Untargeted metabolomics analysis of esophageal squamous cell cancer progression

We collected a total of 75 esophageal tissues, including 15 samples of esophageal squamous cell cancer (ESCC) TNM stage I, 15 samples of ESCC TNM stage II, 15 samples of ESCC TNM stage III, 15 samples of ESCC TNM stage IV, and 15 samples of normal tissues adjacent to the tumor in this study. All esophageal samples were from the Frist Affiliated Hospital of Chongqing Medical University, China, from January 2010 to December 2019. The characteristics of these patients were shown in Table 1. The patients were diagnosed with ESCC by preoperative gastroscopy and subsequently recruited for the study. The esophageal tissues were acquired during the gastroscopic biopsy or surgery and were applied for pathologic biopsy. The tumor segments enrolled meet the following criteria: viable tumor nuclei > 80%, total cellularity > 50%, and necrosis < 20% [21, 22]. Normal tissues adjacent to the tumor were acquired by other 15 ESCC patients, all the patients had no illness of esophagitis, acid reflux, or gastritis, and no patients received the chemoradiation therapy before the surgery. All included patients signed informed consent before they participated in the study. The study was implemented in terms of the Declaration of Helsinki, and the study was approved by the Ethics Committee of the First Affiliated Hospital of Chongqing Medical University. All tissue samples were immediately stored at frozen.

We performed sample preparation and extraction as previously described [23]. We weighed 25 mg of the sample in an EP tube and added 500 μL of extraction solution (acetonitrile: methanol: water = 2:2:1, a standard internal mixture with isotope labeling). After vortexing for 30 s, we homogenized the samples at 35 Hz for 4 min and then sonicated on ice for 5 min. We repeated the homogenization and sonication cycle three times. The samples were then incubated at − 40 °C for 1 h and centrifuged at 12,000 rpm at 4 °C for 15 min. The supernatant was transferred to a fresh tube for analysis. We prepared quality control (QC) samples by mixing aliquots of the supernatant from all samples.

LC–MS/MS analyses were performed using a UHPLC system (Vanquish, Thermo Fisher Scientific) and a UPLC BEH Amide column (2.1 mm × 100 mm, 1.7 μm) combined with a Q Exactive HFX mass spectrometer (Orbitrap MS, Thermo). The mobile phase consists of 25 mmol/L ammonium acetate and 25 ammonium hydroxide aqueous solution (pH = 9.75) (A) and acetonitrile (B). The analysis was performed with an elution gradient as follows: 0–0.5 min, 95% B; m/z. 0.5–7.0 min, 95–65% B; 7.0–8.0 min, 65–40% B; 8.0–9.0 min, 40% B; 9.0–9.1 min, 40–95% B; 9.1–12.0 min, 95% B. The column temperature was 25 °C. The temperature of the auto-sampler was 4 °C, and the injection volume was 3 μL. The QE HFX mass spectrometer was used because it can acquire MS/MS spectra in information dependent acquisition (IDA) mode under the control of acquisition software (Xcalibur, Thermo). In this mode, the acquisition software continuously evaluates the full scan MS spectrum. The ESI source conditions were set as follows: sheath gas flow rate was 50 Arb, the auxiliary gas flow rate was 10 Arb, the capillary temperature was 320 °C, full MS resolution was 60,000, MS/MS resolution was 7500, collision energy was 10/30/in NCE It is 60 in the mode, and the spraying voltage was 3.5 kV (positive) or − 3.2 kV (negative).

We used proteowizard (http://​proteowizard.​sourceforge.​net/​) [24] to convert the original data into mzXML format and used an internal program for processing, which was developed using R and based on package XCMS (version 3.7.1) for peak detection, extraction, alignment, and integration. Then the internal mass-spectrometry 2 (MS2) database (BiotreeDB) was applied to metabolite annotation. The cutoff value of the annotation was set to 0.3.

In this study, principal component analysis (PCA) and orthogonal projection to least squares discriminant analysis (OPLS-DA) were utilized to simplify the metabolomic data. Therefore, after mean-centering and scaling, the UHPLC data were set to default unit variance, multivariate statistical analysis [25‐27] was conducted via SIMCA-P version 16.0 software package (Umetrics Umeå, Sweden). First, we performed unsupervised PCA to observe the inner clusters and find apparent outliers. Then, supervised OPLS-DA was used to distinguish the ESCC samples from the adjacent normal controls visually. The OPLS-DA model eliminates variability unrelated to class separation. The quality and reliability of the model were evaluated by the parameters R²X, R²Y and Q². R²X and R²Y represent the explained data change and R²X indicates the goodness of the fit, R²Y indicates goodness of prediction, while Q² is a sevenfold cross-validation parameter and estimates the predictive ability, with aggregate (cum) values of R2X, R2Y and Q2Y equating to ~ 1 showing a valid model. The cumulative value of the total explained value R2X, R2Y and the predictable change of Q2 suggest that the modeling is correct. To prevent over-fitting, a permutation test (n = 200) in the SIMCA-P software package was applied to the OPLS-DA model [28, 29]. Based on the variable importance of the prediction (VIP) threshold from the OPLS-DA model, the metabolites responsible for ESCC differential from adjacent normal esophageal metabolites can be identified. The Mev (MultiExperiment Viewer) 4.8 software was used to execute one-way ANOVA with standard Bonferroni correction to correct the resultant p values for each metabolite in all cross-comparisons. By VIP > 1.0 and adjust. p value (pFDR) < 0.05 (confidence level), all ESCC TNM stages was screened for significantly different metabolites when compared with the adjacent normal esophagus[30]. The pheatmap package and corrplot package in R 3.6.3 were used to draw heatmap and correlation matrix among ESCC TNM stage I vs. con., stage I vs. stage II, stage II vs. stage III, and stage III vs. stage IV [31, 32].

According to the Kyoto Encyclopedia of Genes and Genomes (KEGG) metabolites compounds database [33], we annotated verified metabolites of ESCC TNM stage I vs. con., stage I vs. stage II, stage II vs. stage III, stage III vs. stage IV and then matched the annotated metabolites with the KEGG pathway database. We fed the most significantly regulated metabolite pathway into gene set enrichment analysis (GSEA) [34]. We next analyzed the metabolic pathway genes mRNA expression between ESCC and normal esophageal tissues in public datasets: gene expression data series (GSE)23400 and The Cancer Genome Atlas (TCGA) [35, 36]. Receiver operating characteristic (ROC) curves applied to assess the ESCC progression predictive value by GSE23400 dataset [37].

From January 2010 through December 2019, 75 esophageal tissues were collected including esophageal squamous cell carcinoma (ESCC) and normal tissues adjacent to the tumor. Compared with either ESCC tumor/node/metastasis (TNM) stages and normal controls, more than 70% ESCC patients were male, almost 80% ESCC located at middle and lower esophagus. 65% ESCC Stage II patients were older than 65, and ESCC Stage IV patients had lowest BMI among all groups (IQR, 17.6–19.2). Characteristics of patients see Table 1.

To assess all ESCC TNM stages and adjacent cancerous tissues metabolic profiles, we measured esophageal tissues via LC–MS/MS. Finally, 712 metabolites over 75 classes were identified, including 145 Glycerophospholipids, 124 Carboxylic acids and (its) derivatives, and 84 Fatty Acyls et al. (Additional file 1 Identified Metabolites: Table S1). Phosphatidylethanolamine (PE) and phosphatidylcholine (PC) were mostly Glycerophospholipids species. The data that support the findings of this study have been deposited into MetaboLights of EMBL-EBI with MTBLS3579 [38].

Principle component analysis (PCA), and orthogonal partial least squares discriminant analysis (OPLS-DA) as data mining methods are used to build multivariate models to discriminate metabolomic profiling among ESCC TNM stages and adjacent cancerous tissues [39]. PCA score plots showed a clear trend of group clusters between the ESCC patients and normal controls (Additional file 1 PCA: Fig. S1). Additionally, to exclude variables with smaller correlations, a supervised OPLS-DA classification model using one PLS component and one orthogonal component was established. The OPLS-DA score plots obtained even clearer class discrimination (Fig. 1A). Goodness of fit (R2 X and R2Y) of ESCC TNM stages versus controls were 0.479 and 0.912, and Q2 of OPLS-DA was 0.815. These results indicated 712 metabolites were well explained by OPLS-DA models. To validate the OPLS-DA models, random permutation tests with 200 permutations were performed (Additional file 1: Validation plots. Fig. S2). Decrease of Q² and R² was observed along with the decrease of X-axis value, suggesting the model did not overfit. The PCA and OPLS-DA plots showed good discrimination between all ESCC TNM stages and normal controls.

To differentiate specific metabolites among all ESCC TNM stages vs. adjacent normal controls, OPLS-DA results were used to screen all metabolites with significant differences. The Variable Importance in Projection (VIP) obtained from OPLS-DA reflects both the loading weights for each sample and the metabolite of the response explained by this sample and can be used for metabolite selection. In this study, the metabolites concentration VIP value (VIP > 1) combined with the metabolite’s concentration adjust p-value (pFDR < 0.05) was used to screen the crucial metabolites. As a result, 145 metabolites had significant differences among all ESCC TNM stages vs. Con.; the ESCC TNM stage I vs. Con. group had 151 metabolites with significant differences; the ESCC TNM stage I vs. stage II group had 100 metabolites with significant differences; the ESCC TNM stage II vs. stage III group had 144 metabolites with significant differences; and the ESCC TNM stage III vs. stage IV group had 120 metabolites with significant differences. Heatmaps were plotted using z-score among all ESCC TNM groups and normal control group.

Pearson correlation coefficient analysis was applied for metabolite-metabolite correlation analysis via z-score in all ESCC TNM Stages vs. Con. This analysis identified metabolites that associated with each other in esophageal carcinoma and normal controls. Specifically, we compared metabolite correlations between each pair of samples (stage I vs. con., stage I vs. stage II, stage II vs. stage III, and stage III vs. stage IV), and the metabolite-metabolite correlations of these three sample combinations showed unique profiles. Metabolites with correlation coefficients p < 0.1 was identified as significantly correlation (Fig. 2A–D).

To validate ESCC progression related metabolic pathway, significantly differential metabolites in all ESCC TNM stage vs. Con were retrieved from the KEGG compound database. The results of the KEGG pathway enrichment analysis were displayed 16 pathways in either ESCC TNM stages and adjacent normal control via adjust p-value (pFDR < 0.05), metabolites count, and enrichment ratio. Among all ESCC TNM stage vs. Con. groups, the Glycerophospholipid metabolism pathway was the most significantly distinct metabolic pathway (Fig. 2E–H).

We obtained the gene set of Glycerophospholipid metabolism genes from the GSEA. A total of 77 genes were retrieved by Glycerophospholipid metabolism (Additional file 1 mRNA Expressions of Glycerophospholipid Metabolism Genes: Table S2). Moreover, we analyzed the 77 genes mRNA expression in normal esophageal tissues and all stages of ESCC tissues in GSE23400 and TCGA dataset and found 16 genes had significantly differential expression (Fig. 3A–Q). Then 16 significantly differential genes of Glycerophospholipid metabolism were performed ROC test from GSE23400. Phosphatidylserine Synthase 1 (PTDSS1) (AUC = 0.980) and Lysophosphatidylcholine Acyltransferase 1 (LPCAT1) (AUC = 0.914) showed a good prediction of ESCC in Fig. 3R.