Construction and evaluation of an integrated predictive model for chronic kidney disease based on the random forest and artificial neural network approaches

Chronic kidney disease (CKD) is recognized as a serious global health problem due to its high prevalence and all-cause mortality. The aim of this research was to identify critical biomarkers and construct an integrated model for the early prediction of CKD. By using existing RNA-seq data and clinical information from CKD patients from the Gene Expression Omnibus (GEO) database, we applied a computational technique that combined the random forest (RF) and artificial neural network (ANN) approaches to identify gene biomarkers and construct an early diagnostic model. We generated ROC curves to compare the model with other markers and evaluated the associations of selected genes with various clinical properties of CKD. Moreover, we highlighted two biomarkers involved in energy metabolism pathways: pyruvate dehydrogenase kinase 4 (PDK4) and zinc finger protein 36 (ZFP36). The downregulation of the identified key genes was subsequently confirmed in both unilateral ureteral obstruction (UUO) and ischemia reperfusion injury (IRI) mouse models, accompanied by decreased energy metabolism. In vitro experiments and single-cell sequencing analysis proved that these key genes were related to the energy metabolism of proximal tubule cells and were involved in the development of CKD. Overall, we constructed a composite prediction model and discovered key genes that might be used as biomarkers and therapeutic targets for CKD. (C) 2022 Elsevier Inc. All rights reserved.