Transcriptional evidence for the "Reverse Warburg Effect" in human breast cancer tumor stroma and metastasis: Similarities with oxidative stress, inflammation, Alzheimer's disease, and "Neuron-Glia Metabolic Coupling"

Introduction

Recently, we identified a loss of stromal caveolin-1 (Cav-1) as a novel biomarker for the cancer-associated fibroblast phenotype in human breast cancers [1]. More specifically, when fibroblasts were isolated from human breast cancers, 8 out of 11 patients showed >2-fold reduction in Cav-1 protein expression, relative normal matched fibroblasts prepared from the same patients [1]. Furthermore, detailed phenotypic analysis of mammary fibroblasts derived from Cav-1 (-/-) null mice revealed that they share numerous properties with cancer-associated fibroblasts, such as constitutively active TGFbeta signaling, and that they have the ability to promote normal mammary epithelial cells to undergo an EMT (epithelial-mesenchymal transition) [2].

To determine if loss of stromal Cav-1 has prognostic value, we performed a series of independent biomarker studies [3,4]. Using a cohort of 160 breast cancer patients, with nearly 20 years of follow-up data, we showed that a loss of stromal Cav-1 (in the fibroblast compartment) is a powerful single independent predictor early tumor recurrence, lymph node metastasis, tamoxifen-resistance, and poor clinical outcome [4]. As the prognostic value of a loss of stromal Cav-1 was independent of epithelial marker status, it appears that a loss of Cav-1 has predictive value in all the different epithelial subtypes of human breast cancer, including ER+, PR+, HER2+, and triple-negative patients [4]. The high predictive value of a loss of stromal Cav-1 was also independently validated by another independent laboratory, using a second independent breast cancer patient cohort [5].

A loss of stromal Cav-1 also appears to play a role in tumor initiation and progression [6]. Using a DCIS patient cohort, in which patients were treated with wide-excision, but without any chemo- or radio-therapy, we also evaluated the prognostic value of stromal Cav-1 [6]. In this DCIS patient cohort, a loss of stromal Cav-1 was specifically associated with DCIS recurrence and invasive progression. 100% of the patients with a loss of stromal Cav-1 underwent recurrence, and 80% of these patients progressed to invasive disease, namely frank invasive ductal carcinoma [6]. Similar results were also independently obtained in human prostate cancers, where a loss of stromal Cav-1 was specifically associated with advanced prostate cancer, tumor progression, and metastatic disease [7].

To begin to understand the mechanism(s) underlying the lethality of a loss of Cav-1 in cancer-associated fibroblasts, we turned to Cav-1 (-/-) deficient mice as a model system.

For this purpose, we isolated bone marrow derived stromal cells from WT and Cav-1 (-/-) deficient mice, as cancer-associated fibroblasts are thought to evolve from mesenchymal stem cells [8]. These cells were then subjected to unbiased proteomic and genome-wide transcriptional analysis. Interestingly, proteomic analysis revealed the upregulation of i) 8 myofibroblast markers (including vimentin, calponin, and tropomyosin), ii) 8 glycolytic enzymes (including PKM2 and LDHA), and iii) 2 markers of oxidative stress (peroxiredoxin1 and catalase) [8]. The glycolytic phenotype of Cav-1 (-/-) null stromal cells was also supported by transcriptional analysis, as most of the proteins that were found to be upregulated by proteomics, were also transcriptionally upregulated [8]. Based on these findings, we proposed a new model to understand the role of the Warburg effect ("aerobic glycolysis") in tumor metabolism. We hypothesized that glycolytic cancer-associated fibroblasts promote tumor growth by the secretion of energy-rich metabolites (such as pyruvate and lactate) that could then be taken up by adjacent epithelial cancer cells, where they would be incorporated into the tumor cell's TCA cycle, leading to enhanced ATP production [8]. This would provide a feed-forward mechanism by which glycolytic fibroblasts could promote tumor growth, progression, and metastasis. Because the Warburg effect was previously thought to be largely confined to tumor cells, and not to the cancer-associated fibroblast compartment, we have termed this new idea "The Reverse Warburg Effect" [8].

In order to determine which transcriptional programs are activated in Cav-1 (-/-) stromal cells, we performed an extensive bioinformatics analysis of our genome-wide profiling data [9]. This informatics analysis revealed that a loss of Cav-1 (-/-) in stromal cells drives ROS production and oxidative stress [9]. This, in turn, results in the activation of key transcription factor, such as HIF and NF-kB, which can then drive aerobic glycolysis and inflammation in the tumor micro-environment [9]. This could provide a molecular basis for understanding the lethality of a loss of stromal Cav-1 in human breast cancer patients.

Here, we have used a bioinformatics approach to determine whether similar "Warburg-like" transcrip-tional profiles exist in the tumor stroma isolated from human breast cancers. For this purpose, we analyzed an existing data set in which the tumor stroma was isolated away from adjacent breast cancer cells using laser-capture micro-dissection [10]. We now provide new evidence for the existence of the "Reverse Warburg Effect" in human tumor stroma from breast cancer patients. More specifically, the tumor stroma of human breast cancers shows a transcriptional shift towards oxidative stress, DNA damage/repair, inflammation, hypoxia, and aerobic glycolysis, supporting with the "Reverse Warburg Effect". Consistent with the idea that oxidative stress in the tumor stroma is a driving factor in promoting tumor progression and metastasis, we also show that the tumor stroma of human breast cancers overlaps significantly with the transcriptional profiles associated with Alzheimer's brain disease.

Finally, the "Reverse Warburg Effect" is strikingly similar to the theory of "Neuon-Glia Metabolic Coupling" [11-18], which was proposed more than 10 years ago to explain metabolic changes associated with normal synaptic transmission, which may be exacerbated during neuronal stress and neuronal degeneration, as in Alzheimer's disease. In "Neuron-Glia Metabolic Coupling", astrocytes undergo aerobic glycolysis, secrete energy-rich metabolites (pyruvate and lactate), and neurons then take up these metabolites and use them in the neuronal TCA cycle to generate high amounts of ATP. Thus, we propose that "The Reverse Warburg Effect" we observe could also be more broadly termed "Epithelial-Stromal Metabolic Coupling".

As such, tumors may be initiating a survival mechanism that is normally used by the brain during stress. Interestingly, myofibroblasts and mesenchymal stem cells are known to often express GFAP (glial fibrillary acidic protein) [19-21], an intermediate filament protein that is thought to be relatively specific for astrocytes in the central nervous system. Here, we see that GFAP is upregulated in the "tumor stroma" and in the stroma of "metastasis-prone" breast cancer patients. Thus, possible similarities between astrocytes and myo-fibroblasts/cancer-associated fibroblasts should be further explored.

Results

Transcriptional comparison of Cav-1 (-/-) stromal cells with human breast cancer stroma

Previously, we subjected Cav-1 (-/-) bone marrow derived stromal cells, and their wild-type counter-parts to genome-wide transcriptional profiling [8]. Because such a large number of gene transcript levels are changed, we focused on the gene transcripts that are upregulated. We speculated that these Cav-1 (-/-) stromal gene profiles might also overlap with the transcriptional stromal profiles obtained from human breast cancers.

To test this hypothesis directly, we obtained the transcriptional profiles of a large data set of human breast cancer patients [10] whose tumors were subjected to laser-capture micro-dissection, to selectively isolate the tumor stroma. Based on this data set [10], we then generated three human breast cancer stromal genes lists:

1) Tumor Stroma vs. Normal Stroma List - Compares the transcriptional profiles of tumor stroma obtained 53 patients to normal stroma obtained from 38 patients. Genes transcripts that were consistently upregulated in tumor stroma were selected and assigned a p-value, with a cut-off of p <0.05 (contains 6,777 genes) (Supplementary Table 1).

2) Recurrence Stroma List - Compares the transcript-tional profiles of tumor stroma obtained from 11 patients with tumor recurrence to the tumor stroma of 42 patients without tumor recurrence. Genes transcripts that were consistently upregulated in the tumor stroma of patients with recurrence were selected and assigned a p-value, with a cut-off of p <0.05 (contains 3,354 genes) (Supplementary Table 2).

3)Lymph-node (LN) Metastasis Stroma List - Compares the transcriptional profiles of tumor stroma obtained from 25 patients with LN metastasis to the tumor stroma of 25 patients without LN metastasis. Genes transcripts that were consistently upregulated in the tumor stroma of patients with LN metastasis were selected and assigned a p-value, with a cut-off of p <0.05 (contains 1,182 genes) ( Supplementary Table 3).

These three gene lists were then individually intersected with the transcriptional profile of Cav-1 (-/-) null stromal cells [8]. The results of these intersections are presented in Figure 1, as Venn diagrams. Most important- ly, significant overlap was seen with all three gene lists. Greater than 2,000 genes were common between the Cav-1 (-/-) stromal gene list and the gene transcripts upregulated in breast cancer tumor stroma (p = 1.6 x 10^-3). Also, more than 1,000 gene transcripts were common between the Cav-1 (-/-) stromal gene list and the gene transcripts upregulated in the breast cancer tumor stroma of patients with tumor recurrence (p = 1 x 10^-3). Finally, nearly 500 genes were commonly upregulated between Cav-1 (-/-) stromal cells and the breast cancer tumor stroma of patients with LN metastasis (p = 4.6 x 10^-6). Thus, the transcriptional profiles of Cav-1 (-/-) stromal cells are most ignificantly related to the tumor stroma of patients with LN-metastasis. Independently, our previous data demonstrated that a loss of stromal Cav-1 protein expression (by immuno-histochemistry) in human breast cancers is specifically associated with a 2.6-fold increase in the number of tumor cell positive lymph nodes (LN-metastasis) [3,4].

Figure 1. Venn diagrams for the transcriptional overlap between Cav-1 (-/-) stromal cells and tumor stroma from breast cancer patients.

The top 100 most significant gene transcripts for all three human breast cancer stromal gene lists, including their transcriptional intersection with Cav-1 (-/-) stromal cells, is included in Supplementary Tables 3, 4, and 5.

As Cav-1 (-/-) stromal cells are a genetic model of activated myofibroblasts [2] which biosynthetically secrete more collagen, and fibrosis is a critical risk factor for poor clinical outcome in human breast cancer patients [3], we also looked at the potential overlap been the expression of collagen gene transcripts (See Table 1). Thirty-five collagen gene transcripts were specifically upregulated in tumor stroma; 16 were upregulated in "recurrence-prone" stroma; and only 1 was upregulated in "metastasis-prone" stroma. In all three cases, there was striking overlap with the collagen gene transcripts upregulated in Cav-1 (-/-) stromal cells, as indicated in bold (24 out of 35 transcripts; 12 out of 16 transcripts; and 1 out of 1 transcript; See Table 1).

Cav-1 (-/-) stromal cells have also been previously subjected to extensive analysis via an unbiased proteomics approach [8,24]. We next intersected these proteomic results with the three human breast cancer stromal gene lists. The results of this intersection are shown in Table 2. Note that many of the proteins that are upregulated in Cav-1 (-/-) stromal cells are also transcriptionally upregulated in the stroma of human breast cancer patients. Most notably, there was a strong association between the metabolic enzymes that were upregulated in Cav-1 (-/-) stromal cells and the "recurrence-prone" and "metastasis-prone" stromal gene lists.

Similarities of the Cav-1 (-/-) stromal gene set with transcriptional profiling data from ER-negative breast cancer

A comparison of the Cav-1 (-/-) stromal cell gene set with other existing transcriptional profiles also shows significant overlap with ER-negative human breast cancer (p = 8.96 x 10^-10; BRCA_ER_NEG [28]). For this overlap analysis, UP genes from the Cav-1 (-/-) stromal data set with a fold-change of > 2.0 (KO/WT) and a P value of < 0.1 were utilized for comparison with existing gene sets in the data base.

Interestingly, these tumors were not laser-capture micro-dissected, so this provides an indication that the Cav-1 (-/-) stromal gene set may also be well represented in the transcriptional profiles obtained from whole tumors. A HeatMap containing these intersecting genes is shown in Figure 3 (205 overlapping genes; FC > 1.5; p < 0.05). See also Supplementary Tables.

These include key overlapping genes associated with metabolism and glycolysis (Acot7, Acsl4, Eno1, Gapdh, Ldhb, Mtrf1l, Pfkl, Pgk1, Pgm2, Pgm3, Slc2a5, Slc2a6), hypoxia (Hyou1), the inflammatory response (Aif1, C3, Ccl5, Crlf3, Ifngr1, Il10ra, Irak1, Irf5, Isg20, Nfib, Nfkbie, Nos3, Tnfaip3, Tnfrsf21, Tnfsf13b, Traf1), myofibroblast differentiation and the extracellular matrix (Actl6a, Capg, Col9a3, Dnmt3b, Mmp9, Myo10, Spock2, Tgfbi, Tgm1, Timp2), as well as DNA-damage and repair (Ddit3, Rad54l). These results are consistent with the existence of the "Reverse Warburg Effect" in ER-negative breast cancers.

Interestingly, it has been previously demonstrated that key secreted inflammatory factors, such as Aif1 (allograft inflammatory factor-1) (upregulated nearly 3-fold in Cav-1 (-/-) stromal cells; Supplementary Tables) promote NFkB-activation, the paracrine growth of ER-negative breast cancer cells [29], and are involved in the pathogenesis of pro-fibrotic diseases, such as scleroderma (systemic sclerosis) [30-32].

Similarly, Aif1 expression is highly-upregulated in the tumor stroma of human breast cancers (See Supplementary Table 1; p = 8.35 x 10^-24).

Figure 2. Venn diagrams for the transcriptional overlap between Alzheimer's disease brain and tumor stroma from breast cancer patients.

Discussion

Here, we provide compelling transcriptional evidence for the "Reverse Warburg Effect" in human breast cancer tumor stroma. Using an unbiased informatics analysis of transcriptional gene profiling, we show that Cav-1 (-/-) stromal cells bear a striking resemblance to the activated tumor stroma of human breast cancers. More specifically, the transcriptional profiles of Cav-1 (-/-) stromal cells were most closely related to the stroma of breast cancer patients that had undergone LN-metastasis. This is consistent with our previous data showing that a loss of stromal Cav-1 protein expression (by immuno-histochemistry) in human breast cancer tumor micro-arrays is specifically associated with increased LN-metastasis [3,4].

Moreover, we provide evidence that the tumor stroma of human breast cancers shows a transcriptional shift towards oxidative stress, DNA damage/repair, inflammation, hypoxia, and aerobic glycolysis. These findings are consistent with the "Reverse Warburg Effect" [8,9]. Notably, the tumor stroma of "metastasis-prone" breast cancer patients was also closely related to the transcriptional profiles derived from the brains of patients with Alzheimer's disease. As such, certain fundamental biological processes are common to both an activated tumor stroma and neuro-degenerative stress. These key biological processes most likely include oxidative stress, NO over-production (peroxynitrite formation), inflammation, hypoxia, and mitochondrial dysfunction, which are all thought to drive Alzheimer's disease pathogenesis.

Thus, we avidly reviewed the literature regarding theories of neuronal functioning, neuronal stress, and neuro-degeneration, in the central nervous system and we stumbled upon the idea of "Neuron-Glia Metabolic Coupling" [11-18] In this model, first proposed over 10 years ago, astrocytes shift towards aerobic glycolyis, secrete pyruvate and lactate, which is then taken-up by adjacent neurons and then "feeds" into the neuronal TCA cycle, resulting in increased neuronal oxidative mitochondrial metabolism, and higher ATP production in neurons. In essence, the astrocytes would function as support cells to "feed" the adjacent neuronal cells. Thus, "Neuron-Glia Metabolic Coupling" and the "Reverse Warburg Effect" are analogous biological processes, where the astrocytes are the cancer-associated fibroblasts and the neurons are the epithelial tumor cells. As such, we propose that the "Reverse Warburg Effect" could also be more generally termed "Epithelial-Stromal Metabolic Coupling" or "Epithelial-Fibroblast Metabolic Coupling".

If these two processes are indeed analogous, then epithelial tumor cells have already learned to behave as neurons, using the stroma as a means of support and nourishment. Figure 4 directly compares "Neuron-Glia Metabolic Coupling" with the "Reverse Warburg effect" schematically.

Myofibroblasts and mesenchymal stem cells are known to often express GFAP (glial fibrillary acidic protein) [19-21], an intermediate filament protein that is thought to be relatively specific for astrocytes in the central nervous system. Table 5 shows that GFAP and other glial-related gene transcripts are indeed upregulated in "tumor stroma" and in the stroma of "metastasis-prone" breast cancer patients. Thus, possible metabolic and functional similarities between CNS astrocytes and myofibroblasts/cancer-associated fibroblasts should be further explored.

Interestingly, in "Neuron-Glia Metabolic Coupling" the glycolytic shift in astrocytes is thought to be mediated by the secretion of glutamate (a neurotransmitter) from neurons. Then, astrocytes take up glutamate via high affinity sodium-dependent glutamate transporters, such as Slc1a2 and Slc1a3. Importantly, one of these two glial-specific glutamate transporters (Slc1a3) is also transcriptionally over-expressed in the stroma of human breast cancer patients (Table 5). As such, the similarities between brain astrocytes, myofibroblasts, mesenchymal stem cells, and tumor stromal cells may be more extensive than we previously appreciated.

Figure 3. Transcriptional overlap of the Cav-1 (-/-) stromal gene set with ER-negative breast cancer. A HeatMap containing 205 intersecting genes is shown (FC >1.5; p <0.05). See also Supplementary Tables. FC, fold-change.

Figure 4. Comparisons between the "Reverse Warburg Effect" and "Neuron-Glia Metabolic Coupling", suggest "Epithelial-Stromal Metabolic Coupling". In "Neuron-Glia Metabolic Coupling", astrocytes take up more glucose, shift towards aerobic glycolyis, secrete pyruvate and lactate, which is then taken up by adjacent neurons and then "feeds" into the neuronal TCA cycle, resulting in increased neuronal oxidative mitochondrial metabolism, and higher ATP production in neurons. In essence, the astrocytes function as support cells to "feed" the adjacent neuronal cells. This schematic diagram shows that "Neuron-Glia Metabolic Coupling" and the "Reverse Warburg Effect" are analogous biological processes, where the astrocytes are the cancer-associated fibroblasts and the neurons are the epithelial tumor cells. Thus, the "Reverse Warburg Effect" could also be more generally termed "Epithelial-Stromal Metabolic Coupling" or "Epithelial-Fibroblast Metabolic Coupling". This figure was partially re-drawn from Bonucelli et al. 2010, with permission [24]. MCT, mono-carboxylate transporter.

Methods of analysis

Venn diagrams. In the Venn diagram of Figure 1, we show the intersections between the set of genes that are upregulated in Cav-1 (-/-) versus wild-type stromal cells [8] and three breast cancer gene sets [10].

(a) the set of stromal genes that are upregulated in breast cancer tumor patients versus normal breast stroma; (b) the set of stromal genes that are upregulated in recurrence positive versus recurrence negative breast cancer patients (c) the set of stromal genes that are upregulated in lymph-node metastasis positive versus lymph-node metastasis negative breast cancer patients.

In the Venn diagram of Figure 2, we show the intersections between the set of genes that are upregulated in Alzheimer's brain disease [22] and the sets of genes (a)-(c) listed above. The p-values determining the significance of upregulation for each gene were computed using a one-sided t-test statistic (Tables 1, 2, and 5). For each pair (X,Y) of sets of genes, we also computed the probability (p-value) that the size of their intersection is less than or equal to the size of the intersection between set X and a randomly-chosen set of size equal to the size of set Y. This probability was calculated by applying the cumulative density function of the hypergeometric distribution on the size of set X, the size of set Y, the observed overlap between X and Y, and the total number of available genes.

Gene set enrichment analysis. For the functional analysis presented in Table 4, we used data from the Molecular Signatures Database (MsigDB [23]) which comprises a collection of gene sets: - collected from various sources such as online pathway databases, publications, and knowledge of domain experts, - comprising genes that share a conserved cis-regulatory motif across the human, mouse, rat, and dog genomes, - identified as co-regulated gene clusters by mining large collections of cancer-oriented microarray data, and - annotated by a common Gene Ontology (GO) term.

For our analysis we used the latest release of MSigDB database v2.5 (April 7, 2008), after converting all the gene names in the database into RefSeq gene IDs. After this preprocessing step, we chose the sub-collection of gene sets that was relevant to our study, and for each gene set X in that sub-collection, we computed the overlap between X and the set of genes Y that are upregulated in lymph-node metastasis positive versus lymph-node metastasis negative breast cancer patients (p-value ≤0.05). Then, we computed the probability (p-value) of the observed overlap between sets X and Y as described in the "Venn diagrams" section.

Supplementary Materials

Supplementary Table 1

Supplementary Table 2

Supplementary Table 3

Supplementary Table 4-10