Mild brain hypothermia (32°–34°C) after human neonatal asphyxia improves neurodevelopmental outcomes. Astrocytes but not neurons have pyruvate carboxylase and an acetate uptake transporter. 13C nuclear magnetic resonance spectroscopy of rodent brain extracts after administering [1-13C]glucose and [1,2-13C]acetate can distinguish metabolic differences between glia and neurons, and tricarboxylic acid cycle entry via pyruvate dehydrogenase and pyruvate carboxylase.
Neonatal rat cerebrocortical slices receiving a 13C-acetate/glucose mixture underwent a 45-min asphyxia simulation via oxygen–glucose-deprivation followed by 6 h of recovery. Protocols in three groups of N = 3 experiments were identical except for temperature management. The three temperature groups were: normothermia (37°C), hypothermia (32°C for 3.75 h beginning at oxygen–-glucose deprivation start), and delayed hypothermia (32°C for 3.75 h, beginning 15 min after oxygen–glucose deprivation start). Multivariate analysis of nuclear magnetic resonance metabolite quantifications included principal component analyses and the L1-penalized regularized regression algorithm known as the least absolute shrinkage and selection operator.
The most significant metabolite difference (P < 0.0056) was [2-13C]glutamine’s higher final/control ratio for the hypothermia group (1.75 ± 0.12) compared with ratios for the delayed (1.12 ± 0.12) and normothermia group (0.94 ± 0.06), implying a higher pyruvate carboxylase/pyruvate dehydrogenase ratio for glutamine formation. Least Absolute Shrinkage and Selection Operator found the most important metabolites associated with adenosine triphosphate preservation: [3,4-13C]glutamate—produced via pyruvate dehydrogenase entry, [2-13C]taurine—an important osmolyte and antioxidant, and phosphocreatine. Final principal component analyses scores plots suggested separate cluster formation for the hypothermia group, but with insufficient data for statistical significance.
Starting mild hypothermia simultaneously with oxygen–glucose deprivation, compared with delayed starting or no hypothermia, has higher pyruvate carboxylase throughput, suggesting that better glial integrity is one important neuroprotection mechanism of earlier hypothermia.
Mild hypothermia after birth asphyxia improves neurologic outcomes, but the mechanisms involved in this effect remain largely unknown
Spectroscopy in a highly controlled brain oxygen–glucose deprivation slice model using neonatal rats was used during three different mild hypothermia protocols
Starting mild hypothermia simultaneously with oxygen–glucose deprivation compared with delayed or no hypothermia is associated with higher pyruvate carboxylase throughput
This suggests that glial integrity is one key component of the neuroprotective effect of mild hypothermia
RANDOMIZED clinical trials with neurological outcomes have led to mild therapeutic hypothermia (approximately 4°C decrease) becoming the standard of care for early treatment of hypoxic–ischemic encephalopathy from birth asphyxia.1,2 Although it is not fully understood why a brain temperature decrease of only approximately 4°C should cause dramatic outcome differences, mechanisms are known in: physiology—decreased intracranial pressure from reduced brain metabolism; biochemistry—possible activation thresholds for injurious biochemical reactions in a 4°C window; and pathology—reduction in complex processes related to delayed neuron cell death after oxygen restoration.
Nuclear magnetic resonance (NMR) spectroscopy, allows identification of an organism’s total set of metabolites, the metabolome, whose group properties are studied with metabolomics, the science of quantifying and understanding dynamic metabolome responses to physiological changes. Because all chemical reactions are temperature dependent, it is reasonable to ask whether temperature changes of 4°C produce detectable early postasphyxia differences in specific brain metabolites or in metabolomic data sets. If postasphyxia differences are detectable, they might help assess tissue viability, predict subsequent neurologic outcomes, and potentially suggest magnetic resonance spectroscopy approaches to individualizing patient management. This 13C NMR investigation is a follow-up to our earlier 1H NMR metabolomics study with the same neonatal brain slices model, in which asphyxia was also simulated by oxygen–glucose deprivation (OGD). That previous study, which examined differences in 1H metabolite patterns,3 could not study neuron-glia metabolic differences in injury and recovery, because such requires the administration of 13C-labeled substrates that exploit neuron-glia enzyme and pathway differences.
In this and the previous study ex vivo brain slices from 7-day-old (P7) rats underwent 45-min OGD protocols approximating the in vivo Vannucci-Rice asphyxia model.4–6 Slices in three groups, treated identically until the beginning of OGD, were treated after OGD with different temperature protocols. One group was always normothermic (37°C), a second group had 3.75 h of mild hypothermia (32°C), which began with OGD, and a third group had 3.75 h of mild hypothermia, which began after a 15-min delay. Multivariate analyses of extracted brain metabolite changes were quantified with high-resolution NMR spectroscopy. Exploring neuron-glia differences was done by administering an equimolar mixture of two differently labeled substrates, [1-13C]glucose and [1,2-13C]acetate, using an experimental design well developed by others.7–13 Because acetate is metabolized almost exclusively by astrocytes,14–16 13C NMR made it possible in the current study to compare treatment-related changes in glial and neuronal nutrient consumption, and in tricarboxylic acid (TCA) cycle entry via pyruvate dehydrogenase (PDH; EC 22.214.171.124) compared with via pyruvate carboxylase (PC; EC 126.96.36.199), which exists primarily in glia. Because 1H NMR was performed at 21.1 Tesla (900 MHz), separate resonance peaks could be identified in 1H spectra for adenosine triphosphate (ATP), adenosine diphosphate, and phosphocreatine (PCr). The hypotheses tested in this study are: (1) that different hypothermia protocols lead to statistically significant differences in the results of principal component analyses (PCA) and partial least squares discriminant analyses (PLS-DA), where it is assumed that metabolites are coregulated by interdependent phenomena; and (2) that key metabolites (biomarkers) can be identified in a univariate analysis, i.e., one that focuses on individual metabolites and assumes that they are independent, with a regularized regression approach known as L1 penalized lasso (leastabsoluteshrinkage andselectionoperation).
Materials and Methods
Cerebrocortical Slice Preparation and Superfusion with [1-13C]glucose and [1,2-13C]acetate
All protocols adhered strictly to the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee of the University of California, San Francisco. Three groups of experiments had protocols that were identical except for temperatures after OGD. Details of the brain-slice preparation were described previously.17–19 Briefly, in each experiment 20 cerebrocortical slices (350 μm thickness) were obtained from 10 isoflurane-anesthetized 7-day-old Sprague–Dawley rat litter-mates of either sex, and then immediately placed in a superfusion chamber containing fresh, oxygenated artificial cerebrospinal fluid (3 ml/min flow) whose nutrient was 10 mm glucose. At 1 h before beginning OGD, 10 mm glucose in oxygenated artificial cerebrospinal fluid was replaced with 5 mm [1-13C]glucose and 5 mm [1,2-13C]acetate, which continued during OGD and thereafter for 6 h. The superfusion chamber was partially submerged in a water bath having a combined circulator-temperature control that maintained the temperature at either 37°C (normothermic) or 32°C (hypothermic). Other details are the same as in previous studies.
Simultaneous administration of [1-13C]glucose and [1,2-13C]acetate7 was used to assess different changes in neuronal and astrocytic TCA-cycle activity, using the conceptual model of glutamate/glutamine recycling shown in figure 1A.13,20 In this approach synaptic glutamate (glu) is taken up by astrocytes and converted to glutamine (gln) by glutamine synthetase (EC 188.8.131.52), an enzyme not in neurons, and then returned to neurons for glutamate restoration. 13C label changes in glutamate are related to 13C label changes in α-ketoglutarate, which are produced in the TCA cycle. Figures 1B and 1C suggest an assumption of many earlier studies, that 13C-glutamate is in rapid equilibrium with its 13C-α-ketoglutarate counterpart, with equal changes occurring in the 13C enrichment of each. This has been found to be approximately true for the brain, but not be exactly true.21,22 In this article, we assume that in each of the groups the approximation is the same and unchanged during the experiments.
Because, as outlined in figure 1, B and C, neurons and astrocytes have different enzymes and roles in neurotransmitter recycling, and also because of rapid transfer of 13C labels from α-ketoglutarate to glutamate, and from oxaloacetate to aspartate, it is possible to use downstream 13C isotopomer concentrations for the computation of changes in the ratio of acetate/glucose metabolism, and changes in the ratio of flux into the TCA cycle through PC relative to flux through PDH.
Three superfusion experiments were performed for each of the following three groups: (1) normothermia (group N), in which a 37°C temperature was maintained during OGD and a subsequent 6-h recovery period; (2) hypothermia (group H), which had a 3.75-h period of hypothermia (32°C) that began immediately before OGD with a rapid temperature decrease from 37°C. Hypothermia continued throughout OGD and afterward throughout 3 h of recovery with oxygenated artificial cerebrospinal fluid. After hypothermia, 1 h of slow rewarming to 37°C was achieved by increasing the bath temperature by 1°C every 12 min. Superfusion with oxygenated artificial cerebrospinal fluid continued at 37°C for the last 2 h of the experiment; (3) 15-min delayed hypothermia (group D), which also underwent 3.75 h of sudden hypothermia (32°C), a subsequent 1-h rewarming period, and then a 2-h recovery period. However, in this group, group D, the temperature decrease was delayed until 15 min after the start of OGD. In each experiment five slices for NMR spectroscopy were removed, washed, and immediately frozen in liquid nitrogen at each of four predetermined sampling times: before the onset of OGD (T0), at the end of OGD (T1), at the end of hypothermia (T2), and at the end of the experiment (T3).
At later times the five slices removed at each time point collectively underwent perchloric acid extraction of small, soluble molecules into D2O in NMR tubes, which then underwent NMR spectroscopy for quantifications of 35 1H metabolite peaks and 23 13C metabolite peaks. Four additional frozen slices were used for non-NMR assays of ATP and glial fibrillary acidic protein (GFAP) staining, the chief outcome variables. Metabolomic assays of 1H/13C metabolite ensembles for the different sampling times were processed for associations between treatment groups and metabolomics signatures. 13C data were used to look for changes in neuron astrocyte in nutrient consumption, and for changes in TCA cycle entry via PDH and PC (see Supplemental Digital Content 1, http://links.lww.com/ALN/A950, for details of OGD, perchloric acid extraction, the ATP-bioassay, and GFAP staining.)
NMR Data Acquisition and Analysis
NMR data were obtained at the Central California 900 MHz NMR Facility, which operates with National Institutes of Health support (GM68933) in QB3 facilities at University of California, Berkeley, California. A 21.1 Tesla Bruker Avance II NMR spectrometer (Bruker Corporation, Billerica, MA) with a 5 mm CPTXI multinuclear cryoprobe optimized for 13C was used to acquire one-dimensional 13C spectra at 226.2 MHz, one-dimensional 1H spectra at 900 MHz, and two-dimensional (2D)1H J-resolved NMR spectra from which one-dimensional projections were obtained along the 1H axis (one-dimensional 1H pJRES spectra). NMR metabolite resonances corresponding to different chemical shifts were identified and preprocessed with Bruker software (TopSpin 3.1 and AMIX), quantified with iNMR® (Nucleomatica, Molfetta, Italy), corrected for relaxation, and normalized to both the weight of the dry powder and the total spectral area. For additional details regarding NMR acquisitions and metabolite quantifications, especially that of ATP from the 1H resonance on adenine’s imidazole ring, see Supplemental Digital Content 1 (http://links.lww.com/ALN/A950), which contains information regarding NMR Methods.
For each 13C molecular position high-spectral resolution, typically 0.002 ppm, permitted the identification of different 13C isotopomers, i.e., molecules where each carbon can be either 12C or 13C. The term is formed from the words isotope and isomer. Isotopomers for a particular carbon position appear as satellite peaks that are typically 0.030 or more ppm on both sides of the central peak, i.e., the peak present when there is only one 13C contributing to the signal. Figure 2 shows representative 13C NMR peaks and isotopomers detected in our system for glutamate, glutamine, and γ-aminobutyric acid. The reference NMR signal was the upfield resonance from natural abundance 13C in 4,4-dimethyl-4-silapentane-1-sulfonic acid. No other natural abundance 13C signals were large enough to be seen.
Calculations: PC/PDH and Acetate/Glucose
Pathway flux ratio calculations come from 13C position changes during migration through the TCA cycle, as depicted in figure 1, B and C, along with noting that two adjacent 13C atoms can stay together during the migration, as in [1,2-13C]acetate’s metabolism to glutamate and glutamine.
The PC/PDH ratio for glutamate and glutamine production was estimated as (C2 − C3)/C4, and for γ-aminobutyric acid production as (C4 − C3)/C2, formulae established and used in previous studies. Similarly, for both glutamate and glutamine production the ratio of acetate’s contribution to glucose’s contribution was estimated as C45/C4. In these equations, C2, C3, and C4 represent singlet resonance peaks alone, while C45 represents the 51 Hz doublet at C4, as illustrated for glutamate in Figure 2. Further details of the calculation involving 13C metabolites are in Supplemental Digital Content 1 (http://links.lww.com/ALN/A950), which contains all supplemental information regarding NMR Methods.
Statistical and Metabolomic Analyses
For both 1H and 13C data, metabolite quantifications for times T1, T2, and T3 were normalized to initial (control) T0 values. In each of the T1, T2, and T3 data sets three paired t tests were done for each metabolite: Normothermia versus hypothermia, hypothermia versus delayed, and normothermia versus delayed. To have the false-positive error rate (α) be less than 5%, significance for each of the nine t tests was defined by P values at or below a Bonferroni-corrected value, α/9, or 0.0056.
Quantifications of identified metabolites were processed with commercial multivariate software (SIMCA-P+ v.11; Umetrics, Inc., San Jose, CA) in which quantifications for each metabolite are expressed in SD units that vary equally on either side of zero. The new data set for each spectrum was then represented by a single point in a multivariate space whose dimensionality is the number of metabolites. Thereafter, PCA and PLS-DA were performed to look for separate clustering in 2D planes defined by principal component axes. For each time point there are nine points available for a 2D principal component plot, these being the three repeat spectra for each treatment group. In PLS-DA computations the integer added as a variable representing group identification introduces a bias toward group clustering.23
In addition to comparing ensembles of quantified 13C metabolites, the chemometric feature of Bruker program AMIX 3.1 was used to check for whole spectra differences in the 13C range 5–190 ppm. Different sized “bucket” ppm intervals were defined and centered on each metabolite’s peak, except that no buckets were created for glucose and acetate, substrates present in the superfusate. Bucket sizes for each 13C position included isotopomers. Intensities for each bucket were treated as independent X variables. A total of 23 X variables was used for a PCA analysis to compare nine spectra. For 1H data, the spectral range from 0.50 to 9.0 ppm was divided into “buckets” 0.05 ppm wide, with exclusion of the intervals 3.17–4.0 and 4.5–5.0 ppm. A PCA analysis was also performed to compare nine 1H spectra.
Metabolite quantifications were also analyzed using a methodology known as the L1 Penalized Multiple Linear Regression with a “lasso” (least absolute selection and shrinkage operator).24 In this approach linear regression models are determined that fit an outcome variable, which in this case was chosen to be the ATP value found in the luciferin bioluminescence assay done for each NMR tube. This approach, used by us and described previously,25 is known to be well suited for situations where the number of covariates (metabolites) is large compared with the number of observations. Reviewing briefly, the lasso is a regularized regression, meaning that the function being minimized also includes a penalty term, in this case L1, the sum of the absolute values of a regression model’s deviations. The L1 penalty term is weighted by a nonnegative tuning parameter λ that multiplies it. When λ is 0, the penalty function does not contribute. When λ is sufficiently large the penalty term dominates and all the coefficients are forced to zero. The analysis was conducted using the glmnet package26 within the R statistics language.27 Analysis results are shown as plots of (1) cross-validated mean-squared-error versus λ (equivalently, the number of metabolites with nonzero coefficients) and (2) metabolite regression coefficients versus the L1 penalty term. From the former one learns appropriate model sizes (number of included metabolites). From the latter one identifies the corresponding metabolites and sees the extent of their respective contributions as the model size changes.
13C and1H Metabolite Comparisons at the End of Recovery for Different Treatment Groups
Table 1 gives the results of pairwise t testing for the ratios of 13C metabolite quantifications at the last time point (T3) relative to the initial time point (T0). All values are given as mean ± standard error (SE). The Bonferroni-corrected upper limit for P value was 0.0056 for having 5% as the upper limit for type I errors. The most significant metabolite difference was therefore [2-13C]glutamine’s higher final/control ratio for the hypothermia group (1.75 ± 0.12) compared with ratios for the delayed (1.12 ± 0.12) and normothermia group (0.94 ± 0.06), with both P values below the upper limit. Because of the role of glutamine C2 in calculating the ratio of PC flux to PDH flux, the hypothermia group had the highest values of PC/PDH ratio for glutamine. Except for glutamine C2, within all treatment groups changes relative to T0 were the same within experimental error for glutamate C2, C3, and C4, and glutamine C3, and C4. The complete set of comparisons appears in figure 3A. Similar pairwise t testing for 1H data among groups had no significant differences as shown in figure 3B.
Greatly Improved13C and1H NMR Spectral Resolution; Quantification of Metabolites; JRES
Figure 2A shows a representative 13C spectrum for a study’s final time point. Figure 2B shows 13C isotopomer details that cannot be seen easily in figure 2A. Table 2 lists the 13C and 1H resonances that were quantified, and their ppm assignments. Figure 4, A–C compare different ppm regions in two 1H NMR 900 MHz spectra, obtained sequentially from the same NMR tube using different NMR methods, standard one-pulse and pJRES, but without changing static magnetic fields or disturbing the tube. The extract was from T0, the initial time point. The pJRES spectrum substantially reduced ppm overlap of resonances for different metabolites. The PCr resonance in figure 4B is totally resolved. Figure 4C shows downfield 1H resonances, including separate ATP and adenosine diphosphate peaks.28 Details of 1H identification and quantification of ATP are with other NMR details in Supplemental Digital Content 1 (http://links.lww.com/ALN/A950).
Ratios of (the total metabolite NMR signal intensity)/(the 4,4-dimethyl-4-silapentane-1-sulfonic acid reference’s signal intensity)/(weight of dry powder for that spectrum, obtained after lyophilization) varied by less than 16% for all 1H or 13C spectra. As well, amounts of dry powder for all groups and time points varied by less than 13%. Together these indicate that within errors total metabolite pools were the same for all groups at all time points.
Acetate/Glucose Consumption, PC/PDH, Time Course of Metabolite Changes
Figure 5A shows ratios for PC versus PDH activity at the end of 6-h recovery. The figure 5B bar graph shows ratios for the last three time points of acetate consumption relative to glucose consumption for glutamate production. Ratios relative to control ranged from 1.0 to 1.5, with the value at T3 significantly being lowest for the delayed hypothermia group (P = 0.0013). Figure 5, C–H show no significant group or time differences in metabolite ratios relative to the initial time point (T0), except for the increased 13C fractional enrichment of lactate with time, and the hypothermia group’s increased glutamine C2 at all time points.
PCA Scores Plots: Before and after Metabolite Identification
Figure 6A shows an SIMCA-P+ PCA Scores Plot for T1, T2, and T3 for all treatment groups. Each point represents a data set with 58 metabolite quantifications: all 23 13C and all 35 1H quantifications. In the upper left quadrant there is clustering of T1 data (end of OGD, a time when the treatment history was the same for all groups). In Scores Plots “R2” and “Q2” are among the statistical parameters used to assess fits for 2D plane and for the principal component axes, principal component 1 and principal component 2. R2 for the 2D plane, the percentage of the total variance explained by the 2D plot was 0.56. Q2, the average of variance percentages explained by in a cross-validation procedure where the software scrambles data among the groups, was 0.43. The Scores Plot in figure 6B (R2 = 0.48, Q2 = 0.59) was produced by a PLS-DA analysis in which group identification integers (either 1, 2, or 3) were used as the outcome variables (Y variables). This maneuver, which uses outcome measures that are not based on physiological measurements, tightens whatever clustering might be present, but does so with a bias toward pulling each group’s points toward a different quadrant.
Figure 6C shows an unsupervised PCA Scores Plot generated by AMIX from the nine 13C spectra for T3. No metabolite assignments or isotopomer identifications were used in this “chemometric” analysis that compares entire spectral shapes by using as multiple variables the NMR quantifications in ppm intervals known as “buckets.” The ppm intervals for glucose and acetate (used in the superfusate) were excluded from the analysis. Interestingly, the normothermia group’s data clustered separately in the upper left quadrant, whereas the other groups had data points in each quadrant. Figure 6D shows a similar PCA Plot for analysis of the 1H data alone. As in figure 6C, the normothermia group’s points are separated from the others.
L1 Penalized Regression with ATP Bioassay as the Outcome Variable
Figure 7, A and B show 1H and 13C results of the L1 penalized lasso regression analyses, performed with “ATP bioassay” as the outcome variable. Each analysis finds linear models for “ATP bioassay” that are linear functions of NMR metabolites, with the number of metabolites in each model depending on λ, as explained in the Methods. In figure 7, A and B values of log(λ) are indicated along the lower x-axes, whereas numbers of metabolites determined by lasso for models corresponding to λ are indicated along the upper horizontal axis. Y-axis values of plotted points provide mean square errors of models corresponding to λ. Lower mean square errors values indicate better predictive performance. Figure 7A indicates that good fits to 1H data can be accomplished with models featuring between three and seven metabolites. Figure 7B indicates that good fits to 13C data can be accomplished with models featuring between 3 and 10 metabolites. Companion plots inserted in figure 7, A and B, show coefficient trajectories as successive metabolites enter the model and relax the L1 penalty by decreasing λ, causing increases in the L1 norm (x axis). For 1H the first three metabolites that were selected—PCr, N-acetylaspartylglutamate, and taurine—stand out, with PCr driving the predictive model. For 13C the embedded coefficient trajectory plot highlights the two dominant metabolites [3,4-13C]glutamate and [2-13C]taurine.
Quantification of GFAP Immunoreactivity
Figure 8 shows representative areas of immunostaining for GFAP and the results of ImageJ†† (developed by Wayne Rasband; National Institutes of Health, Bethesda, MD) quantifications of GFAP immunoreactivity at initial and final times for the normothermia and hypothermia groups. For the normothermia group the observed four-fold increase in GFAP reactivity was statistically significant (P < 0.0001), but no increase was seen in the hypothermia group.
The most important finding in this 13C NMR study relates to the increase of glutamine C2 (i.e., [2-13C]glutamine) being highest in the hypothermia group at the end of recovery. Increased glutamine C2 was also responsible for the glutamine PC/PDH ratio calculation, (C2 − C3)/C4, being higher by a factor of 2 (fig. 5A) for the hypothermia group. The reason for a C3 subtraction in the numerator comes from increases in glutamine C2, which occur without any flux through PC. Glutamine C2 is synthesized from [1-13C]glucose by astrocytic PC after the first turn of the TCA cycle. However, the PDH pathway in figure 1B shows that if glutamate C4’s precursor, [4-13C]α-ketoglutarate, stays in the TCA cycle for a second turn, it begins the second turn as oxaloacetate, with 13C being equally at either C2 or C3. This increases the C2 of glutamate and glutamine, but always with equal increases in C3. Within experimental errors and independent of rate constants and relaxation times, and for all groups, increases from control were the same for glutamine C3, and glutamate C2, C3, and C4. The only treatment-related difference was the larger increase in the hypothermia group’s glutamine C2. Our finding is consistent with a study that found large decreases in PC flux when oxygen deprivation is severe.29 As well it suggests that better preservation of astrocyte metabolism might be a marker or cause of immediate hypothermia’s treatment advantage. Preserved astrocyte metabolism might have also contributed in our earlier study to the immediate hypothermia group having only half the amount of cell death, as measured in enzyme-linked immunosorbent assays (ELISA).19
The glial cell emphasis in our findings fits in with the current growing emphasis in neuroresearch of astrocyte metabolism’s importance to central nervous system neuronal signaling during physiological and pathophysiological states—including general anesthesia.30–34 Historically, astrocytes have been seen primarily as providers of metabolic support for adjacent neurons and regulators of local extracellular environments. However, it is now appreciated that astrocytes control numerous synaptic functions, that they have metabotropic receptors, secrete neurotransmitters, and participate in the coregulation of interdependent neuron signaling, and that they are affected by ischemia and involved in responses. It is possible that more extensive astrocyte research will produce knowledge essential for a mechanistic understanding of mild therapeutic hypothermia.
The increased glutamine and reduced GFAP staining at T3 in the hypothermia group (fig. 8) is also consistent with experimental findings that GFAP expression in astrocytes correlates inversely with the activity of glutamine synthase, a neuroprotective enzyme that produces glutamine by combining glutamate with ammonia. Decreased glutamine levels in brain tissue were found to be associated with increased GFAP expression after insults,35 which was the case for the normothermia group. Much of the increased GFAP expression described in the literature, occurs days and/or weeks after hypoxic or ischemic injury. However, early GFAP increases have been found starting at 4 h after an in vivo insult to P7 rats36 and an in vitro insult to neuronal–glial embryo cells,37 and being substantial at 12 h,38 and 1 day.39 Of greater relevance, however, is the recent finding in human neonates with hypoxic–ischemic encephalopathy that serum GFAP levels were increased during the first week of life, including during the first 6 h, and on magnetic resonance imaging, were predictive of brain injury.40
Another significant finding was the success of the L1-penalized regression algorithm in using the bioassay quantification of ATP to find important NMR biomarkers. The metabolites found by the algorithm are mechanistically significant as well, with PCr being one indicator of available phosphorylation capacity, taurine being associated with osmotic and antioxidant issues in brain edema, and glutamate production being a marker for the health of TCA-cycle turnover. It is not practical to have very large N in preclinical animal studies where each point in a multivariate data set can cost thousands of dollars. The L1 Penalty algorithm is well suited for situations with smaller N, as evidenced by the very recognizable minima in the mean-squared error plots of figure 7, A and B, which provide clear general examples of this algorithm’s important features. As noted in review articles,41,42 the lasso approach has been very widely used, is becoming more popular, and has recently been generalized from “traditional lasso,” our usage, to “group lasso,” a usage for higher-dimensional multivariate data.
The figure 6A PCA Scores Plot provided some methodological comfort, because the analysis, which included all time points, clearly bunched together all T1 data, which was for slices taken at the end of OGD, a time when there were no treatment differences among the groups. The T3 Scores Plots are remarkable for not having enough data to support PCA and PLS-DA distinctions. R2 and Q2 values were respectable relative to the usual cutoff value for significance, 0.50. However, in the PLS-DA analysis similar Scores Plots were generated by permutation testing that had group labels scrambled randomly among all data points. Nevertheless, the R2 values provided tantalizing thoughts of possible clustering that might occur in larger data sets. The AMIX chemometric PCA Scores Plot for T3, which uses ppm bins as variables, had the normothermia group’s data clustered and separated from data of the other two groups (figs. 6C and 6D). It might be possible to improve the AMIX results by introducing more sophisticated preprocessing with better pH corrections or logarithmic scaling,43 or other methods.44–46 However, this would not circumvent the need for many more spectral data sets.
It is natural to ask how the 1H results in this article compare with 1H results in our earlier study of 2010. For many 1H metabolites it is not possible to make such comparisons, because for every 13C metabolite signal listed in table 2, satellite NMR peaks appear in 1H spectra, often overlapping with neighboring signals, causing quantifications different from those obtained when all carbon nuclei are 12C. 1H nuclei near a 13C nucleus can sense the latter’s dipole magnetic field, which causes mirror-image satellite NMR peaks on either side of the unperturbed 1H peak. This can be seen easily for lactate and alanine in figure 4A. As well, nutrients in this study included acetate, thereby putting heavier emphasis on glial metabolism. Although comparisons with earlier 1H quantifications were not possible, 1H quantifications were useful in assessing spectral differences among the treatment groups, in a way similar to chemometric analyses of spectra, which analyze quantifications of spectral regions without knowledge of metabolite identities contributing in those regions. The AMIX PCA (fig. 6D) plot for the 1H data that coexisted with the 13C data provided a confirmation of the 13C AMIX Scores Plot (fig. 6C). In both plots all three points of the normothermia group were separated from the data points of other groups.
Our range for acetate–glucose ratios is comparable with that found for rat cortex in a study of middle cerebral artery occlusion, using the same 13C acetate–glucose substrates.9 Compared with our 0.9–1.5 range, the acetate–glucose ratio for glutamate production in ipsilateral (ischemic) cortex was (2.32 ± 0.93) times the value for contralateral (nonischemic) cortex.9 Changes in acetate–glucose ratios can suggest changes in metabolite traffic, such as the transfer of neuron-produced glutamate to astrocytes for conversion to glutamine, and the return glutamine to neurons for deamination to glutamate, as noted in earlier studies.7 ,16
An assumption mentioned earlier was that there is no change in the proportion of 13C-α-ketoglutarate that is converted to 13C-glutamate. If the kinetics of that conversion were different for the hypothermia group so as to increase glutamine C2, one would expect to have a similar increase in glutamate C2. Figure 5C, however, shows that levels of glutamate C2 were the same in all groups.
This study was limited to cerebrocortical tissue. However, post-OGD injury in the cerebral cortex has been found to demonstrate vulnerability similar to that of the hippocampus in a P7 rat OGD study that studied full coronal brain-slice sections, and also found pathophysiology similar to that of neonatal hypoxic–ischemic encephalopathy being suggested by various biomarkers, including tumor necrosis factor-α, lactate dehydrogenase, and inducible nitric oxide synthase.47 We also appreciate that neuron–astrocyte metabolic interactions have more complexities than we have discussed, especially when, unlike the situation in brain slices, in vivo neurons use substantial aerobic metabolism to maintain the high ATP levels required by extensive neuron activation, and in vivo astrocytes obtain most of their ATP from glycolysis while providing neurons with substantial amounts of lactate for their energy metabolism.
In summary, we used state-of-the-art spectroscopy in a highly controlled brain-slice model to test two hypotheses about finding significant metabolic differences among three different mild hypothermia protocols. We did not validate the first hypothesis, that multivariate data sets could distinguish treatment and outcome groups, due to the combined circumstances of too few data and the putative differences being much smaller than those occurring with severe hypoxia, whose data sets could be distinguished. The second hypothesis, that individual biomarkers could be identified, was strongly validated for one biomarker, glutamine produced from glucose via TCA-cycle entry through PC, an enzyme unique to glia. Also with regard to individual metabolites, the L1 Penalized Regression analysis (lasso), using bioassay ATP as an outcome variable, was very effective in identifying small sets of metabolites appropriate to ATP preservation. We expect that the findings and methodological advances will be useful in future investigations.
Gratefully acknowledged are very helpful and inspiring discussions with Ulrich Günther, Ph.D. (Professor of Biophysical Chemistry, School of Cancer Sciences, College of Medical and Dental Sciences), Mark Viant Ph.D. (Professor of Metabolomics, School of Biosciences), and Christian Ludwig, Ph.D. (Scientific Officer for NMR), along with their hospitality during visits to the Henry Wellcome Center for Biomolecular Spectroscopy (University of Birmingham, Edgbaston, Birmingham, United Kingdom).
Available at: http://rsb.info.nih.gov/ij. Accessed May 15, 2013.