Bootstrap inference and machine learning reveal core differential plasma metabolic connectome signatures in major depressive disorder

Background: Major depressive disorder (MDD) involves molecular alterations and pathway dysregulation. Metabolic interconnections are crucial for normal functioning, yet current analysis focuses on individual pathways or biomarkers, overlooking intricate metabolic biomarker interactions. Methods: Plasma metabolomic data from 182,053 UK Biobank participants [9425 MDD, and 172,628 healthy controls (HC)] were used to construct metabolic correlation networks through bootstrap inference analysis (bootstrap step size: 1000, 3000, 5000, 7000, 9000; n = 1000 times/size). Differential core metabolic network signatures between MDD and HC were identified by machine learning, followed by metabolic pathway analysis. Various deep learning and machine learning models were employed to differentiate MDD from HC groups using the identified network features and baseline characteristics. Results: The MDD metabolic network showed marked reorganization, with a sparser and more streamlined network structure compared to controls (p < 0.05 for both V_net-edge and V_net-node). Analysis of the core network in MDD revealed four key altered pathways, with linoleic acid metabolism being the most influential (p < 0.01, impact = 0.29). An extreme gradient boosting model combining network signatures and baseline features achieved 73 % accuracy, and an AUROC of 0.82 in differentiating MDD from HC groups. Conclusions: This large-scale, metabolomic connectome analysis revealed consistent dysregulated metabolic network features in MDD, providing a robust and distinguishable framework compared to controls. The MDD network exhibits distinct connectivity patterns, particularly within linoleic acid metabolism. Integrating metabolomics as networks, rather than isolated markers, offers a promising approach for elucidating MDD pathophysiology and identifying diagnostic biomarkers.