Massively Parallel Feature Selection Based on Ensemble of Filters and Multiple Robust Consensus Functions for Cancer Gene Identification

Currently, cancer prevails as a prime health matter worldwide. Selecting the appropriate biomarkers for early cancer detection might improve patient care and have often driven revolutions in medicine. Statistics and machine learning techniques have been broadly investigated for biomarker identification, especially feature selection where researchers try to identify the most distinguishing genes that can achieve better predictive performance of cancer subtypes. The robustness of the selected signature remains a crucial goal in personalized medicine. Ensemble and parallel feature selection are promising techniques to overcome this problem in which they have seen an increasing use in biomarker discovery. We focus in this chapter on the principal aspects of using ensemble feature selection in biomarker discovery. Furthermore, we propose a massively parallel meta-ensemble of filters (MPME-FS) to select a robust and parsimonious subset of genes. Two types of filters (ReliefF and Information Gain) are investigated in this study. The performances of the proposed approach in terms of robustness, classification power and the biological meaning of the selected signatures on five publicly available cancer datasets are explored. The results attest that the MPME-FS approach can effectively identify a small subset of biomarkers and improve both robustness and classification accuracy.