Machine learning–based transcriptomic analysis of differentially expressed genes in the epithelium and stroma of keratoconus

This study was conducted to delineate layer-specific transcriptomic alterations in the keratoconus (KC) epithelium and stroma and to identify potential biomarkers using machine learning (ML)-based analysis. Two publicly available microarray datasets were integrated for epithelium analysis. For stroma analysis, RNA sequencing data from 31 KC patients and four normal donors were combined with a public microarray dataset. Differentially expressed genes (DEGs) were identified separately, after which functional enrichment, gene set enrichment, and gene set variation analyses were performed. Four ML algorithms were then applied, and feature importance was ranked using SHapley Additive exPlanations. The top 50% of genes across all the models were intersected, and hub genes were identified on the basis of protein–protein interaction (PPI) networks. The key findings were validated in two independent public datasets and by quantitative real-time polymerase chain reaction. A total of 272 and 5,124 DEGs were identified in the epithelium and stroma, respectively, with 53 DEGs downregulated in both. Enrichment analysis of these shared genes revealed dysregulation of retinoid, steroid hormone, and aldehyde metabolism. Layer-specific analyses revealed the downregulation of immune and inflammatory signalling, stress and redox responses, cellular proliferation, and tissue repair. ML and PPI network analysis were used to identify the epithelial hub genes CXCL10 and SLC1A1, which were both downregulated. In the stroma, 10 immune-related hub genes were identified, with the expression of IL6, IL1B, CXCL8, CXCL10, and ICAM1 being downregulated. In KC, inflammation, stress responses, and tissue repair occur in the epithelium and stroma, and metabolic and neural pathway dysregulation occurs, with more extensive stromal alterations than epithelial changes.