Our primary aim was to characterize the functional and phylogenetic profiles of the subgingival microbiome in individuals with periodontitis. To do this, we obtained 72 million classifiable sequences from deep periodontal pockets of 59 systemically healthy subjects with periodontitis, and 31 million sequences from 25 periodontally healthy individuals. All subjects with periodontitis were classified as Stage 3 based on disease severity and complexity (Table 1), with 17 patients demonstrating the molar-incisor phenotype (equivalent to LAP phenotype). 4 subjects were classified as Stage 3 grade A (S3gA), 22 as S3gB and 33 as S3gC. 25 subjects were classified as Chronic Periodontitis (CP, age range: 56-61 years), 17 as Generalized Aggressive Periodontitis (GAP, age range:24-32 years ) and 17 as Localized Aggressive Periodontitis (LAP, age range: 15-19 years). Their sequences represented 8336 functionally annotated microbial genes, and 454 taxa.
The ‘Anna Karenina’ effect is evident in periodontitis: We began our analysis by creating a catalog of disease-associated genes. Irrespective of disease phenotype, 73.8% of species and 60% of genes were identified in 2 or more individuals in each disease state. We then queried whether this ‘disease profile’ was different from health. We first used a Nonmetric Multidimensional scaling of Bray-Curtis pairwise dissimilarities to compare the phylogenetic as well as the functional signatures of 59 periodontitis samples with data previously collected from 25 medically (American Society of Anesthesiologists classification - ASA 1) and periodontally healthy individuals. The difference between health and each disease phenotype was significantly greater than differences between any two individuals with disease (Figure 1A and 1B, p<0.001, PERMANOVA).
Disease-associated microbiomes demonstrated greater beta-dispersion[39], with only 47% of disease-associated metagenome being shared by 80% or more of individuals with periodontitis (common core metagenome). On the other hand, over 73% of genes were identified in the common core of periodontal health, indicating that dysbiotic periodontitis-associated microbiomes exhibit the so-called Anna Karenina effect. We trained a Random Forest Classifier (RandomForest package in R) to test the impact of the Anna Karenina effect on differences between health and disease. The classifier was able to predict disease with 87% sensitivity and 91% specificity based on functional profiles and with 82% sensitivity and 88% specificity when using phylogenetic metrics (Figures 1C and 1D). Overall, 28% of genes were uniquely observed in disease, and 12% were unique to health, while 26% exhibited significant differential abundances in health and disease.
Disease phenotype explains microbiome variance better than disease grade: Having established that periodontitis differed significantly from health both taxonomically and functionally, we next investigated if differences could be discerned within the periodontitis-associated microbiome using unsupervised cluster analysis (k-means clustering). Bray-Curtis dissimilarity distances (computed from the relative abundances of genes and species in subjects with disease) were used as input and Silhouette width used to estimate number of clusters. We identified three distinct clusters taxonomically and functionally (p=0.0008 and 0.001 respectively, ADONIS test of Bray-Curtis Dissimilarity Index, Figures 2A and 2B). We then investigated the factors that drove these differences using a between-class analysis method that combines principal coordinates analysis with linear discriminant analysis (Figures 2C-2J). Relative abundances of genes and species in subjects with disease were used as input. Disease phenotype, ethnicity and age emerged as the strongest drivers of clustering. Disease phenotype yielded the lowest degree of misclassification both taxonomically and functionally while significant misclassification was evident when using disease grade as a discriminant. Furthermore, disease phenotype explained the strong ethnicity and age-based clustering, since most of the younger individuals and those of African American ethnicity belonged to the LAP group (Figure 2I).
Same players, different teams: Since targeted microbial investigations have previously suggested that localized aggressive periodontitis has a distinct microbial profile while the microbiota of chronic and generalized aggressive periodontitis are similar[40], we tested the hypothesis that GAP and CP are microbially similar while LAP is a taxonomically distinct entity. NMDS revealed significant class separation between the 3 diseases (p<0.001, PERMANOVA of Bray-Curtis Dissimilarity Index, Figure 2B). We then investigated whether GAP and CP are microbially more similar than GAP and LAP by computing pairwise dissimilarities (Bray-Curtis) between each GAP and CP subject, as well as each GAP and LAP subject. Taxonomically, GAP was more similar to LAP than to CP (p<0.001, Dunn’s test for multiple comparisons).
Since the NMDS indicated that these three diseases were microbially discrete entities, we examined the taxonomic features that contributed to class separation. LAP exhibited significantly lower alpha diversity (as measured by the ACE and Chao 1 indices) than the other groups (p<0.05, Dunn’s test, Figure 3A), and the three groups also demonstrated significant differences in beta diversity. While all 3 diseases were dominated by gram-negative anaerobic bacteria, (representing 56.8%, 62.5% and 47.9% of the abundance in CP, GAP, and LAP, respectively (Figure 3B)), the abundances of these groups were significantly greater in GAP when compared to LAP (p=0.03, Wilcoxon nonparametric test). By contrast, gram-positive anaerobic bacteria were significantly higher in CP when compared to either GAP or LAP (p<0.04, Wilcoxon).
Interestingly, 349 out of 416 species were identified in all the three diseases, and only 28 species were unique to any one of the three diseases. Collectively, the unique species constituted less than 0.03% of the abundance in each group. Each disease condition had core taxa that constitute more than 50% of the identified taxa in the condition (Figure3C i-iii). When core taxa of each condition were compared, most species were present in the cores of all 3 conditions (Figure 3C iv). 138 OTUs were found to be significantly differentially abundant between any two disease states (p<0.05, FDR adjusted Wald test - Figure 3E and Supplementary Table 1). Using similarity percentages (SIMPER) analysis, we identified OTUs that explained 70% or more of the class separation. Several OTUs which significantly contributed to the separation were also common core taxa, demonstrating that the diseases differ in the ratios of their predominant shared taxa. 107 OTUs were significant contributors to the separation between GAP and LAP; of these, 65 were part of the common core species of GAP, 63 formed the common core of LAP (Figure 3E, supplementary Table 1). Similarly, 50 species, 39 of which were members of the core microbiomes of GAP and CP, contributed to the separation between GAP and CP. The separation between CP and GAP was driven by Aggregatibacter actinomycetemcomitans, Fusobacterium nucleatum, Treponema socranskii, and several members of the genera Actinomyces, Campylobacter, Prevotella and Capnocytophaga. The separation between LAP and GAP was mainly driven through Porphyromonas gingivalis and members of the genera Neisseria, and Actinomyces.
Since inter-bacterial interactions play a large role in influencing microbial assemblages, we used graph theoretics to assess connectivity between species. The underlying rationale for this analysis is that taxa with the strongest connections demonstrate superior adaptation to their niche. As a corollary, diseases that present similar microenvironments will demonstrate greater co-dependency between member species than diseases that are different. The network topography is summarized in SupplementalTable 2. While GAP and LAP demonstrated robust hubs with 3568 and 2114 edges, CP presented a sparse topography, with only 489 connections, attesting to its phylogenetically idiosyncratic presentation (Figure 4). Zi Pi plots of both CP and LAP demonstrated expansive nodes with several putative keystone species in the network topography, while the node distribution in GAP was equitable (preventing us from creating a Zi Pi plot) and did not demonstrate any candidate keystone species. Together the data suggest that patients with GAP and LAP have a more homogeneous subgingival micro-environment than those with CP, which may explain the taxonomic heterogeneity observed in CP. Based on the clinical observation that 35% of untreated cases of LAP progress to GAP[41], we hypothesize that loss of the influential key players found in LAP creates a state of flux that, when observed cross-sectionally, gives rise to the observation that GAP is a distinct disease phenotype. This theory is further supported by observations that individuals with GAP demonstrate low serum antibody response to the microbial constituents, leading to its continuous periodontal destruction(40). This is unlike the other two phenotypes which can self-arrest with time.
The LAP microbiome is functionally distinguishable from CP and GAP: Since the three disease phenotypes demonstrated several taxa in common, we tested the hypothesis that there would be significant functional overlap in their respective associated microbiomes using the SEED ontology to annotate genes and the KEGG database for pathway identification. A greater degree of class separation was evident based on functional capabilities than on taxonomic profiles (p<0.0001, PERMANOVA of Bray-Curtis Dissimilarity, Figure 5A). 20.61% of the LAP metagenome (1278/6200 genes) was unique, in that, these genes were not present in either GAP or CP (Figure 5B and Supplemental Table 3). 20% of these unique genes did not have functional role assignments, pointing to gaps in our knowledge of the microbiome of localized aggressive periodontitis. 40% of unique genes encoded enzymes for anaerobic degradation of aromatic compounds, methanogenesis, lysine and Acetyl CoA fermentation, and anaerobic respiratory reductases. 20% of the unique genes coded for gram negative cell structures and 5% for gram negative phages. 27.96% (2037/7286) of the LAP metagenome differed significantly from that of GAP (p<0.05, FDR adjusted Wald test, DESeq2). The LAP microbiome demonstrated greater capacity for inositol catabolism, and Lipid A, lipopolysaccharide and peptidoglycan biosynthesis when compared to GAP (p<0.05, FDR adjusted Wald test, DESeq2). Additionally, the LAP biome demonstrated a 4-fold to 144-fold greater enrichment of genes encoding c-type cytochrome and molybdenum cofactor biosynthesis, iron-sulfur clusters, formate dehydrogenase and oxidative stress response. 27.5% (2019/7340) of the LAP metagenome differed from CP. This was attributable to a higher representation of genes encoding acetyl CoA, lactate, mixed-acid and lysine fermentation, methanogenesis, anaerobic respiratory reductases, dehydrogenases, dehydratases and anaerobic toluene and ethylbenzene degradation. Also over-represented in the LAP metagenome were membrane transport functions (type II, III, IV, V and VI secretions systems and ABC transporters), and functions related to quorum sensing and biofilm formation (Autoinducer-2 transport and processing, biofilm adhesins and histidine kinase sensors).
The GAP microbiome–– a functional chimera: Pairwise dissimilarity analysis revealed that the GAP metagenome was intermediate between CP and LAP (p=0.91, Dunn’s multiple comparisons test on Bray-Curtis distances between GAP––CP and GAP––LAP). GAP shared 77% of its metagenome with CP and 64% of its metagenome with LAP. To reduce bias induced by sparse data, we next used core genes in each group to compute these distances. Not surprisingly, we observed a lower similarity within the common core metagenomes, but a more balanced difference, with 56% of genes shared by CP and GAP, and 55.1% by GAP and LAP. Since severe attachment loss in young adults in the presence of clinical inflammation and local factors could represent either early onset chronic periodontitis, a generalized form of the molar-incisor phenotype, or true de novo aggressive periodontitis(41), we examined clustering of the GAP samples alone. NMDS did not reveal significant separation between the 17 samples, suggesting that this chimeric effect cannot be readily attributed to heterogenous diseases.
14% of the core genes shared by LAP and GAP encoded as yet unknown functions (Supplemental Table 4). Among the characterized genes, the predominant shared functionalities were related to an anaerobic lifestyle. These included genes encoding heme and hemin dependent respiration, dehydrogenases, electron donors and acceptors other than oxygen (namely, nitrate, sulfate, hydrogen, and ferric iron), and fermentation. Other shared functions included polyamine metabolism, flagellar biosynthesis and gram-negative cell wall components (including peptidoglycan biosynthesis), response to oxidative and osmotic stress, resistance to antibiotics and toxic compounds, phages and conjugative transposons. The differences in the microbiomes of LAP and GAP were attributable to lower abundances of membrane transport functions (type II, III, IV, V VI, and ABC transporters), quorum sensing and biofilm formation in and higher levels of sporulation and dormancy, phages and transposable elements in GAP. Pathways involved in biofilm stability were also lower in GAP in comparison to LAP, and even lower in CP when compared to GAP.
The functional roles of 19% of the genes shared by CP and GAP were unknown. Both CP and GAP demonstrated a collective capacity for metabolism of amino acid, organic compounds, alcohols and glycogen. The other shared functionalities included capsule and cell wall synthesis, response to oxidative and osmotic stress, resistance to antibiotics and toxic compounds, phages and conjugative transposons. Flagellar components and proteins associated with flagella biosynthesis and assembly machinery are also exclusively enriched in GAP as compared to CP, as were potent inflammatory triggers such as lipopolysaccharides, and peptidoglycans. Other abundant functions of GAP include dormancy and sporulation, invasion and intracellular resistance, iron acquisition and siderophores, multidrug antibiotics efflux pumps.