Identification of plasma biomarkers for chronic kidney disease by integrating proteomics and machine learning

Chronic kidney disease (CKD) is a major global health burden, and early identification of individuals at risk remains challenging. This study aimed to identify plasma proteomic biomarkers associated with incident CKD and develop a machine learning-based prediction model integrating proteomic and clinical data.

We conducted a prospective cohort study of 27,429 UK Biobank participants without baseline CKD after excluding those with estimated glomerular filtration rate <60 mL/min/1.73 m2 or urinary albumin-to-creatinine ratio ≥3 mg/mmol. Among 2,920 plasma proteins, protein-wide Cox regression with Bonferroni correction was followed by least absolute shrinkage and selection operator (LASSO) regression and Light Gradient Boosting Machine (LightGBM)-based feature selection. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration metrics, reclassification indices, and landmark-based prediction analyses, with SHapley Additive exPlanations (SHAP), Kaplan-Meier, and sensitivity analyses.

A total of 205 proteins were significantly associated with incident CKD. LASSO identified 23 candidates, and LightGBM identified four key proteins (EDA2R, GDF15, HAVCR1, and SCARB2). The four-protein model showed comparable performance to the demographic model (AUC, 0.796 vs. 0.805; p = 0.14), while their combination improved discrimination (AUC, 0.830 [95% confidence interval, 0.806-0.855]; p < 0.001). The combined model also showed improved performance across 5-year, 10-year, and >10-year CKD prediction horizons. SHAP and Kaplan-Meier analyses consistently supported the predictive value of these proteins, and sensitivity analyses confirmed robustness.

Integrating plasma proteomics with machine learning improves internal risk prediction for incident CKD. The parsimonious four-protein panel may provide complementary information for CKD risk stratification, although external validation is needed before clinical application.