Abstract: FR-OR34
Combination of Transcriptional Signatures and 3D Images of Epithelial Nuclei in Kidney Biopsies Using a Deep-Learning Outlier Detection Model Uncovers Signatures of Injury in Diabetic Kidney Disease
Session Information
- Diabetic Kidney Disease: From Single Cell to Outcomes
November 04, 2022 | Location: W240, Orange County Convention Center‚ West Building
Abstract Time: 04:57 PM - 05:06 PM
Category: Diabetic Kidney Disease
- 601 Diabetic Kidney Disease: Basic
Authors
- Winfree, Seth, University of Nebraska Medical Center College of Medicine, Omaha, Nebraska, United States
- Barwinska, Daria, Indiana University School of Medicine, Indianapolis, Indiana, United States
- Zhuang, Jun, Indiana University Purdue University Indianapolis, Indianapolis, Indiana, United States
- Talukder, Niloy, Indiana University Purdue University Indianapolis, Indianapolis, Indiana, United States
- Eadon, Michael T., Indiana University School of Medicine, Indianapolis, Indiana, United States
- Dagher, Pierre C., Indiana University School of Medicine, Indianapolis, Indiana, United States
- Hasan, Mohammad Al, Indiana University Purdue University Indianapolis, Indianapolis, Indiana, United States
- El-Achkar, Tarek M., Indiana University School of Medicine, Indianapolis, Indiana, United States
Group or Team Name
- Kidney Precision Medicine Project
Background
Epithelial cell states in renal health and disease have become an area of intense research focus. By integrating multiomics and imaging from samples collected by large consortia (e.g. Kidney Precision Medicine Project) epithelial cell states of injury, maladaptive and adaptive repair, cycling and degeneration have been identified. Recently, we demonstrated the sufficiency of nuclei labeled with a common nuclear stain in classifying the epithelia of the nephron in the human renal cortex. To expand this approach and identify signatures of injury in diabetes, we implemented a deep learning model for outlier detection combining transcriptional profile of cell states with images of renal epithelial nuclei.
Methods
To identify injury classes of epithelia in imaging data we first trained a deep learning encoder with 3D images of nuclei of proximal tubule (PT) and thick ascending limb (TAL) generated by Volumetric Tissue Exploration and Analysis. In parallel, a side-information encoder was trained with transcriptional signatures of normal PT or TAL based on genes previously defined. Lastly, the two encoders were trained jointly to minimize the embedding distance between an image of a given label and its corresponding transcriptional signature for classification.
Results
Our outlier detection model was used to classify 3D nuclei of PT and TAL cells in biopsies from patients with diabetic kidney disease collected by the Kidney Precision Medicine Project. PT and TAL cells were either classified as normal or injured based on an outlier score. The putative injured PT and TAL were shown to positively correlate (p < 0.05 by Pearson) with areas of injury by neighborhood analysis in the original image volumes.
Conclusion
Using a deep learning approach that integrates 3D images of nuclei and transcriptomics we identify a putative signature of injury in biopsies of patients with diabetic kidney disease. Our approach has implications for unbiased assessment of the landscape of cellular injury based only on the morphology and textures of nuclei, leading to a better characterization of the severity of disease and its prognosis.
Funding
- NIDDK Support