A Method for Cancer Genomics Feature Selection Based on LASSO-RFE

A more efficient feature selection method was developed to screen genes corresponding to specific cancers to further investigate their pathogenesis. The LASSO-RFE model, a last absolute shrinkage and selection operator (LASSO) classifier based on the idea of recursive feature elimination (RFE), was constructed. To verify the efficiency of the proposed algorithm, performance tests were conducted by using four kinds of gene expression RNA sequences publicly available in The Cancer Genome Atlas (TCGA). The numerical experiments were used to illustrate that the proposed LASSO-RFE enables a higher accuracy of the classification prediction model and a clearer biological interpretability of the selected gene features compared with three typical feature selection algorithms. The experimental results showed that LASSO-RFE effectively reduced tens of thousands of features in the original data to three dimensions and provided better performance for the classification model than mutual information, L1-SVM and tree-based selection method. This model retains the ability of the common LASSO algorithm to filter and remove redundant and irrelevant features, and enhances the biological interpretability according to RFE, which was compared with the traditional feature reduction methods. In this paper, only a limited number of data cases have been validated, and the application of LASSO-RFE with more recent data remains to be further investigated.