Combining Single-Cell Data with Machine Learning Uncovers the Molecular Landscape of Keratoconus

Keratoconus (KC) is a progressive disorder characterized by corneal ectasia, corneal thinning, and conical protrusion, which can lead to diminished vision and corneal scarring. However, the molecular mechanism underlying KC has not yet been fully elucidated. In this study, we combined bioinformatics tools and experimental assays to uncover cell-type-specific pathogenic regulators and promising diagnostic biomarkers of KC.

We analyzed the KC-associated bulk mRNA expression dataset GSE151631 (19 KC, 7 controls) and the single-cell RNA sequencing dataset GSE285060 (5 KC, 3 controls) from the Gene Expression Omnibus (GEO) database. After quality control, dimensionality reduction, clustering, and annotation of single cells, we assessed each subpopulation's role in the disease and conducted differential expression analysis on the most relevant subpopulation. Feature selection was then performed on the bulk RNA-seq dataset using LASSO regression and SVM-RFE algorithms to identify candidate diagnostic genes. As a preliminary in vivo validation, an acute rabbit model of KC was induced by intrastromal injection of collagenase type II, and corneal tissues were harvested at day 7 to assess key gene expression at mRNA and protein levels by RT-qPCR and immunofluorescence, respectively. We also performed immune infiltration analysis, Gene Set Enrichment Analysis (GSEA), Gene Set Variation Analysis (GSVA), transcriptional regulatory networks construction, pseudotime trajectory analysis, and connectivity map (cMap)-based drug prediction.

Corneal stromal cells exhibited the highest contribution to the disease state. Three key genes (KIAA0930, CCDC85B, and KLHL21) were identified as potential diagnostic biomarkers. Notably, KIAA0930 was upregulated in the rabbit KC model, while CCDC85B and KLHL21 were downregulated, consistent with bioinformatic analysis. Immune infiltration analysis indicated increased cytolytic activity and CD8⁺ T cells in KC samples, with key genes showing significant correlations with various immune cell types and factors. GSEA and GSVA analyses revealed that these three genes are linked to ECM-receptor interaction, MAPK, PI3K-Akt, and WNT/β-catenin signaling pathways. Pseudotime analysis revealed distinct transcriptional states between KC and control cells. Several small-molecule compounds, including menadione and PAC-1, were computationally predicted to show negative connectivity with the KC-associated gene expression signature.

Our study identifies KIAA0930, CCDC85B, and KLHL21 as promising candidate biomarkers for KC diagnosis and highlights corneal stromal cells as a central contributor to KC pathological progression. By combining single-cell transcriptomics and machine learning with preliminary validation in an acute rabbit model, this work provides a molecular framework for early KC diagnosis and therapy, pending validation in chronic models and human cohorts.