Participants were prospectively recruited at the Chung-Ang University Hospital between December 2016 and October 2022. The inclusion criteria for the neoplasm group were newly diagnosed patients with histologically confirmed GC or adenoma. The control group included patients with only gastritis and no evidence of gastric neoplasms on endoscopic examination. The exclusion criteria included individuals with a history of malignancy (including GC), those who had undergone gastric surgery, those who had taken antibiotics or probiotics within the past three months, or those under 20 years of age. The final cohort comprised 141 participants, including 132 patients and nine controls (100 male and 41 female). This study was approved by the Institutional Review Board (IRB) of the Chung-Ang University Hospital (IRB No. C2016047(1790)) and adhered to the Declaration of Helsinki. The informed consent was provided in written form by all participants. Gastric neoplasms were diagnosed and categorized using the Vienna Classification System [9]. Specifically, category 3 (non-invasive low-grade adenoma/dysplasia) was classified as low-grade dysplasia (LG), category 4.1 (high-grade adenoma/dysplasia) and 4.2 (non-invasive carcinoma [carcinoma in situ]) were classified as high-grade dysplasia (HG), and categories 4.3 (suspicion of invasive carcinoma) and 5 (invasive neoplasia) were classified as GC.

The following protocols were used for sample collection. For saliva samples, participants were instructed to swish 10 mL of saliva in their mouths and then expectorate into a specimen tube, which was stored at -80°C until use. Gastric juice samples were collected following a specific protocol. Patients fasted for > 8 hours prior to collection. For those undergoing endoscopic examination or resection, a trap tube was attached between the endoscope and the suction tube. Gastric juice (7–30 mL) was suctioned through the endoscope at the beginning of the procedure. Simultaneously, tissue samples from the antrum and body were collected and stored at -80°C. For patients undergoing surgical gastrectomy, gastric juice and tissue samples were obtained either endoscopically before surgery or during surgery, immediately after making an incision. The endoscopes were washed and disinfected by immersion in a detergent solution containing 7% proteolytic enzyme and 2% glutaraldehyde to prevent contamination. Gastric juice samples were immediately stored at -20°C and transported to the laboratory without preservative reagents. The gastric juice was then diluted to 40 mM with 1 M Tris base and centrifuged at 600 ×g for 10 minutes at 4°C to obtain a supernatant, which was further centrifuged at 1,500 ×g for 10 minutes at 4°C. Finally, the supernatant was collected and stored at -80°C.

Bacterial EVs were isolated from the gastric juice, saliva, serum, and urine of the individuals following a previously described procedure [10,11]. Each gastric juice, saliva, and urine sample was centrifuged at 10,000 ×g for 10 minutes at 4°C. The supernatant was taken and passed through a 0.22 μm membrane filter to eliminate foreign particles and then quantified based on protein concentration [12,13]. For serum, after mixing 100 μL of serum and 900 μL PBS, it was centrifuged at 10,000 ×g for 10 minutes at 4°C to eliminate other components. The supernatant was taken and passed through a 0.22 μm membrane filter to eliminate foreign particles, and it was quantified based on protein concentration.

Bacterial DNA was extracted from the prepared EVs as previously described [11,14]. Isolated EVs (1 μg by protein, each sample) were boiled at 100°C for 40 minutes and centrifuged at 13,000 ×g for 30 minutes, and the supernatants were collected. The collected samples were then subjected to bacterial DNA extraction using a DNA extraction kit (PowerSoil DNA Isolation Kit; MO BIO, Carlsbad, CA, USA) following the manufacturer’s instructions. The isolated DNA was quantified using the QIAxpert system (QIAGEN, Hilden, Germany). Microbial genomic DNA was extracted using a DNeasy PowerSoil kit (QIAGEN) according to the standard protocol provided in the manufacturer’s instructions. The isolated genomic DNA was amplified by targeting the V3–V4 hypervariable regions of the 16S rRNA gene. This amplification was carried out using specific primers (16S_V3_F: 5′-TCGTCGGCAGCGTCAGATGTGTATAAGAGACAGCCTACGGGNGGCWGCAG-3′ and 16S_V4_R: 5′-GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAGGACTACHVGGGTATCTAATCC-3′). Amplicon libraries were prepared from the amplified DNA fragments. The MiSeq Reagent Kit v3 (600-cycle) (Illumina, San Diego, CA, USA) was used to sequence the library reagents. The Nextra XT Index Kit v2 Set A (96 indices, 384 samples) (Illumina) was used for barcodes and adapters. Library preparation for sequencing followed the 16S Metagenomic Sequencing Library Preparation (Part # 15,044,223 Rev. B). All amplicons were sequenced using MiSeq (Illumina) according to the manufacturer’s instructions. The bacterial DNA in each sample was quantified using a QIAxpert system.

Paired-end 16S rRNA gene sequences were processed using the Quantitative Insights into Microbial Ecology (QIIME2 v2021. 4) [15]. Adapter sequences were trimmed using Cutadapt, and the reads were subsequently filtered for quality and the presence of chimeras using DADA2, with specific parameters set as follows: trim-left-f 0, trim-left-r 0, trunclen-f 260, trunc-len-r 200, trunc-q 2, max-ee-f 3, and maxee-r 3 [16]. A naïve Bayes classifier, trained on the V3–V4 region sequences from the SILVA 138 database, was used for taxonomic classification. Sequences identified as chloroplasts or mitochondria were excluded from the analysis.

For alpha diversity analysis, the samples were rarefied to a minimum read count of 2,386 to normalize the data. The p values for alpha diversity were calculated using analysis of variance adjusted for age and sex and were further adjusted using the bootstrap method [17]. Principal coordinate analysis (PCoA) was performed to evaluate the taxon-level clustering of groups based on Bray–Curtis dissimilarity distances. The synthetic minority oversampling technique (SMOTE) algorithm was applied to increase the normal sample size for PCoA [17]. The p value for the PCoA was determined through permutational multivariate analysis of variance using dissimilarity matrices adjusted for age and sex. For the analysis of relative abundances and Multivariable Associations with Linear Models2 (MaAsLin2), operational taxonomic units (OTUs) that exhibited a low count across the dataset were filtered out; specifically, OTUs with a minimum count of less than or equal to four were removed [18]. Furthermore, samples were retained only if they contained OTUs that surpassed the minimum count threshold in more than 20% of the dataset, thereby ensuring the exclusion of samples with sparse OTU representation. To address the low variance, which often indicates a lack of differential abundance across conditions, we implemented a filtering process based on the interquartile range. Features within the lowest 10% of the interquartile range were discarded. Following the filtering process, data normalization was performed to mitigate the effects of varying sequencing depths across the samples. Total sum scaling was used as the normalization technique, where the read count of each sample was divided by the total read count of the sample to ensure comparability across the dataset. The MaAsLin2 method was used to evaluate differences in the OTU counts of bacteria-derived EVs among groups. This analysis was performed using a negative binomial model for group comparisons. Analysis was conducted using MaAsLin2 on MicrobiomeAnalyst (https://www.microbiomeanalyst.ca/) and the MaAsLin2 package in R adjusting for age and sex as covariates. Statistical significance was set at a q-value of less than 0.05. MaAsLin2 was executed using the default parameters in conjunction with the Benjamini–Hochberg false discovery rate method [19]. To validate and complement the results of MaAsLin2, a random forest classifier analysis was performed. To balance the sample sizes across groups, SMOTE oversampling was applied, increasing the sample size of each group (control, LG, HG, and GC) to 100 samples. A random forest classifier was subsequently employed to distinguish between the disease groups. The dataset was split into a training set (70%) and a test set (30%) to rigorously assess the model’s predictive accuracy. For feature selection, microbial biomarkers identified using MaAsLin2 were incorporated into the random forest model. These biomarkers were selected based on their statistically significant associations with the disease groups, serving as key input variables to improve the classifier’s interpretability and overall performance. All statistical analyses were performed in R Software (version 3.6.1; R Foundation for Statistical Computing, Vienna, Austria).

A total of 141 participants were included in the study, including nine controls, 58 with LG, 33 with HG, and 41 with GC. We evaluated samples from four sources: gastric juice, saliva, serum, and urine. A total of 550 samples of all types and conditions were analyzed. The mean age of the study participants was 64.4 years, with males accounting for 70.9% (100 out of 141) of the study population. Table 1 summarizes the clinical characteristics of the enrolled patients, categorized by sample type.

The alpha diversity was assessed in the gastric juice, saliva, serum, and urine samples using various indices (Fig. 1A). In gastric juice, both the Shannon (p = 0.012) and Simpson (p = 0.032) indices showed significant changes in microbial diversity and species dominance with the progression of the pathological condition. Similarly, in the serum, significant changes in microbial diversity were observed in Shannon (p = 0.023) and Simpson (p = 0.040) indices as the disease progressed. In contrast, the saliva and urine samples did not show statistically significant differences in microbial diversity across the different stages of gastric carcinogenesis.

Beta diversity was assessed across different stages of gastric carcinogenesis in all sample types. The results showed significant differences in microbial composition across all sample types during disease progression (p < 0.05, Fig. 1B), with notable differences in gastric juice and urine (p < 0.001, Fig. 1B).

The relative abundances of bacterial communities in various samples (gastric juice, saliva, serum, and urine) were analyzed proportionally, as shown in Figure 2. The bar chart represents the relative abundances of bacterial species identified in the four groups: control, LG, HG, and GC.

In gastric juice samples (Fig. 2A), Cutibacterium acnes dominated in the control group, whereas in the LG, HG, and GC groups, there was a decrease in the abundance of C. acnes and an increase in the abundances of Pseudomonas yamanorum, Pseudomonas antarctica, Ralstonia insidiosa, and Streptococcus oralis. For saliva samples (Fig. 2B), compared with the controls, the LG, HG, and GC groups showed a decrease in C. acnes and S. oralis abundances, along with an increase in P. yamanorum abundance. Notably, Herbaspirillum autotrophicum was detected exclusively in the disease groups (LG, HG, and GC). In the serum samples (Fig. 2C), compared to the controls, the LG, HG, and GC groups showed an increase in C. acnes and R. insidiosa abundances. Urine samples (Fig. 2D) revealed a heterogeneous bacterial composition. However, P. yamanorum was consistently prominent across the disease groups (LG, HG, and GC), indicating a noticeable increase compared to that in the control group.

Specific bacterial strains with significantly different filtered OTU counts in the disease groups (LG, HG, and GC) compared to those in the control group were identified. Figure 3 shows a comparative analysis of the abundances of these bacterial strains across different patient groups for various sample types. Bacterial abundance was represented by filtered counts to assess the correlations between groups. In the gastric juice, the abundance of C. acnes and S. oralis displayed a notable increase with the progression of disease severity (q < 0.01 each, Fig. 3A). In saliva, the levels of C. acnes and S. oralis significantly decreased as the disease progressed, whereas P. antarctica and R. insidiosa increased (q < 0.01 all, Fig. 3B). In the serum, the abundances of P. yamanorum, C. acnes, R. insidiosa, and P. antarctica significantly increased with advancing disease stage (q < 0.01 all, Fig. 3C). Similarly, in urine, C. acnes, R. insidiosa, and P. antarctica showed significantly increasing trends during disease progression (q < 0.01 each, Fig. 3D). To complement the MaAsLin2 results and further assess the discriminative power of these identified bacterial biomarkers across disease stages, the random forest classifier was employed. This analysis evaluated the distribution and classification potential of specific microorganisms associated with different stages of gastric carcinogenesis in each sample type (Supplementary Tables 1-4). The overall classification accuracy was 64.2% for gastric juice and 80.8% for saliva, while serum and urine samples demonstrated classification accuracies of 87.5% and 79.2%, respectively. Meanwhile, it is well known that differences in microbiome diversity and composition are greatly influenced by H. pylori infection status. To assess the impact of H. pylori, all analyses were re-evaluated after excluding samples from H. pylori-positive participants. The overall trends and results remained consistent (data not shown).