188. DEVELOPMENT OF METABOLOMIC BIOMARKER SIGNATURES AND DIAGNOSTIC MODELS FOR MAJOR DEPRESSIVE DISORDER USING MULTIPLE SUPERVISED LEARNING METHODS

Abstract Background Major depressive disorder (MDD) involves complex plasma metabolic changes. This study aims to investigate metabolomic profiles linked to MDD and develop diagnostic models using machine learning. Aims & Objectives Based on metabolomics to construct diagnostic biomarkers for MDD and validate potential causal relationships using Mendelian randomization. Method An analysis was conducted on quantitative data for 630 plasma metabolites from 99 patients with MDD and 50 healthy controls. Metabolites associated with MDD risk were identified after adjustments for potential confounders. The dataset was divided into training and testing sets in a 7:3 ratio. Lasso regression and elastic net methods were employed to screen for metabolites, and multiple machine learning algorithms were used to construct diagnostic models, including XGBoost, support vector machines, and deep learning approaches. Model performance was evaluated through the AUC, calibration curves, and confusion matrix. Mendelian randomization (MR) was performed to infer potential causal links. Results Forty-seven metabolites were identified as significantly associated with MDD risk. Elevated levels of sphingomyelins, ceramides, and hexosylceramides suggest a role for sphingolipid-ceramide metabolism in MDD. The deep learning model, using eight selected metabolites (SM C18:1, SM C18:0, HexCer(d18:1/18:0), HexCer(d18:1/24:1), methionine, cysteine, spermidine, and PC ae C38:1), achieved the highest AUC of 0.83 (95% CI, 0.70-0.95). Significant MR associations were found for two of these metabolites. Discussion & Conclusions This study highlights that altered levels of specific metabolites may indicate an increased risk of MDD and possess potential diagnostic value. The findings underscore the importance of further research to elucidate the underlying mechanisms of their role in MDD development.