Abstract
Current prognostic evaluation in melanoma primarily relies on traditional histopathological and clinical staging evaluation; however, these conventional approaches exhibit limited accuracy and fail to account for individual patient heterogeneity. To address these limitations, we developed a machine learning-driven prognostic signature, with the objectives of identifying pivotal biomarkers and establishing a precision medicine framework for prognostic assessment in melanoma management. Bulk RNA-seq data of 636 melanoma patients were obtained from TCGA and GEO databases, followed by univariate Cox regression to identify prognosis-associated genes. Intersecting results across cohorts identified consistently prognostic genes. Heterogeneity of the selected genes was assessed between primary and metastatic melanoma using scRNA-seq data. The consensus prognosis-related signature was developed by systematically integrating 101 machine learning algorithms, with model performance rigorously evaluated through multidimensional metrics. Finally, molecular experiments validated the prognostic relevance of the model's hub genes, and the biological role of CUL2 was investigated in melanoma. 53 protective prognosis-related genes (PRGs) were identified in melanoma. Single-cell analysis revealed elevated PRGs activity in primary melanoma tissues compared to metastatic lesions. A 14-gene consensus prognosis-related signature was developed using LASSO and RSF algorithms. The model achieved a C-index of 0.908 in the TCGA-SKCM cohort and a mean C-index of 0.758 across four independent validation cohorts. Furthermore, the model outperformed 19 existing prognostic models across multiple cohorts. This study developed a 14-gene consensus prognosis-related signature validated for robust prognostic performance across cohorts. CUL2, identified as a pivotal protective biomarker in melanoma, demonstrates potent tumour-suppressive activity through significant inhibition of proliferation and migration potential.