Two-stage penalized algorithms via integrating prior information improve gene selection from omics data

With the rapid development of cancer biology, considerable cancer driver genes have been determined either experimentally or theoretically, which can be served as prior information. Increasingly accumulated omics data urgently requires efficient statistical learning algorithms to incorporate the prior information for further exploring various cancers. In this paper, four two-stage algorithms that integrate prior information are developed. The first stage of the algorithms integrates the prior information into representative response variables via principal component analysis (PCA), factor analysis or weighted group Lasso penalized logistic regression. In the second stage, penalized linear regression models with Lasso or elastic net are established. The performances of algorithms both in simulated data and 26 real-world cancer datasets are explored. One of the algorithms, which is called PCALasso, has its merits in terms of accuracy and robustness in gene selection. Comparing among eight algorithms, the PCALasso obtains moderately sparse results, correctly screens all desired variables from simulation data, and well identifies actually informative genes from various cancer datasets, which is a promising algorithm for gene selection from omics data.