The Chinese gastric tissue samples were from 80 gastric cardia cancer patients recruited at the Shanxi Cancer Hospital in Taiyuan, Shanxi Province, China, between 1998 and 2001. This study was approved by the Institutional Review Boards of the Shanxi Cancer Hospital and the National Cancer Institute (NCI). All subjects provided written informed consent prior to participation. Cases were histologically confirmed as adenocarcinomas by pathologists at both the Shanxi Cancer Hospital and the NCI. Clinical data was collected by review of medical records. Patients who were <18 years old, with cancer other than GC or with previous treatment for GC were excluded. Tumor tissues and matched non-malignant tissues distant to the tumor were obtained from surgical resections, snap frozen in liquid nitrogen, and stored at −130°C until used. H&E slides were used to determine the percentage of tumor cells in the tissues. Total DNA was extracted using the Allprep RNA/DNA/Protein mini kit (QIAGEN) following the protocol provided by the manufacturer.

The Mexican gastric tissue samples were from 80 gastric non-cardia cancer patients recruited at the Oncology Hospital, Centro Médico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, and the Instituto Nacional de Cancerología, Secretaria de Salud in Mexico City, Mexico, between 2008 and 2013. The study was approved by the ethics committee of each hospital and written informed consent was obtained from all patients prior to enrollment in the study. Cases were histologically confirmed by the pathologist. The clinical and pathological data were recorded in questionnaires. Patients who were <18 years old, with any autoimmune disease, diabetes, or cancer other than GC, and with a previous treatment for GC were excluded from the study. Tumor tissue and matched non-malignant tissue distant to the tumor were obtained from surgical resection specimens, placed immediately in microfuge tubes and submerged in a container with liquid nitrogen, and stored at −70°C until tested. H&E slides were used to determine the percentage of tumor cells in the tissues. Total DNA was extracted by QIAamp DNA mini kit (QIAGEN) using the protocol provided by the manufacturer.

All the non-malignant tissue samples were verified with absence of tumor cells. Tumor tissue samples without tumor cells were excluded from all analyses. The percentage of tumor cells were 70–80% in the Chinese tumor samples and 30–50% in Mexican tumor samples. Examples of H& E slides are shown in Supplementary Figure 1.

The V3–V4 region of the 16S rRNA gene was amplified and sequenced on the Illumina MiSeq platform using the 300 paired-end protocol at the Institute of Genome Sciences, University of Maryland School of Medicine as described previously (Fadrosh et al., 2014).

Sequence reads were processed to remove low quality, short, or chimera reads (Yu et al., 2015). We removed low quality reads (reads with average quality <20 over 30 bp window based on Phred algorithm; paired reads which have at least one read with length <75% of its original length) and chimera reads (by UCHIME). The remaining reads with at least 97% sequence identity were clustered into species-level Operational Taxonomy Units (OTUs) in the software package Quantitative Insights into Microbial Ecology (QIIME 1.8.0) (Caporaso et al., 2010) by using command pick_open_reference_otus.py with usearch61 clustering algorithm and other default settings. The OTUs were assigned to taxa (e.g., genus, family, phylum) using the Greengenes database as reference (version 13_8; DeSantis et al., 2006). OTUs with only one read were excluded from analysis. Samples with <1,000 reads were excluded from analysis. The sequence data were submitted to BioProject database (accession number of 310127) at the National Center for Biotechnology Information website.

Alpha diversity was estimated as number of OTUs, Shannon's Index (Shannon, 1997), and Phylogenetic diversity (PD_whole_tree) (Faith and Baker, 2006) by averaging over 20 rarefied tables of 1,000 reads/sample. Alpha diversity was used to measure the species diversity of each sample. The number of OTUs, also known as richness, is a measure of diversity that does not consider the frequency of OTUs. Shannon's index is estimated by both the number and frequency of the OTUs. PD_whole_tree further takes account of the phylogenetic relationship of OTUs. The phylogenetic tree of OTUs used for PD_whole-tree estimates was prepared in QIIME based on neighbor-joining method. The alpha diversity increased with number of sequence reads sampled (Supplementary Figure 2). The alpha diversity showed differences by sample groups with 1,000 reads/sample; the order of sample groups based on alpha diversity did not change by number of sequence reads.

Beta diversity was measured as unweighted (presence/absence of taxa) and weighted (using taxa relative abundance information) UniFrac distance (Lozupone et al., 2011). Beta diversity measures dissimilarities of two samples in microbial profiles. We calculated both alpha and beta diversity based on rarefied tables of 1,000 reads/sample.

The relative abundance of taxa at different levels (phylum, class, order, family, and genus) was calculated based on the unrarefied table. The taxa relative abundance was estimated as the proportion of OTUs assigned to a taxon.

The Human Microbiome Project (HMP) 16S rRNA V3–V5 data were downloaded for comparison (http://hmpdacc.org/HMQCP/; Human Microbiome Project Consortium, 2012). The HMP sequence reads were processed in the same manner as described above. A total of 2,579 samples from 5 body sites (including oral, nasal cavity, stool, vagina, and skin) of 242 healthy US adults in HMP phase 1 were used for comparison (Aagaard et al., 2013). The study by Lozupone et al. (2013) showed that the difference in population or technologies used should not affect the comparison by body sites. In addition, we limited the comparison of HMP and stomach microbiota data to the highest and least variable taxonomic (phylum)/functional (module) level so that the population/technology differences between these two studies would have limited effect on the comparisons.

We used Phylogenetic Investigation of Communities by Reconstruction of Unobserved States (PICRUSt)1.0.0 (Langille et al., 2013) to predict virtual metagenomes for each sample from the 16S rRNA gene sequence data and used the KEGG database as a reference (Kanehisa et al., 2014) to determine the relative abundance of metabolic pathways and modules within the virtual metagenomes. PICRUSt requires the use of Greengenes reference, version 13_5 to cluster reads into OTUs (DeSantis et al., 2006). Therefore, we re-clustered all the sequence data (including HMP) in QIIME with the command parallel_pick_otus_usearch61_ref.py and the Greengenes reference version 13_5.

The Wilcoxon rank-sum test was used to examine gastric microbiota alpha diversity and taxa relative abundance differences between antrum and corpus in Mexican samples or between Hp+ and Hp− samples. Wilcoxon signed-rank test was used for the differences between non-malignant and matched tumor samples in each population. When examining taxa relative abundance, Bonferroni correction was used to adjust for tests of multiple taxa. Permutational Multivariate Analysis of Variance (PERMANOVA, adonis) was used to compare sample groups by unweighted/weighted UniFrac distance matrix. P < 0.05 were considered significant after adjustment for multiple tests.

In order to compare gastric microbiota with the oral, nasal, stool, skin, and vagina microbiota from the HMP study, we calculated Euclidean/Bray-Curtis distances and generated a matrix from the phylum-level/KEGG module profiles, performed principal coordinate analysis (PCoA) on the Euclidean/Bray-Curtis distance matrix, and then plotted the figure based on the first three principal coordinates to visualize the similarities and differences among different body sites. All statistical analyses were performed in R. Bray-Curtis and Euclidean distance showed similar results, therefore only Bray-Curtis distance is shown.

To address the concerns about possible contamination, we included 2 blank samples as negative controls. We also included 1 vaginal and 1 stool sample as positive controls to evaluate the performance of DNA amplification and sequencing. The two positive controls generated 2,703 and 58,201 reads, respectively, suggesting good performance of DNA amplification and sequencing. The blanks had extremely low number of reads (41 and 43 reads/sample, respectively). Furthermore, the OTUs found in both blanks were extremely rare in the gastric samples with accumulated relative abundance range of 0–0.006. Therefore, our results were unlikely to have been affected by contamination.

The conventional DNA extraction method for microbiome studies often includes an extra cell lysis step (bead-beating) to break the hard-to-break cell membranes of some species. To examine whether some taxa were missed due to the lack of a bead-beating step in our DNA extraction protocols, we evaluated 2 Chinese tissue samples using two different DNA extraction methods: a DNA extraction method with a bead-beating step and commonly used for microbiome study (Flores et al., 2012) and the method used for our Chinese samples in the current study. We found 14 genus-level taxa discovered by the extraction method with the bead-beating step that were not discovered by our DNA extraction method (Supplementary Table 2). However, these taxa were extremely rare with total cumulated relative abundance of 0.007 and 0.038 for two samples, respectively. Therefore, the DNA extraction method should not have adversely affected our findings, although we cannot exclude missing some rare taxa.

After excluding samples with <1,000 reads per sample, 77 non-malignant gastric tissue samples from China and 80 from Mexico were included for analysis and the median (interquartile range) was 10,460 (5,454–19,980) reads per sample. The raw and qualified number of reads for each sample group are shown in Supplementary Table 3. The average age of these Chinese cases was 60.8 years old, 83% were male, and all were diagnosed with gastric cardia adenocarcinoma. The average age of Mexican cases was 64.5 years old, 54% were male, and all tumors were located in the non-cardia regions of the stomach (21 antrum, 24 corpus, 35 unspecified). In addition, 80 tumor samples from China and 54 from Mexico were also included for comparison [median (interquartile range): 9,406 (4,228–15,330) reads/sample] after excluding samples with <1,000 reads per sample or no tumor cells.

The taxonomic and functional profiles are shown in Figure 1 for non-malignant tissue samples. According to the non-malignant gastric tissues, the gastric microbiota for both sample sets was mainly composed of Proteobacteria, followed by Bacteroidetes in Chinese samples or Firmicutes in Mexican samples (Figure 1A). The majority of samples from China (78%) and Mexico (50%) were dominated by Proteobacteria (relative abundance >50%). Nineteen Mexican samples were dominated by Firmicutes and one Chinese sample was dominated by Bacteroidetes. The remaining samples did not have a dominant phylum.

The most abundant genus in the non-malignant microbiota of both Chinese and Mexican gastric cancer patients was Helicobacter (Figure 1B), and 99% of the Helicobacter reads [median (interquartile range): 98.8% (98.7–99.3%)] were classified as Hp. As shown in Table 1, the majority of GC patients' stomachs (94% Chinese and 55% Mexican) were colonized by Hp, and 53% Chinese and 28% Mexican gastric microbiota were dominated by Hp (Hp relative abundance >50%).

The virtual reconstructed functional profiles (KEGG modules) of non-malignant gastric tissue samples predicted by PICRUSt are shown in Figure 1C. The most abundant module functions in gastric microbiota were membrane transport, amino acid metabolism, carbohydrate metabolism, replication and repair, and energy metabolism in both Chinese and Mexican samples. Compared to the variation in taxonomic profiles, the variation in functional profiles among non-malignant gastric tissue samples was more limited (Figure 1A vs. Figure 1C).

Neither Chinese nor Mexican samples showed an association between gastric microbial features and age or gender (data not shown). Within Mexican samples, no significant difference in microbial alpha diversity, beta diversity and taxa relative abundance for the antrum and corpus non-malignant samples was observed (Supplementary Table 4).

After excluding samples with <1,000 reads, 80 tumor tissue samples from China and 54 from Mexico remained for comparison with their matched non-malignant tissues. The taxonomic and functional profiles for these samples are shown in Supplementary Figure 3. The average profiles for both non-malignant and tumor sample groups are shown in Figure 2. Similar to the profiles in non-malignant tissues, the tumor gastric microbiota for both sample sets was also mainly composed of Proteobacteria, followed by Bacteroidetes in Chinese tumor samples or Firmicutes in Mexican tumor samples (Supplementary Figure 3A vs. Figure 2A). The genus with the most abundance in both tumor sample sets was Helicobacter (Supplementary Figure 3B vs. Figure 2B) with average relative abundance of 21% in Chinese samples and 18% in Mexican samples. Compared to non-malignant tissues, tumor tissue had less Proteobacteria, and higher Bacteriodetes, Firmicutes, Fusobacteria, and Spirochaetes in Chinese samples. There was no significant change in Mexican samples in phylum-level taxa (Supplementary Table 5). At the genus level, tumor tissue had lower Helicobacter abundance relative to non-malignant tissue in both Chinese and Mexican samples. Chinese samples showed substantial differences in alpha diversity as well as several other genus taxa. Mexican samples showed differences in Clostridia relative abundance, but did not display differences in alpha diversity measures (Supplementary Table 5). Hp relative abundance was also lower in tumor tissues compared to matched non-malignant tissues in both sample sets (Supplementary Table 5). However, the majority of tumor tissues (94% Chinese and 56% Mexican) were colonized by Hp, and many tumor samples (20% Chinese and 17% Mexican) were dominated by Hp (Hp relative abundance >50%; Table 1).

The most abundant module functions in tumor tissues were membrane transport, amino acid metabolism, carbohydrate metabolism, replication and repair, translation, and energy metabolism in both Chinese and Mexican samples (Supplementary Figure 3C, Figure 2C). In both Chinese and Mexican samples, the functional module of infectious disease was higher in non-malignant than in tumor tissues (Supplementary Table 6). Chinese samples showed substantial differences in other functional modules after Bonferroni correction for multiple comparisons (Supplementary Table 6). Mexican samples did not display differences in relative abundance for other functional modules between tumors and non-malignant tissues (Supplementary Table 6). Functional and taxonomic profiles were correlated (Supplementary Figure 4). For example, high infectious disease function was mainly contributed by Helicobacter as these factors were positively correlated in relative abundance.

We made PCoA plots based on both unweighted or weighted UniFrac distance matrix to visualize similarities and differences among gastric samples. Both plots suggested that gastric samples were primarily clustered by geographic location, rather than by tissue types (Figure 3).

The average relative abundance of the top abundant genera by body sites are shown in Table 2. The top abundant genera are the genera with average relative abundance >0.05 in at least one body site. The top abundant genera in stomach includes Helicobacter and an unknown Enterobacteriaceae genus in either Chinese or Mexico samples, and two additional genera Streptococcus and Lactobacillus in Mexico samples. The top abundant genera in other body sites included Streptococcus, Prevotella, Haemophilus, Veillonella, and Neisseria in oral cavity, Lactobacillus in vagina, Bacteroides, Faecalibacterium, and Alistipes in stool, Staphylococcus, Corynebacterium, and Propionibacterium in both skin and nasal cavity. These top genera in each body site were considered as the genera associated with their corresponding body site (e.g., vagina_associated genus refers to Lactobacillus). We found that the stomach microbiota was enriched with the genera associated with the oral cavity (combined relative abundance of 17.6 and 11.6% in Mexico and China samples, respectively).

Similarities and differences of taxonomic/functional profiles by body sites are shown in Figure 4. The principal coordinates plots based on taxonomic profiles (phylum-level) demonstrated the primary clustering of samples by body sites (Figures 4A,B). The stomach samples, Chinse or Mexico, largely overlapping with oral sample cluster, which was clearer when Helicobacter reads were removed (Figures 4A,B). Compared to the principal coordinates plots based on taxonomic profiles, the plots based on functional profiles (KEGG modules) showed a much clearer pattern of clustering by body sites (Figure 4A vs. Figure 4C, Figure 4B vs. Figure 4D). The stomach samples, with or without Helicobacter reads removed, either Mexico or Chinese samples, did not cluster with the other body sites, but they also did not cluster with each other as closely as the samples in the other body sites. It suggested higher inter-subject dissimilarity in stomach samples than in the other body sites in functional profiles. The inter-subject dissimilarity by body sites based on Bray-Curtis distance of phylum/KEGG module profiles were then evaluated (Supplementary Figure 5). Mexican stomach samples had the highest inter-subject dissimilarity in phylum profiles. The Chinese samples, however, had inter-subject dissimilarity higher than the other body sites only when Helicobacter reads were removed (Supplementary Figure 5A). This might be due to the fact that almost all Chinese samples (94%) had Helicobacter. The inter-subject dissimilarity in functional profiles was much higher in the stomachs of both sample sets than in other body sites (Supplementary Figure 5B).

To further evaluate the gastric tissue microbiota by Hp colonization status, we removed the Helicobacter reads from the Hp+ samples (with Hp) and then compared them to the Hp− samples (without Hp) for alpha diversity, beta diversity, and taxa relative abundance separately for non-malignant and tumor samples. No significant differences were observed among Mexican samples (Supplementary Table 7). A similar comparison could not be performed in the Chinese samples because too few samples were Hp− (n = 5).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.