Abstract
Bioinformatics has grown very quickly for the last 20 years, and it will grow even faster in the future. One of the long-standing open challenges in bioinformatics is biomarker identification and cancer diagnosis from gene expression. In this paper, the authors propose a novel hybrid wrapper/filter feature selection approach to identify the most informative genes for cancer diagnosis, named HWF-GS. It handles selection through two steps. The first one is an iterative filter-based mechanism to generate potential subsets of genes. The second step is the aggregation of the best-selected subsets by means of a wrapper-based consensus process that relies on a particle swarm optimization adapted to feature selection. An ensemble of classifiers (SVM and KNN) is employed to evaluate the selected genes. Experiments on nine publicly available cancer DNA microarray datasets have shown that HWF-GS selects robust signatures with high classification accuracy and competes with and even outperforms other methods in the literature.