Targeting Unique Metabolic Properties of Breast Tumor Initiating Cells

Abstract

Normal stem cells from a variety of tissues display unique metabolic properties compared to their more differentiated progeny. However, relatively little is known about metabolic properties of cancer stem cells, also called tumor initiating cells (TICs). In this study we show that, analogous to some normal stem cells, breast TICs have distinct metabolic properties compared to nontumorigenic cancer cells (NTCs). Transcriptome profiling using RNA-Seq revealed TICs underexpress genes involved in mitochondrial biology and mitochondrial oxidative phosphorylation, and metabolic analyses revealed TICs preferentially perform glycolysis over oxidative phosphorylation compared to NTCs. Mechanistic analyses demonstrated that decreased expression and activity of pyruvate dehydrogenase (Pdh), a key regulator of oxidative phosphorylation, plays a critical role in promoting the proglycolytic phenotype of TICs. Metabolic reprogramming via forced activation of Pdh preferentially eliminated TICs both in vitro and in vivo. Our findings reveal unique metabolic properties of TICs and demonstrate that metabolic reprogramming represents a potential therapeutic strategy for targeting these cells. Stem Cells 2014;32:1734–1745

Introduction

Cells produce energy primarily through two interconnected processes: glycolysis (the catabolism of glucose to pyruvate) and oxidative phosphorylation (the production of energy from glucose via the mitochondrial tricarboxylic acid cycle). Over 85 years ago, the German biochemist Otto Warburg discovered that unlike normal tissues, cancer cells in many kinds of tumors preferentially perform glycolysis in the presence of oxygen as opposed to oxidative phosphorylation 1, 2. His observations, known as the “Warburg effect,” have been repeatedly confirmed but remain incompletely understood 3.

At Warburg's time and for many years thereafter, the prevailing model of tumor biology posited that all cancer cells are equivalent in their ability to proliferate indefinitely. However, in recent years it has become clear that many tumors contain a phenotypically unique subset of cancer cells that appears to be the sole population able to divide indefinitely or to form new tumors. These data have led to the formulation of the cancer stem cell (CSC) hypothesis, which proposes that cancers can be viewed as abnormal organs composed of self-renewing CSCs (a.k.a. tumor initiating cells (TICs) or tumorigenic cancer cells) and their descendent, more differentiated nontumorigenic cancer cells (NTCs) which have limited proliferative potential 4. Published xenograft studies, together with recent lineage tracing studies, have documented the existence of TICs in breast cancers 4, 5 and many other tumor types 6-19. As is the case with normal stem cells, TICs cannot be isolated as entirely pure populations but can be highly enriched compared to NTC populations using cell sorting strategies. Published TIC frequencies in enriched populations from primary tumors vary from 1 in 100 to 1 in 10,000, although are actually likely higher due to cell damage induced by dissociation and cell sorting and inefficiencies of cell transplantation.

In trying to explain the preferential glycolysis he observed in cancer cells, Warburg noted that normal tissues undergo a switch from glycolysis to oxidative phosphorylation during embryogenesis and he therefore equated carcinogenesis with “dedifferentiation” of mature normal cells to more immature malignant cells 20. Based on new insights afforded by the CSC hypothesis, we suggest that instead of “dedifferentiation,” the Warburg effect may arise in part from preferential glycolysis occurring in immature tumor cells (i.e., TICs) that have not differentiated. This concept is in line with well-documented clinical observations of elevated glucose uptake by different types of poorly differentiated tumors compared to well differentiated tumors 21, 22, which may reflect an increased proportion of cancer stem/progenitor cells in poorly differentiated tumors.

We and others have previously found that TIC-enriched population in a variety of tumor types contain lower levels for reactive oxygen species (ROS) than NTCs 23-25 and part of the mechanism for this difference is elevated levels of free radical scavengers in TIC-enriched population. However, another potential mechanism for decreased levels of ROS is decreased production of ROS, which has not been previously explored. Since electron transport during oxidative phosphorylation in mitochondria is the main source of endogenous ROS 26, we hypothesized that TICs preferentially perform glycolysis over oxidative phosphorylation. Intriguingly, long-term hematopoietic stem cells preferentially use glycolysis instead of oxidative phosphorylation for energy production 27, 28. Similarly, embryonic stem cells preferentially perform glycolysis and undergo a switch to oxidative phosphorylation as they differentiate 29, 30.

We set out to test the hypothesis that TICs preferentially perform glycolysis compared to NTCs. We found that TIC-enriched population rely more heavily on glycolysis than NTCs and this is reflected in the gene expression programs of the two cell types. Differences in mitochondrial number and expression of the key mitochondrial enzyme pyruvate dehydrogenase (Pdh) underlie the unique metabolic properties of TICs and forced upregulation of oxidative phosphorylation by activation of Pdh preferentially targets TICs. Our findings suggest that promotion of oxidative phosphorylation is a promising strategy for targeting TICs.

Materials and Methods

Cell Labeling, Flow Cytometry, Cell Sorting, Three-Dimensional Cancer Spheroid Assay, Transplantation, and Limited Dilution Assay

MMTV-Wnt-1 murine breast tumors were harvested, dissociated into single cell suspensions, sorted and processed according to previously described methods 5, 23. Details were also shown in Supporting Information Materials and Methods.

Dichloroacetic Acid Treatment of Tumor Bearing Animals

In vivo treatments were initiated when transplanted tumor reached a size of approximately 0.5 cm diameter. Sodium dichloroacetate (Thermo Fisher, Waltham, MA, http://www.thermofisher.com, #338280100) dissolved in phosphate buffered saline was intraperitoneally injected to tumor-bearing mice once a day until control animals had to be euthanized due to tumor size.

RNA-Seq and Gene Set Enrichment Assay

Methods for generating and analyzing the RNA-Seq data are available in the Supporting Information Methods. Sequencing data are accessible through Gene Expression Omnibus (GSE41286).

Pdh Activity Assay

300,000 cells were sorted for Pdh activity analysis with the MitoProfile Dipstick Assay Kit (MitoScienes, Eugene, OR, http://www.mitosciences.com, #MSP30) according to manufacturer's instruction. Pdh activity was quantified by the intensity of the band using a digital imaging system.

Metabolic Assays

BD Oxygen Biosensor plates (BD Bioscience, San Diego, http://www.bdbiosciences.com, #353830) were used for measurement of oxygen consumption. Cells were incubated at 37°C with or without dichloroacetic acid (DCA) for 12 hours, and oxygen consumption was determined by measuring the fluorescent intensity at 630 nm. From the same wells, 10 µL supernatant was collected for lactate concentration assessment. Lactate was measured using a commercial kit from Trinity Biotech (Bray, County Wicklow, Ireland, http://www.trinitybiotech.com, #735-10) according to manufacturer instruction. Additional details are available in Supporting Information Methods.

Statistical Analyses

Replicate numbers represent biological replicates using distinct tumors from separate mice. Levels of significance were determined by Student's t tests using α = .05.

Results

Optimization of TIC Isolation

We have previously reported that the expression of Thy1, CD24, and CD49f enables the prospective purification of TICs from MMTV-Wnt-1 mammary tumors 5, 23. In order to enable isolation of enough TICs to perform metabolic and biochemical analyses of TICs and NTCs, we set out to further optimize isolation of TICs from these tumors. To begin we examined expression of basal (Krt14 and Trp63) and luminal (Krt8 and Krt18) markers in MMTV-Wnt-1 mammary tumors by immunofluorescence (Supporting Information Fig. S1A and data not shown) and confirmed that these tumors contain both luminal- and basal-like cells. Given that TICs often display a mesenchymal- or basal-like phenotype, we next searched for surface markers that would allow robust separation of basal and luminal breast cancer cells. We specifically focused on CD49f and Epcam, which had previously been used to identify luminal and basal-like cells in human mammary epithelium 31. Expression of these two markers separated malignant cells into two distinct populations: CD49f^high Epcam^low cells and CD49f^low Epcam^high cells (Fig. 1A).

Figure 1

CD49f^highEpcam^low cells in MMTV-Wnt-1 mammary tumors are highly enriched for tumor initiating cells. (A): Fluorescence-activated cell sorting analysis of a representative MMTV-Wnt-1 mammary tumor reveals distinct CD49f^lowEpcam^high and CD49f^highEpcam^low populations. (B): Expression of basal and luminal breast markers in purified CD49f^highEpcam^low and CD49f^lowEpcam^high cells using quantitative reverse transcriptase polymerase chain reaction. (C): In vivo limiting dilution analysis of purified CD49f^highEpcam^low and CD49f^lowEpcam^high cells. Sorted cells were injected into syngeneic recipients at the cell dosages indicated and animals were observed for tumor formation. (D): Representative sorting gates to further subdivide the CD49f^highEpcam^low population. Abbreviations: FITC, Fluorescein isothiocyanate; PE, Phycoerythrin; TIC, tumor initiating cell.

In order to confirm the differentiation state of the two populations, we next examined expression of basal and luminal markers by quantitative reverse transcriptase polymerase chain reaction. CD49f^high Epcam^low cells expressed genes found in mammary basal stem/progenitor cells, including Trp63, Krt14, Vim, and Acta2, while CD49f^low Epcam^high cells overexpressed markers associated with luminal differentiation, including Esr1, Krt18, and Krt19 (Fig. 1B). This expression pattern confirmed that CD49f^high Epcam^low cells displayed a basal-like phenotype and suggested that this population is enriched for TICs.

We next tested the tumor initiating ability of the two populations by transplanting double-sorted CD49f^high Epcam^low and CD49f^low Epcam^high cells into syngeneic mice in a limiting dilution fashion. As shown in Figure 1C, 1 in 79 (95% confidence interval, 1/40 to 1/154) CD49f^high Epcam^low cells could initiate tumors and this population was approximately 300-fold enriched for tumor initiating activity compared to CD49f^low Epcam^high cells (p < .0001). This represents a significant enrichment over the 1 tumorigenic cell in greater than 200 cells observed with our previous isolation strategy based on CD24, Thy1, and CD49f 5 which did not separate TICs and NTCs as clearly (Supporting Information Fig. S1D, S1E) and is among the highest enrichment for solid tumor TICs in primary tumors that has been reported.

We next tested whether subpopulations within the CD49f^high Epcam^low cells were even more highly enriched for TICs. For example, further stratifying this population based on Epcam expression did not result in significant enrichment of tumor initiating activity (Fig. 1D; Supporting Information Table S1; p = 0.20). Similarly, sorting based on expression of CD24 and Thy1 did not yield additional functional enrichment (data not shown). This indicates that TICs are distributed equally throughout the CD49f^high Epcam^low subpopulation and we therefore performed subsequent experiments using CD49f^high Epcam^low cells as the TIC-enriched population and CD49f^low Epcam^high cells as NTCs. Flow cytometric analysis of tumors established from purified CD49f^high Epcam^low cells revealed similar immunophenotypic profiles as the original tumors, including the presence of both CD49f^high Epcam^low and CD49f^low Epcam^high cells (Supporting Information Fig. S1B, S1C). CD49f^high Epcam^low cells isolated from secondary and tertiary tumors retained tumor initiating potential in repeat transplants (data not shown). Thus, CD49f^high Epcam^low cells are highly enriched for TICs and can both self-renew and give rise to nontumorigenic CD49f^low Epcam^high cells. For simplification, we will call this population as TIC for this mouse mammary tumor model in this study.

Transcriptome Profiling of TIC by RNA-Seq

To begin exploring potential metabolic differences between TICs and NTCs, we purified CD49f^high Epcam^low and CD49f^low Epcam^high cells by flow cytometry and performed paired-end RNA sequencing. Sequencing libraries were of high quality and produced an average of over 112 million reads per sample of which approximately 83% could be mapped to the reference genome. While many genes were similarly expressed in the two populations, a significant number was overexpressed by either TICs or NTCs (Supporting Information Fig. S2A, S2B). As expected, CD49f (Itga6) was overexpressed by TIC-enriched CD49f^high Epcam^low cells while Epcam was overexpressed by CD49f^low Epcam^high cells. Known markers of basal/stem cells including Acta2, Vim, Trp63, and Krt5 were overexpressed in the TIC-enriched population while known luminal markers such as Krt18, Krt19, and Esr1 were overexpressed in the NTC population (Fig. 2A; Supporting Information Fig. S2C). Thus, our RNA-Seq data faithfully recapitulated known expression differences between the two tumor cell subpopulations.

Figure 2

Comprehensive transcriptome analysis by RNA-Seq reveals decreased expression of oxidative phosphorylation and mitochondrial genes in TICs compared to NTCs. (A): Relative expression of selected basal/stem and luminal cell markers in TICs and NTCs as measured by RNA-Seq. (B): Random-set enrichment scoring of gene sets from Molecular Signatures Database. Locations of several key gene sets are indicated. (C): Gene set enrichment analysis of metabolism-related gene sets. Analysis includes 11,433 well-measured genes. Vertical bars represent genes from each of the indicated gene sets. (D): Gene expression heat map of tricarboxylic acid cycle enzymes in TICs and NTCs. Abbreviations: GSZ score, gene set z score; NTC, nontumorigenic cancer cell; Ox. Phos., oxidative phosphorylation; TIC, tumor initiating cell.

Next we asked if predefined gene sets, including gene sets related to metabolic processes, were significantly enriched in either TICs or NTCs. To begin, we used random set enrichment scoring to identify gene sets from the Molecular Signatures Database (MSigDB-c2) that were enriched in TICs or NTCs 32. As previously described for human TICs 33, we identified gene sets related to epithelial mesenchymal transition among the most significantly enriched gene sets in TICs. Strikingly, we identified multiple gene sets related to mitochondria and oxidative phosphorylation as being highly enriched in NTCs, suggesting that NTCs may preferentially perform oxidative phosphorylation compared to TICs (Fig 2B; p < 1 × 10⁻⁵). The fact that metabolism-related gene sets were among the most enriched gene sets suggested that metabolic differences might be a fundamental property that distinguishes the two types of cancer cells. In order to further extend this observation, we assessed the expression of additional, curated metabolism-associated gene sets in TICs and NTCs by gene set enrichment analysis (GSEA) 34, 35. This analysis confirmed the significant enrichment of genes involved in oxidative phosphorylation and revealed additional enrichment of nuclear-encoded mitochondrial genes in NTCs (Fig. 2C; Supporting Information Fig. S2D). Similarly, analysis of key enzymes involved in the TCA revealed that the majority were overexpressed in NTCs compared to TICs (Fig. 2D). Glycolytic genes were not significantly overexpressed by either subpopulation but trended toward being overexpressed by TICs. Lack of statistical enrichment of glycolytic genes is not surprising since pyruvate produced by glycolysis is also a key carbon source for mitochondrial oxidative phosphorylation, and therefore glycolytic genes are expressed in cells undergoing oxidative phosphorylation. NTCs also displayed significant overexpression of genes involved in proliferation, suggesting that a higher fraction of these cells are cycling. Thus, deep sequencing of breast TIC and NTC transcriptomes revealed differences in metabolic programs between the two cell types.

Mitochondrial Differences in TICs and NTCs

Elevated expression of nuclear-encoded mitochondrial genes involved in oxidative phosphorylation by NTCs suggested that there were differences in mitochondrial metabolism between TICs and NTCs. To test this possibility, we first assessed the number of mitochondria present in TICs and NTCs by using transmission electron microscopy and found that TICs contained fewer mitochondria than NTCs (Fig. 3A, 3B, p = .03). Comparison of the relative amounts of mitochondrial DNA (mtDNA) to nuclear DNA can also be used to assess relative mitochondrial numbers. Therefore, we determined the ratio of mtDNA to nuclear DNA in the two cell populations using quantitative PCR. The ratio of mtDNA to nuclear DNA was lower in TICs than NTCs, consistent with our ultrastructural analyses (Fig. 3C, p = .02). Finally, we also examined expression of mitochondria-encoded genes in the two cell populations. Transcripts for Cox1, Cox2, and Cytb were significantly overexpressed in NTCs compared to TICs (Fig. 3D).

Figure 3

TICs contain fewer mitochondria than NTCs. (A): Ultrastructural analysis of sorted TICs and NTCs using transmission electron microscopy. Representative images are shown at magnification of ×2,500 in upper panels and ×5,000 in lower panels. Arrows in lower panel indicate examples of mitochondria. (B): Quantification of mitochondrial numbers per section in TIC and NTCs. Results were normalized to the average number of mitochondria in TICs (n = 8; p = .03). (C): Real time quantitative polymerase chain reaction (PCR) analysis of the ratio of mitochondrial Cox1 DNA to nuclear beta-actin DNA. Results were normalized to TICs (n = 4; p = .02). (D): Real time quantitative PCR analysis of RNA expression of the mitochondria-encoded genes Cox1, Cox2, and Cytb. Results were normalized to TICs (n = 4; p = .0002, .01, and .05, respectively). (E): Analysis of mitochondrial activity using the fluorescent dye JC-1. Abbreviations: FITC, Fluorescein isothiocyanate; mtDNA, mitochondrial DNA; NTC, nontumorigenic cancer cell; PE, Phycoerythrin; TIC, tumor initiating cell.

Our ultrastructural analyses also suggested preferential presence of activated mitochondria in NTCs, including clustering of mitochondria, thickened cristae, and “wagon wheel” morphology 36, while little evidence of activation was seen in TICs. To confirm this, we examined mitochondrial membrane potential (ΔΨm) in the two cell populations. Cells that are proglycolytic and contain less active mitochondria display higher ΔΨm than cells performing more oxidative phosphorylation 37. Consistent with our ultrastructural observations, TICs displayed significantly hyperpolarized ΔΨm compared to NTCs as measured by increased Phycoerythrin/Fluorescein isothiocyanate (PE/FITC) ratio of the ΔΨm-sensitive probe JC-1 (Fig. 3E, p < .002). Thus, TICs contain fewer and less active mitochondria than NTCs, suggesting that they display a proglycolytic phenotype.

TICs Display Proglycolytic Metabolism

To directly evaluate glucose metabolism, we compared the balance of glycolysis to oxidative phosphorylation in the two subpopulations by measuring lactate production and oxygen consumption. TIC-enriched cells displayed a significantly higher ratio of lactate production to oxygen consumption than NTC-enriched cells (p = .006; Fig. 4A). To further extend this finding, we compared the glucose transport rates in TICs and NTCs, since proglycolytic cells display higher rates of glucose import than cells relying mostly on oxidative phosphorylation for energy production. Consistent with their proglycolytic metabolic profile, TICs displayed significantly higher glucose uptake than NTCs (Fig. 4B, p = .01). Thus, TICs in MMTV-Wnt-1 mammary tumors are proglycolytic compared to the remaining cancer cells.

Figure 4

TICs display a proglycolytic phenotype. (A): Lactate production and oxygen consumption in TICs and NTCs (n = 3; p = .006). (B): MFI in TICs and NTCs after staining for 40 minutes with 10 µM 2-NBDG (n = 3; p < .01). (C): Representative cell cycle profiles of TICs and NTCs. The percentage of cells in G0–1, S, and G2-M are indicated. (D): Percentage of TICs and NTCs in G0/G1, S, and G2-M fractions. On average, 3.7% of TICs versus 24.2% of NTCs were found in S-G2-M (n = 2; p = .005). Data expressed as mean ± SEM. (E): Double staining of Ki67 (red) and K8/18 (green) in a representative MMTV-Wnt1 mammary tumor. Lower panel displays higher magnification of the region in the white box in the upper panel. (F): Distribution of Ki67 staining among subsets of breast cancer cells. Data expressed as mean ± SEM. Abbreviations: MFI, mean fluorescence intensity; 2-NBDG, 2-[N-(7-Nitrobenz-2-oxa-1,3-diazol-4-yl)amino]-2-deoxy-D-glucose; NTC, nontumorigenic cancer cell; TIC, tumor initiating cell.

Since the pioneering work of Otto Warburg, it has been recognized that highly proliferative cells often preferentially perform glycolysis compared to oxidative phosphorylation. However, TICs are generally thought to represent a relatively quiescent subpopulation of tumor cells and our GSEA analyses seemed to confirm this. To directly address this question, we performed cell cycle analysis on purified CD49f^high Epcam^low and CD49f^low Epcam^high cells. Strikingly, the TIC-enriched population contained a significantly smaller fraction of cells in S-G2-M compared with the NTC-enriched population (3.7% versus 24.2%, respectively; p = .005; Fig. 4C, 4D). This result was confirmed by immunofluorescence analysis of Ki67 and Krt8/18 expression in MMTV-Wnt-1 tumors, which demonstrated that the majority of Ki67 positive cells expressed luminal cytokeratins (Fig. 4E, 4F, p = .002). Thus, TICs are relatively quiescent compared to NTCs and their proglycolytic phenotype is an intrinsic feature of these cells and not simply a reflection of more rapid proliferation.

Decreased Pdh Expression and Activity in TICs

We next sought potential regulators of the metabolic differences between TICs and NTCs and focused on the Pdh complex, which was overexpressed by NTCs in our RNA-Seq data (Fig. 2D). Pdh converts pyruvate, the final product of glycolysis, to acetyl-CoA and thus serves as the key link between the cytosolic glycolytic pathway and mitochondrial oxidative phosphorylation. Genes composing the Pdh complex (Pdha1, Pdhb, and Pdhx) were approximately twofold overexpressed in NTCs compared to TICs (Fig. 5A), suggesting that higher Pdh activity may contribute to the enhanced oxidative phosphorylation observed in NTCs. Analysis of Pdh complex protein levels in sorted TICs and NTCs revealed that protein expression was approximately twofold decreased in TICs (Fig. 5B). Consistent with the observed expression differences, Pdh enzymatic activity was also approximately twofold higher in NTCs, implying decreased import of pyruvate into mitochondria in TICs compared to NTCs (Fig. 5C, 5D). Endogenously, the Pdh complex is negatively regulated by the pyruvate dehydrogenase kinase (Pdk) family, whose members phosphorylate the Pdh complex at multiple serine residues, thus resulting in enzyme inhibition. We observed roughly equivalent Pdh phosphorylation at Ser 232 in both TICs and NTCs (Fig. 5B), suggesting that the observed difference in Pdh activity is mostly due to differences in protein abundance.

Figure 5

Decreased expression and activity of pyruvate dehydrogenase in TICs and NTCs. (A): Real time quantitative polymerase chain reaction (PCR) analysis of pdh subunit RNA expression in TICs and NTCs (n = 3; p < .01). (B): Representative Western blot analysis of Pdh protein expression and phosphorylation in TICs and NTCs (n = 3). (C): Pdh enzyme activity in TICs and NTCs using dipstick assay (see Materials and Methods). (D): Densitometric quantification of bands from Pdh activity assay (n = 3; p < .01). Abbreviations: NTC, nontumorigenic cancer cell; Pdh, pyruvate dehydrogenase; TIC, tumor initiating cell.

Targeting TICs by Metabolic Reprogramming

Given the observed preference for glycolysis and decreased expression and activity of Pdh in TICs, we hypothesized that activation of Pdh activity might specifically target TICs. Pdh activity can be increased by inhibiting Pdk family members using DCA, a small molecule inhibitor that acts as a pyruvate mimetic 38-40. DCA has previously been shown to promote oxidative phosphorylation in bulk cancer cells that display the Warburg effect 37, 41, 42 and we hypothesized that DCA may similarly reprogram the metabolism of TICs. Using a three-dimensional cancer spheroid culture system, we found that treatment of TICs with DCA resulted in a dose dependent decrease of TIC clonogenicity (Fig. 6A, 6B). By comparison, NTCs, which are unable to form tumors in vivo but can give rise to colonies in vitro, were significantly less sensitive to DCA. In order to confirm that DCA treatment resulted in the desired metabolic effects, we measured levels of lactate production and oxygen consumption before and after exposure to the drug. DCA resulted in a significant shift toward oxygen consumption, an effect that was significantly more dramatic in TICs than NTCs given their initially higher levels of lactate production (Fig. 6C).

Figure 6

Glycolysis inhibitors selectively target TICs in vitro and in vivo. (A): Representative images of spheroid formation in presence and absences of 12.5 mM DCA. (B): Dose-dependent inhibition of TIC and NTC spheroid formation by DCA (n = 3, *, p < .01). (C): Ratio of lactate production to oxygen consumption in TICs and NTC in the presence and absence of DCA (n = 3; p < .01). (D): Significant inhibition of TIC, but not NTC, spheroid formation by 2-DG and sodium oxamate (n = 5). Representative images of spheroid formation are shown in Supporting Information Fig. S3A. (E): Growth curves of MMTV-Wnt-1 mammary tumors treated with DCA. Tumors were treated with vehicle or 75 mg/kg DCA once they reached a size of 5 mm. (F): Percent change of TIC-enriched (CD49f^high Epcam^low Lin−) and NTC (CD49f^low Epcam^high Lin−) subpopulations in DCA treated-tumors compared to vehicle-treated tumors as assessed by flow cytometry (n = 3; p < .02). Abbreviations: DCA, dichloroacetic acid; 2-DG, 2-deoxy-D-glucose; NTC, nontumorigenic cancer cell; TIC, tumor initiating cell.

To confirm these observations, we used two additional drugs that target glycolysis and tested if they preferentially target TICs over NTCs. Specifically, we tested the sensitivity of the two cell types to sodium oxamate, which inhibits lactate dehydrogenase and thus makes more pyruvate available for oxidative phosphorylation, and 2-deoxy-d-glucose (2-DG), a glucose analog and well-described glycolytic inhibitor. We found that similar to our results with DCA, TICs were significantly more sensitive to sodium oxamate and 2-DG than NTCs (Fig. 6D; Supporting Information Fig. S3A). Thus, TICs are significantly more dependent on glycolysis than NTCs and inhibition of this metabolic pathway selectively targets TICs.

We next wished to extend our findings to an additional murine mammary tumor model. Therefore, we explored whether TICs from MMTV-PyMT mouse mammary tumors also displayed a proglycolytic phenotype. Based on the recently described characterization of TIC markers in this model 43, we purified CD49f^highCD24⁺ TIC-enriched cells and CD24⁻CD49⁻ NTCs from PyMT tumors. As was the case in MMTV-Wnt-1 tumors, we observed consistently higher expression of mitochondrial and oxidative phosphorylation genes by NTCs (Supporting Information Fig. S4A–S4C) and higher lactate production by TICs (Supporting Information Fig. S4D). Additionally, tumor spheres initiated by CD49f^highCD24⁺ TIC-enriched cells were sensitive to DCA treatment in a dose-dependent manner (Supporting Information Fig. S4E, S4F). Thus, TICs from distinct murine mammary tumors display a proglycolytic phenotype and are sensitive to metabolic reprogramming by DCA.

Given the inhibitory effects of DCA on TICs in vitro we next asked whether promotion of oxidative phosphorylation could also preferentially target TICs in vivo. When mice bearing MMTV-Wnt-1 tumors were treated with DCA, we observed significant tumor growth inhibition (Fig. 6E). In order to assess relative effects on TICs and NTCs, tumors were dissociated, processed, and analyzed by flow cytometry. Compared to control tumors, the abundance of the CD49f^high Epcam^low TIC-enriched population was relatively decreased compared with CD49f^low Epcam^high NTCs in treated tumors (Fig. 6F). Of note, we could observe a similar shift from TICs to NTCs in vitro. While DCA completely inhibited tumor sphere formation when added to TIC cultures at the time of plating (Fig. 6A), if TIC cultures were treated with DCA starting 3 days after plating some spheres survived (Supporting Information Fig. S3B, S3C). Real-time PCR analysis of cells from these spheres revealed an increase of the NTC marker Esr1 and decrease of the TIC marker p63 (Supporting Information Fig. S3D), suggesting that these colonies contained a higher fraction of NTCs after DCA treatment. These findings suggested that promotion of oxidative phosphorylation via DCA preferentially targets and depletes TICs both in vivo and in vitro.

Although CD49f^high Epcam^low cells are highly enriched for TICs, they also contain cells without tumorigenic activity. Thus, it is potentially possible that the metabolic differences we observed are mostly due to non-TICs within the CD49f^high Epcam^low subpopulation. In order to confirm that DCA treatment actually targets TICs, we therefore performed in vivo limiting dilution analyses to measure the fraction of TICs in tumors treated with DCA compared to controls. If DCA targets TICs then the fraction of tumor forming cells should be lower in treated tumors. As shown in Supporting Information Table S3, DCA-treated tumors contained approximately fourfold fewer tumor initiating cells than control tumors. This result confirmed that metabolic reprogramming via DCA directly targeted TICs in vivo.

Metabolic Properties and Reprograming of Human Breast TICs

In order to extend our findings from mouse models to human breast cancer, we explored whether TICs from primary human breast cancer surgical specimens displayed a similar proglycolytic phenotype. We therefore analyzed previously published gene expression data from human TICs (CD44⁺CD24^−/lowLin⁻) and NTCs purified from primary patient specimens 44, 45. We performed GSEA on microarray data from sorted TICs and NTCs from 11 human breast tumors using the same gene sets used in the earlier analyses of murine TICs and NTCs. Congruently, we observed significant enrichment of genes involved in oxidative phosphorylation, mitochondria, and proliferation in human NTCs (Fig. 7A; Supporting Information Fig. S3A–S3D). As in the mouse model, glycolysis genes were not statistically significantly enriched in either cell type, consistent with the glycolytic enzymes being necessary for pyruvate production in both metabolic pathways. Thus, human TICs and NTCs display similar metabolism gene expression profiles as their murine counterparts.

Figure 7

Human breast TICs display proglycolytic phenotypes and are sensitive to DCA. (A): Gene set enrichment analysis of metabolism-related gene sets in microarray data from 11 pairs of TICs and NTCs from primary human breast cancers. Vertical bars represent genes from each of the indicated gene sets. (B): Mitochondrial content in human TICs (CD49f^high Epcam^high) and NTCs (CD49f^low Epcam^low) assessed by MitoTracker (n = 3; p < .001). (C): Real time quantitative polymerase chain reaction analysis of the ratio of mitochondrial Cytb DNA to nuclear beta-actin DNA. Results were normalized to TICs (n = 3; p < .001). (D): Representative images of spheroid formation by human TICs in presence and absences of 25 mM DCA. (E): Dose-dependent inhibition of TIC spheroid formation by DCA (n = 3). (F): Inhibition of in vitro self-renewal in secondary passage for TICs treated with 12.5 mM DCA in the primary passage (n = 3; p < .001). Abbreviations: DCA, dichloroacetic acid; MFI, mean fluorescence intensity; NTC, nontumorogenic cancer cell; Ox. Phos., oxidative phosphorylation; TIC, tumor initiating cell.

Next we examined mitochondrial numbers in human breast TICs and NTCs. To do so we used two independent patient-derived triple negative breast cancer xenografts in which we previously showed CD49f^high Epcam^high cells to be highly enriched for TICs 46 (Supporting Information Fig. S6D). Of note, these markers differ from those used in MMTV-Wnt-1 tumors but were confirmed by in vivo limiting dilution analyses in both cases and are consistent with previously observed differences in marker expression between human and mouse normal and cancer stem cells 15, 47, 48. Assessment of mitochondrial content using MitoTracker (Fig. 7B; Supporting Information Fig. S6) or the ratio of mtDNA to nuclear DNA (Fig. 7C; Supporting Information Fig. S6) revealed significantly lower numbers of mitochondria in the TIC-enriched population. Thus, mitochondrial difference between TICs and NTCs are conserved between murine mammary tumors and human breast cancers.

Finally, we wished to test whether metabolic reprogramming by DCA also targets human breast TICs. CD49f^high Epcam^high human TIC-enriched cells are clonogenic in a three-dimensional cancer spheroid culture system while CD49f^low Epcam^low cells cannot give rise to colonies. Treatment of CD49f^high Epcam^high TIC-enriched cells with DCA resulted in a dose dependent decrease of TIC clonogenicity in the first passage very similar to what we observed with TICs from MMTV-Wnt-1 tumors (Fig. 7D, 7E; Supporting Information Fig. S6). In secondary passage in the absence of DCA, cells treated with DCA in the first passage display dramatically lower sphere formation, indicating decreased in vitro self-renewal (Fig. 7F). Taken together, these results indicate that promotion of oxidative phosphorylation also targets human breast TICs and represents a potential therapeutic strategy for targeting TICs in human patients.

Discussion

In recent years there has been a renewed interest in the unique metabolic properties of cancer cells. Changes in metabolism are one of the hallmarks of cancer 49 and cells in many types of cancers preferentially undergo glycolysis. The vast majority of studies in this area have focused on bulk populations of cancer cell lines. However, a continuously increasing body of evidence documents the existence of cellular hierarchies within many tumor types. Based on previous demonstration that breast TICs contain lower levels of ROS at baseline and since most endogenously produced ROS stem from oxidative phosphorylation, we hypothesized that TICs have unique metabolic properties compared to NTCs. Using a combination of genomic and metabolic approaches we found that TICs are proglycolytic and contain fewer, less active mitochondria. Activity of Pdh, a key enzyme for initiation of oxidative phosphorylation, is significantly reduced in TICs and forced elevation of Pdh activity eliminates TICs in vitro and in vivo in both murine mammary tumors and human breast cancers.

Previous studies have shown that metabolic properties of several types of normal stem cells differ from those of their more differentiated progeny 27, 29, 30, and enhanced rates of glycolysis have been described in embryonic stem cells 29, 50, suggesting that preferential reliance on glycolytic metabolism may be a broadly conserved stem cell property. However, only very limited published data exist regarding metabolic properties of TICs 42, 51-56, and none in breast cancer. Three recently published studies examined metabolic properties of glioma stem cell-enriched cells (GSC-ECs) and nontumorigenic cells in glioma cell lines 51 or primary human samples 53, 56. The study focusing on established cell lines found GSC-ECs were less glycolytic than nontumorigenic cells 51, while the other two studies, using primary glioma samples, observed a proglycolytic phenotype in GSC-ECs 53, 56. Our observation of a proglycolytic phenotype in TICs from primary breast tumors is in line with the latter two studies. The discordant results may reflect differences in metabolic properties between TIC-like cells from cell lines that have been passaged in vitro and TICs isolated from primary tumors. Alternatively, there were methodological differences among the studies, such as differences in isolation of cells (flow cytometry versus enrichment by differential culture conditions), which could have contributed to the results. It will be interesting to examine if the decrease in mitochondrial number and activity that we observed in breast TICs are also prevalent in other tumor types, including gliomas.

One of our most striking findings was the observation that breast TICs are proglycolytic even though they are relatively quiescent. Elevated rates of aerobic glycolysis are usually thought to accompany enhanced proliferation and this phenotype has frequently been documented in rapidly dividing cancer cells and is termed the Warburg effect. Our data indicate that TICs are programmed to undergo enhanced rates of glycolysis, irrespective of their cell cycle state. In agreement with our findings, HSCs have recently been shown to preferentially perform glycolysis over oxidative phosphorylation compared to their more differentiated progeny, even though they are more quiescent 28. Additionally, a recent study comparing metabolic properties of quiescent and proliferating fibroblasts found that quiescent fibroblasts were highly metabolically active, including demonstrating high rates of glycolysis. Furthermore, similar to TICs, total free radical levels were lower in quiescent fibroblasts, potentially due to increased availability of GSH 57.

Our findings raise the question of why TICs display a proglycolytic phenotype. For rapidly dividing cells, a proglycolytic phenotype may be advantageous due to diversion of glycolytic intermediates upstream of pyruvate into other biosynthetic pathways, thus meeting the metabolic requirements of cell proliferation 58. However, compared to NTCs, TICs are proglycolytic in the absence of rapid proliferation suggesting that other explanations are required. We suggest that preference for glycolysis by TICs relates to effects on levels of ROS. Maintenance of low ROS levels is critical for normal stem cells and TICs since elevation of ROS limits self-renewal 23, 59, 60. Since glycolysis results in decreased production of endogenous ROS compared to oxidative phosphorylation 61, proglycolytic metabolism is likely advantageous for this reason. Additionally, glycolytic metabolism and reduced oxygen consumption may confer a survival advantage to TICs during transient hypoxia or anoxia in the tumor microenvironment. Consistent with this idea, tumor hypoxia, induced by antiangiogenic agents, has been shown to increase growth of breast TICs 62.

The fact that breast TICs preferentially perform glycolysis and display decreased Pdh expression and activity led us to hypothesize that exogenous activation of Pdh leading to increased rates of oxidative phosphorylation would induce TIC cell killing. Treatment of TICs with DCA “normalized” the ratio of lactate production to oxygen consumption relative to NTCs and led to significant cell death. NTCs were relatively resistant to DCA, consistent with the observation that these cells undergo higher rates of oxidative phosphorylation at baseline. To our knowledge only two previous studies have examined effects of DCA on TICs or TIC-like cells. These two studies examined effects of DCA treatment on TIC-like (CD133⁺ or sphere forming) cells in glioblastoma 42, 53 and found that DCA treatment inhibited proliferation of these cells in vitro and in vivo. These results are consistent with our findings.

Although we found that DCA can target TICs, single agent treatment with this drug will likely not be sufficient to eliminate TICs in human patients. While we observed significant single agent activity on TICs in vivo, the results were not as dramatic as in vitro. This is likely due in large part to the well-described, relatively high concentrations of DCA required to achieve the desired metabolic reprogramming effects and the difficulty of achieving these in vivo. Human studies have documented achievable plasma doses up to 1.3 mM 63, which is below the IC50 of DCA for several of the Pdk isoforms 64. Thus, combination of DCA with other therapeutic agents is attractive. Indeed, several studies, together with our own (data not shown) have found at least additive effects of DCA with chemotherapy or radiotherapy 65, 66. It will be important to explore the efficacy of more potent metabolic reprogramming agents and combination regimens of DCA with other therapeutics. Our data argue that such studies should specifically evaluate effects on TICs, given that these cells often represent a minority population and NTCs appear to be relatively resistant to metabolic reprogramming.

Although our TIC subpopulations were highly enriched for TICs, they were not entirely pure and this represents a limitation of our study and most other studies of TICs. However, the fraction of tumor initiating cells in the CD49f^high Epcam^low was 1 in 79, which is one of the highest enrichment ratios that have been reported for TICs from primary tumors. Additionally, it is likely that the actual fraction of TICs within CD49f^high Epcam^low cells is higher since tumor dissociation and sorting lead to cell damage and the in vivo transplantation assay is not completely efficient. In support of this concept, the tumor sphere initiating frequency of CD49f^high Epcam^low was less than 1 in 20 (data not shown). Most importantly, we found that treating tumors with DCA in vivo led to a significant reduction in TIC frequency as measured by limiting dilution analysis in secondary transplants. Since limiting dilution analyses represent the gold standard assay for measuring TICs, this functionally confirms that shifting metabolism of TICs toward oxidative phosphorylation targets these cells and is in line with our metabolic analysis of the CD49f^high Epcam^low population.

Our findings have significant implications for studies examining cancer metabolism and the Warburg effect. The vast majority of studies in this area have focused on bulk populations of cancer cells, thus obscuring potential heterogeneity of cancer cell metabolism. Given our findings, it is possible that strategies combining multiple agents that target different aspects of metabolism will be required in order to show dramatic effects in the clinic. In tumors where TICs make up a minority subpopulation, early phase clinical trials of metabolically targeted agents that preferentially target TICs such as DCA may not reveal significant tumor shrinkage and could thus be inadvertently abandoned. Translational endpoints specifically measuring effects of such agents on TICs will be critical for clinical translation.

Conclusions

In summary, we have demonstrated that breast cancer TICs are proglycolytic compared to NTCs. This unique metabolic property of TICs is due to differences in mitochondrial biology and is mechanistically mediated by decreased Pdh activity. Metabolic reprogramming of TICs to increase rates of oxidative phosphorylation preferentially targets these cells and is an attractive therapeutic strategy for developing novel TIC-targeted agents.

Acknowledgments

We thank M. Clarke, N. Denko, B. Mitchell, and S. Plevritis for helpful discussions and I. Papandreou and K. Qu for technical assistance. This work was supported by grants from the Sydney Kimmel Foundation (M.D.), National Institutes of Health (M.D.; P30CA147933), the CRK Faculty Scholar Fund (M.D.), Nadia's Gift (M.D.), and the Virginia and D.K. Ludwig Foundation (M.D.). M.D. is supported by a Doris Duke Clinical Scientist Development Award and the NIH New Innovator Award Program (1-DP2-CA186569).

Author Contributions

W. F. and M.D.: conception and design, collection and assembly of data, and manuscript writing; A.G. and R.V.N.: data analysis and interpretation; M.H., Y.L., C.Y.L., S.C., F.S., A.K.: collection of data and provision of study material.

Disclosure of Potential Conflict of Interest

The authors indicate no potential conflicts of interest.