Computational models applied to metabolomics data hints at the relevance of glutamine metabolism in breast cancer

Metabolomics and gene expression data from 67 fresh tumor tissue samples originally analyzed by Terunuma et al. were included in this study. Untargeted metabolomic quantification was performed by Metabolon Inc. Samples were prepared using the automated MicroLab STAR system from Hamilton company and mass-spectrometry experiments were done in a Waters ACQUITY UPLC and a Thermo-Finnigan linear trap quadrupole mass spectrometer and in a Thermo-Finnigan Trace DSQ fast-scanning quadrupole mass spectrometer [18].

Metabolomics data contains information about 536 metabolites. Log2 was calculated. As quality criteria, data were filtered to include detectable measurements in at least 75% of the samples. Missing values were imputed to a normal distribution using Perseus software [19]. After quality control, 237 metabolites were considered for subsequent analyses.

In terms of gene expression data, the 2000 most variable genes, i.e., those genes with the highest standard deviation, were chosen to build the PGM. This data was from an Affymetrix array and they are available in Gene Expression Omnibus Database under the identifier GSE37751.

As previously described [10, 11, 14, 20], PGMs compatible with high dimensional data were used, using correlation as associative criteria. PGMs were built using metabolomics, gene expression or flux activity data without any a priori information. The grapHD package [21] and R v3.2.5 [22] were employed to build the PGMs. In this case, PGMs use gene expression or metabolomics quantification without any a priori information, making connections based on correlation as associative method. PGM construction is based on two sequential steps: first, finding the spanning tree with the maximum likelihood and, second, reduce the graph edges based on the reduction of Bayesian Information Criterion (BIC) and the preservation of decomposability [23]. The visualization of the resulting network was done using Cytoscape software [24] (Sup Files 1‐3). The resulting networks were divided into branches and ontology analyses were done to assign a majority function/metabolic pathway to each branch, defining in this way different functional nodes in the networks. In the case of genes, gene ontology analyses were performed using the DAVID web tool with “Homo sapiens” as background and GOTERM, KEGG and Biocarta selected as categories [25]. In the case of metabolites, the Integrated Molecular Pathway Level Analysis (IMPaLA) web tool was used to assign a main metabolic pathway to each branch [26].

Once the functional structure was defined, functional node activities were calculated in order to make comparisons between groups, as previously described [10, 11, 14, 20]. Briefly, each functional node activity was calculated as the mean of the expression/quantity of genes/metabolites of each node that are related to the main node function/metabolic pathway.

FBA is a method that allows the estimation of tumor growth rate and the model of the flow of metabolites in a metabolic network. FBA was performed using the library COBRA Toolbox v2.0 [27] available for MATLAB. FBA was calculated using the whole human metabolic reconstruction Recon2 [28]. This metabolic reconstruction includes 2191 genes collected into the Gene Protein Reaction rules (GPRs), 5063 metabolites and 7440 reactions. GPRs represent the relationships between genes and metabolic reactions and they are included into the model as Boolean expressions. GPRs were solved as described in previous studies [11, 14], using a modification of the Barker et al. algorithm [29], which were incorporated into the model by a modified E-flux method [14, 30]. Briefly, the “OR” operators were solved as the sum and the “AND” operators were solved as the minimum. Then, the GPR data were normalized using the Max-min function and introduced into the model as the reaction bounds. As the objective function, the biomass reaction proposed in the Recon2 was used as representative of tumor growth. This biomass reaction was based on experimental measurements of leukemia cells. The 7440 reactions are grouped into 101 metabolic pathways. The formulated mathematical problem was solved by linear programming. Finally, an estimation of tumor growth rate and a value of flux for each reaction in the model were obtained.

Flux activities were previously proposed as a measurement to compare differences at the metabolic pathway level [14]. Briefly, they were calculated as the sum of the fluxes of the reactions included in each pathway defined in Recon2.

The statistical analyses were performed with GraphPad Prism v6. Predictor signatures were built with the BRB Array Tool from Dr. Richard Simon’s team [31]. Variables related to overall survival were ranked on the basis of their p-values for the long-rank test, using the overall survival as endpoint. Leave-one-out cross-validation was used to evaluate the predictive accuracy of the profiles. The cutoff point was established a priori and to test the statistical significance, the p-value of the log-rank test statistic for the risk groups was evaluated using 1000 random permutations. Then, the independence between the defined predictors and clinical parameters was checked using Cox regression models. All p-values are two-sided and are considered statistically significant under 0.05.

Breast cancer cell lines (MCF7, T47D and CAMA1 [ER+], and MDAB231, MDAMB468 and HCC1143 [TNBC]) were cultured in RPMI-1640 medium with phenol red, supplemented with 10% heat-inactivated fetal bovine serum, 100 mg/mL penicillin and 100 mg/mL streptomycin. Cell lines were cultured at 37 °C in a humidified atmosphere with 5% (v/v) CO2 in the air. The MCF7 (ATCC® HTB-22™), T47D (ATCC® HTB-133™) and MDA-MB-231 (ATCC® HTB-26) cell lines were kindly provided by Dr. Nuria Vilaboa (La Paz University Hospital, previously obtained from ATCC in January 2014). The MDAMB468 (ATCC® HTB-132™), CAMA1 (ATCC® HTB-21™) and HCC1143 (ATCC® CRL-2321™) cell lines were obtained from ATCC (July 2014). Cell lines were routinely monitored and authenticated by morphology and growth characteristics, tested for Mycoplasma and frozen, and passaged for fewer than 6 months before experiments. The aminooxyacetic acid (AOA) (Sigma Aldrich C13408) and L-Glutamic acid γ-(p-nitroanilide) hydrochloride (GPNA) (Sigma Aldrich G6133) were obtained from Sigma-Aldrich (St. Louis, MO, USA).

Dose-response curves were designed for AOA and GPNA. As the preparation of GPNA needs an acid medium, HEPES (50 mM) was added to buffer the medium. About 5000 cells were seeded in each well in 96-well plates and after 24 h, drugs were added. After an incubation of 72 h, cell viability was determined using CellTiter 96 AQueous One Solution Cell Proliferation Assay (Promega) kit and absorbance was quantified on a microplate reader (TECAN). As a control untreated cells were used and all the experiments were performed by triplicate.

With the aim of study the relationships between metabolomics, gene expression, and FBA results, metabolomics and gene expression data, analyzed by mass-spectrometry and microarrays GeneChip Human Gene 1.0 ST (Affymetrix) respectively and published by Terunuma et al., were analyzed [18].

A total of 67 tumor fresh tissue samples from patients with breast cancer recruited in Baltimore hospital were studied. This patient cohort comprises 33 ER+ and 34 ER- (of which 14 were also TNBC). The median follow-up was 50 months, and 31 deaths occurred during this time. No significant differences regarding overall survival (OS) were observed between patients with ER+ or ER- tumors. Patient characteristics are shown in Table 1.

After Kaplan-Meier analysis, 29 metabolites were found related to OS (p < 0.05) (Sup Table 1).

Then, an OS predictor using this metabolomics data was built. This metabolite-based signature included five metabolites: glutamine, 2-hydroxypalmitate, deoxycarnitine, butyrylcarnitine and glycerophosphorylcholine (p-value =0.003, hazard ratio [HR] = 0.34, 50:50%) (Fig. 1). A multivariate analysis showed that the signature provided additional prognostic information to clinical data (S1 Table).

Metabolomics data without using any a priori information were analyzed through PGM. Metabolomics database, including information about 536 metabolites, was reduced to 237 metabolites due to quality criteria (detected in at least 75% of the samples). A main metabolic pathway was assigned to each functional node of the resulting network using IMPaLA. IMPaLA is a tool that allows ontology analyses based on metabolic pathways and metabolites instead of other platforms such as DAVID which are based on genes and biological processes. Strikingly, this network had a functional structure, grouping the metabolites into metabolic pathways as it has been previously shown for gene and protein PGMs [10, 11, 14, 20]. This is relevant because this type of networks is built based on the gene expression data or, in this case, the metabolomics data, without any a priori information, meaning that those metabolites that are involved in a concrete metabolic pathway had a correlated quantification profile across the series. Five functional nodes were defined, each with a different overrepresented metabolic pathway (Fig. 2).

The activity of each functional node was calculated as previously described and comparisons between ER+ and ER- were done [10, 11, 14, 20]. Significant differences were found between ER+ and ER- tumors regarding lipid and purine metabolism (p < 0.05) (S1 Fig).

Moreover, the lipid metabolism functional node showed prognostic value in this cohort in a Kaplan-Meier analysis (p-value = 0.008, FDR = 0.04). Then, a predictor signature was built (p = 0.045, HR = 0.48, 50:50%) (Fig. 3). However, a multivariate analysis showed that the predictor do not add additional prognostic information to that provided by clinical features (S2 Table).

On the other hand, a network combining metabolomics and gene expression data was built. Due to the differences between both kinds of data, most of the metabolites were grouped together. However, some metabolites were integrated into gene branches (Fig. 4).

This combined network was then functionally characterized based on the majority function of the genes contained in each branch. The resulting network had eleven functional nodes and a twelfth branch that include the majority of the metabolites (Fig. 4).

Once the main functions were assigned, a literature review was performed to study the relationship between metabolites included in the gene functional nodes and the main function of each functional node. We found out that a relationship with functional nodes had been previously described for 4 of 20 metabolites: succinate, cytidine, histamine and 1,2-propanediol. The relationships between metabolites and their node function are shown in Table 2.

FBA and flux activities were calculated as previously described [14]. Briefly, gene expression data from 67 tumor samples were used, GPR rules were solved and the normalized values were introduced into the metabolic model using modified E-flux algorithm [11, 30]. Finally, FBA was calculated using a biomass reaction representative of tumor growth. No significant differences were found in the tumor growth rate between ER+ and ER- tumors (S2 Fig).

Flux activities showed significant differences between ER+ and ER- in glycerophospholipid metabolism, phosphatidyl inositol metabolism, urea cycle, propanoate metabolism, pyrimidine catabolism and reactive oxygen species (ROS) detoxification (S3 Fig).

In addition, the combination of glutamate metabolism (the pathway that includes the glutamine) and alanine and aspartate metabolism flux activities showed prognostic value in this cohort (p-value = 0.024, HR = 0.41, 50:50%) (Fig. 5). A multivariate analysis showed that this flux activity-based predictor provides prognostic information independently from clinical data (S3 Table).

Flux activities were calculated for each metabolic pathway defined in the Recon2. Then, using flux activities and metabolomics data, a new PGM was built to study association between both types of data. Interestingly, both types of data appeared mixed in the network; with the peculiarity that flux activities appeared usually at the periphery of the network (Fig. 6).

The resulting network was split into several branches to study the relationship of the metabolites with the flux activities included in each branch (Fig. 6). Coherence between both types of data was shown by the PGM, associating flux activities and metabolites related to these flux activities in the same branch. For instance, branch 1 includes glycolysis flux activity and three metabolites previously related to glycolysis (S4 Table). Regarding vitamins and cofactors, it was not possible make comparisons because the IMPaLA label for this category is “Vitamin and co-factor metabolism” and Recon2 labels differentiate between the various vitamins, labeling them as “Vitamin B6 metabolism”, “Vitamin A metabolism”, etc.

As the computational analyses pointed out the relevance of glutamine and its metabolism in breast cancer (glutamine and glutamate metabolism appeared recurrently in the analyses), cell viability assays employing two drugs targeting glutaminolysis (aminooxyacetic acid [AOA], and L-Glutamic acid γ-(p-nitroanilide) hydrochloride [GPNA]) were performed. Dose-response curves of these two drugs confirm that targeting glutamine metabolism affected cell viability (Fig. 7).

In AOA-treated cell lines, the IC₅₀ did not show any differential response between breast cancer subtypes. However, in GPNA-treated cells, IC₅₀ for ER+ cell lines were lower than IC₅₀ for TNBC cells (Table 3).