Multi-omics integration reveals functional signatures of gut microbiome in atherosclerosis

ABSTRACT

Atherosclerosis (AS), a predominant contributor to global cardiovascular disease burden, exhibits complex interplay with gut microbiota dysbiosis. While the associations between microbial imbalance and AS pathogenesis are well-documented, the pathophysiological mechanisms governing microbe-host crosstalk remain incompletely characterized. Current research limitations stem from methodological heterogeneity across studies and the absence of consensus regarding disease-specific microbial signatures. In this study, we conducted an integrated multi-omics analysis to characterize the functional signatures of gut microbiome in AS. We collected all public AS-related 6 microbiome datasets and 8 peripheral blood host transcriptomic datasets from across the world, comprising 456 metagenomic samples and 111 16S rRNA gene sequencing samples for microbial profiling, alongside 118 RNA-Seq samples and 302 microarray samples. We systematically characterized AS microbial taxa and computationally inferred the metabolic potential for the gut microbiome using metabolomic-related data. Metabolite-host gene interactions were further predicted based on the synergistic effects between microbiome and host transcriptome in AS. Five “microbe-metabolite-host gene” tripartite associations related to AS were identified involving 5 microbial genera (Actinomyces, Bacteroides, Eisenbergiella, Gemella, and Veillonella), 2 metabolites (Ethanol and H₂O₂), and 2 host genes (FANCD2 and GPX2), and the reliability of these associations was validated. Five microbial genera demonstrated robust diagnostic potential as noninvasive biomarkers, with 5-fold cross-validation, study-to-study transfer validation, and leave-one-study-out (LOSO) validation confirming good diagnostic performance. Additionally, the specificity of the biomarkers was validated against hypertension, inflammatory bowel disease (IBD), diabetes, and obesity cohorts. Our study unveiled the functional characteristics of gut microbiota interacting with AS host genes and highlighted the potential of gut microbiota as both diagnostic biomarkers and therapeutic targets for AS. However, the findings should be interpreted considering the inherent heterogeneity of the integrated datasets and the preliminary diagnostic value of the biomarkers.

Introduction

Atherosclerosis (AS), a leading cause of cardiovascular diseases worldwide, is characterized by lipid accumulation within the arterial wall and systemic chronic inflammation.¹ It is associated with high morbidity, disability, and mortality rates, with an increasing trend among younger populations.² The gut microbiota has been demonstrated to be intricately linked to host health and has been implicated in the pathogenesis of various diseases, including AS. Substantial evidence has established its regulatory capacity in metabolic homeostasis through microbial-derived metabolite production, modulating critical pathways including immunomodulation, lipid cycling, and glucose metabolism.³ Therefore, elucidating how gut microbiota and their metabolites influence host gene expression patterns in AS is essential for understanding their interactions in AS pathogenesis.

Current investigations into gut microbiota in AS have generated substantial empirical evidence.^4–14 In animal models, germ-free mouse studies have provided insights into the causal relationship between gut microbiota and AS. For instance, one study, raising Apoe^−/− mice, observed that the impact of the gut microbiota on AS is dietary dependent.⁴ Kiouptsi et al. rederived low-density lipoprotein receptor-deficient (Ldlr-/-) mice, and demonstrated a functional role for the commensal microbiota in atherothrombosis.⁵ A comprehensive review discussed the links between the gut microbiota, its metabolites and thromboembolic diseases, concluding that several thromboembolic diseases are associated with an altered gut microbial composition.⁶ In human studies, comparative analyses of microbial communities have demonstrated significant compositional alterations in AS patients, marked by diminished relative abundances of Bacteroides and Prevotella alongside elevated proportions of Streptococcus and Escherichia.⁷ While clinical observations identified a positive correlation between the plaque scores of AS patients and Prevotella copri colonization levels,⁸ contrasting evidence from shotgun metagenomic sequencing revealed reduced Prevotella copri prevalence in AS cohorts.⁹ These contradictory observations may be attributable to inter-cohort variability in genetic predisposition, environmental exposures, and nutritional profiles, presenting challenges for cross-population validation of microbial signatures. Mechanistic investigations have revealed that intestinal dysbiosis facilitates AS pathogenesis through multiple metabolic pathways. Emerging evidence suggests that microbial activity contributes to AS progression mediated by microbial-derived metabolites including trimethylamine-N-oxide (TMAO),¹⁰ bile acids (BAs),¹¹ serum indole-3-acetic acid,¹² and lipopolysaccharides (LPS).¹³ Conversely, protective microbial metabolites such as indole-3-aldehyde have been shown to attenuate inflammatory responses through immunometabolic regulation.¹⁴ However, the precise mechanisms by which gut microbiota and metabolites interact to regulate host gene expression and influence AS development remain incompletely understood.

Currently, microbiome abundance profiles and host transcriptome expression datasets related to AS have been accumulated. However, the availability of AS metabolomic abundance profiles remains limited. Nevertheless, metabolomic-related data can be accessed through established repositories. For instance, MetaCyc represents a non-redundant, experimentally validated database of metabolic pathways and enzymatic reactions,¹⁵ whereas STITCH systematically integrates both experimentally confirmed and computationally predicted chemical-protein interactions.¹⁶ These resources collectively enable systematic investigation of “microbe-metabolite-host gene” interactions.

In this study, we conducted a comprehensive integration of AS multi-omics data and systematically characterized the functional signature of gut microbiota in AS. We collected all publicly available AS fecal microbiome datasets and peripheral blood host transcriptomic datasets, including metagenomic data (n = 456), 16S rRNA sequencing (n = 111), RNA-Seq data (n = 118) and microarray datasets (n = 302). We inferred microbe-metabolite relations and their molecular links to host genes using metabolomic-related data. We also verified the reliability of the identified associations. Additionally, this study validated a panel of noninvasive biomarkers for their potential in AS diagnosis and evaluated their diagnostic specificity. The workflow of this study is presented in Figure 1.

Figure 1.

Workflow of the present study. Step1. Collection and analysis of AS-related microbiome data including metagenome and 16S rRNA gene data. Differentially abundant genera (DAGs) and microbial genes (DEMGs) were identified and only those exhibited concordant directions of change across both data types were retained. Step2. Collection and analysis of AS-related host transcriptome data including RNA-Seq and microarray data. Differentially expressed genes (DEGs) were identified and only those exhibited concordant directions of change across both data types were retained. Step3. Deciphering of functional signatures of AS-related microbiome. (1). Microbiome analysis including alpha/beta diversity analyses, community composition analyses, and phylogenetic tree analyses. (2). Exploring microbial influence on host via metabolites. Firstly, based on the DEMGs, metabolic potential of microbiome was inferred using metabolic reactions from the MetaCyc database, and each metabolite was assigned a PRMT score. Simultaneously, based on the DAGs, leave-one-genus-out (LOGO) analysis was performed to identify genera that significantly drive metabolite changes. Secondly, based on the DEGs, the disease-promoting or disease-mitigating effects of metabolites associated with them were assessed using STITCH database and DM score. Thirdly, the metabolites with their DM scores align in sign with the PRMT scores were retained, and the potential microbe-metabolite-host gene associations were identified. We further validated these associations through literature verification and revealed potential diagnostic biomarkers.

Materials and methods

Data acquisition

Human gut microbiome and transcriptome data associated with AS were systematically retrieved from public databases in October 2023. Detailed procedures for data acquisition and filtering can be found in Supplementary Text. Finally, 6 microbiome datasets were obtained, comprising 249 AS and 207 control samples from metagenomic datasets, along with 51 AS and 60 control samples from 16S rRNA datasets. Furthermore, 8 host transcriptome datasets were included, consisting of 64 AS and 54 control samples from RNA-Seq datasets, and 163 AS and 139 control samples from microarray datasets. These datasets were geographically diverse, originating from research institutions across Asia, America, and Europe. Detailed sample collection procedures are provided in Supplementary Table S1.

Processing of microbiome datasets and differential abundant analysis

Metagenomic and 16S rRNA raw sequencing data underwent independent processing pipelines for quality control, taxonomic profiling, and functional profiling. Resulting microbial taxonomic profiles were aggregated at the genus level. Batch effect correction was applied to both taxonomic and functional profiles (For details, see the Supplementary Text).

Differentially abundant genera (DAGs) and differentially expressed microbial genes (DEMGs) were identified using DESeq2 and the R package “limma”, respectively. Fold-change values and p-values were extracted from each dataset and subjected to random-effects meta-analysis using the MetaDE package.¹⁷ DAGs and DEMGs meeting the threshold criteria of p-value < 0.05 and absolute combined effect size (ES, representing log₂ fold change) > 1¹⁸ were selected for subsequent analyses. Notably, 16S rRNA sequencing and metagenomic data were analyzed independently. DAGs and DEMGs that were consistently present and exhibited concordant directions of change across both data types were retained for further analysis.

Inferring metabolic potential of as fecal microbiome

Metabolic potential of AS fecal microbiome was inferred through using the identified DEMG. We used enzyme Commission (EC) numbers to extract enzymatic reaction data from the MetaCyc database,¹⁵ encompassing substrates, products, and reaction reversibility information. The analysis was restricted to irreversible reactions, and compounds present on both the substrate and product sides were incorporated to establish a potential metabolite catalog. The relative accumulation or depletion of these metabolites in AS patients versus controls was assessed using the predicted relative metabolic turnover (PRMT) method.¹⁹ The detailed calculation process for the PRMT score is described in the Supplementary Text. It reflects the predicted relative turnover of metabolites within the host, with $\mathrm{PRMT}>\mathit{\hspace{0.17em}}0$ indicating accumulation and $\mathrm{PRMT}<\mathit{\hspace{0.17em}}0$ indicating depletion in AS group compared to controls.

A leave-one-genus-out (LOGO) analysis approach was employed to evaluate whether the metabolites are regulated by gut microbe (For details, see the Supplementary Text). A positive z-score indicates that the genus contributed to the relative enrichment of the metabolite in AS versus controls by promoting the biosynthesis of the metabolite or inhibiting its degradation. Conversely, a negative z-score signifies that the genus contributed to the depletion of the metabolite in AS by inhibiting the biosynthesis of the metabolite or facilitating its degradation.

Processing of host transcriptome datasets

Host transcriptome data (microarray and RNA-Seq) underwent platform-specific processing including genomic alignment (GRCh38/hg38), expression quantification, and batch correction. Differentially expressed genes (DEGs) were identified using limma (p-value < 0.05 and an absolute combined ES value > 1). A random-effects meta-analysis (MetaDE) integrated results across datasets (For details, see the Supplementary Text). Concordant DEGs exhibiting consistent directional changes in both microarray and RNA-Seq platforms were retained for further analysis.

Inferring metabolic effects in AS host transcriptome

The information regarding metabolite-host gene interactions with known functional effects (activation or inhibition) and with confidence score exceeding 0.9 were extracted from the STITCH database.¹⁶ These interactions were subsequently filtered to retain only DEGs identified in the host transcriptome. As previously reported,¹⁸ a metabolite may influence disease progression by activating upregulated genes or inhibiting downregulated ones, thereby promoting disease, or by inhibiting upregulated genes or activating downregulated ones, thereby mitigating disease effects. Accordingly, the disease-modifying effects of metabolites on the host transcriptome were evaluated using disease-modifying score calculated through formula $\left(1\right)$:

$$\mathit{DM}=\frac{1}{m}\sum _{\mathit{p}=1}^m\mathit{E}{S}_p-\frac{1}{n}\sum _{\mathit{q}=1}^n\mathit{E}{S}_q\left(1\right)$$

here, $\mathit{DM}$ represents the predicted impact of a metabolite on the host transcriptome, with positive values indicating a disease-promoting effect and negative values reflecting a disease-mitigating effect. $\mathit{ES}$ denotes the combined effect size of host target genes associated with the metabolite. The parameters $\mathit{m}$ and $\mathit{n}$ correspond to the number of genes activated or repressed by the metabolite, respectively.

Performance evaluation of diagnostic biomarkers

In this study, the random forest (RF) model was employed as a diagnostic model, and its classification performance was evaluated using the area under the receiver operating characteristic curve (AUROC), determined via 5-fold cross-validation. Furthermore, comprehensive validation frameworks were implemented, encompassing: (1) Leave-one-study-out (LOSO) validation, involving n iterations where each study cohort was sequentially designated as the test cohort while the remaining n-1 cohorts served as training datasets; and (2) Study-to-study transfer validation, wherein models were initially trained on individual study cohorts followed by evaluation across all remaining independent external cohorts.

To validate the specificity of the diagnostic biomarkers for AS, we selected 4 distinct disease conditions for comparative analysis: hypertension, type 2 diabetes, obesity (all known to be associated with AS risk), and inflammatory bowel disease (IBD) (due to its intestinal relevance). The relevant datasets were retrieved from the gutMDisorder v2.0,²⁰ yielding the following case-control cohorts: 152 cases versus 44 controls for hypertension (PRJEB13870), 63 cases versus 50 controls for type 2 diabetes (PRJNA448494), 152 cases versus 105 controls for obesity (PRJEB12123), and 20 cases versus 20 controls for IBD (PRJEB7949). According to a previous study,²¹ false positive rate (FPR) was used to evaluate the specificity of the diagnostic biomarkers.

Results

Taxonomic and functional profiles of AS fecal microbiome

In this study, 6 microbiome datasets spanning 2 distinct types were analyzed (Table 1), notably confounded by variations in sample origins, sequencing methods, and platforms. These datasets were collected from 4 countries across Asia and Europe (Supplementary Figure S2A). To ensure data quality, we applied strict filtering criteria and removed low-quality samples (Supplementary Figure S2B). We conducted principal component analysis (PCA) at both the genus and species levels to evaluate the impact of these factors on data heterogeneity. As shown in Supplementary Figure S3, at the species level, PCA of all metagenome samples revealed clear separation by study (PERMANOVA $R^2$ = 0.008, p-value = 0.02), country (PERMANOVA $R^2$ = 0.008, p-value = 0.02), and sequencing platform (PERMANOVA $R^2$ = 0.004, p-value = 0.035), while PCA at the genus level showed no remarkable distinction between the groups (p-value > 0.05). All 16S rRNA datasets were generated using identical sequencing platforms and targeted the same hypervariable region, thereby reducing potential data heterogeneity. PCA of all 16S samples revealed clear separation by study and country at both the genus and species levels. However, the cross-dataset heterogeneity was alleviated when analyzed at the genus level. Therefore, subsequent analyses were conducted at the genus level. Although this approach lowered data resolution, it helped reduce heterogeneity.²² Additionally, the average annotation rate of OTUs in the 16S rRNA datasets at the genus level was 61.8%, whereas the average annotation rate at the species level was relatively low, averaging 24.3% (Table 2).

Table 1.

Summary of the datasets used in this study.

Dataset	Sample type	Country	Platform	Data type	No.of AS	No.of Control
PRJNA177201	Fecal	Sweden	Illumina HiSeq 2000	Metagenome	12	13
PRJEB21528	Fecal	China	Illumina HiSeq 2000	Metagenome	214	171
PRJNA615842	Fecal	Italy	Illumina HiSeq 2500	Metagenome	23	23
PRJNA595382	Fecal	Italy	Illumina MiSeq	16S, V3-V4	10	20
PRJDB6472	Fecal	Japan	Illumina MiSeq	16S, V3-V4	30	30
PRJDB7456	Fecal	Japan	Illumina MiSeq	16S, V3-V4	11	10
PRJNA251404	Peripheral blood	USA	Illumina Genome Analyzer llx	RNA-Seq	8	8
PRJNA679890	Peripheral blood	China	Illumina HiSeq 3000	RNA-Seq	5	4
PRJNA821540	Peripheral blood	China	Illumina NovaSeq 6000	RNA-Seq	4	3
PRJNA957897	Peripheral blood	USA	Illumina HiSeq 2500	RNA-Seq	20	12
PRJNA836853	Peripheral blood	Italy	ION_TORRENT	RNA-Seq	27	27
GSE20129	Peripheral blood	USA	GPL6104/ GPL10558	Microarray	49	86
GSE23746	Peripheral blood	USA	GPL2700	Microarray	76	19
GSE37356	Peripheral blood	USA	GPL10558	Microarray	38	34

Table 2.

Taxonomic information of OTU annotations in 16S rRNA datasets.

Dataset	Data type	No.of OTU	No.of genus	No.of species
PRJNA595382	16S	1862	1715(92.1%)	670(36.0%)
PRJDB6472	16S	916	100(10.9%)	29(3.17%)
PRJDB7456	16S	3848	3175(82.5%)	1298(33.7%)

Inter-study batch effects were significantly reduced through the application of the ConQuR method for batch effect correction at the genus level in both metagenomic and 16S rRNA sequencing datasets (Supplementary Figure S4A-D). Microbial alpha diversity, evaluated using Shannon and Simpson indices, demonstrated significantly elevated diversity in AS patients compared to healthy controls within both 16S rRNA and metagenomic datasets (Figure 2A). Taxonomic analysis identified 1,074 genera in metagenomic datasets and 269 genera in 16S rRNA datasets, with 120 genera overlapping between both platforms. These shared genera exhibited significant correlations in relative abundance profiles between metagenomic and 16S rRNA data across disease and control groups, indicating consistency in microbial community characterization (Supplementary Figure S5A,B). Beta diversity analysis, based on Bray-Curtis dissimilarity metrics, revealed significant intergroup differences (PERMANOVA p-value = 0.001, Figure 2B,C). In addition, as demonstrated in Supplementary Figure S6A,B, the relative abundance distribution of genera differed between case and control samples in both metagenome and 16S datasets, with Bacteroides being more abundant in controls. These findings indicated that the microbial community was altered in AS patients, potentially contributing to the occurrence and progression of AS.

Figure 2.

Alterations in the as gut microbiome. (A) Alpha diversity of the as (red) and control (blue) groups in metagenomic and 16S rRNA datasets, assessed using the two-sided Wilcoxon rank-sum test. Significance levels are indicated as follows: ns, p-value > 0.05; *, p-value ≤0.05; **, p-value ≤0.01; ***, p-value ≤0.001; ****, p-value ≤0.0001. (B) and (C) Principal coordinate analysis (PCoA) based on Bray-Curtis distances, demonstrating significant differences in microbial composition between the as and control groups in the metagenomic and 16S rRNA datasets, respectively (p-value = 0.001 for each, calculated by two-sided PERMANOVA test with 999 permutations). (D) the upper bar indicates the 156 genera with significant differences identified in the meta-analysis of metagenomic datasets, along with their corresponding phylum classifications. Among these, 6 genera are highlighted in gray, indicating consistent directional changes observed in both metagenomic and 16S rRNA datasets. The middle bar displays the log₂FC of the 156 significant genera, with 86 genera enriched in as shown in red and 70 genera depleted in as shown in blue. The lower heatmap presents the significance levels of these genera at the genus level in individual metagenomic datasets (gray) and their corresponding log₂FC values (color).

For microbial genes, inter-study batch effects in metagenome and 16S rRNA datasets were separately corrected using the Combat method, substantially reducing variability (Supplementary Figure S4E-H). For the metagenome gene datasets, we performed functional analysis using HUMAnN3, identifying 2,359 EC gene families. For the 16S rRNA datasets, gene prediction was conducted using PICRUSt2, yielding 2,082 EC gene families, with 1,711 overlapping those from the metagenome. The relative abundances of 1,711 genes in the metagenomic datasets were highly correlated with those inferred from the 16S rRNA datasets in both case and control samples, indicating consistency in functional analysis across the 2 data types (Spearman’s R ≥ 0.76, Supplementary Figure S5C, D).

Alterations on the fecal microbiome in AS versus controls

For the 16S rRNA datasets, a total of 57 genus-level taxa were identified as significantly different, while 156 genus-level taxa exhibited significant differences in the metagenomic datasets (Supplementary Table S2). The majority of these taxa belonged to the phylum Firmicutes (Supplementary Figure S7), underscoring its predominant role in the observed microbial shifts. Among these, 6 DAGs (Actinomyces, Eisenbergiella, Gemella, Lactobacillus, Veillonella and Bacteroides) demonstrated consistent directional changes across both the metagenome and 16S rRNA datasets. As illustrated in Figure 2D and Supplementary Figure S6C, Actinomyces, Eisenbergiella, Gemella, Lactobacillus, and Veillonella were significantly enriched, whereas Bacteroides was significantly depleted in AS. Additionally, upon analyzing individual cohorts, inconsistencies were observed, with certain bacterial genera being significantly upregulated in one cohort but either non-significant or displaying opposing trends in another. This variability likely arises from cohort-specific heterogeneity, emphasizing the potential limitations of single-study analyses.

Furthermore, DEMGs associated with AS were identified. In the metagenomic dataset, 1,192 DEMGs were identified, while 41 genes met the same criteria in the 16S rRNA dataset (Supplementary Table S2). Notably, 20 DEMGs exhibited consistent alterations across both data types, underscoring their robust association with disease-related changes. These genes were subsequently utilized for further analysis. Additionally, we examined the contribution of each genus in DAGs to these genes using the metagenomic datasets (Supplementary Figure S8C,D and Supplementary Figure S9). Consequently, Escherichia, Streptococcus, and Klebsiella emerged as the primary contributors to the 16 upregulated DEMGs, which were found to be enriched in AS patients. Conversely, Bacteroidetes were identified as the major contributors to the 4 downregulated DEMGs, which were depleted in AS patients.

Metabolic potential of AS fecal microbiome

Metabolic potential of AS fecal microbiome was inferred based on enzymatic reaction data from the MetaCyc database and using the PRMT method (for details, see Materials and Methods). A total of 84 metabolites that participated in enzymatic reactions with the 20 DEMGs were obtained. To ensure consistency with the metabolite IDs in the STITCH database for subsequent extraction of metabolite-host associations (see Materials and Methods), the metabolite IDs were further converted using the PubChem and ChEBI databases, resulting in a final set of 53 metabolites. Based on these 53 metabolites, 837 metabolite-target interactions with documented functional effects were obtained, including 29 metabolites and 590 human host genes from the STITCH database. As depicted in Supplementary Figure S8A, among the 29 identified metabolites, Orotic acid is included in a group of serum metabolites that have been reported to significantly enhance the predictive accuracy for AS.²³ Additionally, among metabolites enriched in AS, Hydrogen peroxide (H₂O₂) exhibited the highest number of interactions with host genes. Conversely, among metabolites depleted in AS, Ethanol demonstrated the most interactions with host genes. These findings underscore their potential significance in AS-related gene regulation and pathogenesis.

Furthermore, we conducted LOGO analysis to evaluate the impact of 6 DAGs on the accumulation or depletion of metabolites. For each metabolite, the contribution of each genus was quantified as a z-score index, which was calculated by isolating its influence from the PRMT score (for details, see Materials and Methods). As depicted in Supplementary Figure S8B, 6 genera were found to significantly contribute to 26 metabolites, with Bacteroides exhibiting predominant influence across most metabolites. Observations indicated that Bacteroides was significantly depleted in AS, and it promoted the consumption of H₂O₂ (z-score < 0), while H₂O₂ exhibited accumulation in AS (PRMT > 0). Moreover, among the 5 DAGs affecting Ethanol, 4 genera (Actinomyces, Eisenbergiella, Gemella, and Veillonella) were significantly enriched in AS, and they facilitated the consumption of Ethanol (z-score < 0), whereas Ethanol was depleted in AS (PRMT < 0, Supplementary Figure S8A).

Additionally, based on metagenomic data, we used 2 DEMGs (EC-3.5.1.24 and EC-1.3.3.4) as examples to further confirm the associations between genus and metabolite. Following Ning et al.’s method,²¹ multi-omics biological correlation (MOBC) maps were constructed. As shown in Supplementary Figure S8C, EC-3.5.1.24 was depleted in AS patients (Supplementary Table S2). Based on the enzymatic reaction, this reduction inhibited Cholic acid and Glycine accumulation, a finding consistent with prior observations of decreased levels of these metabolites in AS patients (Supplementary Figure S8A). It has been reported that Bile acid-activated TGR5 signaling may play a crucial role in preventing AS,²⁴ and Glycine deficiency has been found to exacerbate hypercholesterolemia and AS progression.²⁵ Notably, Bacteroides was the most significant contributor to EC-3.5.1.24, and its marked depletion in AS aligns with the prior LOGO analysis findings, indicating that Bacteroides facilitated the consumption of Cholic acid and Glycine (Supplementary Figure S8B). Similarly, EC-1.3.3.4 was elevated in AS patients (Supplementary Table S2), as depicted in Supplementary Figure S8D; according to the enzymatic reaction, this elevation promoted the accumulation of H₂O₂, which is consistent with the results showing that H₂O₂ was enriched in AS patients (Supplementary Figure S8A). H₂O₂ is known to accelerate the progression of AS.²⁶ On the other hand, among the top 5 contributors to EC-1.3.3.4, Escherichia was the largest contributor to EC-3.5.1.24 and was enriched in AS patients. Bacteroides was the second largest contributor (Supplementary Figure S8D). As shown in Supplementary Figure S8B, Bacteroides was depleted in AS and promoted the consumption of H₂O₂ (z-score < 0), while H₂O₂ was enriched in AS patients. These findings collectively reinforce the reliability of our predictive framework.

Assessing metabolic effects in human transcriptome

To evaluate metabolic effects in the human transcriptome, 5 RNA-Seq datasets and 3 microarray datasets were analyzed (Table 1). These datasets were derived from 3 countries spanning Asia, America, and Europe (Supplementary Figure S2A). Similarly, rigorous filtering criteria were applied to ensure data quality (Supplementary Figure S2C). To address inter-study batch effects, the Combat method was applied separately to each data type. Prior to correction, samples exhibited notable study-specific clustering (Supplementary Figure S10A,C). Post-correction, batch effects were markedly reduced, as evidenced by enhanced interspersing of samples across studies (Supplementary Figure S10B,D). We next identified genes significantly altered in AS. As a result, a total of 3,729 DEGs were identified within the RNA-Seq datasets, while 834 DEGs were detected in the microarray datasets (Supplementary Table S2). Of these, 84 DEGs demonstrated consistent directional changes across both data types, highlighting their potential relevance to AS pathophysiology.

Based on 29 metabolites identified above, 590 human host genes were found to interact with them through STITCH database, of which three were DEGs. Subsequently, the three DEGs were utilized to calculate the DM score for inferring metabolic effects in the host transcriptome (for details, see Materials and Methods). To enhance the biological interpretability of the results, it is assumed, based on existing literature,¹⁸ that accumulated metabolites promote disease progression, while consumed metabolites alleviate disease progression, both of which are associated with disease onset and progression. Therefore, the assessment of metabolites required their DM scores to align in sign with the PRMT scores. This process ultimately resulted in the selection of 2 metabolites: H₂O₂ and Ethanol. H₂O₂ was predicted to have a PRMT score of 0.64 in the AS microbiome. Correspondingly, it exhibited a potential disease-promoting effect (DM score = 2.08) by inhibiting GPX2, which was significantly downregulated in the host transcriptome. Conversely, Ethanol was predicted to have a PRMT score of −2.74 in the AS microbiome. Consistently, it demonstrated a potential disease-mitigating effect (DM score = −2.10) by activating FANCD2, which was significantly downregulated in the host transcriptome. These findings suggest a synergistic interaction between the microbiome and the host in AS pathogenesis.

Potential “microbe-metabolite-host gene” associations in AS

Metabolites were assessed using the aforementioned microbial and host transcriptomic data, and the potential AS-related “microbe-metabolite-host gene” associations were identified. Under the hypothesis-driven constraint requiring directional congruence between the DM score and the PRMT score, this culminated in the identification of 2 metabolites: H₂O₂ and Ethanol. Based on the results of LOGO analysis, the genera exerting significant effects on H₂O₂ were identified as Bacteroides. For Ethanol, 4 genera (Actinomyces, Eisenbergiella, Gemella, and Veillonella) were determined to have significant effects. Ultimately, 5 “microbe-metabolite-host gene” interactions were identified, comprising 5 genera, 2 metabolites, and 2 DEGs (Figure 3).

Figure 3.

The potential microbe-metabolite-host gene associations. (A) An overview of the potential microbe-metabolite-host gene associations. (B) Detailed information on these associations.

Validation of “microbe-metabolite-host gene” associations

Among the 5 identified “microbe-metabolite-host gene” associations, some were partially elucidated through MOBC maps. Further validation of these associations was conducted via a systematic literature review. As summarized in Table 3, almost all nodes within these associations were validated to be associated with AS. For instance, Bacteroides, a core genus of the human gut microbiome, has shown a decrease in relative abundance in patients with AS.⁷ Furthermore, research conducted by Hirata et al. demonstrated that Bacteroides vulgatus and Bacteroides dorei, which are depleted in AS, can alleviate AS in mice by reducing Endotoxin levels.²⁷ Similarly, Veillonella, another key genus of the gut microbiota, has been recognized as a biomarker for coronary AS.²⁸ For the 2 metabolites, Ethanol at an appropriate dose exerts protective effects against endothelial senescence and AS by activating ALDH2.²⁹ In contrast, H₂O₂ accelerates AS progression by inducing endothelial cell damage and triggering inflammatory signaling pathways, such as the activation of NF-κB, thereby exacerbating arterial wall inflammation.²⁶ For the host gene GPX1, an isoenzyme of GPX2, has been demonstrated to mitigate aortic lesions in ApoE-deficient mice.³⁰ Fancd2 (the mouse homolog of human FANCD2) is suggested to indirectly inhibit AS by enhancing the tolerance and immunoregulatory capacity of CD4LAP Tregs, which are critical for preventing AS through the suppression of pro-inflammatory responses.³¹

Table 3.

Literature evidence for the identified microbe-metabolite-host gene associations.

Type	Node/Edge	Type of evidence	Evidence (PMID)
Genus	Actinomyces	Indirect	39262241
	Bacteroides	Direct Direct Indirect Indirect	32989686 29018189 30571343 31726978
	Eisenbergiella	-	-
	Gemella	Indirect	34869642
	Veillonella	Direct Indirect Indirect	37220524 31726978 31812509
Metabolite	Ethanol	Direct	30885430
		Direct	37286970
		Direct	11067787
		Direct	24582196
		Direct	31906033
	H2O2	Direct Direct	16009356 32245238
Host gene	FANCD2	Indirect	37689092
	GPX2	Indirect	20848490
Genus-Metabolite	Actinomyces – Ethanol	-	-
	Bacteroides – H2O2	Direct	24164536
	Eisenbergiella – Ethanol	-	-
	Gemella – Ethanol	-	-
	Veillonella – Ethanol	-	-
Metabolite-Host gene	Ethanol – FANCD2	Indirect	21919919
	H2O2 – GPX2	Indirect	25261240

We further validated the regulatory relationships between certain nodes. As shown in Table 3, using Bacteroides-H₂O₂-GPX2 association as an example. Previous studies have demonstrated that Bacteroides species scavenge reactive oxygen species (ROS) including H₂O₂, under aerobic conditions.³² GPX2, a H₂O₂ reductase, has been shown to play a key role in alleviating oxidative stress by reducing H₂O₂ levels.³³

Potential AS fecal microbiota diagnostic biomarkers

Based on the identified “microbe-metabolite-host gene” associations, we further investigated the potential utility of these nodes as diagnostic biomarkers for AS. Prioritizing noninvasive diagnostic markers, we focused on 5 microbial genera (Actinomyces, Bacteroides, Eisenbergiella, Gemella, and Veillonella). Using the random forest model and 5-fold cross-validation, we found that these 5 microbial genera exhibited robust diagnostic performance for AS across all the microbiome datasets, achieving an average AUROC of 0.87 (Supplementary Figure S11A). To further assess the transferability and applicability of these microbial features for AS diagnosis, study-to-study transfer validation and LOSO validation were performed. As shown in Supplementary Figure S11A and S11B, the study-to-study transfer validation yielded a mean AUROC value of 0.74, indicating robustness and generalizability of the microbial features across different study cohorts. Additionally, the LOSO validation achieved a mean AUROC of 0.77 (Supplementary Figure S11B). These findings collectively highlight the potential utility of these microbial features as diagnostic biomarkers for AS in diverse populations.

Furthermore, the specificity of these microbial diagnostic biomarkers was evaluated (for details, see Materials and Methods). Four distinct disease conditions, including hypertension, IBD, type 2 diabetes, and obesity, were investigated. The results demonstrated relatively low FPR values across these disease cohorts. As illustrated in Supplementary Figure S11C, study-to-study transfer validation yielded a mean FPR value of 0.20, while LOSO validation produced a mean FPR of 0.36 (Supplementary Figure S11D). These findings collectively suggest that the microbial diagnostic biomarkers demonstrate good specificity.

Discussion

To investigate the impact of gut microbiota on AS, this study presented a large-scale integrative analysis on AS public multi-omics datasets encompassing AS fecal metagenomic profiles, metabolomic-related data, and peripheral blood transcriptomic profiles, and characterized the functional signatures of gut microbiome in AS. Through the application of computational approaches, we identified “microbe-metabolite-host gene” associations and revealed novel noninvasive microbial diagnostic biomarkers for AS.

The analysis of AS gut microbiome datasets revealed increased gut microbiota alpha diversity (Shannon and Simpson indices) in AS patients. This observation contrasts with the prevailing paradigm that associates healthy gut microbiota with greater ecological diversity. While a metagenomic study reported increased alpha diversity in unstable angina patients,³⁴ another study reported non-significant diversity variations between AS patients and controls.⁷ Furthermore, reduced alpha diversity has been observed in IBD patients.²¹ These inconsistent results suggest that diversity alone can’t fully explain microbial community differences between healthy and diseased individuals.³⁴ Future studies should focus on specific microbial taxonomic changes and their links with host responses to better understand the impact of microbiota on disease development and progression.

We identified 5 potential “microbe-metabolite-host gene” associations and assessed their reliability. Previous studies have demonstrated that gut microbiota and their metabolites influence AS development. For instance, Wang et al. found that three pro-atherogenic gut microbiota metabolites (choline, TMAO, betaine) suppress gut microbiota in AS-prone mice and inhibit choline-driven AS.³⁵ Additionally, Li et al. found that Akkermansia muciniphila can mitigate Western diet-induced inflammation and AS in Apoe^−/− mice.³⁶ Our findings indicate that Ethanol may exert a protective effect on AS, which aligns with recent large-cohort epidemiological studies.^37,38 Notably, Ethanol’s beneficial effects on AS exhibited significant dose dependency: moderate consumption may correlate with reduced cardiovascular disease risk, whereas excessive intake may exacerbate AS progression and elevate cardiovascular disease risk.³⁷ It should be emphasized that this study does not advocate alcohol consumption but aims to examine potential factors associated with the relationship between Ethanol and AS. The complex interplay between Ethanol and AS needs further experimental validation.

The identified “microbe-metabolite-host gene” associations encompassed 5 genera, 2 metabolites, and 2 DEGs. Diagnostic utility of these 5 genera was evaluated by employing a random forest model and three validation strategies: within-study validation (5-fold cross-validation), study-to-study transfer validation, and LOSO validation. The 5 microbial genera exhibited satisfactory diagnostic performance. A recent study³⁹ recognized 80 fecal bacterial and fungal features to differentiate AS patients from controls via a random forest model, achieving an AUROC of 0.87. The classification effectiveness was comparable with our diagnostic biomarkers, which had an average AUROC of 0.87 (ranging from 0.73 to 1.00), even though we only had 5 bacterial genera. We further examined the diagnostic potential of other molecules implicated in the interactions. Due to the absence of metabolite profiles, the diagnostic utility of the 2 DEGs was examined. Given that certain studies have limited samples in the host gene transcriptomic data, leave-one-out cross-validation (LOOCV) was implemented within individual studies both for the 2 DEGs and the 5 microbial genera to ensure methodological consistency. As illustrated in Supplementary Figure S12, host genes were inferior to microbial biomarkers in diagnostic utility. This may be due to the constant interaction between microbes and metabolites/host genes weakening disease-related signals. However, this doesn’t mean host genes are universally less diagnostically efficient than microbial biomarkers. Moreover, the microbial biomarkers demonstrate noninvasive properties and their potential associations with pathophysiological processes can be further investigated through the identified “microbe-metabolite-host gene” tripartite interactions. However, these identified biomarkers and associations require further functional validation – for instance, through in vivo strain colonization experiments using gnotobiotic mice, coupled with integrated profiling of paired gut microbiome, targeted/untargeted metabolome, and host transcriptome datasets.

Through systematically integration of AS-related fecal microbiome, metabolomic-related data and peripheral blood transcriptome datasets, we identified the microbial functional signatures in AS and identified 5 potential “microbe-metabolite-host gene” associations involving 5 microbial genera, 2 metabolites, and 2 human host genes. Furthermore, we revealed novel disease-specific noninvasive microbial diagnostic biomarkers for AS. This study provides novel insights into microbe-host interactions in AS and offers a feasible framework for investigating similar interactions for other diseases using unpaired multi-omics datasets. It is important to acknowledge, however, that the interpretation of these findings should consider the study’s limitations outlined below.

Firstly, the integration of multiple AS-associated microbiome and host transcriptome datasets introduced inherent technical variability and batch effects stemming from geographic disparities and heterogeneous sequencing methodologies. We implemented inter-study batch effect correction and restricted microbiome analyses to the genus level, which partially mitigated data heterogeneity. However, this approach limited us to investigate dysbiosis and functional characteristics of the microbiome at higher resolution levels. Secondly, although we identified “microbe-metabolite-host gene” associations potentially implicated in AS pathophysiology, these findings represent correlative relationships rather than causal mechanisms. Thirdly, to enhance the reliability of our results, we focused on intersecting data across multiple sources. While this stringent filtering improved data quality and result robustness, it also reduced the number of associations ultimately identified. Finally, several confounders – including BMI, age, dietary habits, and statin use⁴⁰—may bias microbiome-host interactions. These factors significantly influence both microbiome composition and AS risk. Adjustment or stratified analyses were not performed owing to limited sample availability with accessible metadata. Future studies incorporating such data for appropriate covariate adjustment would strengthen the findings. Collectively, these limitations inherent to the integrational dataset and study design necessitate cautious interpretation of the identified associations and the preliminary diagnostic value of the proposed biomarkers.

Funding Statement

This work was supported by the National Natural Science Foundation of China [62222104, 62172130], Heilongjiang Postdoctoral Fund [LBH-Q20030]. The funders had no role in study design, data collection and analysis, interpretation, and writing of the report.

Acknowledgments

Liang Cheng, Hongbo Shi and Guangde Zhang designed the research. Hongbo Shi, Meitao Wu, and Xiaoliang Wu performed the research and analyzed the results. Meitao Wu, Xiaoliang Wu, Zhuoxin Liu, Shuo Jiang, Gen Li, Yetong Yang, Yanghe Fu and Qiuping Wang analyzed the data. Hongbo Shi and Meitao Wu wrote the paper. All authors read and approved the final paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

All datasets used in this study can be found in online repositories. The names of the repository and accession numbers can be found in the article.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19490976.2025.2542384