A programmed cell death-related model based on machine learning for predicting prognosis and immunotherapy responses in patients with lung adenocarcinoma.

Abstract

lung adenocarcinoma (LUAD) remains one of the most common and lethal malignancies with poor prognosis. Programmed cell death (PCD) is an evolutionarily conserved cell suicide process that regulates tumorigenesis, progression, and metastasis of cancer cells. However, a comprehensive analysis of the role of PCD in LUAD is still unavailable. We analyzed multi-omic variations in PCD-related genes (PCDRGs) for LUAD. We used cross-validation of 10 machine learning algorithms (101 combinations) to synthetically develop and validate an optimal prognostic cell death score (CDS) model based on the PCDRGs expression profile. Patients were classified based on their median CDS values into the high and low-CDS groups. Next, we compared the differences in the genomics, biological functions, and tumor microenvironment of patients between both groups. In addition, we assessed the ability of CDS for predicting the response of patients from the immunotherapy cohort to immunotherapy. Finally, functional validation of key genes in CDS was performed. We constructed CDS based on four PCDRGs, which could effectively and consistently stratify patients with LUAD (patients with high CDS had poor prognoses). The performance of our CDS was superior compared to 77 LUAD signatures that have been previously published. The results revealed significant genetic alterations like mutation count, TMB, and CNV were observed in patients with high CDS. Furthermore, we observed an association of CDS with immune cell infiltration, microsatellite instability, SNV neoantigens. The immune status of patients with low CDS was more active. In addition, CDS could be reliable to predict therapeutic response in multiple immunotherapy cohorts. In vitro experiments revealed that high DNA damage inducible transcript 4 (DDIT4) expression in LUAD cells mediated protumor effects. CDS was constructed based on PCDRGs using machine learning. This model could accurately predict patients' prognoses and their responses to therapy. These results provide new promising tools for clinical management and aid in designing personalized treatment strategies for patients with LUAD.