Integrated machine learning analysis of 30 cell death patterns identifies a novel prognostic signature in glioma

Minhao Huang¹Kai Zhao¹Yongtao Yang¹Kexin Mao²Hangyu Ma¹Tingting Wu¹Guolin Shi¹Wenhu Li¹Yan Li¹Ruiqi Peng²Ying Cheng²Ninghui Zhao^1*

• ¹The Second Affiliated Hospital of Kunming Medical University, Kunming, China
• ²Yunnan University, Kunming, China

Background: Glioma heterogeneity and therapeutic resistance are closely linked to dysregulated programmed cell death (PCD). While individual PCD pathways have been studied, the integrated network of multi-modal PCD interactions and their clinical implications in glioma remain poorly understood. This study aims to decipher the interplay between 30 distinct PCD modalities and the immune microenvironment, developing a robust prognostic signature to guide therapy. Results: Integrated analysis of 2,743 public gliomas samples identified 428 cell death-associated differentially expressed genes, enriched in neuroactive ligand-receptor interactions and extracellular matrix regulation. Unsupervised clustering revealed distinct immune-activated and immune-silent patient subtypes. A pan-death prognostic signature (Cell-Death Score, CDS), constructed via multi-algorithm machine learning and optimized using CoxBoost to incorporate 25 key genes, demonstrated robust performance in training (1-/3-year AUC=0.894/0.943) and validation cohort (C-index=0.717), effectively stratifying high-risk patients (HR=3.21, p<0.0001). High-CDS patients displayed elevated tumor mutational burden, homologous recombination deficiency, and immune checkpoint expression, alongside enhanced sensitivity to 11 therapeutic agents, including gemcitabine. Single-cell trajectory analysis confirmed significant activation of model genes during glioma progression. A clinical nomogram integrating CDS, WHO grade and radiotherapy further improved prognostic utility. Based on in vitro cell line experiments, the expression profiles of 25 key genes demonstrated significant heterogeneity, with partial genes undetectable by qRT-PCR due to expression levels falling below detection thresholds. Among seven genes consistently detected across all four cell lines, tumor cell lines exhibited significantly upregulated expression relative to normal astrocyte counterparts. RNA-seq analysis revealed effective detection of 24/25 key genes in seven paired tumor/adjacent tissue samples, with 20 genes showing higher mean expression in tumor tissues. qRT-PCR validation confirmed upregulated trends for 12 detectable genes in tumor tissues. Spatial transcriptomic analysis further corroborated tumor region-specific overexpression of all 25 key genes compared to adjacent non-tumorous areas. Conclusion: The CDS signature unravels the molecular interplay between glioma cell death heterogeneity, immune dysregulation, and therapeutic resistance. This biomarker system provides both prognostic and therapeutic insights for precision oncology, paving the way for personalized combination therapies in glioma management.