the IVs, N is the sample size, and k is the number of IVs. IVs of F -statistic < 10 were considered weak and were excluded in the subsequent analysis; (4) SNP associated with more than five lipid metabolites were excluded to avoid pleiotropy effect. 1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F-statistic=\frac{{r}^{2}\times (N-1-k)}{(1-{r}^{2})\times k}$$\end{document} We also eliminated metabolites with less than 0.5% variance explained by IVs to ensure adequate statistical power for legitimate causal inferences. Lipid metabolites retained were categorized based on the number of IVs: those with three or more IVs were designated as tier 1, for which sensitivity analyses and robust causal association estimates were available [ 23 ]. Lipid metabolites with less than three IVs were classified as tier 2. The main results were based on tier 1 lipid metabolites, while tier 2 lipid metabolites were presented as suggestive findings. A two-sample MR analysis was carried out to assess the causal effect of lipid metabolites on GDM, using the inverse variance weighted (IVW) method as the primary approach. The MR-Egger regression, weighted median, weighted mode methods and constrained maximum likelihood and model averaging (cML-MA) method were conducted to assess the robustness of the IVW estimate. The absence of horizontal pleiotropy is a prerequisite for the validity of IVW results [ 24 ]. Assuming the instrument strength is independent of direct effect (InSIDE), MR-Egger regression assesses pleiotropy via its intercept term [ 25 ]. The result of the MR-Egger regression is consistent with IVW if the intercept term is equal to zero, which suggests that horizontal pleiotropy does not exist [ 25 ]. The weighted mode estimate demonstrates higher power, less bias, and lower type I error rates compared to MR-Egger regression if the InSIDE hypothesis is violated [ 26 , 27 ]. By removing significant outliers, the MR-PRESSO analysis detects and attempts to reduce horizontal pleiotropy, but it also depends on InSIDE assumptions and requires that at least 50% of the genetic variants are valid IVs [ 28 ]. The weighted median approach can yield a reliable causal estimate when up to 50% of genetic variants were invalid [ 29 ]. To control correlated and uncorrelated pleiotropic effects, we employed cML-MA, which does not rely on the InSIDE assumption [ 30 ]. A series of sensitivity analyses were carried out in order to assess the robustness of causal effects following the main MR analysis. (1) Cochran’s Q test was conducted to measure heterogeneity across IVs [ 31 ]. A fixed-effects model was employed when no heterogeneity was detected among IVs, while a random-effects IVW model was utilized otherwise. (2) As mentioned above, MR-Egger regression estimates an intercept term of the IVW test, with any deviation from zero indicating potential directional pleiotropy [ 24 ]. (3) Additionally, a leave-one-out analysis re-evaluates the MR association by sequentially excluding each individual IV [ 32 ]. Significant associations were validated using summary data from UK Biobank (UKB) obtained from the IEU OpenGWAS platform ( https://gwas.mrcieu.ac.uk/ ). We further performed a reverse MR analysis (i.e., GDM as exposure, and the identified lipid metabolites as outcome) utilizing SNPs that were linked to GDM as IVs to investigate if GDM had any causal impact on the identified lipid metabolites. As summary data released by Karjalainen et al. did not contain all lipid metabolites–SNP associations and was not qualified as an outcome dataset when conducting MR analysis, we used summary data from UKB as mentioned above to verify the reverse causation between significant lipid metabolites and GDM. All IV screening criteria, MR analysis strategies, and sensitivity analysis methods used in the validation and reverse MR analysis were consistent with those applied in the primary MR analysis. We performed colocalization analysis using validation dataset to detect whether the identified lipid metabolites and GDM share common causal variant(s). For each lipid metabolite-GDM pair, we set the window size to ± 250 kb centered on each IV. The evidence for colocalization was evaluated by calculating the posterior probability (PP) for hypothesis 4 (PP.H4), which posits that the same causal variant(s) are responsible for the associations observed between the lipid metabolites and GDM. A threshold of PP.H4 > 0.8 was applied to indicate strong evidence of colocalization. Phe-MR analysis was conducted to evaluate the potential side effects associated with hypothetical interventions aimed at reducing GDM by targeting specific lipid metabolites. Summary GWAS data of 2,447 disease traits with 412,181 Finnish participants were acquired from FinnGen consortium with 21,311,942 variants in the FinnGen release 10 results ( https://www.finngen.fi/en/access_results ). Disease traits were defined in terms of “Phenotype Description” in the original FinnGen data, and were organized into ICD code format to facilitate correspondence to specific diseases in different studies. Disease traits with fewer than 500 cases were excluded due to concerns about statistical power. Additionally, we selected representative phenotypes to minimize inherent redundancy between PheCodes, thereby improving the interpretability of the results. Ultimately, 1,210 non-GDM disease traits were included in the Phe-MR analysis to further investigate the potential side effects of GDM-related metabolites (Fig.  1 ; Table S2). Genetic variants for GDM-related lipidomic biomarkers were derived from the same GWAS as in the main MR analysis. The final Phe-MR results were standardized based on the associations between lipid metabolites and GDM, reflecting a change in lipid metabolite levels equivalent to a 10% reduction in GDM risk. This standardization served three purposes: (1) to provide a clinical context (i.e., what side effects might arise if identified biomarkers are targeted therapeutically for GDM?); (2) to harmonize the directionality among identified lipid metabolites that could either increase or decrease GDM risk; and (3) to facilitate a direct comparison of the magnitude and directionality of the side effects. All analyses were conducted using R version 4.3.0. The strength of causal associations was evaluated using odds ratios (ORs) and corresponding 95% confidence intervals (CIs). An observed P -value < 1.92e−3 (Bonferroni-corrected significance threshold calculated as 0.05 divided by 26 [for 26 tier 1 lipid metabolites]) was considered as statistically significant for a potential causal association [ 33 ], and P -value < 0.05 was taken as suggestive significant threshold in this study. All P -values were two-sided. Tier 1 lipid metabolites that reach the significance threshold were considered to have potential causal relationships with GDM. Lipid metabolites from both tier 1 and tier 2 that met the suggestive threshold were considered as suggestive findings; however, these lipid metabolites lack robust estimation and therefore need to be treated with caution when reporting. For the Phe-MR analysis, a significance threshold of P  < 0.01 was employed, given the primary objective of screening for potential side effects of the target metabolites. This threshold was chosen to minimize false negatives, emphasizing the identification of potential associations for further investigation. Phenotypic variance of specific lipid metabolite explained by relevant IVs were calculated using the “gtx” package (downloaded from https://www.rdocumentation.org/packages/gtx/versions/0.0.8 ). MR analysis and clumping were carried out using ‘TwoSampleMR’ package [ 23 ] and ‘MendelianRandomization’ package [ 34 ]. cML-MA method was conducted using ‘MRcML’ package [ 30 ]. Colocalization was carried out using the “coloc” package ( https://github.com/chr1swallace/coloc ).

A total of 1023 lipid metabolites and 12,408 unique SNPs were identified from three GWASs. After filtering, 793 lipid metabolites were retained, and 409 SNPs qualified as IVs. After harmonizing effect alleles between the exposure and outcome datasets, 396 IVs were available for 790 lipid metabolites. To mitigate potential horizontal pleiotropy, SNPs associated with more than five lipid metabolites were excluded, resulting in 295 lipid metabolites and 239 SNPs included in the main MR analysis (Fig.  2 ). These lipid metabolites were categorized into eight groups: FA, GL, GP, SP, SL, as well as apolipoproteins (Apo), lipid subclasses (e.g., total ceramide), and lipoprotein subclasses (e.g., total lipids in very large HDL). Among these, 26 lipid metabolites belong to the tier 1 group, as they had three or more IVs, enabling sensitivity analysis and providing robust causal association estimates. The remaining lipid metabolites were categorized as tier 2 group. The main results are based on the tier 1 group lipid metabolites, while tier 2 lipid metabolites are presented as suggestive findings. Fig. 2 The flow chart of instrumental variables selection. GWAS genome-wide association study, IVW inverse variance weighted, MR Mendelian randomization; SNP = single nucleotide polymorphism The flow chart of instrumental variables selection. GWAS genome-wide association study, IVW inverse variance weighted, MR Mendelian randomization; SNP = single nucleotide polymorphism Detailed information of the SNP data used as IVs in this MR study is listed in Tables S3 and S4, and an overview of lipides metabolites is available in Table S5. The range of metabolite variation explained by the IVs was found to be 0.514% to 26.864% (Table S5). The F statistics for the genetic instruments associated with blood metabolites ranged from 27.123 to 5977.797 (Table S3), which indicating that they were sufficiently powered to detect MR associations. Associations between tier 1 and tier 2 lipid metabolites and the risk of GDM in the main MR analysis are illustrated in Fig.  3 and Table S6. A genetically determined high ratio of triglycerides to phosphoglycerides (TG_by_PG) was potentially causally associated with an increased risk of GDM (IVW estimate: OR = 2.147, 95% CI: 1.415–3.257, P  = 3.26e-4). Additionally, tier 1 lipid metabolite palmitoyl sphingomyelin (d18:1/16:0) (PSM(d18:1/16:0)) (IVW estimate: OR = 0.747, 95% CI: 0.583–0.956, P  = 0.021), and tier 2 lipid triglycerides in very small very low-density lipoprotein (XS_VLDL_TG) (IVW estimate: OR = 2.498, 95% CI: 1.197–5.215, P  = 0.015) exhibited suggestive causal associations with GDM (Fig.  3 , Table  2 ). Moreover, although the significance threshold was not reached, certain lipid subclasses, such as fatty acids (FAs) and glycerolipids (GLs), appeared to have stronger associations with GDM compared to other lipid subclasses (Fig.  3 ). Fig. 3 Circular Manhattan plot displaying the associations between lipid metabolites and the risk of GDM in the Mendelian randomization analysis. The dashed red line represents the Bonferroni-corrected significance threshold (P < 1.92e−3), while the inner line indicates the suggestive significance threshold (P < 0.05). (Suggestive) Significant lipid metabolites are labeled accordingly. Blue bars denote tier 1 lipid metabolites while grey bars denote tier 2 lipid metabolites. The 295 blood lipid metabolites are categorized and color-coded by lipid classes as shown in Fig.  1 and Table S5. Detailed results for the associations between blood metabolites and GDM by Mendelian randomization analysis are presented in Table S6. GDM gestational diabetes mellitus, TG_by_PG ratio of triglycerides to phosphoglycerides, PSM (d18:1/16:0) palmitoyl sphingomyelin (d18:1/16:0), XS_VLDL_TG triglycerides in very small VLDL Table 2 MR results of significant lipid metabolites and GDM Lipid metabolites nsnp IVW (Fixed-effect) Cochran’s Q test MR-Egger P -value OR (95% CI) Q Q_pval intercept P -value TG_by_PG 4 3.26e-4 2.147 (1.415–3.257) 0.085 0.994 0.011 0.818 PSM(d18:1/16:0)* 3 0.021 0.747 (0.583–0.956) 1.616 0.446 0.036 0.587 XS_VLDL_TG* 2 0.015 2.948 (1.197–5.215) 0.006 0.936 / / CI confidence interval, IVW inverse variance weighted, nsnp number of single nucleotide polymorphisms, OR odds ratio; PSM (d18:1/16:0) = palmitoyl sphingomyelin (d18:1/16:0); TG_by_PG = ratio of triglycerides to phosphoglycerides; XS_VLDL_TG = triglycerides in very small VLDL *: suggestive findings /: Unattainable Circular Manhattan plot displaying the associations between lipid metabolites and the risk of GDM in the Mendelian randomization analysis. The dashed red line represents the Bonferroni-corrected significance threshold (P < 1.92e−3), while the inner line indicates the suggestive significance threshold (P < 0.05). (Suggestive) Significant lipid metabolites are labeled accordingly. Blue bars denote tier 1 lipid metabolites while grey bars denote tier 2 lipid metabolites. The 295 blood lipid metabolites are categorized and color-coded by lipid classes as shown in Fig.  1 and Table S5. Detailed results for the associations between blood metabolites and GDM by Mendelian randomization analysis are presented in Table S6. GDM gestational diabetes mellitus, TG_by_PG ratio of triglycerides to phosphoglycerides, PSM (d18:1/16:0) palmitoyl sphingomyelin (d18:1/16:0), XS_VLDL_TG triglycerides in very small VLDL MR results of significant lipid metabolites and GDM CI confidence interval, IVW inverse variance weighted, nsnp number of single nucleotide polymorphisms, OR odds ratio; PSM (d18:1/16:0) = palmitoyl sphingomyelin (d18:1/16:0); TG_by_PG = ratio of triglycerides to phosphoglycerides; XS_VLDL_TG = triglycerides in very small VLDL *: suggestive findings /: Unattainable The results of Cochran’s IVW Q test indicated significant heterogeneity of IVs associated with five lipid metabolites (L_HDL_TG, N-palmitoyl-sphingosine (d18:1/16:0), 5alpha-androstan-3beta,17beta-diol disulfate, Total Cer, epiandrosterone sulfate), and random-effects IVW models were conducted to evaluate causal associations between these five lipid metabolites and GDM (Table S6, Table S7). For other lipid metabolites, no significant heterogeneity in IVs was observed (Table  2 , Table S7). Additionally, MR-Egger regression intercept analysis showed no significant directional horizontal pleiotropy in the associations between lipid metabolites and GDM (Table  2 , Table S7). Leave-one-out analysis indicated that all tests are driven by multiple IVs (Figure S1). The relationship between TG_by_PG and GDM was validated using summary data from 114,999 individuals from UKB after excluding participants of non-European descent (downloaded from https://gwas.mrcieu.ac.uk/datasets/met-d-TG_by_PG/ ). Similarly, the association betweenXS_VLDL_TG and GDM was validated using summary data from 115,078 individuals in the UKB (data accessed from https://gwas.mrcieu.ac.uk/datasets/met-d-XS_VLDL_TG/ ). However, we were unable to verify the association between PSM(d18:1/16:0) and GDM as no other GWAS summary dataset for PSM(d18:1/16:0) was available for MR analysis. Causal associations between TG_by_PG and GDM was validated using validation data (IVW estimate: OR = 1.125, 95% CI: 1.045–1.211, P  = 0.002), although causal association between XS_VLDL_TG was not verified by the validation dataset (IVW estimate: OR = 1.033, 95% CI: 0.974–1.095, P  = 0.279) (Table S8). Cochran’s IVW Q test showed no significant heterogeneity in TG_by_PG IVs ( Q _pval = 0.977). Though the MR-Egger regression intercept analysis indicated the presence of horizontal pleiotropy, further MR-PRESSO analysis and cML-MA analysis found no significant outliers, and the association remained significant when accounting for associated pleiotropy (cML-MA estimate: OR = 1.126, 95% CI: 1.046–1.212, P  = 0.002) (Table S8). In reverse MR analyses, no significant association was found between GDM and TG_by_PG ( P  = 0.770) or XS_VLDL_TG ( P  = 0.670) (Table S9), indicating no potential reverse causation. Cochran’s IVW Q test showed no significant heterogeneity in GDM IVs ( Q _pval > 0.05), and the MR-Egger regression intercept analysis did not find significant horizontal pleiotropy ( P -value > 0.05) (Table S9). Besides, strong colocalization evidence was observed between TG_by_PG and GDM (PP.H4 = 0.954), as shown in Table S10. In the Phe-MR analysis using the IVW method, a total of 45 associations reached a significance threshold of P  = 0.01. Sensitivity analyses with the methods of Cochran’s IVW Q test and MR-Egger revealed no significant heterogeneity or pleiotropy in these 45 associations (Table S11). In brief, lowering TG_by_PG showed detrimental effects on the risk of peripheral retinal degeneration and intrahepatic cholestasis of pregnancy (ICP), while being beneficial for 13 other diseases. The most common diseases related to TG_by_PG fell into the categories of endocrine / metabolic diseases and genitourinary system (Fig.  4 , Table  3 ). Interestingly, lowering XS_VLDL_TG also had detrimental effects on the risk of ICP, suggesting that reducing TG_by_PG and XS_VLDL_TG as intervention targets for GDM may increase the risk of ICP. Raising PSM(d18:1/16:0) demonstrated deleterious effects on enthesopathies of limb (excluding foot) and osteoporosis with pathological fracture, but was beneficial for perforation of the tympanic membrane (Fig. S2, Table S11). Fig. 4 Potential on-target side effects associated with TG_by_PG, TG and PG intervention revealed by phenome–wide Mendelian randomization analysis. Odds ratios with their 95% confidence intervals represent the effect estimates on the risk of various non-GDM diseases of per 10% reduction in GDM risk by targeting a TG_by_PG, b triglycerides and c phosphoglycerides, respectively. Associations that are above the horizontal line with a black dash indicate deleterious side effects, while those below represent beneficial side effects. The upper red dashed horizontal line (odds ratio = 1.10) represents the point at which decreased GDM risk is counterbalanced by an equal increase in disease risk. CI    confidence interval, OR  odds ratio, TG_by_PG ratio of triglycerides to phosphoglycerides Table 3 Descriptive summary of significant phenome-wide Mendelian randomization findings representing on-target side effects of biomarker intervention Exposures Outcome IVW Cochran’s Q test MR-Egger OR (95% CI) P -value Q Q_pval intercept P -value TG_by_PG Disorders of lipoprotein metabolism 0.980 (0.967–0.993) 0.003 5.464 0.792 0.001 0.791 Metabolic disorders 0.980 (0.968–0.993) 0.003 6.553 0.767 0.003 0.572 Hyperlipidaemia, other/unspecified 0.975 (0.958–0.993) 0.006 7.568 0.911 -0.004 0.553 Familial hypercholesterolemia, with ischemic heart disease 0.957 (0.930–0.984) 0.002 15.038 0.920 0.008 0.495 Migraine 0.991 (0.984–0.997) 0.006 20.735 0.707 0.002 0.421 Psoriasis 0.984 (0.972–0.996) 0.010 25.414 0.551 0.005 0.395 Endometriosis of ovary 0.977 (0.962–0.992) 0.003 18.821 0.926 0.006 0.365 Endometriosis of pelvic peritoneum 0.973 (0.955–0.990) 0.002 26.359 0.606 0.002 0.776 Pilonidal cyst 0.966 (0.945–0.988) 0.003 22.325 0.807 0.005 0.604 Peripheral retinal degeneration 1.061 (1.014–1.109) 0.010 23.896 0.777 − 0.001 0.961 Asthma and allergy 0.976 (0.969–0.993) 0.005 22.455 0.868 − 0.001 0.897 Haemorrhage in early pregnancy 0.982 (0.971–0.994) 0.002 21.860 0.911 0.004 0.433 Intrahepatic Cholestasis of Pregnancy 1.040 (1.015–1.065) 0.001 17.679 0.913 − 0.004 0.655 Vestibular neuronitis 0.969 (0.947–0.992) 0.009 22.956 0.851 0.007 0.472 Inflammatory diseases of female pelvic organs 0.989 (0.981–0.996) 0.004 33.116 0.462 0.002 0.528 Total triglycerides Intrahepatic Cholestasis of Pregnancy* 1.607 (1.048–2.462) 0.029 14.097 0.442 0.021 0.177 Carcinoma in situ of cervix uteri 0.519 (0.332–0.812) 0.004 13.652 0.552 − 0.020 0.213 Familial hypercholesterolemia 0.155 (0.048–0.492) 0.002 0.804 0.938 0.006 0.834 Familial hypercholesterolemia, with ischemic heart disease 0.388 (0.201–0.752) 0.005 4.035 0.854 0.026 0.271 Calcific aortic valvular stenosis, operated 0.602 (0.429–0.842) 0.003 3.664 0.961 0.007 0.552 Coronary artery bypass grafting 0.518 (0.325–0.822) 0.006 5.215 0.516 0.025 0.142 Non-rheumatic valve diseases 0.726 (0.576–0.912) 0.007 9.753 0.638 − 0.006 0.377 Nasal polyp 1.583 (1.219–2.052) 0.001 7.531 0.912 0.001 0.954 Asthma and allergy 0.638 (0.474–0.862) 0.003 11.543 0.827 0.004 0.734 Impacted teeth 0.677 (0.507–0.902) 0.008 7.562 0.961 0.010 0.331 TMD muscular pain linked with fibromyalgia 0.755 (0.612–0.932) 0.009 12.357 0.778 0.002 0.797 Endometriosis of pelvic peritoneum 0.585 (0.421–0.812) 0.001 11.732 0.699 0.007 0.586 Endometriosis ASRM stages 1,2 0.672 (0.512–0.882) 0.004 9.029 0.876 0.007 0.496 Total phosphoglycerides Intrahepatic Cholestasis of Pregnancy* 0.945 (0.528–1.691) 0.848 9.796 0.200 − 0.104 0.116 Cerebral cysts 2.170 (1.270–3.708) 0.005 3.168 0.957 0.023 0.630 Peritonsillar abscess 1.518 (1.137–2.027) 0.005 4.416 0.731 − 0.005 0.885 Endometriosis ASRM stages 1,2 1.597 (1.196–2.132) 0.002 5.070 0.750 − 0.004 0.862 Glomerular disorders 0.569 (0.390–0.830) 0.003 9.195 0.419 − 0.012 0.739 CI confidence interval, IVW inverse variance weighted, OR odds ratio, TG_by_PG ratio of triglycerides to phosphoglycerides * P  > 0.01 Potential on-target side effects associated with TG_by_PG, TG and PG intervention revealed by phenome–wide Mendelian randomization analysis. Odds ratios with their 95% confidence intervals represent the effect estimates on the risk of various non-GDM diseases of per 10% reduction in GDM risk by targeting a TG_by_PG, b triglycerides and c phosphoglycerides, respectively. Associations that are above the horizontal line with a black dash indicate deleterious side effects, while those below represent beneficial side effects. The upper red dashed horizontal line (odds ratio = 1.10) represents the point at which decreased GDM risk is counterbalanced by an equal increase in disease risk. CI    confidence interval, OR  odds ratio, TG_by_PG ratio of triglycerides to phosphoglycerides Descriptive summary of significant phenome-wide Mendelian randomization findings representing on-target side effects of biomarker intervention CI confidence interval, IVW inverse variance weighted, OR odds ratio, TG_by_PG ratio of triglycerides to phosphoglycerides * P  > 0.01 Since the indicator TG_by_PG is composed of two lipids–triglyceride (TG) and phosphoglycerides (PG)–adjusting TG_by_PG involves modifying the levels of these two lipids. Therefore, we conducted Phe-MR analysis on TG and PG separately. Lowering TG showed a detrimental effect on the risk of nasal polyps and beneficial effect on 11 other diseases. Raising PG had detrimental effects on the risk of cerebral cysts, peritonsillar abscess and endometriosis (ASRM stages 1,2), while demonstrating a beneficial effect on glomerular disorders. Lowering TG also showed detrimental effect on the risk of ICP, although this association did not reach the significance threshold of 0.01 (P = 0.029). No significant association was observed between PG levels and ICP ( P  = 0.848) (Fig.  4 , Table  3 ).

Among the 26 tier 1 lipid metabolites included, we identified TG_by_PG as a potential causal biomarker for GDM (IVW OR = 2.147, 95% CI: 1.415–3.257, P  = 3.26e-4), and this association was confirmed by validation and reverse MR results. Additionally, two other lipid metabolites, PSM (d18:1/16:0) (IVW OR = 0.747, 95% CI: 0.583–0.956, P  = 0.021) and XS_VLDL_TG (IVW OR = 2.948, 95% CI: 1.197–5.215, P  = 0.015), were identified as suggestive potential causal biomarkers for GDM. Furthermore, Phe-MR analysis was used to assess potential on-target side effects related to GDM prevention via reducing TG_by_PG and XS_VLDL_TG levels as well as raising PSM (d18:1/16:0) levels. Beyond GDM, lowering TG_by_PG had detrimental effects on two diseases and beneficial effects on 13 diseases. Our findings highlight TG_by_PG as a lipid metabolite of interest in the pathogenesis of GDM, while the associations for XS_VLDL_TG and PSM(d18:1/16:0) are suggestive and warrant further investigation to confirm their potential roles. Our study is the first to indicate a positive correlation between TG_by_PG level and GDM risk. However, previous studies have linked TG_by_PG to overweight/obese BMI [ 35 ] and immune response [ 36 ], and both overweight/obesity [ 37 ] and immune dysregulation [ 38 ] are associated with GDM, suggesting that TG_by_PG may mediate the development of GDM. Furthermore, Sequeiros et al. found that maternal TG_by_PG levels in the second trimester and in cord blood positively correlated with offspring ascending growth profiles (which is associated with higher risk of childhood adiposity) in the first two years post-delivery [ 39 ], highlighting TG_by_PG as a potential target for GDM intervention. The association of TG_by_PG with GDM may be independently linked to its constituent lipid indices, triglycerides (TGs) and phosphoglycerides (PGs). TGs is the primary energy storage in mammals [ 40 , 41 ], while PGs or GPs, which are glycerol-based phospholipids, constitute the main structural components of biological membranes in eukaryotic cells [ 42 ] and serve as sources of signaling molecules [ 43 ]. Observational studies have independently associated both TGs and PGs with GDM. Elevated TG levels have been linked to an increased risk of GDM [ 16 , 44 – 46 ], whereas reduced PG levels are correlated with higher risk of GDM across different ethnic groups [ 16 , 47 – 49 ]. The links between TGs or PGs with GDM can be explained by a higher degree of insulin resistance, which is not adequately compensated by increased β-cell secretion [ 44 ]. Evidence showed that PGs with ester-linkages and ether-linkages have protective roles against oxidative damage [ 50 , 51 ], and it is postulated that oxidative stress, secondary to TG-related dyslipidemia, may lead to pancreatic β-cells dysfunction and suppression of insulin gene expression [ 52 ]. Elevated TG levels or a deficiency in PGs with ester-linkages and ether-linkages could reduce antioxidant protection, making pancreatic β cells more susceptible to oxidative stress [ 50 , 51 ]. Given the low antioxidant capacity of β cells [ 53 ], increased oxidative damage may lead to β cell dysfunction, impairing insulin secretion and exacerbating physical insulin resistance during pregnancy, both key factors in the pathogenesis of GDM [ 51 ]. TG and PG are both important lipid classes influence GDM and are interconnected, as their synthesis pathways share common precursors such as phosphatidic acid (PA) and diacylglycerol (DAG) [ 54 ]. Therefore, the synthesis of PG might compete with TG synthesis for these shared intermediates. And targeting key enzymes involved in TG synthesis, such as phosphatidate phosphatase (PAP) [ 54 ], could be an effect strategy to reduce TG_by_PG level. Lifestyle modifications and several pharmacological agents, including fibrates, niacin, and long-chain omega-3 fatty acids, are known to reduce TG levels [ 55 ]. However, there are currently no clinical drugs or interventions specifically targeting PG, nor do these existing treatments address the shared synthesis pathway of TG and PG. Thus, TG appears to be a more practical intervention target at present. Nonetheless, Phe-MR results suggested that lowing TG_by_PG levels might increase the risk of ICP, and this side effect is mainly caused by reduced TG levels. Which indicated that TG_by_PG may be a potential target for preventing GDM, however, in pregnant women at high risk of ICP, a strategy that increases PG levels rather than reducing TG levels may be more appropriate. Our results indicated that higher levels of TGs in very small very-low-density lipoprotein (VLDL) and lower levels of PSM (d18:1/16:0) is suggestively associated with an increased risk of GDM. A study in a Western European population found an association between XS_VLDL_TG and 2-h OGTT results (OR = 2.319, 95% CI: 1.103–4.874, Wu-Hausman P -value = 0.042), [ 56 ], aligning with our findings. However, although some studies have reported the association between certain sphingomyelin (SM) species and GDM [ 16 , 57 – 59 ], PSM has rarely been reported [ 60 ], and our study is the first to identify the association between PSM (d18:1/16:0) and GDM. Although further research is necessary to confirm these suggestive findings, they are supported by relevant biological mechanisms. The liver constantly synthesizes TGs by utilizing free fatty acids and carbohydrates, secreting them into circulation in the core of VLDL [ 61 ]. However, we did not find associations between other TG-rich lipoproteins (TRLs) (e.g., TG in small high-density lipoprotein) and GDM, as reported by other observational lipidomics studies [ 16 , 46 ]. This may be explained by the relationship between the size of VLDL particles and their susceptibility to lipid oxidation, with smaller dense VLDL being more susceptible to oxidative modification [ 62 ]. The connection between PSM (d18:1/16:0) and GDM may be explained by sphingomyelin (SM) metabolism pathway, which is the most abundant SLs in mammalian tissues [ 63 ] and have been experimentally linked to pancreatic β-cell failure, insulin resistance, and T2DM [ 64 ]. To our knowledge, this is the first systematic MR study utilizing lipid metabolites as exposures to assess their causal effects on GDM risk, highlighting several strengths of our study. Firstly, this two-sample MR study was conducted based on three large-scale GWASs. The large sample size provided sufficient statistical power for making valid causal inferences. Secondly, our results were robust by means of stringent IVs selection criteria and a series of sensitivity analyses. The results were also confirmed by validation analysis and reverse MR analysis. Thirdly, we screened potential intervention targets using Phe-MR analysis in order to comprehensively forecast the on-target side effects of identified lipid metabolites. However, there are several limitations to note. Firstly, our outcomes are pregnancy-specific and limited to women of childbearing age, while the exposure data come from the general population. Although all three GWASs of lipid metabolites were adjusted for sex and age, this study may not have identified sex- or age-specific instrumental variables (IVs) or IVs specifically related to lipid metabolites during pregnancy. Consequently, pregnancy lipidome GWAS data from pregnant women are needed for verification. Secondly, although the stringent IV selection criteria, the number of IVs available for each lipid metabolite was limited. While this approach helps reduce pleiotropy effects and guaranteed the robustness of our causal estimates, it may also inadvertently remove SNPs that influence multiple lipid metabolites through shared biological pathways, thereby potentially diminishing our power to detect causal relationships between lipid metabolites and GDM. Thirdly, the GWAS data used in this MR study came only from European populations (mostly whites), which may have limited the generalizability of our findings when applied to non-European populations and other racial groups. Nevertheless, this restriction minimized population and race stratification bias, though further studies are needed to confirm our findings in populations with different ethnic backgrounds. Moreover, due to strict selection criteria, only 295 of 1,023 lipid metabolites from three large GWAS were included in this study, which represent only a small fraction of the blood lipidome. Further research is required to investigate the associations between additional lipid metabolites and GDM. Lastly, not all the cohorts included in this study were measured using the exactly same blood sample collection methods, such as serum or plasma, fasting or non-fasting blood, which may introduce bias across studies. To validate our findings, further research using larger GWAS with fasting blood samples or individual-level data which allows stratification analysis is necessary. The public health and clinical implications of our study lies in below: by integrating lipidomics with genomics, we provided novel insights that may contribute to the search for promising and safe intervention targets for GDM. Our findings showed a potential causal association between TG_by_PG and GDM, indicating that TG_by_PG could serve as a target for preventive interventions, and in pregnant women at high risk of ICP, a more nuanced approach may involve increasing PG levels rather than lowering TG levels to minimize adverse outcomes. Additionally, XS_VLDL_TG and PSM(d18:1/16:0) were found to be suggestively associated with GDM though further research is necessary to confirm the results. Clinical trials will be essential to confirm the feasibility and safety of these lipid metabolites in the prevention of GDM. Validating these findings will enhance the precision of prevention strategies for GDM.

This systematic MR study identified a potential causal association between TG_by_PG and GDM, as well as suggestive significant associations between XS_VLDL_TG, PSM (d18:1/16:0) and GDM. Side-effect profiles were characterized to inform drug target prioritization, suggesting that TG_by_PG might be a potential target for the prevention and treatment of GDM in caution of the increased risk of ICP.

Gestational diabetes mellitus (GDM), characterized by hyperglycemia developing during pregnancy, is the most frequent pregnancy complication [ 1 ]. GDM increases the risk of multiple short-term and long-term adverse complications in both mother and offspring, including preterm birth, hypertension, obesity, impaired glucose metabolism, and cardiovascular disease [ 2 , 3 ], but relatively little is known about the modifiable targets for prevention and treatment strategies of GDM. Randomized clinical trials have shown that treatment (diet control and exercise) for mild GDM (diagnosed between 24 +0  ~ 30 +6  weeks of pregnancy) does not significantly reduce adverse outcomes including type 2 diabetes mellitus (T2DM), metabolic syndrome, and obesity within seven years postpartum [ 4 , 5 ]. Additional research is necessary to comprehend the causes of GDM and to identify potential intervention targets. The onset of GDM is closely related to lipid metabolism. Metabolic changes, such as physiological insulin resistance [ 6 ] and altered lipid metabolism, [ 7 ] occur during pregnancy, but these changes are more pronounced in women with GDM [ 8 , 9 ], suggesting potential metabolic dysfunction. Lipids play a crucial role in human metabolism, and understanding their relationship with metabolic diseases may help us elucidate the pathological mechanism and identify potential intervention targets. Advances in high-throughput methods have made lipidomics a powerful tool for systematically uncovering lipidomic biomarkers and understanding the pathogenesis of GDM [ 10 ]. An increasing number of observational studies have identified connections between certain lipid metabolites—such as fatty acyls (FAs), glycerolipids (GLs), glycerophospholipids (GPs), sphingolipids (SPs), and lipoproteins—and the risk of GDM [ 11 – 13 ]. Additionally, individual lipid species within certain lipid class may have opposing associations with GDM [ 14 – 16 ], highlighting the necessity of lipidomics research to understand the complex biological effects of individual lipid species on the development of GDM. However, due to the complex correlation and interaction between glucose metabolism and lipid metabolism [ 17 ], it is challenging to determine the causal relationship and the direction of the association between lipid profiles and GDM, especially in observational studies that are prone to confounding and reverse causation. The causal effect of each lipid metabolite on GDM remains unclear, and potential side effects of targeting lipidomic biomarkers for intervention have not been explored. With the rapid increase in genome-wide association studies (GWAS) focusing on blood lipidomes [ 18 – 21 ] and GDM [ 22 ], Mendelian randomization (MR) analysis can be utilized to evaluate the causal relationship between lipidomes and GDM using GWAS summary statistics. In this context, we conducted a two-sample MR study to fulfill two objectives. First, we aimed to examine the causal relationship between lipid metabolites and GDM risk in order to identify potential lipid intervention targets and gain deeper insights into the pathological mechanisms of GDM. Second, prior to clinical trials, we applied a phenome-wide MR (Phe-MR) analysis aimed to assess unanticipated adverse effects of potential lipid biomarker intervention, offering a comprehensive appraisal of their clinical safety.

Supplementary Material 1. Supplementary Material 2. Supplementary Material 1. Supplementary Material 2.