|Home | About | Journals | Submit | Contact Us | Français|
Human clinical trials using 72 hours of mild hypothermia (32°C–34°C) after neonatal asphyxia have found substantially improved neurologic outcomes. As temperature changes differently modulate numerous metabolite fluxes and concentrations, we hypothesized that 1H/31P nuclear magnetic resonance (NMR) spectroscopy of intracellular metabolites can distinguish different insults, treatments, and recovery stages. Three groups of superfused neonatal rat brain slices underwent 45minutes oxygen–glucose deprivation (OGD) and then were: treated for 3hours with mild hypothermia (32°C) that began with OGD, or similarly treated with hypothermia after a 15-minute delay, or not treated (normothermic control group, 37°C). Hypothermia was followed by 3hours of normothermic recovery. Slices collected at different predetermined times were processed, respectively, for 14.1Tesla NMR analysis, enzyme-linked immunosorbent assay (ELISA) cell-death quantification, and superoxide production. Forty-nine NMR-observable metabolites underwent a multivariate analysis. Separated clustering in scores plots was found for treatment and outcome groups. Final ATP (adenosine triphosphate) levels, severely decreased at normothermia, were restored equally by immediate and delayed hypothermia. Cell death was decreased by immediate hypothermia, but was equally substantially greater with normothermia and delayed hypothermia. Potentially important biomarkers in the 1H spectra included PCr-1H (phosphocreatine in the 1H spectrum), ATP-1H (adenosine triphosphate in the 1H spectrum), and ADP-1H (adenosine diphosphate in the 1H spectrum). The findings suggest a potential role for metabolomic monitoring during therapeutic hypothermia.
Two large randomized clinical trials recently found that after neonatal asphyxia, 72hours of mild brain hypothermia (33°C to 34°C), begun no more than 6hours after birth, improves long-term neurologic outcome. In the CoolCap study (Gluckman et al, 2005; Gunn et al, 1998), the scalp and underlying tissues were cooled, whereas in a larger clinical trial (Shankaran et al, 2005), the whole body was cooled. This together with other evidence in human newborns of improved neurologic outcomes (Compagnoni et al, 2002, 2008) has initiated a ‘surge' in mechanism-oriented hypothermia research projects. We believe that important issues related to optimal cooling temperatures, durations, rewarming regimens, and mechanisms—and to the individualization of therapy—would be better addressed if there were a reliable way to identify and assess helpful and harmful intracellular events during the many hours of a hypothermia treatment.
We also believe that magnetic resonance spectroscopy (or nuclear magnetic resonance (NMR) spectroscopy) has enormous potential for ultimately guiding therapeutic hypothermia in humans. One of several NMR approaches involves metabolic profiling, or NMR metabolomics (sometimes referred to as metabonomics), in which ensembles of NMR metabolites, 49 in this study, are subjected to multivariate analyses. As hypothermia differently modulates the fluxes and concentrations of numerous metabolites, it is reasonable to ask whether substantial physiological changes that come from a 4°C temperature change are associated with impressive metabolic variations and changes in the characteristics of specific metabolite ensembles. Such might occur if important chemical activation energy thresholds undergo large changes in the 4°C temperature range, as suggested by the Arrhenius equation (as explained in the Appendix).
Methodological advances have already made it possible to accurately obtain in vivo 14.1Tesla NMR quantifications in mice of 19 1H metabolites (Lei et al, 2010; Mlynarik et al, 2008) and also to use such methods to perform ischemia studies (Berthet et al, 2009). Although safety and other issues augment the challenges of doing high field, in vivo NMR studies of neonates and infants (De Vita et al, 2006; Dagia and Ditchfield, 2008), one can still have the hope that such will ultimately be available for individualizing treatments with therapeutic hypothermia.
As an early step toward in vivo goals, we have performed ex vivo experiments using a neonatal (P7) rodent brain slice model, where insults, treatments, and outcomes can be clearly defined and regulated. Asphyxia was modeled by oxygen–glucose deprivation (OGD), and two hypothermia protocols were studied, along with a normothermia control. Tissues were removed during and after hypothermia, and high-resolution 14.1Tesla 1H/31P NMR spectroscopy was used to determine changes in perchloric acid extracted metabolite concentrations. We asked whether NMR metabolic profiles of metabolite ensembles could accurately differentiate treatments, and good from bad outcomes, these being defined via quantifications of cell death and ATP preservation, with superoxide production helping to provide assessments of OGD insults.
All animal experiments were approved by UCSF's Institutional Animal Care and Use Committee. Slice preparation protocols were as described previously (Espanol et al, 1992; Zeng et al, 2007; Liu et al, 2008, 2009), with 20 350-μm cerebrocortical slices coming from 10 anesthetized 7-day-old Sprague-Dawley rat littermates of either sex. Excised slices were placed immediately in a superfusion chamber containing fresh, oxygenated artificial cerebrospinal fluid (oxy-ACSF) that consisted of a modified Krebs balanced salt solution: 124mmol/L NaCl, 5mmol/L KCl, 1.2mmol/L KH2PO4, 1.2mmol/L MgSO4, 1.2mmol/L CaCl2, 26mmol/L NaHCO3, and 10mmol/L glucose. Continuous bubbling of oxy-ACSF with carbogen (95% O2, 5% CO2) maintained constant values of PO2 (more than ≈600mmHg), PCO2 (40mmHg), and pH (7.4). The oxy-ACSF flow rate was 3mL/min. The chamber was partially submerged in a water bath having a combined circulator-temperature controller (PolyScience Standard Immersion Circulator EW-12101-84, Cole-Parmer Instruments Inc., Vernon Hills, IL, USA; http://www.cole-parmer.com) that maintained temperatures at either 37°C (normothermic) or 32°C (hypothermic).
Oxygen–glucose deprivation was initiated by replacing the post-decapitation oxy-ACSF with glucose-free ACSF that had been deoxygenated by equilibration with an anaerobic gas mixture (95% N2, 5% CO2) at the required temperature. During OGD, the oxygen content of the superfusate was always <2mmHg, as determined by a fiber-optic system (FOXY-R probe and SF2000 Spectrofluorometer, Ocean Optics Inc., Dunedin, FL, USA; http://www.oceanoptics.com). The OGD period of 45minutes was followed by 6hours of superfusion with oxy-ACSF containing 4mmol/L glucose.
Four experiments with 45-minute OGD periods were performed for each of the three groups: (1) Normothermia (Group N), in which 37°C was maintained during OGD and a subsequent 6-hour recovery period. (2) Hypothermia (Group H), in which slices underwent a rapid temperature change from 37°C to 32°C (hypothermia) simultaneously with the onset of OGD. Hypothermia continued throughout OGD (45minutes) and afterwards throughout 3hours of recovery with oxy-ACSF. Following hypothermia was 1hour of slow rewarming to 37°C, achieved by increasing the bath temperature 0.1°C per 1 to 1.5minutes. Superfusion with oxy-ACSF continued at 37°C for the last 2hours of the experiment. (3) Delayed hypothermia (Group D), where the group underwent the same cooling duration and hypothermic treatment as Group H, except for having the onset of hypothermia begin 15minutes after the start of OGD. For all groups, five or six slices were removed from the superfusion chamber for NMR or histologic analysis at each of the four time points: 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). Figure 1 shows a schematic diagram of the protocol.
Perchloric acid extraction and NMR tube loading were performed as described in earlier studies (Espanol et al, 1992; Liu et al, 2008, 2009). In brief, frozen slices from each time point were ground under liquid nitrogen in a precooled mortar and then extracted using 30mL/g (dry mass) of ice-cold 12% perchloric acid. The extracts were neutralized to pH 7.0 with KOH and lyophilized for 24hours (BenchTop 2K lyophilizer, Virtis, Gardiner, NY, USA). Before NMR analysis, lyophilized dry extracts were weighted and stored at −80°C and prepared for 1H spectroscopy by resuspension in 430μL of D2O containing 0.02% 4,4-dimethyl-4-silapentane-1-sulfonic acid, a reference compound for NMR determinations of chemical shifts and signal intensities. Samples were neutralized to pD 7.0 to 7.4, with DCl before being loading into 5mm NMR tubes (Advanced Microtube CMS-005TB matched to D2O, Shigemi, Inc., Allison Park, PA, USA). After completing 1H NMR spectroscopy, NMR tube contents underwent the addition of EDTA and methylene diphosphonate (0.4mmol/L). The former chelates cations that cause NMR line broadening, whereas the latter served as a 31P NMR reference. The pD was readjusted to 7.0 to 7.4 with NaOD before acquiring proton-decoupled 31P NMR spectra.
One-dimensional NMR studies of extracts were performed in the UCSF Magnetic Resonance Laboratory using a 14.1-Tesla (600MHz) Varian UNITY spectrometer, with an INOVA console and a customized, multinuclear Z-SPECT radiofrequency probe that was optimized for this project (3NG600-8, Nalorac Division of Varian, Martinez, CA, USA). Basic one-pulse RF sequences, as described previously (Liu et al, 2009), were used for obtaining 1H spectra at 599.92MHz and 31P spectra at 242.86MHz. The extra homogeneity provided by 5mm Shigemi NMR tubes and the Varian INOVA shimming software typically resulted in spectral linewidths <0.0025p.p.m. full width at half maximum (1.5Hz full width at half maximum) for 1H water protons and 0.01p.p.m. full width at half maximum for 31P in phosphocreatine. Proton spectra were composed from 64 transients, with a 4-second acquisition time and a 1-second interpulse delay. A selective saturation pulse was used to suppress the water peak. Metabolites corresponding to different chemical shifts were identified and quantified with Chenomx NMR Suite Software (version 6.0, Chenomx Inc., Edmonton, AB, Canada) using targeted profiling (Weljie et al, 2006) and their spectral library for 600MHz, whose pulsing parameters were used when acquiring our spectra. For each metabolite, the targeted profiling approach involves manual simultaneous adjustment of all resonance peak clusters for that metabolite. For example, quantification of the 1H doublet resonance peak at 1.33p.p.m. (often referred to simply as ‘lactate') required manipulation of clusters for both lactate and threonine, as the latter overlaps with lactate at that chemical shift. This involved manually fitting portions of the observed spectrum by coordinating simultaneous adjustments of both lactate peaks (the doublet at 1.33p.p.m. and the quartet at 4.12p.p.m., with simultaneous adjustments of all three threonine peaks (the doublet at 1.33p.p.m., the doublet at 3.58p.p.m., and the multiplet at 4.26p.p.m.). All chemical shift assignments for peaks and clusters were then checked against chemical shift assignments published in careful studies (Govindaraju et al, 2000; Willker et al, 1996), and also listed in computer databases at the University of Wisconsin's Biological Magnetic Resonance Data Bank (Standard Compounds; http://www.bmrb.wisc.edu/metabolomics), and at the National Institute of Advanced Industrial Science and Technology (http://www.aist.go.jp/RIODB/SDBS) in Japan. 31P and 1H relaxation time corrections were not introduced, as the same NMR pulse sequence was used for all runs of the same nucleus, and the goal of the analysis was primarily to detect large differences in each metabolite from its control value.
Metabolite signal intensities were computed relative to methylene diphosphonate or 4,4-dimethyl-4-silapentane-1-sulfonic acid, and then normalized to the weight of the dry powder that was mixed with D2O before NMR measurements. 1H NMR spectral analysis provided quantifications of 39 metabolites: 4-aminobutyrate, ADP-1H, ATP-1H, acetate, adenine, alanine, arginine, aspartate, choline, citrate, creatine, phosphocreatine, formate, fumarate, glucose, glutamate, glutamine, glutathione, glycine, inosine, isoleucine, lactate, leucine, malate, N-acetylaspartate, N-acetylglutamate, NADP+, NADPH, phosphocholine, phosphoethanolamine, phenylalanine, phthalate, succinate, taurine, threonine, tryptophan, tyrosine, valine, and myo-inositol. Figures 2A and 2B show a typical 1H NMR spectrum for extracts obtained from slices taken at T0 (just before starting OGD). The insert with spectral expansion illustrates the system's outstanding 1H NMR spectral resolution by showing that the methyl protons in creatine are resolved from the methyl protons in phosphocreatine. 31P NMR spectra that were analyzed similarly, using published chemical shifts (Pettegrew et al, 1993), provided additional metabolite measures: ATP, ADP, phosphocreatine, phosphoethanolamine, glycerophosphoethanolamine, phosphorylcholine, and glycerophosphocholine. Overlapping 31P resonances were quantified by deconvolution with iNMRTM software (Mestrelab Research, Santiago de Compostela, Spain, www.inmr.net.). Figures 2C–2F display typical 31P NMR spectral peaks, also from slices taken at T0.
Hydroethidine (HEt; D1168, Invitrogen, Carlsbad, CA, USA) is freely permeable to living cells and selectively oxidized by the superoxide anion to red fluorescing ethidium (Et). The protocol followed the methods published in an earlier study (Liu et al, 2008). In brief, a parallel superfusion channel with an independent ACSF reservoir containing 5μmol/L HEt was established for this procedure. The channel was identical to that for the original superfusion chamber, except for there being HEt in the superfusate. Thirty minutes before selected time points during perfusion of experiments, three slices were transferred from the original superfusion chamber into the parallel chamber for superfusion with ACSF containing HEt. After 30minutes of incubation, slices were removed and fixed overnight in freshly prepared 4% formaldehyde at 4°C. On the following day, the slices were dehydrated in 30% sucrose, and then washed twice in PBS and immediately embedded in O.C.T. (Sakura Finetek Optimum Cutting Temperature compound, OpticsPlanet Inc., Northbrook, IL, USA), and cut into 10-μm-thick sections on a cryostat (Leica, CM1900, Solms, Germany). Mounting medium containing 1.5μg/mL 4′,6′-diamidino-2-phenylindole (H1200; Vector Laboratories, Burlingame, CA, USA) was applied onto sections to counterstain cell nuclei. Superoxide anion production was assessed in a Leica SP2 Laser scanning confocal microscope at UCSF Diabetes Center's Facility for Microscopy and Cellular Imaging, with excitation wavelengths being 510 to 550nm and emission wavelengths >580nm for Et fluorescence. There was no wavelength overlap for the imaging of 4′,6′-diamidino-2-phenylindole, whose excitation and emission wavelengths are, respectively, 350 and 470nm when bound to DNA. Quantifications of Et-positive cells were performed at five random regions between the pial layer and the injury layer of brain slices. These regions were photographed under a × 63 oil immersion objective, and cell counts included neurons and glia, these not being histologically distinguishable. Counts of Et-positive cells are given as percentages of total cells counted.
Evidence for apoptosis in slices taken at different time points was examined both with fluorescence microscopy of in situ terminal deoxynucleotidyl transferase-mediated uridine 5′-triphosphate-biotin nick end labeling (TUNEL) staining and with quantification using a sensitive, commercial, enzyme-linked immunosorbent assay (ELISA) (Cell Death Detection ELISAPLUS, Roche Applied Science, Indianapolis, IN, USA).
For TUNEL staining, sections were first permeabilized in 0.1% Triton X-100 (Sigma Chemical Co, St Louis, MO, USA) in PBS for 8minutes. The TUNEL reaction mixture was obtained by adding terminal deoxynucleotidyl transferase to the nucleotide mixture according to the instructions in the manufacturer's manual (Roche Diagnostics, Mannheim, Germany). Each section was then incubated with 50μL TUNEL reaction mixture in a humidified, dark chamber at 37°C for 60minutes. After rinsing with PBS, sections were counterstained with 0.5mg/mL propidium iodide to highlight nuclei. The fluorescein wavelength for excitation was 488nm, with emissions being detected above 515nm. Propidium iodide excitation was at wavelengths <535nm, with emission being detected at ~615nm.
The ELISA assay, which was performed according to the manufacturer's instructions, detects apoptotic but not necrotic cell death (Leist et al, 1998) via antibody identification of histone-associated DNA fragments. A quantitative sandwich enzyme immunoassay is performed using mouse monoclonal antibodies directed against DNA and histones. Nucleosomes, which package DNA within chromosomes, are composed of large numbers of linked histones, the proteins onto which DNA is wrapped. During apoptosis, endonuclease cleavage of double-stranded DNA occurs at the most accessible locations, the internucleosomal linker regions, thereby generating mono- and oligonucleosomes having DNA with multiples of 180 basepairs. (As the DNA within nucleosomes is tightly complexed to the histones, it is protected from cleavage during apoptosis.) The quantification by this assay is thus analogous to the quantification of ‘ladder patterns' via DNA electrophoresis.
Briefly, the manufacturer's protocol involved the following: one snap-frozen cerebrocortical slice was homogenized in lysis buffer, incubated for 30minutes, and then centrifuged. The supernatant, which contained apoptotic-related DNA fragments released from the cytoplasm, was transferred into the well of a streptavidin-coated microplate for 2hours of incubation at room temperature. Within the well, an anti-histone antibody and an anti-DNA antibody were used to form and immobilize antibody–nucleosome complexes. After washing and adding peroxidase substrate in the well, the anti-nucleosome complexes were quantified in a spectrophotometer with absorbance at 405nm. To account for variations of slice size and cell numbers, spectrophotometer quantifications for each slice were normalized to the total protein concentration in that cell's lysate. The assay was performed three times for each particular time point, using three different slices, for all time points (T0, T1, T2, and T3) of each treated group.
The NMR metabolite quantifications for all particular time points for all experimental groups underwent within-group and between-group statistical analysis. Different multivariate approaches were used, as further explained below, including: Fisher's protective least squares difference test for multiple comparisons, principal component analyses (PCA), projection to least squares discriminant analyses (PLS-DA), and the L1-penalized regression algorithm (Efron et al, 2004; Tibshirani, 1996).
For each time point, pairwise comparison testing for differences in individual metabolites and cell-death quantifications was performed between treatment groups. Metabolites for the last three time points were normalized to initial values, these being taken as the control in each experiment. Thus, there were nine tests for each metabolite, and also for cell-death quantifications: N (normothermia) versus D (delayed hypothermia), N versus H (hypothermia), and H versus D, that is, three comparisons for each of the times T1, T2, and T3.
SIMCA P+ v.11 software (Umetrics, Inc., San Jose, CA, USA) was used for determining PCA and PLS-DA computations. Briefly, as described in Umetrics publications (Eriksson et al, 2006), the SIMCA P+ approach begins by mathematically creating an N-dimensional space, where N is the number of independent variables in the metabolite ensembles being compared. For comparisons of each time point's ensemble of 1H NMR metabolites, N is 39. When NMR data for a single time point also includes metabolites determined in 31P spectra, N becomes 49. Thus, each set of 39 or 49 numbers for a particular slice-removal time is represented by a single point in a 39- or 49-dimensional space. In the PCA, all metabolite data undergoes centering and unit variance scaling, after which the plane is found whose two orthogonal axes have maximum cluster separations when data points are projected onto it. Two-dimensional scores plots are used to display the clusters in planes where the x and y axes are principal components. In the PLS-DA, an algorithm sharpens the PCA clusters and their separations by creating a new Y variable for each data class (Eriksson et al, 2006). The scores plots in the PCA and PLS-DA analyses have corresponding loadings plots that identify metabolites that contributed most to the separations that appear in the scores plot. In the scores plot, each spectral data set is plotted as a point, which in turn defines a vector from the origin. Each such vector, however, is the vector sum of 39 or 49 smaller vectors defined by the metabolites. Projections onto the plane of the principal axes of those smaller metabolite vectors produces the loadings plot or weight plot. Although any data collection can be put through a PLS-DA analysis, the usual use in this type of study is to focus data sets that best compare outcomes. Thus, we will focus on the PLS-DA analysis for one data set, that for metabolites at the last time point (end of the experiment), with each metabolite value expressed as its ratio to the starting value (after normalization of each slice ensemble to its dry weight).
After Umetrics software processing, NMR-detected metabolites were analyzed independently with L1-penalized regression (Goeman, 2010; Tibshirani, 1996), a multivariate algorithm in which one identifies a priori one or more ‘outcome' measures, and then, in a regression analysis, minimizes L1, the sum of absolute values of the residuals. All data for all time points are used, with identification of metabolite subsets that have the strongest influence on the ‘outcome' variable(s). This was performed twice, first with there being only one outcome variable, βATP, and then again with the only outcome variable chosen to be the cell-death measure quantified by the ELISA analysis. When taking βATP as the outcome variable, the analysis did not include 31P-based metabolites. The motivation for using an L1 penalty (minimizing absolute deviations) in estimating regression coefficients is, in biostatistics terminology, that this specification simultaneously imparts a shrinkage (of the size of the set of final variables) that results both in improved predictive performance, and in a covariate selection that has improved interpretability.
The geometric underpinnings for this algorithm's behavior are detailed in Tibshirani (1996), who terms the procedure LASSO: Least Absolute Shrinkage and Selection Operator. However, the original LASSO algorithm can breakdown when the number of covariates (metabolites) exceeds the number of samples, as here. In such settings, the recently devised, and highly efficient, LARS (Least Angle Regression) algorithm (Efron et al, 2004) can be applied. Indeed, the inherent instabilities associated with small sample sizes and relatively many predictors, often correlated—as in our data sets—accentuates the importance of such shrinkage and selection. In a LARS analysis the tradeoff between model fit, which for continuous outcomes such as βATP or the ELISA quantifications is measured via residual variance, and model complexity, arising here from the number of selected metabolites, is governed by a tuning parameter that weights the contribution of the L1 penalty. Determination of the value of the tuning parameter makes recourse to cross-validation (Hastie et al, 2010). Briefly, the tuning parameter value that achieves optimal fit is selected from evaluations over unseen data furnished by the cross-validation data partitioning approach. The choice of the tuning parameter determines which and how many metabolites to retained.
The vertical bar plot in Figure 3A displays ATP, ADP, and PCr values for slices taken at the three experimental time points (end of OGD, end of hypothermia, and end of experiment) measured relative to initial, pre-insult values (first time point). It shows that at the end of OGD, the time of maximal metabolic stress, ATP, ADP, and PCr decreased, respectively, to ~15%, 40%, and 5% of control. Figure 3B shows that at the end of the recovery period, that is, at the conclusion of the experiment, some 6hours after ending OGD, values of ATP, ADP, and PCr were significantly lower in the normothermia group than in the hypothermia and delayed hypothermia groups, each of which exhibited recovery to control levels. As well, Figure 3B shows that levels of membrane phospholipids PE, GPE, PC, and GPC had similar behavior. However, it should be appreciated that when metabolite values decrease by 40%, as was the case for the final ATP for the normothermia group, NMR spectroscopy cannot distinguish between a uniform 40% loss among all cells, and the situation where 40% of the cells lost all ATP while 60% lost none.
Similar vertical bar plots for 1H metabolites, shown only for the end of the recovery period, are displayed in Figure 3C for 39 metabolites. As in the 31P NMR analysis, the level of ATP and PCr decreased substantially in the normothermia group. In addition, compared with their values in the normothermia group, NAA, glutamate, and taurine increased significantly in both the hypothermia and delayed hypothermia groups. Metabolites that were suppressed in hypothermia and delayed hypothermia groups included acetate, adenine, alanine, choline, isoleucine, lactate, leucine, tyrosine, and valine (inosine and arginine decreased only in the hypothermia group).
Figure 4A shows the PLS-DA scores plot (Eriksson et al, 2006) produced for the final time point by the SIMCA P+ multivariate analysis, which mapped each NMR spectrum or data set into a single point in a multidimensional space, and then looked for planes where the points formed distinct clusters. The data set that resulted in Figure 4A consisted of the ratio of final to initial NMR intensities for each metabolite, including both 1H and 31P metabolites. The scores plot clearly shows a tendency for separate clustering by the three treatment groups: normothermia (blue), hypothermia that starts with OGD (black), and hypothermia that starts after a 15-minute delay (red). The x and y axes of the plot are the two ‘principal component' axes. Figure 4B shows the loadings plot (Eriksson et al, 2006) that corresponds to the scores plot. Metabolites in each quadrant of the loadings plot correspond to points in the same quadrant of the scores plot. Metabolites that are close together on the loadings plot are highly correlated, and metabolites furthest from the origin contributed most to the furthest points in the scores plot. For example, lactate and alanine are close to each other in the lower left quadrant of the plot, as are glycerophosphocholine and glycerophosphoethanolamine, and ATP and ADP in other quadrants, and each of these metabolites was influential in the scores plot.
The TUNEL-positive cells, which contain DNA fragmentation, were too few for cell counting. Figure 5 illustrates the help provided by costaining with propidium iodide in discerning when morphology in TUNEL-positive cells was suggestive of apoptosis (condensed, fragmented nuclei) and necrosis (diffusely stained nuclei). The ELISA DNA fragmentation quantifications, also in Figure 5, show that both at the end of the experiment and also at the time when hypothermia had been completed, cell death was significantly lower in the hypothermia group compared with the normothermia and delayed hypothermia groups, which had statistically similar values. In contrast, at the end of OGD, the normothermia group had significantly greater death than either of the two lower temperature groups, which had statistically similar values.
Cross-validation prediction error (CVPE) profiles from applying L1-penalized regression to βATP and apoptotic cell death are given in Figures 6A and 6B, respectively. (This analysis did not include 31P-based metabolites.) As noted earlier, in this analysis, data are pooled from each of the four repetitions of three groups, each with three time points. Thus, the sample size had 36 data sets, which is close to the number of NMR-determined metabolites. The number of variables chosen for calculating CVPE is plotted along the x axis as ‘the fraction,' which is the ratio of the number of chosen variables to the total number of available variables (39 1H metabolites). The y axis gives the CVPE obtained when the number of independent variables is that indicated by the corresponding value on the x axis. Thus, the ELISA outcome results are best fit when the NMR data set is taken to have six variables (x=0.15, with 39 variables), whereas the βATP outcome results suggest that as many as 12 variables are needed for a best fit. Both profiles are prototypic in that they exhibit initial decreases in CVPE as the value of ‘fraction' starts to increase from 0 as important metabolites are entered; and then show almost no CVPE changes with further increases in ‘fraction,' indicating that several statistical models attain approximately equal and minimal CVPE; and finally show ever increasing values of CVPE as noise variables are included. Several heuristics have been advanced for selecting a specific model based on such profiles. These include taking the number of metabolites to be that associated with minimum CVPE, or by finding a smaller model (fewer metabolites) by taking the smallest number of metabolites associated with a CVPE value that is within one standard error of the minimum CVPE. With βATP as the outcome variable, the top five metabolites were PCr (as seen in 1H spectra), ATP-1H, choline, formate, and phosphoethanolamine. The next five metabolites that would be allowed by going lower in ‘fraction' are malate, fumarate, ADP-1H, threonine, and glutathione. With the ELISA measure of cell death as the outcome variable, the top five metabolites were: ADP-1H, valine, acetate, tryptophan, and formate, with these followed by phosphocreatine and leucine. In the analysis where cell death was the outcome variable, the inclusion of 31P-based metabolites did not change the results.
Figure 7A shows selective, high-magnification confocal fluorescence imaging of superoxide, detected from its oxidation of HEt to Et (red fluorescence). Cross-sections 0.1μm thick are shown through central slice regions, with cell nuclei appearing blue from 4′,6′-diamidino-2-phenylindole staining. Figure 7B plots superoxide quantifications for the hypothermia and normothermia groups, obtained for four time points as described in the Materials and methods. Superoxide, which barely existed in control slices, became incrementally greater with time, and always more so in the normothermia group.
Our brain slice preparation's advantages for modeling neonatal asphyxia include: that its neuron-glia architecture is close to that found in vivo, that the arrangement very conveniently accommodates ultra rapid freeze clamping of metabolites and cellular processes, and that it is easy to have tight control of PO2, PCO2, pH, temperature, and superfusate composition. We chose a hypoxia protocol with continuous superfusate flow, as continuous brain tissue perfusion is a key feature of the widely used Vannucci–Rice in vivo model of neonatal asphyxia (Vannucci et al, 1999; Vannucci and Vannucci, 2005). Glucose is well known to provide enormous anoxic protection to rat astrocytes during early development. In one astrocyte, cell culture study using 15mmol/L glucose, 24hours of total oxygen deprivation produced zero cell death (Callahan et al, 1990). Because of this and preclinical rodent studies that found decreases in both brain and plasma glucose during asphyxia (Basu et al, 2009; Holowach-Thurston et al, 1974), we felt that glucose deprivation was appropriate in our model. A disadvantage of acute brain slice preparations is that treatment times and post-insult recovery times are short compared with those for in vivo studies, where multiple mechanisms contribute to cell death. If, however, early, distinct metabolomic differences are seen clearly, it will be reasonable to look for such differences in vivo.
Knowing that almost all chemical reactions are affected by temperature changes, we felt that a multivariate metabolomic approach with simultaneous monitoring of 30 to 40 metabolite concentrations would lead to group parameters that distinguished metabolite ensembles associated with less injury and better outcomes. Despite our having far fewer data points per treatment group than used in traditional multivariate studies, the scores plots of the PLS-DA showed clear separations, suggesting that multivariate characterizations are potentially useful. This phenomenon has been demonstrated previously in earlier works by us and others (Liu et al, 2009; Serkova and Niemann, 2006a; Serkova et al, 2006b; Jordan et al, 2009). Naturally, one expects that future accumulations of many more data points would lead to overlapping distributions, despite the cluster shrinkage and higher resolution of cluster separations provided by going from the PCA to a PLS-DA analysis. The separation of treatment groups on a metabolomics PLS-DA plot, however, is not the final step.
Although particular treatment groups might form distinguishable clusters on scores plots that possibly end up being consistently associated with better or worse biological outcomes, an analysis is needed that starts with outcome values and assesses the extent to which they were optimized by all or some of the biomarkers. The L1-penalized regression analysis, which for reasons already given was used instead of a PLS-DA or partial least squares regression (Eriksson et al, 2004; Wold et al, 2001), produced impressive differences with the two outcome variables that we chose: early cell death as measured by an ELISA, and ATP preservation, as measured by 31P NMR spectroscopy (while excluding all other 31P NMR data). Although the use of the ELISA assay as an outcome variable worked well in this particular study, interpretations of ELISA quantifications in this study is limited by a lack of calibration. The ELISA changes are measured relative to baseline values, not relative to the maximal apoptotic stimulus. Thus, if overall cell death is small at baseline, for example 10%, a 40% increase in ELISA represents only a 4% increase in overall cell death.
It is easy to conclude that taking ATP preservation as the sole outcome variable would not be very helpful in our model, as the two hypothermia treatment groups totally restored their initial high-energy phosphates while having large differences in early cell death. This is also apparent from the L1-penalized regression, which found that 12 metabolites of 39 1H metabolites were very important to ATP optimization. In contrast, early cell death had significant differences between the two hypothermia treatment groups with only six key metabolites, with two being 1H-detected high-energy phosphates. The data support the notion that early ATP recovery does not imply that all is well at cellular levels. Although ATP preservation after 6hours of recovery, the time when our experiments finished, might not be very helpful as an outcome variable, in experiments that go out to much later times it might nevertheless have associations more meaningful than those of ELISA measurements, which are based on caspase-initiated endonuclease cleavage at internucleosomal linker regions (presumably before ATP depletion). After 6hours of recovery, ELISA measurements might account for only a fraction of the dying neurons. After OGD, eventual ATP depletion from necrosis and caspase-independent pathways can be more prominent at later times, as noted in two earlier studies (Bonfoco et al, 1995; Kalda et al, 1998).
From a metabolomic perspective, the data suggest that the detection of ‘downfield' 1H metabolites (1H resonances to the left of the H2O peak) can add much to metabolic monitoring and predicting. Such, of course, is very easy in extract spectra, and much more challenging in vivo. Nevertheless, data have appeared that show in vivo detection with current techniques is within reach (Vermathen et al, 1999), and that accurate ex vivo detection in tissues is more easily attainable (Fris and Midelfart, 2007). As acquisition of 31P NMR spectra can easily require times that are longer by a factor10, it would be very significant to be able to obtain the same or better analyses with only 1H data.
We considered treating superoxide quantifications as an outcome variable or an additional biomarker measure of insult, but had to hold back because HEt fluorescence detection of superoxide is so sensitive a technique, that noninjurious perturbations, such as isoflurane anesthesia with no insult, can produce HEt images similar to those in this paper (Tanaka et al, 2003). We also considered having an experimental group in which therapeutic hypothermia was initiated at the end of the 45-minute OGD period. Preliminary data (not shown) strongly suggested, as might be surmized from Figure 3, that one would not find this group's ATP results to be distinguishable from the normothermia results.
In conclusion, in the presented OGD brain slice model of neonatal asphyxia, normothermic tissue suffered ATP losses and cell death, but tissue given a 4°C temperature decrease preserved ATP levels, whether the temperature change was applied at the start of OGD or 15minutes later. However, cell death was substantially reduced only when the temperature decrease was simultaneous with the OGD onset. Although the metabolomic analyses found that the different treatment groups were forming separate clusters, larger data samples and improved choices of outcome variables are needed for further mechanistic insight.
Many preclinical studies have found that after brain hypoxia/ischemia, outcomes can be dramatically improved by mild therapeutic hypothermia, with temperature decreases typically of only 4°C. Given that body chemistry is occurring at an absolute temperature of ~309°K, a 4°C decrease is only a 1.3% change. What makes it possible for small temperature changes to be so important?
The answer can be found in the Arrhenius equation for the rate constant K of any chemical reaction A → B: K=(a constant) × exp([−activation energy]/[kT])
The ‘activation energy' normally equals the threshold energy, because contributions from the mean thermal energy are normally extremely small compared with reaction energies (Menzinger and Wolfgang, 1969). In such situations, the activation energy does not change with temperature, and the rate constant has exponential behavior with (1/T). Earlier studies based on temperature independence of the activation energy found that a decrease of 4°C would lower the human cerebral metabolic rate by ~28% (taking Q10=2.3), and the rat cerebral metabolic rate by 63% (taking Q10=12). However, near reaction thresholds the activation energy can be a function of temperature, and dramatic reaction rate changes can occur within a 4°C temperature interval, if chemical reaction thresholds are crossed in that interval. If, for example, the increase of a few degrees allows greater opening of hundreds of hydrogen bonds, reactions that depend on those bond openings will be greatly facilitated.
The authors declare no conflict of interest.
This study was supported by NIH award R01 GM34767.