Plasma metabolomics identifies lipid mediators linking depression and cognitive decline in late-life depression

Late-life depression (LLD) features recurrent episodes and frequently co-exists with cognitive impairment, which predicts worse outcomes and progression to dementia. Evidence indicates a bidirectional depression-cognition relationship, but objective biological tools to capture severity and this interplay are scarce.

This study compared 110 patients with LLD between depressive episodes and remission phases. Based on untargeted liquid chromatography-mass spectrometry (LC-MS) metabolomics analysis of plasma samples, we identified key metabolites and developed four predictive machine learning models (GLMNet, LDA, Naive Bayes, and cTree). Additionally, Spearman rank correlation analysis and mediation analysis were conducted to further investigate the relationships and mediating effects of the key metabolites.

The diagnostic model based on key metabolites selected by the random forest algorithm showed good discriminatory performance in distinguishing LLD Episodes (AUC = 0.824). Tridecanoylcarnitine (Car(13:0)), PC(P-16:0/22:6), and SM(d18:1/22:0) were significantly downregulated during the depressive episode. Tridecanoylcarnitine (Car(13:0)) negatively correlated with depressive severity (p < 0.001) and positively with Mini-Mental State Examination scores. PC(P-16:0/22:6) was associated with both emotional and cognitive impairments. Mediation analysis supported that Tridecanoylcarnitine (Car(13:0)) and PC(P-16:0/22:6) partially mediated the depression-cognition relationship, explaining 22.0 % and 26.9 % of the variance, respectively (p < 0.05).

This study reveals specific lipid metabolic dysregulation in LLD and identifies key metabolites significantly associated with both depressive severity and cognitive function. It further supports their mediating role in the comorbidity between depression and cognitive impairment. These metabolites may serve as potential targets for simultaneously regulating depression and cognition. The machine learning model developed provides a new auxiliary tool for the objective assessment of LLD.