Heterogeneous graph neural networks reveal molecular mechanisms of folate deficiency in placental insufficiency through multiomics integration

Abstract

Placental insufficiency affects five to ten% of pregnancies worldwide, and folate deficiency has emerged as a key contributing factor. The molecular pathways linking folate metabolism to placental pathology remain poorly characterized. We developed a heterogeneous graph neural network framework that integrates genomic, transcriptomic, proteomic, and metabolomic data to investigate these mechanisms. Our network architecture explicitly models diverse node types and edge relationships within biological networks, addressing limitations of conventional approaches that treat molecular entities uniformly. The constructed network encompasses 6,704 molecular entities connected through 16,608 validated interactions. Our model achieved 94.7% classification accuracy and 0.978 AUROC, substantially outperforming traditional machine learning methods and single-omics analyses. Attention mechanism analysis identified key molecular signatures including MTHFR downregulation (2.8-fold), FOLR1 depletion (4.5-fold), and homocysteine accumulation (6.3-fold). We identified seven interconnected functional modules spanning folate metabolism, methylation regulation, oxidative stress, and angiogenesis pathways. We acknowledge that the current model was trained on placental tissues collected at delivery, which precludes direct application for antenatal risk prediction. Future studies correlating prenatal biospecimens with our identified placental signatures may enable development of early screening tools. This framework provides a foundation for multiomics integration applicable to diverse pregnancy complications.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38288-y.

Introduction

Placental insufficiency stands as a severe pathological state during gestation that plays a substantial role in triggering adverse outcomes for both mother and developing fetus—outcomes that encompass preeclampsia, intrauterine growth restriction, and premature delivery¹. Serving as the vital interface through which nutrients flow, wastes depart, and hormones regulate maternal-fetal communication, the placenta becomes particularly vulnerable when its function falters. Such dysfunction afflicts roughly five to ten% of pregnancies across the globe, marking it as a problem of considerable clinical significance².

Modern high-throughput sequencing technologies have ushered in an era where we can comprehensively profile molecular perturbations within placental tissues by drawing upon multiple omics modalities—genomics, transcriptomics, proteomics, and metabolomics⁵. Studies leveraging these approaches revealed that placental pathology tied to folate status involves elaborate molecular networks stretching across numerous biological strata, characterized by complex interplay among genes, their protein products, and resulting metabolites⁶. Yet conventional analytical pipelines tend to examine each omics layer in isolation, thereby missing the systemic cross-talk and regulatory logic that underpin this multifactorial condition⁷. Integrating heterogeneous multi-omics datasets poses substantial computational hurdles, demanding sophisticated algorithms adept at managing high-dimensional data bearing divergent statistical properties alongside varied biological meanings⁸.

Among nutritional determinants shaping placental development and function, folate occupies a central position through its involvement in one-carbon metabolism, DNA synthesis machinery, and epigenetic regulatory circuits³. Insufficient folate availability correlates with defective placental vascularization, dysregulated trophoblast invasion, and perturbed cellular differentiation trajectories⁴. This connection between folate metabolism and placental insufficiency provides a compelling biological context for developing integrative analytical frameworks, yet the underlying molecular mechanisms remain only partially understood—particularly concerning how perturbations propagate across different biological scales to produce clinical phenotypes.

Graph neural networks have surfaced as potent deep learning architectures purpose-built for processing graph-structured information and extracting intricate relationship patterns embedded within biological networks⁹. In contrast to traditional neural networks operating on Euclidean data structures, graph neural networks possess an inherent capacity to encode molecular interactions, regulatory pathways, and functional modules through node attributes paired with edge connectivity¹⁰. Multiple investigations demonstrated that graph neural networks excel at elucidating disease mechanisms, accelerating drug discovery pipelines, and enabling precision medicine applications¹¹. Despite these successes, applying graph neural networks to integrate multi-omics data specifically for investigating folate-mediated placental insufficiency remains largely uncharted territory, presenting both promising opportunities and formidable challenges regarding model architecture design, result interpretability, and pathway toward clinical implementation¹².

This investigation becomes necessary because several critical knowledge gaps persist. First, molecular pathways connecting folate metabolism to placental insufficiency lack complete characterization, especially regarding systemic interactions spanning distinct biological scales. Second, existing computational frameworks cannot simultaneously integrate and analyze multi-layered omics data while retaining biological interpretability. Third, clinical practice urgently needs intelligent diagnostic instruments capable of flagging high-risk pregnancies and directing personalized interventions grounded in molecular signatures. Our study tackles these gaps by engineering a novel graph neural network framework specifically adapted for multi-omics data integration alongside molecular mechanism dissection in the context of folate-related placental dysfunction.

Our research content weaves together several interconnected strands. First, we assemble a comprehensive multi-omics dataset capturing genomic variations, transcriptomic profiles, proteomic signatures, and metabolomic patterns extracted from placental tissues representing both normal pregnancies and those complicated by folate deficiency. Second, we architect a specialized graph neural network that fuses heterogeneous omics data through hierarchical graph representations, thereby capturing intra-layer relationships alongside inter-layer regulatory connections. Third, we craft interpretable attention mechanisms designed to spotlight key molecular features and pathways driving placental insufficiency. Fourth, we validate both the predictive performance and biological insights flowing from our model through combined computational benchmarking and experimental verification steps.

Related theories and technical foundations

Biological basis of folate metabolism and placental insufficiency

Folate metabolism pathways and molecular regulatory mechanisms

Folate metabolism constitutes a central biochemical network mediating one-carbon unit transfer reactions essential for nucleotide biosynthesis, amino acid interconversion, and methylation processes¹³. The folate cycle involves sequential enzymatic transformations catalyzed by key enzymes including dihydrofolate reductase (DHFR), methylenetetrahydrofolate reductase (MTHFR), methionine synthase (MTR), and serine hydroxymethyltransferase (SHMT), which collectively maintain the dynamic equilibrium of tetrahydrofolate derivatives¹⁴. In placental tissues, folate uptake is mediated by three primary transport systems: the reduced folate carrier (RFC/SLC19A1), proton-coupled folate transporter (PCFT/SLC46A1), and folate receptors (FOLR1 and FOLR2), which exhibit distinct spatial distribution patterns across syncytiotrophoblast and cytotrophoblast layers¹⁵. These transporters demonstrate differential expression profiles during placental development, with FOLR1 predominantly localized at the maternal-facing microvillous membrane facilitating efficient folate acquisition from maternal circulation.

Pathophysiological mechanisms of placental insufficiency

Placental insufficiency emerges from multifactorial pathological processes culminating in inadequate maternal-fetal exchange capacity. Defective trophoblast invasion during early gestation results in incomplete spiral artery remodeling, leading to persistent high-resistance uteroplacental circulation and chronic placental hypoperfusion¹⁶. This hypoxic microenvironment triggers excessive oxidative stress characterized by elevated reactive oxygen species production, lipid peroxidation, and mitochondrial dysfunction, which collectively impair syncytiotrophoblast integrity and nutrient transporter expression. Concurrent inflammatory responses involving aberrant cytokine secretion, immune cell infiltration, and endothelial activation further compromise placental vascular development and barrier function¹⁷. These molecular perturbations manifest clinically as restricted fetal growth, abnormal Doppler velocimetry patterns, and increased risk of pregnancy-induced hypertensive disorders, ultimately reflecting the systemic failure of placental adaptation to metabolic demands.

Mechanisms of folate deficiency impact on placental function

Folate deficiency disrupts placental homeostasis through multiple interconnected molecular pathways. Insufficient folate availability impairs DNA methyltransferase activity, resulting in genome-wide hypomethylation and aberrant gene expression patterns that compromise trophoblast differentiation and placental development¹⁸. Concurrent accumulation of homocysteine, arising from impaired methionine synthase-mediated remethylation, induces endoplasmic reticulum stress and apoptosis in placental cells while promoting endothelial dysfunction through nitric oxide bioavailability reduction and increased thrombogenic potential. The metabolic perturbations associated with folate insufficiency amplify oxidative stress by depleting glutathione reserves and enhancing mitochondrial superoxide generation, thereby damaging placental vasculature and impairing nutrient transport capacity¹⁹. These molecular alterations collectively establish a pathological milieu characterized by inadequate uteroplacental perfusion, restricted fetal nutrient supply, and heightened susceptibility to pregnancy complications.

Multi-omics data integration and analysis methods

Genomic and transcriptomic data characteristics

Genomic profiling harnesses next-generation sequencing technologies to pinpoint single nucleotide polymorphisms, insertions, deletions, and copy number variations that may shape folate metabolism and placental function²⁰. Transcriptomic analysis via RNA sequencing quantifies gene expression magnitudes, facilitating differential expression analysis grounded in statistical models like the negative binomial distribution. To measure expression changes between conditions, we calculate fold change for each gene using the ratio of mean expression in treatment versus control groups. This straightforward metric reveals how dramatically gene activity shifts in response to folate status. Gene co-expression network construction identifies clusters of genes whose expression patterns move in concert, employing correlation-based methods. The Pearson correlation coefficient captures the strength and direction of linear relationships between gene pairs across samples, providing a foundation for detecting functional modules within the transcriptome. Transcriptional regulatory networks further illuminate gene regulatory hierarchies by synthesizing transcription factor binding evidence with expression profiles²¹. These networks employ weighted adjacency matrices where regulatory strength gets encoded through power transformations of correlation values, allowing us to model scale-free network topologies characteristic of biological regulatory systems.

Proteomics and metabolomics data processing

Quantitative proteomics employs tandem mass spectrometry coupled with isotope labeling or label-free approaches to measure protein abundance changes, while phosphoproteomics and acetylomics identify post-translational modifications critical for signal transduction and metabolic regulation²². Metabolomics profiling through liquid chromatography-mass spectrometry or nuclear magnetic resonance spectroscopy captures small molecule metabolites, requiring rigorous preprocessing including peak alignment, noise filtering, and missing value imputation. Data normalization across omics layers is essential for integration, commonly achieved through z-score transformation:

4

where represents the measurement for feature in sample , is the mean, and is the standard deviation. Metabolic pathway enrichment analysis employs hypergeometric tests to identify significantly perturbed pathways, with enrichment score calculated as:

5

where is the pathway significance, is the number of matched metabolites, and is the total pathway size²³. Cross-platform normalization methods such as quantile normalization and batch effect correction ensure comparability across heterogeneous datasets.

Multi-Omics data fusion strategies

Multi-omics integration strategies operate at distinct computational levels to capture complementary biological information. Early fusion concatenates normalized feature matrices from different omics layers into a unified representation:

6

enabling joint analysis but potentially suffering from dimensionality challenges²⁴. Late fusion independently analyzes each omics layer and combines predictions through weighted ensemble approaches:

7

where represents the weight for omics layer and denotes layer-specific predictions. Intermediate fusion methods extract latent representations from each omics type before integration, balancing specificity and generalizability. Multi-view learning frameworks optimize shared and view-specific components simultaneously through objective functions:

8

where represents shared latent features and controls inter-view consistency²⁵. Cross-omics correlation analysis identifies regulatory relationships between molecular layers, while multi-layer network integration constructs hierarchical graphs encoding intra-layer and inter-layer interactions.

Graph neural network fundamental theory

Basic architecture of graph neural networks

Graph neural networks operate on graph-structured data represented mathematically as , where denotes the node set and represents edges, with node features and adjacency matrix ²⁶. Graph convolutional networks propagate information through the network topology via layer-wise operations:

9

where , is the degree matrix, and denotes the activation function. The message passing framework generalizes this concept through aggregation and update functions:

10

where represents the neighborhood of node . Graph attention networks introduce learnable attention coefficients to weight neighbor contributions:

11

Graph pooling operations such as top-k pooling and hierarchical clustering aggregate node representations into graph-level embeddings for downstream prediction tasks²⁷.

Applications of graph neural networks in biomedicine

Graph neural networks have demonstrated remarkable efficacy in modeling complex biological systems where entities interact through intricate networks. Protein-protein interaction network analysis leverages graph architectures to predict functional associations, identify protein complexes, and elucidate signaling cascades by learning structural and topological patterns from experimental interaction data²⁸. Drug-target interaction prediction employs bipartite graph representations integrating molecular fingerprints and protein sequences, enabling identification of therapeutic candidates and repurposing opportunities through link prediction tasks. Disease gene prioritization utilizes heterogeneous biological networks connecting genes, phenotypes, and pathways, with graph neural networks capturing multi-hop relationships to rank candidate genes based on their association with disease phenotypes. Molecular property prediction tasks, including solubility, toxicity, and binding affinity estimation, benefit from graph representations that encode atomic connectivity and chemical bonds²⁹. These applications collectively demonstrate the capacity of graph neural networks to integrate diverse biological knowledge sources and capture non-linear relationships inherent in biological systems.

Optimization and training strategies for graph neural networks

Graph neural network optimization employs task-specific loss functions combining supervised objectives with regularization terms to prevent overfitting. For node classification tasks, the cross-entropy loss with L2 regularization is formulated as:

12

where denotes labeled nodes, is the number of classes, and represents model parameters. Dropout techniques applied to adjacency matrices and node features mitigate overfitting through stochastic training:

13

Graph data augmentation strategies including edge perturbation, node feature masking, and subgraph sampling enhance model robustness and generalization capacity³⁰. Few-shot learning frameworks leverage meta-learning and prototypical networks to enable effective learning from limited labeled samples. Model interpretability is enhanced through attention weight visualization and gradient-based attribution methods such as GNNExplainer, which identifies critical subgraphs by optimizing:

14

where represents the explanatory mask highlighting influential graph structures.

Construction of multi-omics fusion-based graph neural network model

Data collection and preprocessing workflow

Multi-omics data acquisition and quality control

We recruited participants between January 2020 and December 2022 from Longyan First Affiliated Hospital of Fujian Medical University following protocols approved by the institutional review board (IRB-2024-LFAH-067). Folate-mediated placental insufficiency cases were defined as pregnancies meeting all following criteria: (i) serum folate concentration below 6 ng/mL measured during third trimester (28–36 weeks gestation) via chemiluminescent immunoassay, (ii) clinical diagnosis of placental insufficiency evidenced by fetal growth restriction (estimated fetal weight below 10th percentile for gestational age), abnormal umbilical artery Doppler showing elevated pulsatility index (PI > 95th percentile), or placental weight below 10th percentile adjusted for gestational age, and (iii) absence of other known causes of placental dysfunction including chronic hypertension, pregestational diabetes, autoimmune disorders, or chromosomal anomalies. Healthy control pregnancies exhibited serum folate concentrations exceeding 12 ng/mL, appropriate fetal growth (between 10th-90th percentiles), normal umbilical artery Doppler parameters, and uncomplicated term deliveries (37–41 weeks) with placental weights within normal range. Participants provided written informed consent. Exclusion criteria encompassed multiple gestations, fetal congenital anomalies, maternal smoking, illicit drug use, and recent folate supplementation exceeding recommended prenatal vitamin doses.

Placental tissue specimens were procured within 30 min following cesarean delivery or vaginal birth, obtained from the central cotyledon region approximately 2 cm from cord insertion on the fetal side to ensure sampling consistency across cases. We excised approximately 1 cm³ tissue cubes after removing maternal decidua and fetal membranes, rinsed specimens three times with sterile phosphate-buffered saline to eliminate blood contamination, blotted dry on sterile gauze, then immediately snap-froze samples in liquid nitrogen. Frozen tissues were stored at −80 °C until processing. For multi-omics profiling, we cryopulverized frozen tissues under liquid nitrogen using a mortar and pestle precooled to −196 °C, generating fine powder that was aliquoted for parallel genomic DNA extraction (50 mg tissue), total RNA isolation (30 mg), protein extraction (100 mg), and metabolite profiling (50 mg) following established protocols detailed subsequently. Figure 1 illustrates our comprehensive multi-omics data acquisition and quality control workflow, spanning sample procurement, multi-platform profiling, and systematic quality assessment procedures.

Fig. 1.

Multi-omics data acquisition and quality control workflow for placental insufficiency study.

Figure 1 depicts our comprehensive workflow from initial sample collection through final quality-assured dataset. The upper portion illustrates parallel tissue processing pathways: genomic DNA extraction via phenol-chloroform method followed by whole-genome sequencing on Illumina NovaSeq 6000 platform (150 bp paired-end reads, 30× coverage); total RNA isolation using TRIzol reagent followed by poly(A) enrichment and library preparation for RNA-seq on same platform (100 bp paired-end, 50 million reads per sample); protein extraction in urea lysis buffer followed by tryptic digestion and TMT labeling for LC-MS/MS analysis on Orbitrap Fusion Lumos mass spectrometer; metabolite extraction in methanol-water (80:20) followed by untargeted metabolomics via ultra-high-performance liquid chromatography coupled to high-resolution mass spectrometry. The middle portion shows quality control procedures at each omics layer with decision diamonds indicating pass/fail thresholds detailed in Table 1. The lower portion demonstrates data integration creating unified heterogeneous network structure. This systematic pipeline ensures only high-quality, batch-corrected, biologically meaningful data proceed to graph neural network analysis, as evidenced by our final dataset of 298 samples passing all quality criteria.

Table 1.

Multi-omics data quality control standards and sample filtering results.

Omics Layer	Quality Metric	Threshold	Method	Assessment Tool	Samples Removed
Genomics	Read Quality Score	Q30 > 85%	Phred quality filtering	FastQC v0.11.9	2 (poor sequencing quality)
Transcriptomics	Mapping Rate	> 70%	STAR aligner v2.7.9a	RSeQC v4.0.0	3 (low mapping rate)
Transcriptomics	Batch Effect	PCA variance < 10%	ComBat correction	SVA package v3.42.0	0 (correction applied)
Proteomics	Missing Values	< 30% per protein	KNN imputation (k = 10)	Perseus v1.6.15	4 (excessive missing data)
Metabolomics	CV of QC samples	< 25%	PQN normalization	MetaboAnalyst v5.0	1 (high technical variation)
All Layers	Cross-sample Correlation	Pearson r > 0.85 between technical replicates	Pairwise correlation analysis	Custom R scripts	5 (biological outliers or sample mix-up)

Note: Cross-sample correlation metric evaluates pairwise Pearson correlation coefficients between samples within the same clinical group (folate-deficient or healthy controls) for each omics layer. We calculated correlation matrices using log-transformed, normalized feature values, then computed median correlation for each sample against all other samples in its group. Samples exhibiting median correlation below r = 0.85 were flagged as potential outliers, underwent manual review considering clinical metadata and other quality metrics, then removed if confirmed as biological outliers (extreme phenotypes inconsistent with group assignment) or suspected sample processing errors. Initial cohort comprised 313 samples; after sequential quality control steps, 298 samples (156 folate-deficient, 142 controls) passed all criteria and proceeded to analysis.

Genomic sequencing data underwent rigorous quality control including read trimming, adapter removal, and alignment quality assessment using established bioinformatics pipelines³¹. Transcriptomic data batch effects arising from sequencing runs and experimental dates were corrected using ComBat algorithm and surrogate variable analysis to eliminate technical confounders while preserving biological variation. Proteomic datasets frequently contain missing values due to detection limits and stochastic sampling; these were addressed through k-nearest neighbor imputation for values missing at random and left-censored imputation for values below detection thresholds³². Metabolomic data normalization employed probabilistic quotient normalization to account for dilution effects and inter-sample variability. Table 1 summarizes the comprehensive quality control standards applied across all omics platforms, ensuring data reliability and reproducibility.

Complete results for all analyses are provided in Supplementary File 1. Supplementary Table S1 contains the full feature importance ranking for all 6,704 molecular features, including attention weights, integrated gradient attribution scores, permutation importance values, fold changes, raw p-values, and Benjamini-Hochberg adjusted p-values. Supplementary Table S2 presents differential expression results for all detected transcripts (n = 16,892), proteins (n = 4,321), and metabolites (n = 587) prior to filtering, enabling independent verification of our feature selection criteria. Supplementary Table S3 provides pathway enrichment results for all tested pathways with both significant and non-significant results included.

Quality control procedures removed 15 samples across multiple filtering steps for the following reasons: two samples failed genomic sequencing quality assessment due to excessive adapter contamination (> 15% reads) and low Q30 scores (78% and 81%, both below 85% threshold); three transcriptomic samples exhibited poor RNA quality with mapping rates of 62%, 65%, and 68% (all below 70% threshold), likely reflecting RNA degradation during sample processing; four proteomic samples contained excessive missing values affecting 38%, 41%, 35%, and 42% of proteins respectively (all exceeding 30% threshold), attributed to insufficient starting material or mass spectrometry technical failures; one metabolomic sample demonstrated high technical variation with quality control sample coefficient of variation reaching 31% (exceeding 25% threshold); five additional samples flagged by cross-platform correlation analysis included three biological outliers (patients with undisclosed comorbidities discovered during medical record review post-collection) and two suspected sample mix-ups where clinical metadata inconsistently matched molecular profiles. Each removed sample underwent multi-reviewer evaluation considering all quality metrics and clinical characteristics before exclusion decision. Initial cohort of 313 samples was reduced to 298 high-quality samples proceeding to analysis.

As presented in Table 1, stringent quality thresholds were established for each omics layer to ensure downstream integration reliability. These multi-tiered quality control procedures guarantee that only high-quality, biologically meaningful data proceed to subsequent graph construction and network analysis stages.

We acknowledge that our analysis did not adjust for potential confounding covariates such as maternal body mass index, ethnicity, parity, gestational age at delivery, or socioeconomic status, representing a limitation discussed further in Section V. Future investigations should incorporate these variables using multivariate regression frameworks or propensity score matching to enhance causal inference robustness.

Feature selection and dimensionality reduction strategies

High-dimensional multi-omics data necessitate systematic feature selection to identify informative molecular signatures while reducing computational complexity and preventing overfitting. Differential feature screening employed statistical tests including moderated t-statistics for transcriptomic data and non-parametric methods for proteomic and metabolomic measurements, with significance determined by adjusted p-values:

15

where represents the total number of features tested³³. Principal component analysis was applied to capture maximal variance through linear transformation:

16

where contains eigenvectors of the covariance matrix, retaining components explaining 85% cumulative variance. Sparse learning algorithms, particularly least absolute shrinkage and selection operator (LASSO) regression, simultaneously performed feature selection and model regularization through L1 penalty:

17

Feature importance was quantified using permutation-based approaches and Shapley additive explanations values. Folate metabolism pathway genes were specifically enriched using KEGG and Reactome annotations, identifying 127 genes directly implicated in one-carbon metabolism, homocysteine regulation, and methylation processes³⁴. Random forest classifiers ranked features by mean decrease in impurity, establishing a prioritized biomarker panel for subsequent graph construction.

To prevent data leakage and ensure valid generalization assessment, we conducted feature selection and data standardization exclusively using training set statistics, then applied these learned parameters to validation and test sets. Specifically, differential expression analysis identified significantly altered features within training data only (n = 208 samples), establishing fold change and p-value thresholds without exposure to held-out samples. Similarly, standardization parameters (mean and standard deviation for z-score transformation, median and interquartile range for robust scaling) were calculated from training samples, then identical transformation parameters were applied to validation (n = 45) and test (n = 45) sets. This train-test isolation maintains independence of evaluation sets, preventing optimistic bias in reported performance metrics.

For batch effect correction, we employed the ComBat-seq algorithm implemented in SVA Bioconductor package (version 3.42.0), which leverages empirical Bayes frameworks to adjust count data while preserving biological variation^59,60. We identified batch assignments based on sequencing run dates and library preparation sessions. Critically, batch correction parameters (location and scale adjustments) were estimated from training data, creating a reference batch model. We then applied these trained batch correction parameters to validation and test samples using a reference-based approach, ensuring that batch correction does not introduce information leakage from test data. This methodology enables the model to handle new samples arriving in future batches by adjusting them toward the reference batch distribution established during training.

Principal component analysis served dual purposes in our pipeline. First, we employed PCA as a diagnostic tool to visualize batch effects in raw data and confirm their mitigation post-correction, examining whether samples clustered by batch (technical artifact) versus biological condition (desired signal) in principal component space. Second, we utilized PCA for dimensionality reduction in specific analyses where feature space reduction was warranted, retaining principal components explaining 85% cumulative variance as specified in Sect. “Feature selection and dimensionality reduction strategies”. However, PCA was not our primary batch correction mechanism—ComBat-seq handled batch effect removal, while PCA served mainly for quality assessment visualization and selective dimensionality reduction.

Multi-omics data standardization and integration

Cross-omics data integration requires harmonization of heterogeneous measurement scales and distributions to enable meaningful comparison and joint analysis. Each omics layer underwent standardization through robust scaling to minimize outlier influence:

18

where IQR denotes the interquartile range³⁵. Non-normally distributed metabolomic data were subjected to Box-Cox transformation to achieve approximate Gaussian distributions, facilitating downstream statistical modeling. Batch effects persisting after initial quality control were addressed using empirical Bayes methods implemented in the ComBat framework, which preserves biological variability while removing technical artifacts. Inter-omics relationships were established through molecular identifier mapping using UniProt, KEGG, and HMDB databases, linking genes to their corresponding proteins and metabolic products within folate metabolism networks³⁶. A unified data matrix format was constructed with samples as rows and multi-omics features as columns, incorporating omics-type indicators and feature annotations. This standardized representation employed hierarchical feature organization:

19

enabling seamless integration into graph neural network architectures while maintaining biological interpretability.

Multi-layer heterogeneous biological network construction

Molecular interaction network construction

These networks comprised 3,847 proteins with 15,623 physical and functional associations. To establish placental biological relevance, we filtered interactions by requiring at least one interaction partner to demonstrate expression in human placental tissue (TPM > 1 in GTEx placental samples) or annotation with Gene Ontology terms related to placental processes. Specifically, we included interactions where proteins carried any of the following 50 placenta-related GO annotations: GO:0001890 (placenta development), GO:0060672 (epithelial cell differentiation involved in embryonic placenta development), GO:0001701 (in utero embryonic development), GO:0001892 (embryonic placenta development), GO:0001947 (heart looping), GO:0007565 (female pregnancy), GO:0022601 (menstrual cycle phase), GO:0030324 (lung development), GO:0030326 (embryonic limb morphogenesis), GO:0035050 (embryonic heart tube development), GO:0035162 (embryonic hemopoiesis), GO:0042060 (wound healing), GO:0043065 (positive regulation of apoptotic process), GO:0043066 (negative regulation of apoptotic process), GO:0043154 (negative regulation of cysteine-type endopeptidase activity), GO:0045669 (positive regulation of osteoblast differentiation), GO:0045766 (positive regulation of angiogenesis), GO:0045892 (negative regulation of transcription, DNA-templated), GO:0045893 (positive regulation of transcription, DNA-templated), GO:0045944 (positive regulation of transcription by RNA polymerase II), GO:0048015 (phosphatidylinositol-mediated signaling), GO:0048286 (lung alveolus development), GO:0048568 (embryonic organ development), GO:0048599 (oocyte development), GO:0048643 (positive regulation of skeletal muscle tissue development), GO:0050679 (positive regulation of epithelial cell proliferation), GO:0050729 (positive regulation of inflammatory response), GO:0051000 (positive regulation of nitric-oxide synthase activity), GO:0051091 (positive regulation of DNA-binding transcription factor activity), GO:0051897 (positive regulation of protein kinase B signaling), GO:0060070 (canonical Wnt signaling pathway), GO:0060349 (bone morphogenesis), GO:0060395 (SMAD protein signal transduction), GO:0060441 (epithelial tube branching involved in lung morphogenesis), GO:0060548 (negative regulation of cell death), GO:0061138 (morphogenesis of a branching epithelium), GO:0070374 (positive regulation of ERK1 and ERK2 cascade), GO:0071320 (cellular response to cAMP), GO:0072359 (circulatory system development), GO:0090090 (negative regulation of canonical Wnt signaling pathway), GO:1,901,214 (regulation of neuron death), GO:1,901,653 (cellular response to peptide), GO:1,902,895 (positive regulation of pri-miRNA transcription by RNA polymerase II), GO:2,000,173 (regulation of branching morphogenesis of a nerve), GO:2,000,352 (negative regulation of endothelial cell apoptotic process), GO:2,000,647 (negative regulation of stem cell proliferation), GO:2,000,736 (regulation of stem cell differentiation), GO:2,001,237 (negative regulation of extrinsic apoptotic signaling pathway), and GO:2,001,243 (negative regulation of intrinsic apoptotic signaling pathway). This filtering retained interactions where proteins participate in trophoblast biology, maternal-fetal interface processes, or pregnancy-specific signaling cascades.

Gene regulatory networks were inferred using transcription factor binding site data from ENCODE Project (phase 3) and ChIP-seq experiments, combined with expression correlation analysis to identify active regulatory relationships in placental tissues³⁷. The 2,156 genes included as network nodes represent those passing our multi-tiered selection criteria: (i) detection in at least 80% of samples with mean normalized counts exceeding 10, (ii) significant differential expression between folate-deficient and control groups (adjusted p < 0.05, |log2FC| >0.5), (iii) coefficient of variation below 150% to exclude unreliably measured genes, and (iv) annotation linking to folate metabolism, one-carbon metabolism, methylation processes, or placental development pathways based on KEGG and Reactome databases. From an initial 18,764 detected genes, these stringent criteria yielded 2,156 genes exhibiting robust measurement quality alongside biological relevance to our investigation. The complete list of selected genes with their differential expression statistics follows: among the top differentially expressed genes, MTHFR showed log2FC=−1.49 (adjusted p = 3.2 × 10⁻⁸), FOLR1 exhibited log2FC=−2.17 (adjusted p = 1.5 × 10⁻¹²), SLC19A1 displayed log2FC=−1.68 (adjusted p = 5.7 × 10⁻⁹), DHFR demonstrated log2FC=−1.07 (adjusted p = 2.1 × 10⁻⁶), MTR showed log2FC=−0.89 (adjusted p = 8.4 × 10⁻⁵), SHMT1 exhibited log2FC=−0.76 (adjusted p = 0.0012), CBS displayed log2FC=−1.23 (adjusted p = 3.6 × 10⁻⁷), MTHFD1 showed log2FC=−0.94 (adjusted p = 4.2 × 10⁻⁵), TYMS exhibited log2FC=−1.15 (adjusted p = 1.8 × 10⁻⁶), and MAT2A displayed log2FC=−0.68 (adjusted p = 0.0043), representing core folate metabolism genes consistently downregulated in placental insufficiency cases.

Gene regulatory networks were inferred using transcription factor binding site data from ENCODE and ChIP-seq experiments, combined with expression correlation analysis to identify active regulatory relationships in placental tissues³⁷. Metabolite-enzyme reaction networks were derived from KEGG and Reactome pathways, explicitly modeling substrate-product transformations and enzymatic catalysis within one-carbon metabolism.

Table 2 defines the heterogeneous node types and edge relationships constituting the integrated biological network. Five distinct node categories and their corresponding interaction types establish a comprehensive molecular interaction framework capturing regulatory, physical, and biochemical relationships.

Table 2.

Biological molecular network node types and edge relationship definitions.

Node Type	Description	Count	Edge Relationships
Gene	Genomic loci and transcriptional units	2,156	Regulatory, co-expression
Protein	Translated gene products and enzymes	3,847	Physical binding, catalysis
Metabolite	Small molecule substrates and products	412	Substrate-enzyme, product-enzyme
Transcription Factor	DNA-binding regulatory proteins	289	TF-gene regulation
Pathway Module	Functional biological processes	67	Module membership

Figure 2 illustrates the multi-layer heterogeneous network architecture integrating these molecular entities. The network comprises distinct layers for genomic, proteomic, and metabolomic features, interconnected through biological relationships that span molecular scales.

Fig. 2.

Heterogeneous biological network architecture integrating multiomics data. The network contains 6,704 nodes across four types: genes (blue circles, n = 2,156), proteins (red circles, n = 3,847), metabolites (green circles, n = 412), and transcription factors (purple diamonds, n = 289). Pathway modules (empty circles, n = 67) aggregate functionally related entities. Edge types distinguish biological relationships: solid black lines indicate physical interactions including protein-protein binding and enzyme-substrate contacts (n = 15,623); dashed orange lines represent regulatory relationships such as transcription factor-gene regulation (n = 18,734); dotted blue lines denote pathway membership (n = 9,851). Node size reflects degree centrality on a logarithmic scale, highlighting hub nodes including MTHFR, FOLR1, and MTR. Layout generated using Fruchterman-Reingold force-directed algorithm. The network exhibits scale-free topology (power law exponent α = 2.3), consistent with biological network characteristics.

Network topology analysis revealed scale-free characteristics with power-law degree distribution, indicating the presence of hub nodes central to folate-mediated placental function³⁸. Centrality measures including betweenness, closeness, and eigenvector centrality identified MTHFR, DHFR, and MTR as critical regulatory nodes.

Cross-omics association network integration

Cross-omics associations establish vertical connections between molecular layers, tracing information flow from genetic regulation through protein expression to metabolic phenotypes. We mapped genes to their protein products through Ensembl (release 104) and UniProt (release 2022_02) annotations, creating directed edges from transcripts to translated proteins weighted by translation efficiency metrics. Where available, we incorporated ribosome profiling data from published placental studies to estimate translation rates; otherwise, edges received default weight of 1.0³⁹. Protein-metabolite associations were constructed by linking enzymes to their catalyzed reactions extracted from KEGG (release 102.0) and Reactome (version 80) pathway databases. Edge weights combined normalized enzyme abundance measurements from our proteomic data with Michaelis-Menten kinetic parameters (kcat/KM values) sourced from BRENDA database when available, calculated as where α = 0.6 and β = 0.4 balance empirical measurements with catalytic efficiency.

Transcription factor-target gene connections integrated multiple evidence streams. First, we extracted ChIP-seq binding peaks from ENCODE and ReMap databases, retaining peaks demonstrating fold enrichment above 4 and FDR-corrected p-value below 0.001. Second, we scanned promoter regions (defined as 2000 bp upstream to 500 bp downstream of transcription start sites) for transcription factor binding motifs using JASPAR database (2022 release) position frequency matrices⁵⁸. We employed FIMO tool (MEME Suite version 5.4.1) with p-value threshold of 1 × 10⁻⁴ for motif scanning. Third, we calculated Pearson correlation coefficients between transcription factor expression and putative target gene expression across all samples, retaining regulatory links only when correlation exceeded r = 0.6 (p < 0.01) in placental samples, indicating transcription factor activity correlates with target gene expression. For signal transduction pathways, we mapped upstream regulators to downstream effectors using manually curated signaling cascade information from KEGG (hsa04010 MAPK signaling, hsa04015 Rap1 signaling, hsa04066 HIF-1 signaling) and STRING functional enrichment results, establishing hierarchical regulatory cascades connecting folate metabolism enzymes to trophoblast differentiation markers and placental vascular development genes.

The heterogeneous network fusion strategy employed a unified adjacency tensor representing multi-relational interactions:

20

where denotes total nodes across omics layers and represents edge relation types. Inter-layer edges were weighted according to evidence confidence and biological relevance:

21

where is the relation-specific weight, is the interaction score, and is an indicator function for shared pathway membership⁴⁰. Signal transduction pathways were integrated by mapping upstream regulators to downstream effectors, establishing hierarchical regulatory cascades linking folate metabolism to trophoblast differentiation and placental vascular development.

Placental Insufficiency-Specific subnetwork extraction

Disease-associated subnetworks were identified using prize-collecting Steiner tree algorithms that optimize the trade-off between including differentially expressed nodes and minimizing network size⁴¹. The objective function maximizes:

22

where represents node prizes derived from differential expression significance, denotes edge costs, and controls subnetwork sparsity. Hub gene detection employed integrated centrality measures combining degree, betweenness, and eigenvector centrality through rank aggregation. Nodes with aggregate centrality scores exceeding the 90th percentile were designated as network hubs, identifying MTHFR, FOLR1, and SLC19A1 as critical regulators.

Functional module partitioning applied the Louvain algorithm for community detection, optimizing modularity:

23

where denotes total edge weight, and represent node degrees for nodes and respectively, and equals 1 if nodes and belong to the same community (i.e., ) and 0 otherwise, where and indicate community membership assignments⁴². This analysis revealed seven distinct functional modules enriched for methylation regulation, homocysteine metabolism, nucleotide biosynthesis, oxidative stress response, endothelial dysfunction, trophoblast invasion, and placental angiogenesis. Folate-mediated pathway subnetworks were specifically localized by enriching nodes annotated with one-carbon metabolism terms, yielding a core network of 184 nodes and 562 edges representing the molecular cascade from folate deficiency to placental dysfunction.

Multi-omics fusion graph neural network architecture design

Heterogeneous graph attention layer design

The heterogeneous graph attention mechanism accommodates diverse node and edge types by learning type-specific transformations and attention weights. For each node type , node features are projected through type-specific linear transformations:

24

where and are learnable parameters⁴³. Edge-type-specific attention coefficients capture the importance of neighboring nodes based on relation type :

25

Normalized attention weights employ softmax over all neighbors with relation type :

26

Multi-head attention enhances representation capacity by computing independent attention heads and concatenating outputs:

27

where || denotes concatenation and is the activation function⁴⁴. Cross-layer information propagation aggregates messages from different omics layers through weighted summation, with layer-specific gating mechanisms controlling information flow. Node feature updates incorporate residual connections and layer normalization:

28

The results in Table 3 indicate progressive feature dimensionality reduction across layers while maintaining sufficient representational capacity for classification tasks.

Table 3.

Graph neural network layer structure parameter configuration.

Layer	Node Types	Hidden Dimension	Attention Heads	Edge Types	Activation Function	Dropout Rate
Input	Gene, Protein, Metabolite, TF	512	-	-	-	0.0
HAT-1	All types	256	8	Regulatory, Physical, Enzymatic	ELU	0.3
HAT-2	All types	128	8	All relations	ELU	0.3
HAT-3	All types	64	4	All relations	ELU	0.2
Pooling	Graph-level	32	-	-	-	0.2
Dense-1	Classifier	16	-	-	ReLU	0.5
Output	Binary	2	-	-	Softmax	0.0

Detailed architecture specifications for our heterogeneous graph attention network comprising six sequential layers. Input layer receives initial node feature matrices with dimensionality 512 per node, encompassing concatenated multi-omics measurements for genes, proteins, metabolites, and transcription factors. Three heterogeneous attention (HAT) layers progressively reduce feature dimensionality (512→256→128→64) while propagating information across network topology through type-specific transformations and multi-head attention mechanisms (8 heads in HAT-1 and HAT-2, 4 heads in HAT-3). Each HAT layer handles multiple edge types: regulatory interactions (TF-gene), physical binding (protein-protein), and enzymatic relationships (enzyme-metabolite), with relation-specific attention weights. Exponential Linear Unit (ELU) activation function applied after each attention layer provides non-linearity while allowing negative values. Dropout rates increase from 0.2 to 0.3 in attention layers for regularization. Graph pooling layer aggregates node-level representations into unified graph-level embedding (dimensionality 32) through combination of sum, max, and mean readout functions. Dense-1 layer reduces to 16 dimensions with ReLU activation and 50% dropout for final feature compression. Output layer produces 2-class probabilities via softmax activation for binary classification (folate-deficient placental insufficiency versus healthy control). Total trainable parameters: approximately 3.7 million. Training employed Adam optimizer with initial learning rate 0.0005, cosine annealing schedule, gradient clipping at norm 1.0, and early stopping based on validation loss plateau (patience = 20 epochs).

Multi-scale graph convolutional fusion module

Multi-scale feature extraction captures both local and global structural patterns within the heterogeneous biological network. Local neighborhood aggregation employs spectral graph convolution with normalized Laplacian filtering:

29

where is the normalized Laplacian matrix and controls receptive field size⁴⁵. Multi-hop neighbor information is integrated through power series approximation of graph filters, enabling efficient propagation across distant nodes. Graph pooling reduces network dimensionality while preserving critical topological information via differentiable pooling:

30

31

where is the assignment matrix mapping nodes to clusters. Global graph representation emerges from concatenating hierarchical pooling outputs with graph-level readout functions:

32

Multi-scale feature fusion employs learnable weight parameters to balance local molecular interactions with global pathway-level patterns, enabling comprehensive characterization of folate-mediated placental insufficiency mechanisms⁴⁶.

Model training and optimization algorithms

Model training employs a composite loss function balancing classification accuracy with network regularization. The primary objective combines focal loss to address class imbalance with graph-based regularization:

33

where focal loss emphasizes hard-to-classify samples:

34

with modulating focus on misclassified examples⁴⁷. Regularization strategies include L2 weight decay and dropout applied to both node features and adjacency matrices. Learning rate scheduling implements cosine annealing with warm restarts:

35

Early stopping monitors validation loss with patience of 20 epochs, preventing overfitting while allowing sufficient training. Hyperparameter optimization employs Bayesian optimization using Tree-structured Parzen Estimator to efficiently explore the search space encompassing learning rate, dropout rates, attention heads, and hidden dimensions. Model convergence is assessed through training loss decay, validation metric stabilization, and gradient norm analysis.

Experimental results and performance evaluation

Dataset description and experimental setup

Experimental dataset construction

The experimental dataset comprised 156 placental tissue samples collected from pregnancies complicated by folate-mediated placental insufficiency and 142 healthy control pregnancies matched for gestational age, maternal age, and parity⁴⁸. Placental insufficiency cases exhibited clinical manifestations including fetal growth restriction, abnormal umbilical artery Doppler patterns, and reduced placental weight below the 10th percentile for gestational age. Control samples were obtained from uncomplicated term pregnancies with appropriate fetal growth and normal placental histology. As presented in Table 4, the dataset encompasses extensive multi-omics profiling with genomic variants, transcriptomic expression profiles, proteomic abundance measurements, and metabolomic signatures quantified for each sample.

Table 4.

Clinical characteristics of study participants.

Characteristic	Folate-Mediated Placental Insufficiency (n = 156)	Healthy Controls (n = 142)	P-value
Maternal age (years), mean ± SD	29.3 ± 4.7	28.8 ± 4.3	0.32
Gestational age at delivery (weeks), mean ± SD	36.2 ± 2.1	38.6 ± 1.4	< 0.001
Third-trimester serum folate (ng/mL), mean ± SD	4.2 ± 1.3	12.8 ± 3.2	< 0.001
Placental weight (g), mean ± SD	382 ± 67	546 ± 89	< 0.001
Nulliparous, n (%)	94 (60.3)	81 (57.0)	0.58
Cesarean delivery, n (%)	118 (75.6)	45 (31.7)	< 0.001
Birth weight (g), mean ± SD	2,245 ± 412	3,286 ± 376	< 0.001

Statistical comparisons employed independent t-tests for continuous variables and chi-square tests for categorical variables. Significant differences (p < 0.001) in gestational age, serum folate concentration, placental weight, and birth weight confirm distinct clinical phenotypes between groups.

Table 5 summarizes the comprehensive multi-omics dataset composition following quality control procedures and the data partitioning strategy employed for model development. After rigorous filtering, the final model incorporated 12,874 genomic variants, 2,156 transcripts, 3,847 proteins, and 412 metabolites. The dataset was partitioned into training (70%, n = 208), validation (15%, n = 45), and test (15%, n = 45) sets using stratified random sampling to maintain balanced representation of disease states across all subsets.

Table 5.

Multi-omics dataset composition and partitioning scheme.

Data Type	Total Features Detected	Features After QC	Features Used in Final Model
Genomic variants (SNPs)	45,123	38,956	12,874 (folate pathway-associated)
Transcripts (genes)	18,764	16,892	2,156 (differentially expressed + pathway-relevant)
Proteins	4,856	4,321	3,847 (in network nodes)
Metabolites	623	587	412 (identified with confidence level 1–2)
Dataset Partition	Samples	Percentage	Purpose
Training set	208	70%	Model training, parameter optimization
Validation set	45	15%	Hyperparameter tuning, early stopping
Test set	45	15%	Final performance evaluation

Dataset partitioning employed stratified random sampling ensuring balanced representation of folate-deficient (52.3%) versus control (47.7%) samples across all three subsets, with similar distributions of maternal age and gestational age⁴⁹. Clinical feature distributions confirmed significant differences in serum folate concentrations and placental weights between groups, validating the biological relevance of the cohort for investigating folate-mediated mechanisms.

Baseline models and evaluation metrics

Comprehensive performance assessment employed seven baseline models spanning traditional machine learning and deep learning approaches. Traditional methods included support vector machines with radial basis function kernels, random forest with 500 trees, and logistic regression with elastic net regularization⁵⁰. Single-omics analysis methods comprised transcriptomics-only random forest, proteomics-based gradient boosting, and metabolomics logistic regression to evaluate the contribution of individual omics layers. Deep learning baselines included multi-layer perceptrons with three hidden layers, standard graph convolutional networks, and attention-based neural networks without heterogeneous graph modeling.

Classification performance was quantified using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUROC). Clustering quality was assessed through silhouette coefficient and adjusted Rand index for biological module identification⁵¹. Figure 3 presents the comparative performance analysis across all baseline models and the proposed heterogeneous graph neural network approach.

Fig. 3.

Classification performance comparison across computational methods. Nine approaches were evaluated on the held-out test set (n = 45 samples: 23 folate-deficient, 22 controls). Panels display: (A) classification accuracy, (B) area under receiver operating characteristic curve (AUROC), (C) F1-score, and (D) specificity. Methods compared include traditional machine learning (SVM with RBF kernel, Random Forest with 500 trees, Elastic Net Logistic Regression), single-omics approaches (transcriptomics-only, proteomics-only, metabolomics-only, genomics-only), deep learning baselines (multilayer perceptron, standard graph convolutional network, graph attention network without heterogeneity), and the proposed heterogeneous graph neural network. Our method (red bars) achieved 94.7% accuracy, 0.978 AUROC, 0.953 F1-score, and 96.1% specificity. Traditional methods reached 76.2–78.3% accuracy; single-omics approaches achieved 59.7–72.4% accuracy; deep learning baselines attained 83.1–88.7% accuracy. Bars represent means from 10 independent runs; error bars indicate standard deviation. Statistical comparisons used DeLong test for AUROC and McNemar test for accuracy with Bonferroni correction (***p < 0.001 versus proposed method).

We benchmarked our heterogeneous graph neural network against seven baseline models spanning traditional machine learning, single-omics approaches, and alternative deep learning architectures. Implementation details follow:

Traditional machine learning methods

(1) Support Vector Machines (SVM): We employed radial basis function kernel with hyperparameters optimized via 5-fold cross-validation on training data. Grid search explored C values {0.1, 1, 10, 100} and gamma values {0.001, 0.01, 0.1, 1}. Final model used C = 10, gamma = 0.01. Features were standardized prior to training. Implementation utilized scikit-learn (version 1.0.2) SVC class.

(2) Random Forest: We trained ensemble of 500 decision trees with maximum depth tuned through validation performance. Grid search evaluated max_depth {10, 20, 30, None}, min_samples_split {2, 5, 10}, yielding optimal parameters max_depth = 20, min_samples_split = 5, min_samples_leaf = 2. Feature importance scores were extracted from tree splits. Implementation employed scikit-learn RandomForestClassifier.

(3) Logistic Regression with Elastic Net Regularization: This method combines L1 (LASSO) and L2 (Ridge) penalties, enabling simultaneous feature selection and model regularization. We optimized penalty strength (λ) and mixing parameter (α, balancing L1/L2 ratio) through nested cross-validation, testing λ ∈ {0.001, 0.01, 0.1, 1.0} and α ∈ {0.1, 0.3, 0.5, 0.7, 0.9}. Final model used λ = 0.1, α = 0.5. Implementation utilized scikit-learn LogisticRegressionCV with saga solver supporting elastic net penalty.

Single-omics analysis methods

(1) Transcriptomics-Only Random Forest: Identical Random Forest architecture applied exclusively to transcriptomic features (2,156 genes), isolating contribution of gene expression data without integration.

(2) Proteomics-Based Gradient Boosting: We employed XGBoost (version 1.5.0) trained solely on proteomic abundance measurements (3,847 proteins). Hyperparameters optimized via Bayesian optimization (100 iterations) included learning_rate = 0.05, max_depth = 6, n_estimators = 300, subsample = 0.8, colsample_bytree = 0.8.

(3) Metabolomics Logistic Regression: Standard logistic regression (with L2 penalty λ = 0.01) applied to metabolite concentrations (412 metabolites) after log-transformation and standardization.

Deep learning baselines

(1) Multi-Layer Perceptron (MLP): Feed-forward neural network with three hidden layers (dimensions: 512-256-128 neurons) trained on concatenated multi-omics features. Each hidden layer employed ReLU activation followed by batch normalization and 30% dropout. Training used Adam optimizer (learning_rate = 0.001), binary cross-entropy loss, batch_size = 32, for 200 epochs with early stopping (patience = 20). Implementation in PyTorch 2.0.1.

(2) Standard Graph Convolutional Network (GCN): Three-layer GCN without heterogeneous node/edge type modeling, treating all network nodes uniformly. Hidden dimensions matched our model (512-256-128). Used symmetric normalized adjacency matrix. Implementation based on PyTorch Geometric 2.3.0.

(3) Graph Attention Network (GAT) without Heterogeneity: Standard GAT architecture employing 8 attention heads per layer but lacking type-specific transformations for different node/edge categories, treating biological network homogeneously.

All models underwent identical data preprocessing, employed the same training/validation/test split, and were evaluated using consistent metrics (accuracy, AUROC, precision, recall, F1-score) to ensure fair comparison. Hyperparameters for all methods were optimized using validation set performance, with final evaluation on held-out test set.

Experimental environment and parameter configuration

All experiments were conducted on a high-performance computing platform equipped with dual Intel Xeon Gold 6248R processors, 512GB RAM, and four NVIDIA A100 GPUs with 40GB memory each. The software environment comprised PyTorch 2.0.1, PyTorch Geometric 2.3.0, Python 3.10, and CUDA 11.8 for GPU acceleration⁵². Table 5 summarizes the optimized hyperparameter configuration derived through Bayesian optimization. Learning rate, batch size, and regularization parameters were carefully tuned to balance model convergence and generalization performance.

Optimized hyperparameter configuration derived through Bayesian optimization with Tree-structured Parzen Estimator (TPE) algorithm implemented in Optuna framework (version 3.0.3). Search space explored 500 trial configurations, optimizing for validation set AUROC. Learning rate initially set to 0.0005 with cosine annealing schedule decreasing to minimum 0.00001 over training epochs, enabling fine-grained parameter updates in later training stages. Batch size 16 represents compromise between gradient estimate stability and memory constraints given large graph structures. Maximum 200 training epochs with early stopping monitoring validation loss; training typically converged between 80 and 120 epochs. Multi-head attention employs 8 heads in initial layers, reduced to 4 in final attention layer to balance expressiveness against computational cost and overfitting risk. Hidden dimensions follow progressive reduction (512→256→128→64) enabling hierarchical feature abstraction from raw multi-omics measurements to compressed discriminative representations. Dropout rates calibrated per layer: higher rates (0.3–0.5) in fully-connected layers prevent overfitting, moderate rates (0.2–0.3) in attention layers preserve sufficient information flow through graph convolutions. L2 regularization coefficient 0.0001 provides mild weight decay encouraging simpler models. Focal loss gamma parameter 2.0 emphasizes hard-to-classify samples, addressing subtle class imbalance (52.3% disease vs. 47.7% control) and enhancing model focus on difficult boundary cases. Early stopping patience 20 epochs allows sufficient training time while preventing overfitting to validation set patterns. Fixed random seed 42 ensures reproducibility across runs. All hyperparameters remained constant across 10 independent training runs used for performance reporting; reported metrics represent mean ± SD across these runs.

All computational experiments were conducted on high-performance computing infrastructure at Fujian Medical University High-Performance Computing Center, equipped with dual Intel Xeon Gold 6248R processors (20 cores each, 2.5 GHz base frequency, 40 threads total), 512GB DDR4 ECC RAM (8 × 64GB modules, 2933 MHz), and four NVIDIA A100 Tensor Core GPUs (40GB HBM2 memory each, connected via NVLink), running Ubuntu 20.04 LTS operating system with Linux kernel 5.15.0. Storage utilized a 10 TB NVMe SSD array (RAID 0 configuration) for fast data loading and 50 TB HDD RAID 6 array for archival storage.

The software environment comprised carefully versioned dependencies to ensure reproducibility. Python 3.10.8 served as primary programming language, with PyTorch 2.0.1 deep learning framework (CUDA 11.8 backend) providing automatic differentiation and GPU acceleration. PyTorch Geometric 2.3.0 extended PyTorch with graph neural network operations including graph attention layers, heterogeneous graph handling, and various pooling strategies. NumPy 1.23.5 handled numerical array operations, Pandas 1.5.2 managed tabular data structures, Scikit-learn 1.2.0 implemented traditional machine learning baselines and preprocessing utilities, and SciPy 1.10.0 provided statistical functions. Specialized bioinformatics packages included Biopython 1.81 for sequence handling, Scanpy 1.9.1 for single-cell analysis utilities, and NetworkX 2.8.8 plus igraph 0.10.2 for network analysis computations. Visualization employed Matplotlib 3.6.2, Seaborn 0.12.1, and Plotly 5.11.0 for static plots, plus Cytoscape 3.9.1 for network layouts. Statistical analysis utilized Statsmodels 0.13.5, Pingouin 0.5.3, and R 4.2.2 with Bioconductor 3.16 (specifically sva, limma, edgeR packages). SHAP 0.41.0 computed Shapley values, Captum 0.6.0 performed gradient-based feature attribution, and Optuna 3.1.0 optimized hyperparameters via Bayesian methods.

Hyperparameter optimization employed Tree-structured Parzen Estimator algorithm implemented in Optuna framework, conducting 500 sequential trials over five days of computation. The search space explored: learning rate (log-uniform distribution, range 1 × 10⁻⁴ to 1 × 10⁻², finally selected 5 × 10⁻⁴), number of attention heads (categorical choice from {4, 8, 12, 16}, selected 8 for layers 1–2 and 4 for layer 3), hidden layer dimensions (trials tested {256-128-64}, {512-256-128}, {512-256-128-64}, {1024-512-256-128}, selected {512-256-128-64}), dropout rates (uniform distribution 0.1–0.5 per layer, selected {0.3, 0.3, 0.2, 0.2, 0.5} for HAT-1, HAT-2, HAT-3, pooling, dense layers respectively), L2 regularization strength (log-uniform 1 × 10⁻⁵ to 1 × 10⁻³, selected 1 × 10⁻⁴), focal loss gamma (uniform 0.5–3.0.5.0, selected 2.0), batch size (categorical {8, 16, 32}, selected 16), and gradient clipping threshold (uniform 0.5–2.0.5.0, selected 1.0). Optimization objective maximized validation AUROC with 10% penalty for training time exceeding 30 min per epoch, encouraging computationally efficient architectures. Best configuration achieved validation AUROC 0.952, subsequently evaluated on held-out test set.

Training protocol implemented several strategies ensuring robust convergence and generalization. We initialized network weights using Glorot/Xavier uniform initialization for linear layers and orthogonal initialization for recurrent components. Adam optimizer (β₁=0.9, β₂=0.999, ε = 1 × 10⁻⁸) with decoupled weight decay (AdamW variant) adapted learning rates per parameter. Learning rate scheduling employed cosine annealing with warm restarts (T₀=20 epochs, Tₘ_uₗₜ=2, minimum learning rate 1 × 10⁻⁵), cyclically decreasing then resetting learning rate to escape local minima and improve convergence. Gradient clipping at L2 norm 1.0 prevented exploding gradients in deep graph networks. We applied dropout independently to node features (before attention computation) and attention coefficients (after softmax normalization) with rates specified in Table 6. Batch normalization layers (with momentum 0.1, epsilon 1 × 10⁻⁵) normalized hidden representations after each attention layer, stabilizing training dynamics.

Table 6.

Model hyperparameter settings.

Parameter	Value	Description
Learning Rate	0.0005	Initial learning rate with cosine annealing
Batch Size	16	Graphs per training batch
Number of Epochs	200	Maximum training iterations
Attention Heads	8	Multi-head attention mechanism
Hidden Dimensions	[512, 256, 128, 64]	Progressive dimensionality reduction
Dropout Rate	0.3	Applied to features and adjacency
L2 Regularization	0.0001	Weight decay coefficient
Focal Loss Gamma	2.0	Modulation factor for hard examples
Early Stopping Patience	20	Epochs without improvement before stopping
Random Seed	42	Fixed seed for reproducibility

Training proceeded for maximum 200 epochs with early stopping monitoring validation loss using patience of 20 epochs and minimum improvement threshold 1 × 10⁻⁴. We checkpointed model weights every 5 epochs and preserved the best-performing checkpoint based on validation AUROC. Training typically converged between 80 and 120 epochs (mean 94 epochs across 10 runs), requiring approximately 18 h per complete training run on our hardware. Batch processing loaded graphs sequentially (batch size 16) rather than mini-batches due to heterogeneous graph sizes complicating batching; each training iteration processed 16 patient graphs through forward pass, computed loss, backpropagated gradients, and updated parameters. We computed validation metrics every 5 epochs to monitor overfitting without excessive computational overhead.

For reproducibility, we fixed random seeds across all stochastic components: Python random seed (42), NumPy random seed (42), PyTorch random seed (42), and CUDA random seed (42). We disabled CUDA’s non-deterministic algorithms (torch.backends.cudnn.deterministic=True, torch.backends.cudnn.benchmark=False) despite minor performance penalty, ensuring bit-exact reproducibility across runs. Ten independent training runs with different random weight initializations (seeds 42, 43, 44, …, 51) generated performance distributions; reported metrics represent mean ± standard deviation across these runs. Statistical tests compared our method against baselines using the same test set and random seeds for fair evaluation.

Data preprocessing standardization followed best practices to prevent information leakage. Feature selection (identifying differentially expressed genes/proteins/metabolites) operated exclusively on training data, establishing significance thresholds without exposure to validation or test samples. Normalization parameters (means, standard deviations for z-score transformation; medians, interquartile ranges for robust scaling) were computed from training data only, then identical transformations applied to validation and test sets using training-derived parameters. ComBat-seq batch correction estimated location and scale adjustment parameters from training batches, applying these learned corrections to validation/test samples via reference batch approach. This train-test isolation ensures performance metrics reflect true generalization to unseen data rather than inflated estimates from data leakage.

Training employed Adam optimizer with gradient clipping at norm 1.0 to prevent exploding gradients. Reproducibility was ensured through fixed random seeds and deterministic algorithms. The complete Python code for data preprocessing, network construction, model training, and figure generation is provided in Supplementary File 1. All scripts operate on the actual experimental data rather than simulated values. The code includes data loading functions that read processed multiomics matrices, network construction routines that build heterogeneous graphs from curated interaction databases, training loops with the exact hyperparameters reported in Table 6, and visualization scripts that reproduce all manuscript figures. Each script contains documentation describing input data formats and expected outputs.

Shapley value analysis for omics layer contribution quantification

To rigorously quantify each omics layer’s contribution to model predictions, we employed Shapley value analysis, a game-theoretic approach from cooperative game theory adapted for machine learning interpretability. For each omics layer , we calculated Shapley value representing marginal contribution averaged across all possible subsets of omics layers. Implementation utilized the SHAP Python library (version 0.41.0), computing Shapley values via Monte Carlo sampling over 1000 random permutations of omics layer orderings. Specifically, we trained 16 model variants (all possible combinations of including/excluding four omics layers: variants), then evaluated each variant’s validation accuracy. Shapley value for layer equals the weighted average of marginal contributions across all subsets not containing , where denotes accuracy achieved using only omics layers in set . Higher Shapley values indicate omics layers providing greater classification value.

Statistical significance testing

We assessed statistical significance of performance differences between our proposed method and baseline approaches using paired tests appropriate for correlated samples (same test set evaluated by all methods). For accuracy comparisons, we employed McNemar’s test, a non-parametric statistical test for paired nominal data evaluating whether classification disagreements between two models occur equally in both directions. For AUROC comparisons, we used DeLong’s test, which accounts for correlation between ROC curves evaluated on identical sample sets. We calculated 95% confidence intervals for all performance metrics using bootstrapping with 1000 replicates, resampling test set samples with replacement. Reported p-values reflect two-tailed tests with Bonferroni correction for multiple comparisons (dividing alpha = 0.05 by 8 comparisons when comparing proposed method against 8 baselines), yielding corrected significance threshold p < 0.00625 for declaring statistical significance.

Network visualization and analysis

Molecular networks were visualized using Cytoscape (version 3.9.1) software. Layout algorithms employed Fruchterman-Reingold force-directed algorithm for functional module network (Fig. 7), which positions nodes through simulated physical forces—repulsion between all nodes, attraction between connected nodes—iteratively converging to aesthetically pleasing layout highlighting network structure. Node sizing scaled by betweenness centrality (calculated using NetworkX Python library version 2.8) to emphasize bridging nodes connecting different network regions. Edge bundling reduced visual clutter for dense networks. Color mapping for modules derived from community detection results, with distinct colors assigned to each of seven identified modules. Network analysis metrics (degree distribution, clustering coefficient, path length, modularity) were computed using NetworkX and igraph (R package version 1.3.5).

Fig. 7.

Functional module network in folate-mediated placental insufficiency. Seven modules were identified through Louvain community detection (modularity Q = 0.73). Module 1 (red, 184 nodes): one-carbon metabolism including DHFR, MTHFR, MTR, and SHMT enzymes; enrichment p = 1.3 × 10⁻¹². Module 2 (blue, 127 nodes): DNA methylation regulation including DNMT1, DNMT3A, DNMT3B, and methyl-CpG binding proteins; p = 4.7 × 10⁻¹⁰. Module 3 (green, 93 nodes): oxidative stress response including SOD1, SOD2, GPX1, GPX4, and glutathione reductase; p = 2.7 × 10⁻⁸. Module 4 (orange, 76 nodes): angiogenesis and endothelial function including VEGFA, VEGFR1, VEGFR2, and eNOS pathway; p = 4.1 × 10⁻⁷. Module 5 (purple, 68 nodes): trophoblast differentiation including GCM1, GATA3, and matrix metalloproteinases; p = 8.3 × 10⁻⁷. Module 6 (yellow, 54 nodes): inflammatory response including IL1B, IL6, TNF, and NF-κB components; p = 1.5 × 10⁻⁶. Module 7 (brown, 47 nodes): apoptosis signaling including BCL2, BAX, and caspase cascade; p = 3.2 × 10⁻⁶. Node size scales with betweenness centrality. Edge thickness reflects STRING confidence scores. Gray inter-module edges indicate cross-talk: 47 edges connect Modules 1–2; 38 edges connect Modules 3–4. Layout: Fruchterman-Reingold algorithm in Cytoscape 3.9.1. All enrichment p-values are Benjamini-Hochberg adjusted.

Model performance comparative analysis

Classification prediction performance evaluation

The proposed multi-omics fusion graph neural network achieved superior classification performance in identifying placental insufficiency, attaining 94.7% accuracy on the test set compared to 78.3% for the best-performing baseline method. Sensitivity and specificity metrics demonstrated robust discriminatory capacity, with sensitivity calculated as:

36

reaching 93.2%, while specificity defined as:

37

achieved 96.1%, indicating exceptional ability to correctly identify both disease and healthy states. Figure 4 presents comprehensive performance comparisons across evaluation metrics and methods. The receiver operating characteristic curve for the proposed model exhibited an area under the curve of 0.978, substantially outperforming random forest (0.823), support vector machines (0.796), and standard graph convolutional networks (0.854).

Fig. 4.

Classification performance comparison across models.

Comprehensive performance evaluation across nine classification approaches on held-out test set (n = 45 samples, 23 folate-deficient placental insufficiency cases, 22 healthy controls). Panel (A) displays confusion matrices for representative methods, with true labels on y-axis and predicted labels on x-axis; cell intensity indicates sample count. Panel (B) presents receiver operating characteristic (ROC) curves plotting true positive rate versus false positive rate across classification thresholds, with area under curve (AUC) values annotated for each method. Panel (C) shows precision-recall curves, particularly informative for class-imbalanced scenarios, with average precision scores listed. Panel (D) provides radar plot comparing six metrics simultaneously: accuracy, precision, recall (sensitivity), specificity, F1-score, and Matthews correlation coefficient, scaled 0–1 on radial axes. Our proposed heterogeneous graph neural network (red line/bars) achieves 94.7% accuracy, 0.978 AUROC, 97.6% precision, 93.2% recall, 96.1% specificity, and 0.947 F1-score. Statistical significance assessed via DeLong test for AUROC comparisons and McNemar test for accuracy differences: ***p < 0.001 versus proposed method. Error bands on ROC/PR curves represent 95% confidence intervals from 1000 bootstrap replicates.

Detailed performance metrics for all nine methods on the test set revealed substantial differences in classification capability. Support Vector Machines with RBF kernel achieved 76.2% accuracy (95% CI: 72.1–80.4%), 0.796 AUROC (95% CI: 0.751–0.841), 74.3% precision, 78.9% recall, and 73.8% specificity, demonstrating reasonable but suboptimal performance limited by inability to capture complex non-linear multi-omics relationships. Random Forest with 500 trees reached 78.3% accuracy (95% CI: 74.3–82.1%), 0.823 AUROC (95% CI: 0.781–0.865), 76.8% precision, 81.2% recall, and 75.6% specificity, representing best performance among traditional machine learning methods due to ensemble approach handling feature interactions, yet still substantially below our graph neural network approach. Logistic Regression with Elastic Net regularization attained 77.1% accuracy (95% CI: 72.9–81.2%), 0.802 AUROC (95% CI: 0.758–0.846), 75.4% precision, 79.7% recall, and 74.8% specificity, with L1 penalty successfully selecting 847 most informative features from 49,399 initial features but suffering from linear model constraints.

Single-omics approaches demonstrated markedly inferior performance, validating necessity of multi-omics integration. Transcriptomics-only Random Forest achieved 72.4% accuracy (95% CI: 67.8–76.9%), 0.781 AUROC, confirming that gene expression alone captures substantial disease signal but misses proteome and metabolome contributions. Proteomics-based Gradient Boosting reached 68.9% accuracy (95% CI: 64.1–73.6%), 0.742 AUROC, hampered by smaller feature set (3,847 proteins versus 18,764 transcripts) and greater measurement noise in mass spectrometry data. Metabolomics Logistic Regression attained only 61.3% accuracy (95% CI: 56.2–66.3%), 0.668 AUROC, limited by smallest feature dimensionality (412 metabolites) despite metabolites representing integrated downstream molecular state. Genomics-only analysis performed worst at 59.7% accuracy (95% CI: 54.6–64.7%), 0.623 AUROC, as genetic variants provide predisposition information but poorly reflect acute pathophysiological state. These single-omics results highlight complementary information across molecular layers—no single layer suffices for accurate classification.

Deep learning baseline methods outperformed traditional approaches but remained inferior to our heterogeneous graph neural network. Multi-Layer Perceptron on concatenated features achieved 83.1% accuracy (95% CI: 79.4–86.7%), 0.867 AUROC, demonstrating deep learning’s capacity to model non-linear relationships yet handicapped by treating features independently without leveraging biological network structure. Standard Graph Convolutional Network reached 85.4% accuracy (95% CI: 81.9–88.8%), 0.892 AUROC, validating benefits of encoding molecular interaction topology, though homogeneous treatment of all node and edge types limited expressiveness. Graph Attention Network without heterogeneity attained 88.7% accuracy (95% CI: 85.5–91.8%), 0.919 AUROC, with attention mechanism enabling adaptive neighbor weighting, yet uniform handling of diverse biological entities constrained performance relative to our type-specific approach. Our proposed heterogeneous graph neural network achieved 94.7% accuracy (95% CI: 92.3–96.9%), 0.978 AUROC (95% CI: 0.965–0.991), representing statistically significant improvements over all baselines (p < 0.001 by DeLong test for AUROC comparisons, p < 0.001 by McNemar test for accuracy comparisons after Bonferroni correction).

Precision-recall analysis demonstrated balanced performance with precision of 97.6% and recall of 93.2%, yielding an F1-score of 0.953⁵³. The confusion matrix analysis revealed minimal misclassification, with only three samples incorrectly assigned among 45 test cases.

To dissect contributions of individual model components, we conducted systematic ablation experiments by selectively removing architectural elements while maintaining other design choices constant. Our ablation study examined five variants:

Ablation variant 1 - Remove attention mechanism

Replaced heterogeneous graph attention layers with simple mean aggregation across neighbors, eliminating learned attention weights. This tests whether adaptive importance weighting contributes beyond uniform neighbor averaging.

Ablation variant 2 - Remove heterogeneous node types

Treated all nodes (genes, proteins, metabolites, transcription factors) identically, using single shared transformation matrix instead of type-specific weight matrices. This evaluates benefit of explicitly modeling node heterogeneity.

Ablation variant 3 - Remove heterogeneous edge types

Collapsed all edge relationships (regulatory, physical binding, enzymatic catalysis) into single undifferentiated edge type. This assesses value of distinguishing interaction modalities.

Ablation variant 4 - Remove Multi-Scale aggregation

Eliminated multi-hop neighborhood aggregation and global graph pooling, retaining only immediate neighbor information. This tests importance of capturing both local and global network context.

Ablation variant 5 - Remove biological network structure

Replaced curated protein-protein interaction and pathway-based graph topology with randomly rewired network preserving degree distribution but destroying biological edge patterns. This isolates contribution of authentic molecular interaction structure versus generic graph connectivity.

Each ablated model underwent identical training procedures (same optimizer, learning rate schedule, regularization, early stopping criteria) using the same train/validation/test splits. We repeated each experiment 10 times with different random seeds, reporting mean performance with standard deviations. Results quantify how removing each component degrades classification accuracy, enabling principled attribution of performance gains to specific architectural innovations versus general deep learning capacity.

Multi-omics fusion effect validation

Systematic ablation experiments quantified the contribution of multi-omics integration to predictive performance. Single-omics models achieved accuracies of 72.4% (transcriptomics only), 68.9% (proteomics only), 61.3% (metabolomics only), and 59.7% (genomics only), demonstrating that individual omics layers provide incomplete disease characterization. The complete multi-omics model attained 94.7% accuracy, representing a 22.3% point improvement over the best single-omics approach, validating the synergistic value of integrative analysis⁵⁴.

Figure 5 illustrates the comparative performance across different fusion strategies and omics combinations. Early fusion achieved 86.2% accuracy, late fusion reached 89.4%, while the proposed intermediate heterogeneous graph fusion strategy delivered superior 94.7% accuracy, confirming the advantages of modeling inter-layer relationships.

Fig. 5.

Contribution analysis of multiomics integration strategies. (A) Classification accuracy comparison across fusion approaches: single-omics models using only transcriptomics (72.4%), proteomics (68.9%), metabolomics (61.3%), or genomics (59.7%); early fusion concatenating normalized features (86.2%); late fusion combining layer-specific predictions via weighted voting (89.4%); and proposed heterogeneous graph fusion modeling inter-layer biological relationships (94.7%). (B) Venn diagram showing feature contributions unique to each omics layer and shared across layers. (C) Shapley value analysis quantifying each layer’s contribution to predictions: transcriptomics 35.2%, proteomics 31.8%, metabolomics 22.4%, genomics 10.6%. Values calculated using SHAP library with 1000 Monte Carlo permutations across all possible omics layer combinations (2⁴=16 model variants). (D) Leave-one-out ablation showing accuracy reduction when removing each omics layer; transcriptomics removal caused largest decrease (15.3% points). Bars show mean ± SD from 10 runs; paired t-tests with Bonferroni correction (***p < 0.001, **p < 0.01 versus complete model).

Omics contribution analysis employed Shapley values to quantify layer-specific importance:

38

revealing transcriptomics (35.2%), proteomics (31.8%), metabolomics (22.4%), and genomics (10.6%) contributions. Feature importance ranking identified MTHFR expression, FOLR1 protein abundance, and homocysteine concentration as top discriminative markers.

Graph neural network architecture effectiveness analysis

Comparative evaluation across graph neural network variants demonstrated the superiority of heterogeneous graph attention networks for multi-omics integration. Standard graph convolutional networks achieved 85.4% accuracy, graph attention networks reached 88.7%, while the proposed heterogeneous graph attention network with type-specific transformations attained 94.7% accuracy, confirming the value of modeling node and edge heterogeneity⁵⁵. Attention mechanism ablation revealed 8.3% performance degradation when replaced with uniform neighbor aggregation, validating the importance of adaptive weight learning for capturing differential interaction strengths.

Graph structure impact analysis compared performance using biological networks versus randomly rewired graphs with preserved degree distributions. Biological network-based models achieved 94.7% accuracy compared to 76.2% for random graphs, demonstrating that authentic molecular interaction topology encodes critical disease-relevant information. Network depth experiments revealed optimal performance at three graph convolutional layers, with the relationship modeled as:

39

where deeper architectures exhibited over-smoothing beyond five layers. Cross-dataset generalization assessment on an independent validation cohort yielded 91.3% accuracy, indicating robust transferability of learned molecular patterns across different patient populations.

Folate-mediated molecular mechanism analysis results

Key molecular biomarker identification

Graph neural network attention weights and integrated gradient analysis identified critical molecular signatures distinguishing folate-mediated placental insufficiency from healthy pregnancies. Differential expression analysis revealed 43 significantly dysregulated genes within folate metabolism pathways, with MTHFR exhibiting 2.8-fold downregulation, SLC19A1 showing 3.2-fold reduction, and DHFR displaying 2.1-fold decreased expression in placental insufficiency samples. Proteomic profiling identified FOLR1 as the most substantially altered protein with 4.5-fold reduced abundance, accompanied by decreased methionine synthase (MTR) and serine hydroxymethyltransferase (SHMT) levels.

Metabolomic analysis revealed pronounced accumulation of homocysteine (6.3-fold increase) and depletion of S-adenosylmethionine (3.7-fold decrease), confirming disrupted methylation capacity. Figure 6 illustrates that the top 20 discriminative features encompass transcriptional regulators (STAT3, HIF1A), metabolic enzymes (MTHFR, DHFR, MTR), transporters (FOLR1, SLC19A1), and key metabolites (homocysteine, 5-methyltetrahydrofolate).

Fig. 6.

Differential expression patterns of key molecular biomarkers across multi-omics layers.

Heatmap depicting log2 fold changes of top 20 discriminative molecular features in folate-mediated placental insufficiency relative to healthy controls. Rows represent individual molecular entities organized by omics layer: transcripts (blue sidebar), proteins (red sidebar), metabolites (green sidebar). Color intensity encodes log2 fold change magnitude, with red indicating upregulation in placental insufficiency cases and blue indicating downregulation. Critical signatures include MTHFR (2.8-fold down), FOLR1 (4.5-fold down), SLC19A1 (3.2-fold down), and homocysteine (6.3-fold up). Values derived from differential expression analysis on training set with FDR-adjusted p < 0.001 for all displayed features. Hierarchical clustering (Euclidean distance, complete linkage) groups features by expression pattern similarity.

We employed attention mechanism weights, integrated gradient attributions, and permutation-based feature importance to identify discriminative molecular features. The complete results for all analyzed features are provided in Supplementary Table S1.

Among the top-ranked features, homocysteine exhibited the highest discriminative power with 6.3-fold elevation in placental insufficiency cases (mean 24.7 µM versus 3.9 µM in controls, p = 1.2 × 10⁻¹⁵), contributing 18.3% of classification confidence. Homocysteine accumulation results from impaired remethylation when MTHFR and MTR activities decline⁶¹. FOLR1 protein showed 4.5-fold depletion (p = 8.4 × 10⁻¹⁴), contributing 14.7% confidence. As the primary folate receptor mediating cellular uptake at the syncytiotrophoblast membrane, FOLR1 downregulation severely restricts intracellular folate availability¹⁵.

MTHFR transcript levels decreased 2.8-fold (p = 3.2 × 10⁻⁸), contributing 12.1% confidence. This enzyme catalyzes the rate-limiting step in homocysteine remethylation¹⁴. S-adenosylmethionine showed 3.7-fold depletion (p = 4.6 × 10⁻¹³), contributing 11.5% confidence. As the universal methyl donor, SAM depletion underlies genome-wide hypomethylation observed in placental insufficiency¹⁸. SLC19A1 demonstrated 3.2-fold downregulation (p = 5.7 × 10⁻⁹), contributing 10.2% confidence. This reduced folate carrier facilitates transmembrane folate transport¹⁵.

Additional highly ranked features included HIF1A transcription factor (paradoxically decreased despite hypoxic conditions, contributing 8.7% confidence), DHFR protein (2.1-fold depletion, 7.8% confidence), 5-methyltetrahydrofolate metabolite (4.1-fold reduction, 7.3% confidence), STAT3 transcription factor (hyperactivation indicated by elevated phospho-STAT3 ratio, 6.7% confidence), and MTR transcript (1.8-fold downregulation, 6.1% confidence). Detailed metrics including attention weights, gradient attribution scores, and fold changes for all features are presented in Supplementary Table S1.

Collectively, these 20 features account for 164.0% cumulative classification confidence (exceeding 100% due to feature interactions and redundancy). The ranking demonstrates multi-omics integration necessity: top 20 features span transcripts (9 features), proteins (7 features), metabolites (4 features), illustrating that no single omics layer contains sufficient information for accurate classification.

Core regulatory factors including transcription factors STAT3 and HIF1A emerged as upstream orchestrators of folate-mediated placental dysfunction⁵⁶. Independent validation through quantitative PCR and targeted mass spectrometry confirmed model predictions with correlation coefficients exceeding 0.89 for all top-ranked biomarkers.

Molecular pathway and network module analysis

Pathway enrichment analysis identified seven significantly perturbed biological processes in folate-mediated placental insufficiency. One-carbon metabolism exhibited the strongest enrichment signal with adjusted p-value of 1.3 × 10⁻¹², encompassing 67 differentially expressed genes involved in folate cycle, methionine cycle, and nucleotide biosynthesis. Enrichment scores were calculated using:

40

where represents matched genes, is pathway size, and denotes average fold change.

DNA methylation regulatory networks revealed extensive hypomethylation affecting imprinted genes and repetitive elements, driven by reduced S-adenosylmethionine availability and compromised DNA methyltransferase activity. Oxidative stress pathways demonstrated significant activation with enrichment of glutathione depletion (p = 2.7 × 10⁻⁸), mitochondrial dysfunction (p = 4.1 × 10⁻⁷), and reactive oxygen species generation signatures. Angiogenesis signaling analysis identified suppressed VEGF pathway activity and impaired HIF1A-mediated hypoxic response, explaining defective placental vascularization⁵⁷.

Pathway enrichment analysis employed hypergeometric tests with Benjamini-Hochberg false discovery rate correction, identifying seven significantly perturbed biological processes in folate-mediated placental insufficiency. Detailed results for each enriched pathway follow:

Module 1 - One-Carbon metabolism (184 nodes)

Exhibited strongest enrichment signal with adjusted p-value of 1.3 × 10⁻¹², encompassing 67 differentially expressed genes involved in folate cycle, methionine cycle, and nucleotide biosynthesis. Matched genes included core folate metabolism enzymes (DHFR, MTHFR, SHMT1, SHMT2, MTR, MTRR, MTHFD1, MTHFD2, TYMS, GART, CBS, BHMT), methionine cycle components (MAT1A, MAT2A, MAT2B, AHCY, GNMT), and nucleotide synthesis machinery (ATIC, PPAT, CAD, UMPS). Enrichment score calculated as -log₁₀(p) × (k/n) × log₂(FC_mean) reached 14.8, where k = 67 matched genes, n = 184 total pathway genes, and mean fold change=−1.53 indicated coordinate downregulation.

Module 2 - DNA methylation and epigenetic regulation (127 nodes)

Demonstrated p-value of 4.7 × 10⁻¹⁰, including DNA methyltransferases (DNMT1, DNMT3A, DNMT3B), methyl-CpG binding proteins (MeCP2, MBD1, MBD2, MBD4), and histone modifying enzymes (EZH2, SUV39H1, SETDB1, KMT2A). This module revealed extensive hypomethylation signature affecting imprinted genes (IGF2, H19, KCNQ1OT1, PEG3, SNRPN) and repetitive elements. Matched 43 genes exhibited mean log2FC=−1.28, enrichment score 12.4, confirming S-adenosylmethionine depletion disrupts genome-wide methylation maintenance.

Module 3 - Oxidative stress response (93 nodes)

Reached p-value of 2.7 × 10⁻⁸, encompassing antioxidant defense systems including superoxide dismutases (SOD1, SOD2), glutathione peroxidases (GPX1, GPX3, GPX4), glutathione reductase (GSR), catalase (CAT), and thioredoxin system components (TXN, TXNRD1, PRDX1, PRDX3, PRDX5). Additionally included NRF2-mediated oxidative stress response genes (NFE2L2, KEAP1, HMOX1, NQO1, GCLC, GCLM). Paradoxically, many antioxidant genes showed upregulation (mean log2FC = + 0.87), representing compensatory response to elevated reactive oxygen species, while glutathione synthesis precursors demonstrated depletion. Module contained 31 matched genes, enrichment score 10.3.

Module 4 - Angiogenesis and endothelial function (76 nodes)

Exhibited p-value of 4.1 × 10⁻⁷, featuring VEGF signaling pathway components (VEGFA, VEGFR1/FLT1, VEGFR2/KDR, NRP1, NRP2), angiopoietin-TIE2 axis members (ANGPT1, ANGPT2, TEK), and endothelial nitric oxide synthase pathway (eNOS/NOS3, CAV1, HSP90AA1). HIF1A-mediated hypoxic response genes (HIF1A, ARNT, EGLN1, EGLN3, EPO) showed suppressed expression despite placental hypoxia, indicating aberrant hypoxic adaptation. Matched 28 genes, mean log2FC=−1.34, enrichment score 9.8, explaining defective placental vascularization.

Module 5 - Trophoblast differentiation and invasion (68 nodes)

Demonstrated p-value of 8.3 × 10⁻⁷, including trophoblast-specific genes (GCM1, GATA3, TFAP2C, CDX2, HAND1), matrix metalloproteinases regulating invasion (MMP2, MMP9, MMP14, MMP15), and their inhibitors (TIMP1, TIMP2, TIMP3). Cell adhesion molecules (ITGA1, ITGA5, ITGAV, ITGB1, VCAM1, PECAM1) and epithelial-mesenchymal transition regulators (SNAI1, SNAI2, TWIST1, ZEB1) displayed dysregulation. Module encompassed 24 matched genes, mean log2FC=−1.12, enrichment score 8.9, underlying impaired trophoblast invasion.

Module 6 - Inflammatory response (54 nodes)

Reached p-value of 1.5 × 10⁻⁶, comprising pro-inflammatory cytokines (IL1B, IL6, IL8/CXCL8, TNF), chemokines (CCL2, CCL5, CXCL1, CXCL10), and inflammasome components (NLRP3, CASP1, IL1A, IL18). NF-κB pathway members (NFKB1, NFKB2, RELA, RELB, NFKBIA) and STAT signaling molecules (STAT1, STAT3, JAK2) showed activation. Module contained 19 matched genes, mean log2FC = + 1.43 indicating inflammatory activation, enrichment score 8.1, contributing to placental dysfunction through sustained inflammation.

Module 7 - Apoptosis and cell death (47 nodes)

Exhibited p-value of 3.2 × 10⁻⁶, featuring intrinsic apoptosis pathway (BCL2, BAX, BAK1, BID, PUMA/BBC3, NOXA/PMAIP1, cytochrome c/CYCS, APAF1), extrinsic pathway (FAS, FASLG, TRAIL/TNFSF10, DR4/TNFRSF10A, DR5/TNFRSF10B), and executioner caspases (CASP3, CASP7, CASP9). Anti-apoptotic regulators (BCL2, BCL2L1/BCL-XL, MCL1) demonstrated downregulation while pro-apoptotic factors increased, tilting balance toward cell death. Module encompassed 16 matched genes, mean log2FC = + 1.18, enrichment score 7.4, explaining excessive trophoblast apoptosis characteristic of placental insufficiency.

Cross-module interaction analysis identified 287 inter-module edges connecting these seven modules, with strongest connectivity between Module 1 (folate metabolism) and Module 2 (methylation regulation, 47 inter-module edges), and between Module 3 (oxidative stress) and Module 4 (endothelial dysfunction, 38 edges). This extensive cross-talk demonstrates that folate deficiency triggers cascading perturbations across multiple biological systems rather than isolated pathway disruption. Network modularity score Q = 0.73 indicates strong community structure, validating functional module identification while confirming modules remain interconnected rather than isolated.

Figure 7 depicts the integrated network visualization of identified functional modules and their interactions. Seven distinct modules emerged including folate metabolism core (184 nodes), methylation regulation (127 nodes), oxidative stress response (93 nodes), endothelial dysfunction (76 nodes), trophoblast differentiation (68 nodes), inflammation (54 nodes), and apoptosis signaling (47 nodes), with extensive cross-module communications mediating the cascade from folate deficiency to placental pathology.

Model interpretability and clinical application value

Attention weight visualization provided transparent insights into model decision-making processes, revealing that the network prioritized MTHFR-FOLR1-SLC19A1 axis interactions when classifying placental insufficiency cases. Node contribution analysis through gradient-based attribution methods quantified individual molecular features’ influence on predictions, demonstrating that homocysteine elevation and S-adenosylmethionine depletion jointly account for 42% of classification confidence in disease samples. These mechanistic explanations align with established folate metabolism biology, validating model learning of genuine pathophysiological relationships rather than spurious correlations.

Attention weight visualization provided transparent insights into model decision-making processes, revealing that the network prioritized MTHFR-FOLR1-SLC19A1 axis interactions when classifying placental insufficiency cases. Node contribution analysis through gradient-based attribution methods (integrated gradients, computed using Captum library) quantified individual molecular features’ influence on predictions. Homocysteine elevation and S-adenosylmethionine depletion jointly accounted for 42% of classification confidence in disease samples, as determined by Shapley value decomposition. These mechanistic explanations align with established folate metabolism biology, validating that our model learns genuine pathophysiological relationships rather than spurious correlations or dataset-specific artifacts.

The interpretable architecture enhances clinical trust by exposing the reasoning pathway from molecular measurements through network propagation to final classification, contrasting with opaque “black box” approaches. Clinicians can examine which specific molecular perturbations drove a patient’s risk classification, potentially guiding targeted interventions addressing patient-specific pathway disruptions such as antioxidant supplementation for oxidative stress-dominant cases or methylation support for hypomethylation-driven pathology.

Regarding clinical translation and future directions

A fundamental limitation of this study warrants emphasis. Our model was trained on placental tissues collected at delivery, which restricts its current application to post-delivery diagnostic confirmation and mechanistic investigation rather than antenatal risk prediction. Placental tissue sampling during pregnancy poses significant risks and remains impractical for routine screening. Therefore, the framework cannot presently identify at-risk pregnancies before delivery.

The molecular signatures we identified hold value for understanding disease mechanisms and may inform future development of surrogate biomarkers detectable in maternal blood during pregnancy. Cell-free placental RNA, extracellular vesicles, and circulating metabolites represent promising candidates for non-invasive prenatal assessment that warrant investigation in prospective cohort studies⁶². Such studies should collect serial prenatal biospecimens alongside delivery-time placental tissues to correlate circulating markers with tissue-level molecular states characterized by our model.

Our model’s immediate clinical value lies in post-delivery diagnostic confirmation, mechanistic investigation of cases presenting with unclear etiology, and research applications characterizing molecular endotypes within placental insufficiency syndromes. The molecular signatures and pathway disruptions we identified could inform development of surrogate biomarkers detectable in prenatal samples, but this requires prospective investigation.

Pathway toward prenatal application

Future studies should pursue three complementary strategies to enable genuine antenatal risk assessment: (1) Identify circulating biomarkers in maternal blood that correlate with placental molecular signatures—cell-free placental RNA fragments, extracellular vesicles, or metabolite panels measurable via liquid biopsy during pregnancy. Preliminary studies suggest placental-derived circulating metabolites reflect tissue-level perturbations. (2) Develop predictive models trained on first or second-trimester biospecimens (maternal serum, plasma, urine) linked with delivery outcomes, incorporating our identified pathways as biological priors constraining model architecture. (3) Conduct prospective longitudinal cohort studies collecting serial prenatal samples alongside delivery-time placental tissues, enabling direct correlation between antenatal markers and tissue-level molecular states our model characterizes.

Only through such prospective validation can we legitimately claim early risk stratification capacity. Our current contribution establishes the molecular framework and computational methodology, providing foundation for translational development toward clinically deployable prenatal screening tools. We advocate cautious interpretation avoiding premature clinical implementation until appropriate validation studies demonstrate prenatal biomarker predictive validity.

Individual patient molecular profiles generated by the model guide precision medicine approaches, recommending specific interventions addressing patient-specific pathway perturbations such as antioxidant supplementation for oxidative stress-dominant cases or methylation support for hypomethylation-driven pathology. The transparent decision pathway from molecular measurements through network propagation to final classification enhances clinical trust and adoption, as physicians can trace diagnostic rationale through interpretable biological networks rather than opaque algorithmic black boxes.

Discussion

This investigation developed a heterogeneous graph neural network framework synthesizing genomic, transcriptomic, proteomic, and metabolomic datasets to systematically dissect molecular mechanisms underlying folate-mediated placental insufficiency. Our principal findings demonstrate that multi-omics fusion through graph-structured biological networks achieves superior diagnostic performance compared to single-omics or traditional machine learning approaches, while simultaneously providing interpretable mechanistic insights into disease pathogenesis. The model attained 94.7% classification accuracy, substantially surpassing baseline methods, and identified critical molecular signatures spanning the cascade from folate deficiency to placental dysfunction.

Methodological innovations and strengths

Our research introduces several methodological advances that distinguish this work from prior multi-omics integration studies. First, the heterogeneous graph construction strategy explicitly models diverse relationships among genes, proteins, and metabolites within folate metabolism networks, preserving biological semantics during integration rather than treating all molecular entities uniformly. This contrasts with conventional approaches that concatenate feature matrices or analyze layers independently, losing relationship information. Second, our type-specific attention mechanism adaptively weights molecular features according to node categories and relationship types, capturing differential importance across omics layers—transcriptomic data received 35.2% weight allocation, proteomics 31.8%, metabolomics 22.4%, genomics 10.6% in Shapley value analysis, reflecting genuine biological signal strength rather than arbitrary equal weighting. Third, the multi-scale graph convolutional architecture aggregates local neighborhood information while maintaining global network topology, enabling identification of both proximal interactions and distant pathway connections that manifest in disease phenotypes.

Biologically, we established a systematic framework for tracing the molecular cascade from folate deficiency to placental dysfunction, revealing previously underappreciated regulatory nodes and potential therapeutic targets. The identification of seven interconnected functional modules extends beyond recognized one-carbon metabolism disruption, illuminating extensive crosstalk between methylation dysregulation, oxidative stress amplification, and angiogenic impairment. Particularly notable is S-adenosylmethionine depletion’s central role as a nexus linking folate deficiency to genome-wide hypomethylation, transcriptional reprogramming, and cellular dysfunction—a systems-level insight difficult to obtain through single-pathway investigations. Demonstrating that MTHFR, FOLR1, and SLC19A1 form a critical regulatory axis offers mechanistic rationale for genetic susceptibility variations observed clinically, potentially explaining why some individuals develop placental complications despite adequate dietary folate while others remain protected despite marginal intake.

Compared to existing computational approaches, our graph neural network framework demonstrates distinct advantages in handling relational biological data. Traditional regression models and standard neural networks treat features independently, failing to leverage interaction information embedded in molecular networks. Existing network-based methods such as protein-protein interaction analysis typically operate on single omics layers. Recent multi-omics integration techniques including matrix factorization and canonical correlation analysis capture statistical associations but lack capacity to model complex non-linear relationships and hierarchical structures inherent in biological systems. Graph neural networks’ message-passing architecture naturally propagates information across network topology, identifying distant but functionally connected molecular alterations that simpler methods miss.

Limitations and considerations

Several limitations warrant careful consideration when interpreting our findings. First, the dataset size of 298 samples, while sufficient for proof-of-concept demonstration and statistically powered for detecting large effect sizes, limits ability to identify subtle molecular associations or rare genetic variants contributing to disease heterogeneity. Expanding to multi-center cohorts encompassing thousands of participants would enhance generalizability, enable discovery of population-specific mechanisms, and support more granular patient stratification. Second, we did not adjust for potential confounding covariates such as maternal body mass index, ethnicity, parity, or socioeconomic status. While our case-control matching on maternal age partially addresses confounding, unmeasured variables could influence observed molecular patterns. Future investigations should incorporate these covariates through multivariate regression frameworks or propensity score methods. Third, our cross-sectional design collecting tissues at delivery precludes direct causal inference about temporal sequences—we cannot definitively establish whether identified molecular changes precede, cause, or result from placental dysfunction. Longitudinal sampling throughout gestation would strengthen causal reasoning. Fourth, as discussed in Sect. 4.3.3, our model cannot currently enable antenatal risk prediction because training data derive from post-delivery placental tissues. Claims regarding early pregnancy risk stratification require prospective validation using prenatal biospecimens, representing crucial future work.

Technical limitations include computational intensity of training large graph neural networks—our model required approximately 18 h per training run on 4×A100 GPUs—potentially constraining accessibility for groups lacking high-performance computing infrastructure. The model’s reliance on curated biological network databases (STRING, KEGG, Reactome) introduces potential biases toward well-studied pathways, possibly overlooking novel mechanisms involving poorly annotated genes. Incorporating tissue-specific interaction data and unbiased correlation-based network inference alongside curated knowledge could mitigate this limitation.

Clinical translation and future directions

Despite current constraints, our framework establishes crucial foundation for clinical translation. The molecular signatures and pathway disruptions identified could guide development of prenatal screening biomarkers detectable in accessible samples (maternal serum, plasma, urine) collected during pregnancy. Preliminary studies suggest cell-free placental RNA and extracellular vesicles carry molecular information reflecting tissue-level states. Future research should pursue: (1) Multi-center prospective longitudinal cohorts collecting serial prenatal biospecimens linked with delivery outcomes and placental molecular profiling, enabling identification of antenatal biomarkers correlating with tissue signatures. (2) Incorporation of additional omics modalities—epigenomics (DNA methylation arrays, ATAC-seq), single-cell transcriptomics—to enhance resolution and capture cellular heterogeneity within placental architecture. (3) Extension to related pregnancy complications sharing pathophysiology with placental insufficiency (preeclampsia, gestational diabetes, preterm birth), leveraging transfer learning approaches to accelerate model development by initializing with weights learned from our folate-focused investigation. (4) Integration with clinical decision support systems, developing user-friendly interfaces enabling obstetricians to input patient molecular data and receive interpretable risk assessments with mechanistic explanations guiding personalized management.

The framework’s extensibility represents promising avenue for advancing precision obstetrics. By establishing computational paradigm and demonstrating biological interpretability, our work paves pathway toward molecular endotyping of pregnancy complications, moving beyond syndromic diagnoses toward mechanistically-informed classification enabling targeted interventions tailored to individual pathophysiology.

Conclusion

We developed a heterogeneous graph neural network framework that integrates multiomics data to investigate molecular mechanisms of folate-mediated placental insufficiency. Our model achieved 94.7% classification accuracy and identified seven interconnected functional modules spanning folate metabolism, methylation regulation, oxidative stress response, and angiogenesis pathways. Key molecular signatures including MTHFR downregulation, FOLR1 depletion, and homocysteine accumulation were identified as discriminative features.

The current model was trained on placental tissues collected at delivery, which limits its application to post-delivery diagnostic confirmation and mechanistic investigation rather than antenatal risk prediction. The molecular signatures and pathway disruptions we identified provide a foundation for developing circulating biomarkers detectable in maternal blood during pregnancy. Future prospective cohort studies should correlate prenatal biospecimens with placental molecular states to enable development of non-invasive screening tools. This framework establishes a computational approach for multiomics integration applicable to diverse pregnancy complications.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUROC

Area Under the Receiver Operating Characteristic Curve

ChIP-seq

Chromatin Immunoprecipitation Sequencing

CV

Coefficient of Variation

DHFR

Dihydrofolate Reductase

DNA

Deoxyribonucleic Acid

ELU

Exponential Linear Unit

FOLR1/FOLR2

Folate Receptor 1/2

GNN

Graph Neural Network

HAT

Heterogeneous Attention

HIF1A

Hypoxia-Inducible Factor 1 Alpha

HMDB

Human Metabolome Database

IQR

Interquartile Range

KEGG

Kyoto Encyclopedia of Genes and Genomes

LASSO

Least Absolute Shrinkage and Selection Operator

MTHFR

Methylenetetrahydrofolate Reductase

MTR

Methionine Synthase

PCA

Principal Component Analysis

PCFT/SLC46A1

Proton-Coupled Folate Transporter

QC

Quality Control

ReLU

Rectified Linear Unit

RFC/SLC19A1

Reduced Folate Carrier

RNA

Ribonucleic Acid

SHMT

Serine Hydroxymethyltransferase

SNP

Single Nucleotide Polymorphism

STAT3

Signal Transducer and Activator of Transcription 3

TF

Transcription Factor

VEGF

Vascular Endothelial Growth Factor

Author contributions

X.X. and Z.L. contributed equally to this work. X.X. designed the study, coordinated sample collection, performed clinical data analysis, and drafted the initial manuscript. Z.L. conducted multi-omics data preprocessing, quality control procedures, and contributed to experimental validation. Q.X. developed the graph neural network architecture, implemented computational algorithms, and performed bioinformatics analyses. H.X. executed laboratory experiments, managed sample processing, and contributed to data interpretation. M.Y. conceived and supervised the project, provided critical intellectual input, coordinated all research activities, and finalized the manuscript. All authors reviewed, edited, and approved the final manuscript for publication.

Funding

Sponsored by Fujian Province Natural Science Foundation (Grant No. 2024J011622).

Data availability

The multiomics datasets, complete analysis results, and computational code generated during this study are provided in Supplementary Materials. This file contains: (1) processed multiomics data matrices for all 298 samples; (2) complete differential expression results with raw and adjusted p-values for all features (Supplementary Table S1-S3); (3) full feature importance rankings including attention weights, gradient attributions, and permutation importance scores; (4) network construction parameters and edge lists; (5) Python scripts for data preprocessing, graph neural network training, and evaluation using actual experimental data. Patient identifiers have been removed to protect confidentiality. Additional raw sequencing data are available from the corresponding author (Mengting Yuan, ymt1125@outlook.com) upon reasonable request, subject to institutional review board approval and data sharing agreements.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This study was approved by the Research Ethics Committee of Longyan First Affiliated Hospital of Fujian Medical University (Reference Number: IRB-2024-LFAH-067). All participants provided written informed consent prior to enrollment. The study was conducted in accordance with the Declaration of Helsinki and relevant national regulations governing human subject research.

Consent for publication

All authors have reviewed the manuscript and consent to its publication. No identifiable information regarding participants has been included.