Participants were selected from the Pregnancy Metabolic Disease and Adverse Pregnancy Outcomes (PMDAPO) study cohort, a dynamic prospective cohort of pregnant women established at Guangzhou Women and Children’s Medical Center. Details of the cohort have been reported previously (Yang et al., 2020; Gan et al., 2022). Pregnant women were included if they received antenatal care and delivered at Guangzhou Women and Children’s Medical Center. Pregnant women were excluded if they had pre-existing diabetes, pre-existing hypertension, pre-eclampsia, artificial insemination and in-vitro fertilization, psychiatric diseases, or resided outside of Guangzhou from three months before pregnancy until delivery.

A total of 1,248 pregnant women with available fecal samples were selected from the PMDAPO cohort between January 2017 and February 2020 (Subcohort 1) to investigate the associations between PM_2.5 exposure, GDM, and impaired glucose homeostasis. After excluding 576 pregnant women whose fecal samples were collected only in the third trimester or fecal sample sequences provided < 10,000 reads, 672 pregnant women were included in the gut microbiota analysis (Subcohort 2). Using 1:1 matched nested case-control study, the GDM-associated blood metabolites (Subcohort 3) and circRNAs (Subcohort 4) were analyzed based on 30 pairs of pregnant women with GDM and matched controls without GDM, randomly selected from the PMDAPO cohort. The control subjects were matched with GDM women for maternal age (± 3 years), gestational weeks (± 3 weeks), gravidity (± 1) and parity (± 1). A schematic overview of the study is shown in Figure 1.

The protocol for this study was approved by the ethics review committee of Guangzhou Medical University, and all participants provided their voluntary signed informed consent.

At the first antenatal visit, information was collected on maternal age, gravidity, parity, last menstrual period, native place, residence address during pregnancy, pre-pregnancy weight, height, disease status, history of abortion, usage of antibiotics and other medication (including immunotoxic drugs and high-dose of commercial probiotics in the past six months or during pregnancy), conventional antibiotic treatment or probiotic supplementation in the preceding 4 weeks. Pre-pregnancy body mass index (BMI) was determined by calculating the pre-pregnancy weight divided by height squared (kg/m²). The season from 3 months before pregnancy to the 13th week of gestation was calculated based on the last menstrual period. Follow-up visits were performed according to routine obstetric examination intervals. Maternal clinical information was obtained from hospital medical records. The gestational age was determined by ultrasound examination.

A 75-g oral glucose tolerance test (OGTT) was conducted between 24 and 28 weeks of gestation. GDM was diagnosed when the following plasma glucose thresholds were met or exceeded at any point in time: fasting blood glucose of 5.1 mmol/L (92 mg/dL), 1-h blood glucose of 10.0 mmol/L (180 mg/dL), and 2-h blood glucose of 8.5 mmol/L (153 mg/dL) (American Diabetes Association Professional Practice Committee, 2022).

Ambient air pollutant exposure for the participants was estimated based on the residential address reported at enrollment (first antenatal visit). Daily (24-h average) pollutant monitoring data on sulfur dioxide (SO₂), nitrogen dioxide (NO₂), and particulate matter < 2.5 μm in aerodynamic diameter (PM_2.5) were collected from 11 National Air Quality Monitoring Stations in Guangzhou, China from the national urban air quality real-time publishing platform (http://106.37.208.233:20035/). The longitude and latitude (X1, Y1) of the pregnant women’s residential addresses and the longitude and latitude (X2, Y2) of each monitoring station were determined using online Coordinates Identification System provided by Baidu Map (http://api.map.baidu.com/lbsapi/getpoint/index.html) (Wu et al., 2018). The squared Euclidean distance was used to identify the nearest monitoring station, and the calculation formula was: D²=(X1−X2)²+(Y1−Y2)². The air pollutant data from the monitoring station closest to each participant’s residence were used as their individual exposure level (Su et al., 2020). Based on the last menstrual period and delivery date of each participant, the average daily exposure concentration of each pollutant during 3 months before pregnancy, the first trimester (1-13 weeks), the second trimester (14-27 weeks), the third trimester (28 weeks to delivery), 3 months before pregnancy to the first trimester, and 3 months before pregnancy to the second trimester was calculated as the exposure level of each participant in each exposure window.

Fecal samples were collected once per participant between 13 and 28 weeks of gestation, following a standardized and detailed protocol as previously described (Liu et al., 2019). The median gestational age at fecal sample collection was 19.3 weeks (IQR: 16.6-24.0). This window coincided with routine antenatal visits and partially overlapped with the standard OGTT screening window (24-28 weeks), which is etiologically relevant to GDM diagnosis and glucose homeostasis. In brief, pregnant women were instructed to follow standardized procedures prior to sample collection. All fecal samples were subpackaged and stored at -80°C within 30 minutes of collection. Total bacterial DNA was extracted from fecal samples using the MOBIO PowerSoil^® DNA Isolation Kit (12888-100) protocol and subsequently profiled through 16S rRNA gene (V4 region) sequencing. The V4 region of the 16S rRNA gene was amplified using universal primers 515F (5’-GTGYCAGCMGCCGCGGTAA-3’) and 806R (5’-GGACTACNVGGGTWTCTAAT-3’), with barcode sequences specific to each sample. The 16S rRNA amplicon sequences were processed using QIIME2 at Promegene Co. Ltd. (Shenzhen, China) in the same batch. Representative sequences were defined as merged sequences with 100% identity. The taxonomy of these representative sequences was identified using the classify-sklearn classification method based on the Greengenes 13.8 database (https://data.qiime2.org/2018.11/common/gg-13-8-99-515-806-nb-classifier.qza). The Greengenes database was selected to maintain methodological consistency with key prior work in this field, facilitating direct comparison and integration of findings (Mutlu et al., 1987; Ferrocino et al., 2018; Liu et al., 2019; Wei et al., 2022; Long et al., 2025).

Fasting blood samples were collected between 15 and 24 weeks of gestation. Serum was separated immediately for metabolite measurement and plasma was used for circRNA measurement. Based on our previous circRNA microarray profiling among pregnant women with GDM, the levels of the eight most relevant circRNAs were validated in 30 pairs of pregnant women (with and without GDM) by RT-qPCR. RT-qPCR was performed according to a standardized and detailed protocol (Yang et al., 2020).

Metabolomic analysis of the serum samples was conducted using untargeted liquid chromatography coupled with mass spectrometry (LC–MS; ACQUITY UPLC & Q-TOF Premier, Waters, Manchester, UK). An ACQUITY UPLC HSS T3 column (2.1 × 100 mm, 1.8 μm, Waters) was used in the LC–MS system, and the column oven was set to 40 °C. Solvent A consisted of water with 0.1% formic acid, while solvent B was acetonitrile with 0.1% formic acid. The flow rate was set to 0.3 mL/min, and the gradient elution was as follows: 0-2 min, 95% A;2-12 min, 95% A;12-15 min, 5% A;15-17 min, 5% A, 17-20 min, 95% A. Data acquisition was conducted in centroid mode using electrospray ionization (ESI) in both positive (ESI+) and negative (ESI−) modes over the mass-to-charge ratio (m/z) range of 50–1500, with a scan time of 0.2 s. For positive and negative modes, the capillary voltages were set to 1.4 kV and 1.3 kV, respectively, and the cone voltages to 40 V and 23 V, respectively. The source temperature was set to 120°C, with a cone gas flow of 50 L/h and a desolvation gas flow of 600 L/h. The desolvation temperature was set to 350°C and the collision energy was ramped from 10 to 40 V.

Data are presented as terms of mean ± SD or n (%). Chi-square tests were used for categorical variables, and t tests for continuous variables. Spearman correlation analysis was conducted to assess correlations among exposure levels of PM_2.5, SO₂ and NO₂. Multivariate logistic regression was employed to estimate the relative risk (RR) and 95% confidence intervals (95% CIs) of GDM associated with each 10 μg/m³ increase in PM_2.5 exposure during specific exposure windows. Linear regression models were used to analyze the associations of PM_2.5 exposure (increments of 10 μg/m³) on blood glucose levels during each exposure window, with fasting, 1-h and 2-h blood glucose levels as the dependent variables, respectively. In the single-pollutant model, independent variables included PM_2.5 exposure concentration, age, season of pregnancy, gravidity, parity, and pre-pregnancy BMI. The multi-pollutant model additionally included SO₂ and NO₂ exposure concentrations, based on the single-pollutant model.

Spearman correlation analysis was used to analyze the correlation between PM_2.5 exposure concentrations and gut microbiota. Hierarchical regression analysis was used to explore whether gut microbiota associated with PM_2.5 exposure moderated the relationship between PM_2.5 exposure and blood glucose levels. The genus significantly associated with PM_2.5 exposure was used as the moderator variable M, PM_2.5 exposure level as the independent variable X, and fasting, 1-h and 2-h blood glucose levels of OGTT as the dependent variables Y, respectively. All variables (X, M, and Y) were standardized. In the first step, the main effects of X and M on Y were examined using multiple linear regression, with X and M as independent variables, and Y as the dependent variable, adjusting for age, gravidity, parity, pre-pregnancy BMI, and exposure concentrations of SO₂ and NO₂. In the second step, the interaction term (X × M) was added to examine the moderating effect of gut microbiota on the association between PM_2.5 exposure and blood glucose levels. The moderation model can be expressed as Y=β₀+β₁X+β₂M+β₃XM, where β₃ quantifies effect modification. A positive β₃ indicates a stronger PM_2.5–glucose association at higher genus abundance, whereas a negative β₃ indicates attenuation.

Metabolite data extraction and preprocessing were performed using MassLynx 4.1 mass spectrometry workstation software. Then, data normalization, principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA) and orthogonal partial least squares discriminant analysis (OPLS-DA) were performed using SIMCA-P+ 13.0 (Umetrics, Umea, Sweden). Metabolite candidates were identified based on their mass-to-charge ratio (m/z) by searching for accurate masses in online databases, including the Human Metabolite Database (HMDB, v3.6), Kyoto Encyclopedia of Genes and Genomes (KEGG), and METLIN Metabolome Database. The metabolite candidates were selected based on the variable importance in projection (VIP) score combined with a two-sided t-test (VIP > 1, P < 0.05). High-throughput metabolic pathway enrichment analysis was performed using the MetaboAnalyst 5.0 platform, and metabolic network mapping, topological analysis and prediction of targets were performed using Cytoscape 3.9.1.

Based on the normal distribution of circRNA expression levels analyzed by the Kolmogorov-Smirnov test, the paired t-test or the Wilcoxon test was applied to evaluate differences in circRNA expression between participants with and without GDM. Spearman correlation analysis was used to analyze the associations among GDM-associated gut microbiota, PM_2.5-related microbiota involved in blood glucose regulation, and GDM-associated metabolites and circRNAs.

All data were duplicated with Epidata3.1, and analyzed with R software (version 4.2.1). A two-sided test was used, with the significance level set at α = 0.05.

Pregnant women in Subcohort 1 and Subcohort 2 were 18-45 years old, with GDM incidence of 19.15% (239/1,248) and 18.90% (127/672), respectively. Compared with those without GDM, pregnant women with GDM in both subcohorts were older and had higher pre-pregnancy BMI, gravidity, parity, fasting blood glucose, 1-h OGTT glucose, and 2-h OGTT glucose (P < 0.05). No statistically significant difference in abortion history was observed between women with and without GDM (P > 0.05) (Table 1). No statistically significant differences were found between the two subcohorts in terms of GDM incidence, age, pre-pregnancy BMI, gravidity or parity (Supplementary Table S1). For Subcohort 2, in which gut microbiota was analyzed, the median gestational age at fecal sample collection was 19.3 weeks (IQR: 16.6-24.0 weeks).

Supplementary Table S2 showed the exposure levels of air pollutants (including PM_2.5, SO₂ and NO₂) in different exposure windows, including 3 months before pregnancy, the first trimester, the second trimester, the third trimester, from 3 months before pregnancy to the 13th week of gestation, and from 3 months before pregnancy to the 27th week of gestation. Spearman correlation analysis showed that the exposure levels of PM_2.5, SO₂ and NO₂ were significantly positively correlated with each other in each exposure window (P < 0.05) (Supplementary Table S3).

The association of PM_2.5 exposure with GDM and impaired glucose homeostasis was analyzed among 1,248 pregnant women (Subcohort 1). Results from the single-pollutant logistic regression model showed that, after adjusting for age, pre-pregnancy BMI, pregnancy season, gravidity and parity, the risk of GDM increased by 36% (RR = 1.36, 95% CI: 1.05, 1.77), 46% (RR = 1.46, 95% CI: 1.14, 1.88) and 67% (RR = 1.67, 95% CI: 1.22, 2.29) for every 10 μg/m³ increase in PM_2.5 exposure during the three months before pregnancy, the first trimester, and from 3 months before pregnancy to the 13th week of gestation, respectively (Figure 2a). After further adjustment for SO₂ and NO₂, the results of the multi-pollutant model showed that the risk of GDM increased by 104% (RR = 2.04, 95% CI: 1.47, 2.84), 149% (RR = 2.49, 95% CI: 1.65, 3.75) and 290% (RR = 3.90, 95% CI: 2.12, 7.18) for every 10 μg/m³ increase in PM_2.5 exposure during the first trimester, from 3 months to the 13th weeks of gestation, and from 3 months to the 27 weeks of gestation, respectively (Figure 2b).

The results of the single-pollutant linear regression model showed that for every 10 μg/m³ increase in PM_2.5 exposure during the first trimester, the second trimester, and from 3 months before pregnancy to the 27th week of gestation, 1-h OGTT glucose increased by 0.35 mmol/L (95% CI: 0.20, 0.49), 0.01 mmol/L (95% CI: -0.16, 0.18), and 0.58 mmol/L (95% CI: 0.34, 0.82), respectively (Figure 2c). After adjustment for SO₂ and NO₂, the results of the multi-pollutant model showed that for every 10 μg/m³ increase in PM_2.5 exposure during the first trimester, from 3 months before pregnancy to the 13th week of gestation, and from 3 months before pregnancy to the 27th week of gestation, fasting blood glucose increased by 0.05 mmol/L (95% CI: 0.01, 0.10), 0.05 mmol/L (95% CI: 0.00, 0.11) and 0.11 mmol/L (95% CI: 0.03, 0.19), respectively. For every 10 μg/m³ increase in PM_2.5 exposure during the first trimester, the second trimester, from 3 months before pregnancy to the 13th week of gestation, and from 3 months before pregnancy to the 27th week of gestation, 1-h OGTT glucose increased by 0.35 mmol/L (95% CI: 0.16, 0.54), 0.37 mmol/L (95% CI: 0.17, 0.57), 0.59 mmol/L (95% CI: 0.34, 0.84) and 0.79 mmol/L (95% CI: 0.42, 1.16), respectively. For every 10 μg/m³ increase in PM_2.5 exposure during the first trimester and from 3 months before pregnancy to the 13th week of gestation, 2-h OGTT glucose increased by 0.36 mmol/L (95% CI: 0.18, 0.53) and 0.32 mmol/L (95% CI: 0.11, 0.54), respectively (Figure 2d).

Based on 672 pregnant women (Subcohort 2), the results of the Wilcoxon test showed that there were no statistically significant differences in gut microbiota α-diversity (such as Simpson index, Shannon index, Chao1 index and Pielou_e index) between pregnant women with GDM and those without GDM (P > 0.05, Supplementary Figure S1). Based on the principal coordinates analysis (PCoA) of Bray-Curtis, the results showed that there were no significant differences in gut microbiota β-diversity between pregnant women with GDM and controls (Supplementary Figure S2).

The results of LEfSe analysis (LDA>2) showed that there were 15 differentially abundant taxa between pregnant women with GDM and those without GDM. At the class and order levels, Bacilli, Clostridiales and Lactobacillales were found to be enriched in non-GDM women. At the family level, Lactobacillales and Clostridiales were also enriched in non-GDM women, while the UCG-010 family was enriched in GDM women. At the genus level, Turicibacter, Lactobacillus, Fusicatenibacter, Clostridium_sensu_stricto_1, Catabacter and Romboutsia were enriched in non-GDM women, while Bacteroides pectinophilus group, UCG-010 and Angelakisella were enriched in GDM women. After adjusting for age, pre-pregnancy BMI, gravidity and parity, the MaAsLin2 analysis showed that there were still significant differences in Bacteroides pectinophilus group and Lactobacillus between GDM and non-GDM women (Figures 3a, b).

The results of Spearman correlation analysis showed that there was no significant correlation between blood glucose levels and the α diversity indices of gut microbiota (P > 0.05). Fasting blood glucose and 1-h OGTT glucose were positively correlated with β- diversity of gut microbiota (P < 0.05; Figure 3c). The results of Spearman correlation analysis also showed that 37 bacteria were correlated with blood glucose levels. Among them, SCFAs-producing bacteria or inflammation-related bacteria Eggerthella, Rothia, Streptococcus and Ruminococcus torques group were negatively correlated with fasting blood glucose, while Bacteroides, Colidextribacter, Lachnospiraceae UCG-008 and Parabacteroides were positively correlated with fasting blood glucose. Anaerostipes was positively correlated with 1-h OGTT glucose, whereas Turicibacter and Odoribacter were negatively correlated with 2-h OGTT glucose. After adjusting for age, pre-pregnancy BMI, gravidity and parity, Eggerthella, Parabacteroides, Candidatus Stoquefichus, Ruminococcus torques group and Anaerostipes were correlated with blood glucose levels (Supplementary Table S4).

Meanwhile, fasting blood glucose was negatively correlated with GDM-associated gut microbiota, including the class Bacilli and the order Lactobacillales, while 2-h OGTT glucose was negatively correlated with the genus Turicibacter and positively correlated with the genus Angelakisella (P < 0.05) (Figure 3d; Supplementary Table S5).

The average daily exposure from 3 months before pregnancy to the 27th week of gestation was used to represent each pregnant woman’s exposure levels to air pollutants. Among 672 pregnant women (Subcohort 2), the average exposure concentrations of PM_2.5, SO₂ and NO₂ were 36.54 μg/m³, 12.11 μg/m³ and 55.57 μg/m³, respectively (Supplementary Table S6). PM_2.5 exposure levels were not significantly correlated with the Simpson index, Shannon index, Chao1 index or Pielou_e index of gut microbiota α diversity (P > 0.05), but significantly positively correlated with β diversity (r = 0.64, P < 0.01) (Supplementary Table S7).

The results of Spearman correlation analysis showed that PM_2.5 exposure levels were significantly correlated with 42 bacterial genera (P < 0.05). Among them, PM_2.5 exposure levels were positively correlated with Bacteroides, Odoribacter, Alistipes, Bilophila, Lachnospira, Sutterella, Terrisporobacter, Mailhella and 14 other bacterial genera. Additionally, PM_2.5 exposure levels were negatively correlated with 22 bacterial genera, such as Rothia, Raoultibacter, Gemella, Solobacterium, Fusicatenibacter, Eubacterium hallii group and Escherichia-Shigella (Figure 4; Supplementary Table S8).

Spearman correlation analysis between blood glucose levels and PM_2.5-associated genera showed that fasting blood glucose was negatively correlated with Rothia (r = -0.09), Raoultibacter (r = -0.10), Eggerthella (r = -0.16), Candidatus Stoquefichus (r = -0.09), Streptococcus (r = -0.09) and UBA1819 (r = -0.11), while it was positively correlated with Bacteroides (r = 0.08), Lachnospiraceae_UCG-008 (r = 0.11), Colidextribacter (r = 0.08) and UCG-003 (r = 0.09) (P < 0.05). Additionally, 2-h OGTT glucose was negatively correlated with Odoribacter (r = -0.11, P < 0.05) (Figure 4; Supplementary Table S9).

The moderating effects of PM_2.5 exposure-related gut microbiota on blood glucose levels were explored using hierarchical regression analysis. To evaluate effect modification, we included an interaction term (PM_2.5 × genus abundance) in the regression models. A statistically significant interaction indicates that the association between PM_2.5 exposure and glucose levels varies across different levels of the gut microbial genus. The main effect was the effect of PM_2.5 exposure (X) and gut microbiota (M) on blood glucose levels (Y) after controlling for age, pre-pregnancy BMI, gravidity, parity, SO₂ and NO₂. The interaction terms were the moderating effect of gut microbiota on the effects of PM_2.5 exposure on blood glucose levels. The results showed that 7 bacterial genera significantly modified the association between PM_2.5 exposure and blood glucose levels (P < 0.05; Table 2). For example, the positive association between PM_2.5 exposure and fasting blood glucose was stronger among women with higher relative abundance of Solobacterium (β: 23.53, 95% CI: 4.33, 42.73, P = 0.02) and Escherichia_Shigella (β: 0.17, 95% CI: 0.00, 0.34, P = 0.04). In contrast, the association between PM_2.5 and1-h OGTT glucose was attenuated among women with higher abundance of Fusicatenibacter (β: -1.10, 95% CI: -2.12, -0.07, P = 0.04) and Ruminococcaceae_UBA1819 (β: -15.85, 95% CI: -29.91, -1.79, P = 0.03). Similarly, higher abundance of Raoultibacter (β: -267.20, 95% CI: -504.32, -30.08, P = 0.03), Anaerofustis (β: -221.50, 95% CI: -411.87, -31.13, P = 0.02), Ruminococcaceae_UBA1819 (β: -19.54, 95% CI: -31.80, -7.27, P < 0.01) and Phascolarctobacterium (β: -1.88, 95% CI: -3.11, -0.65, P < 0.01) attenuated the association between PM_2.5 exposure and 2-h OGTT glucose.

Univariate significance tests were performed to compare all the metabolites, identifying significant differences between pregnant women with GDM and those without GDM. A total of 28 GDM-associated differential blood metabolites were identified (Supplementary Table S10). High-throughput metabolic pathway enrichment analysis based on these 28 differential metabolites of GDM showed that glycerophospholipid metabolism and sphingolipid metabolism may be involved in GDM (Figure 5a; Supplementary Table S11).

Topological data analysis showed that 11 GDM-associated differential metabolites had node degrees higher than the average, including phosphatidylethanolamine (degree = 39), Phosphatidic acid (degree = 26), Dolichyl β-D-glucosyl phosphate (degree = 25), Lysophosphatidylcholine (degree = 24), Phosphatidylcholine (degree = 24), Galactosylceramide (degree = 18), sphinganine-1-phosphate (degree = 10), 2-(S-Glutathionyl)acetyl glutathione (degree = 20), 3-Oxooctanoyl-CoA (degree = 8), Sphingomyelin (degree = 6) and (S)-Hydroxydecanoyl-CoA (degree = 6). The 11 differential metabolites were involved in 9 metabolic pathways, including glycerophospholipid metabolic, Glycosphingolipid metabolism, Phosphatidylinositol phosphate metabolism, Linoleate metabolism, Glycosphingolipid biosynthesis-globo series, Glycosphingolipid biosynthesis-ganglio series, Arachidonic acid metabolism, Saturated fatty acids β-oxidation and Xenobiotics metabolism, respectively (Figure 5b; Supplementary Table S12). Based on topological data analysis, the target genes of the GDM-associated differential metabolites were predicted using key node centrality indicators, including degree centrality (DC), betweenness centrality (BC) and closeness centrality (CC). Genes with DC, BC and CC values above the mean were selected as potential targets of GDM-associated differential metabolites (Supplementary Table S13). PLD1, PLD2, EHHADH and HADHA were selected as key target genes. PLD1 and PLD2 were involved in the glycerophospholipid and phosphatidylinositol phosphate metabolic pathways, while EHHADH and HADHA were involved in the saturated fatty acid β-oxidation pathway.

RT-qPCR analysis showed that several circRNAs, including hsa_circ_0001946, hsa_circ_0000154, hsa_circ_0006732, hsa_circ_0001016 and hsa_circ_0001439 were upregulated in GDM, while hsa_circ_0042852, hsa_circ_0004001 and hsa_circ_0006936 were downregulated in GDM (Supplementary Table S14). Signaling pathway prediction based on KEGG database showed that hsa_circ_0006732 and hsa_circ_0001439 were related to the insulin signaling pathway and the biological process of insulin response.

circRNA binding proteins were predicted by the CircInteractome database. The results showed that hsa_circ_0001946 had four kinds of binding proteins in the lateral region, among which IGF2BP2 had binding sites in both the lateral region and junction region. hsa_circ_0001439 had two kinds of lateral region-binding proteins, among which EIF4A3 had binding sites in both lateral region and junction region. EIF4A3 was identified as a common binding site for hsa_circ_0042852, hsa_circ_0004001, hsa_circ_0006936, hsa_circ_0001439, hsa_circ_0006732, hsa_circ_0000154, and hsa_circ_0001016. IGF2BP3 served as a shared binding protein for hsa_circ_0001946, hsa_circ_0006732 and hsa_circ_0001016 (Figure 6). Both IGF2BP2 and IGF2BP3 were RNA-binding proteins that interact with insulin-like growth factors. AGO proteins were involved in the processing and maturation of small RNAs, while EIF4A3 protein was associated with translation regulation.

Spearman correlation analysis showed that 12 GDM-associated gut microbial genera were significantly correlated with 17 GDM-associated differential metabolites. Among them, the GDM-associated differential metabolites sphinganine-1-phosphate and Sphingomyelin (d18:0/26:1(17Z)), annotated to the sphingolipid metabolic pathway, were positively correlated with Romboutsia and Catabacter, respectively. Sphingomyelin (d18:0/12:0) was negatively correlated with Angelakisella, which was enriched in pregnant women with GDM. Among the GDM-associated differential metabolites involved in the glycerophospholipid metabolic pathway, PE (24:1(15Z)/24:1(15Z)) was positively correlated with Lactobacillus, which was enriched in non-GDM pregnant women. Citicoline and PA (16:0/16:0) were positively correlated with Raoultibacter and Anaerofustis, respectively; both genera showed negative effect modification of the association between PM_2.5 exposure and 2-h OGTT glucose. PE (24:1(15Z)/22:0) and PC (24:0/24:1(15Z)) were negatively correlated with Fusicatenibacter, which was also enriched in non-GDM pregnant women and inhibited the association between PM_2.5 exposure and 1-h OGTT glucose (Figure 7a; Supplementary Table S15).

Spearman correlation analysis showed that seven GDM-associated bacterial genera were correlated with seven GDM-associated circRNAs. Among them, hsa_circ_0042852 and hsa_circ_0000154 were negatively correlated with Catabacter, enriched in non-GDM pregnant women, and with Angelakisella, enriched in GDM pregnant women. Conversely, both circRNAs were positively correlated with Escherichia-Shigella, which was associated with a stronger association between PM_2.5 and fasting blood glucose. hsa_circ_0004001 was negatively correlated with Catabacter and Angelakisella, but was positively correlated with Romboutsia, enriched in non-GDM pregnant women, and with Escherichia-Shigella. hsa_circ_0006936 was negatively correlated with Catabacter, Angelakisella, and Ruminococcaceae_UBA1819, the latter of which showed negative effect modification of PM_2.5 on 2-h OGTT glucose. hsa_circ_0001439 was negatively correlated with Catabacter, Angelakisella and Ruminococcaceae_UBA1819, but positively correlated with Escherichia-Shigella. hsa_circ_0006732 was negatively correlated with Angelakisella and Ruminococcaceae_UBA1819. hsa_circ_0001016 was positively correlated with Oscillospirales_UCG-010, enriched in GDM pregnant women, and with Fusicatenibacter, enriched non-GDM pregnant women, which showed negative effect modification of PM_2.5 on 1-h OGTT glucose (Figure 7b; Supplementary Table S16).

This study has several advantages. To our knowledge, this is the first study that provided epidemiological evidence of the moderating effect of PM_2.5 exposure, gut microbiota, plasma metabolites, circRNAs and GDM. The covariates could be adjusted with the large sample size. Based on the prospective cohort study, we were able to examine the longitudinal associations among PM_2.5 exposure, gut microbiota, plasma metabolites, circRNAs and GDM.

However, several limitations of this study should be acknowledged. Due to the lack of work address information, PM_2.5 exposure level was evaluated based solely on the participants’ residential addresses. And we relied on a single baseline address for exposure assessment and did not capture intra-city residential moves during pregnancy. This could result in non-differential misclassification of exposures with high spatial variability (e.g., air pollution). Future research would benefit from tracking address changes to enhance accuracy. In addition, detailed data on dietary intake and physical activity were not collected, and residual confounding by these factors cannot be ruled out. However, several factors mitigate the potential for confounding by these variables. All participants received standardized antenatal care at the Guangzhou Women and Children’s Medical Centre, which includes routine health education on balanced diet and appropriate physical activity. And their residential addresses during pregnancy were within Guangzhou, meaning they were exposed to similar regional dietary influences (traditional Cantonese diets are characteristically light and balanced), which may reduce extreme variation in dietary patterns. Besides, BMI was included as a covariate in all regression models, which was a strong proxy indicator of long-term energy balance influenced by both diet and physical activity. Third, the gut microbiota was assessed only once during mid-pregnancy. Although this period is key for GDM development, a single sample cannot reflect changes across early or late pregnancy. Therefore, our results provide a snapshot, and microbial or exposure effects at other times might also matter. It should be noted that our previous analysis found that the influence of gestational age on women’s gut microbiota composition was limited (Yang et al., 2020). Future studies featuring longitudinal tracking with multiple samples are required to establish precise temporal links. Fourth, this study did not evaluate interactions between PM_2.5 and other air pollutants (e.g. PM₁₀, CO, O₃), nor did it account for meteorological conditions such as temperature and humidity. The fecal and blood specimens were collected from different pregnant women in the same cohort, which may have introduced some potential impact on the study results. However, the pregnant women were included in the cohort at the same period using the same inclusion and exclusion criteria, and the study results still offer valuable reference for future research. Fifth, this study utilized the Greengenes (v13.8) database for taxonomic annotation. While this database has been widely used in comparable studies to ensure consistency, we acknowledge that it is not the most recent version. Future studies employing updated databases such as SILVA for validation would be a valuable addition. Sixth, gut microbiota profiling was based on 16S rRNA gene (V4) amplicon sequencing, which has inherent limitations in taxonomic resolution (typically up to the genus level) and cannot reliably distinguish species- or strain-level variation. Moreover, 16S amplicon data are compositional and do not directly quantify microbial functional genes or pathways. Thus, functional interpretation should be made with caution. Future studies using shotgun metagenomic sequencing along with updated reference databases and longitudinal sampling would help validate species/strain-level signals and elucidate functional mechanisms.