Integrative analysis of saliva-derived exosomal proteome and lipidome for the diagnosis of esophageal squamous cell carcinoma

Abstract

Background

Early diagnosis of esophageal squamous cell carcinoma (ESCC) is crucial for improving patient prognosis. Currently, the diagnosis of ESCC primarily relies on endoscopic biopsy. Salivary exosomes have shown great potential in non-invasive screening, but their proteomic and lipidomic characteristics remain to be reported.

Methods

Exosomes were isolated from salivary samples of 54 patients with ESCC and 62 healthy controls using ultracentrifugation, and subsequently subjected to non-targeted proteomic and lipidomic analysis by liquid chromatography-tandem mass spectrometry (LC–MS/MS). Differentially expressed proteins and lipids in salivary exosomes were identified through differential analysis, and a comprehensive analysis of multi-omics data was performed using correlation analysis.

Results

Significant proteomic and lipidomic differences were observed between ESCC patients and healthy controls. The proteomic characteristics of ESCC were primarily manifested in immune responses, disruption of tissue structural homeostasis, and enhanced antifungal and antimicrobial humoral immune responses. Through multi-omics analysis, we found that ESCC may regulate fatty acid metabolism by modulating epigenetic modifications, thereby influencing the oral immune microenvironment. Finally, a diagnostic model constructed using 28 lipid features achieved excellent diagnostic performance (Are Under the Curve = 1.000) for ESCC diagnosis.

Conclusions

Our study revealed significant alterations in the proteomic and lipidomic profiles of the oral microenvironment in ESCC patients, which may provide new insights into the development and progression of ESCC. We found that lipid features have high potential for diagnosing ESCC, providing support for further validation in larger cohorts.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-14452-x.

Introduction

According to the 2022 global cancer statistics, esophageal cancer ranks 11th in terms of incidence and 7th in terms of mortality worldwide, while in China, it ranks as the 5th leading cause of cancer deaths [1]. Of paramount importance is that approximately 90% of esophageal cancer cases in China are attributed to esophageal squamous cell carcinoma (ESCC) [2, 3]. Early-stage ESCC patients often remain asymptomatic, leading to delayed diagnosis and treatment, until they present with progressive dysphagia at an advanced stage, when therapeutic options are limited [4]. This dismal scenario contributes to the dismal 5-year survival rate of less than 20%, highlighting the imperative need for early diagnosis and treatment of ESCC [5]. Unfortunately, the current diagnostic landscape is plagued by a significant paucity of effective methods for early detection and diagnosis of ESCC. Clinically, the gold standard for ESCC diagnosis is endoscopic biopsy of pathological tissues from the upper gastrointestinal tract, which can diagnose and even treat ESCC at an early stage, but its high cost and invasiveness limit its use as a widespread screening tool [6, 7]. Thus, the identification of sensitive and non-invasive biomarkers is of paramount importance for revolutionizing patient outcomes.

The advent of liquid biopsy technology has ushered in a new era of promise for early tumor diagnosis. By analyzing patient samples, including blood, saliva, and urine, a plethora of biomarkers can be detected, encompassing circulating tumor cells, circulating tumor DNA, free RNA, exosomes, proteins, and lipids, thereby offering a non-invasive and sensitive approach for cancer detection [8]. In stark contrast to traditional tissue biopsy, liquid biopsy presents a more streamlined and convenient sample collection process, characterized by minimal or non-invasive procedures, thereby facilitating real-time detection and repeat testing [9]. Furthermore, a burgeoning body of research has consistently demonstrated the pivotal role of exosomes in modulating cellular biological functions and their profound implications in pathological conditions, with a particular emphasis on their involvement in cancer development and progression [10, 11]. Exosomes can transfer various molecules between cells, serving as a crucial communication tool, particularly in cancer [12]. Rich in proteins, nucleic acids, and lipids, exosomes can participate in the regulation of multiple cell signaling pathways in the tumor microenvironment, influencing tumor development and progression [13]. Furthermore, tumors can also release exosomes to affect the distant microenvironment [14]. Given the close anatomical relationship between the oral cavity and esophagus as part of the upper gastrointestinal tract, analysis of exosomes isolated from ESCC patient saliva can reveal unique metabolic changes in the oral microenvironment of ESCC patients, thereby facilitating the discovery of more sensitive biomarkers for early diagnosis of esophageal squamous cell carcinoma.

To date, there have been no reports on the proteomics and lipidomics of exosomes in saliva of ESCC patients, which has hindered our understanding of the disease mechanisms and exploration of diagnostic biomarkers. In this pilot study, we conducted proteomics and lipidomics analyses of ESCC saliva exosomes and investigated the potential disease mechanisms through correlation analysis. Furthermore, based on random forest classification, we identified diagnostic biomarkers from the proteomics and lipidomics data, and constructed a diagnostic model comprising 28 lipid features, which demonstrated 100% sensitivity and specificity in the test set.

Materials and methods

Clinical sample collection and clinicopathological characteristics

Between May 2023 and December 2023, we enrolled 54 patients with ESCC and 62 healthy controls (CTRL). All patients underwent endoscopic biopsy with histopathological examination, and were pathologically confirmed as ESCC by at least two experienced pathologists according to the American Joint Committee on Cancer (AJCC) 8th Edition Staging System. Healthy controls underwent endoscopy and imaging examinations and were found to have no tumors or diseases. Following strict screening, all participants were excluded if they had oral system diseases, had received antibiotic treatment within the past six months, or had received adjuvant therapies such as chemotherapy, radiotherapy, or immunotherapy for ESCC. All sample collections and processing were performed at the Central Laboratory of the Fifth Affiliated Hospital of Sun Yat-sen University. All saliva samples were collected around 9 AM, after a one-hour fasting period and rinsing the mouth with lukewarm water (without antibacterial ingredients) for 30 s. Saliva samples were collected using saliva collection tubes and then centrifuged at 1000 × g and 4 °C for 5 min. The supernatant was then stored at −80 °C. This study was approved by the Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University (Ethics No. 2022-K54-1). The clinical and pathological characteristics of all participants are presented in Table 1.

Table 1.

Clinical baseline features of sample data

	ESCC(N = 54)	CTRL(N = 62)
Age (years)	69.19 ± 7.88	61.24 ± 6.75
BMI	21.47 ± 2.38	23.61 ± 2.17
Tumor differentiation degree
High differentiation	8	NA
Middle differentiation	24	NA
Low differentiation	25	NA
pTNM Stage
I	3	NA
II	9	NA
III	45	NA
IV	0	NA
T Stage
T1	6	NA
T2	16	NA
T3	35	NA
N Stage
N0	16	NA
N1	24	NA
N2	14	NA

Isolation and purification of salivary exosomes

To remove apoptotic bodies and cellular debris, the saliva samples stored at −80 °C were thawed and centrifuged at 4 °C at 3000 × g for 20 min. Subsequently, the supernatant was centrifuged at 12,000 × g for 20 min at 4 °C to remove microvesicles. The saliva samples were then diluted to a final volume of 10 mL with phosphate-buffered saline (PBS) and filtered through a 0.22 μm membrane. Next, the samples were subjected to ultracentrifugation at 100,000 × g for 70 min, repeated twice. Finally, the pellets were resuspended in PBS for subsequent protein and lipidomic sequencing analysis.

Acquisition and analysis of proteomic data

Take 20 μL of exosome sample and add an equal volume of pre-chilled lysis buffer. Lyse on ice for 30 min. After heating at 95 °C in a metal bath for 15 min, determine the protein concentration of each sample using a Bicinchoninic Acid Assay (BCA) assay kit. Add trypsin and incubate at 37 °C with constant shaking overnight to ensure complete digestion.

The Evosep One liquid chromatography system (Evosep Biosystems) was employed, which enabled the analysis of 60 samples per day using the standardized gradient (SPD60). To enhance the sensitivity and accuracy of sample analysis, a C18-filled trap column was used for sample pre-treatment. Chromatographic separation was performed using an 8 cm x 100 μm i.d. capillary column. The mobile phase consisted of 0.1% formic acid in water (v/v) as solvent A and 0.1% formic acid in acetonitrile (v/v) as solvent B. The Orbitrap Fusion Lumos Tribrid mass spectrometer, equipped with a Nanospray Flex ion source, was operated in positive ion mode, with a spray voltage of 2200 V, and data-independent acquisition (DIA) mode was employed, with fragmentation performed using higher-energy collisional dissociation (HCD) mode.

Peptide identification was performed using the PEAKS DB search engine combined with PEAKS de novo sequencing, with a De Novo Average Local Confidence (ALC) (%) threshold set at 15%. The false discovery rate (FDR) was controlled at 1% using the target-decoy approach. The raw files were corrected for precursor ion mass accuracy and fragment ion mass accuracy using the refined Tandem Mass Spectrometry (MS/MS) spectra. The precursor mass tolerance was set at 15 ppm, and the fragment mass tolerance was set at 0.03 Da. Trypsin was specified as the enzyme, allowing up to three missed cleavage sites. A maximum of three variable modifications per peptide was allowed, including protein N-terminal acetylation, cysteine carbamidomethylation, methionine oxidation, aspartic acid and glutamic acid deamidation, and glutamine to pyroglutamic acid conversion.

Acquisition and analysis of lipidomic data

Take 20 μL of exosome sample and add an equal volume of pre-chilled lysis buffer. Lyse on ice for 30 min. Add 160 μL of pre-chilled (4 °C) chloroform–methanol solution (3:1), vortex thoroughly, and centrifuge. Separate the liquid into two layers, transfer the lower organic phase to a vacuum concentrator for evaporation, and reconstitute with 40 μL of chloroform–methanol solution (25:65). Perform lipid quantification using LC–MS.

Lipid extracts were separated using a Thermo Scientific Dionex UltiMate 3000 Rapid Separation LC (RSLC) system, equipped with a Waters ACQUITY UPLC HSS T3 analytical column (2.1 × 150 mm, 1.8 μm, 100 Å) and a Waters ACQUITY UPLC HSS T3 VanGuard pre-column (2.1 × 5 mm, 1.8 μm, 100 Å). The mobile phase solvents A and B consisted of acetonitrile: H2O (6:4 v/v) and isopropanol: ACN (9:1 v/v), respectively, both containing 10 mM ammonium acetate and 0.1% acetic acid. The chromatographic separation was performed at a temperature of 55 °C, with a flow rate of 0.35 mL/min, and a gradient elution program as follows: 0–3 min, 30%−35% solvent A; 5–14 min, 65%−98% solvent A; 18–18.1 min, 98%−30% solvent A; and 18.1–22 min, 30% solvent A. Mass spectrometric data were acquired on an Orbitrap Fusion Lumos mass spectrometer equipped with a heated electrospray ionization (HESI) source, operating in both positive and negative ion modes, with separate runs performed for each polarity. The spray voltage was set at 3500 V for both positive and negative ion modes. The ion transfer tube temperature was maintained at 300 °C, and the vaporizer temperature was set at 350 °C. The sheath gas flow rate was optimized at 40 arbitrary units (arb), and the auxiliary gas flow rate was set at 15 arb. In data-dependent acquisition (DDA) mode, lipid identification was performed using a stepped normalized collision energy (NCE) fragmentation method (20%, 30%, 35%) to generate rich MS/MS spectra, with a mass resolving power of 15,000 at m/z 200. The quadrupole isolation window was set at 1.6 m/z, and dynamic exclusion was enabled for 10 s. Four separate general-purpose fragmentation (sGPF) methods were employed to analyze quality control (QC) samples using deep DDA-MS/MS: (run 1) m/z 150–250, 550–650, 950–1050, 1350–1450, and 1750–1850; (run 2) m/z 250–350, 650–750, 1050–1150, 1450–1550, and 1850–1950; (run 3) m/z 350–450, 750–850, 1150–1250, and 1550–1650; (run 4) m/z 450–550, 850–950, 1250–1350, and 1650–1750.

The raw data files were processed using LipidSearch software (version 4.1) (Thermo Fisher Scientific) for lipid molecule identification and quantitation. The peak detection parameters were set as follows: recomputing isotopes was enabled, and the retention time interval was set at 0.01 min. The lipid identification parameters were defined as follows: search type, Product; experiment type, LC–MS; precursor mass tolerance, 10 ppm; fragment mass tolerance, 10 ppm; intensity threshold, 1.0%; target category, all lipid classes. Adduct ions were specified for both positive and negative ion modes: + H, + NH4, + Na, + H-H2O, and + 2H for positive ion mode, and -H, + HCOO, + CH3COO, + Cl, and −2H for negative ion mode. The Top rank filter was enabled, and the Main node filter was set to main isomer peak. The m-Score threshold was set at 5.0, and FA priority was enabled. The ID Quality Filter was applied, with the following criteria: A, lipid class and fatty acid identified; B, lipid class and some fatty acids identified; C, lipid class or fatty acid identified; D, lipid identified by other fragment ions (H2O loss and other non-specific neutral losses). Quantitation was performed with a mass-to-charge ratio (m/z) tolerance of ± 5.0 ppm and a retention time range of −0.5 to + 0.5 min. Peak alignment was performed using the following parameters: alignment method, maximum value; retention time tolerance, 0.25 min; calculate unassigned peak area, enabled; filter type, new filter; Top rank filter, enabled; Main node filter, main isomer peak; m-Score threshold, 5.0; ID Quality Filter, A, B, C.

Nanoparticle-based tracking analysis (NTA)

NTA analysis was performed on every exosome sample extracted to guarantee that the extracted exosomes met the required size and concentration standards. By exploiting the principles of light scattering and Brownian motion, the particle size distribution in the sample was determined. During detection, the trajectories of individual particles undergoing Brownian motion in the suspension were tracked within a predetermined time window, enabling kinetic analysis of each particle. The mean squared displacement and diffusion coefficient of each particle were calculated according to the Stokes–Einstein equation. Subsequently, the fluid dynamic diameter and concentration of exosomes were further calculated based on the computed diffusion coefficient and mean squared displacement.

Transmission Electron Microscopy (TEM)

A 5 μL aliquot of exosome sample was deposited onto a copper grid, which was then incubated at room temperature for 5 min. Following incubation, a drop of 2% uranyl acetate negative staining solution was added, and the sample was subsequently air-dried at room temperature for 2 h. Finally, the morphology of the exosomes was examined using TEM.

Western blot analysis (WB)

The exosomes were lysed under ice-cold conditions, and the protein concentration was determined using a BCA protein assay kit. Subsequently, sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE) was performed using 10% separating gel and 4% stacking gel. A 50 μg aliquot of exosome sample was loaded onto the polyacrylamide gel, and the proteins were separated by electrophoresis. The proteins were then detected using primary antibodies against CD63 (1:1000, EPR5702, Abcam), CD81 (1:1000, EPR4244, Abcam), HSP70 (1:2000, EPR16892, Abcam), and Calnexin (1:1000, EPR3633, Abcam). Alkaline phosphatase-conjugated anti-rabbit IgG (1:1000, BA 1011, Boster) was used as the secondary antibody for detection. Finally, the proteins were visualized using an alkaline phosphatase substrate.

Validation and prognostic analysis based on public databases

We further validated our transcriptomic data results using a public dataset. We integrated transcriptomic expression data from 81 tumor and 11 normal samples from The Cancer Genome Atlas (TCGA) project. Initially, we validated the expression of genes corresponding to the differentially expressed proteins based on the database results (|log₂ fold change|> 1; P < 0.05). To supplement prognostic data, we further assessed the correlation between the corresponding genes and overall survival (OS) using the Mantel-Cox test and visually represented the correlations with Kaplan–Meier curves.

Statistical analysis

In this study, all count data were expressed as frequencies, and all continuous variables were presented as means ± standard deviations. For normally distributed data with equal variances, independent sample t-tests (Student's t-tests) were employed to compare the means between groups. Otherwise, Wilcoxon rank-sum tests were used for non-normally distributed data. To control the FDR at < 0.05, the Benjamini–Hochberg method was applied to correct the P-values obtained from multiple comparisons. Pearson correlation analysis was performed for normally distributed data, while Spearman correlation analysis was used for non-normally distributed data. The R software (version 4.3.1) was utilized to generate volcano plots and heatmaps. Cytoscape (version 3.9.1) was employed to visualize the network graph of inter-group correlation analysis. Orthogonal partial least squares discriminant analysis (OPLS-DA) of lipidomics data was performed using SIMCA software (version 14.1). A random forest algorithm was used to build a diagnostic model, and tenfold cross-validation was performed to prevent model overfitting and ensure the reliability of the model.

Results

Isolation and characterization of salivary exosomes

Figure 1a illustrates the recruitment process, where 54 ESCC patients and 62 CTRL samples were enrolled in this study after rigorous screening. Saliva samples were collected and exosomes were isolated and extracted using ultracentrifugation, followed by separate preparation of protein and lipid samples from the exosomes. Subsequently, multiple methods were employed to characterize the extracted exosomes. The diameter range and particle morphology of the exosomes were determined using NTA and TEM, respectively. The NTA results revealed that the diameter range of the exosomes was 30–200 nm, with an average diameter of 149.2 nm (Fig. 1b). TEM observations showed that the exosomes exhibited a spherical shape with a bilayer membrane structure (Fig. 1c). WB analysis demonstrated that the exosomes were positive for the marker proteins CD81, CD63, and HSP70 in both ESCC patients and healthy controls, whereas the negative marker Calnexin was not detected in the exosome samples (Fig. 1d).

Fig. 1.

Extraction and identification of exosomes. a Recruitment of participants, exosome isolation, and proteomic and lipidomic analyses. b Nanoparticle tracking analysis of exosome size. c Transmission electron microscopy images of exosomes (bar = 200 nm). d Western blot (WB) results of markers (positive markers CD81, CD63, HSP70 and negative marker Calnexin) from exosomes isolated from the saliva of two ESCC patients and two control subjects

Proteomic profiling of exosomes from ESCC and CTRL saliva samples

We first analyzed and processed the proteomic data. The principal component analysis (PCA) results showed a clear separation between ESCC and CTRL. The QC samples from ESCC patients and controls clustered together, indicating good reproducibility of the experimental results (Fig. 2a). Using an untargeted proteomics approach, we filtered out proteins that were not expressed in more than 50% of the ESCC patient and control samples, and identified a total of 227 differentially expressed proteins. We calculated the fold change and FDR-adjusted p-value for each protein and presented the results in a volcano plot (Fig. 2b). Based on the threshold filtering (FDR < 0.05, |log₂FC|≥ 1), we found that 14 proteins were significantly upregulated and 19 proteins were significantly downregulated in ESCC patients compared to controls. The heatmap displayed the expression profiles of the 33 differentially expressed proteins between ESCC patient and control samples (Fig. 2c).

Fig. 2.

Proteomic differences between ESCC and CTRL groups. a PCA score plot of proteomic. b Volcano plot of DEPs. Green dots represent down-regulated proteins in ESCC, while red dots represent up-regulated proteins in ESCC. c Heatmap illustrating the expression profiles of 33 DEPs between ESCC and CTRL groups. d Protein–protein interaction (PPI) network of the DEPs. e Boxplots of the top six proteins with highest node degree in the PPI network between the two groups. f GO enrichment analysis of upregulated DEPs in ESCC. g GO enrichment analysis of downregulated DEPs in ESCC

We retrieved the differentially expressed proteins between the two groups from the STRING database (https://cn.string-db.org/) and constructed a protein–protein interaction (PPI) network, which is depicted in Fig. 2d. The network comprised 26 protein nodes and 39 edges, with an average node degree of 3 and a clustering coefficient of 0.582 (p-value = 8.63e-10), indicating a significant level of interconnectedness among the proteins. Using Cytoscape software, we further analyzed the network and ranked the proteins by node degree. The top 6 proteins, including H2AC20, ALB, GAPDH, PIGR, H2BC14, and MPO, are shown in Fig. 2e, which compares their expression levels between the ESCC and CTRL groups. Notably, H2AC20 and H2BC14 are members of the histone family, playing critical roles in maintaining chromatin structure.

We performed a Gene Ontology (GO) functional enrichment analysis to elucidate the biological significance of the upregulated and downregulated differentially expressed proteins in saliva exosomes, and the top 8 enriched terms in biological processes (BP), cellular components (CC), and molecular functions (MF) are illustrated in Figs. 2f and g. For BP, the upregulated proteins in the ESCC group were primarily enriched in pathways involved in defense responses to fungi and antibacterial humoral immunity. Research has shown that ESCC patients exhibit oral microbiota dysbiosis, with upregulated bacterial expression in the oral cavity, which may be related to the occurrence and development of esophageal squamous cell carcinoma [15–17]. The downregulated proteins were predominantly enriched in biological pathways related to tissue structure maintenance, immunoglobulin production, immune response mediator production, and immune response-regulating cell surface receptor signaling pathways. In terms of CC, the upregulated proteins in the ESCC group were significantly enriched in nucleosomes, DNA packaging complexes, and protein-DNA complexes, whereas the downregulated proteins were primarily enriched in immunoglobulin complexes. Furthermore, in MF, the upregulated proteins in the ESCC group were mainly enriched in chromatin structural constituents and protein heterodimerization activity, whereas the downregulated proteins were primarily enriched in antigen binding.

In our proteomic data analysis, we observed that the expression levels of immunity-related proteins (IGKV3-15, IGKV3-11, IGKV3-20, and IGKC) in saliva samples from patients with poorly differentiated tumors were significantly lower than those in patients with moderately differentiated tumors (Figure S1a). Conversely, the expression levels of histone family members (H4C1, H3-3A, H2AC20, and H2BC14) in saliva samples from patients with moderately and poorly differentiated tumors were significantly higher than those from patients with well-differentiated tumors (Figure S1b). Based on tumor stage, we observed that the expression levels of proteins IGLL5 and IGKC in saliva from patients with stage I ESCC were significantly lower than those in patients with stage II and stage III ESCC. Concurrently, the expression levels of IGHA1, IGHG1, GAPDH, S100A9, and DSG3 in saliva from stage I ESCC patients were significantly lower than those in stage III ESCC patients (Figure S1c and S1d).

Lipidomic analysis of exosomes from ESCC and CTRL saliva samples

Subsequently, we analyzed the lipidomic data of saliva exosomes. PCA revealed a distinct separation between ESCC patients and controls, with QC samples clustering together (Fig. 3a). We performed OPLS-DA multivariate statistical analysis on the lipidomic data and obtained the variable importance in projection (VIP) values for each lipid molecule. Figure 3b displays the top 15 lipid molecules with the highest VIP values. Using SIMCA software, we constructed an OPLS-DA model for ESCC and CTRL samples, which achieved clear group separation (Fig. 3c). The validation plot for the OPLS-DA model is shown in Fig. 3d, with Q2 intercepting the Y-axis below 0.05 and both fitting lines having a slope greater than 0, indicating that the model did not exhibit overfitting.

Fig. 3.

Lipidomic differences between ESCC and CTRL groups. a PCA score plot of lipidomic. b TOP 15 VIP lipids from the OPLS-DA model. c d OPLS-DA models for the differentiation of ESCC and CTRL groups. e Volcano plot of DELs. Green dots represent down-regulated lipids in ESCC, while red dots represent up-regulated lipids in ESCC. f Heatmap illustrating the expression profiles of 17 DELs between ESCC and CTRL groups. g Heatmap of correlation analysis of differentially expressed lipids. *p < 0.05, **p < 0.01, ***p < 0.001

Following the removal of lipids that were not expressed in more than 50% of the samples, a total of 2146 lipid molecules were identified. The fold change and FDR-adjusted p-value were calculated for each lipid, and the results were presented in a volcano plot (Fig. 3e). Based on the threshold filtering criteria (FDR < 0.05, |log₂FC|≥ 1), 17 lipids were found to exhibit significant differences, with 13 lipids being significantly upregulated and 4 lipids being significantly downregulated in the ESCC group. The heatmap displayed the expression profiles of the 17 differentially expressed lipids between the two groups (Fig. 3f). Notably, the upregulated lipids were primarily comprised of fatty acids, phosphatidylcholines, phosphatidic acids, and wax esters, whereas the downregulated lipids were mainly triglycerides and diglycerides. The expression levels of fatty acid lipid molecules in patient saliva exosomes did not show significant differences across different differentiation grades (Figure S1e). However, the expression levels of FA(18:0)_3.44 and FA(18:0)_3.44 in the saliva of stage III ESCC patients were significantly higher than those of stage II patients (Figure S1f).

Correlation analysis of proteomic and lipidomic

Our study unveiled distinct proteomic and lipidomic profiles in saliva exosomes from ESCC patients. To further elucidate the interactions between differentially expressed proteins and lipids, we performed a Pearson correlation analysis on paired samples and visualized the results using a heatmap (Fig. 4a). Subsequent analysis of the correlation results using Cytoscape software revealed a complex interaction network, where FA(16:0)_2.9, FA(16:0)_2.86, FA(18:0)_3.44, and FA(18:0)_3.48 emerged as core lipid molecules, exhibiting significant negative correlations with multiple immunoglobulins (IGLL5, IGLV7-46, IGKV3-11, etc.) and significant positive correlations with histone family members (H2AC20, H2BC14, H4C1, H3-3A) (Fig. 4b). Notably, certain adhesion protein family members (DSC2, GSG3) also displayed significant negative correlations with fatty acids, suggesting a potential interplay between these molecules in the ESCC microenvironment.

Fig. 4.

Correlation analysis of DEPs and DELs in saliva exosomes. a Heatmap depicting the correlation coefficients between DEPs and DELs in Saliva Exosomes. b Network diagram illustrating the correlations between DEPs and DELs. c Scatter plot showing the correlation between core lipids and proteins

Machine learning-based identification of ESCC patients

To assess the diagnostic potential of saliva exosomal proteins as biomarkers for ESCC, we employed a random forest (RF) model to distinguish ESCC patients from healthy controls. We divided the dataset into training and test sets at a ratio of 70:30, comprising 54 ESCC patients and 62 healthy controls. Initially, we constructed an RF model using all protein features, which achieved a sensitivity of 93% and specificity of 90% in the test set (Fig. 5c). Through tenfold cross-validation, we identified a subset of 7 optimal biomarkers, ranked in order of importance as ALB, TPI1, DSG3, KRT9, IGKV3D-20, IGKV3-11, and IGHG1 (Fig. 5a and Figure S2a). Heatmap illustrates the distinct expression profiles of these 7 protein molecules between the two groups (Fig. 5b). We further evaluated the RF model comprising these 7 proteins, which demonstrated 100% sensitivity and specificity in the training set, and 71% sensitivity and 90% specificity in the test set (Fig. 5d and e).

Fig. 5.

Diagnostic model for ESCC based on saliva exosomal proteomics features (a) Feature ranking based on importance values. b Proteins with the highest predictive values in classifying ESCC and CTRL samples by random forest algorithm. c Performance, confusion matrix, and ROC curve of the model built with all exosomal protein features in 30% test set. d Performance, confusion matrix, and ROC curve of the model built with 7 exosomal protein features from (a) in 70% training set. e Performance, confusion matrix, and ROC curve of the model built with 7 exosomal protein features from (a) in 30% test set

We applied a random forest classifier to the lipidomic data, adopting the same approach as for the proteomic analysis. The model, trained on all lipid features, achieved a remarkable 93% sensitivity and 100% specificity in the testing set (Fig. 6c). Following tenfold cross-validation, we identified the top 28 lipid features that yielded the minimum error rate (Fig. 6a and Figure S2b). The heatmap in Fig. 6b illustrates the distinct expression profiles of these 28 lipid molecules between the two groups. Notably, the RF model composed of these 28 lipids outperformed the model built with 7 proteins, achieving 100% sensitivity and specificity in both the training and testing sets (Fig. 6d and e), suggesting the potential of lipidomic markers for ESCC diagnosis.

Fig. 6.

Diagnostic model for ESCC based on saliva exosomal lipidomics features (a) Feature ranking based on importance values. b Lipids with the highest predictive values in classifying ESCC and CTRL samples by random forest algorithm. c Performance, confusion matrix, and ROC curve of the model built with all exosomal lipid features in 30% test set. d Performance, confusion matrix, and ROC curve of the model built with 28 exosomal lipid features from (a) in 70% training set. e Performance, confusion matrix, and ROC curve of the model built with 28 exosomal lipid features from (a) in 30% test set

Ultimately, we established a multi-omics diagnostic model by integrating protein and lipid features, which exhibited outstanding diagnostic performance. Notably, after performing tenfold cross-validation, we found that the top 32 features yielded the lowest error rate, and remarkably, all of these top 32 features were lipid-related (Fig. S2c and d), underscoring the paramount importance of lipidomic markers in accurately discriminating between ESCC patients and healthy controls.

Validating proteomic signatures to identify prognostic biomarkers

Validating the Differentially Expressed Proteins in this study using ESCC transcriptomic data from TCGA database. From Fig. 7a, we found that histone family members (H2ACA0, H2BC14, and H3-3A) were significantly upregulated in ESCC tumor tissues, consistent with the trend observed in saliva exosomes. However, some showed the opposite trend. For example, cadherin family members (DSG3, DSC2) and CALU were significantly upregulated in ESCC tumor tissues. SMARCAD1 was significantly upregulated in both tumor tissues and saliva exosomes from ESCC patients. Furthermore, the Cox hazards models revealed that high SMARCAD1 expression was associated with longer OS (Fig. 7b). Therefore, SMARCAD1 may be involved in regulating the ESCC disease course and guiding the prognosis of ESCC patients.

Fig. 7.

Validating proteomic signatures toidentify prognostic biomarkers (a) The expression of DEPs in this study using ESCC transcriptomic data from TCGA database. b Cox hazards models of SMARCAD1, CALU and H1-3. (Time-month)

Discussion

In this study, we identified unique protein and lipid features in the oral microenvironment of ESCC patients through non-targeted proteomics and lipidomics analysis based on liquid chromatography-tandem mass spectrometry. Furthermore, through multi-omics analysis, we found significant correlations between differentially expressed proteins and lipids, which may be related to the pathogenesis of ESCC.

Nearly all cell types release extracellular vesicles (EVs), which are involved in cancer initiation and progression [18]. EVs are abundant in various bodily fluids, such as blood, saliva, urine, and cerebrospinal fluid. ESCC cells can release exosomes into the bloodstream, modulating distant changes in the body [19]. A study of serum exosomes from 51 ESCC cases and 41 benign controls revealed an association between circulating exosomal miRNA-21 and ESCC progression [20]. Additionally, due to the anatomical relationship between the oral cavity and the esophagus, esophageal reflux may directly lead to alterations in the composition of salivary exosomes in ESCC patients. Furthermore, many researchers believe that oral microorganisms may be one of the causes of esophageal squamous cell carcinoma, influencing its progression by regulating the oral microenvironment [21].

In the saliva-derived exosomes of ESCC patients, we observed downregulated expression of immune-related proteins (IGKV2-28, IGKV3-11, IGLV7-46, IGKV3-20, IGLL5, IGHG1, IGKV3-15, and IGH A1), suggesting that the immune microenvironment in the oral cavity of ESCC patients may be suppressed. It has been reported that cancer cells can modulate immune cells, leading to the formation of an immunosuppressive microenvironment in the extracellular matrix, which can weaken inflammatory responses, facilitate immune evasion, and hinder the efficacy of immunotherapy [22]. This altered extracellular matrix can serve as a protective barrier for cancer cells. In contrast, we observed a significant upregulation of histone family members, including H2AC20, H2BC14, H4C1, and H3-3A, in the saliva-derived exosomes of ESCC patients. Histones, which comprise five main types (H1, H2A, H2B, H3, and H4), are nuclear proteins that play a crucial role in chromatin assembly, gene expression regulation, and genome stability maintenance [23]. Notably, aberrant histone expression and modification can lead to dysregulated gene expression, thereby promoting tumorigenesis and cancer development [24]. H2BC17 is a core histone protein, located in the cell nucleus, wrapped around DNA, and regulates chromosomal structure and gene expression, while dysregulation of histone H2B may be a potential underlying mechanism for the development of cancer [25]. Our study found that the upregulated proteins were enriched in the process of antimicrobial humoral response, consistent with other research findings, suggesting that microbes may be involved in the occurrence and development of ESCC [26]. Through 16S rRNA sequencing, we identified the genus Fusobacterium and Porphyromonas as potential high-risk pathogens for ESCC [27]. The presence of Fusobacterium in ESCC tumors may affect the efficacy of neoadjuvant chemotherapy [28].

In the saliva-derived exosomes of ESCC patients, we identified upregulated expression of lipid subclasses, including FA, PC, PA. Specifically, FA (16:0) and FA (18:0) are involved in fatty acid transportation and alpha-linolenic acid metabolism. Previous studies have shown that FA synthesis is increased in various types of cancer, with multiple cancer signaling pathways converging on FA synthesis [29, 30]. Cancer cells require an accumulation of lipids to form cell membranes and organelles, which is a unique characteristic of uncontrolled cell proliferation [31]. In ESCC patients, we found downregulated expression of TG molecules, including TG (P-7:0_11:0_18:3), TG (O-15:1_3:0_18:3), and TG (P-18:0_16:0). TG serve as the primary energy storage molecules in the body, which can be hydrolyzed by lipases to break down the three ester bonds, ultimately yielding glycerol and fatty acids, and are primarily involved in energy storage and supply processes, as well as lipid signaling transduction pathways [32]. Research has reported that TG plays a role in cancer through its lipase, which is upregulated in melanoma, ovarian cancer, and breast cancer, and that overexpressing lipase genes can promote cancer cell proliferation and invasion [33]. Additionally, a lipidomics study of plasma from liver cancer patients found that multiple TG molecules were downregulated compared to healthy controls. TG serves as a primary energy reserve in the body, which can be broken down into glycerol and fatty acids through lipase-mediated hydrolysis [34]. Our study revealed that increased FA synthesis and decreased TG and diglyceride (DG) levels are a characteristic feature of oral lipids in ESCC patients.

Through correlation analysis, we found that the expression of FA(16:0) and FA(18:0) in salivary exosomes of ESCC patients was significantly negatively correlated with the expression of immunoglobulins. Fatty acid metabolism is an important metabolic pathway in tumor development, playing a crucial role in the activation and function of immune cells [35]. Previous studies have shown that the immune activation of dendritic cells (DCs) in the tumor microenvironment (TME) is decreased, and lipid accumulation is the main reason for DC dysfunction [36]. Treatment of DCs with acyl-CoA cholesterol acyltransferase (ACC) inhibitors reduced lipid levels to normal, restoring DC function [36]. Meanwhile, we also found that FA(16:0) and FA(18:0) exhibit a significant positive correlation with the expression of histones. In the context of histone modifications in adipose tissue, the majority of studies have focused on histones H2 and H3 [37, 38]. Alterations in histones are related to the regulation of gene activities involved in lipid metabolism and play a crucial role in the formation and progression of diseases [39]. Our study found that the oral microenvironment of esophageal squamous cell carcinoma patients may influence lipid metabolism and immune changes through epigenetic modifications.

Some studies have reported the role of plasma tumor biomarkers in the diagnosis of esophageal squamous cell carcinoma. One study analyzed the diagnostic potential of plasma miR-21, miR-223, and miR-375 for ESCC, with AUC values of 0.80, 0.73, and 0.69 [40]. Using reverse transcription-quantitative polymerase chain reaction, the combined detection of 4 miRNAs (miR-21, miR-25, miR-145, and miR-203) in serum showed a sensitivity of 74.2% and a specificity of 88.0% for the diagnosis of ESCC [41]. Serum HOTAIR, a long non-coding RNA (lncRNA), is highly expressed in ESCC patients, with a diagnostic AUC of 0.793 [42]. A diagnostic model constructed using eight highly expressed miRNAs in serum achieved an AUC of 0.83, with a sensitivity of 78% and specificity of 75%, in a retrospective cohort study, and an AUC of 0.92, with a sensitivity of 89% and specificity of 84%, in a prospective cohort study [43]. Saliva samples possess non-invasive and easily accessible advantages in liquid biopsy. In a previous study on saliva-derived exosomal miRNA, a 6-miRNA signature achieved an AUC of 0.968 in identifying ESCC [44]. In this study, we constructed a random forest model using saliva-derived exosomal proteomics and lipidomics data, and performed tenfold cross-validation, successfully distinguishing ESCC patients from healthy individuals. Our findings revealed that the diagnostic performance of the model built with lipid features was superior to that of protein features, highlighting the potential advantages of lipid features in ESCC diagnosis. Furthermore, our established multi-omics diagnostic model combining protein and lipid features also confirmed the superiority of lipid features in esophageal squamous cell carcinoma diagnosis. However, it is necessary to validate our findings in different populations and larger sample sizes.

Notwithstanding the contributions of this study, we acknowledge several limitations that should be considered. Firstly, this study is a cross-sectional study with a relatively small sample size, which may limit its generalizability to a broader population. We included only 5 cases of stage I ESCC patients in our study, which resulted in a bias towards late-stage cases. Additionally, the lack of prognostic information in the sample data is another limitation. Therefore, in future studies, we plan to recruit more patients with esophageal squamous cell carcinoma from multiple centers and employ targeted proteomic and lipidomic analyses to support and validate the results of this study.

Conclusions

In summary, our study revealed unique proteomic and lipidomic alterations in salivary exosomes of ESCC patients, which can help deepen our understanding of the disease's pathogenesis. The potential biomarkers identified through the random forest model provide a promising avenue for developing non-invasive screening tools. In the long run, our study lays the foundation for larger-scale validation experiments, which can help identify new diagnostic targets for ESCC.

Authors’ contributions

Q. C. designed the study and was responsible for quality control. Z. Y. is responsible for the upstream analysis and quality control of proteomics and lipidomics. W. Z. recruited subjects, conducted the experiments, performed the statistical analysis, and drafted the manuscript. Jian Liu participated in subject recruitment and contributed to sample collection. J. X., Z. Z. and Z. G. assisted in sample preparation and experimentation. Q. C. serves as the guarantor and is responsible for the overall content.

Data availability

Data is provided within the supplementary information files. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD062319.

Declarations

Ethics approval and consent to participate

The study was conducted following the Declaration of Helsinki and reviewed and approved by the Ethics Committee for Clinical Research of the Fifth Affiliated Hospital of Sun Yat-sen University (Ethics No. 2022-K54-1). This study was approved by informed consent from all study participants.

Competing interests

The authors declare no competing interests.