Single-cell-informed senescence programs underpin a machine-learning prognostic model robustly validated across melanoma cohorts.

Background: Cellular senescence (CS) shapes tumor evolution and the tumor microenvironment (TME), yet senescence-informed prognostic models for skin cutaneous melanoma (SKCM) remain limited. We aimed to develop and validate a robust senescence-related prognostic signature by integrating single-cell and bulk transcriptomic data with systematic machine-learning screening. Methods: Senescence activity was quantified in the scRNA-seq dataset GSE115978 using a CS-AUC score, followed by cell-type annotation and CS-AUC-associated gene prioritization. In TCGA-SKCM, weighted gene co-expression network analysis (WGCNA) identified senescence-related modules. Genes supported by both single-cell and bulk analyses were intersected to derive 100 candidate genes. Prognostic models were trained in TCGA and evaluated using C-index in six independent GEO cohorts (GSE19234, GSE22153, GSE53118, GSE54467, GSE59455, GSE65904) across 101 machine-learning strategies, selecting the best-performing algorithm. Survival, time-dependent ROC, and PCA were used for validation. Immune infiltration and TME associations were assessed by multi-algorithm deconvolution and ESTIMATE. Key model genes were further analyzed, and GPI was experimentally validated in A375 cells. Results: Single-cell analysis revealed marked heterogeneity of CS-AUC across cell types and identified CS-AUC-correlated genes. WGCNA in TCGA identified a senescence-associated red module, and intersection with the single-cell-derived senescence signals yielded 100 genes. Among 101 candidates, the GBM-based model achieved the best overall validation performance in GEO cohorts. The resulting riskScore significantly stratified overall survival across TCGA and all validation cohorts and showed stable predictive accuracy by time-dependent ROC. High-risk tumors exhibited an immune-depleted TME characterized by lower stromal/immune scores and higher tumor purity, along with altered cancer-immunity cycle activity. Within the GBM signature, GPI displayed the strongest positive correlation with riskScore, was upregulated in tumors, predicted worse survival, and was associated with metabolic/proliferative programs by GSEA. Functionally, GPI knockdown reduced A375 migration and clonogenic growth. Conclusion: We developed an integrative senescence-informed prognostic model for SKCM with strong external validation, immune/TME relevance, and experimental support. GPI emerges as a key risk-associated gene and potential therapeutic target within the senescence-related prognostic framework.