This study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of University of Occupational and Environmental Health, Kitakyushu, Japan (UOHECRB21-057). All participants provided verbal and written informed consent. All participants were assured that they could withdraw consent at any time and would not be disadvantaged as a result.

Forty patients with MDD participated in this study. Patients were recruited from Hospital of University of Occupational and Environmental Health Japan. All patients were diagnosed with MDD using the Diagnostic and Statistical Manual for Mental Disorders, fifth edition [14]. Exclusion criteria included a history of major neurological disease, epilepsy, cerebrovascular accident, head trauma with cognitive sequelae, and intellectual disability. Three of the 40 patients did not take any antidepressant medication, while 37 patients were taking antidepressant drugs. Total antidepressant use was calculated based on imipramine equivalent [15, 16]. Type 2 diabetes mellitus was diagnosed in accordance with the relevant Japan Diabetes Society criteria (i.e., plasma glucose levels: fasting $\ge$126 mg/dL and HbA1c $\ge$6.5%) [17] or if the patient was currently taking diabetes medication. Six HCs with no history of psychiatric or physical diseases were also enrolled as a comparison group. The participants in the present study overlapped with those in our past published study [18]; however, no metabolomics data has been published yet.

The clinical symptoms of patients with MDD were assessed using the Montgomery-Asberg Depression Rating Scale (MADRS) [19]. We collected non-fasting serum blood sample. Patient blood was collected in plain blood tubes at University of Occupational and Environmental Health Japan; serum was separated by centrifugation at 2000 g for 20 min and stored at –80 °C in silicone-coated tubes until analysis.

Serum samples were transferred to Human Metabolome Technologies, Inc. (HMT, Tsuruoka, Japan) for metabolite analysis, using capillary electrophoresis (CE) coupled with Fourier transform mass spectrometry (CE-FTMS) based on the method described in HMT’s $\omega$ Scan package. CE-FTMS analysis was conducted using an Agilent 7100 CE system equipped with the Q Exactive Plus mass spectrometer (Thermo Fisher Scientific Inc., Waltham, MA, USA), an Agilent 1260 isocratic high performance liquid chromatography pump (Agilent Technologies Inc., Santa Clara, CA, USA), an Agilent G1603A CE-MS adapter kit (Agilent Technologies Inc., Santa Clara, CA, USA), and an Agilent G1607A CE-ESI-MS sprayer kit (Agilent Technologies Inc., Santa Clara, CA, USA). The components were controlled using Agilent MassHunter workstation software for 6200 series TOF/6500 series Q-TOF version B.08.00 (Agilent Technologies Inc., Santa Clara, CA, USA) and Xcalibur (Thermo Fisher Scientific Inc.) and were connected by a fused silica capillary (internal diameter, 50 $\mu$m; total length, 80 cm). Commercial electrophoresis buffers (H3301-1001 and I3302-1023 for cation and anion analyses, respectively; HMT, Tsuruoka, Japan) were used as the electrolytes. The spectrometer scanned from m/z 50 to 1000. Peaks were analyzed using the MasterHands automatic integration software (Keio University, Tsuruoka, Yamagata, Japan) to obtain m/z, peak area, and migration time (MT) data. Signal peaks corresponding to isotopomers, adduct ions, and other product ions of known metabolites were excluded, and the remaining peaks were annotated according to the HMT metabolome database based on their m/z values and MTs. The areas of the annotated peaks were then normalized to those of internal standards and sample volumes to obtain the relative levels of each metabolite. The detected metabolites were plotted on metabolic pathway maps using VANTED [20]. Metabolomics analysis assigned a candidate compound to the 496 (290 cation, 206 anion) peaks.

The detected peaks were broadly classified into the following 12 pathways based on the candidate compounds: carbohydrate metabolism, glucose metabolism/gluconeogenesis, energy storage by the tricarboxylic acid (TCA) cycle, energy conversion by the TCA cycle, glutamate metabolism and urea cycle, choline metabolism and methionine cycle, aromatic amino acid metabolism (tryptophan), aromatic amino acid metabolism (phenylalanine and tyrosine), nucleic acid metabolism (purine metabolism - adenosine), nucleic acid metabolism (purine metabolism - guanosine), nucleic acid metabolism (pyrimidine metabolism), and coenzyme metabolism. The classification of metabolic pathways and substance properties was mainly based on the KEGG pathway database [8]. Note that some metabolites were not classified under the 12 pathways. This is because metabolomics analysis also comprehensively detects metabolites which structure can be detected but actual action or metabolic pathway they can be classified into is not known. When a metabolite spanned more than one pathway, it was counted in both. Second, for each of the top 30 positive and negative factor loading metabolites we used the Human Metabolome Database numbers and calculated their pathway impact and match status [21]. Metabolomics pathway analysis combines several advanced pathway enrichment analysis procedures along with the analysis of pathway topological characteristics to help identify the most relevant metabolic pathways involved in given metabolomic study [21].

Statistical analyses were performed using the MATLAB [22], MetaboAnalyst [23], and EZR [24]. First, we used Welch’s t-test to examine the total number of metabolites for MDD with and without diabetes. Next, all metabolites, which metabolomics analysis assigned a candidate compound to the 496 peaks were evaluated using principal component analysis (PCA) [25]. Thereafter, we used naive partial least squares (PLS) analysis [26]. We assessed the separation between groups by PCA and naive PLS using one-way analysis of variance (ANOVA). We extracted the top metabolites with the highest factor loadings. The relationship of individual metabolites with antidepressants (imipramine equivalents) was examined using multiple regression analysis. Age and body mass index (BMI) were corrected as covariates in our multiple regression analysis. We then calculated the pathway impact and match status [21] for individual metabolites specifically, the top 30 positive and negative factor loadings of the first principal component of naive PLS. Thereafter, the potential of blood metabolites to discriminate between patients with MDD with type 2 diabetes and those with MDD without type 2 diabetes was evaluated using receiver operating characteristic (ROC) curve analysis. We targeted metabolites that were detected in all 46 samples as candidate biomarkers and were statistically significant in Welch’s t-test or naive PLS.

Continuous variable data are expressed in terms of mean (standard deviation) or median (interquartile range) values. The tests were two-tailed, and p values 0.05 were considered statistically significant. No outliers were excluded from the analysis. For missing values from the detection limit, only data other than missing values were used in the analysis. Demographics data showed that age and BMI were different among the three groups; hence, they were incorporated into the PCA and naive PLS. We used a dataset of 496 (290 cation, 206 anion) peaks and age and BMI identified in Metabolomics, standardized each item ($\mu$ = 0, $\sigma$ = 1), and calculated PCA and naive PLS. The smoothing parameter $\kappa$ was set to zero. Pearson’s correlation coefficient was used for determining the factor loadings and p values in the naive PLS, for the score of the objective variable and each metabolite level [27]. In our calculation of pathway impact, the enrichment method was parameterized by hypergeometric test and the topology analysis by relative-betweenness centrality.

Table 1 showed the demographic and clinical characteristics of the patients with MDD and HCs. In the MDD group, 13 patients had type 2 diabetes mellitus, while 27 did not. Patients with MDD with type 2 diabetes tended to be older, overweight and had milder depressive symptoms than those without type 2 diabetes.

Welch’s t-test revealed that 22 metabolites showed statistical differences between MDD with and without type 2 diabetes (Table 2). We evaluated all metabolites using PCA and created two-dimensional plots for the first and second principal components. PCA revealed no clear separation among the three groups (first principal component mean $\pm$ standard deviation: MDD without type 2 diabetes mellitus, 0.76 $\pm$ 15.3; MDD with type 2 diabetes mellitus, –0.23 $\pm$ 4.04; HCs, –2.93 $\pm$ 1.19; p value = 0.79, by one-way ANOVA; Fig. 1). The contribution ratio of the first principal component was 28.6%, while that of the second principal component was 8.1%.

Fig. 1.

PCA results. No clear separation among the groups is observed (first principal component mean $\pm$ standard deviation: MDD without type 2 diabetes mellitus, –2.88 $\pm$ 3.85; MDD with type 2 diabetes mellitus, 4.83 $\pm$ 1.88; HCs, 2.52 $\pm$ 0.65, p value 0.001, by one-way ANOVA). The contribution ratio of the first principal component was 28.6%, while that of the second principal component was 8.1%.

Two MDD patients without type 2 diabetes mellitus showed a distinctly different distribution from the other groups. The two patients had a BMI of 18 and 15 kg/m${}^2$, and were underweight. There was no dissociation from the mean for the other background items. We thus removed these two samples and re-analyzed the remaining samples, and there was no difference in the results for PCA or the naive PLS top metabolites. We therefore continued our analysis with sample data, including the outliers.

Since PCA did not show reliable separation, we used naive PLS. As a result, three relatively good and distinct populations were identified (first principal component mean $\pm$ standard deviation: MDD without type 2 diabetes mellitus, –2.88 $\pm$ 3.85; MDD with type 2 diabetes mellitus, 4.83 $\pm$ 1.88; HCs, 2.52 $\pm$ 0.65, p value 0.001, by one-way ANOVA; Fig. 2). The contribution ratio of the first principal component was 70.6%, while that of the second principal component was 29.4%. On the axis of the first principal component, the HCs were in the middle of the two MDD groups and the HCs were separated from the MDD group along the second component.

Fig. 2.

Naive PLS results. Three relatively good and distinct populations are identified (first principal component mean $\pm$ standard deviation: MDD without type 2 diabetes mellitus, –2.88 $\pm$ 3.85; MDD with type 2 diabetes mellitus, 4.83 $\pm$ 1.88; HCs, 2.52 $\pm$ 0.65, p value 0.001, by one-way ANOVA). The contribution ratio of the first principal component is 70.6%, while that of the second principal component is 29.4%. On the axis of the first principal component, the HCs group was between the two MDD groups.

By selecting the top metabolites in the naive PLS factor loadings for which we found statistically significant differences, we identified 30 candidate biomarker metabolites. We presented metabolites with p value 0.05 among the top positive and negative factor loadings in naive PLS (The reason for the two metabolites in one space was insufficient separation and therefore no peak discrimination; these were counted as one metabolite; Table 3). Fourteen species were detected in specimens from all three groups.

Top five positive factor loading were as follows: metformin (factor loading = 0.520, p value 0.001, number detected = 6), 4-hydroxyphenylacetaldehyde (factor loading = 0.478, p value 0.001, number detected = 26), 2-hydroxyoctanoic acid-3 or 8-hydroxyoctanoic acid-3 (factor loading; 0.472, p value 0.001, number detected = 46), Glucurone (factor loading; 0.433, p value = 0.003, number detected = 45) and muscimol (factor loading; 0.420, p value = 0.004, number detected = 46). Top five negative factor loading was as follows: Guanidinoacetic acid (factor loading; –0.405, p value = 0.005, number detected = 46), glycerophosphocholine (factor loading = –0.348, p value = 0.018 number detected = 46), myo-inositol 2-phosphate (factor loading = –0.335, p value = 0.023, number detected = 39), FAD divalent (factor loading = –0.334, p value = 0.023, number detected = 22) and myo-inositol 1-phosphate (factor loading = –0.316, p value = 0.032, number detected = 39).

We found that the background factors of age (factor loading; 0.243, p value = 0.10, number detected = 46) and BMI (factor loading; 0.179, p value = 0.24, number detected = 46) did not contribute much to the differences between groups compared to top factor loading metabolites.

We examined the relationship of individual metabolites in the top factor loading metabolites with antidepressants (imipramine equivalents). Alanine was significantly negatively correlated with imipramine equivalents (standard partial regression coefficient = –0.349, 95% confidence interval: –0.659 to –0.039, standard error = 0.152, t value = –2.295, p value= 0.028; Fig. 3A). N-Acetylglucosamine 6-phosphate was significantly negatively correlated with imipramine equivalents of antidepressants dose (standard partial regression coefficient = –0.530, 95 % confidence interval: –1.038 to –0.023, standard error = 0.236, t value= –2.240, p value = 0.042; Fig. 3B). These p values were adjusted for age and BMI.

Fig. 3.

Relationship between top factor loading metabolites and imipramine equivalence of antidepressants dose in patients with MDD. (A) Alanine is significantly negatively correlated with imipramine equivalents (standard partial regression coefficient = –0.349, p = 0.028). (B) N-acetylglucosamine 6-phosphate is significantly negatively correlated with imipramine equivalents (standard partial regression coefficient = –0.530, p = 0.042). These p values were adjusted for age and BMI.

Among the top 30 positive factor loadings of the first principal component, 14 metabolites were broadly classified into 12 metabolic pathways (Fig. 4A): energy conversion by the TCA cycle (fatty acid catabolism and branched chain amino acid metabolism), six cases; choline metabolism and methionine cycle, three cases; nucleic acid metabolism (purine metabolism - adenosine), two cases; nucleic acid metabolism(purine metabolism - guanosine), two cases; energy storage by the TCA cycle, two cases; aromatic amino acid metabolism (phenylalanine and tyrosine), one case; glucose metabolism/gluconeogenesis, one case; and glutamate metabolism and urea cycle, one case. Among the top 30 negative factor loadings of the first principal component, 20 metabolites were broadly classified into 12 metabolic pathways (Fig. 4B): glutamate metabolism and urea cycle, four cases; choline metabolism and methionine cycle, three cases; carbohydrate metabolism, three cases; coenzyme metabolism, two cases; and glucose metabolism/gluconeogenesis metabolism, one case. Other metabolites could not be classified in the 12 broadly defined pathways.

Fig. 4.

Distributions of the number of cases of the factor loadings of the first principal component in naive PLS. (A) Among the top 30 positive factor loadings of the first principal component, 14 metabolites were classified into metabolic pathways. (B) Among the top 30 negative factor loadings of the first principal component, 20 metabolites were classified into metabolic pathways.

We then calculated the pathway impact and match status [21] for individual metabolites; specifically, for those among the top 30 positive and negative factor loadings of the first principal component of naive PLS, we performed pathway analyses that also included more detailed sub-items. Synthesis and degradation of ketone bodies was the high impact pathway in top 30 for positive factor loading (Table 4; Fig. 5A), and taurine and hypotaurine metabolism were extracted as high impact pathways in top 30 for negative factor loading (Table 4; Fig. 5B). We showed differences between groups of matched metabolites involved in pathway impact (Fig. 6A,B)

Fig. 5.

Pathway impact analysis. (A) Analysis of the top 30 positive factor loadings of the first principal component. Synthesis and degradation of ketone bodies was the high impact pathway in top 30 for positive factor loading (B) Analysis of the top 30 negative factor loadings of the first principal component. Taurine and hypotaurine metabolism were extracted as high impact pathways in top 30 for negative factor loading.

Fig. 6.

Differences between groups of matched metabolites involved in pathway impact. (A) Analysis of the top 30 positive factor loadings of the first principal component. (B) Analysis of the top 30 negative factor loadings of the first principal component.

We performed ROC curve analysis to evaluate the sensitivity and specificity of the candidate biomarkers for distinguishing between MDD patients with or without type 2 diabetes mellitus (Table 5; Fig. 7). We targeted metabolites that were detected in all 46 samples as candidate biomarkers and were statistically significant in Welch’s t-test or naive PLS. We also examined the sensitivity and specificity of the known diabetes mellitus biomarker HbA1c as a comparison. Among the MDD with and without type 2 diabetes groups, HbA1c was the most discriminatory variable compared to any other single metabolite.

Fig. 7.

Candidate biomarkers for distinguishing between MDD patients with or without type 2 diabetes mellitus. A bar chart of the candidate biomarker is presented. Those biomarkers not illustrated here have been shown in Fig. 6.