A non-invasive urinary diagnostic signature for diabetic kidney disease revealed by machine learning and single-cell analysis

by Yonggang Chen, Jintai Luo, Yingying Zheng, Xiaomei Jiang, Zixiang Yang, Xiaobing Liu

Diabetic kidney disease (DKD) poses a significant health burden with inadequate diagnostic sensitivity. This study develops non-invasive biomarkers by integrating urinary and renal single-cell sequencing with machine learning.

This study analyzed DKD single-cell and bulk transcriptomic data from public repositories. We established a computational pipeline to distinguish kidney-originating cells in urinary sediments, enabling the identification of injury-associated gene signatures. These signatures were refined using machine learning to develop a diagnostic model, which was validated in independent cohorts. The biomarkers were further verified in DKD renal tissues at single-cell resolution and across multiple nephropathies. Functional and spatial analyses confirmed biological relevance using transcriptomic and histological validation.

Single-cell analysis of 2,089 urine-derived cells identified eight renal cell types, including injured proximal tubule cells (Inj-PTC) showing upregulated injury markers (HAVCR1, VCAM1) and enriched apoptotic/TGF-β pathways. A machine learning-selected biomarker panel (PDK4, RHCG, FBP1) demonstrated strong diagnostic value (area under the curve, AUC > 0.9), with consistent downregulation across multiple chronic kidney diseases. PDK4 and FBP1 were specifically suppressed in DKD renal Inj-PTC (p  Conclusions

This study identifies a three-gene biomarker panel (PDK4, RHCG, FBP1) as a promising non-invasive diagnostic tool for DKD. While demonstrating excellent diagnostic performance. It represents a tubular injury-associated gene signature that is detectable in urinary cells and shows strong association with DKD in transcriptomic datasets, presenting a promising candidate for a non-invasive diagnostic assay.