Deciphering gut microbial impact in coronary artery disease through multimodal computational approaches

Abstract

Background

Coronary artery disease (CAD) remains a leading global health issue, with growing evidence implicating gut microbiome dysbiosis in its pathogenesis. This study characterized gut microbial alterations in previously reported data of 217 CAD patients compared to 57 healthy controls through advanced computational approaches and bioinformatics analysis. We employed an additional dataset to validate our results as well. Notably, our integrated approach, combining taxonomic profiling, core microbiome and co-occurrence network analyses, stochastic modelling, and machine learning, revealed significant CAD-associated microbial features.

Results

Our findings demonstrated that CAD is associated with specific changes in gut bacterial structure, diversity, composition, and stochastic signature, suggesting a dysbiosis in patients with CAD. The co-occurrence network for healthy controls further illustrated the robust and interconnected nature of the microbial community. Whereas, the CAD group showed a less interconnected and more fragmented microbial community. These changes in microbial population included the depletion of beneficial bacterial taxa and the enrichment of potentially harmful microorganisms. However, assessing the individual abundance of core microbiota at the genus level showed that only a small subset of gut core microbiota was correlated with CAD, although beneficial, may not fully compensate for the loss of other stabilizing taxa in the network. This suggests the importance of complex microbial interactions influencing overall health outcomes. We then identified the potential of gut microbial signatures as novel biomarkers for CAD risk assessment and diagnosis. Among those genera, Holdemanella, Acinetobacter, Fusicatenibacter, Sutterella, Agathobacter, Brevundimonas, Pseudomonas, Subdoligranulum, TM7x, and Delftia contributed the most.

Conclusions

Our study highlights a significant association between gut microbiota dysbiosis and CAD, characterized by distinct alterations in bacterial diversity and composition. Further research is needed to explore microbial dynamics across diverse populations and to clarify the role of microbiota in CAD onset and progression.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12866-025-04544-w.

Background

Coronary artery disease (CAD) is currently the number one killer of human health worldwide [1]. Over the past decades, several risk factors for CAD have been identified, including smoking, hypertension, hypercholesterolemia, diabetes, and obesity [2]. Despite the increasing ability to correct such factors, morbidity and mortality associated with CAD remain high, and they still represent the major cause of mortality. As a consequence, it is necessary to seek new factors that may play a role in the development of CAD and influence the prognosis in this group of patients; among these, particular attention has been focused on the gut microbiota [3].

With a total of more than 10¹³ microorganisms, intestinal microbiota is composed of bacteria, fungi, archaea, protists, and viruses that inhabit the gastrointestinal tract, forming a heterogeneous ecosystem that interacts together and with the host. The gut microbiota of human adults mostly consists of Firmicutes and Bacteroidetes, followed by Proteobacteria and Actinobacteria [4]. There are also pathogenic species such as Salmonella enterica, Bacteroides fragilis, Vibrio cholera, Campylobacter jejuni, and Escherichia coli, but in a low abundance of around 0.1% or less [5]. Lactobacillus, Ruminococcus, Clostridium, Faecalibacterium, Roseburia, Streptococcus (Firmicutes), Bacteroides, Prevotella (Bacteroidetes), Escherichia, Desulfovibrio (Proteobacteria), and Bifidobacterium (Actinobacteria) are among the predominant luminal microbial genera, and they can be identified in stool samples [6]. The composition of gut microbiota strongly varies depending on several environmental and lifestyle factors, such as dietary habits, intestinal infection and the administration of antibiotic drugs, host BMI, and exercise frequency [7].

Gut microbiota maintains a symbiotic relationship with the gut mucosal immune system, with substantial metabolic, immunological, and gut protective functions in the healthy individual [8]. For instance, the microorganisms residing in the human gut can significantly affect human health through their metabolites, particularly short-chain fatty acids (SCFAs) such as propionate, acetate, and butyrate [9, 10]. The SCFAs, which are produced during the fermentation of dietary fibres, can enter various cell types and inhibit histone deacetylases, leading to a suppression of pro-inflammatory responses and enhancement of anti-inflammatory gene expression [9].

Alterations in the gut microbiome may cause gut microbiome dysbiosis, damage the integrity of the intestinal mucosal barrier, and increase the level of systemic inflammation, thereby affecting the health of the host. In recent years, diseases such as inflammatory bowel disease (IBD) [11], asthma [12], depression [13], and diabetes [14] have been associated with the composition of the gut microbiota. There is a piece of evidence suggesting a higher risk of coronary artery disease (CAD) due to a lesser variety of the gut microbiome and a different bacterial composition [15–17]. This implies that gut microbiota may contribute to the development of CAD, opening new perspectives on preventive and therapeutic strategies for CAD via probiotics and prebiotics approaches, respectively [18, 19].

Our study aims to perform an integrated analysis to understand the gut microbiota’s role in CAD pathophysiology comprehensively. We combined high-throughput sequencing data with computational modeling and advanced bioinformatics processing pipelines to (1) generate a list of disease-associated microbial alterations and identifying core bacterial taxa for CAD progression, (2) investigate the stability of microbial networks, with an emphasis on keystone taxa, (3) Describe microbial growth dynamics by using stochastic simulations to explore how temporal changes are captured and (4) develop machine learning models for implicitly identifying microbial biomarkers as diagnostic indicators in CAD. The findings can contribute to the nascent microbiome-directed cardiovascular research field and may pave the way for novel precision medicine approaches to reduce CAD risk (Fig. 1).

Fig. 1.

Graphical abstract illustrating the main steps of our analysis workflow

Methods

Data collection

To investigate the global microbial signatures associated with CAD, we collected and analyzed 16SrRNA sequencing data from four publicly available datasets in the MGnify database [20], including the project numbers PRJDB7456, PRJNA810651, PRJNA984163, and PRJNA779598. The collection, including a total of 274 samples sequenced from the gut targeting the V3-V4 region and containing paired-end and single-end reads, was selected for CAD and healthy controls (217 CAD patients vs. 57 healthy controls). In addition, we utilized another collection comprised of 197 patients diagnosed with severe CAD and 41 healthy controls (project ID PRJNA503710) as validation dataset.

It is worth noting that detailed demographic information such as age, sex, and body mass index (BMI) was not systematically collected for all participants used in these studies, as the primary focus was on microbiome data analysis. As a result, we were unable to provide a comprehensive analysis associated with demographic information. However, to enhance the transparency, we presented a summary demographic table that reports key cohort-level characteristics, including BMI, sex distribution, age range, hypertension history, and diabetes history, for each dataset analyzed (Supplementary Table 1).

Data preprocessing

All quality control, preprocessing, and downstream analyses were performed entirely in R, using different packages. Initial processing included primer removal, filtering of low-quality reads, and trimming performed in the DADA2 pipeline (v. 1.30.0) [21]. Low-quality reads were filtered using the “filterAndTrim” function with the following parameters (maxN = 0, maxEE = c(2, 2), truncQ = 2, minLen = c(250, 220), trimLeft = 20, rm.phix = TRUE, compress = TRUE). Then, error rates were estimated using the “learnErrors” function. The denoised sequence table was constructed after sample inference and merging paired reads. We also dereplicated the reads, merged pair-ends and removed chimeras in DADA2. Taxonomic annotation was performed with DECIPHER version 2.30.0 [22], based on the SILVA version 138.1 [23] as a reference database. Finally, the ASV abundance, taxonomy table, and phylogenetic tree were combined with the sample metadata to create a phyloseq object with 112 unique bacterial taxa assigned to six taxonomic ranks. All singletons and reads from sources other than bacterial origin (i.e. mitochondria, chloroplasts) were removed from the datasets. To account for large differences in read numbers, all samples with less than 5000 reads were removed, which resulted in the removal of 12 samples from the main dataset. Furthermore, all ASVs with relative abundance of < 0.005% were omitted.

Batch effect assessment and correction

To assess the presence of potential batch effects, we performed Principal coordinate analysis (PCoA) based on Bray–Curtis dissimilarity and tested using PERMANOVA (adonis2, 999 permutations) with the batch variable as a covariate. Before correction, the analysis revealed a significant batch contribution (R² = 0.795, p = 0.001), confirming strong dataset-specific variation (Supplementary Fig. 1A). To correct for any detected batch effects, we then applied a variance stabilizing transformation (VST) using the DESeq2 package [24], followed by the “removeBatchEffect” function implemented in limma [25], specifying batch as the adjustment variable and health status (CAD vs. control) as a design factor. This approach effectively normalized the abundance data across batches while preserving biologically meaningful differences between the study groups (Supplementary Fig. 1B). The corrected data were then used for subsequent analyses to ensure that any observed effects were biologically relevant and not confounded by batch effects.

Diversity analysis

For alpha-diversity analyses, all samples were rarefied to 2000 reads per sample to consider for differences in the number of reads across samples. We then estimated the Shannon diversity (H´) of alpha diversity via the “diversity” function from the vegan package [26]. Statistical significance between groups was assessed with the Wilcoxon rank-sum test in R. Beta diversity was further evaluated using Bray–Curtis dissimilarity from the function “vegdist” of the vegan package’’. PCoA analysis was then conducted to visualize group separation. Statistical differences in community structure between CAD and controls were tested with PERMANOVA via the function “adonis” and 999 permutations by employing the vegan package.

Differential abundance analysis

To identify taxa that were differentially abundant between CAD and healthy control groups, we used the DESeq2 package in R to achieve this aim. The OTU table was log-transformed, and low-abundance taxa were filtered out by retaining only those with non-zero counts in at least 20% of samples. The health status was used as the main variable in designing the formula. Genera with and were considered both statistically and biologically significant. In the genus-level differential abundance analysis, we applied False Discovery Rate (FDR) correction using the Benjamini–Hochberg procedure to account for multiple comparisons. This was done to reduce the likelihood of false positives due to the large number of taxa tested. Visualization of significant taxa was further performed using the EnhancedVolcano package in R [27].

Keystone genera and core microbiome identification

To further delineate the community structure and ecological functions of the gut microbiome, we identified core microbiome and keystone genera by employing complementary statistical, computational, and graph-based approaches.

Keystone genera, which disproportionately contribute to the stability of the community, were identified through network centrality measurements based on co-occurrence networks. The co-occurrence network of microbes was constructed using pairwise Spearman correlation coefficients computed based on the relative abundance data for each genus across samples [28]. For a pair of genera , Spearman correlation coefficients was calculated as follows:

which and shows the ranks of the genus in the pair of in a sample , and is the total number of samples. To ensure that only biologically meaningful and statistically reliable associations were retained, correlations were filtered according to two criteria:

First, a correlation threshold of was set to retain positive associations. This is performed by constructing an adjacency matrix as:

Second, the corresponding p-values for each pairwise correlation was computed, and only associations with p < 0.05 were retained after multiple testing correction using the Benjamini–Hochberg FDR method.

The adjacency matrix was employed to construct the undirected co-occurrence network using the igraph package in R [29]. Consequently, three centrality metrics, such as degree, closeness, and betweenness, were taken to quantify the structural prominence of taxa as the nodes of the co-occurrence network [30, 31]. Degree centrality (CD) measures the number of direct connections to the node as the genus indicating its potential to interact with or influence the other taxa, was calculated via . While the Closeness centrality (CC), focus on the shortest path length between the node (genus) and indicates with , utilise to capture how efficiently a genus can spread information or effects throughout the network [31, 32]. Betweenness centrality (CB) value demonstrates the impact of the taxa as the bridge between different parts of the co-occurrence network, which facilitates interaction between otherwise distant microbial groups. This metric is calculated via that is the total number of shortest paths from the node to node , and is the number of those paths that pass through the node . Genera in the upper quartile for all these centrality measurements were identified as keystone taxa, indicating their pivotal roles in the stability of the microbial network and ecological balance. Afterward, files were introduced to Cytoscape 3.10.2 to visualize co-occurrence networks.

The core microbiome composition was determined through taxonomic profiling of all samples, identifying taxa that were persistently present in all samples and showed substantial abundance levels [30]. The abundance of a taxon was quantified as the fraction of samples in which it was found, and its relative abundance was measured by its total contribution to the community. Two metrics were used: (1) Prevalence of the genus calculated by , where is the relative abundance of the genus in sample , and is the indicator function. (2) Mean relative abundance of genus was derived from . Members of the core microbiome were considered if the genus had and . This dual-threshold approach ensured the inclusion of only widely distributed and ecologically valuable taxa. All abundance and prevalence calculations were estimated using the microbiome package [33].

Stochastic simulation of microbial dynamics

To explore microbial dynamics and understand the temporal behavior of taxa associated with CAD and control groups, we employed a stochastic logistic growth model [34]. This approach integrates both deterministic growth dynamics and stochastic fluctuations, capturing the inherent variability in microbial ecosystems and accounting for random perturbations in microbial behavior that may arise from factors like environmental shifts, immune responses, or metabolic changes in the host [35]. The dynamics of the taxon were modelled using the stochastic logistic growth equation:

where represents the abundance of taxon at time , shows the intrinsic growth rate of the taxon, reflecting how rapidly a taxon grows under optimal conditions, is the carrying capacity of the environment for a taxon , is the noise intensity, which reflects the level of the environmental fluctuation, and denotes Gaussian white noise with mean 0 and variance 1 that introduces stochastic variability. To incorporate pairwise interactions between taxa, we extended the model to include interaction terms between microbial genera. So, the updated equation is:

where represents the interaction coefficient between taxa and , capturing synergistic or antagonistic effects in terms of how the growth of one taxon may either enhance or suppress the growth of another. For numerical simulations, the Euler–Maruyama method (EM) was applied to discretize the equations [36]. This method is effective for handling the stochastic nature of the system and approximating the continuous dynamics:

where is a random variable sampled from a standard normal distribution ; That represents the stochastic fluctuations in the system. The simulation was conducted with the following parameters: time step () of 0.1, base growth rate () of 0.01, and carrying capacity () of 1,000. Biologically, the growth rate represents the intrinsic growth rate, capturing how fast a taxon proliferates under optimal conditions, while is the maximum sustainable population under available resources.

For the initial condition, OTU data columns with more than 80% zero values were excluded to avoid bias from the sparse taxa, resulting in a total of 74 taxa for both the CAD and control groups. The column-wise means of taxa abundances for CAD and control groups were used as initial values for the simulation. These initial conditions reflected the typical microbial composition in both groups before dynamic changes were simulated.

The simulation ran for 1000 time-steps, capturing the temporal evolution of microbial abundance in both the CAD and control groups. The simulated trajectories revealed distinct growth patterns for CAD-associated and control-associated taxa, with noticeable divergence between the two groups. Temporal abundance plots were generated to illustrate these differences, emphasizing the altered microbial community dynamics between health states.

Machine learning for diagnosis of the most important taxa

To identify the microbial taxa most relevant for diagnosing health status (CAD vs. control), we applied multiple supervised machine learning models [37]. The OTU data was first preprocessed to improve model performance and avoid sparsity issues; only numeric columns (taxa) with less than 80% zero entries across samples were retained. This filtering step ensured that rare taxa, which might introduce noise or bias, were excluded from the analysis [38].

After data preprocessing, model training was conducted using repeated fivefold cross-validation with 3 repeats. Stratified sampling was applied via the “createDataPartition” function of the caret package [39] to ensure that each fold maintained the same proportion of CAD and control samples, preserving class distribution consistency across folds. A fixed random seed (set.seed(42)) was used to guarantee reproducibility of the sampling process [38]. This stratification minimized sampling bias and ensured balanced representation of both groups during training and evaluation. To further address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied within each cross-validation fold to synthetically augment the minority class. The stratification helped to avoid sampling bias and ensured that both groups were equally represented during model training and evaluation.

We trained three different machine learning models for classification: Random Forest (RF), which is an ensemble method based on decision trees, known for its robustness to overfitting and ability to estimate feature importance. Support Vector Machine (SVM) is a powerful classifier effective in high-dimensional spaces, applied here with a radial basis function (RBF) kernel to capture non-linear patterns. Extreme Gradient Boosting (XGBoost), also, is a high-performance gradient boosting framework optimized for speed and accuracy, particularly well-suited for microbiome-based classification.

The models' performance was evaluated on the test set. We generated Receiver Operating Characteristic (ROC) curves and calculated the Area Under the Curve (AUC) for each model to compare their diagnostic accuracy quantitatively. Higher AUC values indicated better model performance in distinguishing between CAD and control groups. The AUC is defined as:

where is the true positive rate and is the false positive rate across different classification thresholds. Also, for more assessments of the classification quality, we computed the confusion matrix components. True Positives (TP) is used to show the correctly classified CAD cases, while the True Negatives (TN) show the correctly classified control cases. Also, False Positives (FP) introduce the misclassified control samples as CAD in comparison with the False Negative (FN) demonstrates the CAD cases misclassified as controls. Additionally, the following evaluation metrics were derived from the confusion matrix.

Accuracy is the overall proportion of correct predictions, which is defined as:

Additionally, Precision, Recall, and F1-score, which are known as the Positive Predictive Value, True Positive Rate, and harmonic mean of precision and recall, respectively, are:

To interpret dominant model output and identify key discriminatory features (taxa), feature importance scores were extracted. For RF, we computed the Mean Decrease in Gini Impurity for each feature, which quantifies how much each taxon contributes to improving the classification purity across the decision trees. This metric for a given taxon is defined as:

that is the set of tree nodes where is used for splitting, is the proportion of samples reaching the node , and is the decrease in Gini impurity resulting from the split:

so, the Gini impurity for a node is calculated as:

where is the proportion of the class samples at the node, and is the number of classes.

The top 10 most important taxa were selected based on their importance scores in the RF model. These taxa were considered critical microbial biomarkers for differentiating health status. For visualization, we used ggplot2 [40] in R to create a bar plot displaying the feature importance scores of these top taxa. This plot highlighted the most influential microbial signatures associated with CAD and healthy controls, providing an intuitive graphical summary of the diagnostic key features.

Results

Gut microbial diversity differed across CAD and healthy individuals

To assess the alterations in the microbial communities among different health statuses, we measured the alpha diversity (within samples) and beta diversity (between samples). The Shannon index of alpha diversity revealed that CAD patients' gut microbiota was richer compared to the control group (Fig. 2A). Statistical analysis also confirmed that the Shannon diversity was significantly different with the health status (P = 0.0000519; Pairwise Wilcoxon test). Beta diversity was also visualized by PCoA based on Bray–Curtis distance (Fig. 2B). The diversity captured by the top two principal coordinates was around 48%, which reflects extremely significant differences in the composition of the gut microbial community among CAD patients and controls. PERMANOVA based on distance matrices (Adonis) also displayed a significant contribution of the health status (30.6%, P = 0.001; Bray–Curtis). These results suggest that the diversity of gut microbiota is different in CAD patients compared to healthy people.

Fig. 2.

Gut microbial diversity differs across CAD and healthy individuals. (A) Boxplot comparing the Shannon diversity index of the gut microbiome between CAD patients (green) and healthy controls (blue). The CAD group exhibits a higher median Shannon index, suggesting greater microbial diversity compared to healthy controls. The box represents the interquartile range (IQR), the line within the box indicates the median, and the whiskers extend to 1.5 times the IQR. This observation highlights potential differences in microbial diversity between the two groups, which may reflect alterations in gut microbiome composition associated with coronary artery disease. (B) A PCoA plot visualizing microbial community differences between groups based on beta diversity metrics. Green dots represent the coronary artery disease (CAD) group, while blue dots denote the control group. The axes represent the two principal coordinates (PCo1 and PCo2) with their respective percentage variances

CAD and healthy individuals are different in microbial compositions

We subsequently explored the overall microbial composition among the two defined health statuses (CADs compared to healthy controls). Among the eight phyla identified, Firmicutes, Bacteroidetes, Proteobacteria, and Actinobacteria were most prevalent in terms of overall relative abundance, while Verrucomicrobiota, Fusobacteria, Desulfobacterota, and Patescibacteria had lower representation (Supplementary Table 2). Proteobacteria, Bacteroidota, and Verrucomicrobia were more enriched in healthy controls. However, the reverse trend was found for members from the phylum Actinobacteria, as these were the most abundant in CAD patients (Fig. 3A). Further, we did not find any difference in Firmicutes and Fusobacteria between the studied groups. At the genus level, the relative abundance of certain taxa such as Faecalibacterium, Blautia, Subdoligranulum, Streptococcus, Ruminococcus, Lactobacillus, Clostridium_sensu_stricto, Megamonas, Veillonella, Lachnoclostridium and Enterococcus (from Firmicutes), Bacteroides, Prevotellaceae_UCG.001, Prevotella, and Parabacteroides (from Bacteroidota), Escherichia, Delftia, and Klebsiella (from Proteobacteria), and Bifidobacterium (from Actinobacteria) were the most predominant gut microbial genera with the relative abundance higher than 0.01 (Supplementary Table 2).

Fig. 3.

Gut microbiota composition differs across CAD and healthy individuals. (A) Relative abundance of the healthy controls and CAD patients at the phylum level. (B) Stacked bar plots of relative abundances in the gut microbiome of CAD patients (left panel) and healthy controls (right panel) at the genus level. The CAD group exhibits greater microbial heterogeneity and instability, while the control group is more stable and balanced with predominant microbiota

The relative abundance stacked bar plots showed differences in gut microbiome composition between CAD patients and healthy controls. Microbial composition of CADs showed greater instability and variability across different samples (Fig. 3B and Supplementary Table 2). The genera such as Subdoligranulum, Lactobacillus, Prevotella, Prevotellaceae_UCG.001, Alloprevotella, Anaerovoracaceae, Delftia, Brevundimonas, Ralstonia, Desulfovibrio, Holdemanella, Pseudomonas, Weissella, Christensenellaceae_R.7_group, Bacillus, Acinetobacter, Catenibacterium, Romboutsia, and Staphylococcus were among the more dominant taxa in the gut microbiota of the patient with CADs (Supplementary Table 2). Nevertheless, bacterial genera such as Streptococcus, Enterococcus, Clostridium_sensu_stricto, Lachnoclostridium, Escherichia, Bacteroides, Faecalibacterium, Roseburia, Bifidobacterium, Ruminococcus, Fusicatenibacter, Blautia, Alistipes, and Paraprevotella occurred more frequently in control samples (Supplementary Table 2). Taken together, this suggests considerable differences in gut microbiome composition between CAD patients and healthy controls.

Additionally, the differential abundance analysis at the genus level showed that a higher relative abundance of certain taxa, such as Lactobacillus, Prevotella, Prevotellaceae_UCG.001, Alloprevotella, Anaerovoracaceae, Delftia, Brevundimonas, Ralstonia, Desulfovibrio, Holdemanella, Pseudomonas, Weissella, and Christensenellaceae_R.7_group, Bacillus, Acinetobacter, Catenibacterium, Subdoligranulum, and Staphylococcus were significantly associated with CAD patients, among other taxa (Fig. 4 and Supplementary Table 3). In contrast, the gut microbiota of healthy controls was significantly enriched with the genera such as Streptococcus, Enterococcus, Roseburia, Clostridium_sensu_stricto, Lachnoclostridium, Bifidobacterium, Escherichia, Ruminococcus, Bacteroides, Alistipes, Faecalibacterium, Fusicatenibacter, Blautia, and Paraprevotella (Fig. 4 and Supplementary Table 3). Several genera, including Delftia, Brevundimonas, Holdemanella, Pseudomonas, Weissella, Christensenellaceae_R.7_group, Bacillus, Acinetobacter, and Alloprevotella were among the most up-represented taxa in CAD microbiota (). Whereas, Anaerofilum, Paraprevotella, Bacteroides, Veillonella, Lachnospiraceae_ND3007_group, Lachnospiraceae_UCG.004, Erysipelatoclostridium, Fusicatenibacter, and Fusobacterium were over-represented in healthy controls with . These bacteria appear selectively favored in the gut microbiota of the healthy group compared to CAD patients.

Fig. 4.

Differential abundances of genera in gut microbiota of the control group vs. CAD patients. A volcano plot showing taxa significantly associated with health status. The x-axis represents the log2 fold change (|log2FC|> 1), and the y-axis represents the -log10 p-value. Taxa with higher representation in CADs and controls are red dots on the left and right sides of the x-axis, respectively. Dashed lines indicate the thresholds for significance, corresponding to a p-value of < 0.05 (horizontal line) and a log2 fold change of ± 1 (vertical lines)

Health status is represented by a small subset of gut core microbiota

To gain a deeper understanding of the ecological and functional contributions of gut microbial taxa to CAD and healthy control, we defined core microbiota and keystone genera according to prevalence-abundance cutoffs and network centrality metrics, respectively. Core microbiota, meaning taxa consistently represented at high abundance across samples, were recognized by cut points, including prevalence () and relative abundance (). The core microbiota analysis revealed characteristic patterns of common and unique taxa across different health statuses, which are demonstrated in Table 1. Genera including Bacteroides, Escherichia, Faecalibacterium, Streptococcus, and Blautia occurred in the gut microbiome of both healthy controls and CAD patients, indicating their core role in the gut microbiome system. Healthy individuals, however, encompassed genera such as Enterococcus and Clostridium_sensu_stricto in the core microbiota of their guts. The CAD core microbiota was characterized by an enrichment for Subdoligranulum and Bifidobacterium (Table 1). These results suggest that the core microbiota determining the health status of hosts is represented by a small subset of gut genera.

Table 1.

Core microbiota identified in control and CAD Groups

Group	Phylum	Genus	Mean abundance
CAD	Bacteroidota	Bacteroides	0.109742191
	Proteobacteria	Escherichia.Shigella	0.065102979
	Firmicutes	Faecalibacterium	0.047610689
	Firmicutes	Blautia	0.042626661
	Firmicutes	Subdoligranulum	0.037698711
	Firmicutes	Streptococcus	0.034171044
	Actinobacteriota	Bifidobacterium	0.031142071
Control	Bacteroidota	Bacteroides	0.251225313
	Proteobacteria	Escherichia.Shigella	0.108461759
	Firmicutes	Clostridium_sensu_stricto	0.052385875
	Firmicutes	Faecalibacterium	0.051069433
	Firmicutes	Blautia	0.048663454
	Firmicutes	Streptococcus	0.044944278
	Firmicutes	Enterococcus	0.032982619

Network analysis shows a disruption of microbial interactions and community stability in CAD

Keystone genera (hub taxa), conventionally regarded as central to the maintenance of stability and functionality of microbial communities, were determined by betweenness centrality (CB), closeness centrality (CC), and degree centrality (CD) of the co-occurrence network (Table 2). The CB counts the fraction of shortest paths going through a given bacterial taxon to another. The CB of a microbial taxon reflects the importance of control that the taxon exerts over the interactions of other taxa in the network. While CC denotes the proximity of a node to all other nodes, quantifying how many steps away genus is from all genera in a network [41]. Hence, taxa with high CC probably have a potential effect on the microbial community because they can rapidly affect other species. Table 2 demonstrates the high centrality measures in the gut microbiota and highlights the critical nodes in the stability of the microbial network. Notably, we detected a larger number of taxa with higher CB in the healthy controls compared to CADs. This represents keystone taxa with particularly strong interdependencies, which can provide microbial hemostasis for the healthy group. The genera such as Holdemania, Lachnospiraceae_UCG.010, Erysipelotrichaceae_UCG.003, Oscillospira, Romboutsia, Dialister, Alistipes, Monoglobus, and Subdoligranulum had higher CB within the gut microbial community of controls (Table 2; Fig. 5A and Supplementary Fig. 2). In the CAD group, however, the keystone taxa shifted toward a few genera like Lachnospiraceae_UCG.001, Oscillibacter, and UCG.003 with a higher CC, but lower CB compared to genera associated with controls (Table 2; Fig. 5B and Supplementary Fig. 3). The centrality findings in CAD gut microbiota illustrate that although a few bacterial genera are important to the overall connection in the network (high CC), they have probably lost their key roles as keystone (low CB), supporting a disruption of microbial interactions and community stability in CADs.

Table 2.

Centrality metrics of keystone genera in the gut microbiota of healthy and CAD Groups

Group	Phylum	Genus	Betweenness (CB)	Closeness (CC)	Degree (CD)
	Firmicutes	Subdoligranulum	0.11792126	0.420454545	22
	Firmicutes	Erysipelotrichaceae_UCG.003	0.163123967	0.456790123	22
	Firmicutes	Holdemania	0.217250422	0.411111111	10
	Firmicutes	Lachnospiraceae_UCG.010	0.163164433	0.456790123	20
Control	Firmicutes	Romboutsia	0.100264285	0.451219512	30
	Firmicutes	Monoglobus	0.073070476	0.46835443	28
	Firmicutes	Dialister	0.105105105	0.349056604	10
	Bacteroidota	Alistipes	0.078763752	0.486842105	30
	Firmicutes	UCG.003	0.084705171	0.435294118	16
	Firmicutes	Oscillospira	0.133498832	0.420454545	16
	Firmicutes	UCG.003	0.03894758	0.793103448	104
CAD	Firmicutes	Lachnospiraceae_FCS020	0.043087499	0.784090909	104
	Firmicutes	Oscillibacter	0.025911067	0.831325301	114
	Firmicutes	Lachnospiraceae_UCG.001	0.029707499	0.831325301	114

Fig. 5.

Network visualization of microbial co-occurrence patterns in the gut microbiomes of (A) healthy controls and (B) CAD patients. Nodes represent microbial genera, with hub nodes identified based on degree centrality (degree ≥ 50), indicating highly connected taxa within the network. Keystone genera are highlighted in red. Nodes colored green correspond to genera with positive associations predominantly observed in control samples, while blue nodes represent genera with positive associations primarily found in CAD patients. Edges denote significant correlations between genera, illustrating the community structure and interaction patterns in each group. The network was constructed and visualized using Cytoscape

Stochastic simulation reveals altered microbial dynamics in CADs

To assess the temporal dynamics of gut microbial communities under different health conditions, we applied a stochastic differential equation (SDE) model to simulate microbial abundance dynamics in CAD and control groups. This approach captures both deterministic ecological processes and inherent stochasticity reflective of simulated biological systems. Specifically, the simulations incorporated 74 taxa for both the CAD and control groups. These taxa were selected by excluding those with > 80% zero values, ensuring a focus on taxa with sufficient prevalence for analysis.

The simulations were performed by using the Euler–Maruyama method, a numerical integration technique suitable for SDEs [36]. Each group’s microbial community was modelled over 1000 discrete time steps, capturing average genus-level abundance dynamics. The CAD-associated taxa showed a slower initial growth rate but ultimately reached higher total abundances compared to control-associated taxa (Fig. 6). To quantitatively assess the differences in initial abundance between CAD-associated and control-associated taxa, we calculated the mean initial abundance values. The taxa associated with control individuals exhibited a significantly higher mean initial abundance (60,215) compared to those associated with CAD patients (45,911). This difference was statistically significant (p = 0.0175, Student’s t-test), indicating a lower initial abundance and altered microbial dynamics in CAD-associated taxa. This dynamic shift may reflect ecological perturbations within the gut environment of CAD patients, potentially driven by systemic inflammation, altered nutrient flux, or dysregulated host-microbe interactions, which are common factors in cardiovascular diseases [42]. The enhanced steady-state abundance in the samples derived from the CAD patients probably indicates the competitive dominance of inflammation-adapted taxa or reduced microbial turnover. These findings highlight the effectiveness of stochastic modelling in uncovering nonlinear and emergent behavior in microbiome ecosystems that may not be captured through static analyses alone.

Fig. 6.

Stochastic simulation of microbial abundance dynamics in CAD and control groups using the Euler–Maruyama method. The curves represent average taxa abundance over 1000 time-steps, incorporating logistic growth and stochastic perturbations. The CAD group exhibits a consistently higher abundance trajectory, reflecting altered microbial growth patterns compared to controls

Machine learning uncovers the most predictive taxa associated with CAD

To evaluate the diagnostic potential of gut microbial taxa in differentiating CAD patients from healthy controls, we employed three machine learning models, including RF, SVM, and XGBoost. The detailed confusion matrices for each model on the full dataset (providing counts of true positives, true negatives, false positives, and false negatives) are presented in Table 3, highlighting the models' classification patterns. All models were trained using OTU relative abundance data. The performance of the machine learning models was then evaluated via a set of classification metrics, including AUC, accuracy, precision, recall, F1-score, and Cohen’s kappa coefficient.

Table 3.

The confusion matrix of trained ML models on controls and CAD samples used in this study

Model	Health status	Reference
		Control	CAD
RF	Control	32	5
	CAD	9	192
SVM	Control	30	5
	CAD	11	192
XGBoost	Control	32	6
	CAD	9	191

The outputs of the models demonstrate that RF achieved the highest performance across most of the evaluation metrics, with an AUC of 0.951, an accuracy of 0.941, a precision of 0.943, a recall of 0.986, an F1-score of 0.964, and a kappa of 0.810, which can show excellent agreement between predicted and true labels. The XGBoost model also performed well, with an AUC of 0.892, an accuracy of 0.934, and an approximately balanced trade-off between precision and recall, resulting in an F1-score of 0.960 and a kappa value of 0.790. The SVM model, while performing slightly lower than RF and XGBoost, still achieved robust results, with an AUC of 0.891, accuracy of 0.908, precision of 0.940, recall of 0.945, F1-score of 0.942, and kappa of 0.721 (Table 4). These results were further visualized in Fig. 7A, where the receiver operating characteristic (ROC) curve illustrates the trade-off between true positive rate (sensitivity) and false positive rate (1-specificity). Among the three models, the ROC curve for RF is closest to the top-left corner of the plot, indicating its superior ability to discriminate between CAD patients and controls.

Table 4.

The performance metrics of three trained ML models

Model	AUC	Accuracy	Precision	Recall	F1_Score	Kappa
RF	0.951	0.941	0.943	0.986	0.964	0.810
SVM	0.891	0.908	0.940	0.945	0.942	0.721
XGBoost	0.892	0.934	0.942	0.977	0.960	0.790

Fig. 7.

(A) Receiver Operating Characteristic curves comparing the diagnostic performance of three machine learning models- Random Forest (RF), SVM, and XGBoost-trained on OTU data to classify CAD patients and controls. RF demonstrates a higher sensitivity and specificity with an ROC curve (top-left corner of the plot). This suggests that the RF model has the best balance between true positive rate (sensitivity) and false positive rate (specificity) compared to SVM and XGBoost for distinguishing CAD patients from healthy controls. (B) Top 10 microbial genera identified as important features for distinguishing CAD patients from healthy individuals using the RF model. The importance score, derived from the mean decrease in accuracy, highlights Holdemanella as the most predictive genus, followed by Acinetobacter, and Fusicatenibacter. These Genera are ranked in descending order of their contribution to the model's performance

To further interpret the predictive capacity of the gut microbiota, a feature importance analysis derived from the RF model was conducted to identify the top 10 potential taxa for differentiating between the two healthy conditions (Fig. 7B). Holdemanella was ranked as the most important predictive genus, followed by Acinetobacter and Fusicatenibacter. These genera exhibited the highest importance scores, which can highlight their strong potential for classifying the health status. Other notable genera included Sutterella, Agathobacter, Brevundimonas, Pseudomonas, Subdoligranulum, TM7x, and Delftia. These taxa displayed high importance scores based on the Mean Decrease in Accuracy (MDA) metric, which quantifies each feature’s contribution to the classifier’s performance by measuring how much accuracy drops when the feature is randomly permuted [43]. The prominence of these genera suggests their potential as biomarkers for CAD, highlighting the gut microbiome’s role in cardiovascular health and its diagnostic potential.

Validation data

Within the secondary population groups, as illustrated in supplementary Fig. 4, we present a detailed breakdown of the varying microbial compositions at the genus level. CAD patients showed a higher Shannon diversity compared to controls (p-value = 4.73e-15; Supplementary Fig. 4A), and bacterial communities were separated based on healthy status (R2 = 0.04214, p-value = 0.001; Supplementary Fig. 4B). Noteworthy stochastic analysis also confirmed the higher abundance trajectory in the CAD group, reflecting altered microbial growth patterns compared to controls (Supplementary Fig. 4C). Evaluation of three machine learning approaches, including RF, SVM, and XGBoost, for the dataset further demonstrated an excellent agreement between predicted and true labels for all three approaches (AUC > 0.8; Supplementary Fig. 4D). These insights contribute to our understanding of the gut microbial composition and structure within the patients compared to healthy controls.

Discussion

Gut microbiota dysbiosis in patients with CAD

The current study characterized gut microbial alterations in 217 CAD patients compared to 57 healthy controls through advanced bioinformatics analysis. Our integrated approach, combining taxonomic profiling, co-occurrence network analysis, stochastic modelling, and machine learning, revealed significant CAD-associated microbial signatures.

First, we found that the gut microbiota in patients with CAD is significantly perturbed, with an elevated microbial diversity, distinct structure, and discrepant composition at the phylum level [44]. Beta diversity analysis using Bray–Curtis distance and PCoA plot revealed significant differences in the overall composition of the gut microbial community between CAD patients and healthy samples. These differences were observed along the first two principal components in the PCoA plot, indicating profound changes in the gut microbiome structure of CAD patients compared to controls. The finding is in line with previous studies that have reported similar changes in the gut microbiome structure of patients with CAD [45, 46]. The difference in the mean of the Shannon index reinforces the fact that the gut microbiota in CAD patients was more diverse compared with the healthy controls. Although an increased microbial diversity is often considered a positive indicator of microbiome health [47–49], a higher Shannon diversity in CAD patients may reflect the presence of unstable or potentially harmful species resulting from dysbiosis [50]. This suggests that increased microbial diversity in inflammatory and metabolic conditions may be a consequence of disrupted microbiome balance. The simulation of stochastic signatures also revealed a divergence in microbial dynamics in CADs compared to healthy controls. The composition of a healthy individual’s gut microbiome was relatively stable over time. However, the CAD-associated microbiota showed a slower initial growth rate but ultimately reached higher total abundances compared to control-associated taxa. This dynamic shift may reflect ecological perturbations within the gut environment of CAD patients, potentially driven by inflammation or altered metabolic states.

Shifts in microbiota composition are a signature of CAD

Although CAD patients shared a large number of common bacteria with controls, certain bacteria were differently enriched in CAD individuals. A comprehensive comparison between CAD patients and controls regarding the relative abundance of taxa at the phylum level demonstrated the enrichment of Actinobacteria and the reduction of Bacteroidetes and Proteobacteria in CAD compared to the control group. The enrichment of Actinobacteria in CAD patients has been reported previously [1, 4, 51]. Actinobacteria include nitrate-producing Gram-positive bacteria that are associated with cardiometabolic risks [52], retinal artery occlusion [53], obesity and adipogenesis [54], colorectal cancer [55], and depression [56]. In numerous studies, Bacteroidetes decline in the gut microbiome has also been associated with CAD [4, 6, 16], making the decrease in these taxa strikingly CAD-related. The Bacteroidetes in the gut microbiome can carry profound health beneficial effects on the human body and play an essential role in maintaining a healthy gut microbiota. For instance, Bacteroidota can rapidly adapt their carbohydrate metabolism to the available nutrients due to their genetic plasticity, providing benefits to survive and flourish in the gut tract. They also produce SCFAs, which have wide-ranging and many beneficial influences on the host. Bacteroidota can also produce many proteases, thereby facilitating protein metabolism [57]. Moreover, via bile acid metabolism, Bacteroidota play a role in colonization resistance of pathogen organisms, and maintenance of gut integrity [58].

Gut bacterial composition differed in CAD patients compared with healthy controls in the relative abundance of Proteobacteria and Firmicutes; the results were, however, not homogeneous when compared with other studies. For instance, we observed a decreased trend in Proteobacteria in patients with CAD. This result is in agreement with some previous studies [15, 59], while an increased abundance of Proteobacteria has also been reported [1, 51, 60]. One possible explanation is that the Proteobacterial community may shift with the severity of CAD in different participants. Hu et al. [15], compared the structure of the gut microbiota community among three groups based on stenosis in CAD patients, including the healthy controls, participants with evidence of stenosis less than 50% (LT50), and individuals with evidence of stenosis greater than 50% (GT50). They found that the relative abundances of Proteobacteria decreased significantly in individuals with LT50 compared with controls, but not in GT50, indicating that gut bacterial composition may be influenced by the severity of CAD. In addition, dietary, medical, and lifestyle differences, as well as autoimmune disease among participants (e.g., among healthy controls), may have contributed towards heterogeneity between studies [61–63].

Regarding the Firmicutes phylum, there was no significant increase in this phylum across the CAD patients and healthy controls in the current analysis. This comes in line with the findings of Sawicka-Smiarowska et al., [51]. However, most bacteria of the Firmicutes often show an increased abundance in CAD patients [6, 15, 59]. Although we didn’t observe significant changes in the relative abundance of Firmicutes at the phylum level, the most common alterations in CADs were related to the lower taxonomic levels of Firmicutes. For example, the relative abundance of Bacilli was significantly increased in the CAD patients. The high abundance of Firmicutes/Bacilli connects to the indicators of unhealthy lifestyles, such as a high-fat diet [64], abnormal energy balance [65], and obesity [66], which can lead to CAD.

Differential analysis at the genus level also highlighted the significant differences in the abundance of specific microbial taxa between CAD patients and the healthy group. This result comes in line with previous reports [1, 4, 51, 67–70] and supports the notion that the beneficial bacteria are often reduced, whereas the potential pathogens are more abundant in CAD patients than in the healthy participants [71, 72]. For example, members of the Clostridium genus, especially Clostridium sensu stricto, as well as species such as Blautia, and Faecalibacterium prausnitzii are crucial for fermenting dietary fibers, providing energy for intestinal cells, regulating immune function, reducing inflammation, and helping prevent colonization by pathogenic bacteria, all of which contribute to a healthy gut microbiome [73–75]. Increasing abundance of Fusobacteria and Bilophila was also associated with a decreased likelihood of developing atherosclerosis [76]. Collectively suggests that a homeostatic and more protective microbiota exists in healthy controls.

In contrast, the potential pathogens such as Delftia, Lactobacillus, Prevotella, Alloprevotella, Brevundimonas, Ralstonia, Acinetobacter, Pseudomonas, Weissella, and Staphylococcus were more abundant in the CAD group than in the healthy participants. For instance, Lactobacillus reuteri, Delftia, and Acinetobacter are associated with higher HDL [77], obesity [78] and increased mortality in patients with cardiovascular issues, respectively. There is also a case report of infective endocarditis caused by Brevundimonas vesicularis [79]. The Holdemanella levels of gut microbial abundances have the potential to identify individuals with atrial fibrillation [80] and CAD [62].

A recent study has found that high-fat diet-induced obesity could disrupt gut microbiota homeostasis with increases in the abundance of Brevundimonas, Desulfovibrionaceae_unclassified and Ralstonia, ultimately leading to the overproduction of lipopolysaccharides [64]. Further, the growth of Prevotella has been associated with high cardiovascular disease risk [81]. Safarchi et al., [50] demonstrated that a disrupted Bacteroides/Prevotella ratio, along with a reduction in beneficial microbes like Faecalibacterium, is linked to conditions such as IBS and Crohn's disease, potentially increasing inflammation. A depletion in the abundance of Faecalibacterium and Bifidobacterium and an over-presentation of Prevotella in patients with primary hypertension is one of the most important risk factors for CAD [82]. Another study focusing on atherosclerosis risk factors indicated that hypercholesterolemic patients were enriched with Prevotella and a decrease in Clostridium [83]. This finding partly supports our results regarding the altered microbial composition of CAD patients, as we observed three OTUs from Clostridium that were decreased and two OTUs from Prevotella that increased in the CAD group compared with the controls. Collectively indicate that maintaining a balanced gut microbiome is crucial, as dysbiosis in gut microbiota can lead to systemic diseases, highlighting the complex relationship between harmful microbes and the host's immune response [50].

Complex microbial interactions influence overall health outcomes

Assessing the individual abundance of core microbiota at the genus level showed that a small subset of gut core microbiota was correlated with health status, including Subdoligranulum, Bifidobacterium, Enterococcus and Clostridium_sensu_stricto. Other genera, including Bacteroides, Escherichia, Faecalibacterium, Streptococcus, and Blautia, were consistent with the results of the core microbiome analysis between the studied groups, suggesting large contributions of these lineages in gut microbiome architecture. In healthy individuals, the specific core microbiota included Enterococcus and Clostridium_sensu_stricto, which can be associated with maintaining gut barrier integrity and reducing inflammation. Enterococcus has fundamental roles in immune system modulation, breaking down food components, enhancing nutrient absorption, and preventing various gastrointestinal diseases [84, 85]. However, Enterococci, particularly Enterococcus faecalis and Enterococcus faecium, are opportunistic pathogens capable of causing serious infections such as urinary tract infections, endocarditis, and bacteremia [85, 86]. In contrast, the CAD core microbiota was enriched with genera such as Subdoligranulum and Bifidobacterium, which, although beneficial, may not fully compensate for the loss of other stabilizing taxa. Bifidobacterium play a beneficial role in gut health by maintaining intestinal homeostasis, producing metabolites like organic acids that lower pH and inhibit pathogens, and synthesizing vitamins that support overall health. Likewise, the Subdoligranulum, though less studied, may influence metabolic pathways and appetite regulation, potentially impacting muscle health in older adults, particularly with sarcopenia. There is evidence that highlights the importance of these genera in health and disease [87–89]. It seems that members of the core microbiome in both healthy and CAD groups may come from different taxonomic and functional backgrounds. They may be beneficial and/or potentially detrimental, but thrive and decline together, exhibiting co-abundance behaviour.

It is worth noting that in the current study, members of the core microbiome comprised taxa whose prevalence () and relative abundance (). Notably, although high prevalence has been widely used as a criterion for the core microbiome [90], it is speculated that the keystone taxa organizing a community architecture, or providing a key gain-of-function, can be a low-abundance component and is often not easily detected by current conventional analyses. This indicates that the prevalent gut microbiota members do not equally determine the host health status, and a stable interaction can be critical in defining health-relevant microbiomes.

The co-occurrence network for healthy controls further illustrated the robust and interconnected nature of the microbial community. The keystone genera, such as Subdoligranulum, Romboutsia, Dialister, and Alistipes, exhibited a high degree and CB, reflecting their strong connectivity and influence within the microbial community of healthy controls. For instance, the significant suppression of Subdoligranulum relative abundance in the gut microbiota of CAD patients contributes to an imbalanced and unstable microbiome [1, 91]. The positive roles of Romboutsia in metabolism and host health through the breakdown of complex components into simpler metabolites, such as SCFAs, can be associated with a healthy microbial balance [92]. Similarly, Dialister species are recognized for their diverse metabolic pathways, including the fermentation of complex carbohydrates into beneficial SCFAs, which are known for their anti-inflammatory properties and ability to modulate immune responses [93, 94]. The reduction in Dialister levels in patients with acute myocardial infarction indicated an imbalance in the gut microbiota and increased inflammation [94]. Regarding the genus Alistipes, there is a complicated association between Alistipes and heart-related conditions, showing both beneficial and harmful impacts. Although some species of Alistipes, including Alistipes finegoldii, are related to elevated blood pressure and gut dysbiosis, leading to cardiovascular issues, other research indicates that the anti-inflammatory effects of butyrate generated by Alistipes may lower the risk of heart disease by improving endothelial function and decreasing arterial stiffness [95, 96].

A shift in keystone taxa toward genera such as Lachnospiraceae_UCG.001, Lachnospiraceae_UCG.003, and Oscillibacter in CAD patients suggests a disruption in microbial interactions and community stability. Some research has highlighted the involvement of these bacterial genera in various disease conditions, including obesity, inflammatory bowel disease (IBD), and metabolic disorders [97–100]. The co-occurrence network for the CAD group further confirmed a less interconnected and more fragmented microbial community. The replacement of key stabilizing genera with taxa that may not fulfil the same functions may contribute to the dysbiosis observed in the gut microbiota of CAD individuals. This dysbiosis may further exacerbate systemic inflammation and metabolic dysregulation, hallmarks of CAD pathogenesis. Hence, the more stable healthy controls maintain a more consistent network of keystone taxa. In contrast, the less stable ones can be the contributors to CAD in terms of inflammatory or metabolic derangement, as reported in previous studies [101–103].

Gut microbiota acts as a biomarker of CAD

The stochastic simulations, based on an SDE framework discretized via the Euler–Maruyama method, captured the disrupted microbial growth trajectories and interaction dynamics associated with the gut microbiota of CAD patients compared to healthy individuals. This can suggest that the ability of the studied approach to reproduce ecologically plausible patterns and distinguish disease states in 16S rRNA-based microbiome datasets, which is consistent with the earlier probabilistic and stochastic approaches [104, 105]. For instance, a temporal probabilistic model developed for 16S rRNA data showed an overdispersion and zero inflation in bacterial communities of human gut microbiota [106]. Similarly, Xu et al. [104, 105], extended a generalized Lotka-Volterra framework to incorporate stochastic environmental fluctuations in microbial communities and could accurately detect temporal microbial variability. In addition, our finding from stochastic modelling aligns, in part, with observations by Zhang et al. [107], who reported that inflammatory microenvironments can reshape microbial competitive dynamics and lead to resilience shifts. Nevertheless, alternative stochastic models, including agent-based frameworks, have suggested that chronic perturbations may induce prolonged instability rather than steady-state shifts, highlighting model-dependent variability [37, 108]. Such discrepancies could arise from differing assumptions regarding immune feedback loops, nutrient fluxes, and spatial heterogeneity, which emphasize the need for caution when extrapolating stochastic predictions to clinical outcomes [1, 109, 110].

The machine learning classifiers also discriminated CAD status based on microbial profiles. Machine learning analyses corroborated the diagnostic potential of microbial signatures, with Random Forest achieving the highest predictive accuracy. This is consistent with previous microbiome-based diagnostic efforts in other diseases [42, 111]. Nevertheless, performance variability across classifiers, particularly the relatively lower AUC for XGBoost, mirrors Topçuoğlu's observations [112] regarding model sensitivity to sparsity and feature noise. This suggests that classifier selection remains a critical factor when translating microbiome-based predictions into clinical practice [38]. Notably, while certain taxa like Holdemanella have emerged consistently in predictive models and prior studies as potential CAD-associated microbes, their exact role remains controversial. For instance, although Holdemanella has been proposed as a risk-associated genus in CAD [62, 63, 113, 114], some studies have found no significant abundance differences between CAD patients and controls [115]. Additionally, dietary habits may modulate its association with CAD, and recent evidence suggests its involvement in differential responses to statin therapy (a group of medicines that can help lower the level of LDL in the blood), raising the possibility of its use in personalized treatment planning [63]. These findings collectively highlight the need for longitudinal and interventional studies to clarify causal relationships, validate predictive models in independent cohorts, and explore therapeutic implications.

Strengths and limitations

By employing a reliable machine learning approach and combining the potential of 16S rRNA gene sequencing with simulated stochastic analysis, our study sets itself apart. The incorporation of core microbiome and co-occurrence network interpretations contributes to providing a distinctive perspective on the investigation of symbiotic relationships within the complexity of the gut environment. The addition of the validation data in the studied models did not change the results, indicating that our findings were robust. However, one notable limitation is the variability in the microbiome among diverse populations, which could potentially impact our findings. Consequently, the observed results may, in part, be influenced by variations in environmental and laboratory factors between the groups of individuals affected by CAD, as well as by the relatively smaller dataset when compared to other studies on the gut microbiome. Another limitation of this study is that key confounding variables, including BMI, sex, age, hypertension, and diabetes history, were not systematically collected and included as covariates in the differential abundance analysis. Consequently, the potential influence of these demographic factors on microbiome composition and disease outcomes remains unaccounted for. Future studies with more comprehensive data collection are needed to clarify these interactions.

Conclusion

This study provides compelling evidence for the significant association between gut microbiota dysbiosis and CAD. By integrating microbial communities in CAD patients and healthy individuals, we identified distinct patterns of microbial alterations that characterized CAD-associated microbiota. Our findings demonstrated that CAD is associated with specific changes in gut bacterial diversity and composition, indicating a dysbiosis in patients with CAD. Simulation of stochastic signatures also revealed a divergence in microbial dynamics in CADs compared to healthy controls. These changes in microbial community included the depletion of beneficial bacterial taxa and the enrichment of potentially harmful microorganisms. Assessing the individual abundance of core microbiota at the genus level showed that a small subset of gut core microbiota was correlated with different health status, suggesting that the keystone taxa organizing a community architecture, or providing a key gain-of-function, can be a low-abundance component. The co-occurrence network further illustrated the robust and interconnected nature of the microbial community in healthy controls. Whereas, the CAD group showed a less interconnected and more fragmented microbial community, supporting the microbiota dysbiosis. We then introduced the potential of gut microbial signatures as novel biomarkers for CAD risk assessment and diagnosis. The identification of these microbial patterns opens new possibilities for developing microbiota-targeted interventions in CAD prevention and management. Although the number of studies involving the gut microbiome and CAD is increasing, different ethnicities [116], diets [61], and geographical locations [117] in gut microbial composition and CAD development are still rare and need to be expanded in the future. Moreover, we need to know more about whether microbial alterations precede CAD onset and whether microbiota modulation can influence disease progression. These findings contribute to the growing recognition of the gut microbiome as an important factor in cardiovascular health and disease.

Authors’ contributions

Zahra Alaeddini: Methodology, Formal analysis, Data curation, and Writing. Iman Nemati: Writing- original draft, Resources, and Conceptualization. Somayeh Gholizadeh: Writing- review & editing, Project administration, Conceptualization.

Funding

No funding was received for this study.

Data availability

This study is a secondary analysis of data from previously published works. The original datasets that support the findings of this study have been deposited in the European Nucleotide Archive under the primary accession codes PRJDB7456, PRJNA810651, PRJNA984163, PRJNA779598, and PRJNA503710. All scripts are also available in https://github.com/Zahra-Alaeddini/CAD.

Declarations

Ethics approval and consent to participate

This article does not involve any new studies of human or animal subjects performed by any of the authors.

Consent for publication

Not applicable.

Competing interest

The authors declare no competing interests.