CDCSI: a machine learning-based interpretable cell death and cellular senescence index for prognosis improvement,...

Head and neck squamous cell carcinoma (HNSCC) was characterized by pronounced heterogeneity and therapeutic resistance, which posed challenges for effective clinical management. Regulated cell death (RCD) and cellular senescence (CS) were recognized as fundamental hallmarks of cancer, exerting effects on tumorigenesis, progression, and treatment sensitivity. However, the utility of cell death and cellular senescence (CDCS) in predicting prognosis and therapeutic response had not been fully established.

In this study, we integrated five types of multi-omics data with ten clustering algorithms to identify two distinct CDCS subtypes. Subsequently, subtype-associated marker genes and 270 machine learning algorithms were employed to construct a CDCS index (CDCSI). The CDCSI was then validated and compared across various datasets, including clinical features, external datasets, and previously published signatures. Finally, we evaluated the associations between CDCSI and various therapeutic strategies and validated its robustness at multiple levels.

First, we identified two distinct CDCS subtypes in HNSCC using multi-omics data and ten clustering algorithms. These subtypes exhibited different molecular characteristics, prognoses, and treatment responses, with CDCS2 showing broad resistance to radiotherapy, chemotherapy, targeted therapy, and immunotherapy. By analyzing marker genes associated with the CDCS2 subtype, we identified subtype-associated marker genes and integrated 270 machine learning algorithms to construct an individualized predictive model. The Ridge+SuperPC algorithm demonstrated the best performance, resulting in the development of CDCSI, which included six genes (SERPINE1, THBS1, SRPX, FADD, SERINC3, and CSNK2A2). We validated CDCSI across various methods, including clinical features, external datasets, and previously published models, confirming its superior predictive performance. Further analysis revealed that patients with low CDCSI were more likely to respond to radiotherapy, chemotherapy, EGFR-targeted therapy, and immunotherapy, whereas patients with high CDCSI exhibited stronger therapeutic resistance. The interpretability of CDCSI was enhanced through SHAP analysis, and its expression patterns were verified at single-cell and spatial transcriptomic levels, as well as through qPCR and WB for mRNA and protein expression.

In conclusion, through the integration of bioinformatics analyses and machine learning algorithms, we developed CDCSI as a robust and reliable signature for predicting prognosis, characterizing molecular features, and evaluating potential responses to radiotherapy, chemotherapy, targeted therapy, and immunotherapy in HNSCC. These findings suggested that CDCSI may provide an exploratory framework for individualized risk stratification and hypothesis generation regarding therapeutic response in HNSCC, but prospective clinical validation is required before clinical application.