Background
Breast cancer accounts for 25 % of newly diagnosed cancers and 15 % of cancer deaths among women worldwide [
1]. It is a heterogeneous disease [
2] with high diversity in prognosis and response to treatment. Identification of underlying mechanisms contributing to this heterogeneity may reveal new cancer targets and clinically relevant subgroups and has thus been the focus of many recent studies [
3‐
5].
Searching for genetic features causing the variation in breast cancers, Perou et al. used gene expression analyses followed by hierarchical clustering and defined naturally occurring molecular subtypes [
4,
6]. These subtypes are named basal-like, luminal A, luminal B, Erb-B2+ (Her2 enriched), and normal-like, and are found to be associated with tumor characteristics and clinical outcome; patients with basal-like tumors having the shortest and luminal A the longest relapse-free survival [
6]. A centroid-based method called prediction analysis of microarrays 50 (PAM50), which uses the expression of 50 genes to classify breast cancer into these five intrinsic subtypes was later established and is now broadly implemented [
7].
Proteins are the ultimate cellular effectors of pathways and networks within cells, tissues, and organisms. Although protein levels are dependent on mRNA expression, not all mRNA will be translated into protein and further protein levels are also influenced by protein stability. In a study by Myhre et al. only 22 of 52 quantified breast cancer-related proteins were found to correlate with mRNA expression levels [
8] and similar low levels of correlation have been seen in large scale studies [
9,
10]. Protein expression subtypes of breast cancer could give further understanding of underlying mechanisms causing heterogeneity [
11]. Based on the expression of 171 breast cancer-associated proteins detected by reverse phase protein array (RPPA), six breast cancer subtypes, called RPPA subtypes, have been defined [
5]. Four of these subgroups were in high accordance with the gene expression profiles of the PAM50 subtypes and named accordingly; Basal, Her2, luminal A, and luminal A/B. In addition, two new subgroups were defined; reactive I and reactive II, based on expression of proteins possibly produced by the surrounding microenvironment.
The chemical processes controlled by proteins involve metabolites as intermediates or end-products. In metabolomics, metabolite levels are measured to gather the final downstream information of ongoing cellular processes. Which processes are active at a specific time point is strongly influenced by environmental factors like diet and drugs as well as disease state. Well-established metabolic differences have been observed when comparing cancer cells to normal cells. Cancer cell energy production frequently depends on increased glycolysis and production of lactate from glucose regardless of access to oxygen, in contrast to normal cells which produce pyruvate and lactate in aerobic conditions [
12]. Also, to produce macromolecules/biomass, mitochondrial metabolism is reprogrammed [
13]. Altered metabolism has therefore been included as one of the emerging hallmarks of cancer [
14]. In breast cancer, metabolic differences between cancer tissue and normal adjacent tissue have been studied by the magnetic resonance spectroscopy (MRS) method high-resolution magic-angle spinning (HRMAS) MRS [
15]. Using this technique, metabolic profiles and biomarkers predicting long-term survival for locally advanced breast cancer [
16], node involvement of patients with infiltrating ductal carcinoma [
17], and 5-year survival for ER positive patients [
18] have been identified.
Merging transcriptomics and metabolomics led to the discovery of three luminal A subgroups with distinct metabolic profiles and significant differences within gene set expression in a study by Borgan et al. [
19]. The aim of the current study was to establish clusters of breast cancer based on the metabolic expression using an approach similar to Borgan et al., but in a larger cohort of patients including all PAM50 subgroups. This approach reveals the main metabolic differences between untreated breast tumors. In addition, the combination of the metabolic clusters with transcriptomics and protein expression data provide an opportunity for information gain from each -omics technology, giving further characterization of the defined metabolic clusters.
Discussion
In the present work, metabolite, protein, and gene expression data from 228 breast tumors were combined to search for new insight into the heterogeneity of breast cancer. MR metabolite data was used to derive naturally occurring metabolic clusters, which were further combined with data from the proteomics and transcriptomics levels. We identified three significantly different metabolic clusters, Mc1, Mc2, and Mc3, with significant differences in gene expression and protein expression profiles, but not within PAM50 subgroups. The metabolic clusters could therefore contribute with additional information beyond the intrinsic gene sets for understanding breast cancer heterogeneity.
Of the three metabolic clusters, Mc1 was on a separate branch in the dendrogram indicating that the metabolic profile of this cluster was the most different. This cluster is defined by significantly higher levels of GPC and PCho, two choline-containing metabolites involved in the synthesis and degradation of phosphatidylcholine (PtdCho), a major component of cell membranes [
31]. Altered choline metabolism has been considered an emerging hallmark for malignant transformations and has been detected in several cancer types including breast cancer [
32]. PCho in particular has been suggested a biomarker of breast cancer [
33]. Both GPC and PCho are confirmed elevated in tumor tissue compared to adjacent non-involved tissue from breast cancer patients [
17], and a higher GPC/PCho-ratio has been reported in ER negative tumors [
34,
35]. The latter was also observed for our cohort (results not shown); however, there was no significant difference in ER status between the three metabolic clusters. Thus, the high level of GPC and PCho is not resulting from differences in the distribution of estrogen receptor (ER) status. Interestingly, integrated pathway analysis showed that glycerophospholipid metabolism was the most significant pathway, when comparing Mc1 to Mc2. This metabolic pathway had eight hits including the metabolites GPC and PCho and genes
LCAT,
LPCAT2,
PPAP2A,
PPAP2B,
PLD1, and
AGPAT4. Downregulation of the expression of these genes in Mc1 indicate a less active degradation of PtdCho causing an accumulation of GPC and PCho, thus explaining the higher levels of GPC and PCho in Mc1. Furthermore,
LPCAT2 is involved in the reaction where the GPC precursor (acyl-GPC) is converted into PtdCho. Lower expression of this gene may explain why the GPC precursor is directed to the production of GPC instead of PtdCho. The same hits were obtained when Mc1 was compared to Mc3. In addition,
PLA2G5, one of the enzymes degrading PtdCho to acyl-GPC, is downregulated in Mc1 compared to Mc3, further supporting that Mc1 has an altered PtdCho metabolism.
The levels of PCho and GPC were higher in Mc1 compared to the two other clusters, but no significant difference in the expression of choline kinase alpha (
CHKA) could be detected in the SAM analysis. However, univariate analysis confirmed that
CHKA expression was significantly higher in Mc1. This is in agreement with previous findings revealing a positive correlation between levels of PCho and GPC and expression of
CHKA [
34,
36]
.
For Mc1 compared to Mc2 through integrated pathway analysis,
d-glutamine and
d-glutamate metabolism has only two hits, but comes out as significant because of the small number of genes and metabolites within this pathway. Interestingly, the gene
GLS which catalyzes the conversion of glutamine to glutamate is downregulated in Mc1, the cluster with lowest levels of glutamate. Glutamine metabolism is considered a therapeutic target as some cancer cells exhibit high uptake and addiction to this nonessential amino acid [
37]. Since there were no differences in glutamine levels of Mc1 and Mc2, less glutamate in Mc1 could indicate that more glutamine is directed towards other metabolic pathways necessary for proliferation, glutathione needed for reducing power or further that glutamate is rapidly metabolized in cells through the TCA cycle or other mechanisms.
The distribution of protein subtypes (RPPA) was significantly different between the metabolic clusters, whereas no significant differences in the distribution of PAM50 subtypes were found. Thus, the metabolic difference between Mc1, Mc2, and Mc3 is not a result of intrinsic subtypes and might therefore contain additional information for understanding breast cancer heterogeneity. Among the tumors clustered in Mc1, 12 % were classified as RPPA-reactive (either I or II) while 49 % were classified as RPPA-luminal. The reactive RPPA subtypes have a characteristic protein expression pattern probably produced by the microenvironment [
5], indicating less microenvironmental activity within Mc1. Mc1 also had downregulation of several genes involved in processes within the ECM of the stroma compared to both Mc2 and Mc3. As ECM changes can drive cancer behavior [
38], these genetic differences between Mc1 and Mc2 might be of prognostic relevance. In fact, differences in expression of ECM-related genes have been used to stratify breast carcinomas into four groups, where the subgroup ECM1 have the worst prognosis [
39]. ECM classification was not performed on this cohort. However, 34 of 43 genes that clustered with a tendency of being downregulated in ECM1 and ECM2 were also found to be downregulated in Mc1. In addition, only 5 of 46 genes reported to be downregulated in ECM2 compared to ECM1 were downregulated in Mc1 (results extracted from SAM analyses, Additional file
2: Table S6–S7). These results support the contention that Mc1 tumors have an ECM signature similar to the reported ECM2 tumors. ECM2 did not show significant difference in disease outcome compared to ECM3 and ECM4, but had better prognosis than ECM1 tumors [
39].
Mc2 has a metabolic profile with significant higher glucose level and at the same time lower levels of most of the other metabolites compared to one or both of the remaining clusters. High glucose level could reflect lower glucose consumption, inferring a lower demand for energy within these tumors. Glycolysis/gluconeogenesis came out as a significant pathway when Mc1 was compared to Mc2 during integrated pathway analysis with two metabolite hits and five gene hits. For the most significant metabolite, glucose, the levels are higher in Mc2 compared to Mc1. Glucose is the main source of energy for mammalian cells, either through aerobic glycolysis (production of lactate even in the presence of oxygen) or tricarboxylic acid (TCA) cycle and oxidative phosphorylation. For normal proliferating cells and cancer cells, which both have an increased energy demand, a glycolytic switch is often observed (higher glycolytic rate) [
12]. The increased glycolysis is followed by fermentation of the pyruvate to lactate (Warburg effect), in contrast to the conversion of acetyl CoA through the TCA cycle that occurs in normal non-proliferating cells. Increased glucose consumption is commonly used in tumor detection by using a glucose analogue and positron emission tomography (PET) [
40] and has shown to correlate with poor prognosis and tumor aggressiveness [
12]. However, not all breast cancers are detected by PET. Here, we expect lower sensitivity in detection of Mc2 tumors due to the possible difference in glycolytic rate. None of the genes with hits in glycolysis/gluconeogenesis for the comparison of Mc1 and Mc2 could directly explain the high glucose levels of Mc2 tumors, but altered expression of the genes indicates pyruvate being guided towards the TCA cycle rather than lactate production. Two of the alternative fates of pyruvate showed significantly higher levels (alanine) or levels approaching significance (lactate, adjusted
p = 0.056), supporting a higher glycolytic rate in Mc1 and that the pyruvate produced is not directed to metabolism in the TCA cycle. The significantly lower acetate levels in Mc1 compared to Mc2 could be linked to
ALDH1A3 and
ALDH2 downregulation, since the enzymatic product of these genes catalyzes the reversible reaction where acetaldehyde is converted to acetate.
Both DAVID and GSEA showed that many of the genes found to be downregulated in Mc1 and consequently upregulated in Mc2 were related to ECM activity. Mc2 had the highest percentage of RPPA-reactive I with 44 % of Mc2 tumors classified as this protein subtype, also related to stromal changes. Together with the metabolic finding, this implies that Mc2 tumors have cancer cells with low proliferating rate and at the same time ongoing changes within the ECM of the stroma. Mc2 tumors also had a higher frequency of lobular and ductal carcinoma in situ, indicating metabolic differences between histological subtypes of breast cancer which should be further investigated.
Mc3 has the highest lactate levels of all three clusters and higher glycine level than Mc2. These metabolites have been related to poor prognosis in ER positive patients [
18], and higher levels of glycine is also associated with poor prognosis in a study irrespective of ER status [
41]. Although the ER-positive patients are equally distributed among our reported metabolic clusters, Mc3 expressed higher levels of both of these metabolites compared to Mc2. Moestue et al. detected differences in the expression of genes involved in choline degradation that could explain higher glycine concentrations in the poor-prognosis basal-like breast cancer xenograft model compared to luminal-like [
42]. Five of the genes described by Moestue et al. were significantly upregulated in Mc3 compared to Mc1;
AGPAT4,
PPAP2B,
PPAP2A,
LCAT, and
PLD1. Of these,
LCAT and
PLD1 are directly involved in choline metabolism.
LCAT catalyze the conversion of PtdCho to acyl-GPC while
PLD1 catalyzes the conversion of PtdCho to choline. Higher GPC levels, but no difference in choline levels in Mc3 compared to Mc1 indicates that a higher amount of GPC is converted to choline in Mc3, and further contributing to higher glycine levels through choline degradation.
Mc3 shares similarities with a previously reported metabolic subgroup of luminal A tumors with significantly lower levels of glucose, higher levels of alanine, and nearly significantly higher lactate levels [
19]. In Mc3, we also see a significant higher level of lactate. Since one of the main sources of alanine is pyruvate, which also is the source for lactate, it appears that Mc3 is a cluster with a switch in glycolytic activity.
The majority of Mc3 tumors were classified as RPPA-luminal, similar to Mc1. In contrast to Mc1, Mc3 had a higher percentage of RPPA–reactive II tumors, probably linked to changes in stromal content. Also, gene expression wise, this was observed by significantly different gene expressions linked to ECM activity and the gene expression profile of Mc3 was found similar to the previously reported ECM3 or ECM4 subtypes [
39].
In this study, information flow between the transcriptomics, proteomics, and metabolomics levels is illustrated; at the transcriptomics level, only one of the metabolic clusters shows difference in gene expression compared to the two others, while at the proteomics level, there is difference between all three clusters. Combining these findings, Mc1 is expected to have the worst prognosis due to the distinct gene expression profile and the alterations in both glycerophospholipid metabolism and evidence of increased glycolytic rate. However, this has to be validated when 5-year follow-up of this cohort is available. The main metabolic characteristics, especially of Mc1 and Mc3, have been proposed as treatment targets that could improve the therapeutic effect [
43]. Cancer therapy targeting
CHKA, the enzyme responsible for PCho production from choline, causes tumor growth arrest and apoptosis in preclinical models [
44], while treatment targeting glycolytic enzymes in combination with chemotherapy has been shown to re-sensitize cancer cells that had become resistant to treatment [
43]. Metabolic classification as illustrated here could therefore be relevant for developing a more targeted treatment plan. Importantly, the prognostic value of the clusters should be evaluated once 5-year follow-up is available.
Abbreviations
Ala, alanine; Cr, creatine; DCIS, ductal carcinoma in situ; ECM, extracellular matrix; ER, estrogen receptor; FDR, false discovery rate; Gly, glycine; GPC, glycerophosphocholine; GSEA, gene set enrichment analysis; HCA, hierarchical cluster analysis; HR MAS MRS, high resolution magic angle spinning magnetic resonance spectroscopy; Lac, lactate; Mc, metabolic cluster; MRS, magnetic resonance spectroscopy; MSigDB, molecular signatures database; PAM50, prediction analysis of microarrays 50; PCho, phosphocholine; PET, positron emission tomography; PLS-DA, partial least square discriminant analysis; PtdCho, phosphatidylcholine; RPPA, reverse phase protein array; SAM, significance analysis of microarrays; Tau, taurine; TCA, tricarboxylic acid; β-Glc, β-glucose
Acknowledgements
The HR MAS MRS analysis was performed at the MR Core Facility, Norwegian University of Science and Technology (NTNU). MR core facility is funded by the Faculty of Medicine at NTNU and Central Norway Regional Health Authority.
The Oslo Breast Cancer Research Consortium (OSBREAC).
Vessela N Kristensen1,2,3, Torill Sauer4,5, Elin Borgen6, Olav Engebråten7,8,9, Øystein Fodstad7,9, Rolf Kåresen9,10, Bjørn Naume2,8, Gunhild Mari Mælandsmo2,7,11, Hege G Russnes1,2,12, Therese Sørlie1,2, Helle Kristine Skjerven13, Britt Fritzman14.
1Department of Cancer Genetics, Institute for Cancer Research, Oslo University Hospital, Oslo, Norway. 2K.G. Jebsen Centre for Breast Cancer Research, Institute for Clinical Medicine, University of Oslo, Oslo, Norway. 3Department of Clinical Molecular Biology and Laboratory Science (EpiGen), Division of Medicine, Akershus University Hospital, Lørenskog, Norway. 4Department of Pathology, Akershus University Hospital, Lørenskog, Norway. 5Institute of Clinical Medicine, Faculty of Medicine, University of Oslo, Oslo, Norway. 6Department of Pathology, Division of Diagnostics and Intervention, Oslo University Hospital, Oslo, Norway. 7Department of Tumor Biology, Institute for Cancer Research, Oslo University Hospital, Oslo, Norway. 8Department of Oncology, Division of Surgery and Cancer and Transplantation Medicine, Oslo University Hospital, Oslo, Norway. 9Institute of Clinical Medicine, Faculty of Medicine, University of Oslo, Oslo, Norway. 10Department of Breast- and Endocrine Surgery, Division of Surgery, Cancer and Transplantation, Oslo University Hospital, Oslo, Norway. 11Department of Pharmacy, Faculty of Health Sciences, University of Tromsø, Tromsø, Norway. 12Department of Pathology, Oslo University Hospital, Oslo, Norway. 13Breast and Endocrine Surgery, Department of Breast and Endocrine Surgery, Vestre Viken Hospital, Drammen, Norway. 14Østfold Hospital, Østfold, Norway.
Email: Vessela N Kristensen v.n.kristensen@medisin.uio.no, Torill Sauer Torill.sauer@medisin.uio.no, Elin Borgen ebg@ous-hf.no, Olav Engebråten Olav.engebraten@medisin.uio.no, Øystein Fodstad Oystein.Fodstad@rr-research.no, Rolf Kåresen rolf.karesen@medisin.uio.no, Bjørn Naume bjorn.naume@medisin.uio.no, Gunhild Mari Mælandsmo Gunhild.Mari.Malandsmo@rr-research.no, Hege G Russnes Hege.russnes@rr-research.no, Therese Sørlie therese.sorlie@rr-research.no, Helle Kristine Skjerven Helle.skjerven@vestreviken.no, Britt Fritzman Britt.Fritzman@so-hf.no