A Machine Learning Analysis of Big Metabolomics Data for Classifying Depression: Model Development and Validation

Background There have been many metabolomics studies of depression, but these have been limited by their scale. A comprehensive in silico analysis of global metabolite levels in large populations could provide robust insights into the pathological mechanisms underlying depression and candidate clinical biomarkers. Methods Depression-associated metabolomics was studied in two datasets from the UK Biobank database: participants with lifetime depression (n=123,459) and those with current depression (n=94,921). The Whitehall II cohort (n=4,744) was used for external validation. CatBoost machine learning was used for modeling, and Shapley Additive Explanations were used to interpret the model. Five-fold cross-validation was used to validate model performance, training the model on three of the five sets with the remaining two for validation and testing, respectively. The diagnostic performance was assessed using area under receiver operating characteristic (AUC) curves. Results Twenty-four significantly associated metabolic biomarkers were identified in the lifetime depression and current depression datasets and sex-specific analyses, 12 of which overlapped in the two datasets. The addition of metabolic features slightly improved the performance of a diagnostic model using traditional (non-metabolomic) risk factors alone (lifetime depression: AUCs 0.655 versus 0.658 with metabolomics; current depression: AUCs 0.711 versus 0.716 with metabolomics). Conclusions The machine learning model identified 24 metabolic biomarkers associated with depression. If validated, metabolic biomarkers may have future clinical applications as supplementary information to guide early and population-based depression detection.