Local and Interregional Neurochemical Associations Measured by Magnetic Resonance Spectroscopy for Studying Brain Functions and Psychiatric Disorders

Magnetic resonance spectroscopy (MRS) studies have found significant correlations among neurometabolites (e.g., between glutamate and GABA) across individual subjects and altered correlations in neuropsychiatric disorders. In this article, we discuss neurochemical associations among several major neurometabolites which underpin these observations by MRS. We also illustrate the role of spectral editing in eliminating unwanted correlations caused by spectral overlapping. Finally, we describe the prospects of mapping macroscopic neurochemical associations across the brain and characterizing excitation–inhibition balance of neural networks using glutamate- and GABA-editing MRS imaging.


INTRODUCTION
In vivo MRS is the only noninvasive technique that can directly measure brain chemicals in vivo.
Using techniques similar to MRI, MRS can measure concentrations of many neurometabolites as well as metabolic fluxes from localized brain regions (1). Over the past decades, MRS studies have found biochemical abnormalities in essentially all neuropsychiatric disorders, providing important insights into our understanding of etiologies and treatments of various brain diseases. These studies have, in most cases, focused on alterations in the concentrations of individual brain neurometabolites by comparing them among different cohorts and/or effects of treatments.
Many significant correlations among neurometabolites and between individual neurometabolites and non-MRS measures of brain function and disorders have been reported more recently. Altered correlations have been found in neuropsychiatric disorders, revealing abnormal neurochemical associations under pathophysiological conditions [e.g., (2)(3)(4)(5)(6)(7)(8)(9)(10)(11)(12)]. The strong correlations among Nacetylaspartate (NAA)/choline in precentral gyrus, midcingulate cortex, and thalamus found in healthy subjects were absent in patients with amyotrophic lateral sclerosis (4). Correlation between hippocampal Glx (glutamate + glutamine) and NAA has been demonstrated to be a more sensitive biomarker differentiating between healthy controls and schizophrenia patients than either neurometabolite alone (6,7). In patients with subclinical hepatic encephalopathy the occipital lobe phosphodiester measured by 31 P MRS and Glx levels were found to be negatively correlated (13). Interregional correlations of glutamate and GABA levels have also been reported in many studies [e.g., (8,(14)(15)(16)(17)]. These studies have clearly demonstrated that neurochemical associations are abnormally altered in many brain disorders, but the absolute strengths of these correlations measured by different MRS methodologies have been inconsistent or controversial [e.g., (5,14,15,(18)(19)(20)(21)(22)(23)]. In particular, the effects of spectral overlap on the observed neurometabolite correlations have yet to be illustrated although they can significantly confound the intrinsic neurochemical correlations of interest.
Many studies of neurochemical associations rely on spectral fitting [e.g., (24)] to extract neurometabolite concentrations from overlapping signals. When there is significant spectral overlap between two signals overestimate of one signal is statistically correlated with underestimate of the other signal and vice versa, even when there exists no neurochemical correlation between the two signals. In addition, this type of statistical correlations can propagate due to the intensity constraints imposed by LCModel (24) or overlapping with neurochemically correlated signals. When neurometabolite concentrations are correlated with other measurements (e.g., behavior, resting state fMRI functional connectivity, or gene expression), statistical correlations due to spectral overlap among MRS measurements can also affect correlations between MRS measurements and non-MRS measurements.
In this article we review dominant metabolic pathways connecting major neurometabolites (25)(26)(27) which underpin the neurochemical associations detected by MRS correlation studies. Non-MRS neurochemical studies of animal models that found correlated changes in neurometabolite concentrations under various pathophysiological conditions are also discussed. Monte Carlo simulations are performed to demonstrate the existence of statistical correlations that originated from spectral overlap. Finally, we discuss MRS techniques that eliminate spectral overlap and associated statistical correlations. We hope that these discussions will spur interest in developing MRS techniques for mapping neurochemical associations across the brain to facilitate a variety of clinical investigations. In particular, since glutamate and GABA play dominant roles in the excitation-inhibition balance (28,29) and in many neuropsychiatric disorders (30)(31)(32)(33), MRS characterization of glutamate-GABA associations among the nodes of neural networks may provide considerable insight into the interactions between glutamatergic and GABAergic systems and their abnormalities.

METABOLIC PATHWAYS UNDERLYING NEUROCHEMICAL ASSOCIATIONS
Predominant metabolic pathways connecting NAA, glutamate, glutamine, and GABA are reviewed in detail. For clarity, a table summarizing these pathways is provided ( Table 1).

NAA-Glutamate Association
NAA is the most abundant free amino acid derivative in the CNS. NAA is found almost exclusively within the nervous system. In the adult brain, it is mostly confined to neurons. As such, it is of great clinical interest as it has been considered as a neuronal marker for assessment of neuronal viability in a variety of neuropsychiatric disorders using proton MRS. For example, NAA level is markedly reduced within the infarct of stroke patients (34), while a higher NAA level is associated with a better clinical outcome (35). Despite the intense interest in NAA, both its physiological and metabolic roles in normal brain functions as well as in neuropsychiatric disorders remain poorly understood [for a review of the putative role of NAA, see (36)].
Important metabolic associations exist among major neurometabolites observable by MRS through precursor-product relationships and sharing common substrates (26,37,38). Glutamate, the most abundant intracellular amino acids in mammals, is a key component of intermediary metabolism and a precursor of numerous cellular components including proteins as well as neurometabolites such as GABA, N-acetylaspartylglutamate (NAAG), and glutathione (26). As glutamate is also the primary excitatory neurotransmitter in the CNS, it is not surprising that the proton MRS has found abnormal glutamate levels in many neuropsychiatric disorders including multiple sclerosis (39), major depression (40), and bipolar disorder (41) where glutamatergic dysfunction is broadly implicated.
Glutamate is primarily synthesized by transamination from a-ketoglutarate catalyzed by aspartate aminotransferase (25): glutamate + oxaloacetate ↔ a À ketoglutarate + aspartate Glutamate also is produced, to a much lesser extent, from aketoglutarate and ammonium via glutamate dehydrogenase, from glutamine via hydrolysis catalyzed by phosphate-activated glutaminase, by other transamination reactions that use aketoglutarate as receptor of the amino group, and during protein turnover (42).
The transaminases of importance for maintenance of glutamate homeostasis in the brain are mainly aspartate aminotransferase, branched-chain aminotransferase, and alanine aminotransferase with aspartate aminotransferase dominating overwhelmingly, representing >97% of the glutamate-related aminotransferase activities (26). 13 C magnetization transfer MRS experiments have shown that the aspartate aminotransferase reaction is extremely fast in the brain in vivo (43). This rapid transamination by aspartate aminotransferase predominates in the formation of glutamate in the CNS, forming strong metabolic coupling between glutamate and aspartate (25). The tight connection between glutamate and aspartate becomes conspicuous under many pathophysiological conditions where the brain is challenged or perturbed metabolically. For example, during hypoglycemia a decrease in glutamate concentration was accompanied by an increase in aspartate concentration (44)(45)(46). Similarly, barbiturate anesthesia and hypothermia were also found to lower the concentration of aketoglutarate accompanied by reduced glutamate concentration and increased aspartate concentration (47)(48)(49). These correlated changes in glutamate and aspartate were explained by a sizable shift in the aspartate aminotransferase reaction towards aspartate formation at the cost of a reduction in glutamate concentration (25,26). In contrast, both hypocapnia and hypoxic hypoxia are associated with an increase in glutamate concentration and a reduction in aspartate concentration with the aspartate aminotransferase reaction shifting in the opposite direction (25,26,50). A strong metabolic coupling between glutamate and aspartate mediated by the rapid and ubiquitous aspartate aminotransferase reaction also affects NAA, the dominant signal in proton MRS, as NAA is primarily synthesized from acetyl coenzyme A (CoA) and aspartate by NAA synthase (51): acetyl À CoA + aspartate ! CoA + H + + NAA In addition to this indirect connection from glutamate to NAA synthesis via aspartate, both glutamate and NAA are products of NAAG catabolism catalyzed by N-acetylated-alinked-amino dipeptidase (52)(53)(54). The deacetylation of NAA catalyzed by aspartoacylase has also been proposed as a significant metabolic pathway for NAA to act as a reservoir for glutamate in brain (55).

Glutamate-Glutamine Association
In contrast to glutamate, which is predominantly located in glutamatergic neurons, glutamine is primarily an astrocytic chemical. In the MRS literature glutamate + glutamine is often collectively referred to as Glx as at lower magnetic fields it has been difficult to separate the two spectroscopically. Abnormal glutamine concentrations have been found in several brain disorders including cancer, hepatic encephalopathy, and other neuropsychiatric disorders (56,57).
Although the overall glutamate pool in neural tissues rapidly turns over fueled by primarily glucose under normal physiological conditions, glutamate released from nerve terminals is replenished by astroglial glutamine via the glutamate-glutamine neurotransmitter cycle [ Figure 1; (26,59)]. The negatively charged highly hydrophilic glutamate cannot diffuse across cell membranes. The concentration of glutamate in the extracellular space is extremely low due to its rapid uptake into the astroglia facilitated by high-affinity Na +dependent transport systems against a large concentration gradient (60)(61)(62). Once taken up into the astroglial cells, glutamate is converted into glutamine by glutamine synthetase: glutamate + NH 3 + adenosin triphosphate ! glutamine + adenosine diphosphate + P i or oxidized by assimilation into the tricarboxylic acid cycle of astroglial cells (26,63). Once formed, glutamine readily enters nerve terminals by its own low affinity transport system or by simple diffusion. There the phosphate-activated glutaminase converts it into glutamate (26): A large number of neurochemical as well as autoradiographic studies have confirmed that glutamate is selectively taken up by astroglial cells and then converted into glutamine, while glutamine preferentially enters the neurons and is converted into glutamate there (64). In vivo 13 C and 15 N MRS studies have quantitatively measured the glutamate-glutamine cycling flux in rodent and human brains [e.g., (58,(65)(66)(67)(68)(69)]. Results from these studies have demonstrated that the glutamate-glutamine cycle FIGURE 1 | Schematic diagram of the glutamate-glutamine neurotransmitter cycle between neurons and astroglia (58). Glutamate (Glu) is taken up from the synaptic cleft into astroglia. There glutamate is converted to glutamine (Gln) by glutamine synthetase. The inactive glutamine is released by the astroglia, enters the neurons, and then is converted into glutamate by phosphate-activated glutaminase. Glc, glucose; a-KG, a-ketoglutarate; NH 3 , ammonia.
between glutamatergic neurons and astroglia is metabolically significant, providing a major connection between glutamate and glutamine in the brain (70). In addition, over the range of glutamate concentrations found in the nerve terminals, product inhibition appears to be the main mechanism of control of glutaminase activity with glutamate significantly attenuating the activity of glutaminase (71).
The metabolic connection between glutamate and glutamine is manifested in many brain disorders and animal models. For example, elevated ammonia level in the brain is associated with increased glutamine synthesis for ammonia detoxification (67). It has been reported that in hyperammonemia and hepatic encephalopathy the elevation of glutamine level in the brain is accompanied by a reduced glutamate level as glutamate acts as a receptor for the excessive ammonia (72). The changes in glutamate and glutamine levels, however, do not necessarily go in opposite directions. Both glutamate and glutamine are abundant in the brain. Despite efforts to understand the roles of glutamate and glutamine, the reason for maintaining relatively high concentrations of glutamate and glutamine in the brain is still poorly understood. It is possible that a high concentration of glutamate and glutamine facilitates the generally high metabolic activities in the brain because glutamate and glutamine are key components of intermediary metabolism (25). They are also precursors of many other cellular components (26). Their role as "energy reservoirs" is particularly clear when the brain is under metabolic stress. For example, when glucose is scarce, such as in hypoglycemia, both glutamate and glutamine act as energy fuels. As a result, the concentrations of both glutamate and glutamine are reduced in synchrony during hypoglycemia and in many other pathophysiological conditions when normal oxidative metabolism is impaired (45,73).

Glutamate-GABA Association
While glutamate is the major excitatory neurotransmitter in the mammalian brain, GABA is the major inhibitory neurotransmitter. Since the initial detection of reduced GABA levels in epilepsy patients by MRS (74) and a strong correlation between GABA levels and seizure control in epilepsy patients treated by vigabatrin (75), MRS of GABA has greatly advanced both in terms of MRS methodologies and their clinical applications in studying GABAergic abnormalities in neuropsychiatric disorders.
Like glutamate, GABA metabolism proceeds through important intermediates of the tricarboxylic acid cycle. When GABA was first discovered (76) it was realized that GABA was formed from glutamate. Later studies identified that the principal pathway of GABA goes through a-decarboxylation of glutamate via glutamic acid decarboxylase which converts glutamate directly into GABA (77): The source of the GABA precursor is believed to be dominated by neuronal glucose with astroglial glutamine playing a smaller role (78)(79)(80). GABA synthesis from putrescine and other polyamines is metabolically insignificant in the brain although polyamines play an important role in the developing brain (81).
A fundamental aspect of glutamate-GABA association is the excitation-inhibition balance in the brain (32, 82) because of their roles as the dominant excitatory and inhibitory neurotransmitters, respectively, in the CNS. Glutamatergic neurons (e.g., cortical pyramidal neurons) receive a significant degree of GABA Amediated inhibition through interneurons (83). Balanced excitation and inhibition facilitate normal brain functions, and failure to maintain excitation-inhibition balance underlies dysfunction in many brain disorders (32,33,84). Many studies have also revealed altered glutamate and GABA. For example, GABA levels were found to decrease, whereas Glx levels increased with increasing visual input in the occipital cortex of healthy subjects (85). Both glutamate and GABA increased following vigorous exercise (86). In autistic patients the frontal lobe [GABA]/[Glu] ratio was found to be significantly lower, suggesting abnormality in the regulation between GABA and glutamate (87). Abnormalities in MRS measures of glutamate and/or GABA in many other neuropsychiatric disorders have also been reported and reviewed [e.g., (40,(88)(89)(90)].

STATISTICAL CORRELATIONS AMONG MRS SIGNALS DUE TO SPECTRAL OVERLAP
Many neurometabolites have similar resonant frequencies, leading to spectral overlap among MRS signals. Effects of spectral overlap on the Cremer-Rao lower bounds of extracted neurometabolites have been analyzed previously (104). Extracting the concentrations of neurometabolites by spectral fitting is essentially mapping the acquired MRS spectrum (spec) into a vector (conc) consisting of concentrations by inverting the matrix equation spec = basis • conc (24). Here, basis is a matrix consisting of basis spectra of the component neurometabolites, which transforms the concentration vector conc into the fitted spectrum that approximates spec in the sense of least squares. Overlapping neurometabolites become statistically correlated through the covariance matrix (COV) of conc (105): where † denotes Hermitian transposition and s 2 is the noise variance of the measured spectrum spec. The off-diagonal elements (proportional to the square of cross-correlation coefficients) of COV(conc) depend on the frequency separation among the resonances of the neurometabolites in basis. Figure 2 shows Monte Carlo simulations of the correlation between two singlet peaks as a function of their separation in the frequency domain. Figure 2A shows the correlation between conc 1 and conc 2 across individual fits when spectral overlap between the two signals is minimal. The correlation between conc 1 and conc 2 when their frequency separation equals their half-height linewidth is plotted in Figure 2B. Finally, the correlation between conc 1 and conc 2 when their frequency separation equals 0.5* half-height linewidth is plotted in Figure 2C. As shown by Figure 2 statistical correlation between these two neurochemically unrelated signals increases as their spectral overlap increases. Here the large correlation values occur when the two signals happen to exhibit similar chemical shifts, not because one signal influences the other neurochemically. This simple example of two overlapping singlets illustrates a point of caution in the interpretation of neurometabolite correlation results. For multiplets and neurometabolites with multiple resonances, spectral overlap occurs when two resonance lines overlap each other even when the chemical shifts are not very close.
The off-diagonal elements of the covariance matrix become highly significant in the presence of severe spectral overlap such as in short echo time MRS spectra. In addition to overlapping neurometabolites, a strong baseline can also cause statistical correlations (106,107) because the baseline, which arises from macromolecules and/or lipids and residual water, overlaps with essentially all neurometabolite signals. Figure 3 compares spectral fitting of two 3 T short echo time single voxel in vivo spectra by the commercial LCModel software. The spectrum on the left was acquired using short echo time Point RESolved Spectroscopy (PRESS) technique from a cubic voxel in the anterior cingulate cortex of a healthy subject at 3 T (echo time = 35 ms, voxel size = 8 ml). The spectrum on the right was generated by broadening the linewidths of the spectrum on the left by 2.0 Hz and adding random noise to maintain the same signal-to-noise ratio (107). Both spectra in Figure 3 were fitted using LCModel with the same default settings (24). With 2.0 Hz line-broadening the LCModel baseline was conspicuously stronger around the spectral region near 2.35 ppm where glutamate and the aspartyl moiety of NAA resonate. Both NAA and glutamate levels reported by LCModel were lowered by approximately the same amount (~11%) after the line-broadening [n = 10; (107)]. This reduction in the extracted metabolite concentrations was found to be generally more pronounced with greater line-broadening. Although the concentrations of neurometabolites in the two spectra of Figure  3 are identical, the LCModel produced lower neurometabolite levels and a more intense baseline after 2.0 Hz line-broadening. Because of the spectral overlap between baseline and neurometabolites, overestimating (underestimating) the baseline causes underestimating (overestimating) neurometabolites and vice versa. Neurometabolites overlapping with the same broad baseline peak are similarly underestimated (overestimated) due to the broad baseline signals. This in turn, contributes to positive statistical correlations among those neurometabolites regardless of the underlying neurochemical associations.
Spectral fitting techniques such as the LCModel heavily rely on the linewidth difference between neurometabolites and background signals to separate them. Broad neurometabolite peaks in the presence of a strong baseline, as often seen in clinical short echo time MRS data, can lead to significant quantification errors and unwanted statistical correlations because of the large baselinemetabolite covariances. The results in Figure 3 are also corroborated by an earlier study which quantitatively analyzed the estimation uncertainties caused by the baseline using Cramer-Rao lower bound (CRLB) of the baseline (106), confirming that the estimation uncertainty significantly increases with decreased baseline smoothness and increased spectral linewidths.

PROSPECTS FOR MAPPING MACROSCOPIC NEUROCHEMICAL ASSOCIATIONS BY MRS IMAGING
Although group comparison of neurometabolite correlations between healthy controls and patients reveals altered neurochemical associations in brain disorders, determining the absolute strength  (74,(108)(109)(110)(111)(112)(113)(114)(115), which are likely suited for measuring local or intraregional neurochemical associations. Mapping macroscopic neurochemical associations across the human brain has the exciting potential to broadly impact studies of normal brain functions as well as neuropsychiatric disorders (32,116,117). Here we discuss the prospects for measuring interregional excitation-inhibition balance (32,(118)(119)(120) by spectroscopic imaging of spectrally resolved glutamate and GABA. Participant motion is a major issue in scanning many patients of neuropsychiatric disorders. Studying these patients using chemical shift imaging is technically challenging because of the relatively long scan time required for phase encoding. As artifacts in chemical shift images caused by motion are hard to detect, they can lead to erroneous diagnosis and data interpretation (121). It is well-known that participant motion inside a magnetic field causes changes in resonant frequencies.
Incorporating spectral editing techniques based on highly selective radiofrequency pulses into chemical shift imaging therefore can lead to even larger errors due to the additional effects of carrier frequency mismatch on spectral editing yield.
To minimize error due to unavoidable patient movement during extended scan time necessary for phase encoding, we focus on techniques that can resolve glutamate or GABA in a single shot with relatively weak baselines at 7 T. Spectral isolation of glutamate or GABA accompanied with a weak baseline will minimize the unwanted correlations that originated from spectral overlap (107). The emphasis on minimizing spectral overlap for measuring interregional neurometabolite correlations may seem counterintuitive. However, it is necessary because, for example, overlapping with interregionally correlated signals can relay the correlation to overlapped signals. To spectrally resolve glutamate or GABA over an extended brain region, highly frequency-selective pulses popular for single voxel spectral editing cannot be used because of the unavoidable and significant residual B 0 inhomogeneity across a large volume in the brain, especially at high magnetic field strength. Highly frequency-selective pulses are sensitive to patient movement, system instability, and B 0 inhomogeneity as they will miss or partially miss the editing target in part of the slice(s) where resonance frequencies are shifted away (122,123).
Weak or nearly flat baselines are automatically produced at long echo times because of the shorter T 2 values of macromolecules (1). Serendipitously, the strongly coupled glutamate H4 (2.35 ppm) forms an intense pseudo singlet at a relatively long echo time (~100 ms) at 7 T (111,112). The resonances of glutamine H4 at 2.45 ppm and glutathione glutamyl H4 at 2.49 ppm also form pseudo singlets at~100 ms echo time (see Figure 4). Therefore, glutamate can be spectrally resolved in a single shot without using any spectrally selective pulses at~100 ms echo time at 7 T (111). At 7 T the multiplet signal of the aspartyl moiety of NAA at 2.49 ppm still overlaps with glutamine H4 and glutathione glutamyl H4 at~100 ms echo time (111). The overlapping NAA aspartyl moiety signals can be eliminated using a J-suppression pulse acting on the a-H of the aspartyl moiety of NAA at 4.38 ppm (112). The J suppression pulse can be made band-selective with a flat top frequency profile (124) to accommodate variations in B 0 . Either because no frequency selective pulses are needed (111)  LCModel generated a different baseline from the same data after 2.0 Hz line-broadening and noise injection to maintain the same signal-to-noise ratio. The large change in baseline around 2.35 ppm (marked by an arrow) simultaneously reduced fitted NAA and glutamate concentrations, therefore, causing positive correlation between NAA and glutamate even though the only difference between the two spectra is their linewidth (i.e., no correlations). Reprinted from reference (107) with permission from Elsevier.
with band-selective J suppression (112), chemical shift imaging of glutamate is feasible at 7 T in the presence of significant patient motion and residual static magnetic field inhomogeneity across the slice(s). Multiple quantum filtering can be used to edit GABA in a single shot (125)(126)(127)(128). It is also possible to generate a flat baseline using multiple quantum filtering (127). Chemical shift imaging of GABA over a large volume in the brain is challenging even at the high magnetic field strength of 7 T as all available GABA editing techniques rely on frequency selective pulses that differentially act on GABA H3 at 1.91 ppm and GABA H4 at 3.02 ppm. At 7 T, the chemical shift dispersion between GABA H3 and GABA H4 is 328 Hz. This relatively large chemical shift difference makes it possible to use band-selective pulses with a flat top (124) to accommodate changes in B 0 due to patient movement, system instability, and residual B 0 inhomogeneity while still affording a close to uniform editing yield. Figure 5 proposes a band-selective multiple quantum filtering scheme that combines spectral selectivity while allowing signal to shift within the flat passband. This multiple quantum GABA editing scheme is expected to tolerate a large frequency shift with GABA H3 lying within the flat bandwidth and GABA H4 staying outside of the downfield transition band of the band-selective pulse acting on GABA H3.
Although we have focused on discussing chemical shift imaging of spectrally resolved glutamate and GABA without using highly selective editing pulses for the magnetic field strength of 7 T, similar ideas may also be developed for lower magnetic field strengths such as 3 T. For example, multiecho time averaging at 3 T can isolate glutamate with a nearly flat baseline without using any spectral editing pulses (109). Directly combining this approach with conventional phase-encoding for chemical shift imaging would not be feasible due to the large number [e.g., (32)] of echoes required for each phase encoding step. Instead of using evenly spaced echo times, the averaging effect may be obtained using fewer echo times with numerical optimization. Many fast imaging strategies such as echo-planar readout can also greatly accelerate data acquisition of chemical shift imaging experiments.
The coordinated variations of glutamate and GABA across the brain can be assessed using chemical shift images of spectrally resolved glutamate and GABA. The absolute strengths of interregional correlations of spectrally resolved glutamate and GABA have the exciting potential for characterizing excitatory-inhibitory connections among the nodes of neural networks, therefore providing novel FIGURE 4 | Single voxel 7 T spectrum acquired using a single-shot echo time optimized PRESS sequence (112) without signal averaging. Voxel size = 2 × 2 × 2 cm 3 . Echo time = 106 ms. Line broadening = 8 Hz. Number of averages = 1. NAA, N-acetylaspartate; Glu, glutamate; Gln, glutamine; GSH, glutathione; tCr, total creatine; tCho, total choline; mI, myo-inositol. Glutamate H4 at 2.35 ppm was spectrally resolved with an approximately flat baseline.
FIGURE 5 | A proposed scheme for band-selective multiple quantum filtering of GABA. A band-selective refocusing pulse prepares GABA into multiple quantum state while leaving glutathione in single quantum state by avoiding refocusing its cysteinyl a-H at 4.56 ppm. A band-selective 90°pulse converts the double quantum coherence into observable single quantum coherence. The frequency separation between GABA H4 and GABA H3 at 7 T is 164 × 2 = 328 Hz. Asp, aspartate; GSH, glutathione; Cr, creatine. The chemical shifts (in ppm) of the resonance signals are placed below their labels. parameters for gauging interregional excitation-inhibition balance, the disruption of which is implicated in many neuropsychiatric disorders [e.g., (33,103,117,129,130)].

CONCLUSIONS
Previous neurochemical studies of animal models have revealed coordinated changes in major neurometabolites under various pathophysiological conditions which were attributed to the wellstudied metabolic pathways connecting them. Recent MRS findings of neurometabolite correlations in healthy subjects and altered correlations in patients have further corroborated these associations, and their disruption is a hallmark of many neuropsychiatric disorders. To measure the absolute strength of these correlations, it is necessary to use spectral editing techniques to minimize or eliminate statistical correlations among MRS signals that originated from spectral overlap. Finally, chemical shift imaging of spectrally resolved glutamate and GABA is technically feasible at 7 T. It is hoped that the prospects for eliminating the confounding statistical correlations due to spectral overlap will reduce controversies in the field and generate further interest in characterizing local and interregional neurochemical associations especially glutamate-GABA interactions in the brain for studying neuropsychiatric disorders.