Machine learning-based analysis of oral rinse samples to identify candidate proteomic signatures for severe periodontitis: a pilot study

Abstract

This pilot study investigated whether candidate protein signatures from oral rinse samples can distinguish patients with severe periodontitis (stage III/IV) and its subtypes, generalized and localized periodontitis, from non-periodontitis controls. Participants rinsed with phosphate-buffered saline, and samples were analyzed using a Proximity Extension Assay targeting 92 inflammatory and 92 immuno-oncology proteins. A machine learning approach using repeated nested cross-validation and SHAP was implemented to identify protein signatures. The study included 38 patients (18 with localized periodontitis and 20 with generalized periodontitis) and 16 controls. After data preprocessing, 54 samples and 141 proteins were retained. Proteins Gal-1, HGF, TNFSF14, CD27, and ARG1 distinguished periodontitis from controls (ROC-AUC = 0.85, 95% CI 0.82, 0.87). For generalized periodontitis, we found a protein signature including TNFSF14, Gal-1, STAMBP, MUC-16, S100A12, HGF, CASP-8, CD27, LAP TGF-β1, TNFRSF9, and uPA (ROC-AUC = 0.92, 95% CI 0.90, 0.94). For localized periodontitis, we identified ARG1 (ROC-AUC = 0.72, 95% CI 0.68, 0.76). No proteomic signature distinguishing generalized periodontitis from localized periodontitis was identified. This pilot study indicated that oral rinses are suitable for proteomic profiling, and there was a putative protein signature that could differentiate periodontitis, generalized periodontitis, and localized periodontitis from controls. These findings warrant validation in larger independent cohorts, including a clearly defined gingivitis group, before real-world non-invasive screening applications can be considered.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-61760-8.

Introduction

Periodontitis is a chronic, multifactorial inflammatory condition that affects the supporting structures of the teeth^1,2. Globally, its prevalence is estimated to be approximately 62%, whereas severe periodontitis has an age-standardized prevalence of 12.5%^3,4. The disease often progresses silently, with symptoms frequently going unnoticed unless the patient regularly consults a dental professional. Consequently, individuals who do not receive routine dental care may experience periodontal deterioration, potentially leading to symptoms such as periodontal abscesses, tooth mobility, and tooth loss, ultimately impairing their quality of life⁵. Therefore, a simple, non-invasive, and user-friendly screening method for periodontitis, particularly one suitable for self-testing, could be highly valuable, especially for individuals with limited access to dental care.

Saliva is a biofluid rich in biological information relevant to both oral and systemic health. The non-invasive, painless, and easy-to-perform collection makes saliva well-suited for diagnostic applications⁶. An emerging focus within salivary diagnostics is proteomic profiling, which examines the total protein composition of saliva to identify potential disease biomarkers⁷. Several studies have demonstrated that salivary proteomic analysis can distinguish periodontitis patients from healthy controls^8–11. However, collecting whole saliva presents practical challenges. The process is relatively time-consuming, typically requiring 3–5 min of drooling and expectoration per participant, which may hinder large-scale implementation. Furthermore, saliva collection may be challenging for individuals with oral dryness due to conditions such as diabetes mellitus or Sjögren’s syndrome or as a side effect of medication^12,13.

Oral rinse sampling, also referred to as “mouthwash”, offers a promising alternative. The method is quicker, requiring only 30 s, and is better tolerated by individuals experiencing oral dryness, making it more suitable for large cohort studies. For such applications, proteomic analysis of oral rinses should ideally enable the rapid processing of numerous samples while maintaining high sensitivity, high specificity, and cost-effectiveness¹⁴. Sandwich immunoassays, such as the Proximity Extension Assay (PEA; developed by Olink®), are particularly promising in this context. In PEA, a pair of antibodies is used to target a specific protein, each with DNA strand extensions that can be amplified by polymerase chain reaction (PCR) to produce a measurable signal, thereby enabling the sensitive and specific detection of multiple proteins^15–17. Oral rinse sampling for proteomic analysis remains underexplored. In a previous study, we demonstrated that oral rinse samples collected with phosphate-buffered saline (PBS) could be used for metabolomic analysis¹⁸. Fresh aliquots from the same oral rinse samples were used in the current pilot study for proteomic analysis using PEA. The proteomic data were analyzed using machine learning to identify candidate proteomic signatures capable of differentiating patients with severe periodontitis (stage III/IV) and its subtypes, generalized periodontitis (stage III/IV), and localized periodontitis (stage III/IV) from non-periodontitis controls, as well as distinguishing between generalized and localized periodontitis. Ultimately, an oral rinse protocol for periodontitis screening could be implemented beyond dental clinics, extending to medical settings, hospitals, physicians’ offices, homes and pharmacies.

We hypothesized that severe periodontitis (stage III/IV) would exhibit a distinct proteomic signature detectable in oral rinse samples using PEA compared to non-periodontitis controls, with additional differences between the proteomic signatures of localized and generalized forms of periodontitis.

Results

Demographic and clinical characteristics

The study participants consisted of patients with severe periodontitis and non-periodontitis controls and represent a subset of the study population of our previous cross-sectional study on oral and systemic conditions in periodontitis patients and controls¹⁹, which was also used in our previous study on metabolomics¹⁸. In the parent study, patients with severe periodontitis (stage III/IV, grade B or C) were referred to the Department of Periodontology at the Academic Centre of Dentistry Amsterdam (ACTA), and non-periodontitis controls were recruited from the general dentistry clinic at the same institution.

Demographic and clinical characteristics are presented in Table 1. The study population comprised three groups: 16 non-periodontitis controls, 18 patients with localized periodontitis, and 20 patients with generalized periodontitis. The three groups differed significantly in smoking status, obesity (BMI ≥ 30 kg/m²), metabolic syndrome prevalence based on the National Cholesterol Education Program’s Adult Treatment Panel III (NCEP ATP III) criteria, and 10-year cardiovascular mortality risk (all p < 0.05); all other demographic and clinical characteristics were comparable between groups (Table 1). The self-reported comorbidities are presented in Supplementary Table S1.

Table 1.

Demographic and clinical characteristics of the study population.

	Controls (n = 16)	Localized periodontitis (n = 18)	Generalized periodontitis (n = 20)	Overall p value (between 3 groups)	Total periodontitis (n = 38)	p value (Total periodontitis versus Controls)
Age (years)	59.3 ± 10.5	51.9 ± 8.4	56.2 ± 11.2	0.113	54.2 ± 10.1	0.100
Sex				0.261		0.277
Male	8 (50.0)	10 (55.6)	15 (75.0)		25 (65.8)
Female	8 (50.0)	8 (44.4)	5 (25.0)		13 (34.2)
Education				0.587		0.391
Primary	2 (12.5)	1 (5.6)	3 (15.0)		4 (10.5)
Secondary	4 (25.0)	8 (44.4)	9 (45.0)		17 (44.7)
Beyond secondary	10 (62.5)	9 (50.0)	8 (40.0)		17 (44.7)
Smoking status				0.003		0.002
Current	0 (0.0)	6 (33.3)	11 (55.0)		17 (44.7)
Former	6 (37.5)	6 (33.3)	7 (35.0)		13 (34.2)
Never	10 (62.5)	6 (33.3)	2 (10.0)		8 (21.1)
Body Mass Index (kg/m2)	25.4 ± 3.1	27.0 ± 3.6	27.3 ± 4.5	0.354	27.1 ± 4.1	0.150
≥ 30	0 (0.0)	5 (27.8)	5 (25.0)	0.069	10 (26.3)	0.023
Waist circumference (cm)	94.5 ± 7.6	92.6 ± 12.1	99.8 ± 12.1	0.120	96.4 ± 12.5	0.569
Systolic BP (mmHg)	132 ± 18.5	123 ± 13.0	137 ± 20.3	0.054	130 ± 18.4	0.712
Diastolic BP (mmHg)	80 ± 6.6	81 ± 6.9	83 ± 10.2	0.510	82 ± 8.8	0.237
Total cholesterol (mmol/L)	5.5 ± 1.3	5.2 ± 1.2	5.2 ± 1.2	0.769	5.2 ± 1.2	0.466
LDL-C (mmol/L)	3.5 ± 1.2	3.3 ± 1.1	3.2 ± 0.9	0.647	3.2 ± 1.0	0.386
HDL-C (mmol/L)	1.1 ± 0.2	1.1 ± 0.4	1.1 ± 0.3	0.737	1.1 ± 0.3	0.826
Triglycerides (mmol/L)	1.5 ± 0.7	1.9 ± 1.3	2.3 ± 1.7	0.295	2.1 ± 1.5	0.200
Mean HbA1c (%)	5.1 ± 0.3	5.3 ± 0.7 (n = 17)	5.7 ± 1.0	0.164	5.5 ± 0.8	0.140
Mean HbA1c (mmol/mol)	(32.2 ± 3.1)	(34.7 ± 7.4)	(38.6 ± 10.4)		(36.8 ± 9.3)
Metabolic syndrome
NCEP-ATP III	6 (37.5)	6 (33.3)	14 (70.0)	0.044	20 (52.6)	0.310
IDF	9 (56.3)	10 (55.6)	15 (75.0)	0.232	25 (65.8)	0.507
10-year CVD mortality risk
> 5% (‘high’)	3 (18.8)	4 (22.2)	11 (55.0)	0.019	15 (39.5)	0.140
> 10% (‘very high’)	2 (12.5)	3 (16.7)	10 (50.0)	0.011	13 (34.2)	0.104

Data are presented as mean ± standard deviation (SD) or n (%). All periodontitis patients were stage III/IV. For comparisons among the three groups (non-periodontitis controls, localized periodontitis, and generalized periodontitis), ANOVA and χ² tests (linear by linear association) were used. Mann–Whitney and Kruskal–Wallis tests for mean HbA1c were applied due to not normally distributed data. BP, blood pressure; LDL-C, Low-Density Lipoprotein Cholesterol; HDL-C, High-Density Lipoprotein Cholesterol; HbA1c, glycated hemoglobin; NCEP-ATP III, National Cholesterol Education Program Adult Treatment Panel III; IDF, International Diabetes Federation; CVD, Cardiovascular.

Dental and periodontal parameters

The dental and periodontal parameters are presented in Table 2. Patients with generalized periodontitis had the lowest number of teeth (23.9), in comparison to those with localized periodontitis (26.5), and controls (26.6) (p = 0.040). Patients with generalized periodontitis more often had teeth with alveolar bone loss of ≥ 33% of the root length (p < 0.001) and more often had teeth or sites with deep probing pocket depths (≥ 6 mm) (p < 0.001). A total of 16 out of 20 patients with generalized periodontitis and 11 out of 18 patients with localized periodontitis were classified as grade C (p < 0.001).

Table 2.

Dental and periodontal parameters.

	Controls (n = 16)	Localized periodontitis (n = 18)	Generalized periodontitis (n = 20)	Overall P value (between 3 groups)	Total periodontitis (n = 38)	P value (Total periodontitis versus Controls)
# Teeth	26.6 ± 2.3	26.5 ± 2.0	23.9 ± 3.9	0.040	25.11 ± 3.4	0.216
# Teeth with ≥ 33% bone loss	0.0	4.1 ± 2.9	14.3 ± 3.8	< 0.001	9.5 ± 6.2	NA
# Teeth with PPD ≥ 6 mm	0.0	7.3 ± 4.8	12.8 ± 7.3	< 0.001	10.2 ± 6.8	NA
# Sites with PPD ≥ 6 mm	0.0	15.2 ± 12.5	31.6 ± 23.3	< 0.001	23.8 ± 20.5	NA
Full mouth plaque score	NA	53.3 ± 22.9 (n = 17)	61.3 ± 28.6	0.357	57.4 ± 25.9	NA
Full mouth bleeding score	NA	48.8 ± 25.9 (n = 17)	66.1 ± 26.2	0.052	57.6 ± 27.1	NA
Grade
A	0.0	0 (0.0)	0 (0.0)	< 0.001	0 (0.0)	NA
B	0.0	7 (38.9)	4 (20.0)		11 (28.9)
C	0.0	11 (61.1)	16 (80.0)		27 (71.1)

Data are presented as mean ± standard deviation (SD) or n (%). All periodontitis patients were stage III/IV. For comparisons among the three groups (non-periodontitis controls, localized periodontitis, and generalized periodontitis), ANOVA and χ² tests (linear by linear association) were used. Mann–Whitney and Kruskal–Wallis tests for # Teeth, # Teeth with probing pocket depth ≥ 6 mm, # Sites with probing pocket depth ≥ 6 mm were applied due to not normally distributed data. PPD, probing pocket depth.

Data preprocessing

For proteomic signature discovery, two high-throughput proteomic panels were employed, one inflammatory (92 targeted proteins) and one immuno-oncology panel (92 targeted proteins). From the initial dataset of 184 proteins across both panels, 49 proteins were unique to each panel and 43 were present in both panels. Deduplication of the common proteins based on the one with the highest percentage of values above the limit of detection (LOD) resulted in the exclusion of 43 proteins from the total of 184 proteins. The remaining 141 proteins were retained for the machine learning analysis.

Final model selection

Table 3 shows the performance metrics for three models used in the repeated nested cross-validation (CV): Logistic Regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). The Area Under the Receiver Operating Characteristic Curve (ROC-AUC) was the primary metric for comparing model performance, while sensitivity (recall), specificity, balanced accuracy, and precision were secondary metrics. All metrics were aggregated across all outer fold iterations of the repeated nested CV, and their means, standard deviations, and confidence intervals are presented.

Table 3.

Performance metrics for the classification tasks of periodontitis and its subtypes with non-periodontitis controls: (1) periodontitis versus controls (2) generalized periodontitis versus controls, and (3) localized periodontitis versus controls.

Model	LR	RF	XGBoost
Periodontitis versus controls
ROC-AUC	0.85 ± 0.12 [0.82, 0.87]	0.83 ± 0.13 [0.80, 0.85]	0.79 ± 0.16 [0.76, 0.82]
Sensitivity (Recall)	0.76 ± 0.15 [0.73, 0.79]	0.80 ± 0.16 [0.76, 0.83]	0.78 ± 0.17 [0.74, 0.81]
Specificity	0.78 ± 0.24 [0.73, 0.82]	0.68 ± 0.28 [0.63, 0.73]	0.64 ± 0.27 [0.58, 0.69]
Balanced accuracy	0.77 ± 0.13 [0.74, 0.79]	0.74 ± 0.15 [0.71, 0.77]	0.71 ± 0.15 [0.68, 0.74]
Precision	0.90 ± 0.10 [0.88, 0.92]	0.87 ± 0.11 [0.85, 0.89]	0.84 ± 0.11 [0.82, 0.86]
Brier score	0.20*	0.16*	0.18*
Log Loss	0.61*	0.71*	0.57*
ECE	0.16	0.06	0.10
MCE	0.30	0.19	0.30
Generalized periodontitis versus controls
ROC-AUC	0.90 ± 0.11 [0.87, 0.92]	0.92 ± 0.09 [0.90, 0.94]	0.89 ± 0.12 [0.86, 0.91]
Sensitivity (Recall)	0.84 ± 0.17 [0.80, 0.88]	0.83 ± 0.16 [0.79, 0.86]	0.79 ± 0.19 [0.75, 0.83]
Specificity	0.84 ± 0.17 [0.81, 0.88]	0.86 ± 0.15 [0.83, 0.89]	0.86 ± 0.16 [0.82, 0.89]
Balanced accuracy	0.84 ± 0.11 [0.82, 0.87]	0.85 ± 0.09 [0.83, 0.86]	0.82 ± 0.11 [0.80, 0.85]
Precision	0.89 ± 0.12 [0.86, 0.91]	0.90 ± 0.11 [0.88, 0.92]	0.89 ± 0.12 [0.87, 0.92]
Brier score	0.18*	0.12*	0.14*
Log Loss	0.57*	0.39*	0.43*
ECE	0.11	0.03	0.05
MCE	0.26	0.05	0.16
Localized periodontitis versus controls
ROC-AUC	0.72 ± 0.15 [0.68, 0.76]	0.69 ± 0.16 [0.66, 0.73]	0.68 ± 0.15[0.65, 0.72]
Sensitivity (Recall)	0.63 ± 0.25 [0.57, 0.68]	0.64 ± 0.25 [0.59, 0.69]	0.65 ± 0.24 [0.60, 0.71]
Specificity	0.67 ± 0.27 [0.60, 0.72]	0.68 ± 0.25 [0.62, 0.73]	0.64 ± 0.26 [0.58, 0.70]
Balanced accuracy	0.65 ± 0.14 [0.61, 0.68]	0.66 ± 0.14 [0.63, 0.69]	0.65 ± 0.15 [0.61, 0.68]
Precision	0.73 ± 0.19 [0.69, 0.77]	0.72 ± 0.20 [0.67, 0.76]	0.69 ± 0.21 [0.64, 0.74]
Brier score	0.23*	0.25*	0.26*
Log Loss	0.69*	0.85*	0.80*
ECE	0.05	0.14	0.17
MCE	0.22	0.21	0.24

For each classification task, Logistic Regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) were evaluated. Data are presented as mean ± standard deviation (SD) and 95% confidence interval (CI). The models in bold indicate the final model for each classification task, selected based on the highest ROC-AUC. ROC-AUC, Receiver Operating Characteristic—Area Under the Curve; ECE, Expected Calibration Error; MCE, Maximum Calibration Error.

*Reference values for a no-skill (random) model: periodontitis versus controls (Brier: 0.21, Log Loss: 0.61), generalized periodontitis versus controls (Brier: 0.25, Log Loss: 0.69), and localized periodontitis versus controls (Brier: 0.25, Log Loss: 0.69).

For the classification of periodontitis versus controls, LR achieved the highest performance with a ROC-AUC of 0.85 ± 0.12 (95% CI 0.82, 0.87), outperforming RF (ROC-AUC = 0.83 ± 0.13, 95% CI 0.80, 0.85) and XGBoost (ROC-AUC = 0.79 ± 0.16, 95% CI 0.76, 0.82). For the generalized periodontitis versus controls classification, RF had the highest ROC-AUC (0.92 ± 0.09, 95% CI 0.90,0.94), compared to LR (ROC-AUC = 0.90 ± 0.11, 95% CI 0.87, 0.92) and XGBoost (ROC-AUC = 0.89 ± 0.12, 95% CI 0.86, 0.91). The localized periodontitis versus controls classification also achieved the highest ROC-AUC of 0.72 ± 0.15 (95% CI 0.68, 0.76) with LR, while RF and XGBoost had lower ROC-AUCs of 0.69 ± 0.16 (95% CI 0.66, 0.73) and 0.68 ± 0.15 (95% CI 0.65, 0.72), respectively. The classification of localized versus generalized periodontitis did not identify any proteins across all three machine learning models during repeated nested CV, indicating insufficient discriminatory power between localized and generalized periodontitis. The best-performing models for each classification were selected as the final models and are indicated in bold in Table 3. Supplementary Table S2 presents the hyperparameters of the models. Table 3 shows the Brier score, Log Loss, Expected Calibration Error (ECE), and Maximum Calibration Error (MCE), which are the calibration metrics for the three models. For the periodontitis versus controls classification, RF demonstrated the lowest Brier score (0.16), indicating moderate calibration performance, while all models were below the no-skill reference of 0.21 (LR = 0.20, XGBoost = 0.18). For the generalized periodontitis versus controls classification, calibration was stronger: all models substantially outperformed the no-skill Brier score reference (0.25), with the RF achieving the lowest Brier score (0.12). For the localized periodontitis versus controls classification, Brier scores ranged from 0.23 to 0.26, approaching the no-skill (random) reference (0.25), indicating limited calibration performance consistent with the lower overall discriminatory power for this classification and underlining the weak overall reliability of this classification.

Supplementary Table S3 presents the proteins selected from the repeated nested CV with their concurrent selection frequencies in the repeated nested CV for each classification. From the 141 proteins after data preprocessing, the thresholds were calculated using the stability selection framework to control false positive discoveries. The resulting thresholds were 59% for periodontitis versus controls and 53% for generalized periodontitis versus controls and localized periodontitis versus controls. Five proteins exceeded the 59% threshold for distinguishing periodontitis from controls: Galectin-1 (Gal-1, selection frequency = 98%), Tumor Necrosis Factor (ligand) Superfamily Member 14 (TNFSF14, selection frequency = 98%), Hepatocyte Growth Factor (HGF, selection frequency = 92%), Cluster of Differentiation 27 (CD27, selection frequency = 65%), and Arginase 1 (ARG1, selection frequency = 65%). For generalized periodontitis versus controls, eleven proteins surpassed the 53% threshold: TNFSF14 was the most frequently selected feature (selection frequency = 100%), followed by Gal-1 (selection frequency = 99%), Latency-Associated Peptide of Transforming Growth Factor Beta 1 (LAP TGF-β1, selection frequency = 99%), HGF (selection frequency = 85%), Caspase-8 (CASP-8, selection frequency = 75%), Mucin 16 (MUC-16, selection frequency = 74%), STAM Binding Protein (STAMBP, selection frequency = 68%), CD27 (selection frequency = 68%), Tumor Necrosis Factor Receptor Superfamily Member 9 (TNFRSF9, selection frequency = 68%), S100 Calcium Binding Protein A12 (S100A12, selection frequency = 65%), and urokinase-type plasminogen activator (uPA, selection frequency = 56%). For localized periodontitis versus controls, only ARG1 exceeded the 53% threshold, with a selection frequency of 64%. No proteins were selected for localized versus generalized periodontitis, indicating that no robust proteomic signature could distinguish between these groups.

Model interpretation

Figure 1 shows the Shapley additive explanation (SHAP) values for the selected proteins in the final models of each classification, where positive and negative values indicate contributions toward periodontitis and controls classification, respectively. For periodontitis versus controls (Panel A), Gal-1, HGF, and TNFSF14 emerged as the most discriminatory proteins, followed by CD27 and ARG1. For all five proteins, high values directed predictions toward periodontitis, while low values directed predictions toward the controls. The proportion of samples in which each protein contributed to periodontitis was 52% for Gal-1, 50% for HGF, 48% for TNFSF14, 41% for CD27, and 61% for ARG1. For generalized periodontitis versus controls (Panel B), the eleven selected proteins are presented in order of discriminatory power (highest to lowest): TNFSF14, MUC-16, LAP TGF-β1, CD27, S100A12, STAMBP, CASP-8, HGF, Gal-1, TNFRSF9, and uPA. For all proteins, high values directed predictions toward periodontitis, whereas low values directed predictions toward the controls. The proportion of samples in which each protein contributed to generalized periodontitis ranged from 53 to 61% for Gal-1, MUC-16, S100A12, HGF, CD27, TNFRSF9, and uPA, 47% for STAMBP, and approximately 50% for TNFSF14, LAP TGF-β1, and CASP-8. For localized periodontitis versus controls (Panel C), ARG1 was the only protein that predicted localized periodontitis versus controls. High ARG1 values directed predictions toward periodontitis, while low values directed predictions toward the controls, with ARG1 contributing more to the prediction of localized periodontitis (62%). Supplementary Table S3 presents the mean absolute SHAP values for each protein and their percentage contributions to the periodontitis and controls classifications.

Fig. 1.

Shapley additive explanation (SHAP) values for the proteins selected in the final models of each classification. SHAP values show the protein discriminatory power in predictions. The proteins are displayed in descending order of importance, with the most important proteins at the top of each panel. Positive values (right) indicate a contribution to the disease prediction; negative values (left) indicate a contribution to non-periodontitis controls prediction. The horizontal spread of points for each protein indicates the variability in its contribution to the model’s predictions across different samples. A narrow spread indicates that the protein has a uniform effect, whereas a wider spread indicates that the protein has variable effects across individuals. (A) SHAP values of the protein signature for predicting periodontitis versus controls. (B) SHAP values of the protein signature for predicting generalized periodontitis versus controls. (C) SHAP values of the protein signature for predicting localized periodontitis versus controls.

Post-hoc analyses

Figure 2 shows the box plots of the identified proteins. Mann–Whitney U tests showed significant differences in protein levels between controls and periodontitis (unadjusted and adjusted for multiple testing, p < 0.001). In all cases, the protein levels were higher in the periodontitis group than in the controls. In Supplementary Table S3, we present the raw and adjusted p values from the Mann–Whitney tests.

Fig. 2.

Box plots of the identified proteins after the machine learning analysis. The title of each boxplot presents the protein name, raw p value, and adjusted p value of the Mann–Whitney U test. Statistical significance was assessed using the Benjamini–Hochberg correction. The y-axis shows Normalized Protein eXpression (NPX) values on a log₂ scale and x-axis shows the periodontitis status (periodontitis, generalized periodontitis, localized periodontitis, non-periodontitis controls) (A) identified proteins for periodontitis versus controls, (B) identified proteins for generalized periodontitis versus controls, and (C) identified proteins for localized periodontitis versus controls.

Supplementary Figure S1 shows scatter plots of the Normalized Protein eXpression (NPX) values of the identified proteins, the number of sites with PPD ≥ 6 mm in the periodontitis group, and their Spearman correlations. Eight of the twelve candidate proteins showed statistically significant positive correlations after Benjamini–Hochberg adjustment.

Finally, we modelled the relationship between each of the identified proteins and periodontitis status using multivariable linear regressions, while accounting for potential confounding factors (background characteristics listed in Table 1 and comorbidities listed in Supplementary Table S1). Table 4 presents the results of these models for each protein, including potential confounders with a p value ≤ 0.1 from the initial univariate analyses. Notably, all proteins maintained their significant independent associations with the different forms of periodontitis (p ≤ 0.002), except for STAMBP in generalized periodontitis (p = 0.05). Several background characteristics and comorbidities were identified as significant confounders for individual proteins, including triglycerides, smoking, systolic blood pressure, sex, age, respiratory disease, and allergies. The observed associations of the identified putative proteins in periodontitis should be interpreted cautiously due to the small sample sizes and extremely low and uneven prevalence of comorbidities.

Table 4.

Confounder analyses.

	Protein	Univariate analysis	Multivariate analysis
		Potential Confounders (p ≤ 0.1)	Protein Adjusted p value	Confounders (p < 0.05)
Periodontitis versus controls	Gal-1	HbA1c, IBD, fibromyalgia, respiratory disease	< 0.001
	HGF	HbA1c, systolic blood pressure, smoking, IBD, respiratory disease	< 0.001	Systolic blood pressure
	TNFSF14	Triglycerides, IBD, respiratory disease	< 0.001
	CD27	Triglycerides	< 0.001	Triglycerides
	ARG1	Sex, metabolic syndrome IDF, fibromyalgia, respiratory disease, allergies	< 0.001	Sex, respiratory disease
Generalized periodontitis versus controls	TNFSF14	Total cholesterol, triglycerides	< 0.001
	Gal-1	Triglycerides	< 0.001
	STAMBP	Triglycerides, smoking, age	0.05	Triglycerides
	MUC-16	Sex, HbA1c, kidney disease	< 0.001
	S100A12	Smoking, allergies	0.001	Smoking
	HGF	Smoking	< 0.001
	CASP-8	Sex, allergies	< 0.001	Allergies
	CD27	Age, total cholesterol, triglycerides	< 0.001
	LAP TGF-β1	Triglycerides, age	< 0.001	Triglycerides, age
	TNFRSF9	Triglycerides	< 0.001
	uPA		< 0.001
Localized periodontitis versus controls	ARG1	Respiratory disease, allergies	0.002	Respiratory disease, allergies

Multivariable linear regression analysis was used to examine the relationships between identified proteins and periodontitis, generalized periodontitis, localized periodontitis, and non-periodontitis controls, after adjusting for potential confounders. The significance of these relationships is represented by the adjusted p value. The potential confounders were demographic and background characteristics, as shown in Table 1 and comorbidities listed in Supplementary Table S1. Variables were considered as potential confounders and included in the multivariable linear regression model if their p value on univariate linear regression analysis was ≤ 0.1. The analysis was conducted separately for (1) periodontitis versus controls, (2) generalized periodontitis versus controls, and (3) localized periodontitis versus controls. Gal-1, Galectin-1; HGF, Hepatocyte Growth Factor; TNFSF14, Tumor Necrosis Factor (Ligand) Superfamily Member 14, ARG1, Arginase 1; STAMBP, STAM Binding Protein; MUC-16, Mucin 16; S100A12, S100 Calcium Binding Protein A12; CASP-8, Caspase-8; CD27, Cluster of Differentiation 27; LAP TGF-β1, Latency-Associated Peptide of Transforming Growth Factor Beta 1; TNFRSF9, Tumor Necrosis Factor Receptor Superfamily Member 9; uPA, urokinase-type plasminogen activator, HbA1c, glycated hemoglobin; IDF, International Diabetes Federation; IBD, Inflammatory bowel disease.

Discussion

Main findings

In this pilot study, we demonstrated the feasibility of obtaining protein signatures from oral rinse samples with proteomic analysis performed using PEA with two Olink® panels targeting inflammatory and immuno-oncology proteins. Using machine learning on the oral rinse proteomic profiles, we identified candidate proteomic signatures that differentiated periodontitis (stage III/IV) from non-periodontitis controls, and separately, generalized periodontitis and localized periodontitis from controls. All NPX values from the identified proteins were higher in periodontitis and its subtypes than in the controls. Moreover, all correlations between the NPX values of the identified proteins and the number of sites with PPD ≥ 6 mm in the periodontitis group were positive, with higher NPX values associated with a greater number of deep periodontal pockets. Given the pilot nature of this study, small sample size, and absence of an independent validation cohort, these signatures should be regarded as putative candidates rather than established biomarkers, and any screening application would require validation in larger independent cohorts. Nevertheless, our findings highlight the potential of proteomic analysis of oral rinse samples for screening patients with severe periodontitis. In general, once candidate biomarkers are identified using proteomic techniques, less complex and less expensive methods than PEA can be applied for further validation.

Among all machine learning models, the highest performance was achieved for generalized periodontitis versus controls, followed by the classifications of all periodontitis cases versus controls and localized periodontitis versus controls. The intermediate performance of the model for periodontitis patients versus controls is likely due to the inclusion of both localized and generalized periodontitis patients in the periodontitis group. The model distinguishing localized from generalized periodontitis did not identify a protein signature distinguishing the disease extent in patients with severe periodontitis. This finding suggests shared inflammatory and immune-related profiles, despite differences in disease extent. Moreover, oral rinse sampling integrates secretions from the entire oral cavity and may obscure site-specific differences related to disease extent. Additionally, the small sample size of each subgroup limited the statistical power to detect subtle between-group differences. Consequently, the present findings should be interpreted cautiously; currently, the candidate signatures reported here refer to comparisons with non-periodontitis controls.

Machine learning considerations

A comparison of the algorithm performance revealed a consistent pattern: the performance decreased as the model complexity increased. LR achieved the highest performance for most classifications, which can be attributed to its linear decision boundary being well-suited to the data, making it less susceptible to overfitting. RF adds complexity by combining many decision trees, capturing nonlinear patterns, but requiring more data than other models to reliably estimate its parameters. It outperformed LR only for the generalized periodontitis versus controls classification, where a larger separation between the groups may have provided sufficient signal for a more complex model. XGBoost, as a boosted ensemble, was the most complex of the three models and demonstrated the lowest performance across all classifications, reflecting its tendency to overfit small datasets with a large feature number. This pattern supports using simpler algorithms for protein discovery when sample sizes are limited, particularly when the number of features is large.

The calibration metrics revealed a misalignment between discrimination and calibration for certain classifications. This means that models with superior performance in distinguishing periodontitis from controls (discrimination) do not necessarily produce accurate predicted probabilities (calibration). For the classification of periodontitis versus controls, LR was the best discriminator (highest ROC-AUC), whereas RF demonstrated the best probability calibration (lowest Brier score). For the classification of generalized periodontitis versus controls, RF was the best model for both discrimination and calibration. For the classification of localized periodontitis versus controls, the limited calibration performance reflected the overall difficulty and weak reliability of this classification task with small sample sizes. The objective of this study was biomarker discovery and not individual risk prediction based on probabilities; therefore, the ROC-AUC was used as the primary model selection criterion. Future studies aimed at developing a clinical screening tool with probability-based risk stratification should also optimize model calibration. All reported performance and calibration values should be interpreted as internal performance on repeated nested CV rather than external performance on an independent cohort. Even with this resampling strategy, repeated tuning on a small dataset may still yield overly optimistic performance.

Regarding class imbalance, we used class weighting and stratified CV instead of oversampling methods, such as SMOTE or SMOTETomek. Although these oversampling techniques have demonstrated benefits in large-sample, low-dimensional datasets^20,21, they can reduce variability and introduce artificial correlations in high-dimensional data with small sample sizes, without improving classification performance²². Given our dataset of 54 participants and 141 proteins, oversampling could risk the generation of synthetic protein profiles potentially compromising feature selection reliability.

Candidate proteomic signatures

Two panels targeting inflammatory and immuno-oncology proteins were selected because they capture a broad spectrum of proteins involved in immune and inflammatory pathways, as well as wound healing, all of which may play important roles in periodontitis^23,24. The inflammatory panel includes key mediators of immune activation and inflammation across a broad range of acute and chronic inflammatory conditions. The immuno-oncology panel complements the inflammatory panel by covering proteins involved in innate immune and pattern recognition-associated molecules, immunomodulatory mediators, and angiogenesis and tissue remodeling proteins (wound healing).

Five candidate protein biomarkers, Gal-1, HGF, TNFSF14, CD27, and ARG1, showed the strongest cross-validated discriminatory ability between periodontitis patients and controls, whereas ARG1 could also distinguish localized periodontitis patients from controls. Additionally, eleven candidate proteins, TNFSF14, Gal-1, STAMBP, MUC-16, S100A12, HGF, CASP-8, CD27, LAP TGF-β1, TNFRSF9, and uPA, showed the greatest potential to distinguish generalized periodontitis from controls. Notably, several potential confounders were identified as significant confounders for the candidate proteins (Table 4); therefore, the interpretation of these proteins as markers should be cautious. However, these proteins can be considered putative biomarkers for periodontitis, as their potential validity can also be inferred from their known biological functions. Below is a summary of their perceived associations with periodontitis, pending confirmatory validation.

Gal-1 (Galectin-1) is a β-galactoside–binding lectin found on cell surfaces and within cytoplasm²⁵. Under physiological conditions, it helps control excessive inflammation, thereby maintaining immune homeostasis and preventing tissue damage. However, in certain pathological conditions, Gal-1 can also promote pro-inflammatory responses²⁶. Previous studies reported significantly higher Gal-1 levels in gingival crevicular fluid (GCF) from periodontitis and gingivitis patients than in controls^27,28. ROC analysis showed that Gal-1 distinguished periodontitis and gingivitis from controls with 100% accuracy²⁸. In contrast, no significant differences in Gal-1 levels in whole saliva were observed across the study groups²⁷. Our ability to detect discriminatory Gal-1 levels in oral rinse samples is likely due to the presence of GCF in the oral rinse.

HGF (Hepatocyte Growth Factor) is a multifunctional cytokine involved in cell growth, migration, tissue repair, and angiogenesis, but it is also involved in inflammatory responses, tissue destruction, and bone resorption. HGF may play a protective role during the early stages of periodontitis, but it appears to shift toward a destructive, pathogenic role in later stages, contributing to alveolar bone loss²⁹. Previous studies have reported elevated HGF concentrations in saliva, GCF, and/or serum of periodontitis patients compared to controls^30–32. HGF is also detected in oral rinses collected with water, and HGF levels in these rinses increased proportionally with the severity of periodontitis³³. Salivary HGF levels above 0.470 ng/ml distinguished periodontitis patients from controls with an 88.2% accuracy, which increased further when HGF was combined with additional protein biomarkers³².

TNFSF14 (also called LIGHT—Lymphotoxin-like, Inducible, competes with HSV Glycoprotein D for binding to Herpesvirus entry mediator; a receptor expressed on T lymphocytes) is a member of the tumor necrosis factor (TNF) superfamily³⁴. In vitro, LIGHT may locally contribute to soft tissue and bone loss in periodontitis by stimulating inflammatory cytokines and tissue-degrading enzymes in gingival fibroblasts³⁵. A previous study analyzing stimulated saliva samples found no significant differences in LIGHT levels between individuals with periodontitis and/or peri-implantitis versus healthy controls³⁶. Furthermore, LIGHT, detected in GCF, was associated with increased probing depth after periodontal surgery, indicating its potential as a biomarker for identifying individuals at high risk of periodontitis recurrence after surgery³⁵.

ARG1 (Arginase 1) is an enzyme crucial in the urea cycle³⁷. Several studies have reported increased arginase activity in the saliva of periodontitis patients compared to that of controls, which decreased following non-surgical periodontal therapy^38–40. No previous study has directly compared ARG1 levels in localized and generalized periodontitis. ARG1 competes with inducible nitric oxide synthase (iNOS) for L-arginine, and the balance between these two enzymes may reflect the balance between tissue repair and tissue destruction at the site of inflammation⁴¹. We speculate that in localized periodontitis, where the disease is confined to fewer sites, both reparative and destructive immune pathways may be active, producing a detectable ARG1 signal. In generalized periodontitis, the widespread inflammatory burden may favor pro-inflammatory pathways, thereby suppressing the ARG1 signal. However, no mechanistic experiments assessed ARG1/iNOS activity or their functional interplay in this cohort; therefore, this interpretation remains hypothetical. Furthermore, the small sample sizes of each group in the current study and ARG1’s modest discriminatory performance for localized periodontitis suggest that ARG1 is a candidate marker of limited robustness that requires validation in larger cohorts.

S100A12 (S100 calcium binding protein A12), also known as EN-RAGE (Extracellular newly identified receptor for Advanced glycation end products binding protein) or calgranulin C, is a pro-inflammatory protein mainly secreted by activated neutrophils and monocytes/macrophages⁴². Salivary levels of S100A12 have previously been shown to correlate positively with periodontal parameters⁴³. Elevated levels of S100A12 have also been reported in GCF and serum of periodontitis patients⁴⁴. In this study, S100A12 was identified as a discriminatory protein for generalized periodontitis versus controls, but not for overall periodontitis versus controls or localized periodontitis versus controls. This pattern may reflect the greater inflammatory burden in generalized periodontitis, where more affected sites may result in higher S100A12 levels in oral rinses.

CASP-8 (Caspase-8) is a cysteine protease that mediates the extrinsic pathway of apoptosis and may contribute to periodontal tissue damage. An earlier study found no significant differences in GCF CASP-8 levels between periodontitis patients and controls, possibly due to the short half-life and detection window of CASP⁴⁵. This study used the ELISA method, which may have limited the detection of low concentrations of this transient protein. In this context, the higher analytical sensitivity of PEA could overcome this limitation, as it detects proteins below the ELISA detection threshold.

Although several discriminatory proteins identified here have also been reported at elevated levels in saliva or GCF in previous studies, differences in periodontitis across studies are expected and may reflect variations in proteomic techniques, sample types, collection procedures, and patient characteristics. Additionally, we evaluated whether the background characteristics influenced the identified putative protein biomarkers. Previous studies demonstrated that smoking and obesity (BMI ≥ 30 kg/m²) affect salivary protein profiles^46,47. The identified candidate biomarkers remained significantly elevated in periodontitis and its subtypes, even after adjusting for potential confounders. The significant confounders of the individual proteins are shown in Table 4. The multivariable analysis should be interpreted cautiously, as residual confounding cannot be excluded due to the small sample size and the low and uneven distribution of background characteristics and comorbidities between groups.

Strengths and limitations

A major strength of this investigation is that it is the first to apply PEA-based proteomic analysis to oral rinse samples. Previous proteomics studies on periodontitis predominantly used stimulated or unstimulated whole saliva or gingival crevicular fluid^8,32,48. PEA increases the detection of proteins at low concentrations, and the oral rinse sampling method seems to encompass the whole oral cavity and oral fluids. Based on our findings, this combination indicates that oral rinse samples are feasible for PEA-based proteomic analysis and are promising for large-scale studies. While many promising candidate biomarkers in saliva proteomics have been identified in previous studies, follow-up studies to validate these findings in large cohorts and translate them into clinical practice are often lacking⁸. Therefore, the use of oral rinse samples and PEA may offer a practical and sensitive approach for large-scale validation of candidate proteins.

As a pilot study based on a small sample without an independent validation cohort, several limitations should be considered when interpreting the current findings. The cohort of this study was small, and no a priori power calculation was performed. We used unprocessed aliquots from oral rinse samples collected in a previous study¹⁸; therefore, this study should be considered a pilot study. Post-hoc power calculations were not reported, as they are a direct function of the observed test statistic and add no independent information about sample size adequacy^49,50. However, our sample size was consistent with that of published salivary proteomics discovery studies on periodontitis, in which cohorts of 20 to 141 participants are typical⁸. Therefore, despite the small sample sizes, the identified proteins are promising candidates for screening severe periodontitis and warrant further evaluation in independent cohorts before any screening application can be considered. Their perceived functions and reported associations with periodontitis in previous studies support their potential as putative biomarkers.

Regarding external validation and generalizability potential, the combination of a small sample and a large number of proteins presents an overfitting risk despite the repeated nested CV approach, which may limit the clinical translation feasibility and the generalizability of the results. The performance metrics were based on a single cohort and not on an independent external validation set. Furthermore, the study population was recruited from a dental school clinic, which selects for severe, overt periodontitis (stage III/IV) and is not representative of individuals with periodontitis in primary care or community settings, where disease severity, prevalence, and comorbidity profiles may differ substantially. Consequently, the identified candidate signatures and model performance metrics cannot be directly generalized to primary care populations, and validation in broader, community-based cohorts is required before any screening application can be considered.

In the current pilot study, the controls were defined by the absence of periodontitis without formal classification of periodontal health or gingivitis. According to a large European cross-sectional study (n = 3551), 65.7% of adults had bleeding on probing at ≥ 10% of sites⁵¹, suggesting that gingival inflammation is highly prevalent in the general adult population. Therefore, there may have been certain levels of gingival inflammation present in the control group. If so, the identified signatures discriminated periodontitis from a group that already included some degree of gingival inflammation, which would support their specificity for periodontitis rather than gingival inflammation broadly. Nonetheless, the absence of a formally defined gingivitis group remains a limitation for interpreting specificity and future studies should include a defined gingivitis group to evaluate whether the identified proteomic signatures are specific to periodontitis or reflect broader periodontal inflammation.

Regarding pre-analytical variability, factors such as circadian rhythm, hydration status, and oral hygiene practices were not controlled beyond the standardized 1-h fasting period and may have contributed to inter-individual variability in the salivary proteome. Additionally, there may be variability related to oral dryness, residual saliva volume, rinsing vigor, gingival crevicular fluid contribution, and dilution with PBS. The effects of these factors and the reproducibility of the oral rinse protocol need to be investigated in the future across different operators and settings.

Since oral rinse samples contain epithelial cells and leukocytes, the protein composition of the post-thaw supernatant should not be regarded as exclusively extracellular. Cellular disruption during freeze–thaw processing may result in the release of cytoskeletal, nuclear, cytoplasmic, and granule-associated proteins into the soluble fraction. Consequently, some proteins identified in the oral-rinse supernatant may reflect intracellular contributions from shed epithelial cells and neutrophils rather than solely secreted salivary or gingival crevicular fluid proteins^52,53.

Another point to consider is that the identified proteomic signature for periodontitis should be interpreted as a discriminatory disease-associated profile and not as a direct indicator of current disease activity. It is unclear if the identified biomarkers indicate ongoing active tissue destruction or are cumulative results of past destruction. Longitudinal studies, including sampling before and after periodontal treatment and during maintenance therapy, are needed to determine whether these biomarkers track dynamic inflammatory changes and have the potential to be used for monitoring treatment, as presented in recent studies^54,55.

Finally, PEA reports relative quantification units (NPX values) rather than absolute protein concentrations; therefore, the clinical cutoff values for the identified proteins cannot be established from the present data. Future studies using absolute quantification methods, such as enzyme-linked immunosorbent assay (ELISA) are needed to define clinical thresholds and validate the usability of the identified proteins.

Clinical implications

The potential clinical applicability of an oral rinse protocol to screen for periodontitis could be particularly useful for individuals who do not regularly visit a dental professional but receive medical care for a non-communicable disease (e.g., diabetes mellitus, cardiovascular disease). Currently, periodontitis screening requires an oral examination by a dental professional, which limits screening capacity in primary medical care, non-dental settings, homes, and pharmacies. Protein biomarker-based screening using oral rinse samples could enable the identification of at-risk individuals, who can then be referred for periodontal assessment. A proteomic signature could provide an objective and complementary measure to the questionnaires used for screening periodontitis⁵⁶. Our pilot study identified candidate biomarkers; however, clinical translation requires validation in larger independent cohorts, development of point-of-care platforms, evaluation of real-world screening performance, feasibility outside a research laboratory, comparison with established periodontal screening approaches, and cost-effectiveness analysis.

Conclusion

To conclude, this pilot study using PEA and machine learning identified a putative proteomic signature of severe periodontitis (stage III/IV) in oral rinse samples. Our results indicate that oral rinsing is a feasible method for investigating the proteomic profiles of patients with severe periodontitis and its subtypes, localized and generalized periodontitis. The findings provide a basis for further studies on protein biomarkers for severe periodontitis in oral rinse samples. The presented proteomic signature needs validation in large independent cohorts, including a clearly defined gingivitis group, development of point-of-care platforms, and also cost-effectiveness analysis to ensure the robustness of the findings and make clinical implementation feasible.

Methods

Study design and recruitment

The study used a subset of the study population from our previous cross-sectional study (ClinicalTrials.gov Identifier NCT03459638), which was approved by the Medical Ethical Committee of the Amsterdam University Medical Center (2017.490 (A2019.151)-NL62337.029.17), and all measurements were performed at ACTA. In that study, consecutive patients with periodontitis referred to ACTA for diagnosis and treatment were enrolled¹⁹. Importantly, the participants in the present study were included during the final months of the enrollment period, when oral rinses were also collected. All patients had overt periodontitis, were newly referred to our periodontal clinic and had not been treated before. All patients with periodontitis had stage III/IV severity and were classified as grade B or C, indicating that only patients with severe periodontitis were included. In parallel, individuals without periodontitis who attended the ACTA clinic for general dentistry (e.g., routine dental check-ups or restorative procedures) were consecutively recruited as controls.

The only exclusion criterion was age; individuals younger than 40 years were not eligible. This minimum age was set in our parent study¹⁹, in which the recommended age for assessment of cardiovascular risk was above 40 years. All participants received verbal and written information about the purpose of the study and provided informed consent to participate. The study is reported in accordance with the STROBE guidelines⁵⁷.

At the referral visit, periodontitis patients underwent a comprehensive full-mouth periodontal examination conducted by periodontists or residents of the Department of Periodontology. Probing pocket depth (PPD), gingival recessions, and clinical attachment loss (CAL) were measured at six sites per tooth using a manual probe. Molar furcation involvement and tooth mobility were also assessed. Interproximal alveolar bone levels were assessed using dental radiographs (≤ 1 year old).

The initial assessment for periodontitis utilized the case definition criteria established by the Centers for Disease Control and Prevention–American Academy of Periodontology (CDC-AAP). Subjects with a positive diagnosis of periodontitis (≥ 2 interproximal sites with CAL ≥ 3 mm and ≥ 2 interproximal sites with PPD ≥ 4 mm, not on the same tooth, or one site with PPD ≥ 5 mm) were invited to participate⁵⁸. Subsequently, we applied staging (I-IV), grading (A, B, C), and determination of the extent (localized or generalized) per stage for each periodontitis case⁵⁹.

Control subjects were included when they: (1) did not fulfil the criteria for the case definition of periodontitis, (2) had not previously been treated for periodontitis, and (3) did not have interproximal alveolar bone loss on recent bitewing radiographs (≤ 1 year old); a distance of ≤ 3 mm between the cemento-enamel junction to the most coronal part of the radiographic alveolar crest was accepted for a non-periodontitis control subject. The term "non-periodontitis controls" was used for the control subjects because the inclusion criteria for controls were designed to exclude individuals with periodontitis, without applying a formal classification of periodontal health or gingivitis. Therefore, this term precisely describes the study design, without implying a periodontal health classification.

Clinical procedures

Self-reported questionnaires were used to collect information on age, sex, education level (categorized as primary, secondary, or above secondary, serving as a proxy for socio-economic status), smoking habits, height and comorbidities. A clinical examination was performed to assess blood pressure and waist circumference. For blood analysis, approximately 5–6 large drops of capillary blood were collected via a finger stick into a microtube containing 17 USP/mL lithium heparin and analyzed in a chemical lab⁶⁰. Metabolic syndrome was assessed using the two most widely used diagnostic criteria: the National Cholesterol Education Program Adult Treatment Panel III (NCEP ATP III) with age-adjusted blood pressure thresholds and the International Diabetes Federation (IDF). The 10-year risk of CVD mortality was assessed according to the European Systematic COronary Risk (SCORE) model¹⁹.

Oral rinse sample collection

The oral rinse samples used in this study were collected from a subset of participants from a parent study as explained above¹⁹. Participants were asked not to eat or brush their teeth for at least one hour before the oral rinse collection. Then, they were instructed to swallow once before rinsing with 10 mL of sterile Dulbecco’s PBS (ThermoFisher Scientific, Waltham, MA, USA) supplied in a 50 mL centrifugation tube. The patients were asked to rinse vigorously through all sides (left, right, upper, and lower) of the oral cavity for 30 s. After rinsing, the patients expectorated into a 30 mL medicine cup, and the oral rinse was poured into a 50 mL centrifugation tube. Immediately after collection, the oral rinse samples were cooled on ice. Subsequently, each sample was vortexed for 10 s, divided into aliquots of 1.0 mL in 1.5 mL screw cap microtubes, and stored at -80 °C.

Oral rinse sample preparation and proteomic analysis

Oral rinse samples were transported on dry ice to the Arcadia Facility, University Medical Center Utrecht (The Netherlands). Samples were rapidly thawed in a 37 °C water bath, vortexed, and centrifuged for 3 min at 12000×g at 4 °C. Subsequently, 90 µL of each sample was transferred to a 96-well PCR plate and mixed with 10 µL of 10× protease inhibitor (Complete Mini, EDTA-free; Roche) solution in PBS. The mixture was vortexed and centrifuged for 1 min at 400×g. For the Proximity Extension Assay (PEA), 1 µL of the prepared sample was used in the incubation step.

Two panels were used for proteomic profiling: Olink® Target 96 Inflammatory and Target 96 Immuno-oncology panels (Olink® Proteomics AB, Uppsala, Sweden). Both panels employ PEA technology, and a detailed description of this method can be found in previous studies^16,17,61. Olink® provides a classification of measured biomarker proteins based on specific biological processes or disease areas.

In our study, each panel contained 92 targeted proteins; therefore, we had a total of 184 proteins. A complete list of these is provided in Supplementary Table S4. Measured protein concentrations were obtained in NPX, which is a relative quantification unit reported on a log₂ scale. An increase of one NPX corresponds to a doubling of the protein concentration. The LOD for each measured protein, which represents the minimum detectable value, was calculated separately for each Olink® assay and sample plate based on the negative controls included on each plate plus three standard deviations.

Standard statistical analysis

As no formal power calculation was performed, the results should be regarded as preliminary and exploratory. Descriptive statistics (means, standard deviations, ranges, and frequency distributions) were calculated to summarize the demographic and clinical data. Group comparisons between the two categories were conducted using independent samples t-tests for continuous variables and χ² tests for categorical variables. For comparisons among the three groups (non-periodontitis controls, localized periodontitis, and generalized periodontitis), ANOVA and χ² tests (linear by linear association) were used. When assumptions for parametric tests were not met, non-parametric tests (Mann–Whitney U test and Kruskal–Wallis test) were performed. The significance threshold was set at p < 0.05. The standard statistical analysis was carried out using SPSS 25.9.6.0.0 (IBM SPSS, Chicago, IL, USA).

Machine learning

Machine learning was used to identify a robust proteomic signature that distinguished periodontitis status from controls. The initial dataset was characterized by a small sample size relative to the large number of proteins, which may have led to overfitting. Consequently, this could lead to the identification of an unstable proteomic signature that is unable to accurately identify new patients. To address this challenge, we developed a machine learning approach adapted from a previously published approach⁶². The approach consisted of (1) data preprocessing, (2) model training and evaluation using repeated nested CV, (3) final model selection, and (4) model interpretation. For detailed steps, see the flowchart in Fig. 3 and the pseudocode in Supplementary Methods in the Supplementary Materials. The analysis was implemented in Python 3.9.6 using scikit-learn⁶³, XGBoost⁶⁴, SHAP, Boruta-py⁶⁵, NumPy, Pandas, Matplotlib, and Seaborn libraries.

Fig. 3.

Flowchart of the machine learning approach. The approach was designed to identify a robust proteomic signature for periodontitis, generalized periodontitis and localized periodontitis. The pipeline consisted of four steps: (1) data preprocessing, (2) model training and evaluation via repeated nested cross-validation, (3) final model selection, and (4) model interpretation. Each step is visualized below along with its inputs, algorithmic steps, and outputs.

Data preprocessing

The protein data from the inflammatory and immuno-oncology panels consisted of unique proteins present in only one panel and common proteins present in both panels. No filtering based on the percentage of measurements below the LOD was applied. For common proteins, a single value per protein was required to avoid redundancy in highly correlated duplicate measurements. Therefore, we retained each common protein from the panel with the lower percentage of measurements below the LOD. The selection was performed across all samples irrespective of the group label. The resulting set of proteins consisted of all unique proteins and deduplicated common proteins and was used for machine learning analysis.

Model training and evaluation

We evaluated three machine learning algorithms: LR⁶⁶, RF⁶⁷, and XGBoost⁶⁴. LR served as a baseline model, while RF and XGBoost were selected to assess whether tree-based models could capture nonlinear patterns possibly missed by LR^68,69. Feature scaling using MinMax scaling (0–1 range) was required only in LR, while RF and XGBoost, as tree-based algorithms, did not require scaling. Class imbalance was addressed using class weighting as a hyperparameter input.

To obtain reliable performance estimates and a robust protein signature, we implemented repeated nested CV for model training and evaluation⁷⁰. Since a single nested CV may produce unstable feature selection in small samples⁷¹, we employed a repeated nested CV with 20 repetitions. The nested approach was specifically chosen to prevent data leakage and consisted of two loops: an inner loop for model training and hyperparameter tuning, and an outer loop for feature selection and unbiased performance evaluation of the trained model. Within each outer fold, the feature selection using the Boruta algorithm⁶⁵ identified the most discriminatory proteins based on the training portion of that fold. Next, the identified proteins were passed to the inner fold to train the model with the three machine learning algorithms and tune its hyperparameters using cross-validation on the inner training folds only. Finally, the trained model was evaluated using the held-out outer-fold samples, which were never seen by either the feature selection, training, or hyperparameter tuning steps. Repeated nested CV was specifically selected as a simulation-based validation method to evaluate the stability of the model performance and protein selection across a large number of data partitions (100 outer folds in total for the classification of periodontitis versus controls).

The number of folds varied according to the sample size: for periodontitis versus controls, 4-fold outer and 4-fold inner CV were used, and for the subgroup analyses of generalized periodontitis versus controls, localized periodontitis versus controls, and generalized periodontitis versus localized periodontitis, 4-fold outer and 3-fold inner CV were used. All splits were stratified based on the target variable (periodontitis, controls, generalized periodontitis, and localized periodontitis) to address the class imbalance.

Model performance was assessed using the ROC-AUC as the primary performance metric, with sensitivity (recall), specificity, balanced accuracy (average value of specificity and sensitivity), and precision as secondary performance metrics. After completing all repetitions, the performance was reported as mean ± standard deviation with 95% bootstrap confidence intervals (1000 iterations) across all iterations. Binary classification predictions were obtained using a default probability threshold of 0.5. To assess the calibration of the predicted probabilities, the Brier score was used as the primary calibration metric, with Log Loss, ECE, and MCE as secondary calibration metrics. Reference values for a no-skill (random) model were calculated for the Brier score and Log Loss in each classification.

Final model selection

The final model was selected based on the results of the repeated nested CV. The aggregated performance metrics were used to select the machine learning algorithm and optimal hyperparameters. The sets of proteins selected by the Boruta algorithm were used to select the proteins that were consistently selected across multiple data splits.

First, the optimal hyperparameters for each machine learning model were selected based on the most frequent hyperparameter combinations across all outer folds. Second, among the LR, RF, and XGBoost machine learning models, the final model was selected based on the highest mean ROC-AUC performance across all repeated nested CV iterations. Finally, to identify the most consistent proteins, we adjusted the consensus features CV⁷², where only features that appeared consistently in all CV folds were selected. If no proteins appeared in any of the iterations, we used a threshold to determine whether a protein was retained. The threshold calculation was based on the stability selection framework⁷³. Features with selection frequencies across the repeated nested CV folds equal to or greater than the threshold were retained.

Model interpretation

To interpret the predictive power of the proteins in the final model, we computed the SHAP values⁷⁴. SHAP was chosen over model-specific feature importance methods because it provides both the magnitude and directionality of each protein’s contribution and is applicable across different model types, allowing consistent interpretation regardless of the selected algorithm. SHAP values quantify each protein’s contribution to individual predictions: positive values support the prediction of periodontitis, whereas negative values support the prediction of controls.

SHAP values were calculated for all validation sets across the CV folds. For each feature, we calculated the mean absolute SHAP values to assess overall discrimination power, and the percentage of positive and negative contributions to understand the direction of the contribution of each protein, i.e., toward periodontitis or controls.

Post-hoc analyses

Post-hoc analyses were performed after repeated nested CV identified the candidate proteins; they do not constitute independent confirmatory testing of the identified candidate biomarkers. To explore whether there were significant differences between two groups for the selected proteins, we performed Mann–Whitney U tests. Moreover, we performed Spearman’s rank correlation analysis between the NPX values of each selected protein and the number of sites with probing pocket depth PPD ≥ 6 mm for all periodontitis patients (localized and generalized; indicated in different colors). In both analyses, multiple testing corrections were applied using the Benjamini–Hochberg false discovery rate (FDR) method.

Finally, we explored whether the demographic and other background characteristics listed in Table 1 and the comorbidities listed in Supplementary Table S1, could potentially confound the relationship between the identified proteins and periodontitis status. Therefore, we applied confounder analysis separately for each identified protein in the following groups: (1) periodontitis versus controls, (2) generalized periodontitis versus controls, and (3) localized periodontitis versus controls. First, given the small sample size and the large number of confounding variables, we conducted univariate linear regression analyses. More specifically, each variable listed in Table 1 and the comorbidities listed in Supplementary Table S1 were entered separately as independent variables in a simple model, with the selected protein as the dependent variable and periodontitis status as a fixed factor. Next, a multivariable linear regression model was constructed, including only the potential confounders from the univariate analyses with a p value ≤ 0.1, periodontal status as independent variables, and the protein value as the dependent variable. Proteins with a p value < 0.05 after adjusting for confounders were considered significantly linked to periodontitis status independently of potentially confounding factors; however, given the small sample size, these results should be regarded as preliminary.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: Madeline X.F. Kosho, Bruno G. Loos, Elena Stamatelou; Methodology: Madeline X.F. Kosho, Bruno G. Loos, Elena Stamatelou; Formal analysis: Madeline X.F. Kosho, Bruno G. Loos; Elena Stamatelou; Investigation: Madeline X.F. Kosho, Bruno G. Loos, Elena Stamatelou; Data Curation: Madeline X.F. Kosho, Elena Stamatelou; Writing—original draft preparation: Madeline X.F. Kosho, Elena Stamatelou; Writing—review and editing: Madeline X.F. Kosho, Elena Stamatelou, Bruno G. Loos; Visualization: Madeline X.F. Kosho, Elena Stamatelou; Supervision: Bruno G. Loos; Project administration: Madeline X.F. Kosho, Elena Stamatelou; Funding acquisition: Bruno G. Loos.

Data availability

The datasets generated and/or analysed during the current study are available from the corresponding author upon reasonable request. The available variables are listed in Table 1, Supplementary Tables S1 and S4.

Declarations

Competing interests

This study received funding in part by grants from TKI HealthHolland, Sunstar Suisse S.A., Philips Oral Healthcare, Dutch Society of Periodontology (NVvP), ORANGEForce project within the ORANGEHealth consortium funded by Health-Holland.nl, div. Life Sciences & Health with grant number LSHM21064, and through material support by Labonovum BV. The funders were not involved in the study design, collection, analysis, interpretation of data, the writing of this article, or the decision to submit it for publication. Author B.G.L. reported having received research grants from Philips Oral Healthcare (ACTA registration R/010998.01) and reports having received honoraria from Philips Oral Healthcare for speaking at symposia on the broad subject ‘The link between periodontitis and systemic diseases’. All authors declare no other competing interests.