Clinical Subtype Stratification Reveals Gut Microbial Alterations, Candidate Markers, and Co-Occurrence Network Remodeling in Coronary Artery Disease

Figures

• 

  Figure 1. Sample composition and missingness of clinical variables across the five clinical groups. (A) Number of samples included in each group. (B) Missing proportions of each clinical variable across the five groups (%, 0–100%); darker colors indicate higher missing proportions, with exact percentage values annotated in each cell.

• 

  Figure 2. Alpha diversity, beta diversity, and community composition across the five clinical groups. (A–C) Observed OTUs, Chao1, and Shannon indices; (D) Bray–Curtis PCoA based on filtered OTU relative abundances; (E,F) group-mean relative abundances of the top 10 phyla and 15 genera with the highest average abundances. Taxa are ordered in descending order of mean abundance across groups, and the remaining taxa are merged into “Others”.

• 

  Figure 3. Clinical group-associated genera and cross-group core genera. (A) Clinical group-associated genera; the left heatmap displays the Z-scores of the mean relative abundances for the 12 trend genera across the five groups, while the right dot plot shows the Spearman correlation coefficients (ρ) between the CLR-transformed abundances of the corresponding genera and the predefined clinical ordinal ranking, with colors representing the −log10(FDR) values. These 12 genera were selected from the top 12 most significant positive and negative correlations ranked by FDR. (B) Prevalence (%) of highly prevalent genera across the five groups. Note that this ordinal analysis represents cross-sectional associations and does not imply longitudinal disease progression within individuals.

• 

  Figure 4. Comprehensive prioritization and cross-subtype distribution of candidate markers across disease subtypes. (A) Top 8 genera with the highest Marker scores within each subtype. The x-axis represents the Marker score, the dot color indicates the mean CLR difference (direction and magnitude) between the disease group and the Control group, and the dot size reflects the Random Forest feature importance. (B) Heatmap of Marker scores for 16 non-redundant candidate genera, selected based on their maximum scores across all subtypes. Gray cells indicate that the genus was not included in the retained candidate set for the corresponding subtype. Note: The Marker score solely indicates candidate priority and does not represent diagnostic probability or external validation performance. A comprehensive 574-row list detailing individual component values and weight sensitivity analyses is provided in Table S3.

• 

  Figure 5. SPIEC-EASI conditional dependence networks and topological summaries across the five clinical groups. (A–E) Networks for the Control, MCS, SA, UA, and AMI groups, respectively. All five groups utilized an identical candidate set of 40 genera and fixed node coordinates (arranged in a circular layout based on phylum and mean abundance). Node color indicates the phylum-level classification, while node size is proportional to its degree; gray nodes represent isolated nodes. Blue and orange edges denote positive and negative conditional dependencies, respectively, with edge width proportional to the absolute magnitude of the conditional association. N and E in each panel denote the sample size and the number of retained edges, respectively. (F–I) Mean degree, network density, global clustering coefficient, and Louvain modularity. Gray cells indicate that the genus was not included in the retained candidate set for the corresponding subtype. Note: Given that only a single network was fitted per group and sample sizes varied across groups, the observed topological differences were not subjected to statistical inferential testing and are intended solely for descriptive comparisons.

• 

  Figure 6. Internal classification model performance, candidate marker-clinical correlations, and covariate adjustment results. (A) Mean AUCs of random forest (RF), L1-regularized logistic regression, linear support vector machine (SVM), and radial basis function (RBF) kernel SVM in comparisons between each subtype and Control; each cell represents the mean of 15 validation folds obtained from 3 repeats of 5-fold repeated stratified cross-validation. The total sample sizes for the MCS, SA, UA, and AMI tasks were 101, 113, 156, and 131, with input genera counts of 133, 144, 148, and 149, respectively; results represent internal validation performance only. (B) Spearman correlation matrix between 18 candidate genera (the union of the top 15 genera by marker score for each subtype, selected based on the maximum absolute Spearman’s ρ across clinical indicators) and 10 clinical indicators; cell values and colors both represent ρ, with a minimum effective sample size of 20 for each combination. Complete FDR-significant results are presented in Table S5. (C) Up to 3 adjusted associations with FDR < 0.05 shown per subtype; points represent adjusted centered log-ratio (CLR) regression coefficients…