OAPGC: a high-quality oral and airway prokaryotic genome catalog for enhanced ecological resolution and disease inference

Abstract

The human oral cavity and airway harbor diverse microbiomes that are implicated in oral and respiratory diseases, yet comprehensive genomic catalogs remain scarce. Here, we present the Oral and Airway Prokaryotic Genome Catalog (OAPGC), comprising 99,215 high-quality, non-redundant genomes reconstructed from public and newly sequenced metagenomes and isolates. OAPGC was clustered into 2474 species using refined, phylum-specific nucleotide identity thresholds, and 29.5% of them are uncultured. Habitat-driven divergence was evident across 15 oral and 8 airway sites, with airway microbiomes showing greater inter-individual variability and enriched antibiotic resistance genes. Across 25 case–control comparisons covering 12 diseases, disease status explained significant community shifts in 19 datasets, with classifiers achieving an AUC > 0.70 in 20 datasets. Shared microbial signatures were identified for diseases such as periodontitis and pneumonia, including uncultured taxa. We also detected 12.3% of OAPGC species in the gut, whose enrichment was linked to multiple diseases and improved cross-cohort classification performance. OAPGC establishes a foundational, disease-relevant genomic framework for oral and airway microbiome studies.

Introduction

The oral cavity and respiratory tract are intensively exposed to the external environment, with approximately 10⁵–10⁷ bacterial cells inhaled daily¹. The anatomical connection between the two niches supports the topological continuity² of their resident microbiota, with prokaryotes constituting the primary components intimately involved in maintaining oral and respiratory health. Disruptions in the oral prokaryotes have been implicated in local diseases such as periodontitis (PD)³ and dental caries⁴, while the enrichment of oral taxa in the lower respiratory tract has been linked to pulmonary diseases including aspiration pneumonia⁵, lung inflammation⁶, and malignant progression of pulmonary nodule⁷. Alterations in the oral prokaryotes have also been linked to a range of systemic disorders⁸, including rheumatoid arthritis (RA)⁹, cardiovascular disease¹⁰, Alzheimer’s disease (AD)¹¹, and colitis¹². Conversely, the airway prokaryotes are associated with respiratory conditions such as chronic obstructive pulmonary disease (COPD)¹³, asthma¹⁴ and respiratory infections^15,16. Respiratory pathogens may retrogradely translocate into the oral cavity via coughing, though their perturbation is likely minimal due to the high density of the native microbiota¹⁷. Despite the increasing evidence linking the oral and airway microbiota to various diseases, the lack of a comprehensive prokaryotic reference database encompassing sufficient microbial diversity continues to impede accurate characterization of their taxonomic composition and functional potential, thereby limiting mechanistic insights and highlighting the urgent need for a high-quality genome catalog.

With the advent of the big-data era¹⁸, culture-independent metagenome-assembled genomes (MAGs) have become a widely used approach for reconstructing microbial genomes directly from metagenomic samples. This strategy has been successfully applied in large-scale gut microbiome databases, such as the Unified Human Gastrointestinal Genome (UHGG)¹⁹ and HumGut²⁰. In the context of oral and respiratory microbiome genome resources, the expanded Human Oral Microbiome Database (eHOMD)²¹ encompasses 834 taxa from the aerodigestive tract, the majority of which are derived from cultured isolates. More recently, the human reference oral microbiome (HROM)²², together with other genome databases focusing on oral²³ or respiratory²⁴ microbiomes, have been independently published. While these studies have substantially advanced the development of oral and airway prokaryotic genome resources, a vast portion of uncultured diversity within these niches remains unexplored. This ‘microbial dark matter’ may have introduced significant biases or resulted in missing information in previous disease-association studies. Therefore, an inclusive and non-redundant genome catalog that provides comprehensive insights across global, ecologically diverse, and disease-relevant contexts is highly warranted.

In recent years, increasing attention has been paid to the phenomenon of oral-gut transmission of microbes. Oral bacteria may enter the gut through saliva and food intake, or translocate via the bloodstream from oral lesions. Normally, multiple physical, chemical, and biological barriers prevent their colonization in the gut²⁵. Nonetheless, species commonly found in the mouth, such as Prevotella, Streptococcus, and Veillonella, have also been detected in fecal samples from healthy individuals^26,27. In pathological conditions, like inflammatory bowel disease (IBD)²⁸ and cirrhosis patients²⁹, the gut shows a notable enrichment of oral microbes. These observations underscore the potential of oral microbes to shape gut ecology and systemic health, highlighting the need for more comprehensive genomic resources to elucidate their roles.

Taken together, significant knowledge gaps persist in the characterization of human oral and airway prokaryotes, including the characterization of uncultured microbial diversity, the taxonomic composition across distinct anatomical sites, the functional potential of resident microbes, and their relationships with disease and the gut microbiome. To help address these gaps, we present a high-resolution genome catalog of human oral and airway prokaryotes, generated by integrating 12,616 public metagenomic samples, 933 newly sequenced airway metagenomes, and 64,169 isolate genomes from relevant habitats. From these, we reconstructed 99,215 high-quality, non-redundant strain-level genomes, which were subsequently clustered into 2474 species-level clusters. Phylogenetic and functional analyses revealed extensive uncultured diversity and putative metabolic specializations. Leveraging this Oral and Airway Prokaryotic Genome Catalog (OAPGC), we profiled the prokaryotes across distinct ecological niches within the oral cavity and respiratory tract, and explored their associations with disease and geography. Finally, by incorporating the Human Gut Microbiome Reference (GMR)³⁰ and an additional 16,362 public fecal metagenomes, we further extended our analysis to explore the distribution and potential roles of oral prokaryotes within the gut environment.

Results

Construction of the OAPGC from global oral and airway metagenomes

To explore the microbial genome landscape of the human oral and airway microbiomes, we conducted an extensive search of publicly available metagenomic samples from human oral and airway sites. A total of 12,616 metagenomic samples were downloaded from 86 studies, representing individuals from 26 countries across 6 continents (Fig. 1a; Supplementary Fig. 1a; Supplementary Data 1 and 2). These samples span 15 oral cavity and 8 airway sites, predominantly including saliva (38.0%, n = 4798), dental plaque (17.4%, n = 2193), buccal mucosa (8.8%, n = 1104), bronchoalveolar lavage fluid (BALF) (5.9%, n = 748), and sputum (4.1%, n = 514). Using a unified quality control pipeline, these samples generated a total of 55.5 Tbp of non-human data, forming one of the largest human oral and airway microbiome datasets to date. In addition to the public datasets, we incorporated 933 newly sequenced airway metagenomes from the Breathing Well Pneumonia Research (BWPR) cohort of elderly individuals, including 535 oropharyngeal swabs and 398 sputum samples (Supplementary Fig. 1a). After applying the quality control procedures, these samples contributed an additional 6.4 Tbp of non-human data for downstream analysis.

Fig. 1. Overview of the OAPGC.

a Geographic distribution of the study cohorts. Different colors indicate distinct sample types. “Mixed” refers to samples composed of material pooled from multiple anatomical sites. b Workflow for the construction of the OAPGC. c Genome size (Mb) and quality scores of the OAPGC genomes. Colors represent genome sources: isolated genomes (green), high-quality metagenome-assembled genomes (HQMAGs; yellow), and shared genomes (blue). The histograms show the distribution of genome counts. d Proportions of genome integrity categories across different genome sources, based on the revised MIMAG standard. e Read mapping rates (%) of oral (n = 8712) and airway (n = 2329) samples against three genome sources. Mapping rates were calculated as the number of reads assigned to each genome by Kraken2 divided by the total number of reads in the sample.

The construction process of the OAPGC is illustrated in Fig. 1b. Initially, we performed a de novo metagenomic assembly for each sample, generating over 110 million contigs (>2 kb) in total. To maximize the recovery of microbial genomes and improve genome quality, we employed four tools (i.e., MetaBAT 2³¹, MetaBinner³², Semibin2³³, and VAMB³⁴) along with a recently developed multi-coverage binning algorithm based on MetaBAT 2 (namely MetaBAT2-multi)³⁵ for binning of contigs within each sample. In total, these methods produced 2,103,651 raw bins (>200 kbp in genome size) across all samples. Of these, 226,994 high-quality prokaryotic MAGs were obtained (completeness ≥90% and contamination <5%). MetaBAT2-multi and Semibin2 were the most effective algorithms, generating 77,901 and 66,581 high-quality genomes, respectively (Supplementary Fig. 1b). Subsequently, within-sample deduplication resulted in 94,637 high-quality MAGs (HQMAGs; Supplementary Data 3). MetaBAT2-multi and Semibin2 contributed the largest proportion to the final set of HQMAGs (Supplementary Fig. 1c).

To expand the genomic resources of prokaryotic genomes, we next downloaded 64,169 publicly available isolated genomes from the human oral and airway habitats, including 63,079 from the NCBI RefSeq database and 1090 isolated by Li et al.³⁶ After applying the same quality filtering criteria used for HQMAGs, we retained a total of 55,284 high-quality isolated genomes (Supplementary Data 4). An interesting observation is that most of these isolated genomes are clinical pathogens (e.g., Staphylococcus aureus, Streptococcus pneumoniae, and Pseudomonas aeruginosa) derived from the respiratory tract, with a considerable degree of redundancy (Supplementary Fig. 1d, e). To address this, we performed strain-level deduplication at an average nucleotide identity (ANI) threshold of 99.9% after merging the HQMAGs and isolated genomes (Fig. 1b). The final OAPGC consists of 99,215 prokaryotic genomes (Supplementary Data 5). Specifically, the HQMAGs and isolated genomes individually contributed 86,454 and 12,636 genomes to the OAPGC, respectively, with only 125 genomes shared between these two genome sources.

Although isolated genomes comprise a smaller proportion in the OAPGC, we found that their overall genome quality is significantly higher than that of HQMAGs (Fig. 1c; Supplementary Fig. 1f). Specifically, of the 18,351 genomes (18.5% of all genomes) in the OAPGC identified as near-complete using established standards^19,35, the majority (n = 10,836) were contributed by isolated genomes (Fig. 1d). These results suggest that the MAG strategy still has limitations in recovering high-quality prokaryotic genomes compared to traditional culturing methods. On the other hand, we observed that the size of isolated genomes is significantly larger than that of HQMAGs: over half (56.5%) of the isolated genomes having sizes greater than 3 Mbp, while the majority (95.0%) of HQMAGs are smaller than 3 Mbp (Fig. 1c; Supplementary Fig. 1g). This finding aligns with previous studies on the gut microbiota³⁷ and suggests that HQMAGs contain a substantial proportion of non-cultivable prokaryotic genomes with reduced genome sizes. Furthermore, we compared the read mapping rates of HQMAGs and isolated genomes across all samples. HQMAGs achieved a median mapping rate of 81.0% for oral metagenomes and 74.8% for airway metagenomes, significantly higher than those of isolated genomes having median mapping rates of 42.5% and 41.6%, respectively (Fig. 1e). This suggests that HQMAGs better represent the oral and airway microbiomes compared to isolates. The OAPGC, in turn, demonstrated even greater representativity, with read coverage rates for oral and airway metagenomes of 84.4% and 80.4%, respectively.

Taxonomic distribution of OAPGC species

The 95% ANI threshold is commonly regarded as the species boundary for prokaryotes³⁸. However, certain taxa, such as the recently described Candidate Phyla Radiation (CPR), may exhibit deviations from this convention³⁹. To refine the species boundary for oral and airway prokaryotes, we first annotated all OAPGC genomes at the phylum level using the Genome Taxonomy Database (GTDB)⁴⁰, followed by the calculation of pairwise ANI within each phylum. Our analysis revealed that for nearly all phyla, the ANI distribution of genomes exhibited a pronounced “species formation” peak between 95% and 100% ANI, with a distinct gap at 95%, suggesting that 95% ANI serves as a stable species boundary for these phyla (Supplementary Fig. 2a). However, we identified two phyla, Campylobacterota and Patescibacteria, where the peak of the pairwise ANI distribution occurred at 95%, rendering this threshold unsuitable as a species boundary. The ANI distribution gap for these phyla appears to be around 93% for Campylobacterota and 92% for Patescibacteria, indicating that these values may represent the true species boundaries for these groups. The gene-sharing ratio was defined as the proportion of genes shared between genome pairs relative to their total gene content, with higher ratios indicating closer relatedness. At these ANI thresholds, the corresponding pairwise gene-sharing ratios were approximately 75% and 68%, respectively, which also correspond to troughs in the gene-sharing ratio distribution, further supporting their validity as species boundaries (Supplementary Fig. 2b). Similarly, an analysis of all isolated genomes of Campylobacterota from the NCBI RefSeq database confirmed that 93% ANI represents a reliable species boundary (insufficient genome data for Patescibacteria precluded similar analysis) (Supplementary Fig. 2c).

Based on these refined parameters, the 99,215 OAPGC genomes were grouped into 2474 species-level clusters (hereinafter referred to as “species”). Rarefaction analysis indicated that the overall species richness had not yet reached saturation (Fig. 2a), suggesting that additional species likely remain to be discovered. However, most of these are rare taxa, as the accumulation curve began to plateau when considering only clusters with at least two conspecific genomes (which accounted for 63.7% of the total species). Taxonomic classification revealed that the 2474 species spanned 22 phyla, 32 classes, 79 orders, 149 recognized families, and 394 known genera (Supplementary Data 6). The vast majority of species belonged to Bacteria, with only three archaeal species identified—two from Methanobacteriota and one from Thermoplasmatota. Dominant bacterial phyla included Pseudomonadota (consisting of 24.5% of all species), Actinomycetota (22.1%), Bacillota (18.0%), and Bacteroidota (11.0%), followed by Bacillota_A (7.2%), Bacillota_C (4.5%), and Patescibacteria (3.2%) (Fig. 2b). Bacterial orders representing >5% of species include Lactobacillales (14.7%, primarily composed of Streptococcus), Enterobacterales (7.6%, mainly Haemophilus), Mycobacteriales (7.2%, mainly Mycobacterium), Actinomycetales (7.1%, mainly Pauljensenia, Rothia, and Actinomyces), Burkholderiales (7.0%, mainly Neisseria and Burkholderia), Bacteroidales (6.4%, mainly Prevotella and Porphyromonas), Coriobacteriales (5.3%, mainly Lancefieldella), and Pseudomonadales (5.0%, mainly Pseudomonas_E and Acinetobacter).

Fig. 2. Taxonomy of 2474 prokaryotic species in the OAPGC.

a Rarefaction curves showing the number of detected species as the number of prokaryotic genomes increases. Two curves represent all species (red) and species excluding singletons (clusters with only one genome) (blue). b Taxonomic annotation of prokaryotic species. The sunburst plot illustrates the taxonomic hierarchy, with sector sizes proportional to the number of species within each taxonomic rank. The middle circle indicates three sources of species, including species cultured from oral/airway or other habitats, as well as uncultured species. A species is considered cultured if at least one isolated genome from the NCBI database clustered within that species. The two outer rings represent the numbers of isolated genomes (inner ring) and HQMAGs (outer ring) assigned to each species. c Number of prokaryotic species from the three sources. d Distribution of species from the three different sources across phylum, class, order, family, and genus levels, ordered by the percentage of uncultured species in descending order. Taxonomic units containing more than 10 species are shown. The bar plot on the right displays the phylogenetic diversity of uncultured species.

Among the 2474 species cataloged in OAPGC, only 996 (40.3%) are represented by isolates from human oral and airway; the remaining 1478 (59.7%) are known exclusively from HQMAGs. To determine whether these species have been cultured from other niches, we compared the OAPGC species with all reference genomes in the NCBI RefSeq database. This search revealed 747 additional species (30.2% of the total) that matched genomes previously isolated from environments such as the human gut, vagina, or natural environments, leaving 731 species (29.5%) that remain completely uncultured (Fig. 2c). These uncultured species are distributed across all major phyla, with the largest numbers in Actinomycetota, Bacillota, Pseudomonadota, Bacillota_A, Bacteroidota, and Patescibacteria. Phyla that are particularly recalcitrant to cultivation include Bacillota_I (72.1% uncultured), Patescibacteria (60.0%), Campylobacterota (51.9%), and Spirochaetota (50.8%) (Fig. 2d). Phylogenetic diversity (PD), which measures the total branch length spanned by a set of taxa in a phylogenetic tree, was calculated for the 11 phyla represented by ≥10 species. Uncultured taxa contributed on average 52.5% of the total branch length (range 6.4–76.3%), highlighting their substantial contribution to the evolutionary diversity within these phyla. At lower taxonomic ranks, the OAPGC markedly broadens the known diversity of oral and airway prokaryotes. For instance, the order Acholeplasmatales (belonging to Bacillota_I)—frequently detected yet poorly studied in oral and airway microbiome surveys—is represented in OAPGC by 32 species, all of which are currently uncultured. Of these, 28 species were classified into a single novel genus “CALISC01” according to GTDB taxonomy. Phylogenetic tree analysis further revealed that this genus corresponds to Candidatus Bgiplasma, a lineage previously described by Jie et al.²³ (Supplementary Fig. 3). The presence of this genus in the upper aerodigestive tract suggests a potentially unrecognized ecological role that warrants further investigation. Similarly, the order Coriobacteriales, dominated by the genus Lancefieldella, comprises 179 OAPGC species, of which 118 (65.9%) remain uncultured. Given that L. rimae and L. parvula have been associated with periodontal niches and systemic conditions^41,42 the substantial number of uncultured Lancefieldella species identified here may represent a reservoir of yet-to-be-discovered pathogens or immunomodulators. Several other clinically relevant or poorly characterized taxa are also dominated by uncultured representatives, including Granulicatella, Oribacterium, JABCPE02 (a previously undescribed genus within Flavobacteriales), Nanosyncoccus, F0422 (a novel genus within Veillonellales), Treponema, and so on (Fig. 2d). Even within the well-studied clades such as Streptococcus, 73 of 265 species lack isolates; these uncultured species contribute 40.9% of streptococcal PD, underscoring their evolutionary distinctiveness.

In summary, the OAPGC not only expands the genomic inventory of the oral and airway microbiota but also pinpoints phylogenetic gaps that should be prioritized in future cultivation and functional studies.

Functional representation of OAPGC species

To characterize the functional landscape of OAPGC species, we annotated their protein-coding genes using the KEGG (Kyoto Encyclopedia of Genes and Genomes)⁴³ and eggNOG (evolutionary genealogy of genes: Non-supervised Orthologous Groups)⁴⁴ databases, achieving average annotation rates of 69.8% and 72.86%, respectively (Supplementary Fig. 4a). Notably, species from the phylum Patescibacteria exhibited a markedly lower proportion of annotated genes and a reduced number of detected functional orthologs compared to other phyla, underscoring their limited functional characterization. Principal coordinate analysis (PCoA) based on functional profiles revealed a clear separation among major phyla (Supplementary Fig. 4b), with phylogenetic affiliation explaining over 45.0% of the total variance (permutational multivariate analysis of variance [PERMANOVA] p < 0.001), highlighting taxonomy as a primary determinant of functional repertoire in OAPGC species.

The availability of high-quality genomes enables fine-scale functional characterization of OAPGC species. We found that, while core metabolic modules (e.g., glycolysis and glycogen biosynthesis) were conserved across all clades, some essential modules were notably absent in specific phyla (Fig. 3a; Supplementary Fig. 4c). For instance, Patescibacteria and Bacillota_I consistently lacked modules for biosynthesis of diverse amino acid, cofactor, and vitamin, as well as fundamental modules including pentose phosphate pathway, glycogen degradation, nucleotide sugar biosynthesis, and purine (e.g., adenine and guanine) metabolism. These deficiencies may be attributable to their reduced genome sizes and potential parasitic lifestyles, as previously reported⁴⁵. Besides, modules involved in the biosynthesis of essential amino acids, such as aromatic amino acids and branched-chain amino acids (BCAAs), were commonly found in Pseudomonadota and Actinomycetota, but were largely absent in several clades of Bacteroidota, Bacillota, and Bacillota_A. In addition, we examined clade-specific functions relevant to the oral microbiome (Supplementary Fig. 4d). For example, aromatic amino acid (AAA) degradation pathways—capable of producing pro-inflammatory compounds such as indoles and phenols⁴⁶—were enriched in specific clades of Pseudomonadota, Bacillota_A, and Bacteroidota, but were rare in other phyla. Short-chain fatty acid (SCFA) biosynthesis pathways, potentially linked to PD or altered taste perception⁴⁷, were also predominantly found in Bacillota, Bacillota_A, and Bacteroidota clades. Moreover, oxidative stress resistance genes implicated in oral inflammation⁴⁸ were mainly encoded by Pseudomonadota species and were significantly more prevalent in this phylum than in others.

Fig. 3. Functional annotation of OAPGC species.

Heat maps showing the percentage (%) of species with complete KEGG modules (a), CAZymes (b), and antibiotic resistance genes (ARGs) (c) across major families. The phylum to which each family belongs is indicated above the heat maps. Brown circles below denote the number of species within each family. 84 dominant families with at least 5 species or a relative abundance greater than 0.01% are displayed. KEGG models with a prevalence of 50–100% across samples and CAZymes with a prevalence of 50–90% are presented. For ARGs, those observed in at least 10 families are shown. Bac._C, Bacillota_C. Bac._I, Bacillota_I.

We next examined additional functional categories, including carbohydrate-active enzymes (CAZymes)⁴⁹, virulence-related genes, and antibiotic resistance genes (ARGs), which accounted for 15.9%, 0.28%, and 0.12% of total genes, respectively, across the OAPGC species. Among the CAZymes, those involved in complex polysaccharide degradation and peptidoglycan binding or hydrolysis were the most prevalent, though these functions were largely absent in Patescibacteria and Bacillota_I. (Fig. 3b). The most frequently detected ARGs conferred resistance to glycopeptides and were broadly distributed in clades of most phyla except Pseudomonadota (Fig. 3c). Conversely, genes associated with β-lactamase activity and resistance to quinolones and tetracyclines were predominantly found in Pseudomonadota clades, but were comparatively rare in other phyla.

Diversity, composition, and intraspecies divergence across habitat and geography

The OAPGC provides a powerful framework for high-resolution analysis of oral and airway microbiomes. To illustrate this, we compared OAPGC with the recently published HROM. Although OAPGC contains fewer species-level representative genomes than HROM²², 1,101 of the 2,474 OAPGC species were not found in HROM (Supplementary Fig. 5a), and OAPGC predicted a greater number of genes (Supplementary Fig. 5b), which may reflect the use of different ANI thresholds for Campylobacterota and Patescibacteria. In addition, OAPGC outperformed HROM in species-level mapping rates (%, Supplementary Fig. 5c). To further explore this, we profiled the microbiomes of 10,027 samples from 83 cohorts (each with a minimum of 10 samples), rarefying each sample to 1 million reads to mitigate sequencing depth effects, and then calculated relative abundances to investigate their diversity and compositional patterns across different anatomical habitats and geographical origins. Hereafter, we refer to saliva, sputum, BALF, and related materials as sample types representing distinct anatomical habitats. Our analysis revealed that saliva samples exhibited the highest within-sample species richness and diversity, followed by other oral samples (tongue, dental plaque, and tonsils), while the airway samples (sputum, BALF) showed significantly less diverse compared to the oral samples (Fig. 4a; Supplementary Fig. 5d). Interestingly, the between-sample diversity—reflecting interindividual variation within the same habitat—was notably higher in airway microbiomes than in oral communities (Supplementary Fig. 5e). These findings align with previous observations highlighting the ecologically rich and high interpersonal similarity of oral communities⁵⁰, further demonstrating that airway microbiomes appear taxonomically depauperate yet more variable across individuals. Moreover, despite these differences in taxonomy, we found that the functional richness and diversity (based on KEGG profiles) of airway microbiomes closely resembled those of oral communities, with a particularly higher richness of ARGs in airway samples compared to their oral counterparts (Fig. 4b; Supplementary Fig. 5f, g). This suggests that the airway microbiomes possess enhanced functional adaptability and a heightened propensity for therapeutic resistance. However, the functional composition remained highly heterogeneous between oral and airway habitats (Supplementary Fig. 5h), suggesting that similar functional potential is achieved through distinct functional repertoires across habitats.

Fig. 4. Characterization of prokaryotic species across habitats and geography.

a Shannon diversity index and richness (number) of prokaryotic species across projects, with colors representing oral and airway sites. b Richness (number) of prokaryotic ARGs in the different projects spanning different habitats. The central line in each box denotes the median, the upper and lower hinges correspond to the first and third quartiles, and whiskers extend to the smallest and largest values within 1.5× the interquartile range from the hinges. c Taxonomic composition of prokaryotic species at the phylum level (top) and genus level (bottom) in projects across the habitats. Taxa with relative abundance below 0.1% are grouped as “others.” Numbers in parentheses after project names indicate the number of samples for that site within the project. d Effect size (R²) of strain-level divergence in each species explained by geography and habitats, estimated using permutational multivariate analysis of variance (PERMANOVA). Only species with more than 100 strains are shown. Dot colors indicate the genus assigned to each species, and dot sizes correspond to the number of strains within each species. e Phylogenetic analysis of strains within F0428 sp003043955 SGB1675 (left) and Peptostreptococcus sp900759325 SGB1729 (right). Colors of the outer circles denote habitats (left) and geography (right), respectively.

We observed significant differences in taxonomic composition between various oral and airway sites, with more pronounced disparities at lower taxonomic levels (e.g., genus and species) (Supplementary Fig. 6a). Specifically, the phylum Actinobacteria (and its corresponding genus Actinomyces) was more abundant in dental plaque samples, with significantly lower abundance in other samples, while Corynebacterium was predominantly observed in dental plaque samples (Fig. 4c). Neisseria and Veillonella were more abundant in saliva and tongue samples but were present at much lower abundances in other samples. The airway microbiomes (sputum, BALF) were enriched with many clinically relevant genus such as Pseudomonas and Staphylococcus, which were found at low abundances in the oral samples. Notably, Streptococcus and Prevotella were the most abundant genera across all habitats, but their species composition varied significantly by site. For example, S. salivarius and S. parasanguinis predominated in saliva and tongue samples, while S. oralis and S. gordonii were more abundant in dental plaque, and two unidentified Streptococcus species (SGB0167 and SGB0168) were more prevalent in BALF (Supplementary Fig. 6b). Similarly, Prevotella species in saliva and tongue samples included abundant P. jejuni, P. histicola, and P. melaninogenica, while dental plaque was enriched with P. oris, P. nigrescens, and P. intermedia, and multiple Prevotella species were the most abundant in sputum and BALF (Supplementary Fig. 6c).

Compared to the substantial influence of habitats on the oral and airway microbiome variation (effect size >17% at different taxonomic levels, Supplementary Fig. 6a), geographical factors had a relatively minor but significant effect on microbiome composition (effect size = 3.6–4.5%, PERMANOVA p < 0.001, Supplementary Fig. 7a). Furthermore, in saliva and dental plaque (other samples lacked sufficient cross-reginal data for analysis), we also observed significant compositional differences between studies from Europe/North America and Asia (PERMANOVA p < 0.001 at different taxonomic levels), and these differences remained significant after adjusting for experimental heterogeneity across studies (Supplementary Fig. 7b). These findings indicates that geography does influence the oral and airway microbiomes, albeit to a lesser degree than habitat.

Beyond differences in community composition, certain species also exhibited intraspecies divergence potentially shaped by habitat or geography. Strain-level divergence was defined as the genetic distance among strains of a given species, reflecting the degree of sequence variation at the strain level. We assessed the extent to which these distances were associated with habitat or geographic origin. We identified 98 species whose strain-level divergence was strongly associated with habitat, and 113 species whose divergence was significantly linked to geography (Fig. 4d; Supplementary Fig. 7c; Supplementary Data 7). Representative habitat-associated divergence included F0428 sp003043955 SGB1675 (an unclassified Lachnospiraceae species), which displayed distinct strain populations between dental plaque and other samples (Fig. 4e). Representative geography-associated divergence included Peptostreptococcus sp900759325 SGB1729, whose strains from Asia were clearly differentiated from those in North America/Oceania. Such habitat- and geography-driven strain variation may reflect adaptive specialization, warranting further functional investigation.

Oral and airway microbial signatures are associated with various diseases and health

Previous studies have extensively demonstrated the interaction between the oral and airway microbiomes and various diseases^51–53; however, the lack of a unified reference genome database has limited integrated analyses. To address this, we examine microbiome data from 25 case-control comparisons across different oral or airway sites for 12 diseases to explore microbial signatures associated with these diseases. The diseases under investigation included PD (with 7 comparisons across 3 sites), dental caries (DC) (4 comparisons across two sites: saliva and dental plaque), RA (3 comparisons across 3 sites), respiratory infections (including upper respiratory infections [URI] from this study, pneumonia, and COVID-19 infections), COPD (2 sites from this study), and other systemic diseases such as colorectal cancer (CRC), pancreatic cancer (PC), type 1 diabetes mellitus (T1DM), and autism spectrum disorder (ASD). At the species level, we observed that reduced within-sample richness and diversity were most associated with infectious diseases, including pneumonia patients’ throat swabs, COVID-19 patients’ nasal swabs, and URI patients’ sputum (Fig. 5a). Furthermore, RA patients’ dental plaque and saliva, as well as ASD patients’ dental plaque, also exhibited significant reductions in within-sample richness and diversity. In contrast, PD patients’ dental plaque diversity showed an increase in specific instances. Next, we applied PERMANOVA to the species-level profiles to evaluate the effect size of each disease on the overall structure of the oral or airway microbiomes. In 19 of the 25 case-control comparisons, disease status was associated with significant compositional shifts (PERMANOVA p < 0.05), with effect sizes (R²) ranging from 0.2% to 13.9% (Fig. 5a), indicating generally modest but detectable effects. Pneumonia patients exhibited the most substantial variations in their airway microbiome, followed by PD patients’ dental plaque and saliva microbiomes. In contrast, the microbiomes of sputum and throat swabs from stable COPD patients, and saliva from T1DM patients, showed minimal or no detectable changes. Furthermore, we constructed random forest classifiers for each disease to assess the ability of oral or airway microbiome composition to distinguish between cases and controls. While discriminatory performance varied across datasets, 20 of the 25 comparisons achieved an AUC > 0.70 (Fig. 5a), with particularly high performance observed in pneumonia (AUC = 0.98) and several other diseases. These results suggest that microbiome composition contains disease-associated signals, although their generalizability may vary across cohorts.

Fig. 5. OAPGC microbes and diseases.

a Changes in microbial richness (number of species, left) and Shannon diversity index (second from left), effect sizes (R²) based on permutational multivariate analysis of variance (PERMANOVA, third from left), and area under the curve (AUC) of random forest models (right) between disease and control groups across 25 case-control comparisons. Different colors represent different diseases. Statistical significance for diversity changes was assessed by Wilcoxon rank-sum test. Significance of effect sizes was evaluated using the Adonis test with 1000 permutations. **p < 0.01, *p < 0.05. b Heatmap showing 96 prokaryotic signatures with consistent directional changes across multiple case-control comparisons. Heatmap colors represent the log2 fold change of average species relative abundance (%) in disease versus control groups. Species names in red indicate taxa predominantly increased in disease groups across multiple diseases, whereas names in blue indicate taxa predominantly decreased. Blue hollow pentagrams above the heatmap indicate species that can be validated by the meta-analysis. Statistical significance assessed by Wilcoxon rank-sum test with FDR correction is denoted as ⁺q < 0.01, *q < 0.05. c Bar plots showing variables significantly associated with the oropharyngeal and sputum microbiome after adjustment for age, sex, BMI, season of sampling, medication use, and other relevant covariates. Variables are ordered by adjusted R² from PERMANOVA, and colors indicate variable categories. d Collective explanatory power (adjusted R² from PERMANOVA) of significant variables within each category. P-values were calculated using PERMANOVA with 1000 permutations. **p < 0.01; *p < 0.05. DC dental caries, PD,periodontitis, URI upper respiratory infection, COVID coronavirus disease 2019, Pneu pneumonia, COPD chronic obstructive pulmonary disease, CF cystic fibrosis, PC pancreatic ductal carcinoma, CRC colorectal cancer, RA rheumatoid arthritis, T1DM type 1 diabetes mellitus, ASD autism spectrum disorder.

To further identify disease-associated microbial features, we performed Wilcoxon rank-sum tests across 25 case-control comparisons and identified 3462 significant differential signatures involving 1457 species (q < 0.05, FDR corrected; Supplementary Data 8), with the highest number of signatures observed in respiratory infections (URI and pneumonia) and RA (Supplementary Fig. 8a). In addition, we found 96 species that exhibited consistent trends across multiple diseases and anatomical sites, among which 94 showed a significant decrease in patients, while only 2 were significantly enriched (Fig. 5b). Forty-eight of the 94 decreased species were further validated by random-effect meta-analysis (p < 0.05, I² < 80%; Supplementary Fig. 8b), indicating moderate cross-study consistency. For example, Porphyromonas pasteri (SGB2264, SGB2401, and SGB2406) was depleted in three diseases (five comparisons), including PD, respiratory infections, and RA, consistent with previous studies^54–56. Similarly, Neisseria subflava SGB0638 showed decreases across these three diseases, again aligning with previous studies^54,57,58. These findings suggest that the depletion of certain commensal taxa may be a hallmark of oral microbial dysbiosis. Notably, several uncultivated species, such as Porphyromonas SGB0621 and Neisseria SGB0635, displayed similar decreasing trends across multiple diseases, underscoring the potential of OAPGC to uncover novel disease-associated taxa. To further uncover signatures that are consistently linked to specific diseases across various habitats, we focused on PD (plaque, saliva, tongue), RA (plaque, saliva, tonsil), and respiratory infections (URI, COVID-19, and pneumonia among three sample types). Wilcoxon rank-sum tests combined with meta-analyses identified 39 species as site-consistent indicators for PD (Supplementary Fig. 8c–e; Supplementary Data 9), including classic “red complex” members⁵⁹ Porphyromonas gingivalis SGB1520 and Tannerella forsythia SGB1522. In RA, 129 species—including Veillonella infantium_A SGB0504 and Veillonella sp902477915 SGB0505—showed consistent trends across different sites. In respiratory infections, shared microbial signatures included multiple Streptococcus species commonly enriched in pneumonia patients⁶⁰, among which two uncultivated Streptococcus taxa (SGB0232 and SGB0081) could be further validated by meta-analysis. Overall, these results indicate that the OAPGC provides a useful resource for exploring microbial taxa potentially linked to disease, including both well-characterized and previously uncultivated species.

Beyond disease associations, we further investigated the relationship between the airway bacteriome and overall health in the elderly using the BWPR cohort, with a focus on individual and environmental factors. The BWPR cohort recruited 409 relatively healthy individuals, including 401 oropharyngeal samples and 343 sputum samples. At enrollment, extensive metadata were collected through structured questionnaires, yielding 245 variables that could be grouped into 10 major categories: environment, demography, elderly health status, oral hygiene, lifestyle, clinical indices, diseases, vaccines, medicines, and symptoms. PERMANOVA analysis identified 40 variables in the oropharyngeal dataset and 44 variables in the sputum dataset that significantly explained variation in bacteriome composition (Supplementary Data 10). After adjusting for common covariates such as age, sex, BMI, and recent antibiotic use, 37 variables for the oropharyngeal and 43 variables for sputum samples remained significant (PERMANOVA, p < 0.05; Fig. 5c). Among oropharyngeal samples, clinical indices accounted for the largest proportion of explained variance (2.8%), followed by elderly health status (1.1%) and diseases (0.9%). In sputum samples, elderly health status contributed the greatest proportion (2.5%), followed by diseases (2.1%) and clinical indices (1.8%) (Fig. 5d). Within the clinical indices category, albumin (0.7% for oropharyngeal and 0.2% for sputum samples) and the FEV1 (0.5% and 0.6%, respectively) ranked among the top contributors, whereas instrumental activities of daily living (0.5% in both samples) was one of the most influential variables within the elderly health category. These factors reflect both systemic health and pulmonary function, suggesting that airway prokaryotic composition captures aspects of overall health status in older adults. Among disease-related variables, cardiovascular diseases (0.5% for oropharyngeal and 0.3% for sputum samples) ranked prominently, consistent with previous observations^61,62. Taken together, these findings highlight a close link between the airway microbiome and diverse health-related factors in the elderly.

Widespread, low-abundance oral and airway microbes in fecal metagenomes link to diseases

Previous studies have reported that microbes originating from the oral cavity and airway are frequently detected in the gut and may be associated with diseases such as cirrhosis and CRC^29,63. To explore this observation further, we first compared the set of species found in oral/airway and gut habitats by analyzing the OAPGC and the GMR catalog, which comprises 478,588 high-quality gut microbial genomes representing 4404 species from globally distributed samples⁶⁴. The overlap between the OAPGC and GMR was minimal, with only 390 species shared across the two catalogs (Fig. 6a). To further investigate the distribution of oral/airway species in the gut, we downloaded 16,362 fecal metagenomic samples worldwide (Supplementary Data 11) and profiled the 2,474 OAPGC species across these samples. The prevalence of OAPGC species was compared between fecal and oral/airway metagenomes to assess their site-specific preference. Most OAPGC species (87.7%) were more prevalent in oral/airway metagenomes, while 12.3% were more frequently detected in the gut, likely representing gut-adapted taxa occasionally captured from oral/airway samples (Fig. 6a). From the OAPGC-specific species, we identified 1336 species with high prevalence (>20%) as oral/airway-preferred microbes (Supplementary Data 12). Almost all (99.6%) of these species exhibited higher prevalence in the oral/airway than in the gut, while the majority (93.6%) were at least twice as prevalent in the oral cavity as in the gut. Still, their presence in the gut was notable, with a median prevalence of 7.2% (mean 14.3%), and 426 species (31.9%) were found in more than 20% of gut samples (Fig. 6b). Further quantitative analysis revealed that the abundance of oral/airway-preferred microbes in the gut was relatively low, with the mean relative abundance (MRA) being 9.0 × 10⁻⁷ (IQR 1.9 × 10⁻⁷–3.7 × 10⁻⁶) and the cumulative relative abundance being 0.72% (IQR 0.08–0.41%) across 16,326 fecal metagenomes. Most species had an MRA below 0.01%, with only a few species exceeding this threshold, including Streptococcus spp. (SGB0168, SGB0293, SGB0289, SGB0290, and SGB0298), Haemophilus_D SGB1219, Pseudomonas_E aylmerensis SGB1002, and Saccharimonas SGB0035.

Fig. 6. Oral and airway microbes in fecal metagenomes and diseases.

a Scatter plot showing prevalence (%) of prokaryotic species in oral/airway metagenomic samples and fecal metagenomic samples. The red vertical dashed line indicates a 20% prevalence threshold in oral/airway samples. Point colors represent whether species are unique to OAPGC (orange) or shared between OAPGC and GMR (green). Point sizes correspond to the mean relative abundance (MRA, %) of species in fecal samples. Histogram showing species numbers across varying prevalence rates (%) in fecal samples (right) and oral/airway samples (top). The Venn diagram in the upper right corner illustrates the overlap between OAPGC and GMR. b Scatter plot of prevalence (%) and mean relative abundance (%) of the oral/airway-preferred microbes in fecal samples. Adjacent bar plots show the frequency distribution of species. Colors indicate the genus to which each species belongs. Point sizes represent the MRA of species in fecal samples. The red dashed line denotes a 20% prevalence threshold in fecal samples. c Fold change in microbial Shannon diversity index (left) and cumulative abundance (%, second from left), effect sizes based on PERMANOVA (third from left), and AUC of random forest models (right) comparing disease and control groups across 43 case-control comparisons. Different colors represent different diseases. Cumulative abundance (%) was the sum of relative abundance (%) of the oral/airway-preferred microbes. Statistical significance for diversity changes was assessed using the Wilcoxon rank-sum test. Significance of effect sizes was evaluated using the Adonis test with 1000 permutations. **p < 0.01, *p < 0.05. BC breast cancer, CRC colorectal cancer, ACVD atherosclerotic cardiovascular disease, AF atrial fibrillation, BL bone mass loss, CA carotid atherosclerosis, HTN hypertension, pHTN pre-hypertension, OB obesity, MUO metabolically unhealthy obese, T2D type 2 diabetes, pT2D prediabetics, tnT2D treatment-naïve type 2 diabetics, AC acute biliary tract infection, AP acute pancreatitis, CD Crohn's disease, UC ulcerative colitis, IBS irritable bowel syndrome, PCOS polycystic ovary syndrome, AS ankylosing spondylitis, BD Behcet's disease, GD Graves' disease, MG myasthenia gravis, RA rheumatoid arthritis, SLE systemic lupus erythematosus, VKH Vogt -Koyanagi -Harada disease, COVID-19 coronavirus disease 2019, HIV human immunodeficiency virus infection, PT pulmonary tuberculosis, CKD chronic kidney disease, LC liver cirrhosis, ASD autism spectrum disorder, SCZ schizophrenia, PD Parkinson's disease.

Given their consistently low abundance, these oral/airway-preferred microbes are unlikely to be stably colonizing the gut. Instead, their detection in fecal samples likely reflects continuous transit through the gastrointestinal tract following oral or airway colonization. Consistent with this interpretation, we observed a strong positive correlation between the MRA of these 1336 species in oral/airway and fecal metagenomes (Spearman correlation coefficient ρ = 0.666, p = 4.3 × 10⁻¹⁷²), with fecal MRA generally around 1% of the corresponding oral/airway MRA for each species (Supplementary Fig. 9a). To further validate this trend at the individual level, we analyzed several datasets containing paired oral and fecal samples. The correlation between oral and fecal MRA remained consistently positive across datasets, with Spearman correlation coefficients ranging from 0.291 to 0.725 (all p « 0.001; Supplementary Fig. 9b). Notably, in one dataset, the salivary-fecal correlation (ρ = 0.563, p = 2.3 × 10⁻¹¹²) was stronger than that between dental plaque and feces (ρ = 0.291, p = 1.7 × 10⁻²⁷), suggesting that salivary microbes may contribute more directly to those detected in feces. Overall, these findings indicate that the fecal abundance of oral/airway-preferred microbes can partially reflect their abundance in the upper aerodigestive tract.

The distribution of oral/airway-preferred microbes in the gut may be influenced by host geography and health status. In healthy individuals, the richness, diversity, and cumulative abundance of these microbes were significantly higher in fecal samples from Europe and Oceania, and lower in those from Asia and North America (Supplementary Fig. 10a). Similarly, across individuals with disease, these metrics were consistently elevated in patients compared to healthy controls, regardless of geographic origin (Supplementary Fig. 10b), suggesting that the accumulation of oral/airway-preferred microbes in the gut may be a general feature of certain disease states. To further investigate this, we analyzed 6913 fecal metagenomes from Chinese individuals across 43 case-control comparisons covering 29 diseases (Supplementary Data 13). Consistent with the global analysis, 9 comparisons, spanning diseases including liver cirrhosis (LC), atherosclerotic cardiovascular disease (ACVD), CRC, systemic lupus erythematosus (SLE), chronic kidney disease (CKD), and so on, showed significantly increased diversity and cumulative abundance of oral/airway-preferred microbes in patients, while no comparison showed significant decreases (Fig. 6c). Additionally, the overall composition of these microbes significantly differed between cases and controls in 31 of the 44 comparisons (PERMANOVA p < 0.05). Random forest models based on the abundance of oral/airway-preferred microbes achieved AUC > 0.70 in 20 of 44 comparisons, underscoring their potential diagnostic value. In several CRC cohorts, these microbes enabled robust cross-cohort discrimination (average leave-one-study-out [LOSO] AUC = 0.860) and improved classification performance when combined with gut-resident species (average LOSO AUC increased from 0.896 to 0.911; Supplementary Fig. 10c–e). Collectively, these results highlight the potential of oral/airway microbes detected in feces as an indicator of host systemic disease.

Discussion

In this study, we constructed a comprehensive catalog of 99,215 high-quality oral and airway prokaryotic genomes. Taxonomy identification was conducted through phylum-specific ANI thresholds, revealing that the widely applied 95% cutoff is not universally appropriate across taxa. Functional analyses highlighted pronounced interspecies divergence, including traits with potential implications for host health. Community profiling across anatomical niches demonstrated clear site specificity and ecological partitioning. Linking oral and airway microbes to various diseases uncovered shared microbial signatures spanning multiple diseases and distinct anatomical sites. Finally, we traced the biogeographic footprints of oral microbes within the gut environment and explored their potential associations with geography and diseases.

In genomic analysis, a 95% ANI threshold is widely regarded as the standard species boundary⁶⁵. However, several studies^66–68 have argued that, due to the extensive diversity among prokaryotes⁶⁹, a universal cutoff may not be appropriate for all taxa. In this regard, by examining the pairwise ANI distribution within individual phyla, we found that lower thresholds—93% for Campylobacterota and 92% for Patescibacteria—could more accurately delineate species boundaries, a finding further corroborated by gene sharing ratio analyses. In light of this taxonomic analysis, the OAPGC identified 2,474 species, most of which are well-known commensals commonly found in the oral⁷⁰ and respiratory tracts⁷¹. Notably, three methanogenic archaea were also detected, consistent with prior reports of archaea in the oral cavity⁷² and airway⁷³, whose roles in human health merit further investigation. Moreover, while many species identified in the OAPGC have already been cultured from humans or environmental sources, 29.5% remain uncultured, highlighting a vast reservoir of unexplored microbial diversity in the oral and airway habitats. The substantial evolutionary divergence of these uncultured taxa from known species suggests a potential for unique biological functions, which may play critical yet underappreciated roles in host health and disease. Notably, the genome sizes of HQMAGs were generally smaller than those of isolated genomes. Such genome reduction may indicate a reliance on close associations with hosts or other microbes for metabolic complementation⁷⁴. This metabolic dependency likely accounts for the difficulty in obtaining pure cultures. A representative example is Patescibacteria, whose members typically harbor ultrasmall genomes and remain functionally poorly characterized⁷⁵.

Furthermore, we profiled detailed functional attributes of OAPGC clades, revealing taxonomic affiliation as a primary driver of functional divergence. Specifically, certain taxa such as Bacteroidota, Bacillota, and Bacillota_A lack key biosynthetic pathways for AAAs and BCAAs, likely reflecting adaptive strategies to their respective environments⁷⁶. Consistent with previous observations, some airway-associated bacteria have lost amino acid biosynthetic genes, possibly due to their reliance on host-derived⁷⁷ or microbially derived precusors⁷⁸. In contrast, clades such as Pseudomonadota harbor distinctive functions like AAA degradation pathways and resistance to oxidative stress, which may support bacterial survival in niches⁷⁹. Importantly, these traits can adversely affect host physiology, for instance by promoting inflammation, underscoring their potential as candidate disease biomarkers^80,81 or interventional targets⁸². The underlying mechanisms remain to be elucidated, while some subsets of Pseudomonadota have been reported to upregulate vesicle-biogenesis genes to cope with oxidative stress⁸³.

Using the OAPGC, we characterized the taxonomic and functional features of microbiota across distinct niches spanning the oral cavity to the respiratory tract. Notably, although the oral microbiome was suggested as a major source for respiratory microbes⁸⁴, airway microbiota exhibited substantially greater inter-individual heterogeneity. This heterogeneity likely reflects multiple selective filters from the oral cavity to the lower respiratory tract, including microbial immigration (e.g., microaspiration), elimination (e.g., cough), and local microenvironments (e.g., oxygen tension, pH, and nutrient availability)⁸⁵. For example, Zhang et al. reported pronounced inter-individual variation in oral-to-lung microbial transmission, with a bimodal distribution, indicating that distinct oral input patterns shape lung microbiome composition⁸⁶. In addition, respiratory samples harbored a richer reservoir of ARGs, likely reflecting frequent antibiotic exposure associated with respiratory infections. Although oral microbiota are also affected by such treatments, the intensity and duration of exposure may be lower. Moreover, the respiratory tract is more chronically exposed to environmental stressors, such as cigarette smoke and biofuel emissions, which may further shape the respiratory resistome⁸⁷. Beyond anatomical sites, geographical factors also shape both oral and respiratory microbiomes, consistent with previous reports^13,88. These geographic influences may act through variations in diet, lifestyle, and other environmental exposures⁸⁸. The application of OAPGC also extends to elucidating the associations between oral–airway microbiomes and different diseases. Some uncultured species, such as Porphyromonas SGB0621 and Neisseria SGB0635, were found to be depleted across multiple diseases, illustrating the value of OAPGC as a resource to uncover new disease-associated microbial signatures.

Oral microbes are frequently introduced into the gut but are typically restricted by oral–gut barriers including gastric acid, bile, digestive enzymes, and colonization resistance²⁵. Consistent with this, we found that most oral/airway-preferred species were detectable in the gut but at extremely low abundance. Their low abundance further suggests that these species are more likely to undergo transient passage through the gut rather than establishing stable colonization. Moreover, their gut-level MRA was positively correlated with that in the upper aerodigestive tract, indicating their presence in the gut may passively reflect their abundance in upstream sites. Intriguingly, in multiple disease cohorts, oral/airway-preferred microbes showed increased diversity and cumulative abundance in the gut compared to controls, indicating a potentially conserved response to host health status. However, Liao et al.⁸⁹ has suggested that this phenomenon is not due to an actual increase in the transmission of oral species to the gut, but rather to a reduction in gut species, warranting further investigation. Furthermore, in the CRC cohorts, incorporating the abundance of oral/airway-preferred microbes into gut microbiome-based models improved disease discrimination, highlighting the value of profiling oral/airway-preferred microbes in gut samples to improve the diagnostic potential of gut microbiome datasets.

There are some limitations to this study. First, the number of respiratory tract samples remains relatively limited compared with oral samples, which may affect the representativeness of respiratory prokaryotic diversity. Second, although OAPGC integrates a large number of publicly available samples, it may still lack adequate representation of certain regions and populations, such as samples from the lower respiratory tract of healthy individuals and those from Africa. Third, while disease-associated signals were analyzed across multiple cohorts, further validation in independent longitudinal and mechanistic studies is needed. Lastly, the analysis of oral and airway microbes in the gut is constrained by the limited availability of paired oral–gut samples and the cross-sectional nature of the data. Experimental validation is also needed to confirm microbial transmission and its relevance to disease. Although OAPGC does not provide cultured bacterial isolates, it offers a comprehensive genomic resource that can facilitate future cultivation efforts, functional characterization, and the identification of disease-associated biomarkers or therapeutic strategies.

Methods

Ethical statement

The establishment of the Breathing Well Pneumonia Research (BWPR) longitudinal cohort was conducted in strict accordance with the principles of the Declaration of Helsinki and the International Conference on Harmonization Good Clinical Practice (ICH-GCP) guidelines. This study was approved by the Clinical Research Ethics Committee of China-Japan Friendship Hospital (Approval No. 2023-KY-066-1), and written informed consent was obtained from all participants prior to their inclusion in the study.

Self-collected samples

We collected 933 respiratory samples from the BWPR cohort, a longitudinal study designed to track episodes of acute respiratory tract infections (ARTIs) in relatively healthy older adults. Participants were recruited from Continuing Care Retirement Communities (CCRCs) located in the suburban areas of Beijing and Shanghai. Inclusion criteria were: (1) age ≥60 years, and (2) an expected residential duration of at least three years. Exclusion criteria included: (1) history of ARTI in the past three months, and (2) estimated life expectancy of less than six months. Oropharyngeal swabs and induced sputum samples were collected from participants at health baseline and during ARTI episodes. Demographic information, health status, and underlying diseases were obtained through face-to-face interviews and review of medical records. DNA extraction, library preparation, and metagenomic sequencing were performed using the Qiagen DNeasy PowerSoil Pro Kit, QIAseq FX DNA Library Kit, and Illumina NovaSeq X Plus PE150 platform, respectively.

Metagenomic data collection and processing

We systematically searched for whole-metagenome shotgun sequencing of human samples originating from the respiratory tract (e.g., BALF, sputum, nasopharyngeal swabs, cough swabs, etc.) and the oral cavity (e.g., saliva, dental plaque, tongue swabs, buccal mucosa, oropharyngeal swabs, etc.) up to June 2024. The corresponding raw metagenomic sequencing data were retrieved from publicly available repositories, including the Sequence Read Archive (SRA) hosted by NCBI, the European Nucleotide Archive (ENA), and the China National GeneBank (CNGB). Quality control was carried out using fastp v0.23.4⁹⁰ to remove low-quality reads and adapter contamination. Specifically, the parameters ‘-u 30 -q 20 -l 60 -y --trim_poly_g’ were applied to samples with read lengths ≤100 bp, and ‘-u 30 -q 20 -l 90 -y --trim_poly_g’ were used for those with read lengths >100 bp. The BBTools suite v39.00⁹¹ based on the bbduk algorithm, was used to further eliminate low-complexity reads. Non-microbial sequences were removed by aligning reads against the CHM13v2.0 human genome and the Escherichia phage phiX174 genome using Bowtie2 v2.4.4⁹² with the parameters ‘--end-to-end --mm --fast --no-head’. The unmapped reads were then subjected to de novo assembly using MEGAHIT v1.2.9⁹³, with the parameters ‘--k-list 21,41,61,81,101,121’ to identify the optimal assembly configuration.

Metagenomic binning with multiple tools

After assembly, contigs shorter than 2000 bp were discarded. The remaining contigs were subjected to metagenomic binning using five different approaches, including one multi-sample strategy and four single-sample strategies. MetaBAT2-multi, a mash-based multi-sample coverage binning strategy based on MetaBAT2 v2.15³¹, was implemented following our previously published method³⁵. First, pairwise Mash distances were calculated among all assemblies using Mash v2.3⁹⁴ with the parameters ‘-s 100.000 -k 32’. Next, for each sample, the 20 most similar samples were selected based on Mash distance, as this number of closest samples yields optimal performance for multi-coverage binning³⁵. Reads from these selected samples together with the target sample’s own reads were then mapped to its contigs. Sequencing depth profiles were calculated using the jgi_summarize_bam_contig_depths tool³¹, and the resulting depth files were merged using combine.pl (https://github.com/WatsonLab/single_and_multiple_binning/blob/main/scripts/combine.pl). These multi-sample depth profiles were then used for metagenomic binning with MetaBAT2³¹ using the parameters ‘-m 2000 -s 200000’. In addition, four single-sample binning methods were applied based on individual depth profiles: MetaBAT2 (‘-m 2000 -s 200000’), MetaBinner v1.4.4³² (‘-s large’), SemiBin2 v2.0.2³³ (‘easy_bin --minfasta-kbs 200 –environment human_oral -m 2000’), and VAMB v3.0.2³⁴ (‘-e 100 -q 20 40 60 80 –minfasta 200000’).

Genome quality assessment and de-redundancy

All binned genomes were evaluated using CheckM2 v1.0.1⁹⁵ to estimate completeness and contamination, and GUNC v1.0.5⁹⁶ was applied to assess potential chimerism. Genomes with completeness ≥90%, contamination <5%, and a clade separation score <0.45 were classified as High-quality metagenome-assembled genomes (HQMAGs) and retained for downstream analyses. The strain-level de-redundancy within each sample was performed using dRep v3.5.0⁹⁷ with the parameters ‘–pa 0.9 –sa 0.999 –nc 0.3 --S_algorithm fastANI’, corresponding to a threshold of 99.9% average ANI.

Incorporation of cultured genomes and genome integrity evaluation

Cultured genomes from the human oral and respiratory tracts were downloaded from the NCBI RefSeq database and the study by Li et al.³⁶, and subjected to the same quality control procedures to identify high-quality genomes. These genomes were then combined with HQMAGs and dereplicated at the strain level using dRep v3.5.0⁹⁷ with a 99.9% ANI threshold. Altogether, 99,215 prokaryotic genomes were compiled to establish the OAPGC. To assess these genomes' integrity, evaluations were performed based on the revised MIMAG standard⁹⁸. HQMAGs containing 5S, 16S, and 23S rRNA genes, along with at least 18 types of tRNA, are defined as near-complete MAGs. The identification of rRNAs was performed using two methods: ‘cmsearch’ implemented in INFERNAL v1.1.4⁹⁹ and barrnap v1.10.6 under default parameters. A genome was considered to contain a given rRNA type if that type was detected by either method. tRNAs were annotated using tRNAscan-SE v2.0.11⁹⁹ with the parameters ‘–L –B’.

Taxonomy, clustering and phylogenetic analysis

Phylum-level taxonomic annotation of prokaryotic genomes in the OAPGC was performed using the classify_wf model from the Genome Taxonomy Database Toolkit (GTDB-Tk) v2.3.2 with the GTDB r214.1 database⁴⁰. Pairwise ANI between genomes within each phylum was calculated using FastANI v1.33³⁸ with parameters ‘--minFraction 0’. Gene-sharing ratios were computed using MMseqs2 v2.17¹⁰⁰ with parameters ‘--min-seq-id 0.5 --cov-mode 1 -c 0.9 --cluster-mode 2 --cluster-reassign 1 --kmer-per-seq 200 --kmer-per-seq-scale 0.8’. All prokaryotic MAGs were then clustered into species-level genome bins (SGBs, corresponding to species) using dRep v3.5.0⁹⁷, with parameters ‘-pa 0.9 -sa 0.93 -nc 0.3 --S_algorithm fastANI’ for Campylobacterota, ‘-pa 0.9 -sa 0.92 -nc 0.3 --S_algorithm fastANI’ for Patescibacteria, and ‘-pa 0.9 -sa 0.95 -nc 0.3 --S_algorithm fastANI’ for all other phyla, according to the intra-phylum ANI distribution. A total of 2,474 species were finally obtained. Within each SGB, the near-complete MAG with the highest quality score was selected as the representative genome. Taxonomic assignment of species was based on the GTDB-Tk v2.3.2¹⁰¹ annotation as described above. Phylogenetic tree construction was performed using Phylophlan 3 v3.0.3¹⁰², and the resulting tree was visualized using iTOL v6.7.4 (https://itol.embl.de/)¹⁰³. Phylogenetic diversity for each phylum was calculated based on the tree using the ‘pd’ function in the R package picante v1.8.2.

Functional annotation

For prokaryotic genomes in the OAPGC, gene prediction was performed using Prodigal v2.6.3¹⁰⁴ with the parameter ‘-p meta’. The predicted proteins were then aligned against multiple functional databases using DIAMOND v2.0.13¹⁰⁵ against the KEGG database⁴³ with the parameters ‘--query-cover 50 --min-score 60’, CAZy⁴⁹ with ‘--query-cover 50 --min-score 60’, and Virulence Factor Database (VFDB)¹⁰⁶ with ‘--id 90 --subject-cover 80’. The predicted proteins were assigned to the respective functional orthologs or virulence factors based on the best-hit gene. Functional annotation was further performed via eggNOG-mapper v2.1.12 against eggNOG⁴⁴ using the option ‘--query_cover 50 --score 60’. In addition, the putative proteins were screened against the Comprehensive Antibiotic Research Database (CARD) using Resistance Gene Identifier (RGI) v6.0.3 (https://github.com/arpcard/rgi)¹⁰⁷, applying the setting ‘--local --clean -t protein --include_loose -a DIAMOND’.

Quantification of the prokaryotes and functional family

We constructed a custom Kraken2 database by integrating all prokaryotic species from the OAPGC, all viral operational taxonomic units (vOTUs) from the oral and airway viral genome catalog (OAVGC, a self-constructed resource established in our parallel study), and the human reference genome CHM13v2.0. Clean reads from all metagenomic samples were aligned to this custom database using Kraken2 v2.1.3¹⁰⁸ with a confidence score threshold of 0.1. Bracken v2.8¹⁰⁹ was then used to estimate the number of reads assigned to each species. To control for sequencing depth variation, each sample was rarefied to 1 million reads. To reduce false positives from extremely low-abundance species, any species with fewer than 10 mapped reads in a sample was set to zero, meaning that a species was considered present only if it had at least 10 mapped reads. Relative abundances were then calculated by dividing the number of reads assigned to each prokaryotic species by the total number of reads assigned to all prokaryotes in that sample. For functional gene quantification, we adopted a weighted abundance approach. Specifically, the weighted abundance of a given functional ortholog (e.g., KEGG ortholog) in a sample was estimated by multiplying the number of genes annotated with that ortholog in each species by the relative abundance of the corresponding species, and then summing these values across all species.

Gut data collection and processing

We downloaded the GMR⁶⁴, which comprises 6664 human gut microbial species-level genomes. Based on the high-quality genome assessment criteria described above, 4404 high-quality genomes were retained. These genomes were annotated at the phylum level using GTDB-Tk v2.3.2. The 4404 gut genomes were then dereplicated together with the OAPGC species using dRep v3.5.0, applying the same ANI threshold as described above, resulting in a collection of 5849 genomes. A new custom Kraken2 database was constructed, incorporating these 5849 genomes, the OAVGC genomes, and the human reference genome CHM13v2.0. In addition, 16,362 human fecal metagenomic sequencing samples were retrieved from NCBI, ENA, and CNGB (Supplementary Data 11). After processing the raw data using the aforementioned pipeline, the resulting clean reads were mapped against the new Kraken database using Kraken2 v2.1.3 and Bracken v2.8 to profile the prokaryotic species. Normalization and relative abundance calculations were carried out following the procedures outlined above.

Statistical analysis

All statistical analyses and visualizations were performed on the R v4.3.1 platform. Only projects with more than 10 samples were included in downstream analysis.

The Shannon index was calculated via the diversity function in the vegan v2.6.8 package¹¹⁰. The number of observed prokaryotic species was defined as the count of species with relative abundance greater than zero. Beta diversity was estimated using Bray–Curtis distances computed with the vegdist function in vegan v2.6.8¹¹⁰. PCoA was conducted based on the Bray–Curtis distance matrix using the pco function from the ade4 v1.7.22 package¹¹¹. PERMANOVA was performed using the adonis2 function in vegan v2.6.8¹¹⁰ with 1,000 permutations, and the adjusted R² values were obtained with the RsquareAdj function. Wilcoxon rank-sum tests were applied using the wilcox.test function for comparisons of diversity metrics between groups. In PERMANOVA analyses assessing geographic effects, sequencing platform and DNA extraction methods were included as covariates for adjustment. To assess the effects of geography and habitat on strain-level divergence, we selected SGBs containing more than 100 strains. PERMANOVA was performed on pairwise ANI matrices, adjusting for habitat when testing geography, and vice versa.

For case–control comparisons, several cohorts included multiple sample sites from the same participants (e.g., BWPR: oropharynx and sputum; PRJEB6997: dental plaque and saliva; PRJNA932553: dental plaque and saliva; PRJNA678453: dental plaque, saliva, and tongue; PRJNA752888: dental plaque and saliva). For the purposes of analysis, each sample site was treated as a separate case–control comparison, while noting that comparisons from different sites of the same participant are not fully independent. For case–control classification using microbial relative abundance profiles, random forest models were constructed with the randomForest function in the randomForest v4.7.1.1 package¹¹², employing five repeats of fivefold cross-validation. Model performance was evaluated using the receiver operating characteristic curve (ROC) via the roc function from the pROC v1.18.5 package¹¹³. Differential microbial species were identified using Wilcoxon rank-sum tests, and false discovery rate (FDR) was controlled using the Benjamini-Hochberg method implemented in the p.adjust function. To assess directional consistency across diseases and cohorts, we counted how often each species was enriched in cases versus controls across all comparisons. A species was considered to show a consistent trend if it was enriched in one group (cases or controls) at least four more times than in the other group. These consistent signals were further subjected to cross-disease meta-analyses of microbial species, following approaches described in our previous study¹¹⁴ using a random-effects model. Relative abundances of prokaryotic species were log₁₀-transformed, and Hedges’ g effect sizes were calculated using the escalc function in the metafor v4.6.0 package¹¹⁵. Between-study heterogeneity was assessed with the rma function in metafor, reporting Cochrane’s Q test and the I² statistic. P values were derived from the random-effects model.

In analysis of factors contributing to airway bacteriome variance, variables with missing values exceeding 10% of the sample size were removed. The relationship between each variable and the prokaryotic composition was assessed using the adonis2 function with 1,000 permutations, and R² values were adjusted using the RsquareAdj function in vegan v2.6.8¹¹⁰. For each category of variables, only variables significant in univariate analyses (p < 0.05) were included in the PERMANOVA analysis. To reduce potential confounding, the PERMANOVA analyses were additionally adjusted for age, sex, BMI, recent antibiotic use, and other relevant covariates.

In the oral-gut analysis, prevalence was defined as the number of samples in which a species with relative abundance greater than zero was detected. Oral/airway-preferred microbes were defined as species with prevalence >20% that were exclusively present in the OAPGC. Spearman correlations were calculated using the cor.test function with the parameter ‘method = spearman’. To explore the correlations of oral/airway-preferred microbes between paired oral and gut samples from the same individuals, we analyzed data from 1075 publicly available paired oral-fecal samples.

Author contributions

Dr X.Z., Q.Y. and Prof. B.C. had full access to all the data in the study and took responsibility for the integrity of the data and the accuracy of the data analysis. X.Z., Z.W., Q.Y. and B.C. designed the study. Q.Z., K.C., J.Y. and N.G enrolled the participants. Q.Z., Y.N. and X.Y. collected and processed the samples. X.Z., Y.N., S.L., Y.Z., C.W., B.L., R.G., G.X. and H.Y. acquired, analyzed, or interpreted the data. X.Z. and B.C. obtained the funding and provided administrative, technical, or material support. All authors read and approved the final manuscript.

Data availability

The raw metagenomic sequencing data generated in this study have been deposited in ENA under the project ID PRJEB96268 (https://www.ebi.ac.uk/ena/browser/view/PRJEB96268). The data files associated with the OAPGC, including prokaryotic genome sequences, annotation files, and the Kraken library, have been deposited in the Zenodo repository with the accession ID Zenodo 19212794 (https://doi.org/10.5281/zenodo.19212794), 19220405 (https://doi.org/10.5281/zenodo.19220405).

Code availability

The metadata, intermediate results, and analysis and visualization codes used in this study have been made publicly available in the GitHub repository, accessible at https://github.com/NiYawen/OAPGC.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41522-026-01042-3.