A transfer learning framework to elucidate the clinical relevance of altered proximal tubule cell states in kidney disease

Abstract

The application of single-cell technologies in clinical nephrology remains elusive. We generated an atlas of transcriptionally defined cell types and cell states of human kidney disease by integrating single-cell signatures reported in the literature with newly generated signatures obtained from 5 patients with acute kidney injury. We used this information to develop kidney-specific cell-level information ExtractoR (K-CLIER), a transfer learning approach specifically tailored to evaluate the role of cell types/states on bulk RNAseq data. We validated the K-CLIER as a reliable computational framework to obtain a dimensionality reduction and to link clinical data with single-cell signatures. By applying K-CLIER on cohorts of patients with different kidney diseases, we identified the most relevant cell types associated with fibrosis and disease progression. This analysis highlighted the central role of altered proximal tubule cells in chronic kidney disease. Our study introduces a new strategy to exploit the power of single-cell technologies toward clinical applications.

Keywords: Cell biology; Integrative aspects of cell biology; Machine learning; Transcriptomics.