Post-inflammatory behavioural despair in male mice is associated with reduced cortical glutamate-glutamine ratios, and circulating lipid and energy metabolites

Post-inflammatory behaviours in rodents are widely used to model human depression and to test the efficacy of novel anti-depressants. Mice injected with lipopolysaccharide (LPS) display a depressive-like phenotype twenty-four hours after endotoxin administration. Despite the widespread use of this model, the mechanisms that underlie the persistent behavioural changes after the transient peripheral inflammatory response remain elusive. The study of the metabolome, the collection of all the small molecule metabolites in a sample, combined with multivariate statistical techniques provides a way of studying biochemical pathways influenced by an LPS challenge. Adult male CD-1 mice received an intraperitoneal injection of either LPS (0.83 mg/kg) or saline, and were assessed for depressive-like behaviour 24 h later. In a separate mouse cohort, pro-inflammatory cytokine gene expression and 1H nuclear magnetic resonance (NMR) metabolomics measurements were made in brain tissue and blood. Statistical analyses included Independent Sample t-tests for gene expression data, and supervised multi-variate analysis using orthogonal partial least squares discriminant analysis for metabolomics. Both plasma and brain metabolites in male mice were altered following a single peripheral LPS challenge that led to depressive-like behaviour in the forced swim test. The plasma metabolites altered by LPS are involved in energy metabolism, including lipoproteins, glucose, creatine, and isoleucine. In the brain, glutamate, serine, and N-acetylaspartate (NAA) were reduced after LPS, whereas glutamine was increased. Serine-modulated glutamatergic signalling and changes in bioenergetics may mediate the behavioural phenotype induced by LPS. In light of other data supporting a central imbalance of glutamate-glutamine cycling in depression, our results suggest that aberrant central glutaminergic signalling may underpin the depressive-like behaviours that result from both inflammation and non-immune pathophysiology. Normalising glutaminergic signalling, rather than seeking to increase serotonergic signalling, might prove to be a more coherent approach to the development of new treatments for mood disorder.


Section 1: 1 H NMR Spectroscopy
Plasma samples: Plasma samples were defrosted on ice and centrifuged at 17,000 xg for 5 min at 4 °C. An equal volume of plasma was aliquoted (100 μL) into a fresh tube, and diluted to 600 μL in a 75 mM phosphate buffer (5:1 disodium phosphate Na2HPO4, monosodium phosphate NaH2PO4 in 100% D2O, pH 7.4). The volume of plasma used was limited by the volume of plasma collected, and the volume chosen was the maximum volume available for 90% of samples.
Samples with insufficient plasma volume were excluded from the analysis.
Samples were then transferred to 5 mm Borosilicate Glass NMR tubes (Norrell). 1 H NMR spectra were acquired using a 700-MHz Bruker AVII spectrometer operating at 16.4 T equipped with a 1 H (13C/15N) TCI cryoprobe. Sample temperature was stable at 310 K. 1 H NMR spectra were acquired using a one dimensional (1D) Nuclear Overhauser Effect Spectroscopy (NOESY) presaturation scheme for attenuation of the water resonance with a 2 s presaturation.
An additional sequence, the spin-echo Carr-Purcell-Meiboom-Gill (CPMG) sequence, was used for plasma samples to suppress broad signals arising from large molecular weight plasma components with a τ interval of 400 μs, 80 loops, 32 data collections, an acquisition time of 1.5 s, a relaxation delay of 2 s, and a fixed receiver gain. CPMG spectra provide a measurement of small molecular weight metabolites and mobile side chains of lipoproteins in the plasma sample and were used for all further analysis of plasma samples.

Section 2: NMR Data Processing
Processing methods were adapted from published parameters (Jurynczyk et al., 2017;Probert et al., 2018). NMR spectra were imported into MestreNova (Mestrelab Research, Spain) and each spectrum was then processed manually with phase 0 (PH0) correction, baseline correction (Bernstein polynomial fit, order = 3), and referencing to an added standard (TSP referenced to δ0) for brain tissue, and an internal standard (Lactate referenced to δ1.33) for plasma. The individual spectra were then stacked, and binned (sum method, width of each integral region = 0.02ppm). Binning refers to a function where the whole spectrum is divided into bins of equal width, and all the peaks in each bin is integrated to obtain a value representing the area of all the peaks in a bin. Binned values were then exported as a spreadsheet (.xlsx) for further analysis.
Inter-individual variation was reduced by total area normalization (each bin normalized to a ratio of the individual with the lowest total area over total area of the individual) to account for any dilution error. Brain tissue samples were also further normalized to TSP (each bin normalized to a ratio of the individual with the lowest TSP area over TSP area of the individual). Finally, noise areas were removed by two methods. First, spectral relative standard deviation (RSD) values were calculated for each bin, and bins with RSD greater than 100 were removed (Parsons et al., 2009). Second, the addition of the mean and standard deviation was calculated for each bin, and the average taken for a noise region. Bins with a value lower than that of the noise region would be removed. Broadly, areas that were removed include the water peak, regions before 0.7ppm, regions after 9.38ppm, noise region 5.0 to 6.0ppm, and contamination EDTA peaks for plasma samples.

Section 3: NMR Data Analysis
Preliminary exploratory analysis: Normalized bin values were imported into SIMCA (Umetrics, Sweden). Outliers were identified with principal component analysis (PCA) scores plots with pareto scaling (van den Berg et al., 2006). Pareto scaling removes the bias given to large peaks without inflating the noise, thus allowing the detection of changes in small and medium-sized peaks. Supervised multi-variate analysis was conducted using orthogonal partial least squares discriminant analysis (OPLS-DA), which attempts to find a linear relationship between a predictor matrix (spectrometric bin values) and a response matrix (treatment groups) (Triba et al., 2015).
Model building and validation: OPLS-DA models were built in R 3.3.2 (R Core Team, 2016) using the ROPLS package (Thévenot et al., 2015) and an in-house R script using 10-fold cross validation. The total number of samples were divided into 10 groups, with nine groups used to build the model (training set) and the last group used to test the model (testing set). For each iteration of the script, each of the 10 groups was used as the testing set once, thus 10 models were built for each iteration. The 10-fold cross validation was then repeated for a total of 100 iterations, producing an ensemble of 1000 models.
The main outcome of interest is predictive accuracy of the model, reported as mean predictive accuracy of the 1000 models built with the standard error mean (standard deviation/√100 where 100 is the number of iterations). This determines how accurate the model is at predicting which group a sample in the test set belongs to, thus assessing the discriminatory power of the model.
Other parameters included the average specificity of the models, sensitivity of the models, average Q 2 values, R 2 X values, and R 2 Y values. The Q 2 measures the internal predictive ability of the model (the accuracy of the model on the training set) and is used to optimise the model, while the R 2 measures the goodness of fit of the model (Triba et al., 2015), with R 2 X and R 2 Y measuring the fraction of the variation explained by the model of the X and Y variables respectively.
To validate the OPLS-DA models, the same cross-validation process was conducted with samples randomly assigned to treatment groups. This ensemble of OPLS-DA models, representing the null distribution, was used to calculate the accuracy achieved by random chance. If the true OPLS-DA models performed significantly better than the permutation test, then the discrimination observed was unlikely to have occurred by chance and thus the results were considered significant. For significant predictive models, the variable importance in projection (VIP) scores were used to identify the key bins that were important for building the model. Bins with high VIP scores were considered to be significantly different between the treatment groups.
Metabolite identification and direction of change: Metabolites were assigned to peaks in bins with high VIP scores through a combination of literature values (Govindaraju et al., 2000;Misra & Bajpai, 2009), reference to the human metabolome database (HMDB) (Wishart et al., 2013), and confirmation with two-dimensional (2D) correlation spectroscopy (COSY). COSY provides confirmation that peaks occupying different positions on the spectrum belong to the same metabolite through cross-peaks that show a correlation between signals. The direction of change between groups was also determined by comparing means in SPSSv20 (independent samples T-test).  . On all counts (Accuracy, Q 2 , Sensitivity, Specificity, R 2 X, and R 2 Y), the OPLS-DA model was significantly better. *** *** *** ***