Abstract
BACKGROUND: Dysregulation of circadian rhythms is implicated in cancer, but circadian genes impact on survival outcomes and response to immunotherapy remains unclear. This study aimed to systematically investigate these relationships and establish predictive models based on circadian gene expression. METHODS: We integrated transcriptomic and clinical data from 15 published immunotherapy cohorts (n = 768) and TCGA pan-cancer datasets. Circadian risk scores stratified patients for survival analysis. Machine learning models (incorporating 9 algorithms; optimized via 10-fold cross-validation with grid search) were developed using circadian signatures to predict immunotherapy response. Functional enrichment analysis (GO, KEGG, and Reactome) was subsequently performed to identify the underlying biological pathways. RESULTS: We identified 15 circadian genes significantly associated with survival (P < 0.05). The most hazardous gene was CSNK1D (HR = 1.18, 95% CI: 1.13-1.24), while KLF10 conferred strongest protection (HR = 0.93, 95% CI: 0.91-0.95). Remarkably, Kaplan-Meier survival curves revealed significant survival differences between high-risk and low-risk groups stratified by a circadian risk score. Notably, this effect was significant in PCPG (P = 0.03), LUAD (P = 0.02), and COAD (P = 0.01). Additionally, machine learning models, particularly Support Vector Machine (SVM; AUC = 0.913) and Random Forest (RF; AUC = 0.909), effectively predicted immunotherapy response. Feature importance analysis derived from these models highlighted several key circadian genes, including BHLHE40, KLF10, PER1, PER3, CSNK1D, and CSNK1E. Feature importance analysis derived from these models highlighted several key circadian genes, including BHLHE40, KLF10, PER1, PER3, CSNK1D, and CSNK1E. Enrichment analysis linked subgroup divergence to translational regulation, energy metabolism, and neurodegenerative pathways. CONCLUSION: Core circadian genes represent promising candidate biomarkers for both cancer survival risk stratification and immunotherapy response prediction. They further enable the development of high-accuracy machine learning models to predict immunotherapy responses.