Machine learning (ML) model integrating cell death pathways as a prognostic and predictive biomarker for patients with melanoma.

Abstract

9561 Background: Established regulated non-apoptotic cell death mechanisms, including ferroptosis, necroptosis, and pyroptosis, can govern cancer cells' fate. Early reports suggest that these pathways can drive antitumor immune response, and may be associated with survival outcomes for pts treated with checkpoint blocker antibodies (ICB). We developed a ML model based on transcriptomic data (CDPS), integrating these 3 pathways, and interrogated its prognostic and predictive role in pts with melanoma. Methods: An ML score based on 148 non-redundant cell-death–related genes was developed and evaluated in patient-level data from TCGA-SKCM and validated in external cohorts of pts with melanoma (GSE65904, MEL-DFCI-2019, GSE98394, GSE54467). Prognostic performance was assessed using C-index, time-dependent ROC, Kaplan–Meier analysis, and univariable Cox models, with pts stratified by cohort-specific medians. Pathway activity was examined using gene set enrichment analysis (MSigDB Hallmark and Reactome gene sets; pathways were significant if FDR < 0.05). Immune infiltration was estimated with MCP-counter as per Cliff’delta (Mann-Whitney test p < 0.05). Predictive value was tested in pts treated with ICB (GSE78220, GSE91061, GSE168294), which also included on-treatment samples (GSE91061). Results: Across discovery and validation cohorts, a random survival forest model with ridge regression showed consistent prognostic performance (C-index range, 0.58-0.66). 7 genes emerged as dominant contributors during feature selection. The CDPS-high group showed a downregulation of MLKL and GSDMD (key mediators of necroptosis and pyroptosis, respectively), and an upregulation of SLC3A2 (a negative regulator of ferroptosis), suggesting suppression of these regulated cell-death pathways. Across all 5 cohorts, CDPS-high pts exhibited significantly worse overall survival (p < 0.05), corroborated by time-dependent ROC analyses at 1 (0.61 - 0.72), 3 (0.84 – 0.75), and 5 years (0.58-0.76). Functional pathway analyses revealed statistically significant positive enrichment of proliferative, anabolic, and DNA-repair signalling in the CDPS-high group and negative enrichment of immune-related pathways (FDR < 0.05). Similarly, CDPS-high tumors demonstrated broadly reduced innate and adaptive immune cell infiltration, indicating a less favorable tumor immune microenvironment (p < 0.05). CDPS was significantly lower in responders (R) vs non-responders ([NR], p = 0.04) in GSE168294. Notably, a decrease in CDPS levels was observed during treatment for R, whereas remained stable in NR (GSE91061). Conclusions: CDPS exhibited robust and consistent performance by combining selected programmed death pathways (ferroptosis, necroptosis, and pyroptosis), especially in pts treated with ICB. Future efforts will aim to validate CDPS in a prospective cohort of pts with melanoma.