Abstract
OBJECTIVE: To characterise proteomic changes associated with depression risk across different follow-up periods and develop predictive models integrating protein and clinical features.
METHODS: Olink proteomic data from the UK Biobank were analysed using logistic regression and Cox proportional hazards models to identify depression-associated proteins. Time-specific analyses were performed within 1-, 5-, and 10-year follow-up windows. Depression was defined using ICD-10 codes. GO and KEGG enrichment analyses were conducted, and predictive proteins were selected using sequential forward selection prior to XGBoost modelling.
RESULTS: Among 52,121 participants, 2,442 developed depression. Distinct plasma protein signatures were identified across followup windows. Within 1 year, BRK1, MME, LRPAP1, and LRP1 were significantly associated with depression. Within 5 years, MMP12, SPP1, and SPON2 were among 287 unique associated proteins, while TGM2, OMG, and UBAC1 were representative markers among 509 proteins identified within 10 years. LEP was consistently selected across all time windows and in the overall population. The combined protein-clinical XGBoost model achieved the best performance within the 1-year follow-up window (AUC = 0.808, 95% CI: 0.696-0.911), outperforming protein-only and clinical-only models.
CONCLUSION: Circulating plasma proteins have predictive value for depression risk, with LEP emerging as a robust biomarker across multiple time horizons.