GSEA of mouse and human mitochondriomes reveals fatty acid oxidation in astrocytes

The prevalent view in neuroenergetics is that glucose is the main brain fuel, with neurons being mostly oxidative and astrocytes glycolytic. Evidence supporting that astrocyte mitochondria are functional has been overlooked. Here we sought to determine what is unique about astrocyte mitochondria by performing unbiased statistical comparisons of the mitochondriome in astrocytes and neurons. Using MitoCarta, a compendium of mitochondrial proteins, together with transcriptomes of mouse neurons and astrocytes, we generated cell‐specific databases of nuclear genes encoding for mitochondrion proteins, ranked according to relative expression. Standard and in‐house Gene Set Enrichment Analyses (GSEA) of five mouse transcriptomes revealed that genes encoding for enzymes involved in fatty acid oxidation (FAO) and amino acid catabolism are consistently more expressed in astrocytes than in neurons. FAO and oxidative‐metabolism‐related genes are also up‐regulated in human cortical astrocytes versus the whole cortex, and in adult astrocytes versus fetal astrocytes. We thus present the first evidence of FAO in human astrocytes. Further, as shown in vitro, FAO coexists with glycolysis in astrocytes and is inhibited by glutamate. Altogether, these analyses provide arguments against the glucose‐centered view of energy metabolism in astrocytes and reveal mitochondria as specialized organelles in these cells.

. Along these lines, the intrinsic neuronal activity at resting conditions may depend in part on aerobic glycolysis taking place in astrocytes (Raichle, 2015). Lactate availability is necessary to maintain excitatory transmission (Nagase, Takahashi, Watabe, Kubo, & Kato, 2014), and facilitates synapse formation and growth in development (Goyal, Hawrylycz, Miller, Snyder, & Raichle, 2014), albeit the source and levels of lactate are controversial (Dienel, 2017). In fact, the ANLS hypothesis has not been proven in humans. Results in favor of or against the hypothesis have been reported from animal studies in vivo (Ashrafi & Ryan, 2017;Diaz-Garcia et al., 2017;Mächler et al., 2016), and illustrate the complexity of brain metabolism.
Whatever the nature and source of the fuels used by neurons may be, the general consensus is that mitochondria are fundamental to neurons, which are oxidative; while astrocytes are glycolytic cells at the service of neuronal energetic needs. This view underestimates the role that mitochondria may play in astrocytes. However, several observations link excitatory neurotransmission with astrocytic anaplerosis and cataplerosis (i.e., reactions that respectively replenish or deplete TCA intermediaries). For example, glutamate taken up by astrocytes is converted in part to alpha-ketoglutarate by oxidative deamination and is incorporated into the TCA cycle (Belanger & Magistretti, 2009;McKenna, 2007;McKenna et al., 2016;Sonnewald, Westergaard, & Schousboe, 1997), to generate pyruvate and lactate by cataplerotic reactions (Schousboe, Scafidi, Bak, Waagepetersen, & McKenna, 2014;Sonnewald, Westergaard, Petersen, Unsgard, & Schousboe, 1993).
Overall, the view that astrocytes are only glycolytic is simplistic and underestimates astrocyte metabolic functions. To better understand the complexity of neuroenergetics, we investigated astrocyte mitochondrial functions. On the premise that the functions of mitochondria can be inferred from their molecular makeup, we compared the "mitochondriome" of astrocytes and neurons. The mitochondriome was based on the MitoCarta (Pagliarini et al., 2008) and on transcriptome databases of cells isolated from four sources: normal mouse brains, mouse spinal cords before and after injury, cultured postnatal astrocytes, and fetal and adult human brains. Functional validation was carried out in cultured astrocytes. The findings indicate that the metabolism of astrocytes is versatile and complex and that mitochondrial fatty acid oxidation (FAO) in astrocytes is a key element in brain metabolism. Our results therefore imply that core tenets in neuroenergetics, such as glucose is the main brain fuel, ought to be revised.

| Databases and their treatment
We used MitoCarta (Pagliarini et al., 2008), the mouse transcriptome (Anderson et al., 2016;Cahoy et al., 2008;Doyle et al., 2008;Lovatt et al., 2007;Zhang et al., 2014) and the human brain transcriptome databases (Zhang et al., 2016). We applied the alias2Symbol function from Bioconductor organism packages to convert all gene identifiers from all databases to their official gene symbols. We used logarithm base two-fold changes of expression (log2FC) to assess expression differences between conditions. All in-house R scripts are available under request. Organelle gene lists were extracted from the "Cellular Component" ontology of the GO terms using AmiGO 2.0 software. Lists of genes used for each organelle are available in Supporting Information Dataset S5.

| Standard gene set enrichment analysis (GSEA)
We performed GSEA with the software provided by the Massachusetts Institute of Technology (Subramanian et al., 2005). Normalized enrichment score (NES) and false discovery rate (FDR) were used to quantify enrichment magnitude and statistical significance, respectively.

| In-house GSEA
We designed an in-house GSEA that uses the proportion of genes with a log2FC greater than 2 (or lower than 22) within a given gene set to determine the magnitude of enrichment under a particular condition or conditions. The statistical significance of the enrichments were assessed post-hoc with an exact binomial test considering the proportion of genes with log2FC greater than 2 (or lower than 22) with respect to the proportion of over-expressed genes of each cell type in the whole database.

| R E SU LTS
3.1 | Generation of astrocyte and neuron "mitochondriomes" in healthy adult brains To identify the functional signature of astrocyte mitochondria with respect to neurons, we used the Mitocarta database-the most comprehensive inventory of the molecular composition of mitochondria, comprising of data obtained from 14 mammalian tissues (Pagliarini et al., 2008). Mitocarta contains 1,098 proteins, which were converted into 919 genes with the Bioconductor mouse package (Supporting Information Dataset S1). We first manually classified the mitochondrial components identified by MitoCarta (Supporting Information Dataset S1) into 12 functional categories (inner circle, Figure 1). Each category was further divided into groups. We thus had 12 categories with two levels of subdivision ( Figure 1 and Supporting Information Dataset S1).
For example, the functional category Cellular energy homeostasis was divided into ATP-related metabolism, Carbohydrate catabolism, Co-factors metabolism, Ketone body metabolism, Lipid catabolism, Lipid synthesis and Oxidative metabolism. The last group, in turn, was subclassified in Anaplerosis and cataplerosis, Oxidative phosphorylation and TCA. Among all the 919 genes analyzed only 69 fell into more than one functional category. Obviously, categories may contain different numbers of genes. Cellular energy homeostasis is the largest group with 287 genes. This already implies which mitochondrial functions could be more important.
Next, mitochondriomes were generated by extracting nucleusencoded mitochondrial genes and their relative expressions from the transcriptomes of mouse astrocytes and neurons isolated from whole brains or spinal cords. The brain transcriptomes used are available in five repositories (Anderson et al., 2016;Cahoy et al., 2008;Doyle et al., 2008;Lovatt et al., 2007;Zhang et al., 2014), and in the Glia Network website (http://www.networkglia.eu/en). We will henceforth refer to these databases by the name of the first author of the study, viz. Zhang (Zhang et al., 2014), Doyle (Doyle et al., 2008), Cahoy (Cahoy et al., 2008), Lovatt (Lovatt et al., 2007), and Anderson (Anderson et al., 2016). The studies made use of mice at different ages (Zhang, young

| Mouse GSEA
The mitochondriomes of astrocytes and neurons were then compared by standard GSEA (Subramanian et al., 2005), which allows detecting statistically significant expression differences of a given set of genes in different phenotypes or conditions. Because the standard GSEA requires raw expression data, and these are only available in the Cahoy, Doyle and Zhang databases, only these three databases were used at this stage. As in all other analyses in the present manuscript, only To include Lovatt data in the analyses, we took the log 2 (fold change) of the data (log2FC), and performed an in-house GSEA. In brief, this method gives us the percentage of over-expressed genes (log2FC > 2 or< 22) in each functional group and gene-set percentages in neurons and astrocytes are compared with a binomial analysis.
The in-house GSEA showed a high coincidence with the previous standard GSEA. No functional group is over-represented in neurons compared with astrocytes, but three groups are over-represented in astrocytes compared with neurons in at least two transcriptomes: Cellular-energy homeostasis/Lipid catabolism/Fatty acid betaoxidation, Protein homeostasis/Protein catabolism/Amino acid catabolism, and Amine, alcohol and aldehyde metabolism ( Figure 2b).
Since the Lovatt database presents slightly different patterns of gene expression, we performed a correlation analysis of gene expression among databases. The Zhang, Doyle and Cahoy databases present high correlation coefficients between each other, and low correlations with Lovatt ( Figure 3a). One can speculate that, different subpopulations of astrocytes were selected for the Lovatt study, or that older mice brains perform less oxidative metabolism (discussed below).
Overall, the data indicate over-representation of mitochondrial gene expression in astrocytes versus neurons. Importantly, this overrepresentation is not due to transcriptome bias, because the distribu- The results contradict the prevailing view that astrocyte mitochondria are of lesser importance than neuronal mitochondria.

| Fatty acid oxidation in astrocytes
To gain further insight into the distinct functional signature of astrocyte mitochondria, we examined the individual genes over-expressed in astrocytes in at least two transcriptomes (Table 1) , 2007;Sonnewald et al., 1993). Hence glutamine and/or lactate could be supplied to neurons. Neurons could convert glutamine to glutamate via glutaminases Gls and Gls2, which have a higher expression in neurons compared with astrocytes (Table 1). Alternatively, acetyl-CoA from FAO could be diverted to ketogenesis, as astrocytes overexpress Hmgcs2 which encodes HMG-CoA synthase 2, the ratelimiting enzyme of this pathway (Table 1). But some authors claim ketogenesis in the brain of fed animals is minimal (Ebert et al., 2003).
Altogether, the computerized analyses suggest that FAO is a constitutive metabolic pathway in astrocytes, compatible with glycolysis.
To validate this prediction experimentally, we measured FAO in astro-

cytes by quantifying the production of [ 3 H]H 2 O from [ 3 H]palmitic acid.
We chose cultured astrocytes to demonstrate unequivocally that astrocytes perform FAO. First, we confirmed that cultured astrocytes were a valid model by comparing the mitochondriomes of cultured astrocytes versus astrocytes isolated from adult mice (Cahoy et al., 2008) with the in-house GSEA (Supporting Information Dataset S3). We observed that there is no significant difference in Fatty acid betaoxidation between cultured and brain-isolated astrocytes ( Figure 4a).
Second, we detected FAO in astrocytes. In the presence of 10 mM glucose FAO decreased by 50%, but it was not abolished (Figure 4c).
The data thus demonstrate the coexistence and interplay of FAO and glycolysis. Third, we examined FAO in the context of acutely increased glycolysis, as it occurs in vivo during task-elicited excitatory neurotransmission. This entails the release of glutamate that stimulates glucose uptake and glycolysis rate in astrocytes (Pellerin & Magistretti, 1994).
To mimic this scenario, 100 mM glutamate was added 3 hr after the ini-

| Astrocyte mitochondria upon injury
The data thus far indicate that FAO is constitutive in normal astrocytes and finely regulated. In order to study what happens in injury, we compared with standard GSEA the mitochondriomes of spinal-cord astrocytes versus non-astrocytic cells 2 weeks after spinal-cord injury caused by severe crush and lateral compression, a model characterized by the appearance of reactive astrocytes (Anderson et al., 2016). Note that under normal conditions, the expression of mitochondrial genes in spinal cord astrocytes compared with other cell types is similar to astrocytes versus neurons in the healthy forebrain (Supporting Information Figure S1 and Dataset S3). Strikingly, the analysis revealed that mitochondrial functions were globally reduced in astrocytes after injury, with a statistically significant decrease in: Cellular energy homeostasis/Pyruvate metabolism, Lipid metabolism/Fatty acid oxidation, Oxidative metabolism/Oxidative phosphorylation and TCA, Mitochondrial DNA translation, Amino acid catabolism, and Small-molecule transport ( Figure 5a). Although, we cannot assign these changes of the mitochondriome to a specific astrocyte population among those reported to appear in the injured brain (Wanner et al., 2013), our data support that fatty acid and amino acid based oxidative metabolism is a normal function of mitochondria that is impaired in brain injury.

| Astrocyte mitochondria in human brains
At present there is no evidence that FAO takes place in the human brain, probably due to the difficulty of measuring such activity in living subjects or in postmortem samples. Therefore, we took advantage of human brain transcriptomes available in Zhang et al., 2016  Mitochondrion activity is associated with cell differentiation (Wanet, Arnould, Najimi, & Renard, 2015). The databases used so far derive from cells isolated from mice and humans at different ages, suggesting that FAO is present in both young and adult brains. But does it increase or decrease with astrocyte maturation? That is, is mitochondrial activity also crucial for astrocyte maturation? As noted, the comparison between cultured postnatal astrocytes and astrocytes isolated from adult brain serves to argue in favor of a potentiation of astrocyte mitochondrial functions with maturation. To further analyze this hypothesis, we compared, with standard GSEA, mitochondriomes from astrocytes isolated from fetal and adult human brains (Zhang et al., 2016). Finally, we wanted to check if the increased expression of genes encoding for mitochondrial proteins upon maturation is exclusive to this organelle or it is a general trait. To this purpose we examined the distribution of genes encoding for components of different organelles with respect to their relative gene expression in adult versus fetal human astrocytes (in log2FC; Supporting Information Dataset S5). The distribution of "total genes," genes related to the "endoplasmic reticulum" and genes related to the "Golgi apparatus" is even between fetal and adult cells. By contrast, genes encoding for 'nuclear proteins' are over-represented in fetal astrocytes, whereas genes encoding for 'mitochondrial proteins' are over-represented in adult astrocytes (Figure 5d).
This indicates that the increased production of mitochondrial constituents during development is a unique feature of this organelle, at least in human astrocytes.

| D I SCUSSION
This study represents the most extensive and integrated analysis of the expression of nuclear-genes encoding for mitochondrial proteins in the nervous system, to date. We have used data from: (1) four databases generated from mouse brains at different ages using different methodologies either to identify and isolate cells or to analyze gene expression; (2) cultured astrocytes; (3) fetal and adult human astrocytes; and (4) spinal cord astrocytes. The results support the relevance of astrocyte mitochondria and define different mitochondrial functional signatures in astrocytes versus neurons. Importantly, we strongly suggest that FAO is active in human astrocytes.
Our analyses show that the expression of TCA and oxidativephosphorylation-related genes in mice is not different in astrocytes and neurons. Previous studies have reached the same conclusions as our analyses. In fact, Lovatt and colleagues reported that genes related to oxidative metabolism are upregulated in astrocytes versus neurons (Lovatt et al., 2007), but our analyses are more restrictive and do not reach statistical significance. Moreover, activities of complexes I-IV of the mitochondrial ETC have been recently shown to be the same in neurons and astrocytes isolated from mouse brains, despite the fact that complex I is predominantly assembled into super complexes in neurons but not in astrocytes (Lopez-Fabuel et al., 2016). That astrocytes are oxidative in humans is also supported by nuclear magnetic resonance in vivo showing mitochondrial glutamate oxidative metabolism after the addition of acetate (Bluml, Moreno-Torres, Shic, Nguy, & Ross, 2002;Lebon et al., 2002). Despite this evidence, the possibility that astrocytes are oxidative has been overlooked. Indeed, most reviews about brain metabolism, even recent ones, consider glucose as the only energy substrate, FAO being ignored ( -Fabuel et al., 2016;Raichle, 2015), or denied (Schonfeld & Reiser, 2013. However, FAO was calculated to account for 15-20% of total brain oxidative energy production (Ebert et al., 2003;Panov et al., 2014), and direct and indirect evidence argue in its favor (Cabodevilla et al., 2013;Ebert et al., 2003;Edmond, Robbins, Bergstrom, Cole, & de Vellis, 1987;Jernberg et al., 2017;Kawamura, 1988;Panov et al., 2014;Sayre et al., 2017). Nevertheless, which cells perform FAO in whole brains has not been clarified. Our study provides molecular evidence to support that FAO takes place in the CNS, and its compartmentalization is mainly in astrocytes. Unlike previous findings, our conclusions come from an extensive range of transcriptomes of isolated astrocytes, FAO being a significant hit in all the astrocyte mitochondriomes analyzed. Robustness of the conclusions is also supported by the fact that all the FAO enzymes are overexpressed in astrocytes versus neurons at least in two mice transcriptomes, and this coincides with human data. A drawback of our study is that it is mainly based on data from transcriptomic studies, and hence it may not precisely reflect metabolic fluxes (Chubukov et al., 2013). Nevertheless transcriptomic studies have previously led to the identification of metabolic pathways (Corominas et al., 2013;Hu, Gu, Sun, Bai, & Wang, 2016) and they are a sound approach to address brain metabolism in humans (Barros et al., 2017). Moreover, genomic-based models have been extensively used to study metabolic networks (Duarte et al., 2007) and to simulate metabolic-based phenotypes of cancer cells (Gruetter, Seaquist, & Ugurbil, 2001;Lewis & Abdel-Haleem, 2013;Lewis et al., 2010).
Our experiments in vitro confirm the co-existence of aerobic glycolysis and FAO in astrocytes, as forecast in a review (Panov et al., 2014). They also agree with recent findings that suggest FAO occurs in the presence of glucose in mouse hippocampus and cortex (Jernberg et al., 2017). We show that FAO in astrocytes is versatile, because it is inhibited by high concentrations of glutamate, resembling states of high excitatory neuronal activity. It is well established that glutamate up-regulates glycolysis in astrocytes (Bittner et al., 2011;Loaiza, Porras, & Barros, 2003;Pellerin & Magistretti, 1994), which may account for [18F]-Fluorodeoxyglucose PET neuroimaging (Zimmer et al., 2017).
However, the inhibition of FAO by glutamate does not appear to depend on glycolysis, because it also takes place in the absence of glucose. Without ruling out the possible involvement of glutamatergic receptors, glutamate induced inhibition of FAO might be a result of its uptake leading to: (1) an inhibition of the mitochondrial long-chain fatty acid transporter CPT1, as has been shown in permeabilized hepatocytes (Guzm an, Velasco, Castro, & Zammit, 1994); (2) activation of acetyl-CoA carboxylase, increasing levels of malonyl-CoA, and hence inhibiting CPT1, as shown previously in other tissues (Boone, Chan, Kulpa, & Brownsey, 2000); and/or (3) acidification of mitochondria and reduction of oxygen consumption (Azarias et al., 2011). Whatever the precise mechanism of FAO inhibition might be, it seems paradoxical that glutamate can inhibit FAO while entering the TCA and giving rise to lactate to be delivered to neurons (McKenna, 2007;Sonnewald et al., 1993). However, acetyl-CoA entering the TCA could come from other sources than FAO, namely amino acid catabolism, glucose or glycogen, as previously reported (Dienel & Cruz, 2004;Nissen et al., 2017;Nissen, Pajecka, Stridh, Skytt, & Waagepetersen, 2015;Swanson, 1992).
We suggest that FAO is key for brain metabolism. The fine regulated use of glucose, fatty acids and amino acids by astrocytes provide the brain with high flexibility. FAO also ensures ATP production needed under different physiological or pathological conditions. In this regard, our observations that FAO enzymes are downregulated in spinal cord injury agree with a metabolic depression suggested after acute spinal cord trauma (Levine et al., 2016) and with observations that FAO is protective in stroke (Sayre et al., 2017). Thus we propose a role of FAO in brain disease progression and that its restoration could have a therapeutic benefit. Considering that astrocytes become a heterogeneous population after spinal cord injury including reactive and newly generated astrocytes (Wanner et al., 2013), future studies should clarify the fate of FAO in each astrocyte type at different times after injury, and what happens to astrocyte FAO in neurodegenerative diseases.
A most important implication of the existence of a versatile mitochondrial metabolism in astrocytes of adult human brain, including oxidative metabolism by way of FAO and amino acid catabolism, is the interpretation of imaging techniques that monitor neuronal activity based on oxygen consumption or availability. An example is BOLD-MRI, which is widely used to monitor neuronal activity, and that also depends on blood flow. First, if astrocyte FAO consumes high amounts of O 2 and it is inhibited by glutamate, it follows that the increase in glucose consumption in astrocytes triggered by excitatory neuronal activity would be accompanied by a decrease in O 2 consumption. This might explain one of the big enigmas in neuroenergetics and neuroimaging: why are task-elicited increases in blood flow and glucose consumption only marginally matched by increases in O 2 consumption (Belanger et al., 2011;Fox, Raichle, Mintun, & Dence, 1988)? Second, is astrocyte FAO coupled to blood flow and O 2 availability, as measured by the BOLD signal? Third, beyond task-dependent metabolic requirements-the most studied paradigm in the relationship between brain function and energy metabolism; does astrocyte FAO sustain astrocyte functions during intrinsic activity, the activity of the brain in the absence of external inputs (Raichle, 2015)?
Finally, we demonstrated the over-expression of genes encoding for mitochondrial proteins compared with other organelles in adult human astrocytes including FAO, amino acid catabolism and oxidative metabolism-related genes. Our results are an important addition to findings showing that glycogen metabolism, aerobic glycolysis and lactate release are markers of astrocyte differentiation from neural stem cells (Brunet, Allaman, Magistretti, & Pellerin, 2010;Brunet et al., 2004). Our data confirm that astrocytes also follow the golden rule of increased mitochondrial activity and metabolism during differentiation (Wanet et al., 2015), but reveal cell-specific particularities. For example, when neurogenesis is induced by PGC1alfa-mediated mitochondrial biogenesis, there is no upregulation or downregulation of FAO enzymes (Stoll et al., 2015). However, according to our data, astrocyte maturation implies an increase in FAO and amino acid catabolism, in agreement with the developmental upregulation of carnitine palmitoyl transferases and acyl-CoA dehydrogenases in young mice (Jernberg et al., 2017). We thus argue that the key point in deriving astrocytes from iPSCs, or neurons from astrocytes, is mitochondrial specialization, and not the switch from a non-oxidative to an oxidative phenotype.
Interestingly, a meta-analysis of glucose and oxygen consumption across the human lifespan concluded that anaerobic glycolysis appears to decrease dramatically in the elderly, whereas oxygen consumption is slightly higher (Goyal et al., 2014). This raises the possibility that astrocyte FAO gains importance in aging.
In summary, the study reveals astrocyte mitochondria as key, specialized organelles in the adult brain and broadens the perception that astrocytes are merely glycolytic cells. They are glycolytic and fatty acid and amino acid oxidative cells. This metabolic complexity should be taken into consideration when examining the relationship between brain energy metabolism and function in health and disease, as well as in development.