Abstract: TH-PO0595
Artificial Intelligence (AI)-Based Quantification of COL4A5 Deposition on Glomerular Basement Membrane in Mice with Alport Syndrome Defines Therapeutic Threshold for Gene Therapy
Session Information
- Monogenic Kidney Diseases: Glomerular
November 06, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Genetic Diseases of the Kidneys
- 1201 Genetic Diseases of the Kidneys: Monogenic Kidney Diseases
Authors
- Nakai, Hiroyuki, Oregon Health & Science University, Portland, Oregon, United States
- Das, Ranjan, Oregon Health & Science University, Portland, Oregon, United States
- Sairavi, Anusha, Oregon Health & Science University, Portland, Oregon, United States
- Li, Lena, Oregon Health & Science University, Portland, Oregon, United States
- Andeen, Nicole K., Oregon Health & Science University, Portland, Oregon, United States
- Lin, Meei-Hua, Washington University in St Louis School of Medicine, St. Louis, Missouri, United States
- Miner, Jeffrey H., Washington University in St Louis School of Medicine, St. Louis, Missouri, United States
Background
AAV vector-mediated gene therapy holds promise for treating Alport syndrome (AS); however, no prior studies have quantitatively defined the relationship between COL4A5 restoration in the glomerular basement membrane (GBM) and therapeutic efficacy. Quantifying COL4A5 deposition is critical for understanding therapeutic thresholds and optimizing AS gene therapy; however, manual analysis of IF images is time-consuming, subjective, and often lacks reproducibility. Therefore, the development of more objective and robust measures is imperative.
Methods
Four-week-old LSL-Col4a5 X-linked AS mice were intravenously injected with 3×1011 vg of AAV9-CAG-Cre. This early gene therapy intervention led to therapeutic benefits, including improved biomarkers and renal histopathology. At 28 weeks post-injection, 2 treated mice were euthanized, and kidney sections were stained with anti-COL4A5 and anti-agrin antibodies. Using confocal images from a total of 23 double-positive glomeruli, we developed an AI-based pipeline using U-Net with a pre-trained ResNet-50 encoder to simultaneously segment agrin-positive GBM and COL4A5-positive GBM and quantify COL4A5 deposition as a percentage of total GBM length. The results were compared against manually quantified percentages for validation.
Results
Despite the small dataset, the model was successfully trained using 4-fold cross-validation on 8 GPUs over 600 epochs, requiring ~4 h. The average Dice coefficient for individual cross-validated models was 0.67, but ensemble learning across nine models improved robustness and raised it to 0.73. AI-based automated COL4A5 quantification closely matched manual results (manual: 10.6%; AI: 10.0%), showing strong concordance. Using a mathematical model extrapolating 2D data to 3D data and based on the 40% glomerular positivity we observed in treated mice, we estimated that ~9% of GBM had COL4A5 deposition in the 3D context.
Conclusion
Our AI-based quantification method offers a robust and objective alternative to manual analysis, enabling precise measurement of COL4A5 deposition and providing a valuable tool for evaluating gene therapy outcomes. Importantly, this study revealed that the restoration of COL4A5 in as little as 9% of GBM is sufficient to mediate therapeutic effects in AS.