Free amino acid profiling of gastric juice as a method for discovering potential biomarkers of early gastric cancer

Abstract

Background: Prior studies showed that aromatic amino acids (AAAs) could be used as potential gastric juice biomarkers in screening gastric cancer (GC). To identify new biomarkers for early diagnosis of GC, the characteristics of gastric juice free amino acid (GJFAA) profiling was determined. Method: First, gastric juice was collected from 130 consecutive patients who underwent gastroscopy. They were divided into GC group (n = 47) and non-neoplastic gastric disease (NGD) group (n = 83) according to the pathological diagnosis. The concentrations of 34 GJFAAs were examined by amino acid analyzer. Multivariate and univariate analyses were used for comparing the alterations of GJFAA profiles between the two groups. Then candidate differential GJFAAs were verified by LC-MS/MS in another set of patients, which included 32 GC patients and 38 NGD patients. The diagnostic performance of GJFAAs was evaluated by ROC curve. Results: Significant alterations in GJFAA profiles were observed in GC patients compared to NGD patients in the training set. A total of 14 amino acids were screened as differential GJFAAs. Leucine, threonine and serine were the most frequently altered. Combined AUC of the three non-AAAs [0.869 (95% CI, 0.805-0.934)] was superior to the combined three AAAs [0.841 (95% CI, 0.773-0.908)]. In addition, a combined AUC comprisingthe six ones was further improved to 0.871 (95% CI, 0.809-0.933) in the diagnosis of GC. A similar variation trend and diagnostic value were observed in the validation set. Conclusion: This study indicates the potential of GJFAA profiling as a promising approach for the early detection and screening of GC.

Introduction

Gastric cancer (GC) is the fifth most common cancer and third leading cause of cancer-related deaths worldwide [1]. The initial phase of GC remains relatively asymptomatic as most patients have already been in the progressive state at the time of diagnosis, and have lost the chance of receiving radical surgery. Therefore, early diagnosis plays a vital role in the improvement of curative rate of GC. Nowadays, clinical diagnosis still relies on endoscopy and histology, whereas GC can be easily missed during conventional endoscopy, especially early gastric cancer (EGC) due to a minor lesion, endoscopy-related inadequate biopsy sampling or incorrect interpretation [2]. Thus, population-based screening by reliable biomarkers is desperately required. However, the poor sensitivity and specificity of conventional tumor markers, like carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9) make it difficult to diagnose GC, while screened EGC alone [3]. Accordingly, more effective biomarkers are urgently needed.

Recently, utility of metabolism-based technologies opened a new window to discover potential biomarkers for metabolic disorders especially malignant diseases, given that pathological alterations in cell functions are frequently accompanied by metabolic reprogramming [4]. Nevertheless, some practical limitations still remain. These include the necessity to measure a huge number of metabolites, data-redundancy problems, including the false-discovery rate and overfitting, and cost constraints. One approach to overcoming these problems is “focused metabolomics”, which limits the objects of the analysis to those that play roles in general metabolism and share physical similarities [5]. Amino acids are among the most suitable candidates as they are either ingested or synthesized endogenously and play essential physiological roles both as basic metabolites and metabolic regulators [6,7]. To measure amino acids, e.g., plasma free amino acids (PFAAs), which abundantly circulate as a medium linking all organ systems, which are investigated in numerous studies have noted that increased levels of PFAAs in several cancer entities (such as lung, breast, cervix, kidney and digestive organs) compared to normal counterparts [8-15]. However, some investigations have highlighted that a fair number of PFAAs reflecting metabolic disruptions common to many cancers, rather than attacking unique metabolic features in any particular one [16,17]. Gastric juice has direct contact with the gastric epithelium, and it contains metabolic information of gastric epithelial cells. Thus, gastric juice free amino acids (GJFAAs) may be more specific than PFAAs for the diagnosis of GC. But there are very few studies that cover the application of GJFAAs to GC.

Our previous studies have established several endogenous fluorescence spectra of gastric juice for diagnosing and screening of GC [18,19]. Three fluorescent substances (aromatic amino acids, AAAs) in gastric juice were isolated and qualitatively identified using liquid chromatography-tandem mass spectrometry (LC-MS/MS) and (1)H-nuclear magnetic resonance ((1)H-NMR) [20]. AAAs in gastric juice can be applied to distinguish advanced gastric cancer and EGC from non-neoplastic gastric disease (NGD) through quantitative measurement using high-performance liquid chromatography (HPLC) [21]. Lately, we have built and tested a prediction model based on the male-to-female ratio, pH values, and AAA concentrations in gastric juice for diagnosing GC [22]. Interestingly, the results revealed that the level of multi-index (gastric juice index, GJI) begins to rise at the early stage of gastric carcinogenesis. It suggested that GJI has great potential for improving the early detection of GC. However, this work has focused on three AAAs that can produce autofluorescence, and the characteristic signature for the GJFAA profiling has remained poorly understood yet. Hence, this study outlined the metabolic patterns of GJFAA profiles in GC and NGD patients. Then, three most discrepancy GJFAAs were screened out and their diagnostic efficiencies were compared to and combined with three AAAs in gastric juice we focused on before. Finally, the analysis of differential GJFAA metabolic pathways was performed.

Patients and methods

Study design and ethical considerations

This study protocol was approved by the Peking University Third Hospital Medical Ethics Committee (IRB00006761-2016058). Informed consent was obtained from each patient, and the entire clinical investigation was conducted according to the Declaration of Helsinki. Gastric juice samples were collected from consecutive patients who underwent gastroscopy at our institution from December 2015 to April 2017. Altogether 200 patients were included in our study. Based on the sequence of the patients enrolled in the study, the first 130 cases were taken as the training set to compare the differences of GJFAA profiles between the GC and NGD groups and the latter 70 as the validation set to verify the diagnostic performance of the candidate biomarkers screened from the GJFAA profiles. The demographic and clinicopathological data were taken. The flow diagram of this study was presented in Figure 1.

Figure 1.

Flow diagram of the study. GC, gastric cancer; NGD, non-neoplastic gastric diseases; GJFAAs, gastric juice free amino acids; AAAs, aromatic amino acids; non-AAAs, non-aromatic amino acids; ROC, receiver operating characteristic curve; PCA, principal component analysis; OPLS-DA, orthogonal projections to latent structures discriminant analysis; LC-MS/MS, liquid chromatography-tandem mass spectrometry.

Inclusion criteria

Patients (1) suspected with various benign and malignant gastric diseases as diagnosed and confirmed by mucosal biopsy and/or postoperative pathology; (2) > 18 years of age; (3) who had not received any chemoradiation therapy.

Exclusion criteria

Patients (1) whose gastric juice was contaminated by blood, bile, or gross food residue; diluted in the process of gastroscopy, where < 2 mL was insufficient to conduct the experiment; (2) with recurrent, metastatic, neuroendocrine carcinoma, or lymphoma in the stomach; (3) with gastrointestinal submucosal tumor or gastric polyp; (4) with lesions that predominantly localized in the esophagus or duodenum; (5) with organic diseases or concurrent neoplasms diagnosed by clinical examination; (6) who are pregnant or lactating.

Diagnostic criteria

All patients were histologically confirmed by mucosal biopsy and/or postoperative pathology. Biopsies were obtained from the antrum, corpus and other suspicious sites (≥ 2). Abnormal lesions were sampled for histopathological examination. Diagnosis of each subject depended on the most severe lesion. In case of any discrepancy in pathological diagnosis between biopsy and postoperative specimens, the diagnostic results were mainly based on the latter one. GC was confirmed and classified according to the American Joint Committee on Cancer guidelines in 2010 [23]. The diagnosis of gastric dysplasia and chronic gastritis were according to the Padova International Classification [24] and the updated Sydney system [25], respectively.

Sample collection, preservation, and preparation

Gastric juice (5-10 mL) was collected from the patients who underwent gastroscopy after overnight fasting. The samples were centrifuged at 3000 rpm, 4°C for 10 min. The supernatant was then preserved in 2 mL aliquots at -80°C for subsequent analysis.

GJFAA profiling analysis

The concentrations of 34 amino acids and biogenic amines in gastric juice were targeted measured by Sykam automatic amino acid analyzer S 433-D (Sykam GmbH, Eresing, Germany). It is equipped with refrigerated reagent organizer S 7130, autosampler S 5200, solvent delivery system S 2100, amino acid reaction module S 4300, and Clarity Amino data analysis software. Additionally, amino acid standard solution, ninhydrine, washing and regenerate solutions were all provided by the manufacturer. The composition of amino acid standard solution was shown in Table S1, and concentrations of them were 1.0 µmol/mL, except for urea which was 10 µmol/mL.

Gastric juice samples were thawed in a 4°C water bath. The cryopreserved gastric juice of 0.5 ml was added and mixed with 8% sulfosalicylic acid of 1.0 ml, and then centrifuged at 4°C at 12000 rpm for 15 min to precipitate proteins. The supernatant of 0.2 ml was transferred to 10 mL glass tube, evaporated to dryness. The sample was dissolved in 2 mL diluent solution, 100 µL was drawn and filtered by 0.22 µm filter unit (Millipore Corp. Carrigtwohill, Co.Cork, Ireland), which was then transferred to the ampere bottle for detection.

Volumes of 50 µL of each sample were injected into the HPLC system for analysis. Each standard was individually run on the gradient noted in the LC parameters program (reaction temperature 130°C, gradient elution analysis). A LCA K07/Li column (4.6 × 150 mm, 7 um) was used for the separation. The flow rate for ninhydrine solution was 0.25 mL/min, and the flow rate for mobile phase was 0.45 mL/min. The run time for one sample was 110 min, and the equilibration time for the column oven was 30 min. Further details like the gradient program steps were shown in Table S2. Integrated dual-channel photometer at 440 nm and 570 nm wavelengths for GJFAA detection was used. The chromatograms thus obtained were compared with the standards to record the areas of the peaks and used for calibration.

LC-MS/MS analysis

Chemicals and reagents

Leucine, threonine, serine, tyrosine, phenylalanine, tryptophan (purity 99.9%) and their isotope-labeled internal standards (IS) (purity 98.0%) were supplied by Beijing MS Medical Research Co.,Ltd. HPLC-grade formic acid was obtained from Dikma (Lake Forest, USA), methanol and acetonitrile (HPLC grade) were obtained from Fisher Scientific (Fair Lawn, NJ, USA). Water purified by Milli-Q system (Millipore, Bedford, MA, USA) was used for the analysis.

The chromatography system consisted of a Dionex Ultimate -3000 Reagent-Free^TM Ion Chromatograph (Dionex Corporation, Sunnyvale, CA, USA) with an AS auto sampler. The chromatographic system was coupled with an API 3200 Q TRAP instrument (AB SCIEX, USA) operated in a multiple reaction monitoring mode (MRM). Data were acquired and processed using Analyst v1.5.2 software (Applied Biosystems).

Sample treatment

The gastric juice specimens were thawed at room temperature. Before the chromatographic analysis, 50 µL samples were deproteinized by adding 50 µL methanol (contains norvaline), instantaneous centrifugation at 2000 rpm for 10 s was done, followed by centrifugation at 13200 rpm, 4°C for 4 min. 10 µL of the supernatant was further diluted with 50 µL of buffer solution (contains isotope-labeled IS), and then was instantaneously centrifuged at 2000 rpm for 10 s. Added 20 µL derivatization liquid, instantaneous centrifugation at 2000 rpm for 10 s, followed by derivatization at 55°C for 15 min was done. The sample should be cooled in the refrigerator and then 5 µL was injected into the LC-MS/MS system.

Chromatographic conditions

A Phenomenex synergi MSLab45AA-C18 column (5 μm, 4.6 mm’ 150 mm) was used for separation of the samples. The column oven was maintained at 50°C. LC separation was performed using a mobile phase consisting of 0.1% formic acid in water (mobile phase A) and 0.1% formic acid in acetonitrile (mobile phase B). The gradient employed was as follows: 0-1 min maintained at 90% A, 1-12 min linear decrease from 90-30% A, 12-12.1 min decrease from 30%-0 A, and hold at 0% for 2.9 min, 15-15.1 min increase from 0-90% A, and hold at 90% A for 4.9 min. The flow rate was 1.0 mL/min, and the run time for one sample was 20.0 min.

The LC-MS/MS data acquisition for the above six amino acids was conducted in a positive ionization mode. The ion source parameters included were collision gas (CAD medium), curtain gas (CUR 20), ion source gas 1 (GS1 55), ion source gas 2 (GS2 60), and temperature (TEM 500°C). The product ions were monitored in a single reaction monitoring mode. Detection of other parameters were summarized in Table S3.

GJFAA quantification

Standard solutions that contained equivalent concentrations of the six amino acids were prepared by diluting the stock solution using deionized water at 2.5 µmol/L, 12.5 µmol/L, 25.0 µmol/L, 50.0 µmol/L, 100.0 µmol/L, and 200.0 µmol/L for building the calibration curves and were preserved at -20°C. Six-point calibration curves were used and plotted with the peak area ratio using a weighted (1/x²) quadratic fit.

Calibration curves, linearity, and retention time of the quantification system

The calibration curves exhibited excellent linearity for the six amino acids standards, with low interference and high sensitivity. The retention time, linear range of concentration and linear fitting coefficients of them were concluded in Table S4. If some of them fell beyond the measurement range, the samples were re-diluted before injection. In extremely low concentration conditions that was undetectable, the value was replaced by zero.

Helicobacter pylori infection detection

H. pylori infection was determined by Warthin-Starry (WS) staining in all specimens using the H. pylori detection kit (Beijing ShiJi HeLi Biotechnology Co., Ltd, Beijing, China) according to the manufacturer’s instructions. The background tissue was stained to yellow and the nucleus was dyed to brown. If there is H. pylori infection, the bacteria would be stained to black.

Statistical analysis

Firstly, multivariate analysis of pattern recognition was performed by using SIMCA 14.1 software package (MKS Data Analytics Solutions, Umea, Sweden), including principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA). Afterwards, a seven-fold cross-validation was performed to further validate the method, and the goodness of fit parameter (R²) and the goodness of prediction parameter (Q²) values were used to estimate the robustness and the predictive ability of the models. Secondly, univariate analysis was performed by using SPSS 20.0 software (SPSS Inc., Chicago, IL, USA). The concentrations of GJFAAs were expressed as median (interquartile range, IQR), and Mann-Whitney U test was used to assess the differences between the GC and NGD groups. P-values < 0.05 were considered as statistically significant. Furthermore, the criteria for selecting differential GJFAAs were a variable importance in the projection (VIP) value > 1, and a P-value of Mann-Whitney U test < 0.05 between the two groups [26]. Subsequently, the differential GJFAAs were cross-referenced to the pathways by further searching with a commercial database, KE-GG (http://www.kegg.jp/kegg/pathway.html) and a free and web-based tool, MetaboAnalyst (http://www.metaboanalyst.ca/MetaboAnalyst/) [27,28].

Results

No significant difference was found in the disease constitution between the training and validation sets (Table 1). The baseline characteristics of the study population are described in Table 2. No significant difference was observed in H. pylori infection rate, but a significant difference in male/female ratio existed between the GC and NGD groups both in the training set (P = 0.005) and validation set (P = 0.006). However, no significant difference was observed in age between the GC and NGD groups in the training set, but a significant difference existed in the validation set (P = 0.007).

Table 1.

Patients’ disease constitution

Set	Patients	GC		NGD			P value
		AGC	EGC	GD	CAG/IM	CSG
Training	130	26	21	42	24	17	0.197
Validation	70	21	11	13	12	13
Total	200	47	32	55	36	30

All data are shown as count. GC, gastric cancer; NGD, non-neoplastic gastric diseases; AGC, advanced gastric cancer; EGC, early gastric cancer; GD, gastric dysplasia; CAG/IM, chronic atrophic gastritis and/or intestinal metastasis; CSG, chronic superficial gastritis. No statistically significant difference of disease constitution was found between the training and validation sets using Pearson’s Chi-square test.

Table 2.

Patients’ baseline characteristics

Variable	Training set (n = 130)		P value	Validation set (n = 70)		P value
	GC (n = 47)	NGD (n = 83)		GC (n = 32)	NGD (n = 38)
Age (years)	59.7 ± 12.9	63.2 ± 11.8	0.451a	69.6 ± 11.5	60.6 ± 15.0	0.007a,**
Gender (M/F)	32/15	35/48	0.005**	24/8	16/22	0.006**
H. pylori positive	14	13	0.056	7 (4 missed)	5	0.218

Data on age are shown as mean ± standard deviation, other categorical data are shown as count. GC, gastric cancer; NGD, non-neoplastic gastric diseases; M/F, male/female.

^**Statistically significant difference (P < 0.01) using independent t-test (a) or Pearson’s Chi-square test (others).

Different GJFAA profiles between the GC and NGD groups in the training set

First, significant alterations in GJFAA profiles were observed in GC patients when compared to NGD patients in the training set (Figure S1). Next, 26 out of 34 GJFAAs were extracted for subsequent analysis. The other 8 GJFAAs (alpha amino acid, citrulline, alpha aminobutyric acid, cystine, 3-methyl histidine, 1-methyl histidine and carnosine and hydroxyproline) were excluded as their median concentrations were 0. The data scale conversion mode used in both GC and NGD samples were log transformation processing and Ctr-formatted (Mean-Centered Scaling) processing. In addition, the missing values of raw data were filled up by half of the minimum values. Thereafter, the normalized data were imported into the SIMCA software for multivariate analysis.

Identification of GJFAA characteristic metabolic perturbation of GC patients in the training set

Initially, PCA was performed using two principal components: R²X = 0.517 and Q² = 0.339. The score plot (Figure 2A) demonstrated that each sample can be clearly divided, and no abnormal sample should be rejected. Certain samples that could not be clearly distinguished were further examined through subsequent discriminating analyses. Subsequently, a loading plot (Figure 2B) was constructed based on OPLS-DA model using one predictive and one orthogonal component (R²X = 0.455, R²Y = 0.421 and Q²Y = 0.279). Using this model, the clearest separation point was produced between the GC and NGD groups. These values illustrated that huge variation in the statistical data was attributable to the separation between the two groups. Then a seven-fold cross-validation permutation test was applied (Figure 2C), where the R² and Q² intercept values were -0.25 and -0.57, respectively. These results excluded the random effects in the constructed model.

Figure 2.

Score scatter plot of PCA model for group GC vs. NGD (A); Score scatter plot of OPLS-DA model for group GC vs. NGD (B); Permutation test of OPLS-DA model for group GC vs. NGD (C); Volcano plot for group GC vs. NGD (D); ROC curves for the detection of GC (E) and EGC (F) combined with 14 kinds of differential GJFAAs in the training set. GC, gastric cancer; EGC, early gastric cancer; NGD, non-neoplastic gastric diseases; ROC, receiver operating characteristic curve.

Screening out potential gastric juice biomarkers in the training set

In total, 14 amino acids were screened as differential GJFAAs (Table 3). Among them, concentrations of 11 GJFAAs’ were increased significantly, which included threonine, serine, alanine, valine, methionine, isoleucine, leucine, tyrosine, phenylalanine, lysine and arginine, respectively. The other three GJFAAs’ concentrations were decreased significantly, which included phosphoserine, ethanolamine phosphate and urea, respectively. For quick visual identification of the varied GJFAAs, we displayed the statistical results in the form of a volcano plot as depicted in Figure 2D.

Table 3.

Differential GJFAAs between the GC and NGD patients in the training set

GJFAA	Median GC	Median NGD	P value	VIP	FC
PSer	0.028	0.037	0.002	1.054	0.768
PEtN	0.007	0.018	< 0.001	1.028	0.606
Urea	0.178	0.604	< 0.001	1.058	0.486
Thr	0.022	0.007	< 0.001	1.489	2.431
Ser	0.016	0.005	< 0.001	1.420	2.671
Ala	0.033	0.016	< 0.001	1.238	1.973
Val	0.025	0.013	< 0.001	1.025	1.763
Met	0.017	0.007	< 0.001	1.178	2.148
Ile	0.026	0.007	< 0.001	1.343	2.674
Leu	0.075	0.020	< 0.001	1.626	2.697
Tyr	0.066	0.026	< 0.001	1.580	1.926
Phe	0.069	0.032	< 0.001	1.415	1.754
Lys	0.044	0.015	< 0.001	1.091	2.321
Arg	0.036	0.008	< 0.001	1.332	2.722

P value, Statistically significant difference using Mann-Whitney U test; VIP, variable importance in the projection; FC, fold change; GC, gastric cancer; NGD, non-neoplastic gastric diseases; GJFAA, gastric juice free amino acid.

Diagnostic value analysis of 14 differential GJFAAs in the training set

The area under the curve (AUC) of these differential GJFAAs ranged from 0.666 to 0.868 for the diagnosis of GC individually (Table 4). We created logistic regression model by binary logistic regression analysis in the training set. Zero (NGD group) or one (GC group) serves as dichotomous variable, and fourteen differential metabolites as the covariates. The predicted equation was as follows: P = 28.869 * X₁ - 25.640 * X₂ + 12.840 * X₃ + 135.106 * X₄ - 44.802 * X₅ + 83.488 * X₆ - 124.793 * X₇ - 1.921 * X₈ + 37.750 * X₉ + 70.130 * X₁₀ + 18.240 * X₁₁ + 46.077 * X₁₂ + 66.500 * X₁₃ - 54.718 * X₁₄ + 15.324 (Table S5). P referred to the value of predicted probability of each sample based on the levels of 14 candidate metabolites. We used the values of predicted probability as new variables and produced a combined ROC curve with a cutoff value of 30.011. The corresponding sensitivity was 0.851, specificity was 0.892 and accuracy was 0.877. The combined AUC was 0.902 (95% CI, 0.846-0.959) for the diagnosis of GC (Figure 2E). Importantly, their AUCs ranged from 0.649 to 0.857 (Table 4), and their combined AUCs reached up to 0.880 (95% CI, 0.792-0.969) for the diagnosis of EGC (Figure 2F). Particularly, leucine, threonine, and serine were the GJFAAs with the most variance between the two groups, whose fold change was more than 2 and AUC value was greater than 0.8. Moreover, combined AUC of the three non-AAAs [0.869 (95% CI, 0.805-0.934)] for diagnosing GC was slightly higher than the combined three AAAs [0.841 (95% CI, 0.773-0.908)]. Combined detection of the above six GJFAAs could further improve its AUC value to 0.871 (95% CI, 0.809-0.933) (Figure 3).

Table 4.

Discriminating performance of each GJFAA for the GC and EGC patients in the training set

GJFAA	GC		EGC
	AUC (SE)	95% CI	AUC (SE)	95% CI
PSer	0.666 (0.054)	0.561-0.771	0.649 (0.076)	0.500-0.798
PEtN	0.718 (0.052)	0.615-0.820	0.731 (0.076)	0.582-0.879
Urea	0.804 (0.039)	0.729-0.880	0.788 (0.051)	0.688-0.889
Thr	0.835 (0.037)	0.764-0.907	0.833 (0.052)	0.731-0.935
Ser	0.831 (0.037)	0.759-0.903	0.821 (0.051)	0.722-0.920
Ala	0.783 (0.042)	0.702-0.865	0.766 (0.057)	0.654-0.877
Val	0.717 (0.049)	0.621-0.814	0.695 (0.071)	0.556-0.833
Met	0.797 (0.041)	0.718-0.877	0.745 (0.067)	0.613-0.877
Ile	0.812 (0.039)	0.736-0.887	0.827 (0.052)	0.724-0.929
Leu	0.868 (0.033)	0.803-0.933	0.857 (0.044)	0.771-0.942
Tyr	0.833 (0.035)	0.765-0.902	0.853 (0.043)	0.769-0.936
Phe	0.802 (0.041)	0.720-0.883	0.793 (0.059)	0.677-0.910
Lys	0.804 (0.040)	0.725-0.883	0.804 (0.055)	0.697-0.912
Arg	0.772 (0.044)	0.686-0.859	0.787 (0.066)	0.657-0.917

AUC, area under the ROC curve; SE, the standard error of AUC; 95% CI, 95% confidence interval; GC, gastric cancer; EGC, early gastric cancer; GJFAA, gastric juice free amino acid.

Figure 3.

ROC curves for the detection of GC (A-C) and box plots of the levels of the six GJFAAs from the GC and NGD groups (D-I) in the training set. GC, gastric cancer; NGD, non-neoplastic gastric diseases; ROC, receiver operating characteristic curve.

Validation of the six selected GJFAAs in the validation set

In the validation set, the combined AUC of the three non-AAAs [0.900 (95% CI, 0.831-0.969)] was also superior to the combined three AAAs [0.851 (95% CI, 0.761-0.941)] in the diagnosis of GC. Combined AUC comprising of the six GJFAAs was 0.914 (95% CI, 0.850-0.977) for GC (Table 5; Figure 4). The diagnostic panel achieved 71.9% sensitivity, 97.4% specificity, and 85.7% accuracy for the prediction of GC (Table S6). Similarly, combined AUC of the three non-AAAs [0.900 (95% CI, 0.789-1.000)] was still superior over combined three AAAs [0.866 (95% CI, 0.718-1.000)] for the diagnosis of EGC. Combined AUC comprising of the six GJFAAs was 0.909 (95% CI, 0.812-1.000) for EGC (Table 5; Figure 5). The diagnostic panel achieved 72.7% sensitivity, 97.4% specificity, and 91.8% accuracy for the prediction of EGC (Table S6).

Table 5.

Discriminating performance of individual and combined detection of the six GJFAAs for GC and EGC in the validation set

GJFAA	GC		EGC
	AUC (SE)	95% CI	AUC (SE)	95% CI
Leucine	0.872 (0.041)	0.792-0.951	0.895 (0.060)	0.778-1.000
Threonine	0.901 (0.036)	0.831-0.972	0.897 (0.063)	0.774-1.000
Serine	0.871 (0.043)	0.786-0.955	0.866 (0.082)	0.706-1.000
Tyrosine	0.838 (0.047)	0.745-0.931	0.866 (0.072)	0.726-1.000
Phenylalanine	0.830 (0.049)	0.733-0.926	0.842 (0.077)	0.691-0.993
Tryptophan	0.852 (0.045)	0.764-0.941	0.867 (0.073)	0.724-1.000
Combined 3 non-AAAs	0.900 (0.035)	0.831-0.969	0.900 (0.056)	0.789-1.000
Combined 3 AAAs	0.851 (0.046)	0.761-0.941	0.866 (0.076)	0.718-1.000
Combined all	0.914 (0.032)	0.850-0.977	0.909 (0.050)	0.812-1.000

AUC, area under the ROC curve; SE, the standard error of AUC; AAAs, aromatic amino acids; non-AAAs, non-aromatic amino acids; GJFAA, gastric juice free amino acid; GC, gastric cancer; EGC, early gastric cancer.

Figure 4.

Scatter plots of the levels of the six GJFAAs from the GC and NGD patients (A-C) and ROC curves for the detection of GC (D-F) in the validation set. GC, gastric cancer; NGD, non-neoplastic gastric diseases; ROC, receiver operating characteristic curve.

Figure 5.

Scatter plots of the levels of the six GJFAAs from the EGC and NGD patients (A-C) and ROC curves for the detection of EGC (D-F) in the validation set. EGC, early gastric cancer; NGD, non-neoplastic gastric diseases; ROC, receiver operating characteristic curve.

Different metabolites related pathways between the GC and NGD groups in the training set

The pathways that matched based on the KEGG database included aminoacyl-tRNA biosynthesis, biosynthesis of amino acids, and others. More details were described in Tables S7 and S8. Among these pathways, both aminoacyl-tRNA biosynthesis and biosynthesis of amino acids pathways included largest number of significantly varied GJFAAs, including all differential GJFAAs except for phosphor-ethanolamine and urea. To identify the pathway that is most relevant to GC development, metabolomics data were comprehensively analyzed using MetaboAnalyst 3.0, which combined pathway enrichment and topology analyses to process our screening data (Table S9). The enrichment analysis showed that the metabolic pathways with raw P values < 0.001 included aminoacyl-tRNA biosynthesis; valine, leucine and isoleucine biosynthesis; and cysteine and methionine metabolism. The raw P values < 0.001 indicated that the altered pathway significantly contributed to the difference between the two groups. The topology analysis showed that the impact value of the metabolic pathways of aminoacyl-tRNA biosynthesis; glycine serine and threonine metabolism; and phenylalanine metabolism are 0.113, 0.262 and 0.119, respectively. An impact value ≥ 0.1 indicated that the altered pathway affected GC carcinogenesis. Further, a Google map-style interactive visualization system was used to generate Figure 6 to present the pathway analysis results in an intuitive manner.

Figure 6.

Pathway analysis bubble plot for group GC vs. NGD in the training set. GC, gastric cancer; NGD, non-neoplastic gastric diseases.

Discussion

Utility of amino acid analytics in diagnosis of GC

Employing amino acid metabolomics for diagnosing GC had been discussed by many researchers via various types of biofluids and tissue extracts [29-34]. For example, Chen [32] and colleagues used gas chromatography combined with mass spectrometer for fingerprinting urinary metabolites on 293 urine samples to find practical and cost-effective biomarkers. As a result, 17 metabolites are significantly different between patients and healthy controls in the training set. Among them, 14 metabolites emerged diagnostic value better than traditional serum biomarkers by quantitative assay of the validation set. Of these, 10 are amino acids and four are organic metabolites. Their AUC values ranged between 0.669-0.823, and the combination AUC can reach up to 0.893. The diagnostic efficacy displayed in the study was consistent with ours. In our study, the AUC of 14 differential GJFAAs was between 0.666-0.868, and their combined AUC was 0.902, which was slightly higher than that reported by Chen’s study. This may be related to the tissue specificity of gastric juice which was higher than urine.

Wang [33] and colleagues pointed out the metabolic profiling of tissue samples on a large cohort of human GC subjects (n = 125) and normal controls (n = 54) using (1) H-NMR along with multivariate analyses. 48 endogenous distinguishing metabolites (VIP > 1 and P < 0.05) were identified altogether. These modified metabolites revealed disturbances of amino acids, glutaminolysis, glycolysis, tricarboxylic acid cycle and choline metabolism, which were correlated with the growth and progression of GC. The AUC value of OPLS-DA model between the neoplastic and normal tissues was 0.945. Although its predictive power was a little higher than ours, the number of distinguishing metabolites selected by OPLS-DA analysis of our study was much less than that of the study by Wang et al. This is probably caused by our study mainly emphasizing amino acid metabolism, instead of global metabolomic profiling. However, the detection method of our study was relatively simple, the expense was comparatively low, and the sample was substantially rich in content and reproducible.

Additionally, PFAA levels from 56 GC patients and 137 age-matched healthy controls were monitored by Gu [17] and colleagues using amino acid analyzer. In contrast to the healthy control group, the concentrations of threonine, arginine and cysteine were increased significantly in the GC group, and the concentrations of 9 amino acids, such as aspartic acid, glutamic acid, serine, etc., were decreased significantly. Instead, the concentrations of 34 GJFAAs were measured by amino acid analyzer from 47 GC patients and 83 NGD patients in the training set of this study. Among the 14 differential GJFAAs, the levels of 11 amino acids, such as threonine, serine, leucine, etc. were increased significantly in the GC group, and the levels of phosphoserine, ethanolamine phosphate and urea were decreased significantly. This elucidates that a changing concentration trend between the majority of PFAAs and GJFAAs in GC participants was just the opposite. The abnormal accumulation of “useful” metabolites in tumor microenvironment and their intensive metabolism in GC tissues might explain the inconsistency of the metabolic phenotypes between the plasma and gastric juice in GC sufferers.

Screening of differential GJFAAs and their biological significance

To facilitate the clinical application, 14 differential GJFAAs were further screened according to the FC and AUC values. Leucine, threonine, serine, isoleucine, lysine, and urea are the six FAAs that meet both FC > 2.0 and AUC > 0.8 values, and are expected to be the candidate biomarkers for diagnosing GC. We chose the first three for validation, compared and combined their diagnostic capabilities to three AAAs we had concentrated on. More concretely, the combined AUC of the three non-AAAs was superior to the combined three AAAs for diagnosing GC or even EGC in the validation set, too. And the combination of AUC comprised of the above six GJFAAs can be further improved. These findings suggest that this integrated diagnostic panel has great potential for improving early detection of GC.

The biological significance for the tremendous increase of the three non-AAAs’ levels in gastric juice of GC patients were as follows. Leucine belongs to the branched chain amino acids, which can promote protein synthesis and inhibit the decomposition, thereby elevated leucine in gastric juice reflects the metabolic pathway of protein biosynthesis was remarkably active in GC tissues. Threonine is a kind of carbohydrate amino acid. Owing to a large amount of carbohydrate is required to meet the energy metabolism of GC tissues, elevated threonine in gastric juice implies the unusual activation of gluconeogenesis in GC tissues. Serine catabolism produces one carbon unit, which can be used as raw material for the synthesis of purines and pyrimidines, so elevated serine in gastric juice represents the tremendously active metabolism of nucleic acids in GC tissues.

Metabolic pathway analysis of differential GJFAAs

The metabolic pathway of aminoacyl-tRNA biosynthesis was excessively activated, which was significantly responsible for the metabolic modification in GC. It unveiled that the differences of GJFAA profiles between the GC and NGD groups might be closely related to the dramatic activity in cellular protein biosynthesis during the rapid proliferation of the malignant tumor cells. Another three metabolic pathways are concurrent between the eight main metabolic pathways in this study that are selected from 14 differential GJFAAs and six main metabolic pathways in Chen’s study [32] that are selected from 14 differential metabolites in urine, including glycine, serine and threonine metabolism, cystine and methionine metabolism and valine, leucine and isoleucine biosynthesis. This implied that the metabolic patterns of some small molecular substances (especially amino acids) in microenvironment-gastric juice and macroenvironment-urine are more or less the same in the process of gastric carcinogenesis. The metabolic studies based on gastric juice and urine samples manifested a great value for the diagnosis of GC. In addition, combined detection of these might complement each other, as well as enrich our understanding concerning the pathogenesis of GC.

In conclusion, our results suggest that GJFAA profiling might be an ideal tool for improving the prognosis of GC via its early diagnosis and treatment. It also brings deeper insights into the oncogenesis of GC, which will undoubtedly contribute to proposing novel therapeutic targets. Meanwhile, joint examination of the three non-AAAs (leucine, threonine and serine) and three AAAs in gastric juice could further upgrade the diagnostic level of GC especially EGC.

Disclosure of conflict of interest

None.