[Machine learning-based programmed cell death signature model for precise prediction of prognosis and treatment response in melanoma].

OBJECTIVES: The occurrence, metastasis, and drug resistance of melanoma pose major challenges to patient prognosis, and predictive models capable of accurately forecasting patient outcomes and guiding treatment are still lacking. This study aims to develop predictive models for melanoma prognosis and drug sensitivity based on mechanisms of programmed cell death (PCD).

METHODS: Genes related to 19 PCD patterns were collected and integrated from gene set enrichment analysis (GSEA), the Kyoto Encyclopedia of Genes and Genomes (KEGG), relevant reviews, and published studies to establish a comprehensive PCD signature gene set. Transcriptomic and clinical data from The Cancer Genome Atlas (TCGA) skin cutaneous melanoma (SKCM) cohort were obtained, and 3 untreated Gene Expression Omnibus (GEO) datasets (GSE65904, GSE19234, and GSE100797) were included as external validation cohorts. In addition, immunotherapy cohorts PRJEB23709, GSE136961, and GSE215222 were collected to validate the predictive value for immunotherapy. Single-cell transcriptomic data (GSE115978 and GSE215120) were processed using Seurat for quality control, normalization, dimensionality reduction, clustering, and cell annotation; spatial transcriptomic data were obtained from 10× Genomics and combined with Cottrazm for spatial partitioning and SpaCET deconvolution. Cell-cell communication was evaluated using CellChat to assess secreted signaling, extracellular matrix (ECM)-receptor interactions, and cell contact-mediated communication patterns. Based on the TCGA training set, a machine learning strategy comprising 10 algorithms and 101 combinations was used to construct PCD-related prognostic signatures, and the optimal model was selected using the average concordance index (C-index) across multiple cohorts. Differential analysis, GSEA, CIBERSORT, and Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data (ESTIMATE) were further applied to evaluate immune infiltration, calculate T-cell receptor (TCR) clonal diversity, cytolytic activity, and T-cell effector gene expression profiles, and to explore the association between the PCD score (PCDS) and drug sensitivity.

RESULTS: The activities of the 19 PCD pathways differed significantly between normal skin and SKCM, suggesting that dysregulation is involved in melanoma progression. Mutation analysis showed that titin (TTN) and mucin 16 (MUC16) had the highest mutation frequencies. After controlling for collinearity and correlation screening, 314 candidate genes were obtained. Among the 101 model combinations, StepCox (backward)+Ridge performed best (average C-index across 4 cohorts=0.677). The PCDS constructed based on this model stably stratified prognosis: The high-PCDS group showed significantly worse survival in TCGA and 3 external validation cohorts and remained an independent prognostic indicator after adjustment for age, sex, stage, Breslow thickness, Clark level, and ulceration. Immune analyses indicated that high PCDS was associated with immunosuppression: Immune scores were reduced; CD8+ T cells, activated memory CD4+ T cells, plasma cells, and M1 macrophages were decreased, whereas M2 macrophages were increased. PCDS was significantly negatively correlated with immune checkpoint molecules, TCR Shannon index, cytolytic activity score (CYT), and T-cell effector gene expression profile (T-GEP), and was highest in the immune-desert subtype. Spatial transcriptomic and single-cell results suggested spatial and cellular heterogeneity of PCDS: PCDS was lower at boundary regions and negatively correlated with macrophage proportions; myeloid cells had the lowest overall PCDS, macrophages exhibited the strongest communication with T/natural killer (NK) cells, and after immunotherapy, macrophages were enriched with enhanced human leukocyte antigen (HLA)-A/B/C-CD8a-related interactions. In clinical applications, PCDS demonstrated greater robustness compared with 100 previously reported prognostic signatures and complemented hot-cold tumor classification by identifying high-risk patients even among "hot tumors". Validation in immunotherapy cohorts showed that PCDS was negatively associated with therapeutic benefit; high PCDS was more likely to be associated with non-response and poorer prognosis. Drug analyses suggested that PCDS was associated with sensitivity to multiple drugs and could be used for potential treatment stratification.

CONCLUSIONS: PCDS is a cross-cohort robust tool for predicting melanoma prognosis and immunotherapy benefit, reflecting the degree of immunosuppression in the tumor immune microenvironment and myeloid/macrophage-related immunoregulatory features, and provides a basis for individualized risk stratification and potential drug selection. This study provides an in-depth elucidation of the regulatory mechanisms of PCD in the tumor immune microenvironment and offers an important theoretical foundation for personalized treatment decision-making in melanoma patients.