We conducted a cross-sectional study of 397 patients with stomach gastritis, residing in Anhui and surrounding provinces, China, who were prospectively recruited between June 2019 and December 2019. Study subjects were recruited from hospital outpatients and completed the 13C-methacetin breath test at the same time. Written informed consent was obtained from each patient. Experienced endoscopists performed all endoscopic examinations and categorizations according to the Kimura-Takemoto classification (Duc et al., 2011), and obtained gastric juices and biopsies using sterile tubes and forceps. Stool samples were collected using sterile sampling tubes before or 2 h after endoscopy. A total of 176 participants completed baseline and follow-up endoscopic examinations and provided gastric biopsies, juices, and stool samples. The patients who used antibiotics and probiotics within 1 month and pump inhibitors within 2 weeks were excluded. General health information of all participants was acquired using a structured questionnaire. This study was approved by the Ethics Committee of the First Affiliated Hospital of University of science and technology of China (2019-ky064, Anhui, China), and informed consent was obtained from all participants before they enrolled in the study.

In addition, we utilized two validation datasets containing 16S rRNA gene sequencing data of 657 gastric biopsies from patients with gastric cancer or precancerous lesions, which were retrieved from the Sequence Read Archive [BioProject: PRJNA375772 (Coker et al., 2018) and PRJEB26931 (Wang et al., 2020)].

Biopsies were taken from the antrum and the gastric body of the participants for microbiota study and histology diagnosis. The biopsied tissues were examined by experienced pathologists. The AG and IM scores were classified into four grades, i.e., 0 = ‘normal’, 1 = ‘mild’, 2 = ‘moderate’, and 3 = ‘marked’, based on the updated Sydney system (Dixon et al., 1996). Scores ≥ 1 were considered positive. IN was classified according to the World Health Organization classification of digestive system tumors and the Vienna classification (Lewin, 1998). Patients with SG were used as controls. Based on the histological diagnosis, we assembled two comparison groups (PLGC vs. SG) to identify PLGC-associated bacteria. The PLGC group was further subdivided into three subgroups (AG, IM, and IN) to understand how gastrointestinal microbiota changes along with the progression of gastritis. We also evaluated the atrophy grade by the OLGA system (Rugge et al., 2007) and IM grade by the OLGIM system (Capelle et al., 2010).

Two more gastric biopsies were taken from the inflammatory site of the participants’ gastric antrum and body for microbial detection. All gastric biopsies, juices, and stool samples were frozen immediately after collection within 0.5 h and stored at -80°C until analysis. DNA was extracted from gastric biopsy and juice samples using the low-salt CTAB method (Arseneau et al., 2017) and from stool samples using the E.Z.N.A. ^®Stool DNA Kit (D4015; Omega Bio-Tek, Norcross, GA, USA), according to the manufacturer’s instructions. Total DNA was eluted in 50 μl of elution buffer and stored at -80°C until measurement by LC-Bio Technology Co., Ltd, (Hang Zhou, China). Purification of DNA for the bacterial small subunit (16S) rRNA gene sequencing was performed on a total of 528 samples. Of these, 246 samples were excluded due to failed DNA extraction and exclusion criteria, and sequencing was completed for 66 gastric biopsies, 68 gastric juices, and 148 stool samples. The V3-V4 region of the 16S rRNA gene was amplified with primers 341F (5’-CCTACGGGNGGCWGCAG-3’) and 805R (5’-GACTACHVGGGTATCTAATCC-3’) (Logue et al., 2016). Amplicon pools were prepared for sequencing and the size and quantity of the amplicon library were assessed on a 2100 Bioanalyzer system (Agilent, Santa Clara, USA) and with a Library Quantification Kit for Illumina (Kapa Biosciences, Woburn, MA, USA), respectively. The libraries were sequenced on a NovaSeq PE250 platform.

Sequencing quality filtering and analysis were performed using the QIIME2 pipeline (v2020.11) (Bolyen et al., 2019). Paired-end reads were filtered, merged, and dereplicated using VSEARCH and the Deblur plugin, which reduced the feature table and feature sequences (the “features” resulting from Deblur are created by grouping unique sequences). The samples with > 10000 frequency were reserved for subsequent analysis, including public datasets. The SILVA 16S database (v132) was used for taxonomy assignment of sequence datasets and performed by the QIIME2 plugin feature classifier. The metagenome functional profiling of the gastrointestinal microbial communities was estimated by the Phylogenetic Investigation of Communities by Reconstruction of Unobserved States-PICRUSt2 (v2.0.0) (Douglas et al., 2020). Differentially pathways with a false discovery rate adjusted p-value <0.05 were presented.

Alpha diversity analysis was performed using the QIIME2 process and illustrated by Faith’s phylogenetic diversity, Chao1, and Shannon indexes, which were assessed using an ANOVA test for multiple groups. Multiple group comparisons of dissimilarities between groups were made using the permutational multivariate analysis of variance test. Microbial diversity was visualized using the discriminant analysis of principal components based on the pathological diagnosis group. Taxonomic discovery analysis was performed using R package DESeq2 and linear discriminant analysis effect size (LEfSe, http://huttenhower.sph.harvard.edu/galaxy/), and bacteria with > 0.1% relative abundance (RA) were exhibited (Segata et al., 2011). DESeq2 was used in combined public datasets with cohort information added to adjust the batch effect. The identification of co-occurring and co-exclusion bacteria at the genus level was estimated using the SparCC algorithm (Coker et al., 2018), and visualized by Gephi (v0.9.2) (Heymann and Grand, 2013). Network parameters including topological coefficient, closeness, and betweenness were estimated using igraph (v1.2.5) and compared using the Wilcoxon test. Data visualization was performed by the R Project (v4.0.2). Univariate analysis of functional features between multiple diagnosis groups was undertaken with an ANOVA test using STAMP (v2.1.3) (Parks et al., 2014). All p values < 0.05 after multiple comparisons correction using the false discovery rate method were considered significantly different.

The data presented pertain to samples from a total of 148 subjects (148 stool samples, 66 gastric biopsies, and 68 gastric juices) after exclusion and successful sequencing ( Supplementary Figure 1 ). The baseline characteristics of participants and public data are presented in Supplementary Tables 1, 2 . We performed endoscopic diagnosis according to the Kimura-Takemoto classification (Duc et al., 2011) and further pathological diagnosis according to the updated Sydney system (Dixon et al., 1996), operative link on gastritis assessment (OLGA) staging system (Rugge et al., 2007), and the modified operative links of gastric intestinal metaplasia (OLGIM) system (Capelle et al., 2010) ( Supplementary Figures 2A, C ). The histological assessment of AG showed good agreement with the Kimura-Takemoto classification with a moderate correlation ( Supplementary Figure 2B ).

The participants were tested for H. pylori infection using 13C-UBT. Mucosal biopsies with >1% H. pylori RA were classified as H. pylori-infected, the consistency between the conventional detection methods and the results based on 16S rRNA gene sequencing was only 78.94% in our dataset, and 49.02% in validation datasets ( Supplementary Table 3 ). The participants were grouped as H. pylori-positive in biopsies with >1% H. pylori RA, whereas biopsies samples with <1% H. pylori RA were grouped as H. pylori-negative in all the datasets as previously described (Kim et al., 2015).

The overall RA at the phylum level indicated less diverse bacterial taxa among gastric biopsies in the H. pylori-positive group ( Figure 1A ). The taxonomic composition of the mucosa of H. pylori-infected participants compared with H. pylori-negative subjects included significantly increased Epsilonbacteraeota (mean RA from 0.91% to 68.22%, adjusted p < 0.001) and significantly decreased Firmicutes (mean RA from 27.55% to 8.18%, adjusted p < 0.01), Proteobacteria (mean RA from 36.53% to 13.97%, adjusted p < 0.01), and Bacteroidetes (mean RA from 19.70% to 6.16%, adjusted p < 0.05). However, there was no significant change in the proportion of Epsilonbacteraeota and other phyla in the stool and gastric juice samples of the H. pylori-positive group compared with the H. pylori-negative group. The Shannon index of alpha diversity analysis showed a lower bacterial diversity in biopsies of the H. pylori-positive group (Wilcoxon test, p < 0.001), and no significant change in gastric juices and stool samples ( Figure 1B ). We also observed that the H. pylori RA is negatively correlated with mucosal bacterial diversity in validation data ( Supplementary Figure 3 ). Beta diversity analysis with PCoA based on the feature level revealed a significant difference in gastric mucosal microbiota between the H. pylori-positive group and H. pylori-negative group ( Supplementary Figure 4 ).

Biomarker discovery using the R package DESeq2 showed that Helicobacter was enriched in the H. pylori-positive group of gastric biopsies, whereas most genera were depleted (RA > 0.1%, Figure 1C ). In gastric juice and stool samples, there were few changes in bacteria diversity, composition, and genera between H. pylori-positive and H. pylori-negative groups. A heatmap was constructed using data from the top 30 genera among all gastrointestinal samples ( Supplementary Figure 5 ). The high abundance taxa, including Streptococcus, Hemophilus, and Veillonella were dominant in most samples. Halomonas and Shewanella were dominant in gastric biopsies, Sphingomonas, Acinetobacter, and Curvibacter were enriched in gastric juices, while Bacteroides and Faecalibacterium were mainly in stool samples. The high abundance of Helicobacter was detected in all gastric biopsies, in some of the juice samples, and none of the stool samples of the H. pylori-positive group. Helicobacter was also detected at low levels in the stomach and intestine of some H. pylori-negative subjects.

H. pylori-positive subjects were excluded from the disease-associated bacterial diversity analysis because of the significantly reduced microbial diversity induced by H. pylori infection, according to our findings and previous studies (Alarcon et al., 2017; Li et al., 2017). The alpha diversity of gastric mucosal bacteria, including the Chao1 and Faith’s phylogenetic diversity indexes, showed a significant decrease in the IN group compared with the SG group at the features level ( Figures 2A, B ). The decreasing trend in Faith’s phylogenetic diversity index of gastric mucosal microbiota was also found across the different stages of OLGA and OLGIM ( Supplementary Figure 6 ). We also observed that phylogenetic diversity of mucosal bacteria significantly decreased in patients with IN compared with SG, AG, and IM in validation data ( Supplementary Figure 7 ). The collective results suggested the disorder of gastric mucosal microbiota in subjects with gastric precancerous lesions, especially in the stage of intraepithelial neoplasia. However, there was no difference observed in gastric juices and fecal microbial diversity between the different diagnosis groups.

Differences in microbial community structure were further evaluated using discriminant analysis of principal components analysis (Jombart et al., 2010). The analysis separated each histological diagnosis group into different clusters in all the samples ( Figures 2C–E ). Beta diversity analysis using the Weighted-Unifrac distance matrices of gastrointestinal samples was conducted to assess dissimilarities among all groups. And the results showed that only the mucosal microbiota had significant differences between diagnosis groups ( Supplementary Figure 8 ).

We performed a microbial co-occurrence and co-exclusion network and topology analysis at different stages of gastritis and PLGC to explore the interplay of the gastrointestinal microbiota. Overall, the strength of co-occurring interactions among genera increased in IM, and decreased in IN of H. pylori-negative gastric biopsies ( Figure 3A ). We observed that the genera Gemella, Veillonella, Streptococcus, Actinobacillus, and Hemophilus had the higher degree of centrality and strong co-occurrence interaction in gastric biopsies across the PLGC stages. And Acinetobacter co-occurred with a variety of genera in IN. The similar co-occurring interactions trend and central bacteria in each PLGC subgroup of H. pylori-negative gastric biopsies was also observed in validation datasets ( Supplementary Figure 9 ). The co-occurrence interactions between Curvibacter, Sphingomonas, Acinetobacter, and Fusobacterium were observed in gastric juices ( Figure 3B ). In stool samples, several genera of Firmicutes, such as Ruminococcus gnavus and Veillonella, exhibited a high degree of centrality ( Figure 3C ).

We then assessed if the bacteria could potentially contribute to gastric precancerous lesions in patients using DESeq2. The results showed that Prevotella_2 and Sphingomonas were enriched in the AG group, Dorea, Caulobacter, and Bacteroides were enriched in the IM group, and Bradyrhizobium, Sphingomonas, Curvibacter, and Acinetobacter were enriched in the IN group of H. pylori-negative biopsies ( Figure 4A ). Differential bacteria analysis using the LEfSe test between diagnosis groups was also performed. Enriched Bifidobacterium and Klebasiella in SG, and enriched Sphingomonas and Acinetobacter in IN were observed in gastric biopsies ( Supplementary Figure 10 ). In validation datasets, we found that Acinetobacter was enriched in the IN group of both H. pylori-positive and H. pylori-negative biopsies ( Supplementary Figure 11 ). Enriched Prevotella_2 in AG and enriched Dorea in IM were observed in gastric juices ( Figure 4B ). The significantly changed genera in AG and IM groups of biopsies may be influenced by the changing bacteria in gastric juice. In stool samples, Akkermmansia and Catenibacterium in SG, Lactobacillus in AG and IN, and Holdemanella in IM and IN were the enriched genera ( Figure 4C ).

We further constructed random forest models based on the mucosal bacteria with > 0.1% RA in combined validation datasets and tested them in our dataset, to assess the value of gastrointestinal microbiota in predicting the risk of gastric cancer. The model of mucosal microbiota showed excellent performance in distinguishing PLGC and SG with an AUC of 0.794 in H. pylori-positive subjects ( Figure 5A ). Whereas the model of mucosal microbiome in H. pylori-negative subjects yielding an AUC of 0.526 ( Figure 5B ). Additionally, the model based on intestinal microbiota distinguished PLGC from SG with an AUC of 0.65 in our dataset ( Figure 5C ). The important genera that contributed to the models were selected by R package Boruta and listed in Supplementary Table 4 (Miron and Kursa, 2010).

The functional predictions in combined datasets that identified KEGG orthologs, Enzyme Classification, and metaCyc pathways were performed using PICRUSt2 ( Supplementary Table 5 ). The superpathway of N-acetylglucosamine (GLCMANNANAUT-PWY), superpathway of (Kdo)2-lipid A biosynthesis (KDO-NAGLIPASYN-PWY), superpathway of N-acetylneuraminate degradation (P441-PWY), superpathway of demethylmenaquinol-8 biosynthesis (PWY-5861), and Ribonucleoside-triphosphate reductase (EC:1.17.4.2) were found significantly enriched in the IN group of both H. pylori-positive and H. pylori-negative subjects. No significantly changed function was found in gastric juices and stool samples between multiple diagnosis groups.