¹H-NMR spectroscopy identifies potential biomarkers in serum metabolomic signatures for early stage esophageal squamous cell carcinoma

Study subjects and sample collection

Serum samples (65) were collected from the Department of Cardiothoracic Surgery and the Medical Examination Center in the Second Affiliated Hospital of Shantou University Medical College, between December 2016 and June 2018. These samples included those from 40 CTRL and 25 patients in the early stages of ESCC (stage I/II) not treated with chemotherapy, radiotherapy, or chemoradiotherapy. Patient information and clinical characteristics are summarized in Table 1. The clinical stages of ESCC patients were diagnosed using esophagoscopy examination with biopsy, X-ray barium radiography, and chest computed tomography. Tumour staging was based on the American Joint Committee on Cancer’s (AJCC) 7th staging system. CTRL, recruited from our health examination center, were matched with patients with ESCC based on age, gender, BMI, and place of residence. This study was approved by the Ethical Committee of Second Affiliated Hospital of Shantou University Medical College (Registration No. 2016-32) and carried out in accordance with the Declaration of Helsinki. Informed consent was obtained from each subject before participation in the study.

Sample preparation and ¹H NMR spectroscopy

About five mL of peripheral venous blood was collected from each subject between 7 and 8 am into tubes without any anticoagulant, and centrifuged at 3,000 rpm for 10 min. Serum samples were isolated and immediately stored in EP tubes for further analysis. Anticoagulation can reduce or inhibit the activity of protein and enzymes so we did not denature or filter them during NMR acquisition. Then, in order to avoid potential disturbance to the metabolic status during the denaturation or filtration processes, we acquired the serum sample and stored it at −80 °C immediately. In order to reduce the influence of the proteins in blood on the NMR acquisition, Carr-Purcell-Meboom-Gill (CPMG) pulse sequence was used in our study. This pulse sequence can suppress signals from macromolecules and other molecules with constrained molecular motion, although not completely. CPMG is also a common method to acquire NMR spectrum of blood samples and has aided in the metabolomic studies of many kinds of clinical disease (Zhang et al., 2013a; Zhang et al., 2013b; Chen et al., 2015). Frozen serum samples were thawed immediately and vortexed for 10 s at room temperature. Then, 200 µL of phosphate buffer solution (90 mM NaH₂PO₄/Na₂HPO4, pH = 7.4) was added to 400 µL serum for NMR detection. After centrifugation at 10,000 rpm for 10 min at 4 °C, approximately 550 µL of the clear supernatant was transferred into five mm NMR tubes (ST500, NORELL, Inc., Morganton, NC, USA) for sampling. An NMR spectrometer (600.13 MHz, Bruker Avance III, Bruker Corporation, Kalsruhe, Germany) was used to obtain ¹H-NMR spectra under CPMG pulse sequence with a total spin relaxation delay (2 nτ) of 70 ms to weaken broad resonances from high molecular weight compounds and to retain low molecular weight compounds and some lipids. All serum samples were analysed in random order at 298 K. One-dimensional spectrum was used to obtain CPMG spin echo pulse sequence (RD-90°-(τ-180°-τ)n-ACQ) to suppress water signal with a relaxation delay of 5 s. The acquisition parameters were: spectral width, SW = 20 ppm; recycle delay, RD = 4.0 s; t1 = 350 μs; mixing time, tm = 100 ms; number of scans, NS = 32; number of points, TD = 32,768; and acquisition time, AQ = 2.73 s.

¹H NMR spectroscopy analysis

The raw data (free induction decays, FIDs) were input into MestReNova Version 9.0.1 (Mestrelab Research, Santiago de Compostela, Galicia, Spain, 9.0.1) for processing and complexity reduction to facilitate pattern recognition. To enhance the signal-to-noise ratio, all ¹H-NMR spectra were multiplied by a 1.0 Hz exponential line broadening prior to Fourier transformation. The chemical shifts of serum spectra were referenced with the methyl doublet signal of lactate at δ 1.33 ppm as an internal standard. Both phase adjustment and baseline correction were performed manually. Each spectrum (from 9.0 to 0.5 ppm) was segmented into rectangular buckets with equal widths of 0.002 ppm with each excluding residual water (from 5.18 to 4.67 ppm) and urea (from 6.40 to 5.40 ppm) regions. To remove the dilution effect or bulk mass differences among samples due to differing serum weights, the remainder of each bucket was internally normalized to the total sum of the spectral integrals for each compound prior to pattern recognition analysis. In our study, one bin was only 1.2 Hz, therefore, no more than one compound existed in each bucket. In the usual case of more than one bucket per compound, each bucket was statistically analyzed in multivariate statistical analysis, but the sum of the integrals in any peak (not bucket) of some compound in NMR spectra was used for univariate statistical analysis.

Pattern recognition (PR) analysis of serum

To establish a global profile of the feature in ESCC patients and CTRL, we used multivariate analysis to identify consistent variations between ¹H NMR data sets. Serum spectra data were input into the SIMCA-P+ version 14.0 software package (Umetrics Inc., Umea, Sweden, V 14.0) for PR analysis. First, PCA, an unsupervised PR method, was performed using the Parato-scaled normalized ¹H NMR spectra to discover the intrinsic trends and outliers between the two serum sample groups. Then, PLS-DA and OPLS-DA, the two supervised PR methods, were performed to prevent over-fitting of the statistical model and to select potential biomarkers.

Model quality and reliability were assessed using R²X, R²Y, and Q² values, which reflect the explained variance and model predictability. R²X represents the variation explained by the models and R²Y indicates the ‘goodness of fit’ in the data. Q², calculated by a cross-validation procedure, indicates the predictability of the model. To avoid model overfitting, a default seven-round cross-validation procedure was performed in SIMCA-P+ 14.0 to determine the optimal number of principal components. Reliability of the models was further rigorously validated by a permutation analysis (n = 300 times). The variable importance in the projection (VIP) from OPLS-DA models was identified as a coefficient for peak selection. These variables were considered potential biomarker candidates based on class discriminating information –the higher the value, the greater the discriminatory power of the metabolite. VIPs larger than 1.0 usually represent those metabolites with the greatest group discrimination.

Data preprocessing statistical analysis

CV-ANOVA (analysis of variance testing of cross-validated predictive residuals) was performed to identify significantly different features between groups in OPLS-DA models. Univariate statistical significance of p <0.05 was considered to distinguish metabolites. Student’s t (normal distribution) or Mann–Whitney U (if abnormal distribution) tests were performed to analyse the metabolic profiles between ESCC patients and CTRL. The metabolites were recognized according to the Human Metabolome Database (http://www.hmdb.ca/) and receiver operating characteristic (ROC) analysis was conducted in SPSS 19.0 (SPSS Inc., Chicago, IL, United States) to verify potential biomarkers. Area under the ROC curve (AUROC) of >0.80 indicated excellent diagnostic ability.

Metabonomic profiling of serum samples for ESCC patients and CTRL

The one-dimensional ¹H-NMR spectra of serum samples provided an overview of all metabolites from ESCC patients and CTRL (Fig. 1). We did not try to absolutely quantify the metabolites in serum samples by comparing them with an internal standard such as DSS or TMSP. In fact, no current NMR standard is readily available for serum. DSS, TMS or TMSP are not suitable as an internal standard for proteinaceous fluids (such as serum and plasma) due to their potential protein binding capacity (Shimizu, Ikeguchi & Sugai, 1994; Alum et al., 2008). The chemical shift and line-width of DSS or TMSP will show large variation due to protein binding, thus greatly affecting the accurate qualification of the metabolites in serum. In our study, endogenous lactate (δ 1.33 ppm) in serum samples was used as an internal standard reference, and the relative concentrations of the metabolites were calculated and statistically discriminated by integrating their signals in the NMR spectra. Ultimately, approximately 47 metabolites were tagged in the spectra between the two groups, including amino acids, organic acids, energy metabolism, methylguanidine, myo-inositol, trimethylamine N-oxide, glucose components, lipids, carbohydrates, nucleotides and so on.

PCA, PLS-DA, and OPLS-DA pattern recognition analysis of serum metabolomic profiling for ESCC patients and CTRL

To obtain useful metabolomic profiles, unsupervised PCA analysis of ¹H-NMR results showed the difference between ESCC patients and CTRL. Pareto scaling, performed by dividing the mean-centered data by the square root of the standard deviation (SD), was applied to the variables. The two-dimension PCA score plots and loading plots revealed separation trends and group clustering based on ¹H-NMR spectra of the two groups (R²X(PC1+PC2) = 43.3%) (Figs. 2A–2B). However, we were unable to identify a very clear difference between them via PCA scores plot. We performed supervised PLS-DA and the resultant plot showed different metabolic perturbations between the groups using 2D score plots (R²X = 51.9%, R²Y=80.9%, Q² = 81.4%) (Fig. S1). To maximize the group separation and to visualize the metabolic distinctions, the OPLS-DA classification model was performed to investigate metabolomic alterations. We used the OPLS-DA methods to distinguish between ESCC patients and CTRL obviously (Fig. 3 A-D). Taken together, these results suggest that the models were robust and the random permutation tests (300 iterations) indicated that those models were not over-fitted (R²X = 81.9%, R²Y = 95.4%, Q² = 94.0%, CV-ANOVA p = 2. 23 ×10⁻³⁵). Our results indicate that ¹H-NMR-based serum metabolomics have potential application in identifying early ESCC.

Discovery, description, and identification of potential biomarkers and a biomarker panel

A total of 31 differential metabolites were identified as characteristic metabolites (Table 2). We constructed a map of hierarchical cluster analysis (Fig. 4) to visualise the distinction power of biomarkers between the two groups. The rows represent the results of the expression of metabolites, and the columns show serum samples. In the top bar, the light blue color indicates CTRL individuals, and the pink color indicates ESCC patients. Figure 4 indicates that the metabolite profile could distinguish ESCC patients from CTRL. Based on the sensitivity and specificity of this approach, ROC analyses were performed for further prediction of potential biomarkers. Neoplastic diseases involve systematic disturbance of metabolic biochemical pathways. Therefore, a biomarker panel including multiple biomarkers, rather than a single biomarker, can better distinguish the different groups and supply useful information for clinicians. We identified a panel of six biomarkers, glucose, choline, glutamine, glutamate, valine, and dihydrothymine (DHT), that were combined together by binary logistic regression with high AUC values of 0.969 (Fig. 5). Changes in these metabolite biomarkers could be related to tumour burden. In particular, the decline of DHT (Fig. 6, p < 0.01), a new feasible biomarker, is associated with tumor invitation.