The innovations produced and inspired by the Human Genome Project included new approaches to analyzing problems where the number of people or organisms in a study is enormously smaller than number of biomarkers—genes in the case of genomics, proteins in the case of proteomics, transcription factors in the case of transcriptionomics, and metabolites in the case of metabolomics and metabonomics. The new consolidation of fundamental intellectual concepts common to statistical analyses in all of these multivariate disciplines has been given the label of “omics”. Advances in measurement technologies that have occurred in parallel in each of the “omics” specialties have led to an explosion of advances and publications. In metabolic studies the tradition has gone from focusing on a limited number of key metabolites having principal roles in the situation under investigation, to focusing on a metabolomics or metabonomics approach. Because large numbers of accurate metabolite quantifications can now be made in many instances, it is possible that aggregates of metabolite changes might reveal mechanisms or key metabolic reactions that are not discernable in hypothesis-based studies of limited subsets. Numerous review papers have provided overviews of metabolomics and metabonomics from different mechanistic and physiological perspectives [1–13]. None of the reviews, however, have specifically emphasized the complex area of oxidative stress. Our research focus, which is the use of NMR technologies for understanding healthy and unhealthy neurophysiological situations involving oxidative stress, has motivated us to review the metabolomics and metabonomics literature with this focus. In this paper we present our views of the most important conceptual and practical points about oxidative stress and metabolomics and metabonomics, and then discuss important recent advances and studies, going from studies which emphasized a limited number of metabolites to studies that emphasized the behavior of entire metabolomes.

The orbital properties of oxygen’s electrons account for its reactive properties as a radical. An understanding of that began with Michael Faraday’s discovery in the 1840’s that oxygen is paramagnetic (attracted into a magnet.) However, it was not until 1925, after quantum mechanics became more widely appreciated, that the mechanism was fully understood. Robert Mulliken explained it by noting that in its lowest energy state, molecular oxygen has two unpaired electrons. Each electron has an intrinsic angular momentum, or spin, of ½, meaning that in a magnetic field it has only two quantum states (a doublet): along the magnetic field and opposite to it. Together the spin angular momentum of oxygen’s two unpaired, spin ½ electrons, which are in different orbits, add up to a spin angular momentum of 1, which in quantum mechanics is a triplet state. Triplet oxygen is the lowest energy state of molecular oxygen, and is the normal, commonly existing state. There is also an excited orbital state in which the two normally unpaired electrons become paired in the same orbit, with one spin aligned opposite to the other, producing spin 0, which is a singlet. This excited and potent but short-lived oxidant state is known as singlet oxygen. A molecule with an unpaired electron is a radical—except for free molecular oxygen, O₂! Radicals are highly reactive chemically. If oxygen within a molecule has an unpaired electron, that molecule is known as an oxygen radical. Reactive oxygen species (ROS) are chemically-reactive molecules containing oxygen, such as hydrogen peroxide (H₂O₂) and the hydroxyl free radical (HO·). Oxygen’s extremely high affinity for electrons can also allow it to take on an extra electron to form the superoxide anion radical O₂ ⁻·, which is also an ROS molecule. Henceforth we will denote O₂ ⁻· by the shortened term “superoxide”. In living systems there is a constant formation of oxygen radicals, along with a system of reactions that constantly consumes them. When cells or organisms produce oxygen radicals at a greater rate than they can be consumed by the intracellular systems built for that purpose, there is a condition of oxidative stress, and a concern that there will be harmful ROS effects.

In mitochondrial oxidative phosphorylation the final chemical reaction produced by the electron transport chain is

Atomic oxygen collects 2 electrons here, with the basic reaction being

which can be appreciated in the first equation from knowing that an early step in the electron transport chain was

The “energy” produced in the reaction occurs as a proton gradient across the inner mitochondrial membrane, and is used in the phosphorylation reaction that produces ATP.

It is inevitable that the substrates in the high rate basic reaction will sometimes produce superoxide, where the reaction is

Evolution has produced key antioxidant enzymes and scavengers to deal with this, the primary ones being: superoxide dismutase, which exists in different forms in mitochondria and the cytosol; catalase, which is present primarily in peroxisomes; and glutathione peroxidase, which is found both in the cytosol and in mitochondria. Superoxide dismutase catalyzes

while catalase catalyzes

Glutathione is a tripeptide consisting of cysteine, glycine, and glutamate. It exists in reduced (GSH) and oxidized (GSSG) states and catalyzes the decomposition of H₂O₂ and organic peroxides:

Although glutathione, along with SOD are the major intracellular antioxidants, the major blood plasma antioxidants are uric acid, vitamin C (ascorbate), vitamin E (α-tocopherol), thiols, and bilirubin [21]. Uric acid is produced during purine degradation by xanthine oxidase from xanthine and hypoxanthine. Increased uric acid levels are common during hypoxia and ischemia, and are a marker for oxidative stress [21].

An important additional path for superoxide damage was appreciated after it was discovered that the messenger molecule responsible for vasodilation, initially known as EDRF (Endothelium-Dependent Relaxation Factor), could be identified as nitric oxide (NO) [22]). The nitric oxide radical NO combines with superoxide to form peroxynitrite [23].

which among its many reactions produces the lasting marker nitrotyrosine [24,25]. Nitric oxide together with radicals derived from its reactions with either superoxide or oxygen, are known collectively as reactive nitrogen species (RNS) [26]). Because they often act together, RNS are as much a concern as ROS in oxidative stress [27].

During physiological increases in metabolism, for example during exercise, increases in oxygen consumption increases the production of oxidants that can produce damage, such as that coming from muscular fatigue after strenuous exercise. Protective radical scavenging is accomplished not only by reactions just described, but also by scavenging by many other naturally present compounds that contain keto-carbonyl and/or rings, such as vitamin E, vitamin C, and many TCA Cycle metabolites. Although radical damage is the topic of this review, it should not be forgotten that excessive antioxidant activity might be disadvantageous to organisms if it interferes with mechanisms of recovery and adaptation. For example, in ischemic preconditioning, transient increases in reactive oxygen species can provide protection, which in turn can be blocked by antioxidants [28]. Given all of the above, it is easy to see that perturbations of metabolite distributions can result from oxidative stress, oxidative damage, and even from antioxidant therapy. Thus it is appropriate to use multivariate metabolomics analysis techniques to assess and categorize oxidative protection, injury, and recovery.

The front end of all metabolomics projects is accurate data acquisition. All approaches work best when optimum state-of-the-art techniques are use to obtain high signal-to-noise data in which different metabolites appear precisely in the same spectral location with identical characteristics. This is often not trivial, as many NMR resonance peaks shift with pH, and LC/MS or LC/NMR peaks also shift from chromatography changes. With changes in time or circumstance, metabolites in a data set very commonly exhibit large variations in their absolute concentrations, with biological relevance sometimes being more in the percentage variation rather than in having the largest absolute value. For all of these reasons preprocessing is commonly performed in which data variables are centered and scaled with the methods deemed most appropriate in each situation. One approach (Pareto scaling) sets the mean value of each variable as the zero reference, and then scales magnitudes into standard deviation units. However, there are many other available approaches, such as logarithmic or exponential scaling [69]. Appropriate quality controls must be established by investigators for each type of data acquisition and processing.

The next step in metabolomic analyses is the application of pattern recognition algorithms that find statistical differences and attach them to distinct biological phenomena. The “omics” approach generated popularity for certain basic algorithms found in many commercial software packages. Two distinct approaches have been established for processing NMR spectra, as outlined in Figure 4. In the first, known as chemometrics, no attempt is made to identify specific metabolites in the spectra. Spectral patterns and intensities are statistically compared somewhat as in a fingerprint analysis. Axes for variables are divided into “bins”, or short intervals, and intensities for each bin are the statistical variables that are analyzed so as to identify relevant spectral features that will distinguish different data sets. After grouping data sets according to their differences, hopefully with the result being that different classes are very distinct, one then invokes one or more approaches for identifying and quantifying the many metabolite concentrations prominent for each group [70,71]. In the second approach to NMR spectral analysis, known as targeted profiling [1,72], the first job is identifying and quantifying each metabolite in every NMR spectrum, so that metabolite concentrations are the variables, and the next step can be to use “omics”-based statistical methods to search for meaningful differences among them. In targeted profiling one must have a priori knowledge of each metabolite’s complete NMR spectrum, which means knowing the spectral location and relative signal intensity for every ¹H in each metabolite. The advantages of each method have been compared [72,73]. With chemometrics one makes no assumptions about the identity and quantity of metabolites in the spectra, and, with statistical significance, one separates the spectra into different groups. This is a very appropriate and efficient approach to large data samples where metabolite compositions and quantities might be very different. Thus chemometrics is more easily performed, but its results are less easily translated into harder into clinical or mechanistic significance. In targeted profiling one quantifies specific metabolites and distinguishes spectra based on these. Thus targeted profiling is tedious and more slowly accomplished. In situations where the chemical composition of the sample is known, quantifications use information about the entire shape of a metabolite peak, and, depending on spectral circumstances, will provide more accurate or less accurate metabolite quantifications. There are different approaches to targeted profiling. One approach that has been used for many years is the Linear Combination Model (LC Model), where one finds the best fit to the shape of an NMR spectrum by assuming that it is a linear combination of known spectral lineshapes, each being from a particular metabolite [74]. Interactive fitting via computer software is a more modern method [72].

The basic principle behind “omics” methods, which were made practical by the evolution of fast, powerful computers with large data storage capacity, is its being based on a geometrical approach. One copes with having a large number of biomarkers or predictors for one subject by considering the entire set of numbers as the coordinates of a single point in a multidimensional space. So, if one has 200 different metabolites measured for one biological sample, or gene expression quantifications for 2000 genes from one sample, for metabolomic analyses that sample is represented by a single point in a 200 dimensional space for metabolomics, while for genomic analysis it is represented by a single point in a 2000 dimensional space. If in the multidimensional space one plots data sets for two different groups, for example from two groups of patients where each group had a different treatment for the same condition, then the distribution of points for each of the two groups will, within statistical error, overlap in every single plane or line that one passes through the space. If however, a plane can be found where projections to the plane for one treatment group has a cluster that is significantly separated from a cluster in the other treatment group, then a difference exists that can be modeled by the acquisition of a lot more data. The general process here is known as SIMCA, or Soft Independent Modeling of Class Analogy. And orthogonal directions through the point distributions where variance is maximal are known as Principal Component Axes. Usually the number of axes that will sustain variance differences between two groups is very much less than the number of biomarkers. Often two or three principal component axes are sufficient, but even if the number is larger, the dimension of the space where variations appear is much less than the number of biomarkers, 200 or 2000 in the examples we cited.

There are two general types of pattern recognition algorithms in multivariate statistical analyses: unsupervised and supervised. The most common unsupervised method, Principal Component Analysis (PCA) is usually employed first even if one goes on to complex algorithms with supervised analyses. In unsupervised analyses all variables are treated as random, even though some variables might be strongly dependent on others. In supervised analyses one uses prior information to make two groups of variables: predictors, usually denoted as X, and response variables, usually denoted by Y. The point is to test how data in X determine the behavior of data in Y. Many different algorithms are available, with common ones being PLS-DA (Projection to Least Squares-Discriminant Analyses; often with the words before “DA” being “Projection to Latent Systems”), DPLS (Discriminant Partial Least Squares), and more.

For all algorithms two-dimensional Scores Plots are used to display the clusters in planes where the x and y axes are Principal Component axes, and the (x,y) coordinates of a data point comes from projecting the multidimensional vector for that point onto the plane of the chosen Principal Component axes. In the PLS-DA, for example, one algorithm sharpens the PCA clusters and their separations by creating a new Y variable for each data class [75]. The Scores Plots in the PCA and PLS-DA analyses have corresponding Loadings Plots that identify the metabolites that contributed most to separations in the Scores Plot. Like in the Scores Plot, points plotted on the Loadings Plot come from projecting multidimensional vectors onto the plane of chosen Principal Component axes. Each point on a Loadings Plot corresponds to a single metabolite, because the vectors projected onto the chosen plane are the component vectors along each dimension that added to make the single vector that represents the full data set. As an example, an NMR data set of fifty metabolites would contribute one point to a Scores Plot, and fifty points to a Loadings Plot. The construction of a model via SIMCA methods requires the acquisition of training data sets that define the geometrical attributes of different classes in the multidimensional space. When different classes have large separations in a Principal Component subspace, the model can be used to predict the likelihood that new data will belong to a particular class. A recent T-Cell study provided a good example [76].

Recent advances in ¹H NMR data acquisition methods have led to more accurate quantifications of NMR-detectable metabolites. Of special note is the refined technique of Two-Dimensional ¹H J-resolved (JRES) NMR Spectroscopy [77,78]. J-coupling, which arises from the magnetic field property of particles that have spin, is the indirect coupling between two chemically bound nuclear spins that occurs via the electrons of their chemical bond. In 2D JRES NMR Spectroscopy, the two dimensions into which metabolite signals are separated are: one by chemical shift (usually indicated in ppm units), and a second one by the magnitude of the J-coupling to neighboring protons (usually in units of Hertz). Analogous to when a bar magnet is placed in a magnetic field, a spin ½ nucleus such as ¹H or ¹³C can exist in either a low energy or high energy state, depending on whether it is aligned with or against the main (or applied) magnetic field, which is constant for a given spectrometer, with the field strength being somewhere between 7 and 22 Tesla. Which energy state that a nucleus, a proton for example, is in influences, to a very small but measurable degree, the magnetic field experienced by the electrons responsible for chemical bonds, and ultimately by adjacent protons. For the simple case of two protons separated by three bonds, each proton resonance, instead of being a singlet peak, will be split into two peaks, i.e., into a doublet, due to the fact that the other proton can be in one of two states. The separation in Hertz between the two signals in the doublet is referred to as the J-coupling. It is independent of the strength of the applied magnetic field, because it is determined only by the difference in the magnetic field at one nucleus that is experienced by an orientation change in another nucleus. J-coupling can exist between ¹H nuclei on adjacent carbons, for example between the methyl protons on lactate and the ¹H proton on the center carbon (C2), or between ¹H nuclei and adjacent or nearby ¹³C nuclei. The ¹H–¹H J-coupling in lactate is approximately 7 Hz, while the ¹H–¹³C J-coupling between the lactate’s methyl protons (on C3) and ¹³C in [3-¹³C] lactate is approximately 128 Hz. For the rest of this discussion we will be referring only to ¹H-¹H J-coupling.

The phenomenon of J-coupling is used to separate the ¹H signals into a second dimension in the 2D JRES experiment, as illustrated in Figure 5. In Figure 5 it is also apparent that the 2D separation of the resonances helps resolve signals that overlap in normal 1D ¹H spectra. For metabolomics studies, a 2D ¹H JRES spectrum can be fit directly. The choice of NMR data acquisition parameters for a 2D ¹H JRES spectrum permits the extraction of a 1D ¹H spectrum from the 2D dataset in which the J-couplings have been removed. This is because the 1D ¹H spectrum is obtained by projecting the 2D data straight down onto the horizontal ¹H axis, and in the 2D data set, the data in the J-coupling dimension are vertically aligned. Thus for any particular ¹H resonance, data for all J-couplings appear in the projected 1D JRES spectrum at the same ppm. Removal of J-coupling information from 1D ¹H spectra in this way reduces the overlap of resonance peaks from different metabolites, and allows more accurate metabolite quantifications. Improvements in the identification and quantification of glutamate and glutamine are particularly impressive, as seen in Figure 5 and Figure 4. Other impressive comparisons and elucidations can be found in papers by Viant et al [79] and Ludwig and Viant [78]. The notation 1D ¹H pJRES, which is popular among certain authors, is used to refer to the ¹H spectrum produced by projecting a 2D ¹H JRES spectrum onto the ¹H axis. The 2D ¹H JRES and 1D ¹H pJRES approaches are likely to become widely used, as such data can be recorded quickly (~20 min), and 1D and 2D spectra are easy to process and analyze.

Other 2D NMR methods that have been used in metabolomics studies include ¹H-TOCSY [80] and ¹H-¹³C HSQC experiments [81,82]. An advanced 2D NMR technique, known as Targeted Projection NMR Spectroscopy [70] can be used when overlapping signals hamper analysis of TOCSY and HSQC spectra. In this method, a “tilted” 2D projection of a hybrid 3D TOCSY-HSQC spectrum is recorded. The projection can be tailored to resolve specific overlapping sets of resonances [83]. A highly cited 2008 publication by Gowda and colleagues reviewed many NMR experiments available to develop diagnoses from metabolic profiles [84].

In the early 1990’s it was appreciated that increased oxidative stress is important in the pathophysiology of diabetes, because the protein glycosylation that occurs during long exposures to hyperglycemia leads to free radical production [117–119]. In diabetic patients, increases in oxidative damage to lipids and proteins correlates with the increases in complications of the disease [21,119].

A recent 9.4 Tesla ¹H-NMR metabolomic investigation of diabetic retinopathy in humans used a chemometric approach with PCA and PLS-DA analysis to show that spectra for vitreous humor samples clearly separated diabetic patients from non-diabetic patients. Among 14 very clearly distinguished metabolites, decreased ascorbate and increased galactitol were the main biomarker changes for diabetic retinopathy, although increased lactate and glucose were found in samples from all diabetics [120]. The decrease in ascorbate was attributed to higher consumption from higher levels of oxidative stress. A recent preclinical study of diabetes used a streptozotocin rat model to study urine and plasma metabolomics with ¹H NMR at 11.75 Tesla [121]. Spectra were analyzed with chemometrics and profiling that identified 17 metabolites. Plasma TBARS (thiobarbituric acid-reacting substances) were quantified as a measure of oxidative stress. Diabetic rats had altered metabolic pathways and augmented fat degradation products caused by free radicals.

With regard to metabolomic investigations of oxidative stress in the kidney, one human study, in which it was known a priori that cyclosporine nephrotoxicity initially occurs in the proximal tubule, somewhat driven by increased formation of oxygen radicals, found that metabolomics measures with ¹H-NMR and HPLC-MS are sensitive enough to detect toxicity as soon as 4 h after a single oral dose, and could serve as a tool for clinical diagnoses [122]. Several rodent studies used similar metabolomics methods to detect and quantify oxidative stress in the kidney from ischemia/reperfusion injury [123], radiation injury [124], and drug toxicities for acetaminophen [125], cinnabar [126], fenofibrate [127], and ochratoxin [128].

Two very recent liver metabolomics studies provided data on oxidative stress in humans. In the first [129], plasma profiles in patients were studied for three pathological conditions: nonalcoholic fatty liver disease, steatosis, and steatohepatitis. Metabolomic data could separate healthy subjects from those with fatty liver disease with an error rate of approximately 8% and separate those with steatohepatitis from healthy controls with an error rate of 4%. However, metabolomics data could not separate patients with hepatic steatosis from those with steatohepatitis. Marked changes in bile salts and glutathione-related biochemicals were identified in the different group, and a panel of biomarkers was found that can potentially be used to follow response to therapeutic interventions.

In liver transplantation one first causes oxidative stress by inducing ischemia during organ removal from the donor, and then one causes oxidative stress by reoxygenation and perfusion after organ implantation in the recipient. Thus hepatotoxicity in liver transplantation acutely involves considerable oxidative stress. In a very recent human study of 8 liver transplant patients [130], consecutive biopsy findings included that metabolomics time-profiles demonstrated considerable changes in redox-active metabolites up until 18 h postreperfusion, after which there was stabilization. Metabolomics data came from Fourier Transform Ion Cyclotron Resonance Mass Spectrometry (FT-ICR MS) of liver biopsy tissue, and also from microdialysates of extracellular fluid by coulometric electrochemical array detection (CEAD). FT-ICR MS reproducibly detected more than 4,000 peaks, revealing hundreds of significant metabolic differences between pre- and postreperfusion grafts.

In the literature of preclinical experiments there were numerous studies in rodents that showed distinct associations between metabolomic data and experimental manipulations. For example, deficiency of α-tocopheral was studied [131], as were hepato- and nephron-toxicants [132], obesity [133], and acetaminophen toxicity [134].

One important function of the liver is to detoxify poisons that come into an organism, especially via the gastrointestinal system. Xenobiotics are chemicals found in organisms that are not normally produced or expected to be present. The high importance of being able to recognize toxins motivated the formation by certain leaders in metabonomics research of The Consortium for Metabonomic Toxicology (COMET), which is between five major pharmaceutical companies and The Imperial College of London [135]. This group has assessed methodologies and used rodent urine and blood to generate a ¹H NMR metabonomic database for use in evaluating and dealing with xenobiotic toxicity.

Investigations with human tissues and fluids included a LC-MS (liquid chromatography and mass spectroscopy) metabolomics study of plasma lipids that was done for 14 patients having documented silent myocardial ischemia (SMI), as compared to 25 matched control patients [136]. Scores Plots in the multivariate statistic analysis showed a clear separation between the two studied groups. Plasma concentrations of four phospholipids had a tight association with SMI, with 1-linoleoylglycerophosphocholine (C18:2) being significantly decreased in the SMI group, suggesting that such measurements might facilitate early diagnosis. A careful metabolomics study of hypoxia in humans [137], motivated by the deleterious roles hypoxia plays in cardiovascular disorders and numerous diseases, looked for evidence of oxidative stress, and also for evidence that HIF, hypoxia-inducible transcription factor, is activated in vivo in humans during hypoxia. The group used a hypobaric chamber to expose fourteen human subjects to normoxia corresponding to an altitude of 1600 meters, and to hypoxia corresponding to an altitude of 4300 m (alveolar PO₂ ≈ 48 mm Hg). HIF-1 DNA binding and HIF-1α protein were evaluated in circulating leukocytes, and several markers of oxidative stress were among 47 metabolites studied with ¹H NMR plasma metabolomics. Circulating glutathione was reduced by 35% (p = 0.001), and lactate and succinate were increased by 29 and 158%, respectively (p = 0.007 and 0.001), as were urinary 15-F2t-isoprostanes (p = 0.037), showing that hypoxia causes substantial increases in plasma and urinary markers of oxidative stress. Another metabolomics study in humans [138] used LC-MS to look for oxidative stress in the etiology of abdominal aortic aneurysms (AAA), the 13th leading cause of death in the United States with up to 9,000 deaths annually. An AAA is a permanent, localized aortic dilation greater than 3 cm in diameter. AAA enlargement occurs irregularly, and pathology of the involved vascular tissue has been associated previously with oxidative stress, chronic inflammation, elastin fragmentation, apoptosis, loss of extracellular matrix, neovascularization, and a depletion of smooth muscle cells. Metabolites were compared that came from different portions of human aortic aneurysms--luminal and abluminal layers, aneurysm walls, and healthy walls. Compounds related to inflammation and oxidative stress were found, as was a possible role of fatty acid amides. Some metabolites (e.g., hippuric acid) had not been previously associated with such aneurysms.

Several recent animal studies have used metabolomics to investigate oxidative stress in cardiovascular systems. Complementary to the human data just cited are proteomic and metabolomic analyses of atherosclerotic vessels from apolipoprotein E-deficient mice that revealed alterations in inflammation, oxidative stress, and energy metabolism [139] Oxidative stress in atherogenesis was apparent in that study from the upregulation of numerous antioxidant proteins the detection of ROS protein damage, and the increase in albumin degradation. Damage was found in enzymes susceptible to free radical inactivation, including aconitase, cytochrome C reductase, and the Rieske protein of ubiquinol. Such damage can release free iron, thereby promoting subsequent formation of hydroxyl radicals, and perpetuating oxidative stress, as explained early in the review. Proteomic and metabolomics measures were made in a dog model for congestive heart failure and ventricular tachycardia, with the heart paced at 240 beats per min for 24 h in one group, and for 2 weeks in another [140]. Oxidant stress was evident in the 24 h group from the substantial decrease in antioxidant proteins, including SOD and peroxiredoxin. Heat shock protein increases and other protein changes were analyzed, as was the metabolic stress related to congestive heart failure (CHF). That study provided insights to the arrhythmogenic atrial remodeling that results from such physiological stress. In an initial study by others of mice deficient in protein kinase C delta [141], proteomic and metabolomics analysis of smooth muscle cells demonstrated that loss of PKC delta improved the cellular redox state by elevating GSH levels, and thus provided protection against oxidative stress-induced cell death. However, in a second study with mice deficient in protein kinase C delta [28], they showed that the loss of PKC delta caused the loss of protection provided by ischemic preconditioning, because transient increases in reactive oxygen species cause preconditioning, while antioxidants block the beneficial effects of preconditioning. The two studies, besides adding useful specific biochemical information, provide a good example of a general point made earlier, that antioxidant augmentation is not good in situations where it impairs host defenses.

Metabolic derangements and adaptations by cancers have recently been carefully reviewed [142], with two prominent features being: (1) excess glucose metabolism and lactate production with avoidance of the TCA cycle, even in the presence of substantial oxygen availability [Warburg Effect]; and (2) mitochondrial malfunction with augmented ROS production. The role of ROS in tumor progression is controversial, because increased ROS can serve what might appear to be opposite functions, namely induction of apoptosis, or augmentation of mutagenesis and tumor proliferation. Apoptosis can also be induced by radiation therapy of tumors. Indeed, radiation toxicity is attributed to the increased oxidative stress triggered by direct ionization of biological polymers and the generation of reactive oxygen species via radiolysis of water, which produces hydroxyl radicals [143]. In radiation therapy, however, there are adaptive responses. A study of 22 metabolites from healthy human leukocytes that had radiation exposures of 2 Gy, 4 Gy and 8 Gy found that relative to control data at 0 Gy, metabolites were substantially depleted at 2 Gy and 8 Gy, but substantially upregulated at 4 Gy, with many at values greater than control. A recent review [144] discussed metabolomics and oxidative stress in pheochromocytomas, endocrine cell tumors that come from chromaffin cells in the adrenal medulla and can produce enormous amounts of catecholamines.

A wide variety of publications were found that make important clinical and mechanistic associations between oxidative stress and specific metabolites, metabolite groups, and metabolomics analyses. Table A1 provides a compendium of associations between pathological situations with oxidative stress, and metabolites identified in particular studies, all of which are cited in the list of references. Included in the table are references to six publications [145–150] that had valuable information, but matched more than one the categories we established, and were not referenced earlier.

Complex relationships connect oxidative stress to many physiological processes. It is hoped that by developing multivariate analyses techniques it will be able to discern characteristics of metabolite ensembles that distinguish different physiological processes, and in so doing, be able to learn key mechanistic aspects and also develop useful interventions. The goal of studying the metabolomics of oxidative stress is to develop methods that serve as a tool for diagnoses, treatment guidance, and drug discovery. The rapid pace of research in this area gives hope for such progress.