Machine Learning-based Cell Death Signature for Accurate Prediction of Melanoma Prognosis and Clinical Response

Objective: Melanoma development, metastasis, and therapeutic resistance present substantial prognostic challenges, with current approaches lacking robust predictive models for patient outcomes and treatment guidance. Biomarkers for prognostic and drug sensitivity prediction were developed based on programmed cell death (PCD) mechanisms. Methods: Genes related to 19 PCD modalities were manually curated from multiple authoritative sources, including GSEA gene sets, KEGG pathways, relevant reviews, and published studies, to construct a comprehensive PCD signature gene set. Transcriptomic profiles and clinical data of skin cutaneous melanoma (SKCM) were obtained from the TCGA-SKCM cohort, and three independent treatment-naïve GEO cohorts (GSE65904, GSE19234, and GSE100797) were included for external validation. In addition, immunotherapy cohorts (PRJEB23709, GSE136961, and GSE215222) were collected to evaluate the predictive value for immunotherapy benefit. Single-cell RNA-seq datasets (GSE115978 and GSE215120) were processed using Seurat for quality control, normalization, dimensionality reduction, clustering, and cell-type annotation. Spatial transcriptomics data were obtained from the 10x Genomics website and analyzed using Cottrazm for spatial zonation and SpaCET for deconvolution. Cell–cell communication was inferred with CellChat by focusing on secreted signaling, ECM–receptor interactions, and contact-dependent signaling. Using TCGA-SKCM as the training set, a systematic machine-learning strategy comprising 10 algorithms and 101 combinations was implemented to build a PCD-related prognostic signature; the optimal model was selected based on the mean concordance index across multiple cohorts. Differential expression analysis, GSEA, CIBERSORT, and ESTIMATE were further applied to characterize immune infiltration. TCR clonal diversity, cytolytic activity, and the T-cell effector gene expression profile were calculated, and associations between the PCD-derived score and drug sensitivity were explored. Results: The activities of 19 PCD pathways differed significantly between normal skin and SKCM, suggesting that dysregulated PCD contributes to melanoma progression. Mutation analysis identified TTN and MUC16 as the most frequently mutated genes. After controlling collinearity and performing correlation-based filtering, 314 candidate genes were retained for modeling. Among 101 model combinations, the StepCox (backward) + Ridge model achieved the best performance (mean C-index = 0.677 across four cohorts). The resulting PCD score (PCDS) robustly stratified prognosis: patients with high PCDS consistently exhibited significantly worse survival in TCGA and the three external cohorts, and PCDS remained an independent prognostic factor after adjustment for age, sex, stage, Breslow thickness, Clark level, and ulceration. Immune analyses indicated that high PCDS was associated with an immunosuppressive phenotype, characterized by lower immune scores, reduced infiltration of effector immune cells (CD8+ T cells, activated memory CD4+ T cells, plasma cells, and M1 macrophages), and increased M2 macrophages. PCDS was significantly negatively correlated with multiple immune checkpoint molecules, TCR Shannon index, cytolytic activity (CYT), and the T-cell effector gene expression profile (T-GEP), and reached the highest level in the immune-desert phenotype. Spatial transcriptomics and single-cell analyses revealed spatial and cellular heterogeneity of PCDS: PCDS was lower in boundary regions and negatively correlated with macrophage proportions; myeloid cells exhibited the lowest overall PCDS, with macrophages showing the lowest scores and the strongest communication with T/NK cells. After immunotherapy, macrophages were enriched and the interaction probability of HLA-A/B/C–CD8A increased. Clinically, PCDS outperformed 100 previously published melanoma prognostic signatures and complemented the hot–cold tumor classification by identifying high-risk patients even among “hot” tumors. In immunotherapy cohorts, PCDS was negatively associated with treatment benefit; patients with high PCDS were more likely to be non-responders and had poorer survival. Drug-response analyses further indicated that PCDS correlated with sensitivities to multiple compounds, supporting its potential utility for treatment stratification. Overall, PCDS is a robust cross-cohort tool for predicting melanoma prognosis and immunotherapy benefit, reflecting an immunosuppressive tumor microenvironment and myeloid/macrophage-related immunoregulatory features, and may inform individualized risk stratification and potential therapeutic choices.Conclusion: A robust prognostic prediction model was established, and the regulatory mechanisms of programmed cell death within the tumor immune microenvironment were elucidated, providing an essential theoretical foundation for personalized treatment strategies in melanoma patients.