# Screening and Identification of Basement Membrane–Related Gene Signatures for Diagnosis in Keratoconus Through WGCNA and Machine Learning

## TL;DR

This study identifies four key genes linked to basement membranes that could help diagnose keratoconus, a condition causing vision loss, using bioinformatics and machine learning.

## Contribution

The study introduces novel basement membrane-related gene signatures for keratoconus diagnosis and provides insights into its pathogenesis.

## Key findings

- Four key genes (CRY2, RNF19B, PPP1R18, PFKFB3) were identified as potential biomarkers for keratoconus diagnosis.
- Three of the genes (CRY2, RNF19B, PPP1R18) showed strong predictive performance with AUC values over 0.6 in external validation.
- Immune infiltration and functional analyses suggest immune inflammation, metabolism, and apoptosis are involved in keratoconus.

## Abstract

Purpose: Keratoconus (KC) can lead to severe vision loss, impacting daily life. The etiology of KC is not yet clear, and early diagnosis and treatment are crucial for prognosis. This study aimed to explore basement membrane (BM)–related gene signatures for the diagnosis and therapy of KC and provide novel insights into its pathogenesis.

Methods: Based on the public datasets GSE112155 and GSE151631 in the GEO database, we obtained the differentially expressed genes (DEGs) of KC and downloaded BM-related genes based on the GeneCards database. Through a combination of bioinformatics methods, primarily weighted gene coexpression network analysis (WGCNA) and machine learning such as random forest (RF) and support vector machine (SVM), BM-related genes were identified as biomarkers for KC diagnosis. Subsequently, we further validated these findings using unsupervised clustering analysis, nomogram, and ROC curve analysis.

Results: Through the analysis of two KC-related datasets, 227 DEGs were screened out and intersected with BM-related genes to obtain 195 intersecting genes. By applying WGCNA and two machine learning algorithms, we identified four key genes, namely, CRY2, RNF19B, PPP1R18, and PFKFB3. These genes were significantly expressed in the normal control group. According to the ROC analysis, all four genes demonstrated excellent diagnostic performance in internal validation, with AUC values all exceeding 0.8. In external validation, CRY2, RNF19B, and PPP1R18 showed good predictive performance, each with AUC values greater than 0.6. Unsupervised clustering and nomogram also supported the good diagnostic capabilities of these genes. In addition, unsupervised clustering analysis also indicated that these four genes were mainly distributed in subtype A of KC. Immune infiltration analysis and functional enrichment analysis further suggested that immune inflammation, metabolism, and apoptosis were also involved in KC.

Conclusion: Using bioinformatics analysis, we found three novel hub genes, CRY2, RNF19B, and PPP1R18, which are beneficial for the diagnosis and therapy of KC.