Current diagnoses of mood disorders are not cross validated. The aim of the current paper is to explain how machine learning techniques can be used to a) construct a model which ensembles risk/resilience (R/R), adverse outcome pathways (AOPs), staging, and the phenome of mood disorders, and b) disclose new classes based on these feature sets. This study was conducted using data of 67 healthy controls and 105 mood disordered patients. The R/R ratio, assessed as a combination of the paraoxonase 1 (PON1) gene, PON1 enzymatic activity, and early life time trauma (ELT), predicted the high-density lipoprotein cholesterol – paraoxonase 1 complex (HDL-PON1), reactive oxygen and nitrogen species (RONS), nitro-oxidative stress toxicity (NOSTOX), staging (number of depression and hypomanic episodes and suicidal attempts), and phenome (the Hamilton Depression and Anxiety scores and the Clinical Global Impression; current suicidal ideation; quality of life and disability measurements) scores. Partial Least Squares pathway analysis showed that 44.2% of the variance in the phenome was explained by ELT, RONS/NOSTOX, and staging scores. Cluster analysis conducted on all those feature sets discovered two distinct patient clusters, namely 69.5% of the patients were allocated to a class with high R/R, RONS/NOSTOX, staging, and phenome scores, and 30.5% to a class with increased staging and phenome scores. This classification cut across the bipolar (BP1/BP2) and major depression disorder classification and was more distinctive than the latter classifications. We constructed a nomothetic network model which reunited all features of mood disorders into a mechanistically transdiagnostic model.