Identification of Salivary Biomarkers in Colorectal Cancer by Integrating Olink Proteomics and Metabolomics

Abstract

Identifying novel biomarkers is crucial for early detection of colorectal cancer (CRC). Saliva, as a noninvasive sample, holds promise for CRC detection. Here, we used Olink proteomics and untargeted metabolomics to analyze saliva samples from CRC patients and healthy controls with the aim of identifying candidate biomarkers in CRC saliva. Univariate and multivariate analyses revealed 16 differentially expressed proteins (DEPs) and 40 differentially accumulated metabolites (DAMs). Pathway enrichment showed DEPs were mainly involved in cancer transcriptional dysregulation, Toll-like receptor signaling, and chemokine signaling. Metabolomics analysis highlighted significant changes in amino acid metabolites, particularly in pathways such as arginine biosynthesis, histidine metabolism, and cysteine and methionine metabolism. Random forest analysis and ELISA validation identified four potential biomarkers: succinate, l-methionine, GZMB, and MMP12. A combined protein-metabolite diagnostic model was developed using logistic regression, achieving an area under the curve of 0.933 (95% CI: 0.871–0.996) for the discovery cohort and 0.969 (95% CI: 0.918–1.000) for the validation cohort, effectively distinguishing CRC patients from healthy individuals. In conclusion, our study identified and validated a panel of noninvasive saliva-based biomarkers that could improve CRC screening and provide new insights into clinical CRC diagnosis.

Introduction

Colorectal cancer (CRC) is the third most commonly diagnosed cancer and the second leading cause of cancer-related mortality worldwide. Projections suggest that by 2040, 3.2 million new cases and 1.6 million deaths will be attributed to CRC.^1,2 Early diagnosis has been shown to significantly improve survival, with a 5-year survival rate of 90% for patients diagnosed at stage I, in contrast to only 10% for those diagnosed at stage IV.^3,4 Currently, CRC detection primarily relies on blood biomarkers (e.g., CEA and CA19–9) and stool tests (e.g., fecal occult blood and DNA testing). However, blood tests have limited sensitivity and specificity and can be associated with other cancers, such as gastric, pancreatic, and ovarian cancers.⁵ Stool tests are prone to false positives.⁶ While colonoscopy remains the most reliable diagnostic tool for CRC, its invasive procedure, poor patient adherence, and operator skill requirements limit its practical utility.^7,8 Consequently, there is an urgent want for more sensitive, less invasive, or noninvasive detection methods for CRC.

Advances in molecular diagnostics and liquid biopsy technologies have positioned saliva as a promising biofluid for early disease detection and diagnosis, owing to its availability and low cost.⁹ Saliva is a hypotonic fluid primarily secreted by the parotid, submandibular, and sublingual glands. These glands, surrounded by capillaries and acini, facilitate molecular exchange, allowing biomarkers from the bloodstream to enter and be secreted into saliva.¹⁰ The electrolyte composition of saliva closely resembles plasma ultrafiltrate, reflecting the body’s physiological state.¹¹ Saliva has benefits such as noninvasiveness, convenient storage, and cost-effectiveness, while enhancing adherence in groups necessitating regular monitoring, hence facilitating cancer screening initiatives.^12,13 With the advancement of molecular biology, mass spectrometry, and microfluidic chip technologies, efficient and accurate detection of analytes in saliva can now be achieved even at low concentrations. Proteomic and metabolomic studies have identified salivary biomarkers for cancers such as pancreatic, breast, and lung cancers.¹⁴⁻¹⁹ Considerably less attention has been directed toward the identification and validation of salivary biomarkers for CRC detection, representing a significant gap in current diagnostic research.

Against this background, developing salivary biomarkers that can effectively and accurately distinguish between healthy individuals and CRC patients is essential for clinical diagnostics. This study integrated the diagnostic potential of saliva with multiomics approaches for CRC research, offering a more comprehensive understanding of proteomic and metabolomic changes in CRC saliva. We aimed to identify and evaluate possible salivary biomarkers for CRC, thereby offering novel noninvasive diagnostic tools for improving CRC detection.

Materials and Methods

Participants

Between August 2023 and February 2024, we recruited 49 eligible CRC patients from the Gastrointestinal Surgery Department at Guangdong Provincial Hospital of Chinese Medicine, collecting preoperative saliva samples. Additionally, 50 healthy volunteers from Guangzhou No. 11 People’s Hospital were included as controls. A discovery cohort comprising 35 CRC patients and 36 healthy controls (HC) was randomly allocated to investigate potential biomarkers for CRC detection, while an independent validation cohort of 14 CRC patients and 14 HC was employed to evaluate the efficacy of the found biomarkers. The research was approved by the Ethics Committee of Guangdong Provincial Hospital of Chinese Medicine (Approval No. BF2021–181–01). We informed all participants about the anonymization process and data storage security measures for the saliva samples and health data, and all participants provided written informed consent.

The criteria for inclusion were as follows: participants in the CRC group met the diagnostic criteria for CRC, as verified by biopsy, and had not undergone therapy before sampling. The HC group got a colonoscopy to exclude colorectal tumors. The exclusion criteria included: (1) significant systemic diseases (e.g., autoimmune disorders, diabetes, hypertension, chronic liver or kidney disease, cardiovascular diseases); (2) hereditary or inflammatory colorectal tumors; (3) a history of malignant neoplasms; (4) pregnancy or lactation; (5) recent use of antibiotics or probiotics (within the past three months); (6) severe oral health conditions (e.g., diminished salivation, dental pain, gingival bleeding, mucosal lesions, or chronic periodontal disease); and (7) a history of smoking.

Saliva Collection

Samples were obtained between 9:00 and 11:00 a.m. to mitigate the influence of circadian fluctuations in saliva production. Participants refrained from food, drink, and oral hygiene for 1 h before collection. After rinsing the mouth with sterile water and resting for 5 min, 1 mL of unstimulated whole saliva was passively collected using a saliva collection tube (Nest, China), ensuring no sputum or mucus contamination. Samples were promptly stored at 2–8 °C and transported to the lab within an hour. The samples were subsequently centrifuged at 3000g for 15 min at 4 °C, and the supernatant aliquots were preserved at −80 °C.

Proximity Extension Assay

Relative quantification of 92 proteins in saliva supernatants from CRC patients and HC participants was achieved with the Olink Target 96 immuno-oncology panel (Olink Proteomics AB, Sweden). The complete protein names and their abbreviations are summarized in Supporting Table S1. This panel utilizes proximity extension assay (PEA), enabling concurrent detection of 96 analytes (92 protein markers and 4 internal controls) from just 1 μL of a sample. The PEA technology uses oligonucleotide-conjugated antibody pairs for target protein recognition, followed by proximity-driven DNA polymerization that creates specific PCR-amplifiable sequences.²⁰ The microfluidic qPCR device (Signature Q100, Olink Proteomics AB, Sweden) was employed to identify and quantify this DNA sequence. The results were presented as normalized protein expression (NPX) values on a log 2 scale, with elevated values signifying enhanced protein expression.

Sample Preparation for Metabolomics

An 80 μL of saliva supernatant was added to a tube containing 320 μL of precooled methanol (LC-MS grade, Thermo Fisher Scientific). The mixture was shaken at 1500 rpm for 30 min at 4 °C, then sonicated in an ice–water bath for 10 min. Protein precipitation was then achieved through 1-h incubation at −20 °C, with subsequent centrifugation at 17,000g for 20 min at 4 °C. Subsequently, a 200 μL supernatant aliquot was transferred to a new tube and vacuum-dried. Add 200 μL of complex solution (mobile phase B: mobile phase A, 7:3, v/v; formulation details provided in subsequent methods) to each dried sample for reconstitution before LC-MS/MS analysis. Sample filtration was performed using 0.2 μm pore-size centrifugal ultrafiltration devices (PALL Corporation) through 10 min centrifugation at 17,000g under 4 °C conditions. The supernatant (120 μL) was split into two aliquots for LC-MS/MS analysis (positive/negative ion modes). System stability was monitored using a pooled quality control (QC) sample prepared by combining 40 μL supernatant aliquots from each individual sample.

Metabolomics Detection

LC-MS/MS analysis was conducted using a Q-Exactive mass spectrometer (Thermo Fisher Scientific) paired with a UHPLC system (Dionex Ultimate 3000, Thermo Fisher Scientific). Separation was achieved on a UPLC HSS T3 column (2.1 mm × 100 mm, 1.8 μm, Waters) at 35 °C, with an autosampler temperature of 4 °C and an injection volume of 4 μL. In positive ion mode, mobile phase A was distilled water (Watsons, China) with 10 mM ammonium formate (LC-MS grade, Sigma-Aldrich) and 0.1% formic acid (LC-MS grade, Aladdin, China), while mobile phase B was acetonitrile (LC-MS grade, Thermo Fisher Scientific) containing 0.1% formic acid. For negative ion mode, mobile phase A was distilled water with 10 mM ammonium formate, and mobile phase B was acetonitrile. The flow rate was maintained at 0.3 mL/min with the following gradient: 0 min, 2% B; 0–1 min, 2% B; 8 min, 98% B; 8–10 min, 98% B; 10.5 min, 2% B; 10.5–12 min, 2% B. The ESI conditions: sheath gas flow rate at 50 arb, auxiliary gas flow at 15 arb, capillary temperature maintained at 320 °C. Data acquisition used data-dependent acquisition (DDA) mode, covering a scan range of 80–900 m/z, with MS and MS/MS resolutions of 60,000 and 15,000, respectively, normalized collision energies (NCE) of 20:30:40, an 8 s dynamic exclusion, and spray voltages of 3.5 kV (positive) and −3.2 kV (negative). QC samples were analyzed under the same conditions as regular samples, with one QC sample injected after every six samples to ensure analytical stability.

Feature Selection

The “randomForest” package was used to feature selection on differentially expressed proteins (DEPs) and differentially accumulated metabolites (DAMs) data. Feature importance was evaluated by the mean decrease Gini index (MDG), with a higher MDG indicating greater importance. Features were selected based on the following criteria: (1) MDG > 1, and (2) has been reported in CRC biomarkers studies.²¹⁻²⁵

Enzyme-Linked Immunosorbent Assay

Saliva samples from an independent validation cohort (14 CRC patients and 14 HC individuals) were validated using ELISA kits for succinate (Cat# E-BC-K902-M; Elabscience, China), methionine (Cat#LB00372B; LabRe, China), GZMB (Cat# ELH-GZMB; Raybiotech), MMP12 (Cat# ELH-MMP12; Raybiotech), and PDGF-BB (Cat# ELH-PDGFBB; Raybiotech). Assays were performed and analyzed according to the manufacturers’ protocols.

Statistical and Bioinformatics Analysis

Clinical Data

The clinical data were analyzed utilizing SPSS (version 25.0). Continuous variables were evaluated using either Student’s t test or Mann–Whitney U test, while categorical variables were assessed through χ² test or Fisher’s exact method. Two-tailed tests with P < 0.05 were deemed significant.

Proteomics Data

Differential analysis of PEA-based proteomics data between CRC and HC groups was conducted using the “limma” package with a P-value threshold of 0.05. Spearman correlation of DEPs was analyzed using the “corrplot” package. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment for DEPs were performed using “clusterProfiler” with an adjusted P-value cutoff of <0.05. Interaction networks of DEPs with GO and KEGG were analyzed using the STRING database (https://cn.string-db.org/) and Cytoscape (version 3.9.1).

Metabolomics Data

Principal component analysis (PCA) and orthogonal partial least-squares discriminant analysis (OPLS-DA) were conducted using SIMCA (version 14.1), with model stability confirmed by 200 permutation tests. Differential metabolite identification incorporated variable importance in projection (VIP) scores derived from OPLS-DA models and statistical significance assessed by the Wilcoxon rank-sum test. Multiple testing correction was applied using Benjamini-Hochberg FDR adjustment. Subsequent metabolic pathway analysis was conducted using the MetaboAnalyst 6.0 platform (https://www.metaboanalyst.ca/). DAMs and DEPs were uploaded to the KEGG Mapper (https://www.kegg.jp/kegg/mapper/color.html) for protein-metabolite common pathway enrichment analysis.

Logistic Regression Model

The discovery cohort was randomly split into a training set (21 CRC patients, 20 HC individuals) and a test set (9 CRC patients, 9 HC individuals) in a 7:3 ratio using the “caTools” package. Biomarkers validated by ELISA were used to build a protein-metabolite combined model via logistic regression in the training cohort. Receiver operating characteristic (ROC) analysis assessed diagnostic performance in both sets. The generalization ability of the model was evaluated through leave-one-out cross-validation (LOOCV), and model validation was further conducted using ELISA data from an independent validation cohort. Diagnostic models for metabolomics, proteomics, and multiomics were constructed with logistic regression on the whole discovery cohort, and ROC analysis was used to compare their diagnostic performance.

For the preprocessing steps of proteomics and metabolomics data, please refer to the Supporting Methods.

Results

Baseline Characteristics of Participants

The baseline demographic profiles and clinical features of both CRC patients and HC in the discovery cohort are summarized in Table 1. To control for potential confounders, individuals with smoking history were excluded from the study. Statistical analysis revealed comparable alcohol consumption patterns across both groups, without significant disparities (P > 0.05). Gender distribution and body mass index (BMI) showed no intergroup differences (P > 0.05), whereas CRC patients were significantly older than HC subjects (P < 0.001). To further assess the influence of age on omics data, we stratified the cohort into two age groups (<50 and ≥50) and analyzed them using uniform manifold approximation and projection (UMAP). The results showed that the P-values for both UMAP dimensions were greater than 0.05, indicating that there was little difference in the distribution of these dimensions across age groups (Figure S1). Therefore, age appears to have little impact on the overall data distribution and is not a major influencing factor.

Table 1. Demographics and Characteristics of Participants in the Discovery Cohort^a.

characteristics	level	CRC (n = 35)	HC (n = 36)	P-value
gender, n (%)	female	20(57.14)	21(58.33)	0.919
	male	15(42.86)	15(41.67)
age (year)	mean ± SD	51.51 ± 11.47	41.39 ± 7.41	<0.001
BMI (kg/m2)	median (IQR)	21.46 (20.32–23.62)	22.08 (20.89–24.34)	0.252
alcohol status, n (%)	never	34(97.14)	35(97.22)	1.000
	current	1(2.86)	1(2.78)
tumor stage, n (%)	I	7(20.00)
	II	12(34.29)
	III	10(28.57)
	IV	6(17.14)
tumor location, n (%)	left colon segment	11(31.43)
	right colon segment	11(31.43)
	rectum	13(37.14)

^aNote: BMI, body mass index; IQR, interquartile range; SD, standard deviation.

Salivary Proteomic Analysis of CRC Patients

The QC results for the PEA experiment are presented in Supporting Figure S2A. Five samples did not meet the QC criteria and were therefore excluded from further analysis. Thirty-eight proteins were excluded from the analysis because more than 25% of the samples contained less than the limit of detection (LOD) value. The remaining 54 proteins were included in the differential expression analysis (Table S2). A heatmap illustrated their expression levels (Figure 1A). Sixteen proteins exhibited significant differences between the CRC and HC groups (P < 0.05), with one protein (ADA) elevated and 15 downregulated, mainly including CXCL9, GZMB, MMP12, TNF10, and PDGF-BB (Figure 1B,C). The Spearman correlation analysis revealed the most robust positive association between CX3CL1 and CD40 (r = 0.85, P < 0.001) and the most significant negative correlation between ADA and ANGPT1 (r = −0.37, P = 0.002; Figure 1D).

Figure 1.

Proteomic profiles in CRC patients and analysis of differentially expressed proteins (DEPs). (A) Heatmap of 54 proteins that were included in further analysis. (B) Volcano plot of DEPs between CRC and HC. (C) Bar chart of fold changes (FC) for 16 DEPs. (D) Spearman correlation heatmap for DEPs. (E) Bubble plots for GO functional classification of DEPs. The y-axis represents the functional categories (biological process/cellular component/molecular function) within the primary GO classification. (F) Bubble plots for KEGG pathway enrichment of DEPs. (G) Protein-GO-protein interaction network. (H) Protein-pathway-protein interaction network.

We conducted GO and KEGG analyses to investigate the functions of DEPs. The GO investigation showed significant enrichment of DEPs in biological processes (BP) including positive regulation of cell adhesion, lymphocyte activation, leukocyte activation, cell activation, and cell–cell adhesion. The cell component (CC) assessment indicated that most proteins were located on the external side of the plasma membrane (Figure 1E). The KEGG analysis revealed that enriched pathways primarily included cytokine-cytokine receptor interaction, viral protein interaction with cytokine and cytokine receptors, transcriptional dysregulation in cancer, and the Toll-like receptor signaling pathway (Figure 1F). To further investigate the protein–protein interactions (PPI) of DEPs and their association with biological functions, we performed network analyses. The PPI results highlighted CXCL9 and GZMB as key proteins with the highest degree scores, suggesting important roles in CRC. The protein-GO interaction analysis showed that CD83 and CD40 are jointly involved in the positive regulation of cell, leukocyte, and lymphocyte activation (Figure 1G). The protein-KEGG interaction analysis indicated that CD40 and IL8 jointly participate in transcriptional dysregulation in cancer, cytokine-cytokine receptor interactions, and Toll-like receptor signaling (Figure 1H).

Salivary Metabolomics Analysis of CRC Patients

The PCA plot demonstrated tight clustering of QC samples, with Pearson’s correlation coefficients consistently maintaining values above 0.99 (Figure S2B,C). These indicated a high level of instrument stability. The PCA analysis of untargeted metabolites showed most samples within the 95% confidence interval, with good separation between the CRC and HC groups and minimal overlap (Figure 2A), indicating strong discriminatory power. The OPLS-DA analysis highlighted metabolic differences between CRC and HC, with R²Y and Q² values of 0.97 and 0.68, respectively (Figure 2B). The permutation tests (n = 200) revealed R² > Q² and a negative intercept for the Q² regression line (R² = [0.0, 0.917], Q² = [0.0, −0.352]), confirming the model without overfitting (Figure 2C). We conducted an initial screening of metabolites using VIP > 1 and P < 0.05, identifying 150 metabolites that differed significantly between CRC and HC (Figure 2D). After removing drugs and impurities, 40 metabolites were identified through secondary matching (QI score ≥40), among which 36 metabolites were confirmed after FDR correction (Table S3). As the number of metabolites showed little variation before and after FDR correction, all 40 identified metabolites were included in the subsequent analysis to avoid missing any potential findings. In addition, to identify metabolic pathways that suggested strong correlation with DAMs, we performed a topological analysis. Six metabolite-related pathways were observed to be significant (P < 0.05 and impact ≥0.1). These pathways primarily involved arginine biosynthesis, histidine metabolism, and the metabolism of alanine, aspartate, and glutamate (Figure 2E), indicating an important role for DAMs in CRC amino acid metabolism. Furthermore, we conducted a common pathway analysis of DEPs and DAMs, with annotation results presented in Table S4. The DEPs and DAMs were involved in the pathophysiological processes of CRC by regulating eight pathways, including metabolic pathways, purine metabolism, efferocytosis, and neuroactive ligand–receptor interactions. These pathways likely influenced tumor metabolic reprogramming, immune microenvironment regulation, and cellular signaling mechanisms.

Figure 2.

Metabolic profiles distinguishing CRC patients from HC. (A) PCA plot of identified untargeted metabolites. (B) OPLS-DA score plot for CRC and HC. (C) Permutation test plot for the OPLS-DA model. (D) Volcano plot of differentially accumulated metabolites (DAMs) with VIP values greater than 1. (E) Pathway topology analysis metabolites that distinguished CRC from HC.

Selection and Validation of Candidate Biomarkers

To identify saliva biomarkers that accurately differentiate CRC from HC, we performed an integrated analysis of proteomics and metabolomics. In the random forest model, five candidate biomarkers that met the selection criteria were identified—succinic acid, l-methionine, GZMB, MMP12, and PDGF-BB (Figure 3A). The Spearman correlation analysis showed the highest correlation between l-methionine and MMP12 (Figure 3B). By referencing spectra from databases, the MS/MS structures of the selected metabolites (succinic acid and l-methionine) were successfully matched with the corresponding fragments (Figure S3). Furthermore, ELISA validation confirmed that succinic acid (Figure 3C) was upregulated, while l-methionine (Figure 3D), GZMB (Figure 3E), and MMP12 (Figure 3F) were downregulated in CRC patients, aligning with discovery cohort trends. PDGF-BB could not be validated due to ELISA detection limits.

Figure 3.

Validation of four features identified from CRC salivary protein and metabolite profiles. (A) Five features (red) were selected from the top 30 features by Mean Decrease Gini score. (B) Correlation network heatmap of the 5 selected features. (C–F) Scatter plots of Succinic acid (C), l-methionine (D), GZMB (E), and MMP12 (F) validated by ELISA. (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001).

We also analyzed the expression levels of the four validated biomarkers across different CRC stages. While succinic acid and l-methionine levels showed no statistical significance in stage I versus normal controls (P > 0.05), marked differences were observed in stages II–III (P < 0.05). Succinic acid remained significant in stage IV, while l-methionine did not (Figure S4A,B). GZMB and MMP12 demonstrated significant changes in early stage CRC (P < 0.001), suggesting their potential for early diagnosis (Figure S4C,D). A two-by-two comparison of the four biomarkers from stage I to IV revealed no significant differences, indicating that their expression levels may remain relatively stable throughout CRC progression.

Construction of a Protein-Metabolism Combined Model

A panel of two proteins (GZMB and MMP12) and two metabolites (succinic acid and l-methionine) was developed using a logistic regression model. This panel achieved an area under the curve (AUC) of 0.902 (95% CI: 0.809–0.996) in the training set and demonstrated strong diagnostic performance in the test set (AUC of 0.988, 95% CI: 0.953–1.000; Figure 4A). The results of LOOCV showed a strong generalization ability of the model, with an AUC of 0.890 (95% CI: 0.8042–0.9756), accuracy of 0.833, precision of 0.818, and specificity of 0.793 (see Table S5 for details). Additionally, the model maintained high predictive accuracy in the independent validation cohort, with an AUC of 0.969 (95% CI: 0.918–1.000; Figure 4B). The ROC analysis revealed the multiomics model had an AUC of 0.933, surpassing the proteomics (AUC of 0.834) and metabolomics (AUC of 0.820) models (Figure 4C). Confusion matrix analysis using optimal thresholds provided recall, specificity, precision, accuracy, and F1 scores for each model (Table 2 and Figure S5). Additionally, the combination of these four biomarkers surpassed the performance of individual biomarkers (Table S6). These findings suggested that the multiomics panel significantly enhanced CRC diagnostic accuracy, highlighting its potential value for clinical applications.

Figure 4.

Logistic regression classifier distinguishing CRC from HC. (A) ROC curves of the protein-metabolism combined model in the training, test, and whole discovery cohort. (B) ROC curve for the protein-metabolism combined model in an independent validation cohort. (C) ROC curves comparing single-omics models with the combined protein-metabolism model in distinguishing CRC from HC.

Table 2. Performance of Six Logistic Regression Models.

	recall	specificity	precision	accuracy	F1 score	AUC
training set	0.864	0.800	0.826	0.833	0.844	0.902
testing set	1.000	0.889	0.900	0.944	0.947	0.988
validation Cohort	1.000	0.857	0.875	0.929	0.933	0.969
proteomics	0.742	0.828	0.821	0.783	0.780	0.834
metabolomics	0.774	0.828	0.828	0.800	0.800	0.820
combination	0.903	0.828	0.848	0.867	0.875	0.933

Discussion

Saliva, as a vital bodily fluid, is abundant in DNA, RNA, proteins, and metabolites, which can reveal the physiological and pathological conditions of the organism.²⁶ Recent researches indicate that specific miRNAs, immunological components, and metabolites in saliva may be linked to the development of CRC.²⁷⁻³⁰ Although these studies are still in the exploratory phase, saliva demonstrates great potential as a noninvasive tool for CRC detection. Our study focused on analyzing the protein and metabolic profiles of CRC through an integrated approach of proteomics and metabolomics, as well as identifying and validating DEPs and DAMs in saliva samples from CRC patients. The research explored the key pathways associated with these DEPs and DAMs, culminating in the development of a diagnostic model for CRC based on a combination of proteins and metabolites. This model provides a novel, efficient, and noninvasive tool for early CRC detection.

Our study used PEA technology to detect 92 immune-oncology-related proteins in saliva samples from CRC patients and HC, identifying 16 DEPs. These proteins were enriched in pathways related to transcriptional dysregulation in cancer, chemokine signaling, MAPK, and Toll-like receptor signaling, highlighting the role of the tumor microenvironment, inflammation regulation, and signal transduction in CRC development. Notably, GZMB, a key molecule in T cell and NK cell-mediated tumor killing, showed low expression in CRC, which was associated with tumor metastasis, immune dysfunction, and poor prognosis.^23,31 MMP12, a matrix metalloproteinase involved in ECM remodeling and anticancer barrier function, also displayed altered expression in CRC, with its inhibition potentially weakening the anticancer barrier.^32,33 The low levels of GZMB and MMP12 in saliva may reflect the dynamic balance between local tumor immune activation and distal immune suppression in CRC patients. Specifically, high expression of GZMB and MMP12 in CRC tissues indicates active local immune responses and matrix remodeling, while saliva glands may be influenced by systemic immune suppression signals (e.g., IL-10 and TGF-β), leading to significantly reduced levels of these proteins in CRC patients.^24,34,35 This phenomenon underscores the differences in immune regulation between local and systemic environments in CRC.

Moreover, the Spearman correlation analysis between differentially expressed proteins revealed a strong negative correlation between ADA and ANGPT1. ADA, an enzyme involved in adenosine metabolism, was highly expressed in CRC tissue and consistent with the saliva findings, suggesting a disturbance in adenosine metabolism.³⁶⁻³⁸ Adenosine accumulation may promote immune escape, though ADA upregulation only partially alleviates this. ANGPT1, as a vascular stability factor, played a crucial role in maintaining vascular integrity. Its reduced expression likely caused vascular instability, disrupting blood supply to tumors and immune cell infiltration, thereby promoting tumor growth and metastasis. A study found that ANGPT1 expression was reduced in CRC tissues, which is consistent with our results, and its downregulation was associated with CRC tumor metastasis and lower patient survival.³⁹ The negative correlation between ADA and ANGPT1 reflected complex interactions between the immune and vascular systems within the CRC tumor microenvironment, emphasizing the roles of adenosine metabolism and vascular stability in CRC progression. All of the above DEPs provide new insights into CRC immune regulation and potential noninvasive biomarkers, warranting further investigation into how systemic signals influence the immune microenvironment of salivary glands for improved diagnostic strategies.

Our metabolomic analysis revealed significant changes in amino acid metabolites in the saliva of CRC patients, reflecting systemic metabolic dysregulation, particularly an imbalance in amino acid metabolism. Several amino acids, including l-methionine, l-tryptophan, and l-histidine, were significantly reduced, consistent with alterations in amino acid metabolism observed in the blood of CRC patients.⁴⁰⁻⁴² The decrease in l-methionine levels may be linked to its depletion by cancer cells for methylation, a process essential for tumor proliferation and survival. Furthermore, methionine levels are inversely correlated with CRC risk.^43,44 We also observed a marked increase in succinic acid, a key metabolite in the tricarboxylic acid (TCA) cycle, in the saliva of CRC patients, consistent with another CRC saliva metabolomics study.³⁰ This increase reflects the metabolic reprogramming of tumor cells. It has been demonstrated that sudden disruption of the TCA cycle, mutations specific to Warburg-overexpressing pathways, enhanced anaerobic shunting, or stabilization of hypoxia-inducible factor (HIF) can result in the accumulation of succinate within the context of the Warburg effect.⁴⁵ Succinate, derived from tumor cells, has been shown to promote macrophage polarization and facilitate cancer metastasis.⁴⁶ Succinate has been found to activate STAT3, thereby promoting epithelial-mesenchymal transition (EMT) in CRC, which further drives metastatic progression.⁴⁷ Additionally, elevated expression of succinate has been observed in other malignancies, including breast cancer, gastric cancer, and ovarian cancer, suggesting that succinate may hold potential as a candidate biomarker for various types of cancer.⁴⁸⁻⁵⁰ Among the metabolic pathway analysis, changes in arginine biosynthesis were most pronounced. In this study, salivary arginine levels were significantly reduced in CRC patients, although other investigations reported higher plasma arginine levels.⁵¹ Given the connection between saliva and blood circulation, the decrease in salivary arginine likely reflects its metabolic consumption in the tumor microenvironment. The rapid depletion of arginine, facilitated by enzymes such as arginase (ARG) and inducible nitric oxide synthase (iNOS), can suppress T-cell function, thus promoting immune escape and tumor progression.⁵² These salivary metabolic changes offer potential for early CRC detection and disease monitoring, though further research is needed to determine whether these alterations are causes or consequences of CRC progression.

We developed a comprehensive diagnosis model for CRC saliva (succinate, l-methionine, GZMB, and MMP12) by integrating multiomics and machine learning techniques. The diagnostic performance of the combined protein-metabolite model was as follows: AUC 0.933, accuracy 0.867, sensitivity 0.903, and specificity 0.828. The predictive power of the model remained significant in the independent validation cohort, with an AUC of 0.969 (95% CI: 0.918–1.000). However, additional validation across external cohorts from diverse healthcare organizations and populations is required to comprehensively evaluate the model’s validity and stability in real-world clinical settings. Our model outperformed the commonly used biomarkers CEA and CA19–9, in which respectively 46.3 and 68.2% of CRC patients did not show elevated levels. The AUCs for CEA, CA19–9, and their combined assays were 0.789, 0.690, and 0.799, respectively.⁵³ For CRC patients who are negative for traditional biomarkers, this model may help in the earlier identification of cancer, thereby promoting early intervention and improving survival rates. Table S7 provides a systematic overview of CRC biomarkers from various biological fluids, showing that our model’s sensitivity (0.903) and specificity (0.828) exceeded those of markers from most sample types, such as SEPT9 in serum/plasma, miR-21 in saliva, and diacetylspermine and kynurenine in urine. Compared with previous salivary metabolomics studies, the present study significantly improved the feasibility of clinical translation by simplifying the salivary biomarkers from 19 to 4 and improving the AUC (from 0.879 to 0.933).³⁰ The current “gold standard” for CRC diagnosis involves invasive tissue sampling, which limits its use. This model offers a novel, noninvasive approach for early CRC screening using saliva.

This research reveals changes in salivary protein and metabolite profiles in CRC patients, but several limitations need addressing. First, our study was limited by a small sample size and potential confounding factors. While we attempted to control for some of these variables by applying strict inclusion criteria (e.g., excluding individuals with severe oral diseases), factors like mild oral inflammation and regional differences in dietary habits may still influence salivary composition and, consequently, the accuracy of biomarkers. Future research should focus on multicenter collaborations that involve participants from diverse geographic regions, ethnic groups, and healthcare settings to improve sample diversity and the generalizability of the model. Meanwhile, stratified analyses of oral diseases should be performed to systematically explore the relationship between oral health status and salivary biomarkers, thereby enhancing the specificity and clinical relevance of the tests. Furthermore, although 92 immune tumor-related proteins were identified using the PEA technique, technical limitations remain, including the potential omission of other crucial biomarkers. Future studies will validate and complement these findings using techniques like mass spectrometry-based proteomics to provide a more comprehensive analysis of the salivary proteome in CRC patients. Finally, although these biomarkers demonstrate promising diagnostic potential, the causal relationship between salivary biomarkers and CRC remains unclear, and the fundamental principles for detecting biomarkers of distant organs in saliva still need to be elucidated. To address this issue, future research should focus on designing comprehensive experimental studies, including in vitro cellular experiments and in vivo animal models, to determine whether these biomarkers are drivers of CRC development or secondary consequences of the disease. These investigations will provide critical insights into the pathogenesis of CRC and contribute to developing more accurate diagnostic biomarkers.

Conclusions

Our study is the first to combine multiomics analysis with the study of CRC salivary biomarkers, identifying specific proteins and metabolites expressed in CRC patients. The expression levels of succinate, l-methionine, GZMB, and MMP12 were validated in salivary samples from the verification cohort. Logistic regression developed a combined protein-metabolite model that outperformed the single-omics model, providing valuable insights for noninvasive early diagnosis and clinical detection of CRC.

Glossary

Abbreviations

CRC

colorectal cancer

PEA

proximity extension assay

HC

healthy controls

NPX

normalized protein expression

LOD

limit of detection

DEPs

differentially expressed proteins

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

QC

quality control

PCA

principal component analysis

OPLS-DA

orthogonal partial least-squares discriminant analysis

DAMs

differentially accumulated metabolites

VIP

variable importance in projection

MDG

mean decrease Gini index

ROC

receiver operating characteristic

BMI

body mass index

UMAP

uniform manifold approximation and projection

BP

biological processes

CC

cell component

MF

molecular function

PPI

protein–protein interactions

AUC

area under the curve

SUCNR1

succinate receptor 1

ARG

arginase

iNOS

inducible Nitric oxide synthase

Data Availability Statement

The data sets generated during the current study have been deposited to the Science Data Bank repository at https://cstr.cn/31253.11.sciencedb.20464, DOI: 10.57760/sciencedb.20464. All data generated or analyzed during this study are included in Supporting Information files.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jproteome.5c00091.

• Information of 92 proteins (Table S1); common pathway (Table S4); LOOCV results (Table S5); AUC of the salivary biomarkers (Table S6); CRC biomarkers in different sample types (Table S7); scatter plots from UMAP analysis (Figure S1); quality control of experimental data (Figure S2); fragment matching (Figure S3); expression levels of the four validated biomarkers at different stages of CRC (Figure S4); and confusion matrix for six logistic regression models (Figure S5) (PDF)

• NPX data of the 54 proteins (Table S2); summary of differentially accumulated metabolites (Table S3) (XLSX)

Author Contributions

H.S.: Conceptualized the project, analyzed and interpreted the data, and wrote the manuscript. X.G. and W.Z.: Gathered and analyzed the data. H.S., F.L., and X.L.: Conducted the experiments. X.Z., C.W., and W.C.: Participated in the study. W.L., P.T., and L.Z.: Collected and processed samples. B.G. and Q.C.: Provided financial support. Q.C.: Supervised the entire study and revised the manuscript. The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

This research was supported by the State Key Laboratory of Dampness Syndrome of Chinese Medicine Projects (no. SZ2021ZZ27 and no. SZ2023ZZ16).

The authors declare no competing financial interest.

Notes

The study was conducted in accordance with the Declaration of Helsinki, and was approved by the Ethics Committee of Guangdong Provincial Hospital of Traditional Chinese Medicine (approval number: BF2021-181-01). Written and signed informed consents were obtained from all study subjects.