Machine learning-based cell death signature for predicting the prognosis and immunotherapy benefit in stomach adenocarcinoma.

Abstract

Stomach adenocarcinoma (STAD) is a one of most common malignancies with high mortality-to-incidence ratio. Programmed cell death (PCD) exerts vital functions in the progression of cancer. The role of PCD-related genes (PRGs) in STAD are not fully clarified. Using TCGA, GSE15459, GSE26253, GSE62254 and GSE84437 datasets, PCD-related signature (PRS) was constructed with an integrative procedure including 10 machine learning methods. The role of PRS in predicting the immunotherapy benefits was evaluated by several predicting score and 3 immunotherapy datasets (GSE91061, GSE78220, and IMvigor210). The model developed by Lasso + CoxBoost algorithm having a highest average C-index of 0.66 was considered as the optimal PRS. As an independent risk factor for STAD patients, PRS had a good performance in predicting the overall survival rate of patients, with an AUC of 1-, 3-, and 5-year ROC curve being 0.771, 0.751 and 0.827 in TCGA cohort. High PRS score demonstrated a lower gene set score of some immune-activated cells and immune-activated activities. Patient with high PRS score had a higher TIDE score, higher immune escape score, lower PD1&CTLA4 immunophenoscore, lower TMB score, lower response rate and poor prognosis, indicating a less immunotherapy response. The IC50 value of some drugs correlated with chemotherapy and targeted therapy was higher in high PRS score group. Our investigation developed an optimal PRS in STAD and it acted as an indicator for predicting the prognosis, stratifying risk and guiding treatment for STAD patients.