This research was performed based on the prospective specimen collection and retrospective blinded evaluation design principles (Pepe et al., 2008). We recruited 264 AF patients who were hospitalized in Henan Provincial Chest Hospital and Guangshan County People’s Hospital from May 1, 2021 to April 30, 2023 (Figure 1). All enrolled patients received standard treatment according to “Current knowledge and management of atrial fibrillation: consensus of Chinese experts 2021” (Chinese Society of Pacing and Electrophysiology, Chinese Society of Arrhythmias, Atrial Fibrillation Center Union of China, 2022). After rigorous inclusion and exclusion criteria, 529 samples were included for 16S rRNA MiSeq sequencing. Among them, samples from AF patients treated with catheter ablation were collected before the procedure and 6 months after the procedure. Clinical data and participants’ demographics were collected and analyzed (Table 1). Supplementary material showed inclusion, exclusion, and detailed diagnostic criteria of the participants. The study conformed to the principles of the Declaration of Helsinki. The research protocol was approved by the ethics committee of Henan Provincial Chest Hospital (2021-05-05). All participants signed informed consent.

Before taking tongue-coating samples, participants used sterile water to rinse their mouths twice. The posterior middle to anterior middle area of the tongue coating was scraped by a professional operator with a pharyngeal swab. The swab was immediately placed into a cryotube, the virus was inactivated at 56°C for 30 min and then the sample was transferred to the freezer at −80°C.

The sample was suspended in 790 μL of sterile lysis buffer (4 M guanidine thiocyanate; 10% N-lauroyl sarcosine; 5% N-lauroyl sarcosine-0.1 M phosphate buffer [pH 8.0]) in 2 mL screw-cap tube containing 1 g glass beads (0.1 mm BioSpec Products, Inc., United States). This mixture was vortexed vigorously and then incubated at 70°C for 1 h. After incubation by bead beating for 10 min at maximum speed. DNA was extracted by following the manufacturer’s instructions for bacterial DNA extraction using The E.Z.N.A. ® Stool DNA Kit (Omega Bio-tek, Inc., GA), which excepted lysis steps and stored at −20°C for further analysis.

The primers F1 and R2 (5′-CCTACGGGNGGCWGCAG-3′ and 5′-GACTACHVGGGTATCTAATCC-3′) correspond to positions 341 to 805 in the Escherichia coli 16S rRNA gene were used to amplify the V3-V4 region of each sample by PCR. PCR reactions were run in a T100™ Thermal Cycler PCR system (Bio-Rad Laboratories, Inc., United States) using the following program: 3 min of denaturation at 95°C followed by 21 cycles of 0.5 min at 94°C (denaturation), 0.5 min for annealing at 58°C, and 0.5 min at 72°C (elongation), with a final extension at 72°C for 5 min.

The amplicons from different samples were purified by Hieff NGS® DNA Selection Beads (YeasenBiotechCo., Ltd., China). The products were indexed and mixed at equal ratios for sequencing by Shanghai Mobio Biomedical Technology Co., Ltd. using the Miseq platform (Illumina Inc., United States) according to the manufacturer’s instructions. Raw Illumina read data for all samples were deposited in the National Center for Biotechnology Information (PRJNA1061363).

The non-repetitive sequences were extracted in the optimized sequences, which facilitated the reduction of redundant computations in the intermediate process of the analysis. ¹ Single sequences without repeats were removed. ² Operational taxonomy unit (OTU) clustering of non-repeated sequences (excluding single sequences) was performed according to 97% similarity, and chimeras were removed in the clustering process to obtain the OTU representative sequences. All the optimized sequences were mapped to the representative sequences of OTUs, and the sequences with 97% or more similarity to the representative sequences of OTUs were selected to generate a table of OTUs. The OTU representative sequences with a 97% similarity level were taxonomically analyzed at each taxonomic level: phylum, class, order, family, genus, and species were counted for the community composition of each sample. The comparative databases were 16S bacterial and archaeal ribosomal databases Silva² (Release 138). ³

Alpha diversity metrics (Ace estimator, Chao 1 estimator, Shannon-Wiener diversity index, and Simpson diversity index) were assessed by using Mothur v1.42.1. Both Bray-Curtis and weighted and unweighted UniFrac dissimilarity were calculated in QIIME. PCoA, NMDS plots, Adonis, and ANOSIM, which were applied to test for statistical significance between the groups, were generated in the R (version 3.6.0) package vegan 2.5–7. The linear discriminant analysis (LDA) effect size (LEfSe) ⁴ was used to detect taxa with differential abundance among groups.

The discovery OTU frequency profile, validation frequency profile, and independent diagnosis frequency profile were generated by mapping reads from the discovery cohort, validation cohort, and independent diagnosis cohort against these represented sequences, respectively. Then, we used the Wilcoxon rank-sum test to determine the significance (p < 0.05), and OTU biomarkers in the tongue coating microbiomes were selected for further analysis. We constructed a diagnostic model through fivefold cross-validation on a random forest model and evaluated the probability of disease (POD) index and receiver operating characteristic curve. The process was performed as we described in Supplementary material.

PICRUSt2 v2.4.1 ⁵ was used to predict functional abundances based on 16S rRNA gene sequences. Differences between the two groups were compared using Student’s t-test for normal continuous variables, the Wilcoxon rank-sum test for non-normal continuous variables, and the χ2 test or Fisher’s exact test for categorical variables. Differences among groups were assessed by using one-way analysis of variance for normal continuous variables and the Kruskal-Wallis test for non-normal continuous variables. Differences with a p-value < 0.05 (two-sided) were considered statistically significant. We used a multiple comparison false positive correction. To control for the false positive rate associated with multiple comparisons, we corrected the p-values of all comparisons for false discovery rate (FDR). Specifically, we used the Benjamini-Hochberg process for FDR correction. This method first sorted all p-values according to the original significance level and then calculated a threshold for each p-value in the sorted order to control the FDR at a predetermined level. Statistical analyses were performed using IBM SPSS Statistics 26 for Windows (IBM Corp., Armonk, New York).

After excluding, 529 samples were included for further analysis (Figure 1), including 219 AF, 26 ACA, 29 paired BCA, 229 HC samples from Zhengzhou, and 26 AF samples from Guangshan. Tongue-coating samples from AF and HC were randomly divided into the discovery phase and validation phase in a ratio of 3:1. In the discovery phase, we characterized the tongue coating microbiome in 164 AF and 164 matched HC samples, identified the key tongue coating microbial markers, and constructed AF classifiers by a fivefold cross-validation random forest model. In the validation phase, we verified the diagnostic efficacy of the AF classifier based on the tongue coating microbiome in 55 AF samples. Furthermore, 26 tongue-coating samples from Guangshan were applied to validate the cross-regional diagnostic efficacy of the AF classifier. In addition, we divided AF into two groups (79 PAF and 125 psAF samples), and comparatively characterized the oralome of different AF types. Moreover, we also examined the dynamics of the tongue coating microbiota in AF patients before catheter ablation (BCA) (29 samples) and 6 months after catheter ablation (ACA) (26 samples).

In the discovery cohort, the clinical characteristics of AF and HC were shown in Supplementary Table S1. Age and sex were matched between the AF and HC groups (p > 0.05). In the AF groups, compared with HC, total cholesterol (TC) (p < 0.001), triglyceride (TG) (p < 0.05), low-density lipoprotein (LDL) (p < 0.001), lactate dehydrogenase (LDH) (p < 0.05), total protein (TP) (p < 0.001), albumin (ALB) (p < 0.001), direct bilirubin (DBIL) (p < 0.001), red blood cells (RBC) (p < 0.001), blood platelet (PLT) (p < 0.001), and C-reactive protein (CRP) (p < 0.001) were decreased, while glucose (GLU) (p < 0.05), hemoglobin A1c (HbA1c) (p < 0.05), aspartate aminotransferase (AST) (p < 0.001), alanine aminotransferase (ALT) (p < 0.05), gamma-glutamyl transferase (GGT) (p < 0.001), total bilirubin (TBIL) (p < 0.001), indirect bilirubin (IBIL) (p < 0.001), and blood urea nitrogen (UA) (p < 0.001) was increased. We performed a correlation analysis of bacteria in the tongue coating microbiota with data on medication use (anticoagulants, antiarrhythmics, lipid-lowering drugs, antihypertensives, antiplatelets, and amiodarone) and medical history (cerebral infarction, hypertension, and type 2 diabetes mellitus) using Masslin2 (Supplementary Figure S1). We excluded these bacteria associated with clinical indicators from the differential bacteria between AF and healthy individuals.

The Venn diagram displayed that 2,542 of 4,930 OTUs were common to both AF and HC groups, while 1,069 OTUs were unique to the AF (Figure 2a). The tongue coating microbial diversity of AF and controls was evaluated through the Shannon index and Chao 1 for alpha diversity, principal coordinates analysis (PCoA) and nonmetric multidimensional scaling (NMDS) for beta diversity. Rarefaction analysis of the discovery cohort showed that OTU diversity and richness in each group approached saturation (Figures 2b,c). As estimated by the Shannon index (p < 0.001) and Chao 1 index (p < 0.05), tongue coating microbial diversity and richness were increased in the AF versus HC (Figures 2b,c; Supplementary Table S2). Both PCoA and NMDS plots calculated by Bray-Curtis distance revealed that tongue coating microbiome separated AF and controls into two strikingly distinct groups (p = 0.001, ANOSIM) (Figures 2d,e; Supplementary Figures S2a,b), indicating that tongue coating microbiome dysbiosis occurred in AF.

Further analysis of AF and HC composition and alterations uncovered that genera Prevotella, Neisseria, Streptococcus, Veillonella, and Fusobacterium were the five leading genera in both groups (Figure 2f; Supplementary Table S3), and the phylum Firmicutes, Bacteroidota, Proteobacteria, Actinobacteriota, and Fusobacteriota were the most five abundant phyla, accounted for over 95% of sequences on average (Supplementary Figure S2c; Supplementary Table S4). At the genus level, 37 microbial features were enriched in AF, including Streptococcus, Actinomyces, and Leptotrichia, while 15 microbial features were depleted, including Prevotella, Rothia, and Haemophilus (Figure 2g; Supplementary Table S5). At the phylum level, the phylum Firmicutes, Actinobacteriota, Fusobacteriota, Patescibacteria, Campilobacterota, and Spirochaetota were increased, and the phylum Bacteroidota and Proteobacteria were diminished in the AF compared with controls (Supplementary Figure S2d; Supplementary Table S6). We further performed linear discriminant analysis (LDA) effect size (LEfSe) analysis and selected the most representative genera closely correlated with AF in accordance with LDA (LDA > 2.5, p < 0.05) (Supplementary Figure S2e; Supplementary Table S7). These findings indicated that unique tongue coating microbiome profiles were present in AF groups.

To assess the diagnostic value of tongue coating microbial markers for AF, we constructed a random forest classifier model between 164 AF and 164 matched HC. Initially, 4 OTUs that could accurately identify differences between both groups were selected as the optimal marker set through fivefold cross-validation in the random forest model (Figures 3a,b). Then, we calculated the POD index of the discovery cohort by using a 4-OTU set (Supplementary Table S8). The POD index was markedly higher in AF than in HC (Figure 3c), and it reached an AUC of 99.10% (95% CI 98.20 to 100%, p < 0.0001) (Figure 3d). These data indicated that oral microbial markers could specifically identify patients with AF from HC.

Meanwhile, 55 AF in the validation cohort were used to verify the diagnostic efficacy of microbial biomarkers. The POD index (Figure 3e; Supplementary Table S9) was significantly higher in AF, with an AUC value of 98.62% (95% CI 97.10% to 100%) between both groups (p < 0.0001) (Figure 3f). Additionally, we further collected 26 AF tongue-coating samples from Guangshan, which served as an independent diagnostic. The POD index (Figure 3g; Supplementary Table S10) was higher in 26 Guangshan AF versus HC, with an AUC value of 97.97% (95% CI 95.72 to 100%) between both groups (p < 0.0001) (Figure 3h). These results suggested that this classifier based on the oral microbiome for AF has powerful diagnostic efficacy.

According to AF history and manifestation of cardiogram, AF patients were classified as PAF (n = 79) and psAF (n = 125). A Venn diagram showed that 1,701 of 2,970 OTUs were shared between PAF and psAF, while only 492 and 777 OTUs were unique to the PAF and psAF, respectively (Figure 4a). The tongue coating microbial diversity and richness of PAF and psAF were similar (p > 0.05) (Figures 4b,c; Supplementary Table S11). Although there was no significant difference, Shannon and Chao 1 index were decreased in psAF patients. The heatmap, PCoA, and NMDS analysis failed to distinguish different AF patients (p > 0.05, ANOSIM) (Figures 4d–f; Supplementary Figure S2f; Supplementary Table S12). These findings indicated that AF patients possess similar microbial features in tongue coating diversity, regardless of whether they manifested clinically as PAF or psAF.

The average composition and relative abundance of the tongue coating microbiome at the genus and phylum levels were displayed in Figure 4g, Supplementary Figure S2g, and Supplementary Tables S13, S14. Although PAF and psAF shared similar taxonomic characteristics, there were also some specific differential bacteria. Compared with PAF, two genera including Streptococcus and Lactobacillales_P5D1–392 were increased, while three genera including Fusobacterium, Haemophilus, and Leptotrichiaceae_uncultured were depleted in psAF (all p < 0.05) (Figure 4h; Supplementary Table S15). These results demonstrated a further increase in Streptococcus and Lactobacillales_P5D1–392, while a further decrease in Haemophilus during the AF progression. Different analyses at the phylum level are presented in Supplementary Figure S2h and Supplementary Table S16. Overall, these results indicated that the oral microbial abundances and composition in PAF were basically consistent with those in psAF, but some specific AF-related bacteria changed more significantly in psAF.

We explored and compared the tongue coating microbiome characteristics in 26 ACA, 29 matched BCA, and matched HC. The tongue coating microbial diversity in the ACA was similar to that in the BCA but increased compared with that in the HC (p < 0.05) (Figure 5a; Supplementary Table S17). Microbial richness did not differ significantly among the three groups (Supplementary Table S17). A Venn diagram revealed that 1,407 of 1,951 OTUs in ACA were shared with HC, and 1,580 OTUs were shared with BCA (Figure 5b). PCoA and NMDS displayed that the tongue coating microbial distribution in ACA was comparable to that of BCA (p > 0.05, Adonis) but distinct from that of HC (p < 0.05, Adonis) (Figures 5c,d; Supplementary Figure S2i).

The average composition and relative abundance of the tongue coating microbiome for three groups at the genus and phylum levels were displayed in Figure 5f, Supplementary Figure S2k, and Supplementary Tables S18, S19. The most dominant microbes in BCA and HC were Neisseria but were Prevotella in ACA. The differential microbes among the three groups displayed that the microbial characterization of ACA was significantly different from HC but similar to BCA (Figure 5g; Supplementary Table S20). We further performed differential analysis at the genus and phylum levels (Figure 5e; Supplementary Figure S2j; Supplementary Tables S21, S22). The outcomes demonstrated a total of 288 different genera in the three groups, of which the abundance of 15 genera gradually decreased after catheter ablation, including Lactobacillales_P5D1–392, Lautropia, and Lactobacillales_unclassified, while 12 genera gradually increased. We conducted LEfSe analysis and selected the most representative genera among three groups on LDA (LDA > 2.5, p < 0.05) (Supplementary Figure S2l; Supplementary Table S23). These data indicated that dynamic alterations in the tongue coating microbiome might reflect or contribute to post-ablation pathophysiological processes in AF patients.

The enriched pathways in which the tongue coating microbiome might affect the progression of AF were identified in the KEGG database. Compared with HC, 34 functional modules, including biosynthesis of ansamycins, synthesis and degradation of ketone bodies, and galactose metabolism increased significantly, while 27 functions, including lipopolysaccharide biosynthesis, lipoic acid metabolism, and citrate cycle decreased significantly in AF (LDA > 3, p < 0.05) (Figure 6a; Supplementary Table S24). Compared with PAF, 6 functional modules, including drug metabolism (other enzymes), galactose metabolism, and sphingolipid metabolism increased significantly, while 10 functions, including lipopolysaccharide biosynthesis, biotin metabolism, and riboflavin metabolism decreased significantly in AF (LDA > 3, p < 0.05) (Figure 6b; Supplementary Table S25). The results displayed that 14 pathways were enriched in the ACA group, such as biosynthesis of ansamycins, galactose metabolism, as well as glutamine and D glutamate metabolism (LDA > 3, p < 0.05) (Figure 6c; Supplementary Table S26). Compared with HC, many pathways were depleted in the ACA and BCA groups, such as lipoic acid metabolism, biosynthesis of unsaturated fatty acids, and butanoate metabolism. These outcomes implied dysbiosis of functional patterns in AF progression.

The correlation between tongue coating microbiome and clinical indicators was identified in AF patients and controls through Spearman correlation analysis (p < 0.05 and absolute rho > 0.2) (Figure 7; Supplementary Table S27). We found that 2 clinical indicators (GLU and TG) were correlated with no more than 10 OTUs, while 12 clinical indicators (IBIL, BUN, CRP, GLB, ALB, TP, DBIL LDH, PLT, CK, LDL, and RBC) were related to more than 10 OTUs. CK were positively correlated with 11 OTUs, including OTU1856 (Haemophilus) and OTU14 (Prevotella) (p < 0.05), and negatively correlated with 10 OTUs, including OTU4343 (Streptococcus) and OTU21 (Actinomyces) (p < 0.05). CRP was positively correlated with 10 OTUs, including OTU4343 (Streptococcus) and OTU21 (Actinomyces) (p < 0.05), and negatively correlated with 11 OTUs, including OTU1856 (Haemophilus) and OTU14 (Prevotella) (p < 0.05). This explained the interaction of tongue coating microbiome, cardiac enzymes, liver and kidney function, as well as routine blood tests, pathways that might be involved in disease progression.