FCM-SVM-RFE gene feature selection algorithm for leukemia classification from microarray gene expression data

Selecting the most possibly cancer-related genes from huge microarray gene expression data is an important bioinformatics research topic due to its significance to improve human's understandability of the inherent cancer-resulting mechanism. This is actually a feature selection problem. The huge number of genes makes it impossible to execute an exhaustive search. In this work, we propose a Recursive Feature Elimination (RFE) algorithm named FCM-SVM-RFE for the gene selection task. In each step, similar genes are grouped into clusters by the Fuzzy C-Means clustering algorithm, and then a Support Vector Machine (SVM) is modeled in each cluster-induced space, the genes which contribute large to the margin width of the SVM are selected to survive to the next step. This process is repeated until a pre-specified number of genes are selected. FCM-SVM-RFE is compared with SVM-RFE on AML/ALL microarray gene expression data. The experimental results show that FCM-SVM-RFE is more accurate than SVM-RFE to predict the unknown samples. More importantly, FCM-SVM-RFE can find some compact subsets of genes on each of which a SVM with perfect prediction accuracy can be modeled. These "most informative genes" are very helpful for biologists to efficiently and effectively find the inherent cancer-resulting mechanism.