Skip to main content
Cardiovascular Diabetology logoLink to Cardiovascular Diabetology
. 2026 Aug 20;25:335. doi: 10.1186/s12933-026-03340-8

Integrated metagenomic and metabolomic profiling across glycated hemoglobin-defined glycemic states identifies microbe-metabolite networks in glycemic dysregulation and coronary artery disease

Jing Xu 1,#, Yanan Yang 2,#, Jingang Yang 1, Tianjie Wang 1, Tao Tian 1, Hongzhao You 1, Haobo Xu 1, Fasheng Zhu 1, Xueyan Zhao 1, Yonggang Sui 1, Chongming Wu 2,✉, Jiansong Yuan 1,✉, Weixian Yang 1,✉, Yuejin Yang 1,✉
PMCID: PMC13772513  PMID: 43040516

Abstract

Background

Dysglycemia and hyperglycemia often coexist with coronary artery disease (CAD), but how gut microbiota and circulating metabolites differ across glycated hemoglobin (HbA1c)-defined glycemic categories and relate to cardiovascular risk remains incompletely understood. We aimed to characterize cross-sectional microbial and metabolic profiles across HbA1c-defined glycemic categories and to compare the patterns identified at the species and species-level genome bin (SGB) levels.

Methods

In 600 participants from the ACS-GUT CARDIOME cohort, glycemic status was categorized as Normoglycemic, Dysglycemic, or Hyperglycemic according to HbA1c criteria. Fecal metagenomic sequencing and plasma metabolomics/lipidomics were performed. Species-level, SGB-level, polar-metabolite, and lipid features were associated with glycemic status using multivariable models adjusted for CAD status, demographic and anthropometric characteristics, lifestyle factors, and medication exposures. We used Light Gradient Boosting Machine (LightGBM) models to explore whether microbial and metabolic profiles could discriminate HbA1c-defined glycemic states. In individuals with elevated glycemia, we compared those with and without CAD to identify CAD-related multi-omics patterns. We additionally used SGB co-abundance networks, cross-comparison concordance, and matched association analyses to connect the glycemic and CAD findings.

Results

Across the HbA1c-defined categories, selected commensal taxa were less abundant in the elevated-HbA1c groups, accompanied by differences in sugars, amino-acid-related metabolites, and lipids. Species- and SGB-level analyses showed concordant alpha-diversity patterns, whereas significant community-level separation was detected only at the SGB level. Selected SGBs assigned to the same species also showed divergent glycemic associations, indicating that SGB-level profiling provided additional resolution for selected microbial associations beyond species-level summaries. SGB-level microbial models yielded higher area under the receiver operating characteristic curve (AUC) point estimates than species-level models across all three comparisons, with the largest difference observed for Hyperglycemic versus Normoglycemic classification. Combined microbial and metabolic models showed moderate discrimination for comparisons involving the Hyperglycemic group but limited discrimination between Dysglycemic and Normoglycemic participants. Among participants with elevated glycemia, CAD was associated with additional microbial and metabolic differences. An integrative analysis identified concordant microbial and metabolic candidates across glycemic and CAD comparisons and a denser SGB co-abundance network in CAD among matched participants with elevated glycemia.

Conclusions

HbA1c-defined glycemic categories are associated with coordinated differences in gut microbial composition, SGB-level features, and circulating metabolites, with additional CAD-related multi-omics variation. These findings provide hypothesis-generating candidates for future longitudinal, mechanistic, and intervention studies of glycemic dysregulation and cardiometabolic risk.

Keywords: Dysglycemia, Gut microbiota, Metabolomics, Coronary artery disease, Multi-omics

Research Insights

What is currently known about this topic?

Gut microbial and circulating metabolic alterations have been associated with diabetes and cardiovascular disease, but their patterns across intermediate HbA1c-defined categories remain incompletely characterized.

What is the key research question?

What microbial and circulating metabolic features characterize the three HbA1c-defined glycemic categories, and how do the findings compare between species- and SGB-level profiling?

What is new?

Integrated multi-omics analysis identified microbial and metabolic differences across the three HbA1c-defined categories. Species- and SGB-level analyses showed both concordant and resolution-specific patterns, and additional CAD-associated differences were observed among participants with elevated glycemia.

How might this study influence clinical practice?

Microbiome–metabolite signatures may help guide future studies aimed at identifying individuals at risk of glycemic deterioration and cardiovascular complications, but external validation is required before clinical application.

Introduction

Impaired glucose regulation (IGR), spanning intermediate dysglycemia and diabetes, is a major and growing global health problem [1]. Glycated hemoglobin (HbA1c) reflects longer-term glycemic exposure and is widely used to classify glycemic status. The gut microbiota contributes to metabolic homeostasis and has been linked to type 2 diabetes [2, 3]. Microbial metabolites and treatment-related shifts in microbial composition may both influence host glycemic regulation [4, 5].

Previous studies have established important associations between the gut microbiome and diabetes [2, 3]. Microbial and circulating metabolic features across intermediate HbA1c-defined categories, however, remain less well characterized. Because intermediate dysglycemia is heterogeneous and carries a substantial risk of developing diabetes [6], examining Normoglycemic, Dysglycemic, and Hyperglycemic groups together with circulating metabolomic and lipidomic profiles may help identify coordinated microbial and metabolic patterns associated with glycemic heterogeneity.

Glycemic dysregulation frequently coexists with coronary artery disease (CAD), a major cause of morbidity and mortality in people with diabetes [7]. Gut microbial and metabolic alterations have also been associated with atherosclerotic cardiovascular disease [8, 9]. Characterizing CAD-associated multi-omics patterns across different glycemic categories may therefore improve understanding of the biological features shared by these closely related cardiometabolic conditions.

In this study, we integrated fecal metagenomics with plasma metabolomics and lipidomics across three HbA1c-defined categories: Normoglycemic, Dysglycemic, and Hyperglycemic. We characterized microbial profiles at both the species and SGB levels and compared their diversity patterns, glycemia-associated features, clinical correlations, and classification performance. We further examined microbe-metabolite relationships and CAD-associated multi-omics patterns using matched and glycemia-stratified analyses. By linking glycemic heterogeneity with microbial and circulating metabolic profiles, this study provides candidate features and biological insights for future research on glycemic control and cardiovascular risk.

Results

Clinical characteristics of the overall cohort

A total of 600 participants were included in the analysis, categorized into three groups based on HbA1c levels: Normoglycemic (HbA1c < 5.7%, n = 203), Dysglycemic (HbA1c 5.7–6.4%, n = 215), and Hyperglycemic (HbA1c ≥ 6.5%, n = 182). An overview of the study design and group stratification is shown in Fig. 1. As shown in Table S1, median age differed across the three HbA1c-defined groups. The proportion of males was similar across the three groups. Body mass index (BMI), fasting glucose, and HbA1c values were significantly higher in the Dysglycemic and Hyperglycemic groups compared to the Normoglycemic group. These cross-sectional differences reflect the HbA1c-based group definitions and accompanying cardiometabolic heterogeneity. The prevalence of CAD also increased across glycemic categories (50.7%, 76.7%, and 86.3% in the Normoglycemic, Dysglycemic, and Hyperglycemic groups, respectively), and the use of anti-diabetic medications rose in parallel (1.0%, 13.5%, and 56.0%, respectively; Table S1). Given these clinical imbalances, subsequent microbial, polar-metabolite, and lipid association analyses included CAD status and prespecified demographic, anthropometric, lifestyle, and medication covariates.

Fig. 1.

Fig. 1

Overview of study design and multi-omics analytical workflow. The overall cohort included 600 participants with stool metagenomic and clinical data; untargeted plasma metabolomic/lipidomic profiling was conducted in a 287-participant sub-cohort. Participants were stratified into three HbA1c-based glycemic groups, with additional analyses in a CAD sub-cohort and integrative machine-learning models based on multi-omics features. Further methodological details are provided in the Methods section

Microbial community differences and functional associations

We first evaluated microbial diversity and composition at the species level. Species-level alpha-diversity measures differed across the HbA1c-defined groups. Both Chao1 richness and Shannon diversity indices were highest in the Dysglycemic group. Chao1 was significantly lower in the Hyperglycemic group compared to Dysglycemic (p = 0.00019) and Normoglycemic (p = 0.0025), while no significant difference was found between Normoglycemic and Dysglycemic individuals (p = 0.33) (Figure S1A). Similarly, Shannon diversity was higher in the Dysglycemic group than in the Normoglycemic group (p = 0.03) and the Hyperglycemic group (p = 0.017), whereas the Hyperglycemic and Normoglycemic groups did not differ (p = 0.79) (Figure S1B). These cross-sectional findings indicate higher microbial richness and evenness in the Dysglycemic group than in the Hyperglycemic group. Species-level principal coordinates analysis (PCoA) showed visually different group centroids, but pairwise permutational multivariate analysis of variance (PERMANOVA) comparisons were not statistically significant: Normoglycemic versus Dysglycemic, p = 0.097; Normoglycemic versus Hyperglycemic, p = 0.139; and Hyperglycemic versus Dysglycemic, p = 0.731 (Figure S1C).

Taxonomic analysis identified bacterial species with directional abundance patterns across the glycemic categories. Consistently altered taxa were selected by significance in at least one of the three pairwise comparisons (p < 0.05) and concordant fold changes across adjacent categories (> 1.2 for increasing features or < 0.8 for decreasing features). The species displayed in Figure S1D-E included higher-abundance patterns for Bacteroides thetaiotaomicron, Flavonifractor plautii, Bacteroides fragilis, Streptococcus SGB14888, and GGB9818 SGB15459, and lower-abundance patterns for Waltera intestinalis, Clostridium sp. 1001270H_150608_G6, Coprococcus eutactus, Gemmiger SGB15295, GGB33928 SGB15225, and Lachnospiraceae bacterium AM48 27BH (Figure S1D-E). Additional covariate-adjusted species-level associations are provided in Table S2.

Species associated more specifically with the dysglycemia were selected using Dysglycemic vs. Normoglycemic p < 0.05, Hyperglycemic vs. Normoglycemic p < 0.05, Hyperglycemic vs. Dysglycemic p > 0.05, and fold changes > 1.2 or < 0.8 in both the Dysglycemic/Normoglycemic and Hyperglycemic/Normoglycemic comparisons. The eight species displayed in Figure S1F showed lower abundance in Hyperglycemic and Dysglycemic participants than Normoglycemic participants, including GGB9712 SGB15244, Coprococcus eutactus, GGB9758 SGB15368, GGB79734 SGB15291, Collinsella SGB14861, Clostridium SGB4750, Oscillospiraceae bacterium NSJ 64, and GGB2982 SGB3964, while Klebsiella variicola showed higher abundance in Hyperglycemic and Dysglycemic participants than Normoglycemic participants (Figure S1F). Additional covariate-adjusted comparisons are reported in Table S2. These species-level patterns are consistent with previous metagenome-wide studies of type 2 diabetes and impaired glucose control, which reported depletion of commensal or butyrate-producing taxa and enrichment of opportunistic/pathobiont-related bacteria [3, 10].

Genome bins assigned within the same species may capture microbial heterogeneity that is not apparent from species-level summaries. We next examined associations at the SGB level. Alpha diversity measured by Chao1 and Shannon indices followed patterns similar to those observed at the species level (Fig. 2A, B). Pairwise PERMANOVA showed differences in SGB-level community composition for Normoglycemic versus Dysglycemic (p = 0.006), Normoglycemic versus Hyperglycemic (p = 0.002), and Hyperglycemic versus Dysglycemic (p = 0.040; Fig. 2C). In contrast, pathway-level beta-diversity differences were not significant: Normoglycemic versus Dysglycemic, p = 0.370; Normoglycemic versus Hyperglycemic, p = 0.209; and Hyperglycemic versus Dysglycemic, p = 0.514 (Fig. 2D). A phylogenetic tree based on the 150 most abundant SGB-level features further illustrated their taxonomic structure and relative abundance across the three groups (Fig. 2E).

Fig. 2.

Fig. 2

SGB-level gut microbial diversity, composition, and inferred functional potential across HbA1c-defined glycemic categories. A, B Alpha-diversity metrics including Chao1 richness (A) and Shannon diversity (B). (C, D) Principal coordinates analysis (PCoA) based on Bray–Curtis distances at the SGB level (C) and pathway level (D); pairwise PERMANOVA p-values are shown. E Phylogenetic tree of the 150 most abundant SGB-level features colored by phylum; outer rings indicate relative abundance across groups. (F, G) SGB-level features showing directional higher (F) or lower (G) abundance patterns, selected by significance in at least one pairwise comparison (p < 0.05) and consistent fold changes across adjacent categories (> 1.2 for increasing features or < 0.8 for decreasing features). H SGB-level features associated more specifically with the dysglycemia category, defined by Hyperglycemic vs. Normoglycemic p < 0.05, Dysglycemic vs. Normoglycemic p < 0.05, Hyperglycemic vs. Dysglycemic p > 0.05, and fold change > 1.2 or < 0.8. I Functional pathways showing group-associated abundance differences. Complete covariate-adjusted SGB-level and pathway results are provided in Tables S3 and S4. Statistical significance in abundance comparisons was assessed using MaAsLin2; *p < 0.05, **p < 0.01, ***p < 0.001

Directional SGB-level features were selected using the same pairwise-significance and concordant fold-change criteria as the species-level analysis. The patterns visualized in Fig. 2F, G included higher relative abundance of t_SGB10068 (Escherichia coli), t_SGB4761 (Enterocloster citroniae), t_SGB4742 (Hungatella hathewayi), t_SGB4037 (Clostridium butyricum), t_SGB4608 (Ruminococcus torques), t_SGB8003 (Streptococcus vestibularis), t_SGB8047_group (Streptococcus sanguinis), t_SGB8071 (Streptococcus parasanguinis), t_SGB15342 (Faecalibacterium prausnitzii) and t_SGB17256 (Bifidobacterium bifidum), and lower relative abundance of t_SGB4963 (Roseburia sp. BX1005), t_SGB15131 (Lawsonibacter sp. NSJ 51), t_SGB15299 (Gemmiger SGB15299), t_SGB6148 (Romboutsia timonensis), t_SGB4867 (Blautia faecicola), t_SGB1888 (Phocaeicola coprophilus), t_SGB1899 (Phocaeicola plebeius) and t_SGB15317 (Faecalibacterium prausnitzii) across the glycemic categories. Figure 2H further highlighted selected features associated more specifically with the dysglycemia, including lower t_SGB1903 (Phocaeicola plebeius), t_SGB1899 (Phocaeicola plebeius), t_SGB1888 (Phocaeicola coprophilus), t_SGB4651 (Lachnospiraceae bacterium BX3), t_SGB1939 (Parabacteroides sp. TM07 1AC), as well as higher t_SGB8003 (Streptococcus vestibularis), t_SGB8071 (Streptococcus parasanguinis) and t_SGB8028_group (Streptococcus anginosus) (Fig. 2F–H). Additional covariate-adjusted SGB-level results are provided in Table S3. Differences between the two taxonomic resolutions were also apparent among some SGBs assigned to the same species. For example, t_SGB15342 and t_SGB15317, both assigned to Faecalibacterium prausnitzii, showed opposing abundance patterns across the glycemic categories and contrasting correlations with glycemic indicators. The species-level F. prausnitzii feature was not retained among the directional species-level associations displayed in Figure S1D-F. This example indicates that individual SGB-level associations may not always be apparent in the corresponding species-level summary [11, 12].

Complementary pathway analysis revealed differences in multiple functional categories across the HbA1c-defined groups. The pathways displayed in Fig. 2I included carbohydrate-degradation, vitamin- and cofactor-biosynthesis, amino-acid, lipid-related, and fermentation pathways, with examples including D-xylose degradation, heme biosynthesis, allantoin degradation, ubiquinol biosynthesis, lipid A-core biosynthesis, and homolactic fermentation (Fig. 2I). Additional covariate-adjusted pathway results are provided in Table S4. These associations suggest differences in microbial functional potential related to nutrient utilization, cofactor metabolism, and energy production.

In summary, the taxonomic, SGB-level, and pathway analyses identified complementary microbial features across the three HbA1c-defined categories. The combination of directional patterns, Dysglycemic-category-associated features, and covariate-adjusted associations provides a structured set of microbial candidates for future studies of glycemic dysregulation.

Metabolomic and lipidomic differences across glycemic categories

Untargeted metabolomic profiling was performed in the 287-participant sub-cohort, whose baseline characteristics are summarized in Table S5. The polar-metabolite partial least squares-discriminant analysis (PLS-DA) showed considerable overlap among the groups, and the negative Q2 intercept indicated that the model did not support robust group separation (Fig. 3A). Nevertheless, the selected metabolites displayed in Fig. 3B revealed coordinated directional patterns across the HbA1c-defined categories. Several amino acids and related metabolites, including threonine, pipecolate, N-acetyl-beta-alanine, and oxitriptan, were lower in one or both elevated-HbA1c groups than in the Normoglycemic group, whereas 3-hydroxy-L-kynurenine was higher in both elevated-HbA1c groups. Glucose was higher in the Hyperglycemic group than in both other groups, while glyceraldehyde 3-phosphate was lower in the Dysglycemic and Hyperglycemic groups than in the Normoglycemic group, indicating that glucose-related metabolites did not change uniformly. Other displayed metabolites also showed contrasting patterns: 5-acetamidopentanoate was higher in both elevated-HbA1c groups, 2-isopropyl-3-oxosuccinate was higher in the Dysglycemic group than in the Normoglycemic group, and protocatechuatealdehyde was lower in the Hyperglycemic group than in the Normoglycemic group (Fig. 3B). Together, these findings suggest coordinated differences in amino-acid, carbohydrate, and organic-acid metabolism across the glycemic categories. Additional covariate-adjusted polar-metabolite results are provided in Table S6.

Fig. 3.

Fig. 3

Circulating metabolomic and lipidomic differences across HbA1c-defined glycemic categories. A PLS-DA score plot and permutation assessment for polar metabolites. B Circos plot of selected polar metabolites with directional group differences. C PLS-DA score plot and permutation assessment for lipidomic profiles. D Circos plot of selected lipid species with group-associated differences. E Bar plots of N-acetyl-beta-alanine, glyceraldehyde 3-phosphate, aconitate, 3-hydroxy-L-kynurenine and LTA4, selected to highlight metabolites associated more specifically with the dysglycemia category. F Exploratory KEGG pathway enrichment based on differential polar metabolites and lipids; dot size indicates metabolite count, and color represents the enrichment p-value. Complete covariate-adjusted polar-metabolite and lipid results are provided in Tables S6 and S7. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

In lipidomic analysis, the PLS-DA plot likewise showed substantial overlap and limited group separation (Fig. 3C). Lipid abbreviations follow standard shorthand: triacylglycerol (TG), phosphatidic acid (PA), phosphatidylethanolamine (PE), monosialodihexosylganglioside (GM3), sphingomyelin (SM), ceramide phosphoethanolamine (CerPE), fatty acid (FA), hydroxy fatty acid (OHFA), lysophosphatidylethanolamine (LPE), and phosphatidylcholine (PC). The selected lipids displayed in Fig. 3D included TG 52:13, PA 34:8, PE O-34:4, PE 40:2, GM3 30:1, SM 34:1, SM 32:0, and CerPE 34:1 (Fig. 3D). Additional covariate-adjusted lipid results are provided in Table S7.

Polar metabolites associated more specifically with the dysglycemia category were selected using Hyperglycemic vs. Normoglycemic p < 0.05, Dysglycemic vs. Normoglycemic p < 0.05, Hyperglycemic vs. Dysglycemic p > 0.05, and fold changes > 1.2 or < 0.8 in both the Dysglycemic/Normoglycemic and Hyperglycemic/Normoglycemic comparisons. N-acetyl-beta-alanine and glyceraldehyde 3-phosphate were lower in Dysglycemic and Hyperglycemic participants, whereas aconitate, 3-hydroxy-L-kynurenine and leukotriene A4 (LTA4) were higher in Dysglycemic and Hyperglycemic participants (Fig. 3E). Additional covariate-adjusted metabolite associations are provided in Table S6. Clopidogrel carboxylic acid was interpreted as a medication-exposure signal rather than an endogenous metabolic marker.

Exploratory Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment mapped the differential metabolites to pathways involving amino-acid metabolism, glycolysis and gluconeogenesis, the pentose phosphate pathway, galactose metabolism, and lipid-related processes across pairwise group comparisons (Fig. 3F). These results highlight metabolic pathways for future targeted and functional investigation.

Microbial–metabolite–pathway integration and clinical correlations

Correlation analysis integrating microbial species/SGB-level features, circulating metabolites, lipids, and clinical indicators revealed distinct association patterns (Fig. 4A; Figure S2). At the SGB level, t_SGB10068 (Escherichia coli), t_SGB8028_group (Streptococcus anginosus), t_SGB8003 (Streptococcus vestibularis) and t_SGB4037 (Clostridium butyricum) showed positive associations with glucose, HbA1c, and inflammatory or cardiac-stress markers, whereas t_SGB4644 (Clostridium sp. AF36 4), t_SGB5043 (Eubacteriaceae bacterium), t_SGB5090 (Clostridiales bacterium KLE1615), t_SGB4771 (Clostridiaceae unclassified SGB4771), t_SGB14861 (Collinsella SGB14861), t_SGB15291 (GGB79734 SGB15291), t_SGB14974 (Ruminococcaceae bacterium AM28 23LB), t_SGB4914 (Clostridium sp. AF34 10BH), t_SGB6148 (Romboutsia timonensis), t_SGB9663_group (Haemophilus parainfluenzae), t_SGB15317 (Faecalibacterium prausnitzii), and t_SGB4651 (Lachnospiraceae bacterium BX3) showed generally inverse patterns. Glucose and several glucose-related metabolites were positively associated with glucose and HbA1c and inversely associated with estimated glucose disposal rate (eGDR), whereas pipecolate, threonine, and selected commensal-associated features showed generally opposite patterns. Lipid features showed additional associations with glycemic and circulating lipid measures. These contrasting associations suggest that the integrated features capture complementary aspects of glycemic control and cardiometabolic status.

Fig. 4.

Fig. 4

Multi-omics associations across HbA1c-defined glycemic categories. A Correlation heatmap among selected SGB-level features, circulating metabolites, lipids, and clinical indicators. The x-axis shows multi-omics features, and the y-axis shows clinical variables; asterisks denote p < 0.05. B Sankey diagram summarizing putative species-metabolite-pathway links inferred using MetOrigin2 from correlation patterns and database-supported annotations. The left, middle, and right columns represent bacterial species, selected metabolites, and annotated metabolic pathways, respectively. Abbreviations: ApoA, apolipoprotein A; ApoB, apolipoprotein B; BRI, body roundness index; eGDR, estimated glucose disposal rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; hs-CRP, high-sensitivity C-reactive protein; LDH, lactate dehydrogenase; LDL-C, low-density lipoprotein cholesterol; Lp(a), lipoprotein(a); NT-proBNP, N-terminal pro-B-type natriuretic peptide; SCr, serum creatinine; TC, total cholesterol; TG, triglycerides. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

Integration using MetOrigin2 identified putative species-metabolite-pathway connections (Fig. 4B). Escherichia coli, Clostridium butyricum, Streptococcus vestibularis, Streptococcus anginosus, Streptococcus parasanguinis, Coprococcus eutactus, Faecalibacterium prausnitzii, Hungatella hathewayi, Enterococcus faecium, Phocaeicola coprophilus, Klebsiella variicola, Bacteroides thetaiotaomicron, and Bacteroides fragilis were linked to glucose, further mapped to pathways including glycolysis/gluconeogenesis, pentose phosphate pathway, galactose metabolism, starch and sucrose metabolism, amino sugar and nucleotide sugar metabolism, streptomycin biosynthesis, and neomycin, kanamycin and gentamicin biosynthesis. Escherichia coli, Streptococcus vestibularis, Streptococcus anginosus, Klebsiella variicola, Phocaeicola coprophilus, Bacteroides thetaiotaomicron, and Bacteroides fragilis were also linked to aconitate, which was mapped to the citrate cycle, glyoxylate and dicarboxylate metabolism, and other carbon-metabolism pathways. These inferred relationships nominate specific microbe-metabolite-pathway links for targeted experimental validation.

Together, the correlation and MetOrigin2 analyses connect the taxonomic and metabolic findings and provide a focused set of microbial-metabolite relationships for future mechanistic studies.

Machine-learning models for discrimination between glycemic states

We constructed Light Gradient Boosting Machine (LightGBM) classifiers using the 287 participants with paired metagenomic and metabolomic/lipidomic data. Microbial-only, metabolic-only, and combined models were evaluated for each pairwise glycemic comparison. Species- and SGB-level analyses used the same stratified 80% training and 20% held-out test partitions, and all supervised feature selection was confined to the training set.

Using SGB-level microbial features, the microbial-only, metabolic-only, and combined models yielded AUCs of 0.524, 0.508, and 0.503 for Dysglycemic versus Normoglycemic classification, respectively (Fig. 5A–C). Performance was higher for comparisons involving the Hyperglycemic group. The corresponding AUCs were 0.698, 0.814, and 0.839 for Hyperglycemic versus Normoglycemic classification (Fig. 5D–F) and 0.547, 0.718, and 0.721 for Hyperglycemic versus Dysglycemic classification (Fig. 5G–I).

Fig. 5.

Fig. 5

LightGBM classification models based on SGB-level microbial features and metabolic features. A Thirty most important features according to the random-forest model for Dysglycemic versus Normoglycemic classification. B SHapley Additive exPlanations (SHAP) summary plot showing the non-zero SHAP value features for Dysglycemic versus Normoglycemic classification. C Receiver operating characteristic (ROC) curves for Dysglycemic versus Normoglycemic classification. D Thirty most important features according to the random-forest model for Hyperglycemic versus Normoglycemic classification. E SHAP summary plot showing the non-zero SHAP value features for Hyperglycemic versus Normoglycemic classification. F ROC curves for Hyperglycemic versus Normoglycemic classification. G Thirty most important features according to the random-forest model for Hyperglycemic versus Dysglycemic classification. H SHAP summary plot showing the non-zero SHAP value features for Hyperglycemic versus Dysglycemic classification. I ROC curves for Hyperglycemic versus Dysglycemic classification. In the SHAP plots, color indicates the feature value, whereas horizontal position indicates the direction and magnitude of its contribution to the model output. Areas under the ROC curves (AUCs) are shown with 95% confidence intervals in the ROC panels. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

Repeating the analysis with species-level microbial features yielded microbial-only AUCs of 0.474, 0.341, and 0.503 and combined-model AUCs of 0.453, 0.831, and 0.705 for the three comparisons, respectively (Figure S3). Thus, the SGB-level models showed consistently higher AUC point estimates than the corresponding species-level models, with the largest difference observed in the microbial-only Hyperglycemic versus Normoglycemic comparison. However, the combined models performed similarly at the two taxonomic resolutions, indicating that their discrimination was primarily associated with metabolic features. Given the wide and overlapping confidence intervals, these differences were considered exploratory rather than evidence of statistically significant superiority. Feature-importance and SHAP analyses identified selected species- and SGB-level microbial features, lipids, and polar metabolites contributing to each comparison. Overall, discrimination remained limited between Dysglycemic and Normoglycemic participants but was higher for comparisons involving the Hyperglycemic group. These findings suggest that the Dysglycemic category may be characterized by subtler and more heterogeneous multi-omics differences relative to Normoglycemia, whereas the Hyperglycemic category shows more readily detectable alterations, particularly in metabolic features.

Gut microbial and metabolomic features associated with cad in participants with elevated glycemia

Because CAD was highly prevalent among participants with elevated HbA1c [13, 14], we combined the Dysglycemic and Hyperglycemic categories for the matched CAD analysis. This combined group is denoted as D&H in Fig. 6. Within this D&H population, participants with CAD were compared with participants without CAD, denoted as NCA in the figure. Propensity score matching was used to reduce differences in measured baseline characteristics between the two groups. Covariate balance before and after matching was assessed using absolute standardized mean differences (Figure S4).

Fig. 6.

Fig. 6

Gut microbial and metabolic features in matched Dysglycemic and Hyperglycemic participants with or without CAD. D&H denotes the combined Dysglycemic and Hyperglycemic categories; NCA denotes participants without CAD. A, B Chao1 richness and Shannon diversity. C, D PCoA based on Bray–Curtis distances at the SGB and pathway levels; neither comparison was statistically significant by PERMANOVA. E Relative abundance profiles of selected species. F Differentially abundant SGB-level features. G Differential pathways. H Differential polar metabolites. I Differential lipids. Positive and negative log2 fold changes in G–I indicate higher and lower abundance in CAD participants relative to NCA participants, respectively. *p < 0.05, **p < 0.01, ***p < 0.001. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; The “#” in front of metabolites represents the metabolite identified by comparison with standard substances

In the matched D&H cohort, Chao1 richness and Shannon diversity did not differ significantly between CAD and NCA participants, and neither SGB-level nor pathway-level beta diversity differed significantly (PERMANOVA p = 0.67 and p = 0.18, respectively; Fig. 6A–D). The species-level profile showed lower abundance of several Prevotella taxa, Roseburia intestinalis, Phocaeicola coprophilus, and Lachnospira eligens and higher abundance of Sutterella wadsworthensis, Ruminococcus gnavus, Bifidobacterium adolescentis, Alistipes communis, Bacteroides faecis, Streptococcus vestibularis, and Fusobacterium varium in participants with CAD (Fig. 6E). SGB-level and pathway analyses identified additional CAD-associated differences (Fig. 6F, G; Table S8). Ruminococcus gnavus is a mucin-utilizing bacterium that can produce immunologically active polysaccharides and has been associated with inflammatory conditions [15]. The enrichment of R. gnavus, together with Sutterella wadsworthensis, Streptococcus vestibularis, Fusobacterium varium, and Bacteroides faecis, may therefore indicate an inflammation-related microbial pattern in participants with CAD. Among the displayed polar metabolites, aconitate, arogenate, and clopidogrel carboxylic acid were higher, whereas riboflavin, leukotriene B4, 7alpha-hydroxycholesterol, N-acetyl-beta-alanine, threonine, glycocholate, and lysine were lower in participants with CAD (Fig. 6H). Multiple lipid species also differed by CAD status (Fig. 6I), with additional metabolite and lipid results provided in Tables S9-S10. Aconitate is a tricarboxylic-acid-cycle intermediate metabolically related to itaconate synthesis in activated macrophages [16]; its elevation may therefore point to altered energy and immunometabolic regulation in CAD. Clopidogrel carboxylic acid was interpreted as a marker of treatment exposure rather than an endogenous CAD-associated metabolite.

The matched analyses identified a coordinated set of microbial, metabolic, lipid, and pathway features associated with CAD among participants with elevated glycemia. These findings may help guide future studies of the gut microbiome-metabolite interface in patients with combined glycemic dysregulation and CAD.

Stratified multi-omics profiling of CAD across HbA1c-defined glycemic categories

We also examined microbial, metabolite, and clinical associations by CAD status within each HbA1c-defined category (Fig. 7). Alpha-diversity distributions varied across the six glycemia-by-CAD subgroups (Fig. 7A, B). SGB-level beta diversity differed by CAD status in the Normoglycemic (p = 0.005) and Dysglycemic (p = 0.040) categories, but not in the Hyperglycemic category (p = 0.614; Fig. 7C).

Fig. 7.

Fig. 7

Integrative analysis of SGB-level features and circulating metabolites associated with CAD across HbA1c-defined glycemic categories. A, B Chao1 and Shannon diversity across Normoglycemic (N), Dysglycemic (D), and Hyperglycemic (H) categories, stratified by CAD status. C SGB-level PCoA; p-values compare participants with and without CAD within each glycemic category. D, E Relative-abundance profiles of dominant species and SGB-level features. F–H Stratified correlation heatmaps calculated independently within the three glycemic categories. Features were selected using fold-change criteria and MaAsLin2-adjusted p < 0.05. *p < 0.05, **p < 0.01, ***p < 0.001. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

Species- and SGB-level relative-abundance profiles varied across the six glycemia-by-CAD subgroups (Fig. 7D-E). Stratified correlation analyses then showed shared and glycemic-category-specific associations among SGB-level features, circulating metabolites, lipids, and clinical indicators (Fig. 7F–H). Several SGB-level features identified in the matched CAD analysis, including t_SGB6754, t_SGB17234, and t_SGB3813, also appeared in the stratified analysis, suggesting that they may represent informative CAD-associated microbial features across different glycemic backgrounds. Additional SGB-level results are provided in Table S8.

CAD-associated metabolite and lipid differences varied by glycemic category. Aconitate, arogenate, and clopidogrel carboxylic acid were higher in CAD within the Normoglycemic and Hyperglycemic categories but not within the Dysglycemic category. MG O-12:5 was higher in CAD within the Dysglycemic and Hyperglycemic categories; LPE 20:4 and PE 28:3 were lower across multiple categories, whereas PS 26:6 and PS 36:7 were lower only in selected categories (Fig. 7F–H). Additional metabolite and lipid results are provided in Tables S9-S10. Previous studies have linked phospholipid remodeling to metabolic dysregulation [17], providing context for the glycemia-dependent PE and PS patterns observed here.

SGB co-abundance and concordance analyses connecting glycemic and CAD-associated patterns

To connect the glycemic and CAD findings, we applied two distinct directional-concordance criteria. First, we identified five features whose directions across the glycemic-category comparisons were consistent with their CAD-associated differences in the matched Dysglycemic and Hyperglycemic population (D&H). These D&H-concordant features comprised SGB14306 (an unclassified SGB assigned to GGB9342), SGB4761 (Enterocloster citroniae), SGB53517 (an unclassified SGB assigned to GGB33586), aconitate, and N-acetyl-beta-alanine. Second, we identified seven features whose glycemic-associated directions were consistent with their CAD-associated differences specifically within the Normoglycemic category. This Normoglycemic-context set included SGB17256 (Bifidobacterium bifidum), SGB4037 (Clostridium butyricum), SGB4699 (Clostridium symbiosum), SGB8003 (Streptococcus vestibularis), the SGB4540 group (Anaerostipes hadrus), pipecolate, and protocatechuatealdehyde.

We first compared SGB co-abundance networks between the 75 matched D&H CAD-NCA pairs. Among 180 prevalence-filtered SGBs, the CAD network contained more significant edges and connected nodes than the NCA network. Several selected SGBs, including SGB4037, SGB4699, and SGB8003, also showed greater connectivity in the CAD network (Figure S5A-B; Table S11). Directional concordance analysis identified SGB14306, SGB4761, SGB53517, and aconitate as increasing with elevated glycemia and in CAD, whereas N-acetyl-beta-alanine showed the opposite pattern. Additional concordant features included SGB17256, SGB4037, SGB4699, SGB8003, SGB4540 group, pipecolate, and protocatechuatealdehyde (Figure S5C; Table S12). Matched conditional logistic regression further showed positive CAD associations for SGB14306, SGB4761, SGB53517, and aconitate, whereas N-acetyl-beta-alanine, SGB4540 group, and protocatechuatealdehyde showed inverse associations or trends (Figure S5D; Table S12). Together, these analyses identified a focused set of microbial and metabolic features showing consistent patterns across glycemic and CAD-related comparisons. The reduction in N-acetyl-beta-alanine may reflect altered beta-alanine-related metabolism [18], whereas the increase in aconitate may indicate changes in tricarboxylic-acid-cycle and immunometabolic processes [19]. These concordant patterns provide testable hypotheses linking oxidative stress, energy metabolism, and inflammation to the coexistence of glycemic dysregulation and CAD.

Discussion

The present study extends previous microbiome studies of type 2 diabetes [2, 3] by examining three HbA1c-defined glycemic categories and integrating species-level and SGB profiles with pathway, plasma metabolomic/lipidomic, and coronary artery disease (CAD)-stratified analyses. Rather than treating glycemic dysregulation as a binary phenotype, this design allowed us to examine biological differences associated with an intermediate dysglycemic category and to assess whether SGB-level profiling adds information beyond species-level summaries. The findings support a coordinated relationship between gut microbial ecology and host metabolism across glycemic categories and provide a basis for mechanistic and longitudinal investigation.

Previous studies have established a close link between gut microbial composition and disordered glucose metabolism [2]. Notably, species-level richness and diversity were highest in the Dysglycemic group rather than changing monotonically across the three categories. This pattern may reflect ecological reorganization during intermediate dysglycemia rather than a simple progressive loss of diversity, although treatment exposure and other clinical differences may also contribute. Previous studies have similarly shown that microbial richness and community features vary across metabolic phenotypes and the broader cardiometabolic disease spectrum [3, 9, 10, 20, 21]. In our visualization analysis, opportunistic taxa such as Escherichia coli and Klebsiella variicola showed directional higher-abundance patterns across the HbA1c-defined categories, whereas Coprococcus eutactus and Roseburia sp. BX1005 showed lower-abundance patterns. Together, these changes may reflect a shift from commensal functions toward a more inflammation-prone intestinal environment. Experimental studies also illustrate that host metabolites can reshape gut ecology and amplify microbiota-associated inflammatory responses [22], while inflammatory processes may contribute to metabolic and lipid dysregulation in diabetic settings [23]. Short-chain fatty acids support barrier integrity and metabolic regulation through several host pathways, including G-protein-coupled receptor signaling [24]; therefore, lower abundance of putative short-chain-fatty-acid-producing taxa may be biologically relevant even though functional activity and circulating or fecal short-chain fatty acids were not directly measured here. Ruminococcus gnavus, which was enriched among participants with CAD, can produce an immunologically active capsular polysaccharide [25]. At the SGB level, t_SGB6148 (Romboutsia timonensis) showed inverse associations with glycemic and inflammatory traits. More broadly, microbiome studies in immune-related disorders suggest that disease-associated microbial patterns may vary with clinical severity and subtype [26]. Together, these observations suggest that glycemia-associated microbial features may participate in inflammatory and metabolic pathways that warrant direct experimental study.

An important implication of the present analysis is that species-level aggregation may obscure selected microbial associations. Alpha-diversity patterns were highly concordant between the species and SGB levels. However, species-level beta diversity did not differ significantly among the three HbA1c-defined categories, whereas SGB-level community composition showed significant group differences. This observation is consistent with large cross-cohort evidence that type 2 diabetes is associated with strain-specific microbial signatures [27] and with studies showing that strains assigned to the same species can differ substantially in genetic and functional characteristics [12]. Although SGBs do not constitute definitive strain identification, the added genomic resolution may capture associations that are diluted in species-level summaries. Consistent with this interpretation, SGB-level microbial models yielded higher AUC point estimates than species-level models in all three comparisons, with the largest difference observed for Hyperglycemic versus Normoglycemic classification. However, the combined models performed similarly at the two taxonomic resolutions and appeared to be driven mainly by metabolic features; therefore, these results do not establish the general superiority of SGB-level profiling. Because prediabetes is heterogeneous and some individuals develop diabetes within several years [6], the limited discrimination between Dysglycemic and Normoglycemic participants may reflect subtler and more heterogeneous multi-omics differences that are not fully captured by HbA1c alone. In contrast, comparisons involving the Hyperglycemic group showed more readily detectable alterations, particularly in metabolic features. The functional findings reinforce this interpretation: pathway-level beta diversity did not show global separation, whereas multiple individual pathways related to nutrient utilization, cofactor metabolism, and fermentation differed across categories. Thus, glycemic dysregulation may be associated with selective functional changes rather than wholesale restructuring of the inferred microbial functional profile.

Our metabolic findings provide complementary evidence of altered host substrate metabolism across HbA1c-defined categories. Previous metabolomic studies have reported disturbances in glucose and amino-acid metabolism across glucose-tolerance states [28]. In our cohort, higher glucose-related metabolites together with lower concentrations of several amino acids and derivatives are compatible with insulin resistance, in which peripheral glucose uptake and utilization are impaired while hepatic glucose production remains inadequately suppressed. The lower amino-acid signals may therefore reflect altered substrate use, including provision of gluconeogenic precursors, rather than accelerated tricarboxylic-acid-cycle flux alone [28, 29]. Changes across sphingomyelins, phospholipids, gangliosides, and other lipid classes further suggest broader remodeling of circulating lipid metabolism. Large-scale plasma lipidomics has linked lipid-class and molecular-species patterns with cardiometabolic disease risk [30], supporting further evaluation of the lipid signals identified here as potential links between glycemic dysregulation and cardiovascular vulnerability. These pathways also provide plausible connections between host metabolic state and the microbial features identified here and merit targeted quantification and experimental investigation.

Glycemic dysregulation and CAD frequently coexist and may share inflammatory and metabolic pathways. In the matched analysis, depletion of several Prevotella taxa, Roseburia intestinalis, Phocaeicola coprophilus, and Lachnospira eligens is consistent with broader evidence linking intestinal barrier and SCFA-related functions to cardiometabolic health [31, 32]. Conversely, Ruminococcus gnavus, Sutterella wadsworthensis, and Fusobacterium varium were enriched among participants with CAD. Previous observational and experimental work has linked gut microbial composition and its modulation with insulin sensitivity [33, 34], while inflammation-related dysbiosis can influence host inflammatory phenotypes [35]. The stratified analysis further showed CAD-associated SGB-level community differences in the Normoglycemic and Dysglycemic categories but not in the Hyperglycemic category, while CAD-associated metabolite and lipid patterns also varied by glycemic background. These findings suggest that the multi-omics features associated with CAD may be context dependent, consistent with evidence that microbiome and metabolome signatures vary across the cardiometabolic disease spectrum [21]. Formal interaction testing and longitudinal validation will be needed to determine whether glycemic status modifies these CAD-associated patterns. Accompanying differences in aconitate, riboflavin, amino-acid- and bile-acid-related metabolites, and lipid species suggest that CAD in individuals with elevated glycemia may involve interacting microbial, mitochondrial, and lipid-metabolic processes. Clopidogrel carboxylic acid should be interpreted as a marker of antiplatelet treatment exposure rather than an endogenous CAD-associated metabolite. Although residual confounding cannot be excluded, these findings support further investigation of shared gut-metabolic pathways in glycemic dysregulation and CAD. The integrative analysis provides a more direct bridge between the glycemic and CAD findings. The concordant directions of SGB14306, SGB4761, SGB53517, aconitate, and N-acetyl-beta-alanine across the glycemic and CAD comparisons suggest that a subset of microbial and metabolic alterations may be shared across these related cardiometabolic phenotypes. This observation is consistent with previous large-scale studies showing that microbial and circulating metabolic features can extend across the cardiometabolic disease spectrum, with some features shared between metabolic disorders and ischemic heart disease and others associated more specifically with cardiovascular disease [21, 36]. At the community level, the greater number of retained edges and connected nodes in the CAD network was consistent with differences in SGB co-abundance organization between CAD and NCA participants with elevated glycemia. Previous multi-omics studies of CAD have likewise identified disease-related microbial co-abundance groups and coordinated microbial and metabolic differences across CAD severity [37]. Together, our findings suggest that the overlap between glycemic dysregulation and CAD may involve not only individual microbial or metabolic features, but also changes in microbial community relationships and host-microbial metabolism. These patterns provide candidates for future studies examining intestinal barrier function, inflammatory responses, microbial fermentation, and host energy metabolism in the coexistence of glycemic dysregulation and CAD.

Our findings identified microbial and circulating metabolic differences not only across HbA1c-defined categories but also between participants with and without CAD among those with elevated glycemia, suggesting that these signatures may be relevant to both glycemic control and cardiovascular risk. This interpretation is supported by recent post-hoc analyses showing that remission from prediabetes to normoglycemia was associated with approximately half the long-term risk of cardiovascular death or hospitalization for heart failure [38]. Building on our cross-sectional findings, future longitudinal studies could use serial microbial and metabolite profiling to identify patterns associated with maintenance or restoration of normoglycemia, distinguish them from signatures preceding diabetes or cardiovascular disease, and monitor responses to lifestyle or glucose-lowering interventions. If validated prospectively, panels combining conventional risk factors with selected SGBs and metabolites may help identify individuals who would benefit most from intensive prevention and clarify pathways through which improved glycemic control translates into cardiovascular benefit.

A key strength of this study is the integrated evaluation of species-level and SGB-level microbial profiles across three HbA1c-defined categories, together with plasma metabolomic/lipidomic profiling and a matched CAD analysis. This design captures an intermediate glycemic category, resolves microbial associations that may be obscured at the species level, and links gut microbial features with circulating metabolic phenotypes relevant to diabetes and cardiovascular disease. Several limitations should be acknowledged. The cross-sectional design precludes temporal or causal inference, and HbA1c-based classification does not capture all aspects of glucose homeostasis. The single-center, cardiovascular-disease-enriched cohort may limit generalizability, and residual confounding may remain despite multivariable adjustment and matching. Longitudinal studies with external validation and targeted functional analyses are needed to establish the temporal and biological relevance of the identified signatures.

Conclusions

HbA1c-defined glycemic categories were associated with distinct SGB-level microbial features and circulating glucose-, amino-acid-, and lipid-related metabolites. The added information provided by SGB-level profiling and multi-omics integration highlights their potential for refining the biological characterization of glycemic dysregulation and concomitant CAD. These findings provide a foundation for longitudinal and experimental studies evaluating their relevance to glycemic remission, diabetes development, and cardiovascular outcomes. The concordance and network analyses further connect glycemic and CAD-associated patterns and identify a focused set of candidates for prospective and mechanistic validation.

Methods

Participants and study design

This study included 600 participants from the ACS-GUT CARDIOME cohort at Fuwai Hospital, National Center for Cardiovascular Diseases of China, enrolled between March and December 2018, to investigate relationships among gut microbiota, circulating metabolites, and glycemic status [39]. Detailed design, inclusion, and exclusion criteria of the ACS-GUT CARDIOME cohort have been described previously [39]. Glycemic status was classified into three groups using American Diabetes Association criteria based on glycated hemoglobin (HbA1c) levels: Normoglycemic (HbA1c < 5.7%, n = 203), Dysglycemic (HbA1c 5.7–6.4%, n = 215), and Hyperglycemic (HbA1c ≥ 6.5%, n = 182) [40]. The study protocol complied with the Declaration of Helsinki and was approved by the Ethics Committee of Fuwai Hospital (approval number: 2018–995). Written informed consent was obtained from all participants prior to enrollment. Body mass index (BMI) was calculated as weight divided by height squared (kg/m2). Body roundness index (BRI) was used as an anthropometric indicator of body fat distribution and was calculated from height and waist circumference according to the original formula [41], with higher values reflecting a more centrally rounded body shape. Estimated glucose disposal rate (eGDR) was used as an indirect marker of insulin sensitivity and was calculated using a validated equation based on clinical variables [42]; lower eGDR values indicate greater insulin resistance. For plasma metabolomics, after exclusion of samples that were unavailable for measurement, had insufficient volume, or failed analytical quality-control criteria, 287 participants were included in the current metabolomics analysis, comprising 93 normoglycemic, 101 dysglycemic, and 93 hyperglycemic participants. Baseline characteristics of this metabolomics sub-cohort are provided in Table S5.

Diversity and differential analysis

Alpha diversity (Chao1 richness and Shannon diversity) and beta diversity (PCoA based on Bray–Curtis distances) were calculated using the vegan R package (version 2.7-2) and visualized using ggplot2 (version 3.5.2). Pairwise alpha-diversity comparisons used Wilcoxon rank-sum tests, and beta-diversity differences were assessed using PERMANOVA.

PERMANOVA was first used to assess associations between microbial community composition and potential confounders, including CAD status, demographic and anthropometric variables, lifestyle factors, and medication exposure [21]. MaAsLin2 was then used to identify microbial and metabolic features associated with HbA1c-defined glycemic categories while adjusting for CAD status; sex; age; BMI; neck, waist, and hip circumference; smoking; alcohol consumption; and medication exposure [43]. Medication covariates included antiplatelet agents (aspirin, clopidogrel, and ticagrelor), statins, angiotensin-converting enzyme inhibitors or angiotensin receptor blockers, beta-blockers, calcium-channel blockers, nitrates, diuretics, spironolactone, glucose-lowering medications (insulin, metformin, acarbose, and other anti-diabetic drugs), and traditional Chinese medicine. The same covariates were used in microbial, polar-metabolite, and lipid association analyses. Residual confounding remains possible.

Identification of markers for glycemic state differentiation via machine learning models

To minimize the risks of information leakage and selection bias during feature identification, the 287 samples with matched metabolomic and microbiome data were randomly split into an 80% training set and a 20% hold-out test set before any feature selection and missing-value imputation. All data-dependent multi-omics preprocessing, feature selection, and hyperparameter optimization were performed using only the 80% training set. During cross-validation, preprocessing parameters were estimated exclusively from the corresponding fold-training subset and then applied to the fold-validation subset. The 20% hold-out test set remained untouched until the complete modeling pipeline, including the selected features and hyperparameters, had been finalized, after which it was used solely for internal hold-out evaluation. A two-stage feature-selection procedure was used. First, random forest models were trained under five-fold cross-validation using the training set. Mean Decrease Impurity (MDI) importance scores were averaged across folds to rank the features, and the 30 highest-ranked variables were retained. Second, LightGBM models were trained under five-fold cross-validation. For each fold, TreeExplainer was applied to the model fitted on the fold-training subset to calculate SHAP values for the corresponding validation subset. Global out-of-fold (OOF) feature importance was defined as the mean absolute SHAP value across all validation samples. Features with an OOF SHAP importance below 20% of the maximum observed OOF SHAP importance were excluded. The contribution of different omics features was quantified by the AUC through 2,000 iterations of Bootstrap resampling on the hold-out test set.

Statistical analysis

Analyses were conducted using R version 4.4.2 or Python 3.9. Spearman rank correlations were calculated for the heatmaps in Fig. 4A and Figure S2. MetOrigin was used as an exploratory tool to infer putative links among species, metabolites, and pathways [44].

Baseline clinical and biochemical characteristics were summarized as follows: continuous variables were presented as means ± standard deviations (SD) if normally distributed, or as medians with interquartile ranges (IQR) otherwise. Between-group differences in normally distributed continuous variables were assessed using one-way analysis of variance (ANOVA), whereas the Kruskal–Wallis test was used for non-normally distributed variables. Categorical variables were compared using χ2 tests, with Fisher’s exact test applied when expected cell counts in any category were < 5.

Propensity score matching in individuals with elevated glycemia and CAD

To minimize baseline differences between individuals with and without CAD under elevated glycemic conditions, we performed 1:1 propensity score matching (PSM). A total of 75 individuals without CAD were included as the control group, and 75 individuals with CAD were selected from a larger pool based on similarity in baseline characteristics. Propensity scores were estimated using logistic regression with age, gender, body mass index (BMI), LDL cholesterol, total cholesterol, and smoking status as covariates. Missing values in these variables were imputed using mean substitution prior to model fitting. Greedy nearest-neighbor matching without replacement was applied to construct unique matched pairs. Covariate balance was evaluated using absolute standardized mean differences (SMDs) before and after matching, with 0.1 used as a reference threshold. The metagenomic analysis included 75 matched pairs, and 42 of these pairs with available metabolomic/lipidomic data were included in the corresponding multi-omics analysis.

SGB Co-abundance, concordance, and matched association analyses

For exploratory integration of the glycemic and CAD findings, we defined two candidate sets based on directional concordance. D&H-concordant features showed glycemic-category changes consistent with matched CAD-NCA differences in the combined Dysglycemic and Hyperglycemic population, whereas Normoglycemic-context features showed glycemic-category changes consistent with CAD-NCA differences within the Normoglycemic category. SGBs detected in at least 40% of the 150 matched D&H metagenomic samples were included in the network analysis, with the selected candidate SGBs retained irrespective of prevalence. Relative abundances were converted to pseudo-counts, and FastSpar was applied separately to the CAD and NCA groups using 50 iterations and 1,000 bootstrap replicates [45]. Edges with |rho|> 0.30 and Benjamini–Hochberg-adjusted q < 0.05 within each network were retained. Paired Wilcoxon tests were used to assess differential enrichment between matched CAD and NCA participants.

For the concordance analysis, log2 fold changes were calculated from group means, with unpaired Wilcoxon tests used for glycemic-category and Normoglycemic CAD–NCA comparisons and paired Wilcoxon tests used for the matched D&H comparison. Conditional logistic regression stratified by matched pair was used to estimate CAD associations. Features were log-transformed and standardized before modeling, and odds ratios were reported per 1-SD higher feature with 95% confidence intervals. Analyses included 75 metagenomic pairs or 42 metabolomic pairs, as applicable. Multiple-testing correction was performed using the Benjamini–Hochberg method.

Supplementary Information

12933_2026_3340_MOESM1_ESM.jpg (3.7MB, jpg)

Additional file 1. Figure S1. Species-level differences in the gut microbiota across glycemic groups. (A, B) Alpha diversity measured by Chao1 richness (A) and Shannon diversity (B). (C) Species-level PCoA based on Bray-Curtis distances; pairwise PERMANOVA comparisons were not significant (all p > 0.05). (D, E) Species showing directional higher (D) or lower (E) abundance patterns, selected by significance in at least one pairwise comparison (p < 0.05) and consistent fold changes across adjacent categories (> 1.2 for increasing features or < 0.8 for decreasing features). (F) Species associated more specifically with the dysglycemia category, defined by Hyperglycemic vs. Normoglycemic p < 0.05, Dysglycemic vs. Normoglycemic p < 0.05, Hyperglycemic vs. Dysglycemic p > 0.05, and fold change > 1.2 or < 0.8. Complete covariate-adjusted species-level results are provided in Table S2. *p < 0.05, **p < 0.01, ***p < 0.001

12933_2026_3340_MOESM2_ESM.jpg (224.4KB, jpg)

Additional file 2. Figure S2. Correlation heatmap among selected species-level features, circulating metabolites, lipids, and clinical indicators. The x-axis shows multi-omics features, and the y-axis shows clinical variables; asterisks denote p < 0.05. Abbreviations: ApoA, apolipoprotein A; ApoB, apolipoprotein B; BRI, body roundness index; eGDR, estimated glucose disposal rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; hs-CRP, high-sensitivity C-reactive protein; LDH, lactate dehydrogenase; LDL-C, low-density lipoprotein cholesterol; Lp(a), lipoprotein(a); NT-proBNP, N-terminal pro-B-type natriuretic peptide; SCr, serum creatinine; TC, total cholesterol; TG, triglycerides. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

12933_2026_3340_MOESM3_ESM.jpg (586.2KB, jpg)

Additional file 3. Figure S3. LightGBM classification models based on species-level microbial features and metabolic features. (A) Thirty most important features according to the random-forest model for Dysglycemic versus Normoglycemic classification. (B) SHapley Additive exPlanations (SHAP) summary plot showing features with non-zero SHAP values for Dysglycemic versus Normoglycemic classification. (C) Receiver operating characteristic (ROC) curves for Dysglycemic versus Normoglycemic classification. (D) Thirty most important features according to the random-forest model for Hyperglycemic versus Normoglycemic classification. (E) SHAP summary plot showing features with non-zero SHAP values for Hyperglycemic versus Normoglycemic classification. (F) ROC curves for Hyperglycemic versus Normoglycemic classification. (G) Thirty most important features according to the random-forest model for Hyperglycemic versus Dysglycemic classification. (H) SHAP summary plot showing features with non-zero SHAP values for Hyperglycemic versus Dysglycemic classification. (I) ROC curves for Hyperglycemic versus Dysglycemic classification. In the SHAP plots, color indicates the feature value, whereas horizontal position indicates the direction and magnitude of its contribution to the model output. Areas under the ROC curves (AUCs) are shown with 95% confidence intervals in the ROC panels. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

12933_2026_3340_MOESM4_ESM.jpg (191.3KB, jpg)

Additional file 4. Figure S4. Covariate balance before and after propensity score matching in the CAD analyses. (A) Absolute standardized mean differences (SMDs) for the 42 matched metabolomics pairs. (B) Absolute SMDs for the 75 matched metagenomics pairs. Purple and blue bars represent covariate imbalance before and after matching, respectively. The dashed vertical line indicates an absolute SMD of 0.1 as a reference threshold. BMI, body mass index; LDL-C, low-density lipoprotein cholesterol

12933_2026_3340_MOESM5_ESM.jpg (865.8KB, jpg)

Additional file 5. Figure S5. SGB co-abundance networks and cross-comparison concordance in glycemic dysregulation and CAD. (A, B) FastSpar co-abundance networks in 75 matched NCA and CAD participants from the combined Dysglycemic and Hyperglycemic population. The largest connected component of the union network is shown using common node coordinates. Node size represents degree; orange and blue edges indicate positive and negative correlations, respectively. Edges satisfied |rho| > 0.30 and within-network Benjamini-Hochberg-adjusted q < 0.05. Node colors indicate paired differential enrichment between CAD and NCA (red, CAD enriched; blue, NCA enriched; gray, p ≥ 0.05) and are held constant across both networks. Differential-enrichment colors were based on paired Wilcoxon nominal p < 0.05 and are exploratory because no SGB passed global FDR correction. Selected candidate SGBs are labeled. (C) Directional concordance of 12 selected SGB and polar-metabolite features across Dysglycemic versus Normoglycemic (D vs N), Hyperglycemic versus Normoglycemic (H vs N), CAD versus NCA within D&H, and CAD versus NCA within Normoglycemic participants. Colors represent log2 fold change; * p < 0.05 and ** comparison-specific Benjamini-Hochberg-adjusted q < 0.05. (D) Conditional logistic regression estimates for CAD per 1-SD higher log-transformed feature. SGB analyses included 75 matched pairs and polar metabolite analyses included 42 matched pairs; points and horizontal lines represent odds ratios and 95% confidence intervals. D&H, combined Dysglycemic and Hyperglycemic categories; NCA, participants without CAD; OR, odds ratio; SGB, species-level genome bin

Author contributions

Jing Xu, Yanan Yang, Yuejin Yang, Weixian Yang, Jiansong Yuan, and Chongming Wu contributed to the conceptualization of the study. Jing Xu and Yanan Yang developed the methodology, software, and validation. Jing Xu and Yanan Yang performed the formal analysis, investigation, data curation, and visualization. Jingang Yang, Tianjie Wang, Tao Tian, Hongzhao You, Haobo Xu, Fasheng Zhu, Yonggang Sui, and Xueyan Zhao provided resources. Jing Xu and Yanan Yang wrote the original draft. Jing Xu, Yanan Yang, Yuejin Yang, Weixian Yang, Jiansong Yuan, and Chongming Wu reviewed and edited the manuscript. Yuejin Yang, Weixian Yang, Jiansong Yuan, and Chongming Wu supervised the project. Yuejin Yang and Jing Xu managed project administration. Funding was acquired by Jing Xu, Yuejin Yang, and Weixian Yang.

Funding

This work was supported by the Natural Science Foundation of Beijing, China (No. 7264319), the Postdoctoral Fellowship Program of China Postdoctoral Science Foundation (No. GZC20251509 and 2025M772331), the Chinese Academy of Medical Science Innovation Fund for Medical Sciences (No. 2016-I2M-1–009), and the Natural Science Foundation of China (No. 82192902).

Data availability

Metagenomic sequencing data associated with this study are deposited in the BioProject database of the National Genomics Data Center (NGDC, https://ngdc.cncb.ac.cn/bioproject/) under the accession number PRJCA017875 (https://ngdc.cncb.ac.cn/bioproject/browse/PRJCA017875) and are available under controlled access. Due to ethical and legal restrictions, participant-level data are available for research purposes only upon request, subject to approval by the Ethics Committee of Fuwai Hospital and the execution of appropriate data transfer agreements. This includes submitting a research proposal to the corresponding author Yuejin Yang (yangyjfw@126.com), where upon approval, all data analysis need to be done on a local server with protected access, in compliance with Personal Information Protection Law of the People’s Republic of China (PIPL) and Regulations on the Administration of Human Genetic Resources of the People’s Republic of China. Computational analyses were performed using the bioBakery suite of tools; species-level microbial abundances were computed using MetaPhlAn v.4.0 (https://github.com/biobakery/MetaPhlAn). Functional potential profiling was carried out with HUMAnN v.3.0 (https://github.com/biobakery/humann).

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Fuwai Hospital (No. 2018-995). Written informed consent was obtained from all participants.

Competing interests

The authors declare that they have no competing interests.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12933-026-03340-8.

Footnotes

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Jing Xu and Yanan Yang have contributed equally to this work.

Contributor Information

Chongming Wu, Email: chomingwu@163.com.

Jiansong Yuan, Email: jsyuantg@163.com.

Weixian Yang, Email: wxyang2009@sina.com.

Yuejin Yang, Email: yangyjfw@126.com.

References

  • 1.Saeedi P, Petersohn I, Salpea P, Malanda B, Karuranga S, Unwin N, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: results from the international diabetes federation diabetes atlas, 9(th) edition. Diabetes Res Clin Pract. 2019;157:107843. [DOI] [PubMed] [Google Scholar]
  • 2.Gurung M, Li Z, You H, Rodrigues R, Jump DB, Morgun A, et al. Role of gut microbiota in type 2 diabetes pathophysiology. EBioMedicine. 2020;51:102590. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Qin J, Li Y, Cai Z, Li S, Zhu J, Zhang F, et al. A metagenome-wide association study of gut microbiota in type 2 diabetes. Nature. 2012;490(7418):55–60. [DOI] [PubMed] [Google Scholar]
  • 4.Koh A, De Vadder F, Kovatcheva-Datchary P, Bäckhed F. From dietary fiber to host physiology: short-chain fatty acids as key bacterial metabolites. Cell. 2016;165(6):1332–45. [DOI] [PubMed] [Google Scholar]
  • 5.Wu H, Esteve E, Tremaroli V, Khan MT, Caesar R, Mannerås-Holm L, et al. Metformin alters the gut microbiome of individuals with treatment-naive type 2 diabetes, contributing to the therapeutic effects of the drug. Nat Med. 2017;23(7):850–8. [DOI] [PubMed] [Google Scholar]
  • 6.Tabák AG, Herder C, Rathmann W, Brunner EJ, Kivimäki M. Prediabetes: a high-risk state for diabetes development. Lancet. 2012;379(9833):2279–90. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Rawshani A, Rawshani A, Franzén S, Eliasson B, Svensson AM, Miftaraj M, et al. Mortality and cardiovascular disease in type 1 and type 2 diabetes. N Engl J Med. 2017;376(15):1407–18. [DOI] [PubMed] [Google Scholar]
  • 8.Sanna S, van Zuydam NR, Mahajan A, Kurilshikov A, Vich Vila A, Võsa U, et al. Causal relationships among the gut microbiome, short-chain fatty acids and metabolic diseases. Nat Genet. 2019;51(4):600–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Jie Z, Xia H, Zhong SL, Feng Q, Li S, Liang S, et al. The gut microbiome in atherosclerotic cardiovascular disease. Nat Commun. 2017;8(1):845. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Karlsson FH, Tremaroli V, Nookaew I, Bergström G, Behre CJ, Fagerberg B, et al. Gut metagenome in European women with normal, impaired and diabetic glucose control. Nature. 2013;498(7452):99–103. [DOI] [PubMed] [Google Scholar]
  • 11.Chen Y, Li Z, Hu S, Zhang J, Wu J, Shao N, et al. Gut metagenomes of type 2 diabetic patients have characteristic single-nucleotide polymorphism distribution in Bacteroides coprocola. Microbiome. 2017;5(1):15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.De Filippis F, Pasolli E, Tett A, Tarallo S, Naccarati A, De Angelis M, et al. Distinct genetic and functional traits of human intestinal Prevotella copri strains are associated with different habitual diets. Cell Host Microbe. 2019;25(3):444-453.e443. [DOI] [PubMed] [Google Scholar]
  • 13.Rao Kondapally Seshasai S, Kaptoge S, Thompson A, Di Angelantonio E, Gao P, Sarwar N, Whincup PH, Mukamal KJ, Gillum RF, Holme I, et al. Diabetes mellitus fasting glucose and risk of cause-specific death. N Engl J Med. 2011;364(9):829–841. [DOI] [PMC free article] [PubMed]
  • 14.Sarwar N, Gao P, Seshasai SR, Gobin R, Kaptoge S, Di Angelantonio E, et al. Diabetes mellitus, fasting blood glucose concentration, and risk of vascular disease: a collaborative meta-analysis of 102 prospective studies. Lancet. 2010;375(9733):2215–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Kandasamy S, Letchumanan V, Hong KW, Chua K-O, Ab Mutalib NS, Ng ALO, Ming LC, Lim HX, Thurairajasingam S, Law JW-F, et al. The role of human gut microbe Ruminococcus gnavus in inflammatory diseases. Prog Microbes Mol Biol 2023;6(1).
  • 16.Song J, Zhang Y, Frieler RA, Andren A, Wood S, Tyrrell DJ, et al. Itaconate suppresses atherosclerosis by activating a Nrf2-dependent antiinflammatory response in macrophages in mice. J Clin Invest. 2023. 10.1172/JCI173034. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Korbecki J, Bosiacki M, Kupnicka P, Barczak K, Ziętek P, Chlubek D, et al. Biochemistry and diseases related to the interconversion of phosphatidylcholine, phosphatidylethanolamine, and phosphatidylserine. Int J Mol Sci. 2024. 10.3390/ijms251910745. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Li R, Fu S, Lyu Z, Wang B, Hassman R, Shuster J, et al. Tissue-dependent enzymatic control of N-acetyl-β-alanine by PTER. J Biol Chem. 2026;302(7):113191. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Lushchak OV, Piroddi M, Galli F, Lushchak VI. Aconitase post-translational modification as a key in linkage between Krebs cycle, iron homeostasis, redox signaling, and metabolism of reactive oxygen species. Redox Rep. 2014;19(1):8–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Le Chatelier E, Nielsen T, Qin J, Prifti E, Hildebrand F, Falony G, et al. Richness of human gut microbiome correlates with metabolic markers. Nature. 2013;500(7464):541–6. [DOI] [PubMed] [Google Scholar]
  • 21.Fromentin S, Forslund SK, Chechi K, Aron-Wisnewsky J, Chakaroun R, Nielsen T, et al. Microbiome and metabolome features of the cardiometabolic disease spectrum. Nat Med. 2022;28(2):303–14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Zhao C, Hu X, Qiu M, Bao L, Wu K, Meng X, et al. Sialic acid exacerbates gut dysbiosis-associated mastitis through the microbiota-gut-mammary axis by fueling gut microbiota disruption. Microbiome. 2023;11(1):78. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Dong Y, Chauhan M, Betancourt A, Belfort M, Yallampalli C. Adipose tissue inflammation and adrenomedullin overexpression contribute to lipid dysregulation in diabetic pregnancies. J Clin Endocrinol Metab. 2018;103(10):3810–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Rosli NSA, Abd Gani S, Khayat ME, Zaidan UH, Ismail A, Abdul Rahim MBH. Short-chain fatty acids: possible regulators of insulin secretion. Mol Cell Biochem. 2023;478(3):517–30. [DOI] [PubMed] [Google Scholar]
  • 25.Henke MT, Brown EM, Cassilly CD, Vlamakis H, Xavier RJ, Clardy J. Capsular polysaccharide correlates with immune response to the human gut microbe Ruminococcus gnavus. Proc Natl Acad Sci U S A. 2021. 10.1073/pnas.2007595118. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Alshinnawy AS, Badiea EA, Swelam MS, Sayed AA, Mohamed MR. Decoding the gut microbiota in multiple sclerosis using nanopore long-read sequencing: insights into disease severity and subtypes. Mol Neurobiol. 2025;62(11):14245–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Mei Z, Wang F, Bhosle A, Dong D, Mehta R, Ghazi A, et al. Strain-specific gut microbial signatures in type 2 diabetes identified in a cross-cohort analysis of 8,117 metagenomes. Nat Med. 2024;30(8):2265–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Ji W, Xie X, Bai G, He Y, Li L, Zhang L, et al. Metabolomic approaches to dissect dysregulated metabolism in the progression of pre-diabetes to T2DM. Mol Omics. 2024;20(5):333–47. [DOI] [PubMed] [Google Scholar]
  • 29.Gu Y, Zang P, Li L-q, Zhang H-z, Li J, Li J-x, et al. A non-targeted metabolomics study on different glucose tolerance states. Int J Diabetes Dev Ctries. 2019;39(3):478–85. [Google Scholar]
  • 30.Eichelmann F, Sellem L, Wittenbecher C, Jäger S, Kuxhaus O, Prada M, et al. Deep lipidomics in human plasma: cardiometabolic disease risk and effect of dietary fat modulation. Circulation. 2022;146(1):21–35. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Lo Conte M, Cosorich I, Ferrarese R, Antonini Cencicchio M, Nobili A, Palmieri V, et al. Alterations of the intestinal mucus layer correlate with dysbiosis and immune dysregulation in human type 1 diabetes. EBioMedicine. 2023;91:104567. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Li YJ, Chen X, Kwan TK, Loh YW, Singer J, Liu Y, et al. Dietary fiber protects against diabetic nephropathy through short-chain fatty acid-mediated activation of G protein-coupled receptors GPR43 and GPR109A. J Am Soc Nephrol. 2020;31(6):1267–81. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Pedersen HK, Gudmundsdottir V, Nielsen HB, Hyotylainen T, Nielsen T, Jensen BA, et al. Human gut microbes impact host serum metabolome and insulin sensitivity. Nature. 2016;535(7612):376–81. [DOI] [PubMed] [Google Scholar]
  • 34.He X, Chen D, Guo Y, Zhang X, Ma Y, Zhao S. Walnut meal extracts rich in polyphenols mitigate insulin resistance and modulate gut microbiota in high fat diet-fed rats. J Med Food. 2022;25(6):618–29. [DOI] [PubMed] [Google Scholar]
  • 35.Sheikh IA, Bianchi-Smak J, Laubitz D, Schiro G, Midura-Kiela MT, Besselsen DG, et al. Transplant of microbiota from Crohn’s disease patients to germ-free mice results in colitis. Gut Microbes. 2024;16(1):2333483. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Talmor-Barkan Y, Bar N, Shaul AA, Shahaf N, Godneva A, Bussi Y, et al. Metabolomic and microbiome profiling reveals personalized risk factors for coronary artery disease. Nat Med. 2022;28(2):295–302. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Liu H, Chen X, Hu X, Niu H, Tian R, Wang H, et al. Alterations in the gut microbiome and metabolism with coronary artery disease severity. Microbiome. 2019;7(1):68. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Vazquez Arreola E, Gong Q, Hanson RL, Wang J, Sandforth L, He S, et al. Prediabetes remission and cardiovascular morbidity and mortality: post-hoc analyses from the diabetes prevention program outcome study and the daqing diabetes prevention outcome study. Lancet Diabetes Endocrinol. 2026;14(2):137–48. [DOI] [PubMed] [Google Scholar]
  • 39.Xu J, Dai D, Yang Y, Gao S, Yang J, Dong C, et al. Distinct microbial and metabolic shifts characterize acute coronary syndrome and recovery. Imeta. 2025;4(5):e70079. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Classification and Diagnosis of Diabetes. Standards of medical care in diabetes-2021. Diabetes Care. 2021;44(Suppl 1):S15-s33. [DOI] [PubMed] [Google Scholar]
  • 41.Thomas DM, Bredlau C, Bosy-Westphal A, Mueller M, Shen W, Gallagher D, et al. Relationships between body roundness with body fat and visceral adipose tissue emerging from a new geometrical model. Obesity (Silver Spring). 2013;21(11):2264–71. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Williams KV, Erbey JR, Becker D, Arslanian S, Orchard TJ. Can clinical factors estimate insulin resistance in type 1 diabetes? Diabetes. 2000;49(4):626–32. [DOI] [PubMed] [Google Scholar]
  • 43.Mallick H, Rahnavard A, McIver LJ, Ma S, Zhang Y, Nguyen LH, et al. Multivariable association discovery in population-scale meta-omics studies. PLoS Comput Biol. 2021;17(11):e1009442. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Yu G, Xu C, Zhang D, Ju F, Ni Y. MetOrigin: discriminating the origins of microbial metabolites for integrative analysis of the gut microbiome and metabolome. Imeta. 2022;1(1):e10. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Watts SC, Ritchie SC, Inouye M, Holt KE. FastSpar: rapid and scalable correlation estimation for compositional data. Bioinformatics. 2019;35(6):1064–6. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

12933_2026_3340_MOESM1_ESM.jpg (3.7MB, jpg)

Additional file 1. Figure S1. Species-level differences in the gut microbiota across glycemic groups. (A, B) Alpha diversity measured by Chao1 richness (A) and Shannon diversity (B). (C) Species-level PCoA based on Bray-Curtis distances; pairwise PERMANOVA comparisons were not significant (all p > 0.05). (D, E) Species showing directional higher (D) or lower (E) abundance patterns, selected by significance in at least one pairwise comparison (p < 0.05) and consistent fold changes across adjacent categories (> 1.2 for increasing features or < 0.8 for decreasing features). (F) Species associated more specifically with the dysglycemia category, defined by Hyperglycemic vs. Normoglycemic p < 0.05, Dysglycemic vs. Normoglycemic p < 0.05, Hyperglycemic vs. Dysglycemic p > 0.05, and fold change > 1.2 or < 0.8. Complete covariate-adjusted species-level results are provided in Table S2. *p < 0.05, **p < 0.01, ***p < 0.001

12933_2026_3340_MOESM2_ESM.jpg (224.4KB, jpg)

Additional file 2. Figure S2. Correlation heatmap among selected species-level features, circulating metabolites, lipids, and clinical indicators. The x-axis shows multi-omics features, and the y-axis shows clinical variables; asterisks denote p < 0.05. Abbreviations: ApoA, apolipoprotein A; ApoB, apolipoprotein B; BRI, body roundness index; eGDR, estimated glucose disposal rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; hs-CRP, high-sensitivity C-reactive protein; LDH, lactate dehydrogenase; LDL-C, low-density lipoprotein cholesterol; Lp(a), lipoprotein(a); NT-proBNP, N-terminal pro-B-type natriuretic peptide; SCr, serum creatinine; TC, total cholesterol; TG, triglycerides. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

12933_2026_3340_MOESM3_ESM.jpg (586.2KB, jpg)

Additional file 3. Figure S3. LightGBM classification models based on species-level microbial features and metabolic features. (A) Thirty most important features according to the random-forest model for Dysglycemic versus Normoglycemic classification. (B) SHapley Additive exPlanations (SHAP) summary plot showing features with non-zero SHAP values for Dysglycemic versus Normoglycemic classification. (C) Receiver operating characteristic (ROC) curves for Dysglycemic versus Normoglycemic classification. (D) Thirty most important features according to the random-forest model for Hyperglycemic versus Normoglycemic classification. (E) SHAP summary plot showing features with non-zero SHAP values for Hyperglycemic versus Normoglycemic classification. (F) ROC curves for Hyperglycemic versus Normoglycemic classification. (G) Thirty most important features according to the random-forest model for Hyperglycemic versus Dysglycemic classification. (H) SHAP summary plot showing features with non-zero SHAP values for Hyperglycemic versus Dysglycemic classification. (I) ROC curves for Hyperglycemic versus Dysglycemic classification. In the SHAP plots, color indicates the feature value, whereas horizontal position indicates the direction and magnitude of its contribution to the model output. Areas under the ROC curves (AUCs) are shown with 95% confidence intervals in the ROC panels. The “*” in front of metabolites represent metabolites which have been identified by LC-MS2; “#” in front of metabolites represents the metabolite identified by comparison with standard substances

12933_2026_3340_MOESM4_ESM.jpg (191.3KB, jpg)

Additional file 4. Figure S4. Covariate balance before and after propensity score matching in the CAD analyses. (A) Absolute standardized mean differences (SMDs) for the 42 matched metabolomics pairs. (B) Absolute SMDs for the 75 matched metagenomics pairs. Purple and blue bars represent covariate imbalance before and after matching, respectively. The dashed vertical line indicates an absolute SMD of 0.1 as a reference threshold. BMI, body mass index; LDL-C, low-density lipoprotein cholesterol

12933_2026_3340_MOESM5_ESM.jpg (865.8KB, jpg)

Additional file 5. Figure S5. SGB co-abundance networks and cross-comparison concordance in glycemic dysregulation and CAD. (A, B) FastSpar co-abundance networks in 75 matched NCA and CAD participants from the combined Dysglycemic and Hyperglycemic population. The largest connected component of the union network is shown using common node coordinates. Node size represents degree; orange and blue edges indicate positive and negative correlations, respectively. Edges satisfied |rho| > 0.30 and within-network Benjamini-Hochberg-adjusted q < 0.05. Node colors indicate paired differential enrichment between CAD and NCA (red, CAD enriched; blue, NCA enriched; gray, p ≥ 0.05) and are held constant across both networks. Differential-enrichment colors were based on paired Wilcoxon nominal p < 0.05 and are exploratory because no SGB passed global FDR correction. Selected candidate SGBs are labeled. (C) Directional concordance of 12 selected SGB and polar-metabolite features across Dysglycemic versus Normoglycemic (D vs N), Hyperglycemic versus Normoglycemic (H vs N), CAD versus NCA within D&H, and CAD versus NCA within Normoglycemic participants. Colors represent log2 fold change; * p < 0.05 and ** comparison-specific Benjamini-Hochberg-adjusted q < 0.05. (D) Conditional logistic regression estimates for CAD per 1-SD higher log-transformed feature. SGB analyses included 75 matched pairs and polar metabolite analyses included 42 matched pairs; points and horizontal lines represent odds ratios and 95% confidence intervals. D&H, combined Dysglycemic and Hyperglycemic categories; NCA, participants without CAD; OR, odds ratio; SGB, species-level genome bin

Data Availability Statement

Metagenomic sequencing data associated with this study are deposited in the BioProject database of the National Genomics Data Center (NGDC, https://ngdc.cncb.ac.cn/bioproject/) under the accession number PRJCA017875 (https://ngdc.cncb.ac.cn/bioproject/browse/PRJCA017875) and are available under controlled access. Due to ethical and legal restrictions, participant-level data are available for research purposes only upon request, subject to approval by the Ethics Committee of Fuwai Hospital and the execution of appropriate data transfer agreements. This includes submitting a research proposal to the corresponding author Yuejin Yang (yangyjfw@126.com), where upon approval, all data analysis need to be done on a local server with protected access, in compliance with Personal Information Protection Law of the People’s Republic of China (PIPL) and Regulations on the Administration of Human Genetic Resources of the People’s Republic of China. Computational analyses were performed using the bioBakery suite of tools; species-level microbial abundances were computed using MetaPhlAn v.4.0 (https://github.com/biobakery/MetaPhlAn). Functional potential profiling was carried out with HUMAnN v.3.0 (https://github.com/biobakery/humann).


Articles from Cardiovascular Diabetology are provided here courtesy of BMC

RESOURCES