Feature screening for survival trait with application to TCGA high-dimensional genomic data

Background: In high-dimensional survival genomic data, identifying cancer-related genes is a challenging and important subject in the field of bioinformatics. In recent years, many feature screening approaches for survival outcomes with high-dimensional survival genomic data have been developed; however, few studies have systematically compared these methods. The primary purpose of this article is to conduct a series of simulation studies for systematic comparison; the second purpose of this article is to use these feature screening methods to further establish a more accurate prediction model for patient survival based on the survival genomic datasets of The Cancer Genome Atlas (TCGA).

Results: Simulation studies prove that network-adjusted feature screening measurement performs well and outperforms existing popular univariate independent feature screening methods. In the application of real data, we show that the proposed network-adjusted feature screening approach leads to more accurate survival prediction than alternative methods that do not account for gene-gene dependency information. We also use TCGA clinical survival genetic data to identify biomarkers associated with clinical survival outcomes in patients with various cancers including esophageal, pancreatic, head and neck squamous cell, lung, and breast invasive carcinomas.

Conclusions: These applications reveal advantages of the new proposed network-adjusted feature selection method over alternative methods that do not consider gene-gene dependency information. We also identify cancer-related genes that are almost detected in the literature. As a result, the network-based screening method is reliable and credible.