A total of 40 subjects affected by moderate to severe generalized chronic periodontitis were recruited among consecutive patients who were referred between 2014 and 2016 to the Periodontology and Patients with Special Needs units at the School of Medicine and Dentistry (Universidade de Santiago de Compostela, Spain) for an assessment of their oral health status. Patients were selected if they fulfilled the following inclusion criteria: (1) age 30 to 65 years; (2) good general health, no pregnancy or breastfeeding; (3) no intake of systemic antimicrobials during the previous 6 months; (4) no intake of antiinflammatory medication in the previous 4 months; (5) no routine use of oral antiseptics; (6) no presence of implants or orthodontic appliances; (7) no previous periodontal treatment; and (8) the presence of at least 18 natural teeth.

Two experienced dentists and previously calibrated performed all the periodontal diagnoses. The probing pocket depth (PPD) and clinical attachment loss (CAL) (=PPD+ gingival recession) were recorded throughout the mouth on six sites per tooth (excluding third molars) using a PCP-UNC 15 probe. Bleeding on probing (BOP) and bacterial plaque levels (BPL) were recorded for the full mouth on a binary scale (presence/absence) on six sites per tooth. Standardized radiographs of all teeth were obtained to assess the alveolar bone status. Patients were diagnosed as suffering from moderate to severe generalized chronic periodontitis based on the previously established criteria (Armitage, 1999; Page and Eke, 2007). Smoking status was also recorded (Figure 1).

Patients who agreed to participate in the study provided their written informed consent. The study protocol was approved by the Ethics Committee of Clinical Investigations of Galicia (registration number 2015/006). The ‘transparent reporting of a multivariable prediction model for individual prognosis or diagnosis’ (TRIPOD) guidelines were considered (Moons et al., 2015).

Control and periodontal sites were defined within the same periodontal patient as follows: (1) the control sites had a PPD and CAL <4 mm and no radiographic evidence of alveolar bone loss; and (2) the periodontal sites had a PPD and CAL >4 mm and radiographic evidence of bone loss (Figure 1).

All the subgingival plaque samples were collected in the afternoon, approximately 5–7 h after toothbrushing. The samples were obtained from both a control site and a periodontal site (two or three adjacent locations in each site) in each periodontal patient by inserting a total of five sterile endodontic paper points into each location (size 30) for 10 s following isolation and supragingival plaque removal. The samples were placed in 1.5 ml microcentrifuge tubes with 300 μl of a phosphate buffer and frozen at -80°C until further analysis.

Two investigators blinded to the clinical data performed the bacterial analyses by qPCR. The bacteria were separated from the paper points before the DNA extraction by washing with 500 μL of a phosphate buffer solution and vortexing. DNA extraction from the samples was performed in the MagNA Pure Compact Instrument (Roche, Germany) following the manufacturer’s instructions. We used it in combination with the MagNA Pure Bacteria Lysis Buffer (Roche) and the DNA Bacteria purification protocol (Roche) to work with the bacterial DNA isolated from subgingival plaque samples.

The levels of seven pathobionts were quantified: A. actinomycetemcomitans (Aa); F. nucleatum (Fn); P. micra (Pm); P. gingivalis (Pg); P. intermedia (Pi); T. forsythensis (Tf); and T. denticola (Td). Real-time PCR (LightCycler^®480 system, Roche) was performed using TaqMan probes (Universal Probe Library, Roche) and primers were designed to amplify specific target genes from each bacterium (Table 1) (Fernández-Pinero et al., 2013; Fraczyk et al., 2016). The specificity of the probes was analyzed among all the bacteria from this study. We worked with uracil-N-glycosylase, Cod-UNG (ArcticZymes, Norway) (Bou et al., 2014) to remove contaminating amplicons from the previous qPCR. To confirm this, we used the water from the reaction of the qPCR as a negative control (Champlot et al., 2010).

Absolute quantification of the DNA was performed using the standard curve method. A standard dilution series of known amounts of genomic DNA (from 100 pg to 1 fg) were assessed in the same assay. Cp-values obtained from these standards were used to generate a standard curve from which the amount of DNA in the unknown samples was calculated. Each standard contained the same target region that was amplified in the sample. Strains from each microorganism, obtained from the ATCC Bacteriology Collection, were used both to generate the standard curves and as positive controls. A regression analysis was performed to obtain the equation used to interpolate the Cp-values from healthy and diseased sites and thus quantify the corresponding concentrations of genomic DNA of each pathobiont in each sample (pg/μl). The value of R² was used as a measure of the goodness-of-fit of the regression analysis.

The unit of analysis in the study was the sampled subgingival site. The ‘a priori’ sample size calculation was performed using the program G^∗Power 3.1.5 (Faul et al., 2007). The following statistical criteria were established: (1) an effect size of 0.5 for bacterial concentrations between the control and periodontal sites; (2) an alpha error of 0.05; and (3) a statistical power of 90%. A sample size of 36 subjects was required on the basis of these criteria and the application of the Wilcoxon test to evaluate the differential bacterial concentrations.

The independent variables were: (1) the quantitative levels of the species individually (7 variables); and (2) due to its biological significance, the quantitative levels of all possible bacterial clusters (120 variables) (Figure 1) (Supplementary Data Sheet S1). The statistical analyses were performed using the R software (R Core Team, 2016).

The study group consisted of 40 individuals with a diagnosis of moderate-severe generalized chronic periodontitis and an average age of 51.35 years. Of these patients, 18 were male and 22 female. With regard to the comparative analysis of clinical parameters between the control and periodontal sites, the latter showed significantly higher BOP, PPD, and CAL values (Table 2).

In relation to their smoking habits, 19 patients were non-smokers and 21 current smokers, and both patient groups showed similar age and gender characteristics, as well as similar values in terms of the periodontal indices recorded in the full mouth and the sampled subgingival sites.

The detection frequencies of the seven species evaluated were similar between the control and periodontal sites. The median total bacterial concentration of the periodontal sites was four times higher than that detected in the control sites (P < 0.001). Five pathobionts showed significantly higher concentrations in the periodontal sites than the control sites. These species were: Fn, Pm, Pg, Tf, and Td (P < 0.001 for all the bacteria except Fn, P = 0.023). In terms of the 127 bacterial clusters, all except PiAaFn, PiFn, and PiAa showed significantly higher concentrations in the periodontal sites than the control sites (P-values ranged from <0.001 to 0.009). After applying the Benjamini–Hochberg correction, all the significant P-values were maintained (Table 3 and Supplementary Data Sheet S1). With regard to smoking habit, no significant differences were found in the detection frequencies or the bacterial concentrations for each subgingival site between the smokers and non-smokers.

A first description of the relation between pathobiont levels, as well as with the clinical parameters (PPD, CAL) is given in Supplementary Data Sheet S1, by means of their Spearman correlations. Almost all correlations between bacteria and with the PPD and CAL values were positive. The interpretation is that when a subgingival site with periodontal destruction was present, the majority of pathobionts presented larger values, which can be verified in the comparative analysis between control and periodontal sites shown in Table 3. However, the correlation values between the bacteria and these with the clinical parameters were not very high, the strongest associations being observed between Pg and Tf (Rho = 0.699) and several clusters with ≥4 bacteria and CAL (Rho > 0.540) (Supplementary Data Sheet S1).

Applying the statistical criteria set out in the section ‘Materials and Methods,’ a total of eight models were selected (Table 4 and Supplementary Data Sheet S2). In all of them, the predictor variables were represented by one bacterial cluster composed of at least three species. The most common bacteria were: Tf and Pi (both were present in eight models), followed by Td and Fn (both were present in six models) and Aa and Pm (both were present in five and three models, respectively). The concentration values of the optimal threshold ranged from 218.86 pg/μl for the PiTfFn model to 264.16 pg/μl for the TdPiTfAaFnPm model. There was no predictive model formed by two bacterial clusters that fulfilled the established statistical criteria.

The main results on the apparent and bc-measures of discrimination and classification of eight pathobiont-based models are detailed in Table 5 and Supplementary Data Sheet S2. The apparent AUC values of these models ranged from 0.773 to 0.789. The PiTfFn cluster showed an apparent AUC value of 0.773 (sensitivity and specificity = 75.0%; DOR = 9.0). When Td was incorporated in this cluster, a new predictive model emerged with better apparent AUC, sensitivity and DOR values (0.783, 77.5%, and 10.3, respectively). When Pm and AaPm were incorporated in the TdPiTfFn cluster, we obtained the two best predictive models with apparent AUC values of 0.788 and 0.789, respectively, as well as sensitivity and specificity values of 77.5% and DOR of 11.8 for both models. However, when Aa and TdAa were incorporated in the PiTfFn cluster, the two predictive models that appeared did not have better discrimination and classification measures than those detected in clusters of lower numbers of bacteria, such as PiTfFn and TdPiTfFn.

The TdPiTfAa cluster showed an apparent AUC value of 0.785 (sensitivity and specificity = 75.0%; DOR = 9.0). When Pm was incorporated in this cluster, a new predictive model appeared with better apparent AUC, specificity and DOR values (0.787, 80.0%, and 12.0, respectively). The optimism values obtained by the bootstrap methods in relation to the performance measures of these models ranged from -0.001 to 0.003 in the AUC values and 1.52–3.07% in the classification indices.

Figures 3, 4 show the ROC curves of the most predictive clusters with the lowest number of species – PiTfFn and TdPiTfAa – and the highest number of bacteria – TdPiTfAaFnPm and TdPiTfAaPm. Figures 5, 6 show the diagnostic nomograms of the four previous models. As noted previously, the discrimination and classification performance values of all the nomograms were high, indicating their good accuracy. Overall, in eight nomograms, the high concentration levels of the different bacterial clusters were associated with an increased probability of having a subgingival site with periodontal destruction in a patient with chronic periodontitis. Supplementary Files show these graphics of the eight models with an AUC ≥0.76 and sensitivity and specificity values ≥ 75% (Supplementary Data Sheets S3, S4, and Figures S1–S8).

In terms of study design, a control group (healthy subjects) is compared to a group of periodontitis patients in numerous qPCR studies of periodontal pathogens at the subgingival level (Braga et al., 2010; Kretschmar et al., 2012). The control group may also include gingivitis patients (Scapoli et al., 2015), or they may constitute a separate group (Masunaga et al., 2010). In addition, there are other authors who have studied subgingival samples microbiologically using the qPCR from different periodontal patient groups according to the degree of periodontal affection (mild, moderate, or severe) (Zorina et al., 2011) or the type of disease (chronic vs. aggressive) (Casarin et al., 2010). There are also qPCR studies that have been performed exclusively on subjects with periodontitis (Kotsilkov et al., 2015; Milne et al., 2015).

Although certain subject-associated factors can increase an individual’s overall susceptibility to periodontal diseases, it appears that not all sites within a subject are equally susceptible at one time (Griffiths, 2003). This site-specific clinical pattern can be attributed to differences in the bacterial species composition of subgingival plaque (Byrne et al., 2009). However, there is little evidence in the literature of the influence of pathobionts analyzed by the qPCR in relation to the site-specific severity of chronic periodontitis (Mineoka et al., 2008; Teixeira et al., 2009; Al-hebshi et al., 2014). As a consequence, we propose using a qPCR technique to study the relationship between pathobiont levels from subgingival sites with different degrees of disease (control sites vs. periodontal sites) within the same periodontal patient. This methodology (paired design) means that the effects of inter-individual variations in factors other than bacterial composition are avoided.

Regarding the species studied, we decided to quantify them initially according to Socransky’s bacterial complexes (Socransky et al., 1998), with those most frequently studied being: three species of the Red Complex (Pg, Tf, and Td), the main three species of the Orange Complex (Fn, Pi, and Pm) and a species from the Green Complex (Aa). Consequently, our intention was to focus on various well-known pathobionts that could be used as biomarkers in a possible diagnostic tool for dental practitioners in order to improve disease control and reduce the future risk of progression seen in everyday clinical practice.

In numerous qPCR papers, the authors report the number of DNA copies as a measure of the quantification of the bacterial load rather than cell numbers, because the copy number of the marker gene, particularly the 16S rRNA gene, varies from one species to another and is unknown for some of them (Al-hebshi et al., 2014, 2015; Yang et al., 2016). This implies that available bacterial counts for some of the tested genera/species have probably been over-estimated (Al-hebshi et al., 2014). In the present series, we used a new qPCR technique with TaqMan probes and Cod-UNG enzymes for the molecular detection of these pathobionts. This patented technique allows the easy design of probes for the detection of any pathogen while maintaining excellent sensitivity and specificity (Bou et al., 2014). In the present study, the absolute quantification of the DNA of each species was performed using the standard curve method. Moreover, the target genes studied were different single-copy genes per genome, which may allow absolute counts of DNA copies to be reported in terms of cell numbers.

From an analytical point of view, the majority of authors in their respective qPCR studies applied a ‘simple approach.’ As a result, the subgingival levels of pathobionts in control volunteers and periodontal patients were compared for testing using q-PCR, whether or not a particular bacterium is absent/present and/or diminished/elevated at the individual level (Braga et al., 2010; Polonyi et al., 2013; Kotsilkov et al., 2015; Milne et al., 2015; Scapoli et al., 2015; Golynska et al., 2016). In parallel with current trends in big data analytics for genomic medicine (He et al., 2017), further evidence is required about applying multivariate analytical approaches to qPCR data. This would enable there to be a better understanding of the complex bacterial interactions present in chronic periodontitis. This type of analytical approach was used in our study.

Al-hebshi et al. (2014) used the qPCR to investigate seven pathobionts (the species of the Red Complex, the genera Fusobacterium spp. and Prevotella spp., as well as Aa) in healthy and periodontal subgingival sites within the same periodontal patient. This study was similar to ours methodologically. The Al-hebshi’s series revealed detection percentages of 100% for Pm, Td, Tf, Fusobacterium spp. and Prevotella spp., 97.5% for Pg, and 67.5% for Aa. We agree with them on the high detection rates found for the Red Complex species, Fn and Pm (80–100%), confirming that the mere presence of the bacteria is not an indicator of the clinical situation (Byrne et al., 2009; Charalampakis et al., 2013). However, in contrast to Al-hebshi et al. (2014), our series of periodontal patients showed lower detection frequencies for Pi (69%) and, especially, Aa (17.5%), although those values were not conditioned by the subgingival site tested. Initially, our results on the low prevalence of Aa confirmed its marginal role in chronic periodontitis in our study population, which reflects the findings of other authors (Gatto et al., 2014); however, we will later see how its role changes in the multivariate analysis.

In Al-hebshi’s study (Al-hebshi et al., 2014), the levels of Pm, Prevotella spp., Pg, Tf, and Td were higher in subgingival sites with periodontal destruction than the control sites. However, after correcting for multiple comparisons, these differences only remained statistically significant for Pm and Tf. This was probably due to the small sample size (only 20 periodontal patients). In the present study, our results revealed that periodontal sites had significantly higher concentrations for the three Red Complex species (concentration increments of 27, 14, and 3 times for Pg, Td, and Tf, respectively) and Pm and Fn (concentration increments of 2.7 and 1.5 times, respectively).

Our findings demonstrated that all the pathobionts (except Aa and Pi) were present at the control sites (non-active sites) and that an increase in their concentrations was required to trigger activity in the periodontal site; Pg was the bacteria that had a higher increment. The relevance of the Red Complex species together and Pg alone in the pathogenesis of periodontitis has been proved in numerous in vitro studies (Bao et al., 2014; Lin et al., 2014; Willi et al., 2014). A bacterial consortium named the Orange Complex, which includes Fn and Pm, was indicated as preceding the Red Complex with respect to colonization and proliferation (Socransky et al., 1998). In relation to Fn and Pm, there is growing evidence based on in vitro research on their role as periodontal pathogens (Nonnenmacher et al., 2003; Lee and Baek, 2013; Wang et al., 2013; Kim and Lee, 2014; Marchesan et al., 2016).

Although evaluating the influence of smoking was not a major objective of the present study, our findings in this regard are consistent with those of Tomita et al. (2013), as we did not detect significant differences in the profiles of the seven species present in the sampled subgingival sites (control and periodontal sites). These sites showed similar clinical characteristics between smokers and non-smokers. Nevertheless, these results should be interpreted with caution since, in larger series, other authors have found significantly higher prevalences of Td and Tf and significantly higher levels of Aa, Pg, and Tf in smokers with chronic periodontitis (Gatto et al., 2014; Guglielmetti et al., 2014).

In periodontology, traditional clinical criteria are often inadequate for determining sites of active disease, measuring the degree of susceptibility to future progression, and monitoring the response to therapy (Giannobile et al., 2009; Korte and Kinney, 2016). In this sense, there is a need for research on innovative diagnostic tests based on biomarkers that focus on the early recognition of the microbial challenge to the host, providing a benefit in terms of disease control at the site-specific level.

The first step to assess whether a test could be clinically useful is the analysis of its predictive ability (Deeks, 1999). In this sense, there are some papers in the literature that have focused on the analysis of the predictive ability of the pathobiont levels quantified by the qPCR to distinguish different degrees of periodontal affection in subgingival samples from different individuals (Ramseier et al., 2009; Saygun et al., 2011; Kinney et al., 2014; Al-hebshi et al., 2015; Torrungruang et al., 2015). Several of these studies coincide in terms of detecting that the subgingival levels of certain bacteria, such as the Red Complex species, C. rectus, Pi, Pm and even a new bacterial phylum of oral Synergistetes, showed a good diagnostic power for distinguishing between healthy/gingivitis patients and those with periodontitis (AUC values ranged from >0. 74 to >0.80) (Ramseier et al., 2009; Al-hebshi et al., 2015).

However, we have found very few papers like ours on bacteria-based predictive models for diagnosing periodontitis severity or progression using a paired design (Byrne et al., 2009; Nomura et al., 2012; Charalampakis et al., 2013; Gatto et al., 2014). Charalampakis et al. (2013) identified sites at risk of future progression during 2 years of maintenance in 50 patients with chronic periodontitis. This research was based on longitudinal clinical and microbiological monitoring using DNA–DNA hybridization involving 25 bacterial species. In this series, both members of the Red Complex as well as of the ‘B Complex’ (Pi, Fn and C. rectus) were equally able to identify periodontitis from non-periodontitis sites and severe periodontitis sites from non-severe periodontitis sites; these bacterial complexes had AUC values above 0.75. In accordance with these results (Charalampakis et al., 2013), in our study the best predictive models with AUC values ≥0.76 (sensitivity and specificity ≥75%), which could be regarded as good (Hosmer et al., 2013), included the concentration levels of different bacterial clusters formed by at least three bacteria. These findings were ratified by the analysis of the internal validation that was performed, since the optimism values obtained by the bootstrap techniques were not high. Therefore, in accordance with the findings of previous authors (Byrne et al., 2009; Charalampakis et al., 2013), our results confirm that the use of bacterial quantitative data rather than dichotomous data (presence/absence) provides more statistical power for detecting any predictive relationship with disease severity.

Interestingly, and unlike Charalampakis’s study, the bacteria present in all the predictive clusters were Tf and Pi (both in eight models), followed by Td and Fn (both in six models) and Aa and Pm (both in five and three models, respectively). All these clusters were constituted by bacteria that belong to different Socransky’s complexes and have different pathogenic roles in disease severity in terms of individual concentration levels. Such an outcome was previously observed in our own series. In fact, although Pi and Aa had non-significant, higher concentrations in the periodontal sites than the control sites, their levels contributed to increasing the predictive ability of the cluster in which they were incorporated. On the other hand, although Pg was the species that showed the highest concentration increment in the periodontal sites, surprisingly this species did not have a high power for predicting periodontitis severity in comparison with other pathobionts, since it was not included in any of the best predictive models. From a predictive perspective, these findings contribute to Pg being regarded as a ‘keystone’ pathobiont which is a microorganism that can change the environment to alter proportions or levels of other symbionts and pathobionts within the ecological niche; these microorganisms are the main ones responsible for triggering the destructive cascade that provokes the activation of inflammation and subsequent bone destruction (Costalonga and Herzberg, 2014).

In accordance with previous authors (Byrne et al., 2009; Charalampakis et al., 2013), we demonstrated that the detection of a specific threshold of concentration levels of certain bacterial clusters may serve as a predictor of periodontitis severity at site-specific level. Two main predictive clusters were identified: PiTfFn (AUC = 0.773; sensitivity and specificity = 75.0%; DOR = 9.0) and TdPiTfAa (AUC = 0.785; sensitivity and specificity = 75.0%; DOR = 9.0). In relation to PiTfFn, when other bacteria or clusters such as Td, Pm, and AaPm were incorporated in this cluster, new predictive models emerged with better predictive parameters. The two best predictive clusters were TdPiTfFnPm and TdPiTfAaFnPm, with AUC values of 0.788 and 0.789, respectively (sensitivity and specificity values of 77.5% and DOR of 11.8 for both models). In relation to TdPiTfAa, when Pm was incorporated in this cluster, a new predictive model appeared with better AUC, specificity and DOR values (0.787, 80.0%, and 12.0, respectively).

To the best of our knowledge, this study presents the first evidence that these new bacterial clusters are capable of predicting the severity of chronic periodontitis, with several species detected outside the Red Complex that contribute to the detection of periodontal destruction at the specific-site level. To date, the findings obtained by open-ended molecular approaches support the hypothesis that chronic periodontitis is initiated by polymicrobial synergy and the dysbiosis of the entire microbial community (PSD model). The condition is characterized not only by a greater involvement of the ‘established’ pathobionts, as studied in this series, but also by the coexistence of other pathobionts with an unknown role such as Anaeroglobus, Bulleidia, Desulfobulbus, Filifactor alocis, Mogibacterium, Phocaeicola, Schwartzia, or TM7 (Camelo-Castillo et al., 2015a,b). Given this new pathogenic approach of the disease, it would be very interesting to study the predictive ability of these new pathobionts and other symbionts in chronic periodontitis, as well as their relationship with host mediators.

Nomograms are simplified representations of complicated statistical models, and their clinical value relates to the fact that they map the predicted probabilities into points on a scale from 0 to 100 in a user-friendly graphical interface (Iasonos et al., 2008). To our knowledge, this is the first study providing several nomograms based on the concentration levels of certain bacterial clusters to predict the probability of having a periodontal site in a patient with chronic periodontitis.

Our nomograms, which are derived from the eight best-fitting models, fulfilled the requirements of discrimination. Overall, in all the nomograms, higher concentration levels of different bacterial clusters were associated with an increased probability of having a periodontal site in a periodontal patient. The use of only a few variables is desirable in nomograms to increase their utility in clinical practice (Altman and Royston, 2000). Consequently, of the nomograms developed in the present study, we highlight those based on the clusters formed by the lower number of species, PiTfFn and TdPiTfAa. The application of these tools in the field of clinical activity would improve the identification of sites with periodontal destruction that are at possible risk of future progression, contributing to the decision-making process and treatment planning dilemmas.

Our research has some limitations. The most important weakness is that the prediction of the study’s accuracy is only measured in the samples that generated the model equations. As a consequence, to evaluate the reproducibility of the models and control for the possibility of overfitting, we validated the prediction rule internally (discrimination and classification measures) using bootstrap methods on the original derivation dataset by sampling with replacements for 1000 iterations (Steyerberg et al., 2001a,b). This internal validation method revealed optimal results on the control of overfitting in the models.

The evaluation of these cluster-based predictive models and nomograms is a potential future research direction. Firstly, it would greatly benefit the strength of our study if the predictive accuracy of the predictive models derived from our series could be measured in a large ‘external’ or independent cohort of patients to verify whether our findings are universally applicable or if they are conditioned by differences associated with geographic regions and ethnic groups (Herrera et al., 2008; Wara-aswapati et al., 2009; Psoter et al., 2011). An appropriate calibration analysis of the predictive models should also be performed. Secondly, the potential prognostic value at the site-specific level of these cluster-based predictive models with regard to disease progression and the response to treatment in periodontal patients should be exploited in longitudinal studies, as should their relationship with host mediators and their potential predictive accuracy in saliva samples.