modeling susceptibility to periodontitis

8
http://jdr.sagepub.com/ Journal of Dental Research http://jdr.sagepub.com/content/92/1/45 The online version of this article can be found at: DOI: 10.1177/0022034512465435 2013 92: 45 originally published online 25 October 2012 J DENT RES M.L. Laine, V. Moustakis, L. Koumakis, G. Potamias and B.G. Loos Modeling Susceptibility to Periodontitis Published by: http://www.sagepublications.com On behalf of: International and American Associations for Dental Research can be found at: Journal of Dental Research Additional services and information for http://jdr.sagepub.com/cgi/alerts Email Alerts: http://jdr.sagepub.com/subscriptions Subscriptions: http://www.sagepub.com/journalsReprints.nav Reprints: http://www.sagepub.com/journalsPermissions.nav Permissions: What is This? - Oct 25, 2012 OnlineFirst Version of Record - Dec 10, 2012 Version of Record >> at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission. jdr.sagepub.com Downloaded from © 2013 International & American Associations for Dental Research at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission. jdr.sagepub.com Downloaded from © 2013 International & American Associations for Dental Research at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission. jdr.sagepub.com Downloaded from © 2013 International & American Associations for Dental Research

Upload: b-g

Post on 16-Feb-2017

214 views

Category:

Documents


0 download

TRANSCRIPT

http://jdr.sagepub.com/Journal of Dental Research

http://jdr.sagepub.com/content/92/1/45The online version of this article can be found at:

 DOI: 10.1177/0022034512465435

2013 92: 45 originally published online 25 October 2012J DENT RESM.L. Laine, V. Moustakis, L. Koumakis, G. Potamias and B.G. Loos

Modeling Susceptibility to Periodontitis  

Published by:

http://www.sagepublications.com

On behalf of: 

International and American Associations for Dental Research

can be found at:Journal of Dental ResearchAdditional services and information for    

  http://jdr.sagepub.com/cgi/alertsEmail Alerts:

 

http://jdr.sagepub.com/subscriptionsSubscriptions:  

http://www.sagepub.com/journalsReprints.navReprints:  

http://www.sagepub.com/journalsPermissions.navPermissions:  

What is This? 

- Oct 25, 2012OnlineFirst Version of Record  

- Dec 10, 2012Version of Record >>

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

45

RESEARCH REPORTSClinical

DOI: 10.1177/0022034512465435

Received July 20, 2012; Last revision September 29, 2012; Accepted September 29, 2012

A supplemental appendix to this article is published elec-tronically only at http://jdr.sagepub.com/supplemental.

© International & American Associations for Dental Research

M.L. Laine1*, V. Moustakis2,3, L. Koumakis2, G. Potamias2, and B.G. Loos1

1Academic Centre for Dentistry Amsterdam (ACTA), University of Amsterdam, and VU University Amsterdam, Department of Periodontology, The Netherlands; 2Foundation for Research and Technology – Hellas (FORTH), Institute of Computer Science, Bioinformatics Laboratory, Science and Technology Park of Crete, Heraklion 71110, Greece; and 3Technical University of Crete, Department of Production Engineering and Management, Chania, Greece; *correspond-ing author, [email protected]

J Dent Res 92(1):45-50, 2013

AbstrActChronic inflammatory diseases like periodontitis have a complex pathogenesis and a multifactorial etiology, involving complex interactions between multiple genetic loci and infectious agents. We aimed to investigate the influence of genetic poly-morphisms and bacteria on chronic periodontitis risk. We determined the prevalence of 12 single-nucleotide polymorphisms (SNPs) in immune response candidate genes and 7 bacterial species of potential relevance to periodontitis etiology, in chronic periodontitis patients and non-periodontitis control individuals (N = 385). Using decision tree analysis, we identified the presence of bacterial spe-cies Tannerella forsythia, Porphyromonas gingiva-lis, Aggregatibacter actinomycetemcomitans, and SNPs TNF -857 and IL-1A -889 as discriminators between periodontitis and non-periodontitis. The model reached an accuracy of 80%, sensitivity of 85%, specificity of 73%, and AUC of 73%. This pilot study shows that, on the basis of 3 periodontal pathogens and SNPs, patterns may be recognized to identify patients at risk for periodontitis. Modern bioinformatics tools are valuable in modeling the multifactorial and complex nature of periodontitis.

KEY WOrDs: periodontal disease, genetic poly-morphism, microbiology, bioinformatics, disease susceptibility, single-nucleotide polymorphism (SNP).

IntrODuctIOn

Periodontitis is an infectious disease of the supporting tissues of the teeth. It is well-accepted that this disease has a genetic component. The

search for genetic markers and candidate disease-modifying genes in peri-odontitis has targeted a large variety of single-nucleotide polymorphisms (SNPs) in genes encoding molecules of the host defense system (Zhang et al., 2011; Laine et al., 2012). Many association studies have attempted to correlate different SNPs with periodontal disease. Polymorphisms in several immune response genes may be associated with chronic periodon-titis susceptibility as a single genetic factor in certain populations (Loos et al., 2005; Laine et al., 2012). However, analogous to other complex inflammatory diseases, it is assumed that periodontitis is a polygenic dis-order. Gene-gene interactions may play a major role, and therefore single SNP genotyping results may give very limited information on the genetic susceptibility for periodontitis.

Infection with specific micro-organisms in the subgingival plaque (mainly Gram-negative strict anaerobic rods) is essential for periodontal disease ini-tiation and progression (Haffajee and Socransky, 1994). Immune responses against these microbiological agents are partly genetically determined (Le Bourhis et al., 2007), and therefore host response genes and the intrinsic genetic variations are important determinants for susceptibility to periodonti-tis (Michalowicz, 1994; Laine et al., 2001). Thus, not only gene-gene interac-tions but also gene-environment interactions form a complex network in which the disease can initiate and progress.

We hypothesized that the combination of bacterial species and SNPs might present higher discrimination between periodontitis and periodontal health than a single bacterial species or a single SNP. Our aim was to combine mul-tiple genetic and microbiological data to understand periodontitis as a com-plex disease in which the combination of these risk factors confers susceptibility. The present pilot study modeled 11 candidate gene polymor-phisms and the prevalence of 7 bacterial species, with a tool from the knowl-edge discovery field (machine learning techniques). First, we present decision tree models to distinguish and predict periodontal disease from non-periodon-titis on the basis of bacterial culture and multiple SNP genotyping data sepa-rately. In the second stage, we apply the obtained patterns of the first stage and combine microbiological patterns with SNP data and, vice versa, the SNP patterns with microbiological data.

Modeling susceptibility to Periodontitis

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

46 Laine et al. J Dent Res 92(1) 2013

MAtErIAls & MEthODs

Participant Data

Periodontitis patients and non-periodontitis individuals from previously published studies were included in the present study (Loos et al., 2003; Laine et al., 2005; Bizzarro et al., 2007). Study participants were recruited for the original studies in a similar manner and entered the studies consecutively. Records used in the present study included: (a) demographic data (age and gender); (b) smoking data; (c) periodontal disease status; (d) genetic data (single-nucleotide polymorphisms = SNPs); and (e) microbiological data. All the individuals are of North European Caucasian origin (N = 385). Patients were diagnosed with severe periodontitis if ≥ 7 teeth with ≥ 50% bone loss were found on radiographs. Patients with mild-to-moderate periodon-titis showed < 7 teeth with < 50% with bone loss on radiographs. For the purpose of the current study, no further distinction between mild-to-moderate and severe periodontitis was made, and all patients were entered into one group, “periodontitis.” Control individuals had no more than 1 tooth per quadrant miss-ing (3rd molar excluded) and showed, on dental-bitewing radio-graphs ≤ 1 yr old, a distance between the cemento-enamel junction and alveolar bone crest of ≤ 3 mm on all teeth. Participants who were current smokers or had stopped smoking ≤ 1 yr ago were defined as ‘smokers’, and those who had never smoked or had stopped smoking > 1 yr ago were defined as ‘non-smokers’. The previous studies, which contributed to the data, were approved by the Medical Ethical Committee of the VU Medical Center, VU University, Amsterdam, and the Academic Medical Center, University of Amsterdam, and all participants gave informed consent to participate in a study on the genetic susceptibility to periodontitis. The database was specifically constructed for complete anonymity of participants.

single-nucleotide Polymorphisms

For the purpose of the current study, DNA aliquots from all partici-pants were used for the following 12 SNPs: IL-1A -889 (dbSNP ID rs1800587), IL-1B -31 (rs1143627) and +3954 (rs1143634), IL-1RN +2018 (rs419590), IL-10 -1082 (rs1800896) and -819 (rs1800871), CD14 -260 (rs2569190), LTA +252 (rs909253) and +368 (rs746868), and TNF -308 (rs1800629), -857 (rs1799724), and -863 (rs1800630). The technical procedures have been described previously (Nicu et al., 2009; Ursum et al., 2010). These SNPs were selected from candidate genes encoding molecules involved in the immune response to bacteria in periodontitis. LTA +368 was not in Hardy-Weinberg equilibrium in the control group and was excluded from further analyses. The other 11 SNPs were in Hardy-Weinberg equilibrium in the control group.

Microbiological Data

The presence of Porphyromonas gingivalis, Aggregatibacter actinomycetemcomitans, Tannerella forsythia, Prevotella inter-media, Fusobacterium nucleatum, Peptostreptococcus micros, and Campylobacter rectus species was detected by standard culture techniques as described previously (van Winkelhoff et al., 2002). The individuals were considered positive for these

species if the proportions were at or above these thresholds: 0.1% for P. gingivalis, 0.01% for A. actinomycetemcomitans, 3.0% for T. forsythia, 2.5% for P. intermedia, 3.0% for F. nucleatum, 3.0% for P. micros, and 2.0% for C. rectus (Rams et al., 1996; Meijndert et al., 2010).

Decision tree learning

The decision tree learning (C4.5 software Weka J48; Quinlan, 1993) was applied, with data from the 11 SNPs and positivity (≥ threshold) of the 7 microbial species as variables, and periodon-titis and periodontal health as different classes. The C4.5 algo-rithm builds a decision tree from the top: First, the most discriminative variable (microbial species or SNP) for classify-ing between periodontitis and non-periodontitis is selected (Huang et al., 2009). Then, the algorithm searches for the next best informative variable (microbial species or SNP) of the tree to improve the model. Feature selection is a part of the decision tree algorithm. Interactions between and among features are taken into account. Randomized V-cross-folding validation is performed with leave-one-out (LOO), which means that all cases but one are used to ‘learn’ a tree, and then the tree is tested with the left-out case. The process is repeated as many times as the number of records, and the final results aggregate successes and misses. To measure the performance of the models, we cal-culated accuracy, sensitivity, specificity, and the area under the receiver operating characteristic (ROC) curve (AUC). AUC calculations by Weka J48 correspond to the Mann-Whitney U-statistics, equivalent to the AUC calculation (Bamber, 1975; Hanley and McNeil, 1982; Cortes and Mohri, 2004).

In the first stage, a decision tree was formed on the basis of (a) microbial and (b) genetic data separately. In the next stage, the patterns formed in the first stage for microbial (a) and genetic data (b) were combined in 2 ways—the microbial pat-terns with the genetic data and thereafter the genetic patterns with microbial data—and the decision trees were formed.

Other statistical Analyses

Differences in mean ages of the periodontitis group and the non-periodontitis control group were tested by a t test (PASW Statistics18, SPSS Inc., Hong Kong). For possible differences in gender, smoking distribution, and prevalence of SNPs and the 7 bacterial species, a chi-square test was used. AUC was calculated with continuous values (proportions of bacteria) and accuracy, sensitivity, and specificity with the thresholds. This study con-forms to the Strobe guidelines (http://www.strobe-statement.org).

rEsults

Demographic Data

Mean ages of the controls (44 yrs; range, 22-90 yrs; N = 155) and the patients (46 yrs; range, 21-77 yrs; N = 230) were not significantly different. The gender distribution was also similar in the control (58% female) and patient (58% female) groups. Significantly more periodontitis patients (56%) were smokers in comparison with controls (25%, P < 0.001, OR 0.3, 95%CI 0.2-0.4). All controls and patients were systemically healthy.

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

J Dent Res 92(1) 2013 Modeling Susceptibility to Periodontitis 47

Microbiological Data

All species except F. nucleatum and C. rectus were more frequently isolated from the patients with periodontitis than from the control individuals (Table 1). The highest odds ratios were found for P. gingivalis and T. forsythia, 8.3 and 8.9, respectively. T. forsythia gave the highest accuracy as a single factor and AUC (72.2% and 73.7%, respectively), followed by decreased accuracy and AUC of P. gingivalis (63.6% and 68.1%, Table 2). All bacterial species showed poor sensitivity as a single factor (10.5-64.6%). Sensitivity and specificity of T. forsythia positivity reached 64.9% and 82.6%, respectively, and corresponding figures for P. gingivalis were 45.2% and 91.0%. Other bacterial species showed poor accuracy and AUC between periodontitis and non-periodontitis.

single-nucleotide Polymorphisms

Genotype frequencies of the 11 SNPs were not significantly dif-ferent between patients and controls (Table 2).

Decision tree learning

Bacterial species T. forsythia, P. gingivalis, and A. actinomy-cetemcomitans formed a decision tree with accuracy of 80.3%, sensitivity of 84.8%, specificity of 73.5%, and AUC of 68.2% (Fig. 1, Table 2). T. forsythia formed the root of the tree as the most informative variable, followed by P. gingivalis and A. acti-nomycetemcomitans. Four patterns were formed: 3 patterns for periodontitis consisting of (1) T. forsythia-positive individuals,

table 1. (A) Mean Proportions (range) of Bacterial Species of the Total Cultivable Microflora, and Percentages of Individuals with Bacteria ≥ Thresholds; and (B) Genotype Frequencies of the SNPs in Control Individuals and Periodontitis Patients (N = 385)

A. Bacterial Species Control Individuals N = 155 (%) Patients N = 230 (%) p ValueOdds Ratio

(95%CI)

P. gingivalis Mean (range) 9.8 (0.1-67.2) 30.7 (0.3-86.0) 0.003 Prevalence ≥ threshold 9.0 45.2 < 0.0001 8.3 (4.3-15.3)

A. actinomycetemcomitans Mean (range) 1.6 (0.01-10.1) 4.1 (0.01-46.3) NS* Prevalence ≥ threshold 7.1 25.7 < 0.0001 5.0 (2.3-8.9)

T. forsythia Mean (range) 3.0 (0.03-30.0) 9.8 (0.2-81.4) < 0.0001 Prevalence ≥ threshold 17.4 65.2 < 0.0001 8.9 (5.4-14.6)

P. intermedia Mean (range) 4.6 (0.04-48.0) 4.7 (0.02-29.0) NS Prevalence ≥ threshold 23.9 34.5 0.03 1.7 (1.1-2.7)

F. nucleatum Mean (range) 4.7 (0.1-42.0) 5.4 (0.02-40.6) NS Prevalence ≥ threshold 49.0 52.2 NS NS

P. micros Mean (range) 7.2 (0.01-47.9) 7.4 (0.03-40.2) NS Prevalence ≥ threshold 45.2 60.0 0.005 1.8 (1.2-2.7)

C. rectus Mean (range) 1.6 (0.07-6.0) 3.2 (0.2-12.0) NS Prevalence ≥ threshold 5.2 10.5 NS NS

B. SNP Genotype** Frequencies (%)

11 12 22 11 12 22

IL-10 -1082 G>A 27.1 51.0 21.9 29.7 42.9 27.4 NS IL-10 -819 C>T 62.1 34.0 3.9 64.2 29.3 6.5 NS IL-1A -889 T>C 43.5 44.2 12.3 42.6 49.1 8.3 NS IL-1B -31 T>C 46.2 44.7 9.1 47.9 39.4 12.7 NS IL-1B +3954 C>T 49.0 43.2 7.7 54.6 37.0 8.4 NS IL-1RN +2018 T>C 56.1 38.7 5.2 55.5 37.9 6.6 NS CD14 -260 C>T 24.8 51.6 23.5 24.9 52.1 23.0 NS LTA +252 A>G 41.7 49.7 8.6 50.7 40.0 9.3 NS TNF -863 C>A 66.5 28.4 5.1 66.1 30.8 3.1 NS TNF -857 C>T 80.6 16.8 2.6 79.8 18.4 1.8 NS TNF -308 G>A 70.8 26.6 2.6 66.1 30.0 3.9 NS

*NS = not significant.**11 genotype = homozygous for common allele, 12 = heterozygous, 22 = homozygous for rare allele.

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

48 Laine et al. J Dent Res 92(1) 2013

(2) T. forsythia-positive and P. gingivalis-negative individuals, and (3) T. forsythia-negative, P. gingivalis-negative and A. actinomycetemcomitans-positive individuals, and one pattern for non-periodontitis consisting of T. forsythia-negative, P. gingivalis-negative, and A. actinomycetemcomitans-negative individuals.

A complex decision tree with poor performance was formed on the basis of 8 SNPs (Appendix Fig. 1), with an overall accu-racy of 55.8%, sensitivity of 72.6%, specificity of 31.0%, and

AUC of 45.4 %. In total, 21 patterns were generated: 7 associated with no periodontitis and 14 associated with peri-odontitis.

In stage 2, we explored whether (A) the addition of the SNP data to the micro-bial patterns, or (B) the addition of the microbial data to the SNP patterns would give us more accurate models. For experi-ment A, a decision tree with accuracy of 80.0%, sensitivity of 84.8%, specificity of 72.9%, and AUC of 72.9% was formed (Fig. 2, Table 2). Here, the microbial pat-terns of Fig. 1 formed the root of the deci-sion tree, followed by SNPs TNF -857 and IL-1A -889. These variables yielded 8 different patterns: 4 associated with peri-odontitis and 4 associated with no peri-odontitis (Fig. 2). In experiment B, the decision tree with the SNP patterns and microbial data reached accuracy of 79.0%, sensitivity of 79.1%, specificity of 78.7%, and AUC of 72.0 (Table 2, Appendix Fig. 2). However, SNP patterns were not strong enough to form the root of the decision tree. The presence of T. for-sythia was the most predictive variable

and formed the root, followed by P. gingivalis, A. actinomycetem-comitans, SNP patterns, and F. nucleatum. These variables formed 10 different patterns: 4 associated with no periodontitis and 6 associated with periodontitis (Appendix Fig. 2).

DIscussIOn

We evaluated the possibility of distinguishing between disease and health using bioinformatic knowledge, and applied a

table 2. (A) Results of Leave-one-out Cross-validation* and (B) Single Bacterial Species**

A. Decision Tree Models Accuracy Sensitivity Specificity AUC

Microbiological data 80.3 84.8 73.5 68.2SNP data 55.8 72.6 31.0 45.4Microbial patterns and SNP data 80.0 84.8 72.9 72.6SNP patterns and microbial data 79.0 79.1 78.7 72.0

B. Single Factors Accuracy Sensitivity Specificity AUC

P. gingivalis 63.6 45.2 91.0 68.1A. actinomycetemcomitans 52.7 25.4 92.9 59.2T. forsythia 72.2 64.9 82.6 73.7P. intermedia 51.2 34.6 76.1 55.4F. nucleatum 51.7 52.6 51.0 51.8P. micros 57.9 60.5 54.8 57.7C. rectus 44.4 10.5 94.8 52.7

*In each case, a decision-tree model was built fusing N-1 records (N = 385). The record left out was used for testing, and the average is reported. AUC was calculated with the Weka J48 (decision tree) platform and corresponds to the Mann-Whitney U statistics, equivalent to AUC calcu-lations (Bamber, 1975; Hanley and McNeil, 1982; Cortes and Mohri, 2004).

**AUC was calculated with continuous values (proportions of bacteria), and accuracy, sensitivity, and specificity were calculated based on the threshold values reported in Materials & Methods, with PASW Statistics18 (SPSS Inc., Hong Kong).

Figure 1. The bacterial decision tree of the periodontitis patients and controls. Overall Leave-one-out (LOO) cross-validation accuracy was 80.3% (specificity 73.5% and sensitivity 84.8%). Orange boxes represent patterns associated with periodontitis, and the green box represents a pattern associated with periodontal health. Numbers in the boxes indicate the numbers of individuals clinically diagnosed as periodontitis patients or as controls (a minimum of five cases per branch-leaf). Other bacterial species did not enter into the decision tree. *Positive = species ≥ threshold; **Negative = species < threshold; ***A.a. = A. actinomycetemcomitans.

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

J Dent Res 92(1) 2013 Modeling Susceptibility to Periodontitis 49

machine learning technique, i.e., decision tree. We demonstrate that a decision tree can be a potential tool for predicting and modeling periodontitis. The best accuracy (80%) and AUC (73%) in distinguishing between periodontitis and no periodon-titis was obtained when microbial and genetic SNP data were combined. In other words, 308 out of 385 individuals were cor-rectly classified by a combination of microbial patterns of T. forsythia, P. gingivalis, and A. actinomycetemcomitans and genetic data from SNPs TNF -857 and IL-1A -889. The decision tree, formed solely on the basis of bacterial species, selected T. forsythia as the most discriminative factor, followed by P. gin-givalis and A. actinomycetemcomitans, and gave an accuracy of 80%, specificity of 74%, sensitivity of 85%, and AUC of 68%.

On the basis of anaerobic culture, several single bacterial species have been associated with periodontitis (Haffajee and Socransky, 1994; van Winkelhoff et al., 2002), and our study confirms the association of T. forsythia, P. gingivalis, A. actino-mycetemcomitans, P. intermedia, and P. micros, as single fac-tors, with periodontitis. In the present study, we used previously published thresholds for the detection of bacterial species (Rams et al., 1996; Meijndert et al., 2010). When the decision tree analysis was run without the thresholds, a decision tree with very poor accuracy was achieved (data not shown). The targeted bacterial species were found not only in periodontitis, but also in low levels in non-periodontitis individuals (van Winkelhoff et al., 2002; Kuboniwa et al., 2004), and therefore a minimal infectious dose was applied on the basis of anaerobic culture (Rams et al., 1996; Meijndert et al., 2010). T. forsythia was the strongest species associated with periodontitis as a single factor and achieved accuracy of 72% and AUC of 74%. But sensitivity of T. forsythia as a single factor predicting periodontitis reached only 65%. Clearly, our results show that a combination of differ-ent species in subgingival plaque is important in periodontitis. We found that the combination of T. forsythia, P. gingivalis, and A. actinomycetemcomitans gave higher accuracy and sensitivity

than information on the basis of only a single bacterial species. Previously, combinations of T. forsythia, P. gingivalis, and Treponema denticola (red cluster), and P. gingivalis, A. actino-mycetemcomitans, and P. intermedia have been reported to be discriminative for periodontitis (Socransky et al., 1998; Hyvarinen et al., 2009).

None of the SNPs studied, as a single factor, were associated with periodontitis, and a complex decision tree with a poor LOO cross-validation performance was formed. Thus, the current genetic variations are not distinctive for periodontitis. However, the best accuracy and AUC were reached with combination of the microbial data with the SNPs TNF -857 and IL-1A -889. TNF -857 and IL-1A -889 are polymorphisms in regulatory regions of cytokine genes. Functional effects of these polymor-phisms on gene transcription and cytokine production are still controversial (Higuchi et al., 1998; Uglialoro et al., 1998; Dominici et al., 2002; Um et al., 2011).

Using decision tree learning algorithms, we have shown that carriership of multiple periodontal pathogens in conjunction with certain genetic variations can achieve good accuracy and AUC in discriminating between periodontitis and no periodon-titis. In the analysis, no distinction was made between different forms of periodontitis, since the present study focused on risk factors for periodontitis vs. periodontal health; moreover, the current database was too small to build the decision trees for subgroups of periodontitis. Previously, with the decision tree learning approach, multiple SNPs have been reported to have better accuracy than single SNPs in identifying groups of indi-viduals at high risk for breast cancer and diabetes (Listgarten et al., 2004; Weedon et al., 2006).

It is evident that the validity of the analysis depends on the study design, including sample size, phenotype of periodontitis, choice of the SNPs and microbial species, as well as the detection method of microbial species. New techniques in bacterial detec-tion, like 454 pyrosequencing, have identified many new candidate

Figure 2. Decision tree based on bacterial patterns and SNPs. Overall LOO cross-validation accuracy is 80.0% (specificity 72.9% and sensitivity 84.8%). Orange boxes indicate that the pattern is associated with periodontal disease; green boxes indicate that the pattern is associated with periodontal health. Numbers in the boxes indicate the numbers of individuals clinically diagnosed as periodontitis patients or as controls (a minimum of two cases per branch-leaf). Genotype 11 = homozygous for common allele, 12 = heterozygous, 22 = homozygous for rare allele. Abbreviations: *T.f. = T. forsythia; P.g. = P. gingivalis; A.a. = A. actinomycetemcomitans. - = species < threshold; + = species ≥ threshold.

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

50 Laine et al. J Dent Res 92(1) 2013

bacterial species that may play even more important roles in the pathogenesis of periodontitis than the classic periodontal patho-gens included in the present study (Griffen et al., 2011). Furthermore, other candidate genes and SNPs may be important in the pathophysiology of periodontitis. Recent studies have reported novel candidate genes (ANRIL, COX2, and GLT6D1) and poly-morphisms for aggressive periodontitis (Schaefer et al., 2010a,b, 2011); future studies need to be performed to investigate whether these genes also play a major role in chronic periodontitis.

In conclusion, we identified, by decision tree analysis, a complex of potential causative factors for periodontitis: the simultaneous presence of T. forsythia, P. gingivalis, and A. acti-nomycetemcomitans, and SNPs TNF -857 and IL-1A -889 as discriminators between periodontitis and periodontal health. On the basis of these variables, patterns may be recognized to iden-tify individuals at risk for periodontitis. We have shown that the bioinformatics tool decision tree learning is valuable in model-ing the multifactorial and complex nature of periodontitis.

AcKnOWlEDgMEnts

Work reported herein was supported in part by the INFOBIOMED, Network of Excellence, IST-507585, http://www.infobiomed.org. Results, views, and opinions expressed herein do not necessarily correspond to official INFOBIOMED or European Commission positions, and the responsibility lies entirely with the authors. We thank S. Peña, S. Morre, B. Crusius, A.J. van Winkelhoff, F. Lessman, and C. van der Palen for their valuable contributions. The authors declare no potential conflicts of interest with respect to the authorship and/or publi-cation of this article.

rEFErEncEsBamber D (1975). The area above the ordinal dominance graph and the area

below the receiver operating characteristic graph. J Math Psychol 12:387-415.

Bizzarro S, van der Velden U, ten Heggeler JM, Leivadaros E, Hoek FJ, Gerdes VE, et al. (2007). Periodontitis is characterized by elevated PAI-1 activity. J Clin Periodontol 34:574-580.

Cortes C, Mohri M (2004). AUC optimization vs. error rate minimization. Research Report, AT&T Labs. Accessed on 10/1/2012 from: http://books.nips.cc/papers/files/nips16/NIPS2003_AA40.pdf

Dominici R, Cattaneo M, Malferrari G, Archi D, Mariani C, Grimaldi LM, et al. (2002). Cloning and functional analysis of the allelic polymor-phism in the transcription regulatory region of interleukin-1 alpha. Immunogenetics 54:82-86.

Griffen AL, Beall CJ, Campbell JH, Firestone ND, Kumar PS, Yang ZK, et al. (2011). Distinct and complex bacterial profiles in human periodontitis and health revealed by 16S pyrosequencing. ISME J 6:1176-1185.

Haffajee AD, Socransky SS (1994). Microbial etiological agents of destruc-tive periodontal diseases. Periodontol 2000 5:78-111.

Hanley JA, McNeil BJ (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 143:29-36.

Higuchi T, Seki N, Kamizono S, Yamada A, Kimura A, Kato H, et al. (1998). Polymorphism of the 5′-flanking region of the human tumor necrosis factor (TNF)-alpha gene in Japanese. Tissue Antigens 51:605-612.

Huang LC, Hsu SY, Lin E (2009). A comparison of classification methods for predicting Chronic Fatigue Syndrome based on genetic data. J Transl Med 7:81.

Hyvarinen K, Laitinen S, Paju S, Hakala A, Suominen-Taipale L, Skurnik M, et al. (2009). Detection and quantification of five major periodontal pathogens by single copy gene-based real-time PCR. Innate Immun 15:195-204.

Kuboniwa M, Amano A, Kimura KR, Sekine S, Kato S, Yamamoto Y, et al. (2004). Quantitative detection of periodontal pathogens using real-time polymerase chain reaction with TaqMan probes. Oral Microbiol Immunol 19:168-176.

Laine ML, Farre MA, Gonzalez G, van Dijk LJ, Ham AJ, Winkel EG, et al. (2001). Polymorphisms of the interleukin-1 gene family, oral microbial pathogens, and smoking in adult periodontitis. J Dent Res 80:1695-1699.

Laine ML, Morre SA, Murillo LS, van Winkelhoff AJ, Pena AS (2005). CD14 and TLR4 gene polymorphisms in adult periodontitis. J Dent Res 84:1042-1046.

Laine ML, Crielaard W, Loos BG (2012). Genetic susceptibility to periodon-titis. Periodontol 2000 58:37-68.

Le Bourhis L, Benko S, Girardin SE (2007). Nod1 and Nod2 in innate immunity and human inflammatory disorders. Biochem Soc Trans 35(Pt 6):1479-1484.

Listgarten J, Damaraju S, Poulin B, Cook L, Dufour J, Driga A, et al. (2004). Predictive models for breast cancer susceptibility from multiple single nucleotide polymorphisms. Clin Cancer Res 10:2725-2737.

Loos BG, Leppers-Van de, Straat FG, Van de, Winkel JG, Van der Velden U (2003). Fcgamma receptor polymorphisms in relation to periodontitis. J Clin Periodontol 30:595-602.

Loos BG, John RP, Laine ML (2005). Identification of genetic risk factors for periodontitis and possible mechanisms of action. J Clin Periodontol 32(Suppl 6):159-179.

Meijndert L, van der Reijden WA, Raghoebar GM, Meijer HJ, Vissink A (2010). Microbiota around teeth and dental implants in periodontally healthy, partially edentulous patients: is pre-implant microbiological testing relevant? Eur J Oral Sci 118:357-363.

Michalowicz BS (1994). Genetic and heritable risk factors in periodontal disease. J Periodontol 65(5 Suppl):479S-488S.

Nicu EA, Laine ML, Morre SA, Van der Velden U, Loos BG (2009). Soluble CD14 in periodontitis. Innate Immun 15:121-128.

Quinlan JR (1993). C4.5: Programs for machine learning. San Mateo, CA: Morgan Kaufmann Publishers.

Rams TE, Listgarten MA, Slots J (1996). Utility of 5 major putative peri-odontal pathogens and selected clinical parameters to predict periodon-tal breakdown in patients on maintenance care. J Clin Periodontol 23:346-354.

Schaefer AS, Richter GM, Nothnagel M, Laine ML, Noack B, Glas J, et al. (2010a). COX-2 is associated with periodontitis in Europeans. J Dent Res 89:384-388.

Schaefer AS, Richter GM, Nothnagel M, Manke T, Dommisch H, Jacobs G, et al. (2010b). A genome-wide association study identifies GLT6D1 as a susceptibility locus for periodontitis. Hum Mol Genet 19:553-562.

Schaefer AS, Richter GM, Dommisch H, Reinartz M, Nothnagel M, Noack B, et al. (2011). CDKN2BAS is associated with periodontitis in differ-ent European populations and is activated by bacterial infection. J Med Genet 48:38-47.

Socransky SS, Haffajee AD, Cugini MA, Smith C, Kent RL Jr (1998). Microbial complexes in subgingival plaque. J Clin Periodontol 25:134-144.

Uglialoro AM, Turbay D, Pesavento PA, Delgado JC, McKenzie FE, Gribben JG, et al. (1998). Identification of three new single nucleotide polymorphisms in the human tumor necrosis factor-alpha gene pro-moter. Tissue Antigens 52:359-367.

Um JY, Rim HK, Kim SJ, Kim HL, Hong SH (2011). Functional polymor-phism of IL-1 alpha and its potential role in obesity in humans and mice. PLoS One 6:e29524.

Ursum J, van der Weijden MA, van Schaardenburg D, Prins AP, Dijkmans BA, Twisk JW, et al. (2010). IL10 GGC haplotype is positively and HLA-DQA1*05-DQB1*02 is negatively associated with radiographic progression in undifferentiated arthritis. J Rheumatol 37:1431-1438.

van Winkelhoff AJ, Loos BG, van der Reijden WA, van der Velden U (2002). Porphyromonas gingivalis, Bacteroides forsythus and other putative periodontal pathogens in subjects with and without periodontal destruction. J Clin Periodontol 29:1023-1028.

Weedon MN, McCarthy MI, Hitman G, Walker M, Groves CJ, Zeggini E, et al. (2006). Combining information from common type 2 diabetes risk polymorphisms improves disease prediction. PLoS Med 3:e374.

Zhang J, Sun X, Xiao L, Xie C, Xuan D, Luo G (2011). Gene polymor-phisms and periodontitis. Periodontol 2000 56:102-124.

at Clayton State University Library on October 11, 2014 For personal use only. No other uses without permission.jdr.sagepub.comDownloaded from

© 2013 International & American Associations for Dental Research

566

CORRIGENDUM

DOI: 10.1177/0022034513486610

© International & American Associations for Dental Research

Laine ML, Moustakis V, Koumakis L, Potamias G, Loos BG (2013). Modeling susceptibility to periodontitis. J Dent Res 92:45-50. (doi: 10.1177/0022034512465435)

The decision tree of Figure 2 contained an error: Tf+ or Tf+/Pg- or Tf-/Pg-/Aa+* should be Tf+ or Tf-/Pg+ or Tf-/Pg-/Aa+*. The corrected figure is printed below.

Genotype 22Genotype 11 Genotype 12

TNF -857

Tf-, Pg- & Aa-*

Periodontitis 0Control 3

Periodontitis 73Control 13

Genotype 11 Genotype 12

IL-A -889

Genotype 12

Periodontitis 85Control 21

Genotype 12 Genotype 22

Periodontitis 5Control 6

IL-A -889

TNF -857

Periodontitis 20Control 0

Tf+ or Tf-/Pg+ or Tf-/Pg-/Aa+*

Subgingival bacteria

Genotype 11

Periodontitis 26Control 92

Genotype 11

Periodontitis 2Control 13

Periodontitis 17Control 5

Figure 2. Decision tree based on bacterial patterns and SNPs. Overall LOO cross-validation accuracy is 80.0% (specificity 72.9% and sensitivity 84.8%). Orange boxes indicate that the pattern is associated with periodontal disease; green boxes indicate that the pattern is associated with peri-odontal health. Numbers in the boxes indicate the numbers of individuals clinically diagnosed as periodontitis patients or as controls (a minimum of two cases per branch-leaf). Genotype 11 = homozygous for common allele, 12 = heterozygous, 22 = homozygous for rare allele. Abbreviations: *T.f. = T. forsythia; P.g. = P. gingivalis; A.a. = A. actinomycetemcomitans. – = species < threshold; + = species ≥ threshold.