Abstract Background Gestational diabetes mellitus (GDM), characterized by insulin resistance (IR) and β-cell dysfunction, is one of the most common complications of pregnancy with unmet needs of prevention methods. Objective To investigate the causal role of insulin resistance and metabolic pathways in the pathogenesis of GDM with our proposed high-dimensional systematic Mendelian randomization (hdsMR) framework. Methods Cases with GDM and controls with normal glucose tolerance were recruited at the University of Hong Kong–Shenzhen Hospital from 2015 to 2018. A total of 566 participants (aged > 18 years), including 274 with GDM, were enrolled after excluding subjects with major chronic diseases or long-term use of medications affecting glycolipid metabolism. Clinical characteristics and serum samples were collected during the GDM screening stage, and the genome and metabolome were tested. A novel hdsMR framework was proposed to estimate the causal role of IR index (Homeostasis Model Assessment of Insulin Resistance, HOMA-IR) and metabolic pathways in the pathogenesis of GDM. Results Our hdsMR method confirmed that HOMA-IR was causal to GDM (odds ratio, 1.17; 95% confidence interval, 1.04–1.32) and revealed that two metabolic pathways (glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway) mediated 14.6% and 8.4%, respectively, between HOMA-IR and GDM. In an independent validation cohort comprising 255 pre-diabetic individuals, we showed that both pathways could be intervened through diet (P < 0.05). Furthermore, glyoxylate and dicarboxylate metabolism pathway was significantly associated with adverse pregnancy outcomes in GDM. Conclusions These results indicated that targeting specific metabolic pathways through dietary intervention is worth exploring as a possible GDM prevention approach, and hdsMR is more efficient in finding causal mediating metabolic pathways than traditional MR methods. Supplementary Information The online version contains supplementary material available at 10.1186/s12933-025-02746-0. Keywords: Gestational diabetes mellitus, Mendelian randomization, Insulin resistance, Adverse pregnancy outcome, Metabolic pathway Research insights What is currently known about this topic? * Gestational diabetes mellitus (GDM), characterized by metabolic disorders, including increased insulin resistance (IR) and β-cell defects, is one of the most common complications of pregnancy with unmet needs for prevention methods. However, the metabolic mechanism between IR and the pathogenesis of GDM remains unclear. What is the key research question? * What is the relationship among metabolome, IR, and GDM? What is new? * This study proposed a novel high-dimensional systematic Mendelian randomization framework. We identified the causal mediation effect of two metabolic pathways (glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway) from HOMA-IR to GDM and their impact on adverse pregnancy outcomes in GDM subjects. How might this study influence clinical practice? * These results indicated that targeting specific metabolic pathways through dietary modifications could be explored as a possible GDM prevention approach. Background Gestational diabetes mellitus (GDM) affects 14.2% of pregnant women globally and is associated with a variety of adverse pregnancy outcomes (APOs), including preeclampsia, macrosomia, preterm labor, stillbirth, and neonatal hyperinsulinemia [[50]1–[51]3]. Risk factors and underlying mechanisms of GDM have been studied extensively [[52]4–[53]7]. Nevertheless, there is still a scarcity of effective prevention methods to reduce the occurrence of GDM. GDM characterized by metabolic disorders include increased insulin resistance (IR) and β-cell defects [[54]2]. As shown in previous studies, high IR increases the risk of GDM [[55]8], whose resulting metabolic changes may be a potential target for GDM intervention. However, the metabolic mechanism between IR and the pathogenesis of GDM remains unclear. Given that GDM is a metabolic disease, there have been many studies on the relationship between maternal metabolites and GDM [[56]7, [57]9–[58]12]. In addition to this, there is also evidence that modifications of metabolites involved in IR during pregnancy contribute to GDM development [[59]13, [60]14], which provides a novel source of prevention and treatment targets for GDM. Recent research has focused on evaluating the causal relationship between serum metabolites and GDM by Mendelian randomization (MR), but few metabolites were found [[61]15–[62]17], which indicates the limited effects of single metabolites. This also illustrates that traditional MR approaches face challenges in handling high-dimensional metabolomic data and identifying pathway-level mediators. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways, collaborative clusters with the defined biological function, have a more stable and sufficient influence on elucidating disease pathogenesis than individual molecules [[63]18, [64]19]. This provides methodological direction to better quantify the role of metabolic pathways in mediating disease causation. The aim of this study was to propose a high-dimensional systematic MR (hdsMR) framework that integrates MR and pathway quantification to assess causal relationships among IR, metabolic pathways and GDM. Building on these findings, we further examined whether these causal metabolic pathways influence APOs in women with GDM. Finally, using an independent external dataset of pre-diabetic individuals, we validated the potential for dietary interventions to modulate these metabolic pathways in glycemic management. Methods Study participants and ethical approval The present study was conducted at the University of Hong Kong–Shenzhen Hospital from 2015 to 2018. A total of 566 pregnant women who met the following criteria were recruited: (a) Pregnant women aged 18 years or older; (b) performed oral glucose tolerance test (OGTT); and (c) singleton pregnancy. Women with preexisting diabetes mellitus or chronic diseases (including cardiovascular, cerebrovascular, hepatic, renal, or autoimmune disorders), and those receiving long-term medications that affect glycolipid metabolism (e.g., glucocorticoids) were excluded. The study was approved by the Institutional Review Board of the University of Hong Kong–Shenzhen Hospital ([2017]13) and conducted in accordance with the principles of the Declaration of Helsinki as revised in 2013. All of the participants signed written informed consent prior to enrolment. The workflow of the study is illustrated in Fig. [65]1. Fig. 1. [66]Fig. 1 [67]Open in a new tab Overview of hdsMR framework. Insulin-related factors (FINS and HOMA-IR) were screened as GDM risk factors by GLM analysis. Metabolic pathways were quantified by PC1 scores obtained by PCA for the metabolites based on the KEGG database. MR analysis was carried out for the following three parts: (1) MR1 was used to identify causal risk factors of GDM and HOMA-IR; (2) MR2 was used to identify causal metabolic pathways of GDM, and 16 pathways were likely causal for GDM; (3) MR3 was bidirectional MR analysis to identify HOMA-IR–associated pathways, and HOMA-IR was observed to have potential causal impacts on two pathways. Finally, mediation analysis based on two-step MR was performed to estimate the effects of HOMA-IR on GDM via pathways. Diagnosis of GDM The standards recommended by the International Association of the Diabetes and Pregnancy Study Group (IADPSG) were used to diagnose GDM based on 2-h 75-g OGTT [[68]20]. The pregnant women took 75 g of glucose between 24 and 28 weeks of gestation (OGTT week), and their venous plasma glucose level was measured at fasting and at 1 and 2 h after glucose administration. GDM was diagnosed if fasting plasma glucose (FPG) was ≥ 5.1 mmol/L, 1-h plasma glucose (1 h-PG) was ≥ 10.0 mmol/L, or 2-h plasma glucose (2 h-PG) was ≥ 8.5 mmol/L. All values for the OGTT lower than the thresholds were considered normal. Data collection Clinical characteristics Serum samples and clinical characteristics, including age, height, pre-pregnancy weight, pregnancy weight gain, gestational week, FPG, 1 h-PG, 2 h-PG, hemoglobin A1c (HbA1c), total cholesterol (TC), triglycerides (TG), low-density lipoprotein–cholesterol (LDL-C), high-density lipoprotein–cholesterol (HDL-C), fasting insulin (FINS), and total bile acid, were collected during the GDM screening stage. Details of the measurement of these biochemical indicators have been described in previous research [[69]21]. HOMA-β and HOMA-IR were computed from FPG and FINS, as shown in Eqs. ([70]1) and (2) [[71]22].The aera under curve of glucose (G[AUC]) from the 75-g OGTT was calculated as shown in Eq. ([72]3) [[73]23]. Triglyceride-glucose (TyG) index was calculated as Eq. ([74]4) [[75]24]. graphic file with name d33e703.gif 1 graphic file with name d33e709.gif 2 graphic file with name d33e715.gif 3 graphic file with name d33e721.gif 4 The definition of APOs was based on Simmons et al. and included any of the following, such as birth before 37 weeks of gestation, birth weight of 4500 g or greater, birth trauma, neonatal respiratory distress, phototherapy, stillbirth or neonatal death, or shoulder dystocia [[76]25]. Among the GDM population, 64 individuals had APOs. Serum metabolomics Targeted metabolomic analysis was performed using Metabo-Profile (Shanghai, China). Detailed serum sample preparation, chemical materials for targeted metabolomics, and mass spectrometry (MS) acquisition and chromatographic conditions have been described previously by Luo et al [[77]21]. Raw data files generated by ultra-high performance liquid chromatography (UPLC)-MS/MS were processed with Masslynx software (v4.1, Waters, Milford, MA, USA). A total of 200 metabolites were detected. Limit of detection was applied to fill in missing values of quantitative metabolomic data. Single-nucleotide polymorphisms (SNP) genotyping and quality control Genotyping was performed using Infinium Asian Screening Array-24 v1.0 BeadChip (Illumina, Inc., San Diego, CA, United States). The genotype data underwent strict quality control, and PLINK files were generated for subsequent analysis via PLINK (Version 1.9) [[78]26]. SNPs with minor allele frequency < 0.01 and those not in Hardy–Weinberg equilibrium (HWE) were removed. Specifically, SNPs with HWE P < 1 × 10^−6 in controls and P < 1 × 10^−10 across all samples were excluded. After filtering, genotype imputation was done with 1000 Genomes Project data. To reduce linkage disequilibrium (LD), SNP pairs with r^2 > 0.2 were removed. Principal component analysis (PCA) was then applied to adjust for population substructure, retaining 566 samples and 479,053 SNPs for further analysis. External metabolic validation data We collected serum metabolite profiles from a randomized, controlled, single-blind dietary intervention study ([79]NCT03222791) [[80]27]. In that work, researchers conducted a 6-month dietary intervention in pre-diabetic individuals and 225 participants were randomly assigned to either a personalized postprandial glucose–targeting diet (PPT) (n = 113) or a Mediterranean diet (MED) (n = 112). The results demonstrated that diet intervention had a positive impact on glycemic control through changes in serum metabolites. The data can be accessed at [81]https://static-content.springer.com/esm/art%3A10.1038%2Fs41467-023- 41042-x/MediaObjects/41467_2023_41042_MOESM9_ESM.xlsx. Metabolite landscape construction of the study population MetaboAnalyst 6.0 ([82]https://www.metaboanalyst.ca) [[83]28] was used to analyze the metabolites in the NGT and GDM groups by univariate and multivariate analyses online. The differentially expressed metabolites (DEMs) were detected through a volcano plot, which combined the results from fold change (FC) analysis (FC > 1) and t tests (FDR < 0.05) into a single graph based on both biological significance and statistical significance. PCA was utilized for unsupervised clustering of metabolites among all samples, and one GDM subject that deviated from other samples was removed. Metabolite set enrichment analysis (MSEA) was performed to directly investigate the biological functions of the DEMs in the KEGG database with P < 0.05 regarded as statistically significant. Framework of hdsMR Metabolic pathways quantization The latest KEGG pathways and metabolites contained in the pathways were downloaded by the KEGGREST package in R (10.18129/B9.bioc.KEGGREST). Of the 200 metabolites detected, 92 were mapped to 205 pathways in the KEGG database. A hypergeometric test was performed to calculate the P values corrected by Bonferroni for those matched pathways, which was completed by the magrittr package in R (10.32614/CRAN.package.magrittr). FDR < 0.05 was taken as the significant enrichment, and 52 of the 92 mapped pathways met this condition. Then, we performed PCA on the metabolites that each significant pathway contains and extracted the PC1 score as the quantitative characterization of the pathway. PCA was conducted by prcomp function in R ([84]https://www.r-project.org/). Genetic variants associated with insulin-related factors and metabolic pathways To test for the association between genetic variants and GDM, a logistic regression model was used in PLINK 1.9 [[85]26], including family history of diabetes, pre-gestational body mass index (BMI), changes in BMI, age, and principal component (PC) factors from population stratification as covariates. As for the associations of genetic variants with insulin-related factors (FINS, HOMA-IR) and metabolic pathways, a linear regression (additive model) was performed, adjusting for family history of diabetes, pre-gestational BMI, changes in BMI, age, and top 10 genotype-based PCs. The adjust parameter was used to obtain the P values of the multiple test correction for the abovementioned association analysis, providing Bonferroni-corrected P values along with FDR and other parameters. Causal associations among insulin-related factors, metabolic pathways, and GDM MR uses genetic variants to assess the causal relationships using observational data. A genetic variant can be considered an instrumental variable (IV) for a given exposure if it satisfies the following IV assumptions [[86]29, [87]30]: (1) it is associated with the exposure; (2) it is not associated with the outcome due to confounding pathways; and (3) it does not affect the outcome, except potentially via the exposure. In this study, we performed MR using data from a single sample (known as one-sample MR), in which genetic variants, exposure, and outcome were measured in the same individuals [[88]31]. All of the analyses were conducted using the MendelianRandomization (10.32614/CRAN.package.MendelianRandomization) and TwoSampleMR ([89]https://mrcieu.github.io/TwoSampleMR/) packages in R. MR analysis was carried out for three parts in our study as follows: (1) MR1 analysis step was to find out whether insulin-related factors were causal for GDM; (2) MR2 analysis step was to find metabolic pathways causal for GDM; and (3) MR3 analysis step was a bidirectional MR between insulin-related traits and metabolic pathways that were all associated with GDM. SNPs selected for MR1 were based on a suggestive threshold of Bonferroni-corrected P values < 5 × 10^−8 [[90]32], and those for MR2 and MR3 were based on P values < 1 × 10^−5 [[91]33]. SNPs selected for IVs that are associated with GDM (P < 1 × 10^−5) in PhenoScannerV2 ([92]http://www.phenoscanner.medschl.cam.ac.uk/) were excluded when performing MR1 and MR2. F statistic was used to evaluate the effects of weak IVs [[93]34], which was calculated with the following formula: Inline graphic , where N is the sample size in GWAS analysis, k is the number of IVs, and R^2 is the extent to which the IVs explain the exposure. R^2 was obtained from the get_r_from_pn function of the TwoSampleMR package. The SNPs with an F statistic greater than 10 were considered strong IVs and remained for the analysis. To avoid LD, the IVs were clumped using the criterion r^2 < 0.2 with a clumping window of 50 kb for independence. The inverse-variance weighting (IVW) method, either in a fixed-effect framework (IVs ≤ 3) or in a multiplicative random-effect meta-analysis framework (IVs > 3) [[94]35], was used to generate an overall estimate of the causal effect in each MR analysis. For each single SNP that remained after clumping, the Wald ratio was used, which is the most basic method. Other methods such as MR Egger, Weighted median, Simple mode, and Weighted mode were also completed in each MR analysis. The suggested threshold of P < 0.05 was used as a significance level for MR results. Sensitivity analysis Several sensitivity analyses were conducted to assess the robustness of results to potential violations of the MR assumptions [[95]36]. First, heterogeneity tests, which could use MR Egger and IVW, were estimated by the Cochran Q test. When the P value of Cochran Q-test results was below 0.05, the heterogeneity of the MR results was indicated. Second, the intercept term in MR Egger regression was used as an indication of whether directional horizontal pleiotropy was driving the results of the MR analysis [[96]37]. Furthermore, leave-one-out analysis was performed to identify whether a single SNP was driving the association. In the leave-one-out analysis, the MR was performed again, but leaving out each SNP one by one. If the result changed greatly after the elimination of an SNP, it indicated that there was an SNP with a great influence on the result. Mediation analysis For insulin-related factors that causally associate with both metabolic pathways and GDM, a mediation analysis based on two-step MR [[97]38] was used to quantify the effects of the risk factors on GDM via pathways. Total effect of exposure on outcome included both direct and indirect effects through mediators. A univariate MR model was carried out to estimate the effect of the exposure on the mediator. To estimate the indirect effect, results from two-step MR were used. The Product method was chosen to estimate the beta of the indirect effect, and the Delta method was used to estimate the standard error (SE) and confidence interval (CI). Statistical analysis Clinical characteristics between the normal glucose tolerance (NGT) and GDM groups or between the normal pregnancy outcome (NPO) and APO groups were compared using the t test or Wilcoxon test for continuous variables and the chi-square test for categorical variables with R version 4.3.2. Spearman correlation analysis was used to assess relationships between continuous variables. A generalized linear model (GLM) was used to investigate the relationship between HOMA-IR and APOs. In the validation study, PC1 scores of metabolic pathways were derived using the hdsMR method, and changes in PC1 before and after dietary intervention were statistically evaluated via a paired T-test. Two-sided P values lower than 0.05 were considered significant. Results were visualized using the ggplot2 package in R (10.32614/CRAN.package.ggplot2). Results Clinical characteristics and metabolite landscape of the study population A total of 566 (aged > 18 years) participants with GDM or NGT were recruited in line with the diagnostic criteria of the IADPSG at the University of Hong Kong–Shenzhen Hospital from 2015 to 2018 (Fig. [98]1). The clinical characteristics of the 566 study participants from the two groups are summarized in Table [99]1. Compared with the NGT group, the GDM group had higher levels of HbA1c, FINS, LDL-C, FPG, 1 h-PG, 2 h-PG, G[AUC], HOMA-β, HOMA-IR, TyG and family history of diabetes, and lower levels of TC and changes in BMI during pregnancy (P < 0.05). Table 1. Clinical characteristics of the study participants with or without GDM NGT (n = 292) GDM (n = 274) P Age (years) 29 (28, 31) 29 (28, 30) 0.168 Gestational age (week) 27 (26, 28) 27 (26, 28) 0.641 Pre-gestational BMI (kg/m^2) 20.65 ± 2.53 20.97 ± 2.68 0.148 Changes in BMI (kg/m^2) 5.43 ± 1.33 4.72 ± 1.53  < 0.001 FPG (mmol/L) 4.39 ± 0.28 4.67 ± 0.47  < 0.001 1 h-PG (mmol/L) 7.27 ± 1.40 9.70 ± 1.40  < 0.001 2 h-PG (mmol/L) 6.33 ± 0.97 8.56 ± 1.36  < 0.001 HbA1c (%) 5.15 ± 0.30 5.26 ± 0.30  < 0.001 Total cholesterol (mmol/L) 6.40 ± 1.18 5.92 ± 1.13  < 0.001 Triglycerides (mmol/L) 2.49 ± 1.26 2.60 ± 1.28 0.309 LDL-C (mmol/L) 3.14 ± 0.85 3.31 ± 0.94 0.025 HDL-C (mmol/L) 2.01 ± 0.37 1.96 ± 0.41 0.121 G[AUC] (mmol/L·h) 12.63 ± 1.74 16.31 ± 1.70  < 0.001 Fasting insulin (mU/L) 6.05 (4.15, 9.22) 8.69 (5.83, 12.20)  < 0.001 HOMA-β 144.65 (88.68, 215.74) 155.01 (109.43, 248.92) 0.013 HOMA-IR 1.20 (0.79, 1.79) 1.84 (1.12, 2.63)  < 0.001 Total bile acid (μmol/L) 2.27 (1.67, 2.96) 2.22 (1.53, 3.07) 0.802 Family history of diabetes, n (%) 22 (7.5) 68 (24.6)  < 0.001 TyG 9.0 ± 0.5 9.1 ± 0.4 0.014 [100]Open in a new tab BMI body mass index; FPG fasting plasma glucose; 1 h-PG one hour postprandial glucose; 2 h-PG two hours postprandial glucose; HbA1c hemoglobin A1c; LDL-C low-density lipoprotein cholesterol; HDL-C high-density lipoprotein cholesterol; G[AUC] area under the curve of glucose from the 75-g OGTT; HOMA-β homeostasis model assessment index of β-cell secretion; HOMA-IR homeostasis model assessment of insulin resistance; TyG triglyceride-glucose Classical metabolome analysis showed that serum metabolism in GDM patients was significantly different from that in pregnant women with NGT (Fig. [101]2). The detected metabolites were grouped into 13 chemical classes (Fig. [102]2A). Namely, 37 DEMs and corresponding metabolic pathways were obtained (Fig. [103]2C, [104]D; Tables [105]S1 and [106]S2). Weighted correlation network analysis (WGCNA) detected one module (MEbrown) that positively correlated (P < 0.05) with GDM and many clinical traits, such as 2 h-PG, FINS, HOMA-β, and HOMA-IR (Fig. [107]2E; Tables [108]S3 and [109]S4). Fig. 2. [110]Fig. 2 [111]Open in a new tab Metabolite profile characteristics between the NGT and GDM groups. A The statistics of metabolite compositions detected in serum samples from the GDM and NGT groups. B Sample distributions by PCA based on the metabolite profiles from all participants. C Differentially expressed metabolites (DEMs) between the GDM and NGT groups shown in a volcano plot. D Functional enrichments of DEMs and recognition of GDM-associated biological pathways. E Metabolite modules newly identified by WGCNA and their association with clinical indicators, especially their relationships with GDM outcome. NGT, normal glucose tolerance; GDM, gestational diabetes mellitus. Identification of the causal role of HOMA-IR and metabolic pathways in GDM GLM analysis revealed that FINS (odds ratio [OR] 1.20; 95% CI 1.14–1.26; P < 0.001) and HOMA-IR (OR 2.55; 95% CI 2.04–3.25; P < 0.001) were both significantly associated with GDM (Fig. [112]S1). Thus, we used the hdsMR framework (Fig. [113]1) for causal analysis among these insulin-related factors (i.e., FINS and HOMA-IR), metabolic pathways, and GDM (Fig. [114]3A). We identified 23 SNPs associated with FINS and HOMA-IR (P[adj] < 5 × 10^−8; Table [115]S5). HOMA-IR was confirmed as a causal GDM risk factor in the MR1 analysis step by the IVW test (OR 1.17; 95% CI 1.04–1.32; P = 0.007; Fig. [116]3B). In the MR2 analysis step, 16 metabolic pathways characterized with PC1 scores were causal for GDM (P < 0.05) by the IVW test (Table [117]S6; Fig. [118]3C). In the MR3 analysis step, bidirectional MR analysis was performed on HOMA-IR and 16 GDM-associated pathways (Fig. [119]3D). We found that HOMA-IR had a potential causal risk effect on glyoxylate and dicarboxylate metabolism pathway (IVW: OR 1.08; 95% CI 1.08–1.15; P = 0.006) and a causal protective effect on lysine degradation pathway (IVW: OR 0.95; 95% CI 0.91–0.98; P = 0.006). By contrast, the causal effect of these two metabolic pathways on HOMA-IR was not confirmed (Fig. [120]3D). Fig. 3. [121]Fig. 3 [122]Open in a new tab Key metabolic pathways with causal relationship for GDM outcome and HOMA-IR identified by one-sample MR. A The basic MR model used in the GDM study for inferring the causal relationship among HOMA-IR, key pathways, and GDM outcome, which includes three MR determinations. B Association between HOMA-IR and GDM determined by one-sample MR, which indicates the relation between biochemistry exposure and outcome. C Association between metabolic pathways and GDM determined by one-sample MR, which supports the relation between biological exposure and the same outcome. Here, 16 key metabolic pathways were selected. D Association between HOMA-IR and metabolic pathways determined by bidirectional MR, which identified the causal relation direction from HOMA-IR to two key metabolic pathways. The robustness of the abovementioned MR analyses was confirmed by the results of sensitivity analyses. The IVs had no horizontal pleiotropy in the association of HOMA-IR and any of the two key metabolic pathways, as measured by MR-Egger intercept P > 0.05 (Table [123]S7). There was no evidence of heterogeneity in each MR analysis step, according to Cochran’s Q-test P > 0.05 by both the MR Egger and IVW methods (Table [124]S9). Furthermore, the results of leave-one-out analysis showed that no single SNP was driving the association (Fig. [125]S2). However, there was no evidence for colocalization between the exposure and outcome at any of the MR analysis steps based on a Bayesian algorithm (PP.H4.abf < 60%; Table [126]S7), which was possibly due to violation of its single-causal-variable hypothesis. Besides, the HOMA-IR value was increased in the GDM group (Fig. [127]4D), and correlations between the two metabolic pathways and GDM/HOMA-IR were consistent with the MR estimation. PC1 score of glyoxylate and dicarboxylate metabolism pathway was higher in GDM and that of lysine degradation pathway was lower (Fig. [128]4D). HOMA-IR showed a risk trend with glyoxylate and dicarboxylate metabolism pathway and a protective trend with the lysine degradation pathway (Fig. [129]4E). Fig. 4. [130]Fig. 4 [131]Open in a new tab Mediating role of the key metabolic pathways from HOMA-IR to GDM identified by two-step MR. A A two-step MR model for mediation analysis used in this study. B Mediation analysis of glyoxylate and dicarboxylate metabolism pathway, indicating its upregulation in GDM. C Mediation analysis of lysine degradation pathway, indicating its downregulation in GDM. D The consistent alteration between HOMA-IR and metabolic pathway for GDM. HOMA-IR was increased in GDM (left panel). The PC1 score of glyoxylate and dicarboxylate metabolism pathway was increased in GDM (middle panel). Meanwhile, the PC1 score of the lysine degradation pathway was decreased in GDM (right panel). E Significant association between PC1 scores of two key metabolic pathways and HOMA-IR. F The PC1 scores of glyoxylate and dicarboxylate metabolism pathway decreased after both MED and PPT diet intervention (left panel). Meanwhile, the PC1 scores of lysine degradation pathway increased after both MED and PPT diet intervention (right panel). MED diet, the standard of care Mediterranean diet; PPT diet, a personalized postprandial glucose–targeting diet. Mediation effect of the key metabolic pathways between HOMA-IR and GDM Next, we conducted a two-step MR analysis to infer the causal mediating effect of HOMA-IR on GDM through two key metabolic pathways (Fig. [132]4A). Glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway significantly mediated the association between HOMA-IR and GDM, explaining 14.6% (indirect effect = 1.02; 95% CI 1.01–1.10; Fig. [133]4B) and 8.4% (indirect effect = 1.013; 95% CI 1.00–1.03; Fig. [134]4C) of the total effect, respectively. We also performed conventional MR analysis on the metabolite level (Fig. [135]S3A; Table [136]S8) to investigate the causal relationships among HOMA-IR, metabolites (mediators), and GDM (outcomes). Although eight metabolites were causally related to GDM and one was related to HOMA-IR, none of them mediated HOMA-IR to GDM (Fig. [137]S3C, D). External validation of the key metabolic pathways on GDM development and prevention To explore the interactions between metabolic pathways as well as their relationship with GDM, we constructed a regularized partial correlation network using the metabolites involved in glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway (Fig. [138]S4). This network showed that dense pathway crosstalk occurred between the two pathways, indicating that they may regulate each other. According to the STRING database [[139]39], the GLYCTK gene involved in the glyoxylate and dicarboxylate metabolism pathway was linked to the well-known GDM risk gene HKDC1 [[140]40] (Fig. [141]S5A). Genes EZH2 and PRDM16, involved in the lysine degradation pathway, were linked to the known GDM-related gene PPARG [[142]41] (Fig. [143]S5B). We also validated the effects of these two pathways on diabetes prevention with dietary intervention using an external independent dataset from pre-diabetic individuals. PC1 scores of the two key pathways significantly changed after the dietary intervention (P < 0.05), namely the PC1 score of glyoxylate and dicarboxylate metabolism pathway decreased, while that of the lysine degradation pathway increased (Fig. [144]4F). Association between glyoxylate and dicarboxylate metabolism pathway and APOs in GDM We further evaluated the impact of the two pathways on pregnancy outcomes in GDM patients and found that the glyoxylate and dicarboxylate metabolism pathway was not only associated with GDM but also contributed to APOs (Fig. [145]5). The APO subjects had higher TC, TG, FINS, and HOMA-IR compared with the NPO subjects in the GDM group (Table [146]2; Fig. [147]5A). HOMA-IR also showed a risk effect on APOs in GLM analysis after adjusting for TG, TC, age, changes in BMI, and pre-gestational BMI (Fig. [148]5B). Through evaluating the association between pathway activation and pregnancy outcomes, glyoxylate and dicarboxylate metabolism pathway was significantly related to APOs (Fisher’s exact test: OR 2.13; 95% CI 1.16–3.94; P = 0.01) (Table [149]S9). The trend of the average activity of glyoxylate and dicarboxylate metabolism pathway in the APO group was also higher than that in the NPO group (Fig. [150]5C), consistent with its risk role in GDM development. Fig. 5. [151]Fig. 5 [152]Open in a new tab Role of glyoxylate and dicarboxylate metabolism pathway in the association with HOMA-IR and APO of GDM. A HOMA-IR differences of GDM individuals between APO and NPO groups. NPO, normal pregnancy outcome; APO, adverse pregnancy outcome. B Risk model for APO outcome of GDM individuals. C Activity differences in glyoxylate and dicarboxylate metabolism pathway observed between the APO and NPO groups. D The enrichment of KEGG pathways related to the genes involved in glyoxylate and dicarboxylate metabolism and lysine degradation pathways (top panel), and the associations of those pathways with GDM and APO (bottom panel). Table 2. Characteristics of the NPO and APO groups for GDM patients NPO (n = 209) APO (n = 64) P Age (years) 30 (28, 31) 29 (28, 30) 0.001 Gestational age (week) 26 (25, 26) 26 (25, 27) 0.111 Pre-gestational BMI (kg/m^2) 20.90 ± 2.70 21.24 ± 2.59 0.375 Changes in BMI (kg/m^2) 4.73 ± 1.55 4.70 ± 1.50 0.921 FPG (mmol/L) 4.66 ± 0.46 4.71 ± 0.51 0.522 1 h-PG (mmol/L) 9.69 ± 1.38 9.71 ± 1.46 0.914 2 h-PG (mmol/L) 8.53 ± 1.35 8.66 ± 1.40 0.507 HbA1c (%) 5.24 ± 0.30 5.31 ± 0.30 0.155 Total cholesterol (mmol/L) 5.87 ± 1.14 6.16 ± 1.20 0.032 Triglycerides (mmol/L) 2.48 ± 1.04 2.97 ± 1.90 0.014 LDL-C (mmol/L) 3.28 ± 0.94 3.48 ± 0.92 0.071 HDL-C (mmol/L) 1.96 ± 0.40 1.90 ± 0.42 0.323 G[AUC] (mmol/L·h) 16.28 ± 1.65 16.40 ± 1.88 0.621 Fasting insulin (mU/L) 8.62 (5.70, 12.15) 10.18 (7.58, 14.19) 0.024 HOMA-β 152.93 (110.13, 238.94) 190.34 (116.81, 268.75) 0.071 HOMA-IR 1.81 (1.11, 2.60) 2.06 (1.56, 3.03) 0.030 Total bile acid (μmol/L) 2.17 (1.55, 3.00) 2.53 (1.51, 3.54) 0.359 Family history of diabetes, n (%) 43 (20.6) 22 (34.4) 0.023 [153]Open in a new tab BMI body mass index; FPG fasting plasma glucose; 1 h-PG one hour postprandial glucose; 2 h-PG two hours postprandial glucose; HbA1c hemoglobin A1c; LDL-C low-density lipoprotein cholesterol; HDL-C high-density lipoprotein cholesterol; G[AUC] area under the curve of glucose from the 75-g OGTT; HOMA-β homeostasis model assessment index of β-cell secretion; HOMA-IR homeostasis model assessment of insulin resistance Besides, based on the STRING and KEGG databases, these two key pathways had cross talk with 17 biological pathways through their involved genes (Fig. [154]5D, top). According to previous studies (Table [155]S10; Fig. [156]2D), 11 of the 17 pathways were related to GDM, but one pathway was related to APO [[157]42], which only cross-talked with glyoxylate and dicarboxylate metabolism pathway (Fig. [158]5D, bottom; Methods). Discussion In this study, we elucidated the role of metabolic pathways in GDM development by providing a high-dimensional causal mediation analytic framework for investigating the relationship among metabolome, IR, and GDM. The proposed hdsMR framework applied the PC1 score of PCA to quantify metabolomic biological pathways, which overcame the high dimensionality and instability of the traditional metabolomic MR analysis. Under the hdsMR framework, we confirmed that HOMA-IR had a causal risk effect on GDM, indicating the essential role of IR in the pathogenesis of GDM, consistent with classical epidemiological studies [[159]8, [160]43–[161]47]. We found that the association between HOMA-IR and GDM could be mediated by glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway, the former being a risk factor for GDM and the latter being a protective factor for GDM. The two pathways were also enriched by DEMs through the KEGG pathway enrichment analysis (Fig. [162]2D). In addition, both key pathways were validated to be potential targets to prevent GDM through dietary intervention in an external independent dataset [[163]27]. Finally, we showed glyoxylate and dicarboxylate metabolism pathway as a risk factor for APOs in GDM. Traditional metabolomic MR studies have usually been based on metabolites to describe changes associated with disease states [[164]15–[165]17], but the single-molecule level is not systematic enough to gain an in-depth understanding of biological significance. Therefore, we focused on finding new ways to infer causality between metabolism and disease. Gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA) have been applied for pathway quantification to reduce the dimensionality of multiple metabolites [[166]48]. Our hdsMR framework utilized PCA [[167]49] to quantify metabolic pathways to efficiently infer causality and mediating factors among metabolome, IR, and GDM, based on the existing MR mediation methods [[168]38, [169]50]. Compared to traditional MR mediation methods that yielded no significant findings, our hdsMR framework identified two biologically plausible causal mediation pathways, demonstrating both superior statistical power and clearer mechanistic interpretations. The consistency of results across multiple sensitivity analysis approaches (Cochran Q test, MR-Egger and Leave-one-out) supported the robustness of our findings. Especially leave-one-out tests confirmed that the MR results were not influenced by any single SNP. Based on rigorous screening processes and sensitivity tests, we ensured the precision and validity of the selected genetic variants. To validate the effectiveness of our hdsMR method, we conducted data simulation comparison with traditional MR analysis. Our simulation results (Fig. [170]S6) demonstrated that the hdsMR method performed better than traditional MR analysis. Across different noise levels, our approach showed higher accuracy. HOMA-IR is an index used to evaluate the level of IR in individuals [[171]22]. In observational studies, high HOMA-IR before or at the early stage of pregnancy was associated with an increased risk of GDM [[172]8, [173]43–[174]46]. Consistent with this, we provided robust evidence that HOMA-IR was a causal risk factor for GDM (P = 0.007), indicating that HOMA-IR could be used for GDM risk screening in early pregnancy and even before pregnancy. Moreover, previous studies have indicated that strategies to improve IR may help reduce the risk of GDM [[175]44]. We found that the causal relationship of HOMA-IR with GDM could be mediated by the glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway. Of the two mediating pathways, the glyoxylate and dicarboxylate metabolism pathway describes a variety of reactions involving glyoxylate or dicarboxylates in the KEGG database [[176]18], which has previously been related to obesity [[177]51], an important risk factor for GDM [[178]52]. Glyoxylate and dicarboxylate metabolism was upregulated at the peri-implantation period of early pregnancy in mice [[179]53]. The evidence indicates that this pathway is physiologically upregulated during a normal pregnancy. In our study, we showed that glyoxylate and dicarboxylate metabolism was a direct causal risk for GDM. Our findings reveal that GDM is associated with an exaggerated activation of this pathway. Previous studies have shown that the activity of this pathway is increased in patients with type 2 diabetes (T2D) [[180]51]. In animal studies, it was associated with a high glucose intake in prediabetes [[181]54] and was related to IR in T2D rats [[182]55]. Overall, the amplified pathway activity may reflect a response to IR. In addition, we found that glyoxylate and dicarboxylate metabolism was a risk factor for APOs in GDM. This pathway has been reported to be involved in the development of fetal growth restriction [[183]56], which leads to a higher risk of mortality and neonatal complications with long-term consequences [[184]57]. This pathway was also altered in other APOs, such as neonates of preeclamptic pregnancies and preterm infants with bronchopulmonary dysplasia [[185]58, [186]59]. Taken together, the inhibition of glyoxylate and dicarboxylate metabolism represents a promising intervention target to prevent GDM and APO, but further experimental studies are warranted to illustrate the mechanism. The other pathway that mediates HOMA-IR to GDM development is the lysine degradation pathway, which occurs in the liver [[187]60] and was shown to be a protective factor for GDM in our study. It was highly associated with metabolic syndrome of prediabetes according to metabolic pathway analysis in a population-based study [[188]61]. Changes in the concentrations of metabolites for lysine degradation correlated with pre-diabetic state in an animal study [[189]62]. L-lysine is the initial substrate of the lysine degradation pathway [[190]18]. In T2D patients, L-lysine supplementation has been reported to reduce the protein glycation [[191]63], which can be linked to long-term hyperglycemia [[192]64]. L-lysine is abundant in legumes [[193]65], so increasing their intake might lower the risk of GDM [[194]66, [195]67]. The abovementioned studies suggest that the lysine degradation pathway highly correlates with abnormal glucose metabolism, which may become a potential dietary target for GDM prevention and therapy. Furthermore, we showed that both key pathways could be altered after diet interventions according to an independent metabolome data assessment from pre-diabetes individuals [[196]27]. Based on evidence suggesting shared pathogenic mechanisms and intervention between GDM and pre-diabetic individuals, we selected a metabolic dataset reflecting 6-month dietary intervention changes in pre-diabetic individuals to approximate whether the two metabolic pathways we identified could achieve glycemic management through dietary intervention. The original research based on this dataset showed that diets demonstrated a positive impact on both serum metabolites and glycemic control. Therefore, we analyzed the changes in metabolic pathways and found that the glyoxylate and dicarboxylate metabolism pathway was downregulated, while the lysine degradation pathway was upregulated after a PPT or MED intervention. Our findings suggest that early screening for HOMA-IR could identify pregnant women at high risk for GDM. For this high-risk population, dietary interventions (e.g., PPT diet or MED diet) or prebiotic supplements targeting both metabolic pathways [[197]27, [198]66, [199]67] may be beneficial in reducing their risk of conversion to GDM. In addition, evidence from population studies and animal studies indicate that dietary interventions may mitigate the risk of GDM through the improvements in IR [[200]68, [201]69]. The role of other lifestyles such as regular physical activity and adequate sleep duration on these metabolic pathways needs to be further investigated. There are some limitations to our study. First, as a cross-sectional study, we only collected samples at the time of the OGTT, lacking samples from early pregnancy. Although MR methods were used to exclude confounding, future prospective studies of the association between early indices of IR and GDM are still needed. Second, we used one-sample MR with not a large sample size, which resulted in low statistical power; however, we performed various sensitivity analyses to ensure the reliability of our MR analysis, and the low power may cause false negative results that would not weaken our positive findings. Third, as this study was conducted in a single-center Chinese population, the generalizability of the results may be limited. Fourth, although our study advances the current understanding of metabolic pathways mediating GDM and its perinatal outcomes, the critical evaluation of long-term health outcomes (e.g., T2D, metabolic syndrome, and cardiovascular complications) in mothers and offspring necessitates future large-scale birth cohort studies with extended follow-up periods. Collectively, in this study, we performed hdsMR analysis and discovered the mediating roles of metabolic biological pathways between IR and GDM. Our findings demonstrated that both mediating pathways, namely glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway, could be intervened by diet, which provides evidence for potential prevention methods of GDM. The proposed hdsMR framework overcomes the high dimensionality and instability of the traditional metabolomic MR analysis and is useful for investigating the underlying biological mechanism of diseases. Conclusions This study exhibited the utility of the hdsMR framework for estimating the causal role of HOMA-IR and metabolic pathways in the pathogenesis of GDM. We identified the causal mediation effect of two metabolic pathways (glyoxylate and dicarboxylate metabolism pathway and lysine degradation pathway) from HOMA-IR to GDM and their impact on APOs in GDM subjects. We further highlighted that targeting specific metabolic pathways through dietary modifications could be explored as a possible GDM prevention approach, and hdsMR was more efficient in finding causal mediating metabolic pathways than traditional MR methods. Supplementary Information [202]Supplementary Material 1.^ (870.4KB, pdf) [203]Supplementary Material 2.^ (175KB, xls) Acknowledgements