Machine Learning-Driven Transcriptome Analysis of Keratoconus for Predictive Biomarker Identification

• Published in: Biomedicines Vol. 13; no. 5; p. 1032
• Main Authors: Chang, Shao-Hsuan, Yeh, Lung-Kun, Hung, Kuo-Hsuan, Chiu, Yen-Jung, Hsieh, Chia-Hsun, Ma, Chung-Pei
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 24.04.2025; MDPI
• Subjects: Accuracy; Artificial intelligence; Biomarkers; Biomechanics; Chemical elements; Collagen; Cornea; CXCR4 protein; Cytoskeleton; Datasets; Deep learning; Development and progression; Disease; Epidermis; Extracellular matrix; Gene expression; Gene set enrichment analysis; Genetic aspects; Genetic testing; Genetic transcription; Health aspects; Immune response; Inflammation; Keratoconus; Learning algorithms; Machine learning; Medical diagnosis; Metabolism; Metabolites; Pathogenesis; Pathophysiology; Regression analysis; Risk assessment; Statistical methods; Therapeutic targets; transcriptome; Transcriptomes; Taiwan; United States > US; keratoconus; machine learning; inflammation; biomarkers; transcriptome
• ISSN: 2227-9059; 2227-9059
• DOI: 10.3390/biomedicines13051032

Background: Keratoconus (KTCN) is a multifactorial disease characterized by progressive corneal degeneration. Recent studies suggest that a gene expression analysis of corneas may uncover potential novel biomarkers involved in corneal matrix remodeling. However, identifying reliable combinations of biomarkers that are linked to disease risk or progression remains a significant challenge. Objective: This study employed multiple machine learning algorithms to analyze the transcriptomes of keratoconus patients, identifying feature gene combinations and their functional associations, with the aim of enhancing the understanding of keratoconus pathogenesis. Methods: We analyzed the GSE77938 (PRJNA312169) dataset for differential gene expression (DGE) and performed gene set enrichment analysis (GSEA) using Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways to identify enriched pathways in keratoconus (KTCN) versus controls. Machine learning algorithms were then used to analyze the gene sets, with SHapley Additive exPlanations (SHAP) applied to assess the contribution of key feature genes in the model’s predictions. Selected feature genes were further analyzed through Gene Ontology (GO) enrichment to explore their roles in biological processes and cellular functions. Results: Machine learning models, including XGBoost, Random Forest, Logistic Regression, and SVM, identified a set of important feature genes associated with keratoconus, with 15 notable genes appearing across multiple models, such as IL1R1, JUN, CYBB, CXCR4, KRT13, KRT14, S100A8, S100A9, and others. The under-expressed genes in KTCN were involved in the mechanical resistance of the epidermis (KRT14, KRT15) and in inflammation pathways (S100A8/A9, IL1R1, CYBB, JUN, and CXCR4), as compared to controls. The GO analysis highlighted that the S100A8/A9 complex and its associated genes were primarily involved in biological processes related to the cytoskeleton organization, inflammation, and immune response. Furthermore, we expanded our analysis by incorporating additional datasets from PRJNA636666 and PRJNA1184491, thereby offering a broader representation of gene features and increasing the generalizability of our results across diverse cohorts. Conclusions: The differing gene sets identified by XGBoost and SVM may reflect distinct but complementary aspects of keratoconus pathophysiology. Meanwhile, XGBoost captured key immune and chemotactic regulators (e.g., IL1R1, CXCR4), suggesting upstream inflammatory signaling pathways. SVM highlighted structural and epithelial differentiation markers (e.g., KRT14, S100A8/A9), possibly reflecting downstream tissue remodeling and stress responses. Our findings provide a novel research platform for the evaluation of keratoconus using machine learning-based approaches, offering valuable insights into its pathogenesis and potential therapeutic targets.