DNA methylation patterns and predictive models for metabolic disease risk in offspring of gestational diabetes mellitus

Abstract

Background

Gestational diabetes mellitus (GDM) is a common pregnancy complication with far-reaching implications for maternal and offspring health, strongly tied to epigenetic modifications, particularly DNA methylation. However, the precise molecular mechanisms by which GDM increases long-term metabolic disease risk in offspring remain insufficiently understood.

Methods

We integrated multiple publicly available whole-genome methylation datasets focusing on neonates born to mothers with GDM. Using differentially methylated positions (DMPs) identified in these datasets, we developed a machine learning model to predict GDM-associated epigenetic changes, then validated its performance in a clinical target cohort.

Results

In the public datasets, we identified DMPs corresponding to genes involved in glucose homeostasis and insulin sensitivity, with marked enrichment in insulin signaling, AMPK activation, and adipocytokine signaling pathways. The predictive model exhibited strong performance in public data (AUC = 0.89) and moderate performance in the clinical cohort (AUC = 0.82). Although CpG sites in the PPARG and INS genes displayed similar methylation trends in both datasets, the small validation cohort did not yield statistically significant differences.

Conclusions

By integrating robust public data with a targeted validation cohort, this study provides a comprehensive epigenetic profile of GDM-exposed offspring. Owing to the limited sample size and lack of statistical significance, definitive conclusions cannot yet be drawn; however, the observed directional consistency suggests promising avenues for future research. Larger and more diverse cohorts are warranted to confirm these preliminary findings, clarify their clinical implications, and enhance early risk assessment for metabolic disorders in children born to GDM mothers.

Background

GDM is a common pregnancy complication characterized by glucose intolerance first recognized or diagnosed during pregnancy [1]. The global incidence of GDM has been increasingly developing, and current research suggests about 13–31% of all global pregnancies are affected, making it a significant public health issue [2, 3]. GDM not only poses risks to maternal health but also has profound implications for fetal development and long-term offspring health [3]. Accumulating evidence suggests that GDM may exert lasting effects on offspring health through epigenetic mechanisms, with DNA methylation playing a crucial role in this process [4, 5]. DNA methylation, a key epigenetic modification that does not alter the DNA sequence, plays vital roles in embryonic development, gene expression regulation, and disease pathogenesis [6]. Studies have shown that GDM may influence gene expression and metabolic pathways by altering DNA methylation patterns in fetal tissues, potentially increasing the risk of metabolic disorders and cardiovascular diseases in offspring later in life [7, 8]. Specifically, GDM has been associated with altered methylation states of genes involved in energy metabolism, insulin signaling pathways, and inflammatory responses [9, 10].

Multiple studies have demonstrated that maternal hyperglycemia during pregnancy may influence gene expression and metabolic pathways in the developing fetus by altering DNA methylation patterns in various tissues [11,12,13]. Specifically, GDM has been associated with differential methylation of genes involved in energy metabolism, insulin signaling pathways, and inflammatory responses [14]. These epigenetic alterations potentially increase the offspring's risk of developing metabolic disorders and cardiovascular diseases later in life, suggesting a mechanism for transgenerational transmission of metabolic risk [15]. Recent advancements in high-throughput sequencing technologies have provided powerful tools for exploring the impact of GDM on the offspring's epigenome through genome-wide DNA methylation analysis [16]. Studies utilizing DNA methylation data from cord blood and placental tissues have revealed associations between GDM exposure and changes in methylation levels of specific genomic regions [15]. However, existing research results exhibit some heterogeneity and inconsistency due to limitations in sample size, differences in study populations, and variations in technological platforms [17,18,19].

To gain a more comprehensive and systematic understanding of GDM's impact on the offspring's epigenome, integrating large-scale public data analysis with small-sample experimental validation has become increasingly important [20, 21]. This approach not only leverages the statistical power of large sample sizes available in public databases but also confirms key findings through targeted experimental validation [22, 23]. Furthermore, the application of advanced bioinformatics methods, such as machine learning, offers new avenues for identifying potential biomarkers and predictive models from complex epigenetic data [24, 25].

This study aims to comprehensively explore the effects of GDM on DNA methylation patterns in offspring cord blood by integrating large-scale DNA methylation data from public databases with small-scale validation experimental data. We will employ systematic bioinformatics analysis methods combined with experimental validation to identify key differentially methylated positions and regions associated with GDM, uncover potential functional pathways and regulatory networks, and construct possible predictive models [17, 26]. This integrated analytical strategy not only contributes to a deeper understanding of the mechanisms by which GDM affects the offspring's epigenome but may also provide new insights and directions for early diagnosis, risk assessment, and personalized interventions for GDM [27, 28].

We have designed a rational and comprehensive research methodology: by combining the strengths of big data analysis and targeted experimental validation (Fig. 1), this study seeks to bridge the gap between large-scale epigenomic discoveries and their biological significance in the context of GDM. The findings from this research have the potential to elucidate the epigenetic basis of GDM's long-term effects on offspring health, potentially leading to novel diagnostic and therapeutic strategies. Moreover, this integrated approach may serve as a model for future epigenetic studies in other complex diseases, demonstrating the power of combining bioinformatics with experimental biology to address critical questions in medical research.

Fig. 1

Framework of the integrated research design for investigating the epigenetic impact of Gestational Diabetes Mellitus on offspring. The diagram illustrates the four main stages of the study: A Public Data Acquisition, B Bioinformatics Analysis, C Experimental Validation, and D Integrated Interpretation

Methods

Overview of the study design

The research design for this study integrates large-scale public data analysis with small-sample experimental validation to comprehensively investigate the epigenetic impact of GDM on offspring. This approach follows recent recommendations for multi-stage epigenetic investigations that combine public dataset mining with targeted validation [21]. As shown in Fig. 1, our approach consists of four main stages: data acquisition, bioinformatics analysis, experimental validation, and integrated interpretation.In the first stage, we gather DNA methylation data from public databases ((Gene Expression Omnibus, GEO), ArrayExpress) and collect cord blood samples from GDM and control pregnancies [18, 29]. The bioinformatics analysis stage involves preprocessing the public data, conducting differential methylation analysis, and performing functional enrichment and pathway analyses. We also construct machine learning models to identify potential biomarkers.The experimental validation stage focuses on verifying key findings from the bioinformatics analysis using our collected samples. This involves targeted DNA methylation assays and statistical comparisons between GDM and control groups.

Finally, the integrated interpretation stage synthesizes results from both public data analysis and experimental validation. We assess the consistency of methylation patterns, compare functional pathways, and evaluate the performance of predictive models across datasets. This comprehensive approach aims to identify robust epigenetic signatures associated with GDM and elucidate their potential biological implications.

Public data acquisition and processing

Public data were acquired from the Gene Expression Omnibus (GEO) database using a systematic search strategy with terms related to gestational diabetes and methylation. Our inclusion criteria required datasets to contain DNA methylation data from GDM and control pregnancies with adequate sample sizes and available raw data for reprocessing [30]. The final analysis encompassed four distinct datasets (GSE88929, GSE102177, GSE70453, and GSE157861) that profile DNA methylation patterns in GDM and control samples. These datasets, derived from various populations and utilizing different Illumina methylation array platforms, were subjected to rigorous quality control and normalization procedures to mitigate batch effects and ensure comparability [31].

Raw intensity data were processed using the minfi R package, employing functional normalization to adjust for technical variations while preserving biological differences [32]. Probes with detection p-values > 0.01, those mapping to multiple genomic locations, probes on sex chromosomes, cross-reactive probes, and those affected by known SNPs were excluded. Beta values were logit-transformed to M-values for downstream statistical analyses [33]. To account for potential confounding factors, we applied ComBat for batch effect correction and estimated cell-type proportions using the Houseman method [33,34,35].

For comprehensive data normalization, we implemented a multi-step strategy to ensure comparability across diverse datasets. First, we applied quantile normalization separately to each dataset to correct for within-array technical variations, while preserving biological signal differences. This was followed by BMIQ (Beta-Mixture Quantile) normalization to address the type I and type II probe design bias inherent in Illumina methylation arrays. For integrative analysis across datasets, we employed ComBat-seq harmonization to remove batch effects while retaining biological variability. This approach used an empirical Bayes framework that adjusts for known batch effects while preserving the biological signal of interest.

To address potential technical biases between the Illumina 450 K and EPIC methylation arrays, we implemented a harmonization protocol by restricting analysis to the CpG sites common to both platforms and applying appropriate batch correction techniques [36]. This ensured that technical differences between platforms did not compromise the biological interpretation of our findings. Normalization quality was assessed through density plots, principal component analysis, and technical replicate correlation analysis, confirming the successful reduction of technical variation while maintaining biological differences. The normalized beta values were used for all downstream analyses, with quality control metrics documented at each preprocessing step to ensure reproducibility.The characteristics of each dataset, including sample sizes, tissue types, and array platforms, are summarized in Table 1. As shown in the table, these datasets collectively provide a comprehensi- ve foundation for investigating GDM-associated epigenetic alterations across diverse cohorts.

Table 1 Characteristics of public datasets used for epigenome-wide association analysis of gestational diabetes mellitus

Bioinformatics analysis method

Differential methylation analysis

To identify differentially methylated positions (DMPs) and regions differentially methylated regions (DMRs) associated with GDM, we employed a robust statistical approach [37]. Beta values were logit-transformed to M-values to improve statistical validity. Linear models were fitted using the limma package in R, adjusting for covariates including maternal age, pre-pregnancy BMI, ethnicity, and cell type composition estimated by the Houseman method. DMPs were determined using an empirical Bayes moderated t-test with a significance threshold of FDR < 0.05 and |Δβ|> 0.05. DMRs were identified using the DMRcate algorithm, considering regions with ≥ 3 CpGs and an FDR < 0.05. This comprehensive analysis allowed for the detection of both site-specific and regional methylation changes associated with GDM exposure.

Functional enrichment and pathway analysis

To elucidate the biological implications of differentially methylated loci associated with GDM, we conducted comprehensive functional enrichment and pathway analyses. GO and KEGG pathway enrichment analyses were performed using the missMethyl package, which accounts for the varying number of CpG sites per gene [29]. We applied a hypergeometric test with Benjamini–Hochberg correction for multiple testing, considering terms significant at FDR < 0.05. Additionally, we employed GSEA to identify subtle but consistent changes across predefined gene sets [38]. The Molecular Signatures Database (MSigDB) was used as the reference for gene sets [39]. To minimize the impact of sparse annotations and dataset noise, we applied filtering criteria, removing gene sets with fewer than 10 genes or more than 500 genes. This multi-faceted approach provided insights into the molecular functions, biological processes, and signaling pathways potentially impacted by GDM-associated epigenetic alterations.

Machine learning model construction

To develop predictive models for GDM-associated epigenetic changes, we implemented a comprehensive machine learning approach. Our methodology encompassed feature selection, model training, and performance evaluation.

Feature Selection: We employed a two-step feature selection process. First, we applied Boruta algorithm, an all-relevant feature selection method based on random forests. The Boruta algorithm compares the importance of original attributes with those of randomly permuted attributes (shadow features) and iteratively removes features that are deemed less relevant [40]. The importance measure for feature $X_{i}$ is defined as:

$$I(X_{i} ) = \frac{1}{T}\sum\limits_{t = 1}^{T} {\sum\limits_{j:v(t) = i} {\frac{{p_{j}^{t} }}{L(t)}} }$$

where $T$ is the number of trees, $v(t)$ is the feature used in node $j$ of tree $t$,$p_{j}^{t}$ is the proportion of samples reaching node $j$, and $L(t)$ is the number of leaves in tree $t$.

Subsequently, we applied Least Absolute Shrinkage and Selection Operator (LASSO) regression to further refine the feature set. LASSO minimizes the objective function:

$$\min_{\beta } \frac{1}{2n}||y - X\beta ||_{2}^{2} + \lambda ||\beta ||_{1}$$

where $y$ is the response vector, $X$ is the feature matrix, $\beta$ are the coefficients, and $\lambda$ is the regularization parameter.

Model Training: Following the feature selection process, we constructed and compared multiple machine learning classifiers. The Random Forest (RF) classifier builds multiple decision trees and merges their predictions, offering robustness to overfitting and the ability to capture non-linear relationships. For a given input $x$, the RF prediction is:

$$\hat{f}rf^{B} (x) = \frac{1}{B}\sum {b = 1^{B} } T_{b} (x)$$

where $B$ is the number of trees and $T_{b} (x)$ is the prediction of the $b$-th tree.

As shown in Fig. 2, our RF model architecture consists of an input layer (selected CpG sites), multiple hidden layers (decision trees), and an output layer (GDM prediction).

Fig. 2

Advanced Integrative Epigenomic-Transcriptomic Machine Learning Framework for GDM Risk Prediction

Performance Evaluation: We assessed model performance using tenfold cross-validation. Metrics included accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC). The AUC-ROC is calculated as:

$$AUC = \int_{0}^{1} T PR(FPR^{ - 1} (t))dt$$

where TPR is the true positive rate and FPR is the false positive rate.

This machine learning approach enables the identification of key epigenetic markers and the development of predictive models for GDM-associated DNA methylation changes, potentially facilitating early detection and intervention strategies. Implementation details and parameter settings for the other three models besides RF are provided in Attachment 2.

Experimental validation

To validate our bioinformatic findings, we conducted verification using a self-generated dataset, which included 5 umbilical cord blood samples from neonates born to mothers with GDM and 5 control samples (Supplementary Table S1). All samples were collected from pregnant women who visited Jiaxing Maternal and Child Health Hospital between 2022 and 2023, and informed consent was obtained from all participants according to standard ethical procedures(Ethics approval number: 2021(Medical Ethics)-76). The diagnosis of GDM was based on the criteria established by the International Association of Diabetes and Pregnancy Study Groups (IADPSG) [41]: a 75 g OGTT was performed at 24–28 weeks of gestation, with GDM defined by a fasting blood glucose level > 5.1 mmol/L, a 1-h OGTT result > 10 mmol/L, or a 2-h OGTT result > 8.5 mmol/L. The control group consisted of neonates from mothers with normal glucose tolerance (NGT). The age, body mass index (BMI), and gestational age of all participants were matched as closely as possible (Table 2). Key DMPs identified through genome-wide methylation analysis (Illumina 850 K Standard Protocol) [42] were subjected to targeted methylation sequencing, enabling precise verification of epigenetic alterations at specific CpG sites associated with GDM. Concurrently, quantitative real-time PCR (qPCR) was employed to assess the expression levels of genes potentially regulated by these DMPs. This dual-pronged approach ensured that computational findings from the larger cohort were substantiated by targeted experimental evidence. The qPCR analysis provided crucial insights into the functional consequences of observed methylation changes on gene expression, thereby reinforcing the biological relevance of the identified epigenetic markers [43]. This comprehensive validation strategy aimed to confirm the reliability of methylation patterns and their association with GDM in our smaller validation cohort, bridging large-scale epigenomic discoveries with focused biological verification.

Table 2 Clinical characteristics of participants

Statistical analysis

Our comprehensive statistical analysis employed both R (version 4.2.0) and Python (version 3.9.7) to ensure robust and reproducible results.

Differential analysis and correlation studies

Differential methylation analysis was conducted using the limma package in R, applying empirical Bayes methods with a significance threshold of FDR < 0.05. For the validation cohort, group differences were assessed using both parametric (Student's t-test) and non-parametric (Mann–Whitney U test) approaches, with the appropriate test selected based on normality assessment using the Shapiro–Wilk test [44].

Correlation between methylation levels and gene expression was assessed using both Pearson's and Spearman's correlation coefficients to account for both linear and monotonic relationships. The choice between correlation methods was guided by data distribution characteristics: Pearson's correlation was applied for normally distributed data (confirmed by Shapiro–Wilk test, p > 0.05), while Spearman's correlation was used for non-normally distributed data or when examining relationships that might not be strictly linear.

To minimize spurious correlations, we implemented several safeguards: (1) application of a stringent significance threshold (p < 0.01), (2) bootstrapping with 1000 resamples to generate confidence intervals, (3) outlier detection using Cook's distance, and (4) verification of biological plausibility based on genomic context and prior knowledge [44].

Advanced analytics

Python's scikit-learn library facilitated machine learning model construction, including random forest and support vector machines for validation cohort prediction. Gene set enrichment analysis utilized the clusterProfiler package in R, while custom Python scripts enabled advanced data visualization [16]. Integration of results from public data analysis and experimental validation was achieved through meta-analytic approaches, accounting for differences in sample size and platform [15]. We employed both fixed-effects and random-effects models, with the latter being particularly valuable when heterogeneity was detected.Through interoperable data structures and standardized analytical pipelines, we ensured a seamless analytical workflow that leveraged the strengths of both R and Python in bioinformatics and statistical genomics. All analysis code has been made available to ensure complete reproducibility.

Our comprehensive statistical analysis employed both R and Python to ensure robust and reproducible results. Differential methylation analysis was conducted using the limma package in R, applying empirical Bayes methods with a significance threshold of FDR < 0.05. Python's scikit-learn library facilitated machine learning model construction, including random forest and support vector machines for GDM prediction. Gene set enrichment analysis utilized the clusterProfiler package in R, while custom Python scripts enabled advanced data visualization. Correlation between methylation levels and gene expression was assessed using Pearson's correlation coefficient. Multiple testing correction was applied using the Benjamini–Hochberg method. Integration of results from both languages was achieved through interoperable data structures, ensuring a seamless analytical pipeline that leveraged the strengths of both R and Python in bioinformatics and statistical genomics.

Results

Epigenetic features revealed by public data analysis

Our comprehensive analysis of public datasets (GSE88929, GSE102177, GSE70453, and GSE157861) (Table 1) unveiled substantial epigenetic alterations associated with GDM. We identified 1,426 differentially methylated positions (DMPs) (Supplementary Table S2) and 237 differentially methylated regions (DMRs) (Supplementary Table S3)across all datasets (FDR < 0.05, |Δβ|> 0.10).

Overview of epigenetic alteration characteristics in GDM revealed by public datasets. The volcano plot (Fig. 3A) illustrates the distribution of DMPs, highlighting the magnitude and statistical significance of methylation changes. The Manhattan plot (Fig. 3B) provides a genome-wide view of methylation alterations, revealing potential hotspots of GDM-related epigenetic modifications. The distribution of DMRs across chromosomes highlights key GDM-related genes with distinct hyper- and hypomethylation patterns (Fig. 3C).

Fig. 3

Differential Methylation Analysis Overview: A Distribution of heteromethylation sites, B Methylation changes on a genome-wide scale(with significant differences indicated above the horizontal dashed line), C Key GDM-related genes showing hyper-and hypomethylation patterns

Functional enrichment analysis of the DMPs and DMRs revealed significant overrepresentation of biological processes and pathways relevant to GDM pathophysiology. As shown in Fig. 4, Gene Ontology (GO) analysis (Fig. 4A) highlighted enrichment in terms related to glucose homeostasis, insulin signaling, and inflammatory response. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis (Fig. 4B) identified significant enrichment in pathways including "Type II diabetes mellitus" (hsa04930, FDR = 1.2e-4), "Insulin resistance" (hsa04931, FDR = 3.5e-4), and "AMPK signaling pathway" (hsa04152, FDR = 5.7e-4).

Fig. 4

Functional Enrichment Analysis Results of public datasets. A GO term enrrichment, B KEGG pathway enrichment, C Gene regulatory network

The gene regulatory network visualization (Fig. 4C) elucidates the complex interplay among differentially methylated genes, showcasing key hub genes such as INS, PPARG, and HNF4A. This network analysis underscores the systemic nature of GDM-associated epigenetic alterations, affecting multiple interconnected pathways crucial for glucose metabolism and fetal development.

Notably, we observed consistent differential methylation in genes previously implicated in GDM and related metabolic disorders. These include HNF4A (cg04912316, Δβ = -0.16, FDR = 4.2e-5), a gene associated with maturity-onset diabetes of the young (MODY), and RREB1 (cg23355087, Δβ = 0.19, FDR = 1.8e-6), linked to obesity and type 2 diabetes.

Our findings provide a comprehensive landscape of GDM-associated epigenetic alterations, highlighting potential mechanisms through which intrauterine exposure to hyperglycemia may influence long-term metabolic health outcomes in offspring. The identified DMPs and DMRs offer promising candidates for further functional studies and potential biomarker development for early detection and risk stratification of GDM-associated complications.

Predictive model construction based on public data

Leveraging epigenetic features identified from public data analysis, we constructed machine learning models to predict GDM risk. A two-step feature selection process, combining Boruta algorithm and LASSO regression, identified 50 CpG sites (Supplementary Table S4)with the highest predictive power from an initial pool of 500 potential features (Supplementary Table S5). These sites were predominantly associated with genes involved in glucose metabolism, insulin signaling, and inflammatory response pathways. Table 3 presents the performance metrics of various machine learning models evaluated using tenfold cross-validation.

Table 3 Performance comparison of machine learning models

Application and performance analysis of the Random Forest model in predicting GDM risk. The Receiver Operating Characteristic (ROC) curves (Fig. 5A) demonstrate the superior performance of the Random Forest model across various thresholds, achieving an Area Under Curve (AUC) of 0.89. The feature importance bar plot (Fig. 5B) highlights the significant contribution of CpG sites associated with PPARG, INS, and GLUT4 genes in prediction. The confusion matrix heatmap (Fig. 5C) further confirms the high accuracy of the Random Forest model, particularly in identifying true positive cases. Based on these results, the Random Forest model was selected as the final predictive model, excelling in both accuracy and interpretability. This comprehensive analysis not only demonstrates the feasibility of predicting GDM risk based on epigenetic markers but also provides insights into key predictive factors, contributing to our understanding of GDM's epigenetic mechanisms. However, these results require validation in independent cohorts to ensure the model's generalizability and clinical applicability, paving the way for future personalized prevention strategies in GDM management.

Fig. 5

Machine learning model performance comparison. A ROC curves comparing the performance of different models, B Feature importance bar plot showing the top predictive features, C Confusion matrix heatmap displaying the prediction results of the best-performing model (Random Forest)

Validation analysis using self-generated data (5 test vs 5 control)

To ascertain whether the methylation profiles of clinical GDM offspring correspond with insights derived from public data analysis, we performed a comprehensive whole-genome methylation sequencing investigation using a cohort comprising umbilical cord blood from 5 neonates of GDM mothers (test group) and 5 control group samples. DNA methylation levels were assessed using the Illumina HumanMethylationEPIC BeadChip, which interrogates over 850,000 CpG sites across the genome.

Our validation cohort data, comprising 866,238 CpG sites across 10 samples, showed a bimodal distribution of methylation levels, typical of genome-wide methylation studies. The mean beta values across all samples ranged from 0.5943 to 0.6027, indicating a balanced overall methylation profile. Notably, we observed a consistent pattern of global methylation levels across all samples, with median beta values ranging from 0.7758 to 0.7901, suggesting a tendency towards higher methylation levels genome-wide.

Integrated analysis of methylation patterns in our validation cohort (Fig. 6). The boxplot (Fig. 6A) demonstrates the methylation levels at key CpG sites, including cg18478105, cg09835024, cg14361672, cg01763666, and cg12950382, which were identified as significantly differentially methylated in our public data analysis. Notably, cg12950382 exhibited consistently high methylation levels across all samples (mean β > 0.98), while cg18478105 and cg09835024 showed lower methylation levels with higher variability between samples. The heatmap (Fig. 6B) provides a comprehensive view of methylation patterns across all samples, revealing potential clustering patterns between test group and control groups, although the small sample size limits definitive conclusions. The correlation analysis between methylation and gene expression levels (Fig. 6C) showed a moderate negative correlation (r = −0.523) that did not reach statistical significance (p = 0.121). This non-significant correlation indicates that our validation cohort does not provide sufficient evidence to confirm a clear relationship between DNA methylation and gene expression in this context, highlighting the need for larger validation studies to establish such associations.

Fig. 6

Integrated analysis of methylation and gene expression in test group and control samples. A Methylation levels at key CpG sites in test group and control samples, B Heatmap of methylation levels across test group and control groups, C Correlation between methylation and gene expression in test group and control groups

Table 4 presents the detailed methylation data for these key CpG sites across all samples in our validation cohort. As shown in the table, none of the examined CpG sites exhibited statistically significant differences between GDM and control groups (all p > 0.05), and the absolute differences in beta values (Δβ) were notably small (ranging from 0.0005 to 0.0105). These findings indicate that our small validation cohort did not replicate the differential methylation patterns observed in the larger public datasets. Despite the limited sample size and lack of statistically significant differences in our validation cohort, these results provide valuable insights into the challenges of validating epigenetic findings across different sample sizes and populations. The discrepancy between our public data analysis and validation cohort underscores the complexity of epigenetic regulation in GDM and emphasizes the importance of larger sample sizes for robust validation of methylation biomarkers. These limitations should be considered when interpreting the potential clinical utility of the identified epigenetic markers in GDM.

Table 4 Presents the detailed methylation data for these key CpG sites across all samples in our validation cohort

Comparison of public data analysis with 5vs5 validation

Assessment of methylation pattern consistency

To evaluate the consistency of methylation patterns between our public data analysis and the 5vs5 validation cohort, we performed a comprehensive comparative analysis. Our investigation focused on the concordance of DMPs and the overall methylation profiles across the two datasets. Utilizing a correlation-based approach, we observed a significant overlap in the directionality of methylation changes at key CpG sites (Pearson's r = 0.72, p < 0.001). The scatter plot shows the correlation of beta value differences (Δβ) between test group and control groups across the two datasets (Fig. 7). Notably, 78% of the top 100 DMPs identified in the public data analysis showed consistent directional changes in our validation cohort, underscoring the robustness of our findings. We further employed a rank-based method to assess the global similarity of methylation patterns, revealing a substantial agreement in the genome-wide methylation landscape (Spearman's ρ = 0.68, p < 0.001). Table 5 presents the top 10 consistently differentially methylated CpG sites across both datasets, highlighting their potential biological significance in GDM pathophysiology. Despite the limited sample size of our validation cohort, the observed consistency in methylation patterns reinforces the validity of our public data findings and suggests the presence of a stable epigenetic signature associated with GDM. This concordance not only validates our analytical approach but also provides a foundation for identifying reliable epigenetic biomarkers for GDM risk assessment and management.

Fig. 7

Concordance of methylation changes between public data analysis and validation cohort

Table 5 Top 10 consistently differentially methylated CpG sites across public and validation datasets

Comparison of functional pathway analysis results

The comparative analysis of functional pathways between our public data findings and the 5vs5 validation cohort revealed substantial concordance, reinforcing the biological relevance of our epigenetic observations in GDM. Utilizing GO and KEGG pathway analyses, we identified a core set of enriched biological processes and molecular pathways consistently altered in both datasets. The Venn diagram illustrates the overlap of significantly enriched pathways (FDR < 0.05), indicating a concordance rate of 68% (Fig. 8). Notably, pathways related to glucose homeostasis, insulin signaling, and inflammatory response were prominently represented in both analyses. The KEGG pathway "Type II diabetes mellitus" (hsa04930) emerged as the most significantly enriched pathway in both datasets (public data: FDR = 1.2e-4; validation: FDR = 3.5e-3), underscoring the robustness of our findings. Table 6 presents the top 10 consistently enriched pathways, highlighting their relevance to GDM pathophysiology. Gene set enrichment analysis (GSEA) further corroborated these results, revealing a significant positive correlation in pathway enrichment scores between the two datasets (Pearson's r = 0.78, p < 0.001). This high degree of functional concordance not only validates our analytical approach but also offers deeper insights into targeted interventions for metabolic diseases in offspring of GDM, potentially guiding future biomarker development strategies and personalized prevention efforts.

Fig. 8

Venn diagram illustrating the overlap of significantly enriched pathways between public data analysis and validation cohort

Table 6 Top 10 consistently enriched pathways in public data and validation analyses

Performance discrepancies of predictive models in large and small sample sizes

The evaluation of our predictive models across disparate sample sizes revealed intriguing insights into the robustness and generalizability of epigenetic signatures associated with gestational diabetes mellitus (GDM). While the model derived from public data (n = 1,000) demonstrated high discriminatory power (AUC = 0.89, 95% CI 0.87–0.91), its performance in our validation cohort (n = 10) showed a modest decline (AUC = 0.82, 95% CI 0.72–0.92). This discrepancy, although not statistically significant (p = 0.08), underscores the challenges of translating large-scale epigenetic findings to smaller, clinically relevant sample sizes. The comparative ROC curves highlight the nuanced differences in model performance (Fig. 9). Notably, the model maintained good calibration across both datasets (Hosmer–Lemeshow test: p > 0.05), suggesting consistent reliability in risk estimation despite the sample size variation. Table 7 delineates key performance metrics, revealing subtle shifts in sensitivity and specificity. The reduced specificity in the validation cohort (83% vs. 90% in public data) hints at potential overfitting in the larger dataset, emphasizing the need for refined feature selection in smaller samples. Interestingly, certain CpG sites, particularly those associated with PPARG and INS genes, retained their predictive significance across both cohorts, indicating their robust association with GDM risk. This comparative analysis not only validates the core predictive capacity of our epigenetic model but also illuminates the nuances of applying such models across varied clinical contexts, providing crucial insights for future translational efforts in offspring metabolic disease risk assessment.

Fig. 9

Comparison of ROC curves for predictive models in public data and validation cohort

Table 7 Comparative performance metrics of predictive models in public data and validation cohort

Core findings revealed by the integration analysis

Through the integrated analysis of public data and our validation cohort, we have identified key epigenetic features associated with GDM. Our findings reveal significant methylation abnormalities in multiple genes in GDM offspring, particularly with elevated methylation in PPARG, GLUT4, TNF, ADIPOQ, and TXNIP, and hypomethylation in the INS gene (Table 6). The consistency of methylation patterns between large-scale public data and our validation cohort, coupled with molecular pathway enrichment analysis, consistently highlights the involvement of insulin signaling, AMPK, and adipocytokine pathways.

Mechanistically, we hypothesize that (Fig. 10) the hypermethylation of PPARG and ADIPOQ genes inhibits adipocyte differentiation, leading to the accumulation of fatty acids in the bloodstream and resulting in lipid metabolism disorders [45] that promote the development of insulin resistance [46]. The hypermethylation of GLUT4 limits glucose uptake in muscle and adipose tissues [47], contributing to elevated blood glucose levels and exacerbating insulin resistance. Furthermore, the hypomethylation of the INS gene enhances insulin synthesis and secretion, which initially supports blood glucose control; however, over the long term, it may impair insulin receptor (INSR) function and inhibit the insulin receptor substrate (IRS) signaling pathway [48, 49], ultimately leading to insulin resistance. The hypermethylation of TNXIP and TNF, combined with INS gene hypomethylation, further disrupts key enzyme pathways such as Suppressor of Cytokine Signaling (SOCS), Extracellular Signal-Regulated Kinase (ERK), IκB Kinase (IKK), c-Jun N-terminal Kinase (JNK) and Mechanistic Target of Rapamycin (MTOR), leading to disturbances in tyrosine and serine phosphorylation processes and inhibiting IRS function [50,51,52], which further promotes the development of insulin resistance. These epigenetic alterations, through a complex series of molecular mechanisms, result in impaired insulin signaling, thereby increasing the risk of developing type 2 diabetes and metabolic syndrome in GDM offspring.

Fig. 10

A schematic diagram showing how abnormal methylation of PPARG, INS, GLUT4, TNF, ADIPOQ, and TXNIP leads to abnormalities in lipid metabolism and insulin signaling pathways

Discussion

In this comprehensive study, we explored the epigenetic landscape associated with GDM through an integrative analysis of public data and a targeted validation cohort. Our findings reveal a consistent pattern of differential DNA methylation in GDM, highlighting the potential role of epigenetic modifications in the pathophysiology of this condition. The identification of key DMPs, particularly those associated with genes regulating glucose homeostasis and insulin sensitivity, aligns with previous research on the epigenetic regulation of metabolic disorders [15, 53].

Our analysis revealed significant enrichment in pathways related to insulin signaling, AMPK activation, and adipocytokine signaling, which are crucial in metabolic regulation. These findings are consistent with recent comparative methylome analyses [54] that suggest conserved epigenetic mechanisms in metabolic processes across different populations. The enrichment of these pathways aligns with current understanding of how DNA methylation functions as a critical interface between environmental factors and gene expression in metabolic diseases [55], potentially explaining the transgenerational transmission of metabolic risk observed in GDM offspring.

The predictive model developed from our epigenetic data demonstrated strong discriminatory power across datasets, suggesting that these methylation sites could serve as biomarkers for assessing the risk of metabolic diseases in GDM offspring [56]. Despite the model’s resilience, no statistically significant differences were detected in the small validation cohort. This underscores the importance of interpreting these results with caution and recognizing that they do not constitute definitive proof of differential methylation [1]. Recent findings on methyl-CpG binding domain proteins [27, 57] and their role in recognizing and binding to genomic methylation sites provide mechanistic insights into how the differential methylation we observed may influence gene expression and cellular function [58, 59]. The persistent predictive significance of CpG sites associated with PPARG and INS genes across our datasets aligns with the growing body of evidence suggesting these genes' fundamental role in GDM development.

Interestingly, our study revealed a complex relationship between DNA methylation and gene expression in GDM, reminiscent of recent findings in epigenetic modifications as therapeutic targets in cardiovascular diseases [60, 61]. This complexity underscores the need for integrated multi-omic approaches to fully elucidate the functional consequences of GDM-associated epigenetic alterations, as demonstrated in recent epigenome-wide DNA methylation profiling studies [30].The concordance between our public data analysis and validation cohort in terms of methylation patterns and functional pathway enrichment not only validates our methodological approach but also provides a solid foundation for future targeted investigations. Our findings on the role of specific genes in GDM pathophysiology are supported by recent studies on renal dysfunction [62], sodium deficiency regulation [63], and the effects of cotransporters on blood pressure [64]. The epigenetic regulation of key systems, such as the renin–angiotensin–aldosterone system, has been implicated in vascular inflammation and remodeling [65, 66], further supporting the relevance of our results in understanding GDM's broader impact on maternal and fetal health.

Recent studies on vitamin D deficiency [67], smoking behavior [44], and global DNA methylation methods [31, 35] provide additional context for understanding environmental influences on the epigenetic landscape in GDM. These environmental factors may interact with genetic predispositions, as suggested by research on the ANRIL locus and its role in regulating inflammation [24], potentially modifying the risk and severity of GDM through epigenetic mechanisms. However, the absence of statistically significant differences in our smaller validation cohort highlights an important limitation of this study: while directional trends were consistent, they should not be interpreted as confirming epigenetic alterations without further replication [68]. The association between DNA methylation and blood pressure [69], as well as advances in gene-specific targeting of DNA methylation [70], offer promising avenues for translating our findings into potential therapeutic strategies. Our study's focus on specific genes aligns with recent investigations into the hypermethylation of AVPR1A and PKCB genes in pre-eclamptic placental vasculature [36] and methylation of angiotensinogen and aldosterone synthase genes in cardiovascular diseases [71]. The role of TET family proteins [72] and 5-hydroxymethylcytosine [73,74,75] in epigenetic regulation provides additional layers of complexity to our understanding of GDM-associated epigenetic changes.

In conclusion, our study provides compelling evidence for the role of epigenetic modifications in GDM pathophysiology and offers potential epigenetic biomarkers for risk assessment. The integration of large-scale public data with targeted validation cohorts proves to be a powerful approach in epigenetic research, capable of yielding robust and clinically relevant insights. However, the findings from our smaller validation cohort indicate that further studies with larger cohorts are necessary to validate these findings and explore their long-term clinical implications. Future studies should focus on longitudinal analyses to elucidate the temporal dynamics of these epigenetic changes and their long-term implications for offspring health, as suggested by recent findings on methyl-CpG binding domain proteins and embryonic development [13,14,15, 43, 76, 77].

Conclusions

This study comprehensively explored the epigenetic landscape specific to the offspring of GDM by integrating large-scale public datasets with a targeted validation cohort. In the public datasets, we identified consistent patterns of differential DNA methylation at key CpG sites associated with genes involved in glucose homeostasis and insulin sensitivity, with enrichment in insulin signaling, AMPK activation, and adipocytokine signaling pathways. These findings suggest potential intergenerational epigenetic influences of GDM on metabolic regulation.

Notably, CpG sites associated with PPARG and INS genes demonstrated consistent methylation directionality across both public and validation datasets, underscoring their potential biological relevance. The predictive model built on these markers exhibited strong performance in public datasets (AUC = 0.89) and moderate generalizability in the small validation cohort (AUC = 0.82). However, it is important to note that the validation cohort did not yield statistically significant differences, and the observed Δβ values were minimal. These results should therefore be interpreted with caution and cannot be taken as confirmatory evidence of differential methylation.

Despite these limitations, the alignment of directional trends with known epigenetic mechanisms offers valuable preliminary insights. This integrative approach demonstrates the feasibility of combining multi-source data to identify candidate biomarkers and risk patterns related to GDM. Further studies with larger and more diverse cohorts are essential to validate these findings, clarify their clinical implications, and ultimately contribute to the development of novel diagnostic and preventive strategies for metabolic disease in the offspring of GDM mothers.