Multiple Oxygen Tension Environments Reveal Diverse Patterns of Transcriptional Regulation in Primary Astrocytes

The central nervous system normally functions at O2 levels which would be regarded as hypoxic by most other tissues. However, most in vitro studies of neurons and astrocytes are conducted under hyperoxic conditions without consideration of O2-dependent cellular adaptation. We analyzed the reactivity of astrocytes to 1, 4 and 9% O2 tensions compared to the cell culture standard of 20% O2, to investigate their ability to sense and translate this O2 information to transcriptional activity. Variance of ambient O2 tension for rat astrocytes resulted in profound changes in ribosomal activity, cytoskeletal and energy-regulatory mechanisms and cytokine-related signaling. Clustering of transcriptional regulation patterns revealed four distinct response pattern groups that directionally pivoted around the 4% O2 tension, or demonstrated coherent ascending/decreasing gene expression patterns in response to diverse oxygen tensions. Immune response and cell cycle/cancer-related signaling pathway transcriptomic subsets were significantly activated with increasing hypoxia, whilst hemostatic and cardiovascular signaling mechanisms were attenuated with increasing hypoxia. Our data indicate that variant O2 tensions induce specific and physiologically-focused transcript regulation patterns that may underpin important physiological mechanisms that connect higher neurological activity to astrocytic function and ambient oxygen environments. These strongly defined patterns demonstrate a strong bias for physiological transcript programs to pivot around the 4% O2 tension, while uni-modal programs that do not, appear more related to pathological actions. The functional interaction of these transcriptional ‘programs’ may serve to regulate the dynamic vascular responsivity of the central nervous system during periods of stress or heightened activity.


Introduction
Normal brain function has an absolute requirement for a continuous supply of O 2 , yet it is considered particularly susceptible to oxidative damage. This has been postulated to arise because of the high levels of central nervous system (CNS) O 2 consumption, which is employed to generate ATP primarily through oxidative phosphorylation [1]. This profound energy dependence requires high levels of mitochondrial activity which, along with the presence of excitotoxic and oxidizable neurotransmitters and large, transient Ca 2+ fluxes, contribute to a high degree of potential oxidative stress. Such stress is imposed not only on neurons but also on glia, and although both neurons and glia have antioxidant defense mechanisms [2], the simplest approach to avoiding oxidative stress is to keep brain O 2 levels as low as possible without compromising oxidative phosphorylation. Astrocytes are of particular importance in this regard, as they are now known to regulate local blood supply to match local neuronal activity with remarkable speed and precision [3,4]. Thus, they are of major importance in the control of cerebral blood flow and, hence, brain O 2 levels.
The strategy of maintaining brain O 2 at low but sufficient levels results in reported pO 2 values ranging from ca. 20-30 mmHg despite arterial pO 2 levels being ca. 90 mmHg. Indeed, some estimates have suggested that some 50% of brain regions normally exist at a pO 2 of below 10 mmHg [1,5]. This value therefore represents normoxia for neurons and glia, yet would be considered hypoxic by most other tissues. The brain itself can experience relative hypoxia either at altitude, or as a consequence of cardiorespiratory disorders which result in perturbed O 2 collection in the lungs or distribution by the vasculature, e.g. sleep apnea. If such hypoxia persists, this can predispose individuals to CNS damage, and markedly increases the likelihood of developing progressive dementias such as Alzheimer's disease (AD) [6]. Indeed, we and others have previously shown that hypoxia in vitro leads to increased production of amyloid peptides (Ab) associated with AD [6,7,8].
Local O 2 levels, either hyperoxic or hypoxic, are now recognized as a major determinant of gene expression in all tissues examined to date. Numerous transcription factors (e.g. hypoxia inducible factor (HIF) and nuclear factor kappa B (NF-kB)) are known to be activated in hypoxia, and control the expression of large numbers of genes [9]. Such altered gene transcription and expression is of fundamental importance in the development of multiple tissue disorders such as cancers and dementias [10]. Given the importance of the relative abundance, or paucity, of O 2 to such activity, and the uniquely low levels of O 2 which can be regarded as normoxic in the brain, it is perhaps surprising that no systematic, unbiased, study of physiologically relevant levels of O 2 on gene transcription in primary cortical astrocytic tissue has been previously conducted. We have shown previously that the relative abundance of oxygen (and oxygen-derived reactive species) can exert significant effects upon CNS pathological protein metabolism [6][7][8], as well as strongly regulate the pharmacogenomic 'signatures' of transmembrane receptor signaling systems [11,12]. As the relative oxygen levels in the CNS are likely to demonstrate a high dynamic flux, an understanding of the responsiveness at the cellular level to this, may assist in the development of CNStargeted anti-neurodegenerative therapeutic strategies. Understanding how the CNS system reacts and functions at diverse oxygen tensions may also benefit the rational development of drug-like agents with specific 'context-sensitive' efficacies and potencies [11,13].
Here, we report the effects of various O 2 levels on gene transcription and protein expression profiles in cortical astrocytes, employing levels of O 2 which can be regarded as physiologically normoxic, hypoxic and, importantly hyperoxic, a condition which is commonly imposed on these and other primary cultured cells when typically studied in vitro. Our data indicates that clearly defined, specific 'morphometric' response patterns to varying O 2 tensions exist, that bear distinct and functionally relevant phenotypes to CNS health. An enhanced appreciation of the alterations in the CNS environment during altered O 2 tension will likely assist the development of novel therapies that may demonstrate specific efficacy in separate areas of the brain experiencing divergent levels of perfusion.

Primary astrocyte cultures
Primary cultures of rat cortical astrocytes were obtained as previously described [14]. All animal care and experimental procedures, performed under code (A(SP)A-86), followed United Kingdom Home Office Animals Scientific Procedures guidelines. Animal care and experimental procedures followed United Kingdom Home Office regulations (code A(SP)A-86) and were conducted under the project licence PPL 40/3356 held by Professor C. S. Peers, following the official United Kingdom Home Office Animals Scientific Procedures guidelines. In brief, cerebral cortices were removed from 6-8-day-old Wistar rats and placed in ice-cold phosphate-buffered solution (PBS) containing: 8 mM NaH 2 PO 4 , 2.7 mM KCl, 138 mM NaCl, and 2.7 mM KH 2 PO 4 . Multiplydissected cortices were dispersed into the same buffer containing 0.25 mg/mL trypsin, at 37uC for 15 min. Digestion was halted by the addition of an equal volume of buffer supplemented with 16 mg/mL soy bean trypsin inhibitor (type I-S; Sigma, Poole, Dorset, UK), 20 U/mL DNase I (EC 3.1.21.1 type II from bovine pancreas; Sigma) and 1.6 mM MgSO 4 . The tissue was then pelleted by centrifugation at 10006g for 1 min and the supernatant was poured off before resuspending the cell pellet in 6.8 mL of buffer solution containing 100 mg/mL soy bean trypsin inhibitor, 125 U/mL DNase I and 10 mM MgSO 4 . The tissue was subsequently triturated and, after allowing larger pieces of tissue to settle, the cell suspension was pipetted into media (Eagle's minimal essential medium supplemented with 10% fetal calf serum (v/v) and 1% (v/v) penicillin-streptomycin (Gibco, Paisley, UK)). The cell suspension was then aliquoted into 75 cm 2 flasks. Cells were then maintained in a humidified incubator at 37uC (95% air; 5% CO 2 ). Four to six hours following plating, cells were washed twice with fresh media to remove non-adherent cells. This resulted in a culture of cortical astrocytes, as confirmed by visual inspection the following day and later by glial fibrillary acidic protein immunohistochemistry (data not shown). Any cortical astrocyte culture that was not homogenous was disposed of and not used in this study. Culture medium was exchanged every 7 days and cells were grown in culture for up to 14 days. In some cases, cells were exposed to chronic hypoxia; 24 hr prior to experimentation cells were transferred to a humidified incubator or hypoxic workstation both equilibrated with 1, 4 or 9% O 2 , 5% CO 2 , and the remaining percentage gas N 2 . Control cells were maintained in a 95% air, 5% CO 2 incubator for the same time period.
Once cortical astrocytes had reached approximately 90% confluence (75 cm 2 flask) they were subjected to hypoxia or normoxia, as above, washed with PBS, removed from the flask base with 0.05% trypsin-EDTA (Gibco) and then gently centrifuged (5006g). The cell pellet was then re-suspended in PBS and centrifuged twice more to remove any traces of media. The cell pellet was then: a) if required for RNA analysis; triturated in 8-10 volumes of RNAlater (Applied Biosystems), frozen and stored at 280uC until analysis; b) if required for the analysis of proteins; resuspended in either 10-15 volumes of M-PER reagent (Pierce-Perbio, UK) supplemented with a Complete mini protease inhibitor tablet (Roche Applied Science) or an equivalent volume of a chaotropic solution (7 M Urea, 4% CHAPS, 30 mM Tris at pH 8.5). These samples were frozen and stored at 280uC until analysis.

RNA extraction
RNA isolation was carried out using the Qiagen RNeasy Mini Kit for cell extraction (Qiagen, Inc. Valencia CA). The cells were lysed in the proprietary buffer and then centrifuged. The supernatant was transferred to a second tube and centrifuged again to clear any remaining cellular debris. The supernatant was added to 95% ethanol, mixed and added to the proprietary binding columns. The columns were centrifuged, washed several times and the bound RNA was eluted using water. The RNA quality and quantity was checked using an Agilent 2100 Bio-analyzer and the RNA 6000 nano-chips. As an index of RNA quality we assessed the mean 28S/18S ribosomal RNA values for the samples. The ideal ratio of 28S/18S for intact RNA is 2.0, our measured 28S/ 18S ratio was 1.9860.06 (mean 6 standard error of mean).

Microarray hybridization and analysis
Total RNA was used to generate biotin labeled cRNA using the Illumina TotalPrep RNA Amplification Kit (Ambion; Austin, TX, cat #IL1791). In brief, 0.5 mg of total RNA was first converted into single-stranded cDNA with reverse transcriptase using an oligo-dT primer containing the T7 RNA polymerase promoter site and then copied to produce double-stranded cDNA molecules. The double stranded cDNA was cleaned and concentrated with the supplied columns and used in an overnight in vitro transcription reaction where single-stranded RNA (cRNA) was generated and labeled by incorporation of biotin-16-UTP. A total of 0.75 mg of biotin-labeled cRNA was hybridized at 58uC for 16 h to Illumina's Sentrix MouseRef-8 Expression BeadChips (Illumina, San Diego, CA). Each BeadChip has 24,000 well-annotated RefSeq transcripts with approximately 30-fold redundancy. The arrays were washed, blocked and the labeled cRNA was detected by staining with streptavidin-Cy3. The arrays were scanned using an Illumina BeadStation 5006 Genetic Analysis Systems scanner and the image data extracted using the Illumina BeadStudio software, Version 3.0.

Microarray data analysis
Microarray data were analyzed using DIANE 6.0, a spreadsheet-based microarray analysis program based on SAS JMP7.0 system. Raw microarray data were subjected to filtering and normalization and tested for significant changes as described previously [15]. Raw hybridization intensity data were logtransformed and normalized to yield z-scores, which in turn were used to calculate a z-ratio value for each gene transcript with respect to the control tissues. The z-ratio was calculated as the difference between the observed gene z-scores for the experimental and the control comparisons, and dividing by the standard deviation associated with the distribution of these differences [15]. Z-ratio values $+1.5 or #21.5 were chosen as cut-off values, defining increased and decreased expression, respectively. Array data were analyzed using DIANE 6.0, a spreadsheet-based microarray analysis program based on the SAS JMP7.0 system. Raw microarray data were subjected to filtering and Z normalization and tested for significant changes as described previously [15]. Briefly, sample quality was analyzed by scatter plot followed by gene filtering as follows. A false discovery rate (FDR) cut-off , 0.01 was applied, which controls for the expected proportion of falsely rejected hypotheses. Subsequent remaining genomic data, possessing a z-ratio of $1.5 (6), were further analyzed using a two-way ANOVA design with significance set at p#0.05. Significantly-regulated genes that fulfilled all these criteria were assigned a selector score of 63 (+, upregulated: 2, downregulated) to indicate that they possess a z-ratio of $1.5(6), FDR of #0.01 and an ANOVA p#0.05. Array data for each experimental oxygen tension condition was additionally subjected to hierarchically k-means clustering in DIANE 6.0/Ilumina BeadStudio Version 1.5 to investigate the presence of oxygen tension-related transitional transcriptomic patterns. We have deposited the raw data at GEO/ArrayExpress under accession number GSE29296, we can confirm all details are MIAME compliant.

Bioinformatic geneset analysis
After identifying individual genes that were significantly regulated by different oxygen tensions, the gene lists were analyzed further using multiple forms of functional annotational clustering, i.e. principal component analysis and k-means clustering (using DIANE 6.0/JMP7.0), parametric geneset enrichment analysis (PAGE) [16] using the Broad Institute Molecular Signatures Database (MSigDB: http://www.broadinstitute.org/gsea/msigdb/) canonical signaling pathway (CanPath) or functional network prediction analysis using Ingenuity Pathway Analysis (IPA: http:// www.ingenuity.com/). For PAGE/CanPath analyses similar genomic statistical criteria were employed to ensure significant CanPath and PAGE gene collection 'population'. For each form of grouping, at least two genes were required to fill the requisite group/pathway at a significance of at least p,0.05. For CanPath functional grouping, a single amenable numeric index was created (hybrid score) by multiplying the relative enrichment (to a speciesspecific background set) factor (R) with the negative log 10 of the probability of enrichment. For MSigDB annotation a cumulate Z score (positive or negative) for a specific PAGE gene collection was calculated using the sum of the z ratios of the individual genes clustering significantly into that PAGE collection. In addition to parametric geneset enrichment and IPA CanPath/functional network analysis, we also performed latent semantic indexing (LSI) textual analysis using GeneIndexer (https://computablegenomix. com/geneindexer/) as described previously [11,12,17].

Principal component analysis and global gene regulation
At the most basic analytical level, it was evident from the primary genomic principal component analysis (PCA) that the transcriptomic profile of primary rat astrocytes is highly distinct at differing culturing oxygen tensions ( Figure 1A). Each of the transcription profiles occupied a close-knit and discrete grouping in three dimensional PCA space. The PCA component breakdown into the respective X, Y and Z vectors is depicted in the inset panel in Figure 1A. The inset panel in Figure 1A indicates the quality of extracted RNA (18S, 28S RNA) across samples. A representative quantitation of the 28S/18S ratio is included in the inset. The individual, significantly regulated, gene lists for each experimental oxygen tension, compared to the standard control condition (atmospheric 20% O 2 ), are listed in Tables S1 (1% O 2 ), S2 (4% O 2 ) and S3 (9% O 2 ). Employing a z ratio cut-off (61.5 compared to the mean) the z ratios of the genes significantly altered (relative to 20% O 2 control) by the different oxygen tensions are depicted in the heatmap in Figure 1B. In addition to the z ratio heatmap, using the three-way significant criteria described in Materials and Methods a selector score for all the genes that satisfy this criteria is indicated (upregulated +3, downregulated 23). All of the experimental oxygen tensions (1, 4, 9% O 2 ) caused the significant alteration of hundreds of genes to generally similar z-ratio extents, ranging from a 11-fold (logarithmic) increase to a similar level of decrease. The highest number of significantly altered genes was noted in response to the most extreme level of oxygen tension (1% O 2 ) reduction compared to control ( Figure 1C). Upon inspection of the relative z ratio extent of gene regulation, each of the experimental oxygen tensions induced a similar mean up or downregulation of genes ( Figure 1D), however the 1% oxygen tension did generate the greatest mean elevation in gene transcription (1% O 2 -2.58, 4% O 2 -2.11, 9% O 2 -2.09). In each ambient oxygen tension paradigm however, it was noted that a greater number of genes were significantly downregulated, compared to those upregulated when compared to control (20% O 2 ) conditions.
Inspecting the gene regulation at the 4% O 2 tension compared to the 20% O 2 tension (Table S2), typical hypoxia-regulated factors such as carbonic anhydrase 3, lipocalin-2 [24] and prostacyclin synthase [25] were all strongly upregulated (Table  S2). As with the 1% O 2 tension geneset, there was a strong ribosomally-directed component, as well as a profound heat shock/ chaperone-directed component. Among the relatively novel genes demonstrated to be responsive to this O 2 level, we noted a significant transcriptional potentiation of jagged 1 (Jag1) [26] and pre-proenkephalin (Penk-rs) [27]. Significant downregulation of several transcripts linked to functional hypoxia was also observed, e.g. aurora kinase B [28] and uncoupling protein 2 [29]. In addition, the lipid raft associated protein, flotilin-1, was also noted to be down regulated at the 4% O 2 level relative to 20% O 2 . At the 9% O 2 tension, markers of functional hypoxia were again evident, i.e. carbonic anhydrase 3, prostacyclin synthase and Hig-1 (Table S3). In addition to these factors, we also demonstrated a profound upregulation of the WNT1 inducible signaling pathway protein 2 (Wisp2) [30] and basigin (Bsg) [31]. With regards to the downregulated factors at the 9% O 2 tension level, we noted a strong chronological timing component, as Per1, Per2 and Cry1 were all significantly downregulated. The heatshock/immune function gene, Schlafen-3 was also significantly downregulated.

Multiple oxygen tension-dependent transcriptional activity relationships
We have demonstrated that the O 2 responsive genesets, at our three experimental tensions, indicate significant changes in both classical and novel oxygen-sensitive factors in the primary cortical astrocytes. We next investigated the relationships between the identities of the genes uniquely regulated, or regulated by more than one oxygen tension using Venn diagram analysis. The resultant proportional Venn diagram for the gene set intersections between the specific oxygen tension-regulated dataset is depicted in Figure 2A. The gene identities in the respective unique (A, B, C) and intersected subsets (D, E, F, G) are listed in Table S4. Using this simple Venn diagram separation, it was evident that for each experimental oxygen tension, a relatively unique overall transcriptomic phenotype existed ( Figure 2A). For example, at each specific experimental oxygen tension the majority of the significantly regulated transcripts were unique to that oxygen tension, i.e. 63.2%, 54.2% and 56.4% of the transcripts were uniquely regulated at 1%, 4% or 9% O 2 respectively (Figure 2A, sets A, B, C according to associated key). This primarily unique oxygen tension response reinforces the highly distinguished PCA clustering observed in Figure 1A. The number of shared (between at least two individual tensions) significantly-regulated transcripts between the three experimental oxygen tensions were relatively similar ( Figure 2A). The smallest intersection set was the group of transcripts common to all the experimental oxygen conditions ( Figure 2A).
With respect to the analyses of the genes compartmentalized into the respective intersections, we rationalized the subsets into the following hypothetical groups. Specific subsets of transcripts may regulate astrocyte responsiveness to broad-range dynamic changes in ambient oxygen tensions: 'multi-tension' responsive genes ( Figure 2A, intersection G: altered in response to 1%, 4% and 9% O 2 relative to 20% O 2 ); 'low tension-responsive' genes ( Figure 2A, Table S5, intersection D: significantly regulated in 1-4% O 2 ); 'intermediate-tension' oxygen responsive ( Figure 2A, Table S6, intersection F: significantly regulated in 4-9% O 2 ); 'cross-tension' responsive ( Figure 2A, Table S7, intersection E: significantly regulated in 1-9% O 2 ). In each of the Venn intersections in Figure 2, the majority of transcripts common to two or more oxygen tensions demonstrated similar directions of regulation (upor downregulation). The percentage of significantly regulated transcripts identified at all oxygen tensions ( Figure 2A, intersection G) that demonstrated diverse polarities of regulation was only 16.7% (10 transcripts out of 60: Figure 2B). Similar percentages of transcripts, common between at least two oxygen tensions but with differing regulation polarities, were observed for the other Venn diagram intersections in Figure 2: 1-4% -14 out of 128 (10.9%); 1-9% -28 out of 121 (23.1%); 4-9% -17 out of 116 (14.6%). Therefore it seemed that reversal of transcriptional regulation in astrocytes between different oxygen tensions is a relatively rare event. The large number of transcripts that are uniquely and significantly regulated at only one O 2 tension ( Figure 2A, sets A, B, C) may represent discrete functional gene groups that control steady-state cellular function during constant O 2 tensions. Conversely, the significantly-regulated transcripts in the intersections, D, E, F and G may therefore control dynamic responses to fluctuations in ambient O 2 tensions.
The cross-tension responsive gene subset (regulated at both 1% and 9% O 2 : Table S7) was again strongly divided between up and downregulation with respect to ribosomal activity. The upregulated cross-tension subset again demonstrated a strong ribosomal phenotype (20.4% of transcripts ribosomally-linked). The downregulated cross-tension responsive subset however contained no ribosomally-related genes. The cross-tension subset was heavily populated by genes involved in energy regulation (Tpi1-triosephosphate isomerase, Ldh-lactate dehydrogenase, Pygb-glycogen phosphorylase, Gapdh, Gpi1-glucose phosphate isomerase, Pfkl-phosphofructokinase) and protein chaperoning (Hsp27-heat shock protein 27 kDa, Gadd45b-growth arrest and DNA-damage-inducible 45-b, Hsp1-heat shock protein 1, Hsp8-heat shock protein 8). Interestingly, as with the upregulation of peptidyl prolylisomerase in the low tension geneset, again a small group of transcripts, involved in proline-directed peptide modification, were also upregulated (Dpp7-dipeptidylpeptidase 7, P4hb-prolyl-4 hydroxylase). With respect to the downregulated cross-tension responsive subset we noted that stanniocalcin (Stc1), typically upregulated in peripheral tissues such as the heart [47], was downregulated in the cortical astrocytes. However, it should be noted that in the Westberg study [47], the stanniocalcin transcriptional upregulation was induced by a hypoxic preconditioning action, rather than a chronic exposure, as in our paradigm. Amongst the downregulated transcripts in the cross-tension subset (Table S7), there were functional groups involved with transcriptional regulation (Zfp426zinc finger protein 426, zinc finger, Morc2-CW-type with coiledcoil domain 1, Zfp36-zinc finger protein 36), protein trafficking (Snx14-sorting nexin 14, Srpr-signal recognition particle receptor) and structural/connective component regulators (Adamts9-A disintegrin-like and metalloprotease (reprolysin type) with thrombospondin type 1 motif 9, Srrm2-serine/arginine repetitive matrix 2, Fbn2-fibrillin 2, Agrn-agrin). Downregulation of several receptorrelated factors linked to the amelioration of cell damage or hypoxic stress was also observed in the cross-tension subset, e.g. aryl hydrocarbon receptor (Ahr), which was recently linked to hypoxic regulation of carbonic anhydrases [48]. In addition to downregulation of this receptor we also noted reductions in the levels of the latrophilin (Lphn1) and platelet-derived growth factor (PDGF) (Pdgfra) receptors. It has been previously shown that a-latrotoxin, the active component in black widow spider venom, can selectively stimulate astrocytic cells, inducing death [49], therefore Lphn1 downregulation would be beneficial in times of stress. G proteincoupled receptors can transactivate PDGF receptors [50], and GPCR ligands such as angiotensin appear to control astrocyte growth via PDGF receptors [51]. Therefore, during periodic hypertension or ischemic stress, alterations in PDGF receptor signaling may control astrocyte proliferation or activation of downstream transcription factors such as myc. Reinforcing this concept, was our observation that N-myc downstream-regulated 4 (Ndr4) was also significantly downregulated in this cross-tension subset.

Parametric geneset enrichment analysis
To fully exploit the depth of transcriptomic information contained within any specific geneset, it is crucial to perform as great a functional correlational analysis between individual genes as possible. It is clear that concerted gene group responses underpin multiple and diverse areas of biology [15,52]. Therefore computational analysis and clustering of genes into specific and unbiased functional groups is vital to fully appreciate the subtleties of transcriptional responses to environmental or experimental perturbations and place a physiological relevance to the experimental regulation of individual genes.
To investigate higher-order transcriptional functionality, parametric geneset enrichment analysis (PAGE) was performed using the MSigDB gene collections (http://www.broad.mit.edu/gsea/ msigdb/index.jsp). Significant gene population of a given PAGE geneset collection is created by the inclusion of multiple genes ($2 per PAGE pathway collection: p#0.05) with the resultant z ratio scores being summated to create a total PAGE pathway collection overall Z-score. Therefore signaling pathway groups, possessing many co-upregulated genes will possess a strong positive Z-score, and those significantly populated by multiple correlated downregulated genes will possess a negative Z-score. The complete significantly-populated PAGE pathway collection diagrams generated by the three specific O 2 tension genesets (compared to 20% O 2 ) are represented in Figure S1, and summarized in Tables S8, S9, S10 (Figure S1A-1% O 2 , Table S8; Figure S1B-4% O 2 , Table S9; Figure S1C-9% O 2 , Table S10). To reduce the inherent complexity of multiple PAGE analyses, we grouped related functional pathways (10 highest-scoring) to highlight their specific relationship to the divergent oxygen tensions (Figure 3). PAGE pathways that demonstrated significant (p#0.05) transcript population were grouped into functional areas controlling: oxygen tension responses ( Figure 3A); receptor-mediated processes ( Figure 3B); energy metabolism ( Figure 3C); neurophysiology and pathology ( Figure 3D); cytokine physiology ( Figure 3E) and transcriptional/ translational activity ( Figure 3F). Demonstrating the strong in vivo physiological correlation of our astrocytic culture conditions, we identified multiple oxygen-responsive pathways strongly represented in our datasets ( Figure 3A: MENSE HYPOXIA UP, HYPOXIA REVIEW, HYPOXIA REG UP, HYPOXIA FIBRO UP, HIF1 TARGETS). In addition to the positive cumulative pathway Z-scores of these sets, we also observed a negative score (indicative of multiple down-regulated genes operating in a concerted manner) for PAGE collections typically downregulated with hypoxia (MANALO_HYPOXIA_DN). While hypoxia and oxidative stresses may be considered biophysical phenomena, many of the functional effects of such conditions appear to bear signaling similarities to receptor-mediated effects [12]. In line with this, a strong linkage of the oxygen tension-modulated sets was observed with receptor functions (Figure 3B), including G protein coupled (GPCRS_CLASS_A_RHODOPSIN_LIKE, CXCR4PATHWAY), integrin-related (ST_INTEGRIN_SIGNALING_PATHWAY) and receptor tyrosine kinase receptors (EPHA4PATHWAY). Oxidative stresses may disrupt cellular energetics by causing a mechanistic shift away from mitochondrial oxidative phosphorylation to other forms of cellular energy production [11]. Indeed, we observed a strong clustering of upregulated energy-regulatory pathways associated with alternative energy production mechanisms ( Figure 3C: GLYCOLYSIS_AND_GLUCONEOGENESIS, PPARAPATHWAY), with a concomitant down-regulation of the mitochondrial oxidative phosphorylation functions (ELECTRON_TRANSPORT, MOOTHA_ VOXPHOS, MITOCHONDRIA_PATHWAY). As oxidative and hypoxic events are often associated with the aging process and pathology [2,7,8,11] we extracted PAGE pathways associated with these functions ( Figure 3D). We noted that there was strong positive transcript population of age-related pathways (AGED_MOUSE_ CORTEX_UP, ALZHEIMERS_DISEASE_UP), with a coherent negative population of converse pathways (ALZHEIMERS_DISEA-SE_DN, AGEING_BRAIN_DN). In addition to oxidative events, nervous system aging has also been strongly associated with attenuation of functional stem cell capacity [53]. In accordance with this, we noted that hypoxic conditions downregulated transcripts controlling the upregulation of stem cell populations ( Figure 3D; STEM_CELL_COMMON_DN, STEMCELL_EMBRYONIC_UP), indicating a potential 'pro-aging' phenotype of the datasets. Interestingly, we also noted that there were several other agingrelated groups (HG PROGERIA DN, OLDAGE DN: Table S9, Table  S10) that were profoundly populated (with negative cumulative Zscores) suggesting that an effect of higher oxygen tension may be to prematurely age cell populations.
In addition to oxidative events and alterations in stem cell populations, advancing age and neurological disorders also are linked to cytokine-related inflammatory processes [54]. Specific cytokine functional groups were also demonstrated to possess strong positive transcript population ( Figure 3E: IL3PATHWAY, IL6PATHWAY), while interferon-controlling transcript PAGE collections demonstrated significant negative transcript population ( Figure 3E: IFNAPATHWAY, IFNALPHA_HCC_UP). The various oxygen tensions appear to induce a potent stress response in the astrocytic cells, perhaps indicating their importance as damagemitigating entities in the central nervous system. We therefore extracted PAGE pathways linked to stress responses and translational activity, functions that may be required to replace/correct tissue damage ( Figure 3F). With the various oxygen tensions we identified a strong positive transcript population of protein synthetic (RIBOSOMAL_PROTEINS, TRNA_SYNTHETASES) and stress-response pathways (STRESS_P53_SPECIFIC_UP), cellular development pathways were inhibited (negative transcript population of SHH_UP) and cell cycle inhibitory pathways were antagonized (SMITH_HTERT_DN) ( Figure 3F).
We next analyzed the transitional relationships of PAGE collections between oxygen tensions (Figure 4), in a similar manner to the primary gene transcripts. The Venn diagram distribution of the PAGE pathway collections demonstrated a greater functional similarity, compared to gene transcripts, between PAGE collections across the different O 2 tensions. Hence, the percentages of PAGE collections unique to each tension were lower than those for the gene distribution (Figure 2A (Table S11, E) also failed to demonstrate any strong ribosomal or energy regulation bias but did contain several pathways linked to cellular signaling (P38MAPKPATHWAY, CALCINEURIN NF AT SIGNALING, PTDINSPATHWAY). In sharp contrast to the primary gene distribution (Figure 2A), there were proportionately a greater number of multi-tension PAGE collections common between all three experimental oxygen tensions (Table S11, G). Out of these, many were directly related to hypoxic responses, ribosomal activity or glucose regulation, suggesting that the astrocytes may possess several transcriptomic 'programs' that can coherently respond across a wide range of ambient oxygen tensions. In a profound contrast to this multi-tension group, the other major intersections of upregulated pathways did not share such a diverse functional 'signature'. The specific transitional nature of PAGE collection population for this multitension set is depicted in Figure 4B As there seemed to be important transitional changes, e.g. pivoting around 4% O 2 , in signaling pathways (composed of multiple, functionally-related transcripts) between the diverse oxygen environments we decided to evaluate the importance of these transitional regulatory patterns using k-means clustering [55] across the multiple O 2 tensions.

Morphometric analysis of oxygen tension transitional responses in primary cortical astrocytes
Hierarchical k-means clustering was performed using the primary transcriptomic datasets generated from the three different oxygen tension conditions compared to the 20% control O 2 . Simultaneous clustering of all three experimental datasets, using DIANE-6.0/ SAS-JMP7.0, into clusters that possessed similar directions of transcription changes across the three experimental oxygen tensions (relative to 20% O 2 ) generated 40 specific gene clusters ( Figure 5A, Table S12). These 40 k-means clusters were highly divergent with respect to the number of significantly-regulated genes contained within them, i.e. ranging between only one gene to 197 genes ( Figure 5B). The average number of transcripts in each cluster was 40.967.98 (standard error of mean). Due to this high diversity of k-means clusters present within the transcript datasets, we chose to simplify the multiple k-means clusters through pattern analysis. The initial 40 clusters were therefore first contracted into multiple, morphometrically-classed clusters (morphometric clusters: A-K), based on the transitional gene regulation polarity across the O 2 tensions, and then into color-coded, larger group clusters (ONE-TWO-THREE-FOUR) ( Figure 6A, B). The oxygen tension transitional patterning of these morphometric and group clusters, across the experimental oxygen tensions, is depicted in Figure 6B. Hence group cluster ONE demonstrates transcript elevation at 4% O 2 compared to 1% and 9% O 2 . Group cluster TWO demonstrates transcript reduction at 4% O 2 compared to 1% and 9% O 2 . Group cluster THREE represents consistent potentiation of transcription with increasing hypoxia, while group cluster FOUR demonstrates the opposite pattern to cluster THREE. To demonstrate the strong morphometric relationship of all the original 40 k-means clusters that were compressed into the group clusters ONE-FOUR, we averaged (6 standard error) the mean z ratios across the three experimental tensions for all the genes in the k-means clusters ( Figure 7A-D). This analysis demonstrated a strong and significant coherency of the four distinct group clusters of oxygen-sensitive transcripts. We additionally performed western blot analysis, for the gene products of exemplars from each group cluster (ONE to FOUR), upon protein lysates from astrocytes exposed to the various oxygen tensions, to confirm the transcriptional variation seen with our array data ( Figure 7E-I). Each of the exemplar gene products assessed (c-fos, Calr, Vim, Calm) demonstrated a similar oxygen tension response pattern of protein expression, when compared to the group cluster mean gene transcript regulation ( Figure 7E-I).

Physiological and phenotypic characterization of oxygentension sensitive astrocyte transcriptome clusters
To create a generalized assessment of the physiological signaling significance of the specific group clusters (and regulated genes therein) identified within the astrocyte datasets (ONE to FOUR), we performed an unbiased canonical signaling pathway analysis (Ingenuity Pathway Analysis v. 8.5). Venn diagram delineation of the significantly populated (p#0.05, at least two genes per pathway) canonical pathways demonstrated that the four group clusters contained primarily distinct signaling functions, with only a small degree of cross-over ( Figure 8A: Table S13). Clustering of the significantly-regulated signaling pathways unique to group cluster ONE (Table S14), were primarily concerned with cytoskeletal architecture control, cell death/survival control and lipid metabolism ( Figure 8B). The pathways unique to group cluster TWO (the morphometric inverse of group cluster ONE), unsurprisingly demonstrated a distinct signaling phenotype, i.e. the signaling pathways were associated with energy regulation, nucleotide metabolism and inhibitory neurotransmission ( Figure 8C: Table S14). As with group clusters ONE and TWO, groups THREE and FOUR were characterized by their opposite oxygen tension-regulated behavior. Superficially however, we noted some similarities in the signaling pathway output of these groups, i.e. both involved cellular signaling pathways associated with immune regulation and receptor signaling ( Figure 8D, E). The specific predicted nature of the immune regulation and receptor signaling pathways however demonstrated a significant functional divergence (Table S14). For example, the receptor signaling profile of cluster THREE consisted of glucocorticoid, endothelin-1, angiopoietin, erythropoeitin, neurotrophin and estrogen receptor signaling, while the receptor profile in group FOUR consisted of dopamine, melatonin, glutamate, chemokine (Cxcr4), insulin-like growth factor-1, corticotropinreleasing hormone and androgen receptor signaling instead (Table S14). The largest gross functional differences observed between group clusters THREE and FOUR involved the significantly larger representation of cancer-related pathways in group THREE, compared to FOUR, as well as the presence of hemostatic, bonemetabolism and cardiovascular signaling groups specifically in group cluster FOUR. It therefore appeared that, along with their morphometric divergences, i.e. group clusters ONE/TWO versus THREE/FOUR, their predicted signaling profile was quite distinct as well. In general terms it appeared that group clusters ONE and TWO were more strongly associated with more physiological processes associated with astrocytes, while group clusters THREE and FOUR were more linked to pathology-associated actions. This partitioning may be due to morphometric basis of the dynamic responsiveness of these groups, i.e. ONE and TWO demonstrate bimodal regulation and a fulcrum around 4% O 2 , while THREE and FOUR seem merely responsive to progressive and nonfluctuating changes in oxygen tension. To test this hypothesis we utilized GeneIndexer to perform latent semantic indexing (LSI), a bioinformatic process that allows multidimensional correlation of user-defined 'interrogation' terms with the significantly regulated transcripts within a dataset [11,12,17]. Using LSI we interrogated the transcripts contained in the four group clusters (ONE to FOUR) with terms related to potential astrocyte functions as well as neurological activity (terms 1-9: Figure 9). LSI correlation scores for genes in the group cluster lists with the input interrogation terms were only considered if they were .0.1 (i.e. demonstrating implicit correlation). Two-way matrices of the implicitly correlated transcripts within the group clusters (indicated with colored blocks in the matrix) with the specific input interrogation terms are depicted in Figure S2 (ONE), Figure S3 (TWO), Figure S4 (THREE) and Figure S5 (FOUR). The ten most highly correlated (greatest LSI correlation score across multiple interrogation terms) transcripts from each group cluster LSI matrix are indicated in  transcripts, that were in the top ten highest correlations to the input GeneIndexer interrogation terms ( Figure 9A-D), also formed key nodes within the most statistically reliable, functional gene networks created using the Ingenuity Pathway Analysis (IPA) network-creating algorithm ( Figure S6; Table S15). This may suggest that the LSI analysis has extracted highly relevant and functionally co-related genes in each of the group clusters. For example, in cluster ONE, both Cst-3 and Timp1 were in the top ten of LSI correlations and were present in the highest ranking functional interaction network ( Figure 9A, Figure S6A). In cluster TWO, both Adk and Sirt2 were identified by both techniques ( Figure 9B, Figure S6B). For cluster THREE, we co-identified three separate transcripts using combined LSI and IPA-network analysis, i.e. Vps34/Pik3c3, Ddit3 and Col18a1 ( Figure 9C, Figure  S6C). With respect to group cluster FOUR, using both LSI and IPA network analysis, we co-identified Dcn, Cxcl12 and Ucp2 ( Figure 9D, Figure S6D).
In agreement with our previous analyses, the general predicted functions of the highest scoring gene networks (generated using IPA algorithms) also displayed a variance between clusters ONE/ TWO and THREE/FOUR. Clusters ONE and TWO demonstrated a functional linkage, as both were associated with physiological homeostatic and synthetic regulatory mechanisms (top predicted network functions: ONE-cellular assembly and organization, cellular function and maintenance, protein synthesis: TWO-cellular development, protein synthesis, carbohydrate metabolism) ( Figure  S6; Table S15). In contrast, the highest scoring gene interaction networks present in clusters THREE and FOUR, were more closely associated with pathology and homeostatic disruption (Top predicted network functions: THREE-metabolic disease, renal and urological disease, cell death: FOUR-endocrine system disorders, cellular assembly and organization, metabolic disease) ( Figure S6; Table S15).
To further investigate these potential gross functional differences in the four group clusters we compared, using a novel methodology, the cumulative cluster transcript LSI scores. Cumulative transcript LSI scores were calculated as follows: the number of implicitly correlated transcripts in each group cluster linked to a specific interrogation term (Table S16), were multiplied by the mean correlation score of the transcripts implicitly associated with the input interrogation term (Table S17). It was evident from this comparison that clusters ONE and TWO demonstrated a much stronger association with astrocyte-related functions as well as neurological activities, compared to group clusters THREE and FOUR ( Figure 10A). Averaging across the multiple transcript/interrogation term correlations, we demonstrated that group clusters ONE and TWO possessed a statistically significantly greater cumulative transcript LSI score compared clusters THREE and FOUR ( Figure 10B). Therefore it seems that the group clusters characterized by bi-modal sensitivities to oxygen tension, that pivot around 4% O 2 , appear to be more associated with physiological responses when compared to the THREE/FOUR uni-modal clusters. These findings may suggest that distinct, and physiologically specific, transcriptional programs can be elicited by subtle differences in the dynamic nature of oxygen tension modulation in astrocytes. Therefore, from our transcriptional-transitional profile analysis, it appears that cortical astrocytes potentially possess multiple modes of response to their ambient oxygen tensions and that these can entrain distinct, functional activities mediated by diverse, yet synergistic, cellular signaling systems.

Discussion
In this study we have investigated, in a multidimensional manner, how alterations in ambient oxygen tension can affect the transcriptional activity of primary rat astrocytes. With the application, and transition between, multiple ambient oxygen tension environments, we have noted that primary cortical astrocytic cells generate considerable transcriptomic changes. Many of the patterns of transcriptional activity suggest the presence of underlying Figure 9. Latent Semantic Indexing analysis of group clusters ONE, TWO, THREE, FOUR with astrocytic and neurophysiological activities. Significantly regulated transcripts from the four respective group clusters (ONE-FOUR) were individually analyzed by latent semantic indexing (LSI) using GeneIndexer with the following user-defined interrogation terms: 1-neurodegeneration; 2-Alzheimer's; 3-aging; 4-ischemia; 5-neuroprotective; 6-cognition; 7-hyperoxia; 8-hypoxia; 9-astrocyte. (A) Block matrix of the ten highest LSI correlating genes from group ONE according to each interrogation term (numbered 1-9 atop the block). Four selected genes were verified from each group cluster LSI block. The quantitative and statistical analysis of at least three separate westerns for each chosen protein is indicated in each panel. Statistical significance of the variation of the 1% and 9% datapoints from the 4% O 2 datapoint is demonstrated in a manner similar to that previously performed ( Figure 7A-D). This similar form of statistical analysis was subsequently performed for panels B-D. (B) Block matrix of the ten highest LSI correlating genes from group TWO. (C) Block matrix of the ten highest LSI correlating genes from group THREE. (D) Block matrix of the ten highest LSI correlating genes from group FOUR. Input interrogation terms not implicitly correlating with any significantly regulated transcripts in panels (C) and (D) were omitted from the matrix. Statistical significance was estimated with a Student's t-test using GraphPad Prism version 5.02 (GraphPad Software Inc.). p#0.05 = *, p#0.01 = **, p#0.001 = ***. doi:10.1371/journal.pone.0021638.g009 'programs' of co-regulated transcripts, that are differentially sensitive to multiple levels of ambient oxygen tension.
Astrocytes form a specific population of glial cells that not only create part of the blood brain barrier, but also represent a point of control of the global cerebral vascular blood flow. Astrocytes are often found in close intimate contact with microvessels in the brain and their ability to release vasodilatory or vasoconstrictory factors in response to neuronal activity allows them to divert away, or increase perfusion of blood to different areas of the brain [4]. Additionally, in response to local vascular injury it would be beneficial to reduce the flow to that region if there is considerable vascular damage. In the converse scenario, astrocytes may also be required to increase perfusion to areas of the brain that may be temporarily ischemic. With respect to these functional attributes it is not surprising that astrocytes have been demonstrated to be exquisitely sensitive to ambient oxygen tensions [56]. We have exposed primary cortical rat astrocytes to multiple levels of ambient oxygen tension in our experimental paradigms, i.e. 1%, 4%, 9% and the standard atmospheric oxygen tension, 20% O 2 . These levels were selected as we considered 4% O 2 to potentially represent physiological normoxia for the CNS [1,5], and so 1% O 2 represents relative CNS hypoxia, elative hyperoxia was examined at 9% O 2 (considered normoxic in other tissues) and 20%, which is the standard condition for tissue culture. We have demonstrated that there are highly significant, and contextually intricate, variances in the transcriptomic responses of cortical astrocytes at these different oxygen tensions. This finding is crucial with respect to many of the current experimental conditions employed by the majority of research groups. Thus, in vitro culture of neurons and astrocytes is commonly conducted using humidified atmospheres of 95% air, 5% CO 2 , i.e. 20% O 2 , clearly a hyperoxic situation compared to in vivo ''normoxia''. Thus, we contend that there will be profound differences in cellular physiology between cells maintained at the ambient oxygen tension and oxygen tensions more likely to occur in the central nervous tissue.
With regards to specific transcriptomic changes induced by different oxygen environments, relative to the standard 20% O 2 culture condition, we noted the profound downregulation of Gjb2 (gap junction membrane channel protein beta 2), in response to 1% O 2 exposure. This gene encodes for the protein connexin 26 which has been linked to congenital deafness that is often induced by congenital anoxia [57]. Interestingly, with respect to the involvement of hypoxic conditions in potential neurodegenerative mechanisms, we detected a profound downregulation of the beta amyloid binding precursor (Bbp) [58] and the neuronal regeneration-related protein (Nrep) [59] with exposure to 1% O 2 . The modulation of expression of the lipid raft marker, flotillin-1, may also be important both for synaptic transmission (post-synaptic densities are commonly associated with lipid-rich areas) and organization/control of amyloid precursor protein metabolism, as the Ab-forming gamma secretase enzyme complex is also enriched in lipid raft areas [60]. In addition, flotillin-1 was also recently identified as a functional binding partner for the reversible oxygen binding protein neuroglobin [61].
Correlating to the oxygen tension-modulated alteration of clockrelated transcripts, Per1, Per2 and Cry1 and those linked to protein chaperoning and immune function (e.g. Schlafen-3) (downregulated at 9% O 2 compared to 20% O 2 control: Table S3) we also observed a significant population of many cell cycle/cancerrelated and immune-related canonical signaling pathways (Table  S14), suggesting strong causative links between oxygen tension exposure in astrocytes and potential cancer/cell development/cell cycle regulatory issues [62]. Indeed, a re-entry into the cell cycle of even post-mitotic neurons has been suggested as a potential proneurodegenerative mechanism [63]. When investigating transcripts that demonstrated cross-oxygen tension regulation patterns, we noted several factors that could functionally control neurological functions via their responsivity to ambient oxygen conditions. In our observed low tension-responsive set (Table S4), we identified the upregulation of peptidyl prolylisomerase gene (Pin1), which has been strongly implicated in modification of the pathology of Alzheimer's disease [60,63]; connective tissue growth factor (Ctgf) whose expression and effects upon cellular morphology and rearrangement are oxygen sensitive [64] and metallothienein (Mt1a), that has been previously shown to control anti-apoptotic mechanisms induced by anoxic/hypoxic conditions [65]. In our observed intermediate tension-responsive set (Table S4, F), several other important neurological factors were revealed, e.g. latexin (Lxn) and growth differentiation factor 10 (Gdf10) were both seen to be upregulated in this subset. Latexin, a zinc-dependent metallocarboxypeptidase, has recently been demonstrated to possess the capacity to form long-lasting and highlystable multioligomeric stable structures in response to stress, in a similar manner to beta-amyloid protein [66], while Gdf10 (a controller of cellular development, also known as bone morphogenic protein 3B) expression is positively regulated by both Sox-9 and hypoxic mechanisms [67].
Performing PAGE analysis, to generate a higher order of appreciation of the oxygen-dependent transcriptomic responses, we were able to appreciate the broader picture of the functional relevance of the multiple genetic observations recorded. One interesting finding was the clear demonstration of a protein synthetic response 'pivot' for cortical astrocytes, ( Figure S1B, compared to S1A and S1C). In addition to this protein synthetic pivot, a strong tumor-suppressor and cytokine-related functional phenotype was dominant at the 4% O 2 tension ( Figure 3). Importantly, this oxygen tension may indeed be indicative of the true normoxic condition in central nervous astrocytes [1]. We selected 4% O 2 as an experimental oxygen tension since the majority of brain tissue normally exists at or below this level: 1% O 2 therefore represents hypoxia to most nervous tissue whilst 9% O 2 may represent small arteriolar levels and 20% O 2 represents levels experienced by most cells in conventional tissue culture. Using morphometric-based clustering of the transcriptomic data ( Figures 6, 7), we were able to further characterize important response profiles of the astrocytes in a manner that indicates a potentially wide degree of functional sensitivity in these cells. It was interesting to note that each of the larger morphometric group clusters we statistically identified (ONE, TWO, THREE, FOUR), when analyzed for their potential cell signaling activity, demonstrated relatively unique phenotypes ( Figure 8). The two most different, from a signaling perspective, were interestingly the two 'pivoting' groups (ONE and TWO). Group ONE (and the regulated transcripts therein) seemed likely to manage cytoskeletal dynamics and lipid metabolism, which could potentially indicate a strong role of these genes in modulating neurotransmission in the central nervous system [60,61]. Group TWO, also possessing a similar regulation polarity 'pivot' around the 4% O 2 tension, was primarily linked to energy regulation and inhibitory mechanisms in neurotransmission (e.g. GABA receptor signaling, Table S14). As with our previous data, it appears that astrocytes display a form of 'stability sensor' for 4% O 2 , and thus coordinate distinct signaling pathways dependent on any subsequent change in ambient O 2 tension with unique and diverse genetic programs (Figures 9, 10).
Therefore, using successive unbiased clustering techniques, we have identified four morphometrically-distinct transcription 'programs' in cortical astrocytes (ONE-FOUR) ( Figure 6). As we have seen, these four transcript clusters presented two distinct morphometric patterns, i.e. those with a transcriptional polarity pivot around the 4% O 2 tension (presumed normoxic: (ONE and TWO) and those with a uni-modal transcriptional progression (THREE and FOUR) (Figure 7). A simple interpretation of such transcription models may suggest that the bi-modal groups represent 'programs' of transcripts important for homeostasis of the astrocytic function, while the uni-modal programs indicate responses to continuous or uncontrolled oxidative changes in the CNS. Potentially reinforcing this hypothesis we found, with application of multiple and diverse bioinformatic techniques, that the bi-modal clusters (ONE and TWO) actually demonstrated a functional homeostatic and physiological phenotype (Figures 8, 10) while the uni-modal clusters (THREE and FOUR) possessed a more pathology-related phenotype. Therefore it may be likely that rather than regulating individual transcripts randomly across multiple oxygen environments, astrocytes needing to regulate local vascular tone and neuronal activity, possess concerted transcriptional 'programs' that are engaged in a transitional-activitydependent manner. Such clustering of transcriptional responses would likely aid rapid minute-to-minute control of complex cellular functions such as release of vasoconstrictory agents and neuroprotective behavior. Analysis of the nature of the specific transcripts clustered together in these different group clusters, indicated several potentially interesting aspects to the oxygen tension regulation of cortical astrocyte function. For example, in cluster ONE and TWO respectively, we confirmed the presence and regulation of both Gclc (glutamate-cysteine ligase catalytic subunit) and Gclm (Glutamate-cysteine ligase modifier subunit). These two factors have both recently been intricately implicated in astrocytic stress response mechanisms involving the Nrf2 signaling pathway [68]. The silent mating type information regulation 2, homolog (Sirt2), confirmed in cluster TWO, has been previously demonstrated to play a crucial role in stress response activities associated with aging [69]. Sirt2 has also been recently implicated in oxygen tension-related homeostasis, through a regulatory control of the 14-3-3 scaffolding protein and the apoptosis-related factor Bad [70]. In addition, we identified and confirmed, in cluster THREE, the regulation profile of chordin-like 1 (Chrdl1). Chrdl1 is a bone morphogenetic protein-4 antagonist [71], that has also been demonstrated to act as a regulator of blood vessel dynamics and angiogenesis in retinal pericytes [72]. The progressive increase of expression of Chrdl1 with ascending hypoxia, therefore may also indicate a similar role in cortical astrocytic tissue. In cluster FOUR we found that Cxcl12 expression was successively reduced by increasing hypoxia. The chemokine CXC motif, ligand 12 (Cxcl12), also known as stromal cell-derived factor 1, is a member of the secreted small cytokine family that is considered to be a strong developmental factor in the central nervous system. With respect to its astrocytic activity, Cxcl12 has also recently been implicated in cytotoxic pathways in astrocytes linked to NF-kB and cyclooxygenase-2 [73], suggesting a potentially complex functional role for this protein in the CNS.
While our data has assisted in generating a broader appreciation of the dynamic astrocyte responses to ambient oxygen tensions, we have also demonstrated that gross significant changes can be created, between potentially physiological-CNS oxygen tensions, and the cell culture standard of 20% O 2 . We repeatedly identified that progressive reductions in oxygen tension were correlated with concerted increases in the ribosomal component of the astrocyte transcriptome, suggesting that there may be profound global differences in the resultant protein expression profiles between the ambient oxygen tension employed in typical cell culture experimental paradigms. This may predict that physiological ligand responses may be significantly different between these conditions. Such distinctions may form the basis of many future investigations that could illuminate such potential proteomic shifts between culture conditions. In addition to this general aspect, significant and specific changes in the energy regulatory profile, response to inflammatory cytokines and immune modulators, the control of cytoskeletal dynamics, receptor tyrosine kinase activity, cancer/cellcycle-related and cardiovascular signaling activity were observed in response to dynamic transitional changes in astrocyte oxygen environment. While we have attempted to comprehensively assess astrocyte responsivity to oxygen tensions it is however likely that even in primary cultures of rat cortical astrocytes there is astrocytic cell heterogeneity that may affect specific cellular responses. Hence extracted astrocytes that previously possessed a close proximity to the cerebral vasculature may display variable responses to O 2 tension and transcriptional activity. In future investigations specific microdissection of vessel-proximal or vessel-distal astrocytes may further illuminate the intricacies of oxygen-dependent astrocyte functions. There may therefore be a complex interplay between oxygen tension gradients, created by physical juxtaposition of tissues with distinct metabolic activities, e.g. neurons and microvasculature, and control of astrocytic transcriptional functionality. For example, recent research has suggested that glial cells, neurons and microvascular endothelial cells could be considered to constitute a singular 'neurovascular unit'. This tight functional association of tissues, sensitive to factors pre-disposing to pathology (e.g. hypoxia, Ab release, inflammation) may constitute a system for stressor detection but may also be responsible for pathological ''spreading effects'' characteristic of progressive neurodegenerative disorders such as AD [74].
In addition to illuminating these potential functional aspects of astrocytic function, our current findings though may still bear an important message for in vitro culturing techniques, as it is clear that complex and coherent changes in multiple genomic and signaling systems are likely to occur and may impact any resultant findings from cells/tissue in such conditions.  Figure S2 Latent Semantic Indexing gene-interrogation term matrix for cluster ONE. Each colored block represents a latent semantic indexing correlation score ($0.1) for the specific gene-interrogation term pair in the matrix. The user-defined interrogation terms used were as follows: 1-neurodegeneration; 2-Alzheimer's; 3-aging; 4-ischemia; 5-neuroprotective; 6-cognition; 7-hyperoxia; 8-hypoxia; 9-astrocyte. (TIF) Figure S3 Latent Semantic Indexing gene-interrogation term matrix for cluster TWO. Each colored block represents a latent semantic indexing correlation score ($0.1) for the specific gene-interrogation term pair in the matrix. The user-defined interrogation terms used were as follows: 1-neurodegeneration; 2-Alzheimer's; 3-aging; 4-ischemia; 5-neuroprotective; 6-cognition; 7hyperoxia; 8-hypoxia; 9-astrocyte. (TIF) Figure S4 Latent Semantic Indexing gene-interrogation term matrix for cluster THREE. Each colored block represents a latent semantic indexing correlation score ($0.1) for the specific gene-interrogation term pair in the matrix. The userdefined interrogation terms used were as follows: 1-neurodegeneration; 2-Alzheimer's; 3-aging; 4-ischemia; 5-neuroprotective; 6-cognition; 7hyperoxia; 8-hypoxia; 9-astrocyte. (TIF) Figure S5 Latent Semantic Indexing gene-interrogation term matrix for cluster FOUR. Each colored block represents a latent semantic indexing correlation score ($0.1) for the specific gene-interrogation term pair in the matrix. The user-defined interrogation terms used were as follows: 1-neurodegeneration; 2-Alzheimer's; 3-aging; 4-ischemia; 5-neuroprotective; 6-cognition; 7-hyperoxia; 8-hypoxia; 9-astrocyte. (TIF) Figure S6 Functional network analysis of astrocyte group clusters ONE, TWO, THREE and FOUR. The highest scoring functional interaction network, ranked by the greatest inclusion of genes identified both in the input dataset and those found in the functional predicted network. Network scores were calculated using Ingenuity Pathway Analysis version 8.5. the highest scoring networks for group cluster ONE, TWO, THREE and FOUR are depicted in panels (A), (B), (C) and (D) respectively. (TIF)

Supporting Information
Table S1 Significantly regulated (p,0.05) genes in rat primary astrocytes exposed to 1% ambient O 2 tension compared to 20% O 2 tension. Z ratios were calculated as described in Materials and Methods. (DOC) Table S2 Significantly regulated (p,0.05) genes in rat primary astrocytes exposed to 4% ambient O 2 tension compared to 20% O 2 tension. Z ratios were calculated as described in Materials and Methods. (DOC) Table S3 Significantly regulated (p,0.05) genes in rat primary astrocytes exposed to 9% ambient O 2 tension compared to 20% O 2 tension. Z ratios were calculated as described in Materials and Methods. (DOC)   Significantlyregulated genes were used to populate the specified MSigDB collections. '# genes in collection' describes the total gene count of the specific MSigDB collection and the '# exp genes in collection' describes the number of genes from the input experimental set that were able to significantly populate the specific MSigDB collection. The Z score is calculated based upon the cumulative z ratios of the respective genes from the experimental dataset that populated the specific MSigDB collection. (DOC) Table S9 Significantly populated PAGE gene collections created with transcripts responding to 4% O 2 tension compared to control 20% O 2 condition. Significantlyregulated genes were used to populate the specified MSigDB collections. '# genes in collection' describes the total gene count of the specific MSigDB collection and the '# exp genes in collection' describes the number of genes from the input experimental set that were able to significantly populate the specific MSigDB collection. The Z score is calculated based upon the cumulative z ratios of the respective genes from the experimental dataset that populated the specific MSigDB collection. (DOC) Table S10 Significantly populated PAGE gene collections created with transcripts responding to 9% O 2 tension compared to control 20% O 2 condition. Significantly-regulated genes were used to populate the specified MSigDB collections. '# genes in collection' describes the total gene count of the specific MSigDB collection and the '# exp genes in collection' describes the number of genes from the input experimental set that were able to significantly populate the specific MSigDB collection. The Z score is calculated based upon the cumulative z ratios of the respective genes from the experimental dataset that populated the specific MSigDB collection. (DOC)   for the specific oxygen tension versus 20% O 2 for each transcript are displayed alongside the k-means generated cluster number. (DOC) Table S13 Venn diagram analysis of canonical signaling pathway analysis for the four classified group clusters. The four-way Venn diagram describes the relationships between the statistically significant canonical signaling pathways populated by the multiple transcripts contained in the four group clusters, ONE, TWO, THREE, FOUR. The intersection subsets, A to O, of significantly populated signaling pathways are depicted in Figure 7. (DOC) Table S14 Physiological relevance of group clusters ONE-, TWO-, THREE-and FOUR-unique signaling pathways. For each canonical signaling pathway significantly populated using the group cluster input genes a hybrid score was created that was the generated by multiplication of the negative log 10 of the probability of enrichment of the genes in the respective signaling pathway with the gene enrichment ratio relative to a species-specific background geneset. For each significantly-populated signaling pathway the genes from the input dataset that generated the associated hybrid score are denoted.

(DOC)
Table S15 Highest scoring gene networks present within group clusters ONE, TWO, THREE and FOUR. For each group cluster geneset, functional interaction networks were created using Ingenuity Pathway Analysis (IPA). Network scores are created by the structural strength of the network in addition to the number of focus genes (genes present in network also present in the input experimental dataset: focus genes are indicated in bold). The IPA-predicted primary functions of the network are indicated for each of the four group clusters of genes. (DOC) Table S16 Number of implicitly correlated genes, from group clusters ONE, TWO, THREE and FOUR, associated with each LSI interrogation term. The number of genes from each group cluster dataset that demonstrated at least an implicit correlation (score $0.1) to the specific LSI interrogation term 1-9 is represented. (DOC) Table S17 Mean Latent Semantic Indexing correlation scores for group clusters ONE, TWO, THREE and FOUR. The mean LSI correlation score of all the genes that demonstrated an individual implicit correlation (score $0.1) to the specific LSI interrogation term 1-9. (DOC)