Abstract

   Bipolar disorder is a major mental illness characterized by severe
   swings in mood and activity levels which occur with variable amplitude
   and frequency. Attempts have been made to identify mood states and
   biological features associated with mood changes to compensate for
   current clinical diagnosis, which is mainly based on patients’
   subjective reports. Here, we used infradian (a cycle > 24 h) cyclic
   locomotor activity in a mouse model useful for the study of bipolar
   disorder as a proxy for mood changes. We show that metabolome patterns
   in peripheral blood could retrospectively predict the locomotor
   activity levels. We longitudinally monitored locomotor activity in the
   home cage, and subsequently collected peripheral blood and performed
   metabolomic analyses. We then constructed cross-validated linear
   regression models based on blood metabolome patterns to predict
   locomotor activity levels of individual mice. Our analysis revealed a
   significant correlation between actual and predicted activity levels,
   indicative of successful predictions. Pathway analysis of metabolites
   used for successful predictions showed enrichment in mitochondria
   metabolism-related terms, such as “Warburg effect” and “citric acid
   cycle.” In addition, we found that peripheral blood metabolome patterns
   predicted expression levels of genes implicated in bipolar disorder in
   the hippocampus, a brain region responsible for mood regulation,
   suggesting that the brain–periphery axis is related to
   mood-change-associated behaviors. Our results may serve as a basis for
   predicting individual mood states through blood metabolomics in bipolar
   disorder and other mood disorders and may provide potential insight
   into systemic metabolic activity in relation to mood changes.

Introduction

   Mood naturally changes over time with variable amplitude and frequency,
   sometimes in an infradian (a cycle > 24 h) manner [[41]1, [42]2]. This
   is often accompanied by changes in physical activity [[43]3]. In
   bipolar disorder, unusual swings in mood and activity levels from
   depressive to manic states are core phenomenological and clinical
   features. Such exaggerated fluctuations in mood and the related
   activity can cause significant distress and/or social and occupational
   impairment, often leading to high direct and indirect health care costs
   [[44]4, [45]5]. Currently, the clinical assessment and management of
   these conditions mainly rely on clinicians’ interview and patients’
   subjective description. To overcome the potential subjective biases,
   developing objective and quantitative measures is expected to
   compensate for the current procedures. To this end, substantial efforts
   have been invested into predicting mood states of patients with bipolar
   disorder based on biological features, such as heartbeat [[46]6],
   spontaneous speech [[47]7], and neural activity measured with
   functional magnetic resonance imaging [[48]8]. Meanwhile, molecular
   omics approaches, such as metabolomics and proteomics, with which
   hundreds of molecules can be targeted simultaneously, are being applied
   to gain more complete profiles of a wide range of biological conditions
   and diseases [[49]9, [50]10]. However, such molecular omics-based
   approaches have not been well examined to predict mood states. While
   mood disorders are considered to be brain diseases, the use of
   peripheral samples has been desirable for clinical purposes, as the
   opportunity to obtain human brain tissues is limited [[51]11].

   Previous studies have attempted to detect biomarkers for mood disorders
   in the blood by comparing the metabolomes of patients to control
   subjects [[52]11–[53]13]. It is still not well understood whether or to
   what extent metabolomic alterations in peripheral blood are related to
   the shift of mood states. This may, in part, be due to the lack of
   animal models that exhibit behavioral phenotypes associated with
   infradian changes in mood states. We previously found that mice with
   the heterozygous knockout of the alpha-isoform of
   calcium/calmodulin-dependent protein kinase II (Camk2a^+/− mice)
   exhibit various dysregulated behaviors, including exaggerated cyclic
   mood-change-associated activity in an infradian manner, in which
   locomotor activity (LA) spontaneously and recurrently increases and
   decreases in a period of approximately 10–20 days [[54]14, [55]15].
   Importantly, changes in the LA were correlated with changes in
   depression-like and anxiety-like behaviors [[56]14], suggesting that
   Camk2a^+/− mice are useful as an animal model which shows similar
   infradian oscillations of mood to those found in patients with bipolar
   disorder. CAMK2A is one gene suggested to be a candidate for bipolar
   disorder [[57]16–[58]18], and decreased CAMK2A mRNA has been found in
   the prefrontal cortex of patients with the disorder [[59]19].
   Camk2a^+/− mice also harbor brain endophenotypes relevant to bipolar
   disorder, such as neuronal hyperexcitability [[60]20] and immaturity
   [[61]15, [62]21–[63]23] in the hippocampus and decreased brain pH
   [[64]24]. These findings suggest that Camk2a^+/− mice provide a model
   for bipolar disorder which has good face and construct validity
   [[65]25].

   Using the Camk2a^+/− mice, we previously showed that gene expression
   patterns in the hippocampus, a brain region implicated in mood
   regulation, could retrospectively predict the LA level of individual
   mice by using statistical learning algorithm [[66]14]. We used
   cross-validated multivariate regression models to ensure the
   generalization ability of our prediction models. Generalization ability
   is the capacity to predict unseen samples, and ensuring and improving
   the generalization ability of models are believed to be crucial for
   practical use, including identification of subpopulations among
   patients with mood disorders [[67]26]. In the present study, using this
   strategy, we sought to determine whether metabolome patterns in
   peripheral blood can predict individual LA levels of the mice, which
   would have potential for future translational applications. We also
   investigated whether gene expression levels in the brain can be
   predicted by peripheral blood metabolome patterns, aiming to gain
   insights into the link between systemic metabolic pathways and the
   regulation of brain gene expression in relation to the infradian mood
   changes.

Materials and methods

Animals

   Thirty-seven adult (> 8 weeks old) male Camk2a^+/− mice were used
   [[68]14, [69]15, [70]27]. Every effort was made to minimize the number
   of animals used.

Locomotor activity monitoring and blood sampling

   The LA data of Camk2a^+/− mice analyzed in this study were the same as
   was used in a previous study [[71]14]. Mice were singly housed with a
   12 h light/dark cycle (lights on at 7:00 a.m.) and access to food and
   water ad libitum. The monitoring of LA and blood sampling was performed
   as previously described [[72]14]. Briefly, LA was monitored for
   72–82 days through a system that automatically analyzes the distance
   traveled by a mouse in its home cage [[73]15]. Peripheral blood and
   brain tissue [[74]14], was collected at zeitgeber time (ZT) 6–7 (ZT0,
   lights on; ZT12, lights off) from mice with short or long distances
   traveled (assessed by distance traveled during the 24 h before ZT0 on
   the sampling day; Fig. [75]1a, b). In this way, mice were selected for
   the sampling such that their 24 h LA levels varied among the 37 mice
   (Fig. [76]1c). The 24 h LA was defined as distance traveled during the
   24 h between ZT0 on the day before the sampling and ZT0 on the day of
   sampling. The 3 h LA was defined as distance traveled in every 3 h
   window before sampling (ZT6) (Fig. [77]1b).

Fig. 1.

   [78]Fig. 1
   [79]Open in a new tab

   Experimental workflow. a LA was longitudinally monitored, and blood was
   taken from each mouse ZT6–7. b 24 h LA, distance traveled during the
   24 h before ZT0 on the sampling day; 3 h LAs, distances traveled of
   every 3 h window before sampling (ZT6). Black and white bars indicate
   dark and light period, respectively. c Variations in 24 h LA. Each dot
   indicates an individual mouse. Solid lines indicate average 24 h LA and
   dashed lines indicate the standard deviations for 35 mice. LA,
   locomotor activity; ZT, zeitgeber time

   Mice were removed from the home cage, and immediately euthanatized by
   cervical dislocation. Blood was collected from the neck in tubes with
   Na-heparin and centrifuged (2200×g at 4 °C for 10 min) to prepare
   plasma. Plasma samples were stored at − 80 °C until use.

Gas chromatography mass spectrometry (GC/MS) analysis of blood samples

   GC/MS analysis was performed as described previously [[80]28] using a
   GCMS-QP2010 Ultra (Shimazu, Kyoto, Japan) with a fused silica capillary
   column (CP-SIL 8 CB low bleed/MS; 30 m × 0.25 mm inner diameter, film
   thickness 0.25 μm; Agilent Technologies, Palo Alto, CA, USA). Data
   processing was performed as described previously [[81]28].
   2-isopropylmalic acid was used as an internal standard. Two mice whose
   metabolite contents could not be measured due to unknown reasons were
   excluded from further analyses. Hence, LA and metabolome data from 35
   mutant mice were processed for correlation and prediction analyses
   (Additional file [82]2: Table S1), as described below.

Brain gene expression data

   Expression microarray data of the hippocampal dentate gyrus [[83]29]
   was obtained from the 35 mice mentioned above. We used data that has
   been deposited at the Gene Expression Omnibus (GEO; accession number
   [84]GSE68293) [[85]14]. The log[2]-transformed expression values were
   used for the prediction analyses.

Construction of models to predict LA

   The linear regression algorithm was used to predict LA (24 h LA and 3 h
   LAs; Fig. [86]1) and gene expression levels of individual mice from
   peripheral blood metabolome patterns using MATLAB [[87]14]. The
   prediction accuracy was evaluated by the leave-one-out cross-validation
   method and the feature selection for the prediction was conducted by
   the nested cross-validation method. Calculation details are included in
   the Additional file [88]1: Supplementary Materials and Methods.

Pathway enrichment analysis

   To determine the characteristics of metabolites of interest, we used
   the default settings in MetaboAnalyst (version 4.0;
   [89]http://www.metaboanalyst.ca/) [[90]30], a comprehensive, web-based
   tool for metabolomics analysis and interpretation. Some feature names
   that had not been recognized in the query were modified if necessary,
   as suggested in the instructions.

Results

Locomotor activity can be predicted from blood metabolomic profiles

   Of the 106 metabolites tested, the concentrations of 16 features were
   correlated with 24 h LA (P < 0.05, Pearson’s correlation; Additional
   file [91]1: Figure S1). None of these survived false discovery rate
   correction for multiple tests (FDR; q value < 0.1). To determine
   whether the metabolome patterns in peripheral blood could
   retrospectively predict the 24 h LA of individual mice, we performed
   linear regression analysis with the nested cross-validation method
   [[92]14]. Independent sets of mice were used for feature selection from
   entire set of 106 metabolites and model fitting in the inner
   cross-validation loop and testing the model performance in the outer
   cross-validation loop. Statistical evaluation of the prediction
   accuracy of the model revealed a significant correlation between the
   actual and the predicted 24 h LA (Fig. [93]2a), indicating that the
   metabolome patterns in the peripheral blood quantitatively predict the
   24 h LA of individual mice. Thirty-three out of the 106 metabolites
   were selected to build the successful prediction model (Fig. [94]2b,
   Additional file [95]1: Figure S1).

Fig. 2.

   [96]Fig. 2
   [97]Open in a new tab

   Metabolome patterns in the peripheral blood can predict 24 h LA in
   Camk2a^+/− mice. a Scatter plot showing a significant correlation
   between predicted and actual 24 h LA (N = 35 mice). b Feature
   preference for constructing the 24 h LA prediction model. c Results of
   pathway enrichment analysis for the metabolites used for constructing
   24 h LA prediction model. The statistically enriched terms with raw
   P-value below 0.05 are shown. LA, locomotor activity

Mitochondrial alterations implicated in infradian oscillatory LA

   The prediction algorithm we used identified metabolome signatures
   related to infradian oscillatory LA by weighting metabolites according
   to their individual predictive strength (Fig. [98]2b). Next, we
   examined the weighted metabolites that were selected to predict the
   24 h LA to gain insight into metabolic alterations related to changes
   in infradian oscillatory LA. Pathway enrichment analysis using a
   bioinformatics tool MetboAnalyst [[99]30] revealed that the set of
   metabolites exhibited enrichments in the Warburg effect, the citric
   acid cycle, and mitochondria-related pathway. These survived multiple
   comparison correction (FDR q value < 0.05; Fig. [100]2c).
   Interestingly, all the terms were to mitochondrial functions (see
   Discussion), suggesting that mitochondrial function may change
   concomitantly with infradian oscillatory LA.

Peripheral blood metabolome predicts LA levels for longer time periods

   To investigate whether metabolome patterns in peripheral blood could
   predict the LA of the past several days, we constructed models for
   predicting the LA within the 3 days before sampling using a sliding
   time window (window size: 3 h, step size: 3 h; yielding 24 time
   windows) (3 h LAs; Fig. [101]1b). The actual and the predicted 3 h LAs
   were similar within the past 3 days in most mice, while differences
   between them were apparent in some mice (Fig. [102]3a, Additional file
   [103]1: Figure S2). Overall, statistical evaluation of prediction
   accuracy of the models detected significant correlations between the
   actual and the predicted 3 h LAs in 5 of the 24 time windows after FDR
   correction (time windows of 6–9 h: r = 0.46, P = 0.0049; 9–12 h:
   r = 0.49, P = 0.0025; 21–24 h: r = 0.41, P = 0.016; 51–54 h: r = 0.41,
   P = 0.016; 63–66 h: r = 0.51, P = 0.0017; Fig. [104]3b–g). The oldest
   time window with a significant correlation was from 63 to 66 h before
   sampling. These results suggest that multivariate patterns of the
   metabolome in the peripheral blood hold information about LA of at
   least the past 3 days. Out of the five time windows mentioned above,
   metabolites selected were similar in 3 h LA predictions of 6–9 h and
   9–12 h time windows (as well as the 24 h LA prediction), and differed
   in other 3 h time windows (Additional file [105]1: Figure S3).
   Metabolites selected for the prediction of the oldest 3 h time window
   (63–66 h LA) were associated with aspartate-related pathways
   (unadjusted P < 0.05; Additional file [106]1: Figure S4), which
   differed from those for 24 h LA prediction (Fig. [107]2c).

Fig. 3.

   [108]Fig. 3
   [109]Open in a new tab

   Prediction of 3 h LAs during the 3 days before sampling. a The
   prediction results of 6 mice are shown; the results of the remaining 29
   mice are shown in Additional file [110]1: Figure S2. b Correlation
   coefficients between the actual and the predicted 3 h LAs in each time
   window. Dashed lines indicate P value of 0.05 and asterisks indicate
   significant correlations between actual and predicted 3 h LAs after FDR
   correction (q value < 0.1). Black and white bars indicate dark and
   light periods, respectively. c–g Scatter plots of predicted and actual
   3 h LAs of each mouse at the time windows indicated in b. LA, locomotor
   activity

Hippocampal expression levels of the Arntl gene can be predicted from
peripheral blood metabolome

   The regulation of mood is thought to involve the hippocampus [[111]31,
   [112]32] and hippocampal expression of some circadian genes are closely
   related to infradian oscillation of LA in Camk2a^+/− mice [[113]14].
   Consequently, using a prediction approach, we asked whether metabolome
   patterns in peripheral blood are related to hippocampal expression of
   circadian genes in relation to mood-change-associated activity. We used
   microarray-based gene expression data in the hippocampus of Camk2a^+/−
   mice that we had previously obtained from the same mice whose blood
   metabolome data was analyzed above [[114]14]. Cross-validated
   prediction models based on the blood metabolome data were constructed
   to predict expression levels of genes, focusing on seven
   circadian-related genes (Lonrf1, Cys1, Hist1h1c, Tef, Ak4, Arntl, and
   Sfpq), which were among the genes that were selected to predict the
   24 h LA of Camk2a^+/− mice [[115]14]. Of the seven genes tested, Arntl
   and Sfpq expression levels could be predicted by blood metabolome
   patterns, which was determined by correlations between actual and
   predicted expression levels (P < 0.05), and Arntl survived correction
   for multiple testing (Fig. [116]4a, b). Metabolites used for the
   successful prediction of hippocampal Arntl expression levels showed
   enrichment for mitochondria-related pathways (Fig. [117]4c, d).

Fig. 4.

   [118]Fig. 4
   [119]Open in a new tab

   Prediction of hippocampal gene expression from peripheral blood
   metabolome. a Genes processed for this prediction analysis were
   provided from a previous study by Hagihara et al. [[120]14]. The
   correlation coefficients between actual and predicted expression levels
   of each gene are shown. Dashed lines indicatea a P-value of 0.05 and
   asterisks indicates significant correlations after Bonferroni
   correction for multiple testing (significance level:
   P < 0.00714 = 0.05/7). b Scatter plot showing significant correlations
   between actual and predicted expression level of Arntl. The correlation
   was not due to outliers. c Feature preference for constructing the
   prediction model of Arntl expression level. d Results of pathway
   enrichment analysis for the metabolites used for constructing the
   prediction model of Arntl expression level. The statistical enriched
   terms with raw P-value below 0.05 are shown

Discussion

   In this study, we demonstrated for the first time that peripheral blood
   metabolomics in combination with a statistical learning algorithm can
   predict individual levels of mood-change-related activity using an
   animal model showing exaggerated infradian rhythm. It could be
   suggested that this result is simply a consequence of the metabolite
   secretions affected by LA immediately before sampling. However,
   considering that blood metabolome patterns could not predict LA during
   the 0–6 h immediately before sampling (Fig. [121]3b), this cannot be
   considered a major factor in determining the blood metabolome patterns.
   Thus, blood metabolome patterns may hold information about infradian
   states of LA oscillation during specific time windows that are several
   hours or days apart from the sampling time.

   Our prediction approach highlighted potential alterations in
   mitochondria metabolism concomitant with infradian mood change-related
   activity. The Warburg effect is a shift in metabolism towards aerobic
   glycolysis away from oxidative phosphorylation even in the presence of
   oxygen which is induced through mitochondrial dysfunction. It was first
   described in cancer, and studies of both patients’ brains and of mouse
   models have recently been suggested that it could be involved in
   several psychiatric disorders [[122]24, [123]33]. Alterations in
   metabolites related to the citric acid cycle have been found in
   peripheral blood and cerebrospinal fluid of patients with bipolar
   disorder [[124]11, [125]34]. While these studies detected trait-related
   alterations and support the mitochondria hypothesis of bipolar disorder
   [[126]35–[127]38], our present study suggests that mitochondrial
   alterations are related to state-related alterations, or changes in
   mood states. However, the causal relationship between the
   mitochondria-related metabolic alterations in the blood and changes in
   LA levels remains unclear. Regarding state-related alterations, a
   recent meta-analysis indicated that blood levels of some inflammatory
   cytokines, such as IL-6 and sIL-6R, were increased in manic and
   euthymic states but not the depressive state compared to controls
   [[128]39]. Other study has shown that blood levels of IL-6 and IL-2 are
   positively correlated with mood symptoms, as indicated by the Young
   Mania Rating Scale [[129]40]. These findings suggest that inflammatory
   conditions could alter depending on affective states in bipolar
   disorder, with an enhanced inflammatory-related signature in manic
   states. Supporting these observations, we found that pyruvate, a
   metabolite suggested to be an endogenous anti-inflammatory molecule
   [[130]41], was negatively correlated with LA in Camk2a^+/− mice, and
   relatively low in manic-like states (Additional file [131]1: Figure
   S1). Although increased blood levels of pyruvate has implicated
   mitochondrial dysfunction in bipolar disorder, its significance on the
   pathogenesis of the disorder has remained unclear [[132]11].
   Considering the findings of the present study and those of studies
   discussed above, mitochondrial destabilization may be intricately
   associated with pro- and anti-inflammatory balance in relation to
   changes in mood states as well as to the pathogenesis of the disorder
   [[133]42].

   We found that blood metabolome patterns predicted 3 h LA at 3 days
   before sampling (63–66 h LA; Fig. [134]3b, c), and metabolites selected
   for the prediction showed enrichments in aspartate-related pathways
   (Additional file [135]1: Figure S4). Alterations in aspartate
   metabolism have been suggested in the brain and blood of patients with
   bipolar disorder [[136]11, [137]43]; however, their relations with mood
   changes are unknown. Considering that metabolites selected for the
   prediction of 3 h LA at 63–66 h before sampling and the related
   metabolic pathways were different from those for 24 h LA prediction
   (Fig. [138]2, Additional file [139]2: Table S1), levels of a distinct
   set of blood metabolites and the metabolic pathways might retain
   information about LA during different and specific time windows.

   We also found that the peripheral blood metabolome can predict
   expression levels of the Arntl gene in the hippocampus, a brain region
   associated with mood regulation [[140]31, [141]32], and that the
   metabolites that successfully predicted these levels showed enrichment
   for mitochondria-related pathways. These results suggest that systemic
   alterations in mitochondrial metabolism and hippocampal expression of
   Arntl, a candidate gene for bipolar disorder [[142]44–[143]46], may
   serve as a link for brain–periphery interactions in relation to
   mood-change-associated locomotor activity. Blood metabolites that
   change according to infradian cyclic locomotor activity could be
   associated not merely with the physical activity levels, but also
   potentially with hippocampal function in mood regulation via
   concomitant expression changes of the bipolar disorder candidate gene.

   In conclusion, the present study provided evidence that metabolome
   patterns in peripheral blood could be used as predictors of states of
   mood-change-related activity in this mouse model. As a limitation of
   this study, a recent history of physical activity might change
   peripheral metabolome patterns independent of the mood state, as
   intense or moderate exercise might affect mitochondria-related pathways
   in the brain, heart, and other organs in mice [[144]47, [145]48]. Among
   Camk2a^+/− mice, we previously showed that levels of infradian
   oscillatory LA in the home cages were associated with anxiety-like and
   depression-like behaviors [[146]14]. Although it is most likely that
   the LA state in the home cage reflects a certain state of mood in
   Camk2a^+/− mice, the effects of physical activity independent of the
   mood state on blood metabolome patterns should be considered as
   potential confounding factors in this study. This study, although
   limited to an animal model, may have potential for translational
   applications to human studies, as peripheral blood can be collected
   from living subjects in a minimally invasive way. States of mood and
   the related behaviors in human, especially patients with mood
   disorders, could potentially be retrospectively and prospectively
   predicted with blood samples through methods analogous to those used
   here. This may help to develop novel biologically-based methods for the
   diagnosis and treatment of bipolar disorder. To this end, further
   studies are needed to determine the most appropriate samples to collect
   (e.g., blood, saliva, urine, or skin), the appropriate processing
   methods (e.g., metabolomics, proteomics, or transcriptomics) and to
   optimize the prediction algorithm [[147]49].

Supplementary information

   [148]13041_2019_527_MOESM1_ESM.pdf^ (807.9KB, pdf)

   Additional file 1. Supplementary Materials and Methods. Figure S1.
   Correlations between 24 h LA and levels of metabolites that were used
   for the prediction of 24 h LA. Figure S2. Prediction results of 3 h LAs
   from metabolome patterns in peripheral blood of Camk2a^+/− mice.
   Prediction of 3 h LAs during the 3 days before sampling for 29 mice.
   LA, locomotor activity. Figure S3. Metabolites used in prediction
   models that significantly predicted 3 h LA. Hierarchical clustering of
   metabolites based on their frequencies selected to build
   cross-validated prediction models for each time window of LA using R
   software ([149]http://www.r-project.org/). LA, locomotor activity.
   Figure S4. Pathway enrichment analysis of the metabolites used for
   constructing a prediction model of 3 h LA at 63–66 h before sampling.
   The statistically enriched terms with raw P-values below 0.05 are
   shown.
   [150]13041_2019_527_MOESM2_ESM.xlsx^ (74.7KB, xlsx)

   Additional file 2: Table S1. The raw data used for this study.

Acknowledgements