Computational Pathology of Proteinuric Diseases

Project Summary/Abstract: Computational Pathology for Proteinuric Glomerulopathies The presently employed morphology-based classification system of focal segmental glomerulosclerosis (FSGS) and minimal change disease (MCD) does not adequately capture the clinical and molecular heterogeneity of these diseases and impairs the ability of clinicians to precisely define a patient’s disease, or predict outcome or effective intervention. The goal of this research is to advance the work of the glomerular disease research community by identifying biologically-relevant surrogates and subclasses of FSGS/MCD using computational pathology and machine learning methods. Prospective, longitudinal, multi-dimensional data sets that include digital kidney biopsies and molecular and clinical information can be analyzed by advanced “computer vision” methods and machine learning analytical approaches. This project will employ these rich resources and methods, which offer an unprecedented opportunity to leverage information derived from kidney tissue to improve the diagnosis, outcome prediction, and identification of glomerular disease mechanisms. The interdisciplinary team assembled to conduct this study has vast experience and a long-standing history of collaboration. In our preliminary studies we have demonstrated that (i) structural changes associate with outcomes and molecular mechanisms and improve clinical outcome prediction beyond current clinical approaches, and that (ii) computer vision technology can be used to accurately and efficiently detect normal and abnormal kidney structures and quantify textural (i.e., the spatial relationship between pixel values) and morphological (i.e., shape, size) information from kidney tissue. We will test our central hypothesis that inherent in the complexity of the structural changes in the renal parenchyma is information predictive of underlying disease biology and clinical outcome. We will pursue this hypothesis (1) by testing the clinical and molecular relevance of automatic detection and quantification of known morphologic biomarkers of clinical outcomes and mechanisms and of groups of patients with similar DL-derived morphologic characteristics; (2) by extracting next-generation pathomic features from DL-derived morphologic parameters and by testing their associations—individually and combined into pathomic profiles—with clinical outcomes and gene expression; and (3) by building machine learning-based models that integrate computer vision-derived pathology data with gene expression and clinical data to predict individual patient clinical outcomes, and to assess the additive prediction value of each data domain. Finally, we will group patients using the biomarkers identified to be most predictive of clinical outcomes. Ultimately, our work will contribute to a foundation for the deployment of a comprehensive artificial intelligence-guided precision medicine program for FSGS/MCD, applying descriptive, predictive, and ultimately prescriptive analytics as support tools for practicing pathologists and nephrologists.