Gene network analysis

To reduce outliers and simplify the calculation burden, we first retained 4677 prognostic-associated mRNAs selected by the univariate Cox proportional hazards regression (P < 0.05). A weighted correlation network analysis (WGCNA) was further performed on prognostic-associated mRNAs to construct scale-free gene co-expression networks, by using the R package “WGCNA”, with min-ModuleSize of 20, mergeCutHeight of 0.25, and unsigned topological overlap measure (TOM) [25]. An appropriate soft-threshold power was selected according to standard scale-free distribution (scale independence and mean connectivity). The module eigengenes (MEs) based on the first principal component were calculated for each module to measure the correlation between modules and clinical information including immunotype. The experimental protein-protein interactions (PPIs) data for genes within the immunotype-related module were retrieved from the STRING database, by specifying the interaction score of more than 0.4 (http://string-db.org/). The network was visualized within Cytoscape software (version 3.5.1.), and the hub genes were selected based on the Matthews correlation coefficient (MCC) score by using “cytoHubba” plug.