Machine learning reveals diverse cell death patterns in colorectal cancer prognosis and therapy

Abstract Background Colorectal cancer is a common gastrointestinal malignancy with a high mortality rate. Cancer cells must resist all forms of cell death to progress, and these death-related genes have been shown to have a fundamental impact on cancer development and metastasis. However, these cell death-related genes are still rarely explored in colorectal cancer. Methods Bulk transcriptomic and clinical data were obtained from The Cancer Genome Atlas (TCGA-COAD, n = 378) and the Gene Expression Omnibus (GSE38832, n = 122). Single-cell RNA-seq data (GSE132465) were used for cell-type-specific analysis. Cell-death-related genes were identified by integrating differentially expressed genes, prognosis-associated genes from univariate Cox regression, and hdWGCNA-derived modules. A cell-death-related signature (CDRS) was constructed and validated using 101 combinations of 10 machine-learning algorithms. The final model was obtained through Lasso regression and subsequent stepwise Cox proportional hazards regression, and it included 8 genes: AGAP3 , GTPBP4 , KLK6 , OSBPL1A , SCD , SEZ6L2 , SNTB1 , and STXBP1 . Performance was evaluated by Kaplan–Meier survival analysis, time-dependent ROC curves, and C-index in training, internal test, and external test cohorts. In vitro experiments were performed in DLD1 and HCT116 cell lines, with expression validated by qRT-PCR, Western blot, and immunohistochemistry on tissue microarrays and seven paired fresh tissues. Results We have identified 67 genes linked to the process of cell death. A computational model, designated as the cell-death-related signature (CDRS), was developed using machine learning techniques for the purpose of predicting the prognosis of colorectal cancer patients. The CDRS performed well in predicting patient outcomes and showed strong prognostic validity. We also observed differences in pathways enrichment, immune infiltration, drug sensitivity, and other factors between high risk and low risk groups based on CDRS scores. Additionally, we confirmed the expression of CDRS in colorectal cancer using qRT-PCR and found that GTPBP4, a gene with a high weight coefficient in CDRS, was markedly overexpressed in colorectal cancer tissues in comparison to paracancerous tissues. Knockdown of GTPBP4 led to reduced proliferation, migration, and invasion ability of colorectal cancer cells. Conclusion The CDRS signature we constructed is based on retrospective transcriptome data and shows excellent prognostic stratification ability for patients with colorectal cancer. It may serve as a stable prognostic stratification tool for CRC patients and can be used as a reference for individualized risk grouping studies. However, its potential value as a clinical diagnostic marker still needs to be further confirmed through large-scale prospective clinical validation.