Salivary metabolic profile landscape for gastric cancer screening: a metabolomic approach

Abstract

This study employed non-targeted liquid chromatography-mass spectrometry (LC-MS) to investigate the salivary metabolomic profiles of 177 participants, including 147 patients with gastric cancer (GC) and 30 controls with common gastritis. Following rigorous quality control measures, a total of 333 high-confidence metabolites were identified from an initial pool of 368 positive-mode metabolites and 178 negative-mode metabolites. Principal component analysis (PCA) and linear discriminant analysis (LDA) revealed significant differences between the GC and control groups, with stage I GC clearly distinguishable from more advanced stages. Differential metabolite screening was conducted using partial least squares discriminant analysis (PLS-DA) in conjunction with t-tests and fold-change analysis. After controlling for confounding factors such as age, smoking, and alcohol use, 38 salivary metabolites were identified as potential diagnostic markers. Notably, the univariate and multivariate diagnostic models demonstrated excellent discriminative performance in distinguishing GC patients from controls. PLS-DA validated by permutation testing, along with univariate and multivariate ROC analyses, exhibited excellent classification performance based on the 38 salivary metabolites. Metabolic analysis revealed significant downregulation of purines, pyrimidines, amino acids, and carbohydrates in the saliva of GC patients, while sebacic acid and GABA were found to be upregulated. Tyrosine was identified as the most significantly altered metabolite between early and advanced stages of GC. These findings underscore the substantial impact of gastric cancer on the salivary metabolome and suggest the potential of saliva as a promising tool for mass screening of GC.

Trial registration DGLES, NCT01420588, Registered 19 August 2011, https://clinicaltrials.gov/ct2/show/NCT01420588.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12935-025-04107-z.

Introduction

Gastric cancer (GC) remains a significant global health concern, ranking as the fifth most commonly diagnosed cancer and the third leading cause of cancer-related mortality worldwide. According to Global Cancer Statistics 2022, nearly one million new cases of GC are reported annually, resulting in approximately 659,853 deaths, reflecting the high mortality associated with this disease [1]. Early gastric cancer (EGC), which is confined to the mucosa or submucosa, is often asymptomatic, making it challenging to detect and contributing to a low detection rate. In contrast, advanced gastric cancer (AGC) is more frequently diagnosed due to the presence of more pronounced clinical symptoms, such as weight loss, abdominal pain, and gastrointestinal bleeding. Unfortunately, the majority of cases are identified at advanced stages (e.g., 51.1% at stage III), where prognosis is poor, and the 5-year survival rate declines drastically compared to early-stage detection (e.g., 96% for stage II) [2, 3]. This pattern of late-stage diagnosis highlights the critical need for improved diagnostic tools to enable the early detection of gastric cancer, which could significantly enhance patient outcomes.

Current diagnostic methods for gastric cancer (GC) predominantly rely on endoscopy combined with histopathological examination, which is considered the gold standard for detection. While endoscopy is highly sensitive, it is invasive, costly, and often uncomfortable for patients, limiting its feasibility for widespread use, particularly in routine screening [4]. Tumor biomarkers, such as α-fetoprotein for liver cancer, have demonstrated effectiveness in identifying specific types of malignancies [5]. However, blood-based biomarkers for GC, including carcinoembryonic antigen (CEA) and pepsinogen, often lack the sensitivity and specificity required for reliable detection in massively rapid screening [6].

In recent years, liquid biopsy techniques have emerged as promising non-invasive diagnostic tools, leveraging circulating tumor cells (CTC), circulating tumor DNA (ctDNA), and exosomes to analyze tumor-associated molecules in the bloodstream. These methods offer a less invasive alternative for cancer detection. However, several technical challenges limit their reliability. CTC detection is hindered by inconsistent positivity thresholds and the heterogeneity of tumor cells, complicating accurate identification [7]. While ctDNA provides insights into tumor-specific mutations, it is susceptible to false positives due to the inability to differentiate between tumor-specific and somatic mutations unrelated to cancer [8]. Similarly, exosome detection faces challenges related to protein degradation, low isolation efficiency, and substantial heterogeneity among exosomal populations [9]. Despite their potential, liquid biopsy techniques require further refinement and validation to address these limitations, particularly in the context of early-stage gastric cancer detection.

Metabolomics focuses on the quantitative analysis of metabolites, offering real-time and precise insights into dynamic physiological and pathological processes in the human body [10]. This unique approach enables metabolomics to capture phenotypic shifts that are directly reflective of cellular and systemic changes, making it particularly valuable in cancer research. Metabolic reprogramming is a hallmark of cancer, underpinning tumor development and progression. Tumor cells undergo significant metabolic adaptations to meet increased energy demands, synthesize macromolecules for proliferation, and remodel the extracellular matrix [11]. Key metabolic pathways, such as glycolysis and glutaminolysis, are reprogrammed to sustain self-sufficiency in growth signals and resist apoptosis [12, 13]. As the tumor advances, metabolic alterations further promote angiogenesis, enable tissue invasion [14], and create an immunosuppressive microenvironment that facilitates immune evasion, hallmarks of the middle and late stages of cancer progression [15, 16]. These peripheral circulating metabolic changes, induced by tumors, may play a critical role in advancing cancer diagnostic and therapeutic strategies [17, 18].

Additionally, previous studies have revealed consistent genomic, proteomic, and metabolomic heterogeneity between cancer cells and controls through comprehensive multi-omics analyses of tumor tissues and single cells [19]. The use of metabolic profiling for cancer screening offers several compelling advantages. First, metabolic changes often precede structural abnormalities detectable through conventional imaging, enabling the identification of tumors at earlier stages. Second, metabolic profiling provides a dynamic approach to monitoring tumor behavior, as variations in metabolic signatures can reflect treatment responses or disease progression in real-time [20]. This capability is invaluable for developing personalized treatment strategies tailored to individual tumor biology [21]. Peripheral metabolome only offer potential biomarkers for early detection but also provide dynamic insights into tumor progression and treatment responses, enabling more personalized and effective cancer management.

When evaluating biomarker potential, the practicality and stability of the biofluid itself are critical considerations. Urine collection typically requires obtaining midstream urine to avoid contamination of the urethral orifice [22]. Moreover, the abundance of urinary metabolites is susceptible to factors such as circadian rhythms, mental state, hormonal fluctuations, and fluid intake [23]. Additionally, although plasma metabolism is relatively stable, it involves invasive procedures and complex sample processing. Saliva, which contains a diverse array of metabolites reflective of systemic health, presents a non-invasive and accessible alternative for capturing these metabolic changes. Despite saliva composition is significantly influenced by the metabolism of resident oral flora, desquamated host cells, and immune cells [24], studies have established strong correlations between metabolites in blood and saliva [25], facilitated by transcellular and paracellular transport pathways into the highly vascularized salivary glands. Compared to blood sampling, saliva collection is less invasive, more convenient, and easier to perform, making it particularly suitable for mass screening initiatives or frequent monitoring.

Research into salivary biomarkers has revealed significant differences in the concentrations of specific metabolites across various desease. For example, studies have explored the effects of oral cancer on the salivary proteome [26, 27], the impact of chronic inflammation on salivary cytokines and chemokines [28], and the diagnostic potential of salivary exosomes and other biological metabolites [29, 30]. This innovative approach aligns with the principles of personalized medicine and addresses the urgent need for improved diagnostic tools in the detection and management of gastric cancer. By analyzing salivary metabolites, researchers can identify unique metabolic signatures associated with stomach cancer, potentially improving patient outcomes and reducing the reliance on invasive diagnostic procedures.

The objective of this study was to characterize the salivary metabolic profiles of GC patients in comparison to matched healthy controls using non-targeted LC-MS metabolomics. Additionally, the study aimed to differentiate the metabolic changes associated with early- and late-stage GC and evaluate the influence of HER2 expression on salivary metabolism.

Methods

Subject recruited

Saliva samples for LC-MS analysis were collected from a clinical cohort at the First Affiliated Hospital of Anhui Medical University between January 1, 2017, and January 1, 2018. All patients were confirmed to have gastric cancer through gastroscopic examination or surgical verification of tissue samples. Patients who had undergone targeted therapy or chemotherapy prior to sample collection were excluded. Additionally, all participants were free from other types of cancer, as well as liver or kidney dysfunction, diabetes, thyroid disorders, metabolic disturbances, and recent respiratory or oral diseases.

Ethical approval for the study was obtained from the Medical Ethics Committee of the First Affiliated Hospital of Anhui Medical University. Written informed consent was obtained from all participants before saliva collection and the acquisition of clinical information.

Samples collection

All participants were instructed to refrain from eating, smoking, and brushing their teeth for at least 10 h prior to saliva collection. Additionally, participants were asked to rinse their mouths with clean water three times within the 5 min preceding the collection process. Saliva samples were collected into 5 mL cryopreservation tubes, which were immediately sealed. The samples were stored at −80 °C within 30 min of collection and transported in foam boxes containing dry ice to maintain stability.

Upon analysis, the saliva samples were removed from the − 80 °C freezer and thawed on ice at 4 °C. A 100 µL aliquot of each sample was transferred into an EP tube, followed by the addition of 400 µL of extraction solution (methanol: acetonitrile = 1:1). The mixture was vortexed for 30 s and subjected to ultrasonic treatment for 10 min while incubated in ice water. The samples were then incubated at −20 °C for 1 h to precipitate proteins, followed by centrifugation at 4 °C at 12,000 rpm for 15 min.

The supernatant (430 µL) was carefully transferred to a new EP tube, and the extracted substances were dried using an unheated vacuum concentrator. To reconstitute the dried samples, 100 µL of reconstitution solution (acetonitrile: water = 1:1) was added. The mixture was vortexed for 30 s, ultrasonicated for 10 min in a water bath at 4 °C, and centrifuged at 12,000 rpm for 15 min at 4 °C. Finally, 60 µL of the supernatant was transferred into fresh 2 mL LC-MS glass vials for analysis.

A pooled quality control (QC) sample was prepared by combining 10 µL aliquots from each sample. Methanol, acetonitrile, ammonium acetate, and formic acid used in the extraction process were purchased from CNW Technologies GmbH (Düsseldorf, Germany).

LC-MS process

The LC-MS analysis was conducted using an Agilent 1290 UHPLC system (Agilent Technologies) coupled with high-resolution mass spectrometers, including the Triple TOF 6600 and QTOF 6550 models (AB Sciex and Agilent Technologies). Data acquisition and system control were performed using Analyst TF 1.7 software. Chromatographic separation was achieved with a UPLC BEH Amide column (1.7 μm × 2.1 mm × 100 mm, Waters).

Both positive and negative ionization modes were utilized during the analysis, with injection volumes of 2 µL and 3 µL, respectively. The mobile phase consisted of water containing 25 mM ammonium acetate and 25 mM ammonia (mobile phase A) and acetonitrile (mobile phase B). The gradient elution program started at 5% mobile phase A and 95% mobile phase B, held for 0.5 min, increased to 35% A over 7 min, and further to 60% A by 8 min. The gradient was then reverted to the initial conditions (5% A, 95% B) at 9.1 min and maintained until the 12-minute mark.

The mass spectrometry analysis employed an information-dependent acquisition (IDA) strategy. Ions with signal intensities exceeding 100 were selected for MS2 fragmentation at a collision energy of 30 eV. The acquisition captured 15 spectra every 50 ms. The electrospray ionization (ESI) source was configured with the following settings: Temperature: 650 °C, Nebulizer and auxiliary gas pressures: 60 psi, Curtain gas pressure: 35 psi, and Spray voltages: 5000 V (positive mode) and − 4000 V (negative mode).

Statistical analysis

The raw mass spectrometry data were converted into mzXML format using ProteoWizard software. Preprocessing of the data, including retention time correction, peak identification, extraction, integration, and alignment, was performed using the R package XCMS (v3.2). For preprocessing, minfrac was set to 0.5 and cutoff was set to 0.6. Peak annotation was carried out using the R package CAMERA, and metabolite identification was achieved by referencing an internal secondary mass spectrometry database.

Differential metabolite analysis and salivary biomarker identification were conducted using the Statistical Analysis and Biomarker Analysis modules in MetaboAnalyst 6.0. Linear discriminant analysis (LDA) was plotted using the R platform (version 4.4.0) with several packages, including mlr, mlr3viz, MASS, tidyverse, and ggplot [31]. Prior to differential analysis, 10% of background noise variables were removed based on the relative standard deviation (RSD), and mass spectrometry data were normalized using LOG10 transformation and range scaling.

Statistical significance for differential metabolites was assessed using Student’s t-test and Mann-Whitney U test. To control for the false discovery rate (FDR) in multiple hypothesis testing, the Benjamini-Hochberg method was applied. Classification performance of the univariate and multivariate models was evaluated by estimating the area under(AUC) the receiver operating characteristic curve.

Results

Data quality assessment and sample distribution

The demographic and clinical characteristics of the study participants are summarized in Table 1. A total of 177 participants were enrolled, consisting of 147 GC patients and 30 control subjects with general gastritis. Non-targeted salivary LC-MS analysis identified 368 metabolites in positive electrospray ionization (ESI+) mode and 178 metabolites in negative electrospray ionization (ESI-) mode, based on comparisons with an in-house metabolite database. To ensure data quality, dipeptides and metabolites with MS² scores below 0.6 were excluded prior to differential metabolite analysis. Additionally, duplicate metabolites detected in both ESI + and ESI- modes with low MS² scores were removed. This preprocessing resulted in a final dataset of 333 high-confidence metabolites for subsequent analysis.

Table 1.

Clinical characteristics of 190 subjects

Characters	Gastric cancer	Gastritis
Total	148	34
Age (years), mean (SD)	62.06 (10.67)	53.71 (10.29)
Gender, male (n, %)	77.6%	76.5%
Smoking habit # (n, %)	19.9%	38.2%
Dringking habit # (n, %)	16.0%	41.2%
Pathologic stage Carcinoma in situ Stage 1 Stage 2 Stage 3 Stage 4 Noon Report*	1 18 42 76 8 2	/ / / / / /

^# Defined as Defined as having maintained the habit for a long time and recently quit. * Stage records were missing in 3 cases of gastric cancer

An unsupervised principal component analysis (PCA) model was applied to assess the quality of the metabolomic data. As shown in Fig. 1a, the quality control (QC) samples clustered centrally in the PCA score plot, indicating the robustness and reproducibility of the analysis. To further explore the group separation, a supervised linear discriminant analysis (LDA) model was employed, a commonly used technique in medical research to minimize intra-group variability and accentuate group differences. The LDA model incorporated 333 metabolites for classification, with 80% of the data used for training and 20% for validation, incorporating 50 iterations of K-fold cross-validation. The resulting 2D visualization revealed a distinct separation between the benign control group and all GC samples. Notably, stage I GC samples were clearly distinguishable from more advanced stages (II, III, and IV), which were more tightly clustered. These results suggest that the salivary metabolic profiles of the 333 metabolites exhibit a well-structured distribution across patient groups, aligning with clinical expectations. The LDA model based on these metabolites is illustrated in Fig. 1b.

Fig. 1.

Data Quality and Sample Distribution. (a) PCA plot showing the unsupervised distribution of benign controls, gastric cancer, and QC samples. The clustering of QC samples approximately at the center between cancer and control samples indicates consistent sequencing performance and strong data quality. (b) LDA plot generated using machine learning for a supervised 2D visualization of sample distribution. This plot evaluates the abundance of all 333 metabolites and their differential distribution across the continuum from benign to advanced gastric cancer

Screening of differential salivary metabolites

The flow chart of the differential metabolite screening strategy is shown in Fig. 2a. To identify metabolites distinguishing GC from gastritis, the dataset was divided into a training set and a test set. Systematic sampling was applied to select 54 samples from the total of 177 participants for the test set. Specifically, 50 tumor samples were selected from the 147 GC cases using sample numbers that were multiples of 3, and 4 control samples were chosen using multiples of 6. The remaining 123 samples were designated as the training set.

Fig. 2.

Differential Metabolites Screening and Diagnostic Model. (a) Study cohort and differential metabolites screening strategy. The 177 original samples were divided into a training set and a validation set. Differential metabolites were identified in the training set using T-test and PLS-DA analysis, and their classification performance was evaluated in the validation set. (b) A volcano plot depicting differential metabolites was constructed using t-test (p ≤ 0.05) and fold change (F_C ≥ 2) as the selection criteria. (c) PLS-DA score plot illustrating metabolic differences between the gastric cancer group (green) and the control group (red). Each point represents an individual sample. The first two principal components, PC1 and PC2, explain 30.7% and 11.4% of the total variance, respectively. The model’s R² and Q² values are 0.87 and 0.68, indicating good model fit and classfication performance. (d) Heatmap of hierarchical clustering based on differential metabolites identified between GC and control groups. The x-axis represents samples (green: GC group; red: control group), and the y-axis represents significantly altered metabolites. Color gradients reflect the relative abundance of metabolites (red: upregulated; blue: downregulated). The clustering reveals distinct separation between the GC and control groups, demonstrating significant metabolic differences. (e) Receiver operating characteristic (ROC) curve of N-a-Acetyl-arginine for distinguishing gastric cancer patients from controls. The area under the curve (AUC) is 0.96 (95% CI: 0.92–0.96), indicating excellent diagnostic accuracy. (f) Multivariate ROC curves illustrating the diagnostic performance of models incorporating different numbers of metabolites. The curves represent models with 2 metabolites, 3 metabolites, 5 metabolites, 10 metabolites, 20 metabolites, and all 38 differential metabolites. The AUC values ranged from 0.931 (2 metabolites) to 0.963 (all differential metabolites), indicating that a small number of key metabolites is sufficient to achieve high diagnostic accuracy, with limited performance improvement from incorporating additional metabolites

A partial least squares-discriminant analysis (PLS-DA) model, combined with t-tests and fold-change analysis, was employed for differential metabolite screening. The intersection of these methods was used to refine the results. Using Student’s t-test with a false discovery rate (FDR) threshold of < 0.05 and a cutoff of LOG2|fold change| >1, a total of 57 metabolites were identified from the initial set of 333 metabolites, as shown in the volcano plot (Fig. 2b). The PLS-DA model identified 87 differential metabolites with a variable importance in projection (VIP) score ≥ 1 (Fig. 2c). The permutation testing (1000 iterations) demonstrated an accuracy of 0.96859, an 2 value of 0.86526, and a 2 value of 0.68241 and yielded an empirical p-value of 0.001, confirming the model’s predictive performance was statistically significant and substantially outperformed random models. The PLS-DA score plot and permutation test results are presented in Fig. 2c and Figure S1a, 1b.

The intersection of the two approaches resulted in 44 overlapping differential metabolites. After cross-referencing with the Human Metabolome Database, metabolites such as (3-Carboxypropyl) trimethylammonium, coumarin, O-succinyl-L-homoserine, D-quinovose, promethazine, and raffinose were excluded due to the inability to definitively determine their endogenous origins. This refinement resulted in a final set of 38 differential metabolites. Among these, 38 metabolites, including purines, pyrimidines, essential amino acids, carbohydrates, and their derivatives, were significantly downregulated in GC. Conversely, two metabolites—sebacic acid and 4-aminobutyric acid (GABA)—were upregulated in GC (Table S1). The hierarchical clustering heatmap of these differential metabolites are shown in Fig. 2d.

SMPDB (Small Molecule Pathway Database) enrichment analysis of the remaining 38 metabolites identified significant metabolic disturbances in GC. These included disruptions in purine and pyrimidine biosynthesis, carbohydrate and energy metabolism (notably lactose synthesis, β-alanine metabolism, glycolysis, gluconeogenesis, and galactose metabolism), amino acid metabolism (e.g., histidine metabolism, glutamate and aspartic acid metabolism), and the ammonia cycle. These metabolic changes play critical roles in supporting tumor growth by enhancing the supply of essential building blocks, such as nucleotides and amino acids, optimizing energy production through glycolysis and the Warburg effect, and facilitating metabolic waste management via ammonia recycling. The results of this enrichment analysis are visualized in the enriched bubble plot shown in Figure S2.

Effects of smoking and alcohol consumption on salivary metabolites

To determine whether the observed changes in differential metabolites were tumor-derived, we analyzed the impact of demographic factors (gender, age), lifestyle habits (smoking, alcohol consumption), and tumor characteristics on salivary metabolites. This evaluation was conducted to ensure that the identified metabolites were primarily influenced by tumor biology than hazard factors. No significant differences in gender distribution were observed between the GC and control groups. However, the average age of GC patients was significantly higher than that of the controls by 8.35 years (p < 0.001). Notably, smoking and alcohol consumption were more prevalent in the gastritis control group, with higher rates of smoking (χ² = 6.718, p < 0.01) and alcohol use (χ² = 12.925, p < 0.001).

Given the diverse anatomical distribution of gastric tumors (e.g., gastric angle, fundus, and body) and the heterogeneity of histological subtypes (e.g., adenocarcinoma, mucinous carcinoma, neuroendocrine carcinoma, and signet-ring cell carcinoma), classification based on tumor location or histological type was deemed infeasible. Consequently, these factors were excluded from further analyses. To minimize the confounding effects of disease progression on salivary metabolites, analyses related to smoking and alcohol consumption were restricted to salivary samples from stage III GC patients. Patients were categorized into groups based on their smoking and drinking habits, and a two-sample t-test (raw p ≤ 0.05) was used to evaluate differences in metabolite levels.

Significant differences were identified in metabolites between drinkers and non-drinkers, including 4-aminobenzoate, pyrrolidine, 3’-O-methyladenosine, N-acetyl-D-glucosamine, and L-saccharopine. Similarly, comparisons between smokers and non-smokers revealed significant variations in pyrrolidine, N-acetyl-D-glucosamine, choline, 4-aminobenzoate, oxytocin, and N-methyl-1-deoxynojirimycin. Among these, oxytocin was the only metabolite overlapping with those identified as differentially expressed between GC and control groups. A comparison of oxytocin’s fold change revealed distinct values between GC/control groups (F_C = 0.15) and non-smoker/smoker groups (F_C = 0.57). These findings suggest that smoking and alcohol consumption have a limited impact on the salivary metabolomic profile used for GC diagnosis, reinforcing the tumor-derived nature of these metabolic changes.

Development and validation of diagnostic models

To identify effective diagnostic biomarkers, univariate and multivariate receiver operating characteristic (ROC) models were constructed and validated using the test datasets. In the univariate ROC analysis, 21 metabolites achieved an area under the curve (AUC) greater than 0.80, highlighting their strong discriminative power (Table S1). Among them, N-α-acetylarginine exhibited the most significant difference between gastric cancer and benign controls, with an AUC of 0.96 (95% CI: 0.92–0.99, p < 0.01) (Fig. 2e).

For multivariate ROC analysis, a Support Vector Machine (SVM) algorithm was used to rank features, resulting in six classification models incorporating 2, 3, 5, 10, 20, and 38 differential metabolites, respectively. The model using two metabolites achieved an AUC of 0.94 (95% CI: 0.86–0.99), while the model with 38 metabolites achieved the highest diagnostic accuracy, with an AUC of 0.96 (95% CI: 0.90–0.99) (Fig. 2f). The modest improvement in AUC indicates that the addition of more metabolites beyond a small subset provides limited incremental diagnostic value. This suggests that the model effectively captures key diagnostic information with fewer metabolites, while additional features may introduce noise or redundancy, reducing the contribution of critical variables.

During test set evaluation, N-acetyl-arginine was used as a single feature for classification, achieving correct classification in 48 out of 50 tumor samples and 3 out of 4 control samples, based on an optimized threshold determined from the training set. Additionally, a multivariate ROC model incorporating all 38 differential metabolites demonstrated similar accuracy, correctly classifying 47 out of 50 tumor samples and 3 out of 4 control samples. These results underscore the strong diagnostic potential of highly discriminative metabolites, such as N-acetyl-arginine, and highlight the limited benefit of more complex models in this dataset.

Nevertheless, it should be noted that the reported classification accuracy of 96% needs to be interpreted with caution, given the imbalanced nature of the dataset. A high overall accuracy may be achieved even by a trivial model that predicts the majority class only, which would already yield a baseline performance close to the majority class proportion (50/54 = 92.59%). To contextualize this, we note that both the training-set PLS-DA model and test-set SVM ROC are inherently robust to class imbalance—metrics like AUC-ROC, which summarize performance across thresholds, remain strong despite imbalance. Permutation testing further validated our results: only 1 out of 1000 permuted models achieved a training accuracy equal to or exceeding the original model (empirical p = 0.001). While the accuracy improvement over chance was moderate, this confirms the classifiers effectively capture meaningful biological signals rather than relying on majority class bias.

Final, in small-sample and class-imbalanced datasets, multivariable classifiers often underperform compared to univariate models due to feature redundancy, noise amplification, and unstable feature selection—issues that may explain why our 38-metabolite model showed limited accuracy gains over the single-feature N-acetyl-arginine model. The use of K-means clustering to summarize metabolic variability reduced overfitting by grouping the 40 metabolites into 5 clusters, with cluster centroids representing abstract metabolic patterns. However, this clustering approach may have led to some loss of biologically relevant information. Moreover, the relatively small sample size posed challenges for robust multivariable modeling, as limited data can amplify noise and destabilize feature selection. Expanding cohort sizes and improving detection sensitivity in future studies may help fully leverage additional metabolites and enhance diagnostic performance.

Influence of tumor staging and HER2 expression on salivary metabolomics

To investigate the potential of salivary metabolites as biomarkers for GC progression and HER2 expression levels, we analyzed metabolite differences across tumor stages and HER2 categories. In the LDA model, stage II GC overlapped with stages III and IV while maintaining a noticeable separation from stage I. Hence, we determined to classify stage I gastric cancer as early and set the remaining stage gastric cancer samples as advanced.

To address the imbalance between early-stage (n = 19) and late-stage (n = 126) GC samples, a random sampling and iterative testing strategy was employed. In each iteration, 19 late-stage samples were randomly paired with 19 early-stage samples to form balanced groups. Statistical significance for each metabolite was evaluated using the Mann-Whitney U test (raw p-value < 0.1), with the process repeated 50 times. A metabolite was considered significantly different if its raw p-value was below this threshold in at least 25 of the 50 iterations. This approach balanced statistical robustness and biological relevance, reducing the likelihood of false positives caused by random sampling variability.

After excluding (3-carboxypropyl)trimethylammonium cation, whose endogenous origin could not be confirmed, and 3,4-methylenedioxyamphetamine, which is potentially drug-derived, 6 metabolites were identified as significantly different between early- and late-stage GC. These include L-tyrosine, nicotinate, L-phenylalanine, DL-α-hydroxybutyric acid, L-lysine, and sarcosine. All eight were upregulated in late-stage GC. These metabolites, involved in amino acid metabolism (L-tyrosine, L-lysine, sarcosine), energy pathways (DL-α-hydroxybutyric acid), and signaling molecules (nicotinate), reflect metabolic adaptations that support tumor progression in advanced GC stages. Among them, L-tyrosine exhibited the strongest discriminatory power (AUC = 0.73), while nicotinate and DL-α-hydroxybutyric acid also demonstrated potential with AUC values exceeding 0.70. To control for multiple comparisons, p-values were adjusted using the Benjamini-Hochberg method; however, none of the metabolites remained significant after adjustment. This lack of significance may be attributed to the limited sample size and inherent variability of salivary metabolite levels, underscoring the need for larger cohorts.

HER2 expression was evaluated in stage III GC samples to avoid confounding effects of cancer stage. Samples with HER2-negative, 1+, and 2 + expression levels were categorized as low expression (n = 51), while HER2 3 + samples were classified as high expression (n = 18). Using the same screening strategy, 13 significantly different metabolites were identified. However, 3-aminosalicylic acid, nicotine, D-quinovose, D-allose, phenol, 3-hydroxy-3-methylglutaric acid, propionic acid, and indoleacetic acid were excluded from the analysis due to their potential origin from external sources, such as gut microbiota, plant-derived metabolites, or environmental pollutants. After excluding these metabolites, L-fucose, alpha-D-galactose 1-phosphate, glucosamine, phosphocreatine, and alpha-N-acetyl-L-glutamine were retained as potential HER2-related differential metabolites. These metabolites, which are involved in key metabolic pathways such as glycosylation, amino acid metabolism, and energy regulation, may reflect tumor-specific metabolic alterations associated with HER2 expression.

Significant differences in salivary metabolites were identified between early- and advanced-stage GC as well as between high and low HER2 expression levels. However, these differences were less pronounced compared to the metabolic distinctions observed between GC and gastritis. The metabolic disparities between GC and gastritis involved large-scale alterations, reflecting the transition from a benign to a malignant state. In contrast, the differences associated with GC staging and HER2 expression levels were more subtle. This variation may reflect the limited sensitivity of detection methods or the relatively small sample size, which reduced the statistical power to detect significant intra-disease metabolic changes. While salivary metabolomics demonstrated strong potential for distinguishing GC from gastritis, its ability to differentiate GC stages or HER2 expression levels was limited in this study.

Discussion

Gastric cancer arises from a series of long-term gastric disorders. Early diagnosis of malignant transformation from chronic inflammation remains a clinical challenge [32]. Real-time monitoring and early detection are critical for improving clinical outcomes. Saliva, a non-invasive and easily accessible biological fluid, has not yet been widely used in disease diagnosis. Although previous studies have explored salivary components for early GC detection, such as proteins [33], cell-free DNA [34], and extracellular RNA [35], comprehensive salivary metabolomic analysis remains underutilized.

Emerging liquid biopsy methods (e.g., circulating tumor cells, circulating tumor DNA) in recent years are considered to have drawbacks including difficulties in enrichment, limited sensitivity, poor specificity, and a lack of standardized operating procedures and data processing [36]. In this study, we observed that during the metabolic reprogramming of GC, the abundance of minimal functional units that constitute biological macromolecules (e.g., nucleotides, amino acids) exhibits significant characteristic changes. The potential value of this phenomenon may lie in breaking free from the reliance of traditional liquid biopsy or emerging circulating tumor markers on tumor-specific molecular expression. Unlike emerging liquid biopsy methods, metabolomic approaches use specific metabolic profiles composed of broad-based changes in the abundance of foundational metabolites to match and diagnose the metabolic features of corresponding diseases, demonstrating stronger universal applicability.

In this study, compared to the salivary metabolic profiles of benign gastritis, the most prominent feature of GC was the synergistic dysregulation of multiple categories of metabolites, with all nucleotides, most amino acids, and carbohydrates, among other metabolites, generally downregulated, whereas only GABA and sebacic acid showed significant upregulation. Among the differential metabolites, N-α-acetyl-L-arginine was the most significantly downregulated metabolite in GC. As a core driver of tumor metabolism, arginine promotes carcinogenesis through multiple pathways such as providing energy for tumor proliferation [37], regulating anti-inflammatory responses, supporting pyrimidine synthesis and histone acetylation [38], and facilitating immune escape [39]. Notably, significant downregulation of arginine in gastric cancer saliva has also been reported in earlier studies [40]. Additionally, the significant downregulation of metabolites such as glutamine, histidine, aspartic acid, and various purines and pyrimidines further reflects the typical characteristics of tumor metabolic reprogramming. Furthermore, two metabolites—GABA and sebacic acid—were specifically upregulated in GC. GABA, a classical neurotransmitter, accumulates in a tumor-derived manner due to overexpression of glutamate decarboxylase 1 (GAD1), driving oncogenic signaling (e.g., β-catenin activation, CD8⁺ T cell suppression) [41, 42], whereas sebacic acid exhibited divergent abundance changes in early studies of various tumors, and its specific metabolic mechanisms remain unclear, thus its role in GC requires further investigation [43–46].

Additionally, changes in the abundance of caffeine and its metabolic derivatives (1,7-dimethylxanthine, 1,7-dimethyluric acid) in saliva also warrant attention. In this study, caffeine levels in the saliva of patients with GC were significantly downregulated (F_C = 0.47, p < 0.05), and showed significant positive correlations with 1,7-dimethylxanthine (R = 0.67, p < 0.001) and 1,7-dimethyluric acid (R = 0.53, p < 0.001). Although caffeine and its primary metabolite xanthine are primarily metabolized by the hepatic CYP1A enzyme [50], combined with substantial existing evidence—for example, the involvement of P450 family genes in the pathogenesis of various tumors including GC [47, 48]—and the reported elevated caffeine consumption in other cancer studies (suggesting its potential as a biomarker for evaluating cytochrome P450 (CYP) family gene activity [49]), we propose that future research could explore the use of caffeine as an exogenous biomarker probe for cancer screening, conducting prospective cohort studies to validate its application value.

Compared to the contrast between GC and benign gastritis, the differences in metabolic profiles between early-stage GC and advanced-stage GC are more subtle. In the saliva of patients with advanced GC, the levels of amino acids such as tyrosine and lysine, as well as organic acids like DL-α-hydroxybutyric acid, are significantly downregulated; whereas in the saliva of HER2-positive patients, the levels of sugar derivatives such as glucosamine and L-fucose, along with energy metabolites including phosphocreatine and α-N-acetyl-L-glutamine, are notably upregulated. These metabolic differences suggest that as gastric cancer (GC) progresses, the disease exhibits characteristic metabolic patterns—including the downregulation of tyrosine and lysine which may be linked to enhanced key tumoral pathway activity [50] and increased tumor invasiveness in advanced GC [51]. Additionally, metabolites including L-fucose, α-N-acetyl-L-glutamine, and α-D-galactose 1-phosphate likely reflect that HER2-positive GC cells not only strengthen cell adhesion capacity but also exhibit a significantly increased demand for energy metabolites [52]. It should be noted, however, that since the false discovery rate of all differentially metabolites between early-stage GC and advanced-stage GC were not significant, although these findings hint at potential biological associations, their statistical significance still requires further confirmation through expanding the sample size and independent validation.

Notably, the metabolic differences between early-stage and advanced-stage GC exhibit distinct features compared to those between GC and benign controls, with limited statistical significance. We attribute this limitation to two potential factors: the metabolic differences between GC and benign gastritis are inherently more pronounced than those driving disease progression (i.e., the transition from early- to advanced-stage GC), and salivary metabolites reflect mixed signals from both tumor and non-tumor tissues, which dilutes stage-specific signals.

Although saliva is a biological fluid susceptible to physiological and environmental interference, the findings of this study still demonstrate its core strengths in gastric cancer diagnosis—specifically, its ease of accessibility and outstanding diagnostic performance. Both univariate and multivariate ROC models demonstrated high sensitivity and specificity in distinguishing gastric cancer (GC) patients from benign controls, preliminarily validating the potential of salivary metabolomics in gastric cancer screening. However, it is important to note that there was a notable imbalance in sample size between the GC group and the benign control group in this study. Future improvements could focus on expanding the sample size, incorporating cross-validation and external cohort validation to assess the generalization ability of models under different sample distributions. Additionally, changes in salivary metabolomic profiles in GC require further validation through comprehensive analysis with blood metabolomics and transcriptomics of GC tissues.

Conclusion

This study highlights the significant potential of saliva metabolomics as a non-invasive and promising approach for the early diagnosis of GC. Early detection and timely intervention are critical for improving clinical outcomes, and our findings suggest that monitoring salivary metabolites can provide valuable insights into the onset and progression of GC.

The identified differential metabolites, particularly N-α-acetyl-L-arginine, demonstrated excellent diagnostic performance, emphasizing their utility in distinguishing GC from benign controls. While both univariate and multivariate ROC models showed high diagnostic accuracy, our analyses revealed that metabolic changes occurring during tumor progression differ substantially from those distinguishing benign and malignant states. These findings reinforce the utility of salivary metabolomics as a non-invasive diagnostic tool for GC. Finally, the major limitation of this study is the significant imbalance between the gastric cancer and control groups, as well as between the early and late stages, despite efforts to mitigate this issue. Further studies in larger and more diverse cohorts are required to validate the diagnostic model and explore its applicability in clinical and population screening contexts.

Author contributions

CY Bao, LJ Xiang, SY Wang, and YL Wang performed the collection of saliva samples as well as the collection of pathological data. CY Bao was the led of data calculation and all figures. XF Hu and A Murtazt participated to the disscussion of the Results. CY Bao, W Zhang and Yali Bi were the major contributors to writing the manuscript. Zhang, Liu and Bi were the study’s designer and supervisor. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No. 81472750, 82061138004, 12304319), Anhui Science and Technology Major Project (Grant No. 202103a07020014, 202203a07020013).

Data availability

The 546 LC-MS metabolites and clinical infomation are freely accessible on China National Bioinformatics Database, OMIX database, https://ngdc.cncb.ac.cn/omix, Data number: OMIX007116.

Declarations

Ethics approval and consent to participate

The ethics of the study was approved by the ethics committee of Anhui Medical University (No. 20140141). All saliva samples and clinical information were collected with the informed consent of the patient. Clinical trials are conducted in strict accordance with the Declaration of Helsinki. Consent for publication: All patients agreed and signed informed consent. The clinical information presented in this study is anonymous.

Consent for publication

All authors have read and given their consent for this article to be published Competing interests.

Competing interests

The authors declare no competing interests.