The metabolic co-regulator PGC1α suppresses prostate cancer metastasis

Cellular transformation and cancer progression is accompanied by changes in the metabolic landscape. Master co-regulators of metabolism orchestrate the modulation of multiple metabolic pathways through transcriptional programs, and hence constitute a probabilistically parsimonious mechanism for general metabolic rewiring. Here we show that the transcriptional co-activator PGC1α suppresses prostate cancer progression and metastasis. A metabolic co-regulator data mining analysis unveiled that PGC1α is down-regulated in prostate cancer and associated to disease progression. Using genetically engineered mouse models and xenografts, we demonstrated that PGC1α opposes prostate cancer progression and metastasis. Mechanistically, the use of integrative metabolomics and transcriptomics revealed that PGC1α activates an Oestrogen-related receptor alpha (ERRα)-dependent transcriptional program to elicit a catabolic state and metastasis suppression. Importantly, a signature based on the PGC1α-ERRα pathway exhibited prognostic potential in prostate cancer, thus uncovering the relevance of monitoring and manipulating this pathway for prostate cancer stratification and treatment.

metabolic enzymes 5 , it is widely accepted that the main trigger for metabolic reprogramming is the alteration in cancer genes that remodel the signalling landscape 2 . Numerous reports provide evidence of the pathways regulating one or a few enzymes within a metabolic pathway in cancer. However, the means of coordinated regulation of complex metabolic networks remain poorly documented.
Master transcriptional co-regulators of metabolism control a variety of genes that are in charge of remodelling the metabolic landscape, and their impact in cellular and systemic physiology has been studied for decades. It is worth noting that these co-regulators, through their capacity to interact and regulate diverse transcription factors, exhibit a unique capacity to control complex and extensive transcriptional networks, making them ideal candidates to promote or oppose oncogenic metabolic programs.
The tumour suppressor PTEN is a negative regulator of cell growth, transformation and metabolism [6][7][8][9] . PTEN and its main downstream pathway, PI-3-Kinase, have been extensively implicated in prostate cancer (PCa) pathogenesis and progression [10][11][12] . This tumour suppressor is progressively lost through the progression of PCa, and complete loss of PTEN is predominant in advanced disease and metastasis 8 . Genetically engineered mouse models (GEMMs) recapitulate many of the features of PCa progression. However, the molecular and metabolic bases for PCa metastasis remain poorly understood [13][14][15][16] . Indeed, complete loss of PTEN in the mouse prostate does not result in metastasis 11 , in turn suggesting that additional critical events are required in this process.
In this study, we designed a bioinformatics analysis to interrogate multiple PCa datasets encompassing hundreds of well-annotated specimens. This approach allowed us to define a master regulator of PCa metabolism that is crucial for the progression of the disease. Our results identify the Peroxisome proliferator-activated receptor gamma co-activator 1 alpha (PGC1α) as a suppressor of PCa metastasis. This transcriptional co-activator exerts its function through the regulation of Oestrogen-related receptor alpha (ERRα) activity, in concordance with the activation of a catabolic program and the inhibition of PCa metastasis.

A bioinformatics screen identifies PGC1A as metabolic co-regulator associated to prostate cancer progression
We approached the study of PCa metabolism applying criteria to ensure the selection of relevant master regulators that contribute to the metabolic switch. We focused on transcriptional co-regulators of metabolism 17 that i) were consistently altered in several publicly available PCa datasets [18][19][20][21][22][23][24] , and ii) were associated with reduced time to recurrence and disease aggressiveness. We first evaluated the expression levels of the metabolic coregulators in a study comprising 150 PCa specimens and 29 non-pathological prostate tissues (or controls) 22 . The analysis revealed 10 co-regulators in the set of study with significant differential expression in PCa compared to non-neoplastic prostate tissue (Fig.  1a, Supplementary Fig. 1A). We next extended this observation to four additional datasets 18,21,23,24 in which there was available data for non-tumoural and PCa tissues. Only the alteration in PPARGC1A (PGC1A), PPARGC1B (PGC1B) and HDAC1 expression was further confirmed in the majority or all sets (Fig. 1b, Supplementary Fig. 1B). Among these, PGC1A was the sole co-regulator with altered expression associated to Gleason score ( Supplementary Fig. 1C, D) and DFS (Fig. 1c).
In order to rule out that cellular proliferation could contribute to the alteration of metabolic regulators, we carried out an additional analysis in which we compared the expression of PGC1A in PCa versus a benign hyper-proliferative condition (benign prostate hyperplasia or BPH). The results corroborated that the decrease in PGC1A expression is associated to a cancerous state rather than to a proliferative condition (Supplementary Fig. 1E).
We observed that the expression of PGC1A was progressively decreased from primary tumours to metastasis (Fig. 1d, Supplementary Fig. 1F). Strikingly, genomic analysis revealed shallow deletions of PGC1A exquisitely restricted to metastatic PCa specimens [19][20][21][22]25 (Fig. 1e), in full agreement with the notion that there is a selective pressure to reduce the expression of this transcriptional co-activator as the disease progresses.
From our analysis, PGC1α emerges as the major master metabolic co-regulator altered in PCa, with an expression pattern reminiscent of a tumour suppressor.

Pgc1a deletion in the murine prostate epithelium promotes prostate cancer metastasis
PGC1α has been widely studied in the context of systemic metabolism 26 , whereas its activity in cancer is just beginning to be understood [27][28][29][30][31][32][33] . To ascertain the role of PGC1α in PCa in vivo, we conditionally deleted this metabolic co-regulator in the prostate epithelium 34 , alone or in combination with loss of the tumour suppressor Pten 11 ( Fig. 2a-d,  Supplementary Fig. 2A, B). Pgc1a deletion alone or in the context of Pten heterozygosity did not result in any differential tissue mass or histological alteration, which led us to conclude that it is not an initiating event (Fig. 2b, d). However, compound loss of both Pten and Pgc1a resulted in significantly larger prostate mass (Fig. 2c), together with a remarkable increase in the rate of invasive cancer (Fig. 2d). Histological analysis of the prostate revealed the existence of vascular invasion in double mutant mice (DKO), but not in Ptendeleted (Pten KO) prostates (Supplementary Fig. 2C). PGC1α regulates the inflammatory response, which could influence and contribute to the phenotype observed 35 . However, we did not observe significant differences in the infiltration of polimorphonuclear netrophils (PMN) and lympho-plasmacytic infiltrates in our experimental settings ( Supplementary  Fig. 2D). PGC1α has been also shown to induce angiogenesis in coherence with the induction of VEGF-A expression 36 . Pgc1a status in our GEMM did not alter VEGF-A expression and microvessel density (Supplementary Fig. 2E, F). We therefore excluded the possibility that regulation of angiogenesis or inflammation downstream PGC1α could drive the phenotype characterised in this study.
PCa GEMMs faithfully recapitulate many of the features of the human disease 37 . A reduced number of mouse models with clinically relevant mutations show increased metastatic potential [13][14][15][16] . Strikingly, histopathological analysis of our mouse model in the context of Pten loss revealed that DKO mice -but not Pten KO counterparts -presented evidence of and Supplementary Fig. 2G). Metastatic dissemination was in agreement with the observation of Pan-cytokeratin (PanCK) and Androgen Receptor (AR)-positive PCa cell deposits in the lymph nodes of DKO mice (Fig. 2g). Of note, 33% of Pten KO mice presented small groups of PanCK-positive cells in LN (without metastatic lesions) (Supplementary Fig. 2H), suggesting that even if these cells are able to reach the LN, they lack capacity to establish clinical metastasis. Interestingly, bone analysis revealed disseminated groups (but not clinical metastasis) of PanCK-positive cells in DKO but not in Pten KO mice ( Supplementary Fig. 2I-K). Analysis of a small cohort of Pten pc−/− ; Pgc1a pc+/− mice demonstrated that heterozygous loss of Pgc1a is sufficient to promote aggressiveness, vascular invasion and metastasis (Supplementary Fig. 2L-N). This observation supports the notion that single copy loss of PGC1A (as observed in metastatic human PCa specimens, Fig. 1e) could be a key contributing factor to the metastatic phenotype.
The cooperative effect observed in our mouse model between loss of Pten and Pgc1a was supported by the direct correlation of the two transcripts in patient specimens and the association of PGC1A down-regulation with PTEN genomic loss (TCGA provisional data 19,20 , Supplementary Fig. 2O).
In summary, our data in GEMMs and patient datasets formally demonstrates that the downregulation of PGC1α in PCa is an unprecedented causal event for the progression of the disease and its metastatic dissemination.

PGC1α suppresses prostate cancer growth and metastasis
In order to characterize the prostate tumour suppressive activity of PGC1α, we first evaluated its expression level in well-established PCa cell lines 38 . Using previously reported PGC1α-positive and negative melanoma cells 28 , we could demonstrate that PCa cell lines lack detectable expression of the transcriptional co-activator at the protein level (Fig. 3a). In agreement with this notion, PGC1α-silencing in these cells failed to impact on the expression of its well-established targets 39 (Supplementary Fig. 3A). Importantly, through the analysis of publicly available datasets 22 , we could demonstrate that the transcript levels of PGC1A in metastatic cell lines are comparable to those observed in human metastatic PCa specimens and vastly reduced compared to PGC1α-positive melanoma cells (Fig. 3a,  Supplementary Fig. 3B). Despite our efforts to optimize the detection of the protein with different commercial antibodies, we could not identify an immunoreactive band that would correspond to PGC1α, in contrast with other reports 40,41 . Yet, we cannot rule out that in non-basal conditions, stimulation of other factors such as AR 41 or AMPK 40 could lead to the up-regulation and allow detection of PGC1α in PCa cells.
Due to the lack of PGC1α detection in PCa cellular systems, we aimed at reconstituting the expression of this gene to levels achievable in the cancer cell lines previously reported 28 . By means of lentiviral delivery of inducible Pgc1α and doxycycline titration, we reached expression levels of this protein in three PCa cell lines (AR-dependent -LnCaP -and independent -PC3 and DU145) equivalent to that observed in the PGC1α-expressing melanoma cell line MeWo (Fig. 3b, Supplementary Fig. 3C, D). Next, we evaluated the cellular outcome of expressing PGC1α in PCa cell lines. Interestingly, expression of Pgc1α in this context resulted in a reduction in bi-dimensional and three-dimensional growth (Fig.  3c, d, Supplementary Fig. E), cellular proliferation (Fig. 3e, Supplementary Fig. 3F) and cell cycle progression (Supplementary Fig. 3G). Of note, we excluded the possibility that doxycycline treatment could influence the result of the growth analysis ( Supplementary  Fig. 3H). The Pgc1α phenotype was recapitulated in vivo, where ectopic expression of this gene decreased tumour formation and growth (Fig. 3f, Supplementary Fig. 3I-K). In agreement with the GEMM data, we did not observe a contribution of angiogenesis to the phenotype (Supplementary Fig. 3L-N).
We observed in GEMMs that Pgc1a-loss resulted in metastatic dissemination (Fig. 2). We next sought to study whether Pgc1α expression could oppose a pre-existing metastatic phenotype. To this end, we carried out xenotransplant assays in immunocompromised mice using luciferase-expressing Pgc1α-inducible PC3 cells. Intra-cardiac injection of these cells (Fig. 3g) revealed that Pgc1α expression blunted metastatic growth in the lung, and led to a remarkable decrease in bone colonisation ( Fig. 3h-i). As an additional approach, we sought to analyse metastatic tumour re-initiation capacity by means of local injection of PCa cells at the metastatic site. Since PCa exhibits osteotropic nature 42 , we carried out intra-tibial injection of cells and the appearance of tumour masses in the bone was monitored 43 (Fig.  3j). The results demonstrated that PGC1α exerts a potent anti-metastatic activity both in bone tumour mass and metastatic foci (Fig. 3k). These data provide evidence of the antimetastatic potential of PGC1α.

PGC1α determines the oncogenic metabolic wiring in prostate cancer
PGC1α regulates gene expression through the interaction with diverse transcription factors 26 . In order to define the transcriptional program associated to the tumour suppressive activity of PGC1α, we performed gene expression profiling from Pgc1α-expressing vs. nonexpressing PC3 cells. We identified 174 probes with significantly altered signal encoding genes predominantly related to functions such as mitochondrial catabolic programs and energy-producing processes 26,44 (Supplementary Table 1, Fig. 4a), which we validated by qRTPCR ( Fig. 4b-d, Supplementary Fig. 4).
In order to demonstrate that the tumour suppressive activity of PGC1α was indeed accompanied by a global metabolic wiring, we carried out integrative metabolomics. We analysed cell line, xenograft and GEMM tissue extracts using liquid-chromatography highresolution mass spectrometry (LC-HRMS). LC-HRMS metabolomics and subsequent biochemical assays confirmed that oxidative processes such as fatty acid β-oxidation (  Fig. 5e, Supplementary Fig. 5D) were increased in response to Pgc1α expression. In order to quantitatively define the use of glucose in the TCA cycle, we carried out stable 13 C-U 6 -Glucose isotope labelling. This experimental approach provided definitive evidence of the increased oxidation of glucose in the mitochondria in Pgc1α expressing cells (Fig. 5f). This metabolic wiring was consistent with elevated oxygen consumption (basal and ATPproducing) and ATP levels upon Pgc1α expression ( We next reasoned that over-activation of mitochondrial oxidative processes would lead to decreased anabolic routes. On the one hand, we monitored the incorporation of carbons from 13 C-U 6 -Glucose into fatty acids (through the export of citrate from TCA to the cytoplasm 45 and conversion to acetyl CoA that is used for de novo lipid synthesis).
Interestingly, we found a significant decrease in 13 C incorporation into palmitate (reflected as 13 C carbon pairs) when Pgc1α was expressed (Fig. 5j, Supplementary Fig. 5J). On the other hand, we monitored lactate production as a readout of aerobic glycolysis or "the Warburg effect" 2 , which has been associated to the anabolic switch. As predicted, Pgc1αexpressing cells exhibited reduced extracellular lactate levels (Supplementary Fig. 5K). Of note, lactate production and respiration were unaltered by doxycycline challenge in nontransduced PC3 cells (Supplementary Fig. 5L, M). Taken together, our data provide a metabolic basis for the tumour suppressive potential of PGC1α in PCa, according to which this metabolic co-regulator controls the balance between catabolic and anabolic processes (Fig. 5k).

An ERRα-dependent transcriptional program mediates the prostate tumour suppressive activity of PGC1α
We next aimed to identify the transcription factor that mediated the activity of PGC1α, and hence we performed a promoter enrichment analysis. The results revealed a predominant abundance in genes regulated by ERRα (Fig. 6a). We corroborated these results with Gene Set Enrichment Analysis (GSEA; Normalized Enrichment Score=2.02; Nominal p value=0.0109) 46 . This transcription factor controls a wide array of metabolic functions, from oxidative processes to mitochondrial biogenesis 44 . We have shown that PGC1α is indeed capable of regulating functions attributed to ERRα, such as mitochondrial oxidative metabolism (Fig. 4, 5 and Supplementary Fig. 4, 5). In addition, we observed that Pgc1a expression led to increased mitochondrial volume (Supplementary Fig. 6A). In order to ascertain to which extent the growth inhibitory and anti-metastatic activity of PGC1α required its ability to interact with ERRα, we took advantage of a mutant variant of the coactivator (PGC1α L2L3M ) that is unable to interact with this and other nuclear receptors 46,47 . The expression of PGC1α L2L3M in PC3 cells (Supplementary Fig. 6B) failed to up-regulate target genes, to reprogram oxidative metabolism, to inhibit cell growth, and, importantly, to suppress bone metastasis in intratibial xenografts ( Fig. 6b-f, Supplementary Fig. 6C). To further discriminate between PGC1a functions that depend on ERRa or other nuclear receptors, we undertook a targeted silencing approach, and we transduced Pgc1α-inducible PC3 cells with an ERRα-targeting or a scramble shRNA (Supplementary Fig. 6D). In coherence with the L2L3M mutant data, ERRα silencing partially blunted the effects of Pgc1α on gene expression and cell growth (Fig. 6g, Supplementary Fig. 6H). In vivo, silencing of ERRα in the presence of the ectopically expressed transcriptional co-activator resulted in a significant increase in bone metastasis incidence from 40% (in Pgc1αexpressing cells transduced with scramble shRNA) to full penetrance (Fig. 6h). Of note, the requirement of ERRα for the effect of PGC1α was recapitulated in vitro with a reverse agonist of the transcription factor, namely XCT790 48 (Supplementary Fig. 6F-I).
It is worth noting that other metabolic pathways have been suggested to sustain the metastatic phenotype. Oxidative stress has been shown to limit metastatic potential in breast cancer and melanoma 29,49 . PGC1α regulates the expression of antioxidant genes, while the enhancement of mitochondrial metabolism can lead to the production of reactive oxygen species (ROS) 28,29,49 (Fig. 4b and Supplementary Table 1). We therefore tested whether ROS production was modified in our experimental settings and if it could contribute to the phenotype observed. Mitochondrial and cellular ROS production were not consistently altered by Pgc1α expression in vitro (Supplementary Fig. 6J). In addition, lipid peroxidation (which serves as readout of ROS production) was unaffected in our xenograft study (Supplementary Fig. 6K). These results are coherent with the inability of antioxidants to rescue the proliferative defect elicited by Pgc1α (Supplementary Fig. 6L).
Our data provides a molecular mechanism by which ERRα activation downstream PGC1α promotes a metabolic rewiring that suppresses PCa proliferation and metastasis.

A PGC1α-ERRα transcriptional signature harbours prognostic potential
We have shown that reduced PGC1A expression in PCa exhibits prognostic potential (Fig.   1c). Since our data demonstrates that transcriptional regulation downstream ERRα is key for the tumour suppressive activity of this co-activator, we reasoned that the association of PGC1α with aggressiveness and DFS should be recapitulated when monitoring ERRα target genes (Fig. 7a). We started the analysis from the list of genes positively regulated by PGC1α in our cellular system (153 genes, Fig. 7b). As predicted, the analysis in two independent patient datasets confirmed that the average signal of the PGC1α gene list was positively correlated with time to PCa recurrence (Fig. 7c). In addition, we observed a decrease in the expression of the aforementioned gene list associated to disease initiation and progression (Supplementary Fig. 7A). Importantly, comparable results were obtained when we performed the analysis with the subset of ERRα-target genes within the PGC1α gene set (73 genes, Supplementary Table 6, Fig. 7b, d and Supplementary Fig. 7B). We next sought to curate the gene list in order to consolidate a prognostic PGC1α-ERRα gene set. We therefore focused on genes that exhibited a strong correlation with PGC1A in patient datasets. We selected genes that were significantly correlated with the co-activator (R>0.2; p<0.05) in at least 3 out of 5 studies. The results unveiled a PGC1α transcriptional signature in patients consisting of 17 genes, the majority of which i) were directly correlated with the transcriptional co-activator in the queried datasets, ii) exhibited decreased expression in PCa vs. BPH and iii) were further down-regulated in metastatic disease (Supplementary Table   7, Supplementary Fig. 7C, D). Nearly 60% of these genes were regulated by ERRα (10 genes out of 17) and were selected for further analysis as a PGC1α-ERRα curated gene set (Supplementary Table 7). The results revealed reduced PGC1α-ERRα curated gene set expression as the disease progressed (Fig. 7e). We next analysed the association of the PGC1α-ERRα curated gene set with disease recurrence. To this end, we compared patients harbouring primary tumours with signal values in the first quartile (Q1) versus the rest (Q2-Q4). Patients with signature positive tumours exhibited reduced DFS in two independent datasets (Fig. 7f). A Hazard ratio (HR) of 4.2 (Taylor) and 17.8 (TCGA) was defined for signature-positive patients, while signature-negative individuals presented reduced risk of recurrence, with HR of 0.23 (Taylor) and 0.05 (TCGA). Furthermore, the frequency of patients with signature-positive signal values was absent or low in the normal prostate group and further increased in metastasis compared to primary tumours (Supplementary Fig. 7E).
Taken together, ERRα-regulated metabolic transcriptional program is associated to the activity of PGC1α in PCa. This interplay is conserved in patient specimens and defines a gene signature that harbours prognostic potential.

Discussion
In this study we provide a comprehensive analysis of master transcriptional co-regulators of metabolism in PCa. Through the use of human data mining analysis, GEMMs and cellular systems, our study presents evidence demonstrating that PGC1α exerts a tumour suppressive activity opposing PCa metastasis. Interestingly, three out of ten significantly altered coregulators (PGC1A, PGC1B and NRIP1, Fig. 1a) in the Taylor 22 PCa dataset (2 out of 3 consistently altered throughout databases, Fig. 1b) converge in the regulation of a common transcriptional metabolic program, led by ERRα 44 and that is associated to the phenotype observed in this study. These data strongly suggest that such pathway is of critical importance for the control of aggressiveness properties in PCa. Indeed, our results demonstrate that a gene set composed of ERRα target genes that are under the control of PGC1α expression 1) is progressively down-regulated in PCa and metastatic disease, and 2) presents prognostic potential for the identification of patients at risk of early recurrence.
The study of the tumour suppressive potential of Pgc1α in mouse models allowed us to characterise a clinically relevant PCa GEMM presenting enhanced metastatic dissemination. PGC1α is added to the short list of genetic events that drive metastasis in this model [13][14][15][16] , and the first to be explicitly linked to the regulation of the metabolic switch. Overall, our finding is of importance for the future study of the requirements for PCa metastasis and therefore for therapeutic purposes.
The sole alteration of PGC1α expression in PCa has profound impact on the oncogenic metabolic switch 50 . This data is in line with the reported activities of this protein in metabolism and mitochondrial biogenesis 26 . Of note, despite the widely accepted fact that the reported metabolic switch 50 has comparable consequences in all cancer scenarios, the study of PGC1α in other tumour types has also revealed a selective pressure towards oxidative processes [27][28][29] . Previous work from others and us defined PGC1α signalling as a selective advantage for breast cancer and melanoma cells 4,[27][28][29]51 . The contribution of this co-activator to cellular proliferation differs between tumour types and experimental systems, promoting growth in melanoma 28 while irrelevant to breast cancer cells 29 . Interestingly, in breast circulating tumour cells, PGC1α expression supports metastatic capacity 29 . The molecular pathways regulating these diverse biological features converge in the activation of ERRα and Peroxisome proliferator-activated receptors (PPAR). While PPAR activation mediates the increase in fatty acid β-oxidation 4 , ERRα is responsible for the overall increase in oxidative metabolism and mitochondrial biogenesis 44 . Similarly, the activation of an antioxidant transcriptional program has been suggested to contribute to anoikis and cancer cell dissemination in a PGC1α-dependent and independent manner 27,28,49,52 . In PCa, however, we demonstrate that the oxidative metabolic program elicited by PGC1α prevents tumour growth and metastatic dissemination, in the absence of overt changes in ROS production, inflammatory response or angiogenic signals. These findings support the notion that the optimal metabolic wiring for tumour growth and metastasis might differ depending on the tumour type, the mutational landscape of the tumour and, potentially, the microenvironment. This would lead to opposite activities of PGC1α depending on the cancer setting, from metastatic promoter 29 to metastasis suppressor (as we demonstrate in the present work).
In summary, our study identifies PGC1α as a master regulator of PCa metabolism that opposes the dissemination of the disease. Therefore, a PGC1α-regulated ERRα-dependent transcriptional program might open new avenues in the identification of metabolic transcriptional signatures that can be exploited for patient stratification and the use of metabolism-modulatory therapies.

Animals
All mouse experiments were carried out following the ethical guidelines established by the Biosafety and Welfare Committee at CIC bioGUNE and The Institutional Animal Care and Use Committee of IRB Barcelona. The procedures employed were carried out following the recommendations from AAALAC. Xenograft experiments were performed as previously described 54 , injecting 10 6 cells per condition in two flanks per mouse. PC3 TRIPZ-HA-Pgc1a cells were injected in each flank of nude mice and 24 h post-injections mice were fed with chow or doxycycline diet (Research diets, D12100402). GEMM experiments were carried out as reported in a mixed background 11,14,55,56 (where the founder colony was cross-bred for at least three generations prior to the expansion of experimental cohorts in order to ensure and homogenous mixed background). The Pten loxP and Pgc1a loxP conditional knockout alleles have been described elsewhere 11,34 . Prostate epithelium specific deletion was effected by the Pb-Cre4 11 . Mice were fasted for 6 h prior to tissue harvest (9 am-3 pm) in order to prevent metabolic alterations due to immediate food intake.
For intra-tibial and intra-cardiac injections BALB/c nude male mice (Harlan) of 9-11 weeks of age were used. Before the injections, PC3 Tripz-HA-Pgc1a (WT, L2L3M, shSC, shERRA) cell lines were pre-treated for 48h with PBS or doxycycline (0.5μg/ml). Mice injected with cells treated with doxycycline were also pre-treated for 48h with 1mg/ml of doxycycline in drinking water. After the injections this group of mice was left on continuous doxycycline treatment (1mg/ml in drinking water). Before the injections mice were anesthetized with mixture of Kethamine (80 mg/kg) and Xilacine (8 mg/kg). For intra-tibial injections, 1×10 4 cells were resuspended in final volume of 5 μl of cold PBS and injected as described previously 57 . For intra-cardiac injections 2×10 5 cells were resuspended in final volume of 100 μl of cold PBS and injected as described previously 57 . Upon the injections tumour development was followed on weekly basis by BLI using the IVIS-200 imaging system from Xenogen. Quantification of bioluminescent images was done with Living Image 2.60.1 software. the development of metastasis was confirmed by doing in vivo or ex vivo (upon necropsy) bioluminescent images of organs of interest (metastasis positivity in lesion incidence analysis was defined as tibias with luciferase signals greater than 50.000 units). When comparing cell lines independently transduced with the luciferase-expressing vector (Fig. 6h), photon flux values per limb where presented as normalized signal (corrected by basal signal, obtained within 24 hours after injection): Normalized photon flux = [day 14 signal/day 0 signal] × 1000. For metastasis-free survival curves metastatic event was scored when measured value of bioluminescence bypasses 1/10 of the day 0 value.

Patient samples
All samples were obtained from the Basque Biobank for research (BIOEF, Basurto University hospital) upon informed consent and with evaluation and approval from the corresponding ethics committee (CEIC code OHEUN11-12 and OHEUN14-14).

Cellular, molecular and metabolic assays
Cell number quantification with crystal violet 58 was performed as referenced.
RNA was extracted using NucleoSpin® RNA isolation kit from Macherey-Nagel (ref: 740955.240C). For patients and animal tissues a Trizol-based implementation of the NucleoSpin® RNA isolation kit protocol was used as reported 59 . For all cases, 1μg of total RNA was used for cDNA synthesis using qScript cDNA Supermix from Quanta (ref. 95048). Quantitative Real Time PCR (qRTPCR) was performed as previously described 51 .
FAO was performed as previously described 51 . Lactate production was performed as referenced 60 using Trinity Biotech lactate measurement kit.
Oxygen consumption rate (OCR) was measured with a XF24 extracellular flux analyser (Seahorse Bioscience) 61 . Briefly, 50.000 cells per well were seeded in a XF24 plate, and OCR measurements were normalized to cell number analysed by crystal violet. Cells were initially plated in 10% FBS DMEM media for 24 hours, and 1h before measurements, media was changed to DMEM serum and bicarbonate free, with glutamine and glucose (10mM). Mitochondrial stress test was carried out using the following concentrations of injected compound: Oligomycin (1μM).
For mitochondrial ATP assays, 50.000 PC3 and DU145 cells plated onto 13-mm coverslips and transfected with a mitochondrial targeted luciferase chimera (mtLuc). Cells were perfused in the luminometer at 37°C with KRB solution containing 25 μM luciferin and 1 mM CaCl 2 and supplemented with 5.5 mM glucose. Under these conditions, the light output of a coverslip of transfected cells was in the range of 5.000-20.000 cps for the luciferase construct vs. a background lower than 100 cps. Luminescence was entirely dependent on the presence of luciferin and was proportional to the perfused luciferin concentration between 20 and 200 μM.
Mitochondrial morphology was assessed by using a cDNA encoding mitochondrial matrixtargeted DsRed (mtDsRed). Cells were seeded onto 24-mm diameter coverslip (thickness between 0.16-0.19 mm) (Thermo Scientific) and 24 h later cells were transfected with 2μg mtDSred (Lipofectamine LTX reagent; Invitrogen). mtDsRed expression was assessed 36 h later. All the acquisitions were performed with a confocal Nikon Eclipse T i system and fluorescent images were captured by using NisElements 3.2.
Lipid peroxidation based on MDA detection was assayed in xenograft samples following the manufacture instructions (MAK085 Sigma-Aldrich).
ROS production was determined by Mitosox and DCF staining as previously described 62 .

Histopathological analysis
After euthanasia, histological evaluation of a Haematoxylin and eosin (H&E) stained section from formalin-fixed paraffin embedded tissues of the following organs was performed: prostate gland, lymph nodes, long bones from lower limbs and other solid organs such as lungs and liver.
Following the consensus reported by Ittmann et al. 63 , prostate gland alterations were classified into 4 categories: gland within normal limits; high grade prostatic intraepithelial neoplasia (HGPIN); HGPIN with focal micro-invasion; and invasive carcinoma. Lymphovascular invasion was assessed in all cases were micro-invasion foci or invasive carcinoma were observed.
Lymph node (LN) metastasis and the presence of groups of PCa cells in bone marrow (BM) were determined after haematoxylin-eosin (H&E) staining (LN) and immunohistochemical identification of cytokeratin (CK) and androgen receptor (AR) -expressing cells using a pan-CK rabbit polyclonal antibody (Dako, Carpinteria, CA) and AR rabbit polyclonal antibody (Santa Cruz Biotechnology, sc-816) (LN and BM). In the case of BM, cases were classified as "dissemination negative" when none or few scattered (less than 5) CK-expressing cells were identified and "dissemination positive" when more than 5 or small groups of these cells were observed.
To assess the inflammatory component in the prostate tissues we performed a semiquantitative analysis in the glandular and the stromal areas separately for each of the specimens. We first determined the type of inflammatory cell present in each tissue compartment: polymorphonuclear neutrophils versus lympho-plasmacytic infiltrates. Then we performed a quantification of these cells using the following score system: 0-no inflammatory cells, 1-few cells, 2-moderate amount of cells and 3-high amount of cells. Scores in between were also determined as 0.5, 1.5 and 2.5. If both types of cells were present in one compartment, we chose the highest as the final score.
Proliferation was assessed in paraffin embedded xenografts samples by using Ki67 antibody (MA5-14520, Thermo Scientific). Microvessel density was determined and quantified in GEMMs and xenograft samples by the immunodetection of CD31 (Rabbit anti-CD31; Ref. ab28364 Abcam).

Metabolomics
LCHR-MS metabolomics and stable isotope 13 C-U 6 -Glucose labelling was performed as reported [64][65][66] . Briefly, PC3 TRIPZ-HA-Flag-Pgc1a cells treated or untreated for 72h with 0.5μg/ml doxycycline were plated at 500.000 cells/well in 6-well plates. For LCHR-MS metabolomics and grown maintaining the doxycycline regime for 42h before harvesting, while for stable isotope 13 C-U 6 -Glucose labelling experiments, 24h after seeding cells were washed and exposed to media with serum, without glucose and pyruvate and supplemented 2mM 13 C-U 6 -Glucose. 16h after that, cells were washed and another 13 C-U 6 -Glucose pulse was performed for 2h before harvesting.
Promoter enrichment analysis was assessed with the Transcription Factors (TFs) dataset from MSigDB (The Molecular Signature Database; http://www.broadinstitute.org/gsea/ msigdb/collections.jsp). TFs dataset contain genes that share a transcription factor-binding site defined in the TRANSFAC (version 7.4, http://www.gene-regulation.com/) database. Each of these gene set was annotated by a TRANSFAC record. A hypergeometric test was used to detect enriched dataset categories.
The GSEA was performed using the GenePattern web tool from the Broad Institute (http:// genepattern.broadinstitute.org). The list of PGC1α upregulated genes ranked by their fold change was uploaded and analysed against a list of ERRα target genes 46 . The number of permutations carried out were 1000 and the threshold was 0.05.

Bioinformatic analysis
Database normalization: all the datasets used for the data mining analysis were downloaded from GEO, and subjected to background correction, log 2 transformation and quartile normalization. In the case of using a pre-processed dataset, this normalization was reviewed and corrected if required.
Frequency of alteration of metabolic co-regulators ( Fig. 1 and Fig. S1A): expression levels of the selected co-regulators were obtained from the dataset reported by Taylor et al 22 . A matrix containing signal values and clinical information was prepared in order to ascertain the up-or down-regulation. We computed the relative expression of an individual gene and tumour to the expression distribution in a reference population (patients without prostate tumour or metastasis). The returned value indicates the number of standard deviations away from the mean of expression in the reference population (Z-score). Using a fold change of +2 and −2 as a threshold, we determined the number of samples from the cancer dataset that were up-or down-regulated. p-values were calculated by comparing the means of normal of cancerous biopsies.
Quartile analysis in DFS: Patients biopsies from primary tumours were organized into four quartiles according to the expression of the gene of interest in two datasets. The recurrence of the disease was selected as the event of interest. Kaplan-Meier estimator was used to perform the test as it takes into account right-censoring, which occurs if a patient withdraws from a study. On the plot, small vertical tick-marks indicate losses, where a patient's survival time has been right-censored. With this estimator we obtained a survival curve, a graphical representation of the occurrence of the event in the different groups, and a p-value that estimate the statistical power of the differences observed.
For PGC1A genomic analysis, data from prostate cancer patients with copy number alteration information in Taylor  Correlation analysis: Pearson correlation test was applied to analyse the relationship between paired genes. From this analysis, Pearson's coefficient (R) indicates the existing linear correlation (dependence) between two variables X and Y, giving a value between +1 and -1 (both included), where 1 is total positive correlation, 0 is no correlation, and -1 is total negative correlation. The p-value indicates the significance of this R coefficient.

Statistics and Reproducibility
No statistical method was used to predetermine sample size. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. Unless otherwise stated, data analysed by parametric tests is represented by the mean ± s.e.m. of pooled experiments and median ± interquartile range for experiments analysed by non-parametric tests. n values represent the number of independent experiments performed, the number of individual mice or patient specimens. For each independent in vitro experiment, at least three technical replicates were used (exceptions: in western blot analysis technical replicates are presented, in untargeted metabolomics two technical replicates were used and for, 13 C-U 6 -Glucose isotope labelling one technical replicate was used) and a minimum number of three experiments were done to ensure adequate statistical power. For data mining analysis, ANOVA test was used for multicomponent comparisons and Student T test for two component comparisons. In the in vitro experiments, normal distribution was confirmed or assumed (for n<5) and Student T test was applied for two component comparisons. For in vivo experiments, as well as for experimental analysis of human biopsies (from Basurto U. Hospital) a non-parametric Mann-Whitney exact test was used, without using approximate algorithms to avoid different outcomes of statistics packages 68 . To this end, we applied the formulas described 69 for small-sized groups and Graphpad Prism for large-sized groups. In the statistical analyses involving fold changes, unequal variances were assumed. For contingency analysis, Fisher exact test as used for 2-group comparison (metastasis incidence) and Chi Square when analyzing more than 2 groups (analysis of PGC1α-ERRα signature frequency in PCa human specimens). The confidence level used for all the statistical analyses was of 95% (alpha value = 0.05). Two-tail statistical analysis was applied for experimental design without predicted result, and one-tail for validation or hypothesis-driven experiments.

Supplementary Material
Refer to Web version on PubMed Central for supplementary material.       experiments). e-f, Evaluation of the metastatic capacity of PC3 Pgc1α WT (upper panels) or PC3 Pgc1α L2L3M (lower panels) expressing cells using intra-tibial xenotransplant assays (e, photon flux quantification, WT, n=6 mice and L2L3M, n=7 mice, 2 hind limb per mice; f, incidence of metastatic lesions presented as histograms). Representative luciferase images are presented referred to the quantification plots. For photon flux analysis, average signal from two limbs per mouse is presented. For incidence analysis, mice with at least one limb yielding luciferase signal > 50.000 units were considered metastasis-positive. (i) and (ii) depict tibia photon flux images from specimens that are proximal to the median signal in − Dox and + Dox, respectively. g, Relative cell number quantification upon ERRα silencing in PC3 Pgc1α expressing cells. Data is represented as cell number at day 4 relative to − Dox cells (n=3, independent experiments). h, Evaluation of metastatic capacity of Pgc1αexpressing PC3 cells transduced with shSC or shERRα using intra-tibial implantation for 14 days (n=8 mice; 2 injections per mice; incidence of metastatic lesions presented as histograms   Table 7). c-d, Correlation between time to recurrence and the average signal of the genes within the PGC1α-upregulated gene set (c) or the PGC1α-dependent ERRα-upregulated gene set (d) in the indicated datasets 19,20,22