Machine learning-based screening of anoikis-related genes in melanoma diagnosis and their clinicopathological significance

Introduction: Cutaneous melanoma, a highly aggressive malignant tumor, continues to pose significant challenges in early diagnosis and prognosis assessment. Although anoikis-related genes (ARGs) play crucial roles in tumor progression, systematic molecular diagnostic biomarkers remain lacking. Objective: This study aims to elucidate the molecular mechanisms and diagnostic value of ARGs in cutaneous melanoma using advanced bioinformatics and machine learning approaches. Methods: Multiple melanoma datasets from The Cancer Genome Atlas and GEO databases (GSE3189, GSE15605, GSE19234, GSE65904, and GSE66839) were integrated to analyze ARG expression patterns. Differential expression analysis identified candidate genes significantly associated with cutaneous skin melanoma. A total of 113 machine learning models were employed to screen and validate candidate genes, with receiver operating characteristic curves evaluating diagnostic value. Core genes were identified through protein-protein interaction (PPI) networks. The immune infiltration correlations of core genes were analyzed, and their tissue-level expression was validated using immunohistochemistry. Results: Twenty-two ARGs were significantly enriched in cancer signaling pathways, including phosphoinositide 3-kinase-protein kinase B, hypoxia-inducible factor 1, and programmed death-ligand 1. Nine candidate genes were identified through differential and intersection analysis, and eight model genes were further screened using machine learning. Seven genes exhibited high diagnostic efficacy in both training and external validation sets (area under the curve [AUC] >0.7). Survival analysis revealed that high expression of carcinoembryonic antigen-related cell adhesion molecule (CEACAM)5, CEACAM6, epidermal growth factor receptor (EGFR), stratifin (SFN), and Polo-like kinase 1 (PLK1) correlated with poor prognosis. PPI analysis confirmed these as core regulatory factors, and immune infiltration analysis revealed strong associations with dendritic cells, T cells, and macrophages. The five-gene combination yielded excellent diagnostic performance, with AUCs of 0.969 (training) and 0.971 (validation). Immunohistochemistry confirmed their elevated expression in melanoma tissues. Conclusion: This study identified five key ARGs, CEACAM5, CEACAM6, EGFR, SFN, and PLK1, as significantly upregulated in cutaneous melanoma and associated with poor prognosis. Their combined diagnostic power suggests strong potential as clinical pathological biomarkers.