Systematic verification of saliva-based biosignatures for oral cancer via targeted metabolomics

Abstract

Oral squamous cell carcinoma (OSCC) remains a major global health challenge, and early detection is essential for improving patient outcomes. To identify reliable salivary biomarkers, we performed a targeted metabolomics study using ¹³C₂/¹²C₂-dansylation, integrating candidate metabolites from published OSCC metabolomics datasets and systematically verifying them through LC-MRM-MS. A panel of 26 metabolites was quantified in saliva samples from 299 subjects, including healthy controls (n = 98), oral potentially malignant disorders (OPMD I, n = 50; OPMD II, n = 53), and OSCC patients (n = 98). Among these, cadaverine, N-acetylcadaverine, choline, glycine, and tryptophan were significantly dysregulated across disease groups. Cadaverine, N-acetylcadaverine and choline were consistently elevated in OSCC and showed progressive increases with disease stages, whereas glycine levels declined. A four-metabolite panel (cadaverine, N-acetylcadaverine, choline, glycine) demonstrated strong diagnostic performance in distinguishing OSCC from healthy controls (AUC = 0.91). Correlation analyses further revealed coordinated regulation among polyamine-related metabolites, suggesting a reprogramming of ornithine–polyamine metabolism in OSCC progression. In conclusion, we establish a verified salivary metabolite panel with high discriminatory power for OSCC detection. These findings highlight polyamine-associated metabolic alterations as a hallmark of malignant transformation in the oral cavity and support salivary metabolomics as a clinically accessible tool for non-invasive screening and risk stratification.

1. Introduction

Oral cancer represents a significant global health burden, ranking among the leading causes of cancer-related mortality worldwide [1,2]. The majority of cases arise from the mucosal surfaces of the oral cavity, oropharynx, and larynx, with oral squamous cell carcinoma (OSCC) accounting for 92–95% of all diagnoses [3]. OSCC incidence is particularly high in Taiwan, Melanesia, South-Central Asia, and parts of Europe, where risk factors such as betel nut chewing, tobacco smoking, and alcohol consumption are prevalent [4,5]. Despite advances in surgery, chemoradiotherapy, targeted therapies, and photodynamic therapy, the five-year survival rate for OSCC remains only 40–60% [6–8]. This poor prognosis is largely attributable to late-stage detection, highlighting the urgent need for reliable early diagnostic strategies. Non-invasive biomarkers hold particular promise for improving early diagnosis, prognosis, and treatment monitoring in OSCC.

Altered cellular metabolism is now recognized as a hallmark of cancer. To better define the molecular alterations underlying OSCC and to identify biomarkers with diagnosis and clinical utility, numerous multi-omics studies have been performed using blood, tissue, and saliva specimens (metabolomics related articles are summarized in Table S1 (https://doi.org/10.38212/2224-6614.3586)). Our group previously applied differential ¹³C₂-/¹²C₂-dansyl labeling to clinical OSCC tissue samples, leading to the identification of a three-metabolite biomarker panel (putrescine, glycyl-leucine, and phenylalanine) that discriminated OSCC from adjacent normal tissues. By integrating transcriptomic and metabolomic data, we further demonstrated significant disruptions in the polyamine pathway in OSCC tissues [9]. These findings suggest that tumor-associated metabolic perturbations may be reflected in proximal biofluid such as saliva, making it an attractive medium for oral cancer noninvasive screening.

Saliva offers unique advantages as a diagnostic biofluid: it can be collected non-invasively, reflects both systemic and oral health, and is in direct contact with OSCC lesions. More than 1200 metabolites have been catalogued in the Human Metabolome Database (HMDB) for saliva [10], and salivary metabolomics has been explored in multiple cancers, including oral [11–14], breast [13,15,16], lung [17,18], pancreatic [13,19,20], and colorectal cancers [21]. Several salivary biomarker panels have shown promising diagnostic performance [11,12,22]; however, limited overlap exists across studies, likely reflecting heterogeneity in patient cohorts, sample collection protocols, and analytical platforms [23,24]. Standardized methodologies and robust quantification strategies are therefore needed to validate candidate biomarkers across independent studies.

Technical innovations in metabolomics have greatly enhanced the sensitivity and reproducibility of salivary biomarker discovery. Liquid chromatography-mass spectrometry (LC-MS) remains the preferred platform for high-throughput metabolite profiling due to its versatility and robustness [25]. Chemical isotope labeling (CIL), such as dansylation, improves resolution of chromatographic separation and ionization efficiency, thereby enabling accurate quantification of low-abundance metabolites [26,27]. The incorporation of stable isotope-labeled internal standards further reduces variability across runs and platforms, ensuring metabolite quantification accurately and precisely.

Among the metabolic pathways altered in OSCC, polyamine metabolism has received particular attention. Polyamines such as putrescine, cadaverine, and their acetylated derivatives regulate fundamental processes including cell proliferation, apoptosis evasion, and invasion, and their dysregulation has been reported in multiple cancers [15,16,18–21,28–31]. Although cadaverine is traditionally attributed to bacterial lysine decarboxylase activity, it can also be synthesized by mammalian ornithine decarboxylase [32]. Our prior integrative omics analyses confirmed marked disruptions in polyamine pathways in OSCC tissues [9], reinforcing its central role in oral tumorigenesis. Importantly, tumor-associated metabolites may diffuse into saliva, positioning them as accessible biomarkers for non-invasive detection.

In this study, we developed a multiple-reaction monitoring (MRM)–based quantitative method to profile 26 salivary metabolites previously implicated in OSCC and oral potentially malignant disorders (OPMDs). By systematically comparing metabolite signatures across healthy controls, OPMD, and OSCC patients, we sought to (i) identify robust biomarkers that discriminate OSCC from both healthy and premalignant states, (ii) assess their utility for early-stage OSCC detection, and (iii) gain mechanistic insights into OSCC-associated metabolic alterations through correlation analyses of salivary metabolite concentrations. This work highlights the translational potential of salivary metabolite biomarkers for noninvasive OSCC detection and risk stratification.

2. Materials and methods

2.1. Subject enrollment and clinical sample collection

This study was conducted in accordance with the principles of the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of the Chang Gung Medical Foundation, Taoyuan, Taiwan (IRB Nos. 201802130B0 and 102-5685). Written informed consent was obtained from all participants prior to sample collection. Saliva samples were collected from 98 healthy volunteers, 50 patients with OPMD I, 53 patients with OPMD II, and 98 patients with OSCC at Kaohsiung Chang Gung Memorial Hospital. Patient demographics and clinical characteristics are summarized in Table S2 (https://doi.org/10.38212/2224-6614.3586).

To minimize confounding factors, donors were instructed to refrain from eating, drinking, or smoking for at least 1 h prior to saliva collection. Approximately 5 mL of unstimulated whole saliva was collected into 50 mL centrifuge tubes. Samples were centrifuged at 3000 × g for 15 min at 4 °C, and the clarified supernatants were pooled. Metabolites were extracted by adding five volumes of methanol and incubating at −20 °C overnight. Proteins were removed by centrifugation at 10,000 rpm for 30 min at 4 °C, and the resulting supernatants were aliquoted, dried, and stored at −80 °C until analysis.

2.2. Reagents

Dansyl chloride, sodium bicarbonate (NaHCO₃), sodium carbonate (Na₂CO₃), sodium hydroxide (NaOH), formic acid (FA), metabolite standards, and the amino acid standards (AAS 18) were purchased from Sigma–Aldrich (St. Louis, MO, USA). Acetonitrile and MS-grade water were also obtained from Sigma–Aldrich. ¹³C₂-dansyl chloride was purchased from The Metabolomics Innovation Centre (TMIC; Alberta, Canada).

2.3. Dansylation labeling of amine/phenol metabolites

Metabolites from individual samples were labeled with ¹²C₂-dansyl chloride (“light” labeling) [26]. Briefly, dried extracts from 100 μL saliva were resuspended in 25 μL H₂O, followed by sequential addition of 12.5 μL sodium carbonate/sodium bicarbonate buffer (0.5 mol/L, pH 9.5), 12.5 μL acetonitrile, and 25 μL of freshly prepared ¹²C₂-dansyl chloride solution (18 mg/mL in acetonitrile). Reactions were incubated at 40 °C with shaking for 45 min, then quenched with 5 μL of 250 mM NaOH. The pH was adjusted to 3–4 with formic acid. Twenty-six metabolite standards (Table 1) were labeled with ¹³C₂-dansyl chloride (“heavy” labeling) under the same conditions. Standards were dissolved in H₂O at 2 mM prior to labeling.

Table 1.

Salivary concentrations of 26 candidate metabolite biomarkers across four subject groups.

No.	Metabolite	Heathy control (n = 98)		OPMD I (n = 50)		OPMD II (n = 53)		OSCC (n = 98)
		Concentration (uM)	>LOD	Concentration (uM)	>LOD	Concentration (uM)	>LOD	Concentration (uM)	>LOD
1	4-Aminobutyrate (GABA)	3.37 ± 3.38	97	5.61 ± 9.48	49	4.37 ± 5.28	51	4.90 ± 6.25	96
2	5-Aminovaleric acid	441.19 ± 443.91	96	400.53 ± 380.23	45	394.14 ± 462.10	45	478.79 ± 551.66	86
3	Alanine	35.86 ± 24.63	98	39.86 ± 29.09	50	38.07 ± 37.04	53	56.74 ± 80.26	98
4	Arginine	14.97 ± 12.93	90	19.59 ± 17.36	49	17.17 ± 13.13	50	16.49 ± 16.67	92
5	Aspartic acid	10.32 ± 7.45	97	13.49 ± 16.69	50	9.54 ± 8.29	53	11.20 ± 10.19	95
6	Cadaverine	74.04 ± 97.85	94	83.09 ± 106.53	49	98.12 ± 145.28	53	293.49 ± 578.89	98
7	Choline	1.14 ± 1.44	59	1.20 ± 1.65	30	1.31 ± 1.56	33	2.83 ± 3.02	75
8	Desaminotyrosine	11.09 ± 25.56	64	6.86 ± 8.65	29	6.53 ± 8.99	33	26.72 ± 59.12	70
9	Ethanolamine	8.71 ± 15.13	38	14.57 ± 14.51	30	12.86 ± 19.91	24	13.59 ± 18.95	47
10	Glutamic acid	20.11 ± 22.42	93	30.38 ± 35.07	50	18.53 ± 28.44	50	28.14 ± 55.65	92
11	Glutamine	17.84 ± 21.62	94	21.64 ± 30.74	49	19.06 ± 26.70	48	15.17 ± 21.38	87
12	Glycine	166.94 ± 127.40	98	155.11 ± 144.66	50	149.55 ± 151.50	53	110.76 ± 121.34	98
13	Isoleucine	10.56 ± 15.52	81	9.09 ± 16.88	39	11.04 ± 24.62	44	10.96 ± 16.39	84
14	Leucine	23.60 ± 34.15	77	20.11 ± 35.09	39	26.14 ± 53.39	44	28.15 ± 40.46	84
15	Lysine	75.54 ± 63.37	98	67.50 ± 67.39	50	84.84 ± 120.25	53	60.55 ± 71.99	98
16	N-Acetylcadaverine	1.07 ± 1.68	66	0.82 ± 1.10	33	1.09 ± 1.24	37	2.41 ± 3.25	77
17	Ornithine	31.61 ± 34.09	72	29.39 ± 42.22	36	27.75 ± 37.69	33	23.65 ± 41.87	55
18	Phenylalanine	25.11 ± 25.54	91	26.99 ± 31.12	48	25.78 ± 28.86	47	24.76 ± 27.78	89
19	Proline	3.83 ± 10.97	13	5.37 ± 11.93	10	3.71 ± 10.27	8	4.02 ± 11.63	13
20	Putrescine	668.54 ± 689.04	97	722.56 ± 687.46	50	659.56 ± 768.57	53	822.53 ± 981.82	98
21	Serine	8.86 ± 6.03	63	9.65 ± 7.91	29	11.39 ± 11.56	33	9.26 ± 7.71	60
22	Taurine	107.58 ± 69.70	98	107.99 ± 71.10	50	124.67 ± 106.92	53	104.35 ± 84.74	98
23	Threonine	3.91 ± 4.16	76	4.69 ± 4.48	44	4.97 ± 7.61	40	5.51 ± 7.01	78
24	Tryptophan	0.94 ± 2.48	8	1.27 ± 3.90	5	0.98 ± 4.17	2	1.54 ± 3.73	14
25	Tyrosine	60.38 ± 47.72	98	64.01 ± 53.83	50	62.08 ± 51.65	53	50.74 ± 47.32	97
26	Valine	26.80 ± 44.95	83	19.77 ± 39.13	45	25.64 ± 63.63	40	32.29 ± 45.89	83

2.4. LC/MRM-MS analysis

Each saliva sample was analyzed in triplicate by LC-MRM-MS. Samples were spiked with a cocktail of 26 ³C₂-dansyl standards (Table S3 (https://doi.org/10.38212/2224-6614.3586)). Labeled samples were diluted 60-fold, and 10 μL of sample was mixed with 10 μL of standards at concentrations ranging from 0.03 to 10 pmol/μL according to their endogenous amounts (Table S3 (https://doi.org/10.38212/2224-6614.3586)). For each LC-MS run, injection of 2 μL mixture was analyzed using an ABSciex 6500⁺ QTRAP with an electrospray ionization source controlled by the Analyst 1.5.1 software (AB Sciex, Singapore). Separation was achieved using a Waters ACQUITY BEH C18 column (1 × 100 mm, 1.7 μm, 130 Å) at 60 μL/min with a binary gradient of solvent A (0.1% FA in H₂O) and solvent B (0.1% FA in acetonitrile): 0 min, 5% B; 2 min, 15% B; 3 min, 35% B; 13 min, 70% B; 25 min, 99% B; 28 min, 99% B; 30.1 min, 5% B; 32.5 min, 5% B. MS acquisition was performed in positive ion mode with the following parameters: ion spray voltage, 4500 V; curtain gas, 30 psi (UHP nitrogen); heater temperature, 250 °C; operating pressure, 2.7 × 10⁻⁵ Torr. Q1 and Q3, unit resolution (0.6–0.8 Da full width at half height). Three MRM ion pairs were monitored per metabolite, optimized for declustering potential (DP), entrance potential (EP), collision energy (CE), and collision cell exit potential (CXP). Scheduled MRM was used with a 1 s cycle time and 6 min detection window. The transitions and parameters of the 26 metabolite targets were listed in Table S3 (https://doi.org/10.38212/2224-6614.3586).

2.5. Data processing

MRM data were processed in MultiQuant (v2.1; AB Sciex) using the MQ4 algorithm for peak integration. Calibration curves of each metabolite were generated as previously described [33,34]. Briefly, varying amounts of ¹³C₂-dansylated standard (10 pmol/μL with serial two-fold dilutions down to 0.039 pmol/μL) were mixed with a fixed amount of ¹²C₂-dansylated pooled saliva to generate a 10-point calibration curve. For individual quantification, a fixed amount of the heavy standard cocktail was spiked into each clinical sample. The composition of this cocktail was adjusted to match the endogenous metabolite concentrations (Table S3 (https://doi.org/10.38212/2224-6614.3586)), ensuring accuracy quantitation [35]. Metabolite concentrations were determined based on peak area ratios relative to calibration curves. Three independent technical replicates were performed for each sample. Linear regression was performed with 1/x weighting factor to improve accuracy across a wide dynamic range. For each metabolite, one unique MRM transition was used for quantification, while two additional transitions (might include dansyl-related fragments) were used for retention time confirmation and interference detection. All integrated peaks were manually inspected to ensure accurate peak detection and integration. The limit of detection (LOD) was defined as the lowest concentration at which the heavy quantifier was consistently detected in all replicates, with accuracy within 85–115%. The limit of quantitation (LOQ) was defined as the lowest concentration that could be measured reliably (signal-to-noise ratio > 30, error ≤ 20%, CV < 20%) and with an intensity at least three times greater than the blank [33,35]. LOD and LOQ values for each metabolite are listed in Table S4 (https://doi.org/10.38212/2224-6614.3586). For downstream statistical analyses, undetected values were imputed as half of the LOD [35]. The final metabolite concentrations in saliva were reported in μM.

2.6. Statistical analysis

To compare differences in variable distributions across groups (healthy, OPMD I, OPMD II, OSCC, and OSCC stages), we applied the Kruskal–Wallis test followed by Dunn's multiple comparison test. The Kruskal–Wallis test is a nonparametric alternative to one-way analysis of variance (ANOVA). Dunn's test provides nonparametric pairwise multiple comparisons when the Kruskal–Wallis test indicates a significant difference. For pairwise group comparisons shown in tables, we used the Mann–Whitney U test. Diagnostic performance of individual and combined biomarkers was evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC) values [36]. Differences between control and test samples were statistically analyzed using SPSS software (version 12.0.1C) and presented using GraphPad Prism (version 7.00). Correlation and scatter plots were analyzed by MetaboAnalyst 6.0.

2.7. Random forest model construction with selected biomarker panels

Random forest models were developed using the scikit-learn Python library to identify discriminative metabolite features and classify OSCC versus healthy controls. Data were split into training (90%, n = 176) and testing (10%, n = 20) sets with balanced class representation. Model performance was assessed using ROC and AUC. To interpret the model, we applied Shapley additive explanations (SHAP) to quantify each feature's contribution to the predictions. This analysis highlighted the relative importance of individual metabolites in disease classification, providing interpretable insights into the model's decision process.

3. Results

3.1. Systematic literature survey and tissue metabolome integration for OSCC biomarker selection and LC-MRM/MS method development

To identify potential salivary metabolite biomarkers capable of distinguishing healthy controls, OPMD (Oral Potentially Malignant Disorder), and OSCC patients—including those with early-stage OSCC—we integrated our tissue metabolomics data [9] with findings from the literature (Table S1 (https://doi.org/10.38212/2224-6614.3586)). A systematic survey of studies published between 2006 and 2025 yielded 53 highly relevant papers, forming a reference database of OSCC-associated metabolites. Among these, 32 studies analyzed saliva, 15 examined tissue, 6 used plasma, 6 serum, and 1 urine, employing a range of analytical platforms. Earlier studies primarily utilized NMR, CE-MS, or GC-MS, whereas more recent reports increasingly applied LC-MS (including reversed-phase and HILIC separations), which has become the dominant approach. Integration across studies identified 1259 metabolites, including 436 lipid species. The remaining 823 metabolites were categorized by functional groups amenable to chemical derivatization: 336 amine/phenol, 387 carboxyl, 364 hydroxyl, and 392 ketone/aldehyde metabolites, with several overlapping groups. As our dansylation approach targets amine/phenol groups, we selected the 30 most frequently reported candidates (Table S1 (https://doi.org/10.38212/2224-6614.3586)). Four metabolites were excluded due to poor assay performance (oxidation interference: methionine, cysteine; multi-tag labeling: spermidine; low labeling efficiency: hypoxanthine). Ultimately, 26 metabolites were retained for downstream LC-MRM/MS analysis (Fig. 1). Fig. S1 (https://doi.org/10.38212/2224-6614.3586) presents the calibration curves of the 26 metabolites for salivary quantification, as well as the linearity and overall performance of the quantification method.

Fig. 1.

Workflow for the selection of candidate salivary metabolites as potentially OSCC biomarkers and their subsequent verification by LC-MRMMS.

A total of 299 saliva samples—comprising 98 healthy controls, 50 OPMD I, 53 OPMD II, and 98 OSCC patients (Table S2A (https://doi.org/10.38212/2224-6614.3586))—were analyzed to evaluate the discriminatory performance of the 26 candidate biomarkers across disease groups and OSCC stages (Table S2B (https://doi.org/10.38212/2224-6614.3586)), using ROC curves, sensitivity, and specificity metrics.

3.2. Multiplexed quantification of salivary metabolites across clinical groups for OSCC biomarker prioritization

All 26 candidate metabolites were quantified in the 299 saliva specimens (Table 1). Proline and tryptophan were detected in only a minority of samples (44/299 and 14/299, respectively; Table S4 (https://doi.org/10.38212/2224-6614.3586)). Nevertheless, tryptophan concentrations were significantly elevated in OSCC compared with controls, consistent with prior CE-MS reports [12,13,28]. Group comparisons (Table 2) revealed that cadaverine, choline, desaminotyrosine, N-acetylcadaverine, and tryptophan were significantly elevated in OSCC relative to healthy controls, whereas glutamine, glycine, lysine, and ornithine were significantly reduced. Notably, cadaverine, N-acetylcadaverine, choline, tryptophan, and glycine exhibited highly significant differences (p < 0.0001). Cadaverine showed the largest fold increase (3.96-fold), with its downstream metabolite N-acetylcadaverine also elevated (2.26-fold). These findings align with prior reports linking cadaverine to OSCC and periodontal disease [13,29], implicating polyamine pathway dysregulation and microbial contributions. Importantly, this is the first study to demonstrate significant upregulation of N-acetylcadaverine in OSCC saliva. It was also reported recently to be associated with an increased risk of harboring high-grade of intraductal papillary mucinous neoplasms pancreatic ductal adenocarcinoma [37]. Desaminotyrosine, also of microbial origin [38], was elevated in OSCC consistent with earlier CE-MS findings [12,28,39]. Glycine showed the most significant decrease, consistent with Kamarajan et al. [40], though other studies reported either increases [14] or no significant changes [13,41,42]. Ornithine, a key polyamine precursor, was downregulated in line with previous findings [43]. Lysine, another upstream polyamine metabolite, also decreased, further suggesting suppression of this pathway in OSCC saliva.

Table 2.

Discriminative performance of the 26 salivary metabolites for between-group comparisons: healthy controls vs. OSCC, OPMD I vs. OSCC, and OPMD II vs. OSCC.

No.	Metabolite	Healthy control vs. OSCC			OPMD I vs. OSCC			OPMD II vs. OSCC
		Fold-change	p-value	AUC	Fold-change	p-value	AUC	Fold-change	p-value	AUC
1	4-Aminobutyrate (GABA)	1.45	0.175	0.556	0.87	0.587	0.528	1.12	0.745	0.516
2	5-Aminovaleric acid	1.09	0.614	0.520	1.20	0.994	0.502	1.21	0.490	0.534
3	Alanine	1.58	0.461	0.530	1.42	0.990	0.501	1.49	0.302	0.551
4	Arginine	1.10	0.964	0.502	0.84	0.104	0.582	0.96	0.347	0.546
5	Aspartic acid	1.08	0.889	0.506	0.83	0.277	0.555	1.17	0.324	0.549
6	Cadaverine	3.96	< 0.001	0.698	3.53	0.001	0.672	2.99	0.006	0.637
7	Choline	2.48	< 0.001	0.685	2.37	< 0.001	0.688	2.17	0.001	0.667
8	Desaminotyrosine	2.41	0.025	0.594	3.89	0.007	0.632	4.09	0.004	0.638
9	Ethanolamine	1.56	0.110	0.566	0.93	0.135	0.575	1.06	0.748	0.516
10	Glutamic acid	1.40	0.894	0.506	0.93	0.022	0.615	1.52	0.373	0.544
11	Glutamine	0.85	0.048	0.582	0.70	0.076	0.589	0.80	0.116	0.578
12	Glycine	0.66	< 0.001	0.672	0.71	0.021	0.616	0.74	0.031	0.606
13	Isoleucine	1.04	0.828	0.503	1.21	0.570	0.541	0.99	0.960	0.510
14	Leucine	1.19	0.310	0.532	1.40	0.156	0.560	1.08	0.577	0.518
15	Lysine	0.80	0.008	0.609	0.90	0.227	0.561	0.71	0.111	0.579
16	N-Acetylcadaverine	2.26	< 0.001	0.650	2.95	0.001	0.669	2.22	0.026	0.609
17	Ornithine	0.75	0.002	0.625	0.80	0.082	0.581	0.85	0.259	0.554
18	Phenylalanine	0.99	0.600	0.518	0.92	0.501	0.532	0.96	0.395	0.537
19	Proline	1.05	0.876	0.504	0.75	0.186	0.543	1.08	0.798	0.508
20	Putrescine	1.23	0.669	0.518	1.14	0.830	0.511	1.25	0.391	0.542
21	Serine	1.04	0.868	0.507	0.96	0.775	0.514	0.81	0.402	0.541
22	Taurine	0.97	0.265	0.546	0.97	0.473	0.536	0.84	0.123	0.576
23	Threonine	1.41	0.099	0.570	1.17	0.940	0.504	1.11	0.285	0.553
24	Tryptophan	1.64	< 0.001	0.541	1.21	< 0.001	0.536	1.58	< 0.001	0.559
25	Tyrosine	0.84	0.170	0.557	0.79	0.089	0.586	0.82	0.087	0.585
26	Valine	1.20	0.198	0.553	1.63	0.070	0.591	1.26	0.144	0.572

Fold change of metabolite levels in OSCC group to healthy control, OPMD I, and OPMD II groups. p-value were calculated by Mann–Whitney test (two-tailed) between two groups, and bold present for significant between two groups (p-value< 0.05).

Saliva metabolomics has been increasingly applied for the early diagnosis and monitoring of OSCC [44]. When comparing OPMD I/II with OSCC, cadaverine, choline, desaminotyrosine, N-acetylcadaverine and tryptophan were elevated, while glycine was decreased (Fig. 2 and Table 2). A similar increase in cadaverine was reported by Song et al. [29], and Kitabatake et al. [45] also observed slight tryptophan upregulation. Notably, choline, desaminotyrosine, N-acetylcadaverine and glycine were identified here for the first time as differentially regulated between OSCC and OPMD, suggesting their potential as discriminative biomarkers. Dunn's multiple comparisons test across all four groups (Fig. 2 and Fig. S2 (https://doi.org/10.38212/2224-6614.3586)) revealed minimal differences between healthy controls and OPMD I/II. Only ethanolamine was slightly increased in OPMD I, and glutamic acid decreased in OPMD II compared to OPMD I (p < 0.05). Despite prior reports describing differential regulation between OSCC and OPMD, including increased levels of putrescine [29,46], threonine [11], arginine [47], and leucine [11], as well as decreased levels of proline [11,48,49], tryptophan [48], valine [11], phenylalanine [11], serine [29], isoleucine [11], and ornithine [50]. These discrepancies highlight the variability across studies and emphasize the need for standardized protocols, particularly the use of stable isotope–labeled internal standards for absolute quantification and detailed reporting of collection methods.

Fig. 2.

LC-MRM/MS analysis of salivary metabolites across groups (healthy controls, OPMD I, OPMD II, and OSCC). Differentially regulation was observed in (A) cadaverine, (B) choline, (C) desaminotyrosine, (D) ethanolamine, (E) glutamic acid, (F) glycine, (G) lysine, (H) N-acetylcadaverine, (I) ornithine, and (J) tryptophan. Statistical significance was determined by Dunn's multiple comparisons test (*p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001).

3.3. Metabolite biomarker candidates to distinguish different stages of OSCC in individual saliva specimens

In the OSCC cohort (n = 98), 26 metabolite concentration levels were analyzed by stage: healthy controls (n = 98), early-stage OSCC (n = 51; stage I = 24, stage II = 27), and late-stage OSCC (n = 47; stage III = 17, stage IV = 30). Cadaverine, choline, glycine, and N-acetylcadaverine significantly distinguished early-stage OSCC from healthy controls (Table 3), and these metabolites also remained significant in late-stage disease. Additional alterations emerged in late-stage OSCC, including increases in alanine, desaminotyrosine, lysine, and threonine, alongside decreases in glutamine, ornithine, and tyrosine (Table 3). Concentration distributions of significant metabolites are shown in Fig. 3, with non-significant metabolites in Fig. S3 (https://doi.org/10.38212/2224-6614.3586). ROC analyses of cadaverine, glycine, choline, and N-acetylcadaverine in saliva from healthy controls (n = 98) and OSCC patients (n = 98) are shown in Fig. S4 (https://doi.org/10.38212/2224-6614.3586). The diagnostic performance of salivary metabolite biomarkers was evaluated using cutoff values determined by the Youden index, as summarized in Table S5 (https://doi.org/10.38212/2224-6614.3586). When applying cutoff values derived from the OSCC vs. healthy control comparison, individual metabolites detected only 45% of OSCC cases (Fig. 4A and Table S5A (https://doi.org/10.38212/2224-6614.3586)). In contrast, using cutoff values from early-stage OSCC vs. healthy controls improved the detection rate to 70%, albeit with a higher false-positive rate (Fig. 4B and Table S5B (https://doi.org/10.38212/2224-6614.3586)). These findings prompted further investigation into biomarker panels and machine learning–based models to enhance OSCC detection.

Table 3.

Discriminative performance of the 26 salivary metabolites for stage-specific comparisons: healthy controls vs. Stage I + II OSCC, and healthy controls vs. Stage III + IV OSCC.

No.	Metabolite	Healthy control vs. Stage I + II			Healthy control vs Stage III + IV
		Fold-change	p-value	AUC	Fold-change	p-value	AUC
1	4-Aminobutyrate (GABA)	1.22	0.078	0.588	1.70	0.682	0.521
2	5-Aminovaleric acid	1.16	0.935	0.505	1.00	0.352	0.547
3	Alanine	1.03	0.251	0.557	2.18	0.014	0.626
4	Arginine	1.18	0.566	0.529	1.01	0.596	0.527
5	Aspartic acid	0.97	0.164	0.570	1.21	0.217	0.564
6	Cadaverine	2.35	< 0.001	0.691	5.72	< 0.001	0.705
7	Choline	1.88	0.044	0.605	3.13	< 0.001	0.776
8	Desaminotyrosine	1.23	0.094	0.585	3.69	0.047	0.604
9	Ethanolamine	1.74	0.483	0.535	1.36	0.053	0.600
10	Glutamic acid	1.61	0.987	0.501	1.17	0.810	0.512
11	Glutamine	1.08	0.471	0.536	0.60	0.010	0.632
12	Glycine	0.76	0.009	0.631	0.55	< 0.001	0.715
13	Isoleucine	1.18	0.739	0.517	0.89	0.809	0.512
14	Leucine	1.35	0.326	0.549	1.02	0.793	0.513
15	Lysine	1.02	0.387	0.543	0.56	< 0.001	0.681
16	N-Acetylcadaverine	1.68	0.008	0.633	2.88	0.001	0.668
17	Ornithine	0.95	0.149	0.570	0.52	< 0.001	0.685
18	Phenylalanine	1.16	0.433	0.539	0.80	0.115	0.581
19	Proline	1.24	0.584	0.517	0.84	0.364	0.527
20	Putrescine	1.25	0.417	0.541	1.21	0.887	0.507
21	Serine	1.08	0.673	0.521	1.01	0.867	0.508
22	Taurine	1.06	0.946	0.503	0.88	0.052	0.600
23	Threonine	1.38	0.542	0.533	1.44	0.034	0.610
24	Tryptophan	1.44	0.239	0.533	1.86	0.087	0.550
25	Tyrosine	0.95	0.761	0.515	0.72	0.048	0.602
26	Valine	1.25	0.283	0.554	1.16	0.304	0.553

Fold change of metabolite levels in Stage I + II and Stage III + IV groups to healthy control group. p-value were calculated by Mann–Whitney test (two-tailed) between two groups, and bold present for significant between two groups (p-value < 0.05).

Fig. 3.

Differential regulation of salivary metabolites across OSCC stages. Shown are (A) alanine, (B) cadaverine, (C) choline, (D) glycine, (E) lysine, (F) ornithine, (G) glutamine and (H) N-acetylcadaverine. Statistical significance was determined by Dunn's multiple comparisons test (*p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001).

Fig. 4.

Diagnostic Performance and model interpretation of the four-metabolite panel for OSCC detection. (A) Diagnostic accuracy of the four-metabolite panel (from top to bottom: cadaverine, N-acetylcadaverine, choline and glycine) in distinguishing OSCC patients from healthy controls. Cutoff values were determined using the Youden index from ROC analysis (OSCC vs healthy controls). Metabolite concentrations above the cutoff are marked in red. (B) Diagnostic performance for early stage detection, with cutoff values from OSCC stage I + II vs healthy controls. (C) ROC curve of the random forest model, and AUC reached 0.91. (D) SHAP summary plot illustrating the contribution of the four metabolites to OSCC prediction. Each point represents a metabolite feature in a single sample. Colors indicate metabolite concentrations (blue: low; red: high). The x-axis shows the SHAP values, reflecting both the magnitude and the direction of each feature's effect on the model output.

Although these metabolite biomarkers have been previously reported in the literature, our study data indicate that not all molecules exhibit satisfactory diagnostic AUC values for oral cancer. Without this study, we would not have been able to determine which of these molecules are more reliable than others. This underscores the importance of developing a targeted metabolomics methodology to systematically verify oral cancer metabolite biomarkers using the same analytical approach on the same batch of saliva specimens.

3.4. Random forest–derived salivary biomarker panel for OSCC discrimination

Classification models were constructed using random forests to identify key features and distinguish OSCC samples. Among the 26 monitored metabolites, the most relevant four metabolites, cadaverine, glycine, choline, and N-acetylcadaverine, were selected based on Gini impurity scores reported through a preliminary model considering all 26 metabolites, consistent with early-stage marker selection. Using fewer than these four metabolites reduced model performance. The final model consisting of the four metabolites was trained on the randomly selected 90% of the samples and evaluated on the remaining 10%, achieving a testing AUC of 0.91 in ROC analysis (Fig. 4C). Next, we applied SHAP analysis to interpret the model. The four selected metabolites showed clear concentration gradients between OSCC and healthy controls, indicating distinct metabolic alterations associated with disease status (Fig. 4D). As shown in Table 2, glycine had a fold change below one, reflecting higher concentrations in healthy controls compared to OSCC samples. This pattern was consistent with the SHAP result, where higher glycine levels decreased the likelihood of OSCC classification.

3.5. Correlation analysis of salivary metabolite concentrations reveals pathway clustering and polyamine dysregulation in OSCC

Correlation analysis across all 299 saliva samples revealed metabolite clustering consistent with shared biochemical pathways (Fig. 5). Cluster 1 (cadaverine, N-acetylcadaverine, desaminotyrosine) likely reflects microbial contributions. Cluster 2 (lysine, glycine, ornithine, proline, putrescine) represents polyamine metabolism. Cluster 3 (serine, threonine, aspartate, glutamate) aligns with glycine/serine/threonine metabolism. Cluster 4 (valine, isoleucine, leucine, phenylalanine, tyrosine, tryptophan) corresponds to branched-chain amino acid and aromatic amino acid metabolism. Polyamine metabolism, well known to be activated during tumorigenesis, emerged as a central node. The strong correlations among lysine, ornithine, and glycine suggest a tightly linked regulatory circuit. Intriguingly, alanine clustered with cadaverine, raising the possibility of microbial involvement in its metabolism. Overall, these findings highlight coordinated alterations in polyamine metabolism and microbiota-derived metabolites in OSCC saliva, underscoring their potential as early diagnostic biomarkers.

Fig. 5.

Correlation analysis of salivary concentrations of 26 metabolites across all 299 subjects.

4. Discussion and conclusion

4.1. Experimental verification of OSCC-associated salivary metabolite biomarker candidates: comparison with literature evidence

Saliva is an easily accessible biofluid that reflects systemic physiology and has emerged as a promising diagnostic medium. Salivary metabolomics has been widely applied to cancer detection, prognosis, and treatment monitoring [51–53], with disease-specific signatures reported in oral [11–14], breast [13,15,16], lung [17,18], pancreatic [13,19,20], bladder [27,54] and colorectal cancers [21,53,55]. In our cohort, cadaverine, N-acetylcadaverine and choline were the most elevated metabolites in OSCC saliva. Cadaverine, classically produced by bacterial lysine decarboxylase, can also be generated by mammalian ornithine decarbosylase (ODC) [32]. We observed a consistent pattern of decreased substrates (lysine, ornithine) with increased products (cadaverine, putrescine), indicating altered polyamine metabolism in OSCC. Polyamines are known to drive proliferation, suppress apoptosis, and promote invasion [56], and their upregulation has been documented in multiple cancers [15,16,18–21,28–31]. Importantly, isotope-labeled quantification of polyamine derivatives in head and neck cancer [31] underscores their robustness as non-invasive biomarkers.

In our previous tissue-based study, putrescine and phenylalanine were identified as significant metabolites associated with OSCC. In the current saliva-based analysis, although both metabolites were included in the targeted assay, neither exhibited statistically significant differences between OSCC patients and healthy controls. As supported by multiple studies, metabolic alterations observed in tumor tissues do not necessarily translate directly to adjacent biofluids. Tissue metabolomics reflects the intrinsic biochemical state of the tumor microenvironment, whereas the detectability of metabolites in saliva requires additional biological processes, including secretion, diffusion, or transport across epithelial or vascular barriers. Furthermore, saliva contains abundant high-molecular-weight and dominant components that may mask or dilute low-abundance metabolites, leading to reduced detectability or diminished fold changes compared with tissue profiles. Therefore, verifying whether tissue-derived biomarkers can be reliably detected in saliva represents a crucial step in translational metabolomics. Our findings underscore that not all tissue biomarker candidates are preserved in salivary profiles, highlighting the importance of evaluating biofluid-specific metabolic signatures rather than assuming direct correspondence.

Choline was also significantly increased in our dataset, consistent with reports in OSCC [12,13,22,28,39,41]. As a strong organic base and essential dietary nutrient, choline plays a central role in membrane turnover. Aberrant choline metabolism is now recognized as a hallmark of oncogenesis [57]. Elevated phosphocholine and glycerophosphocholine levels in OSCC saliva further support dysregulated choline metabolism during tumor progression [58].

By contrast, amino acid changes were less consistent across studies. In our data, glycine, lysine, and glutamine were decreased, whereas prior reports described variable directions depending on cohort and platform (Table S1 (https://doi.org/10.38212/2224-6614.3586)) [13,14,39–42]. Such discrepancies highlight the sensitivity of salivary metabolomes to biological and methodological factors, including stimulation, sex, BMI, periodontal status, and collection protocols [24,59,60]. Notably, polyamine elevation appears specific to cancer rather than periodontal disease [13,61]. To reduce variability, we applied stable isotope-labeled internal standards for absolute quantification and implemented standardized collection protocols with a uniform male cohort. These measures strengthen the conclusion that polyamine and choline pathway metabolites are robustly dysregulated in OSCC saliva. Collectively, our findings support their reliability as non-invasive biomarkers and highlight the advantages of our platform over conventional methods.

4.2. Metabolite biomarkers differentiating OSCC and OPMD

Saliva metabolomics has been increasingly applied to the early diagnosis and monitoring of OSCC [44]. Several studies have also examined OPMD saliva samples to identify premalignant biomarkers, as summarized by Nazar et al. [62]. In our study, most of the 26 selected metabolites in OPMD saliva showed concentrations comparable to healthy controls. Consequently, metabolites significantly altered in OSCC vs healthy controls exhibited the same directional trends when OSCC was compared with OPMD. Specifically, cadaverine, choline, desaminotyrosine, N-acetylcadaverine and tryptophan were elevated, while glycine was decreased. A similar increase in cadaverine was reported by Song et al. [29]. Kitabatake et al. [45] also observed modest tryptophan elevation in OPMD relative to heathy controls, alongside increases in glutamine and leucine and a decrease in valine. Notably, choline, desaminotyrosine, N-acetylcadaverine and glycine were identified here for the first time as differentially regulated between OSCC and OPMD, suggesting their potential as discriminative biomarkers. By contrast, other metabolites previously reported as altered between OSCC and OPMD were not significantly different in our dataset, emphasizing variability across studies and the importance of standardized methodologies. Our results indicates that OPMD I and II share salivary metabolite profiles similar to healthy controls, thereby highlighting their distinct separation from OSCC.

For early-stage OSCC detection, cadaverine, choline, glycine and N-acetylcadaverine demonstrated strong discriminatory performance. These metabolites differentiated OSCC not only from healthy controls and OPMD (I/II) but also distinguished early-stage of OSCC (stage I-II) from controls. A random forest machine-learning model further validated this four-metabolite panel, achieving high predictive accuracy. Increased cadaverine, choline and N-acetylcadaverine, combined with decreased glycine, effectively separated stage I–II OSCC from healthy individuals, underscoring their utility for early diagnosis. Of these, only choline has previously been reported as significantly elevated in stage I–II OSCC [44], whereas cadaverine, N-acetylcadaverine, and glycine are novel findings for early-stage detection. The concurrent elevation of cadaverine and N-acetylcadaverine across both OSCC versus controls and OSCC versus OPMD comparisons highlights dysregulated polyamine metabolism and its potential link to oral microbiota activity. Collectively, these findings indicate that cadaverine, choline, N-acetylcadaverine, and glycine represent promising salivary biomarkers for OSCC, particularly at early clinical stages.

4.3. Salivary microbiota and metabolite profiles in OSCC

Cadaverine, N-acetylcadaverine and desaminotyrosine were upregulated in OSCC saliva, potentially reflecting alterations in the oral microbiota [38]. Accumulating evidences indicates that oral microorganisms contribute not only to local diseases such as dental caries [63] and periodontitis [64] but also to systemic diseases, including cancer [65,66]. Multiple studies have reported significant microbiome alterations in OSCC saliva. Enrichment of Fusobacterium nucleatum, Porphyromonas endodontalis, Prevotella melaninogenica [67], Deferribacterota, Vibrio and Lactococcus [68] has been observed, whereas reductions in Veillonella parvula NCTC11810 [69], Cyanobacteria, Bifidobacterium, and Faecalibacterium [68] have also been documented. Microbiota composition is shaped by host diet, and lifestyle, and immune responses, which collectively influence disease susceptibility. Enhanced levels of Tannerella forsythia, Porphyromonas gingivalis, and Candida albicans have been reported in OSCC patients [70]. Similarly, Mager et al. demonstrated increased abundance of Streptococcus mitis, P. gingivalis, and P. melaninogenica suggesting their potential as diagnostic indicators of OSCC [71].

Functional studies further demonstrate that oral microbiota actively contribute to carinogenesis. Transplantation of saliva from chronically restraint stressed (CRS) mice into germ-free mice induced dysbiotic oral microbiota and increased kynurenine biosynthesis, promoting HNSCC tumorigenesis. CRS-induced dysbiosis and their metabolites also disrupt oral and gut barrier function, facilitating cancer progression [72]. Microorganisms may promote cancer development by producing carcinogenic compounds such as acetaldehyde [73], inducing chronic inflammation, and directly interfering with host cell cycle and signaling pathways [74]. Indeed, elevated Prevotella, Veillonella, Porphyromonas and Capnocytophaga species are consistently detected in OSCC saliva, with Porphyromonas linked to immune suppression and F. nucleatum shown to trigger pro-inflammatory chemokine responses in oral epithelial cells [75]. Collectively, these findings provide a framework for exploiting the oral microbiome to monitor oral cancer initiation, progression, and recurrence [75], and may also inform therapeutic outcomes [76].

5. Conclusion

In summary, our findings indicate that not all previously reported metabolite biomarkers exhibit satisfactory diagnostic AUC values for oral cancer. These results underscore the necessity of establishing a targeted metabolomics framework capable of systematically validating candidate biomarkers under a uniform analytical approach, thereby ensuring their reliability and clinical applicability. This study demonstrates that salivary metabolite alterations in OSCC are characterized by dysregulated polyamine and choline metabolism, exemplified by the accumulation of cadaverine, N-acetyl-cadaverine, and choline. These metabolite signatures not only underscore their pathogenic relevance but also provide a promising basis for liquid biopsy–based early detection and patient stratification. Emerging evidence further suggests that oral microbial dysbiosis contributes to carcinogenesis through multiple mechanisms, including the production of carcinogenic compounds, modulation of polyamine metabolism, induction of chronic inflammation, and suppression of antitumor immunity. The integration of microbiome and metabolome analyses offers new opportunities for clinical translation. Salivary metabolite profiling, combined with microbial community signatures, could serve as a noninvasive diagnostic tool for early OSCC detection and disease monitoring. Moreover, microbiome-targeted interventions—such as probiotics, dietary modulation, or microbiota-directed therapeutics—may represent novel strategies to restore metabolic balance and reprogram the tumor microenvironment. Positioning the salivary microbiome–metabolome axis as both a biomarker source and a therapeutic target has the potential to transform OSCC management, advancing precision oncology through noninvasive and personalized approaches.

Funding Statement

This study was supported by the Ministry of Science and Technology, Taiwan (NSTC-114-2113- M-182-002, 113-2113-M-182-001, 114-2622-M-182- 001, 114-2113-M-007-024 and 113-2314-B-182-026-MY2), and the Chang Gung Memorial Hospital, Taiwan (CORPD1P0061, and BMRPD78). The mass spectrometry instrumentation and data analysis were supported by the Proteomics Core Facility at the Molecular Medicine Research Center, Chang Gung University, Taoyuan, Taiwan