ABNORMAL GLUTAMATERGIC NEUROTRANSMISSION AND NEURONALGLIAL INTERACTIONS IN ACUTE MANIA

Subjects

Following approval by McLean Hospital IRB, 22 control, 39 bipolar I disorder, currently manic, and 20 schizophrenia subjects with acute psychosis were recruited (NC, BD and SZ groups, respectively). All patients were hospitalized in a psychiatric inpatient unit. Subjects with significant neurological or medical problems, current substance abuse or history of substance dependence were excluded. All patients had negative urine toxicology tests. We used an Informed Consent Survey which asks 10 simple questions about the study such as “What illness is being studied?” and “If you don’t want to, do you have to be in this study?” Every subject answered these questions correctly, and answers were documented. Patients were assessed using the Structured Clinical Interview for DSM-IV (SCID) [40], Young Mania Rating Scale (YMRS), Montgomery-Asberg Depression Rating Scale (MADRS), and Positive and Negative Syndrome Scale (PANSS) on scan day. The SCID substance abuse/dependence assessment did not include tobacco products; tobacco smokers were not excluded from the study. Healthy control subjects were assessed using the SCID; no family history information was obtained from healthy controls. The demographic and clinical variables and statistical comparisons are in Table 1.

Although severity of mania and psychosis cannot be directly compared, all patients were hospitalized with acute episodes, suggesting that all had pronounced psychopathology. All BD subjects met full DSM-IV criteria for current manic episode; there was no additional cutoff for YRMS scores. No patients met criteria for rapid cycling. Usable data were obtained from 21 NC, 15 BD and 17 SZ (eight diagnosed with schizoaffective disorder, not currently in a mood episode) subjects. Inability to tolerate the scanning environment due to psychiatric condition was the primary reason for dropout (N=22 for BD, 2 for SZ). BD subjects who did not complete the study were more symptomatic than those who did (YMRS:26.9; MADRS:18.4; PANSS:82.0; compare with Table 1). Chlorpromazine (CPZ) equivalents were calculated for patients taking antipsychotic medications[41].

MRI/MRS Scans

As part of one IRB protocol, all subjects were recruited for two MRI scans on the same or consecutive days (depending on scanner availability). Subjects underwent a structural MRI in a Siemens 3 Tesla Trio scanner (Erlangen, Germany) with a quadrature radio-frequency coil using 3D inversion prepared sagittal FLASH sequence. TE/TR/TI times=2.74ms/2.1s/1.1s, echo spacing 6.3ms, 12° flip. Field of view (FOV) was 256mm with 1×1mm pixels, and 1.33 mm slices. Subjects with structural abnormalities were excluded.

¹H MRS acquisitions were conducted on a 4 Tesla full-body MR scanner (Varian/UnityInova, Varian Inc., Palo Alto, CA), using a 16-rung, single-tuned, volumetric-birdcage coil (Robarts Research Institute, London, Canada). 2D gradient-recalled echo images (12s) were acquired in three dimensions to ensure optimal patient positioning. Manual global shimming of unsuppressed water signal yielded an unfiltered global water linewidth of ≤30Hz. T1-weighted sagittal images (TE/TR=6.2/11.4ms, FOV=24×24×8cm, in-plane resolution=0.94×1.88mm, slice thickness=5mm, readout points=512, flip angle=11°) and axial images (similar parameters as above, except for in-plane resolution =0.94×0.94mm, slice thickness=2.5mm) were next acquired, allowing for clear differentiation between cortical grey matter (GM) and white matter (WM), and serving as anatomical guide to positioning MRS voxels.

A 2×2×2cm single voxel was then placed on the ACC. We were interested in studying the subgenual ACC but susceptibility artifacts precluded collecting such data. We therefore placed the voxel, on midsagittal T1-weighted images, anterior to genu of the corpus callosum, with the ventral edge aligned with the dorsal corner of the genu (Figure 1), and positioned the voxel on the midline on axial images. A point-resolved spectroscopy sequence (PRESS) modified for J-resolved ¹H MRS, employing a four-pulse WET (water suppression enhanced through T1-effects) sequence [42] was used. Manual shimming within the voxel resulted in water linewidths ≤ 11 Hz. Following tip angle, water-suppression, and RF pulse power optimization using automated methods, the 2D J-resolved sequence collected 48 TE-stepped spectra from the voxel, with the echo-time ranging from 30ms to 500 ms in 10ms increments. Specifically, the second TE period in the PRESS sequence was progressively prolonged by symmetrically flanking the last 180 degree pulse prior to signal acquisition with two 5 millisecond delay periods. Acquisition parameters were: TR=2.0s, acquisition bandwidth=2kHz, repetitions=16; nominal voxel volume=8cc; approximate scan duration=28 minutes.

This process was repeated on a 2×2×2cm POC voxel. On midsagittal images we aligned the ventral edge of the voxel with the dorsal corner of the splenium of the corpus callosum, covering posterior cingulate and retrosplenial cortices, and positioned the voxel on the midline in axial images (Figure 1). Because POC spectral quality was superior, 8 repetitions were obtained in approximately 14 minutes. Total time in the magnet was 75–90 minutes.

The water resonance full width at half maxima were 9.9±1.2 and 8.7±0.8Hz for NC, 9.3±1.5 and 8.6±0.7Hz for BD and 9.4±2.3 and 8.9±0.9Hz for SZ subjects (in ACC and POC, respectively; no between-group differences). Coil loading was not quantified; this affects all metabolites similarly, and would not modify the Gln/Glu ratio.

MRS Data Processing and Analysis

All MRS processing was blinded to diagnosis. Real (not magnitude) spectra were used (Figure 2). For each voxel, the 48 TE-stepped free-induction decays (FIDs) were zero-filled to 128 FIDs in f1, Gaussian-filtered to minimize residual ringing from longer TE NAA and Cr signal, and Fourier-Transformed in f1 to produce a series of 128 J-resolved spectral FIDs over 100Hz bandwidth. A blinded MR physicist (JEJ) excluded spectra from 5 subjects due to low signal-to-noise and/or spectral resolution: 2 ACC and 1 POC in NC, 1 POC in BD and 1 ACC in SZ.

We used the commercial spectral-fitting package, LCModel (version 6.0–1) [43] for 2D fitting. 2D fitting improves spectral fitting by utilizing all available signal rather than fitting select single J-resolved extractions [35, 39]. For 2D fitting, LCModel utilized GAMMA-simulated [44] theoretical J-resolved basis sets to fit every J-resolved spectral extraction. We generated basis sets for N-acetylaspartate, N-acetylaspartylglutamate, Alanine, Gamma-aminobutyric acid, Aspartate, Choline, Creatine, Glucose, Glutamate, Glutamine, Glutathione, Glycerophosphocholine, Glycine, Myo-inositol, Scyllo-inositol, Lactate, Phosphocreatine, Phosphocholine, Serine, and Taurine. We first used GAMMA to generate 48 theoretical, TE-stepped spectra ranging from 30ms to 500ms in 10ms increments [20]. Each GAMMA spectrum was modeled at 2kHz spectral bandwidth with 1024 complex pairs and a 2Hz Lorentzian lineshape, zero and first-order phased with no baseline roll. We also modeled a formate peak at 8.45ppm and a TSS reference peak at 0.0ppm to mimic the standard LCModel stock solution required for basis-set generation [45]. For each GAMMA-simulated metabolite TE-series, we zero-filled complex time-points in each FID to 128 TE points, apodized with an exponential filter to approximate metabolite T2-decay at 4T, and Fourier-Transformed in the TE dimension to obtain basis sets sampled at nominal 0.78Hz J-resolution for 100Hz bandwidth. These J-resolved metabolite FIDs were converted into 128 separate LCModel basis sets for each metabolite. This process resulted in artifact-free and theoretically-correct J-resolved basis sets for each J-resolved spectral extraction from in vivo spectra. For every metabolite, we fit its spectrum over 50Hz bandwidth (−25Hz to +25Hz; 64 TE points), deriving the integrated area under the 2D-surface. Spectral regions outside this range did not contain significant metabolite signal but contributed additional noise.

LCModel provides Cramer-Rao Lower Bounds (CRLBs), an estimate of the variance associated with fitting. In our case, each 2D spectrum produced 64 CRLBs for each metabolite. These ranged from less than 5% in regions of high certainty (typically close to J=0Hz; Figure 3), to greater than 50% in regions of low certainty. We report the mode, median and mean CRLBs for each metabolite from the clinical study (Table 2). The results of LCModel fitting for each metabolite are reported as arbitrary units (AUs). Glu and Gln measures in the test-retest study are shown in Figure 4.

Image Segmentation

Tissue-segmentation of T1-weighted images into GM, WM, and cerebrospinal fluid (CSF) used FMRIB's Automated Segmentation Tool (Oxford, UK). Poor image quality precluded segmentation in 30 voxels. The percentage of GM in ACC and POC was 79±5% and 70±8% (NC), 76±8% and 69±8% (BD), and 77±3% and 69±7% (SZ),while that of WM was 17±6% and 28±8% (NC), 16±8% and 29±8% (BD), and 16±4% and 29±8% (SZ), respectively.

For the 73 voxels where data were available, we computed Pearson’s R for percent GM and MRS measures. GM was positively correlated with all MRS measures (NAA: R=0.417, p=0.002; Cr: R=0.586, p<0.001; Cho: R=0.563, p<0.001; mI: R=0.574, p<0.001; Glu: R=0.555, p<0.001; Gln: R=0.270, p=0.046), consistent with prior literature [46–48]. No relationship existed between Gln/Glu and GM (R=−0.009, p=0.947), suggesting that Glu and Gln levels covary between GM, WM, and CSF.

Statistical Analysis

Demographic and clinical variables were compared across groups using analysis of variance (ANOVA) and chi-square tests (Table 1). There was no effect of sex on any metabolite measures. A series of correlation coefficients (uncorrected for multiple comparisons) showed that age, age at disease onset, and CPZ equivalents were not related to metabolite measures, except for a negative correlation between age and NAA levels (R=−0.208, p=0.048).

Given our a priori hypothesis, we considered the Gln/Glu ratio as our primary outcome measure, and did not analyze Glu or Gln separately. In addition, we examined NAA, Cr, Cho and mI levels. The effect of diagnosis on metabolite measures was investigated using repeated measures ANOVAs with brain region (ACC vs. POC) as within-subjects factor, and diagnosis as between-subjects factor. Main effects and the brain region × diagnosis interaction were explored and Dunnett’s post-hoc tests were used. There was a statistical trend difference in age between diagnostic groups (Table 1). Therefore, we repeated the metabolite analyses with age as covariate.

A series of correlation coefficients showed no relationship between metabolite measures and MADRS, YMRS, and PANSS scores (uncorrected for multiple comparisons). Finally, we examined medication effects on metabolite levels using one-way ANOVAs. This was done in BD and SZ groups separately because different numbers of subjects took each medication in the SZ and BD groups and group differences in metabolite levels could confound the medication effect.

Medication and Subgroup Effects

Since all BD subjects and all but one SZ subject were taking second generation antipsychotics, we could not examine the effect of these medications on metabolite measures. Gln/Glu ratios and metabolite levels were no different between patients who were taking benzodiazepines and those who were not (F<1.308, p>0.266 for BD and F<1.294, p>0.266 for SZ). Similarly, grouping by treatment with anticonvulsants (F<4.120, p>0.055 for BD and F<2.300, p>0.142 for SZ) or lithium (F<1.243, p>0.277 for BD and F<2.157, p>0.155 for SZ) had no effect on any measure (Figure 6).

Excluding the 8 patients with schizoaffective disorder from the SZ group, there was still a significant main effect of diagnosis on Gln/Glu (F(30,2)=6.925, p=0.003) and a significant Gln/Glu elevation in BD but not SZ (Dunnett’s 2-sided comparison to NC, p=0.002 and p=0.323, respectively). In a one-way ANOVA, Gln/Glu and metabolite levels were no different between 8 schizoaffective disorder and 9 schizophrenia patients (p-values between p=0.624 (Cho) and p=0.890 (NAA)).

Glutamatergic Neurotransmission in Bipolar Disorder

Although glutamate plays a crucial role in psychiatric disorders, brain Glu levels do not reflect glutamatergic activity because Glu is also involved in cellular metabolism, and because static Glu measures could mask dynamic changes in Glu synthesis/release/metabolism. Simultaneous Glu and Gln measurements better capture such changes, but even these measures cannot provide direct evidence of synaptic abnormalities because brain Glu and Gln are mostly intracellular. Nonetheless, specific changes in Gln/Glu taken together with known abnormalities in BD may be valuable indicators of glutamatergic function.

Our finding of elevated Gln/Glu ratio in mania is similar to that reported in animal ischemia models where excessive Glu release leads to excitotoxicity [14, 15], and in first-episode schizophrenia where it may indicate “greater than normal glutamatergic activity” [32]. In mania, cerebral metabolic rate elevations accompany racing thoughts, distractibility, irritability, and insomnia [49] and our findings are consistent with excessive glutamatergic neurotransmission driven by neuronal activity. In this context, the balance between Glu-Gln conversion in glia and Gln-Glu conversion in neurons appears disrupted due to a breakdown in neuronal-glial interactions. We cannot pinpoint the site of pathology, but well-documented glial number reductions in BD are interesting [3, 50] because glial cells modulate neuronal Glu release, and provide Gln to neurons for Glu synthesis [51]. The transport of Gln from glial cells to neurons is poorly understood, but may involve a neuronal Gln transporter [52].

We report elevated Gln/Glu in mania, while others have found lower Gln and higher Glu in depression [30]. The modulation of cortical Gln/Glu by mood state suggests that BD is characterized not by a simple excess or deficiency in glutamatergic signaling, but by an inability to maintain homeostasis. Therefore, therapies reducing variability in glutamatergic neurotransmission, e.g. by stabilizing Glu release through actions on presynaptic metabotropic Glu receptors may be as effective in BD as in SZ [53].

Importantly, our findings are consistent with our a priori hypothesis of abnormal glial-neuronal interactions in BD, but by no means provide definitive evidence for it. Elevated Gln/Glu ratios could arise from intermediary metabolism abnormalities where both molecules play a role, or from intraneuronal abnormalities, e.g. in glutaminase which converts Gln to Glu.

Neurochemical Profile of Schizophrenia and Mania

Although there is postmortem, biochemical and clinical evidence of glutamatergic abnormalities in schizophrenia [31], we found only a trend toward alterations in brain Gln/Glu ratios in the SZ group. Several factors are relevant: first, our work focused on one specific aspect of glutamatergic neurotransmission, Gln/Glu ratio. This is not an index of total glutamatergic function; rather a specific probe of Glu-Gln cycling. Therefore, significant abnormalities in glutamatergic function in schizophrenia may still exist at the level of the NMDA receptor or in second messenger signaling without any alteration in Gln/Glu ratios. Second, previous studies reported that first-episode schizophrenia patients have elevated Gln but normal Glu levels in ACC [32], and that chronically ill patients show lower than normal Gln and Glu levels [33]. Our finding of no statistical Gln/Glu change in SZ is consistent with this literature because we studied chronically ill patients. It is intriguing that several SZ subjects had elevated Gln/Glu ratios (Figure 5) but correlation of illness duration with Gln/Glu in SZ was not significant (not shown). Studies explicitly designed to test the effect of illness duration on Gln/Glu are needed.

Both BD and SZ showed lower than normal NAA in the ACC, although the effect was statistically significant only in the SZ group. NAA levels reflect neuronal integrity and function, and reductions are reported in several brain regions in BD and SZ [54–56]. Our replication reassures us that our cohort and methodology were consistent with previous studies. By contrast, we did not identify changes in Cho or mI. Abnormalities in both of these metabolites have been reported in some but not all BD studies [57, 58]. Possible explanations for discrepancies include subjects’ clinical status (mania vs. euthymia/depression), medication effects, and differences in data acquisition/analysis.

Strengths and Limitations

This study has several strengths: high field strength, novel data acquisition and analysis techniques, and reliable metabolite measures. We evaluated the improvements in Glu and Gln quantification afforded by our approach in two ways: first, we demonstrated that our measures were reliable in a test-retest study. Our reliability measures were comparable to those previously reported for Cr, Cho, NAA, and mI, and improved upon those reported at 1.5T and 4T for Glu and Gln [59–61]. To our knowledge, no other BD or SZ study has reported such reliability estimates. Second, we examined CRLBs in patient data. CRLBs are considered excellent <20% [62]. In our study, mean CRLBs were <20% for all metabolites except for mI (POC) and Gln (both voxels) (Table 2). Both mI and Gln had lower mode/median than mean CRLBs, indicating that the mean was inflated by outliers. The CRLBs obtained for Gln in our study are lower than those in a recent in vivo human spectroscopic imaging study at 3T/4T (mid-30%) [46] and a single voxel study at 3T (>35%) [63]. Another 2D MRS study at 3T obtained a mean Gln CRLB of 14% from 27 healthy volunteers in a 15.625cc voxel [36]. Thus, our quantification compares favorably with other studies.

One limitation in this study is that all patients were taking medications. Psychotropic medications have complex and variable effects on glutamatergic neurotransmission, including on Glu receptors [64, 65]. For example, 3.6 months of lithium but not valproate treatment in depressed BD outpatients caused a reduction in GM Glx [66]. In animal studies, lithium and valproate stimulate Glu release and inhibit uptake, although chronic treatment has opposite effects [67, 68]. We found that lithium, anticonvulsants, or benzodiazepines did not impact MRS measures but this is preliminary due to small N. As for antipsychotic medications, 8 BD patients were taking risperidone, 4 olanzapine, 3 quetiapine, 2 aripiprazole, and 1 ziprasidone (3 patients on combinations). Five SZ patients were taking olanzapine, 4 clozapine, 3 aripiprazole, 2 risperidone, 2 quetiapine, and 1 haloperidol. Although many more BD patients were taking risperidone, all had Gln/Glu ratios similar to others with their diagnosis (BD=0.586, SZ=0.412; see Table 3). Likewise, 5 SZ patients took clozapine (Gln/Glu=0.589), but this did not explain our findings. Although we uncovered no evidence that psychotropic medications affect Gln/Glu ratios, it is likely that medications modulate glutamatergic neurotransmission [69]. Two factors may obscure this effect: first, the sample size in most neuroimaging studies is small and interindividual variability overwhelms all but prominent diagnosis effects. Medication-effect studies are needed to examine these changes [66]. Second, we scanned acutely ill subjects, and 7 BD and 4 SZ patients were unmedicated prior to hospitalization. Therefore, medication effects may be weaker than diagnosis effects.

We did not utilize absolute quantification; such technically demanding and time-consuming attempts were not feasible in this study. We also avoided internal referencing to Cr because of the recent reports suggesting Cr abnormalities in BD and SZ [70, 71]. Finally, internal referencing to the water resonance [72] was not possible because we did not acquire water unsuppressed spectra in this study.

Grouping patients with schizophrenia and schizoaffective disorder together is another limitation. Although our analyses revealed no differences in MRS measures between these diagnoses, future studies may find such differences. Finally, because GABAergic interneurons are abnormal in BD and SZ [73, 74], measuring GABA levels would be important. While J-resolved MRSI allowed quantification of GABA previously [39], this measure did not meet quality-control criteria in our study, possibly because of reduced information in our two-voxel design. GABA-dedicated acquisition protocols may provide further information in this regard [22, 28].