Integrative transcriptomic and machine learning analysis identifies CDH17 and HOXC13 as robust candidate prognostic biomarkers in uveal melanoma.

BACKGROUND: Uveal melanoma (UM), the most common form of ocular melanoma, represents poor prognosis, with approximately 50% of patients developing metastatic disease that significantly reduces survival rates. There is an crtitical need for accurate prognostic biomarkers to better stratify patients by risk and guide personalized treatment strategies.

METHODS: We employed an integrative, multi-step strategy to analyze UVM transcriptomic data, combining Cox proportional hazards survival analysis, stage-dependent differential expression, and a multi-layer perceptron (MLP) model to identify candidate prognostic genes. Receiver operating characteristic (ROC) analysis and Kaplan-Meier survival analysis were used to evaluate predictive performance, and external validation was performed using the GSE22138 cohort. Functional relevance was assessed through pan-cancer expression, DNA methylation, immune-infiltration, and protein expression analyses.

RESULTS: Integration of Cox regression, stage-dependent differential expression, and MLP modeling identified four robust prognostic biomarkers in UVM: CDH17, HOXC13, FABP5P7, and LINC02188. CDH17, a cadherin involved in cell-cell adhesion, promotes invasion and metastasis via integrin signaling and is oncogenic in gastrointestinal and melanoma tumors. HOXC13, a HOX transcription factor regulating proliferation, differentiation, and EMT, drives metastasis in skin melanoma and other cancers. FABP5P7, a pseudogene of the fatty acid-binding protein family, may act as a ceRNA influencing lipid metabolism and tumor progression. LINC02188, a long noncoding RNA, participates in ceRNA networks and has been linked to proliferation and migration in multiple cancers. Given the established oncogenic roles and protein-coding nature, CDH17 and HOXC13 were further prioritized for in-depth validation, including ROC analysis and pan-cancer characterization. ROC analysis demonstrated discriminative ability for patient outcomes (AUC = 0.76 for CDH17 and 0.65 for HOXC13), and metastasis-free survival trends were validated in the independent cohort. Pan-cancer analysis revealed elevated expression and frequent gene amplification in multiple tumor types, while immune-infiltration and methylation patterns suggested potential mechanistic roles. Both genes have established oncogenic functions in cutaneous melanoma, strengthening their biological plausibility in UM.

CONCLUSIONS: Our integrative approach identifies CDH17 and HOXC13 as biologically relevant, stage-associated prognostic biomarkers in UM. These findings provide a foundation for mechanistic studies and potential translational applications, including therapeutic targeting and risk-stratified patient management.