Deep visual proteomics reveals distinct proximal tubular and glomerular injury programs in experimental diabetic kidney disease
Background: Diabetic kidney disease (DKD) is the leading cause of chronic kidney disease (CKD). However, most proteomic studies of DKD rely on bulk kidney tissue, which cannot distinguish the contribution or response of individual nephron compartments to injury. Methods: Diabetes was induced in male mice by streptozotocin (STZ) injections. After 16-weeks, the mice and vehicle-treated controls were characterized physiologically, biochemically, and histologically. A deep learning-powered Deep Visual Proteomics (DVP) pipeline, validated against manual annotation, was adapted to isolate proximal tubule (PT) and glomeruli from Megalin stained kidney sections by automated laser microdissection. Bulk kidney, PT, and glomerular proteomes were generated by data independent acquisition mass spectrometry. PT-enriched candidates were prioritized using a composite scoring approach and compared with human tubulointerstitial proteomic data from the Kidney Precision Medicine Project. Results: STZ mice developed sustained hyperglycaemia and albuminuria, alongside elevated markers of tubular injury and interstitial fibrosis. Segmentation models isolated PT and glomeruli with high fidelity (Dice coefficients 0.878 and 0.914; area correlations r=0.993 and r=0.996). Compartment-resolved proteomics determined that PT and glomeruli underwent largely distinct, non-overlapping remodelling: PT exhibited loss of proteostatic, cell cycle, and structural programs with compensatory mitochondrial and lipid metabolic upregulation, whereas glomeruli showed broad loss of oxidative metabolic capacity without any compensatory metabolic program. Fourteen of the top twenty prioritized PT candidates, including LARS2 and ANXA2, changed in the same direction in human CKD tubulointerstitial proteomic data. The STZ PT proteome correlated significantly with this human dataset, while the glomerular comparison did not. Conclusions: Compartment-resolved and deep learning-guided visual proteomics can uncover divergent, biologically coherent PT and glomerular injury programs in DKD that are masked in bulk tissue analysis. A PT injury signature was uncovered that is partially conserved in human CKD, identifying novel candidate mechanisms and biomarkers for future exploration.