Malignancy-associated metabolic profiling of human glioma cell lines using ¹H NMR spectroscopy

The five cell lines used in this study were derived from glioma tissues of different malignant degrees (shown in Table 1). Aqueous phase extracts of these cell lines were subjected to NMR analysis. The NMR spectra acquired 48 hours after seeding are shown in Figure 1. As the ¹H NMR spectra from six or more measurements for each cell line displayed almost identical spectral profiles, only one of the replicates is shown in the Figure 1. These observations indicated that metabolic profiles of the independently grown cell lines were highly reproducible under the same culture conditions. The resonance assignments of major metabolites in the spectra were performed based on both literature data[13‐15] and the Human Metabolome Database (http://​www.​hmdb.​ca). As a result, more than 30 metabolites were identified, which provided adequate information for assessing variations in metabolic profiles within the five cell lines.

To determine the differences in the metabolic profiles among the five cell lines, we performed principal component analysis (PCA) to analyze the normalized ¹H NMR data. The results showed that all samples were located in the Hotelling’s T2 oval of the 95% confidence interval, and the cell extracts from the same cell line clustered together (Figure 2A). Moreover, cell lines (CHG5, SHG44) from low-grade glioma were separated from those derived from high-grade glioma (U118, U87 and U251) along the first principal component (PC1).

To confirm the separation of groups from PCA, we utilized the fuzzy c-means clustering (FCM) method to analysis metabolic profile clustering by calculating memberships of each feature to each of the user-defined numbers of clusters. In this method, a membership value near 1 indicates strong membership and close to 0 indicates weak or no membership. The PCA scores plot (Figure 2A) provided visual information on the distribution of metabolic profiles, which suggested the number of clusters could be set to two. When the cluster number was two, CHG5 and SHG44 were grouped into one cluster whereas U118, U87 and U251 were grouped together into another cluster (Figure 2B). We further used projection to latent structure with discriminant analysis (PLS-DA) to evaluate the reliability of the two-cluster model and obtained three predictive principal components using the discrimination calculation. We then performed permutation tests with 999 iterations to assess the possibility of model over fitting. The obtained parameters were R2X (cum) = 0.560, R2Y (cum) = 0.964, Q2 (cum) = 0.933, which indicated that the model was highly reliable (Figure 2C and2D). These results indicated that the two-cluster model possessed high discrimination and predictive capability. Notably, the result from the two-cluster model was fully consistent with the grouping of cell lines based on the pathological grade of the corresponding glioma tissues.

After continuous subculture in vitro, the malignant degree of the glioma cell lines could become inconsistent with the pathological grade of their glioma tissues. To further identify the diversity of the malignant behaviors of cell lines in the two groups, we firstly detected the proliferation abilities of five cell lines and did not observe significant difference in proliferation abilities among the cell lines (Additional file1: Figure S1). Then we conducted hematoxylin and eosin (HE) staining, immunohistochemical staining of GFAP and MMP-9, and a transwell assay to evaluate the differentiation level and invasion ability. The HE staining results showed three major shapes (circle, star and spindle) and obvious atypia of all these cell lines (Figure 3A). Immunohistochemical staining of GFAP was used to evaluate the differentiation level of the glioma cells. As shown in Figure 3B and3E, cell lines of the high-grade group (U118, U87 and U251) had considerably lower GFAP positive expression compared with those of the low-grade group(CHG5 and SHG44), indicative of lower differentiation levels of cell lines of the high-grade group. Both the MMP-9 immunohistochemical staining and the transwell experiment were used to assess the invasion abilities of glioma cells. The results showed that the cell lines in the high-grade group possessed a generally higher invasion ability than those in the low-grade group (Figure 3C,3D,3F and3G).

The comprehensive analyses of GFAP and MMP-9 expression levels, and invasion ability indicated that the cell lines from glioma tissues of different pathological grades showed obvious variations in malignant behavior. Compared with the low-grade group, the high-grade group showed a lower degree of differentiation and higher invasion ability. Thus, the two-cluster model of metabolic profiles was closely associated with the malignant behaviors of glioma cell lines cultured in vitro.

To identify the characteristic metabolites responsible for the separation between the low-grade group and the high-grade group, we applied OPLS-DA to analysis the difference in metabolic profiles of cell lines in the two groups. The scores and loading plots with correlation coefficients are shown in Figure 4A. The loadings plot was used to identify the significant characteristic metabolites responsible for the clustering patterns. Resonance assignments of these characteristic metabolites were according to literature references and the human metabolome database (http://​www.​hmdb.​ca). Approximately 17 characteristic metabolites were identified (Table 2), including: leucine (Leu), valine (Val), isoleucine (Ile), lysine (Lys), glutamate (Glu), glutamine(Gln), glutathione (GSH), threonine (Thr), tyrosine (Tyr), phenylalanine (Phe), taurine (Tau), sn-glycero-3-phosphocholine (GPC), myo-inositol (Myo), creatine (Cr), lactate (Lac), formate (For) and acetate (Ac).

The importance of the 17 metabolites in distinguishing metabolic profiles was ranked according to their VIP scores of the OPLS-DA model (Figure 4B). The results indicated that taurine had the highest correlation with the grouping of the glioma cells, followed by glutamine and lactate.

Additionally, we calculated the relative integrals of the characteristic metabolites for the two groups. Compared with the low-grade group, the high-grade group showed distinctly decreased levels of metabolites such as Phe, Cr, Tau, Lac, Ac and For, and significantly increased levels of other metabolites, including Val, Leu, Ile, Lys, Glu, Gln, GSH, GPC, Myo, Thr and Tyr (Figure 4C).

To further clarify the potential relationships among the characteristic metabolites and the interrupted metabolic pathways in the two groups, we calculated Pearson’s correlation coefficients of the relative integrals of the characteristic metabolites (Figure 5A). Both the low-grade group and the high-grade group displayed quite different correlation patterns. Figure 5B show a network connecting the differently correlated metabolites in the two groups. Both groups displayed few similar correlations but many different correlations. Positive correlations of the branched chain amino acids (BCAA, including Ile, Val, Leu) in the low-grade group were also observed in the high-grade group. However, BCAAs in the high-grade group showed positive correlations with glutamine, tyrosine, phenylalanine, GPC; but negative correlation with myo-inositol, glutamate, threonine, glutathione, lactate, which disappeared in the low-grade group. Moreover, the positive correlations of BCAA with acetate observed in the low-grade group, disappeared in the high-grade group. Lysine was negatively correlated with taurine and creatine in the low-grade group, but disappeared in the high-grade group. Notably, the correlation patterns of the metabolites in the high-grade group were significantly different from those in the low group. Based on the different correlations of characteristic metabolites and the results from Kyoto encyclopedia of genes and genomes (KEGG) analysis, we summarized the potential disordered metabolic pathways that are associated with the malignant behaviors. As shown in Figure 6, the disordered metabolic pathways are mostly involved in TCA cycle anaplerotic flux, amino acid metabolism, anti-oxidant mechanism and choline metabolism.

In addition, we utilized the GeneGo pathway analysis platform to generate a protein-metabolite interactional network that could associate the characteristic metabolites with malignant behavior markers GFAP and MMP-9. The GeneGo analysis platform provides the prebuilt networks of protein-compound/metabolite assembled by GeneGo scientific annotators based on proven literature evidence[19]. Nine of the seventeen metabolites (valine, glutamine, glutathione, tyrosine, phenylalanine, leucine, acetate, choline, creatine) could directly connect to GFAP and MMP-9 pathway networks via the shortest path network option (Additional file2: Figure S2). The altered levels of the nine metabolites were closely associated with the changed functions of GFAP and MMP-9, suggesting that these metabolites could be explored as potential biomarkers for monitoring malignant transformation of glioma cells.

Five astrocytoma cell lines from glioma tissues with different pathological grades (CHG5, SHG44, U87, U118, U251, Table 1) were used in the present work. Cell lines (CHG5 and SHG44) were kindly provided by Professor XW Bian of the Third Military Medical University, China. Glioblastoma U87 and U118 cell lines were obtained from the American Type Culture Collection (ATCC). Glioblastoma U251 cell line was obtained by the China Center for Typical Culture Collection (CCTCC). All the cell lines were maintained in DMEM supplemented with 100 units/ml penicillin, 100 μg/ml streptomycin and 10% fetal bovine serum (FBS, Hyclone) at 37°C in a humidified atmosphere of 5% CO₂.

Five cell lines plated in 96-well plates (5 × 10³ per well) were cultured for 24 h and 48 h. Cell samples were then incubated with CellTiter 96 AQueous solution (MTS, 20 μl/well) and culture medium (DMEM, 100 μl/well) for 4 h. Next, colored MTS products were detected by absorbance at 490 nm on a Molecular Devices Microplate Reader (BioTek, USA).

Cells were maintained on cover slips of 6-well plates in DMEM containing 10% fetal bovine serum as described above. At confluence, cells on cover slips were fixed with 4% paraformaldehyde for 15 min at room temperature, and were then stained with haematoxylin and eosin. Immunohistochemistry was performed as follows. Fixed cover slips were rinsed with 1% normal calf serum in PBS for 15 min, and permeabilized with 0.3% Triton X-100/PBS for 15 minutes. The cells were then incubated with primary antibodies specifically against GFAP(GA-5,MAIXIN-BiO) and MMP-9(56-1A4, MAIXIN-BiO) at 4°C overnight. The slides were washed with phosphate buffer solution (PBS) including 0.1% Triton X-100, incubated with biotinylated anti-mouse antibody (1:100) at 37°C for 1 hour, incubated with fluorescein isothiocyanate-labeled streptavidin conjugate (1:100) for 1 hour, and finally washed with PBS including 0.1% Triton X-100 three times, mounted, and analyzed under a microscope. The positive cells were counted and scored at a magnification of × 400 under a light microscope in five different fields for each coverslips. One-way analysis of variance (ANOVA) was used to determine the statistical significance of the differences among the five cell lines.

The invasion assay was performed using transwell cell culture chambers (24 wells, 8-μm pore size; BD Biosciences). 1 × 10⁵ tumor cells were resuspended in 200 μl of serum-free DMEM and added to the corresponding upper inserts, respectively. DMEM (600 μl) with 10% FBS was added to the lower chamber. After 24 h, invaded cells were fixed and stained with crystal violet (0.2% in 2% ethanol) for 20 min. Cells on the upper side of the insert membrane were removed with cotton rods. The invaded cells were counted at a magnification of × 100 under a light microscope in nine different fields for each insert. The measurements were repeated at least three times. One-way ANOVA was used to determine the statistical significance of the differences among five cell lines.

Before metabolite extraction, 1 × 10⁶ cells were seeded in 10-cm diameter culture dishes and incubated for 48 h at 37°C and 5% CO₂. About 5 × 10⁶ cells were then harvested and quenched by a direct cell quenching method, as described by Teng et al.[40]. Intracellular metabolites were extracted using a dual phase extraction procedure adopted from Viant et al.[41]. A mixture of methanol, chloroform and water in the volume ratio of 4:4:2.85 was used to generate a two-phase extract.

In the present study, only the aqueous intracellular extracts were used. Before NMR analysis, solvents were completely removed using a Nitrogen Blowing Concentrator. Each aqueous sample was reconstituted in 500 μl of D₂O. Then 50 μl of D₂O containing 1.5 M KH₂PO₄ and 0.1% sodium 3-(trimethylsilyl)propionate-2,2,3,3-d4 (TSP) was added. D₂O was used for field frequency lock, and TSP was used to provide the chemical shift reference (d0.00). Subsequently, all the samples were vortexed, and centrifuged at 12000 g for 15 min at 4°C to remove any insoluble components. Finally the collected supernatants (500 μl) were transferred to 5 mm NMR tubes[42].

All ¹H NMR experiments were conducted on a Bruker Avance III 600 MHz spectrometer at 25°C. Solvent-suppressed 1D NOESY spectra were acquired using the pulse sequence [(RD)-90°-t₁-90°-τ_m-90°-ACQ]. t₁ was 6.6 μs. Water suppression was achieved by irradiation of the water resonance during the recycle delay ( RD ) of 4 s and the mixing time (τ_m) of 120 ms. The spectral width was 10 kHz with an acquisition time per scan of 1.64 s, and 256 transients were collected into 32 K data points for each spectrum. The free induction decay (FID) was zero-filled to 64 K and an exponential line-broadening function of 0.3 Hz was applied to the FID before Fourier transformation. Both phase and baseline corrections were carefully performed. The ¹H NMR spectra were referenced to the methyl group of TSP (δ 0.00).

NMR spectra were reduced to 2587 integrated regions with a width of 0.003 ppm (bin) corresponding to the region of δ 9.5-0.8 using the MestRova6.5 software (Mestrelab Research S.L, Spain). The region of δ 5.5-4.5 was removed to eliminate artifacts related to the residual water resonance. (The remaining integrals for each NMR spectra were normalized to the sum of the spectral intensity to compensate for the differences in sample concentration)[43].

Before multivariate statistical analysis, the integral values were mean-centered and pareto-scaled[44]. To check general separation and identify the outliers, PCA was performed on NMR data sets of all cell samples using the SIMCA-P V12.0 software package (Umetrics AB, Umea, Sweden). Then, PLS-DA and OPLS-DA, were subsequently used to improve the separation[45]. The PLS-DA model was cross-validated to measure the robustness by a permutation analysis with 999 times.

FCM is a clustering method that allows one piece of data to belong to two or more clusters and is extensively used in pattern recognition[46, 47]. The FCM analysis was conducted by minimization of the following objective function:

Where m is any number greater than 1, u_ik is the degree of membership of x_i in the cluster k, x_i is the i-th measured data, v_k is the center data of the k cluster. We used the FCM clustering algorithm through the Fuzzy Logic Toolbox in MATLAB (Version MATLAB2011b, MathWorks, USA).

Two criteria were used to identify the characteristic metabolites. One was the VIP score of the OPLS-DA model[48] and the other one was the correlation coefficient(r) of the variable relative to the predictive component(t[1]) in the OPLS-DA model[49]. The critical values of correlation coefficients were determined by the degrees of freedom in the OPLS-DA model. Characteristic metabolites with a VIP > 1 and |r| > the critical values were identified.

For relative quantification of characteristic metabolites, the relative integrals of metabolites were used for comparison between two groups. The average changes and standard error were calculated[43].

Pearsons’s correlation coefficients of cell samples in two groups were further calculated to display the relationships between the relative integrals of spectral peaks in a certain biological profile, as described previously[43]. A heatmap was used to display the correlation matrices. For all correlations, a p-value was calculated based on a t-test to check the statistical significance. The significance threshold was set to the usual value of 0.05 and corrected according to the number of potential correlations[50].

The network of the metabolites with significance correlations was displayed by MATLAB Bioinformatics toolbox. Significant positive correlations were shown in red, while significant negative correlations were in blue. In addition, GeneGo MetaCore was used to analyze the network that describes the interaction between the characteristic metabolites and GFAP/MMP-9[19].