Phenotype prediction is one of the central issues in genetics and medical sciences research. Due to the advent of high-throughput screening technologies, microarray-based cancer classification has become a standard procedure to identify cancer-related gene signatures. Since gene expression profiling in transcriptome is of high dimensionality, it is a challenging task to discover a biologically functional signature over different cell lines. In this article, we present an innovative framework for finding a small portion of discriminative genes for a specific disease phenotype classification by using information theory. The framework is a data-driven approach and considers feature relevance, redundancy, and interdependence in the context of feature pairs. Its effectiveness has been validated by using a brain cancer benchmark, where the gene expression profiling matrix is derived from Affymetrix Human Genome U95Av2 GeneChip ®. ?x00AE; . Three multivariate filters based on information theory have also been used for comparison. To show the strengths of the framework, three performance measures, two sets of enrichment analysis, and a stability index have been used in our experiments. The results show that the framework is robust and able to discover a gene signature having a high level of classification performance and being more statistically significant enriched.