Merging transcriptomics and metabolomics - advances in breast cancer profiling

Transcriptomics and metabolomics in cancer research have traditionally been considered as two separate fields. Different levels of the molecular processes are studied, aiming at improving cancer treatment by understanding the underlying mechanisms of the disease. Breast cancer treatment decisions today are mainly based on tumor size, histological characterization, grading and receptor status, as well as axillary lymph node status, and age of the patients [1]. However, patients with similar diagnosis and treatment can experience large differences in the progression and relapse of their disease. The various -omics fields, transcriptomics in particular, have provided an understanding of breast cancer as a group of molecularly distinct neoplastic disorders [2]. Clinical use of molecular characterization of breast cancer has the potential to stratify breast cancer patients for more individual treatment, but has so far only been implemented to a limited extent.

The field of transcriptomics, using DNA microarrays that enable measurements of thousands of RNA transcripts in a single experiment, has had a huge impact on breast cancer research over the last decade [2]. One of the important findings has been the classification of breast cancer into five subtypes (luminal A, luminal B, basal-like, ERBB2 enriched and normal-like) based on gene expression profiles of so called intrinsic genes [3, 4]. This molecular subtyping of breast cancer, has been reproduced in several studies and is also associated with clinical outcome across datasets [5, 6].

Metabolomics studies the metabolites and how they are affected by specific cellular processes. The possibility of using in vivo magnetic resonance spectroscopy (MRS) as a diagnostic, prognostic or predictive tool in the clinic simultaneously with an MR imaging (MRI) examination, makes MRS techniques attractive methods for molecular classification of disease. Metabolic profiling of intact biological samples using high resolution magic angle spinning magnetic resonance spectroscopy (HR MAS MRS) enables measurement of multiple cellular metabolites simultaneously. The method has been utilized in a wide range of biological applications [7], and studies of cancers have proven HR MAS MRS to be a promising tool in cancer diagnosis and treatment monitoring [8]. Importantly, the sample is kept intact throughout the HR MAS MRS analysis and can subsequently be analyzed by gene expression analysis.

Profiling gene expression and metabolite content in the same breast carcinoma samples enables comparisons of molecular findings at different levels. Gene expression data and metabolite data from MRS techniques of different samples from the same breast cancer cell line or xenograft model have been combined previously [9‐12], but these studies have mainly focused on specific genes involved in choline metabolism, known as the Kennedy pathway. Combining transcriptomic and metabolomic profiling of the same sample allow us to capture a comprehensive picture at a given moment in time. Such studies could reveal differences and similarities between groups of samples at different molecular levels and provide a fundament for enhanced knowledge of the biological dynamics of breast cancers.

The aim of this study was to combine gene expression microarrays and HR MAS MRS for more refined profiling of breast tumors, and to explore some of the potentials and limitations of the experimental procedures. This study focuses on the most prevalent type of breast cancer, invasive ductal carcinomas (IDC) with oestrogen (ER) receptor positive disease, and the largest molecular subgroup within these tumors, luminal A. ER positive breast cancer accounts for approximately 2/3 of the cases, and although they overall have a relatively good prognosis, some patients experience relapse and do not respond to antieostrogen treatment. So far there are no biomarkers available to identify those patients and no targets for therapy, and identification of molecular markers of such tumors, possibly by metabolic profiling, are therefore of high relevance. Identifying subgroups of patients within this group is an important goal to make it possible to further individualize cancer diagnosis and treatment.

Two main strategies of combining information from the microarray data with the HR MAS MR spectra are presented, as shown in the workflow illustrated in Figure 1. The first strategy uses multidimensional gene expression microarray data and the HR MAS MR spectra to identify molecular subgroups of tumors. The second strategy compares the quantified concentration of specific metabolites to the expression level of each transcript in the filtered microarray data. In a parallel study, we performed additional experiments to assess the impact of sample treatment during HR MAS MRS on RNA integrity and gene expression.

In this study, we have shown the feasibility of merging transcriptomics and metabolomics data from the very same tumor tissue sample. Two strategies of combining microarray data and HR MAS MR spectra are presented, providing a framework for how information from these different molecular levels can be combined and analyzed. We also identified a set of transcripts which showed slightly altered expression after the HR MAS MRS procedure, but overall this variation was smaller than the biological variation in tumors from one patient to another.

In the first strategy to combine gene expression microarray data and HR MAS MR spectra, the expression of "intrinsic" genes was used to classify the samples into established molecular subtypes. The HR MAS MR spectra of the majority of tumor samples, which were classified as luminal A, were further explored to investigate whether metabolic characteristics could define subgroups within a transcriptionally homogenous set of samples. The use of spin echo acquired spectral profiles ensured a more extensive use of metabolic information than using calculated tissue metabolite concentrations, which is limited by several peak areas being non-quantifiable. Three subgroups of luminal A tumors were identified (Figure 2 and 3). The fact that samples cluster together differently with respect to the transcriptional and metabolic profiles (results not shown), indicates that microarrays and HR MAS MRS reflect different traits of the tumors. Lower levels of glucose, which may reflect high energy consumption, and higher levels of alanine in A2 compared with the other luminal A samples indicate that the A2 subgroup has a higher Warburg effect [23]. The lactate signal in A2 also appears to be higher than in the other groups, although not at the significance threshold level. From the GO enrichment analysis, the A2 group was found to be enriched for processes related to cell cycle and DNA repair, compared to A1. The presented subclassification of luminal A might have identified a subgroup of patients (A2) with a more aggressive breast cancer, based on the metabolic and transcriptional profile. However, since this group is small and no long term clinical follow-up is yet available for these patients, a larger cohort with clinical data needs to be analyzed in order to validate whether this finding has clinical impact. It should be noted that intrinsic molecular subtyping is sensitive for selection bias in the cohort analyzed, because of the required gene centring of the microarray data prior to classification. All samples were therefore included in the intrinsic molecular classification, resulting in 10% ER negative samples which is slightly lower than the typical ER negative frequency in IDC. All samples classified as luminal A were ER positive and the majority were PgR positive (3 samples were PgR negative and 1 sample had no IHC data for PgR), which supports the classification since luminal A samples are mostly ER/PgR positive and typically 40-50% of IDCs are classified as luminal A [5]. Even though these preliminary results revealing metabolic subgroups within luminal A tumors need to be reproduced in a larger cohort, they suggest that microarray and HR MAS MRS data complement each other, which can be exploited both in subclassification and for constructing predictors of outcome or treatment response.

The second strategy to combine gene expression microarrays and HR MAS MRS was performed by correlating metabolite concentration and gene expression. In this approach, we have not focused on any specific pathways, but correlated the metabolite concentrations to all transcripts on the microarray that showed some variation across samples. We excluded samples with ER negative status from this analysis to avoid detecting associations related to ER-status, which is known to have a profound effect on the transcriptional profile [24]. The three ER positive samples that were not classified as luminal A were classified as normal-like. Since the gene expression of these three samples also correlated to the published luminal A centroid [17], they were included in the correlation analyses to increase power. The gene transcripts that correlated the most to the concentration of taurine and myo-inositol were enriched for GO-terms associated to extracellular processes, which could reflect a tumor-stroma interaction. "Cell adhesion" was also one of the enriched GO-terms for the gene transcripts that correlated to myo-inositol, which supports this hypothesis. Taurine and myo-inositol are known to be involved in osmoregulation and volume regulation [25]. It should be noted that the concentrations of taurine and myo-inositol correlated negatively to tumor percentage (Additional file 2: Scatterplot of metabolite concentrations and tumor percentage), which could contribute to the apparent association of these metabolites to extracellular processes. Gene transcripts that correlated the most to choline concentration were enriched for cell cycle related GO-terms which indicate that the choline level in these samples reflects proliferation. Choline is involved in glycerophospholipid metabolism and is a nutrient taken up by the cells as well as a breakdown product from phosphatidylcholine. The total choline signal can be detected by in vivo MRS, and is elevated in breast cancer compared to normal mammary tissue and benign lesions [26]. No significantly enriched GO-terms were found in the genes that correlated the most to the two other choline metabolites involved in the total choline peak, PCho and GPC. Glucose correlated to genes that were significantly enriched for the GO-term "immune system process". Glucose concentration has been shown to have an inverse relationship to the number of proliferating cells [27] and tumor cell density [28]. For creatine, only the GO-term "mannosyltransferase activity", which is a glycosylation process, was significantly enriched among the genes that correlated to the metabolite. The genes that correlated to the amino acid glycine, were also only significantly enriched for one GO-term, "respiratory chain", suggesting a possible association between aerobic respiration and glycine levels in these tumor samples. However, glycine was the only metabolite that showed significant positive correlation to tumor percentage (Additional file 2: Scatterplot of metabolite concentrations and tumor percentage.), which could have influenced this relationship. It is worth noting that few transcripts showed a strong correlation to the eight metabolites, as can be seen in the examples in Figure 5. Since metabolite concentrations reflect the sum of many different pathways, correlating the expression of single genes to metabolites is probably not the optimal way to compare the transcriptional and metabolic profiles of tumors. The fact that the most correlated genes were not directly associated with the metabolic pathways of the metabolites they correlated to also emphasizes the complexity of the relationships between gene expression levels and metabolite concentrations. Improved quantification of tissue metabolite concentrations using ERETIC [27] or dieretic [29] and more refined approaches for data analysis, possibly involving curated metabolic pathways, should be explored in the future when larger datasets of microarray and HR MAS MRS data from the same tumor samples can be obtained, with corresponding clinical information.

MRS and microarray experiments have not previously been performed on the same breast cancer sample from the same patient. A study by Tzika et al. combined gene expression microarrays and HR MAS MRS on the same sample of brain tumor and control biopsies [30]. However, no results were reported of combining these two types of information, except for stating that a number of transcripts correlated well to the measured metabolites.

Breast cancer biology is highly complex, which is reflected at many different molecular levels. Using gene expression microarray and HR MAS MRS data from the very same tumor sample can reduce the biological variance which gives a higher power to study the transcriptional and metabolic levels in a combined approach. Even though HR MAS MRS leaves the tumor tissue intact, the procedure exposes the tissue to several potential stresses, including hypoxic conditions and lack of nutrients by being embedded in a surrounding buffer at 4°C for approximately an hour, as well as high centrifugal force and magnetic field during the HR MAS MRS acquisition. In our parallel study to address this issue, total RNA integrity was not significantly affected by HR MAS MRS (p-value = 0.86), and findings in a similar evaluation in prostate tissue support this result [31]. The pairs of tumor samples from each patient that had or had not been analyzed by HR MAS MRS cluster together (Additional file 3: Plots illustrating the effect of HR MAS MRS on the transcriptome), indicating that patient to patient variation is larger than the effect of HR MAS MRS. Even though 1199 transcripts were defined as differentially expressed (fdr < 0.01) by HR MAS MRS, these transcripts showed a small fold change. In their study of prostate tissue, Santos et al. reported no differential expression caused by HR MAS MRS [31]. However, Santos et al. used cDNA microarray data from unpaired samples to test for differential expression. The patient heterogeneity might have been too high to achieve the power needed to detect possible changes in gene expression in their study.