Abstract
OBJECTIVE: Anoikis, a form of programmed cell death triggered by detachment from the extracellular matrix, plays a crucial role in metastasis and immune escape in melanoma. We aimed to identify anoikis-related prognostic markers using integrated machine learning and single-cell analysis. METHODS: We integrated single-cell RNA sequencing data from the GEO dataset GSE215120 and transcriptomic profiles from multiple melanoma cohorts, including TCGA, GSE19234, GSE22153, and GSE65904. Batch effects in single-cell data were corrected using the Harmony algorithm. Cell subpopulations were annotated via t-SNE dimensionality reduction and canonical markers, and AUCell was employed to compute the enrichment scores of anoikis-related genes across cell subtypes. A total of 150 anoikis-related genes were identified, and 101 machine learning algorithms and their combinations (including Cox regression, random survival forest, and gradient boosting machine) were systematically evaluated to identify the optimal prognostic model. Model performance was validated in independent cohorts using the concordance index (C-index), Kaplan-Meier survival analysis, and time-dependent ROC curves. Tumor microenvironment characteristics were assessed using ESTIMATE, CIBERSORT, and GSVA. The clinical relevance and functional role of SLC3A2 were further validated using the BEST database and in vitro experiments, including shRNA-mediated knockdown, colony formation, and Transwell migration assays. RESULTS: Single-cell analysis revealed significantly elevated anoikis scores in endothelial cells, fibroblasts, and melanocytes. High-scoring subpopulations exhibited more active cell-cell communication networks centered on endothelial cells. The "random survival forest + gradient boosting machine" model demonstrated optimal prognostic performance across the TCGA training cohort and validation cohorts (GSE19234, GSE22153, GSE65904), with a C-index of 0.774. Patients in the high-risk group had significantly shorter overall survival, and the model achieved strong predictive accuracy with AUCs ranging from 0.64 to 0.81 for 1-, 3-, and 5-year survival. Tumor microenvironment analysis indicated reduced immune infiltration (CD8⁺ T cells, B cells) in the high-risk group, suggestive of an immunosuppressive phenotype. SLC3A2 was highly expressed in melanoma and correlated with advanced T stage, drug resistance, and poor prognosis. Knockdown of SLC3A2 suppressed melanoma cell proliferation and migration in vitro. CONCLUSION: This study highlights the pivotal role of anoikis resistance in melanoma heterogeneity and immune microenvironment remodeling. The machine learning-based prognostic model we constructed holds clinical translational potential, and SLC3A2 was validated as a potential therapeutic target, offering new strategies for precision treatment of melanoma.