Fig 1.
Flow diagram processing patients’ plasma samples from three independent clinical centers.
Fig 2.
Plasma metabolome-based PLS regression model for PHQ-9 and HAMD-17 prediction.
(A) For the medication-free plasma sample set of 26 patients, we conducted LC-MS-based metabolome analysis of water-soluble metabolites in plasma. Based on the 123 kinds of metabolites, the multivariate data sets were centered, scaled to Pareto and subjected to a multivariate statistical analysis using the SIMCA P+ ver. 14.0 software program (Umetrics, Sweden). The partial-least-squares (PLS) PHQ-9-regression model was created to identify the metabolites that associated with the severity of depression (SOD). The x-axis indicates the observed score of PHQ-9, while the y-axis indicates the predicted value of PHQ-9. (B) The HAMD-17-regression model was created based on the same data set 1. The x-axis indicates the observed score of HAMD-17, while the y-axis indicates the predicted value of HAMD-17. (C) Plasma samples from 23 medicated MDD patients (Data set 2) were subjected to LC-MS metabolome analysis and a HAMD-17-regression model was created as in (B). (D) Plasma samples of 41 patients from the third data set including both the medicated and non-medicated plasma samples, which were diagnosed with MDD (n = 27, black) and with bipolar disorder (n = 14, gray), were subjected to LC-MS metabolome/HAMD-17 regression analysis. R2, the square of Spearman’s correlation coefficient; RMSEE, root mean square error of estimation; RMSEcv, root mean square error of cross validation.
Table 1.
Plasma metabolites primarily contributed to the respective PLS-regression model.
The variable importance in projection (VIP) denotes the degree of contribution to the PLS regression model, whose scores greater than 1.0 can be considered important in given model. Metabolites displayed in bold style represent to commonly contribute to the three independent PLS regression models, thereby suggesting strongly associated with the underlying pathophysiology of depression.
Table 2.
Correlation between sub-scales of PHQ-9/HAMD-17 and plasma metabolites in medication-free cohort study (Data set-1).
Metabolites which have moderate correlation (absolute correlation coefficient value >0.2) with reciprocally synonymous sub-scale of PHQ-9/HAMD-17 are listed. Gray shaded denotes negatively correlation with the respective sub-scales.
Fig 3.
Correlation networks of HAMD-17 sub-scale and metabolites in medication-free cohort samples (Data set 1).
Red solid/dot lines, positive correlation; blue solid/dot lines, negative correlation; Black nodes, HAMD-17 subscales; blue nodes, metabolites. Edge-width scales as the correlation coefficients in absolute value. The bolder the lines, the stronger the correlation between connected nodes is.
Table 3.
Metabolites commonly associated with HAMD_SI.
In three distinct cohort studies, four metabolites moderately and commonly correlate with HAMD_SI. Gray shaded denotes negatively correlation with the respective sub-scales.
Fig 4.
Classification and regression models for predicting suicidality and a grade of suicidal ideation in depressive patients.
(A) Receiver-operating curve based on ten kinds of logistic regression models for classifying suicidality in depressive patients. Based on mixed data containing three independent data sets (Data set-1, Data set-2, and Data set-3), 10 kinds of distinct training data are created and subjected to building classification models discriminating between the depressive patients who have a sub-scale of suicidal ideation (HAMD_SI: HAMD-17_11> = 1) and not (HAMD_SI = 0) (details are described in S1 Table). Curves in red denote highly predictive models for test data set (true rate >0.7). (B) Significant correlation (R = 0.22, p = 0.028) between scored HAMD_SI and predictive scores of multiple linear regression model, based on standardized intensities of plasma citrate and kynurenine. A fitted linear regression line was depicted with 95% confidence area (gray shaded). Parameters of the model were described in S2 Table.