A comprehensive analysis of prognosis prediction models based on pathway-level, gene-level and clinical information for glioblastoma

The present study aimed to develop a pathway-based prognosis prediction model for glioblastoma (GBM). Univariate and multivariate Cox regression analysis were used to identify prognosis-related genes and clinical factors using mRNA-seq data of GBM samples from The Cancer Genome Atlas (TCGA) database. The expression matrix of prognosis-related genes was transformed into pathway deregulation score (PDS) based on the Gene Set Enrichment Analysis (GSEA) repository using Pathifier software. With PDS scores as input, L1-penalized estimation-based Cox-proportional hazards (PH) model was used to identify prognostic pathways. Consequently, a prognosis prediction model based on these prognostic pathways was constructed for classifying patients in the TCGA set or each of the three validation sets into two risk groups. The survival difference between these risk groups was then analyzed using Kaplan-Meier survival analysis and log-rank test. In addition, a gene-based prognostic model was constructed using the Cox-PH model. The model of prognostic pathway combined with clinical factors was also evaluated. In total, 148 genes were discovered to be associated with prognosis. The Cox-PH model identified 13 prognostic pathways. Subsequently, a prognostic model based on the 13 pathways was constructed, and was demonstrated to successfully differentiate overall survival in the TCGA set and in three independent sets. However, the gene-based prognosis model was validated in only two of the three independent sets. Furthermore, the pathway+clinic factor-based model exhibited better predictive results compared with the pathway-based model. In conclusion, the present study suggests a promising prognosis prediction model of 13 pathways for GBM, which may be superior to the gene-level information-based prognostic model.