Integration of machine learning algorithms and experimental validation for discovery of endoplasmic reticulum stress-related biomarkers in diabetic nephropathy

Diabetic nephropathy (DN) is a leading cause of end-stage renal disease, with endoplasmic reticulum (ER) stress playing a critical role. This study aimed to identify ER stress-associated biomarkers and explore their immunometabolic associations in DN. Differentially expressed genes (DEGs) were identified using GEO datasets. WGCNA, LASSO, and SVM-RFE algorithms were integrated to select key ER stress-related signature genes. An exploratory diagnostic model was constructed via an artificial neural network (ANN) with a 6-5-2 architecture and explicitly binarized inputs, and validated using ROC curves. The model's performance was additionally compared against simpler baseline classifiers, and clinical applicability was evaluated using calibration and decision curve analyses (DCA). Potential mechanisms were explored through GSEA. Candidate biomarkers were validated in HK-2 cells and DN rats. Molecular docking predicted potential therapeutic agents targeting these biomarkers. From 107 DEGs, six ER stress-related signature genes were identified. The exploratory diagnostic model demonstrated high predictive performance, with AUC values of 0.977 (merged training cohort) and 0.947 (independent testing cohort). Further analysis revealed that two specific genes, C1QA and PTGDS , were significantly correlated with immune responses in DN. Experimental validation confirmed their dysregulation in high glucose-stimulated HK-2 cells and DN rat models. Molecular docking demonstrated that triptolide, quercetin, genistein, and resveratrol exhibited strong binding affinities for C1QA and PTGDS proteins, providing purely hypothesis-generating insights without establishing therapeutic efficacy. C1QA and PTGDS might be promising diagnostic biomarkers and potential indicators for DN. These findings provide novel insights into ER stress-related associations and offer a theoretical proof-of-concept for personalized treatment strategies in DN management.