The L1/2 regularization network Cox model for analysis of genomic data

Methods based on a L1/2 penalty have been utilized to solve the variable selection problem associated with the Cox proportional hazards model. One limitation of the existing methods for survival analysis is that these ignore the regulatory networks and pathways information. To merge prior pathway information into the analysis of genomic data, we proposed a network-based regularization method for the L1/2 penalty and applied it to high-dimensional survival analysis data. This method used a L1/2 regularized solver and network that penalizes a Cox proportional hazards model with respect to the sparsity of the regression and the smoothness between the coefficients in a given network. Based on the limited simulation studies and real breast cancer gene expression datasets, the experimental results showed that our method achieves a higher predictive accuracy than previous methods. Even though fewer genes were selected compared to those using previous methods, results showed stronger associations with cancer. The results of the analysis were also validated using GeneCards.