Predicting gene expression from pathology images in diabetic kidney disease
Vijaya Kolachalama   (Boston, MA)
Diabetic kidney disease (DKD) is currently staged by clinical manifestations such as microalbuminuria, overt proteinuria and chronic kidney disease (CKD) and end-stage kidney disease (ESKD). This conventional understanding is questioned by the striking heterogeneity in DKD. In some cases, DKD progresses rapidly towards ESKD and in others, the renal function remains stable despite suboptimal diabetic control. This complex landscape of DKD challenges the current paradigm of grouping patients based on albuminuria or eGFR, and begs the question to better characterize the patients with DKD based on the pathway alterations within their kidneys. Although recent explosion in genetic approaches have generated a wealth of information regarding pathway alterations in the single cells of kidneys, there is little translation of such findings to the clinic for direct management of DKD patients. We propose to develop a deep learning framework to associate histomorphologic images with deep -omics data from the same subject to further gain insight on the relationship between gene pathway derived alterations and the histologic features of DKD. Our hypothesis is grounded by the connection between dynamic genomic alterations and tissue-level changes, and we contend that our framework will reveal novel histologically-informed molecular targets associated with early and late changes in DKD severity. Using the publicly available -omics and whole slide imaging data from the Kidney Precision Medicine Project, we will build a deep neural network that can predict fold-changes in pathway alterations in specific cell types within the kidney (Aim 1). We will then build another deep neural network that can associate whole slide image-level features with clinical manifestations in DKD patients (Aim 2). Successful completion of this project will develop a computational framework that will predict pathway alterations from the histopathological features from kidney biopsies of DKD patients. Such information will allow clinical translation of genetic information to improve precision in the management of DKD patients. An interdisciplinary team of investigators with a track record of collaboration will pursue these tasks, and establish a proof-of-principle phenotype-genotype framework for DKD, which can allow us to apply for future NIH funding.