Metabolomic biosignature differentiates melancholic depressive patients from healthy controls

Abstract Background Major depressive disorder (MDD) is a heterogeneous disease at the level of clinical symptoms, and this heterogeneity is likely reflected at the level of biology. Two clinical subtypes within MDD that have garnered interest are “melancholic depression” and “anxious depression”. Metabolomics enables us to characterize hundreds of small molecules that comprise the metabolome, and recent work suggests the blood metabolome may be able to inform treatment decisions for MDD, however work is at an early stage. Here we examine a metabolomics data set to (1) test whether clinically homogenous MDD subtypes are also more biologically homogeneous, and hence more predictiable, (2) devise a robust machine learning framework that preserves biological meaning, and (3) describe the metabolomic biosignature for melancholic depression. Results With the proposed computational system we achieves around 80 % classification accuracy, sensitivity and specificity for melancholic depression, but only ~72 % for anxious depression or MDD, suggesting the blood metabolome contains more information about melancholic depression.. We develop an ensemble feature selection framework (EFSF) in which features are first clustered, and learning then takes place on the cluster centroids, retaining information about correlated features during the feature selection process rather than discarding them as most machine learning methods will do. Analysis of the most discriminative feature clusters revealed differences in metabolic classes such as amino acids and lipids as well as pathways studied extensively in MDD such as the activation of cortisol in chronic stress. Conclusions We find the greater clinical homogeneity does indeed lead to better prediction based on biological measurements in the case of melancholic depression. Melancholic depression is shown to be associated with changes in amino acids, catecholamines, lipids, stress hormones, and immune-related metabolites. The proposed computational framework can be adapted to analyze data from many other biomedical applications where the data has similar characteristics.