Differentially expressed gene identification based on separability index

The identification of differentially expressed genes is central to microarray data analysis. Presented in this paper is an approach to differentially expressed gene identification based on a Separability Index (SI). Features are selected by identifying the optimal number of top ranking genes which result in maximum class separability. The approach was implemented on a training dataset comprising 400 samples from three types of cancers: colon, breast and lung cancer. The top 4222 genes resulted in a maximum separability of 91%. These genes were then used to classify a testing dataset comprising 250 samples, using a K-nearest neighbour (K-NN) classifier, achieving an accuracy of 92%. This outperformed a K-NN classifier trained on features selected based on p < 1:8311 × 10^-7 (Bonferroni corrected p-value cut-off criterion of p < 0:01), which achieved an accuracy of 89.6%. The performance is attributed to the non-arbitrary nature of the maximum SI selection criterion, which is an inherent property of the data, as opposed to the arbitrary assignment of a p-value cut-off. Hierarchical clustering was used to identify clusters of genes, amongst the 4222 genes, with similar expression patterns for each of the three cancers. These clusters were then examined for functional enrichment and significant biological pathways, which were identified for all three cancer types.