Integrate multiple machine learning algorithms to establish a prognostic model of risk signature for programmed cell death and hypoxia and conduct multi-omics analysis of risk signatures in colorectal cancer

Abstract

Background Colorectal cancer (CRC) poses a significant challenge due to its high heterogeneity, making accurate prognosis prediction complex. Hypoxia plays a central role in influencing cell death mechanisms, tumorigenesis and progression. However, the prognostic significance of the interplay between hypoxia and cell death in CRC needs further investigation. Methods We employed a robust computational framework to explore the relationship between hypoxia and 18 cell death patterns in a global cohort of 1294 CRC patients from four multicenter cohorts. Thirteen commonly used machine learning algorithms were employed to develop optimal-performing hypoxia-associated programmed cell death risk signatures. Additionally, risk signature genes were screened at both the single-cell and spatial transcriptome levels using AddModuleScore analysis to enhance the identification of the signature. Results The risk signature, composed of 63 influential genes, conducted significant performance in predicting in CRC patients. The high-risk signature correlates significantly with pathways related to tumor occurrence, progression, and increased immune cells infiltration. The risk signature scores closely correlated with various immune cells proportions. At the single-cell level, high-risk epithelial cells were closely linked to pathways involving tumor occurrence, development, and drug resistance. High-risk epithelial cells exhibit enhanced communication with different cell types, acting as stronger Senders, Mediators, Receivers, and Influencers, promoting tumor progression. There are significant differences in copy number variation (CNV) and developmental trajectory between high-risk and low-risk epithelial cells. Moreover, we identified these risk signatures at the spatial transcriptome level, revealing their high expression throughout the entire tumor tissue. Conclusion Our robust machine learning framework highlights the prognostic potential of hypoxia-associated programmed cell death risk signatures in CRC. Integrating these signatures into prognosis prediction offers a unique opportunity for clinical intelligence and innovative management approaches.